| File: | src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/CodeGen/SelectionDAGNodes.h |
| Warning: | line 1122, column 10 Called C++ object pointer is null |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
| 1 | //===-- X86ISelLowering.cpp - X86 DAG Lowering Implementation -------------===// | |||
| 2 | // | |||
| 3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. | |||
| 4 | // See https://llvm.org/LICENSE.txt for license information. | |||
| 5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception | |||
| 6 | // | |||
| 7 | //===----------------------------------------------------------------------===// | |||
| 8 | // | |||
| 9 | // This file defines the interfaces that X86 uses to lower LLVM code into a | |||
| 10 | // selection DAG. | |||
| 11 | // | |||
| 12 | //===----------------------------------------------------------------------===// | |||
| 13 | ||||
| 14 | #include "X86ISelLowering.h" | |||
| 15 | #include "MCTargetDesc/X86ShuffleDecode.h" | |||
| 16 | #include "X86.h" | |||
| 17 | #include "X86CallingConv.h" | |||
| 18 | #include "X86FrameLowering.h" | |||
| 19 | #include "X86InstrBuilder.h" | |||
| 20 | #include "X86IntrinsicsInfo.h" | |||
| 21 | #include "X86MachineFunctionInfo.h" | |||
| 22 | #include "X86TargetMachine.h" | |||
| 23 | #include "X86TargetObjectFile.h" | |||
| 24 | #include "llvm/ADT/SmallBitVector.h" | |||
| 25 | #include "llvm/ADT/SmallSet.h" | |||
| 26 | #include "llvm/ADT/Statistic.h" | |||
| 27 | #include "llvm/ADT/StringExtras.h" | |||
| 28 | #include "llvm/ADT/StringSwitch.h" | |||
| 29 | #include "llvm/Analysis/BlockFrequencyInfo.h" | |||
| 30 | #include "llvm/Analysis/EHPersonalities.h" | |||
| 31 | #include "llvm/Analysis/ObjCARCUtil.h" | |||
| 32 | #include "llvm/Analysis/ProfileSummaryInfo.h" | |||
| 33 | #include "llvm/Analysis/VectorUtils.h" | |||
| 34 | #include "llvm/CodeGen/IntrinsicLowering.h" | |||
| 35 | #include "llvm/CodeGen/MachineFrameInfo.h" | |||
| 36 | #include "llvm/CodeGen/MachineFunction.h" | |||
| 37 | #include "llvm/CodeGen/MachineInstrBuilder.h" | |||
| 38 | #include "llvm/CodeGen/MachineJumpTableInfo.h" | |||
| 39 | #include "llvm/CodeGen/MachineLoopInfo.h" | |||
| 40 | #include "llvm/CodeGen/MachineModuleInfo.h" | |||
| 41 | #include "llvm/CodeGen/MachineRegisterInfo.h" | |||
| 42 | #include "llvm/CodeGen/TargetLowering.h" | |||
| 43 | #include "llvm/CodeGen/WinEHFuncInfo.h" | |||
| 44 | #include "llvm/IR/CallingConv.h" | |||
| 45 | #include "llvm/IR/Constants.h" | |||
| 46 | #include "llvm/IR/DerivedTypes.h" | |||
| 47 | #include "llvm/IR/DiagnosticInfo.h" | |||
| 48 | #include "llvm/IR/Function.h" | |||
| 49 | #include "llvm/IR/GlobalAlias.h" | |||
| 50 | #include "llvm/IR/GlobalVariable.h" | |||
| 51 | #include "llvm/IR/Instructions.h" | |||
| 52 | #include "llvm/IR/Intrinsics.h" | |||
| 53 | #include "llvm/IR/IRBuilder.h" | |||
| 54 | #include "llvm/MC/MCAsmInfo.h" | |||
| 55 | #include "llvm/MC/MCContext.h" | |||
| 56 | #include "llvm/MC/MCExpr.h" | |||
| 57 | #include "llvm/MC/MCSymbol.h" | |||
| 58 | #include "llvm/Support/CommandLine.h" | |||
| 59 | #include "llvm/Support/Debug.h" | |||
| 60 | #include "llvm/Support/ErrorHandling.h" | |||
| 61 | #include "llvm/Support/KnownBits.h" | |||
| 62 | #include "llvm/Support/MathExtras.h" | |||
| 63 | #include "llvm/Target/TargetOptions.h" | |||
| 64 | #include <algorithm> | |||
| 65 | #include <bitset> | |||
| 66 | #include <cctype> | |||
| 67 | #include <numeric> | |||
| 68 | using namespace llvm; | |||
| 69 | ||||
| 70 | #define DEBUG_TYPE"x86-isel" "x86-isel" | |||
| 71 | ||||
| 72 | STATISTIC(NumTailCalls, "Number of tail calls")static llvm::Statistic NumTailCalls = {"x86-isel", "NumTailCalls" , "Number of tail calls"}; | |||
| 73 | ||||
| 74 | static cl::opt<int> ExperimentalPrefLoopAlignment( | |||
| 75 | "x86-experimental-pref-loop-alignment", cl::init(4), | |||
| 76 | cl::desc( | |||
| 77 | "Sets the preferable loop alignment for experiments (as log2 bytes)" | |||
| 78 | "(the last x86-experimental-pref-loop-alignment bits" | |||
| 79 | " of the loop header PC will be 0)."), | |||
| 80 | cl::Hidden); | |||
| 81 | ||||
| 82 | static cl::opt<int> ExperimentalPrefInnermostLoopAlignment( | |||
| 83 | "x86-experimental-pref-innermost-loop-alignment", cl::init(4), | |||
| 84 | cl::desc( | |||
| 85 | "Sets the preferable loop alignment for experiments (as log2 bytes) " | |||
| 86 | "for innermost loops only. If specified, this option overrides " | |||
| 87 | "alignment set by x86-experimental-pref-loop-alignment."), | |||
| 88 | cl::Hidden); | |||
| 89 | ||||
| 90 | static cl::opt<bool> MulConstantOptimization( | |||
| 91 | "mul-constant-optimization", cl::init(true), | |||
| 92 | cl::desc("Replace 'mul x, Const' with more effective instructions like " | |||
| 93 | "SHIFT, LEA, etc."), | |||
| 94 | cl::Hidden); | |||
| 95 | ||||
| 96 | static cl::opt<bool> ExperimentalUnorderedISEL( | |||
| 97 | "x86-experimental-unordered-atomic-isel", cl::init(false), | |||
| 98 | cl::desc("Use LoadSDNode and StoreSDNode instead of " | |||
| 99 | "AtomicSDNode for unordered atomic loads and " | |||
| 100 | "stores respectively."), | |||
| 101 | cl::Hidden); | |||
| 102 | ||||
| 103 | /// Call this when the user attempts to do something unsupported, like | |||
| 104 | /// returning a double without SSE2 enabled on x86_64. This is not fatal, unlike | |||
| 105 | /// report_fatal_error, so calling code should attempt to recover without | |||
| 106 | /// crashing. | |||
| 107 | static void errorUnsupported(SelectionDAG &DAG, const SDLoc &dl, | |||
| 108 | const char *Msg) { | |||
| 109 | MachineFunction &MF = DAG.getMachineFunction(); | |||
| 110 | DAG.getContext()->diagnose( | |||
| 111 | DiagnosticInfoUnsupported(MF.getFunction(), Msg, dl.getDebugLoc())); | |||
| 112 | } | |||
| 113 | ||||
| 114 | X86TargetLowering::X86TargetLowering(const X86TargetMachine &TM, | |||
| 115 | const X86Subtarget &STI) | |||
| 116 | : TargetLowering(TM), Subtarget(STI) { | |||
| 117 | bool UseX87 = !Subtarget.useSoftFloat() && Subtarget.hasX87(); | |||
| 118 | X86ScalarSSEf64 = Subtarget.hasSSE2(); | |||
| 119 | X86ScalarSSEf32 = Subtarget.hasSSE1(); | |||
| 120 | MVT PtrVT = MVT::getIntegerVT(TM.getPointerSizeInBits(0)); | |||
| 121 | ||||
| 122 | // Set up the TargetLowering object. | |||
| 123 | ||||
| 124 | // X86 is weird. It always uses i8 for shift amounts and setcc results. | |||
| 125 | setBooleanContents(ZeroOrOneBooleanContent); | |||
| 126 | // X86-SSE is even stranger. It uses -1 or 0 for vector masks. | |||
| 127 | setBooleanVectorContents(ZeroOrNegativeOneBooleanContent); | |||
| 128 | ||||
| 129 | // For 64-bit, since we have so many registers, use the ILP scheduler. | |||
| 130 | // For 32-bit, use the register pressure specific scheduling. | |||
| 131 | // For Atom, always use ILP scheduling. | |||
| 132 | if (Subtarget.isAtom()) | |||
| 133 | setSchedulingPreference(Sched::ILP); | |||
| 134 | else if (Subtarget.is64Bit()) | |||
| 135 | setSchedulingPreference(Sched::ILP); | |||
| 136 | else | |||
| 137 | setSchedulingPreference(Sched::RegPressure); | |||
| 138 | const X86RegisterInfo *RegInfo = Subtarget.getRegisterInfo(); | |||
| 139 | setStackPointerRegisterToSaveRestore(RegInfo->getStackRegister()); | |||
| 140 | ||||
| 141 | // Bypass expensive divides and use cheaper ones. | |||
| 142 | if (TM.getOptLevel() >= CodeGenOpt::Default) { | |||
| 143 | if (Subtarget.hasSlowDivide32()) | |||
| 144 | addBypassSlowDiv(32, 8); | |||
| 145 | if (Subtarget.hasSlowDivide64() && Subtarget.is64Bit()) | |||
| 146 | addBypassSlowDiv(64, 32); | |||
| 147 | } | |||
| 148 | ||||
| 149 | // Setup Windows compiler runtime calls. | |||
| 150 | if (Subtarget.isTargetWindowsMSVC() || Subtarget.isTargetWindowsItanium()) { | |||
| 151 | static const struct { | |||
| 152 | const RTLIB::Libcall Op; | |||
| 153 | const char * const Name; | |||
| 154 | const CallingConv::ID CC; | |||
| 155 | } LibraryCalls[] = { | |||
| 156 | { RTLIB::SDIV_I64, "_alldiv", CallingConv::X86_StdCall }, | |||
| 157 | { RTLIB::UDIV_I64, "_aulldiv", CallingConv::X86_StdCall }, | |||
| 158 | { RTLIB::SREM_I64, "_allrem", CallingConv::X86_StdCall }, | |||
| 159 | { RTLIB::UREM_I64, "_aullrem", CallingConv::X86_StdCall }, | |||
| 160 | { RTLIB::MUL_I64, "_allmul", CallingConv::X86_StdCall }, | |||
| 161 | }; | |||
| 162 | ||||
| 163 | for (const auto &LC : LibraryCalls) { | |||
| 164 | setLibcallName(LC.Op, LC.Name); | |||
| 165 | setLibcallCallingConv(LC.Op, LC.CC); | |||
| 166 | } | |||
| 167 | } | |||
| 168 | ||||
| 169 | if (Subtarget.getTargetTriple().isOSMSVCRT()) { | |||
| 170 | // MSVCRT doesn't have powi; fall back to pow | |||
| 171 | setLibcallName(RTLIB::POWI_F32, nullptr); | |||
| 172 | setLibcallName(RTLIB::POWI_F64, nullptr); | |||
| 173 | } | |||
| 174 | ||||
| 175 | // If we don't have cmpxchg8b(meaing this is a 386/486), limit atomic size to | |||
| 176 | // 32 bits so the AtomicExpandPass will expand it so we don't need cmpxchg8b. | |||
| 177 | // FIXME: Should we be limiting the atomic size on other configs? Default is | |||
| 178 | // 1024. | |||
| 179 | if (!Subtarget.hasCmpxchg8b()) | |||
| 180 | setMaxAtomicSizeInBitsSupported(32); | |||
| 181 | ||||
| 182 | // Set up the register classes. | |||
| 183 | addRegisterClass(MVT::i8, &X86::GR8RegClass); | |||
| 184 | addRegisterClass(MVT::i16, &X86::GR16RegClass); | |||
| 185 | addRegisterClass(MVT::i32, &X86::GR32RegClass); | |||
| 186 | if (Subtarget.is64Bit()) | |||
| 187 | addRegisterClass(MVT::i64, &X86::GR64RegClass); | |||
| 188 | ||||
| 189 | for (MVT VT : MVT::integer_valuetypes()) | |||
| 190 | setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Promote); | |||
| 191 | ||||
| 192 | // We don't accept any truncstore of integer registers. | |||
| 193 | setTruncStoreAction(MVT::i64, MVT::i32, Expand); | |||
| 194 | setTruncStoreAction(MVT::i64, MVT::i16, Expand); | |||
| 195 | setTruncStoreAction(MVT::i64, MVT::i8 , Expand); | |||
| 196 | setTruncStoreAction(MVT::i32, MVT::i16, Expand); | |||
| 197 | setTruncStoreAction(MVT::i32, MVT::i8 , Expand); | |||
| 198 | setTruncStoreAction(MVT::i16, MVT::i8, Expand); | |||
| 199 | ||||
| 200 | setTruncStoreAction(MVT::f64, MVT::f32, Expand); | |||
| 201 | ||||
| 202 | // SETOEQ and SETUNE require checking two conditions. | |||
| 203 | for (auto VT : {MVT::f32, MVT::f64, MVT::f80}) { | |||
| 204 | setCondCodeAction(ISD::SETOEQ, VT, Expand); | |||
| 205 | setCondCodeAction(ISD::SETUNE, VT, Expand); | |||
| 206 | } | |||
| 207 | ||||
| 208 | // Integer absolute. | |||
| 209 | if (Subtarget.hasCMov()) { | |||
| 210 | setOperationAction(ISD::ABS , MVT::i16 , Custom); | |||
| 211 | setOperationAction(ISD::ABS , MVT::i32 , Custom); | |||
| 212 | if (Subtarget.is64Bit()) | |||
| 213 | setOperationAction(ISD::ABS , MVT::i64 , Custom); | |||
| 214 | } | |||
| 215 | ||||
| 216 | // Funnel shifts. | |||
| 217 | for (auto ShiftOp : {ISD::FSHL, ISD::FSHR}) { | |||
| 218 | // For slow shld targets we only lower for code size. | |||
| 219 | LegalizeAction ShiftDoubleAction = Subtarget.isSHLDSlow() ? Custom : Legal; | |||
| 220 | ||||
| 221 | setOperationAction(ShiftOp , MVT::i8 , Custom); | |||
| 222 | setOperationAction(ShiftOp , MVT::i16 , Custom); | |||
| 223 | setOperationAction(ShiftOp , MVT::i32 , ShiftDoubleAction); | |||
| 224 | if (Subtarget.is64Bit()) | |||
| 225 | setOperationAction(ShiftOp , MVT::i64 , ShiftDoubleAction); | |||
| 226 | } | |||
| 227 | ||||
| 228 | if (!Subtarget.useSoftFloat()) { | |||
| 229 | // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this | |||
| 230 | // operation. | |||
| 231 | setOperationAction(ISD::UINT_TO_FP, MVT::i8, Promote); | |||
| 232 | setOperationAction(ISD::STRICT_UINT_TO_FP, MVT::i8, Promote); | |||
| 233 | setOperationAction(ISD::UINT_TO_FP, MVT::i16, Promote); | |||
| 234 | setOperationAction(ISD::STRICT_UINT_TO_FP, MVT::i16, Promote); | |||
| 235 | // We have an algorithm for SSE2, and we turn this into a 64-bit | |||
| 236 | // FILD or VCVTUSI2SS/SD for other targets. | |||
| 237 | setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom); | |||
| 238 | setOperationAction(ISD::STRICT_UINT_TO_FP, MVT::i32, Custom); | |||
| 239 | // We have an algorithm for SSE2->double, and we turn this into a | |||
| 240 | // 64-bit FILD followed by conditional FADD for other targets. | |||
| 241 | setOperationAction(ISD::UINT_TO_FP, MVT::i64, Custom); | |||
| 242 | setOperationAction(ISD::STRICT_UINT_TO_FP, MVT::i64, Custom); | |||
| 243 | ||||
| 244 | // Promote i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have | |||
| 245 | // this operation. | |||
| 246 | setOperationAction(ISD::SINT_TO_FP, MVT::i8, Promote); | |||
| 247 | setOperationAction(ISD::STRICT_SINT_TO_FP, MVT::i8, Promote); | |||
| 248 | // SSE has no i16 to fp conversion, only i32. We promote in the handler | |||
| 249 | // to allow f80 to use i16 and f64 to use i16 with sse1 only | |||
| 250 | setOperationAction(ISD::SINT_TO_FP, MVT::i16, Custom); | |||
| 251 | setOperationAction(ISD::STRICT_SINT_TO_FP, MVT::i16, Custom); | |||
| 252 | // f32 and f64 cases are Legal with SSE1/SSE2, f80 case is not | |||
| 253 | setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom); | |||
| 254 | setOperationAction(ISD::STRICT_SINT_TO_FP, MVT::i32, Custom); | |||
| 255 | // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64 | |||
| 256 | // are Legal, f80 is custom lowered. | |||
| 257 | setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom); | |||
| 258 | setOperationAction(ISD::STRICT_SINT_TO_FP, MVT::i64, Custom); | |||
| 259 | ||||
| 260 | // Promote i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have | |||
| 261 | // this operation. | |||
| 262 | setOperationAction(ISD::FP_TO_SINT, MVT::i8, Promote); | |||
| 263 | // FIXME: This doesn't generate invalid exception when it should. PR44019. | |||
| 264 | setOperationAction(ISD::STRICT_FP_TO_SINT, MVT::i8, Promote); | |||
| 265 | setOperationAction(ISD::FP_TO_SINT, MVT::i16, Custom); | |||
| 266 | setOperationAction(ISD::STRICT_FP_TO_SINT, MVT::i16, Custom); | |||
| 267 | setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom); | |||
| 268 | setOperationAction(ISD::STRICT_FP_TO_SINT, MVT::i32, Custom); | |||
| 269 | // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64 | |||
| 270 | // are Legal, f80 is custom lowered. | |||
| 271 | setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom); | |||
| 272 | setOperationAction(ISD::STRICT_FP_TO_SINT, MVT::i64, Custom); | |||
| 273 | ||||
| 274 | // Handle FP_TO_UINT by promoting the destination to a larger signed | |||
| 275 | // conversion. | |||
| 276 | setOperationAction(ISD::FP_TO_UINT, MVT::i8, Promote); | |||
| 277 | // FIXME: This doesn't generate invalid exception when it should. PR44019. | |||
| 278 | setOperationAction(ISD::STRICT_FP_TO_UINT, MVT::i8, Promote); | |||
| 279 | setOperationAction(ISD::FP_TO_UINT, MVT::i16, Promote); | |||
| 280 | // FIXME: This doesn't generate invalid exception when it should. PR44019. | |||
| 281 | setOperationAction(ISD::STRICT_FP_TO_UINT, MVT::i16, Promote); | |||
| 282 | setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom); | |||
| 283 | setOperationAction(ISD::STRICT_FP_TO_UINT, MVT::i32, Custom); | |||
| 284 | setOperationAction(ISD::FP_TO_UINT, MVT::i64, Custom); | |||
| 285 | setOperationAction(ISD::STRICT_FP_TO_UINT, MVT::i64, Custom); | |||
| 286 | ||||
| 287 | setOperationAction(ISD::LRINT, MVT::f32, Custom); | |||
| 288 | setOperationAction(ISD::LRINT, MVT::f64, Custom); | |||
| 289 | setOperationAction(ISD::LLRINT, MVT::f32, Custom); | |||
| 290 | setOperationAction(ISD::LLRINT, MVT::f64, Custom); | |||
| 291 | ||||
| 292 | if (!Subtarget.is64Bit()) { | |||
| 293 | setOperationAction(ISD::LRINT, MVT::i64, Custom); | |||
| 294 | setOperationAction(ISD::LLRINT, MVT::i64, Custom); | |||
| 295 | } | |||
| 296 | } | |||
| 297 | ||||
| 298 | if (Subtarget.hasSSE2()) { | |||
| 299 | // Custom lowering for saturating float to int conversions. | |||
| 300 | // We handle promotion to larger result types manually. | |||
| 301 | for (MVT VT : { MVT::i8, MVT::i16, MVT::i32 }) { | |||
| 302 | setOperationAction(ISD::FP_TO_UINT_SAT, VT, Custom); | |||
| 303 | setOperationAction(ISD::FP_TO_SINT_SAT, VT, Custom); | |||
| 304 | } | |||
| 305 | if (Subtarget.is64Bit()) { | |||
| 306 | setOperationAction(ISD::FP_TO_UINT_SAT, MVT::i64, Custom); | |||
| 307 | setOperationAction(ISD::FP_TO_SINT_SAT, MVT::i64, Custom); | |||
| 308 | } | |||
| 309 | } | |||
| 310 | ||||
| 311 | // Handle address space casts between mixed sized pointers. | |||
| 312 | setOperationAction(ISD::ADDRSPACECAST, MVT::i32, Custom); | |||
| 313 | setOperationAction(ISD::ADDRSPACECAST, MVT::i64, Custom); | |||
| 314 | ||||
| 315 | // TODO: when we have SSE, these could be more efficient, by using movd/movq. | |||
| 316 | if (!X86ScalarSSEf64) { | |||
| 317 | setOperationAction(ISD::BITCAST , MVT::f32 , Expand); | |||
| 318 | setOperationAction(ISD::BITCAST , MVT::i32 , Expand); | |||
| 319 | if (Subtarget.is64Bit()) { | |||
| 320 | setOperationAction(ISD::BITCAST , MVT::f64 , Expand); | |||
| 321 | // Without SSE, i64->f64 goes through memory. | |||
| 322 | setOperationAction(ISD::BITCAST , MVT::i64 , Expand); | |||
| 323 | } | |||
| 324 | } else if (!Subtarget.is64Bit()) | |||
| 325 | setOperationAction(ISD::BITCAST , MVT::i64 , Custom); | |||
| 326 | ||||
| 327 | // Scalar integer divide and remainder are lowered to use operations that | |||
| 328 | // produce two results, to match the available instructions. This exposes | |||
| 329 | // the two-result form to trivial CSE, which is able to combine x/y and x%y | |||
| 330 | // into a single instruction. | |||
| 331 | // | |||
| 332 | // Scalar integer multiply-high is also lowered to use two-result | |||
| 333 | // operations, to match the available instructions. However, plain multiply | |||
| 334 | // (low) operations are left as Legal, as there are single-result | |||
| 335 | // instructions for this in x86. Using the two-result multiply instructions | |||
| 336 | // when both high and low results are needed must be arranged by dagcombine. | |||
| 337 | for (auto VT : { MVT::i8, MVT::i16, MVT::i32, MVT::i64 }) { | |||
| 338 | setOperationAction(ISD::MULHS, VT, Expand); | |||
| 339 | setOperationAction(ISD::MULHU, VT, Expand); | |||
| 340 | setOperationAction(ISD::SDIV, VT, Expand); | |||
| 341 | setOperationAction(ISD::UDIV, VT, Expand); | |||
| 342 | setOperationAction(ISD::SREM, VT, Expand); | |||
| 343 | setOperationAction(ISD::UREM, VT, Expand); | |||
| 344 | } | |||
| 345 | ||||
| 346 | setOperationAction(ISD::BR_JT , MVT::Other, Expand); | |||
| 347 | setOperationAction(ISD::BRCOND , MVT::Other, Custom); | |||
| 348 | for (auto VT : { MVT::f32, MVT::f64, MVT::f80, MVT::f128, | |||
| 349 | MVT::i8, MVT::i16, MVT::i32, MVT::i64 }) { | |||
| 350 | setOperationAction(ISD::BR_CC, VT, Expand); | |||
| 351 | setOperationAction(ISD::SELECT_CC, VT, Expand); | |||
| 352 | } | |||
| 353 | if (Subtarget.is64Bit()) | |||
| 354 | setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal); | |||
| 355 | setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal); | |||
| 356 | setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal); | |||
| 357 | setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand); | |||
| 358 | ||||
| 359 | setOperationAction(ISD::FREM , MVT::f32 , Expand); | |||
| 360 | setOperationAction(ISD::FREM , MVT::f64 , Expand); | |||
| 361 | setOperationAction(ISD::FREM , MVT::f80 , Expand); | |||
| 362 | setOperationAction(ISD::FREM , MVT::f128 , Expand); | |||
| 363 | ||||
| 364 | if (!Subtarget.useSoftFloat() && Subtarget.hasX87()) { | |||
| 365 | setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom); | |||
| 366 | setOperationAction(ISD::SET_ROUNDING , MVT::Other, Custom); | |||
| 367 | } | |||
| 368 | ||||
| 369 | // Promote the i8 variants and force them on up to i32 which has a shorter | |||
| 370 | // encoding. | |||
| 371 | setOperationPromotedToType(ISD::CTTZ , MVT::i8 , MVT::i32); | |||
| 372 | setOperationPromotedToType(ISD::CTTZ_ZERO_UNDEF, MVT::i8 , MVT::i32); | |||
| 373 | ||||
| 374 | if (Subtarget.hasBMI()) { | |||
| 375 | // Promote the i16 zero undef variant and force it on up to i32 when tzcnt | |||
| 376 | // is enabled. | |||
| 377 | setOperationPromotedToType(ISD::CTTZ_ZERO_UNDEF, MVT::i16, MVT::i32); | |||
| 378 | } else { | |||
| 379 | setOperationAction(ISD::CTTZ, MVT::i16, Custom); | |||
| 380 | setOperationAction(ISD::CTTZ , MVT::i32 , Custom); | |||
| 381 | setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i16 , Legal); | |||
| 382 | setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32 , Legal); | |||
| 383 | if (Subtarget.is64Bit()) { | |||
| 384 | setOperationAction(ISD::CTTZ , MVT::i64 , Custom); | |||
| 385 | setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Legal); | |||
| 386 | } | |||
| 387 | } | |||
| 388 | ||||
| 389 | if (Subtarget.hasLZCNT()) { | |||
| 390 | // When promoting the i8 variants, force them to i32 for a shorter | |||
| 391 | // encoding. | |||
| 392 | setOperationPromotedToType(ISD::CTLZ , MVT::i8 , MVT::i32); | |||
| 393 | setOperationPromotedToType(ISD::CTLZ_ZERO_UNDEF, MVT::i8 , MVT::i32); | |||
| 394 | } else { | |||
| 395 | for (auto VT : {MVT::i8, MVT::i16, MVT::i32, MVT::i64}) { | |||
| 396 | if (VT == MVT::i64 && !Subtarget.is64Bit()) | |||
| 397 | continue; | |||
| 398 | setOperationAction(ISD::CTLZ , VT, Custom); | |||
| 399 | setOperationAction(ISD::CTLZ_ZERO_UNDEF, VT, Custom); | |||
| 400 | } | |||
| 401 | } | |||
| 402 | ||||
| 403 | for (auto Op : {ISD::FP16_TO_FP, ISD::STRICT_FP16_TO_FP, ISD::FP_TO_FP16, | |||
| 404 | ISD::STRICT_FP_TO_FP16}) { | |||
| 405 | // Special handling for half-precision floating point conversions. | |||
| 406 | // If we don't have F16C support, then lower half float conversions | |||
| 407 | // into library calls. | |||
| 408 | setOperationAction( | |||
| 409 | Op, MVT::f32, | |||
| 410 | (!Subtarget.useSoftFloat() && Subtarget.hasF16C()) ? Custom : Expand); | |||
| 411 | // There's never any support for operations beyond MVT::f32. | |||
| 412 | setOperationAction(Op, MVT::f64, Expand); | |||
| 413 | setOperationAction(Op, MVT::f80, Expand); | |||
| 414 | setOperationAction(Op, MVT::f128, Expand); | |||
| 415 | } | |||
| 416 | ||||
| 417 | setLoadExtAction(ISD::EXTLOAD, MVT::f32, MVT::f16, Expand); | |||
| 418 | setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f16, Expand); | |||
| 419 | setLoadExtAction(ISD::EXTLOAD, MVT::f80, MVT::f16, Expand); | |||
| 420 | setLoadExtAction(ISD::EXTLOAD, MVT::f128, MVT::f16, Expand); | |||
| 421 | setTruncStoreAction(MVT::f32, MVT::f16, Expand); | |||
| 422 | setTruncStoreAction(MVT::f64, MVT::f16, Expand); | |||
| 423 | setTruncStoreAction(MVT::f80, MVT::f16, Expand); | |||
| 424 | setTruncStoreAction(MVT::f128, MVT::f16, Expand); | |||
| 425 | ||||
| 426 | setOperationAction(ISD::PARITY, MVT::i8, Custom); | |||
| 427 | if (Subtarget.hasPOPCNT()) { | |||
| 428 | setOperationPromotedToType(ISD::CTPOP, MVT::i8, MVT::i32); | |||
| 429 | } else { | |||
| 430 | setOperationAction(ISD::CTPOP , MVT::i8 , Expand); | |||
| 431 | setOperationAction(ISD::CTPOP , MVT::i16 , Expand); | |||
| 432 | setOperationAction(ISD::CTPOP , MVT::i32 , Expand); | |||
| 433 | if (Subtarget.is64Bit()) | |||
| 434 | setOperationAction(ISD::CTPOP , MVT::i64 , Expand); | |||
| 435 | else | |||
| 436 | setOperationAction(ISD::CTPOP , MVT::i64 , Custom); | |||
| 437 | ||||
| 438 | setOperationAction(ISD::PARITY, MVT::i16, Custom); | |||
| 439 | setOperationAction(ISD::PARITY, MVT::i32, Custom); | |||
| 440 | if (Subtarget.is64Bit()) | |||
| 441 | setOperationAction(ISD::PARITY, MVT::i64, Custom); | |||
| 442 | } | |||
| 443 | ||||
| 444 | setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom); | |||
| 445 | ||||
| 446 | if (!Subtarget.hasMOVBE()) | |||
| 447 | setOperationAction(ISD::BSWAP , MVT::i16 , Expand); | |||
| 448 | ||||
| 449 | // X86 wants to expand cmov itself. | |||
| 450 | for (auto VT : { MVT::f32, MVT::f64, MVT::f80, MVT::f128 }) { | |||
| 451 | setOperationAction(ISD::SELECT, VT, Custom); | |||
| 452 | setOperationAction(ISD::SETCC, VT, Custom); | |||
| 453 | setOperationAction(ISD::STRICT_FSETCC, VT, Custom); | |||
| 454 | setOperationAction(ISD::STRICT_FSETCCS, VT, Custom); | |||
| 455 | } | |||
| 456 | for (auto VT : { MVT::i8, MVT::i16, MVT::i32, MVT::i64 }) { | |||
| 457 | if (VT == MVT::i64 && !Subtarget.is64Bit()) | |||
| 458 | continue; | |||
| 459 | setOperationAction(ISD::SELECT, VT, Custom); | |||
| 460 | setOperationAction(ISD::SETCC, VT, Custom); | |||
| 461 | } | |||
| 462 | ||||
| 463 | // Custom action for SELECT MMX and expand action for SELECT_CC MMX | |||
| 464 | setOperationAction(ISD::SELECT, MVT::x86mmx, Custom); | |||
| 465 | setOperationAction(ISD::SELECT_CC, MVT::x86mmx, Expand); | |||
| 466 | ||||
| 467 | setOperationAction(ISD::EH_RETURN , MVT::Other, Custom); | |||
| 468 | // NOTE: EH_SJLJ_SETJMP/_LONGJMP are not recommended, since | |||
| 469 | // LLVM/Clang supports zero-cost DWARF and SEH exception handling. | |||
| 470 | setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom); | |||
| 471 | setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom); | |||
| 472 | setOperationAction(ISD::EH_SJLJ_SETUP_DISPATCH, MVT::Other, Custom); | |||
| 473 | if (TM.Options.ExceptionModel == ExceptionHandling::SjLj) | |||
| 474 | setLibcallName(RTLIB::UNWIND_RESUME, "_Unwind_SjLj_Resume"); | |||
| 475 | ||||
| 476 | // Darwin ABI issue. | |||
| 477 | for (auto VT : { MVT::i32, MVT::i64 }) { | |||
| 478 | if (VT == MVT::i64 && !Subtarget.is64Bit()) | |||
| 479 | continue; | |||
| 480 | setOperationAction(ISD::ConstantPool , VT, Custom); | |||
| 481 | setOperationAction(ISD::JumpTable , VT, Custom); | |||
| 482 | setOperationAction(ISD::GlobalAddress , VT, Custom); | |||
| 483 | setOperationAction(ISD::GlobalTLSAddress, VT, Custom); | |||
| 484 | setOperationAction(ISD::ExternalSymbol , VT, Custom); | |||
| 485 | setOperationAction(ISD::BlockAddress , VT, Custom); | |||
| 486 | } | |||
| 487 | ||||
| 488 | // 64-bit shl, sra, srl (iff 32-bit x86) | |||
| 489 | for (auto VT : { MVT::i32, MVT::i64 }) { | |||
| 490 | if (VT == MVT::i64 && !Subtarget.is64Bit()) | |||
| 491 | continue; | |||
| 492 | setOperationAction(ISD::SHL_PARTS, VT, Custom); | |||
| 493 | setOperationAction(ISD::SRA_PARTS, VT, Custom); | |||
| 494 | setOperationAction(ISD::SRL_PARTS, VT, Custom); | |||
| 495 | } | |||
| 496 | ||||
| 497 | if (Subtarget.hasSSEPrefetch() || Subtarget.has3DNow()) | |||
| 498 | setOperationAction(ISD::PREFETCH , MVT::Other, Legal); | |||
| 499 | ||||
| 500 | setOperationAction(ISD::ATOMIC_FENCE , MVT::Other, Custom); | |||
| 501 | ||||
| 502 | // Expand certain atomics | |||
| 503 | for (auto VT : { MVT::i8, MVT::i16, MVT::i32, MVT::i64 }) { | |||
| 504 | setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, VT, Custom); | |||
| 505 | setOperationAction(ISD::ATOMIC_LOAD_SUB, VT, Custom); | |||
| 506 | setOperationAction(ISD::ATOMIC_LOAD_ADD, VT, Custom); | |||
| 507 | setOperationAction(ISD::ATOMIC_LOAD_OR, VT, Custom); | |||
| 508 | setOperationAction(ISD::ATOMIC_LOAD_XOR, VT, Custom); | |||
| 509 | setOperationAction(ISD::ATOMIC_LOAD_AND, VT, Custom); | |||
| 510 | setOperationAction(ISD::ATOMIC_STORE, VT, Custom); | |||
| 511 | } | |||
| 512 | ||||
| 513 | if (!Subtarget.is64Bit()) | |||
| 514 | setOperationAction(ISD::ATOMIC_LOAD, MVT::i64, Custom); | |||
| 515 | ||||
| 516 | if (Subtarget.hasCmpxchg16b()) { | |||
| 517 | setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, MVT::i128, Custom); | |||
| 518 | } | |||
| 519 | ||||
| 520 | // FIXME - use subtarget debug flags | |||
| 521 | if (!Subtarget.isTargetDarwin() && !Subtarget.isTargetELF() && | |||
| 522 | !Subtarget.isTargetCygMing() && !Subtarget.isTargetWin64() && | |||
| 523 | TM.Options.ExceptionModel != ExceptionHandling::SjLj) { | |||
| 524 | setOperationAction(ISD::EH_LABEL, MVT::Other, Expand); | |||
| 525 | } | |||
| 526 | ||||
| 527 | setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom); | |||
| 528 | setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom); | |||
| 529 | ||||
| 530 | setOperationAction(ISD::INIT_TRAMPOLINE, MVT::Other, Custom); | |||
| 531 | setOperationAction(ISD::ADJUST_TRAMPOLINE, MVT::Other, Custom); | |||
| 532 | ||||
| 533 | setOperationAction(ISD::TRAP, MVT::Other, Legal); | |||
| 534 | setOperationAction(ISD::DEBUGTRAP, MVT::Other, Legal); | |||
| 535 | setOperationAction(ISD::UBSANTRAP, MVT::Other, Legal); | |||
| 536 | ||||
| 537 | // VASTART needs to be custom lowered to use the VarArgsFrameIndex | |||
| 538 | setOperationAction(ISD::VASTART , MVT::Other, Custom); | |||
| 539 | setOperationAction(ISD::VAEND , MVT::Other, Expand); | |||
| 540 | bool Is64Bit = Subtarget.is64Bit(); | |||
| 541 | setOperationAction(ISD::VAARG, MVT::Other, Is64Bit ? Custom : Expand); | |||
| 542 | setOperationAction(ISD::VACOPY, MVT::Other, Is64Bit ? Custom : Expand); | |||
| 543 | ||||
| 544 | setOperationAction(ISD::STACKSAVE, MVT::Other, Expand); | |||
| 545 | setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand); | |||
| 546 | ||||
| 547 | setOperationAction(ISD::DYNAMIC_STACKALLOC, PtrVT, Custom); | |||
| 548 | ||||
| 549 | // GC_TRANSITION_START and GC_TRANSITION_END need custom lowering. | |||
| 550 | setOperationAction(ISD::GC_TRANSITION_START, MVT::Other, Custom); | |||
| 551 | setOperationAction(ISD::GC_TRANSITION_END, MVT::Other, Custom); | |||
| 552 | ||||
| 553 | if (!Subtarget.useSoftFloat() && X86ScalarSSEf64) { | |||
| 554 | // f32 and f64 use SSE. | |||
| 555 | // Set up the FP register classes. | |||
| 556 | addRegisterClass(MVT::f32, Subtarget.hasAVX512() ? &X86::FR32XRegClass | |||
| 557 | : &X86::FR32RegClass); | |||
| 558 | addRegisterClass(MVT::f64, Subtarget.hasAVX512() ? &X86::FR64XRegClass | |||
| 559 | : &X86::FR64RegClass); | |||
| 560 | ||||
| 561 | // Disable f32->f64 extload as we can only generate this in one instruction | |||
| 562 | // under optsize. So its easier to pattern match (fpext (load)) for that | |||
| 563 | // case instead of needing to emit 2 instructions for extload in the | |||
| 564 | // non-optsize case. | |||
| 565 | setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f32, Expand); | |||
| 566 | ||||
| 567 | for (auto VT : { MVT::f32, MVT::f64 }) { | |||
| 568 | // Use ANDPD to simulate FABS. | |||
| 569 | setOperationAction(ISD::FABS, VT, Custom); | |||
| 570 | ||||
| 571 | // Use XORP to simulate FNEG. | |||
| 572 | setOperationAction(ISD::FNEG, VT, Custom); | |||
| 573 | ||||
| 574 | // Use ANDPD and ORPD to simulate FCOPYSIGN. | |||
| 575 | setOperationAction(ISD::FCOPYSIGN, VT, Custom); | |||
| 576 | ||||
| 577 | // These might be better off as horizontal vector ops. | |||
| 578 | setOperationAction(ISD::FADD, VT, Custom); | |||
| 579 | setOperationAction(ISD::FSUB, VT, Custom); | |||
| 580 | ||||
| 581 | // We don't support sin/cos/fmod | |||
| 582 | setOperationAction(ISD::FSIN , VT, Expand); | |||
| 583 | setOperationAction(ISD::FCOS , VT, Expand); | |||
| 584 | setOperationAction(ISD::FSINCOS, VT, Expand); | |||
| 585 | } | |||
| 586 | ||||
| 587 | // Lower this to MOVMSK plus an AND. | |||
| 588 | setOperationAction(ISD::FGETSIGN, MVT::i64, Custom); | |||
| 589 | setOperationAction(ISD::FGETSIGN, MVT::i32, Custom); | |||
| 590 | ||||
| 591 | } else if (!Subtarget.useSoftFloat() && X86ScalarSSEf32 && | |||
| 592 | (UseX87 || Is64Bit)) { | |||
| 593 | // Use SSE for f32, x87 for f64. | |||
| 594 | // Set up the FP register classes. | |||
| 595 | addRegisterClass(MVT::f32, &X86::FR32RegClass); | |||
| 596 | if (UseX87) | |||
| 597 | addRegisterClass(MVT::f64, &X86::RFP64RegClass); | |||
| 598 | ||||
| 599 | // Use ANDPS to simulate FABS. | |||
| 600 | setOperationAction(ISD::FABS , MVT::f32, Custom); | |||
| 601 | ||||
| 602 | // Use XORP to simulate FNEG. | |||
| 603 | setOperationAction(ISD::FNEG , MVT::f32, Custom); | |||
| 604 | ||||
| 605 | if (UseX87) | |||
| 606 | setOperationAction(ISD::UNDEF, MVT::f64, Expand); | |||
| 607 | ||||
| 608 | // Use ANDPS and ORPS to simulate FCOPYSIGN. | |||
| 609 | if (UseX87) | |||
| 610 | setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand); | |||
| 611 | setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom); | |||
| 612 | ||||
| 613 | // We don't support sin/cos/fmod | |||
| 614 | setOperationAction(ISD::FSIN , MVT::f32, Expand); | |||
| 615 | setOperationAction(ISD::FCOS , MVT::f32, Expand); | |||
| 616 | setOperationAction(ISD::FSINCOS, MVT::f32, Expand); | |||
| 617 | ||||
| 618 | if (UseX87) { | |||
| 619 | // Always expand sin/cos functions even though x87 has an instruction. | |||
| 620 | setOperationAction(ISD::FSIN, MVT::f64, Expand); | |||
| 621 | setOperationAction(ISD::FCOS, MVT::f64, Expand); | |||
| 622 | setOperationAction(ISD::FSINCOS, MVT::f64, Expand); | |||
| 623 | } | |||
| 624 | } else if (UseX87) { | |||
| 625 | // f32 and f64 in x87. | |||
| 626 | // Set up the FP register classes. | |||
| 627 | addRegisterClass(MVT::f64, &X86::RFP64RegClass); | |||
| 628 | addRegisterClass(MVT::f32, &X86::RFP32RegClass); | |||
| 629 | ||||
| 630 | for (auto VT : { MVT::f32, MVT::f64 }) { | |||
| 631 | setOperationAction(ISD::UNDEF, VT, Expand); | |||
| 632 | setOperationAction(ISD::FCOPYSIGN, VT, Expand); | |||
| 633 | ||||
| 634 | // Always expand sin/cos functions even though x87 has an instruction. | |||
| 635 | setOperationAction(ISD::FSIN , VT, Expand); | |||
| 636 | setOperationAction(ISD::FCOS , VT, Expand); | |||
| 637 | setOperationAction(ISD::FSINCOS, VT, Expand); | |||
| 638 | } | |||
| 639 | } | |||
| 640 | ||||
| 641 | // Expand FP32 immediates into loads from the stack, save special cases. | |||
| 642 | if (isTypeLegal(MVT::f32)) { | |||
| 643 | if (UseX87 && (getRegClassFor(MVT::f32) == &X86::RFP32RegClass)) { | |||
| 644 | addLegalFPImmediate(APFloat(+0.0f)); // FLD0 | |||
| 645 | addLegalFPImmediate(APFloat(+1.0f)); // FLD1 | |||
| 646 | addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS | |||
| 647 | addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS | |||
| 648 | } else // SSE immediates. | |||
| 649 | addLegalFPImmediate(APFloat(+0.0f)); // xorps | |||
| 650 | } | |||
| 651 | // Expand FP64 immediates into loads from the stack, save special cases. | |||
| 652 | if (isTypeLegal(MVT::f64)) { | |||
| 653 | if (UseX87 && getRegClassFor(MVT::f64) == &X86::RFP64RegClass) { | |||
| 654 | addLegalFPImmediate(APFloat(+0.0)); // FLD0 | |||
| 655 | addLegalFPImmediate(APFloat(+1.0)); // FLD1 | |||
| 656 | addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS | |||
| 657 | addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS | |||
| 658 | } else // SSE immediates. | |||
| 659 | addLegalFPImmediate(APFloat(+0.0)); // xorpd | |||
| 660 | } | |||
| 661 | // Handle constrained floating-point operations of scalar. | |||
| 662 | setOperationAction(ISD::STRICT_FADD, MVT::f32, Legal); | |||
| 663 | setOperationAction(ISD::STRICT_FADD, MVT::f64, Legal); | |||
| 664 | setOperationAction(ISD::STRICT_FSUB, MVT::f32, Legal); | |||
| 665 | setOperationAction(ISD::STRICT_FSUB, MVT::f64, Legal); | |||
| 666 | setOperationAction(ISD::STRICT_FMUL, MVT::f32, Legal); | |||
| 667 | setOperationAction(ISD::STRICT_FMUL, MVT::f64, Legal); | |||
| 668 | setOperationAction(ISD::STRICT_FDIV, MVT::f32, Legal); | |||
| 669 | setOperationAction(ISD::STRICT_FDIV, MVT::f64, Legal); | |||
| 670 | setOperationAction(ISD::STRICT_FP_EXTEND, MVT::f64, Legal); | |||
| 671 | setOperationAction(ISD::STRICT_FP_ROUND, MVT::f32, Legal); | |||
| 672 | setOperationAction(ISD::STRICT_FP_ROUND, MVT::f64, Legal); | |||
| 673 | setOperationAction(ISD::STRICT_FSQRT, MVT::f32, Legal); | |||
| 674 | setOperationAction(ISD::STRICT_FSQRT, MVT::f64, Legal); | |||
| 675 | ||||
| 676 | // We don't support FMA. | |||
| 677 | setOperationAction(ISD::FMA, MVT::f64, Expand); | |||
| 678 | setOperationAction(ISD::FMA, MVT::f32, Expand); | |||
| 679 | ||||
| 680 | // f80 always uses X87. | |||
| 681 | if (UseX87) { | |||
| 682 | addRegisterClass(MVT::f80, &X86::RFP80RegClass); | |||
| 683 | setOperationAction(ISD::UNDEF, MVT::f80, Expand); | |||
| 684 | setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand); | |||
| 685 | { | |||
| 686 | APFloat TmpFlt = APFloat::getZero(APFloat::x87DoubleExtended()); | |||
| 687 | addLegalFPImmediate(TmpFlt); // FLD0 | |||
| 688 | TmpFlt.changeSign(); | |||
| 689 | addLegalFPImmediate(TmpFlt); // FLD0/FCHS | |||
| 690 | ||||
| 691 | bool ignored; | |||
| 692 | APFloat TmpFlt2(+1.0); | |||
| 693 | TmpFlt2.convert(APFloat::x87DoubleExtended(), APFloat::rmNearestTiesToEven, | |||
| 694 | &ignored); | |||
| 695 | addLegalFPImmediate(TmpFlt2); // FLD1 | |||
| 696 | TmpFlt2.changeSign(); | |||
| 697 | addLegalFPImmediate(TmpFlt2); // FLD1/FCHS | |||
| 698 | } | |||
| 699 | ||||
| 700 | // Always expand sin/cos functions even though x87 has an instruction. | |||
| 701 | setOperationAction(ISD::FSIN , MVT::f80, Expand); | |||
| 702 | setOperationAction(ISD::FCOS , MVT::f80, Expand); | |||
| 703 | setOperationAction(ISD::FSINCOS, MVT::f80, Expand); | |||
| 704 | ||||
| 705 | setOperationAction(ISD::FFLOOR, MVT::f80, Expand); | |||
| 706 | setOperationAction(ISD::FCEIL, MVT::f80, Expand); | |||
| 707 | setOperationAction(ISD::FTRUNC, MVT::f80, Expand); | |||
| 708 | setOperationAction(ISD::FRINT, MVT::f80, Expand); | |||
| 709 | setOperationAction(ISD::FNEARBYINT, MVT::f80, Expand); | |||
| 710 | setOperationAction(ISD::FMA, MVT::f80, Expand); | |||
| 711 | setOperationAction(ISD::LROUND, MVT::f80, Expand); | |||
| 712 | setOperationAction(ISD::LLROUND, MVT::f80, Expand); | |||
| 713 | setOperationAction(ISD::LRINT, MVT::f80, Custom); | |||
| 714 | setOperationAction(ISD::LLRINT, MVT::f80, Custom); | |||
| 715 | ||||
| 716 | // Handle constrained floating-point operations of scalar. | |||
| 717 | setOperationAction(ISD::STRICT_FADD , MVT::f80, Legal); | |||
| 718 | setOperationAction(ISD::STRICT_FSUB , MVT::f80, Legal); | |||
| 719 | setOperationAction(ISD::STRICT_FMUL , MVT::f80, Legal); | |||
| 720 | setOperationAction(ISD::STRICT_FDIV , MVT::f80, Legal); | |||
| 721 | setOperationAction(ISD::STRICT_FSQRT , MVT::f80, Legal); | |||
| 722 | setOperationAction(ISD::STRICT_FP_EXTEND, MVT::f80, Legal); | |||
| 723 | // FIXME: When the target is 64-bit, STRICT_FP_ROUND will be overwritten | |||
| 724 | // as Custom. | |||
| 725 | setOperationAction(ISD::STRICT_FP_ROUND, MVT::f80, Legal); | |||
| 726 | } | |||
| 727 | ||||
| 728 | // f128 uses xmm registers, but most operations require libcalls. | |||
| 729 | if (!Subtarget.useSoftFloat() && Subtarget.is64Bit() && Subtarget.hasSSE1()) { | |||
| 730 | addRegisterClass(MVT::f128, Subtarget.hasVLX() ? &X86::VR128XRegClass | |||
| 731 | : &X86::VR128RegClass); | |||
| 732 | ||||
| 733 | addLegalFPImmediate(APFloat::getZero(APFloat::IEEEquad())); // xorps | |||
| 734 | ||||
| 735 | setOperationAction(ISD::FADD, MVT::f128, LibCall); | |||
| 736 | setOperationAction(ISD::STRICT_FADD, MVT::f128, LibCall); | |||
| 737 | setOperationAction(ISD::FSUB, MVT::f128, LibCall); | |||
| 738 | setOperationAction(ISD::STRICT_FSUB, MVT::f128, LibCall); | |||
| 739 | setOperationAction(ISD::FDIV, MVT::f128, LibCall); | |||
| 740 | setOperationAction(ISD::STRICT_FDIV, MVT::f128, LibCall); | |||
| 741 | setOperationAction(ISD::FMUL, MVT::f128, LibCall); | |||
| 742 | setOperationAction(ISD::STRICT_FMUL, MVT::f128, LibCall); | |||
| 743 | setOperationAction(ISD::FMA, MVT::f128, LibCall); | |||
| 744 | setOperationAction(ISD::STRICT_FMA, MVT::f128, LibCall); | |||
| 745 | ||||
| 746 | setOperationAction(ISD::FABS, MVT::f128, Custom); | |||
| 747 | setOperationAction(ISD::FNEG, MVT::f128, Custom); | |||
| 748 | setOperationAction(ISD::FCOPYSIGN, MVT::f128, Custom); | |||
| 749 | ||||
| 750 | setOperationAction(ISD::FSIN, MVT::f128, LibCall); | |||
| 751 | setOperationAction(ISD::STRICT_FSIN, MVT::f128, LibCall); | |||
| 752 | setOperationAction(ISD::FCOS, MVT::f128, LibCall); | |||
| 753 | setOperationAction(ISD::STRICT_FCOS, MVT::f128, LibCall); | |||
| 754 | setOperationAction(ISD::FSINCOS, MVT::f128, LibCall); | |||
| 755 | // No STRICT_FSINCOS | |||
| 756 | setOperationAction(ISD::FSQRT, MVT::f128, LibCall); | |||
| 757 | setOperationAction(ISD::STRICT_FSQRT, MVT::f128, LibCall); | |||
| 758 | ||||
| 759 | setOperationAction(ISD::FP_EXTEND, MVT::f128, Custom); | |||
| 760 | setOperationAction(ISD::STRICT_FP_EXTEND, MVT::f128, Custom); | |||
| 761 | // We need to custom handle any FP_ROUND with an f128 input, but | |||
| 762 | // LegalizeDAG uses the result type to know when to run a custom handler. | |||
| 763 | // So we have to list all legal floating point result types here. | |||
| 764 | if (isTypeLegal(MVT::f32)) { | |||
| 765 | setOperationAction(ISD::FP_ROUND, MVT::f32, Custom); | |||
| 766 | setOperationAction(ISD::STRICT_FP_ROUND, MVT::f32, Custom); | |||
| 767 | } | |||
| 768 | if (isTypeLegal(MVT::f64)) { | |||
| 769 | setOperationAction(ISD::FP_ROUND, MVT::f64, Custom); | |||
| 770 | setOperationAction(ISD::STRICT_FP_ROUND, MVT::f64, Custom); | |||
| 771 | } | |||
| 772 | if (isTypeLegal(MVT::f80)) { | |||
| 773 | setOperationAction(ISD::FP_ROUND, MVT::f80, Custom); | |||
| 774 | setOperationAction(ISD::STRICT_FP_ROUND, MVT::f80, Custom); | |||
| 775 | } | |||
| 776 | ||||
| 777 | setOperationAction(ISD::SETCC, MVT::f128, Custom); | |||
| 778 | ||||
| 779 | setLoadExtAction(ISD::EXTLOAD, MVT::f128, MVT::f32, Expand); | |||
| 780 | setLoadExtAction(ISD::EXTLOAD, MVT::f128, MVT::f64, Expand); | |||
| 781 | setLoadExtAction(ISD::EXTLOAD, MVT::f128, MVT::f80, Expand); | |||
| 782 | setTruncStoreAction(MVT::f128, MVT::f32, Expand); | |||
| 783 | setTruncStoreAction(MVT::f128, MVT::f64, Expand); | |||
| 784 | setTruncStoreAction(MVT::f128, MVT::f80, Expand); | |||
| 785 | } | |||
| 786 | ||||
| 787 | // Always use a library call for pow. | |||
| 788 | setOperationAction(ISD::FPOW , MVT::f32 , Expand); | |||
| 789 | setOperationAction(ISD::FPOW , MVT::f64 , Expand); | |||
| 790 | setOperationAction(ISD::FPOW , MVT::f80 , Expand); | |||
| 791 | setOperationAction(ISD::FPOW , MVT::f128 , Expand); | |||
| 792 | ||||
| 793 | setOperationAction(ISD::FLOG, MVT::f80, Expand); | |||
| 794 | setOperationAction(ISD::FLOG2, MVT::f80, Expand); | |||
| 795 | setOperationAction(ISD::FLOG10, MVT::f80, Expand); | |||
| 796 | setOperationAction(ISD::FEXP, MVT::f80, Expand); | |||
| 797 | setOperationAction(ISD::FEXP2, MVT::f80, Expand); | |||
| 798 | setOperationAction(ISD::FMINNUM, MVT::f80, Expand); | |||
| 799 | setOperationAction(ISD::FMAXNUM, MVT::f80, Expand); | |||
| 800 | ||||
| 801 | // Some FP actions are always expanded for vector types. | |||
| 802 | for (auto VT : { MVT::v4f32, MVT::v8f32, MVT::v16f32, | |||
| 803 | MVT::v2f64, MVT::v4f64, MVT::v8f64 }) { | |||
| 804 | setOperationAction(ISD::FSIN, VT, Expand); | |||
| 805 | setOperationAction(ISD::FSINCOS, VT, Expand); | |||
| 806 | setOperationAction(ISD::FCOS, VT, Expand); | |||
| 807 | setOperationAction(ISD::FREM, VT, Expand); | |||
| 808 | setOperationAction(ISD::FCOPYSIGN, VT, Expand); | |||
| 809 | setOperationAction(ISD::FPOW, VT, Expand); | |||
| 810 | setOperationAction(ISD::FLOG, VT, Expand); | |||
| 811 | setOperationAction(ISD::FLOG2, VT, Expand); | |||
| 812 | setOperationAction(ISD::FLOG10, VT, Expand); | |||
| 813 | setOperationAction(ISD::FEXP, VT, Expand); | |||
| 814 | setOperationAction(ISD::FEXP2, VT, Expand); | |||
| 815 | } | |||
| 816 | ||||
| 817 | // First set operation action for all vector types to either promote | |||
| 818 | // (for widening) or expand (for scalarization). Then we will selectively | |||
| 819 | // turn on ones that can be effectively codegen'd. | |||
| 820 | for (MVT VT : MVT::fixedlen_vector_valuetypes()) { | |||
| 821 | setOperationAction(ISD::SDIV, VT, Expand); | |||
| 822 | setOperationAction(ISD::UDIV, VT, Expand); | |||
| 823 | setOperationAction(ISD::SREM, VT, Expand); | |||
| 824 | setOperationAction(ISD::UREM, VT, Expand); | |||
| 825 | setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT,Expand); | |||
| 826 | setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Expand); | |||
| 827 | setOperationAction(ISD::EXTRACT_SUBVECTOR, VT,Expand); | |||
| 828 | setOperationAction(ISD::INSERT_SUBVECTOR, VT,Expand); | |||
| 829 | setOperationAction(ISD::FMA, VT, Expand); | |||
| 830 | setOperationAction(ISD::FFLOOR, VT, Expand); | |||
| 831 | setOperationAction(ISD::FCEIL, VT, Expand); | |||
| 832 | setOperationAction(ISD::FTRUNC, VT, Expand); | |||
| 833 | setOperationAction(ISD::FRINT, VT, Expand); | |||
| 834 | setOperationAction(ISD::FNEARBYINT, VT, Expand); | |||
| 835 | setOperationAction(ISD::SMUL_LOHI, VT, Expand); | |||
| 836 | setOperationAction(ISD::MULHS, VT, Expand); | |||
| 837 | setOperationAction(ISD::UMUL_LOHI, VT, Expand); | |||
| 838 | setOperationAction(ISD::MULHU, VT, Expand); | |||
| 839 | setOperationAction(ISD::SDIVREM, VT, Expand); | |||
| 840 | setOperationAction(ISD::UDIVREM, VT, Expand); | |||
| 841 | setOperationAction(ISD::CTPOP, VT, Expand); | |||
| 842 | setOperationAction(ISD::CTTZ, VT, Expand); | |||
| 843 | setOperationAction(ISD::CTLZ, VT, Expand); | |||
| 844 | setOperationAction(ISD::ROTL, VT, Expand); | |||
| 845 | setOperationAction(ISD::ROTR, VT, Expand); | |||
| 846 | setOperationAction(ISD::BSWAP, VT, Expand); | |||
| 847 | setOperationAction(ISD::SETCC, VT, Expand); | |||
| 848 | setOperationAction(ISD::FP_TO_UINT, VT, Expand); | |||
| 849 | setOperationAction(ISD::FP_TO_SINT, VT, Expand); | |||
| 850 | setOperationAction(ISD::UINT_TO_FP, VT, Expand); | |||
| 851 | setOperationAction(ISD::SINT_TO_FP, VT, Expand); | |||
| 852 | setOperationAction(ISD::SIGN_EXTEND_INREG, VT,Expand); | |||
| 853 | setOperationAction(ISD::TRUNCATE, VT, Expand); | |||
| 854 | setOperationAction(ISD::SIGN_EXTEND, VT, Expand); | |||
| 855 | setOperationAction(ISD::ZERO_EXTEND, VT, Expand); | |||
| 856 | setOperationAction(ISD::ANY_EXTEND, VT, Expand); | |||
| 857 | setOperationAction(ISD::SELECT_CC, VT, Expand); | |||
| 858 | for (MVT InnerVT : MVT::fixedlen_vector_valuetypes()) { | |||
| 859 | setTruncStoreAction(InnerVT, VT, Expand); | |||
| 860 | ||||
| 861 | setLoadExtAction(ISD::SEXTLOAD, InnerVT, VT, Expand); | |||
| 862 | setLoadExtAction(ISD::ZEXTLOAD, InnerVT, VT, Expand); | |||
| 863 | ||||
| 864 | // N.b. ISD::EXTLOAD legality is basically ignored except for i1-like | |||
| 865 | // types, we have to deal with them whether we ask for Expansion or not. | |||
| 866 | // Setting Expand causes its own optimisation problems though, so leave | |||
| 867 | // them legal. | |||
| 868 | if (VT.getVectorElementType() == MVT::i1) | |||
| 869 | setLoadExtAction(ISD::EXTLOAD, InnerVT, VT, Expand); | |||
| 870 | ||||
| 871 | // EXTLOAD for MVT::f16 vectors is not legal because f16 vectors are | |||
| 872 | // split/scalarized right now. | |||
| 873 | if (VT.getVectorElementType() == MVT::f16) | |||
| 874 | setLoadExtAction(ISD::EXTLOAD, InnerVT, VT, Expand); | |||
| 875 | } | |||
| 876 | } | |||
| 877 | ||||
| 878 | // FIXME: In order to prevent SSE instructions being expanded to MMX ones | |||
| 879 | // with -msoft-float, disable use of MMX as well. | |||
| 880 | if (!Subtarget.useSoftFloat() && Subtarget.hasMMX()) { | |||
| 881 | addRegisterClass(MVT::x86mmx, &X86::VR64RegClass); | |||
| 882 | // No operations on x86mmx supported, everything uses intrinsics. | |||
| 883 | } | |||
| 884 | ||||
| 885 | if (!Subtarget.useSoftFloat() && Subtarget.hasSSE1()) { | |||
| 886 | addRegisterClass(MVT::v4f32, Subtarget.hasVLX() ? &X86::VR128XRegClass | |||
| 887 | : &X86::VR128RegClass); | |||
| 888 | ||||
| 889 | setOperationAction(ISD::FNEG, MVT::v4f32, Custom); | |||
| 890 | setOperationAction(ISD::FABS, MVT::v4f32, Custom); | |||
| 891 | setOperationAction(ISD::FCOPYSIGN, MVT::v4f32, Custom); | |||
| 892 | setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom); | |||
| 893 | setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom); | |||
| 894 | setOperationAction(ISD::VSELECT, MVT::v4f32, Custom); | |||
| 895 | setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom); | |||
| 896 | setOperationAction(ISD::SELECT, MVT::v4f32, Custom); | |||
| 897 | ||||
| 898 | setOperationAction(ISD::LOAD, MVT::v2f32, Custom); | |||
| 899 | setOperationAction(ISD::STORE, MVT::v2f32, Custom); | |||
| 900 | ||||
| 901 | setOperationAction(ISD::STRICT_FADD, MVT::v4f32, Legal); | |||
| 902 | setOperationAction(ISD::STRICT_FSUB, MVT::v4f32, Legal); | |||
| 903 | setOperationAction(ISD::STRICT_FMUL, MVT::v4f32, Legal); | |||
| 904 | setOperationAction(ISD::STRICT_FDIV, MVT::v4f32, Legal); | |||
| 905 | setOperationAction(ISD::STRICT_FSQRT, MVT::v4f32, Legal); | |||
| 906 | } | |||
| 907 | ||||
| 908 | if (!Subtarget.useSoftFloat() && Subtarget.hasSSE2()) { | |||
| 909 | addRegisterClass(MVT::v2f64, Subtarget.hasVLX() ? &X86::VR128XRegClass | |||
| 910 | : &X86::VR128RegClass); | |||
| 911 | ||||
| 912 | // FIXME: Unfortunately, -soft-float and -no-implicit-float mean XMM | |||
| 913 | // registers cannot be used even for integer operations. | |||
| 914 | addRegisterClass(MVT::v16i8, Subtarget.hasVLX() ? &X86::VR128XRegClass | |||
| 915 | : &X86::VR128RegClass); | |||
| 916 | addRegisterClass(MVT::v8i16, Subtarget.hasVLX() ? &X86::VR128XRegClass | |||
| 917 | : &X86::VR128RegClass); | |||
| 918 | addRegisterClass(MVT::v4i32, Subtarget.hasVLX() ? &X86::VR128XRegClass | |||
| 919 | : &X86::VR128RegClass); | |||
| 920 | addRegisterClass(MVT::v2i64, Subtarget.hasVLX() ? &X86::VR128XRegClass | |||
| 921 | : &X86::VR128RegClass); | |||
| 922 | ||||
| 923 | for (auto VT : { MVT::v2i8, MVT::v4i8, MVT::v8i8, | |||
| 924 | MVT::v2i16, MVT::v4i16, MVT::v2i32 }) { | |||
| 925 | setOperationAction(ISD::SDIV, VT, Custom); | |||
| 926 | setOperationAction(ISD::SREM, VT, Custom); | |||
| 927 | setOperationAction(ISD::UDIV, VT, Custom); | |||
| 928 | setOperationAction(ISD::UREM, VT, Custom); | |||
| 929 | } | |||
| 930 | ||||
| 931 | setOperationAction(ISD::MUL, MVT::v2i8, Custom); | |||
| 932 | setOperationAction(ISD::MUL, MVT::v4i8, Custom); | |||
| 933 | setOperationAction(ISD::MUL, MVT::v8i8, Custom); | |||
| 934 | ||||
| 935 | setOperationAction(ISD::MUL, MVT::v16i8, Custom); | |||
| 936 | setOperationAction(ISD::MUL, MVT::v4i32, Custom); | |||
| 937 | setOperationAction(ISD::MUL, MVT::v2i64, Custom); | |||
| 938 | setOperationAction(ISD::MULHU, MVT::v4i32, Custom); | |||
| 939 | setOperationAction(ISD::MULHS, MVT::v4i32, Custom); | |||
| 940 | setOperationAction(ISD::MULHU, MVT::v16i8, Custom); | |||
| 941 | setOperationAction(ISD::MULHS, MVT::v16i8, Custom); | |||
| 942 | setOperationAction(ISD::MULHU, MVT::v8i16, Legal); | |||
| 943 | setOperationAction(ISD::MULHS, MVT::v8i16, Legal); | |||
| 944 | setOperationAction(ISD::MUL, MVT::v8i16, Legal); | |||
| 945 | ||||
| 946 | setOperationAction(ISD::SMULO, MVT::v16i8, Custom); | |||
| 947 | setOperationAction(ISD::UMULO, MVT::v16i8, Custom); | |||
| 948 | ||||
| 949 | setOperationAction(ISD::FNEG, MVT::v2f64, Custom); | |||
| 950 | setOperationAction(ISD::FABS, MVT::v2f64, Custom); | |||
| 951 | setOperationAction(ISD::FCOPYSIGN, MVT::v2f64, Custom); | |||
| 952 | ||||
| 953 | for (auto VT : { MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v2i64 }) { | |||
| 954 | setOperationAction(ISD::SMAX, VT, VT == MVT::v8i16 ? Legal : Custom); | |||
| 955 | setOperationAction(ISD::SMIN, VT, VT == MVT::v8i16 ? Legal : Custom); | |||
| 956 | setOperationAction(ISD::UMAX, VT, VT == MVT::v16i8 ? Legal : Custom); | |||
| 957 | setOperationAction(ISD::UMIN, VT, VT == MVT::v16i8 ? Legal : Custom); | |||
| 958 | } | |||
| 959 | ||||
| 960 | setOperationAction(ISD::UADDSAT, MVT::v16i8, Legal); | |||
| 961 | setOperationAction(ISD::SADDSAT, MVT::v16i8, Legal); | |||
| 962 | setOperationAction(ISD::USUBSAT, MVT::v16i8, Legal); | |||
| 963 | setOperationAction(ISD::SSUBSAT, MVT::v16i8, Legal); | |||
| 964 | setOperationAction(ISD::UADDSAT, MVT::v8i16, Legal); | |||
| 965 | setOperationAction(ISD::SADDSAT, MVT::v8i16, Legal); | |||
| 966 | setOperationAction(ISD::USUBSAT, MVT::v8i16, Legal); | |||
| 967 | setOperationAction(ISD::SSUBSAT, MVT::v8i16, Legal); | |||
| 968 | setOperationAction(ISD::USUBSAT, MVT::v4i32, Custom); | |||
| 969 | setOperationAction(ISD::USUBSAT, MVT::v2i64, Custom); | |||
| 970 | ||||
| 971 | setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom); | |||
| 972 | setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom); | |||
| 973 | setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom); | |||
| 974 | ||||
| 975 | for (auto VT : { MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v2i64 }) { | |||
| 976 | setOperationAction(ISD::SETCC, VT, Custom); | |||
| 977 | setOperationAction(ISD::STRICT_FSETCC, VT, Custom); | |||
| 978 | setOperationAction(ISD::STRICT_FSETCCS, VT, Custom); | |||
| 979 | setOperationAction(ISD::CTPOP, VT, Custom); | |||
| 980 | setOperationAction(ISD::ABS, VT, Custom); | |||
| 981 | ||||
| 982 | // The condition codes aren't legal in SSE/AVX and under AVX512 we use | |||
| 983 | // setcc all the way to isel and prefer SETGT in some isel patterns. | |||
| 984 | setCondCodeAction(ISD::SETLT, VT, Custom); | |||
| 985 | setCondCodeAction(ISD::SETLE, VT, Custom); | |||
| 986 | } | |||
| 987 | ||||
| 988 | for (auto VT : { MVT::v16i8, MVT::v8i16, MVT::v4i32 }) { | |||
| 989 | setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom); | |||
| 990 | setOperationAction(ISD::BUILD_VECTOR, VT, Custom); | |||
| 991 | setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom); | |||
| 992 | setOperationAction(ISD::VSELECT, VT, Custom); | |||
| 993 | setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom); | |||
| 994 | } | |||
| 995 | ||||
| 996 | for (auto VT : { MVT::v2f64, MVT::v2i64 }) { | |||
| 997 | setOperationAction(ISD::BUILD_VECTOR, VT, Custom); | |||
| 998 | setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom); | |||
| 999 | setOperationAction(ISD::VSELECT, VT, Custom); | |||
| 1000 | ||||
| 1001 | if (VT == MVT::v2i64 && !Subtarget.is64Bit()) | |||
| 1002 | continue; | |||
| 1003 | ||||
| 1004 | setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom); | |||
| 1005 | setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom); | |||
| 1006 | } | |||
| 1007 | ||||
| 1008 | // Custom lower v2i64 and v2f64 selects. | |||
| 1009 | setOperationAction(ISD::SELECT, MVT::v2f64, Custom); | |||
| 1010 | setOperationAction(ISD::SELECT, MVT::v2i64, Custom); | |||
| 1011 | setOperationAction(ISD::SELECT, MVT::v4i32, Custom); | |||
| 1012 | setOperationAction(ISD::SELECT, MVT::v8i16, Custom); | |||
| 1013 | setOperationAction(ISD::SELECT, MVT::v16i8, Custom); | |||
| 1014 | ||||
| 1015 | setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal); | |||
| 1016 | setOperationAction(ISD::FP_TO_UINT, MVT::v4i32, Custom); | |||
| 1017 | setOperationAction(ISD::FP_TO_SINT, MVT::v2i32, Custom); | |||
| 1018 | setOperationAction(ISD::FP_TO_UINT, MVT::v2i32, Custom); | |||
| 1019 | setOperationAction(ISD::STRICT_FP_TO_SINT, MVT::v4i32, Legal); | |||
| 1020 | setOperationAction(ISD::STRICT_FP_TO_SINT, MVT::v2i32, Custom); | |||
| 1021 | ||||
| 1022 | // Custom legalize these to avoid over promotion or custom promotion. | |||
| 1023 | for (auto VT : {MVT::v2i8, MVT::v4i8, MVT::v8i8, MVT::v2i16, MVT::v4i16}) { | |||
| 1024 | setOperationAction(ISD::FP_TO_SINT, VT, Custom); | |||
| 1025 | setOperationAction(ISD::FP_TO_UINT, VT, Custom); | |||
| 1026 | setOperationAction(ISD::STRICT_FP_TO_SINT, VT, Custom); | |||
| 1027 | setOperationAction(ISD::STRICT_FP_TO_UINT, VT, Custom); | |||
| 1028 | } | |||
| 1029 | ||||
| 1030 | setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal); | |||
| 1031 | setOperationAction(ISD::STRICT_SINT_TO_FP, MVT::v4i32, Legal); | |||
| 1032 | setOperationAction(ISD::SINT_TO_FP, MVT::v2i32, Custom); | |||
| 1033 | setOperationAction(ISD::STRICT_SINT_TO_FP, MVT::v2i32, Custom); | |||
| 1034 | ||||
| 1035 | setOperationAction(ISD::UINT_TO_FP, MVT::v2i32, Custom); | |||
| 1036 | setOperationAction(ISD::STRICT_UINT_TO_FP, MVT::v2i32, Custom); | |||
| 1037 | ||||
| 1038 | setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Custom); | |||
| 1039 | setOperationAction(ISD::STRICT_UINT_TO_FP, MVT::v4i32, Custom); | |||
| 1040 | ||||
| 1041 | // Fast v2f32 UINT_TO_FP( v2i32 ) custom conversion. | |||
| 1042 | setOperationAction(ISD::SINT_TO_FP, MVT::v2f32, Custom); | |||
| 1043 | setOperationAction(ISD::STRICT_SINT_TO_FP, MVT::v2f32, Custom); | |||
| 1044 | setOperationAction(ISD::UINT_TO_FP, MVT::v2f32, Custom); | |||
| 1045 | setOperationAction(ISD::STRICT_UINT_TO_FP, MVT::v2f32, Custom); | |||
| 1046 | ||||
| 1047 | setOperationAction(ISD::FP_EXTEND, MVT::v2f32, Custom); | |||
| 1048 | setOperationAction(ISD::STRICT_FP_EXTEND, MVT::v2f32, Custom); | |||
| 1049 | setOperationAction(ISD::FP_ROUND, MVT::v2f32, Custom); | |||
| 1050 | setOperationAction(ISD::STRICT_FP_ROUND, MVT::v2f32, Custom); | |||
| 1051 | ||||
| 1052 | // We want to legalize this to an f64 load rather than an i64 load on | |||
| 1053 | // 64-bit targets and two 32-bit loads on a 32-bit target. Similar for | |||
| 1054 | // store. | |||
| 1055 | setOperationAction(ISD::LOAD, MVT::v2i32, Custom); | |||
| 1056 | setOperationAction(ISD::LOAD, MVT::v4i16, Custom); | |||
| 1057 | setOperationAction(ISD::LOAD, MVT::v8i8, Custom); | |||
| 1058 | setOperationAction(ISD::STORE, MVT::v2i32, Custom); | |||
| 1059 | setOperationAction(ISD::STORE, MVT::v4i16, Custom); | |||
| 1060 | setOperationAction(ISD::STORE, MVT::v8i8, Custom); | |||
| 1061 | ||||
| 1062 | setOperationAction(ISD::BITCAST, MVT::v2i32, Custom); | |||
| 1063 | setOperationAction(ISD::BITCAST, MVT::v4i16, Custom); | |||
| 1064 | setOperationAction(ISD::BITCAST, MVT::v8i8, Custom); | |||
| 1065 | if (!Subtarget.hasAVX512()) | |||
| 1066 | setOperationAction(ISD::BITCAST, MVT::v16i1, Custom); | |||
| 1067 | ||||
| 1068 | setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, MVT::v2i64, Custom); | |||
| 1069 | setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, MVT::v4i32, Custom); | |||
| 1070 | setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, MVT::v8i16, Custom); | |||
| 1071 | ||||
| 1072 | setOperationAction(ISD::SIGN_EXTEND, MVT::v4i64, Custom); | |||
| 1073 | ||||
| 1074 | setOperationAction(ISD::TRUNCATE, MVT::v2i8, Custom); | |||
| 1075 | setOperationAction(ISD::TRUNCATE, MVT::v2i16, Custom); | |||
| 1076 | setOperationAction(ISD::TRUNCATE, MVT::v2i32, Custom); | |||
| 1077 | setOperationAction(ISD::TRUNCATE, MVT::v4i8, Custom); | |||
| 1078 | setOperationAction(ISD::TRUNCATE, MVT::v4i16, Custom); | |||
| 1079 | setOperationAction(ISD::TRUNCATE, MVT::v8i8, Custom); | |||
| 1080 | ||||
| 1081 | // In the customized shift lowering, the legal v4i32/v2i64 cases | |||
| 1082 | // in AVX2 will be recognized. | |||
| 1083 | for (auto VT : { MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v2i64 }) { | |||
| 1084 | setOperationAction(ISD::SRL, VT, Custom); | |||
| 1085 | setOperationAction(ISD::SHL, VT, Custom); | |||
| 1086 | setOperationAction(ISD::SRA, VT, Custom); | |||
| 1087 | } | |||
| 1088 | ||||
| 1089 | setOperationAction(ISD::ROTL, MVT::v4i32, Custom); | |||
| 1090 | setOperationAction(ISD::ROTL, MVT::v8i16, Custom); | |||
| 1091 | ||||
| 1092 | // With 512-bit registers or AVX512VL+BW, expanding (and promoting the | |||
| 1093 | // shifts) is better. | |||
| 1094 | if (!Subtarget.useAVX512Regs() && | |||
| 1095 | !(Subtarget.hasBWI() && Subtarget.hasVLX())) | |||
| 1096 | setOperationAction(ISD::ROTL, MVT::v16i8, Custom); | |||
| 1097 | ||||
| 1098 | setOperationAction(ISD::STRICT_FSQRT, MVT::v2f64, Legal); | |||
| 1099 | setOperationAction(ISD::STRICT_FADD, MVT::v2f64, Legal); | |||
| 1100 | setOperationAction(ISD::STRICT_FSUB, MVT::v2f64, Legal); | |||
| 1101 | setOperationAction(ISD::STRICT_FMUL, MVT::v2f64, Legal); | |||
| 1102 | setOperationAction(ISD::STRICT_FDIV, MVT::v2f64, Legal); | |||
| 1103 | } | |||
| 1104 | ||||
| 1105 | if (!Subtarget.useSoftFloat() && Subtarget.hasSSSE3()) { | |||
| 1106 | setOperationAction(ISD::ABS, MVT::v16i8, Legal); | |||
| 1107 | setOperationAction(ISD::ABS, MVT::v8i16, Legal); | |||
| 1108 | setOperationAction(ISD::ABS, MVT::v4i32, Legal); | |||
| 1109 | setOperationAction(ISD::BITREVERSE, MVT::v16i8, Custom); | |||
| 1110 | setOperationAction(ISD::CTLZ, MVT::v16i8, Custom); | |||
| 1111 | setOperationAction(ISD::CTLZ, MVT::v8i16, Custom); | |||
| 1112 | setOperationAction(ISD::CTLZ, MVT::v4i32, Custom); | |||
| 1113 | setOperationAction(ISD::CTLZ, MVT::v2i64, Custom); | |||
| 1114 | ||||
| 1115 | // These might be better off as horizontal vector ops. | |||
| 1116 | setOperationAction(ISD::ADD, MVT::i16, Custom); | |||
| 1117 | setOperationAction(ISD::ADD, MVT::i32, Custom); | |||
| 1118 | setOperationAction(ISD::SUB, MVT::i16, Custom); | |||
| 1119 | setOperationAction(ISD::SUB, MVT::i32, Custom); | |||
| 1120 | } | |||
| 1121 | ||||
| 1122 | if (!Subtarget.useSoftFloat() && Subtarget.hasSSE41()) { | |||
| 1123 | for (MVT RoundedTy : {MVT::f32, MVT::f64, MVT::v4f32, MVT::v2f64}) { | |||
| 1124 | setOperationAction(ISD::FFLOOR, RoundedTy, Legal); | |||
| 1125 | setOperationAction(ISD::STRICT_FFLOOR, RoundedTy, Legal); | |||
| 1126 | setOperationAction(ISD::FCEIL, RoundedTy, Legal); | |||
| 1127 | setOperationAction(ISD::STRICT_FCEIL, RoundedTy, Legal); | |||
| 1128 | setOperationAction(ISD::FTRUNC, RoundedTy, Legal); | |||
| 1129 | setOperationAction(ISD::STRICT_FTRUNC, RoundedTy, Legal); | |||
| 1130 | setOperationAction(ISD::FRINT, RoundedTy, Legal); | |||
| 1131 | setOperationAction(ISD::STRICT_FRINT, RoundedTy, Legal); | |||
| 1132 | setOperationAction(ISD::FNEARBYINT, RoundedTy, Legal); | |||
| 1133 | setOperationAction(ISD::STRICT_FNEARBYINT, RoundedTy, Legal); | |||
| 1134 | setOperationAction(ISD::FROUNDEVEN, RoundedTy, Legal); | |||
| 1135 | setOperationAction(ISD::STRICT_FROUNDEVEN, RoundedTy, Legal); | |||
| 1136 | ||||
| 1137 | setOperationAction(ISD::FROUND, RoundedTy, Custom); | |||
| 1138 | } | |||
| 1139 | ||||
| 1140 | setOperationAction(ISD::SMAX, MVT::v16i8, Legal); | |||
| 1141 | setOperationAction(ISD::SMAX, MVT::v4i32, Legal); | |||
| 1142 | setOperationAction(ISD::UMAX, MVT::v8i16, Legal); | |||
| 1143 | setOperationAction(ISD::UMAX, MVT::v4i32, Legal); | |||
| 1144 | setOperationAction(ISD::SMIN, MVT::v16i8, Legal); | |||
| 1145 | setOperationAction(ISD::SMIN, MVT::v4i32, Legal); | |||
| 1146 | setOperationAction(ISD::UMIN, MVT::v8i16, Legal); | |||
| 1147 | setOperationAction(ISD::UMIN, MVT::v4i32, Legal); | |||
| 1148 | ||||
| 1149 | setOperationAction(ISD::UADDSAT, MVT::v4i32, Custom); | |||
| 1150 | ||||
| 1151 | // FIXME: Do we need to handle scalar-to-vector here? | |||
| 1152 | setOperationAction(ISD::MUL, MVT::v4i32, Legal); | |||
| 1153 | ||||
| 1154 | // We directly match byte blends in the backend as they match the VSELECT | |||
| 1155 | // condition form. | |||
| 1156 | setOperationAction(ISD::VSELECT, MVT::v16i8, Legal); | |||
| 1157 | ||||
| 1158 | // SSE41 brings specific instructions for doing vector sign extend even in | |||
| 1159 | // cases where we don't have SRA. | |||
| 1160 | for (auto VT : { MVT::v8i16, MVT::v4i32, MVT::v2i64 }) { | |||
| 1161 | setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, VT, Legal); | |||
| 1162 | setOperationAction(ISD::ZERO_EXTEND_VECTOR_INREG, VT, Legal); | |||
| 1163 | } | |||
| 1164 | ||||
| 1165 | // SSE41 also has vector sign/zero extending loads, PMOV[SZ]X | |||
| 1166 | for (auto LoadExtOp : { ISD::SEXTLOAD, ISD::ZEXTLOAD }) { | |||
| 1167 | setLoadExtAction(LoadExtOp, MVT::v8i16, MVT::v8i8, Legal); | |||
| 1168 | setLoadExtAction(LoadExtOp, MVT::v4i32, MVT::v4i8, Legal); | |||
| 1169 | setLoadExtAction(LoadExtOp, MVT::v2i64, MVT::v2i8, Legal); | |||
| 1170 | setLoadExtAction(LoadExtOp, MVT::v4i32, MVT::v4i16, Legal); | |||
| 1171 | setLoadExtAction(LoadExtOp, MVT::v2i64, MVT::v2i16, Legal); | |||
| 1172 | setLoadExtAction(LoadExtOp, MVT::v2i64, MVT::v2i32, Legal); | |||
| 1173 | } | |||
| 1174 | ||||
| 1175 | // i8 vectors are custom because the source register and source | |||
| 1176 | // source memory operand types are not the same width. | |||
| 1177 | setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom); | |||
| 1178 | ||||
| 1179 | if (Subtarget.is64Bit() && !Subtarget.hasAVX512()) { | |||
| 1180 | // We need to scalarize v4i64->v432 uint_to_fp using cvtsi2ss, but we can | |||
| 1181 | // do the pre and post work in the vector domain. | |||
| 1182 | setOperationAction(ISD::UINT_TO_FP, MVT::v4i64, Custom); | |||
| 1183 | setOperationAction(ISD::STRICT_UINT_TO_FP, MVT::v4i64, Custom); | |||
| 1184 | // We need to mark SINT_TO_FP as Custom even though we want to expand it | |||
| 1185 | // so that DAG combine doesn't try to turn it into uint_to_fp. | |||
| 1186 | setOperationAction(ISD::SINT_TO_FP, MVT::v4i64, Custom); | |||
| 1187 | setOperationAction(ISD::STRICT_SINT_TO_FP, MVT::v4i64, Custom); | |||
| 1188 | } | |||
| 1189 | } | |||
| 1190 | ||||
| 1191 | if (!Subtarget.useSoftFloat() && Subtarget.hasSSE42()) { | |||
| 1192 | setOperationAction(ISD::UADDSAT, MVT::v2i64, Custom); | |||
| 1193 | } | |||
| 1194 | ||||
| 1195 | if (!Subtarget.useSoftFloat() && Subtarget.hasXOP()) { | |||
| 1196 | for (auto VT : { MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v2i64, | |||
| 1197 | MVT::v32i8, MVT::v16i16, MVT::v8i32, MVT::v4i64 }) | |||
| 1198 | setOperationAction(ISD::ROTL, VT, Custom); | |||
| 1199 | ||||
| 1200 | // XOP can efficiently perform BITREVERSE with VPPERM. | |||
| 1201 | for (auto VT : { MVT::i8, MVT::i16, MVT::i32, MVT::i64 }) | |||
| 1202 | setOperationAction(ISD::BITREVERSE, VT, Custom); | |||
| 1203 | ||||
| 1204 | for (auto VT : { MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v2i64, | |||
| 1205 | MVT::v32i8, MVT::v16i16, MVT::v8i32, MVT::v4i64 }) | |||
| 1206 | setOperationAction(ISD::BITREVERSE, VT, Custom); | |||
| 1207 | } | |||
| 1208 | ||||
| 1209 | if (!Subtarget.useSoftFloat() && Subtarget.hasAVX()) { | |||
| 1210 | bool HasInt256 = Subtarget.hasInt256(); | |||
| 1211 | ||||
| 1212 | addRegisterClass(MVT::v32i8, Subtarget.hasVLX() ? &X86::VR256XRegClass | |||
| 1213 | : &X86::VR256RegClass); | |||
| 1214 | addRegisterClass(MVT::v16i16, Subtarget.hasVLX() ? &X86::VR256XRegClass | |||
| 1215 | : &X86::VR256RegClass); | |||
| 1216 | addRegisterClass(MVT::v8i32, Subtarget.hasVLX() ? &X86::VR256XRegClass | |||
| 1217 | : &X86::VR256RegClass); | |||
| 1218 | addRegisterClass(MVT::v8f32, Subtarget.hasVLX() ? &X86::VR256XRegClass | |||
| 1219 | : &X86::VR256RegClass); | |||
| 1220 | addRegisterClass(MVT::v4i64, Subtarget.hasVLX() ? &X86::VR256XRegClass | |||
| 1221 | : &X86::VR256RegClass); | |||
| 1222 | addRegisterClass(MVT::v4f64, Subtarget.hasVLX() ? &X86::VR256XRegClass | |||
| 1223 | : &X86::VR256RegClass); | |||
| 1224 | ||||
| 1225 | for (auto VT : { MVT::v8f32, MVT::v4f64 }) { | |||
| 1226 | setOperationAction(ISD::FFLOOR, VT, Legal); | |||
| 1227 | setOperationAction(ISD::STRICT_FFLOOR, VT, Legal); | |||
| 1228 | setOperationAction(ISD::FCEIL, VT, Legal); | |||
| 1229 | setOperationAction(ISD::STRICT_FCEIL, VT, Legal); | |||
| 1230 | setOperationAction(ISD::FTRUNC, VT, Legal); | |||
| 1231 | setOperationAction(ISD::STRICT_FTRUNC, VT, Legal); | |||
| 1232 | setOperationAction(ISD::FRINT, VT, Legal); | |||
| 1233 | setOperationAction(ISD::STRICT_FRINT, VT, Legal); | |||
| 1234 | setOperationAction(ISD::FNEARBYINT, VT, Legal); | |||
| 1235 | setOperationAction(ISD::STRICT_FNEARBYINT, VT, Legal); | |||
| 1236 | setOperationAction(ISD::FROUNDEVEN, VT, Legal); | |||
| 1237 | setOperationAction(ISD::STRICT_FROUNDEVEN, VT, Legal); | |||
| 1238 | ||||
| 1239 | setOperationAction(ISD::FROUND, VT, Custom); | |||
| 1240 | ||||
| 1241 | setOperationAction(ISD::FNEG, VT, Custom); | |||
| 1242 | setOperationAction(ISD::FABS, VT, Custom); | |||
| 1243 | setOperationAction(ISD::FCOPYSIGN, VT, Custom); | |||
| 1244 | } | |||
| 1245 | ||||
| 1246 | // (fp_to_int:v8i16 (v8f32 ..)) requires the result type to be promoted | |||
| 1247 | // even though v8i16 is a legal type. | |||
| 1248 | setOperationPromotedToType(ISD::FP_TO_SINT, MVT::v8i16, MVT::v8i32); | |||
| 1249 | setOperationPromotedToType(ISD::FP_TO_UINT, MVT::v8i16, MVT::v8i32); | |||
| 1250 | setOperationPromotedToType(ISD::STRICT_FP_TO_SINT, MVT::v8i16, MVT::v8i32); | |||
| 1251 | setOperationPromotedToType(ISD::STRICT_FP_TO_UINT, MVT::v8i16, MVT::v8i32); | |||
| 1252 | setOperationAction(ISD::FP_TO_SINT, MVT::v8i32, Legal); | |||
| 1253 | setOperationAction(ISD::FP_TO_UINT, MVT::v8i32, Custom); | |||
| 1254 | setOperationAction(ISD::STRICT_FP_TO_SINT, MVT::v8i32, Legal); | |||
| 1255 | ||||
| 1256 | setOperationAction(ISD::SINT_TO_FP, MVT::v8i32, Legal); | |||
| 1257 | setOperationAction(ISD::STRICT_SINT_TO_FP, MVT::v8i32, Legal); | |||
| 1258 | ||||
| 1259 | setOperationAction(ISD::STRICT_FP_ROUND, MVT::v4f32, Legal); | |||
| 1260 | setOperationAction(ISD::STRICT_FADD, MVT::v8f32, Legal); | |||
| 1261 | setOperationAction(ISD::STRICT_FADD, MVT::v4f64, Legal); | |||
| 1262 | setOperationAction(ISD::STRICT_FSUB, MVT::v8f32, Legal); | |||
| 1263 | setOperationAction(ISD::STRICT_FSUB, MVT::v4f64, Legal); | |||
| 1264 | setOperationAction(ISD::STRICT_FMUL, MVT::v8f32, Legal); | |||
| 1265 | setOperationAction(ISD::STRICT_FMUL, MVT::v4f64, Legal); | |||
| 1266 | setOperationAction(ISD::STRICT_FDIV, MVT::v8f32, Legal); | |||
| 1267 | setOperationAction(ISD::STRICT_FDIV, MVT::v4f64, Legal); | |||
| 1268 | setOperationAction(ISD::STRICT_FP_EXTEND, MVT::v4f64, Legal); | |||
| 1269 | setOperationAction(ISD::STRICT_FSQRT, MVT::v8f32, Legal); | |||
| 1270 | setOperationAction(ISD::STRICT_FSQRT, MVT::v4f64, Legal); | |||
| 1271 | ||||
| 1272 | if (!Subtarget.hasAVX512()) | |||
| 1273 | setOperationAction(ISD::BITCAST, MVT::v32i1, Custom); | |||
| 1274 | ||||
| 1275 | // In the customized shift lowering, the legal v8i32/v4i64 cases | |||
| 1276 | // in AVX2 will be recognized. | |||
| 1277 | for (auto VT : { MVT::v32i8, MVT::v16i16, MVT::v8i32, MVT::v4i64 }) { | |||
| 1278 | setOperationAction(ISD::SRL, VT, Custom); | |||
| 1279 | setOperationAction(ISD::SHL, VT, Custom); | |||
| 1280 | setOperationAction(ISD::SRA, VT, Custom); | |||
| 1281 | } | |||
| 1282 | ||||
| 1283 | // These types need custom splitting if their input is a 128-bit vector. | |||
| 1284 | setOperationAction(ISD::SIGN_EXTEND, MVT::v8i64, Custom); | |||
| 1285 | setOperationAction(ISD::SIGN_EXTEND, MVT::v16i32, Custom); | |||
| 1286 | setOperationAction(ISD::ZERO_EXTEND, MVT::v8i64, Custom); | |||
| 1287 | setOperationAction(ISD::ZERO_EXTEND, MVT::v16i32, Custom); | |||
| 1288 | ||||
| 1289 | setOperationAction(ISD::ROTL, MVT::v8i32, Custom); | |||
| 1290 | setOperationAction(ISD::ROTL, MVT::v16i16, Custom); | |||
| 1291 | ||||
| 1292 | // With BWI, expanding (and promoting the shifts) is the better. | |||
| 1293 | if (!Subtarget.useBWIRegs()) | |||
| 1294 | setOperationAction(ISD::ROTL, MVT::v32i8, Custom); | |||
| 1295 | ||||
| 1296 | setOperationAction(ISD::SELECT, MVT::v4f64, Custom); | |||
| 1297 | setOperationAction(ISD::SELECT, MVT::v4i64, Custom); | |||
| 1298 | setOperationAction(ISD::SELECT, MVT::v8i32, Custom); | |||
| 1299 | setOperationAction(ISD::SELECT, MVT::v16i16, Custom); | |||
| 1300 | setOperationAction(ISD::SELECT, MVT::v32i8, Custom); | |||
| 1301 | setOperationAction(ISD::SELECT, MVT::v8f32, Custom); | |||
| 1302 | ||||
| 1303 | for (auto VT : { MVT::v16i16, MVT::v8i32, MVT::v4i64 }) { | |||
| 1304 | setOperationAction(ISD::SIGN_EXTEND, VT, Custom); | |||
| 1305 | setOperationAction(ISD::ZERO_EXTEND, VT, Custom); | |||
| 1306 | setOperationAction(ISD::ANY_EXTEND, VT, Custom); | |||
| 1307 | } | |||
| 1308 | ||||
| 1309 | setOperationAction(ISD::TRUNCATE, MVT::v16i8, Custom); | |||
| 1310 | setOperationAction(ISD::TRUNCATE, MVT::v8i16, Custom); | |||
| 1311 | setOperationAction(ISD::TRUNCATE, MVT::v4i32, Custom); | |||
| 1312 | setOperationAction(ISD::BITREVERSE, MVT::v32i8, Custom); | |||
| 1313 | ||||
| 1314 | for (auto VT : { MVT::v32i8, MVT::v16i16, MVT::v8i32, MVT::v4i64 }) { | |||
| 1315 | setOperationAction(ISD::SETCC, VT, Custom); | |||
| 1316 | setOperationAction(ISD::STRICT_FSETCC, VT, Custom); | |||
| 1317 | setOperationAction(ISD::STRICT_FSETCCS, VT, Custom); | |||
| 1318 | setOperationAction(ISD::CTPOP, VT, Custom); | |||
| 1319 | setOperationAction(ISD::CTLZ, VT, Custom); | |||
| 1320 | ||||
| 1321 | // The condition codes aren't legal in SSE/AVX and under AVX512 we use | |||
| 1322 | // setcc all the way to isel and prefer SETGT in some isel patterns. | |||
| 1323 | setCondCodeAction(ISD::SETLT, VT, Custom); | |||
| 1324 | setCondCodeAction(ISD::SETLE, VT, Custom); | |||
| 1325 | } | |||
| 1326 | ||||
| 1327 | if (Subtarget.hasAnyFMA()) { | |||
| 1328 | for (auto VT : { MVT::f32, MVT::f64, MVT::v4f32, MVT::v8f32, | |||
| 1329 | MVT::v2f64, MVT::v4f64 }) { | |||
| 1330 | setOperationAction(ISD::FMA, VT, Legal); | |||
| 1331 | setOperationAction(ISD::STRICT_FMA, VT, Legal); | |||
| 1332 | } | |||
| 1333 | } | |||
| 1334 | ||||
| 1335 | for (auto VT : { MVT::v32i8, MVT::v16i16, MVT::v8i32, MVT::v4i64 }) { | |||
| 1336 | setOperationAction(ISD::ADD, VT, HasInt256 ? Legal : Custom); | |||
| 1337 | setOperationAction(ISD::SUB, VT, HasInt256 ? Legal : Custom); | |||
| 1338 | } | |||
| 1339 | ||||
| 1340 | setOperationAction(ISD::MUL, MVT::v4i64, Custom); | |||
| 1341 | setOperationAction(ISD::MUL, MVT::v8i32, HasInt256 ? Legal : Custom); | |||
| 1342 | setOperationAction(ISD::MUL, MVT::v16i16, HasInt256 ? Legal : Custom); | |||
| 1343 | setOperationAction(ISD::MUL, MVT::v32i8, Custom); | |||
| 1344 | ||||
| 1345 | setOperationAction(ISD::MULHU, MVT::v8i32, Custom); | |||
| 1346 | setOperationAction(ISD::MULHS, MVT::v8i32, Custom); | |||
| 1347 | setOperationAction(ISD::MULHU, MVT::v16i16, HasInt256 ? Legal : Custom); | |||
| 1348 | setOperationAction(ISD::MULHS, MVT::v16i16, HasInt256 ? Legal : Custom); | |||
| 1349 | setOperationAction(ISD::MULHU, MVT::v32i8, Custom); | |||
| 1350 | setOperationAction(ISD::MULHS, MVT::v32i8, Custom); | |||
| 1351 | ||||
| 1352 | setOperationAction(ISD::SMULO, MVT::v32i8, Custom); | |||
| 1353 | setOperationAction(ISD::UMULO, MVT::v32i8, Custom); | |||
| 1354 | ||||
| 1355 | setOperationAction(ISD::ABS, MVT::v4i64, Custom); | |||
| 1356 | setOperationAction(ISD::SMAX, MVT::v4i64, Custom); | |||
| 1357 | setOperationAction(ISD::UMAX, MVT::v4i64, Custom); | |||
| 1358 | setOperationAction(ISD::SMIN, MVT::v4i64, Custom); | |||
| 1359 | setOperationAction(ISD::UMIN, MVT::v4i64, Custom); | |||
| 1360 | ||||
| 1361 | setOperationAction(ISD::UADDSAT, MVT::v32i8, HasInt256 ? Legal : Custom); | |||
| 1362 | setOperationAction(ISD::SADDSAT, MVT::v32i8, HasInt256 ? Legal : Custom); | |||
| 1363 | setOperationAction(ISD::USUBSAT, MVT::v32i8, HasInt256 ? Legal : Custom); | |||
| 1364 | setOperationAction(ISD::SSUBSAT, MVT::v32i8, HasInt256 ? Legal : Custom); | |||
| 1365 | setOperationAction(ISD::UADDSAT, MVT::v16i16, HasInt256 ? Legal : Custom); | |||
| 1366 | setOperationAction(ISD::SADDSAT, MVT::v16i16, HasInt256 ? Legal : Custom); | |||
| 1367 | setOperationAction(ISD::USUBSAT, MVT::v16i16, HasInt256 ? Legal : Custom); | |||
| 1368 | setOperationAction(ISD::SSUBSAT, MVT::v16i16, HasInt256 ? Legal : Custom); | |||
| 1369 | setOperationAction(ISD::UADDSAT, MVT::v8i32, Custom); | |||
| 1370 | setOperationAction(ISD::USUBSAT, MVT::v8i32, Custom); | |||
| 1371 | setOperationAction(ISD::UADDSAT, MVT::v4i64, Custom); | |||
| 1372 | setOperationAction(ISD::USUBSAT, MVT::v4i64, Custom); | |||
| 1373 | ||||
| 1374 | for (auto VT : { MVT::v32i8, MVT::v16i16, MVT::v8i32 }) { | |||
| 1375 | setOperationAction(ISD::ABS, VT, HasInt256 ? Legal : Custom); | |||
| 1376 | setOperationAction(ISD::SMAX, VT, HasInt256 ? Legal : Custom); | |||
| 1377 | setOperationAction(ISD::UMAX, VT, HasInt256 ? Legal : Custom); | |||
| 1378 | setOperationAction(ISD::SMIN, VT, HasInt256 ? Legal : Custom); | |||
| 1379 | setOperationAction(ISD::UMIN, VT, HasInt256 ? Legal : Custom); | |||
| 1380 | } | |||
| 1381 | ||||
| 1382 | for (auto VT : {MVT::v16i16, MVT::v8i32, MVT::v4i64}) { | |||
| 1383 | setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, VT, Custom); | |||
| 1384 | setOperationAction(ISD::ZERO_EXTEND_VECTOR_INREG, VT, Custom); | |||
| 1385 | } | |||
| 1386 | ||||
| 1387 | if (HasInt256) { | |||
| 1388 | // The custom lowering for UINT_TO_FP for v8i32 becomes interesting | |||
| 1389 | // when we have a 256bit-wide blend with immediate. | |||
| 1390 | setOperationAction(ISD::UINT_TO_FP, MVT::v8i32, Custom); | |||
| 1391 | setOperationAction(ISD::STRICT_UINT_TO_FP, MVT::v8i32, Custom); | |||
| 1392 | ||||
| 1393 | // AVX2 also has wider vector sign/zero extending loads, VPMOV[SZ]X | |||
| 1394 | for (auto LoadExtOp : { ISD::SEXTLOAD, ISD::ZEXTLOAD }) { | |||
| 1395 | setLoadExtAction(LoadExtOp, MVT::v16i16, MVT::v16i8, Legal); | |||
| 1396 | setLoadExtAction(LoadExtOp, MVT::v8i32, MVT::v8i8, Legal); | |||
| 1397 | setLoadExtAction(LoadExtOp, MVT::v4i64, MVT::v4i8, Legal); | |||
| 1398 | setLoadExtAction(LoadExtOp, MVT::v8i32, MVT::v8i16, Legal); | |||
| 1399 | setLoadExtAction(LoadExtOp, MVT::v4i64, MVT::v4i16, Legal); | |||
| 1400 | setLoadExtAction(LoadExtOp, MVT::v4i64, MVT::v4i32, Legal); | |||
| 1401 | } | |||
| 1402 | } | |||
| 1403 | ||||
| 1404 | for (auto VT : { MVT::v4i32, MVT::v8i32, MVT::v2i64, MVT::v4i64, | |||
| 1405 | MVT::v4f32, MVT::v8f32, MVT::v2f64, MVT::v4f64 }) { | |||
| 1406 | setOperationAction(ISD::MLOAD, VT, Subtarget.hasVLX() ? Legal : Custom); | |||
| 1407 | setOperationAction(ISD::MSTORE, VT, Legal); | |||
| 1408 | } | |||
| 1409 | ||||
| 1410 | // Extract subvector is special because the value type | |||
| 1411 | // (result) is 128-bit but the source is 256-bit wide. | |||
| 1412 | for (auto VT : { MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v2i64, | |||
| 1413 | MVT::v4f32, MVT::v2f64 }) { | |||
| 1414 | setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Legal); | |||
| 1415 | } | |||
| 1416 | ||||
| 1417 | // Custom lower several nodes for 256-bit types. | |||
| 1418 | for (MVT VT : { MVT::v32i8, MVT::v16i16, MVT::v8i32, MVT::v4i64, | |||
| 1419 | MVT::v8f32, MVT::v4f64 }) { | |||
| 1420 | setOperationAction(ISD::BUILD_VECTOR, VT, Custom); | |||
| 1421 | setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom); | |||
| 1422 | setOperationAction(ISD::VSELECT, VT, Custom); | |||
| 1423 | setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom); | |||
| 1424 | setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom); | |||
| 1425 | setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom); | |||
| 1426 | setOperationAction(ISD::INSERT_SUBVECTOR, VT, Legal); | |||
| 1427 | setOperationAction(ISD::CONCAT_VECTORS, VT, Custom); | |||
| 1428 | setOperationAction(ISD::STORE, VT, Custom); | |||
| 1429 | } | |||
| 1430 | ||||
| 1431 | if (HasInt256) { | |||
| 1432 | setOperationAction(ISD::VSELECT, MVT::v32i8, Legal); | |||
| 1433 | ||||
| 1434 | // Custom legalize 2x32 to get a little better code. | |||
| 1435 | setOperationAction(ISD::MGATHER, MVT::v2f32, Custom); | |||
| 1436 | setOperationAction(ISD::MGATHER, MVT::v2i32, Custom); | |||
| 1437 | ||||
| 1438 | for (auto VT : { MVT::v4i32, MVT::v8i32, MVT::v2i64, MVT::v4i64, | |||
| 1439 | MVT::v4f32, MVT::v8f32, MVT::v2f64, MVT::v4f64 }) | |||
| 1440 | setOperationAction(ISD::MGATHER, VT, Custom); | |||
| 1441 | } | |||
| 1442 | } | |||
| 1443 | ||||
| 1444 | // This block controls legalization of the mask vector sizes that are | |||
| 1445 | // available with AVX512. 512-bit vectors are in a separate block controlled | |||
| 1446 | // by useAVX512Regs. | |||
| 1447 | if (!Subtarget.useSoftFloat() && Subtarget.hasAVX512()) { | |||
| 1448 | addRegisterClass(MVT::v1i1, &X86::VK1RegClass); | |||
| 1449 | addRegisterClass(MVT::v2i1, &X86::VK2RegClass); | |||
| 1450 | addRegisterClass(MVT::v4i1, &X86::VK4RegClass); | |||
| 1451 | addRegisterClass(MVT::v8i1, &X86::VK8RegClass); | |||
| 1452 | addRegisterClass(MVT::v16i1, &X86::VK16RegClass); | |||
| 1453 | ||||
| 1454 | setOperationAction(ISD::SELECT, MVT::v1i1, Custom); | |||
| 1455 | setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v1i1, Custom); | |||
| 1456 | setOperationAction(ISD::BUILD_VECTOR, MVT::v1i1, Custom); | |||
| 1457 | ||||
| 1458 | setOperationPromotedToType(ISD::FP_TO_SINT, MVT::v8i1, MVT::v8i32); | |||
| 1459 | setOperationPromotedToType(ISD::FP_TO_UINT, MVT::v8i1, MVT::v8i32); | |||
| 1460 | setOperationPromotedToType(ISD::FP_TO_SINT, MVT::v4i1, MVT::v4i32); | |||
| 1461 | setOperationPromotedToType(ISD::FP_TO_UINT, MVT::v4i1, MVT::v4i32); | |||
| 1462 | setOperationPromotedToType(ISD::STRICT_FP_TO_SINT, MVT::v8i1, MVT::v8i32); | |||
| 1463 | setOperationPromotedToType(ISD::STRICT_FP_TO_UINT, MVT::v8i1, MVT::v8i32); | |||
| 1464 | setOperationPromotedToType(ISD::STRICT_FP_TO_SINT, MVT::v4i1, MVT::v4i32); | |||
| 1465 | setOperationPromotedToType(ISD::STRICT_FP_TO_UINT, MVT::v4i1, MVT::v4i32); | |||
| 1466 | setOperationAction(ISD::FP_TO_SINT, MVT::v2i1, Custom); | |||
| 1467 | setOperationAction(ISD::FP_TO_UINT, MVT::v2i1, Custom); | |||
| 1468 | setOperationAction(ISD::STRICT_FP_TO_SINT, MVT::v2i1, Custom); | |||
| 1469 | setOperationAction(ISD::STRICT_FP_TO_UINT, MVT::v2i1, Custom); | |||
| 1470 | ||||
| 1471 | // There is no byte sized k-register load or store without AVX512DQ. | |||
| 1472 | if (!Subtarget.hasDQI()) { | |||
| 1473 | setOperationAction(ISD::LOAD, MVT::v1i1, Custom); | |||
| 1474 | setOperationAction(ISD::LOAD, MVT::v2i1, Custom); | |||
| 1475 | setOperationAction(ISD::LOAD, MVT::v4i1, Custom); | |||
| 1476 | setOperationAction(ISD::LOAD, MVT::v8i1, Custom); | |||
| 1477 | ||||
| 1478 | setOperationAction(ISD::STORE, MVT::v1i1, Custom); | |||
| 1479 | setOperationAction(ISD::STORE, MVT::v2i1, Custom); | |||
| 1480 | setOperationAction(ISD::STORE, MVT::v4i1, Custom); | |||
| 1481 | setOperationAction(ISD::STORE, MVT::v8i1, Custom); | |||
| 1482 | } | |||
| 1483 | ||||
| 1484 | // Extends of v16i1/v8i1/v4i1/v2i1 to 128-bit vectors. | |||
| 1485 | for (auto VT : { MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v2i64 }) { | |||
| 1486 | setOperationAction(ISD::SIGN_EXTEND, VT, Custom); | |||
| 1487 | setOperationAction(ISD::ZERO_EXTEND, VT, Custom); | |||
| 1488 | setOperationAction(ISD::ANY_EXTEND, VT, Custom); | |||
| 1489 | } | |||
| 1490 | ||||
| 1491 | for (auto VT : { MVT::v1i1, MVT::v2i1, MVT::v4i1, MVT::v8i1, MVT::v16i1 }) | |||
| 1492 | setOperationAction(ISD::VSELECT, VT, Expand); | |||
| 1493 | ||||
| 1494 | for (auto VT : { MVT::v2i1, MVT::v4i1, MVT::v8i1, MVT::v16i1 }) { | |||
| 1495 | setOperationAction(ISD::SETCC, VT, Custom); | |||
| 1496 | setOperationAction(ISD::STRICT_FSETCC, VT, Custom); | |||
| 1497 | setOperationAction(ISD::STRICT_FSETCCS, VT, Custom); | |||
| 1498 | setOperationAction(ISD::SELECT, VT, Custom); | |||
| 1499 | setOperationAction(ISD::TRUNCATE, VT, Custom); | |||
| 1500 | ||||
| 1501 | setOperationAction(ISD::BUILD_VECTOR, VT, Custom); | |||
| 1502 | setOperationAction(ISD::CONCAT_VECTORS, VT, Custom); | |||
| 1503 | setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom); | |||
| 1504 | setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom); | |||
| 1505 | setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom); | |||
| 1506 | setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom); | |||
| 1507 | } | |||
| 1508 | ||||
| 1509 | for (auto VT : { MVT::v1i1, MVT::v2i1, MVT::v4i1, MVT::v8i1 }) | |||
| 1510 | setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom); | |||
| 1511 | } | |||
| 1512 | ||||
| 1513 | // This block controls legalization for 512-bit operations with 32/64 bit | |||
| 1514 | // elements. 512-bits can be disabled based on prefer-vector-width and | |||
| 1515 | // required-vector-width function attributes. | |||
| 1516 | if (!Subtarget.useSoftFloat() && Subtarget.useAVX512Regs()) { | |||
| 1517 | bool HasBWI = Subtarget.hasBWI(); | |||
| 1518 | ||||
| 1519 | addRegisterClass(MVT::v16i32, &X86::VR512RegClass); | |||
| 1520 | addRegisterClass(MVT::v16f32, &X86::VR512RegClass); | |||
| 1521 | addRegisterClass(MVT::v8i64, &X86::VR512RegClass); | |||
| 1522 | addRegisterClass(MVT::v8f64, &X86::VR512RegClass); | |||
| 1523 | addRegisterClass(MVT::v32i16, &X86::VR512RegClass); | |||
| 1524 | addRegisterClass(MVT::v64i8, &X86::VR512RegClass); | |||
| 1525 | ||||
| 1526 | for (auto ExtType : {ISD::ZEXTLOAD, ISD::SEXTLOAD}) { | |||
| 1527 | setLoadExtAction(ExtType, MVT::v16i32, MVT::v16i8, Legal); | |||
| 1528 | setLoadExtAction(ExtType, MVT::v16i32, MVT::v16i16, Legal); | |||
| 1529 | setLoadExtAction(ExtType, MVT::v8i64, MVT::v8i8, Legal); | |||
| 1530 | setLoadExtAction(ExtType, MVT::v8i64, MVT::v8i16, Legal); | |||
| 1531 | setLoadExtAction(ExtType, MVT::v8i64, MVT::v8i32, Legal); | |||
| 1532 | if (HasBWI) | |||
| 1533 | setLoadExtAction(ExtType, MVT::v32i16, MVT::v32i8, Legal); | |||
| 1534 | } | |||
| 1535 | ||||
| 1536 | for (MVT VT : { MVT::v16f32, MVT::v8f64 }) { | |||
| 1537 | setOperationAction(ISD::FNEG, VT, Custom); | |||
| 1538 | setOperationAction(ISD::FABS, VT, Custom); | |||
| 1539 | setOperationAction(ISD::FMA, VT, Legal); | |||
| 1540 | setOperationAction(ISD::STRICT_FMA, VT, Legal); | |||
| 1541 | setOperationAction(ISD::FCOPYSIGN, VT, Custom); | |||
| 1542 | } | |||
| 1543 | ||||
| 1544 | for (MVT VT : { MVT::v16i1, MVT::v16i8, MVT::v16i16 }) { | |||
| 1545 | setOperationPromotedToType(ISD::FP_TO_SINT , VT, MVT::v16i32); | |||
| 1546 | setOperationPromotedToType(ISD::FP_TO_UINT , VT, MVT::v16i32); | |||
| 1547 | setOperationPromotedToType(ISD::STRICT_FP_TO_SINT, VT, MVT::v16i32); | |||
| 1548 | setOperationPromotedToType(ISD::STRICT_FP_TO_UINT, VT, MVT::v16i32); | |||
| 1549 | } | |||
| 1550 | setOperationAction(ISD::FP_TO_SINT, MVT::v16i32, Legal); | |||
| 1551 | setOperationAction(ISD::FP_TO_UINT, MVT::v16i32, Legal); | |||
| 1552 | setOperationAction(ISD::STRICT_FP_TO_SINT, MVT::v16i32, Legal); | |||
| 1553 | setOperationAction(ISD::STRICT_FP_TO_UINT, MVT::v16i32, Legal); | |||
| 1554 | setOperationAction(ISD::SINT_TO_FP, MVT::v16i32, Legal); | |||
| 1555 | setOperationAction(ISD::UINT_TO_FP, MVT::v16i32, Legal); | |||
| 1556 | setOperationAction(ISD::STRICT_SINT_TO_FP, MVT::v16i32, Legal); | |||
| 1557 | setOperationAction(ISD::STRICT_UINT_TO_FP, MVT::v16i32, Legal); | |||
| 1558 | ||||
| 1559 | setOperationAction(ISD::STRICT_FADD, MVT::v16f32, Legal); | |||
| 1560 | setOperationAction(ISD::STRICT_FADD, MVT::v8f64, Legal); | |||
| 1561 | setOperationAction(ISD::STRICT_FSUB, MVT::v16f32, Legal); | |||
| 1562 | setOperationAction(ISD::STRICT_FSUB, MVT::v8f64, Legal); | |||
| 1563 | setOperationAction(ISD::STRICT_FMUL, MVT::v16f32, Legal); | |||
| 1564 | setOperationAction(ISD::STRICT_FMUL, MVT::v8f64, Legal); | |||
| 1565 | setOperationAction(ISD::STRICT_FDIV, MVT::v16f32, Legal); | |||
| 1566 | setOperationAction(ISD::STRICT_FDIV, MVT::v8f64, Legal); | |||
| 1567 | setOperationAction(ISD::STRICT_FSQRT, MVT::v16f32, Legal); | |||
| 1568 | setOperationAction(ISD::STRICT_FSQRT, MVT::v8f64, Legal); | |||
| 1569 | setOperationAction(ISD::STRICT_FP_EXTEND, MVT::v8f64, Legal); | |||
| 1570 | setOperationAction(ISD::STRICT_FP_ROUND, MVT::v8f32, Legal); | |||
| 1571 | ||||
| 1572 | setTruncStoreAction(MVT::v8i64, MVT::v8i8, Legal); | |||
| 1573 | setTruncStoreAction(MVT::v8i64, MVT::v8i16, Legal); | |||
| 1574 | setTruncStoreAction(MVT::v8i64, MVT::v8i32, Legal); | |||
| 1575 | setTruncStoreAction(MVT::v16i32, MVT::v16i8, Legal); | |||
| 1576 | setTruncStoreAction(MVT::v16i32, MVT::v16i16, Legal); | |||
| 1577 | if (HasBWI) | |||
| 1578 | setTruncStoreAction(MVT::v32i16, MVT::v32i8, Legal); | |||
| 1579 | ||||
| 1580 | // With 512-bit vectors and no VLX, we prefer to widen MLOAD/MSTORE | |||
| 1581 | // to 512-bit rather than use the AVX2 instructions so that we can use | |||
| 1582 | // k-masks. | |||
| 1583 | if (!Subtarget.hasVLX()) { | |||
| 1584 | for (auto VT : {MVT::v4i32, MVT::v8i32, MVT::v2i64, MVT::v4i64, | |||
| 1585 | MVT::v4f32, MVT::v8f32, MVT::v2f64, MVT::v4f64}) { | |||
| 1586 | setOperationAction(ISD::MLOAD, VT, Custom); | |||
| 1587 | setOperationAction(ISD::MSTORE, VT, Custom); | |||
| 1588 | } | |||
| 1589 | } | |||
| 1590 | ||||
| 1591 | setOperationAction(ISD::TRUNCATE, MVT::v8i32, Legal); | |||
| 1592 | setOperationAction(ISD::TRUNCATE, MVT::v16i16, Legal); | |||
| 1593 | setOperationAction(ISD::TRUNCATE, MVT::v32i8, HasBWI ? Legal : Custom); | |||
| 1594 | setOperationAction(ISD::TRUNCATE, MVT::v16i64, Custom); | |||
| 1595 | setOperationAction(ISD::ZERO_EXTEND, MVT::v32i16, Custom); | |||
| 1596 | setOperationAction(ISD::ZERO_EXTEND, MVT::v16i32, Custom); | |||
| 1597 | setOperationAction(ISD::ZERO_EXTEND, MVT::v8i64, Custom); | |||
| 1598 | setOperationAction(ISD::ANY_EXTEND, MVT::v32i16, Custom); | |||
| 1599 | setOperationAction(ISD::ANY_EXTEND, MVT::v16i32, Custom); | |||
| 1600 | setOperationAction(ISD::ANY_EXTEND, MVT::v8i64, Custom); | |||
| 1601 | setOperationAction(ISD::SIGN_EXTEND, MVT::v32i16, Custom); | |||
| 1602 | setOperationAction(ISD::SIGN_EXTEND, MVT::v16i32, Custom); | |||
| 1603 | setOperationAction(ISD::SIGN_EXTEND, MVT::v8i64, Custom); | |||
| 1604 | ||||
| 1605 | if (HasBWI) { | |||
| 1606 | // Extends from v64i1 masks to 512-bit vectors. | |||
| 1607 | setOperationAction(ISD::SIGN_EXTEND, MVT::v64i8, Custom); | |||
| 1608 | setOperationAction(ISD::ZERO_EXTEND, MVT::v64i8, Custom); | |||
| 1609 | setOperationAction(ISD::ANY_EXTEND, MVT::v64i8, Custom); | |||
| 1610 | } | |||
| 1611 | ||||
| 1612 | for (auto VT : { MVT::v16f32, MVT::v8f64 }) { | |||
| 1613 | setOperationAction(ISD::FFLOOR, VT, Legal); | |||
| 1614 | setOperationAction(ISD::STRICT_FFLOOR, VT, Legal); | |||
| 1615 | setOperationAction(ISD::FCEIL, VT, Legal); | |||
| 1616 | setOperationAction(ISD::STRICT_FCEIL, VT, Legal); | |||
| 1617 | setOperationAction(ISD::FTRUNC, VT, Legal); | |||
| 1618 | setOperationAction(ISD::STRICT_FTRUNC, VT, Legal); | |||
| 1619 | setOperationAction(ISD::FRINT, VT, Legal); | |||
| 1620 | setOperationAction(ISD::STRICT_FRINT, VT, Legal); | |||
| 1621 | setOperationAction(ISD::FNEARBYINT, VT, Legal); | |||
| 1622 | setOperationAction(ISD::STRICT_FNEARBYINT, VT, Legal); | |||
| 1623 | setOperationAction(ISD::FROUNDEVEN, VT, Legal); | |||
| 1624 | setOperationAction(ISD::STRICT_FROUNDEVEN, VT, Legal); | |||
| 1625 | ||||
| 1626 | setOperationAction(ISD::FROUND, VT, Custom); | |||
| 1627 | } | |||
| 1628 | ||||
| 1629 | for (auto VT : {MVT::v32i16, MVT::v16i32, MVT::v8i64}) { | |||
| 1630 | setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, VT, Custom); | |||
| 1631 | setOperationAction(ISD::ZERO_EXTEND_VECTOR_INREG, VT, Custom); | |||
| 1632 | } | |||
| 1633 | ||||
| 1634 | setOperationAction(ISD::ADD, MVT::v32i16, HasBWI ? Legal : Custom); | |||
| 1635 | setOperationAction(ISD::SUB, MVT::v32i16, HasBWI ? Legal : Custom); | |||
| 1636 | setOperationAction(ISD::ADD, MVT::v64i8, HasBWI ? Legal : Custom); | |||
| 1637 | setOperationAction(ISD::SUB, MVT::v64i8, HasBWI ? Legal : Custom); | |||
| 1638 | ||||
| 1639 | setOperationAction(ISD::MUL, MVT::v8i64, Custom); | |||
| 1640 | setOperationAction(ISD::MUL, MVT::v16i32, Legal); | |||
| 1641 | setOperationAction(ISD::MUL, MVT::v32i16, HasBWI ? Legal : Custom); | |||
| 1642 | setOperationAction(ISD::MUL, MVT::v64i8, Custom); | |||
| 1643 | ||||
| 1644 | setOperationAction(ISD::MULHU, MVT::v16i32, Custom); | |||
| 1645 | setOperationAction(ISD::MULHS, MVT::v16i32, Custom); | |||
| 1646 | setOperationAction(ISD::MULHS, MVT::v32i16, HasBWI ? Legal : Custom); | |||
| 1647 | setOperationAction(ISD::MULHU, MVT::v32i16, HasBWI ? Legal : Custom); | |||
| 1648 | setOperationAction(ISD::MULHS, MVT::v64i8, Custom); | |||
| 1649 | setOperationAction(ISD::MULHU, MVT::v64i8, Custom); | |||
| 1650 | ||||
| 1651 | setOperationAction(ISD::SMULO, MVT::v64i8, Custom); | |||
| 1652 | setOperationAction(ISD::UMULO, MVT::v64i8, Custom); | |||
| 1653 | ||||
| 1654 | setOperationAction(ISD::BITREVERSE, MVT::v64i8, Custom); | |||
| 1655 | ||||
| 1656 | for (auto VT : { MVT::v64i8, MVT::v32i16, MVT::v16i32, MVT::v8i64 }) { | |||
| 1657 | setOperationAction(ISD::SRL, VT, Custom); | |||
| 1658 | setOperationAction(ISD::SHL, VT, Custom); | |||
| 1659 | setOperationAction(ISD::SRA, VT, Custom); | |||
| 1660 | setOperationAction(ISD::SETCC, VT, Custom); | |||
| 1661 | ||||
| 1662 | // The condition codes aren't legal in SSE/AVX and under AVX512 we use | |||
| 1663 | // setcc all the way to isel and prefer SETGT in some isel patterns. | |||
| 1664 | setCondCodeAction(ISD::SETLT, VT, Custom); | |||
| 1665 | setCondCodeAction(ISD::SETLE, VT, Custom); | |||
| 1666 | } | |||
| 1667 | for (auto VT : { MVT::v16i32, MVT::v8i64 }) { | |||
| 1668 | setOperationAction(ISD::SMAX, VT, Legal); | |||
| 1669 | setOperationAction(ISD::UMAX, VT, Legal); | |||
| 1670 | setOperationAction(ISD::SMIN, VT, Legal); | |||
| 1671 | setOperationAction(ISD::UMIN, VT, Legal); | |||
| 1672 | setOperationAction(ISD::ABS, VT, Legal); | |||
| 1673 | setOperationAction(ISD::CTPOP, VT, Custom); | |||
| 1674 | setOperationAction(ISD::ROTL, VT, Custom); | |||
| 1675 | setOperationAction(ISD::ROTR, VT, Custom); | |||
| 1676 | setOperationAction(ISD::STRICT_FSETCC, VT, Custom); | |||
| 1677 | setOperationAction(ISD::STRICT_FSETCCS, VT, Custom); | |||
| 1678 | } | |||
| 1679 | ||||
| 1680 | for (auto VT : { MVT::v64i8, MVT::v32i16 }) { | |||
| 1681 | setOperationAction(ISD::ABS, VT, HasBWI ? Legal : Custom); | |||
| 1682 | setOperationAction(ISD::CTPOP, VT, Subtarget.hasBITALG() ? Legal : Custom); | |||
| 1683 | setOperationAction(ISD::CTLZ, VT, Custom); | |||
| 1684 | setOperationAction(ISD::SMAX, VT, HasBWI ? Legal : Custom); | |||
| 1685 | setOperationAction(ISD::UMAX, VT, HasBWI ? Legal : Custom); | |||
| 1686 | setOperationAction(ISD::SMIN, VT, HasBWI ? Legal : Custom); | |||
| 1687 | setOperationAction(ISD::UMIN, VT, HasBWI ? Legal : Custom); | |||
| 1688 | setOperationAction(ISD::UADDSAT, VT, HasBWI ? Legal : Custom); | |||
| 1689 | setOperationAction(ISD::SADDSAT, VT, HasBWI ? Legal : Custom); | |||
| 1690 | setOperationAction(ISD::USUBSAT, VT, HasBWI ? Legal : Custom); | |||
| 1691 | setOperationAction(ISD::SSUBSAT, VT, HasBWI ? Legal : Custom); | |||
| 1692 | } | |||
| 1693 | ||||
| 1694 | if (Subtarget.hasDQI()) { | |||
| 1695 | setOperationAction(ISD::SINT_TO_FP, MVT::v8i64, Legal); | |||
| 1696 | setOperationAction(ISD::UINT_TO_FP, MVT::v8i64, Legal); | |||
| 1697 | setOperationAction(ISD::STRICT_SINT_TO_FP, MVT::v8i64, Legal); | |||
| 1698 | setOperationAction(ISD::STRICT_UINT_TO_FP, MVT::v8i64, Legal); | |||
| 1699 | setOperationAction(ISD::FP_TO_SINT, MVT::v8i64, Legal); | |||
| 1700 | setOperationAction(ISD::FP_TO_UINT, MVT::v8i64, Legal); | |||
| 1701 | setOperationAction(ISD::STRICT_FP_TO_SINT, MVT::v8i64, Legal); | |||
| 1702 | setOperationAction(ISD::STRICT_FP_TO_UINT, MVT::v8i64, Legal); | |||
| 1703 | ||||
| 1704 | setOperationAction(ISD::MUL, MVT::v8i64, Legal); | |||
| 1705 | } | |||
| 1706 | ||||
| 1707 | if (Subtarget.hasCDI()) { | |||
| 1708 | // NonVLX sub-targets extend 128/256 vectors to use the 512 version. | |||
| 1709 | for (auto VT : { MVT::v16i32, MVT::v8i64} ) { | |||
| 1710 | setOperationAction(ISD::CTLZ, VT, Legal); | |||
| 1711 | } | |||
| 1712 | } // Subtarget.hasCDI() | |||
| 1713 | ||||
| 1714 | if (Subtarget.hasVPOPCNTDQ()) { | |||
| 1715 | for (auto VT : { MVT::v16i32, MVT::v8i64 }) | |||
| 1716 | setOperationAction(ISD::CTPOP, VT, Legal); | |||
| 1717 | } | |||
| 1718 | ||||
| 1719 | // Extract subvector is special because the value type | |||
| 1720 | // (result) is 256-bit but the source is 512-bit wide. | |||
| 1721 | // 128-bit was made Legal under AVX1. | |||
| 1722 | for (auto VT : { MVT::v32i8, MVT::v16i16, MVT::v8i32, MVT::v4i64, | |||
| 1723 | MVT::v8f32, MVT::v4f64 }) | |||
| 1724 | setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Legal); | |||
| 1725 | ||||
| 1726 | for (auto VT : { MVT::v64i8, MVT::v32i16, MVT::v16i32, MVT::v8i64, | |||
| 1727 | MVT::v16f32, MVT::v8f64 }) { | |||
| 1728 | setOperationAction(ISD::CONCAT_VECTORS, VT, Custom); | |||
| 1729 | setOperationAction(ISD::INSERT_SUBVECTOR, VT, Legal); | |||
| 1730 | setOperationAction(ISD::SELECT, VT, Custom); | |||
| 1731 | setOperationAction(ISD::VSELECT, VT, Custom); | |||
| 1732 | setOperationAction(ISD::BUILD_VECTOR, VT, Custom); | |||
| 1733 | setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom); | |||
| 1734 | setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom); | |||
| 1735 | setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom); | |||
| 1736 | setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom); | |||
| 1737 | } | |||
| 1738 | ||||
| 1739 | for (auto VT : { MVT::v16i32, MVT::v8i64, MVT::v16f32, MVT::v8f64 }) { | |||
| 1740 | setOperationAction(ISD::MLOAD, VT, Legal); | |||
| 1741 | setOperationAction(ISD::MSTORE, VT, Legal); | |||
| 1742 | setOperationAction(ISD::MGATHER, VT, Custom); | |||
| 1743 | setOperationAction(ISD::MSCATTER, VT, Custom); | |||
| 1744 | } | |||
| 1745 | if (HasBWI) { | |||
| 1746 | for (auto VT : { MVT::v64i8, MVT::v32i16 }) { | |||
| 1747 | setOperationAction(ISD::MLOAD, VT, Legal); | |||
| 1748 | setOperationAction(ISD::MSTORE, VT, Legal); | |||
| 1749 | } | |||
| 1750 | } else { | |||
| 1751 | setOperationAction(ISD::STORE, MVT::v32i16, Custom); | |||
| 1752 | setOperationAction(ISD::STORE, MVT::v64i8, Custom); | |||
| 1753 | } | |||
| 1754 | ||||
| 1755 | if (Subtarget.hasVBMI2()) { | |||
| 1756 | for (auto VT : { MVT::v8i16, MVT::v4i32, MVT::v2i64, | |||
| 1757 | MVT::v16i16, MVT::v8i32, MVT::v4i64, | |||
| 1758 | MVT::v32i16, MVT::v16i32, MVT::v8i64 }) { | |||
| 1759 | setOperationAction(ISD::FSHL, VT, Custom); | |||
| 1760 | setOperationAction(ISD::FSHR, VT, Custom); | |||
| 1761 | } | |||
| 1762 | ||||
| 1763 | setOperationAction(ISD::ROTL, MVT::v32i16, Custom); | |||
| 1764 | setOperationAction(ISD::ROTR, MVT::v8i16, Custom); | |||
| 1765 | setOperationAction(ISD::ROTR, MVT::v16i16, Custom); | |||
| 1766 | setOperationAction(ISD::ROTR, MVT::v32i16, Custom); | |||
| 1767 | } | |||
| 1768 | }// useAVX512Regs | |||
| 1769 | ||||
| 1770 | // This block controls legalization for operations that don't have | |||
| 1771 | // pre-AVX512 equivalents. Without VLX we use 512-bit operations for | |||
| 1772 | // narrower widths. | |||
| 1773 | if (!Subtarget.useSoftFloat() && Subtarget.hasAVX512()) { | |||
| 1774 | // These operations are handled on non-VLX by artificially widening in | |||
| 1775 | // isel patterns. | |||
| 1776 | ||||
| 1777 | setOperationAction(ISD::FP_TO_UINT, MVT::v8i32, | |||
| 1778 | Subtarget.hasVLX() ? Legal : Custom); | |||
| 1779 | setOperationAction(ISD::FP_TO_UINT, MVT::v4i32, | |||
| 1780 | Subtarget.hasVLX() ? Legal : Custom); | |||
| 1781 | setOperationAction(ISD::STRICT_FP_TO_UINT, MVT::v8i32, | |||
| 1782 | Subtarget.hasVLX() ? Legal : Custom); | |||
| 1783 | setOperationAction(ISD::STRICT_FP_TO_UINT, MVT::v4i32, | |||
| 1784 | Subtarget.hasVLX() ? Legal : Custom); | |||
| 1785 | setOperationAction(ISD::STRICT_FP_TO_UINT, MVT::v2i32, Custom); | |||
| 1786 | setOperationAction(ISD::UINT_TO_FP, MVT::v8i32, | |||
| 1787 | Subtarget.hasVLX() ? Legal : Custom); | |||
| 1788 | setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, | |||
| 1789 | Subtarget.hasVLX() ? Legal : Custom); | |||
| 1790 | setOperationAction(ISD::STRICT_UINT_TO_FP, MVT::v8i32, | |||
| 1791 | Subtarget.hasVLX() ? Legal : Custom); | |||
| 1792 | setOperationAction(ISD::STRICT_UINT_TO_FP, MVT::v4i32, | |||
| 1793 | Subtarget.hasVLX() ? Legal : Custom); | |||
| 1794 | ||||
| 1795 | if (Subtarget.hasDQI()) { | |||
| 1796 | // Fast v2f32 SINT_TO_FP( v2i64 ) custom conversion. | |||
| 1797 | // v2f32 UINT_TO_FP is already custom under SSE2. | |||
| 1798 | assert(isOperationCustom(ISD::UINT_TO_FP, MVT::v2f32) &&((void)0) | |||
| 1799 | isOperationCustom(ISD::STRICT_UINT_TO_FP, MVT::v2f32) &&((void)0) | |||
| 1800 | "Unexpected operation action!")((void)0); | |||
| 1801 | // v2i64 FP_TO_S/UINT(v2f32) custom conversion. | |||
| 1802 | setOperationAction(ISD::FP_TO_SINT, MVT::v2f32, Custom); | |||
| 1803 | setOperationAction(ISD::FP_TO_UINT, MVT::v2f32, Custom); | |||
| 1804 | setOperationAction(ISD::STRICT_FP_TO_SINT, MVT::v2f32, Custom); | |||
| 1805 | setOperationAction(ISD::STRICT_FP_TO_UINT, MVT::v2f32, Custom); | |||
| 1806 | } | |||
| 1807 | ||||
| 1808 | for (auto VT : { MVT::v2i64, MVT::v4i64 }) { | |||
| 1809 | setOperationAction(ISD::SMAX, VT, Legal); | |||
| 1810 | setOperationAction(ISD::UMAX, VT, Legal); | |||
| 1811 | setOperationAction(ISD::SMIN, VT, Legal); | |||
| 1812 | setOperationAction(ISD::UMIN, VT, Legal); | |||
| 1813 | setOperationAction(ISD::ABS, VT, Legal); | |||
| 1814 | } | |||
| 1815 | ||||
| 1816 | for (auto VT : { MVT::v4i32, MVT::v8i32, MVT::v2i64, MVT::v4i64 }) { | |||
| 1817 | setOperationAction(ISD::ROTL, VT, Custom); | |||
| 1818 | setOperationAction(ISD::ROTR, VT, Custom); | |||
| 1819 | } | |||
| 1820 | ||||
| 1821 | // Custom legalize 2x32 to get a little better code. | |||
| 1822 | setOperationAction(ISD::MSCATTER, MVT::v2f32, Custom); | |||
| 1823 | setOperationAction(ISD::MSCATTER, MVT::v2i32, Custom); | |||
| 1824 | ||||
| 1825 | for (auto VT : { MVT::v4i32, MVT::v8i32, MVT::v2i64, MVT::v4i64, | |||
| 1826 | MVT::v4f32, MVT::v8f32, MVT::v2f64, MVT::v4f64 }) | |||
| 1827 | setOperationAction(ISD::MSCATTER, VT, Custom); | |||
| 1828 | ||||
| 1829 | if (Subtarget.hasDQI()) { | |||
| 1830 | for (auto VT : { MVT::v2i64, MVT::v4i64 }) { | |||
| 1831 | setOperationAction(ISD::SINT_TO_FP, VT, | |||
| 1832 | Subtarget.hasVLX() ? Legal : Custom); | |||
| 1833 | setOperationAction(ISD::UINT_TO_FP, VT, | |||
| 1834 | Subtarget.hasVLX() ? Legal : Custom); | |||
| 1835 | setOperationAction(ISD::STRICT_SINT_TO_FP, VT, | |||
| 1836 | Subtarget.hasVLX() ? Legal : Custom); | |||
| 1837 | setOperationAction(ISD::STRICT_UINT_TO_FP, VT, | |||
| 1838 | Subtarget.hasVLX() ? Legal : Custom); | |||
| 1839 | setOperationAction(ISD::FP_TO_SINT, VT, | |||
| 1840 | Subtarget.hasVLX() ? Legal : Custom); | |||
| 1841 | setOperationAction(ISD::FP_TO_UINT, VT, | |||
| 1842 | Subtarget.hasVLX() ? Legal : Custom); | |||
| 1843 | setOperationAction(ISD::STRICT_FP_TO_SINT, VT, | |||
| 1844 | Subtarget.hasVLX() ? Legal : Custom); | |||
| 1845 | setOperationAction(ISD::STRICT_FP_TO_UINT, VT, | |||
| 1846 | Subtarget.hasVLX() ? Legal : Custom); | |||
| 1847 | setOperationAction(ISD::MUL, VT, Legal); | |||
| 1848 | } | |||
| 1849 | } | |||
| 1850 | ||||
| 1851 | if (Subtarget.hasCDI()) { | |||
| 1852 | for (auto VT : { MVT::v4i32, MVT::v8i32, MVT::v2i64, MVT::v4i64 }) { | |||
| 1853 | setOperationAction(ISD::CTLZ, VT, Legal); | |||
| 1854 | } | |||
| 1855 | } // Subtarget.hasCDI() | |||
| 1856 | ||||
| 1857 | if (Subtarget.hasVPOPCNTDQ()) { | |||
| 1858 | for (auto VT : { MVT::v4i32, MVT::v8i32, MVT::v2i64, MVT::v4i64 }) | |||
| 1859 | setOperationAction(ISD::CTPOP, VT, Legal); | |||
| 1860 | } | |||
| 1861 | } | |||
| 1862 | ||||
| 1863 | // This block control legalization of v32i1/v64i1 which are available with | |||
| 1864 | // AVX512BW. 512-bit v32i16 and v64i8 vector legalization is controlled with | |||
| 1865 | // useBWIRegs. | |||
| 1866 | if (!Subtarget.useSoftFloat() && Subtarget.hasBWI()) { | |||
| 1867 | addRegisterClass(MVT::v32i1, &X86::VK32RegClass); | |||
| 1868 | addRegisterClass(MVT::v64i1, &X86::VK64RegClass); | |||
| 1869 | ||||
| 1870 | for (auto VT : { MVT::v32i1, MVT::v64i1 }) { | |||
| 1871 | setOperationAction(ISD::VSELECT, VT, Expand); | |||
| 1872 | setOperationAction(ISD::TRUNCATE, VT, Custom); | |||
| 1873 | setOperationAction(ISD::SETCC, VT, Custom); | |||
| 1874 | setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom); | |||
| 1875 | setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom); | |||
| 1876 | setOperationAction(ISD::SELECT, VT, Custom); | |||
| 1877 | setOperationAction(ISD::BUILD_VECTOR, VT, Custom); | |||
| 1878 | setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom); | |||
| 1879 | setOperationAction(ISD::CONCAT_VECTORS, VT, Custom); | |||
| 1880 | setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom); | |||
| 1881 | } | |||
| 1882 | ||||
| 1883 | for (auto VT : { MVT::v16i1, MVT::v32i1 }) | |||
| 1884 | setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom); | |||
| 1885 | ||||
| 1886 | // Extends from v32i1 masks to 256-bit vectors. | |||
| 1887 | setOperationAction(ISD::SIGN_EXTEND, MVT::v32i8, Custom); | |||
| 1888 | setOperationAction(ISD::ZERO_EXTEND, MVT::v32i8, Custom); | |||
| 1889 | setOperationAction(ISD::ANY_EXTEND, MVT::v32i8, Custom); | |||
| 1890 | ||||
| 1891 | for (auto VT : { MVT::v32i8, MVT::v16i8, MVT::v16i16, MVT::v8i16 }) { | |||
| 1892 | setOperationAction(ISD::MLOAD, VT, Subtarget.hasVLX() ? Legal : Custom); | |||
| 1893 | setOperationAction(ISD::MSTORE, VT, Subtarget.hasVLX() ? Legal : Custom); | |||
| 1894 | } | |||
| 1895 | ||||
| 1896 | // These operations are handled on non-VLX by artificially widening in | |||
| 1897 | // isel patterns. | |||
| 1898 | // TODO: Custom widen in lowering on non-VLX and drop the isel patterns? | |||
| 1899 | ||||
| 1900 | if (Subtarget.hasBITALG()) { | |||
| 1901 | for (auto VT : { MVT::v16i8, MVT::v32i8, MVT::v8i16, MVT::v16i16 }) | |||
| 1902 | setOperationAction(ISD::CTPOP, VT, Legal); | |||
| 1903 | } | |||
| 1904 | } | |||
| 1905 | ||||
| 1906 | if (!Subtarget.useSoftFloat() && Subtarget.hasVLX()) { | |||
| 1907 | setTruncStoreAction(MVT::v4i64, MVT::v4i8, Legal); | |||
| 1908 | setTruncStoreAction(MVT::v4i64, MVT::v4i16, Legal); | |||
| 1909 | setTruncStoreAction(MVT::v4i64, MVT::v4i32, Legal); | |||
| 1910 | setTruncStoreAction(MVT::v8i32, MVT::v8i8, Legal); | |||
| 1911 | setTruncStoreAction(MVT::v8i32, MVT::v8i16, Legal); | |||
| 1912 | ||||
| 1913 | setTruncStoreAction(MVT::v2i64, MVT::v2i8, Legal); | |||
| 1914 | setTruncStoreAction(MVT::v2i64, MVT::v2i16, Legal); | |||
| 1915 | setTruncStoreAction(MVT::v2i64, MVT::v2i32, Legal); | |||
| 1916 | setTruncStoreAction(MVT::v4i32, MVT::v4i8, Legal); | |||
| 1917 | setTruncStoreAction(MVT::v4i32, MVT::v4i16, Legal); | |||
| 1918 | ||||
| 1919 | if (Subtarget.hasBWI()) { | |||
| 1920 | setTruncStoreAction(MVT::v16i16, MVT::v16i8, Legal); | |||
| 1921 | setTruncStoreAction(MVT::v8i16, MVT::v8i8, Legal); | |||
| 1922 | } | |||
| 1923 | ||||
| 1924 | setOperationAction(ISD::TRUNCATE, MVT::v16i32, Custom); | |||
| 1925 | setOperationAction(ISD::TRUNCATE, MVT::v8i64, Custom); | |||
| 1926 | setOperationAction(ISD::TRUNCATE, MVT::v16i64, Custom); | |||
| 1927 | } | |||
| 1928 | ||||
| 1929 | if (Subtarget.hasAMXTILE()) { | |||
| 1930 | addRegisterClass(MVT::x86amx, &X86::TILERegClass); | |||
| 1931 | } | |||
| 1932 | ||||
| 1933 | // We want to custom lower some of our intrinsics. | |||
| 1934 | setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom); | |||
| 1935 | setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::Other, Custom); | |||
| 1936 | setOperationAction(ISD::INTRINSIC_VOID, MVT::Other, Custom); | |||
| 1937 | if (!Subtarget.is64Bit()) { | |||
| 1938 | setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::i64, Custom); | |||
| 1939 | } | |||
| 1940 | ||||
| 1941 | // Only custom-lower 64-bit SADDO and friends on 64-bit because we don't | |||
| 1942 | // handle type legalization for these operations here. | |||
| 1943 | // | |||
| 1944 | // FIXME: We really should do custom legalization for addition and | |||
| 1945 | // subtraction on x86-32 once PR3203 is fixed. We really can't do much better | |||
| 1946 | // than generic legalization for 64-bit multiplication-with-overflow, though. | |||
| 1947 | for (auto VT : { MVT::i8, MVT::i16, MVT::i32, MVT::i64 }) { | |||
| 1948 | if (VT == MVT::i64 && !Subtarget.is64Bit()) | |||
| 1949 | continue; | |||
| 1950 | // Add/Sub/Mul with overflow operations are custom lowered. | |||
| 1951 | setOperationAction(ISD::SADDO, VT, Custom); | |||
| 1952 | setOperationAction(ISD::UADDO, VT, Custom); | |||
| 1953 | setOperationAction(ISD::SSUBO, VT, Custom); | |||
| 1954 | setOperationAction(ISD::USUBO, VT, Custom); | |||
| 1955 | setOperationAction(ISD::SMULO, VT, Custom); | |||
| 1956 | setOperationAction(ISD::UMULO, VT, Custom); | |||
| 1957 | ||||
| 1958 | // Support carry in as value rather than glue. | |||
| 1959 | setOperationAction(ISD::ADDCARRY, VT, Custom); | |||
| 1960 | setOperationAction(ISD::SUBCARRY, VT, Custom); | |||
| 1961 | setOperationAction(ISD::SETCCCARRY, VT, Custom); | |||
| 1962 | setOperationAction(ISD::SADDO_CARRY, VT, Custom); | |||
| 1963 | setOperationAction(ISD::SSUBO_CARRY, VT, Custom); | |||
| 1964 | } | |||
| 1965 | ||||
| 1966 | if (!Subtarget.is64Bit()) { | |||
| 1967 | // These libcalls are not available in 32-bit. | |||
| 1968 | setLibcallName(RTLIB::SHL_I128, nullptr); | |||
| 1969 | setLibcallName(RTLIB::SRL_I128, nullptr); | |||
| 1970 | setLibcallName(RTLIB::SRA_I128, nullptr); | |||
| 1971 | setLibcallName(RTLIB::MUL_I128, nullptr); | |||
| 1972 | } | |||
| 1973 | ||||
| 1974 | // Combine sin / cos into _sincos_stret if it is available. | |||
| 1975 | if (getLibcallName(RTLIB::SINCOS_STRET_F32) != nullptr && | |||
| 1976 | getLibcallName(RTLIB::SINCOS_STRET_F64) != nullptr) { | |||
| 1977 | setOperationAction(ISD::FSINCOS, MVT::f64, Custom); | |||
| 1978 | setOperationAction(ISD::FSINCOS, MVT::f32, Custom); | |||
| 1979 | } | |||
| 1980 | ||||
| 1981 | if (Subtarget.isTargetWin64()) { | |||
| 1982 | setOperationAction(ISD::SDIV, MVT::i128, Custom); | |||
| 1983 | setOperationAction(ISD::UDIV, MVT::i128, Custom); | |||
| 1984 | setOperationAction(ISD::SREM, MVT::i128, Custom); | |||
| 1985 | setOperationAction(ISD::UREM, MVT::i128, Custom); | |||
| 1986 | } | |||
| 1987 | ||||
| 1988 | // On 32 bit MSVC, `fmodf(f32)` is not defined - only `fmod(f64)` | |||
| 1989 | // is. We should promote the value to 64-bits to solve this. | |||
| 1990 | // This is what the CRT headers do - `fmodf` is an inline header | |||
| 1991 | // function casting to f64 and calling `fmod`. | |||
| 1992 | if (Subtarget.is32Bit() && | |||
| 1993 | (Subtarget.isTargetWindowsMSVC() || Subtarget.isTargetWindowsItanium())) | |||
| 1994 | for (ISD::NodeType Op : | |||
| 1995 | {ISD::FCEIL, ISD::STRICT_FCEIL, | |||
| 1996 | ISD::FCOS, ISD::STRICT_FCOS, | |||
| 1997 | ISD::FEXP, ISD::STRICT_FEXP, | |||
| 1998 | ISD::FFLOOR, ISD::STRICT_FFLOOR, | |||
| 1999 | ISD::FREM, ISD::STRICT_FREM, | |||
| 2000 | ISD::FLOG, ISD::STRICT_FLOG, | |||
| 2001 | ISD::FLOG10, ISD::STRICT_FLOG10, | |||
| 2002 | ISD::FPOW, ISD::STRICT_FPOW, | |||
| 2003 | ISD::FSIN, ISD::STRICT_FSIN}) | |||
| 2004 | if (isOperationExpand(Op, MVT::f32)) | |||
| 2005 | setOperationAction(Op, MVT::f32, Promote); | |||
| 2006 | ||||
| 2007 | // We have target-specific dag combine patterns for the following nodes: | |||
| 2008 | setTargetDAGCombine(ISD::VECTOR_SHUFFLE); | |||
| 2009 | setTargetDAGCombine(ISD::SCALAR_TO_VECTOR); | |||
| 2010 | setTargetDAGCombine(ISD::INSERT_VECTOR_ELT); | |||
| 2011 | setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT); | |||
| 2012 | setTargetDAGCombine(ISD::CONCAT_VECTORS); | |||
| 2013 | setTargetDAGCombine(ISD::INSERT_SUBVECTOR); | |||
| 2014 | setTargetDAGCombine(ISD::EXTRACT_SUBVECTOR); | |||
| 2015 | setTargetDAGCombine(ISD::BITCAST); | |||
| 2016 | setTargetDAGCombine(ISD::VSELECT); | |||
| 2017 | setTargetDAGCombine(ISD::SELECT); | |||
| 2018 | setTargetDAGCombine(ISD::SHL); | |||
| 2019 | setTargetDAGCombine(ISD::SRA); | |||
| 2020 | setTargetDAGCombine(ISD::SRL); | |||
| 2021 | setTargetDAGCombine(ISD::OR); | |||
| 2022 | setTargetDAGCombine(ISD::AND); | |||
| 2023 | setTargetDAGCombine(ISD::ADD); | |||
| 2024 | setTargetDAGCombine(ISD::FADD); | |||
| 2025 | setTargetDAGCombine(ISD::FSUB); | |||
| 2026 | setTargetDAGCombine(ISD::FNEG); | |||
| 2027 | setTargetDAGCombine(ISD::FMA); | |||
| 2028 | setTargetDAGCombine(ISD::STRICT_FMA); | |||
| 2029 | setTargetDAGCombine(ISD::FMINNUM); | |||
| 2030 | setTargetDAGCombine(ISD::FMAXNUM); | |||
| 2031 | setTargetDAGCombine(ISD::SUB); | |||
| 2032 | setTargetDAGCombine(ISD::LOAD); | |||
| 2033 | setTargetDAGCombine(ISD::MLOAD); | |||
| 2034 | setTargetDAGCombine(ISD::STORE); | |||
| 2035 | setTargetDAGCombine(ISD::MSTORE); | |||
| 2036 | setTargetDAGCombine(ISD::TRUNCATE); | |||
| 2037 | setTargetDAGCombine(ISD::ZERO_EXTEND); | |||
| 2038 | setTargetDAGCombine(ISD::ANY_EXTEND); | |||
| 2039 | setTargetDAGCombine(ISD::SIGN_EXTEND); | |||
| 2040 | setTargetDAGCombine(ISD::SIGN_EXTEND_INREG); | |||
| 2041 | setTargetDAGCombine(ISD::ANY_EXTEND_VECTOR_INREG); | |||
| 2042 | setTargetDAGCombine(ISD::SIGN_EXTEND_VECTOR_INREG); | |||
| 2043 | setTargetDAGCombine(ISD::ZERO_EXTEND_VECTOR_INREG); | |||
| 2044 | setTargetDAGCombine(ISD::SINT_TO_FP); | |||
| 2045 | setTargetDAGCombine(ISD::UINT_TO_FP); | |||
| 2046 | setTargetDAGCombine(ISD::STRICT_SINT_TO_FP); | |||
| 2047 | setTargetDAGCombine(ISD::STRICT_UINT_TO_FP); | |||
| 2048 | setTargetDAGCombine(ISD::SETCC); | |||
| 2049 | setTargetDAGCombine(ISD::MUL); | |||
| 2050 | setTargetDAGCombine(ISD::XOR); | |||
| 2051 | setTargetDAGCombine(ISD::MSCATTER); | |||
| 2052 | setTargetDAGCombine(ISD::MGATHER); | |||
| 2053 | setTargetDAGCombine(ISD::FP16_TO_FP); | |||
| 2054 | setTargetDAGCombine(ISD::FP_EXTEND); | |||
| 2055 | setTargetDAGCombine(ISD::STRICT_FP_EXTEND); | |||
| 2056 | setTargetDAGCombine(ISD::FP_ROUND); | |||
| 2057 | ||||
| 2058 | computeRegisterProperties(Subtarget.getRegisterInfo()); | |||
| 2059 | ||||
| 2060 | MaxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores | |||
| 2061 | MaxStoresPerMemsetOptSize = 8; | |||
| 2062 | MaxStoresPerMemcpy = 8; // For @llvm.memcpy -> sequence of stores | |||
| 2063 | MaxStoresPerMemcpyOptSize = 4; | |||
| 2064 | MaxStoresPerMemmove = 8; // For @llvm.memmove -> sequence of stores | |||
| 2065 | MaxStoresPerMemmoveOptSize = 4; | |||
| 2066 | ||||
| 2067 | // TODO: These control memcmp expansion in CGP and could be raised higher, but | |||
| 2068 | // that needs to benchmarked and balanced with the potential use of vector | |||
| 2069 | // load/store types (PR33329, PR33914). | |||
| 2070 | MaxLoadsPerMemcmp = 2; | |||
| 2071 | MaxLoadsPerMemcmpOptSize = 2; | |||
| 2072 | ||||
| 2073 | // Set loop alignment to 2^ExperimentalPrefLoopAlignment bytes (default: 2^4). | |||
| 2074 | setPrefLoopAlignment(Align(1ULL << ExperimentalPrefLoopAlignment)); | |||
| 2075 | ||||
| 2076 | // An out-of-order CPU can speculatively execute past a predictable branch, | |||
| 2077 | // but a conditional move could be stalled by an expensive earlier operation. | |||
| 2078 | PredictableSelectIsExpensive = Subtarget.getSchedModel().isOutOfOrder(); | |||
| 2079 | EnableExtLdPromotion = true; | |||
| 2080 | setPrefFunctionAlignment(Align(16)); | |||
| 2081 | ||||
| 2082 | verifyIntrinsicTables(); | |||
| 2083 | ||||
| 2084 | // Default to having -disable-strictnode-mutation on | |||
| 2085 | IsStrictFPEnabled = true; | |||
| 2086 | } | |||
| 2087 | ||||
| 2088 | // This has so far only been implemented for 64-bit MachO. | |||
| 2089 | bool X86TargetLowering::useLoadStackGuardNode() const { | |||
| 2090 | return Subtarget.isTargetMachO() && Subtarget.is64Bit(); | |||
| 2091 | } | |||
| 2092 | ||||
| 2093 | bool X86TargetLowering::useStackGuardXorFP() const { | |||
| 2094 | // Currently only MSVC CRTs XOR the frame pointer into the stack guard value. | |||
| 2095 | return Subtarget.getTargetTriple().isOSMSVCRT() && !Subtarget.isTargetMachO(); | |||
| 2096 | } | |||
| 2097 | ||||
| 2098 | SDValue X86TargetLowering::emitStackGuardXorFP(SelectionDAG &DAG, SDValue Val, | |||
| 2099 | const SDLoc &DL) const { | |||
| 2100 | EVT PtrTy = getPointerTy(DAG.getDataLayout()); | |||
| 2101 | unsigned XorOp = Subtarget.is64Bit() ? X86::XOR64_FP : X86::XOR32_FP; | |||
| 2102 | MachineSDNode *Node = DAG.getMachineNode(XorOp, DL, PtrTy, Val); | |||
| 2103 | return SDValue(Node, 0); | |||
| 2104 | } | |||
| 2105 | ||||
| 2106 | TargetLoweringBase::LegalizeTypeAction | |||
| 2107 | X86TargetLowering::getPreferredVectorAction(MVT VT) const { | |||
| 2108 | if ((VT == MVT::v32i1 || VT == MVT::v64i1) && Subtarget.hasAVX512() && | |||
| 2109 | !Subtarget.hasBWI()) | |||
| 2110 | return TypeSplitVector; | |||
| 2111 | ||||
| 2112 | if (!VT.isScalableVector() && VT.getVectorNumElements() != 1 && | |||
| 2113 | VT.getVectorElementType() != MVT::i1) | |||
| 2114 | return TypeWidenVector; | |||
| 2115 | ||||
| 2116 | return TargetLoweringBase::getPreferredVectorAction(VT); | |||
| 2117 | } | |||
| 2118 | ||||
| 2119 | static std::pair<MVT, unsigned> | |||
| 2120 | handleMaskRegisterForCallingConv(unsigned NumElts, CallingConv::ID CC, | |||
| 2121 | const X86Subtarget &Subtarget) { | |||
| 2122 | // v2i1/v4i1/v8i1/v16i1 all pass in xmm registers unless the calling | |||
| 2123 | // convention is one that uses k registers. | |||
| 2124 | if (NumElts == 2) | |||
| 2125 | return {MVT::v2i64, 1}; | |||
| 2126 | if (NumElts == 4) | |||
| 2127 | return {MVT::v4i32, 1}; | |||
| 2128 | if (NumElts == 8 && CC != CallingConv::X86_RegCall && | |||
| 2129 | CC != CallingConv::Intel_OCL_BI) | |||
| 2130 | return {MVT::v8i16, 1}; | |||
| 2131 | if (NumElts == 16 && CC != CallingConv::X86_RegCall && | |||
| 2132 | CC != CallingConv::Intel_OCL_BI) | |||
| 2133 | return {MVT::v16i8, 1}; | |||
| 2134 | // v32i1 passes in ymm unless we have BWI and the calling convention is | |||
| 2135 | // regcall. | |||
| 2136 | if (NumElts == 32 && (!Subtarget.hasBWI() || CC != CallingConv::X86_RegCall)) | |||
| 2137 | return {MVT::v32i8, 1}; | |||
| 2138 | // Split v64i1 vectors if we don't have v64i8 available. | |||
| 2139 | if (NumElts == 64 && Subtarget.hasBWI() && CC != CallingConv::X86_RegCall) { | |||
| 2140 | if (Subtarget.useAVX512Regs()) | |||
| 2141 | return {MVT::v64i8, 1}; | |||
| 2142 | return {MVT::v32i8, 2}; | |||
| 2143 | } | |||
| 2144 | ||||
| 2145 | // Break wide or odd vXi1 vectors into scalars to match avx2 behavior. | |||
| 2146 | if (!isPowerOf2_32(NumElts) || (NumElts == 64 && !Subtarget.hasBWI()) || | |||
| 2147 | NumElts > 64) | |||
| 2148 | return {MVT::i8, NumElts}; | |||
| 2149 | ||||
| 2150 | return {MVT::INVALID_SIMPLE_VALUE_TYPE, 0}; | |||
| 2151 | } | |||
| 2152 | ||||
| 2153 | MVT X86TargetLowering::getRegisterTypeForCallingConv(LLVMContext &Context, | |||
| 2154 | CallingConv::ID CC, | |||
| 2155 | EVT VT) const { | |||
| 2156 | if (VT.isVector() && VT.getVectorElementType() == MVT::i1 && | |||
| 2157 | Subtarget.hasAVX512()) { | |||
| 2158 | unsigned NumElts = VT.getVectorNumElements(); | |||
| 2159 | ||||
| 2160 | MVT RegisterVT; | |||
| 2161 | unsigned NumRegisters; | |||
| 2162 | std::tie(RegisterVT, NumRegisters) = | |||
| 2163 | handleMaskRegisterForCallingConv(NumElts, CC, Subtarget); | |||
| 2164 | if (RegisterVT != MVT::INVALID_SIMPLE_VALUE_TYPE) | |||
| 2165 | return RegisterVT; | |||
| 2166 | } | |||
| 2167 | ||||
| 2168 | return TargetLowering::getRegisterTypeForCallingConv(Context, CC, VT); | |||
| 2169 | } | |||
| 2170 | ||||
| 2171 | unsigned X86TargetLowering::getNumRegistersForCallingConv(LLVMContext &Context, | |||
| 2172 | CallingConv::ID CC, | |||
| 2173 | EVT VT) const { | |||
| 2174 | if (VT.isVector() && VT.getVectorElementType() == MVT::i1 && | |||
| 2175 | Subtarget.hasAVX512()) { | |||
| 2176 | unsigned NumElts = VT.getVectorNumElements(); | |||
| 2177 | ||||
| 2178 | MVT RegisterVT; | |||
| 2179 | unsigned NumRegisters; | |||
| 2180 | std::tie(RegisterVT, NumRegisters) = | |||
| 2181 | handleMaskRegisterForCallingConv(NumElts, CC, Subtarget); | |||
| 2182 | if (RegisterVT != MVT::INVALID_SIMPLE_VALUE_TYPE) | |||
| 2183 | return NumRegisters; | |||
| 2184 | } | |||
| 2185 | ||||
| 2186 | return TargetLowering::getNumRegistersForCallingConv(Context, CC, VT); | |||
| 2187 | } | |||
| 2188 | ||||
| 2189 | unsigned X86TargetLowering::getVectorTypeBreakdownForCallingConv( | |||
| 2190 | LLVMContext &Context, CallingConv::ID CC, EVT VT, EVT &IntermediateVT, | |||
| 2191 | unsigned &NumIntermediates, MVT &RegisterVT) const { | |||
| 2192 | // Break wide or odd vXi1 vectors into scalars to match avx2 behavior. | |||
| 2193 | if (VT.isVector() && VT.getVectorElementType() == MVT::i1 && | |||
| 2194 | Subtarget.hasAVX512() && | |||
| 2195 | (!isPowerOf2_32(VT.getVectorNumElements()) || | |||
| 2196 | (VT.getVectorNumElements() == 64 && !Subtarget.hasBWI()) || | |||
| 2197 | VT.getVectorNumElements() > 64)) { | |||
| 2198 | RegisterVT = MVT::i8; | |||
| 2199 | IntermediateVT = MVT::i1; | |||
| 2200 | NumIntermediates = VT.getVectorNumElements(); | |||
| 2201 | return NumIntermediates; | |||
| 2202 | } | |||
| 2203 | ||||
| 2204 | // Split v64i1 vectors if we don't have v64i8 available. | |||
| 2205 | if (VT == MVT::v64i1 && Subtarget.hasBWI() && !Subtarget.useAVX512Regs() && | |||
| 2206 | CC != CallingConv::X86_RegCall) { | |||
| 2207 | RegisterVT = MVT::v32i8; | |||
| 2208 | IntermediateVT = MVT::v32i1; | |||
| 2209 | NumIntermediates = 2; | |||
| 2210 | return 2; | |||
| 2211 | } | |||
| 2212 | ||||
| 2213 | return TargetLowering::getVectorTypeBreakdownForCallingConv(Context, CC, VT, IntermediateVT, | |||
| 2214 | NumIntermediates, RegisterVT); | |||
| 2215 | } | |||
| 2216 | ||||
| 2217 | EVT X86TargetLowering::getSetCCResultType(const DataLayout &DL, | |||
| 2218 | LLVMContext& Context, | |||
| 2219 | EVT VT) const { | |||
| 2220 | if (!VT.isVector()) | |||
| 2221 | return MVT::i8; | |||
| 2222 | ||||
| 2223 | if (Subtarget.hasAVX512()) { | |||
| 2224 | // Figure out what this type will be legalized to. | |||
| 2225 | EVT LegalVT = VT; | |||
| 2226 | while (getTypeAction(Context, LegalVT) != TypeLegal) | |||
| 2227 | LegalVT = getTypeToTransformTo(Context, LegalVT); | |||
| 2228 | ||||
| 2229 | // If we got a 512-bit vector then we'll definitely have a vXi1 compare. | |||
| 2230 | if (LegalVT.getSimpleVT().is512BitVector()) | |||
| 2231 | return EVT::getVectorVT(Context, MVT::i1, VT.getVectorElementCount()); | |||
| 2232 | ||||
| 2233 | if (LegalVT.getSimpleVT().isVector() && Subtarget.hasVLX()) { | |||
| 2234 | // If we legalized to less than a 512-bit vector, then we will use a vXi1 | |||
| 2235 | // compare for vXi32/vXi64 for sure. If we have BWI we will also support | |||
| 2236 | // vXi16/vXi8. | |||
| 2237 | MVT EltVT = LegalVT.getSimpleVT().getVectorElementType(); | |||
| 2238 | if (Subtarget.hasBWI() || EltVT.getSizeInBits() >= 32) | |||
| 2239 | return EVT::getVectorVT(Context, MVT::i1, VT.getVectorElementCount()); | |||
| 2240 | } | |||
| 2241 | } | |||
| 2242 | ||||
| 2243 | return VT.changeVectorElementTypeToInteger(); | |||
| 2244 | } | |||
| 2245 | ||||
| 2246 | /// Helper for getByValTypeAlignment to determine | |||
| 2247 | /// the desired ByVal argument alignment. | |||
| 2248 | static void getMaxByValAlign(Type *Ty, Align &MaxAlign) { | |||
| 2249 | if (MaxAlign == 16) | |||
| 2250 | return; | |||
| 2251 | if (VectorType *VTy = dyn_cast<VectorType>(Ty)) { | |||
| 2252 | if (VTy->getPrimitiveSizeInBits().getFixedSize() == 128) | |||
| 2253 | MaxAlign = Align(16); | |||
| 2254 | } else if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { | |||
| 2255 | Align EltAlign; | |||
| 2256 | getMaxByValAlign(ATy->getElementType(), EltAlign); | |||
| 2257 | if (EltAlign > MaxAlign) | |||
| 2258 | MaxAlign = EltAlign; | |||
| 2259 | } else if (StructType *STy = dyn_cast<StructType>(Ty)) { | |||
| 2260 | for (auto *EltTy : STy->elements()) { | |||
| 2261 | Align EltAlign; | |||
| 2262 | getMaxByValAlign(EltTy, EltAlign); | |||
| 2263 | if (EltAlign > MaxAlign) | |||
| 2264 | MaxAlign = EltAlign; | |||
| 2265 | if (MaxAlign == 16) | |||
| 2266 | break; | |||
| 2267 | } | |||
| 2268 | } | |||
| 2269 | } | |||
| 2270 | ||||
| 2271 | /// Return the desired alignment for ByVal aggregate | |||
| 2272 | /// function arguments in the caller parameter area. For X86, aggregates | |||
| 2273 | /// that contain SSE vectors are placed at 16-byte boundaries while the rest | |||
| 2274 | /// are at 4-byte boundaries. | |||
| 2275 | unsigned X86TargetLowering::getByValTypeAlignment(Type *Ty, | |||
| 2276 | const DataLayout &DL) const { | |||
| 2277 | if (Subtarget.is64Bit()) { | |||
| 2278 | // Max of 8 and alignment of type. | |||
| 2279 | Align TyAlign = DL.getABITypeAlign(Ty); | |||
| 2280 | if (TyAlign > 8) | |||
| 2281 | return TyAlign.value(); | |||
| 2282 | return 8; | |||
| 2283 | } | |||
| 2284 | ||||
| 2285 | Align Alignment(4); | |||
| 2286 | if (Subtarget.hasSSE1()) | |||
| 2287 | getMaxByValAlign(Ty, Alignment); | |||
| 2288 | return Alignment.value(); | |||
| 2289 | } | |||
| 2290 | ||||
| 2291 | /// It returns EVT::Other if the type should be determined using generic | |||
| 2292 | /// target-independent logic. | |||
| 2293 | /// For vector ops we check that the overall size isn't larger than our | |||
| 2294 | /// preferred vector width. | |||
| 2295 | EVT X86TargetLowering::getOptimalMemOpType( | |||
| 2296 | const MemOp &Op, const AttributeList &FuncAttributes) const { | |||
| 2297 | if (!FuncAttributes.hasFnAttribute(Attribute::NoImplicitFloat)) { | |||
| 2298 | if (Op.size() >= 16 && | |||
| 2299 | (!Subtarget.isUnalignedMem16Slow() || Op.isAligned(Align(16)))) { | |||
| 2300 | // FIXME: Check if unaligned 64-byte accesses are slow. | |||
| 2301 | if (Op.size() >= 64 && Subtarget.hasAVX512() && | |||
| 2302 | (Subtarget.getPreferVectorWidth() >= 512)) { | |||
| 2303 | return Subtarget.hasBWI() ? MVT::v64i8 : MVT::v16i32; | |||
| 2304 | } | |||
| 2305 | // FIXME: Check if unaligned 32-byte accesses are slow. | |||
| 2306 | if (Op.size() >= 32 && Subtarget.hasAVX() && | |||
| 2307 | (Subtarget.getPreferVectorWidth() >= 256)) { | |||
| 2308 | // Although this isn't a well-supported type for AVX1, we'll let | |||
| 2309 | // legalization and shuffle lowering produce the optimal codegen. If we | |||
| 2310 | // choose an optimal type with a vector element larger than a byte, | |||
| 2311 | // getMemsetStores() may create an intermediate splat (using an integer | |||
| 2312 | // multiply) before we splat as a vector. | |||
| 2313 | return MVT::v32i8; | |||
| 2314 | } | |||
| 2315 | if (Subtarget.hasSSE2() && (Subtarget.getPreferVectorWidth() >= 128)) | |||
| 2316 | return MVT::v16i8; | |||
| 2317 | // TODO: Can SSE1 handle a byte vector? | |||
| 2318 | // If we have SSE1 registers we should be able to use them. | |||
| 2319 | if (Subtarget.hasSSE1() && (Subtarget.is64Bit() || Subtarget.hasX87()) && | |||
| 2320 | (Subtarget.getPreferVectorWidth() >= 128)) | |||
| 2321 | return MVT::v4f32; | |||
| 2322 | } else if (((Op.isMemcpy() && !Op.isMemcpyStrSrc()) || Op.isZeroMemset()) && | |||
| 2323 | Op.size() >= 8 && !Subtarget.is64Bit() && Subtarget.hasSSE2()) { | |||
| 2324 | // Do not use f64 to lower memcpy if source is string constant. It's | |||
| 2325 | // better to use i32 to avoid the loads. | |||
| 2326 | // Also, do not use f64 to lower memset unless this is a memset of zeros. | |||
| 2327 | // The gymnastics of splatting a byte value into an XMM register and then | |||
| 2328 | // only using 8-byte stores (because this is a CPU with slow unaligned | |||
| 2329 | // 16-byte accesses) makes that a loser. | |||
| 2330 | return MVT::f64; | |||
| 2331 | } | |||
| 2332 | } | |||
| 2333 | // This is a compromise. If we reach here, unaligned accesses may be slow on | |||
| 2334 | // this target. However, creating smaller, aligned accesses could be even | |||
| 2335 | // slower and would certainly be a lot more code. | |||
| 2336 | if (Subtarget.is64Bit() && Op.size() >= 8) | |||
| 2337 | return MVT::i64; | |||
| 2338 | return MVT::i32; | |||
| 2339 | } | |||
| 2340 | ||||
| 2341 | bool X86TargetLowering::isSafeMemOpType(MVT VT) const { | |||
| 2342 | if (VT == MVT::f32) | |||
| 2343 | return X86ScalarSSEf32; | |||
| 2344 | if (VT == MVT::f64) | |||
| 2345 | return X86ScalarSSEf64; | |||
| 2346 | return true; | |||
| 2347 | } | |||
| 2348 | ||||
| 2349 | bool X86TargetLowering::allowsMisalignedMemoryAccesses( | |||
| 2350 | EVT VT, unsigned, Align Alignment, MachineMemOperand::Flags Flags, | |||
| 2351 | bool *Fast) const { | |||
| 2352 | if (Fast) { | |||
| 2353 | switch (VT.getSizeInBits()) { | |||
| 2354 | default: | |||
| 2355 | // 8-byte and under are always assumed to be fast. | |||
| 2356 | *Fast = true; | |||
| 2357 | break; | |||
| 2358 | case 128: | |||
| 2359 | *Fast = !Subtarget.isUnalignedMem16Slow(); | |||
| 2360 | break; | |||
| 2361 | case 256: | |||
| 2362 | *Fast = !Subtarget.isUnalignedMem32Slow(); | |||
| 2363 | break; | |||
| 2364 | // TODO: What about AVX-512 (512-bit) accesses? | |||
| 2365 | } | |||
| 2366 | } | |||
| 2367 | // NonTemporal vector memory ops must be aligned. | |||
| 2368 | if (!!(Flags & MachineMemOperand::MONonTemporal) && VT.isVector()) { | |||
| 2369 | // NT loads can only be vector aligned, so if its less aligned than the | |||
| 2370 | // minimum vector size (which we can split the vector down to), we might as | |||
| 2371 | // well use a regular unaligned vector load. | |||
| 2372 | // We don't have any NT loads pre-SSE41. | |||
| 2373 | if (!!(Flags & MachineMemOperand::MOLoad)) | |||
| 2374 | return (Alignment < 16 || !Subtarget.hasSSE41()); | |||
| 2375 | return false; | |||
| 2376 | } | |||
| 2377 | // Misaligned accesses of any size are always allowed. | |||
| 2378 | return true; | |||
| 2379 | } | |||
| 2380 | ||||
| 2381 | /// Return the entry encoding for a jump table in the | |||
| 2382 | /// current function. The returned value is a member of the | |||
| 2383 | /// MachineJumpTableInfo::JTEntryKind enum. | |||
| 2384 | unsigned X86TargetLowering::getJumpTableEncoding() const { | |||
| 2385 | // In GOT pic mode, each entry in the jump table is emitted as a @GOTOFF | |||
| 2386 | // symbol. | |||
| 2387 | if (isPositionIndependent() && Subtarget.isPICStyleGOT()) | |||
| 2388 | return MachineJumpTableInfo::EK_Custom32; | |||
| 2389 | ||||
| 2390 | // Otherwise, use the normal jump table encoding heuristics. | |||
| 2391 | return TargetLowering::getJumpTableEncoding(); | |||
| 2392 | } | |||
| 2393 | ||||
| 2394 | bool X86TargetLowering::useSoftFloat() const { | |||
| 2395 | return Subtarget.useSoftFloat(); | |||
| 2396 | } | |||
| 2397 | ||||
| 2398 | void X86TargetLowering::markLibCallAttributes(MachineFunction *MF, unsigned CC, | |||
| 2399 | ArgListTy &Args) const { | |||
| 2400 | ||||
| 2401 | // Only relabel X86-32 for C / Stdcall CCs. | |||
| 2402 | if (Subtarget.is64Bit()) | |||
| 2403 | return; | |||
| 2404 | if (CC != CallingConv::C && CC != CallingConv::X86_StdCall) | |||
| 2405 | return; | |||
| 2406 | unsigned ParamRegs = 0; | |||
| 2407 | if (auto *M = MF->getFunction().getParent()) | |||
| 2408 | ParamRegs = M->getNumberRegisterParameters(); | |||
| 2409 | ||||
| 2410 | // Mark the first N int arguments as having reg | |||
| 2411 | for (auto &Arg : Args) { | |||
| 2412 | Type *T = Arg.Ty; | |||
| 2413 | if (T->isIntOrPtrTy()) | |||
| 2414 | if (MF->getDataLayout().getTypeAllocSize(T) <= 8) { | |||
| 2415 | unsigned numRegs = 1; | |||
| 2416 | if (MF->getDataLayout().getTypeAllocSize(T) > 4) | |||
| 2417 | numRegs = 2; | |||
| 2418 | if (ParamRegs < numRegs) | |||
| 2419 | return; | |||
| 2420 | ParamRegs -= numRegs; | |||
| 2421 | Arg.IsInReg = true; | |||
| 2422 | } | |||
| 2423 | } | |||
| 2424 | } | |||
| 2425 | ||||
| 2426 | const MCExpr * | |||
| 2427 | X86TargetLowering::LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI, | |||
| 2428 | const MachineBasicBlock *MBB, | |||
| 2429 | unsigned uid,MCContext &Ctx) const{ | |||
| 2430 | assert(isPositionIndependent() && Subtarget.isPICStyleGOT())((void)0); | |||
| 2431 | // In 32-bit ELF systems, our jump table entries are formed with @GOTOFF | |||
| 2432 | // entries. | |||
| 2433 | return MCSymbolRefExpr::create(MBB->getSymbol(), | |||
| 2434 | MCSymbolRefExpr::VK_GOTOFF, Ctx); | |||
| 2435 | } | |||
| 2436 | ||||
| 2437 | /// Returns relocation base for the given PIC jumptable. | |||
| 2438 | SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table, | |||
| 2439 | SelectionDAG &DAG) const { | |||
| 2440 | if (!Subtarget.is64Bit()) | |||
| 2441 | // This doesn't have SDLoc associated with it, but is not really the | |||
| 2442 | // same as a Register. | |||
| 2443 | return DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), | |||
| 2444 | getPointerTy(DAG.getDataLayout())); | |||
| 2445 | return Table; | |||
| 2446 | } | |||
| 2447 | ||||
| 2448 | /// This returns the relocation base for the given PIC jumptable, | |||
| 2449 | /// the same as getPICJumpTableRelocBase, but as an MCExpr. | |||
| 2450 | const MCExpr *X86TargetLowering:: | |||
| 2451 | getPICJumpTableRelocBaseExpr(const MachineFunction *MF, unsigned JTI, | |||
| 2452 | MCContext &Ctx) const { | |||
| 2453 | // X86-64 uses RIP relative addressing based on the jump table label. | |||
| 2454 | if (Subtarget.isPICStyleRIPRel()) | |||
| 2455 | return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx); | |||
| 2456 | ||||
| 2457 | // Otherwise, the reference is relative to the PIC base. | |||
| 2458 | return MCSymbolRefExpr::create(MF->getPICBaseSymbol(), Ctx); | |||
| 2459 | } | |||
| 2460 | ||||
| 2461 | std::pair<const TargetRegisterClass *, uint8_t> | |||
| 2462 | X86TargetLowering::findRepresentativeClass(const TargetRegisterInfo *TRI, | |||
| 2463 | MVT VT) const { | |||
| 2464 | const TargetRegisterClass *RRC = nullptr; | |||
| 2465 | uint8_t Cost = 1; | |||
| 2466 | switch (VT.SimpleTy) { | |||
| 2467 | default: | |||
| 2468 | return TargetLowering::findRepresentativeClass(TRI, VT); | |||
| 2469 | case MVT::i8: case MVT::i16: case MVT::i32: case MVT::i64: | |||
| 2470 | RRC = Subtarget.is64Bit() ? &X86::GR64RegClass : &X86::GR32RegClass; | |||
| 2471 | break; | |||
| 2472 | case MVT::x86mmx: | |||
| 2473 | RRC = &X86::VR64RegClass; | |||
| 2474 | break; | |||
| 2475 | case MVT::f32: case MVT::f64: | |||
| 2476 | case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64: | |||
| 2477 | case MVT::v4f32: case MVT::v2f64: | |||
| 2478 | case MVT::v32i8: case MVT::v16i16: case MVT::v8i32: case MVT::v4i64: | |||
| 2479 | case MVT::v8f32: case MVT::v4f64: | |||
| 2480 | case MVT::v64i8: case MVT::v32i16: case MVT::v16i32: case MVT::v8i64: | |||
| 2481 | case MVT::v16f32: case MVT::v8f64: | |||
| 2482 | RRC = &X86::VR128XRegClass; | |||
| 2483 | break; | |||
| 2484 | } | |||
| 2485 | return std::make_pair(RRC, Cost); | |||
| 2486 | } | |||
| 2487 | ||||
| 2488 | unsigned X86TargetLowering::getAddressSpace() const { | |||
| 2489 | if (Subtarget.is64Bit()) | |||
| 2490 | return (getTargetMachine().getCodeModel() == CodeModel::Kernel) ? 256 : 257; | |||
| 2491 | return 256; | |||
| 2492 | } | |||
| 2493 | ||||
| 2494 | static bool hasStackGuardSlotTLS(const Triple &TargetTriple) { | |||
| 2495 | return TargetTriple.isOSGlibc() || TargetTriple.isOSFuchsia() || | |||
| 2496 | (TargetTriple.isAndroid() && !TargetTriple.isAndroidVersionLT(17)); | |||
| 2497 | } | |||
| 2498 | ||||
| 2499 | static Constant* SegmentOffset(IRBuilderBase &IRB, | |||
| 2500 | int Offset, unsigned AddressSpace) { | |||
| 2501 | return ConstantExpr::getIntToPtr( | |||
| 2502 | ConstantInt::get(Type::getInt32Ty(IRB.getContext()), Offset), | |||
| 2503 | Type::getInt8PtrTy(IRB.getContext())->getPointerTo(AddressSpace)); | |||
| 2504 | } | |||
| 2505 | ||||
| 2506 | Value *X86TargetLowering::getIRStackGuard(IRBuilderBase &IRB) const { | |||
| 2507 | // glibc, bionic, and Fuchsia have a special slot for the stack guard in | |||
| 2508 | // tcbhead_t; use it instead of the usual global variable (see | |||
| 2509 | // sysdeps/{i386,x86_64}/nptl/tls.h) | |||
| 2510 | if (hasStackGuardSlotTLS(Subtarget.getTargetTriple())) { | |||
| 2511 | if (Subtarget.isTargetFuchsia()) { | |||
| 2512 | // <zircon/tls.h> defines ZX_TLS_STACK_GUARD_OFFSET with this value. | |||
| 2513 | return SegmentOffset(IRB, 0x10, getAddressSpace()); | |||
| 2514 | } else { | |||
| 2515 | unsigned AddressSpace = getAddressSpace(); | |||
| 2516 | Module *M = IRB.GetInsertBlock()->getParent()->getParent(); | |||
| 2517 | // Specially, some users may customize the base reg and offset. | |||
| 2518 | int Offset = M->getStackProtectorGuardOffset(); | |||
| 2519 | // If we don't set -stack-protector-guard-offset value: | |||
| 2520 | // %fs:0x28, unless we're using a Kernel code model, in which case | |||
| 2521 | // it's %gs:0x28. gs:0x14 on i386. | |||
| 2522 | if (Offset == INT_MAX2147483647) | |||
| 2523 | Offset = (Subtarget.is64Bit()) ? 0x28 : 0x14; | |||
| 2524 | ||||
| 2525 | StringRef GuardReg = M->getStackProtectorGuardReg(); | |||
| 2526 | if (GuardReg == "fs") | |||
| 2527 | AddressSpace = X86AS::FS; | |||
| 2528 | else if (GuardReg == "gs") | |||
| 2529 | AddressSpace = X86AS::GS; | |||
| 2530 | return SegmentOffset(IRB, Offset, AddressSpace); | |||
| 2531 | } | |||
| 2532 | } | |||
| 2533 | return TargetLowering::getIRStackGuard(IRB); | |||
| 2534 | } | |||
| 2535 | ||||
| 2536 | void X86TargetLowering::insertSSPDeclarations(Module &M) const { | |||
| 2537 | // MSVC CRT provides functionalities for stack protection. | |||
| 2538 | if (Subtarget.getTargetTriple().isWindowsMSVCEnvironment() || | |||
| 2539 | Subtarget.getTargetTriple().isWindowsItaniumEnvironment()) { | |||
| 2540 | // MSVC CRT has a global variable holding security cookie. | |||
| 2541 | M.getOrInsertGlobal("__security_cookie", | |||
| 2542 | Type::getInt8PtrTy(M.getContext())); | |||
| 2543 | ||||
| 2544 | // MSVC CRT has a function to validate security cookie. | |||
| 2545 | FunctionCallee SecurityCheckCookie = M.getOrInsertFunction( | |||
| 2546 | "__security_check_cookie", Type::getVoidTy(M.getContext()), | |||
| 2547 | Type::getInt8PtrTy(M.getContext())); | |||
| 2548 | if (Function *F = dyn_cast<Function>(SecurityCheckCookie.getCallee())) { | |||
| 2549 | F->setCallingConv(CallingConv::X86_FastCall); | |||
| 2550 | F->addAttribute(1, Attribute::AttrKind::InReg); | |||
| 2551 | } | |||
| 2552 | return; | |||
| 2553 | } | |||
| 2554 | ||||
| 2555 | StringRef GuardMode = M.getStackProtectorGuard(); | |||
| 2556 | ||||
| 2557 | // glibc, bionic, and Fuchsia have a special slot for the stack guard. | |||
| 2558 | if ((GuardMode == "tls" || GuardMode.empty()) && | |||
| 2559 | hasStackGuardSlotTLS(Subtarget.getTargetTriple())) | |||
| 2560 | return; | |||
| 2561 | TargetLowering::insertSSPDeclarations(M); | |||
| 2562 | } | |||
| 2563 | ||||
| 2564 | Value *X86TargetLowering::getSDagStackGuard(const Module &M) const { | |||
| 2565 | // MSVC CRT has a global variable holding security cookie. | |||
| 2566 | if (Subtarget.getTargetTriple().isWindowsMSVCEnvironment() || | |||
| 2567 | Subtarget.getTargetTriple().isWindowsItaniumEnvironment()) { | |||
| 2568 | return M.getGlobalVariable("__security_cookie"); | |||
| 2569 | } | |||
| 2570 | return TargetLowering::getSDagStackGuard(M); | |||
| 2571 | } | |||
| 2572 | ||||
| 2573 | Function *X86TargetLowering::getSSPStackGuardCheck(const Module &M) const { | |||
| 2574 | // MSVC CRT has a function to validate security cookie. | |||
| 2575 | if (Subtarget.getTargetTriple().isWindowsMSVCEnvironment() || | |||
| 2576 | Subtarget.getTargetTriple().isWindowsItaniumEnvironment()) { | |||
| 2577 | return M.getFunction("__security_check_cookie"); | |||
| 2578 | } | |||
| 2579 | return TargetLowering::getSSPStackGuardCheck(M); | |||
| 2580 | } | |||
| 2581 | ||||
| 2582 | Value * | |||
| 2583 | X86TargetLowering::getSafeStackPointerLocation(IRBuilderBase &IRB) const { | |||
| 2584 | if (Subtarget.getTargetTriple().isOSContiki()) | |||
| 2585 | return getDefaultSafeStackPointerLocation(IRB, false); | |||
| 2586 | ||||
| 2587 | // Android provides a fixed TLS slot for the SafeStack pointer. See the | |||
| 2588 | // definition of TLS_SLOT_SAFESTACK in | |||
| 2589 | // https://android.googlesource.com/platform/bionic/+/master/libc/private/bionic_tls.h | |||
| 2590 | if (Subtarget.isTargetAndroid()) { | |||
| 2591 | // %fs:0x48, unless we're using a Kernel code model, in which case it's %gs: | |||
| 2592 | // %gs:0x24 on i386 | |||
| 2593 | int Offset = (Subtarget.is64Bit()) ? 0x48 : 0x24; | |||
| 2594 | return SegmentOffset(IRB, Offset, getAddressSpace()); | |||
| 2595 | } | |||
| 2596 | ||||
| 2597 | // Fuchsia is similar. | |||
| 2598 | if (Subtarget.isTargetFuchsia()) { | |||
| 2599 | // <zircon/tls.h> defines ZX_TLS_UNSAFE_SP_OFFSET with this value. | |||
| 2600 | return SegmentOffset(IRB, 0x18, getAddressSpace()); | |||
| 2601 | } | |||
| 2602 | ||||
| 2603 | return TargetLowering::getSafeStackPointerLocation(IRB); | |||
| 2604 | } | |||
| 2605 | ||||
| 2606 | //===----------------------------------------------------------------------===// | |||
| 2607 | // Return Value Calling Convention Implementation | |||
| 2608 | //===----------------------------------------------------------------------===// | |||
| 2609 | ||||
| 2610 | bool X86TargetLowering::CanLowerReturn( | |||
| 2611 | CallingConv::ID CallConv, MachineFunction &MF, bool isVarArg, | |||
| 2612 | const SmallVectorImpl<ISD::OutputArg> &Outs, LLVMContext &Context) const { | |||
| 2613 | SmallVector<CCValAssign, 16> RVLocs; | |||
| 2614 | CCState CCInfo(CallConv, isVarArg, MF, RVLocs, Context); | |||
| 2615 | return CCInfo.CheckReturn(Outs, RetCC_X86); | |||
| 2616 | } | |||
| 2617 | ||||
| 2618 | const MCPhysReg *X86TargetLowering::getScratchRegisters(CallingConv::ID) const { | |||
| 2619 | static const MCPhysReg ScratchRegs[] = { X86::R11, 0 }; | |||
| 2620 | return ScratchRegs; | |||
| 2621 | } | |||
| 2622 | ||||
| 2623 | /// Lowers masks values (v*i1) to the local register values | |||
| 2624 | /// \returns DAG node after lowering to register type | |||
| 2625 | static SDValue lowerMasksToReg(const SDValue &ValArg, const EVT &ValLoc, | |||
| 2626 | const SDLoc &Dl, SelectionDAG &DAG) { | |||
| 2627 | EVT ValVT = ValArg.getValueType(); | |||
| 2628 | ||||
| 2629 | if (ValVT == MVT::v1i1) | |||
| 2630 | return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, Dl, ValLoc, ValArg, | |||
| 2631 | DAG.getIntPtrConstant(0, Dl)); | |||
| 2632 | ||||
| 2633 | if ((ValVT == MVT::v8i1 && (ValLoc == MVT::i8 || ValLoc == MVT::i32)) || | |||
| 2634 | (ValVT == MVT::v16i1 && (ValLoc == MVT::i16 || ValLoc == MVT::i32))) { | |||
| 2635 | // Two stage lowering might be required | |||
| 2636 | // bitcast: v8i1 -> i8 / v16i1 -> i16 | |||
| 2637 | // anyextend: i8 -> i32 / i16 -> i32 | |||
| 2638 | EVT TempValLoc = ValVT == MVT::v8i1 ? MVT::i8 : MVT::i16; | |||
| 2639 | SDValue ValToCopy = DAG.getBitcast(TempValLoc, ValArg); | |||
| 2640 | if (ValLoc == MVT::i32) | |||
| 2641 | ValToCopy = DAG.getNode(ISD::ANY_EXTEND, Dl, ValLoc, ValToCopy); | |||
| 2642 | return ValToCopy; | |||
| 2643 | } | |||
| 2644 | ||||
| 2645 | if ((ValVT == MVT::v32i1 && ValLoc == MVT::i32) || | |||
| 2646 | (ValVT == MVT::v64i1 && ValLoc == MVT::i64)) { | |||
| 2647 | // One stage lowering is required | |||
| 2648 | // bitcast: v32i1 -> i32 / v64i1 -> i64 | |||
| 2649 | return DAG.getBitcast(ValLoc, ValArg); | |||
| 2650 | } | |||
| 2651 | ||||
| 2652 | return DAG.getNode(ISD::ANY_EXTEND, Dl, ValLoc, ValArg); | |||
| 2653 | } | |||
| 2654 | ||||
| 2655 | /// Breaks v64i1 value into two registers and adds the new node to the DAG | |||
| 2656 | static void Passv64i1ArgInRegs( | |||
| 2657 | const SDLoc &Dl, SelectionDAG &DAG, SDValue &Arg, | |||
| 2658 | SmallVectorImpl<std::pair<Register, SDValue>> &RegsToPass, CCValAssign &VA, | |||
| 2659 | CCValAssign &NextVA, const X86Subtarget &Subtarget) { | |||
| 2660 | assert(Subtarget.hasBWI() && "Expected AVX512BW target!")((void)0); | |||
| 2661 | assert(Subtarget.is32Bit() && "Expecting 32 bit target")((void)0); | |||
| 2662 | assert(Arg.getValueType() == MVT::i64 && "Expecting 64 bit value")((void)0); | |||
| 2663 | assert(VA.isRegLoc() && NextVA.isRegLoc() &&((void)0) | |||
| 2664 | "The value should reside in two registers")((void)0); | |||
| 2665 | ||||
| 2666 | // Before splitting the value we cast it to i64 | |||
| 2667 | Arg = DAG.getBitcast(MVT::i64, Arg); | |||
| 2668 | ||||
| 2669 | // Splitting the value into two i32 types | |||
| 2670 | SDValue Lo, Hi; | |||
| 2671 | Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, Dl, MVT::i32, Arg, | |||
| 2672 | DAG.getConstant(0, Dl, MVT::i32)); | |||
| 2673 | Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, Dl, MVT::i32, Arg, | |||
| 2674 | DAG.getConstant(1, Dl, MVT::i32)); | |||
| 2675 | ||||
| 2676 | // Attach the two i32 types into corresponding registers | |||
| 2677 | RegsToPass.push_back(std::make_pair(VA.getLocReg(), Lo)); | |||
| 2678 | RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), Hi)); | |||
| 2679 | } | |||
| 2680 | ||||
| 2681 | SDValue | |||
| 2682 | X86TargetLowering::LowerReturn(SDValue Chain, CallingConv::ID CallConv, | |||
| 2683 | bool isVarArg, | |||
| 2684 | const SmallVectorImpl<ISD::OutputArg> &Outs, | |||
| 2685 | const SmallVectorImpl<SDValue> &OutVals, | |||
| 2686 | const SDLoc &dl, SelectionDAG &DAG) const { | |||
| 2687 | MachineFunction &MF = DAG.getMachineFunction(); | |||
| 2688 | X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>(); | |||
| 2689 | ||||
| 2690 | // In some cases we need to disable registers from the default CSR list. | |||
| 2691 | // For example, when they are used for argument passing. | |||
| 2692 | bool ShouldDisableCalleeSavedRegister = | |||
| 2693 | CallConv == CallingConv::X86_RegCall || | |||
| 2694 | MF.getFunction().hasFnAttribute("no_caller_saved_registers"); | |||
| 2695 | ||||
| 2696 | if (CallConv == CallingConv::X86_INTR && !Outs.empty()) | |||
| 2697 | report_fatal_error("X86 interrupts may not return any value"); | |||
| 2698 | ||||
| 2699 | SmallVector<CCValAssign, 16> RVLocs; | |||
| 2700 | CCState CCInfo(CallConv, isVarArg, MF, RVLocs, *DAG.getContext()); | |||
| 2701 | CCInfo.AnalyzeReturn(Outs, RetCC_X86); | |||
| 2702 | ||||
| 2703 | SmallVector<std::pair<Register, SDValue>, 4> RetVals; | |||
| 2704 | for (unsigned I = 0, OutsIndex = 0, E = RVLocs.size(); I != E; | |||
| 2705 | ++I, ++OutsIndex) { | |||
| 2706 | CCValAssign &VA = RVLocs[I]; | |||
| 2707 | assert(VA.isRegLoc() && "Can only return in registers!")((void)0); | |||
| 2708 | ||||
| 2709 | // Add the register to the CalleeSaveDisableRegs list. | |||
| 2710 | if (ShouldDisableCalleeSavedRegister) | |||
| 2711 | MF.getRegInfo().disableCalleeSavedRegister(VA.getLocReg()); | |||
| 2712 | ||||
| 2713 | SDValue ValToCopy = OutVals[OutsIndex]; | |||
| 2714 | EVT ValVT = ValToCopy.getValueType(); | |||
| 2715 | ||||
| 2716 | // Promote values to the appropriate types. | |||
| 2717 | if (VA.getLocInfo() == CCValAssign::SExt) | |||
| 2718 | ValToCopy = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), ValToCopy); | |||
| 2719 | else if (VA.getLocInfo() == CCValAssign::ZExt) | |||
| 2720 | ValToCopy = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), ValToCopy); | |||
| 2721 | else if (VA.getLocInfo() == CCValAssign::AExt) { | |||
| 2722 | if (ValVT.isVector() && ValVT.getVectorElementType() == MVT::i1) | |||
| 2723 | ValToCopy = lowerMasksToReg(ValToCopy, VA.getLocVT(), dl, DAG); | |||
| 2724 | else | |||
| 2725 | ValToCopy = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), ValToCopy); | |||
| 2726 | } | |||
| 2727 | else if (VA.getLocInfo() == CCValAssign::BCvt) | |||
| 2728 | ValToCopy = DAG.getBitcast(VA.getLocVT(), ValToCopy); | |||
| 2729 | ||||
| 2730 | assert(VA.getLocInfo() != CCValAssign::FPExt &&((void)0) | |||
| 2731 | "Unexpected FP-extend for return value.")((void)0); | |||
| 2732 | ||||
| 2733 | // Report an error if we have attempted to return a value via an XMM | |||
| 2734 | // register and SSE was disabled. | |||
| 2735 | if (!Subtarget.hasSSE1() && X86::FR32XRegClass.contains(VA.getLocReg())) { | |||
| 2736 | errorUnsupported(DAG, dl, "SSE register return with SSE disabled"); | |||
| 2737 | VA.convertToReg(X86::FP0); // Set reg to FP0, avoid hitting asserts. | |||
| 2738 | } else if (!Subtarget.hasSSE2() && | |||
| 2739 | X86::FR64XRegClass.contains(VA.getLocReg()) && | |||
| 2740 | ValVT == MVT::f64) { | |||
| 2741 | // When returning a double via an XMM register, report an error if SSE2 is | |||
| 2742 | // not enabled. | |||
| 2743 | errorUnsupported(DAG, dl, "SSE2 register return with SSE2 disabled"); | |||
| 2744 | VA.convertToReg(X86::FP0); // Set reg to FP0, avoid hitting asserts. | |||
| 2745 | } | |||
| 2746 | ||||
| 2747 | // Returns in ST0/ST1 are handled specially: these are pushed as operands to | |||
| 2748 | // the RET instruction and handled by the FP Stackifier. | |||
| 2749 | if (VA.getLocReg() == X86::FP0 || | |||
| 2750 | VA.getLocReg() == X86::FP1) { | |||
| 2751 | // If this is a copy from an xmm register to ST(0), use an FPExtend to | |||
| 2752 | // change the value to the FP stack register class. | |||
| 2753 | if (isScalarFPTypeInSSEReg(VA.getValVT())) | |||
| 2754 | ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy); | |||
| 2755 | RetVals.push_back(std::make_pair(VA.getLocReg(), ValToCopy)); | |||
| 2756 | // Don't emit a copytoreg. | |||
| 2757 | continue; | |||
| 2758 | } | |||
| 2759 | ||||
| 2760 | // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64 | |||
| 2761 | // which is returned in RAX / RDX. | |||
| 2762 | if (Subtarget.is64Bit()) { | |||
| 2763 | if (ValVT == MVT::x86mmx) { | |||
| 2764 | if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) { | |||
| 2765 | ValToCopy = DAG.getBitcast(MVT::i64, ValToCopy); | |||
| 2766 | ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, | |||
| 2767 | ValToCopy); | |||
| 2768 | // If we don't have SSE2 available, convert to v4f32 so the generated | |||
| 2769 | // register is legal. | |||
| 2770 | if (!Subtarget.hasSSE2()) | |||
| 2771 | ValToCopy = DAG.getBitcast(MVT::v4f32, ValToCopy); | |||
| 2772 | } | |||
| 2773 | } | |||
| 2774 | } | |||
| 2775 | ||||
| 2776 | if (VA.needsCustom()) { | |||
| 2777 | assert(VA.getValVT() == MVT::v64i1 &&((void)0) | |||
| 2778 | "Currently the only custom case is when we split v64i1 to 2 regs")((void)0); | |||
| 2779 | ||||
| 2780 | Passv64i1ArgInRegs(dl, DAG, ValToCopy, RetVals, VA, RVLocs[++I], | |||
| 2781 | Subtarget); | |||
| 2782 | ||||
| 2783 | // Add the second register to the CalleeSaveDisableRegs list. | |||
| 2784 | if (ShouldDisableCalleeSavedRegister) | |||
| 2785 | MF.getRegInfo().disableCalleeSavedRegister(RVLocs[I].getLocReg()); | |||
| 2786 | } else { | |||
| 2787 | RetVals.push_back(std::make_pair(VA.getLocReg(), ValToCopy)); | |||
| 2788 | } | |||
| 2789 | } | |||
| 2790 | ||||
| 2791 | SDValue Flag; | |||
| 2792 | SmallVector<SDValue, 6> RetOps; | |||
| 2793 | RetOps.push_back(Chain); // Operand #0 = Chain (updated below) | |||
| 2794 | // Operand #1 = Bytes To Pop | |||
| 2795 | RetOps.push_back(DAG.getTargetConstant(FuncInfo->getBytesToPopOnReturn(), dl, | |||
| 2796 | MVT::i32)); | |||
| 2797 | ||||
| 2798 | // Copy the result values into the output registers. | |||
| 2799 | for (auto &RetVal : RetVals) { | |||
| 2800 | if (RetVal.first == X86::FP0 || RetVal.first == X86::FP1) { | |||
| 2801 | RetOps.push_back(RetVal.second); | |||
| 2802 | continue; // Don't emit a copytoreg. | |||
| 2803 | } | |||
| 2804 | ||||
| 2805 | Chain = DAG.getCopyToReg(Chain, dl, RetVal.first, RetVal.second, Flag); | |||
| 2806 | Flag = Chain.getValue(1); | |||
| 2807 | RetOps.push_back( | |||
| 2808 | DAG.getRegister(RetVal.first, RetVal.second.getValueType())); | |||
| 2809 | } | |||
| 2810 | ||||
| 2811 | // Swift calling convention does not require we copy the sret argument | |||
| 2812 | // into %rax/%eax for the return, and SRetReturnReg is not set for Swift. | |||
| 2813 | ||||
| 2814 | // All x86 ABIs require that for returning structs by value we copy | |||
| 2815 | // the sret argument into %rax/%eax (depending on ABI) for the return. | |||
| 2816 | // We saved the argument into a virtual register in the entry block, | |||
| 2817 | // so now we copy the value out and into %rax/%eax. | |||
| 2818 | // | |||
| 2819 | // Checking Function.hasStructRetAttr() here is insufficient because the IR | |||
| 2820 | // may not have an explicit sret argument. If FuncInfo.CanLowerReturn is | |||
| 2821 | // false, then an sret argument may be implicitly inserted in the SelDAG. In | |||
| 2822 | // either case FuncInfo->setSRetReturnReg() will have been called. | |||
| 2823 | if (Register SRetReg = FuncInfo->getSRetReturnReg()) { | |||
| 2824 | // When we have both sret and another return value, we should use the | |||
| 2825 | // original Chain stored in RetOps[0], instead of the current Chain updated | |||
| 2826 | // in the above loop. If we only have sret, RetOps[0] equals to Chain. | |||
| 2827 | ||||
| 2828 | // For the case of sret and another return value, we have | |||
| 2829 | // Chain_0 at the function entry | |||
| 2830 | // Chain_1 = getCopyToReg(Chain_0) in the above loop | |||
| 2831 | // If we use Chain_1 in getCopyFromReg, we will have | |||
| 2832 | // Val = getCopyFromReg(Chain_1) | |||
| 2833 | // Chain_2 = getCopyToReg(Chain_1, Val) from below | |||
| 2834 | ||||
| 2835 | // getCopyToReg(Chain_0) will be glued together with | |||
| 2836 | // getCopyToReg(Chain_1, Val) into Unit A, getCopyFromReg(Chain_1) will be | |||
| 2837 | // in Unit B, and we will have cyclic dependency between Unit A and Unit B: | |||
| 2838 | // Data dependency from Unit B to Unit A due to usage of Val in | |||
| 2839 | // getCopyToReg(Chain_1, Val) | |||
| 2840 | // Chain dependency from Unit A to Unit B | |||
| 2841 | ||||
| 2842 | // So here, we use RetOps[0] (i.e Chain_0) for getCopyFromReg. | |||
| 2843 | SDValue Val = DAG.getCopyFromReg(RetOps[0], dl, SRetReg, | |||
| 2844 | getPointerTy(MF.getDataLayout())); | |||
| 2845 | ||||
| 2846 | Register RetValReg | |||
| 2847 | = (Subtarget.is64Bit() && !Subtarget.isTarget64BitILP32()) ? | |||
| 2848 | X86::RAX : X86::EAX; | |||
| 2849 | Chain = DAG.getCopyToReg(Chain, dl, RetValReg, Val, Flag); | |||
| 2850 | Flag = Chain.getValue(1); | |||
| 2851 | ||||
| 2852 | // RAX/EAX now acts like a return value. | |||
| 2853 | RetOps.push_back( | |||
| 2854 | DAG.getRegister(RetValReg, getPointerTy(DAG.getDataLayout()))); | |||
| 2855 | ||||
| 2856 | // Add the returned register to the CalleeSaveDisableRegs list. | |||
| 2857 | if (ShouldDisableCalleeSavedRegister) | |||
| 2858 | MF.getRegInfo().disableCalleeSavedRegister(RetValReg); | |||
| 2859 | } | |||
| 2860 | ||||
| 2861 | const X86RegisterInfo *TRI = Subtarget.getRegisterInfo(); | |||
| 2862 | const MCPhysReg *I = | |||
| 2863 | TRI->getCalleeSavedRegsViaCopy(&DAG.getMachineFunction()); | |||
| 2864 | if (I) { | |||
| 2865 | for (; *I; ++I) { | |||
| 2866 | if (X86::GR64RegClass.contains(*I)) | |||
| 2867 | RetOps.push_back(DAG.getRegister(*I, MVT::i64)); | |||
| 2868 | else | |||
| 2869 | llvm_unreachable("Unexpected register class in CSRsViaCopy!")__builtin_unreachable(); | |||
| 2870 | } | |||
| 2871 | } | |||
| 2872 | ||||
| 2873 | RetOps[0] = Chain; // Update chain. | |||
| 2874 | ||||
| 2875 | // Add the flag if we have it. | |||
| 2876 | if (Flag.getNode()) | |||
| 2877 | RetOps.push_back(Flag); | |||
| 2878 | ||||
| 2879 | X86ISD::NodeType opcode = X86ISD::RET_FLAG; | |||
| 2880 | if (CallConv == CallingConv::X86_INTR) | |||
| 2881 | opcode = X86ISD::IRET; | |||
| 2882 | return DAG.getNode(opcode, dl, MVT::Other, RetOps); | |||
| 2883 | } | |||
| 2884 | ||||
| 2885 | bool X86TargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const { | |||
| 2886 | if (N->getNumValues() != 1 || !N->hasNUsesOfValue(1, 0)) | |||
| 2887 | return false; | |||
| 2888 | ||||
| 2889 | SDValue TCChain = Chain; | |||
| 2890 | SDNode *Copy = *N->use_begin(); | |||
| 2891 | if (Copy->getOpcode() == ISD::CopyToReg) { | |||
| 2892 | // If the copy has a glue operand, we conservatively assume it isn't safe to | |||
| 2893 | // perform a tail call. | |||
| 2894 | if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue) | |||
| 2895 | return false; | |||
| 2896 | TCChain = Copy->getOperand(0); | |||
| 2897 | } else if (Copy->getOpcode() != ISD::FP_EXTEND) | |||
| 2898 | return false; | |||
| 2899 | ||||
| 2900 | bool HasRet = false; | |||
| 2901 | for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end(); | |||
| 2902 | UI != UE; ++UI) { | |||
| 2903 | if (UI->getOpcode() != X86ISD::RET_FLAG) | |||
| 2904 | return false; | |||
| 2905 | // If we are returning more than one value, we can definitely | |||
| 2906 | // not make a tail call see PR19530 | |||
| 2907 | if (UI->getNumOperands() > 4) | |||
| 2908 | return false; | |||
| 2909 | if (UI->getNumOperands() == 4 && | |||
| 2910 | UI->getOperand(UI->getNumOperands()-1).getValueType() != MVT::Glue) | |||
| 2911 | return false; | |||
| 2912 | HasRet = true; | |||
| 2913 | } | |||
| 2914 | ||||
| 2915 | if (!HasRet) | |||
| 2916 | return false; | |||
| 2917 | ||||
| 2918 | Chain = TCChain; | |||
| 2919 | return true; | |||
| 2920 | } | |||
| 2921 | ||||
| 2922 | EVT X86TargetLowering::getTypeForExtReturn(LLVMContext &Context, EVT VT, | |||
| 2923 | ISD::NodeType ExtendKind) const { | |||
| 2924 | MVT ReturnMVT = MVT::i32; | |||
| 2925 | ||||
| 2926 | bool Darwin = Subtarget.getTargetTriple().isOSDarwin(); | |||
| 2927 | if (VT == MVT::i1 || (!Darwin && (VT == MVT::i8 || VT == MVT::i16))) { | |||
| 2928 | // The ABI does not require i1, i8 or i16 to be extended. | |||
| 2929 | // | |||
| 2930 | // On Darwin, there is code in the wild relying on Clang's old behaviour of | |||
| 2931 | // always extending i8/i16 return values, so keep doing that for now. | |||
| 2932 | // (PR26665). | |||
| 2933 | ReturnMVT = MVT::i8; | |||
| 2934 | } | |||
| 2935 | ||||
| 2936 | EVT MinVT = getRegisterType(Context, ReturnMVT); | |||
| 2937 | return VT.bitsLT(MinVT) ? MinVT : VT; | |||
| 2938 | } | |||
| 2939 | ||||
| 2940 | /// Reads two 32 bit registers and creates a 64 bit mask value. | |||
| 2941 | /// \param VA The current 32 bit value that need to be assigned. | |||
| 2942 | /// \param NextVA The next 32 bit value that need to be assigned. | |||
| 2943 | /// \param Root The parent DAG node. | |||
| 2944 | /// \param [in,out] InFlag Represents SDvalue in the parent DAG node for | |||
| 2945 | /// glue purposes. In the case the DAG is already using | |||
| 2946 | /// physical register instead of virtual, we should glue | |||
| 2947 | /// our new SDValue to InFlag SDvalue. | |||
| 2948 | /// \return a new SDvalue of size 64bit. | |||
| 2949 | static SDValue getv64i1Argument(CCValAssign &VA, CCValAssign &NextVA, | |||
| 2950 | SDValue &Root, SelectionDAG &DAG, | |||
| 2951 | const SDLoc &Dl, const X86Subtarget &Subtarget, | |||
| 2952 | SDValue *InFlag = nullptr) { | |||
| 2953 | assert((Subtarget.hasBWI()) && "Expected AVX512BW target!")((void)0); | |||
| 2954 | assert(Subtarget.is32Bit() && "Expecting 32 bit target")((void)0); | |||
| 2955 | assert(VA.getValVT() == MVT::v64i1 &&((void)0) | |||
| 2956 | "Expecting first location of 64 bit width type")((void)0); | |||
| 2957 | assert(NextVA.getValVT() == VA.getValVT() &&((void)0) | |||
| 2958 | "The locations should have the same type")((void)0); | |||
| 2959 | assert(VA.isRegLoc() && NextVA.isRegLoc() &&((void)0) | |||
| 2960 | "The values should reside in two registers")((void)0); | |||
| 2961 | ||||
| 2962 | SDValue Lo, Hi; | |||
| 2963 | SDValue ArgValueLo, ArgValueHi; | |||
| 2964 | ||||
| 2965 | MachineFunction &MF = DAG.getMachineFunction(); | |||
| 2966 | const TargetRegisterClass *RC = &X86::GR32RegClass; | |||
| 2967 | ||||
| 2968 | // Read a 32 bit value from the registers. | |||
| 2969 | if (nullptr == InFlag) { | |||
| 2970 | // When no physical register is present, | |||
| 2971 | // create an intermediate virtual register. | |||
| 2972 | Register Reg = MF.addLiveIn(VA.getLocReg(), RC); | |||
| 2973 | ArgValueLo = DAG.getCopyFromReg(Root, Dl, Reg, MVT::i32); | |||
| 2974 | Reg = MF.addLiveIn(NextVA.getLocReg(), RC); | |||
| 2975 | ArgValueHi = DAG.getCopyFromReg(Root, Dl, Reg, MVT::i32); | |||
| 2976 | } else { | |||
| 2977 | // When a physical register is available read the value from it and glue | |||
| 2978 | // the reads together. | |||
| 2979 | ArgValueLo = | |||
| 2980 | DAG.getCopyFromReg(Root, Dl, VA.getLocReg(), MVT::i32, *InFlag); | |||
| 2981 | *InFlag = ArgValueLo.getValue(2); | |||
| 2982 | ArgValueHi = | |||
| 2983 | DAG.getCopyFromReg(Root, Dl, NextVA.getLocReg(), MVT::i32, *InFlag); | |||
| 2984 | *InFlag = ArgValueHi.getValue(2); | |||
| 2985 | } | |||
| 2986 | ||||
| 2987 | // Convert the i32 type into v32i1 type. | |||
| 2988 | Lo = DAG.getBitcast(MVT::v32i1, ArgValueLo); | |||
| 2989 | ||||
| 2990 | // Convert the i32 type into v32i1 type. | |||
| 2991 | Hi = DAG.getBitcast(MVT::v32i1, ArgValueHi); | |||
| 2992 | ||||
| 2993 | // Concatenate the two values together. | |||
| 2994 | return DAG.getNode(ISD::CONCAT_VECTORS, Dl, MVT::v64i1, Lo, Hi); | |||
| 2995 | } | |||
| 2996 | ||||
| 2997 | /// The function will lower a register of various sizes (8/16/32/64) | |||
| 2998 | /// to a mask value of the expected size (v8i1/v16i1/v32i1/v64i1) | |||
| 2999 | /// \returns a DAG node contains the operand after lowering to mask type. | |||
| 3000 | static SDValue lowerRegToMasks(const SDValue &ValArg, const EVT &ValVT, | |||
| 3001 | const EVT &ValLoc, const SDLoc &Dl, | |||
| 3002 | SelectionDAG &DAG) { | |||
| 3003 | SDValue ValReturned = ValArg; | |||
| 3004 | ||||
| 3005 | if (ValVT == MVT::v1i1) | |||
| 3006 | return DAG.getNode(ISD::SCALAR_TO_VECTOR, Dl, MVT::v1i1, ValReturned); | |||
| 3007 | ||||
| 3008 | if (ValVT == MVT::v64i1) { | |||
| 3009 | // In 32 bit machine, this case is handled by getv64i1Argument | |||
| 3010 | assert(ValLoc == MVT::i64 && "Expecting only i64 locations")((void)0); | |||
| 3011 | // In 64 bit machine, There is no need to truncate the value only bitcast | |||
| 3012 | } else { | |||
| 3013 | MVT maskLen; | |||
| 3014 | switch (ValVT.getSimpleVT().SimpleTy) { | |||
| 3015 | case MVT::v8i1: | |||
| 3016 | maskLen = MVT::i8; | |||
| 3017 | break; | |||
| 3018 | case MVT::v16i1: | |||
| 3019 | maskLen = MVT::i16; | |||
| 3020 | break; | |||
| 3021 | case MVT::v32i1: | |||
| 3022 | maskLen = MVT::i32; | |||
| 3023 | break; | |||
| 3024 | default: | |||
| 3025 | llvm_unreachable("Expecting a vector of i1 types")__builtin_unreachable(); | |||
| 3026 | } | |||
| 3027 | ||||
| 3028 | ValReturned = DAG.getNode(ISD::TRUNCATE, Dl, maskLen, ValReturned); | |||
| 3029 | } | |||
| 3030 | return DAG.getBitcast(ValVT, ValReturned); | |||
| 3031 | } | |||
| 3032 | ||||
| 3033 | /// Lower the result values of a call into the | |||
| 3034 | /// appropriate copies out of appropriate physical registers. | |||
| 3035 | /// | |||
| 3036 | SDValue X86TargetLowering::LowerCallResult( | |||
| 3037 | SDValue Chain, SDValue InFlag, CallingConv::ID CallConv, bool isVarArg, | |||
| 3038 | const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl, | |||
| 3039 | SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals, | |||
| 3040 | uint32_t *RegMask) const { | |||
| 3041 | ||||
| 3042 | const TargetRegisterInfo *TRI = Subtarget.getRegisterInfo(); | |||
| 3043 | // Assign locations to each value returned by this call. | |||
| 3044 | SmallVector<CCValAssign, 16> RVLocs; | |||
| 3045 | CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs, | |||
| 3046 | *DAG.getContext()); | |||
| 3047 | CCInfo.AnalyzeCallResult(Ins, RetCC_X86); | |||
| 3048 | ||||
| 3049 | // Copy all of the result registers out of their specified physreg. | |||
| 3050 | for (unsigned I = 0, InsIndex = 0, E = RVLocs.size(); I != E; | |||
| 3051 | ++I, ++InsIndex) { | |||
| 3052 | CCValAssign &VA = RVLocs[I]; | |||
| 3053 | EVT CopyVT = VA.getLocVT(); | |||
| 3054 | ||||
| 3055 | // In some calling conventions we need to remove the used registers | |||
| 3056 | // from the register mask. | |||
| 3057 | if (RegMask) { | |||
| 3058 | for (MCSubRegIterator SubRegs(VA.getLocReg(), TRI, /*IncludeSelf=*/true); | |||
| 3059 | SubRegs.isValid(); ++SubRegs) | |||
| 3060 | RegMask[*SubRegs / 32] &= ~(1u << (*SubRegs % 32)); | |||
| 3061 | } | |||
| 3062 | ||||
| 3063 | // Report an error if there was an attempt to return FP values via XMM | |||
| 3064 | // registers. | |||
| 3065 | if (!Subtarget.hasSSE1() && X86::FR32XRegClass.contains(VA.getLocReg())) { | |||
| 3066 | errorUnsupported(DAG, dl, "SSE register return with SSE disabled"); | |||
| 3067 | if (VA.getLocReg() == X86::XMM1) | |||
| 3068 | VA.convertToReg(X86::FP1); // Set reg to FP1, avoid hitting asserts. | |||
| 3069 | else | |||
| 3070 | VA.convertToReg(X86::FP0); // Set reg to FP0, avoid hitting asserts. | |||
| 3071 | } else if (!Subtarget.hasSSE2() && | |||
| 3072 | X86::FR64XRegClass.contains(VA.getLocReg()) && | |||
| 3073 | CopyVT == MVT::f64) { | |||
| 3074 | errorUnsupported(DAG, dl, "SSE2 register return with SSE2 disabled"); | |||
| 3075 | if (VA.getLocReg() == X86::XMM1) | |||
| 3076 | VA.convertToReg(X86::FP1); // Set reg to FP1, avoid hitting asserts. | |||
| 3077 | else | |||
| 3078 | VA.convertToReg(X86::FP0); // Set reg to FP0, avoid hitting asserts. | |||
| 3079 | } | |||
| 3080 | ||||
| 3081 | // If we prefer to use the value in xmm registers, copy it out as f80 and | |||
| 3082 | // use a truncate to move it from fp stack reg to xmm reg. | |||
| 3083 | bool RoundAfterCopy = false; | |||
| 3084 | if ((VA.getLocReg() == X86::FP0 || VA.getLocReg() == X86::FP1) && | |||
| 3085 | isScalarFPTypeInSSEReg(VA.getValVT())) { | |||
| 3086 | if (!Subtarget.hasX87()) | |||
| 3087 | report_fatal_error("X87 register return with X87 disabled"); | |||
| 3088 | CopyVT = MVT::f80; | |||
| 3089 | RoundAfterCopy = (CopyVT != VA.getLocVT()); | |||
| 3090 | } | |||
| 3091 | ||||
| 3092 | SDValue Val; | |||
| 3093 | if (VA.needsCustom()) { | |||
| 3094 | assert(VA.getValVT() == MVT::v64i1 &&((void)0) | |||
| 3095 | "Currently the only custom case is when we split v64i1 to 2 regs")((void)0); | |||
| 3096 | Val = | |||
| 3097 | getv64i1Argument(VA, RVLocs[++I], Chain, DAG, dl, Subtarget, &InFlag); | |||
| 3098 | } else { | |||
| 3099 | Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), CopyVT, InFlag) | |||
| 3100 | .getValue(1); | |||
| 3101 | Val = Chain.getValue(0); | |||
| 3102 | InFlag = Chain.getValue(2); | |||
| 3103 | } | |||
| 3104 | ||||
| 3105 | if (RoundAfterCopy) | |||
| 3106 | Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val, | |||
| 3107 | // This truncation won't change the value. | |||
| 3108 | DAG.getIntPtrConstant(1, dl)); | |||
| 3109 | ||||
| 3110 | if (VA.isExtInLoc()) { | |||
| 3111 | if (VA.getValVT().isVector() && | |||
| 3112 | VA.getValVT().getScalarType() == MVT::i1 && | |||
| 3113 | ((VA.getLocVT() == MVT::i64) || (VA.getLocVT() == MVT::i32) || | |||
| 3114 | (VA.getLocVT() == MVT::i16) || (VA.getLocVT() == MVT::i8))) { | |||
| 3115 | // promoting a mask type (v*i1) into a register of type i64/i32/i16/i8 | |||
| 3116 | Val = lowerRegToMasks(Val, VA.getValVT(), VA.getLocVT(), dl, DAG); | |||
| 3117 | } else | |||
| 3118 | Val = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), Val); | |||
| 3119 | } | |||
| 3120 | ||||
| 3121 | if (VA.getLocInfo() == CCValAssign::BCvt) | |||
| 3122 | Val = DAG.getBitcast(VA.getValVT(), Val); | |||
| 3123 | ||||
| 3124 | InVals.push_back(Val); | |||
| 3125 | } | |||
| 3126 | ||||
| 3127 | return Chain; | |||
| 3128 | } | |||
| 3129 | ||||
| 3130 | //===----------------------------------------------------------------------===// | |||
| 3131 | // C & StdCall & Fast Calling Convention implementation | |||
| 3132 | //===----------------------------------------------------------------------===// | |||
| 3133 | // StdCall calling convention seems to be standard for many Windows' API | |||
| 3134 | // routines and around. It differs from C calling convention just a little: | |||
| 3135 | // callee should clean up the stack, not caller. Symbols should be also | |||
| 3136 | // decorated in some fancy way :) It doesn't support any vector arguments. | |||
| 3137 | // For info on fast calling convention see Fast Calling Convention (tail call) | |||
| 3138 | // implementation LowerX86_32FastCCCallTo. | |||
| 3139 | ||||
| 3140 | /// CallIsStructReturn - Determines whether a call uses struct return | |||
| 3141 | /// semantics. | |||
| 3142 | enum StructReturnType { | |||
| 3143 | NotStructReturn, | |||
| 3144 | RegStructReturn, | |||
| 3145 | StackStructReturn | |||
| 3146 | }; | |||
| 3147 | static StructReturnType | |||
| 3148 | callIsStructReturn(ArrayRef<ISD::OutputArg> Outs, bool IsMCU) { | |||
| 3149 | if (Outs.empty()) | |||
| 3150 | return NotStructReturn; | |||
| 3151 | ||||
| 3152 | const ISD::ArgFlagsTy &Flags = Outs[0].Flags; | |||
| 3153 | if (!Flags.isSRet()) | |||
| 3154 | return NotStructReturn; | |||
| 3155 | if (Flags.isInReg() || IsMCU) | |||
| 3156 | return RegStructReturn; | |||
| 3157 | return StackStructReturn; | |||
| 3158 | } | |||
| 3159 | ||||
| 3160 | /// Determines whether a function uses struct return semantics. | |||
| 3161 | static StructReturnType | |||
| 3162 | argsAreStructReturn(ArrayRef<ISD::InputArg> Ins, bool IsMCU) { | |||
| 3163 | if (Ins.empty()) | |||
| 3164 | return NotStructReturn; | |||
| 3165 | ||||
| 3166 | const ISD::ArgFlagsTy &Flags = Ins[0].Flags; | |||
| 3167 | if (!Flags.isSRet()) | |||
| 3168 | return NotStructReturn; | |||
| 3169 | if (Flags.isInReg() || IsMCU) | |||
| 3170 | return RegStructReturn; | |||
| 3171 | return StackStructReturn; | |||
| 3172 | } | |||
| 3173 | ||||
| 3174 | /// Make a copy of an aggregate at address specified by "Src" to address | |||
| 3175 | /// "Dst" with size and alignment information specified by the specific | |||
| 3176 | /// parameter attribute. The copy will be passed as a byval function parameter. | |||
| 3177 | static SDValue CreateCopyOfByValArgument(SDValue Src, SDValue Dst, | |||
| 3178 | SDValue Chain, ISD::ArgFlagsTy Flags, | |||
| 3179 | SelectionDAG &DAG, const SDLoc &dl) { | |||
| 3180 | SDValue SizeNode = DAG.getIntPtrConstant(Flags.getByValSize(), dl); | |||
| 3181 | ||||
| 3182 | return DAG.getMemcpy( | |||
| 3183 | Chain, dl, Dst, Src, SizeNode, Flags.getNonZeroByValAlign(), | |||
| 3184 | /*isVolatile*/ false, /*AlwaysInline=*/true, | |||
| 3185 | /*isTailCall*/ false, MachinePointerInfo(), MachinePointerInfo()); | |||
| 3186 | } | |||
| 3187 | ||||
| 3188 | /// Return true if the calling convention is one that we can guarantee TCO for. | |||
| 3189 | static bool canGuaranteeTCO(CallingConv::ID CC) { | |||
| 3190 | return (CC == CallingConv::Fast || CC == CallingConv::GHC || | |||
| 3191 | CC == CallingConv::X86_RegCall || CC == CallingConv::HiPE || | |||
| 3192 | CC == CallingConv::HHVM || CC == CallingConv::Tail || | |||
| 3193 | CC == CallingConv::SwiftTail); | |||
| 3194 | } | |||
| 3195 | ||||
| 3196 | /// Return true if we might ever do TCO for calls with this calling convention. | |||
| 3197 | static bool mayTailCallThisCC(CallingConv::ID CC) { | |||
| 3198 | switch (CC) { | |||
| 3199 | // C calling conventions: | |||
| 3200 | case CallingConv::C: | |||
| 3201 | case CallingConv::Win64: | |||
| 3202 | case CallingConv::X86_64_SysV: | |||
| 3203 | // Callee pop conventions: | |||
| 3204 | case CallingConv::X86_ThisCall: | |||
| 3205 | case CallingConv::X86_StdCall: | |||
| 3206 | case CallingConv::X86_VectorCall: | |||
| 3207 | case CallingConv::X86_FastCall: | |||
| 3208 | // Swift: | |||
| 3209 | case CallingConv::Swift: | |||
| 3210 | return true; | |||
| 3211 | default: | |||
| 3212 | return canGuaranteeTCO(CC); | |||
| 3213 | } | |||
| 3214 | } | |||
| 3215 | ||||
| 3216 | /// Return true if the function is being made into a tailcall target by | |||
| 3217 | /// changing its ABI. | |||
| 3218 | static bool shouldGuaranteeTCO(CallingConv::ID CC, bool GuaranteedTailCallOpt) { | |||
| 3219 | return (GuaranteedTailCallOpt && canGuaranteeTCO(CC)) || | |||
| 3220 | CC == CallingConv::Tail || CC == CallingConv::SwiftTail; | |||
| 3221 | } | |||
| 3222 | ||||
| 3223 | bool X86TargetLowering::mayBeEmittedAsTailCall(const CallInst *CI) const { | |||
| 3224 | if (!CI->isTailCall()) | |||
| 3225 | return false; | |||
| 3226 | ||||
| 3227 | CallingConv::ID CalleeCC = CI->getCallingConv(); | |||
| 3228 | if (!mayTailCallThisCC(CalleeCC)) | |||
| 3229 | return false; | |||
| 3230 | ||||
| 3231 | return true; | |||
| 3232 | } | |||
| 3233 | ||||
| 3234 | SDValue | |||
| 3235 | X86TargetLowering::LowerMemArgument(SDValue Chain, CallingConv::ID CallConv, | |||
| 3236 | const SmallVectorImpl<ISD::InputArg> &Ins, | |||
| 3237 | const SDLoc &dl, SelectionDAG &DAG, | |||
| 3238 | const CCValAssign &VA, | |||
| 3239 | MachineFrameInfo &MFI, unsigned i) const { | |||
| 3240 | // Create the nodes corresponding to a load from this parameter slot. | |||
| 3241 | ISD::ArgFlagsTy Flags = Ins[i].Flags; | |||
| 3242 | bool AlwaysUseMutable = shouldGuaranteeTCO( | |||
| 3243 | CallConv, DAG.getTarget().Options.GuaranteedTailCallOpt); | |||
| 3244 | bool isImmutable = !AlwaysUseMutable && !Flags.isByVal(); | |||
| 3245 | EVT ValVT; | |||
| 3246 | MVT PtrVT = getPointerTy(DAG.getDataLayout()); | |||
| 3247 | ||||
| 3248 | // If value is passed by pointer we have address passed instead of the value | |||
| 3249 | // itself. No need to extend if the mask value and location share the same | |||
| 3250 | // absolute size. | |||
| 3251 | bool ExtendedInMem = | |||
| 3252 | VA.isExtInLoc() && VA.getValVT().getScalarType() == MVT::i1 && | |||
| 3253 | VA.getValVT().getSizeInBits() != VA.getLocVT().getSizeInBits(); | |||
| 3254 | ||||
| 3255 | if (VA.getLocInfo() == CCValAssign::Indirect || ExtendedInMem) | |||
| 3256 | ValVT = VA.getLocVT(); | |||
| 3257 | else | |||
| 3258 | ValVT = VA.getValVT(); | |||
| 3259 | ||||
| 3260 | // FIXME: For now, all byval parameter objects are marked mutable. This can be | |||
| 3261 | // changed with more analysis. | |||
| 3262 | // In case of tail call optimization mark all arguments mutable. Since they | |||
| 3263 | // could be overwritten by lowering of arguments in case of a tail call. | |||
| 3264 | if (Flags.isByVal()) { | |||
| 3265 | unsigned Bytes = Flags.getByValSize(); | |||
| 3266 | if (Bytes == 0) Bytes = 1; // Don't create zero-sized stack objects. | |||
| 3267 | ||||
| 3268 | // FIXME: For now, all byval parameter objects are marked as aliasing. This | |||
| 3269 | // can be improved with deeper analysis. | |||
| 3270 | int FI = MFI.CreateFixedObject(Bytes, VA.getLocMemOffset(), isImmutable, | |||
| 3271 | /*isAliased=*/true); | |||
| 3272 | return DAG.getFrameIndex(FI, PtrVT); | |||
| 3273 | } | |||
| 3274 | ||||
| 3275 | EVT ArgVT = Ins[i].ArgVT; | |||
| 3276 | ||||
| 3277 | // If this is a vector that has been split into multiple parts, and the | |||
| 3278 | // scalar size of the parts don't match the vector element size, then we can't | |||
| 3279 | // elide the copy. The parts will have padding between them instead of being | |||
| 3280 | // packed like a vector. | |||
| 3281 | bool ScalarizedAndExtendedVector = | |||
| 3282 | ArgVT.isVector() && !VA.getLocVT().isVector() && | |||
| 3283 | VA.getLocVT().getSizeInBits() != ArgVT.getScalarSizeInBits(); | |||
| 3284 | ||||
| 3285 | // This is an argument in memory. We might be able to perform copy elision. | |||
| 3286 | // If the argument is passed directly in memory without any extension, then we | |||
| 3287 | // can perform copy elision. Large vector types, for example, may be passed | |||
| 3288 | // indirectly by pointer. | |||
| 3289 | if (Flags.isCopyElisionCandidate() && | |||
| 3290 | VA.getLocInfo() != CCValAssign::Indirect && !ExtendedInMem && | |||
| 3291 | !ScalarizedAndExtendedVector) { | |||
| 3292 | SDValue PartAddr; | |||
| 3293 | if (Ins[i].PartOffset == 0) { | |||
| 3294 | // If this is a one-part value or the first part of a multi-part value, | |||
| 3295 | // create a stack object for the entire argument value type and return a | |||
| 3296 | // load from our portion of it. This assumes that if the first part of an | |||
| 3297 | // argument is in memory, the rest will also be in memory. | |||
| 3298 | int FI = MFI.CreateFixedObject(ArgVT.getStoreSize(), VA.getLocMemOffset(), | |||
| 3299 | /*IsImmutable=*/false); | |||
| 3300 | PartAddr = DAG.getFrameIndex(FI, PtrVT); | |||
| 3301 | return DAG.getLoad( | |||
| 3302 | ValVT, dl, Chain, PartAddr, | |||
| 3303 | MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI)); | |||
| 3304 | } else { | |||
| 3305 | // This is not the first piece of an argument in memory. See if there is | |||
| 3306 | // already a fixed stack object including this offset. If so, assume it | |||
| 3307 | // was created by the PartOffset == 0 branch above and create a load from | |||
| 3308 | // the appropriate offset into it. | |||
| 3309 | int64_t PartBegin = VA.getLocMemOffset(); | |||
| 3310 | int64_t PartEnd = PartBegin + ValVT.getSizeInBits() / 8; | |||
| 3311 | int FI = MFI.getObjectIndexBegin(); | |||
| 3312 | for (; MFI.isFixedObjectIndex(FI); ++FI) { | |||
| 3313 | int64_t ObjBegin = MFI.getObjectOffset(FI); | |||
| 3314 | int64_t ObjEnd = ObjBegin + MFI.getObjectSize(FI); | |||
| 3315 | if (ObjBegin <= PartBegin && PartEnd <= ObjEnd) | |||
| 3316 | break; | |||
| 3317 | } | |||
| 3318 | if (MFI.isFixedObjectIndex(FI)) { | |||
| 3319 | SDValue Addr = | |||
| 3320 | DAG.getNode(ISD::ADD, dl, PtrVT, DAG.getFrameIndex(FI, PtrVT), | |||
| 3321 | DAG.getIntPtrConstant(Ins[i].PartOffset, dl)); | |||
| 3322 | return DAG.getLoad( | |||
| 3323 | ValVT, dl, Chain, Addr, | |||
| 3324 | MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI, | |||
| 3325 | Ins[i].PartOffset)); | |||
| 3326 | } | |||
| 3327 | } | |||
| 3328 | } | |||
| 3329 | ||||
| 3330 | int FI = MFI.CreateFixedObject(ValVT.getSizeInBits() / 8, | |||
| 3331 | VA.getLocMemOffset(), isImmutable); | |||
| 3332 | ||||
| 3333 | // Set SExt or ZExt flag. | |||
| 3334 | if (VA.getLocInfo() == CCValAssign::ZExt) { | |||
| 3335 | MFI.setObjectZExt(FI, true); | |||
| 3336 | } else if (VA.getLocInfo() == CCValAssign::SExt) { | |||
| 3337 | MFI.setObjectSExt(FI, true); | |||
| 3338 | } | |||
| 3339 | ||||
| 3340 | SDValue FIN = DAG.getFrameIndex(FI, PtrVT); | |||
| 3341 | SDValue Val = DAG.getLoad( | |||
| 3342 | ValVT, dl, Chain, FIN, | |||
| 3343 | MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI)); | |||
| 3344 | return ExtendedInMem | |||
| 3345 | ? (VA.getValVT().isVector() | |||
| 3346 | ? DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VA.getValVT(), Val) | |||
| 3347 | : DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), Val)) | |||
| 3348 | : Val; | |||
| 3349 | } | |||
| 3350 | ||||
| 3351 | // FIXME: Get this from tablegen. | |||
| 3352 | static ArrayRef<MCPhysReg> get64BitArgumentGPRs(CallingConv::ID CallConv, | |||
| 3353 | const X86Subtarget &Subtarget) { | |||
| 3354 | assert(Subtarget.is64Bit())((void)0); | |||
| 3355 | ||||
| 3356 | if (Subtarget.isCallingConvWin64(CallConv)) { | |||
| 3357 | static const MCPhysReg GPR64ArgRegsWin64[] = { | |||
| 3358 | X86::RCX, X86::RDX, X86::R8, X86::R9 | |||
| 3359 | }; | |||
| 3360 | return makeArrayRef(std::begin(GPR64ArgRegsWin64), std::end(GPR64ArgRegsWin64)); | |||
| 3361 | } | |||
| 3362 | ||||
| 3363 | static const MCPhysReg GPR64ArgRegs64Bit[] = { | |||
| 3364 | X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9 | |||
| 3365 | }; | |||
| 3366 | return makeArrayRef(std::begin(GPR64ArgRegs64Bit), std::end(GPR64ArgRegs64Bit)); | |||
| 3367 | } | |||
| 3368 | ||||
| 3369 | // FIXME: Get this from tablegen. | |||
| 3370 | static ArrayRef<MCPhysReg> get64BitArgumentXMMs(MachineFunction &MF, | |||
| 3371 | CallingConv::ID CallConv, | |||
| 3372 | const X86Subtarget &Subtarget) { | |||
| 3373 | assert(Subtarget.is64Bit())((void)0); | |||
| 3374 | if (Subtarget.isCallingConvWin64(CallConv)) { | |||
| 3375 | // The XMM registers which might contain var arg parameters are shadowed | |||
| 3376 | // in their paired GPR. So we only need to save the GPR to their home | |||
| 3377 | // slots. | |||
| 3378 | // TODO: __vectorcall will change this. | |||
| 3379 | return None; | |||
| 3380 | } | |||
| 3381 | ||||
| 3382 | bool isSoftFloat = Subtarget.useSoftFloat(); | |||
| 3383 | if (isSoftFloat || !Subtarget.hasSSE1()) | |||
| 3384 | // Kernel mode asks for SSE to be disabled, so there are no XMM argument | |||
| 3385 | // registers. | |||
| 3386 | return None; | |||
| 3387 | ||||
| 3388 | static const MCPhysReg XMMArgRegs64Bit[] = { | |||
| 3389 | X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3, | |||
| 3390 | X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7 | |||
| 3391 | }; | |||
| 3392 | return makeArrayRef(std::begin(XMMArgRegs64Bit), std::end(XMMArgRegs64Bit)); | |||
| 3393 | } | |||
| 3394 | ||||
| 3395 | #ifndef NDEBUG1 | |||
| 3396 | static bool isSortedByValueNo(ArrayRef<CCValAssign> ArgLocs) { | |||
| 3397 | return llvm::is_sorted( | |||
| 3398 | ArgLocs, [](const CCValAssign &A, const CCValAssign &B) -> bool { | |||
| 3399 | return A.getValNo() < B.getValNo(); | |||
| 3400 | }); | |||
| 3401 | } | |||
| 3402 | #endif | |||
| 3403 | ||||
| 3404 | namespace { | |||
| 3405 | /// This is a helper class for lowering variable arguments parameters. | |||
| 3406 | class VarArgsLoweringHelper { | |||
| 3407 | public: | |||
| 3408 | VarArgsLoweringHelper(X86MachineFunctionInfo *FuncInfo, const SDLoc &Loc, | |||
| 3409 | SelectionDAG &DAG, const X86Subtarget &Subtarget, | |||
| 3410 | CallingConv::ID CallConv, CCState &CCInfo) | |||
| 3411 | : FuncInfo(FuncInfo), DL(Loc), DAG(DAG), Subtarget(Subtarget), | |||
| 3412 | TheMachineFunction(DAG.getMachineFunction()), | |||
| 3413 | TheFunction(TheMachineFunction.getFunction()), | |||
| 3414 | FrameInfo(TheMachineFunction.getFrameInfo()), | |||
| 3415 | FrameLowering(*Subtarget.getFrameLowering()), | |||
| 3416 | TargLowering(DAG.getTargetLoweringInfo()), CallConv(CallConv), | |||
| 3417 | CCInfo(CCInfo) {} | |||
| 3418 | ||||
| 3419 | // Lower variable arguments parameters. | |||
| 3420 | void lowerVarArgsParameters(SDValue &Chain, unsigned StackSize); | |||
| 3421 | ||||
| 3422 | private: | |||
| 3423 | void createVarArgAreaAndStoreRegisters(SDValue &Chain, unsigned StackSize); | |||
| 3424 | ||||
| 3425 | void forwardMustTailParameters(SDValue &Chain); | |||
| 3426 | ||||
| 3427 | bool is64Bit() const { return Subtarget.is64Bit(); } | |||
| 3428 | bool isWin64() const { return Subtarget.isCallingConvWin64(CallConv); } | |||
| 3429 | ||||
| 3430 | X86MachineFunctionInfo *FuncInfo; | |||
| 3431 | const SDLoc &DL; | |||
| 3432 | SelectionDAG &DAG; | |||
| 3433 | const X86Subtarget &Subtarget; | |||
| 3434 | MachineFunction &TheMachineFunction; | |||
| 3435 | const Function &TheFunction; | |||
| 3436 | MachineFrameInfo &FrameInfo; | |||
| 3437 | const TargetFrameLowering &FrameLowering; | |||
| 3438 | const TargetLowering &TargLowering; | |||
| 3439 | CallingConv::ID CallConv; | |||
| 3440 | CCState &CCInfo; | |||
| 3441 | }; | |||
| 3442 | } // namespace | |||
| 3443 | ||||
| 3444 | void VarArgsLoweringHelper::createVarArgAreaAndStoreRegisters( | |||
| 3445 | SDValue &Chain, unsigned StackSize) { | |||
| 3446 | // If the function takes variable number of arguments, make a frame index for | |||
| 3447 | // the start of the first vararg value... for expansion of llvm.va_start. We | |||
| 3448 | // can skip this if there are no va_start calls. | |||
| 3449 | if (is64Bit() || (CallConv != CallingConv::X86_FastCall && | |||
| 3450 | CallConv != CallingConv::X86_ThisCall)) { | |||
| 3451 | FuncInfo->setVarArgsFrameIndex( | |||
| 3452 | FrameInfo.CreateFixedObject(1, StackSize, true)); | |||
| 3453 | } | |||
| 3454 | ||||
| 3455 | // 64-bit calling conventions support varargs and register parameters, so we | |||
| 3456 | // have to do extra work to spill them in the prologue. | |||
| 3457 | if (is64Bit()) { | |||
| 3458 | // Find the first unallocated argument registers. | |||
| 3459 | ArrayRef<MCPhysReg> ArgGPRs = get64BitArgumentGPRs(CallConv, Subtarget); | |||
| 3460 | ArrayRef<MCPhysReg> ArgXMMs = | |||
| 3461 | get64BitArgumentXMMs(TheMachineFunction, CallConv, Subtarget); | |||
| 3462 | unsigned NumIntRegs = CCInfo.getFirstUnallocated(ArgGPRs); | |||
| 3463 | unsigned NumXMMRegs = CCInfo.getFirstUnallocated(ArgXMMs); | |||
| 3464 | ||||
| 3465 | assert(!(NumXMMRegs && !Subtarget.hasSSE1()) &&((void)0) | |||
| 3466 | "SSE register cannot be used when SSE is disabled!")((void)0); | |||
| 3467 | ||||
| 3468 | if (isWin64()) { | |||
| 3469 | // Get to the caller-allocated home save location. Add 8 to account | |||
| 3470 | // for the return address. | |||
| 3471 | int HomeOffset = FrameLowering.getOffsetOfLocalArea() + 8; | |||
| 3472 | FuncInfo->setRegSaveFrameIndex( | |||
| 3473 | FrameInfo.CreateFixedObject(1, NumIntRegs * 8 + HomeOffset, false)); | |||
| 3474 | // Fixup to set vararg frame on shadow area (4 x i64). | |||
| 3475 | if (NumIntRegs < 4) | |||
| 3476 | FuncInfo->setVarArgsFrameIndex(FuncInfo->getRegSaveFrameIndex()); | |||
| 3477 | } else { | |||
| 3478 | // For X86-64, if there are vararg parameters that are passed via | |||
| 3479 | // registers, then we must store them to their spots on the stack so | |||
| 3480 | // they may be loaded by dereferencing the result of va_next. | |||
| 3481 | FuncInfo->setVarArgsGPOffset(NumIntRegs * 8); | |||
| 3482 | FuncInfo->setVarArgsFPOffset(ArgGPRs.size() * 8 + NumXMMRegs * 16); | |||
| 3483 | FuncInfo->setRegSaveFrameIndex(FrameInfo.CreateStackObject( | |||
| 3484 | ArgGPRs.size() * 8 + ArgXMMs.size() * 16, Align(16), false)); | |||
| 3485 | } | |||
| 3486 | ||||
| 3487 | SmallVector<SDValue, 6> | |||
| 3488 | LiveGPRs; // list of SDValue for GPR registers keeping live input value | |||
| 3489 | SmallVector<SDValue, 8> LiveXMMRegs; // list of SDValue for XMM registers | |||
| 3490 | // keeping live input value | |||
| 3491 | SDValue ALVal; // if applicable keeps SDValue for %al register | |||
| 3492 | ||||
| 3493 | // Gather all the live in physical registers. | |||
| 3494 | for (MCPhysReg Reg : ArgGPRs.slice(NumIntRegs)) { | |||
| 3495 | Register GPR = TheMachineFunction.addLiveIn(Reg, &X86::GR64RegClass); | |||
| 3496 | LiveGPRs.push_back(DAG.getCopyFromReg(Chain, DL, GPR, MVT::i64)); | |||
| 3497 | } | |||
| 3498 | const auto &AvailableXmms = ArgXMMs.slice(NumXMMRegs); | |||
| 3499 | if (!AvailableXmms.empty()) { | |||
| 3500 | Register AL = TheMachineFunction.addLiveIn(X86::AL, &X86::GR8RegClass); | |||
| 3501 | ALVal = DAG.getCopyFromReg(Chain, DL, AL, MVT::i8); | |||
| 3502 | for (MCPhysReg Reg : AvailableXmms) { | |||
| 3503 | // FastRegisterAllocator spills virtual registers at basic | |||
| 3504 | // block boundary. That leads to usages of xmm registers | |||
| 3505 | // outside of check for %al. Pass physical registers to | |||
| 3506 | // VASTART_SAVE_XMM_REGS to avoid unneccessary spilling. | |||
| 3507 | TheMachineFunction.getRegInfo().addLiveIn(Reg); | |||
| 3508 | LiveXMMRegs.push_back(DAG.getRegister(Reg, MVT::v4f32)); | |||
| 3509 | } | |||
| 3510 | } | |||
| 3511 | ||||
| 3512 | // Store the integer parameter registers. | |||
| 3513 | SmallVector<SDValue, 8> MemOps; | |||
| 3514 | SDValue RSFIN = | |||
| 3515 | DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(), | |||
| 3516 | TargLowering.getPointerTy(DAG.getDataLayout())); | |||
| 3517 | unsigned Offset = FuncInfo->getVarArgsGPOffset(); | |||
| 3518 | for (SDValue Val : LiveGPRs) { | |||
| 3519 | SDValue FIN = DAG.getNode(ISD::ADD, DL, | |||
| 3520 | TargLowering.getPointerTy(DAG.getDataLayout()), | |||
| 3521 | RSFIN, DAG.getIntPtrConstant(Offset, DL)); | |||
| 3522 | SDValue Store = | |||
| 3523 | DAG.getStore(Val.getValue(1), DL, Val, FIN, | |||
| 3524 | MachinePointerInfo::getFixedStack( | |||
| 3525 | DAG.getMachineFunction(), | |||
| 3526 | FuncInfo->getRegSaveFrameIndex(), Offset)); | |||
| 3527 | MemOps.push_back(Store); | |||
| 3528 | Offset += 8; | |||
| 3529 | } | |||
| 3530 | ||||
| 3531 | // Now store the XMM (fp + vector) parameter registers. | |||
| 3532 | if (!LiveXMMRegs.empty()) { | |||
| 3533 | SmallVector<SDValue, 12> SaveXMMOps; | |||
| 3534 | SaveXMMOps.push_back(Chain); | |||
| 3535 | SaveXMMOps.push_back(ALVal); | |||
| 3536 | SaveXMMOps.push_back( | |||
| 3537 | DAG.getTargetConstant(FuncInfo->getRegSaveFrameIndex(), DL, MVT::i32)); | |||
| 3538 | SaveXMMOps.push_back( | |||
| 3539 | DAG.getTargetConstant(FuncInfo->getVarArgsFPOffset(), DL, MVT::i32)); | |||
| 3540 | llvm::append_range(SaveXMMOps, LiveXMMRegs); | |||
| 3541 | MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, DL, | |||
| 3542 | MVT::Other, SaveXMMOps)); | |||
| 3543 | } | |||
| 3544 | ||||
| 3545 | if (!MemOps.empty()) | |||
| 3546 | Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOps); | |||
| 3547 | } | |||
| 3548 | } | |||
| 3549 | ||||
| 3550 | void VarArgsLoweringHelper::forwardMustTailParameters(SDValue &Chain) { | |||
| 3551 | // Find the largest legal vector type. | |||
| 3552 | MVT VecVT = MVT::Other; | |||
| 3553 | // FIXME: Only some x86_32 calling conventions support AVX512. | |||
| 3554 | if (Subtarget.useAVX512Regs() && | |||
| 3555 | (is64Bit() || (CallConv == CallingConv::X86_VectorCall || | |||
| 3556 | CallConv == CallingConv::Intel_OCL_BI))) | |||
| 3557 | VecVT = MVT::v16f32; | |||
| 3558 | else if (Subtarget.hasAVX()) | |||
| 3559 | VecVT = MVT::v8f32; | |||
| 3560 | else if (Subtarget.hasSSE2()) | |||
| 3561 | VecVT = MVT::v4f32; | |||
| 3562 | ||||
| 3563 | // We forward some GPRs and some vector types. | |||
| 3564 | SmallVector<MVT, 2> RegParmTypes; | |||
| 3565 | MVT IntVT = is64Bit() ? MVT::i64 : MVT::i32; | |||
| 3566 | RegParmTypes.push_back(IntVT); | |||
| 3567 | if (VecVT != MVT::Other) | |||
| 3568 | RegParmTypes.push_back(VecVT); | |||
| 3569 | ||||
| 3570 | // Compute the set of forwarded registers. The rest are scratch. | |||
| 3571 | SmallVectorImpl<ForwardedRegister> &Forwards = | |||
| 3572 | FuncInfo->getForwardedMustTailRegParms(); | |||
| 3573 | CCInfo.analyzeMustTailForwardedRegisters(Forwards, RegParmTypes, CC_X86); | |||
| 3574 | ||||
| 3575 | // Forward AL for SysV x86_64 targets, since it is used for varargs. | |||
| 3576 | if (is64Bit() && !isWin64() && !CCInfo.isAllocated(X86::AL)) { | |||
| 3577 | Register ALVReg = TheMachineFunction.addLiveIn(X86::AL, &X86::GR8RegClass); | |||
| 3578 | Forwards.push_back(ForwardedRegister(ALVReg, X86::AL, MVT::i8)); | |||
| 3579 | } | |||
| 3580 | ||||
| 3581 | // Copy all forwards from physical to virtual registers. | |||
| 3582 | for (ForwardedRegister &FR : Forwards) { | |||
| 3583 | // FIXME: Can we use a less constrained schedule? | |||
| 3584 | SDValue RegVal = DAG.getCopyFromReg(Chain, DL, FR.VReg, FR.VT); | |||
| 3585 | FR.VReg = TheMachineFunction.getRegInfo().createVirtualRegister( | |||
| 3586 | TargLowering.getRegClassFor(FR.VT)); | |||
| 3587 | Chain = DAG.getCopyToReg(Chain, DL, FR.VReg, RegVal); | |||
| 3588 | } | |||
| 3589 | } | |||
| 3590 | ||||
| 3591 | void VarArgsLoweringHelper::lowerVarArgsParameters(SDValue &Chain, | |||
| 3592 | unsigned StackSize) { | |||
| 3593 | // Set FrameIndex to the 0xAAAAAAA value to mark unset state. | |||
| 3594 | // If necessary, it would be set into the correct value later. | |||
| 3595 | FuncInfo->setVarArgsFrameIndex(0xAAAAAAA); | |||
| 3596 | FuncInfo->setRegSaveFrameIndex(0xAAAAAAA); | |||
| 3597 | ||||
| 3598 | if (FrameInfo.hasVAStart()) | |||
| 3599 | createVarArgAreaAndStoreRegisters(Chain, StackSize); | |||
| 3600 | ||||
| 3601 | if (FrameInfo.hasMustTailInVarArgFunc()) | |||
| 3602 | forwardMustTailParameters(Chain); | |||
| 3603 | } | |||
| 3604 | ||||
| 3605 | SDValue X86TargetLowering::LowerFormalArguments( | |||
| 3606 | SDValue Chain, CallingConv::ID CallConv, bool IsVarArg, | |||
| 3607 | const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl, | |||
| 3608 | SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const { | |||
| 3609 | MachineFunction &MF = DAG.getMachineFunction(); | |||
| 3610 | X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>(); | |||
| 3611 | ||||
| 3612 | const Function &F = MF.getFunction(); | |||
| 3613 | if (F.hasExternalLinkage() && Subtarget.isTargetCygMing() && | |||
| 3614 | F.getName() == "main") | |||
| 3615 | FuncInfo->setForceFramePointer(true); | |||
| 3616 | ||||
| 3617 | MachineFrameInfo &MFI = MF.getFrameInfo(); | |||
| 3618 | bool Is64Bit = Subtarget.is64Bit(); | |||
| 3619 | bool IsWin64 = Subtarget.isCallingConvWin64(CallConv); | |||
| 3620 | ||||
| 3621 | assert(((void)0) | |||
| 3622 | !(IsVarArg && canGuaranteeTCO(CallConv)) &&((void)0) | |||
| 3623 | "Var args not supported with calling conv' regcall, fastcc, ghc or hipe")((void)0); | |||
| 3624 | ||||
| 3625 | // Assign locations to all of the incoming arguments. | |||
| 3626 | SmallVector<CCValAssign, 16> ArgLocs; | |||
| 3627 | CCState CCInfo(CallConv, IsVarArg, MF, ArgLocs, *DAG.getContext()); | |||
| 3628 | ||||
| 3629 | // Allocate shadow area for Win64. | |||
| 3630 | if (IsWin64) | |||
| 3631 | CCInfo.AllocateStack(32, Align(8)); | |||
| 3632 | ||||
| 3633 | CCInfo.AnalyzeArguments(Ins, CC_X86); | |||
| 3634 | ||||
| 3635 | // In vectorcall calling convention a second pass is required for the HVA | |||
| 3636 | // types. | |||
| 3637 | if (CallingConv::X86_VectorCall == CallConv) { | |||
| 3638 | CCInfo.AnalyzeArgumentsSecondPass(Ins, CC_X86); | |||
| 3639 | } | |||
| 3640 | ||||
| 3641 | // The next loop assumes that the locations are in the same order of the | |||
| 3642 | // input arguments. | |||
| 3643 | assert(isSortedByValueNo(ArgLocs) &&((void)0) | |||
| 3644 | "Argument Location list must be sorted before lowering")((void)0); | |||
| 3645 | ||||
| 3646 | SDValue ArgValue; | |||
| 3647 | for (unsigned I = 0, InsIndex = 0, E = ArgLocs.size(); I != E; | |||
| 3648 | ++I, ++InsIndex) { | |||
| 3649 | assert(InsIndex < Ins.size() && "Invalid Ins index")((void)0); | |||
| 3650 | CCValAssign &VA = ArgLocs[I]; | |||
| 3651 | ||||
| 3652 | if (VA.isRegLoc()) { | |||
| 3653 | EVT RegVT = VA.getLocVT(); | |||
| 3654 | if (VA.needsCustom()) { | |||
| 3655 | assert(((void)0) | |||
| 3656 | VA.getValVT() == MVT::v64i1 &&((void)0) | |||
| 3657 | "Currently the only custom case is when we split v64i1 to 2 regs")((void)0); | |||
| 3658 | ||||
| 3659 | // v64i1 values, in regcall calling convention, that are | |||
| 3660 | // compiled to 32 bit arch, are split up into two registers. | |||
| 3661 | ArgValue = | |||
| 3662 | getv64i1Argument(VA, ArgLocs[++I], Chain, DAG, dl, Subtarget); | |||
| 3663 | } else { | |||
| 3664 | const TargetRegisterClass *RC; | |||
| 3665 | if (RegVT == MVT::i8) | |||
| 3666 | RC = &X86::GR8RegClass; | |||
| 3667 | else if (RegVT == MVT::i16) | |||
| 3668 | RC = &X86::GR16RegClass; | |||
| 3669 | else if (RegVT == MVT::i32) | |||
| 3670 | RC = &X86::GR32RegClass; | |||
| 3671 | else if (Is64Bit && RegVT == MVT::i64) | |||
| 3672 | RC = &X86::GR64RegClass; | |||
| 3673 | else if (RegVT == MVT::f32) | |||
| 3674 | RC = Subtarget.hasAVX512() ? &X86::FR32XRegClass : &X86::FR32RegClass; | |||
| 3675 | else if (RegVT == MVT::f64) | |||
| 3676 | RC = Subtarget.hasAVX512() ? &X86::FR64XRegClass : &X86::FR64RegClass; | |||
| 3677 | else if (RegVT == MVT::f80) | |||
| 3678 | RC = &X86::RFP80RegClass; | |||
| 3679 | else if (RegVT == MVT::f128) | |||
| 3680 | RC = &X86::VR128RegClass; | |||
| 3681 | else if (RegVT.is512BitVector()) | |||
| 3682 | RC = &X86::VR512RegClass; | |||
| 3683 | else if (RegVT.is256BitVector()) | |||
| 3684 | RC = Subtarget.hasVLX() ? &X86::VR256XRegClass : &X86::VR256RegClass; | |||
| 3685 | else if (RegVT.is128BitVector()) | |||
| 3686 | RC = Subtarget.hasVLX() ? &X86::VR128XRegClass : &X86::VR128RegClass; | |||
| 3687 | else if (RegVT == MVT::x86mmx) | |||
| 3688 | RC = &X86::VR64RegClass; | |||
| 3689 | else if (RegVT == MVT::v1i1) | |||
| 3690 | RC = &X86::VK1RegClass; | |||
| 3691 | else if (RegVT == MVT::v8i1) | |||
| 3692 | RC = &X86::VK8RegClass; | |||
| 3693 | else if (RegVT == MVT::v16i1) | |||
| 3694 | RC = &X86::VK16RegClass; | |||
| 3695 | else if (RegVT == MVT::v32i1) | |||
| 3696 | RC = &X86::VK32RegClass; | |||
| 3697 | else if (RegVT == MVT::v64i1) | |||
| 3698 | RC = &X86::VK64RegClass; | |||
| 3699 | else | |||
| 3700 | llvm_unreachable("Unknown argument type!")__builtin_unreachable(); | |||
| 3701 | ||||
| 3702 | Register Reg = MF.addLiveIn(VA.getLocReg(), RC); | |||
| 3703 | ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT); | |||
| 3704 | } | |||
| 3705 | ||||
| 3706 | // If this is an 8 or 16-bit value, it is really passed promoted to 32 | |||
| 3707 | // bits. Insert an assert[sz]ext to capture this, then truncate to the | |||
| 3708 | // right size. | |||
| 3709 | if (VA.getLocInfo() == CCValAssign::SExt) | |||
| 3710 | ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue, | |||
| 3711 | DAG.getValueType(VA.getValVT())); | |||
| 3712 | else if (VA.getLocInfo() == CCValAssign::ZExt) | |||
| 3713 | ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue, | |||
| 3714 | DAG.getValueType(VA.getValVT())); | |||
| 3715 | else if (VA.getLocInfo() == CCValAssign::BCvt) | |||
| 3716 | ArgValue = DAG.getBitcast(VA.getValVT(), ArgValue); | |||
| 3717 | ||||
| 3718 | if (VA.isExtInLoc()) { | |||
| 3719 | // Handle MMX values passed in XMM regs. | |||
| 3720 | if (RegVT.isVector() && VA.getValVT().getScalarType() != MVT::i1) | |||
| 3721 | ArgValue = DAG.getNode(X86ISD::MOVDQ2Q, dl, VA.getValVT(), ArgValue); | |||
| 3722 | else if (VA.getValVT().isVector() && | |||
| 3723 | VA.getValVT().getScalarType() == MVT::i1 && | |||
| 3724 | ((VA.getLocVT() == MVT::i64) || (VA.getLocVT() == MVT::i32) || | |||
| 3725 | (VA.getLocVT() == MVT::i16) || (VA.getLocVT() == MVT::i8))) { | |||
| 3726 | // Promoting a mask type (v*i1) into a register of type i64/i32/i16/i8 | |||
| 3727 | ArgValue = lowerRegToMasks(ArgValue, VA.getValVT(), RegVT, dl, DAG); | |||
| 3728 | } else | |||
| 3729 | ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue); | |||
| 3730 | } | |||
| 3731 | } else { | |||
| 3732 | assert(VA.isMemLoc())((void)0); | |||
| 3733 | ArgValue = | |||
| 3734 | LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, InsIndex); | |||
| 3735 | } | |||
| 3736 | ||||
| 3737 | // If value is passed via pointer - do a load. | |||
| 3738 | if (VA.getLocInfo() == CCValAssign::Indirect && !Ins[I].Flags.isByVal()) | |||
| 3739 | ArgValue = | |||
| 3740 | DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue, MachinePointerInfo()); | |||
| 3741 | ||||
| 3742 | InVals.push_back(ArgValue); | |||
| 3743 | } | |||
| 3744 | ||||
| 3745 | for (unsigned I = 0, E = Ins.size(); I != E; ++I) { | |||
| 3746 | if (Ins[I].Flags.isSwiftAsync()) { | |||
| 3747 | auto X86FI = MF.getInfo<X86MachineFunctionInfo>(); | |||
| 3748 | if (Subtarget.is64Bit()) | |||
| 3749 | X86FI->setHasSwiftAsyncContext(true); | |||
| 3750 | else { | |||
| 3751 | int FI = MF.getFrameInfo().CreateStackObject(4, Align(4), false); | |||
| 3752 | X86FI->setSwiftAsyncContextFrameIdx(FI); | |||
| 3753 | SDValue St = DAG.getStore(DAG.getEntryNode(), dl, InVals[I], | |||
| 3754 | DAG.getFrameIndex(FI, MVT::i32), | |||
| 3755 | MachinePointerInfo::getFixedStack(MF, FI)); | |||
| 3756 | Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, St, Chain); | |||
| 3757 | } | |||
| 3758 | } | |||
| 3759 | ||||
| 3760 | // Swift calling convention does not require we copy the sret argument | |||
| 3761 | // into %rax/%eax for the return. We don't set SRetReturnReg for Swift. | |||
| 3762 | if (CallConv == CallingConv::Swift || CallConv == CallingConv::SwiftTail) | |||
| 3763 | continue; | |||
| 3764 | ||||
| 3765 | // All x86 ABIs require that for returning structs by value we copy the | |||
| 3766 | // sret argument into %rax/%eax (depending on ABI) for the return. Save | |||
| 3767 | // the argument into a virtual register so that we can access it from the | |||
| 3768 | // return points. | |||
| 3769 | if (Ins[I].Flags.isSRet()) { | |||
| 3770 | Register Reg = FuncInfo->getSRetReturnReg(); | |||
| 3771 | if (!Reg) { | |||
| 3772 | MVT PtrTy = getPointerTy(DAG.getDataLayout()); | |||
| 3773 | Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(PtrTy)); | |||
| 3774 | FuncInfo->setSRetReturnReg(Reg); | |||
| 3775 | } | |||
| 3776 | SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[I]); | |||
| 3777 | Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain); | |||
| 3778 | break; | |||
| 3779 | } | |||
| 3780 | } | |||
| 3781 | ||||
| 3782 | unsigned StackSize = CCInfo.getNextStackOffset(); | |||
| 3783 | // Align stack specially for tail calls. | |||
| 3784 | if (shouldGuaranteeTCO(CallConv, | |||
| 3785 | MF.getTarget().Options.GuaranteedTailCallOpt)) | |||
| 3786 | StackSize = GetAlignedArgumentStackSize(StackSize, DAG); | |||
| 3787 | ||||
| 3788 | if (IsVarArg) | |||
| 3789 | VarArgsLoweringHelper(FuncInfo, dl, DAG, Subtarget, CallConv, CCInfo) | |||
| 3790 | .lowerVarArgsParameters(Chain, StackSize); | |||
| 3791 | ||||
| 3792 | // Some CCs need callee pop. | |||
| 3793 | if (X86::isCalleePop(CallConv, Is64Bit, IsVarArg, | |||
| 3794 | MF.getTarget().Options.GuaranteedTailCallOpt)) { | |||
| 3795 | FuncInfo->setBytesToPopOnReturn(StackSize); // Callee pops everything. | |||
| 3796 | } else if (CallConv == CallingConv::X86_INTR && Ins.size() == 2) { | |||
| 3797 | // X86 interrupts must pop the error code (and the alignment padding) if | |||
| 3798 | // present. | |||
| 3799 | FuncInfo->setBytesToPopOnReturn(Is64Bit ? 16 : 4); | |||
| 3800 | } else { | |||
| 3801 | FuncInfo->setBytesToPopOnReturn(0); // Callee pops nothing. | |||
| 3802 | // If this is an sret function, the return should pop the hidden pointer. | |||
| 3803 | if (!Is64Bit && !canGuaranteeTCO(CallConv) && | |||
| 3804 | !Subtarget.getTargetTriple().isOSMSVCRT() && | |||
| 3805 | argsAreStructReturn(Ins, Subtarget.isTargetMCU()) == StackStructReturn) | |||
| 3806 | FuncInfo->setBytesToPopOnReturn(4); | |||
| 3807 | } | |||
| 3808 | ||||
| 3809 | if (!Is64Bit) { | |||
| 3810 | // RegSaveFrameIndex is X86-64 only. | |||
| 3811 | FuncInfo->setRegSaveFrameIndex(0xAAAAAAA); | |||
| 3812 | } | |||
| 3813 | ||||
| 3814 | FuncInfo->setArgumentStackSize(StackSize); | |||
| 3815 | ||||
| 3816 | if (WinEHFuncInfo *EHInfo = MF.getWinEHFuncInfo()) { | |||
| 3817 | EHPersonality Personality = classifyEHPersonality(F.getPersonalityFn()); | |||
| 3818 | if (Personality == EHPersonality::CoreCLR) { | |||
| 3819 | assert(Is64Bit)((void)0); | |||
| 3820 | // TODO: Add a mechanism to frame lowering that will allow us to indicate | |||
| 3821 | // that we'd prefer this slot be allocated towards the bottom of the frame | |||
| 3822 | // (i.e. near the stack pointer after allocating the frame). Every | |||
| 3823 | // funclet needs a copy of this slot in its (mostly empty) frame, and the | |||
| 3824 | // offset from the bottom of this and each funclet's frame must be the | |||
| 3825 | // same, so the size of funclets' (mostly empty) frames is dictated by | |||
| 3826 | // how far this slot is from the bottom (since they allocate just enough | |||
| 3827 | // space to accommodate holding this slot at the correct offset). | |||
| 3828 | int PSPSymFI = MFI.CreateStackObject(8, Align(8), /*isSpillSlot=*/false); | |||
| 3829 | EHInfo->PSPSymFrameIdx = PSPSymFI; | |||
| 3830 | } | |||
| 3831 | } | |||
| 3832 | ||||
| 3833 | if (CallConv == CallingConv::X86_RegCall || | |||
| 3834 | F.hasFnAttribute("no_caller_saved_registers")) { | |||
| 3835 | MachineRegisterInfo &MRI = MF.getRegInfo(); | |||
| 3836 | for (std::pair<Register, Register> Pair : MRI.liveins()) | |||
| 3837 | MRI.disableCalleeSavedRegister(Pair.first); | |||
| 3838 | } | |||
| 3839 | ||||
| 3840 | return Chain; | |||
| 3841 | } | |||
| 3842 | ||||
| 3843 | SDValue X86TargetLowering::LowerMemOpCallTo(SDValue Chain, SDValue StackPtr, | |||
| 3844 | SDValue Arg, const SDLoc &dl, | |||
| 3845 | SelectionDAG &DAG, | |||
| 3846 | const CCValAssign &VA, | |||
| 3847 | ISD::ArgFlagsTy Flags, | |||
| 3848 | bool isByVal) const { | |||
| 3849 | unsigned LocMemOffset = VA.getLocMemOffset(); | |||
| 3850 | SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset, dl); | |||
| 3851 | PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(DAG.getDataLayout()), | |||
| 3852 | StackPtr, PtrOff); | |||
| 3853 | if (isByVal) | |||
| 3854 | return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl); | |||
| 3855 | ||||
| 3856 | return DAG.getStore( | |||
| 3857 | Chain, dl, Arg, PtrOff, | |||
| 3858 | MachinePointerInfo::getStack(DAG.getMachineFunction(), LocMemOffset)); | |||
| 3859 | } | |||
| 3860 | ||||
| 3861 | /// Emit a load of return address if tail call | |||
| 3862 | /// optimization is performed and it is required. | |||
| 3863 | SDValue X86TargetLowering::EmitTailCallLoadRetAddr( | |||
| 3864 | SelectionDAG &DAG, SDValue &OutRetAddr, SDValue Chain, bool IsTailCall, | |||
| 3865 | bool Is64Bit, int FPDiff, const SDLoc &dl) const { | |||
| 3866 | // Adjust the Return address stack slot. | |||
| 3867 | EVT VT = getPointerTy(DAG.getDataLayout()); | |||
| 3868 | OutRetAddr = getReturnAddressFrameIndex(DAG); | |||
| 3869 | ||||
| 3870 | // Load the "old" Return address. | |||
| 3871 | OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, MachinePointerInfo()); | |||
| 3872 | return SDValue(OutRetAddr.getNode(), 1); | |||
| 3873 | } | |||
| 3874 | ||||
| 3875 | /// Emit a store of the return address if tail call | |||
| 3876 | /// optimization is performed and it is required (FPDiff!=0). | |||
| 3877 | static SDValue EmitTailCallStoreRetAddr(SelectionDAG &DAG, MachineFunction &MF, | |||
| 3878 | SDValue Chain, SDValue RetAddrFrIdx, | |||
| 3879 | EVT PtrVT, unsigned SlotSize, | |||
| 3880 | int FPDiff, const SDLoc &dl) { | |||
| 3881 | // Store the return address to the appropriate stack slot. | |||
| 3882 | if (!FPDiff) return Chain; | |||
| 3883 | // Calculate the new stack slot for the return address. | |||
| 3884 | int NewReturnAddrFI = | |||
| 3885 | MF.getFrameInfo().CreateFixedObject(SlotSize, (int64_t)FPDiff - SlotSize, | |||
| 3886 | false); | |||
| 3887 | SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, PtrVT); | |||
| 3888 | Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx, | |||
| 3889 | MachinePointerInfo::getFixedStack( | |||
| 3890 | DAG.getMachineFunction(), NewReturnAddrFI)); | |||
| 3891 | return Chain; | |||
| 3892 | } | |||
| 3893 | ||||
| 3894 | /// Returns a vector_shuffle mask for an movs{s|d}, movd | |||
| 3895 | /// operation of specified width. | |||
| 3896 | static SDValue getMOVL(SelectionDAG &DAG, const SDLoc &dl, MVT VT, SDValue V1, | |||
| 3897 | SDValue V2) { | |||
| 3898 | unsigned NumElems = VT.getVectorNumElements(); | |||
| 3899 | SmallVector<int, 8> Mask; | |||
| 3900 | Mask.push_back(NumElems); | |||
| 3901 | for (unsigned i = 1; i != NumElems; ++i) | |||
| 3902 | Mask.push_back(i); | |||
| 3903 | return DAG.getVectorShuffle(VT, dl, V1, V2, Mask); | |||
| 3904 | } | |||
| 3905 | ||||
| 3906 | SDValue | |||
| 3907 | X86TargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI, | |||
| 3908 | SmallVectorImpl<SDValue> &InVals) const { | |||
| 3909 | SelectionDAG &DAG = CLI.DAG; | |||
| 3910 | SDLoc &dl = CLI.DL; | |||
| 3911 | SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs; | |||
| 3912 | SmallVectorImpl<SDValue> &OutVals = CLI.OutVals; | |||
| 3913 | SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins; | |||
| 3914 | SDValue Chain = CLI.Chain; | |||
| 3915 | SDValue Callee = CLI.Callee; | |||
| 3916 | CallingConv::ID CallConv = CLI.CallConv; | |||
| 3917 | bool &isTailCall = CLI.IsTailCall; | |||
| 3918 | bool isVarArg = CLI.IsVarArg; | |||
| 3919 | const auto *CB = CLI.CB; | |||
| 3920 | ||||
| 3921 | MachineFunction &MF = DAG.getMachineFunction(); | |||
| 3922 | bool Is64Bit = Subtarget.is64Bit(); | |||
| 3923 | bool IsWin64 = Subtarget.isCallingConvWin64(CallConv); | |||
| 3924 | StructReturnType SR = callIsStructReturn(Outs, Subtarget.isTargetMCU()); | |||
| 3925 | bool IsSibcall = false; | |||
| 3926 | bool IsGuaranteeTCO = MF.getTarget().Options.GuaranteedTailCallOpt || | |||
| 3927 | CallConv == CallingConv::Tail || CallConv == CallingConv::SwiftTail; | |||
| 3928 | X86MachineFunctionInfo *X86Info = MF.getInfo<X86MachineFunctionInfo>(); | |||
| 3929 | bool HasNCSR = (CB && isa<CallInst>(CB) && | |||
| 3930 | CB->hasFnAttr("no_caller_saved_registers")); | |||
| 3931 | bool HasNoCfCheck = (CB && CB->doesNoCfCheck()); | |||
| 3932 | bool IsIndirectCall = (CB && isa<CallInst>(CB) && CB->isIndirectCall()); | |||
| 3933 | const Module *M = MF.getMMI().getModule(); | |||
| 3934 | Metadata *IsCFProtectionSupported = M->getModuleFlag("cf-protection-branch"); | |||
| 3935 | ||||
| 3936 | MachineFunction::CallSiteInfo CSInfo; | |||
| 3937 | if (CallConv == CallingConv::X86_INTR) | |||
| 3938 | report_fatal_error("X86 interrupts may not be called directly"); | |||
| 3939 | ||||
| 3940 | bool IsMustTail = CLI.CB && CLI.CB->isMustTailCall(); | |||
| 3941 | if (Subtarget.isPICStyleGOT() && !IsGuaranteeTCO && !IsMustTail) { | |||
| 3942 | // If we are using a GOT, disable tail calls to external symbols with | |||
| 3943 | // default visibility. Tail calling such a symbol requires using a GOT | |||
| 3944 | // relocation, which forces early binding of the symbol. This breaks code | |||
| 3945 | // that require lazy function symbol resolution. Using musttail or | |||
| 3946 | // GuaranteedTailCallOpt will override this. | |||
| 3947 | GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee); | |||
| 3948 | if (!G || (!G->getGlobal()->hasLocalLinkage() && | |||
| 3949 | G->getGlobal()->hasDefaultVisibility())) | |||
| 3950 | isTailCall = false; | |||
| 3951 | } | |||
| 3952 | ||||
| 3953 | ||||
| 3954 | if (isTailCall && !IsMustTail) { | |||
| 3955 | // Check if it's really possible to do a tail call. | |||
| 3956 | isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv, | |||
| 3957 | isVarArg, SR != NotStructReturn, | |||
| 3958 | MF.getFunction().hasStructRetAttr(), CLI.RetTy, | |||
| 3959 | Outs, OutVals, Ins, DAG); | |||
| 3960 | ||||
| 3961 | // Sibcalls are automatically detected tailcalls which do not require | |||
| 3962 | // ABI changes. | |||
| 3963 | if (!IsGuaranteeTCO && isTailCall) | |||
| 3964 | IsSibcall = true; | |||
| 3965 | ||||
| 3966 | if (isTailCall) | |||
| 3967 | ++NumTailCalls; | |||
| 3968 | } | |||
| 3969 | ||||
| 3970 | if (IsMustTail && !isTailCall) | |||
| 3971 | report_fatal_error("failed to perform tail call elimination on a call " | |||
| 3972 | "site marked musttail"); | |||
| 3973 | ||||
| 3974 | assert(!(isVarArg && canGuaranteeTCO(CallConv)) &&((void)0) | |||
| 3975 | "Var args not supported with calling convention fastcc, ghc or hipe")((void)0); | |||
| 3976 | ||||
| 3977 | // Analyze operands of the call, assigning locations to each operand. | |||
| 3978 | SmallVector<CCValAssign, 16> ArgLocs; | |||
| 3979 | CCState CCInfo(CallConv, isVarArg, MF, ArgLocs, *DAG.getContext()); | |||
| 3980 | ||||
| 3981 | // Allocate shadow area for Win64. | |||
| 3982 | if (IsWin64) | |||
| 3983 | CCInfo.AllocateStack(32, Align(8)); | |||
| 3984 | ||||
| 3985 | CCInfo.AnalyzeArguments(Outs, CC_X86); | |||
| 3986 | ||||
| 3987 | // In vectorcall calling convention a second pass is required for the HVA | |||
| 3988 | // types. | |||
| 3989 | if (CallingConv::X86_VectorCall == CallConv) { | |||
| 3990 | CCInfo.AnalyzeArgumentsSecondPass(Outs, CC_X86); | |||
| 3991 | } | |||
| 3992 | ||||
| 3993 | // Get a count of how many bytes are to be pushed on the stack. | |||
| 3994 | unsigned NumBytes = CCInfo.getAlignedCallFrameSize(); | |||
| 3995 | if (IsSibcall) | |||
| 3996 | // This is a sibcall. The memory operands are available in caller's | |||
| 3997 | // own caller's stack. | |||
| 3998 | NumBytes = 0; | |||
| 3999 | else if (IsGuaranteeTCO && canGuaranteeTCO(CallConv)) | |||
| 4000 | NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG); | |||
| 4001 | ||||
| 4002 | int FPDiff = 0; | |||
| 4003 | if (isTailCall && | |||
| 4004 | shouldGuaranteeTCO(CallConv, | |||
| 4005 | MF.getTarget().Options.GuaranteedTailCallOpt)) { | |||
| 4006 | // Lower arguments at fp - stackoffset + fpdiff. | |||
| 4007 | unsigned NumBytesCallerPushed = X86Info->getBytesToPopOnReturn(); | |||
| 4008 | ||||
| 4009 | FPDiff = NumBytesCallerPushed - NumBytes; | |||
| 4010 | ||||
| 4011 | // Set the delta of movement of the returnaddr stackslot. | |||
| 4012 | // But only set if delta is greater than previous delta. | |||
| 4013 | if (FPDiff < X86Info->getTCReturnAddrDelta()) | |||
| 4014 | X86Info->setTCReturnAddrDelta(FPDiff); | |||
| 4015 | } | |||
| 4016 | ||||
| 4017 | unsigned NumBytesToPush = NumBytes; | |||
| 4018 | unsigned NumBytesToPop = NumBytes; | |||
| 4019 | ||||
| 4020 | // If we have an inalloca argument, all stack space has already been allocated | |||
| 4021 | // for us and be right at the top of the stack. We don't support multiple | |||
| 4022 | // arguments passed in memory when using inalloca. | |||
| 4023 | if (!Outs.empty() && Outs.back().Flags.isInAlloca()) { | |||
| 4024 | NumBytesToPush = 0; | |||
| 4025 | if (!ArgLocs.back().isMemLoc()) | |||
| 4026 | report_fatal_error("cannot use inalloca attribute on a register " | |||
| 4027 | "parameter"); | |||
| 4028 | if (ArgLocs.back().getLocMemOffset() != 0) | |||
| 4029 | report_fatal_error("any parameter with the inalloca attribute must be " | |||
| 4030 | "the only memory argument"); | |||
| 4031 | } else if (CLI.IsPreallocated) { | |||
| 4032 | assert(ArgLocs.back().isMemLoc() &&((void)0) | |||
| 4033 | "cannot use preallocated attribute on a register "((void)0) | |||
| 4034 | "parameter")((void)0); | |||
| 4035 | SmallVector<size_t, 4> PreallocatedOffsets; | |||
| 4036 | for (size_t i = 0; i < CLI.OutVals.size(); ++i) { | |||
| 4037 | if (CLI.CB->paramHasAttr(i, Attribute::Preallocated)) { | |||
| 4038 | PreallocatedOffsets.push_back(ArgLocs[i].getLocMemOffset()); | |||
| 4039 | } | |||
| 4040 | } | |||
| 4041 | auto *MFI = DAG.getMachineFunction().getInfo<X86MachineFunctionInfo>(); | |||
| 4042 | size_t PreallocatedId = MFI->getPreallocatedIdForCallSite(CLI.CB); | |||
| 4043 | MFI->setPreallocatedStackSize(PreallocatedId, NumBytes); | |||
| 4044 | MFI->setPreallocatedArgOffsets(PreallocatedId, PreallocatedOffsets); | |||
| 4045 | NumBytesToPush = 0; | |||
| 4046 | } | |||
| 4047 | ||||
| 4048 | if (!IsSibcall && !IsMustTail) | |||
| 4049 | Chain = DAG.getCALLSEQ_START(Chain, NumBytesToPush, | |||
| 4050 | NumBytes - NumBytesToPush, dl); | |||
| 4051 | ||||
| 4052 | SDValue RetAddrFrIdx; | |||
| 4053 | // Load return address for tail calls. | |||
| 4054 | if (isTailCall && FPDiff) | |||
| 4055 | Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall, | |||
| 4056 | Is64Bit, FPDiff, dl); | |||
| 4057 | ||||
| 4058 | SmallVector<std::pair<Register, SDValue>, 8> RegsToPass; | |||
| 4059 | SmallVector<SDValue, 8> MemOpChains; | |||
| 4060 | SDValue StackPtr; | |||
| 4061 | ||||
| 4062 | // The next loop assumes that the locations are in the same order of the | |||
| 4063 | // input arguments. | |||
| 4064 | assert(isSortedByValueNo(ArgLocs) &&((void)0) | |||
| 4065 | "Argument Location list must be sorted before lowering")((void)0); | |||
| 4066 | ||||
| 4067 | // Walk the register/memloc assignments, inserting copies/loads. In the case | |||
| 4068 | // of tail call optimization arguments are handle later. | |||
| 4069 | const X86RegisterInfo *RegInfo = Subtarget.getRegisterInfo(); | |||
| 4070 | for (unsigned I = 0, OutIndex = 0, E = ArgLocs.size(); I != E; | |||
| 4071 | ++I, ++OutIndex) { | |||
| 4072 | assert(OutIndex < Outs.size() && "Invalid Out index")((void)0); | |||
| 4073 | // Skip inalloca/preallocated arguments, they have already been written. | |||
| 4074 | ISD::ArgFlagsTy Flags = Outs[OutIndex].Flags; | |||
| 4075 | if (Flags.isInAlloca() || Flags.isPreallocated()) | |||
| 4076 | continue; | |||
| 4077 | ||||
| 4078 | CCValAssign &VA = ArgLocs[I]; | |||
| 4079 | EVT RegVT = VA.getLocVT(); | |||
| 4080 | SDValue Arg = OutVals[OutIndex]; | |||
| 4081 | bool isByVal = Flags.isByVal(); | |||
| 4082 | ||||
| 4083 | // Promote the value if needed. | |||
| 4084 | switch (VA.getLocInfo()) { | |||
| 4085 | default: llvm_unreachable("Unknown loc info!")__builtin_unreachable(); | |||
| 4086 | case CCValAssign::Full: break; | |||
| 4087 | case CCValAssign::SExt: | |||
| 4088 | Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg); | |||
| 4089 | break; | |||
| 4090 | case CCValAssign::ZExt: | |||
| 4091 | Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg); | |||
| 4092 | break; | |||
| 4093 | case CCValAssign::AExt: | |||
| 4094 | if (Arg.getValueType().isVector() && | |||
| 4095 | Arg.getValueType().getVectorElementType() == MVT::i1) | |||
| 4096 | Arg = lowerMasksToReg(Arg, RegVT, dl, DAG); | |||
| 4097 | else if (RegVT.is128BitVector()) { | |||
| 4098 | // Special case: passing MMX values in XMM registers. | |||
| 4099 | Arg = DAG.getBitcast(MVT::i64, Arg); | |||
| 4100 | Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg); | |||
| 4101 | Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg); | |||
| 4102 | } else | |||
| 4103 | Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg); | |||
| 4104 | break; | |||
| 4105 | case CCValAssign::BCvt: | |||
| 4106 | Arg = DAG.getBitcast(RegVT, Arg); | |||
| 4107 | break; | |||
| 4108 | case CCValAssign::Indirect: { | |||
| 4109 | if (isByVal) { | |||
| 4110 | // Memcpy the argument to a temporary stack slot to prevent | |||
| 4111 | // the caller from seeing any modifications the callee may make | |||
| 4112 | // as guaranteed by the `byval` attribute. | |||
| 4113 | int FrameIdx = MF.getFrameInfo().CreateStackObject( | |||
| 4114 | Flags.getByValSize(), | |||
| 4115 | std::max(Align(16), Flags.getNonZeroByValAlign()), false); | |||
| 4116 | SDValue StackSlot = | |||
| 4117 | DAG.getFrameIndex(FrameIdx, getPointerTy(DAG.getDataLayout())); | |||
| 4118 | Chain = | |||
| 4119 | CreateCopyOfByValArgument(Arg, StackSlot, Chain, Flags, DAG, dl); | |||
| 4120 | // From now on treat this as a regular pointer | |||
| 4121 | Arg = StackSlot; | |||
| 4122 | isByVal = false; | |||
| 4123 | } else { | |||
| 4124 | // Store the argument. | |||
| 4125 | SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT()); | |||
| 4126 | int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex(); | |||
| 4127 | Chain = DAG.getStore( | |||
| 4128 | Chain, dl, Arg, SpillSlot, | |||
| 4129 | MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI)); | |||
| 4130 | Arg = SpillSlot; | |||
| 4131 | } | |||
| 4132 | break; | |||
| 4133 | } | |||
| 4134 | } | |||
| 4135 | ||||
| 4136 | if (VA.needsCustom()) { | |||
| 4137 | assert(VA.getValVT() == MVT::v64i1 &&((void)0) | |||
| 4138 | "Currently the only custom case is when we split v64i1 to 2 regs")((void)0); | |||
| 4139 | // Split v64i1 value into two registers | |||
| 4140 | Passv64i1ArgInRegs(dl, DAG, Arg, RegsToPass, VA, ArgLocs[++I], Subtarget); | |||
| 4141 | } else if (VA.isRegLoc()) { | |||
| 4142 | RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg)); | |||
| 4143 | const TargetOptions &Options = DAG.getTarget().Options; | |||
| 4144 | if (Options.EmitCallSiteInfo) | |||
| 4145 | CSInfo.emplace_back(VA.getLocReg(), I); | |||
| 4146 | if (isVarArg && IsWin64) { | |||
| 4147 | // Win64 ABI requires argument XMM reg to be copied to the corresponding | |||
| 4148 | // shadow reg if callee is a varargs function. | |||
| 4149 | Register ShadowReg; | |||
| 4150 | switch (VA.getLocReg()) { | |||
| 4151 | case X86::XMM0: ShadowReg = X86::RCX; break; | |||
| 4152 | case X86::XMM1: ShadowReg = X86::RDX; break; | |||
| 4153 | case X86::XMM2: ShadowReg = X86::R8; break; | |||
| 4154 | case X86::XMM3: ShadowReg = X86::R9; break; | |||
| 4155 | } | |||
| 4156 | if (ShadowReg) | |||
| 4157 | RegsToPass.push_back(std::make_pair(ShadowReg, Arg)); | |||
| 4158 | } | |||
| 4159 | } else if (!IsSibcall && (!isTailCall || isByVal)) { | |||
| 4160 | assert(VA.isMemLoc())((void)0); | |||
| 4161 | if (!StackPtr.getNode()) | |||
| 4162 | StackPtr = DAG.getCopyFromReg(Chain, dl, RegInfo->getStackRegister(), | |||
| 4163 | getPointerTy(DAG.getDataLayout())); | |||
| 4164 | MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg, | |||
| 4165 | dl, DAG, VA, Flags, isByVal)); | |||
| 4166 | } | |||
| 4167 | } | |||
| 4168 | ||||
| 4169 | if (!MemOpChains.empty()) | |||
| 4170 | Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains); | |||
| 4171 | ||||
| 4172 | if (Subtarget.isPICStyleGOT()) { | |||
| 4173 | // ELF / PIC requires GOT in the EBX register before function calls via PLT | |||
| 4174 | // GOT pointer (except regcall). | |||
| 4175 | if (!isTailCall) { | |||
| 4176 | // Indirect call with RegCall calling convertion may use up all the | |||
| 4177 | // general registers, so it is not suitable to bind EBX reister for | |||
| 4178 | // GOT address, just let register allocator handle it. | |||
| 4179 | if (CallConv != CallingConv::X86_RegCall) | |||
| 4180 | RegsToPass.push_back(std::make_pair( | |||
| 4181 | Register(X86::EBX), DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), | |||
| 4182 | getPointerTy(DAG.getDataLayout())))); | |||
| 4183 | } else { | |||
| 4184 | // If we are tail calling and generating PIC/GOT style code load the | |||
| 4185 | // address of the callee into ECX. The value in ecx is used as target of | |||
| 4186 | // the tail jump. This is done to circumvent the ebx/callee-saved problem | |||
| 4187 | // for tail calls on PIC/GOT architectures. Normally we would just put the | |||
| 4188 | // address of GOT into ebx and then call target@PLT. But for tail calls | |||
| 4189 | // ebx would be restored (since ebx is callee saved) before jumping to the | |||
| 4190 | // target@PLT. | |||
| 4191 | ||||
| 4192 | // Note: The actual moving to ECX is done further down. | |||
| 4193 | GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee); | |||
| 4194 | if (G && !G->getGlobal()->hasLocalLinkage() && | |||
| 4195 | G->getGlobal()->hasDefaultVisibility()) | |||
| 4196 | Callee = LowerGlobalAddress(Callee, DAG); | |||
| 4197 | else if (isa<ExternalSymbolSDNode>(Callee)) | |||
| 4198 | Callee = LowerExternalSymbol(Callee, DAG); | |||
| 4199 | } | |||
| 4200 | } | |||
| 4201 | ||||
| 4202 | if (Is64Bit && isVarArg && !IsWin64 && !IsMustTail) { | |||
| 4203 | // From AMD64 ABI document: | |||
| 4204 | // For calls that may call functions that use varargs or stdargs | |||
| 4205 | // (prototype-less calls or calls to functions containing ellipsis (...) in | |||
| 4206 | // the declaration) %al is used as hidden argument to specify the number | |||
| 4207 | // of SSE registers used. The contents of %al do not need to match exactly | |||
| 4208 | // the number of registers, but must be an ubound on the number of SSE | |||
| 4209 | // registers used and is in the range 0 - 8 inclusive. | |||
| 4210 | ||||
| 4211 | // Count the number of XMM registers allocated. | |||
| 4212 | static const MCPhysReg XMMArgRegs[] = { | |||
| 4213 | X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3, | |||
| 4214 | X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7 | |||
| 4215 | }; | |||
| 4216 | unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs); | |||
| 4217 | assert((Subtarget.hasSSE1() || !NumXMMRegs)((void)0) | |||
| 4218 | && "SSE registers cannot be used when SSE is disabled")((void)0); | |||
| 4219 | RegsToPass.push_back(std::make_pair(Register(X86::AL), | |||
| 4220 | DAG.getConstant(NumXMMRegs, dl, | |||
| 4221 | MVT::i8))); | |||
| 4222 | } | |||
| 4223 | ||||
| 4224 | if (isVarArg && IsMustTail) { | |||
| 4225 | const auto &Forwards = X86Info->getForwardedMustTailRegParms(); | |||
| 4226 | for (const auto &F : Forwards) { | |||
| 4227 | SDValue Val = DAG.getCopyFromReg(Chain, dl, F.VReg, F.VT); | |||
| 4228 | RegsToPass.push_back(std::make_pair(F.PReg, Val)); | |||
| 4229 | } | |||
| 4230 | } | |||
| 4231 | ||||
| 4232 | // For tail calls lower the arguments to the 'real' stack slots. Sibcalls | |||
| 4233 | // don't need this because the eligibility check rejects calls that require | |||
| 4234 | // shuffling arguments passed in memory. | |||
| 4235 | if (!IsSibcall && isTailCall) { | |||
| 4236 | // Force all the incoming stack arguments to be loaded from the stack | |||
| 4237 | // before any new outgoing arguments are stored to the stack, because the | |||
| 4238 | // outgoing stack slots may alias the incoming argument stack slots, and | |||
| 4239 | // the alias isn't otherwise explicit. This is slightly more conservative | |||
| 4240 | // than necessary, because it means that each store effectively depends | |||
| 4241 | // on every argument instead of just those arguments it would clobber. | |||
| 4242 | SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain); | |||
| 4243 | ||||
| 4244 | SmallVector<SDValue, 8> MemOpChains2; | |||
| 4245 | SDValue FIN; | |||
| 4246 | int FI = 0; | |||
| 4247 | for (unsigned I = 0, OutsIndex = 0, E = ArgLocs.size(); I != E; | |||
| 4248 | ++I, ++OutsIndex) { | |||
| 4249 | CCValAssign &VA = ArgLocs[I]; | |||
| 4250 | ||||
| 4251 | if (VA.isRegLoc()) { | |||
| 4252 | if (VA.needsCustom()) { | |||
| 4253 | assert((CallConv == CallingConv::X86_RegCall) &&((void)0) | |||
| 4254 | "Expecting custom case only in regcall calling convention")((void)0); | |||
| 4255 | // This means that we are in special case where one argument was | |||
| 4256 | // passed through two register locations - Skip the next location | |||
| 4257 | ++I; | |||
| 4258 | } | |||
| 4259 | ||||
| 4260 | continue; | |||
| 4261 | } | |||
| 4262 | ||||
| 4263 | assert(VA.isMemLoc())((void)0); | |||
| 4264 | SDValue Arg = OutVals[OutsIndex]; | |||
| 4265 | ISD::ArgFlagsTy Flags = Outs[OutsIndex].Flags; | |||
| 4266 | // Skip inalloca/preallocated arguments. They don't require any work. | |||
| 4267 | if (Flags.isInAlloca() || Flags.isPreallocated()) | |||
| 4268 | continue; | |||
| 4269 | // Create frame index. | |||
| 4270 | int32_t Offset = VA.getLocMemOffset()+FPDiff; | |||
| 4271 | uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8; | |||
| 4272 | FI = MF.getFrameInfo().CreateFixedObject(OpSize, Offset, true); | |||
| 4273 | FIN = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout())); | |||
| 4274 | ||||
| 4275 | if (Flags.isByVal()) { | |||
| 4276 | // Copy relative to framepointer. | |||
| 4277 | SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset(), dl); | |||
| 4278 | if (!StackPtr.getNode()) | |||
| 4279 | StackPtr = DAG.getCopyFromReg(Chain, dl, RegInfo->getStackRegister(), | |||
| 4280 | getPointerTy(DAG.getDataLayout())); | |||
| 4281 | Source = DAG.getNode(ISD::ADD, dl, getPointerTy(DAG.getDataLayout()), | |||
| 4282 | StackPtr, Source); | |||
| 4283 | ||||
| 4284 | MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN, | |||
| 4285 | ArgChain, | |||
| 4286 | Flags, DAG, dl)); | |||
| 4287 | } else { | |||
| 4288 | // Store relative to framepointer. | |||
| 4289 | MemOpChains2.push_back(DAG.getStore( | |||
| 4290 | ArgChain, dl, Arg, FIN, | |||
| 4291 | MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI))); | |||
| 4292 | } | |||
| 4293 | } | |||
| 4294 | ||||
| 4295 | if (!MemOpChains2.empty()) | |||
| 4296 | Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains2); | |||
| 4297 | ||||
| 4298 | // Store the return address to the appropriate stack slot. | |||
| 4299 | Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, | |||
| 4300 | getPointerTy(DAG.getDataLayout()), | |||
| 4301 | RegInfo->getSlotSize(), FPDiff, dl); | |||
| 4302 | } | |||
| 4303 | ||||
| 4304 | // Build a sequence of copy-to-reg nodes chained together with token chain | |||
| 4305 | // and flag operands which copy the outgoing args into registers. | |||
| 4306 | SDValue InFlag; | |||
| 4307 | for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) { | |||
| 4308 | Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first, | |||
| 4309 | RegsToPass[i].second, InFlag); | |||
| 4310 | InFlag = Chain.getValue(1); | |||
| 4311 | } | |||
| 4312 | ||||
| 4313 | if (DAG.getTarget().getCodeModel() == CodeModel::Large) { | |||
| 4314 | assert(Is64Bit && "Large code model is only legal in 64-bit mode.")((void)0); | |||
| 4315 | // In the 64-bit large code model, we have to make all calls | |||
| 4316 | // through a register, since the call instruction's 32-bit | |||
| 4317 | // pc-relative offset may not be large enough to hold the whole | |||
| 4318 | // address. | |||
| 4319 | } else if (Callee->getOpcode() == ISD::GlobalAddress || | |||
| 4320 | Callee->getOpcode() == ISD::ExternalSymbol) { | |||
| 4321 | // Lower direct calls to global addresses and external symbols. Setting | |||
| 4322 | // ForCall to true here has the effect of removing WrapperRIP when possible | |||
| 4323 | // to allow direct calls to be selected without first materializing the | |||
| 4324 | // address into a register. | |||
| 4325 | Callee = LowerGlobalOrExternal(Callee, DAG, /*ForCall=*/true); | |||
| 4326 | } else if (Subtarget.isTarget64BitILP32() && | |||
| 4327 | Callee->getValueType(0) == MVT::i32) { | |||
| 4328 | // Zero-extend the 32-bit Callee address into a 64-bit according to x32 ABI | |||
| 4329 | Callee = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i64, Callee); | |||
| 4330 | } | |||
| 4331 | ||||
| 4332 | // Returns a chain & a flag for retval copy to use. | |||
| 4333 | SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue); | |||
| 4334 | SmallVector<SDValue, 8> Ops; | |||
| 4335 | ||||
| 4336 | if (!IsSibcall && isTailCall && !IsMustTail) { | |||
| 4337 | Chain = DAG.getCALLSEQ_END(Chain, | |||
| 4338 | DAG.getIntPtrConstant(NumBytesToPop, dl, true), | |||
| 4339 | DAG.getIntPtrConstant(0, dl, true), InFlag, dl); | |||
| 4340 | InFlag = Chain.getValue(1); | |||
| 4341 | } | |||
| 4342 | ||||
| 4343 | Ops.push_back(Chain); | |||
| 4344 | Ops.push_back(Callee); | |||
| 4345 | ||||
| 4346 | if (isTailCall) | |||
| 4347 | Ops.push_back(DAG.getTargetConstant(FPDiff, dl, MVT::i32)); | |||
| 4348 | ||||
| 4349 | // Add argument registers to the end of the list so that they are known live | |||
| 4350 | // into the call. | |||
| 4351 | for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) | |||
| 4352 | Ops.push_back(DAG.getRegister(RegsToPass[i].first, | |||
| 4353 | RegsToPass[i].second.getValueType())); | |||
| 4354 | ||||
| 4355 | // Add a register mask operand representing the call-preserved registers. | |||
| 4356 | const uint32_t *Mask = [&]() { | |||
| 4357 | auto AdaptedCC = CallConv; | |||
| 4358 | // If HasNCSR is asserted (attribute NoCallerSavedRegisters exists), | |||
| 4359 | // use X86_INTR calling convention because it has the same CSR mask | |||
| 4360 | // (same preserved registers). | |||
| 4361 | if (HasNCSR) | |||
| 4362 | AdaptedCC = (CallingConv::ID)CallingConv::X86_INTR; | |||
| 4363 | // If NoCalleeSavedRegisters is requested, than use GHC since it happens | |||
| 4364 | // to use the CSR_NoRegs_RegMask. | |||
| 4365 | if (CB && CB->hasFnAttr("no_callee_saved_registers")) | |||
| 4366 | AdaptedCC = (CallingConv::ID)CallingConv::GHC; | |||
| 4367 | return RegInfo->getCallPreservedMask(MF, AdaptedCC); | |||
| 4368 | }(); | |||
| 4369 | assert(Mask && "Missing call preserved mask for calling convention")((void)0); | |||
| 4370 | ||||
| 4371 | // If this is an invoke in a 32-bit function using a funclet-based | |||
| 4372 | // personality, assume the function clobbers all registers. If an exception | |||
| 4373 | // is thrown, the runtime will not restore CSRs. | |||
| 4374 | // FIXME: Model this more precisely so that we can register allocate across | |||
| 4375 | // the normal edge and spill and fill across the exceptional edge. | |||
| 4376 | if (!Is64Bit && CLI.CB && isa<InvokeInst>(CLI.CB)) { | |||
| 4377 | const Function &CallerFn = MF.getFunction(); | |||
| 4378 | EHPersonality Pers = | |||
| 4379 | CallerFn.hasPersonalityFn() | |||
| 4380 | ? classifyEHPersonality(CallerFn.getPersonalityFn()) | |||
| 4381 | : EHPersonality::Unknown; | |||
| 4382 | if (isFuncletEHPersonality(Pers)) | |||
| 4383 | Mask = RegInfo->getNoPreservedMask(); | |||
| 4384 | } | |||
| 4385 | ||||
| 4386 | // Define a new register mask from the existing mask. | |||
| 4387 | uint32_t *RegMask = nullptr; | |||
| 4388 | ||||
| 4389 | // In some calling conventions we need to remove the used physical registers | |||
| 4390 | // from the reg mask. | |||
| 4391 | if (CallConv == CallingConv::X86_RegCall || HasNCSR) { | |||
| 4392 | const TargetRegisterInfo *TRI = Subtarget.getRegisterInfo(); | |||
| 4393 | ||||
| 4394 | // Allocate a new Reg Mask and copy Mask. | |||
| 4395 | RegMask = MF.allocateRegMask(); | |||
| 4396 | unsigned RegMaskSize = MachineOperand::getRegMaskSize(TRI->getNumRegs()); | |||
| 4397 | memcpy(RegMask, Mask, sizeof(RegMask[0]) * RegMaskSize); | |||
| 4398 | ||||
| 4399 | // Make sure all sub registers of the argument registers are reset | |||
| 4400 | // in the RegMask. | |||
| 4401 | for (auto const &RegPair : RegsToPass) | |||
| 4402 | for (MCSubRegIterator SubRegs(RegPair.first, TRI, /*IncludeSelf=*/true); | |||
| 4403 | SubRegs.isValid(); ++SubRegs) | |||
| 4404 | RegMask[*SubRegs / 32] &= ~(1u << (*SubRegs % 32)); | |||
| 4405 | ||||
| 4406 | // Create the RegMask Operand according to our updated mask. | |||
| 4407 | Ops.push_back(DAG.getRegisterMask(RegMask)); | |||
| 4408 | } else { | |||
| 4409 | // Create the RegMask Operand according to the static mask. | |||
| 4410 | Ops.push_back(DAG.getRegisterMask(Mask)); | |||
| 4411 | } | |||
| 4412 | ||||
| 4413 | if (InFlag.getNode()) | |||
| 4414 | Ops.push_back(InFlag); | |||
| 4415 | ||||
| 4416 | if (isTailCall) { | |||
| 4417 | // We used to do: | |||
| 4418 | //// If this is the first return lowered for this function, add the regs | |||
| 4419 | //// to the liveout set for the function. | |||
| 4420 | // This isn't right, although it's probably harmless on x86; liveouts | |||
| 4421 | // should be computed from returns not tail calls. Consider a void | |||
| 4422 | // function making a tail call to a function returning int. | |||
| 4423 | MF.getFrameInfo().setHasTailCall(); | |||
| 4424 | SDValue Ret = DAG.getNode(X86ISD::TC_RETURN, dl, NodeTys, Ops); | |||
| 4425 | DAG.addCallSiteInfo(Ret.getNode(), std::move(CSInfo)); | |||
| 4426 | return Ret; | |||
| 4427 | } | |||
| 4428 | ||||
| 4429 | if (HasNoCfCheck && IsCFProtectionSupported && IsIndirectCall) { | |||
| 4430 | Chain = DAG.getNode(X86ISD::NT_CALL, dl, NodeTys, Ops); | |||
| 4431 | } else if (CLI.CB && objcarc::hasAttachedCallOpBundle(CLI.CB)) { | |||
| 4432 | // Calls with a "clang.arc.attachedcall" bundle are special. They should be | |||
| 4433 | // expanded to the call, directly followed by a special marker sequence and | |||
| 4434 | // a call to a ObjC library function. Use the CALL_RVMARKER to do that. | |||
| 4435 | assert(!isTailCall &&((void)0) | |||
| 4436 | "tail calls cannot be marked with clang.arc.attachedcall")((void)0); | |||
| 4437 | assert(Is64Bit && "clang.arc.attachedcall is only supported in 64bit mode")((void)0); | |||
| 4438 | ||||
| 4439 | // Add target constant to select ObjC runtime call just before the call | |||
| 4440 | // target. RuntimeCallType == 0 selects objc_retainAutoreleasedReturnValue, | |||
| 4441 | // RuntimeCallType == 0 selects objc_unsafeClaimAutoreleasedReturnValue when | |||
| 4442 | // epxanding the pseudo. | |||
| 4443 | unsigned RuntimeCallType = | |||
| 4444 | objcarc::hasAttachedCallOpBundle(CLI.CB, true) ? 0 : 1; | |||
| 4445 | Ops.insert(Ops.begin() + 1, | |||
| 4446 | DAG.getTargetConstant(RuntimeCallType, dl, MVT::i32)); | |||
| 4447 | Chain = DAG.getNode(X86ISD::CALL_RVMARKER, dl, NodeTys, Ops); | |||
| 4448 | } else { | |||
| 4449 | Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, Ops); | |||
| 4450 | } | |||
| 4451 | ||||
| 4452 | InFlag = Chain.getValue(1); | |||
| 4453 | DAG.addNoMergeSiteInfo(Chain.getNode(), CLI.NoMerge); | |||
| 4454 | DAG.addCallSiteInfo(Chain.getNode(), std::move(CSInfo)); | |||
| 4455 | ||||
| 4456 | // Save heapallocsite metadata. | |||
| 4457 | if (CLI.CB) | |||
| 4458 | if (MDNode *HeapAlloc = CLI.CB->getMetadata("heapallocsite")) | |||
| 4459 | DAG.addHeapAllocSite(Chain.getNode(), HeapAlloc); | |||
| 4460 | ||||
| 4461 | // Create the CALLSEQ_END node. | |||
| 4462 | unsigned NumBytesForCalleeToPop; | |||
| 4463 | if (X86::isCalleePop(CallConv, Is64Bit, isVarArg, | |||
| 4464 | DAG.getTarget().Options.GuaranteedTailCallOpt)) | |||
| 4465 | NumBytesForCalleeToPop = NumBytes; // Callee pops everything | |||
| 4466 | else if (!Is64Bit && !canGuaranteeTCO(CallConv) && | |||
| 4467 | !Subtarget.getTargetTriple().isOSMSVCRT() && | |||
| 4468 | SR == StackStructReturn) | |||
| 4469 | // If this is a call to a struct-return function, the callee | |||
| 4470 | // pops the hidden struct pointer, so we have to push it back. | |||
| 4471 | // This is common for Darwin/X86, Linux & Mingw32 targets. | |||
| 4472 | // For MSVC Win32 targets, the caller pops the hidden struct pointer. | |||
| 4473 | NumBytesForCalleeToPop = 4; | |||
| 4474 | else | |||
| 4475 | NumBytesForCalleeToPop = 0; // Callee pops nothing. | |||
| 4476 | ||||
| 4477 | // Returns a flag for retval copy to use. | |||
| 4478 | if (!IsSibcall) { | |||
| 4479 | Chain = DAG.getCALLSEQ_END(Chain, | |||
| 4480 | DAG.getIntPtrConstant(NumBytesToPop, dl, true), | |||
| 4481 | DAG.getIntPtrConstant(NumBytesForCalleeToPop, dl, | |||
| 4482 | true), | |||
| 4483 | InFlag, dl); | |||
| 4484 | InFlag = Chain.getValue(1); | |||
| 4485 | } | |||
| 4486 | ||||
| 4487 | // Handle result values, copying them out of physregs into vregs that we | |||
| 4488 | // return. | |||
| 4489 | return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins, dl, DAG, | |||
| 4490 | InVals, RegMask); | |||
| 4491 | } | |||
| 4492 | ||||
| 4493 | //===----------------------------------------------------------------------===// | |||
| 4494 | // Fast Calling Convention (tail call) implementation | |||
| 4495 | //===----------------------------------------------------------------------===// | |||
| 4496 | ||||
| 4497 | // Like std call, callee cleans arguments, convention except that ECX is | |||
| 4498 | // reserved for storing the tail called function address. Only 2 registers are | |||
| 4499 | // free for argument passing (inreg). Tail call optimization is performed | |||
| 4500 | // provided: | |||
| 4501 | // * tailcallopt is enabled | |||
| 4502 | // * caller/callee are fastcc | |||
| 4503 | // On X86_64 architecture with GOT-style position independent code only local | |||
| 4504 | // (within module) calls are supported at the moment. | |||
| 4505 | // To keep the stack aligned according to platform abi the function | |||
| 4506 | // GetAlignedArgumentStackSize ensures that argument delta is always multiples | |||
| 4507 | // of stack alignment. (Dynamic linkers need this - Darwin's dyld for example) | |||
| 4508 | // If a tail called function callee has more arguments than the caller the | |||
| 4509 | // caller needs to make sure that there is room to move the RETADDR to. This is | |||
| 4510 | // achieved by reserving an area the size of the argument delta right after the | |||
| 4511 | // original RETADDR, but before the saved framepointer or the spilled registers | |||
| 4512 | // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4) | |||
| 4513 | // stack layout: | |||
| 4514 | // arg1 | |||
| 4515 | // arg2 | |||
| 4516 | // RETADDR | |||
| 4517 | // [ new RETADDR | |||
| 4518 | // move area ] | |||
| 4519 | // (possible EBP) | |||
| 4520 | // ESI | |||
| 4521 | // EDI | |||
| 4522 | // local1 .. | |||
| 4523 | ||||
| 4524 | /// Make the stack size align e.g 16n + 12 aligned for a 16-byte align | |||
| 4525 | /// requirement. | |||
| 4526 | unsigned | |||
| 4527 | X86TargetLowering::GetAlignedArgumentStackSize(const unsigned StackSize, | |||
| 4528 | SelectionDAG &DAG) const { | |||
| 4529 | const Align StackAlignment = Subtarget.getFrameLowering()->getStackAlign(); | |||
| 4530 | const uint64_t SlotSize = Subtarget.getRegisterInfo()->getSlotSize(); | |||
| 4531 | assert(StackSize % SlotSize == 0 &&((void)0) | |||
| 4532 | "StackSize must be a multiple of SlotSize")((void)0); | |||
| 4533 | return alignTo(StackSize + SlotSize, StackAlignment) - SlotSize; | |||
| 4534 | } | |||
| 4535 | ||||
| 4536 | /// Return true if the given stack call argument is already available in the | |||
| 4537 | /// same position (relatively) of the caller's incoming argument stack. | |||
| 4538 | static | |||
| 4539 | bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags, | |||
| 4540 | MachineFrameInfo &MFI, const MachineRegisterInfo *MRI, | |||
| 4541 | const X86InstrInfo *TII, const CCValAssign &VA) { | |||
| 4542 | unsigned Bytes = Arg.getValueSizeInBits() / 8; | |||
| 4543 | ||||
| 4544 | for (;;) { | |||
| 4545 | // Look through nodes that don't alter the bits of the incoming value. | |||
| 4546 | unsigned Op = Arg.getOpcode(); | |||
| 4547 | if (Op == ISD::ZERO_EXTEND || Op == ISD::ANY_EXTEND || Op == ISD::BITCAST) { | |||
| 4548 | Arg = Arg.getOperand(0); | |||
| 4549 | continue; | |||
| 4550 | } | |||
| 4551 | if (Op == ISD::TRUNCATE) { | |||
| 4552 | const SDValue &TruncInput = Arg.getOperand(0); | |||
| 4553 | if (TruncInput.getOpcode() == ISD::AssertZext && | |||
| 4554 | cast<VTSDNode>(TruncInput.getOperand(1))->getVT() == | |||
| 4555 | Arg.getValueType()) { | |||
| 4556 | Arg = TruncInput.getOperand(0); | |||
| 4557 | continue; | |||
| 4558 | } | |||
| 4559 | } | |||
| 4560 | break; | |||
| 4561 | } | |||
| 4562 | ||||
| 4563 | int FI = INT_MAX2147483647; | |||
| 4564 | if (Arg.getOpcode() == ISD::CopyFromReg) { | |||
| 4565 | Register VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg(); | |||
| 4566 | if (!VR.isVirtual()) | |||
| 4567 | return false; | |||
| 4568 | MachineInstr *Def = MRI->getVRegDef(VR); | |||
| 4569 | if (!Def) | |||
| 4570 | return false; | |||
| 4571 | if (!Flags.isByVal()) { | |||
| 4572 | if (!TII->isLoadFromStackSlot(*Def, FI)) | |||
| 4573 | return false; | |||
| 4574 | } else { | |||
| 4575 | unsigned Opcode = Def->getOpcode(); | |||
| 4576 | if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r || | |||
| 4577 | Opcode == X86::LEA64_32r) && | |||
| 4578 | Def->getOperand(1).isFI()) { | |||
| 4579 | FI = Def->getOperand(1).getIndex(); | |||
| 4580 | Bytes = Flags.getByValSize(); | |||
| 4581 | } else | |||
| 4582 | return false; | |||
| 4583 | } | |||
| 4584 | } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) { | |||
| 4585 | if (Flags.isByVal()) | |||
| 4586 | // ByVal argument is passed in as a pointer but it's now being | |||
| 4587 | // dereferenced. e.g. | |||
| 4588 | // define @foo(%struct.X* %A) { | |||
| 4589 | // tail call @bar(%struct.X* byval %A) | |||
| 4590 | // } | |||
| 4591 | return false; | |||
| 4592 | SDValue Ptr = Ld->getBasePtr(); | |||
| 4593 | FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr); | |||
| 4594 | if (!FINode) | |||
| 4595 | return false; | |||
| 4596 | FI = FINode->getIndex(); | |||
| 4597 | } else if (Arg.getOpcode() == ISD::FrameIndex && Flags.isByVal()) { | |||
| 4598 | FrameIndexSDNode *FINode = cast<FrameIndexSDNode>(Arg); | |||
| 4599 | FI = FINode->getIndex(); | |||
| 4600 | Bytes = Flags.getByValSize(); | |||
| 4601 | } else | |||
| 4602 | return false; | |||
| 4603 | ||||
| 4604 | assert(FI != INT_MAX)((void)0); | |||
| 4605 | if (!MFI.isFixedObjectIndex(FI)) | |||
| 4606 | return false; | |||
| 4607 | ||||
| 4608 | if (Offset != MFI.getObjectOffset(FI)) | |||
| 4609 | return false; | |||
| 4610 | ||||
| 4611 | // If this is not byval, check that the argument stack object is immutable. | |||
| 4612 | // inalloca and argument copy elision can create mutable argument stack | |||
| 4613 | // objects. Byval objects can be mutated, but a byval call intends to pass the | |||
| 4614 | // mutated memory. | |||
| 4615 | if (!Flags.isByVal() && !MFI.isImmutableObjectIndex(FI)) | |||
| 4616 | return false; | |||
| 4617 | ||||
| 4618 | if (VA.getLocVT().getFixedSizeInBits() > | |||
| 4619 | Arg.getValueSizeInBits().getFixedSize()) { | |||
| 4620 | // If the argument location is wider than the argument type, check that any | |||
| 4621 | // extension flags match. | |||
| 4622 | if (Flags.isZExt() != MFI.isObjectZExt(FI) || | |||
| 4623 | Flags.isSExt() != MFI.isObjectSExt(FI)) { | |||
| 4624 | return false; | |||
| 4625 | } | |||
| 4626 | } | |||
| 4627 | ||||
| 4628 | return Bytes == MFI.getObjectSize(FI); | |||
| 4629 | } | |||
| 4630 | ||||
| 4631 | /// Check whether the call is eligible for tail call optimization. Targets | |||
| 4632 | /// that want to do tail call optimization should implement this function. | |||
| 4633 | bool X86TargetLowering::IsEligibleForTailCallOptimization( | |||
| 4634 | SDValue Callee, CallingConv::ID CalleeCC, bool isVarArg, | |||
| 4635 | bool isCalleeStructRet, bool isCallerStructRet, Type *RetTy, | |||
| 4636 | const SmallVectorImpl<ISD::OutputArg> &Outs, | |||
| 4637 | const SmallVectorImpl<SDValue> &OutVals, | |||
| 4638 | const SmallVectorImpl<ISD::InputArg> &Ins, SelectionDAG &DAG) const { | |||
| 4639 | if (!mayTailCallThisCC(CalleeCC)) | |||
| 4640 | return false; | |||
| 4641 | ||||
| 4642 | // If -tailcallopt is specified, make fastcc functions tail-callable. | |||
| 4643 | MachineFunction &MF = DAG.getMachineFunction(); | |||
| 4644 | const Function &CallerF = MF.getFunction(); | |||
| 4645 | ||||
| 4646 | // If the function return type is x86_fp80 and the callee return type is not, | |||
| 4647 | // then the FP_EXTEND of the call result is not a nop. It's not safe to | |||
| 4648 | // perform a tailcall optimization here. | |||
| 4649 | if (CallerF.getReturnType()->isX86_FP80Ty() && !RetTy->isX86_FP80Ty()) | |||
| 4650 | return false; | |||
| 4651 | ||||
| 4652 | CallingConv::ID CallerCC = CallerF.getCallingConv(); | |||
| 4653 | bool CCMatch = CallerCC == CalleeCC; | |||
| 4654 | bool IsCalleeWin64 = Subtarget.isCallingConvWin64(CalleeCC); | |||
| 4655 | bool IsCallerWin64 = Subtarget.isCallingConvWin64(CallerCC); | |||
| 4656 | bool IsGuaranteeTCO = DAG.getTarget().Options.GuaranteedTailCallOpt || | |||
| 4657 | CalleeCC == CallingConv::Tail || CalleeCC == CallingConv::SwiftTail; | |||
| 4658 | ||||
| 4659 | // Win64 functions have extra shadow space for argument homing. Don't do the | |||
| 4660 | // sibcall if the caller and callee have mismatched expectations for this | |||
| 4661 | // space. | |||
| 4662 | if (IsCalleeWin64 != IsCallerWin64) | |||
| 4663 | return false; | |||
| 4664 | ||||
| 4665 | if (IsGuaranteeTCO) { | |||
| 4666 | if (canGuaranteeTCO(CalleeCC) && CCMatch) | |||
| 4667 | return true; | |||
| 4668 | return false; | |||
| 4669 | } | |||
| 4670 | ||||
| 4671 | // Look for obvious safe cases to perform tail call optimization that do not | |||
| 4672 | // require ABI changes. This is what gcc calls sibcall. | |||
| 4673 | ||||
| 4674 | // Can't do sibcall if stack needs to be dynamically re-aligned. PEI needs to | |||
| 4675 | // emit a special epilogue. | |||
| 4676 | const X86RegisterInfo *RegInfo = Subtarget.getRegisterInfo(); | |||
| 4677 | if (RegInfo->hasStackRealignment(MF)) | |||
| 4678 | return false; | |||
| 4679 | ||||
| 4680 | // Also avoid sibcall optimization if either caller or callee uses struct | |||
| 4681 | // return semantics. | |||
| 4682 | if (isCalleeStructRet || isCallerStructRet) | |||
| 4683 | return false; | |||
| 4684 | ||||
| 4685 | // Do not sibcall optimize vararg calls unless all arguments are passed via | |||
| 4686 | // registers. | |||
| 4687 | LLVMContext &C = *DAG.getContext(); | |||
| 4688 | if (isVarArg && !Outs.empty()) { | |||
| 4689 | // Optimizing for varargs on Win64 is unlikely to be safe without | |||
| 4690 | // additional testing. | |||
| 4691 | if (IsCalleeWin64 || IsCallerWin64) | |||
| 4692 | return false; | |||
| 4693 | ||||
| 4694 | SmallVector<CCValAssign, 16> ArgLocs; | |||
| 4695 | CCState CCInfo(CalleeCC, isVarArg, MF, ArgLocs, C); | |||
| 4696 | ||||
| 4697 | CCInfo.AnalyzeCallOperands(Outs, CC_X86); | |||
| 4698 | for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) | |||
| 4699 | if (!ArgLocs[i].isRegLoc()) | |||
| 4700 | return false; | |||
| 4701 | } | |||
| 4702 | ||||
| 4703 | // If the call result is in ST0 / ST1, it needs to be popped off the x87 | |||
| 4704 | // stack. Therefore, if it's not used by the call it is not safe to optimize | |||
| 4705 | // this into a sibcall. | |||
| 4706 | bool Unused = false; | |||
| 4707 | for (unsigned i = 0, e = Ins.size(); i != e; ++i) { | |||
| 4708 | if (!Ins[i].Used) { | |||
| 4709 | Unused = true; | |||
| 4710 | break; | |||
| 4711 | } | |||
| 4712 | } | |||
| 4713 | if (Unused) { | |||
| 4714 | SmallVector<CCValAssign, 16> RVLocs; | |||
| 4715 | CCState CCInfo(CalleeCC, false, MF, RVLocs, C); | |||
| 4716 | CCInfo.AnalyzeCallResult(Ins, RetCC_X86); | |||
| 4717 | for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) { | |||
| 4718 | CCValAssign &VA = RVLocs[i]; | |||
| 4719 | if (VA.getLocReg() == X86::FP0 || VA.getLocReg() == X86::FP1) | |||
| 4720 | return false; | |||
| 4721 | } | |||
| 4722 | } | |||
| 4723 | ||||
| 4724 | // Check that the call results are passed in the same way. | |||
| 4725 | if (!CCState::resultsCompatible(CalleeCC, CallerCC, MF, C, Ins, | |||
| 4726 | RetCC_X86, RetCC_X86)) | |||
| 4727 | return false; | |||
| 4728 | // The callee has to preserve all registers the caller needs to preserve. | |||
| 4729 | const X86RegisterInfo *TRI = Subtarget.getRegisterInfo(); | |||
| 4730 | const uint32_t *CallerPreserved = TRI->getCallPreservedMask(MF, CallerCC); | |||
| 4731 | if (!CCMatch) { | |||
| 4732 | const uint32_t *CalleePreserved = TRI->getCallPreservedMask(MF, CalleeCC); | |||
| 4733 | if (!TRI->regmaskSubsetEqual(CallerPreserved, CalleePreserved)) | |||
| 4734 | return false; | |||
| 4735 | } | |||
| 4736 | ||||
| 4737 | unsigned StackArgsSize = 0; | |||
| 4738 | ||||
| 4739 | // If the callee takes no arguments then go on to check the results of the | |||
| 4740 | // call. | |||
| 4741 | if (!Outs.empty()) { | |||
| 4742 | // Check if stack adjustment is needed. For now, do not do this if any | |||
| 4743 | // argument is passed on the stack. | |||
| 4744 | SmallVector<CCValAssign, 16> ArgLocs; | |||
| 4745 | CCState CCInfo(CalleeCC, isVarArg, MF, ArgLocs, C); | |||
| 4746 | ||||
| 4747 | // Allocate shadow area for Win64 | |||
| 4748 | if (IsCalleeWin64) | |||
| 4749 | CCInfo.AllocateStack(32, Align(8)); | |||
| 4750 | ||||
| 4751 | CCInfo.AnalyzeCallOperands(Outs, CC_X86); | |||
| 4752 | StackArgsSize = CCInfo.getNextStackOffset(); | |||
| 4753 | ||||
| 4754 | if (CCInfo.getNextStackOffset()) { | |||
| 4755 | // Check if the arguments are already laid out in the right way as | |||
| 4756 | // the caller's fixed stack objects. | |||
| 4757 | MachineFrameInfo &MFI = MF.getFrameInfo(); | |||
| 4758 | const MachineRegisterInfo *MRI = &MF.getRegInfo(); | |||
| 4759 | const X86InstrInfo *TII = Subtarget.getInstrInfo(); | |||
| 4760 | for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) { | |||
| 4761 | CCValAssign &VA = ArgLocs[i]; | |||
| 4762 | SDValue Arg = OutVals[i]; | |||
| 4763 | ISD::ArgFlagsTy Flags = Outs[i].Flags; | |||
| 4764 | if (VA.getLocInfo() == CCValAssign::Indirect) | |||
| 4765 | return false; | |||
| 4766 | if (!VA.isRegLoc()) { | |||
| 4767 | if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags, | |||
| 4768 | MFI, MRI, TII, VA)) | |||
| 4769 | return false; | |||
| 4770 | } | |||
| 4771 | } | |||
| 4772 | } | |||
| 4773 | ||||
| 4774 | bool PositionIndependent = isPositionIndependent(); | |||
| 4775 | // If the tailcall address may be in a register, then make sure it's | |||
| 4776 | // possible to register allocate for it. In 32-bit, the call address can | |||
| 4777 | // only target EAX, EDX, or ECX since the tail call must be scheduled after | |||
| 4778 | // callee-saved registers are restored. These happen to be the same | |||
| 4779 | // registers used to pass 'inreg' arguments so watch out for those. | |||
| 4780 | if (!Subtarget.is64Bit() && ((!isa<GlobalAddressSDNode>(Callee) && | |||
| 4781 | !isa<ExternalSymbolSDNode>(Callee)) || | |||
| 4782 | PositionIndependent)) { | |||
| 4783 | unsigned NumInRegs = 0; | |||
| 4784 | // In PIC we need an extra register to formulate the address computation | |||
| 4785 | // for the callee. | |||
| 4786 | unsigned MaxInRegs = PositionIndependent ? 2 : 3; | |||
| 4787 | ||||
| 4788 | for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) { | |||
| 4789 | CCValAssign &VA = ArgLocs[i]; | |||
| 4790 | if (!VA.isRegLoc()) | |||
| 4791 | continue; | |||
| 4792 | Register Reg = VA.getLocReg(); | |||
| 4793 | switch (Reg) { | |||
| 4794 | default: break; | |||
| 4795 | case X86::EAX: case X86::EDX: case X86::ECX: | |||
| 4796 | if (++NumInRegs == MaxInRegs) | |||
| 4797 | return false; | |||
| 4798 | break; | |||
| 4799 | } | |||
| 4800 | } | |||
| 4801 | } | |||
| 4802 | ||||
| 4803 | const MachineRegisterInfo &MRI = MF.getRegInfo(); | |||
| 4804 | if (!parametersInCSRMatch(MRI, CallerPreserved, ArgLocs, OutVals)) | |||
| 4805 | return false; | |||
| 4806 | } | |||
| 4807 | ||||
| 4808 | bool CalleeWillPop = | |||
| 4809 | X86::isCalleePop(CalleeCC, Subtarget.is64Bit(), isVarArg, | |||
| 4810 | MF.getTarget().Options.GuaranteedTailCallOpt); | |||
| 4811 | ||||
| 4812 | if (unsigned BytesToPop = | |||
| 4813 | MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn()) { | |||
| 4814 | // If we have bytes to pop, the callee must pop them. | |||
| 4815 | bool CalleePopMatches = CalleeWillPop && BytesToPop == StackArgsSize; | |||
| 4816 | if (!CalleePopMatches) | |||
| 4817 | return false; | |||
| 4818 | } else if (CalleeWillPop && StackArgsSize > 0) { | |||
| 4819 | // If we don't have bytes to pop, make sure the callee doesn't pop any. | |||
| 4820 | return false; | |||
| 4821 | } | |||
| 4822 | ||||
| 4823 | return true; | |||
| 4824 | } | |||
| 4825 | ||||
| 4826 | FastISel * | |||
| 4827 | X86TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo, | |||
| 4828 | const TargetLibraryInfo *libInfo) const { | |||
| 4829 | return X86::createFastISel(funcInfo, libInfo); | |||
| 4830 | } | |||
| 4831 | ||||
| 4832 | //===----------------------------------------------------------------------===// | |||
| 4833 | // Other Lowering Hooks | |||
| 4834 | //===----------------------------------------------------------------------===// | |||
| 4835 | ||||
| 4836 | static bool MayFoldLoad(SDValue Op) { | |||
| 4837 | return Op.hasOneUse() && ISD::isNormalLoad(Op.getNode()); | |||
| 4838 | } | |||
| 4839 | ||||
| 4840 | static bool MayFoldIntoStore(SDValue Op) { | |||
| 4841 | return Op.hasOneUse() && ISD::isNormalStore(*Op.getNode()->use_begin()); | |||
| 4842 | } | |||
| 4843 | ||||
| 4844 | static bool MayFoldIntoZeroExtend(SDValue Op) { | |||
| 4845 | if (Op.hasOneUse()) { | |||
| 4846 | unsigned Opcode = Op.getNode()->use_begin()->getOpcode(); | |||
| 4847 | return (ISD::ZERO_EXTEND == Opcode); | |||
| 4848 | } | |||
| 4849 | return false; | |||
| 4850 | } | |||
| 4851 | ||||
| 4852 | static bool isTargetShuffle(unsigned Opcode) { | |||
| 4853 | switch(Opcode) { | |||
| 4854 | default: return false; | |||
| 4855 | case X86ISD::BLENDI: | |||
| 4856 | case X86ISD::PSHUFB: | |||
| 4857 | case X86ISD::PSHUFD: | |||
| 4858 | case X86ISD::PSHUFHW: | |||
| 4859 | case X86ISD::PSHUFLW: | |||
| 4860 | case X86ISD::SHUFP: | |||
| 4861 | case X86ISD::INSERTPS: | |||
| 4862 | case X86ISD::EXTRQI: | |||
| 4863 | case X86ISD::INSERTQI: | |||
| 4864 | case X86ISD::VALIGN: | |||
| 4865 | case X86ISD::PALIGNR: | |||
| 4866 | case X86ISD::VSHLDQ: | |||
| 4867 | case X86ISD::VSRLDQ: | |||
| 4868 | case X86ISD::MOVLHPS: | |||
| 4869 | case X86ISD::MOVHLPS: | |||
| 4870 | case X86ISD::MOVSHDUP: | |||
| 4871 | case X86ISD::MOVSLDUP: | |||
| 4872 | case X86ISD::MOVDDUP: | |||
| 4873 | case X86ISD::MOVSS: | |||
| 4874 | case X86ISD::MOVSD: | |||
| 4875 | case X86ISD::UNPCKL: | |||
| 4876 | case X86ISD::UNPCKH: | |||
| 4877 | case X86ISD::VBROADCAST: | |||
| 4878 | case X86ISD::VPERMILPI: | |||
| 4879 | case X86ISD::VPERMILPV: | |||
| 4880 | case X86ISD::VPERM2X128: | |||
| 4881 | case X86ISD::SHUF128: | |||
| 4882 | case X86ISD::VPERMIL2: | |||
| 4883 | case X86ISD::VPERMI: | |||
| 4884 | case X86ISD::VPPERM: | |||
| 4885 | case X86ISD::VPERMV: | |||
| 4886 | case X86ISD::VPERMV3: | |||
| 4887 | case X86ISD::VZEXT_MOVL: | |||
| 4888 | return true; | |||
| 4889 | } | |||
| 4890 | } | |||
| 4891 | ||||
| 4892 | static bool isTargetShuffleVariableMask(unsigned Opcode) { | |||
| 4893 | switch (Opcode) { | |||
| 4894 | default: return false; | |||
| 4895 | // Target Shuffles. | |||
| 4896 | case X86ISD::PSHUFB: | |||
| 4897 | case X86ISD::VPERMILPV: | |||
| 4898 | case X86ISD::VPERMIL2: | |||
| 4899 | case X86ISD::VPPERM: | |||
| 4900 | case X86ISD::VPERMV: | |||
| 4901 | case X86ISD::VPERMV3: | |||
| 4902 | return true; | |||
| 4903 | // 'Faux' Target Shuffles. | |||
| 4904 | case ISD::OR: | |||
| 4905 | case ISD::AND: | |||
| 4906 | case X86ISD::ANDNP: | |||
| 4907 | return true; | |||
| 4908 | } | |||
| 4909 | } | |||
| 4910 | ||||
| 4911 | static bool isTargetShuffleSplat(SDValue Op) { | |||
| 4912 | unsigned Opcode = Op.getOpcode(); | |||
| 4913 | if (Opcode == ISD::EXTRACT_SUBVECTOR) | |||
| 4914 | return isTargetShuffleSplat(Op.getOperand(0)); | |||
| 4915 | return Opcode == X86ISD::VBROADCAST || Opcode == X86ISD::VBROADCAST_LOAD; | |||
| 4916 | } | |||
| 4917 | ||||
| 4918 | SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) const { | |||
| 4919 | MachineFunction &MF = DAG.getMachineFunction(); | |||
| 4920 | const X86RegisterInfo *RegInfo = Subtarget.getRegisterInfo(); | |||
| 4921 | X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>(); | |||
| 4922 | int ReturnAddrIndex = FuncInfo->getRAIndex(); | |||
| 4923 | ||||
| 4924 | if (ReturnAddrIndex == 0) { | |||
| 4925 | // Set up a frame object for the return address. | |||
| 4926 | unsigned SlotSize = RegInfo->getSlotSize(); | |||
| 4927 | ReturnAddrIndex = MF.getFrameInfo().CreateFixedObject(SlotSize, | |||
| 4928 | -(int64_t)SlotSize, | |||
| 4929 | false); | |||
| 4930 | FuncInfo->setRAIndex(ReturnAddrIndex); | |||
| 4931 | } | |||
| 4932 | ||||
| 4933 | return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy(DAG.getDataLayout())); | |||
| 4934 | } | |||
| 4935 | ||||
| 4936 | bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M, | |||
| 4937 | bool hasSymbolicDisplacement) { | |||
| 4938 | // Offset should fit into 32 bit immediate field. | |||
| 4939 | if (!isInt<32>(Offset)) | |||
| 4940 | return false; | |||
| 4941 | ||||
| 4942 | // If we don't have a symbolic displacement - we don't have any extra | |||
| 4943 | // restrictions. | |||
| 4944 | if (!hasSymbolicDisplacement) | |||
| 4945 | return true; | |||
| 4946 | ||||
| 4947 | // FIXME: Some tweaks might be needed for medium code model. | |||
| 4948 | if (M != CodeModel::Small && M != CodeModel::Kernel) | |||
| 4949 | return false; | |||
| 4950 | ||||
| 4951 | // For small code model we assume that latest object is 16MB before end of 31 | |||
| 4952 | // bits boundary. We may also accept pretty large negative constants knowing | |||
| 4953 | // that all objects are in the positive half of address space. | |||
| 4954 | if (M == CodeModel::Small && Offset < 16*1024*1024) | |||
| 4955 | return true; | |||
| 4956 | ||||
| 4957 | // For kernel code model we know that all object resist in the negative half | |||
| 4958 | // of 32bits address space. We may not accept negative offsets, since they may | |||
| 4959 | // be just off and we may accept pretty large positive ones. | |||
| 4960 | if (M == CodeModel::Kernel && Offset >= 0) | |||
| 4961 | return true; | |||
| 4962 | ||||
| 4963 | return false; | |||
| 4964 | } | |||
| 4965 | ||||
| 4966 | /// Determines whether the callee is required to pop its own arguments. | |||
| 4967 | /// Callee pop is necessary to support tail calls. | |||
| 4968 | bool X86::isCalleePop(CallingConv::ID CallingConv, | |||
| 4969 | bool is64Bit, bool IsVarArg, bool GuaranteeTCO) { | |||
| 4970 | // If GuaranteeTCO is true, we force some calls to be callee pop so that we | |||
| 4971 | // can guarantee TCO. | |||
| 4972 | if (!IsVarArg && shouldGuaranteeTCO(CallingConv, GuaranteeTCO)) | |||
| 4973 | return true; | |||
| 4974 | ||||
| 4975 | switch (CallingConv) { | |||
| 4976 | default: | |||
| 4977 | return false; | |||
| 4978 | case CallingConv::X86_StdCall: | |||
| 4979 | case CallingConv::X86_FastCall: | |||
| 4980 | case CallingConv::X86_ThisCall: | |||
| 4981 | case CallingConv::X86_VectorCall: | |||
| 4982 | return !is64Bit; | |||
| 4983 | } | |||
| 4984 | } | |||
| 4985 | ||||
| 4986 | /// Return true if the condition is an signed comparison operation. | |||
| 4987 | static bool isX86CCSigned(unsigned X86CC) { | |||
| 4988 | switch (X86CC) { | |||
| 4989 | default: | |||
| 4990 | llvm_unreachable("Invalid integer condition!")__builtin_unreachable(); | |||
| 4991 | case X86::COND_E: | |||
| 4992 | case X86::COND_NE: | |||
| 4993 | case X86::COND_B: | |||
| 4994 | case X86::COND_A: | |||
| 4995 | case X86::COND_BE: | |||
| 4996 | case X86::COND_AE: | |||
| 4997 | return false; | |||
| 4998 | case X86::COND_G: | |||
| 4999 | case X86::COND_GE: | |||
| 5000 | case X86::COND_L: | |||
| 5001 | case X86::COND_LE: | |||
| 5002 | return true; | |||
| 5003 | } | |||
| 5004 | } | |||
| 5005 | ||||
| 5006 | static X86::CondCode TranslateIntegerX86CC(ISD::CondCode SetCCOpcode) { | |||
| 5007 | switch (SetCCOpcode) { | |||
| 5008 | default: llvm_unreachable("Invalid integer condition!")__builtin_unreachable(); | |||
| 5009 | case ISD::SETEQ: return X86::COND_E; | |||
| 5010 | case ISD::SETGT: return X86::COND_G; | |||
| 5011 | case ISD::SETGE: return X86::COND_GE; | |||
| 5012 | case ISD::SETLT: return X86::COND_L; | |||
| 5013 | case ISD::SETLE: return X86::COND_LE; | |||
| 5014 | case ISD::SETNE: return X86::COND_NE; | |||
| 5015 | case ISD::SETULT: return X86::COND_B; | |||
| 5016 | case ISD::SETUGT: return X86::COND_A; | |||
| 5017 | case ISD::SETULE: return X86::COND_BE; | |||
| 5018 | case ISD::SETUGE: return X86::COND_AE; | |||
| 5019 | } | |||
| 5020 | } | |||
| 5021 | ||||
| 5022 | /// Do a one-to-one translation of a ISD::CondCode to the X86-specific | |||
| 5023 | /// condition code, returning the condition code and the LHS/RHS of the | |||
| 5024 | /// comparison to make. | |||
| 5025 | static X86::CondCode TranslateX86CC(ISD::CondCode SetCCOpcode, const SDLoc &DL, | |||
| 5026 | bool isFP, SDValue &LHS, SDValue &RHS, | |||
| 5027 | SelectionDAG &DAG) { | |||
| 5028 | if (!isFP) { | |||
| 5029 | if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) { | |||
| 5030 | if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) { | |||
| 5031 | // X > -1 -> X == 0, jump !sign. | |||
| 5032 | RHS = DAG.getConstant(0, DL, RHS.getValueType()); | |||
| 5033 | return X86::COND_NS; | |||
| 5034 | } | |||
| 5035 | if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) { | |||
| 5036 | // X < 0 -> X == 0, jump on sign. | |||
| 5037 | return X86::COND_S; | |||
| 5038 | } | |||
| 5039 | if (SetCCOpcode == ISD::SETGE && RHSC->isNullValue()) { | |||
| 5040 | // X >= 0 -> X == 0, jump on !sign. | |||
| 5041 | return X86::COND_NS; | |||
| 5042 | } | |||
| 5043 | if (SetCCOpcode == ISD::SETLT && RHSC->isOne()) { | |||
| 5044 | // X < 1 -> X <= 0 | |||
| 5045 | RHS = DAG.getConstant(0, DL, RHS.getValueType()); | |||
| 5046 | return X86::COND_LE; | |||
| 5047 | } | |||
| 5048 | } | |||
| 5049 | ||||
| 5050 | return TranslateIntegerX86CC(SetCCOpcode); | |||
| 5051 | } | |||
| 5052 | ||||
| 5053 | // First determine if it is required or is profitable to flip the operands. | |||
| 5054 | ||||
| 5055 | // If LHS is a foldable load, but RHS is not, flip the condition. | |||
| 5056 | if (ISD::isNON_EXTLoad(LHS.getNode()) && | |||
| 5057 | !ISD::isNON_EXTLoad(RHS.getNode())) { | |||
| 5058 | SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode); | |||
| 5059 | std::swap(LHS, RHS); | |||
| 5060 | } | |||
| 5061 | ||||
| 5062 | switch (SetCCOpcode) { | |||
| 5063 | default: break; | |||
| 5064 | case ISD::SETOLT: | |||
| 5065 | case ISD::SETOLE: | |||
| 5066 | case ISD::SETUGT: | |||
| 5067 | case ISD::SETUGE: | |||
| 5068 | std::swap(LHS, RHS); | |||
| 5069 | break; | |||
| 5070 | } | |||
| 5071 | ||||
| 5072 | // On a floating point condition, the flags are set as follows: | |||
| 5073 | // ZF PF CF op | |||
| 5074 | // 0 | 0 | 0 | X > Y | |||
| 5075 | // 0 | 0 | 1 | X < Y | |||
| 5076 | // 1 | 0 | 0 | X == Y | |||
| 5077 | // 1 | 1 | 1 | unordered | |||
| 5078 | switch (SetCCOpcode) { | |||
| 5079 | default: llvm_unreachable("Condcode should be pre-legalized away")__builtin_unreachable(); | |||
| 5080 | case ISD::SETUEQ: | |||
| 5081 | case ISD::SETEQ: return X86::COND_E; | |||
| 5082 | case ISD::SETOLT: // flipped | |||
| 5083 | case ISD::SETOGT: | |||
| 5084 | case ISD::SETGT: return X86::COND_A; | |||
| 5085 | case ISD::SETOLE: // flipped | |||
| 5086 | case ISD::SETOGE: | |||
| 5087 | case ISD::SETGE: return X86::COND_AE; | |||
| 5088 | case ISD::SETUGT: // flipped | |||
| 5089 | case ISD::SETULT: | |||
| 5090 | case ISD::SETLT: return X86::COND_B; | |||
| 5091 | case ISD::SETUGE: // flipped | |||
| 5092 | case ISD::SETULE: | |||
| 5093 | case ISD::SETLE: return X86::COND_BE; | |||
| 5094 | case ISD::SETONE: | |||
| 5095 | case ISD::SETNE: return X86::COND_NE; | |||
| 5096 | case ISD::SETUO: return X86::COND_P; | |||
| 5097 | case ISD::SETO: return X86::COND_NP; | |||
| 5098 | case ISD::SETOEQ: | |||
| 5099 | case ISD::SETUNE: return X86::COND_INVALID; | |||
| 5100 | } | |||
| 5101 | } | |||
| 5102 | ||||
| 5103 | /// Is there a floating point cmov for the specific X86 condition code? | |||
| 5104 | /// Current x86 isa includes the following FP cmov instructions: | |||
| 5105 | /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu. | |||
| 5106 | static bool hasFPCMov(unsigned X86CC) { | |||
| 5107 | switch (X86CC) { | |||
| 5108 | default: | |||
| 5109 | return false; | |||
| 5110 | case X86::COND_B: | |||
| 5111 | case X86::COND_BE: | |||
| 5112 | case X86::COND_E: | |||
| 5113 | case X86::COND_P: | |||
| 5114 | case X86::COND_A: | |||
| 5115 | case X86::COND_AE: | |||
| 5116 | case X86::COND_NE: | |||
| 5117 | case X86::COND_NP: | |||
| 5118 | return true; | |||
| 5119 | } | |||
| 5120 | } | |||
| 5121 | ||||
| 5122 | ||||
| 5123 | bool X86TargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info, | |||
| 5124 | const CallInst &I, | |||
| 5125 | MachineFunction &MF, | |||
| 5126 | unsigned Intrinsic) const { | |||
| 5127 | Info.flags = MachineMemOperand::MONone; | |||
| 5128 | Info.offset = 0; | |||
| 5129 | ||||
| 5130 | const IntrinsicData* IntrData = getIntrinsicWithChain(Intrinsic); | |||
| 5131 | if (!IntrData) { | |||
| 5132 | switch (Intrinsic) { | |||
| 5133 | case Intrinsic::x86_aesenc128kl: | |||
| 5134 | case Intrinsic::x86_aesdec128kl: | |||
| 5135 | Info.opc = ISD::INTRINSIC_W_CHAIN; | |||
| 5136 | Info.ptrVal = I.getArgOperand(1); | |||
| 5137 | Info.memVT = EVT::getIntegerVT(I.getType()->getContext(), 48); | |||
| 5138 | Info.align = Align(1); | |||
| 5139 | Info.flags |= MachineMemOperand::MOLoad; | |||
| 5140 | return true; | |||
| 5141 | case Intrinsic::x86_aesenc256kl: | |||
| 5142 | case Intrinsic::x86_aesdec256kl: | |||
| 5143 | Info.opc = ISD::INTRINSIC_W_CHAIN; | |||
| 5144 | Info.ptrVal = I.getArgOperand(1); | |||
| 5145 | Info.memVT = EVT::getIntegerVT(I.getType()->getContext(), 64); | |||
| 5146 | Info.align = Align(1); | |||
| 5147 | Info.flags |= MachineMemOperand::MOLoad; | |||
| 5148 | return true; | |||
| 5149 | case Intrinsic::x86_aesencwide128kl: | |||
| 5150 | case Intrinsic::x86_aesdecwide128kl: | |||
| 5151 | Info.opc = ISD::INTRINSIC_W_CHAIN; | |||
| 5152 | Info.ptrVal = I.getArgOperand(0); | |||
| 5153 | Info.memVT = EVT::getIntegerVT(I.getType()->getContext(), 48); | |||
| 5154 | Info.align = Align(1); | |||
| 5155 | Info.flags |= MachineMemOperand::MOLoad; | |||
| 5156 | return true; | |||
| 5157 | case Intrinsic::x86_aesencwide256kl: | |||
| 5158 | case Intrinsic::x86_aesdecwide256kl: | |||
| 5159 | Info.opc = ISD::INTRINSIC_W_CHAIN; | |||
| 5160 | Info.ptrVal = I.getArgOperand(0); | |||
| 5161 | Info.memVT = EVT::getIntegerVT(I.getType()->getContext(), 64); | |||
| 5162 | Info.align = Align(1); | |||
| 5163 | Info.flags |= MachineMemOperand::MOLoad; | |||
| 5164 | return true; | |||
| 5165 | } | |||
| 5166 | return false; | |||
| 5167 | } | |||
| 5168 | ||||
| 5169 | switch (IntrData->Type) { | |||
| 5170 | case TRUNCATE_TO_MEM_VI8: | |||
| 5171 | case TRUNCATE_TO_MEM_VI16: | |||
| 5172 | case TRUNCATE_TO_MEM_VI32: { | |||
| 5173 | Info.opc = ISD::INTRINSIC_VOID; | |||
| 5174 | Info.ptrVal = I.getArgOperand(0); | |||
| 5175 | MVT VT = MVT::getVT(I.getArgOperand(1)->getType()); | |||
| 5176 | MVT ScalarVT = MVT::INVALID_SIMPLE_VALUE_TYPE; | |||
| 5177 | if (IntrData->Type == TRUNCATE_TO_MEM_VI8) | |||
| 5178 | ScalarVT = MVT::i8; | |||
| 5179 | else if (IntrData->Type == TRUNCATE_TO_MEM_VI16) | |||
| 5180 | ScalarVT = MVT::i16; | |||
| 5181 | else if (IntrData->Type == TRUNCATE_TO_MEM_VI32) | |||
| 5182 | ScalarVT = MVT::i32; | |||
| 5183 | ||||
| 5184 | Info.memVT = MVT::getVectorVT(ScalarVT, VT.getVectorNumElements()); | |||
| 5185 | Info.align = Align(1); | |||
| 5186 | Info.flags |= MachineMemOperand::MOStore; | |||
| 5187 | break; | |||
| 5188 | } | |||
| 5189 | case GATHER: | |||
| 5190 | case GATHER_AVX2: { | |||
| 5191 | Info.opc = ISD::INTRINSIC_W_CHAIN; | |||
| 5192 | Info.ptrVal = nullptr; | |||
| 5193 | MVT DataVT = MVT::getVT(I.getType()); | |||
| 5194 | MVT IndexVT = MVT::getVT(I.getArgOperand(2)->getType()); | |||
| 5195 | unsigned NumElts = std::min(DataVT.getVectorNumElements(), | |||
| 5196 | IndexVT.getVectorNumElements()); | |||
| 5197 | Info.memVT = MVT::getVectorVT(DataVT.getVectorElementType(), NumElts); | |||
| 5198 | Info.align = Align(1); | |||
| 5199 | Info.flags |= MachineMemOperand::MOLoad; | |||
| 5200 | break; | |||
| 5201 | } | |||
| 5202 | case SCATTER: { | |||
| 5203 | Info.opc = ISD::INTRINSIC_VOID; | |||
| 5204 | Info.ptrVal = nullptr; | |||
| 5205 | MVT DataVT = MVT::getVT(I.getArgOperand(3)->getType()); | |||
| 5206 | MVT IndexVT = MVT::getVT(I.getArgOperand(2)->getType()); | |||
| 5207 | unsigned NumElts = std::min(DataVT.getVectorNumElements(), | |||
| 5208 | IndexVT.getVectorNumElements()); | |||
| 5209 | Info.memVT = MVT::getVectorVT(DataVT.getVectorElementType(), NumElts); | |||
| 5210 | Info.align = Align(1); | |||
| 5211 | Info.flags |= MachineMemOperand::MOStore; | |||
| 5212 | break; | |||
| 5213 | } | |||
| 5214 | default: | |||
| 5215 | return false; | |||
| 5216 | } | |||
| 5217 | ||||
| 5218 | return true; | |||
| 5219 | } | |||
| 5220 | ||||
| 5221 | /// Returns true if the target can instruction select the | |||
| 5222 | /// specified FP immediate natively. If false, the legalizer will | |||
| 5223 | /// materialize the FP immediate as a load from a constant pool. | |||
| 5224 | bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT, | |||
| 5225 | bool ForCodeSize) const { | |||
| 5226 | for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) { | |||
| 5227 | if (Imm.bitwiseIsEqual(LegalFPImmediates[i])) | |||
| 5228 | return true; | |||
| 5229 | } | |||
| 5230 | return false; | |||
| 5231 | } | |||
| 5232 | ||||
| 5233 | bool X86TargetLowering::shouldReduceLoadWidth(SDNode *Load, | |||
| 5234 | ISD::LoadExtType ExtTy, | |||
| 5235 | EVT NewVT) const { | |||
| 5236 | assert(cast<LoadSDNode>(Load)->isSimple() && "illegal to narrow")((void)0); | |||
| 5237 | ||||
| 5238 | // "ELF Handling for Thread-Local Storage" specifies that R_X86_64_GOTTPOFF | |||
| 5239 | // relocation target a movq or addq instruction: don't let the load shrink. | |||
| 5240 | SDValue BasePtr = cast<LoadSDNode>(Load)->getBasePtr(); | |||
| 5241 | if (BasePtr.getOpcode() == X86ISD::WrapperRIP) | |||
| 5242 | if (const auto *GA = dyn_cast<GlobalAddressSDNode>(BasePtr.getOperand(0))) | |||
| 5243 | return GA->getTargetFlags() != X86II::MO_GOTTPOFF; | |||
| 5244 | ||||
| 5245 | // If this is an (1) AVX vector load with (2) multiple uses and (3) all of | |||
| 5246 | // those uses are extracted directly into a store, then the extract + store | |||
| 5247 | // can be store-folded. Therefore, it's probably not worth splitting the load. | |||
| 5248 | EVT VT = Load->getValueType(0); | |||
| 5249 | if ((VT.is256BitVector() || VT.is512BitVector()) && !Load->hasOneUse()) { | |||
| 5250 | for (auto UI = Load->use_begin(), UE = Load->use_end(); UI != UE; ++UI) { | |||
| 5251 | // Skip uses of the chain value. Result 0 of the node is the load value. | |||
| 5252 | if (UI.getUse().getResNo() != 0) | |||
| 5253 | continue; | |||
| 5254 | ||||
| 5255 | // If this use is not an extract + store, it's probably worth splitting. | |||
| 5256 | if (UI->getOpcode() != ISD::EXTRACT_SUBVECTOR || !UI->hasOneUse() || | |||
| 5257 | UI->use_begin()->getOpcode() != ISD::STORE) | |||
| 5258 | return true; | |||
| 5259 | } | |||
| 5260 | // All non-chain uses are extract + store. | |||
| 5261 | return false; | |||
| 5262 | } | |||
| 5263 | ||||
| 5264 | return true; | |||
| 5265 | } | |||
| 5266 | ||||
| 5267 | /// Returns true if it is beneficial to convert a load of a constant | |||
| 5268 | /// to just the constant itself. | |||
| 5269 | bool X86TargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm, | |||
| 5270 | Type *Ty) const { | |||
| 5271 | assert(Ty->isIntegerTy())((void)0); | |||
| 5272 | ||||
| 5273 | unsigned BitSize = Ty->getPrimitiveSizeInBits(); | |||
| 5274 | if (BitSize == 0 || BitSize > 64) | |||
| 5275 | return false; | |||
| 5276 | return true; | |||
| 5277 | } | |||
| 5278 | ||||
| 5279 | bool X86TargetLowering::reduceSelectOfFPConstantLoads(EVT CmpOpVT) const { | |||
| 5280 | // If we are using XMM registers in the ABI and the condition of the select is | |||
| 5281 | // a floating-point compare and we have blendv or conditional move, then it is | |||
| 5282 | // cheaper to select instead of doing a cross-register move and creating a | |||
| 5283 | // load that depends on the compare result. | |||
| 5284 | bool IsFPSetCC = CmpOpVT.isFloatingPoint() && CmpOpVT != MVT::f128; | |||
| 5285 | return !IsFPSetCC || !Subtarget.isTarget64BitLP64() || !Subtarget.hasAVX(); | |||
| 5286 | } | |||
| 5287 | ||||
| 5288 | bool X86TargetLowering::convertSelectOfConstantsToMath(EVT VT) const { | |||
| 5289 | // TODO: It might be a win to ease or lift this restriction, but the generic | |||
| 5290 | // folds in DAGCombiner conflict with vector folds for an AVX512 target. | |||
| 5291 | if (VT.isVector() && Subtarget.hasAVX512()) | |||
| 5292 | return false; | |||
| 5293 | ||||
| 5294 | return true; | |||
| 5295 | } | |||
| 5296 | ||||
| 5297 | bool X86TargetLowering::decomposeMulByConstant(LLVMContext &Context, EVT VT, | |||
| 5298 | SDValue C) const { | |||
| 5299 | // TODO: We handle scalars using custom code, but generic combining could make | |||
| 5300 | // that unnecessary. | |||
| 5301 | APInt MulC; | |||
| 5302 | if (!ISD::isConstantSplatVector(C.getNode(), MulC)) | |||
| 5303 | return false; | |||
| 5304 | ||||
| 5305 | // Find the type this will be legalized too. Otherwise we might prematurely | |||
| 5306 | // convert this to shl+add/sub and then still have to type legalize those ops. | |||
| 5307 | // Another choice would be to defer the decision for illegal types until | |||
| 5308 | // after type legalization. But constant splat vectors of i64 can't make it | |||
| 5309 | // through type legalization on 32-bit targets so we would need to special | |||
| 5310 | // case vXi64. | |||
| 5311 | while (getTypeAction(Context, VT) != TypeLegal) | |||
| 5312 | VT = getTypeToTransformTo(Context, VT); | |||
| 5313 | ||||
| 5314 | // If vector multiply is legal, assume that's faster than shl + add/sub. | |||
| 5315 | // TODO: Multiply is a complex op with higher latency and lower throughput in | |||
| 5316 | // most implementations, so this check could be loosened based on type | |||
| 5317 | // and/or a CPU attribute. | |||
| 5318 | if (isOperationLegal(ISD::MUL, VT)) | |||
| 5319 | return false; | |||
| 5320 | ||||
| 5321 | // shl+add, shl+sub, shl+add+neg | |||
| 5322 | return (MulC + 1).isPowerOf2() || (MulC - 1).isPowerOf2() || | |||
| 5323 | (1 - MulC).isPowerOf2() || (-(MulC + 1)).isPowerOf2(); | |||
| 5324 | } | |||
| 5325 | ||||
| 5326 | bool X86TargetLowering::isExtractSubvectorCheap(EVT ResVT, EVT SrcVT, | |||
| 5327 | unsigned Index) const { | |||
| 5328 | if (!isOperationLegalOrCustom(ISD::EXTRACT_SUBVECTOR, ResVT)) | |||
| 5329 | return false; | |||
| 5330 | ||||
| 5331 | // Mask vectors support all subregister combinations and operations that | |||
| 5332 | // extract half of vector. | |||
| 5333 | if (ResVT.getVectorElementType() == MVT::i1) | |||
| 5334 | return Index == 0 || ((ResVT.getSizeInBits() == SrcVT.getSizeInBits()*2) && | |||
| 5335 | (Index == ResVT.getVectorNumElements())); | |||
| 5336 | ||||
| 5337 | return (Index % ResVT.getVectorNumElements()) == 0; | |||
| 5338 | } | |||
| 5339 | ||||
| 5340 | bool X86TargetLowering::shouldScalarizeBinop(SDValue VecOp) const { | |||
| 5341 | unsigned Opc = VecOp.getOpcode(); | |||
| 5342 | ||||
| 5343 | // Assume target opcodes can't be scalarized. | |||
| 5344 | // TODO - do we have any exceptions? | |||
| 5345 | if (Opc >= ISD::BUILTIN_OP_END) | |||
| 5346 | return false; | |||
| 5347 | ||||
| 5348 | // If the vector op is not supported, try to convert to scalar. | |||
| 5349 | EVT VecVT = VecOp.getValueType(); | |||
| 5350 | if (!isOperationLegalOrCustomOrPromote(Opc, VecVT)) | |||
| 5351 | return true; | |||
| 5352 | ||||
| 5353 | // If the vector op is supported, but the scalar op is not, the transform may | |||
| 5354 | // not be worthwhile. | |||
| 5355 | EVT ScalarVT = VecVT.getScalarType(); | |||
| 5356 | return isOperationLegalOrCustomOrPromote(Opc, ScalarVT); | |||
| 5357 | } | |||
| 5358 | ||||
| 5359 | bool X86TargetLowering::shouldFormOverflowOp(unsigned Opcode, EVT VT, | |||
| 5360 | bool) const { | |||
| 5361 | // TODO: Allow vectors? | |||
| 5362 | if (VT.isVector()) | |||
| 5363 | return false; | |||
| 5364 | return VT.isSimple() || !isOperationExpand(Opcode, VT); | |||
| 5365 | } | |||
| 5366 | ||||
| 5367 | bool X86TargetLowering::isCheapToSpeculateCttz() const { | |||
| 5368 | // Speculate cttz only if we can directly use TZCNT. | |||
| 5369 | return Subtarget.hasBMI(); | |||
| 5370 | } | |||
| 5371 | ||||
| 5372 | bool X86TargetLowering::isCheapToSpeculateCtlz() const { | |||
| 5373 | // Speculate ctlz only if we can directly use LZCNT. | |||
| 5374 | return Subtarget.hasLZCNT(); | |||
| 5375 | } | |||
| 5376 | ||||
| 5377 | bool X86TargetLowering::isLoadBitCastBeneficial(EVT LoadVT, EVT BitcastVT, | |||
| 5378 | const SelectionDAG &DAG, | |||
| 5379 | const MachineMemOperand &MMO) const { | |||
| 5380 | if (!Subtarget.hasAVX512() && !LoadVT.isVector() && BitcastVT.isVector() && | |||
| 5381 | BitcastVT.getVectorElementType() == MVT::i1) | |||
| 5382 | return false; | |||
| 5383 | ||||
| 5384 | if (!Subtarget.hasDQI() && BitcastVT == MVT::v8i1 && LoadVT == MVT::i8) | |||
| 5385 | return false; | |||
| 5386 | ||||
| 5387 | // If both types are legal vectors, it's always ok to convert them. | |||
| 5388 | if (LoadVT.isVector() && BitcastVT.isVector() && | |||
| 5389 | isTypeLegal(LoadVT) && isTypeLegal(BitcastVT)) | |||
| 5390 | return true; | |||
| 5391 | ||||
| 5392 | return TargetLowering::isLoadBitCastBeneficial(LoadVT, BitcastVT, DAG, MMO); | |||
| 5393 | } | |||
| 5394 | ||||
| 5395 | bool X86TargetLowering::canMergeStoresTo(unsigned AddressSpace, EVT MemVT, | |||
| 5396 | const SelectionDAG &DAG) const { | |||
| 5397 | // Do not merge to float value size (128 bytes) if no implicit | |||
| 5398 | // float attribute is set. | |||
| 5399 | bool NoFloat = DAG.getMachineFunction().getFunction().hasFnAttribute( | |||
| 5400 | Attribute::NoImplicitFloat); | |||
| 5401 | ||||
| 5402 | if (NoFloat) { | |||
| 5403 | unsigned MaxIntSize = Subtarget.is64Bit() ? 64 : 32; | |||
| 5404 | return (MemVT.getSizeInBits() <= MaxIntSize); | |||
| 5405 | } | |||
| 5406 | // Make sure we don't merge greater than our preferred vector | |||
| 5407 | // width. | |||
| 5408 | if (MemVT.getSizeInBits() > Subtarget.getPreferVectorWidth()) | |||
| 5409 | return false; | |||
| 5410 | ||||
| 5411 | return true; | |||
| 5412 | } | |||
| 5413 | ||||
| 5414 | bool X86TargetLowering::isCtlzFast() const { | |||
| 5415 | return Subtarget.hasFastLZCNT(); | |||
| 5416 | } | |||
| 5417 | ||||
| 5418 | bool X86TargetLowering::isMaskAndCmp0FoldingBeneficial( | |||
| 5419 | const Instruction &AndI) const { | |||
| 5420 | return true; | |||
| 5421 | } | |||
| 5422 | ||||
| 5423 | bool X86TargetLowering::hasAndNotCompare(SDValue Y) const { | |||
| 5424 | EVT VT = Y.getValueType(); | |||
| 5425 | ||||
| 5426 | if (VT.isVector()) | |||
| 5427 | return false; | |||
| 5428 | ||||
| 5429 | if (!Subtarget.hasBMI()) | |||
| 5430 | return false; | |||
| 5431 | ||||
| 5432 | // There are only 32-bit and 64-bit forms for 'andn'. | |||
| 5433 | if (VT != MVT::i32 && VT != MVT::i64) | |||
| 5434 | return false; | |||
| 5435 | ||||
| 5436 | return !isa<ConstantSDNode>(Y); | |||
| 5437 | } | |||
| 5438 | ||||
| 5439 | bool X86TargetLowering::hasAndNot(SDValue Y) const { | |||
| 5440 | EVT VT = Y.getValueType(); | |||
| 5441 | ||||
| 5442 | if (!VT.isVector()) | |||
| 5443 | return hasAndNotCompare(Y); | |||
| 5444 | ||||
| 5445 | // Vector. | |||
| 5446 | ||||
| 5447 | if (!Subtarget.hasSSE1() || VT.getSizeInBits() < 128) | |||
| 5448 | return false; | |||
| 5449 | ||||
| 5450 | if (VT == MVT::v4i32) | |||
| 5451 | return true; | |||
| 5452 | ||||
| 5453 | return Subtarget.hasSSE2(); | |||
| 5454 | } | |||
| 5455 | ||||
| 5456 | bool X86TargetLowering::hasBitTest(SDValue X, SDValue Y) const { | |||
| 5457 | return X.getValueType().isScalarInteger(); // 'bt' | |||
| 5458 | } | |||
| 5459 | ||||
| 5460 | bool X86TargetLowering:: | |||
| 5461 | shouldProduceAndByConstByHoistingConstFromShiftsLHSOfAnd( | |||
| 5462 | SDValue X, ConstantSDNode *XC, ConstantSDNode *CC, SDValue Y, | |||
| 5463 | unsigned OldShiftOpcode, unsigned NewShiftOpcode, | |||
| 5464 | SelectionDAG &DAG) const { | |||
| 5465 | // Does baseline recommend not to perform the fold by default? | |||
| 5466 | if (!TargetLowering::shouldProduceAndByConstByHoistingConstFromShiftsLHSOfAnd( | |||
| 5467 | X, XC, CC, Y, OldShiftOpcode, NewShiftOpcode, DAG)) | |||
| 5468 | return false; | |||
| 5469 | // For scalars this transform is always beneficial. | |||
| 5470 | if (X.getValueType().isScalarInteger()) | |||
| 5471 | return true; | |||
| 5472 | // If all the shift amounts are identical, then transform is beneficial even | |||
| 5473 | // with rudimentary SSE2 shifts. | |||
| 5474 | if (DAG.isSplatValue(Y, /*AllowUndefs=*/true)) | |||
| 5475 | return true; | |||
| 5476 | // If we have AVX2 with it's powerful shift operations, then it's also good. | |||
| 5477 | if (Subtarget.hasAVX2()) | |||
| 5478 | return true; | |||
| 5479 | // Pre-AVX2 vector codegen for this pattern is best for variant with 'shl'. | |||
| 5480 | return NewShiftOpcode == ISD::SHL; | |||
| 5481 | } | |||
| 5482 | ||||
| 5483 | bool X86TargetLowering::shouldFoldConstantShiftPairToMask( | |||
| 5484 | const SDNode *N, CombineLevel Level) const { | |||
| 5485 | assert(((N->getOpcode() == ISD::SHL &&((void)0) | |||
| 5486 | N->getOperand(0).getOpcode() == ISD::SRL) ||((void)0) | |||
| 5487 | (N->getOpcode() == ISD::SRL &&((void)0) | |||
| 5488 | N->getOperand(0).getOpcode() == ISD::SHL)) &&((void)0) | |||
| 5489 | "Expected shift-shift mask")((void)0); | |||
| 5490 | EVT VT = N->getValueType(0); | |||
| 5491 | if ((Subtarget.hasFastVectorShiftMasks() && VT.isVector()) || | |||
| 5492 | (Subtarget.hasFastScalarShiftMasks() && !VT.isVector())) { | |||
| 5493 | // Only fold if the shift values are equal - so it folds to AND. | |||
| 5494 | // TODO - we should fold if either is a non-uniform vector but we don't do | |||
| 5495 | // the fold for non-splats yet. | |||
| 5496 | return N->getOperand(1) == N->getOperand(0).getOperand(1); | |||
| 5497 | } | |||
| 5498 | return TargetLoweringBase::shouldFoldConstantShiftPairToMask(N, Level); | |||
| 5499 | } | |||
| 5500 | ||||
| 5501 | bool X86TargetLowering::shouldFoldMaskToVariableShiftPair(SDValue Y) const { | |||
| 5502 | EVT VT = Y.getValueType(); | |||
| 5503 | ||||
| 5504 | // For vectors, we don't have a preference, but we probably want a mask. | |||
| 5505 | if (VT.isVector()) | |||
| 5506 | return false; | |||
| 5507 | ||||
| 5508 | // 64-bit shifts on 32-bit targets produce really bad bloated code. | |||
| 5509 | if (VT == MVT::i64 && !Subtarget.is64Bit()) | |||
| 5510 | return false; | |||
| 5511 | ||||
| 5512 | return true; | |||
| 5513 | } | |||
| 5514 | ||||
| 5515 | bool X86TargetLowering::shouldExpandShift(SelectionDAG &DAG, | |||
| 5516 | SDNode *N) const { | |||
| 5517 | if (DAG.getMachineFunction().getFunction().hasMinSize() && | |||
| 5518 | !Subtarget.isOSWindows()) | |||
| 5519 | return false; | |||
| 5520 | return true; | |||
| 5521 | } | |||
| 5522 | ||||
| 5523 | bool X86TargetLowering::shouldSplatInsEltVarIndex(EVT VT) const { | |||
| 5524 | // Any legal vector type can be splatted more efficiently than | |||
| 5525 | // loading/spilling from memory. | |||
| 5526 | return isTypeLegal(VT); | |||
| 5527 | } | |||
| 5528 | ||||
| 5529 | MVT X86TargetLowering::hasFastEqualityCompare(unsigned NumBits) const { | |||
| 5530 | MVT VT = MVT::getIntegerVT(NumBits); | |||
| 5531 | if (isTypeLegal(VT)) | |||
| 5532 | return VT; | |||
| 5533 | ||||
| 5534 | // PMOVMSKB can handle this. | |||
| 5535 | if (NumBits == 128 && isTypeLegal(MVT::v16i8)) | |||
| 5536 | return MVT::v16i8; | |||
| 5537 | ||||
| 5538 | // VPMOVMSKB can handle this. | |||
| 5539 | if (NumBits == 256 && isTypeLegal(MVT::v32i8)) | |||
| 5540 | return MVT::v32i8; | |||
| 5541 | ||||
| 5542 | // TODO: Allow 64-bit type for 32-bit target. | |||
| 5543 | // TODO: 512-bit types should be allowed, but make sure that those | |||
| 5544 | // cases are handled in combineVectorSizedSetCCEquality(). | |||
| 5545 | ||||
| 5546 | return MVT::INVALID_SIMPLE_VALUE_TYPE; | |||
| 5547 | } | |||
| 5548 | ||||
| 5549 | /// Val is the undef sentinel value or equal to the specified value. | |||
| 5550 | static bool isUndefOrEqual(int Val, int CmpVal) { | |||
| 5551 | return ((Val == SM_SentinelUndef) || (Val == CmpVal)); | |||
| 5552 | } | |||
| 5553 | ||||
| 5554 | /// Return true if every element in Mask is the undef sentinel value or equal to | |||
| 5555 | /// the specified value.. | |||
| 5556 | static bool isUndefOrEqual(ArrayRef<int> Mask, int CmpVal) { | |||
| 5557 | return llvm::all_of(Mask, [CmpVal](int M) { | |||
| 5558 | return (M == SM_SentinelUndef) || (M == CmpVal); | |||
| 5559 | }); | |||
| 5560 | } | |||
| 5561 | ||||
| 5562 | /// Val is either the undef or zero sentinel value. | |||
| 5563 | static bool isUndefOrZero(int Val) { | |||
| 5564 | return ((Val == SM_SentinelUndef) || (Val == SM_SentinelZero)); | |||
| 5565 | } | |||
| 5566 | ||||
| 5567 | /// Return true if every element in Mask, beginning from position Pos and ending | |||
| 5568 | /// in Pos+Size is the undef sentinel value. | |||
| 5569 | static bool isUndefInRange(ArrayRef<int> Mask, unsigned Pos, unsigned Size) { | |||
| 5570 | return llvm::all_of(Mask.slice(Pos, Size), | |||
| 5571 | [](int M) { return M == SM_SentinelUndef; }); | |||
| 5572 | } | |||
| 5573 | ||||
| 5574 | /// Return true if the mask creates a vector whose lower half is undefined. | |||
| 5575 | static bool isUndefLowerHalf(ArrayRef<int> Mask) { | |||
| 5576 | unsigned NumElts = Mask.size(); | |||
| 5577 | return isUndefInRange(Mask, 0, NumElts / 2); | |||
| 5578 | } | |||
| 5579 | ||||
| 5580 | /// Return true if the mask creates a vector whose upper half is undefined. | |||
| 5581 | static bool isUndefUpperHalf(ArrayRef<int> Mask) { | |||
| 5582 | unsigned NumElts = Mask.size(); | |||
| 5583 | return isUndefInRange(Mask, NumElts / 2, NumElts / 2); | |||
| 5584 | } | |||
| 5585 | ||||
| 5586 | /// Return true if Val falls within the specified range (L, H]. | |||
| 5587 | static bool isInRange(int Val, int Low, int Hi) { | |||
| 5588 | return (Val >= Low && Val < Hi); | |||
| 5589 | } | |||
| 5590 | ||||
| 5591 | /// Return true if the value of any element in Mask falls within the specified | |||
| 5592 | /// range (L, H]. | |||
| 5593 | static bool isAnyInRange(ArrayRef<int> Mask, int Low, int Hi) { | |||
| 5594 | return llvm::any_of(Mask, [Low, Hi](int M) { return isInRange(M, Low, Hi); }); | |||
| 5595 | } | |||
| 5596 | ||||
| 5597 | /// Return true if the value of any element in Mask is the zero sentinel value. | |||
| 5598 | static bool isAnyZero(ArrayRef<int> Mask) { | |||
| 5599 | return llvm::any_of(Mask, [](int M) { return M == SM_SentinelZero; }); | |||
| 5600 | } | |||
| 5601 | ||||
| 5602 | /// Return true if the value of any element in Mask is the zero or undef | |||
| 5603 | /// sentinel values. | |||
| 5604 | static bool isAnyZeroOrUndef(ArrayRef<int> Mask) { | |||
| 5605 | return llvm::any_of(Mask, [](int M) { | |||
| 5606 | return M == SM_SentinelZero || M == SM_SentinelUndef; | |||
| 5607 | }); | |||
| 5608 | } | |||
| 5609 | ||||
| 5610 | /// Return true if Val is undef or if its value falls within the | |||
| 5611 | /// specified range (L, H]. | |||
| 5612 | static bool isUndefOrInRange(int Val, int Low, int Hi) { | |||
| 5613 | return (Val == SM_SentinelUndef) || isInRange(Val, Low, Hi); | |||
| 5614 | } | |||
| 5615 | ||||
| 5616 | /// Return true if every element in Mask is undef or if its value | |||
| 5617 | /// falls within the specified range (L, H]. | |||
| 5618 | static bool isUndefOrInRange(ArrayRef<int> Mask, int Low, int Hi) { | |||
| 5619 | return llvm::all_of( | |||
| 5620 | Mask, [Low, Hi](int M) { return isUndefOrInRange(M, Low, Hi); }); | |||
| 5621 | } | |||
| 5622 | ||||
| 5623 | /// Return true if Val is undef, zero or if its value falls within the | |||
| 5624 | /// specified range (L, H]. | |||
| 5625 | static bool isUndefOrZeroOrInRange(int Val, int Low, int Hi) { | |||
| 5626 | return isUndefOrZero(Val) || isInRange(Val, Low, Hi); | |||
| 5627 | } | |||
| 5628 | ||||
| 5629 | /// Return true if every element in Mask is undef, zero or if its value | |||
| 5630 | /// falls within the specified range (L, H]. | |||
| 5631 | static bool isUndefOrZeroOrInRange(ArrayRef<int> Mask, int Low, int Hi) { | |||
| 5632 | return llvm::all_of( | |||
| 5633 | Mask, [Low, Hi](int M) { return isUndefOrZeroOrInRange(M, Low, Hi); }); | |||
| 5634 | } | |||
| 5635 | ||||
| 5636 | /// Return true if every element in Mask, beginning | |||
| 5637 | /// from position Pos and ending in Pos + Size, falls within the specified | |||
| 5638 | /// sequence (Low, Low + Step, ..., Low + (Size - 1) * Step) or is undef. | |||
| 5639 | static bool isSequentialOrUndefInRange(ArrayRef<int> Mask, unsigned Pos, | |||
| 5640 | unsigned Size, int Low, int Step = 1) { | |||
| 5641 | for (unsigned i = Pos, e = Pos + Size; i != e; ++i, Low += Step) | |||
| 5642 | if (!isUndefOrEqual(Mask[i], Low)) | |||
| 5643 | return false; | |||
| 5644 | return true; | |||
| 5645 | } | |||
| 5646 | ||||
| 5647 | /// Return true if every element in Mask, beginning | |||
| 5648 | /// from position Pos and ending in Pos+Size, falls within the specified | |||
| 5649 | /// sequential range (Low, Low+Size], or is undef or is zero. | |||
| 5650 | static bool isSequentialOrUndefOrZeroInRange(ArrayRef<int> Mask, unsigned Pos, | |||
| 5651 | unsigned Size, int Low, | |||
| 5652 | int Step = 1) { | |||
| 5653 | for (unsigned i = Pos, e = Pos + Size; i != e; ++i, Low += Step) | |||
| 5654 | if (!isUndefOrZero(Mask[i]) && Mask[i] != Low) | |||
| 5655 | return false; | |||
| 5656 | return true; | |||
| 5657 | } | |||
| 5658 | ||||
| 5659 | /// Return true if every element in Mask, beginning | |||
| 5660 | /// from position Pos and ending in Pos+Size is undef or is zero. | |||
| 5661 | static bool isUndefOrZeroInRange(ArrayRef<int> Mask, unsigned Pos, | |||
| 5662 | unsigned Size) { | |||
| 5663 | return llvm::all_of(Mask.slice(Pos, Size), | |||
| 5664 | [](int M) { return isUndefOrZero(M); }); | |||
| 5665 | } | |||
| 5666 | ||||
| 5667 | /// Helper function to test whether a shuffle mask could be | |||
| 5668 | /// simplified by widening the elements being shuffled. | |||
| 5669 | /// | |||
| 5670 | /// Appends the mask for wider elements in WidenedMask if valid. Otherwise | |||
| 5671 | /// leaves it in an unspecified state. | |||
| 5672 | /// | |||
| 5673 | /// NOTE: This must handle normal vector shuffle masks and *target* vector | |||
| 5674 | /// shuffle masks. The latter have the special property of a '-2' representing | |||
| 5675 | /// a zero-ed lane of a vector. | |||
| 5676 | static bool canWidenShuffleElements(ArrayRef<int> Mask, | |||
| 5677 | SmallVectorImpl<int> &WidenedMask) { | |||
| 5678 | WidenedMask.assign(Mask.size() / 2, 0); | |||
| 5679 | for (int i = 0, Size = Mask.size(); i < Size; i += 2) { | |||
| 5680 | int M0 = Mask[i]; | |||
| 5681 | int M1 = Mask[i + 1]; | |||
| 5682 | ||||
| 5683 | // If both elements are undef, its trivial. | |||
| 5684 | if (M0 == SM_SentinelUndef && M1 == SM_SentinelUndef) { | |||
| 5685 | WidenedMask[i / 2] = SM_SentinelUndef; | |||
| 5686 | continue; | |||
| 5687 | } | |||
| 5688 | ||||
| 5689 | // Check for an undef mask and a mask value properly aligned to fit with | |||
| 5690 | // a pair of values. If we find such a case, use the non-undef mask's value. | |||
| 5691 | if (M0 == SM_SentinelUndef && M1 >= 0 && (M1 % 2) == 1) { | |||
| 5692 | WidenedMask[i / 2] = M1 / 2; | |||
| 5693 | continue; | |||
| 5694 | } | |||
| 5695 | if (M1 == SM_SentinelUndef && M0 >= 0 && (M0 % 2) == 0) { | |||
| 5696 | WidenedMask[i / 2] = M0 / 2; | |||
| 5697 | continue; | |||
| 5698 | } | |||
| 5699 | ||||
| 5700 | // When zeroing, we need to spread the zeroing across both lanes to widen. | |||
| 5701 | if (M0 == SM_SentinelZero || M1 == SM_SentinelZero) { | |||
| 5702 | if ((M0 == SM_SentinelZero || M0 == SM_SentinelUndef) && | |||
| 5703 | (M1 == SM_SentinelZero || M1 == SM_SentinelUndef)) { | |||
| 5704 | WidenedMask[i / 2] = SM_SentinelZero; | |||
| 5705 | continue; | |||
| 5706 | } | |||
| 5707 | return false; | |||
| 5708 | } | |||
| 5709 | ||||
| 5710 | // Finally check if the two mask values are adjacent and aligned with | |||
| 5711 | // a pair. | |||
| 5712 | if (M0 != SM_SentinelUndef && (M0 % 2) == 0 && (M0 + 1) == M1) { | |||
| 5713 | WidenedMask[i / 2] = M0 / 2; | |||
| 5714 | continue; | |||
| 5715 | } | |||
| 5716 | ||||
| 5717 | // Otherwise we can't safely widen the elements used in this shuffle. | |||
| 5718 | return false; | |||
| 5719 | } | |||
| 5720 | assert(WidenedMask.size() == Mask.size() / 2 &&((void)0) | |||
| 5721 | "Incorrect size of mask after widening the elements!")((void)0); | |||
| 5722 | ||||
| 5723 | return true; | |||
| 5724 | } | |||
| 5725 | ||||
| 5726 | static bool canWidenShuffleElements(ArrayRef<int> Mask, | |||
| 5727 | const APInt &Zeroable, | |||
| 5728 | bool V2IsZero, | |||
| 5729 | SmallVectorImpl<int> &WidenedMask) { | |||
| 5730 | // Create an alternative mask with info about zeroable elements. | |||
| 5731 | // Here we do not set undef elements as zeroable. | |||
| 5732 | SmallVector<int, 64> ZeroableMask(Mask.begin(), Mask.end()); | |||
| 5733 | if (V2IsZero) { | |||
| 5734 | assert(!Zeroable.isNullValue() && "V2's non-undef elements are used?!")((void)0); | |||
| 5735 | for (int i = 0, Size = Mask.size(); i != Size; ++i) | |||
| 5736 | if (Mask[i] != SM_SentinelUndef && Zeroable[i]) | |||
| 5737 | ZeroableMask[i] = SM_SentinelZero; | |||
| 5738 | } | |||
| 5739 | return canWidenShuffleElements(ZeroableMask, WidenedMask); | |||
| 5740 | } | |||
| 5741 | ||||
| 5742 | static bool canWidenShuffleElements(ArrayRef<int> Mask) { | |||
| 5743 | SmallVector<int, 32> WidenedMask; | |||
| 5744 | return canWidenShuffleElements(Mask, WidenedMask); | |||
| 5745 | } | |||
| 5746 | ||||
| 5747 | // Attempt to narrow/widen shuffle mask until it matches the target number of | |||
| 5748 | // elements. | |||
| 5749 | static bool scaleShuffleElements(ArrayRef<int> Mask, unsigned NumDstElts, | |||
| 5750 | SmallVectorImpl<int> &ScaledMask) { | |||
| 5751 | unsigned NumSrcElts = Mask.size(); | |||
| 5752 | assert(((NumSrcElts % NumDstElts) == 0 || (NumDstElts % NumSrcElts) == 0) &&((void)0) | |||
| 5753 | "Illegal shuffle scale factor")((void)0); | |||
| 5754 | ||||
| 5755 | // Narrowing is guaranteed to work. | |||
| 5756 | if (NumDstElts >= NumSrcElts) { | |||
| 5757 | int Scale = NumDstElts / NumSrcElts; | |||
| 5758 | llvm::narrowShuffleMaskElts(Scale, Mask, ScaledMask); | |||
| 5759 | return true; | |||
| 5760 | } | |||
| 5761 | ||||
| 5762 | // We have to repeat the widening until we reach the target size, but we can | |||
| 5763 | // split out the first widening as it sets up ScaledMask for us. | |||
| 5764 | if (canWidenShuffleElements(Mask, ScaledMask)) { | |||
| 5765 | while (ScaledMask.size() > NumDstElts) { | |||
| 5766 | SmallVector<int, 16> WidenedMask; | |||
| 5767 | if (!canWidenShuffleElements(ScaledMask, WidenedMask)) | |||
| 5768 | return false; | |||
| 5769 | ScaledMask = std::move(WidenedMask); | |||
| 5770 | } | |||
| 5771 | return true; | |||
| 5772 | } | |||
| 5773 | ||||
| 5774 | return false; | |||
| 5775 | } | |||
| 5776 | ||||
| 5777 | /// Returns true if Elt is a constant zero or a floating point constant +0.0. | |||
| 5778 | bool X86::isZeroNode(SDValue Elt) { | |||
| 5779 | return isNullConstant(Elt) || isNullFPConstant(Elt); | |||
| 5780 | } | |||
| 5781 | ||||
| 5782 | // Build a vector of constants. | |||
| 5783 | // Use an UNDEF node if MaskElt == -1. | |||
| 5784 | // Split 64-bit constants in the 32-bit mode. | |||
| 5785 | static SDValue getConstVector(ArrayRef<int> Values, MVT VT, SelectionDAG &DAG, | |||
| 5786 | const SDLoc &dl, bool IsMask = false) { | |||
| 5787 | ||||
| 5788 | SmallVector<SDValue, 32> Ops; | |||
| 5789 | bool Split = false; | |||
| 5790 | ||||
| 5791 | MVT ConstVecVT = VT; | |||
| 5792 | unsigned NumElts = VT.getVectorNumElements(); | |||
| 5793 | bool In64BitMode = DAG.getTargetLoweringInfo().isTypeLegal(MVT::i64); | |||
| 5794 | if (!In64BitMode && VT.getVectorElementType() == MVT::i64) { | |||
| 5795 | ConstVecVT = MVT::getVectorVT(MVT::i32, NumElts * 2); | |||
| 5796 | Split = true; | |||
| 5797 | } | |||
| 5798 | ||||
| 5799 | MVT EltVT = ConstVecVT.getVectorElementType(); | |||
| 5800 | for (unsigned i = 0; i < NumElts; ++i) { | |||
| 5801 | bool IsUndef = Values[i] < 0 && IsMask; | |||
| 5802 | SDValue OpNode = IsUndef ? DAG.getUNDEF(EltVT) : | |||
| 5803 | DAG.getConstant(Values[i], dl, EltVT); | |||
| 5804 | Ops.push_back(OpNode); | |||
| 5805 | if (Split) | |||
| 5806 | Ops.push_back(IsUndef ? DAG.getUNDEF(EltVT) : | |||
| 5807 | DAG.getConstant(0, dl, EltVT)); | |||
| 5808 | } | |||
| 5809 | SDValue ConstsNode = DAG.getBuildVector(ConstVecVT, dl, Ops); | |||
| 5810 | if (Split) | |||
| 5811 | ConstsNode = DAG.getBitcast(VT, ConstsNode); | |||
| 5812 | return ConstsNode; | |||
| 5813 | } | |||
| 5814 | ||||
| 5815 | static SDValue getConstVector(ArrayRef<APInt> Bits, APInt &Undefs, | |||
| 5816 | MVT VT, SelectionDAG &DAG, const SDLoc &dl) { | |||
| 5817 | assert(Bits.size() == Undefs.getBitWidth() &&((void)0) | |||
| 5818 | "Unequal constant and undef arrays")((void)0); | |||
| 5819 | SmallVector<SDValue, 32> Ops; | |||
| 5820 | bool Split = false; | |||
| 5821 | ||||
| 5822 | MVT ConstVecVT = VT; | |||
| 5823 | unsigned NumElts = VT.getVectorNumElements(); | |||
| 5824 | bool In64BitMode = DAG.getTargetLoweringInfo().isTypeLegal(MVT::i64); | |||
| 5825 | if (!In64BitMode && VT.getVectorElementType() == MVT::i64) { | |||
| 5826 | ConstVecVT = MVT::getVectorVT(MVT::i32, NumElts * 2); | |||
| 5827 | Split = true; | |||
| 5828 | } | |||
| 5829 | ||||
| 5830 | MVT EltVT = ConstVecVT.getVectorElementType(); | |||
| 5831 | for (unsigned i = 0, e = Bits.size(); i != e; ++i) { | |||
| 5832 | if (Undefs[i]) { | |||
| 5833 | Ops.append(Split ? 2 : 1, DAG.getUNDEF(EltVT)); | |||
| 5834 | continue; | |||
| 5835 | } | |||
| 5836 | const APInt &V = Bits[i]; | |||
| 5837 | assert(V.getBitWidth() == VT.getScalarSizeInBits() && "Unexpected sizes")((void)0); | |||
| 5838 | if (Split) { | |||
| 5839 | Ops.push_back(DAG.getConstant(V.trunc(32), dl, EltVT)); | |||
| 5840 | Ops.push_back(DAG.getConstant(V.lshr(32).trunc(32), dl, EltVT)); | |||
| 5841 | } else if (EltVT == MVT::f32) { | |||
| 5842 | APFloat FV(APFloat::IEEEsingle(), V); | |||
| 5843 | Ops.push_back(DAG.getConstantFP(FV, dl, EltVT)); | |||
| 5844 | } else if (EltVT == MVT::f64) { | |||
| 5845 | APFloat FV(APFloat::IEEEdouble(), V); | |||
| 5846 | Ops.push_back(DAG.getConstantFP(FV, dl, EltVT)); | |||
| 5847 | } else { | |||
| 5848 | Ops.push_back(DAG.getConstant(V, dl, EltVT)); | |||
| 5849 | } | |||
| 5850 | } | |||
| 5851 | ||||
| 5852 | SDValue ConstsNode = DAG.getBuildVector(ConstVecVT, dl, Ops); | |||
| 5853 | return DAG.getBitcast(VT, ConstsNode); | |||
| 5854 | } | |||
| 5855 | ||||
| 5856 | /// Returns a vector of specified type with all zero elements. | |||
| 5857 | static SDValue getZeroVector(MVT VT, const X86Subtarget &Subtarget, | |||
| 5858 | SelectionDAG &DAG, const SDLoc &dl) { | |||
| 5859 | assert((VT.is128BitVector() || VT.is256BitVector() || VT.is512BitVector() ||((void)0) | |||
| 5860 | VT.getVectorElementType() == MVT::i1) &&((void)0) | |||
| 5861 | "Unexpected vector type")((void)0); | |||
| 5862 | ||||
| 5863 | // Try to build SSE/AVX zero vectors as <N x i32> bitcasted to their dest | |||
| 5864 | // type. This ensures they get CSE'd. But if the integer type is not | |||
| 5865 | // available, use a floating-point +0.0 instead. | |||
| 5866 | SDValue Vec; | |||
| 5867 | if (!Subtarget.hasSSE2() && VT.is128BitVector()) { | |||
| 5868 | Vec = DAG.getConstantFP(+0.0, dl, MVT::v4f32); | |||
| 5869 | } else if (VT.isFloatingPoint()) { | |||
| 5870 | Vec = DAG.getConstantFP(+0.0, dl, VT); | |||
| 5871 | } else if (VT.getVectorElementType() == MVT::i1) { | |||
| 5872 | assert((Subtarget.hasBWI() || VT.getVectorNumElements() <= 16) &&((void)0) | |||
| 5873 | "Unexpected vector type")((void)0); | |||
| 5874 | Vec = DAG.getConstant(0, dl, VT); | |||
| 5875 | } else { | |||
| 5876 | unsigned Num32BitElts = VT.getSizeInBits() / 32; | |||
| 5877 | Vec = DAG.getConstant(0, dl, MVT::getVectorVT(MVT::i32, Num32BitElts)); | |||
| 5878 | } | |||
| 5879 | return DAG.getBitcast(VT, Vec); | |||
| 5880 | } | |||
| 5881 | ||||
| 5882 | static SDValue extractSubVector(SDValue Vec, unsigned IdxVal, SelectionDAG &DAG, | |||
| 5883 | const SDLoc &dl, unsigned vectorWidth) { | |||
| 5884 | EVT VT = Vec.getValueType(); | |||
| 5885 | EVT ElVT = VT.getVectorElementType(); | |||
| 5886 | unsigned Factor = VT.getSizeInBits() / vectorWidth; | |||
| 5887 | EVT ResultVT = EVT::getVectorVT(*DAG.getContext(), ElVT, | |||
| 5888 | VT.getVectorNumElements() / Factor); | |||
| 5889 | ||||
| 5890 | // Extract the relevant vectorWidth bits. Generate an EXTRACT_SUBVECTOR | |||
| 5891 | unsigned ElemsPerChunk = vectorWidth / ElVT.getSizeInBits(); | |||
| 5892 | assert(isPowerOf2_32(ElemsPerChunk) && "Elements per chunk not power of 2")((void)0); | |||
| 5893 | ||||
| 5894 | // This is the index of the first element of the vectorWidth-bit chunk | |||
| 5895 | // we want. Since ElemsPerChunk is a power of 2 just need to clear bits. | |||
| 5896 | IdxVal &= ~(ElemsPerChunk - 1); | |||
| 5897 | ||||
| 5898 | // If the input is a buildvector just emit a smaller one. | |||
| 5899 | if (Vec.getOpcode() == ISD::BUILD_VECTOR) | |||
| 5900 | return DAG.getBuildVector(ResultVT, dl, | |||
| 5901 | Vec->ops().slice(IdxVal, ElemsPerChunk)); | |||
| 5902 | ||||
| 5903 | SDValue VecIdx = DAG.getIntPtrConstant(IdxVal, dl); | |||
| 5904 | return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, ResultVT, Vec, VecIdx); | |||
| 5905 | } | |||
| 5906 | ||||
| 5907 | /// Generate a DAG to grab 128-bits from a vector > 128 bits. This | |||
| 5908 | /// sets things up to match to an AVX VEXTRACTF128 / VEXTRACTI128 | |||
| 5909 | /// or AVX-512 VEXTRACTF32x4 / VEXTRACTI32x4 | |||
| 5910 | /// instructions or a simple subregister reference. Idx is an index in the | |||
| 5911 | /// 128 bits we want. It need not be aligned to a 128-bit boundary. That makes | |||
| 5912 | /// lowering EXTRACT_VECTOR_ELT operations easier. | |||
| 5913 | static SDValue extract128BitVector(SDValue Vec, unsigned IdxVal, | |||
| 5914 | SelectionDAG &DAG, const SDLoc &dl) { | |||
| 5915 | assert((Vec.getValueType().is256BitVector() ||((void)0) | |||
| 5916 | Vec.getValueType().is512BitVector()) && "Unexpected vector size!")((void)0); | |||
| 5917 | return extractSubVector(Vec, IdxVal, DAG, dl, 128); | |||
| 5918 | } | |||
| 5919 | ||||
| 5920 | /// Generate a DAG to grab 256-bits from a 512-bit vector. | |||
| 5921 | static SDValue extract256BitVector(SDValue Vec, unsigned IdxVal, | |||
| 5922 | SelectionDAG &DAG, const SDLoc &dl) { | |||
| 5923 | assert(Vec.getValueType().is512BitVector() && "Unexpected vector size!")((void)0); | |||
| 5924 | return extractSubVector(Vec, IdxVal, DAG, dl, 256); | |||
| 5925 | } | |||
| 5926 | ||||
| 5927 | static SDValue insertSubVector(SDValue Result, SDValue Vec, unsigned IdxVal, | |||
| 5928 | SelectionDAG &DAG, const SDLoc &dl, | |||
| 5929 | unsigned vectorWidth) { | |||
| 5930 | assert((vectorWidth == 128 || vectorWidth == 256) &&((void)0) | |||
| 5931 | "Unsupported vector width")((void)0); | |||
| 5932 | // Inserting UNDEF is Result | |||
| 5933 | if (Vec.isUndef()) | |||
| 5934 | return Result; | |||
| 5935 | EVT VT = Vec.getValueType(); | |||
| 5936 | EVT ElVT = VT.getVectorElementType(); | |||
| 5937 | EVT ResultVT = Result.getValueType(); | |||
| 5938 | ||||
| 5939 | // Insert the relevant vectorWidth bits. | |||
| 5940 | unsigned ElemsPerChunk = vectorWidth/ElVT.getSizeInBits(); | |||
| 5941 | assert(isPowerOf2_32(ElemsPerChunk) && "Elements per chunk not power of 2")((void)0); | |||
| 5942 | ||||
| 5943 | // This is the index of the first element of the vectorWidth-bit chunk | |||
| 5944 | // we want. Since ElemsPerChunk is a power of 2 just need to clear bits. | |||
| 5945 | IdxVal &= ~(ElemsPerChunk - 1); | |||
| 5946 | ||||
| 5947 | SDValue VecIdx = DAG.getIntPtrConstant(IdxVal, dl); | |||
| 5948 | return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResultVT, Result, Vec, VecIdx); | |||
| 5949 | } | |||
| 5950 | ||||
| 5951 | /// Generate a DAG to put 128-bits into a vector > 128 bits. This | |||
| 5952 | /// sets things up to match to an AVX VINSERTF128/VINSERTI128 or | |||
| 5953 | /// AVX-512 VINSERTF32x4/VINSERTI32x4 instructions or a | |||
| 5954 | /// simple superregister reference. Idx is an index in the 128 bits | |||
| 5955 | /// we want. It need not be aligned to a 128-bit boundary. That makes | |||
| 5956 | /// lowering INSERT_VECTOR_ELT operations easier. | |||
| 5957 | static SDValue insert128BitVector(SDValue Result, SDValue Vec, unsigned IdxVal, | |||
| 5958 | SelectionDAG &DAG, const SDLoc &dl) { | |||
| 5959 | assert(Vec.getValueType().is128BitVector() && "Unexpected vector size!")((void)0); | |||
| 5960 | return insertSubVector(Result, Vec, IdxVal, DAG, dl, 128); | |||
| 5961 | } | |||
| 5962 | ||||
| 5963 | /// Widen a vector to a larger size with the same scalar type, with the new | |||
| 5964 | /// elements either zero or undef. | |||
| 5965 | static SDValue widenSubVector(MVT VT, SDValue Vec, bool ZeroNewElements, | |||
| 5966 | const X86Subtarget &Subtarget, SelectionDAG &DAG, | |||
| 5967 | const SDLoc &dl) { | |||
| 5968 | assert(Vec.getValueSizeInBits().getFixedSize() < VT.getFixedSizeInBits() &&((void)0) | |||
| 5969 | Vec.getValueType().getScalarType() == VT.getScalarType() &&((void)0) | |||
| 5970 | "Unsupported vector widening type")((void)0); | |||
| 5971 | SDValue Res = ZeroNewElements ? getZeroVector(VT, Subtarget, DAG, dl) | |||
| 5972 | : DAG.getUNDEF(VT); | |||
| 5973 | return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, VT, Res, Vec, | |||
| 5974 | DAG.getIntPtrConstant(0, dl)); | |||
| 5975 | } | |||
| 5976 | ||||
| 5977 | /// Widen a vector to a larger size with the same scalar type, with the new | |||
| 5978 | /// elements either zero or undef. | |||
| 5979 | static SDValue widenSubVector(SDValue Vec, bool ZeroNewElements, | |||
| 5980 | const X86Subtarget &Subtarget, SelectionDAG &DAG, | |||
| 5981 | const SDLoc &dl, unsigned WideSizeInBits) { | |||
| 5982 | assert(Vec.getValueSizeInBits() < WideSizeInBits &&((void)0) | |||
| 5983 | (WideSizeInBits % Vec.getScalarValueSizeInBits()) == 0 &&((void)0) | |||
| 5984 | "Unsupported vector widening type")((void)0); | |||
| 5985 | unsigned WideNumElts = WideSizeInBits / Vec.getScalarValueSizeInBits(); | |||
| 5986 | MVT SVT = Vec.getSimpleValueType().getScalarType(); | |||
| 5987 | MVT VT = MVT::getVectorVT(SVT, WideNumElts); | |||
| 5988 | return widenSubVector(VT, Vec, ZeroNewElements, Subtarget, DAG, dl); | |||
| 5989 | } | |||
| 5990 | ||||
| 5991 | // Helper function to collect subvector ops that are concatenated together, | |||
| 5992 | // either by ISD::CONCAT_VECTORS or a ISD::INSERT_SUBVECTOR series. | |||
| 5993 | // The subvectors in Ops are guaranteed to be the same type. | |||
| 5994 | static bool collectConcatOps(SDNode *N, SmallVectorImpl<SDValue> &Ops) { | |||
| 5995 | assert(Ops.empty() && "Expected an empty ops vector")((void)0); | |||
| 5996 | ||||
| 5997 | if (N->getOpcode() == ISD::CONCAT_VECTORS) { | |||
| 5998 | Ops.append(N->op_begin(), N->op_end()); | |||
| 5999 | return true; | |||
| 6000 | } | |||
| 6001 | ||||
| 6002 | if (N->getOpcode() == ISD::INSERT_SUBVECTOR) { | |||
| 6003 | SDValue Src = N->getOperand(0); | |||
| 6004 | SDValue Sub = N->getOperand(1); | |||
| 6005 | const APInt &Idx = N->getConstantOperandAPInt(2); | |||
| 6006 | EVT VT = Src.getValueType(); | |||
| 6007 | EVT SubVT = Sub.getValueType(); | |||
| 6008 | ||||
| 6009 | // TODO - Handle more general insert_subvector chains. | |||
| 6010 | if (VT.getSizeInBits() == (SubVT.getSizeInBits() * 2) && | |||
| 6011 | Idx == (VT.getVectorNumElements() / 2)) { | |||
| 6012 | // insert_subvector(insert_subvector(undef, x, lo), y, hi) | |||
| 6013 | if (Src.getOpcode() == ISD::INSERT_SUBVECTOR && | |||
| 6014 | Src.getOperand(1).getValueType() == SubVT && | |||
| 6015 | isNullConstant(Src.getOperand(2))) { | |||
| 6016 | Ops.push_back(Src.getOperand(1)); | |||
| 6017 | Ops.push_back(Sub); | |||
| 6018 | return true; | |||
| 6019 | } | |||
| 6020 | // insert_subvector(x, extract_subvector(x, lo), hi) | |||
| 6021 | if (Sub.getOpcode() == ISD::EXTRACT_SUBVECTOR && | |||
| 6022 | Sub.getOperand(0) == Src && isNullConstant(Sub.getOperand(1))) { | |||
| 6023 | Ops.append(2, Sub); | |||
| 6024 | return true; | |||
| 6025 | } | |||
| 6026 | } | |||
| 6027 | } | |||
| 6028 | ||||
| 6029 | return false; | |||
| 6030 | } | |||
| 6031 | ||||
| 6032 | static std::pair<SDValue, SDValue> splitVector(SDValue Op, SelectionDAG &DAG, | |||
| 6033 | const SDLoc &dl) { | |||
| 6034 | EVT VT = Op.getValueType(); | |||
| 6035 | unsigned NumElems = VT.getVectorNumElements(); | |||
| 6036 | unsigned SizeInBits = VT.getSizeInBits(); | |||
| 6037 | assert((NumElems % 2) == 0 && (SizeInBits % 2) == 0 &&((void)0) | |||
| 6038 | "Can't split odd sized vector")((void)0); | |||
| 6039 | ||||
| 6040 | SDValue Lo = extractSubVector(Op, 0, DAG, dl, SizeInBits / 2); | |||
| 6041 | SDValue Hi = extractSubVector(Op, NumElems / 2, DAG, dl, SizeInBits / 2); | |||
| 6042 | return std::make_pair(Lo, Hi); | |||
| 6043 | } | |||
| 6044 | ||||
| 6045 | // Split an unary integer op into 2 half sized ops. | |||
| 6046 | static SDValue splitVectorIntUnary(SDValue Op, SelectionDAG &DAG) { | |||
| 6047 | EVT VT = Op.getValueType(); | |||
| 6048 | ||||
| 6049 | // Make sure we only try to split 256/512-bit types to avoid creating | |||
| 6050 | // narrow vectors. | |||
| 6051 | assert((Op.getOperand(0).getValueType().is256BitVector() ||((void)0) | |||
| 6052 | Op.getOperand(0).getValueType().is512BitVector()) &&((void)0) | |||
| 6053 | (VT.is256BitVector() || VT.is512BitVector()) && "Unsupported VT!")((void)0); | |||
| 6054 | assert(Op.getOperand(0).getValueType().getVectorNumElements() ==((void)0) | |||
| 6055 | VT.getVectorNumElements() &&((void)0) | |||
| 6056 | "Unexpected VTs!")((void)0); | |||
| 6057 | ||||
| 6058 | SDLoc dl(Op); | |||
| 6059 | ||||
| 6060 | // Extract the Lo/Hi vectors | |||
| 6061 | SDValue Lo, Hi; | |||
| 6062 | std::tie(Lo, Hi) = splitVector(Op.getOperand(0), DAG, dl); | |||
| 6063 | ||||
| 6064 | EVT LoVT, HiVT; | |||
| 6065 | std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(VT); | |||
| 6066 | return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, | |||
| 6067 | DAG.getNode(Op.getOpcode(), dl, LoVT, Lo), | |||
| 6068 | DAG.getNode(Op.getOpcode(), dl, HiVT, Hi)); | |||
| 6069 | } | |||
| 6070 | ||||
| 6071 | /// Break a binary integer operation into 2 half sized ops and then | |||
| 6072 | /// concatenate the result back. | |||
| 6073 | static SDValue splitVectorIntBinary(SDValue Op, SelectionDAG &DAG) { | |||
| 6074 | EVT VT = Op.getValueType(); | |||
| 6075 | ||||
| 6076 | // Sanity check that all the types match. | |||
| 6077 | assert(Op.getOperand(0).getValueType() == VT &&((void)0) | |||
| 6078 | Op.getOperand(1).getValueType() == VT && "Unexpected VTs!")((void)0); | |||
| 6079 | assert((VT.is256BitVector() || VT.is512BitVector()) && "Unsupported VT!")((void)0); | |||
| 6080 | ||||
| 6081 | SDLoc dl(Op); | |||
| 6082 | ||||
| 6083 | // Extract the LHS Lo/Hi vectors | |||
| 6084 | SDValue LHS1, LHS2; | |||
| 6085 | std::tie(LHS1, LHS2) = splitVector(Op.getOperand(0), DAG, dl); | |||
| 6086 | ||||
| 6087 | // Extract the RHS Lo/Hi vectors | |||
| 6088 | SDValue RHS1, RHS2; | |||
| 6089 | std::tie(RHS1, RHS2) = splitVector(Op.getOperand(1), DAG, dl); | |||
| 6090 | ||||
| 6091 | EVT LoVT, HiVT; | |||
| 6092 | std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(VT); | |||
| 6093 | return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, | |||
| 6094 | DAG.getNode(Op.getOpcode(), dl, LoVT, LHS1, RHS1), | |||
| 6095 | DAG.getNode(Op.getOpcode(), dl, HiVT, LHS2, RHS2)); | |||
| 6096 | } | |||
| 6097 | ||||
| 6098 | // Helper for splitting operands of an operation to legal target size and | |||
| 6099 | // apply a function on each part. | |||
| 6100 | // Useful for operations that are available on SSE2 in 128-bit, on AVX2 in | |||
| 6101 | // 256-bit and on AVX512BW in 512-bit. The argument VT is the type used for | |||
| 6102 | // deciding if/how to split Ops. Ops elements do *not* have to be of type VT. | |||
| 6103 | // The argument Builder is a function that will be applied on each split part: | |||
| 6104 | // SDValue Builder(SelectionDAG&G, SDLoc, ArrayRef<SDValue>) | |||
| 6105 | template <typename F> | |||
| 6106 | SDValue SplitOpsAndApply(SelectionDAG &DAG, const X86Subtarget &Subtarget, | |||
| 6107 | const SDLoc &DL, EVT VT, ArrayRef<SDValue> Ops, | |||
| 6108 | F Builder, bool CheckBWI = true) { | |||
| 6109 | assert(Subtarget.hasSSE2() && "Target assumed to support at least SSE2")((void)0); | |||
| 6110 | unsigned NumSubs = 1; | |||
| 6111 | if ((CheckBWI && Subtarget.useBWIRegs()) || | |||
| 6112 | (!CheckBWI && Subtarget.useAVX512Regs())) { | |||
| 6113 | if (VT.getSizeInBits() > 512) { | |||
| 6114 | NumSubs = VT.getSizeInBits() / 512; | |||
| 6115 | assert((VT.getSizeInBits() % 512) == 0 && "Illegal vector size")((void)0); | |||
| 6116 | } | |||
| 6117 | } else if (Subtarget.hasAVX2()) { | |||
| 6118 | if (VT.getSizeInBits() > 256) { | |||
| 6119 | NumSubs = VT.getSizeInBits() / 256; | |||
| 6120 | assert((VT.getSizeInBits() % 256) == 0 && "Illegal vector size")((void)0); | |||
| 6121 | } | |||
| 6122 | } else { | |||
| 6123 | if (VT.getSizeInBits() > 128) { | |||
| 6124 | NumSubs = VT.getSizeInBits() / 128; | |||
| 6125 | assert((VT.getSizeInBits() % 128) == 0 && "Illegal vector size")((void)0); | |||
| 6126 | } | |||
| 6127 | } | |||
| 6128 | ||||
| 6129 | if (NumSubs == 1) | |||
| 6130 | return Builder(DAG, DL, Ops); | |||
| 6131 | ||||
| 6132 | SmallVector<SDValue, 4> Subs; | |||
| 6133 | for (unsigned i = 0; i != NumSubs; ++i) { | |||
| 6134 | SmallVector<SDValue, 2> SubOps; | |||
| 6135 | for (SDValue Op : Ops) { | |||
| 6136 | EVT OpVT = Op.getValueType(); | |||
| 6137 | unsigned NumSubElts = OpVT.getVectorNumElements() / NumSubs; | |||
| 6138 | unsigned SizeSub = OpVT.getSizeInBits() / NumSubs; | |||
| 6139 | SubOps.push_back(extractSubVector(Op, i * NumSubElts, DAG, DL, SizeSub)); | |||
| 6140 | } | |||
| 6141 | Subs.push_back(Builder(DAG, DL, SubOps)); | |||
| 6142 | } | |||
| 6143 | return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Subs); | |||
| 6144 | } | |||
| 6145 | ||||
| 6146 | /// Insert i1-subvector to i1-vector. | |||
| 6147 | static SDValue insert1BitVector(SDValue Op, SelectionDAG &DAG, | |||
| 6148 | const X86Subtarget &Subtarget) { | |||
| 6149 | ||||
| 6150 | SDLoc dl(Op); | |||
| 6151 | SDValue Vec = Op.getOperand(0); | |||
| 6152 | SDValue SubVec = Op.getOperand(1); | |||
| 6153 | SDValue Idx = Op.getOperand(2); | |||
| 6154 | unsigned IdxVal = Op.getConstantOperandVal(2); | |||
| 6155 | ||||
| 6156 | // Inserting undef is a nop. We can just return the original vector. | |||
| 6157 | if (SubVec.isUndef()) | |||
| 6158 | return Vec; | |||
| 6159 | ||||
| 6160 | if (IdxVal == 0 && Vec.isUndef()) // the operation is legal | |||
| 6161 | return Op; | |||
| 6162 | ||||
| 6163 | MVT OpVT = Op.getSimpleValueType(); | |||
| 6164 | unsigned NumElems = OpVT.getVectorNumElements(); | |||
| 6165 | SDValue ZeroIdx = DAG.getIntPtrConstant(0, dl); | |||
| 6166 | ||||
| 6167 | // Extend to natively supported kshift. | |||
| 6168 | MVT WideOpVT = OpVT; | |||
| 6169 | if ((!Subtarget.hasDQI() && NumElems == 8) || NumElems < 8) | |||
| 6170 | WideOpVT = Subtarget.hasDQI() ? MVT::v8i1 : MVT::v16i1; | |||
| 6171 | ||||
| 6172 | // Inserting into the lsbs of a zero vector is legal. ISel will insert shifts | |||
| 6173 | // if necessary. | |||
| 6174 | if (IdxVal == 0 && ISD::isBuildVectorAllZeros(Vec.getNode())) { | |||
| 6175 | // May need to promote to a legal type. | |||
| 6176 | Op = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, WideOpVT, | |||
| 6177 | DAG.getConstant(0, dl, WideOpVT), | |||
| 6178 | SubVec, Idx); | |||
| 6179 | return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, OpVT, Op, ZeroIdx); | |||
| 6180 | } | |||
| 6181 | ||||
| 6182 | MVT SubVecVT = SubVec.getSimpleValueType(); | |||
| 6183 | unsigned SubVecNumElems = SubVecVT.getVectorNumElements(); | |||
| 6184 | assert(IdxVal + SubVecNumElems <= NumElems &&((void)0) | |||
| 6185 | IdxVal % SubVecVT.getSizeInBits() == 0 &&((void)0) | |||
| 6186 | "Unexpected index value in INSERT_SUBVECTOR")((void)0); | |||
| 6187 | ||||
| 6188 | SDValue Undef = DAG.getUNDEF(WideOpVT); | |||
| 6189 | ||||
| 6190 | if (IdxVal == 0) { | |||
| 6191 | // Zero lower bits of the Vec | |||
| 6192 | SDValue ShiftBits = DAG.getTargetConstant(SubVecNumElems, dl, MVT::i8); | |||
| 6193 | Vec = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, WideOpVT, Undef, Vec, | |||
| 6194 | ZeroIdx); | |||
| 6195 | Vec = DAG.getNode(X86ISD::KSHIFTR, dl, WideOpVT, Vec, ShiftBits); | |||
| 6196 | Vec = DAG.getNode(X86ISD::KSHIFTL, dl, WideOpVT, Vec, ShiftBits); | |||
| 6197 | // Merge them together, SubVec should be zero extended. | |||
| 6198 | SubVec = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, WideOpVT, | |||
| 6199 | DAG.getConstant(0, dl, WideOpVT), | |||
| 6200 | SubVec, ZeroIdx); | |||
| 6201 | Op = DAG.getNode(ISD::OR, dl, WideOpVT, Vec, SubVec); | |||
| 6202 | return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, OpVT, Op, ZeroIdx); | |||
| 6203 | } | |||
| 6204 | ||||
| 6205 | SubVec = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, WideOpVT, | |||
| 6206 | Undef, SubVec, ZeroIdx); | |||
| 6207 | ||||
| 6208 | if (Vec.isUndef()) { | |||
| 6209 | assert(IdxVal != 0 && "Unexpected index")((void)0); | |||
| 6210 | SubVec = DAG.getNode(X86ISD::KSHIFTL, dl, WideOpVT, SubVec, | |||
| 6211 | DAG.getTargetConstant(IdxVal, dl, MVT::i8)); | |||
| 6212 | return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, OpVT, SubVec, ZeroIdx); | |||
| 6213 | } | |||
| 6214 | ||||
| 6215 | if (ISD::isBuildVectorAllZeros(Vec.getNode())) { | |||
| 6216 | assert(IdxVal != 0 && "Unexpected index")((void)0); | |||
| 6217 | NumElems = WideOpVT.getVectorNumElements(); | |||
| 6218 | unsigned ShiftLeft = NumElems - SubVecNumElems; | |||
| 6219 | unsigned ShiftRight = NumElems - SubVecNumElems - IdxVal; | |||
| 6220 | SubVec = DAG.getNode(X86ISD::KSHIFTL, dl, WideOpVT, SubVec, | |||
| 6221 | DAG.getTargetConstant(ShiftLeft, dl, MVT::i8)); | |||
| 6222 | if (ShiftRight != 0) | |||
| 6223 | SubVec = DAG.getNode(X86ISD::KSHIFTR, dl, WideOpVT, SubVec, | |||
| 6224 | DAG.getTargetConstant(ShiftRight, dl, MVT::i8)); | |||
| 6225 | return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, OpVT, SubVec, ZeroIdx); | |||
| 6226 | } | |||
| 6227 | ||||
| 6228 | // Simple case when we put subvector in the upper part | |||
| 6229 | if (IdxVal + SubVecNumElems == NumElems) { | |||
| 6230 | SubVec = DAG.getNode(X86ISD::KSHIFTL, dl, WideOpVT, SubVec, | |||
| 6231 | DAG.getTargetConstant(IdxVal, dl, MVT::i8)); | |||
| 6232 | if (SubVecNumElems * 2 == NumElems) { | |||
| 6233 | // Special case, use legal zero extending insert_subvector. This allows | |||
| 6234 | // isel to optimize when bits are known zero. | |||
| 6235 | Vec = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVecVT, Vec, ZeroIdx); | |||
| 6236 | Vec = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, WideOpVT, | |||
| 6237 | DAG.getConstant(0, dl, WideOpVT), | |||
| 6238 | Vec, ZeroIdx); | |||
| 6239 | } else { | |||
| 6240 | // Otherwise use explicit shifts to zero the bits. | |||
| 6241 | Vec = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, WideOpVT, | |||
| 6242 | Undef, Vec, ZeroIdx); | |||
| 6243 | NumElems = WideOpVT.getVectorNumElements(); | |||
| 6244 | SDValue ShiftBits = DAG.getTargetConstant(NumElems - IdxVal, dl, MVT::i8); | |||
| 6245 | Vec = DAG.getNode(X86ISD::KSHIFTL, dl, WideOpVT, Vec, ShiftBits); | |||
| 6246 | Vec = DAG.getNode(X86ISD::KSHIFTR, dl, WideOpVT, Vec, ShiftBits); | |||
| 6247 | } | |||
| 6248 | Op = DAG.getNode(ISD::OR, dl, WideOpVT, Vec, SubVec); | |||
| 6249 | return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, OpVT, Op, ZeroIdx); | |||
| 6250 | } | |||
| 6251 | ||||
| 6252 | // Inserting into the middle is more complicated. | |||
| 6253 | ||||
| 6254 | NumElems = WideOpVT.getVectorNumElements(); | |||
| 6255 | ||||
| 6256 | // Widen the vector if needed. | |||
| 6257 | Vec = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, WideOpVT, Undef, Vec, ZeroIdx); | |||
| 6258 | ||||
| 6259 | unsigned ShiftLeft = NumElems - SubVecNumElems; | |||
| 6260 | unsigned ShiftRight = NumElems - SubVecNumElems - IdxVal; | |||
| 6261 | ||||
| 6262 | // Do an optimization for the the most frequently used types. | |||
| 6263 | if (WideOpVT != MVT::v64i1 || Subtarget.is64Bit()) { | |||
| 6264 | APInt Mask0 = APInt::getBitsSet(NumElems, IdxVal, IdxVal + SubVecNumElems); | |||
| 6265 | Mask0.flipAllBits(); | |||
| 6266 | SDValue CMask0 = DAG.getConstant(Mask0, dl, MVT::getIntegerVT(NumElems)); | |||
| 6267 | SDValue VMask0 = DAG.getNode(ISD::BITCAST, dl, WideOpVT, CMask0); | |||
| 6268 | Vec = DAG.getNode(ISD::AND, dl, WideOpVT, Vec, VMask0); | |||
| 6269 | SubVec = DAG.getNode(X86ISD::KSHIFTL, dl, WideOpVT, SubVec, | |||
| 6270 | DAG.getTargetConstant(ShiftLeft, dl, MVT::i8)); | |||
| 6271 | SubVec = DAG.getNode(X86ISD::KSHIFTR, dl, WideOpVT, SubVec, | |||
| 6272 | DAG.getTargetConstant(ShiftRight, dl, MVT::i8)); | |||
| 6273 | Op = DAG.getNode(ISD::OR, dl, WideOpVT, Vec, SubVec); | |||
| 6274 | ||||
| 6275 | // Reduce to original width if needed. | |||
| 6276 | return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, OpVT, Op, ZeroIdx); | |||
| 6277 | } | |||
| 6278 | ||||
| 6279 | // Clear the upper bits of the subvector and move it to its insert position. | |||
| 6280 | SubVec = DAG.getNode(X86ISD::KSHIFTL, dl, WideOpVT, SubVec, | |||
| 6281 | DAG.getTargetConstant(ShiftLeft, dl, MVT::i8)); | |||
| 6282 | SubVec = DAG.getNode(X86ISD::KSHIFTR, dl, WideOpVT, SubVec, | |||
| 6283 | DAG.getTargetConstant(ShiftRight, dl, MVT::i8)); | |||
| 6284 | ||||
| 6285 | // Isolate the bits below the insertion point. | |||
| 6286 | unsigned LowShift = NumElems - IdxVal; | |||
| 6287 | SDValue Low = DAG.getNode(X86ISD::KSHIFTL, dl, WideOpVT, Vec, | |||
| 6288 | DAG.getTargetConstant(LowShift, dl, MVT::i8)); | |||
| 6289 | Low = DAG.getNode(X86ISD::KSHIFTR, dl, WideOpVT, Low, | |||
| 6290 | DAG.getTargetConstant(LowShift, dl, MVT::i8)); | |||
| 6291 | ||||
| 6292 | // Isolate the bits after the last inserted bit. | |||
| 6293 | unsigned HighShift = IdxVal + SubVecNumElems; | |||
| 6294 | SDValue High = DAG.getNode(X86ISD::KSHIFTR, dl, WideOpVT, Vec, | |||
| 6295 | DAG.getTargetConstant(HighShift, dl, MVT::i8)); | |||
| 6296 | High = DAG.getNode(X86ISD::KSHIFTL, dl, WideOpVT, High, | |||
| 6297 | DAG.getTargetConstant(HighShift, dl, MVT::i8)); | |||
| 6298 | ||||
| 6299 | // Now OR all 3 pieces together. | |||
| 6300 | Vec = DAG.getNode(ISD::OR, dl, WideOpVT, Low, High); | |||
| 6301 | SubVec = DAG.getNode(ISD::OR, dl, WideOpVT, SubVec, Vec); | |||
| 6302 | ||||
| 6303 | // Reduce to original width if needed. | |||
| 6304 | return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, OpVT, SubVec, ZeroIdx); | |||
| 6305 | } | |||
| 6306 | ||||
| 6307 | static SDValue concatSubVectors(SDValue V1, SDValue V2, SelectionDAG &DAG, | |||
| 6308 | const SDLoc &dl) { | |||
| 6309 | assert(V1.getValueType() == V2.getValueType() && "subvector type mismatch")((void)0); | |||
| 6310 | EVT SubVT = V1.getValueType(); | |||
| 6311 | EVT SubSVT = SubVT.getScalarType(); | |||
| 6312 | unsigned SubNumElts = SubVT.getVectorNumElements(); | |||
| 6313 | unsigned SubVectorWidth = SubVT.getSizeInBits(); | |||
| 6314 | EVT VT = EVT::getVectorVT(*DAG.getContext(), SubSVT, 2 * SubNumElts); | |||
| 6315 | SDValue V = insertSubVector(DAG.getUNDEF(VT), V1, 0, DAG, dl, SubVectorWidth); | |||
| 6316 | return insertSubVector(V, V2, SubNumElts, DAG, dl, SubVectorWidth); | |||
| 6317 | } | |||
| 6318 | ||||
| 6319 | /// Returns a vector of specified type with all bits set. | |||
| 6320 | /// Always build ones vectors as <4 x i32>, <8 x i32> or <16 x i32>. | |||
| 6321 | /// Then bitcast to their original type, ensuring they get CSE'd. | |||
| 6322 | static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, const SDLoc &dl) { | |||
| 6323 | assert((VT.is128BitVector() || VT.is256BitVector() || VT.is512BitVector()) &&((void)0) | |||
| 6324 | "Expected a 128/256/512-bit vector type")((void)0); | |||
| 6325 | ||||
| 6326 | APInt Ones = APInt::getAllOnesValue(32); | |||
| 6327 | unsigned NumElts = VT.getSizeInBits() / 32; | |||
| 6328 | SDValue Vec = DAG.getConstant(Ones, dl, MVT::getVectorVT(MVT::i32, NumElts)); | |||
| 6329 | return DAG.getBitcast(VT, Vec); | |||
| 6330 | } | |||
| 6331 | ||||
| 6332 | // Convert *_EXTEND_VECTOR_INREG to *_EXTEND opcode. | |||
| 6333 | static unsigned getOpcode_EXTEND(unsigned Opcode) { | |||
| 6334 | switch (Opcode) { | |||
| 6335 | case ISD::ANY_EXTEND: | |||
| 6336 | case ISD::ANY_EXTEND_VECTOR_INREG: | |||
| 6337 | return ISD::ANY_EXTEND; | |||
| 6338 | case ISD::ZERO_EXTEND: | |||
| 6339 | case ISD::ZERO_EXTEND_VECTOR_INREG: | |||
| 6340 | return ISD::ZERO_EXTEND; | |||
| 6341 | case ISD::SIGN_EXTEND: | |||
| 6342 | case ISD::SIGN_EXTEND_VECTOR_INREG: | |||
| 6343 | return ISD::SIGN_EXTEND; | |||
| 6344 | } | |||
| 6345 | llvm_unreachable("Unknown opcode")__builtin_unreachable(); | |||
| 6346 | } | |||
| 6347 | ||||
| 6348 | // Convert *_EXTEND to *_EXTEND_VECTOR_INREG opcode. | |||
| 6349 | static unsigned getOpcode_EXTEND_VECTOR_INREG(unsigned Opcode) { | |||
| 6350 | switch (Opcode) { | |||
| 6351 | case ISD::ANY_EXTEND: | |||
| 6352 | case ISD::ANY_EXTEND_VECTOR_INREG: | |||
| 6353 | return ISD::ANY_EXTEND_VECTOR_INREG; | |||
| 6354 | case ISD::ZERO_EXTEND: | |||
| 6355 | case ISD::ZERO_EXTEND_VECTOR_INREG: | |||
| 6356 | return ISD::ZERO_EXTEND_VECTOR_INREG; | |||
| 6357 | case ISD::SIGN_EXTEND: | |||
| 6358 | case ISD::SIGN_EXTEND_VECTOR_INREG: | |||
| 6359 | return ISD::SIGN_EXTEND_VECTOR_INREG; | |||
| 6360 | } | |||
| 6361 | llvm_unreachable("Unknown opcode")__builtin_unreachable(); | |||
| 6362 | } | |||
| 6363 | ||||
| 6364 | static SDValue getEXTEND_VECTOR_INREG(unsigned Opcode, const SDLoc &DL, EVT VT, | |||
| 6365 | SDValue In, SelectionDAG &DAG) { | |||
| 6366 | EVT InVT = In.getValueType(); | |||
| 6367 | assert(VT.isVector() && InVT.isVector() && "Expected vector VTs.")((void)0); | |||
| 6368 | assert((ISD::ANY_EXTEND == Opcode || ISD::SIGN_EXTEND == Opcode ||((void)0) | |||
| 6369 | ISD::ZERO_EXTEND == Opcode) &&((void)0) | |||
| 6370 | "Unknown extension opcode")((void)0); | |||
| 6371 | ||||
| 6372 | // For 256-bit vectors, we only need the lower (128-bit) input half. | |||
| 6373 | // For 512-bit vectors, we only need the lower input half or quarter. | |||
| 6374 | if (InVT.getSizeInBits() > 128) { | |||
| 6375 | assert(VT.getSizeInBits() == InVT.getSizeInBits() &&((void)0) | |||
| 6376 | "Expected VTs to be the same size!")((void)0); | |||
| 6377 | unsigned Scale = VT.getScalarSizeInBits() / InVT.getScalarSizeInBits(); | |||
| 6378 | In = extractSubVector(In, 0, DAG, DL, | |||
| 6379 | std::max(128U, (unsigned)VT.getSizeInBits() / Scale)); | |||
| 6380 | InVT = In.getValueType(); | |||
| 6381 | } | |||
| 6382 | ||||
| 6383 | if (VT.getVectorNumElements() != InVT.getVectorNumElements()) | |||
| 6384 | Opcode = getOpcode_EXTEND_VECTOR_INREG(Opcode); | |||
| 6385 | ||||
| 6386 | return DAG.getNode(Opcode, DL, VT, In); | |||
| 6387 | } | |||
| 6388 | ||||
| 6389 | // Match (xor X, -1) -> X. | |||
| 6390 | // Match extract_subvector(xor X, -1) -> extract_subvector(X). | |||
| 6391 | // Match concat_vectors(xor X, -1, xor Y, -1) -> concat_vectors(X, Y). | |||
| 6392 | static SDValue IsNOT(SDValue V, SelectionDAG &DAG) { | |||
| 6393 | V = peekThroughBitcasts(V); | |||
| 6394 | if (V.getOpcode() == ISD::XOR && | |||
| 6395 | ISD::isBuildVectorAllOnes(V.getOperand(1).getNode())) | |||
| 6396 | return V.getOperand(0); | |||
| 6397 | if (V.getOpcode() == ISD::EXTRACT_SUBVECTOR && | |||
| 6398 | (isNullConstant(V.getOperand(1)) || V.getOperand(0).hasOneUse())) { | |||
| 6399 | if (SDValue Not = IsNOT(V.getOperand(0), DAG)) { | |||
| 6400 | Not = DAG.getBitcast(V.getOperand(0).getValueType(), Not); | |||
| 6401 | return DAG.getNode(ISD::EXTRACT_SUBVECTOR, SDLoc(Not), V.getValueType(), | |||
| 6402 | Not, V.getOperand(1)); | |||
| 6403 | } | |||
| 6404 | } | |||
| 6405 | SmallVector<SDValue, 2> CatOps; | |||
| 6406 | if (collectConcatOps(V.getNode(), CatOps)) { | |||
| 6407 | for (SDValue &CatOp : CatOps) { | |||
| 6408 | SDValue NotCat = IsNOT(CatOp, DAG); | |||
| 6409 | if (!NotCat) return SDValue(); | |||
| 6410 | CatOp = DAG.getBitcast(CatOp.getValueType(), NotCat); | |||
| 6411 | } | |||
| 6412 | return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(V), V.getValueType(), CatOps); | |||
| 6413 | } | |||
| 6414 | return SDValue(); | |||
| 6415 | } | |||
| 6416 | ||||
| 6417 | void llvm::createUnpackShuffleMask(EVT VT, SmallVectorImpl<int> &Mask, | |||
| 6418 | bool Lo, bool Unary) { | |||
| 6419 | assert(VT.getScalarType().isSimple() && (VT.getSizeInBits() % 128) == 0 &&((void)0) | |||
| 6420 | "Illegal vector type to unpack")((void)0); | |||
| 6421 | assert(Mask.empty() && "Expected an empty shuffle mask vector")((void)0); | |||
| 6422 | int NumElts = VT.getVectorNumElements(); | |||
| 6423 | int NumEltsInLane = 128 / VT.getScalarSizeInBits(); | |||
| 6424 | for (int i = 0; i < NumElts; ++i) { | |||
| 6425 | unsigned LaneStart = (i / NumEltsInLane) * NumEltsInLane; | |||
| 6426 | int Pos = (i % NumEltsInLane) / 2 + LaneStart; | |||
| 6427 | Pos += (Unary ? 0 : NumElts * (i % 2)); | |||
| 6428 | Pos += (Lo ? 0 : NumEltsInLane / 2); | |||
| 6429 | Mask.push_back(Pos); | |||
| 6430 | } | |||
| 6431 | } | |||
| 6432 | ||||
| 6433 | /// Similar to unpacklo/unpackhi, but without the 128-bit lane limitation | |||
| 6434 | /// imposed by AVX and specific to the unary pattern. Example: | |||
| 6435 | /// v8iX Lo --> <0, 0, 1, 1, 2, 2, 3, 3> | |||
| 6436 | /// v8iX Hi --> <4, 4, 5, 5, 6, 6, 7, 7> | |||
| 6437 | void llvm::createSplat2ShuffleMask(MVT VT, SmallVectorImpl<int> &Mask, | |||
| 6438 | bool Lo) { | |||
| 6439 | assert(Mask.empty() && "Expected an empty shuffle mask vector")((void)0); | |||
| 6440 | int NumElts = VT.getVectorNumElements(); | |||
| 6441 | for (int i = 0; i < NumElts; ++i) { | |||
| 6442 | int Pos = i / 2; | |||
| 6443 | Pos += (Lo ? 0 : NumElts / 2); | |||
| 6444 | Mask.push_back(Pos); | |||
| 6445 | } | |||
| 6446 | } | |||
| 6447 | ||||
| 6448 | /// Returns a vector_shuffle node for an unpackl operation. | |||
| 6449 | static SDValue getUnpackl(SelectionDAG &DAG, const SDLoc &dl, EVT VT, | |||
| 6450 | SDValue V1, SDValue V2) { | |||
| 6451 | SmallVector<int, 8> Mask; | |||
| 6452 | createUnpackShuffleMask(VT, Mask, /* Lo = */ true, /* Unary = */ false); | |||
| 6453 | return DAG.getVectorShuffle(VT, dl, V1, V2, Mask); | |||
| 6454 | } | |||
| 6455 | ||||
| 6456 | /// Returns a vector_shuffle node for an unpackh operation. | |||
| 6457 | static SDValue getUnpackh(SelectionDAG &DAG, const SDLoc &dl, EVT VT, | |||
| 6458 | SDValue V1, SDValue V2) { | |||
| 6459 | SmallVector<int, 8> Mask; | |||
| 6460 | createUnpackShuffleMask(VT, Mask, /* Lo = */ false, /* Unary = */ false); | |||
| 6461 | return DAG.getVectorShuffle(VT, dl, V1, V2, Mask); | |||
| 6462 | } | |||
| 6463 | ||||
| 6464 | /// Return a vector_shuffle of the specified vector of zero or undef vector. | |||
| 6465 | /// This produces a shuffle where the low element of V2 is swizzled into the | |||
| 6466 | /// zero/undef vector, landing at element Idx. | |||
| 6467 | /// This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3). | |||
| 6468 | static SDValue getShuffleVectorZeroOrUndef(SDValue V2, int Idx, | |||
| 6469 | bool IsZero, | |||
| 6470 | const X86Subtarget &Subtarget, | |||
| 6471 | SelectionDAG &DAG) { | |||
| 6472 | MVT VT = V2.getSimpleValueType(); | |||
| 6473 | SDValue V1 = IsZero | |||
| 6474 | ? getZeroVector(VT, Subtarget, DAG, SDLoc(V2)) : DAG.getUNDEF(VT); | |||
| 6475 | int NumElems = VT.getVectorNumElements(); | |||
| 6476 | SmallVector<int, 16> MaskVec(NumElems); | |||
| 6477 | for (int i = 0; i != NumElems; ++i) | |||
| 6478 | // If this is the insertion idx, put the low elt of V2 here. | |||
| 6479 | MaskVec[i] = (i == Idx) ? NumElems : i; | |||
| 6480 | return DAG.getVectorShuffle(VT, SDLoc(V2), V1, V2, MaskVec); | |||
| 6481 | } | |||
| 6482 | ||||
| 6483 | static const Constant *getTargetConstantFromBasePtr(SDValue Ptr) { | |||
| 6484 | if (Ptr.getOpcode() == X86ISD::Wrapper || | |||
| 6485 | Ptr.getOpcode() == X86ISD::WrapperRIP) | |||
| 6486 | Ptr = Ptr.getOperand(0); | |||
| 6487 | ||||
| 6488 | auto *CNode = dyn_cast<ConstantPoolSDNode>(Ptr); | |||
| 6489 | if (!CNode || CNode->isMachineConstantPoolEntry() || CNode->getOffset() != 0) | |||
| 6490 | return nullptr; | |||
| 6491 | ||||
| 6492 | return CNode->getConstVal(); | |||
| 6493 | } | |||
| 6494 | ||||
| 6495 | static const Constant *getTargetConstantFromNode(LoadSDNode *Load) { | |||
| 6496 | if (!Load || !ISD::isNormalLoad(Load)) | |||
| 6497 | return nullptr; | |||
| 6498 | return getTargetConstantFromBasePtr(Load->getBasePtr()); | |||
| 6499 | } | |||
| 6500 | ||||
| 6501 | static const Constant *getTargetConstantFromNode(SDValue Op) { | |||
| 6502 | Op = peekThroughBitcasts(Op); | |||
| 6503 | return getTargetConstantFromNode(dyn_cast<LoadSDNode>(Op)); | |||
| 6504 | } | |||
| 6505 | ||||
| 6506 | const Constant * | |||
| 6507 | X86TargetLowering::getTargetConstantFromLoad(LoadSDNode *LD) const { | |||
| 6508 | assert(LD && "Unexpected null LoadSDNode")((void)0); | |||
| 6509 | return getTargetConstantFromNode(LD); | |||
| 6510 | } | |||
| 6511 | ||||
| 6512 | // Extract raw constant bits from constant pools. | |||
| 6513 | static bool getTargetConstantBitsFromNode(SDValue Op, unsigned EltSizeInBits, | |||
| 6514 | APInt &UndefElts, | |||
| 6515 | SmallVectorImpl<APInt> &EltBits, | |||
| 6516 | bool AllowWholeUndefs = true, | |||
| 6517 | bool AllowPartialUndefs = true) { | |||
| 6518 | assert(EltBits.empty() && "Expected an empty EltBits vector")((void)0); | |||
| 6519 | ||||
| 6520 | Op = peekThroughBitcasts(Op); | |||
| 6521 | ||||
| 6522 | EVT VT = Op.getValueType(); | |||
| 6523 | unsigned SizeInBits = VT.getSizeInBits(); | |||
| 6524 | assert((SizeInBits % EltSizeInBits) == 0 && "Can't split constant!")((void)0); | |||
| 6525 | unsigned NumElts = SizeInBits / EltSizeInBits; | |||
| 6526 | ||||
| 6527 | // Bitcast a source array of element bits to the target size. | |||
| 6528 | auto CastBitData = [&](APInt &UndefSrcElts, ArrayRef<APInt> SrcEltBits) { | |||
| 6529 | unsigned NumSrcElts = UndefSrcElts.getBitWidth(); | |||
| 6530 | unsigned SrcEltSizeInBits = SrcEltBits[0].getBitWidth(); | |||
| 6531 | assert((NumSrcElts * SrcEltSizeInBits) == SizeInBits &&((void)0) | |||
| 6532 | "Constant bit sizes don't match")((void)0); | |||
| 6533 | ||||
| 6534 | // Don't split if we don't allow undef bits. | |||
| 6535 | bool AllowUndefs = AllowWholeUndefs || AllowPartialUndefs; | |||
| 6536 | if (UndefSrcElts.getBoolValue() && !AllowUndefs) | |||
| 6537 | return false; | |||
| 6538 | ||||
| 6539 | // If we're already the right size, don't bother bitcasting. | |||
| 6540 | if (NumSrcElts == NumElts) { | |||
| 6541 | UndefElts = UndefSrcElts; | |||
| 6542 | EltBits.assign(SrcEltBits.begin(), SrcEltBits.end()); | |||
| 6543 | return true; | |||
| 6544 | } | |||
| 6545 | ||||
| 6546 | // Extract all the undef/constant element data and pack into single bitsets. | |||
| 6547 | APInt UndefBits(SizeInBits, 0); | |||
| 6548 | APInt MaskBits(SizeInBits, 0); | |||
| 6549 | ||||
| 6550 | for (unsigned i = 0; i != NumSrcElts; ++i) { | |||
| 6551 | unsigned BitOffset = i * SrcEltSizeInBits; | |||
| 6552 | if (UndefSrcElts[i]) | |||
| 6553 | UndefBits.setBits(BitOffset, BitOffset + SrcEltSizeInBits); | |||
| 6554 | MaskBits.insertBits(SrcEltBits[i], BitOffset); | |||
| 6555 | } | |||
| 6556 | ||||
| 6557 | // Split the undef/constant single bitset data into the target elements. | |||
| 6558 | UndefElts = APInt(NumElts, 0); | |||
| 6559 | EltBits.resize(NumElts, APInt(EltSizeInBits, 0)); | |||
| 6560 | ||||
| 6561 | for (unsigned i = 0; i != NumElts; ++i) { | |||
| 6562 | unsigned BitOffset = i * EltSizeInBits; | |||
| 6563 | APInt UndefEltBits = UndefBits.extractBits(EltSizeInBits, BitOffset); | |||
| 6564 | ||||
| 6565 | // Only treat an element as UNDEF if all bits are UNDEF. | |||
| 6566 | if (UndefEltBits.isAllOnesValue()) { | |||
| 6567 | if (!AllowWholeUndefs) | |||
| 6568 | return false; | |||
| 6569 | UndefElts.setBit(i); | |||
| 6570 | continue; | |||
| 6571 | } | |||
| 6572 | ||||
| 6573 | // If only some bits are UNDEF then treat them as zero (or bail if not | |||
| 6574 | // supported). | |||
| 6575 | if (UndefEltBits.getBoolValue() && !AllowPartialUndefs) | |||
| 6576 | return false; | |||
| 6577 | ||||
| 6578 | EltBits[i] = MaskBits.extractBits(EltSizeInBits, BitOffset); | |||
| 6579 | } | |||
| 6580 | return true; | |||
| 6581 | }; | |||
| 6582 | ||||
| 6583 | // Collect constant bits and insert into mask/undef bit masks. | |||
| 6584 | auto CollectConstantBits = [](const Constant *Cst, APInt &Mask, APInt &Undefs, | |||
| 6585 | unsigned UndefBitIndex) { | |||
| 6586 | if (!Cst) | |||
| 6587 | return false; | |||
| 6588 | if (isa<UndefValue>(Cst)) { | |||
| 6589 | Undefs.setBit(UndefBitIndex); | |||
| 6590 | return true; | |||
| 6591 | } | |||
| 6592 | if (auto *CInt = dyn_cast<ConstantInt>(Cst)) { | |||
| 6593 | Mask = CInt->getValue(); | |||
| 6594 | return true; | |||
| 6595 | } | |||
| 6596 | if (auto *CFP = dyn_cast<ConstantFP>(Cst)) { | |||
| 6597 | Mask = CFP->getValueAPF().bitcastToAPInt(); | |||
| 6598 | return true; | |||
| 6599 | } | |||
| 6600 | return false; | |||
| 6601 | }; | |||
| 6602 | ||||
| 6603 | // Handle UNDEFs. | |||
| 6604 | if (Op.isUndef()) { | |||
| 6605 | APInt UndefSrcElts = APInt::getAllOnesValue(NumElts); | |||
| 6606 | SmallVector<APInt, 64> SrcEltBits(NumElts, APInt(EltSizeInBits, 0)); | |||
| 6607 | return CastBitData(UndefSrcElts, SrcEltBits); | |||
| 6608 | } | |||
| 6609 | ||||
| 6610 | // Extract scalar constant bits. | |||
| 6611 | if (auto *Cst = dyn_cast<ConstantSDNode>(Op)) { | |||
| 6612 | APInt UndefSrcElts = APInt::getNullValue(1); | |||
| 6613 | SmallVector<APInt, 64> SrcEltBits(1, Cst->getAPIntValue()); | |||
| 6614 | return CastBitData(UndefSrcElts, SrcEltBits); | |||
| 6615 | } | |||
| 6616 | if (auto *Cst = dyn_cast<ConstantFPSDNode>(Op)) { | |||
| 6617 | APInt UndefSrcElts = APInt::getNullValue(1); | |||
| 6618 | APInt RawBits = Cst->getValueAPF().bitcastToAPInt(); | |||
| 6619 | SmallVector<APInt, 64> SrcEltBits(1, RawBits); | |||
| 6620 | return CastBitData(UndefSrcElts, SrcEltBits); | |||
| 6621 | } | |||
| 6622 | ||||
| 6623 | // Extract constant bits from build vector. | |||
| 6624 | if (ISD::isBuildVectorOfConstantSDNodes(Op.getNode())) { | |||
| 6625 | unsigned SrcEltSizeInBits = VT.getScalarSizeInBits(); | |||
| 6626 | unsigned NumSrcElts = SizeInBits / SrcEltSizeInBits; | |||
| 6627 | ||||
| 6628 | APInt UndefSrcElts(NumSrcElts, 0); | |||
| 6629 | SmallVector<APInt, 64> SrcEltBits(NumSrcElts, APInt(SrcEltSizeInBits, 0)); | |||
| 6630 | for (unsigned i = 0, e = Op.getNumOperands(); i != e; ++i) { | |||
| 6631 | const SDValue &Src = Op.getOperand(i); | |||
| 6632 | if (Src.isUndef()) { | |||
| 6633 | UndefSrcElts.setBit(i); | |||
| 6634 | continue; | |||
| 6635 | } | |||
| 6636 | auto *Cst = cast<ConstantSDNode>(Src); | |||
| 6637 | SrcEltBits[i] = Cst->getAPIntValue().zextOrTrunc(SrcEltSizeInBits); | |||
| 6638 | } | |||
| 6639 | return CastBitData(UndefSrcElts, SrcEltBits); | |||
| 6640 | } | |||
| 6641 | if (ISD::isBuildVectorOfConstantFPSDNodes(Op.getNode())) { | |||
| 6642 | unsigned SrcEltSizeInBits = VT.getScalarSizeInBits(); | |||
| 6643 | unsigned NumSrcElts = SizeInBits / SrcEltSizeInBits; | |||
| 6644 | ||||
| 6645 | APInt UndefSrcElts(NumSrcElts, 0); | |||
| 6646 | SmallVector<APInt, 64> SrcEltBits(NumSrcElts, APInt(SrcEltSizeInBits, 0)); | |||
| 6647 | for (unsigned i = 0, e = Op.getNumOperands(); i != e; ++i) { | |||
| 6648 | const SDValue &Src = Op.getOperand(i); | |||
| 6649 | if (Src.isUndef()) { | |||
| 6650 | UndefSrcElts.setBit(i); | |||
| 6651 | continue; | |||
| 6652 | } | |||
| 6653 | auto *Cst = cast<ConstantFPSDNode>(Src); | |||
| 6654 | APInt RawBits = Cst->getValueAPF().bitcastToAPInt(); | |||
| 6655 | SrcEltBits[i] = RawBits.zextOrTrunc(SrcEltSizeInBits); | |||
| 6656 | } | |||
| 6657 | return CastBitData(UndefSrcElts, SrcEltBits); | |||
| 6658 | } | |||
| 6659 | ||||
| 6660 | // Extract constant bits from constant pool vector. | |||
| 6661 | if (auto *Cst = getTargetConstantFromNode(Op)) { | |||
| 6662 | Type *CstTy = Cst->getType(); | |||
| 6663 | unsigned CstSizeInBits = CstTy->getPrimitiveSizeInBits(); | |||
| 6664 | if (!CstTy->isVectorTy() || (CstSizeInBits % SizeInBits) != 0) | |||
| 6665 | return false; | |||
| 6666 | ||||
| 6667 | unsigned SrcEltSizeInBits = CstTy->getScalarSizeInBits(); | |||
| 6668 | unsigned NumSrcElts = SizeInBits / SrcEltSizeInBits; | |||
| 6669 | ||||
| 6670 | APInt UndefSrcElts(NumSrcElts, 0); | |||
| 6671 | SmallVector<APInt, 64> SrcEltBits(NumSrcElts, APInt(SrcEltSizeInBits, 0)); | |||
| 6672 | for (unsigned i = 0; i != NumSrcElts; ++i) | |||
| 6673 | if (!CollectConstantBits(Cst->getAggregateElement(i), SrcEltBits[i], | |||
| 6674 | UndefSrcElts, i)) | |||
| 6675 | return false; | |||
| 6676 | ||||
| 6677 | return CastBitData(UndefSrcElts, SrcEltBits); | |||
| 6678 | } | |||
| 6679 | ||||
| 6680 | // Extract constant bits from a broadcasted constant pool scalar. | |||
| 6681 | if (Op.getOpcode() == X86ISD::VBROADCAST_LOAD && | |||
| 6682 | EltSizeInBits <= VT.getScalarSizeInBits()) { | |||
| 6683 | auto *MemIntr = cast<MemIntrinsicSDNode>(Op); | |||
| 6684 | if (MemIntr->getMemoryVT().getScalarSizeInBits() != VT.getScalarSizeInBits()) | |||
| 6685 | return false; | |||
| 6686 | ||||
| 6687 | SDValue Ptr = MemIntr->getBasePtr(); | |||
| 6688 | if (const Constant *C = getTargetConstantFromBasePtr(Ptr)) { | |||
| 6689 | unsigned SrcEltSizeInBits = C->getType()->getScalarSizeInBits(); | |||
| 6690 | unsigned NumSrcElts = SizeInBits / SrcEltSizeInBits; | |||
| 6691 | ||||
| 6692 | APInt UndefSrcElts(NumSrcElts, 0); | |||
| 6693 | SmallVector<APInt, 64> SrcEltBits(1, APInt(SrcEltSizeInBits, 0)); | |||
| 6694 | if (CollectConstantBits(C, SrcEltBits[0], UndefSrcElts, 0)) { | |||
| 6695 | if (UndefSrcElts[0]) | |||
| 6696 | UndefSrcElts.setBits(0, NumSrcElts); | |||
| 6697 | SrcEltBits.append(NumSrcElts - 1, SrcEltBits[0]); | |||
| 6698 | return CastBitData(UndefSrcElts, SrcEltBits); | |||
| 6699 | } | |||
| 6700 | } | |||
| 6701 | } | |||
| 6702 | ||||
| 6703 | // Extract constant bits from a subvector broadcast. | |||
| 6704 | if (Op.getOpcode() == X86ISD::SUBV_BROADCAST_LOAD) { | |||
| 6705 | auto *MemIntr = cast<MemIntrinsicSDNode>(Op); | |||
| 6706 | SDValue Ptr = MemIntr->getBasePtr(); | |||
| 6707 | // The source constant may be larger than the subvector broadcast, | |||
| 6708 | // ensure we extract the correct subvector constants. | |||
| 6709 | if (const Constant *Cst = getTargetConstantFromBasePtr(Ptr)) { | |||
| 6710 | Type *CstTy = Cst->getType(); | |||
| 6711 | unsigned CstSizeInBits = CstTy->getPrimitiveSizeInBits(); | |||
| 6712 | unsigned SubVecSizeInBits = MemIntr->getMemoryVT().getStoreSizeInBits(); | |||
| 6713 | if (!CstTy->isVectorTy() || (CstSizeInBits % SubVecSizeInBits) != 0 || | |||
| 6714 | (SizeInBits % SubVecSizeInBits) != 0) | |||
| 6715 | return false; | |||
| 6716 | unsigned CstEltSizeInBits = CstTy->getScalarSizeInBits(); | |||
| 6717 | unsigned NumSubElts = SubVecSizeInBits / CstEltSizeInBits; | |||
| 6718 | unsigned NumSubVecs = SizeInBits / SubVecSizeInBits; | |||
| 6719 | APInt UndefSubElts(NumSubElts, 0); | |||
| 6720 | SmallVector<APInt, 64> SubEltBits(NumSubElts * NumSubVecs, | |||
| 6721 | APInt(CstEltSizeInBits, 0)); | |||
| 6722 | for (unsigned i = 0; i != NumSubElts; ++i) { | |||
| 6723 | if (!CollectConstantBits(Cst->getAggregateElement(i), SubEltBits[i], | |||
| 6724 | UndefSubElts, i)) | |||
| 6725 | return false; | |||
| 6726 | for (unsigned j = 1; j != NumSubVecs; ++j) | |||
| 6727 | SubEltBits[i + (j * NumSubElts)] = SubEltBits[i]; | |||
| 6728 | } | |||
| 6729 | UndefSubElts = APInt::getSplat(NumSubVecs * UndefSubElts.getBitWidth(), | |||
| 6730 | UndefSubElts); | |||
| 6731 | return CastBitData(UndefSubElts, SubEltBits); | |||
| 6732 | } | |||
| 6733 | } | |||
| 6734 | ||||
| 6735 | // Extract a rematerialized scalar constant insertion. | |||
| 6736 | if (Op.getOpcode() == X86ISD::VZEXT_MOVL && | |||
| 6737 | Op.getOperand(0).getOpcode() == ISD::SCALAR_TO_VECTOR && | |||
| 6738 | isa<ConstantSDNode>(Op.getOperand(0).getOperand(0))) { | |||
| 6739 | unsigned SrcEltSizeInBits = VT.getScalarSizeInBits(); | |||
| 6740 | unsigned NumSrcElts = SizeInBits / SrcEltSizeInBits; | |||
| 6741 | ||||
| 6742 | APInt UndefSrcElts(NumSrcElts, 0); | |||
| 6743 | SmallVector<APInt, 64> SrcEltBits; | |||
| 6744 | auto *CN = cast<ConstantSDNode>(Op.getOperand(0).getOperand(0)); | |||
| 6745 | SrcEltBits.push_back(CN->getAPIntValue().zextOrTrunc(SrcEltSizeInBits)); | |||
| 6746 | SrcEltBits.append(NumSrcElts - 1, APInt(SrcEltSizeInBits, 0)); | |||
| 6747 | return CastBitData(UndefSrcElts, SrcEltBits); | |||
| 6748 | } | |||
| 6749 | ||||
| 6750 | // Insert constant bits from a base and sub vector sources. | |||
| 6751 | if (Op.getOpcode() == ISD::INSERT_SUBVECTOR) { | |||
| 6752 | // If bitcasts to larger elements we might lose track of undefs - don't | |||
| 6753 | // allow any to be safe. | |||
| 6754 | unsigned SrcEltSizeInBits = VT.getScalarSizeInBits(); | |||
| 6755 | bool AllowUndefs = EltSizeInBits >= SrcEltSizeInBits; | |||
| 6756 | ||||
| 6757 | APInt UndefSrcElts, UndefSubElts; | |||
| 6758 | SmallVector<APInt, 32> EltSrcBits, EltSubBits; | |||
| 6759 | if (getTargetConstantBitsFromNode(Op.getOperand(1), SrcEltSizeInBits, | |||
| 6760 | UndefSubElts, EltSubBits, | |||
| 6761 | AllowWholeUndefs && AllowUndefs, | |||
| 6762 | AllowPartialUndefs && AllowUndefs) && | |||
| 6763 | getTargetConstantBitsFromNode(Op.getOperand(0), SrcEltSizeInBits, | |||
| 6764 | UndefSrcElts, EltSrcBits, | |||
| 6765 | AllowWholeUndefs && AllowUndefs, | |||
| 6766 | AllowPartialUndefs && AllowUndefs)) { | |||
| 6767 | unsigned BaseIdx = Op.getConstantOperandVal(2); | |||
| 6768 | UndefSrcElts.insertBits(UndefSubElts, BaseIdx); | |||
| 6769 | for (unsigned i = 0, e = EltSubBits.size(); i != e; ++i) | |||
| 6770 | EltSrcBits[BaseIdx + i] = EltSubBits[i]; | |||
| 6771 | return CastBitData(UndefSrcElts, EltSrcBits); | |||
| 6772 | } | |||
| 6773 | } | |||
| 6774 | ||||
| 6775 | // Extract constant bits from a subvector's source. | |||
| 6776 | if (Op.getOpcode() == ISD::EXTRACT_SUBVECTOR) { | |||
| 6777 | // TODO - support extract_subvector through bitcasts. | |||
| 6778 | if (EltSizeInBits != VT.getScalarSizeInBits()) | |||
| 6779 | return false; | |||
| 6780 | ||||
| 6781 | if (getTargetConstantBitsFromNode(Op.getOperand(0), EltSizeInBits, | |||
| 6782 | UndefElts, EltBits, AllowWholeUndefs, | |||
| 6783 | AllowPartialUndefs)) { | |||
| 6784 | EVT SrcVT = Op.getOperand(0).getValueType(); | |||
| 6785 | unsigned NumSrcElts = SrcVT.getVectorNumElements(); | |||
| 6786 | unsigned NumSubElts = VT.getVectorNumElements(); | |||
| 6787 | unsigned BaseIdx = Op.getConstantOperandVal(1); | |||
| 6788 | UndefElts = UndefElts.extractBits(NumSubElts, BaseIdx); | |||
| 6789 | if ((BaseIdx + NumSubElts) != NumSrcElts) | |||
| 6790 | EltBits.erase(EltBits.begin() + BaseIdx + NumSubElts, EltBits.end()); | |||
| 6791 | if (BaseIdx != 0) | |||
| 6792 | EltBits.erase(EltBits.begin(), EltBits.begin() + BaseIdx); | |||
| 6793 | return true; | |||
| 6794 | } | |||
| 6795 | } | |||
| 6796 | ||||
| 6797 | // Extract constant bits from shuffle node sources. | |||
| 6798 | if (auto *SVN = dyn_cast<ShuffleVectorSDNode>(Op)) { | |||
| 6799 | // TODO - support shuffle through bitcasts. | |||
| 6800 | if (EltSizeInBits != VT.getScalarSizeInBits()) | |||
| 6801 | return false; | |||
| 6802 | ||||
| 6803 | ArrayRef<int> Mask = SVN->getMask(); | |||
| 6804 | if ((!AllowWholeUndefs || !AllowPartialUndefs) && | |||
| 6805 | llvm::any_of(Mask, [](int M) { return M < 0; })) | |||
| 6806 | return false; | |||
| 6807 | ||||
| 6808 | APInt UndefElts0, UndefElts1; | |||
| 6809 | SmallVector<APInt, 32> EltBits0, EltBits1; | |||
| 6810 | if (isAnyInRange(Mask, 0, NumElts) && | |||
| 6811 | !getTargetConstantBitsFromNode(Op.getOperand(0), EltSizeInBits, | |||
| 6812 | UndefElts0, EltBits0, AllowWholeUndefs, | |||
| 6813 | AllowPartialUndefs)) | |||
| 6814 | return false; | |||
| 6815 | if (isAnyInRange(Mask, NumElts, 2 * NumElts) && | |||
| 6816 | !getTargetConstantBitsFromNode(Op.getOperand(1), EltSizeInBits, | |||
| 6817 | UndefElts1, EltBits1, AllowWholeUndefs, | |||
| 6818 | AllowPartialUndefs)) | |||
| 6819 | return false; | |||
| 6820 | ||||
| 6821 | UndefElts = APInt::getNullValue(NumElts); | |||
| 6822 | for (int i = 0; i != (int)NumElts; ++i) { | |||
| 6823 | int M = Mask[i]; | |||
| 6824 | if (M < 0) { | |||
| 6825 | UndefElts.setBit(i); | |||
| 6826 | EltBits.push_back(APInt::getNullValue(EltSizeInBits)); | |||
| 6827 | } else if (M < (int)NumElts) { | |||
| 6828 | if (UndefElts0[M]) | |||
| 6829 | UndefElts.setBit(i); | |||
| 6830 | EltBits.push_back(EltBits0[M]); | |||
| 6831 | } else { | |||
| 6832 | if (UndefElts1[M - NumElts]) | |||
| 6833 | UndefElts.setBit(i); | |||
| 6834 | EltBits.push_back(EltBits1[M - NumElts]); | |||
| 6835 | } | |||
| 6836 | } | |||
| 6837 | return true; | |||
| 6838 | } | |||
| 6839 | ||||
| 6840 | return false; | |||
| 6841 | } | |||
| 6842 | ||||
| 6843 | namespace llvm { | |||
| 6844 | namespace X86 { | |||
| 6845 | bool isConstantSplat(SDValue Op, APInt &SplatVal, bool AllowPartialUndefs) { | |||
| 6846 | APInt UndefElts; | |||
| 6847 | SmallVector<APInt, 16> EltBits; | |||
| 6848 | if (getTargetConstantBitsFromNode(Op, Op.getScalarValueSizeInBits(), | |||
| 6849 | UndefElts, EltBits, true, | |||
| 6850 | AllowPartialUndefs)) { | |||
| 6851 | int SplatIndex = -1; | |||
| 6852 | for (int i = 0, e = EltBits.size(); i != e; ++i) { | |||
| 6853 | if (UndefElts[i]) | |||
| 6854 | continue; | |||
| 6855 | if (0 <= SplatIndex && EltBits[i] != EltBits[SplatIndex]) { | |||
| 6856 | SplatIndex = -1; | |||
| 6857 | break; | |||
| 6858 | } | |||
| 6859 | SplatIndex = i; | |||
| 6860 | } | |||
| 6861 | if (0 <= SplatIndex) { | |||
| 6862 | SplatVal = EltBits[SplatIndex]; | |||
| 6863 | return true; | |||
| 6864 | } | |||
| 6865 | } | |||
| 6866 | ||||
| 6867 | return false; | |||
| 6868 | } | |||
| 6869 | } // namespace X86 | |||
| 6870 | } // namespace llvm | |||
| 6871 | ||||
| 6872 | static bool getTargetShuffleMaskIndices(SDValue MaskNode, | |||
| 6873 | unsigned MaskEltSizeInBits, | |||
| 6874 | SmallVectorImpl<uint64_t> &RawMask, | |||
| 6875 | APInt &UndefElts) { | |||
| 6876 | // Extract the raw target constant bits. | |||
| 6877 | SmallVector<APInt, 64> EltBits; | |||
| 6878 | if (!getTargetConstantBitsFromNode(MaskNode, MaskEltSizeInBits, UndefElts, | |||
| 6879 | EltBits, /* AllowWholeUndefs */ true, | |||
| 6880 | /* AllowPartialUndefs */ false)) | |||
| 6881 | return false; | |||
| 6882 | ||||
| 6883 | // Insert the extracted elements into the mask. | |||
| 6884 | for (const APInt &Elt : EltBits) | |||
| 6885 | RawMask.push_back(Elt.getZExtValue()); | |||
| 6886 | ||||
| 6887 | return true; | |||
| 6888 | } | |||
| 6889 | ||||
| 6890 | /// Create a shuffle mask that matches the PACKSS/PACKUS truncation. | |||
| 6891 | /// A multi-stage pack shuffle mask is created by specifying NumStages > 1. | |||
| 6892 | /// Note: This ignores saturation, so inputs must be checked first. | |||
| 6893 | static void createPackShuffleMask(MVT VT, SmallVectorImpl<int> &Mask, | |||
| 6894 | bool Unary, unsigned NumStages = 1) { | |||
| 6895 | assert(Mask.empty() && "Expected an empty shuffle mask vector")((void)0); | |||
| 6896 | unsigned NumElts = VT.getVectorNumElements(); | |||
| 6897 | unsigned NumLanes = VT.getSizeInBits() / 128; | |||
| 6898 | unsigned NumEltsPerLane = 128 / VT.getScalarSizeInBits(); | |||
| 6899 | unsigned Offset = Unary ? 0 : NumElts; | |||
| 6900 | unsigned Repetitions = 1u << (NumStages - 1); | |||
| 6901 | unsigned Increment = 1u << NumStages; | |||
| 6902 | assert((NumEltsPerLane >> NumStages) > 0 && "Illegal packing compaction")((void)0); | |||
| 6903 | ||||
| 6904 | for (unsigned Lane = 0; Lane != NumLanes; ++Lane) { | |||
| 6905 | for (unsigned Stage = 0; Stage != Repetitions; ++Stage) { | |||
| 6906 | for (unsigned Elt = 0; Elt != NumEltsPerLane; Elt += Increment) | |||
| 6907 | Mask.push_back(Elt + (Lane * NumEltsPerLane)); | |||
| 6908 | for (unsigned Elt = 0; Elt != NumEltsPerLane; Elt += Increment) | |||
| 6909 | Mask.push_back(Elt + (Lane * NumEltsPerLane) + Offset); | |||
| 6910 | } | |||
| 6911 | } | |||
| 6912 | } | |||
| 6913 | ||||
| 6914 | // Split the demanded elts of a PACKSS/PACKUS node between its operands. | |||
| 6915 | static void getPackDemandedElts(EVT VT, const APInt &DemandedElts, | |||
| 6916 | APInt &DemandedLHS, APInt &DemandedRHS) { | |||
| 6917 | int NumLanes = VT.getSizeInBits() / 128; | |||
| 6918 | int NumElts = DemandedElts.getBitWidth(); | |||
| 6919 | int NumInnerElts = NumElts / 2; | |||
| 6920 | int NumEltsPerLane = NumElts / NumLanes; | |||
| 6921 | int NumInnerEltsPerLane = NumInnerElts / NumLanes; | |||
| 6922 | ||||
| 6923 | DemandedLHS = APInt::getNullValue(NumInnerElts); | |||
| 6924 | DemandedRHS = APInt::getNullValue(NumInnerElts); | |||
| 6925 | ||||
| 6926 | // Map DemandedElts to the packed operands. | |||
| 6927 | for (int Lane = 0; Lane != NumLanes; ++Lane) { | |||
| 6928 | for (int Elt = 0; Elt != NumInnerEltsPerLane; ++Elt) { | |||
| 6929 | int OuterIdx = (Lane * NumEltsPerLane) + Elt; | |||
| 6930 | int InnerIdx = (Lane * NumInnerEltsPerLane) + Elt; | |||
| 6931 | if (DemandedElts[OuterIdx]) | |||
| 6932 | DemandedLHS.setBit(InnerIdx); | |||
| 6933 | if (DemandedElts[OuterIdx + NumInnerEltsPerLane]) | |||
| 6934 | DemandedRHS.setBit(InnerIdx); | |||
| 6935 | } | |||
| 6936 | } | |||
| 6937 | } | |||
| 6938 | ||||
| 6939 | // Split the demanded elts of a HADD/HSUB node between its operands. | |||
| 6940 | static void getHorizDemandedElts(EVT VT, const APInt &DemandedElts, | |||
| 6941 | APInt &DemandedLHS, APInt &DemandedRHS) { | |||
| 6942 | int NumLanes = VT.getSizeInBits() / 128; | |||
| 6943 | int NumElts = DemandedElts.getBitWidth(); | |||
| 6944 | int NumEltsPerLane = NumElts / NumLanes; | |||
| 6945 | int HalfEltsPerLane = NumEltsPerLane / 2; | |||
| 6946 | ||||
| 6947 | DemandedLHS = APInt::getNullValue(NumElts); | |||
| 6948 | DemandedRHS = APInt::getNullValue(NumElts); | |||
| 6949 | ||||
| 6950 | // Map DemandedElts to the horizontal operands. | |||
| 6951 | for (int Idx = 0; Idx != NumElts; ++Idx) { | |||
| 6952 | if (!DemandedElts[Idx]) | |||
| 6953 | continue; | |||
| 6954 | int LaneIdx = (Idx / NumEltsPerLane) * NumEltsPerLane; | |||
| 6955 | int LocalIdx = Idx % NumEltsPerLane; | |||
| 6956 | if (LocalIdx < HalfEltsPerLane) { | |||
| 6957 | DemandedLHS.setBit(LaneIdx + 2 * LocalIdx + 0); | |||
| 6958 | DemandedLHS.setBit(LaneIdx + 2 * LocalIdx + 1); | |||
| 6959 | } else { | |||
| 6960 | LocalIdx -= HalfEltsPerLane; | |||
| 6961 | DemandedRHS.setBit(LaneIdx + 2 * LocalIdx + 0); | |||
| 6962 | DemandedRHS.setBit(LaneIdx + 2 * LocalIdx + 1); | |||
| 6963 | } | |||
| 6964 | } | |||
| 6965 | } | |||
| 6966 | ||||
| 6967 | /// Calculates the shuffle mask corresponding to the target-specific opcode. | |||
| 6968 | /// If the mask could be calculated, returns it in \p Mask, returns the shuffle | |||
| 6969 | /// operands in \p Ops, and returns true. | |||
| 6970 | /// Sets \p IsUnary to true if only one source is used. Note that this will set | |||
| 6971 | /// IsUnary for shuffles which use a single input multiple times, and in those | |||
| 6972 | /// cases it will adjust the mask to only have indices within that single input. | |||
| 6973 | /// It is an error to call this with non-empty Mask/Ops vectors. | |||
| 6974 | static bool getTargetShuffleMask(SDNode *N, MVT VT, bool AllowSentinelZero, | |||
| 6975 | SmallVectorImpl<SDValue> &Ops, | |||
| 6976 | SmallVectorImpl<int> &Mask, bool &IsUnary) { | |||
| 6977 | unsigned NumElems = VT.getVectorNumElements(); | |||
| 6978 | unsigned MaskEltSize = VT.getScalarSizeInBits(); | |||
| 6979 | SmallVector<uint64_t, 32> RawMask; | |||
| 6980 | APInt RawUndefs; | |||
| 6981 | uint64_t ImmN; | |||
| 6982 | ||||
| 6983 | assert(Mask.empty() && "getTargetShuffleMask expects an empty Mask vector")((void)0); | |||
| 6984 | assert(Ops.empty() && "getTargetShuffleMask expects an empty Ops vector")((void)0); | |||
| 6985 | ||||
| 6986 | IsUnary = false; | |||
| 6987 | bool IsFakeUnary = false; | |||
| 6988 | switch (N->getOpcode()) { | |||
| 6989 | case X86ISD::BLENDI: | |||
| 6990 | assert(N->getOperand(0).getValueType() == VT && "Unexpected value type")((void)0); | |||
| 6991 | assert(N->getOperand(1).getValueType() == VT && "Unexpected value type")((void)0); | |||
| 6992 | ImmN = N->getConstantOperandVal(N->getNumOperands() - 1); | |||
| 6993 | DecodeBLENDMask(NumElems, ImmN, Mask); | |||
| 6994 | IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1); | |||
| 6995 | break; | |||
| 6996 | case X86ISD::SHUFP: | |||
| 6997 | assert(N->getOperand(0).getValueType() == VT && "Unexpected value type")((void)0); | |||
| 6998 | assert(N->getOperand(1).getValueType() == VT && "Unexpected value type")((void)0); | |||
| 6999 | ImmN = N->getConstantOperandVal(N->getNumOperands() - 1); | |||
| 7000 | DecodeSHUFPMask(NumElems, MaskEltSize, ImmN, Mask); | |||
| 7001 | IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1); | |||
| 7002 | break; | |||
| 7003 | case X86ISD::INSERTPS: | |||
| 7004 | assert(N->getOperand(0).getValueType() == VT && "Unexpected value type")((void)0); | |||
| 7005 | assert(N->getOperand(1).getValueType() == VT && "Unexpected value type")((void)0); | |||
| 7006 | ImmN = N->getConstantOperandVal(N->getNumOperands() - 1); | |||
| 7007 | DecodeINSERTPSMask(ImmN, Mask); | |||
| 7008 | IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1); | |||
| 7009 | break; | |||
| 7010 | case X86ISD::EXTRQI: | |||
| 7011 | assert(N->getOperand(0).getValueType() == VT && "Unexpected value type")((void)0); | |||
| 7012 | if (isa<ConstantSDNode>(N->getOperand(1)) && | |||
| 7013 | isa<ConstantSDNode>(N->getOperand(2))) { | |||
| 7014 | int BitLen = N->getConstantOperandVal(1); | |||
| 7015 | int BitIdx = N->getConstantOperandVal(2); | |||
| 7016 | DecodeEXTRQIMask(NumElems, MaskEltSize, BitLen, BitIdx, Mask); | |||
| 7017 | IsUnary = true; | |||
| 7018 | } | |||
| 7019 | break; | |||
| 7020 | case X86ISD::INSERTQI: | |||
| 7021 | assert(N->getOperand(0).getValueType() == VT && "Unexpected value type")((void)0); | |||
| 7022 | assert(N->getOperand(1).getValueType() == VT && "Unexpected value type")((void)0); | |||
| 7023 | if (isa<ConstantSDNode>(N->getOperand(2)) && | |||
| 7024 | isa<ConstantSDNode>(N->getOperand(3))) { | |||
| 7025 | int BitLen = N->getConstantOperandVal(2); | |||
| 7026 | int BitIdx = N->getConstantOperandVal(3); | |||
| 7027 | DecodeINSERTQIMask(NumElems, MaskEltSize, BitLen, BitIdx, Mask); | |||
| 7028 | IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1); | |||
| 7029 | } | |||
| 7030 | break; | |||
| 7031 | case X86ISD::UNPCKH: | |||
| 7032 | assert(N->getOperand(0).getValueType() == VT && "Unexpected value type")((void)0); | |||
| 7033 | assert(N->getOperand(1).getValueType() == VT && "Unexpected value type")((void)0); | |||
| 7034 | DecodeUNPCKHMask(NumElems, MaskEltSize, Mask); | |||
| 7035 | IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1); | |||
| 7036 | break; | |||
| 7037 | case X86ISD::UNPCKL: | |||
| 7038 | assert(N->getOperand(0).getValueType() == VT && "Unexpected value type")((void)0); | |||
| 7039 | assert(N->getOperand(1).getValueType() == VT && "Unexpected value type")((void)0); | |||
| 7040 | DecodeUNPCKLMask(NumElems, MaskEltSize, Mask); | |||
| 7041 | IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1); | |||
| 7042 | break; | |||
| 7043 | case X86ISD::MOVHLPS: | |||
| 7044 | assert(N->getOperand(0).getValueType() == VT && "Unexpected value type")((void)0); | |||
| 7045 | assert(N->getOperand(1).getValueType() == VT && "Unexpected value type")((void)0); | |||
| 7046 | DecodeMOVHLPSMask(NumElems, Mask); | |||
| 7047 | IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1); | |||
| 7048 | break; | |||
| 7049 | case X86ISD::MOVLHPS: | |||
| 7050 | assert(N->getOperand(0).getValueType() == VT && "Unexpected value type")((void)0); | |||
| 7051 | assert(N->getOperand(1).getValueType() == VT && "Unexpected value type")((void)0); | |||
| 7052 | DecodeMOVLHPSMask(NumElems, Mask); | |||
| 7053 | IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1); | |||
| 7054 | break; | |||
| 7055 | case X86ISD::VALIGN: | |||
| 7056 | assert((VT.getScalarType() == MVT::i32 || VT.getScalarType() == MVT::i64) &&((void)0) | |||
| 7057 | "Only 32-bit and 64-bit elements are supported!")((void)0); | |||
| 7058 | assert(N->getOperand(0).getValueType() == VT && "Unexpected value type")((void)0); | |||
| 7059 | assert(N->getOperand(1).getValueType() == VT && "Unexpected value type")((void)0); | |||
| 7060 | ImmN = N->getConstantOperandVal(N->getNumOperands() - 1); | |||
| 7061 | DecodeVALIGNMask(NumElems, ImmN, Mask); | |||
| 7062 | IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1); | |||
| 7063 | Ops.push_back(N->getOperand(1)); | |||
| 7064 | Ops.push_back(N->getOperand(0)); | |||
| 7065 | break; | |||
| 7066 | case X86ISD::PALIGNR: | |||
| 7067 | assert(VT.getScalarType() == MVT::i8 && "Byte vector expected")((void)0); | |||
| 7068 | assert(N->getOperand(0).getValueType() == VT && "Unexpected value type")((void)0); | |||
| 7069 | assert(N->getOperand(1).getValueType() == VT && "Unexpected value type")((void)0); | |||
| 7070 | ImmN = N->getConstantOperandVal(N->getNumOperands() - 1); | |||
| 7071 | DecodePALIGNRMask(NumElems, ImmN, Mask); | |||
| 7072 | IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1); | |||
| 7073 | Ops.push_back(N->getOperand(1)); | |||
| 7074 | Ops.push_back(N->getOperand(0)); | |||
| 7075 | break; | |||
| 7076 | case X86ISD::VSHLDQ: | |||
| 7077 | assert(VT.getScalarType() == MVT::i8 && "Byte vector expected")((void)0); | |||
| 7078 | assert(N->getOperand(0).getValueType() == VT && "Unexpected value type")((void)0); | |||
| 7079 | ImmN = N->getConstantOperandVal(N->getNumOperands() - 1); | |||
| 7080 | DecodePSLLDQMask(NumElems, ImmN, Mask); | |||
| 7081 | IsUnary = true; | |||
| 7082 | break; | |||
| 7083 | case X86ISD::VSRLDQ: | |||
| 7084 | assert(VT.getScalarType() == MVT::i8 && "Byte vector expected")((void)0); | |||
| 7085 | assert(N->getOperand(0).getValueType() == VT && "Unexpected value type")((void)0); | |||
| 7086 | ImmN = N->getConstantOperandVal(N->getNumOperands() - 1); | |||
| 7087 | DecodePSRLDQMask(NumElems, ImmN, Mask); | |||
| 7088 | IsUnary = true; | |||
| 7089 | break; | |||
| 7090 | case X86ISD::PSHUFD: | |||
| 7091 | case X86ISD::VPERMILPI: | |||
| 7092 | assert(N->getOperand(0).getValueType() == VT && "Unexpected value type")((void)0); | |||
| 7093 | ImmN = N->getConstantOperandVal(N->getNumOperands() - 1); | |||
| 7094 | DecodePSHUFMask(NumElems, MaskEltSize, ImmN, Mask); | |||
| 7095 | IsUnary = true; | |||
| 7096 | break; | |||
| 7097 | case X86ISD::PSHUFHW: | |||
| 7098 | assert(N->getOperand(0).getValueType() == VT && "Unexpected value type")((void)0); | |||
| 7099 | ImmN = N->getConstantOperandVal(N->getNumOperands() - 1); | |||
| 7100 | DecodePSHUFHWMask(NumElems, ImmN, Mask); | |||
| 7101 | IsUnary = true; | |||
| 7102 | break; | |||
| 7103 | case X86ISD::PSHUFLW: | |||
| 7104 | assert(N->getOperand(0).getValueType() == VT && "Unexpected value type")((void)0); | |||
| 7105 | ImmN = N->getConstantOperandVal(N->getNumOperands() - 1); | |||
| 7106 | DecodePSHUFLWMask(NumElems, ImmN, Mask); | |||
| 7107 | IsUnary = true; | |||
| 7108 | break; | |||
| 7109 | case X86ISD::VZEXT_MOVL: | |||
| 7110 | assert(N->getOperand(0).getValueType() == VT && "Unexpected value type")((void)0); | |||
| 7111 | DecodeZeroMoveLowMask(NumElems, Mask); | |||
| 7112 | IsUnary = true; | |||
| 7113 | break; | |||
| 7114 | case X86ISD::VBROADCAST: | |||
| 7115 | // We only decode broadcasts of same-sized vectors, peeking through to | |||
| 7116 | // extracted subvectors is likely to cause hasOneUse issues with | |||
| 7117 | // SimplifyDemandedBits etc. | |||
| 7118 | if (N->getOperand(0).getValueType() == VT) { | |||
| 7119 | DecodeVectorBroadcast(NumElems, Mask); | |||
| 7120 | IsUnary = true; | |||
| 7121 | break; | |||
| 7122 | } | |||
| 7123 | return false; | |||
| 7124 | case X86ISD::VPERMILPV: { | |||
| 7125 | assert(N->getOperand(0).getValueType() == VT && "Unexpected value type")((void)0); | |||
| 7126 | IsUnary = true; | |||
| 7127 | SDValue MaskNode = N->getOperand(1); | |||
| 7128 | if (getTargetShuffleMaskIndices(MaskNode, MaskEltSize, RawMask, | |||
| 7129 | RawUndefs)) { | |||
| 7130 | DecodeVPERMILPMask(NumElems, MaskEltSize, RawMask, RawUndefs, Mask); | |||
| 7131 | break; | |||
| 7132 | } | |||
| 7133 | return false; | |||
| 7134 | } | |||
| 7135 | case X86ISD::PSHUFB: { | |||
| 7136 | assert(VT.getScalarType() == MVT::i8 && "Byte vector expected")((void)0); | |||
| 7137 | assert(N->getOperand(0).getValueType() == VT && "Unexpected value type")((void)0); | |||
| 7138 | assert(N->getOperand(1).getValueType() == VT && "Unexpected value type")((void)0); | |||
| 7139 | IsUnary = true; | |||
| 7140 | SDValue MaskNode = N->getOperand(1); | |||
| 7141 | if (getTargetShuffleMaskIndices(MaskNode, 8, RawMask, RawUndefs)) { | |||
| 7142 | DecodePSHUFBMask(RawMask, RawUndefs, Mask); | |||
| 7143 | break; | |||
| 7144 | } | |||
| 7145 | return false; | |||
| 7146 | } | |||
| 7147 | case X86ISD::VPERMI: | |||
| 7148 | assert(N->getOperand(0).getValueType() == VT && "Unexpected value type")((void)0); | |||
| 7149 | ImmN = N->getConstantOperandVal(N->getNumOperands() - 1); | |||
| 7150 | DecodeVPERMMask(NumElems, ImmN, Mask); | |||
| 7151 | IsUnary = true; | |||
| 7152 | break; | |||
| 7153 | case X86ISD::MOVSS: | |||
| 7154 | case X86ISD::MOVSD: | |||
| 7155 | assert(N->getOperand(0).getValueType() == VT && "Unexpected value type")((void)0); | |||
| 7156 | assert(N->getOperand(1).getValueType() == VT && "Unexpected value type")((void)0); | |||
| 7157 | DecodeScalarMoveMask(NumElems, /* IsLoad */ false, Mask); | |||
| 7158 | break; | |||
| 7159 | case X86ISD::VPERM2X128: | |||
| 7160 | assert(N->getOperand(0).getValueType() == VT && "Unexpected value type")((void)0); | |||
| 7161 | assert(N->getOperand(1).getValueType() == VT && "Unexpected value type")((void)0); | |||
| 7162 | ImmN = N->getConstantOperandVal(N->getNumOperands() - 1); | |||
| 7163 | DecodeVPERM2X128Mask(NumElems, ImmN, Mask); | |||
| 7164 | IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1); | |||
| 7165 | break; | |||
| 7166 | case X86ISD::SHUF128: | |||
| 7167 | assert(N->getOperand(0).getValueType() == VT && "Unexpected value type")((void)0); | |||
| 7168 | assert(N->getOperand(1).getValueType() == VT && "Unexpected value type")((void)0); | |||
| 7169 | ImmN = N->getConstantOperandVal(N->getNumOperands() - 1); | |||
| 7170 | decodeVSHUF64x2FamilyMask(NumElems, MaskEltSize, ImmN, Mask); | |||
| 7171 | IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1); | |||
| 7172 | break; | |||
| 7173 | case X86ISD::MOVSLDUP: | |||
| 7174 | assert(N->getOperand(0).getValueType() == VT && "Unexpected value type")((void)0); | |||
| 7175 | DecodeMOVSLDUPMask(NumElems, Mask); | |||
| 7176 | IsUnary = true; | |||
| 7177 | break; | |||
| 7178 | case X86ISD::MOVSHDUP: | |||
| 7179 | assert(N->getOperand(0).getValueType() == VT && "Unexpected value type")((void)0); | |||
| 7180 | DecodeMOVSHDUPMask(NumElems, Mask); | |||
| 7181 | IsUnary = true; | |||
| 7182 | break; | |||
| 7183 | case X86ISD::MOVDDUP: | |||
| 7184 | assert(N->getOperand(0).getValueType() == VT && "Unexpected value type")((void)0); | |||
| 7185 | DecodeMOVDDUPMask(NumElems, Mask); | |||
| 7186 | IsUnary = true; | |||
| 7187 | break; | |||
| 7188 | case X86ISD::VPERMIL2: { | |||
| 7189 | assert(N->getOperand(0).getValueType() == VT && "Unexpected value type")((void)0); | |||
| 7190 | assert(N->getOperand(1).getValueType() == VT && "Unexpected value type")((void)0); | |||
| 7191 | IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1); | |||
| 7192 | SDValue MaskNode = N->getOperand(2); | |||
| 7193 | SDValue CtrlNode = N->getOperand(3); | |||
| 7194 | if (ConstantSDNode *CtrlOp = dyn_cast<ConstantSDNode>(CtrlNode)) { | |||
| 7195 | unsigned CtrlImm = CtrlOp->getZExtValue(); | |||
| 7196 | if (getTargetShuffleMaskIndices(MaskNode, MaskEltSize, RawMask, | |||
| 7197 | RawUndefs)) { | |||
| 7198 | DecodeVPERMIL2PMask(NumElems, MaskEltSize, CtrlImm, RawMask, RawUndefs, | |||
| 7199 | Mask); | |||
| 7200 | break; | |||
| 7201 | } | |||
| 7202 | } | |||
| 7203 | return false; | |||
| 7204 | } | |||
| 7205 | case X86ISD::VPPERM: { | |||
| 7206 | assert(N->getOperand(0).getValueType() == VT && "Unexpected value type")((void)0); | |||
| 7207 | assert(N->getOperand(1).getValueType() == VT && "Unexpected value type")((void)0); | |||
| 7208 | IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1); | |||
| 7209 | SDValue MaskNode = N->getOperand(2); | |||
| 7210 | if (getTargetShuffleMaskIndices(MaskNode, 8, RawMask, RawUndefs)) { | |||
| 7211 | DecodeVPPERMMask(RawMask, RawUndefs, Mask); | |||
| 7212 | break; | |||
| 7213 | } | |||
| 7214 | return false; | |||
| 7215 | } | |||
| 7216 | case X86ISD::VPERMV: { | |||
| 7217 | assert(N->getOperand(1).getValueType() == VT && "Unexpected value type")((void)0); | |||
| 7218 | IsUnary = true; | |||
| 7219 | // Unlike most shuffle nodes, VPERMV's mask operand is operand 0. | |||
| 7220 | Ops.push_back(N->getOperand(1)); | |||
| 7221 | SDValue MaskNode = N->getOperand(0); | |||
| 7222 | if (getTargetShuffleMaskIndices(MaskNode, MaskEltSize, RawMask, | |||
| 7223 | RawUndefs)) { | |||
| 7224 | DecodeVPERMVMask(RawMask, RawUndefs, Mask); | |||
| 7225 | break; | |||
| 7226 | } | |||
| 7227 | return false; | |||
| 7228 | } | |||
| 7229 | case X86ISD::VPERMV3: { | |||
| 7230 | assert(N->getOperand(0).getValueType() == VT && "Unexpected value type")((void)0); | |||
| 7231 | assert(N->getOperand(2).getValueType() == VT && "Unexpected value type")((void)0); | |||
| 7232 | IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(2); | |||
| 7233 | // Unlike most shuffle nodes, VPERMV3's mask operand is the middle one. | |||
| 7234 | Ops.push_back(N->getOperand(0)); | |||
| 7235 | Ops.push_back(N->getOperand(2)); | |||
| 7236 | SDValue MaskNode = N->getOperand(1); | |||
| 7237 | if (getTargetShuffleMaskIndices(MaskNode, MaskEltSize, RawMask, | |||
| 7238 | RawUndefs)) { | |||
| 7239 | DecodeVPERMV3Mask(RawMask, RawUndefs, Mask); | |||
| 7240 | break; | |||
| 7241 | } | |||
| 7242 | return false; | |||
| 7243 | } | |||
| 7244 | default: llvm_unreachable("unknown target shuffle node")__builtin_unreachable(); | |||
| 7245 | } | |||
| 7246 | ||||
| 7247 | // Empty mask indicates the decode failed. | |||
| 7248 | if (Mask.empty()) | |||
| 7249 | return false; | |||
| 7250 | ||||
| 7251 | // Check if we're getting a shuffle mask with zero'd elements. | |||
| 7252 | if (!AllowSentinelZero && isAnyZero(Mask)) | |||
| 7253 | return false; | |||
| 7254 | ||||
| 7255 | // If we have a fake unary shuffle, the shuffle mask is spread across two | |||
| 7256 | // inputs that are actually the same node. Re-map the mask to always point | |||
| 7257 | // into the first input. | |||
| 7258 | if (IsFakeUnary) | |||
| 7259 | for (int &M : Mask) | |||
| 7260 | if (M >= (int)Mask.size()) | |||
| 7261 | M -= Mask.size(); | |||
| 7262 | ||||
| 7263 | // If we didn't already add operands in the opcode-specific code, default to | |||
| 7264 | // adding 1 or 2 operands starting at 0. | |||
| 7265 | if (Ops.empty()) { | |||
| 7266 | Ops.push_back(N->getOperand(0)); | |||
| 7267 | if (!IsUnary || IsFakeUnary) | |||
| 7268 | Ops.push_back(N->getOperand(1)); | |||
| 7269 | } | |||
| 7270 | ||||
| 7271 | return true; | |||
| 7272 | } | |||
| 7273 | ||||
| 7274 | // Wrapper for getTargetShuffleMask with InUnary; | |||
| 7275 | static bool getTargetShuffleMask(SDNode *N, MVT VT, bool AllowSentinelZero, | |||
| 7276 | SmallVectorImpl<SDValue> &Ops, | |||
| 7277 | SmallVectorImpl<int> &Mask) { | |||
| 7278 | bool IsUnary; | |||
| 7279 | return getTargetShuffleMask(N, VT, AllowSentinelZero, Ops, Mask, IsUnary); | |||
| 7280 | } | |||
| 7281 | ||||
| 7282 | /// Compute whether each element of a shuffle is zeroable. | |||
| 7283 | /// | |||
| 7284 | /// A "zeroable" vector shuffle element is one which can be lowered to zero. | |||
| 7285 | /// Either it is an undef element in the shuffle mask, the element of the input | |||
| 7286 | /// referenced is undef, or the element of the input referenced is known to be | |||
| 7287 | /// zero. Many x86 shuffles can zero lanes cheaply and we often want to handle | |||
| 7288 | /// as many lanes with this technique as possible to simplify the remaining | |||
| 7289 | /// shuffle. | |||
| 7290 | static void computeZeroableShuffleElements(ArrayRef<int> Mask, | |||
| 7291 | SDValue V1, SDValue V2, | |||
| 7292 | APInt &KnownUndef, APInt &KnownZero) { | |||
| 7293 | int Size = Mask.size(); | |||
| 7294 | KnownUndef = KnownZero = APInt::getNullValue(Size); | |||
| 7295 | ||||
| 7296 | V1 = peekThroughBitcasts(V1); | |||
| 7297 | V2 = peekThroughBitcasts(V2); | |||
| 7298 | ||||
| 7299 | bool V1IsZero = ISD::isBuildVectorAllZeros(V1.getNode()); | |||
| 7300 | bool V2IsZero = ISD::isBuildVectorAllZeros(V2.getNode()); | |||
| 7301 | ||||
| 7302 | int VectorSizeInBits = V1.getValueSizeInBits(); | |||
| 7303 | int ScalarSizeInBits = VectorSizeInBits / Size; | |||
| 7304 | assert(!(VectorSizeInBits % ScalarSizeInBits) && "Illegal shuffle mask size")((void)0); | |||
| 7305 | ||||
| 7306 | for (int i = 0; i < Size; ++i) { | |||
| 7307 | int M = Mask[i]; | |||
| 7308 | // Handle the easy cases. | |||
| 7309 | if (M < 0) { | |||
| 7310 | KnownUndef.setBit(i); | |||
| 7311 | continue; | |||
| 7312 | } | |||
| 7313 | if ((M >= 0 && M < Size && V1IsZero) || (M >= Size && V2IsZero)) { | |||
| 7314 | KnownZero.setBit(i); | |||
| 7315 | continue; | |||
| 7316 | } | |||
| 7317 | ||||
| 7318 | // Determine shuffle input and normalize the mask. | |||
| 7319 | SDValue V = M < Size ? V1 : V2; | |||
| 7320 | M %= Size; | |||
| 7321 | ||||
| 7322 | // Currently we can only search BUILD_VECTOR for UNDEF/ZERO elements. | |||
| 7323 | if (V.getOpcode() != ISD::BUILD_VECTOR) | |||
| 7324 | continue; | |||
| 7325 | ||||
| 7326 | // If the BUILD_VECTOR has fewer elements then the bitcasted portion of | |||
| 7327 | // the (larger) source element must be UNDEF/ZERO. | |||
| 7328 | if ((Size % V.getNumOperands()) == 0) { | |||
| 7329 | int Scale = Size / V->getNumOperands(); | |||
| 7330 | SDValue Op = V.getOperand(M / Scale); | |||
| 7331 | if (Op.isUndef()) | |||
| 7332 | KnownUndef.setBit(i); | |||
| 7333 | if (X86::isZeroNode(Op)) | |||
| 7334 | KnownZero.setBit(i); | |||
| 7335 | else if (ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(Op)) { | |||
| 7336 | APInt Val = Cst->getAPIntValue(); | |||
| 7337 | Val = Val.extractBits(ScalarSizeInBits, (M % Scale) * ScalarSizeInBits); | |||
| 7338 | if (Val == 0) | |||
| 7339 | KnownZero.setBit(i); | |||
| 7340 | } else if (ConstantFPSDNode *Cst = dyn_cast<ConstantFPSDNode>(Op)) { | |||
| 7341 | APInt Val = Cst->getValueAPF().bitcastToAPInt(); | |||
| 7342 | Val = Val.extractBits(ScalarSizeInBits, (M % Scale) * ScalarSizeInBits); | |||
| 7343 | if (Val == 0) | |||
| 7344 | KnownZero.setBit(i); | |||
| 7345 | } | |||
| 7346 | continue; | |||
| 7347 | } | |||
| 7348 | ||||
| 7349 | // If the BUILD_VECTOR has more elements then all the (smaller) source | |||
| 7350 | // elements must be UNDEF or ZERO. | |||
| 7351 | if ((V.getNumOperands() % Size) == 0) { | |||
| 7352 | int Scale = V->getNumOperands() / Size; | |||
| 7353 | bool AllUndef = true; | |||
| 7354 | bool AllZero = true; | |||
| 7355 | for (int j = 0; j < Scale; ++j) { | |||
| 7356 | SDValue Op = V.getOperand((M * Scale) + j); | |||
| 7357 | AllUndef &= Op.isUndef(); | |||
| 7358 | AllZero &= X86::isZeroNode(Op); | |||
| 7359 | } | |||
| 7360 | if (AllUndef) | |||
| 7361 | KnownUndef.setBit(i); | |||
| 7362 | if (AllZero) | |||
| 7363 | KnownZero.setBit(i); | |||
| 7364 | continue; | |||
| 7365 | } | |||
| 7366 | } | |||
| 7367 | } | |||
| 7368 | ||||
| 7369 | /// Decode a target shuffle mask and inputs and see if any values are | |||
| 7370 | /// known to be undef or zero from their inputs. | |||
| 7371 | /// Returns true if the target shuffle mask was decoded. | |||
| 7372 | /// FIXME: Merge this with computeZeroableShuffleElements? | |||
| 7373 | static bool getTargetShuffleAndZeroables(SDValue N, SmallVectorImpl<int> &Mask, | |||
| 7374 | SmallVectorImpl<SDValue> &Ops, | |||
| 7375 | APInt &KnownUndef, APInt &KnownZero) { | |||
| 7376 | bool IsUnary; | |||
| 7377 | if (!isTargetShuffle(N.getOpcode())) | |||
| 7378 | return false; | |||
| 7379 | ||||
| 7380 | MVT VT = N.getSimpleValueType(); | |||
| 7381 | if (!getTargetShuffleMask(N.getNode(), VT, true, Ops, Mask, IsUnary)) | |||
| 7382 | return false; | |||
| 7383 | ||||
| 7384 | int Size = Mask.size(); | |||
| 7385 | SDValue V1 = Ops[0]; | |||
| 7386 | SDValue V2 = IsUnary ? V1 : Ops[1]; | |||
| 7387 | KnownUndef = KnownZero = APInt::getNullValue(Size); | |||
| 7388 | ||||
| 7389 | V1 = peekThroughBitcasts(V1); | |||
| 7390 | V2 = peekThroughBitcasts(V2); | |||
| 7391 | ||||
| 7392 | assert((VT.getSizeInBits() % Size) == 0 &&((void)0) | |||
| 7393 | "Illegal split of shuffle value type")((void)0); | |||
| 7394 | unsigned EltSizeInBits = VT.getSizeInBits() / Size; | |||
| 7395 | ||||
| 7396 | // Extract known constant input data. | |||
| 7397 | APInt UndefSrcElts[2]; | |||
| 7398 | SmallVector<APInt, 32> SrcEltBits[2]; | |||
| 7399 | bool IsSrcConstant[2] = { | |||
| 7400 | getTargetConstantBitsFromNode(V1, EltSizeInBits, UndefSrcElts[0], | |||
| 7401 | SrcEltBits[0], true, false), | |||
| 7402 | getTargetConstantBitsFromNode(V2, EltSizeInBits, UndefSrcElts[1], | |||
| 7403 | SrcEltBits[1], true, false)}; | |||
| 7404 | ||||
| 7405 | for (int i = 0; i < Size; ++i) { | |||
| 7406 | int M = Mask[i]; | |||
| 7407 | ||||
| 7408 | // Already decoded as SM_SentinelZero / SM_SentinelUndef. | |||
| 7409 | if (M < 0) { | |||
| 7410 | assert(isUndefOrZero(M) && "Unknown shuffle sentinel value!")((void)0); | |||
| 7411 | if (SM_SentinelUndef == M) | |||
| 7412 | KnownUndef.setBit(i); | |||
| 7413 | if (SM_SentinelZero == M) | |||
| 7414 | KnownZero.setBit(i); | |||
| 7415 | continue; | |||
| 7416 | } | |||
| 7417 | ||||
| 7418 | // Determine shuffle input and normalize the mask. | |||
| 7419 | unsigned SrcIdx = M / Size; | |||
| 7420 | SDValue V = M < Size ? V1 : V2; | |||
| 7421 | M %= Size; | |||
| 7422 | ||||
| 7423 | // We are referencing an UNDEF input. | |||
| 7424 | if (V.isUndef()) { | |||
| 7425 | KnownUndef.setBit(i); | |||
| 7426 | continue; | |||
| 7427 | } | |||
| 7428 | ||||
| 7429 | // SCALAR_TO_VECTOR - only the first element is defined, and the rest UNDEF. | |||
| 7430 | // TODO: We currently only set UNDEF for integer types - floats use the same | |||
| 7431 | // registers as vectors and many of the scalar folded loads rely on the | |||
| 7432 | // SCALAR_TO_VECTOR pattern. | |||
| 7433 | if (V.getOpcode() == ISD::SCALAR_TO_VECTOR && | |||
| 7434 | (Size % V.getValueType().getVectorNumElements()) == 0) { | |||
| 7435 | int Scale = Size / V.getValueType().getVectorNumElements(); | |||
| 7436 | int Idx = M / Scale; | |||
| 7437 | if (Idx != 0 && !VT.isFloatingPoint()) | |||
| 7438 | KnownUndef.setBit(i); | |||
| 7439 | else if (Idx == 0 && X86::isZeroNode(V.getOperand(0))) | |||
| 7440 | KnownZero.setBit(i); | |||
| 7441 | continue; | |||
| 7442 | } | |||
| 7443 | ||||
| 7444 | // INSERT_SUBVECTOR - to widen vectors we often insert them into UNDEF | |||
| 7445 | // base vectors. | |||
| 7446 | if (V.getOpcode() == ISD::INSERT_SUBVECTOR) { | |||
| 7447 | SDValue Vec = V.getOperand(0); | |||
| 7448 | int NumVecElts = Vec.getValueType().getVectorNumElements(); | |||
| 7449 | if (Vec.isUndef() && Size == NumVecElts) { | |||
| 7450 | int Idx = V.getConstantOperandVal(2); | |||
| 7451 | int NumSubElts = V.getOperand(1).getValueType().getVectorNumElements(); | |||
| 7452 | if (M < Idx || (Idx + NumSubElts) <= M) | |||
| 7453 | KnownUndef.setBit(i); | |||
| 7454 | } | |||
| 7455 | continue; | |||
| 7456 | } | |||
| 7457 | ||||
| 7458 | // Attempt to extract from the source's constant bits. | |||
| 7459 | if (IsSrcConstant[SrcIdx]) { | |||
| 7460 | if (UndefSrcElts[SrcIdx][M]) | |||
| 7461 | KnownUndef.setBit(i); | |||
| 7462 | else if (SrcEltBits[SrcIdx][M] == 0) | |||
| 7463 | KnownZero.setBit(i); | |||
| 7464 | } | |||
| 7465 | } | |||
| 7466 | ||||
| 7467 | assert(VT.getVectorNumElements() == (unsigned)Size &&((void)0) | |||
| 7468 | "Different mask size from vector size!")((void)0); | |||
| 7469 | return true; | |||
| 7470 | } | |||
| 7471 | ||||
| 7472 | // Replace target shuffle mask elements with known undef/zero sentinels. | |||
| 7473 | static void resolveTargetShuffleFromZeroables(SmallVectorImpl<int> &Mask, | |||
| 7474 | const APInt &KnownUndef, | |||
| 7475 | const APInt &KnownZero, | |||
| 7476 | bool ResolveKnownZeros= true) { | |||
| 7477 | unsigned NumElts = Mask.size(); | |||
| 7478 | assert(KnownUndef.getBitWidth() == NumElts &&((void)0) | |||
| 7479 | KnownZero.getBitWidth() == NumElts && "Shuffle mask size mismatch")((void)0); | |||
| 7480 | ||||
| 7481 | for (unsigned i = 0; i != NumElts; ++i) { | |||
| 7482 | if (KnownUndef[i]) | |||
| 7483 | Mask[i] = SM_SentinelUndef; | |||
| 7484 | else if (ResolveKnownZeros && KnownZero[i]) | |||
| 7485 | Mask[i] = SM_SentinelZero; | |||
| 7486 | } | |||
| 7487 | } | |||
| 7488 | ||||
| 7489 | // Extract target shuffle mask sentinel elements to known undef/zero bitmasks. | |||
| 7490 | static void resolveZeroablesFromTargetShuffle(const SmallVectorImpl<int> &Mask, | |||
| 7491 | APInt &KnownUndef, | |||
| 7492 | APInt &KnownZero) { | |||
| 7493 | unsigned NumElts = Mask.size(); | |||
| 7494 | KnownUndef = KnownZero = APInt::getNullValue(NumElts); | |||
| 7495 | ||||
| 7496 | for (unsigned i = 0; i != NumElts; ++i) { | |||
| 7497 | int M = Mask[i]; | |||
| 7498 | if (SM_SentinelUndef == M) | |||
| 7499 | KnownUndef.setBit(i); | |||
| 7500 | if (SM_SentinelZero == M) | |||
| 7501 | KnownZero.setBit(i); | |||
| 7502 | } | |||
| 7503 | } | |||
| 7504 | ||||
| 7505 | // Forward declaration (for getFauxShuffleMask recursive check). | |||
| 7506 | // TODO: Use DemandedElts variant. | |||
| 7507 | static bool getTargetShuffleInputs(SDValue Op, SmallVectorImpl<SDValue> &Inputs, | |||
| 7508 | SmallVectorImpl<int> &Mask, | |||
| 7509 | const SelectionDAG &DAG, unsigned Depth, | |||
| 7510 | bool ResolveKnownElts); | |||
| 7511 | ||||
| 7512 | // Attempt to decode ops that could be represented as a shuffle mask. | |||
| 7513 | // The decoded shuffle mask may contain a different number of elements to the | |||
| 7514 | // destination value type. | |||
| 7515 | static bool getFauxShuffleMask(SDValue N, const APInt &DemandedElts, | |||
| 7516 | SmallVectorImpl<int> &Mask, | |||
| 7517 | SmallVectorImpl<SDValue> &Ops, | |||
| 7518 | const SelectionDAG &DAG, unsigned Depth, | |||
| 7519 | bool ResolveKnownElts) { | |||
| 7520 | Mask.clear(); | |||
| 7521 | Ops.clear(); | |||
| 7522 | ||||
| 7523 | MVT VT = N.getSimpleValueType(); | |||
| 7524 | unsigned NumElts = VT.getVectorNumElements(); | |||
| 7525 | unsigned NumSizeInBits = VT.getSizeInBits(); | |||
| 7526 | unsigned NumBitsPerElt = VT.getScalarSizeInBits(); | |||
| 7527 | if ((NumBitsPerElt % 8) != 0 || (NumSizeInBits % 8) != 0) | |||
| 7528 | return false; | |||
| 7529 | assert(NumElts == DemandedElts.getBitWidth() && "Unexpected vector size")((void)0); | |||
| 7530 | unsigned NumSizeInBytes = NumSizeInBits / 8; | |||
| 7531 | unsigned NumBytesPerElt = NumBitsPerElt / 8; | |||
| 7532 | ||||
| 7533 | unsigned Opcode = N.getOpcode(); | |||
| 7534 | switch (Opcode) { | |||
| 7535 | case ISD::VECTOR_SHUFFLE: { | |||
| 7536 | // Don't treat ISD::VECTOR_SHUFFLE as a target shuffle so decode it here. | |||
| 7537 | ArrayRef<int> ShuffleMask = cast<ShuffleVectorSDNode>(N)->getMask(); | |||
| 7538 | if (isUndefOrInRange(ShuffleMask, 0, 2 * NumElts)) { | |||
| 7539 | Mask.append(ShuffleMask.begin(), ShuffleMask.end()); | |||
| 7540 | Ops.push_back(N.getOperand(0)); | |||
| 7541 | Ops.push_back(N.getOperand(1)); | |||
| 7542 | return true; | |||
| 7543 | } | |||
| 7544 | return false; | |||
| 7545 | } | |||
| 7546 | case ISD::AND: | |||
| 7547 | case X86ISD::ANDNP: { | |||
| 7548 | // Attempt to decode as a per-byte mask. | |||
| 7549 | APInt UndefElts; | |||
| 7550 | SmallVector<APInt, 32> EltBits; | |||
| 7551 | SDValue N0 = N.getOperand(0); | |||
| 7552 | SDValue N1 = N.getOperand(1); | |||
| 7553 | bool IsAndN = (X86ISD::ANDNP == Opcode); | |||
| 7554 | uint64_t ZeroMask = IsAndN ? 255 : 0; | |||
| 7555 | if (!getTargetConstantBitsFromNode(IsAndN ? N0 : N1, 8, UndefElts, EltBits)) | |||
| 7556 | return false; | |||
| 7557 | for (int i = 0, e = (int)EltBits.size(); i != e; ++i) { | |||
| 7558 | if (UndefElts[i]) { | |||
| 7559 | Mask.push_back(SM_SentinelUndef); | |||
| 7560 | continue; | |||
| 7561 | } | |||
| 7562 | const APInt &ByteBits = EltBits[i]; | |||
| 7563 | if (ByteBits != 0 && ByteBits != 255) | |||
| 7564 | return false; | |||
| 7565 | Mask.push_back(ByteBits == ZeroMask ? SM_SentinelZero : i); | |||
| 7566 | } | |||
| 7567 | Ops.push_back(IsAndN ? N1 : N0); | |||
| 7568 | return true; | |||
| 7569 | } | |||
| 7570 | case ISD::OR: { | |||
| 7571 | // Handle OR(SHUFFLE,SHUFFLE) case where one source is zero and the other | |||
| 7572 | // is a valid shuffle index. | |||
| 7573 | SDValue N0 = peekThroughBitcasts(N.getOperand(0)); | |||
| 7574 | SDValue N1 = peekThroughBitcasts(N.getOperand(1)); | |||
| 7575 | if (!N0.getValueType().isVector() || !N1.getValueType().isVector()) | |||
| 7576 | return false; | |||
| 7577 | SmallVector<int, 64> SrcMask0, SrcMask1; | |||
| 7578 | SmallVector<SDValue, 2> SrcInputs0, SrcInputs1; | |||
| 7579 | if (!getTargetShuffleInputs(N0, SrcInputs0, SrcMask0, DAG, Depth + 1, | |||
| 7580 | true) || | |||
| 7581 | !getTargetShuffleInputs(N1, SrcInputs1, SrcMask1, DAG, Depth + 1, | |||
| 7582 | true)) | |||
| 7583 | return false; | |||
| 7584 | ||||
| 7585 | size_t MaskSize = std::max(SrcMask0.size(), SrcMask1.size()); | |||
| 7586 | SmallVector<int, 64> Mask0, Mask1; | |||
| 7587 | narrowShuffleMaskElts(MaskSize / SrcMask0.size(), SrcMask0, Mask0); | |||
| 7588 | narrowShuffleMaskElts(MaskSize / SrcMask1.size(), SrcMask1, Mask1); | |||
| 7589 | for (int i = 0; i != (int)MaskSize; ++i) { | |||
| 7590 | // NOTE: Don't handle SM_SentinelUndef, as we can end up in infinite | |||
| 7591 | // loops converting between OR and BLEND shuffles due to | |||
| 7592 | // canWidenShuffleElements merging away undef elements, meaning we | |||
| 7593 | // fail to recognise the OR as the undef element isn't known zero. | |||
| 7594 | if (Mask0[i] == SM_SentinelZero && Mask1[i] == SM_SentinelZero) | |||
| 7595 | Mask.push_back(SM_SentinelZero); | |||
| 7596 | else if (Mask1[i] == SM_SentinelZero) | |||
| 7597 | Mask.push_back(i); | |||
| 7598 | else if (Mask0[i] == SM_SentinelZero) | |||
| 7599 | Mask.push_back(i + MaskSize); | |||
| 7600 | else | |||
| 7601 | return false; | |||
| 7602 | } | |||
| 7603 | Ops.push_back(N0); | |||
| 7604 | Ops.push_back(N1); | |||
| 7605 | return true; | |||
| 7606 | } | |||
| 7607 | case ISD::INSERT_SUBVECTOR: { | |||
| 7608 | SDValue Src = N.getOperand(0); | |||
| 7609 | SDValue Sub = N.getOperand(1); | |||
| 7610 | EVT SubVT = Sub.getValueType(); | |||
| 7611 | unsigned NumSubElts = SubVT.getVectorNumElements(); | |||
| 7612 | if (!N->isOnlyUserOf(Sub.getNode())) | |||
| 7613 | return false; | |||
| 7614 | uint64_t InsertIdx = N.getConstantOperandVal(2); | |||
| 7615 | // Handle INSERT_SUBVECTOR(SRC0, EXTRACT_SUBVECTOR(SRC1)). | |||
| 7616 | if (Sub.getOpcode() == ISD::EXTRACT_SUBVECTOR && | |||
| 7617 | Sub.getOperand(0).getValueType() == VT) { | |||
| 7618 | uint64_t ExtractIdx = Sub.getConstantOperandVal(1); | |||
| 7619 | for (int i = 0; i != (int)NumElts; ++i) | |||
| 7620 | Mask.push_back(i); | |||
| 7621 | for (int i = 0; i != (int)NumSubElts; ++i) | |||
| 7622 | Mask[InsertIdx + i] = NumElts + ExtractIdx + i; | |||
| 7623 | Ops.push_back(Src); | |||
| 7624 | Ops.push_back(Sub.getOperand(0)); | |||
| 7625 | return true; | |||
| 7626 | } | |||
| 7627 | // Handle INSERT_SUBVECTOR(SRC0, SHUFFLE(SRC1)). | |||
| 7628 | SmallVector<int, 64> SubMask; | |||
| 7629 | SmallVector<SDValue, 2> SubInputs; | |||
| 7630 | if (!getTargetShuffleInputs(peekThroughOneUseBitcasts(Sub), SubInputs, | |||
| 7631 | SubMask, DAG, Depth + 1, ResolveKnownElts)) | |||
| 7632 | return false; | |||
| 7633 | ||||
| 7634 | // Subvector shuffle inputs must not be larger than the subvector. | |||
| 7635 | if (llvm::any_of(SubInputs, [SubVT](SDValue SubInput) { | |||
| 7636 | return SubVT.getFixedSizeInBits() < | |||
| 7637 | SubInput.getValueSizeInBits().getFixedSize(); | |||
| 7638 | })) | |||
| 7639 | return false; | |||
| 7640 | ||||
| 7641 | if (SubMask.size() != NumSubElts) { | |||
| 7642 | assert(((SubMask.size() % NumSubElts) == 0 ||((void)0) | |||
| 7643 | (NumSubElts % SubMask.size()) == 0) && "Illegal submask scale")((void)0); | |||
| 7644 | if ((NumSubElts % SubMask.size()) == 0) { | |||
| 7645 | int Scale = NumSubElts / SubMask.size(); | |||
| 7646 | SmallVector<int,64> ScaledSubMask; | |||
| 7647 | narrowShuffleMaskElts(Scale, SubMask, ScaledSubMask); | |||
| 7648 | SubMask = ScaledSubMask; | |||
| 7649 | } else { | |||
| 7650 | int Scale = SubMask.size() / NumSubElts; | |||
| 7651 | NumSubElts = SubMask.size(); | |||
| 7652 | NumElts *= Scale; | |||
| 7653 | InsertIdx *= Scale; | |||
| 7654 | } | |||
| 7655 | } | |||
| 7656 | Ops.push_back(Src); | |||
| 7657 | Ops.append(SubInputs.begin(), SubInputs.end()); | |||
| 7658 | if (ISD::isBuildVectorAllZeros(Src.getNode())) | |||
| 7659 | Mask.append(NumElts, SM_SentinelZero); | |||
| 7660 | else | |||
| 7661 | for (int i = 0; i != (int)NumElts; ++i) | |||
| 7662 | Mask.push_back(i); | |||
| 7663 | for (int i = 0; i != (int)NumSubElts; ++i) { | |||
| 7664 | int M = SubMask[i]; | |||
| 7665 | if (0 <= M) { | |||
| 7666 | int InputIdx = M / NumSubElts; | |||
| 7667 | M = (NumElts * (1 + InputIdx)) + (M % NumSubElts); | |||
| 7668 | } | |||
| 7669 | Mask[i + InsertIdx] = M; | |||
| 7670 | } | |||
| 7671 | return true; | |||
| 7672 | } | |||
| 7673 | case X86ISD::PINSRB: | |||
| 7674 | case X86ISD::PINSRW: | |||
| 7675 | case ISD::SCALAR_TO_VECTOR: | |||
| 7676 | case ISD::INSERT_VECTOR_ELT: { | |||
| 7677 | // Match against a insert_vector_elt/scalar_to_vector of an extract from a | |||
| 7678 | // vector, for matching src/dst vector types. | |||
| 7679 | SDValue Scl = N.getOperand(Opcode == ISD::SCALAR_TO_VECTOR ? 0 : 1); | |||
| 7680 | ||||
| 7681 | unsigned DstIdx = 0; | |||
| 7682 | if (Opcode != ISD::SCALAR_TO_VECTOR) { | |||
| 7683 | // Check we have an in-range constant insertion index. | |||
| 7684 | if (!isa<ConstantSDNode>(N.getOperand(2)) || | |||
| 7685 | N.getConstantOperandAPInt(2).uge(NumElts)) | |||
| 7686 | return false; | |||
| 7687 | DstIdx = N.getConstantOperandVal(2); | |||
| 7688 | ||||
| 7689 | // Attempt to recognise an INSERT*(VEC, 0, DstIdx) shuffle pattern. | |||
| 7690 | if (X86::isZeroNode(Scl)) { | |||
| 7691 | Ops.push_back(N.getOperand(0)); | |||
| 7692 | for (unsigned i = 0; i != NumElts; ++i) | |||
| 7693 | Mask.push_back(i == DstIdx ? SM_SentinelZero : (int)i); | |||
| 7694 | return true; | |||
| 7695 | } | |||
| 7696 | } | |||
| 7697 | ||||
| 7698 | // Peek through trunc/aext/zext. | |||
| 7699 | // TODO: aext shouldn't require SM_SentinelZero padding. | |||
| 7700 | // TODO: handle shift of scalars. | |||
| 7701 | unsigned MinBitsPerElt = Scl.getScalarValueSizeInBits(); | |||
| 7702 | while (Scl.getOpcode() == ISD::TRUNCATE || | |||
| 7703 | Scl.getOpcode() == ISD::ANY_EXTEND || | |||
| 7704 | Scl.getOpcode() == ISD::ZERO_EXTEND) { | |||
| 7705 | Scl = Scl.getOperand(0); | |||
| 7706 | MinBitsPerElt = | |||
| 7707 | std::min<unsigned>(MinBitsPerElt, Scl.getScalarValueSizeInBits()); | |||
| 7708 | } | |||
| 7709 | if ((MinBitsPerElt % 8) != 0) | |||
| 7710 | return false; | |||
| 7711 | ||||
| 7712 | // Attempt to find the source vector the scalar was extracted from. | |||
| 7713 | SDValue SrcExtract; | |||
| 7714 | if ((Scl.getOpcode() == ISD::EXTRACT_VECTOR_ELT || | |||
| 7715 | Scl.getOpcode() == X86ISD::PEXTRW || | |||
| 7716 | Scl.getOpcode() == X86ISD::PEXTRB) && | |||
| 7717 | Scl.getOperand(0).getValueSizeInBits() == NumSizeInBits) { | |||
| 7718 | SrcExtract = Scl; | |||
| 7719 | } | |||
| 7720 | if (!SrcExtract || !isa<ConstantSDNode>(SrcExtract.getOperand(1))) | |||
| 7721 | return false; | |||
| 7722 | ||||
| 7723 | SDValue SrcVec = SrcExtract.getOperand(0); | |||
| 7724 | EVT SrcVT = SrcVec.getValueType(); | |||
| 7725 | if (!SrcVT.getScalarType().isByteSized()) | |||
| 7726 | return false; | |||
| 7727 | unsigned SrcIdx = SrcExtract.getConstantOperandVal(1); | |||
| 7728 | unsigned SrcByte = SrcIdx * (SrcVT.getScalarSizeInBits() / 8); | |||
| 7729 | unsigned DstByte = DstIdx * NumBytesPerElt; | |||
| 7730 | MinBitsPerElt = | |||
| 7731 | std::min<unsigned>(MinBitsPerElt, SrcVT.getScalarSizeInBits()); | |||
| 7732 | ||||
| 7733 | // Create 'identity' byte level shuffle mask and then add inserted bytes. | |||
| 7734 | if (Opcode == ISD::SCALAR_TO_VECTOR) { | |||
| 7735 | Ops.push_back(SrcVec); | |||
| 7736 | Mask.append(NumSizeInBytes, SM_SentinelUndef); | |||
| 7737 | } else { | |||
| 7738 | Ops.push_back(SrcVec); | |||
| 7739 | Ops.push_back(N.getOperand(0)); | |||
| 7740 | for (int i = 0; i != (int)NumSizeInBytes; ++i) | |||
| 7741 | Mask.push_back(NumSizeInBytes + i); | |||
| 7742 | } | |||
| 7743 | ||||
| 7744 | unsigned MinBytesPerElts = MinBitsPerElt / 8; | |||
| 7745 | MinBytesPerElts = std::min(MinBytesPerElts, NumBytesPerElt); | |||
| 7746 | for (unsigned i = 0; i != MinBytesPerElts; ++i) | |||
| 7747 | Mask[DstByte + i] = SrcByte + i; | |||
| 7748 | for (unsigned i = MinBytesPerElts; i < NumBytesPerElt; ++i) | |||
| 7749 | Mask[DstByte + i] = SM_SentinelZero; | |||
| 7750 | return true; | |||
| 7751 | } | |||
| 7752 | case X86ISD::PACKSS: | |||
| 7753 | case X86ISD::PACKUS: { | |||
| 7754 | SDValue N0 = N.getOperand(0); | |||
| 7755 | SDValue N1 = N.getOperand(1); | |||
| 7756 | assert(N0.getValueType().getVectorNumElements() == (NumElts / 2) &&((void)0) | |||
| 7757 | N1.getValueType().getVectorNumElements() == (NumElts / 2) &&((void)0) | |||
| 7758 | "Unexpected input value type")((void)0); | |||
| 7759 | ||||
| 7760 | APInt EltsLHS, EltsRHS; | |||
| 7761 | getPackDemandedElts(VT, DemandedElts, EltsLHS, EltsRHS); | |||
| 7762 | ||||
| 7763 | // If we know input saturation won't happen (or we don't care for particular | |||
| 7764 | // lanes), we can treat this as a truncation shuffle. | |||
| 7765 | bool Offset0 = false, Offset1 = false; | |||
| 7766 | if (Opcode == X86ISD::PACKSS) { | |||
| 7767 | if ((!(N0.isUndef() || EltsLHS.isNullValue()) && | |||
| 7768 | DAG.ComputeNumSignBits(N0, EltsLHS, Depth + 1) <= NumBitsPerElt) || | |||
| 7769 | (!(N1.isUndef() || EltsRHS.isNullValue()) && | |||
| 7770 | DAG.ComputeNumSignBits(N1, EltsRHS, Depth + 1) <= NumBitsPerElt)) | |||
| 7771 | return false; | |||
| 7772 | // We can't easily fold ASHR into a shuffle, but if it was feeding a | |||
| 7773 | // PACKSS then it was likely being used for sign-extension for a | |||
| 7774 | // truncation, so just peek through and adjust the mask accordingly. | |||
| 7775 | if (N0.getOpcode() == X86ISD::VSRAI && N->isOnlyUserOf(N0.getNode()) && | |||
| 7776 | N0.getConstantOperandAPInt(1) == NumBitsPerElt) { | |||
| 7777 | Offset0 = true; | |||
| 7778 | N0 = N0.getOperand(0); | |||
| 7779 | } | |||
| 7780 | if (N1.getOpcode() == X86ISD::VSRAI && N->isOnlyUserOf(N1.getNode()) && | |||
| 7781 | N1.getConstantOperandAPInt(1) == NumBitsPerElt) { | |||
| 7782 | Offset1 = true; | |||
| 7783 | N1 = N1.getOperand(0); | |||
| 7784 | } | |||
| 7785 | } else { | |||
| 7786 | APInt ZeroMask = APInt::getHighBitsSet(2 * NumBitsPerElt, NumBitsPerElt); | |||
| 7787 | if ((!(N0.isUndef() || EltsLHS.isNullValue()) && | |||
| 7788 | !DAG.MaskedValueIsZero(N0, ZeroMask, EltsLHS, Depth + 1)) || | |||
| 7789 | (!(N1.isUndef() || EltsRHS.isNullValue()) && | |||
| 7790 | !DAG.MaskedValueIsZero(N1, ZeroMask, EltsRHS, Depth + 1))) | |||
| 7791 | return false; | |||
| 7792 | } | |||
| 7793 | ||||
| 7794 | bool IsUnary = (N0 == N1); | |||
| 7795 | ||||
| 7796 | Ops.push_back(N0); | |||
| 7797 | if (!IsUnary) | |||
| 7798 | Ops.push_back(N1); | |||
| 7799 | ||||
| 7800 | createPackShuffleMask(VT, Mask, IsUnary); | |||
| 7801 | ||||
| 7802 | if (Offset0 || Offset1) { | |||
| 7803 | for (int &M : Mask) | |||
| 7804 | if ((Offset0 && isInRange(M, 0, NumElts)) || | |||
| 7805 | (Offset1 && isInRange(M, NumElts, 2 * NumElts))) | |||
| 7806 | ++M; | |||
| 7807 | } | |||
| 7808 | return true; | |||
| 7809 | } | |||
| 7810 | case X86ISD::VTRUNC: { | |||
| 7811 | SDValue Src = N.getOperand(0); | |||
| 7812 | EVT SrcVT = Src.getValueType(); | |||
| 7813 | // Truncated source must be a simple vector. | |||
| 7814 | if (!SrcVT.isSimple() || (SrcVT.getSizeInBits() % 128) != 0 || | |||
| 7815 | (SrcVT.getScalarSizeInBits() % 8) != 0) | |||
| 7816 | return false; | |||
| 7817 | unsigned NumSrcElts = SrcVT.getVectorNumElements(); | |||
| 7818 | unsigned NumBitsPerSrcElt = SrcVT.getScalarSizeInBits(); | |||
| 7819 | unsigned Scale = NumBitsPerSrcElt / NumBitsPerElt; | |||
| 7820 | assert((NumBitsPerSrcElt % NumBitsPerElt) == 0 && "Illegal truncation")((void)0); | |||
| 7821 | for (unsigned i = 0; i != NumSrcElts; ++i) | |||
| 7822 | Mask.push_back(i * Scale); | |||
| 7823 | Mask.append(NumElts - NumSrcElts, SM_SentinelZero); | |||
| 7824 | Ops.push_back(Src); | |||
| 7825 | return true; | |||
| 7826 | } | |||
| 7827 | case X86ISD::VSHLI: | |||
| 7828 | case X86ISD::VSRLI: { | |||
| 7829 | uint64_t ShiftVal = N.getConstantOperandVal(1); | |||
| 7830 | // Out of range bit shifts are guaranteed to be zero. | |||
| 7831 | if (NumBitsPerElt <= ShiftVal) { | |||
| 7832 | Mask.append(NumElts, SM_SentinelZero); | |||
| 7833 | return true; | |||
| 7834 | } | |||
| 7835 | ||||
| 7836 | // We can only decode 'whole byte' bit shifts as shuffles. | |||
| 7837 | if ((ShiftVal % 8) != 0) | |||
| 7838 | break; | |||
| 7839 | ||||
| 7840 | uint64_t ByteShift = ShiftVal / 8; | |||
| 7841 | Ops.push_back(N.getOperand(0)); | |||
| 7842 | ||||
| 7843 | // Clear mask to all zeros and insert the shifted byte indices. | |||
| 7844 | Mask.append(NumSizeInBytes, SM_SentinelZero); | |||
| 7845 | ||||
| 7846 | if (X86ISD::VSHLI == Opcode) { | |||
| 7847 | for (unsigned i = 0; i != NumSizeInBytes; i += NumBytesPerElt) | |||
| 7848 | for (unsigned j = ByteShift; j != NumBytesPerElt; ++j) | |||
| 7849 | Mask[i + j] = i + j - ByteShift; | |||
| 7850 | } else { | |||
| 7851 | for (unsigned i = 0; i != NumSizeInBytes; i += NumBytesPerElt) | |||
| 7852 | for (unsigned j = ByteShift; j != NumBytesPerElt; ++j) | |||
| 7853 | Mask[i + j - ByteShift] = i + j; | |||
| 7854 | } | |||
| 7855 | return true; | |||
| 7856 | } | |||
| 7857 | case X86ISD::VROTLI: | |||
| 7858 | case X86ISD::VROTRI: { | |||
| 7859 | // We can only decode 'whole byte' bit rotates as shuffles. | |||
| 7860 | uint64_t RotateVal = N.getConstantOperandAPInt(1).urem(NumBitsPerElt); | |||
| 7861 | if ((RotateVal % 8) != 0) | |||
| 7862 | return false; | |||
| 7863 | Ops.push_back(N.getOperand(0)); | |||
| 7864 | int Offset = RotateVal / 8; | |||
| 7865 | Offset = (X86ISD::VROTLI == Opcode ? NumBytesPerElt - Offset : Offset); | |||
| 7866 | for (int i = 0; i != (int)NumElts; ++i) { | |||
| 7867 | int BaseIdx = i * NumBytesPerElt; | |||
| 7868 | for (int j = 0; j != (int)NumBytesPerElt; ++j) { | |||
| 7869 | Mask.push_back(BaseIdx + ((Offset + j) % NumBytesPerElt)); | |||
| 7870 | } | |||
| 7871 | } | |||
| 7872 | return true; | |||
| 7873 | } | |||
| 7874 | case X86ISD::VBROADCAST: { | |||
| 7875 | SDValue Src = N.getOperand(0); | |||
| 7876 | if (!Src.getSimpleValueType().isVector()) { | |||
| 7877 | if (Src.getOpcode() != ISD::EXTRACT_VECTOR_ELT || | |||
| 7878 | !isNullConstant(Src.getOperand(1)) || | |||
| 7879 | Src.getOperand(0).getValueType().getScalarType() != | |||
| 7880 | VT.getScalarType()) | |||
| 7881 | return false; | |||
| 7882 | Src = Src.getOperand(0); | |||
| 7883 | } | |||
| 7884 | Ops.push_back(Src); | |||
| 7885 | Mask.append(NumElts, 0); | |||
| 7886 | return true; | |||
| 7887 | } | |||
| 7888 | case ISD::ZERO_EXTEND: | |||
| 7889 | case ISD::ANY_EXTEND: | |||
| 7890 | case ISD::ZERO_EXTEND_VECTOR_INREG: | |||
| 7891 | case ISD::ANY_EXTEND_VECTOR_INREG: { | |||
| 7892 | SDValue Src = N.getOperand(0); | |||
| 7893 | EVT SrcVT = Src.getValueType(); | |||
| 7894 | ||||
| 7895 | // Extended source must be a simple vector. | |||
| 7896 | if (!SrcVT.isSimple() || (SrcVT.getSizeInBits() % 128) != 0 || | |||
| 7897 | (SrcVT.getScalarSizeInBits() % 8) != 0) | |||
| 7898 | return false; | |||
| 7899 | ||||
| 7900 | bool IsAnyExtend = | |||
| 7901 | (ISD::ANY_EXTEND == Opcode || ISD::ANY_EXTEND_VECTOR_INREG == Opcode); | |||
| 7902 | DecodeZeroExtendMask(SrcVT.getScalarSizeInBits(), NumBitsPerElt, NumElts, | |||
| 7903 | IsAnyExtend, Mask); | |||
| 7904 | Ops.push_back(Src); | |||
| 7905 | return true; | |||
| 7906 | } | |||
| 7907 | } | |||
| 7908 | ||||
| 7909 | return false; | |||
| 7910 | } | |||
| 7911 | ||||
| 7912 | /// Removes unused/repeated shuffle source inputs and adjusts the shuffle mask. | |||
| 7913 | static void resolveTargetShuffleInputsAndMask(SmallVectorImpl<SDValue> &Inputs, | |||
| 7914 | SmallVectorImpl<int> &Mask) { | |||
| 7915 | int MaskWidth = Mask.size(); | |||
| 7916 | SmallVector<SDValue, 16> UsedInputs; | |||
| 7917 | for (int i = 0, e = Inputs.size(); i < e; ++i) { | |||
| 7918 | int lo = UsedInputs.size() * MaskWidth; | |||
| 7919 | int hi = lo + MaskWidth; | |||
| 7920 | ||||
| 7921 | // Strip UNDEF input usage. | |||
| 7922 | if (Inputs[i].isUndef()) | |||
| 7923 | for (int &M : Mask) | |||
| 7924 | if ((lo <= M) && (M < hi)) | |||
| 7925 | M = SM_SentinelUndef; | |||
| 7926 | ||||
| 7927 | // Check for unused inputs. | |||
| 7928 | if (none_of(Mask, [lo, hi](int i) { return (lo <= i) && (i < hi); })) { | |||
| 7929 | for (int &M : Mask) | |||
| 7930 | if (lo <= M) | |||
| 7931 | M -= MaskWidth; | |||
| 7932 | continue; | |||
| 7933 | } | |||
| 7934 | ||||
| 7935 | // Check for repeated inputs. | |||
| 7936 | bool IsRepeat = false; | |||
| 7937 | for (int j = 0, ue = UsedInputs.size(); j != ue; ++j) { | |||
| 7938 | if (UsedInputs[j] != Inputs[i]) | |||
| 7939 | continue; | |||
| 7940 | for (int &M : Mask) | |||
| 7941 | if (lo <= M) | |||
| 7942 | M = (M < hi) ? ((M - lo) + (j * MaskWidth)) : (M - MaskWidth); | |||
| 7943 | IsRepeat = true; | |||
| 7944 | break; | |||
| 7945 | } | |||
| 7946 | if (IsRepeat) | |||
| 7947 | continue; | |||
| 7948 | ||||
| 7949 | UsedInputs.push_back(Inputs[i]); | |||
| 7950 | } | |||
| 7951 | Inputs = UsedInputs; | |||
| 7952 | } | |||
| 7953 | ||||
| 7954 | /// Calls getTargetShuffleAndZeroables to resolve a target shuffle mask's inputs | |||
| 7955 | /// and then sets the SM_SentinelUndef and SM_SentinelZero values. | |||
| 7956 | /// Returns true if the target shuffle mask was decoded. | |||
| 7957 | static bool getTargetShuffleInputs(SDValue Op, const APInt &DemandedElts, | |||
| 7958 | SmallVectorImpl<SDValue> &Inputs, | |||
| 7959 | SmallVectorImpl<int> &Mask, | |||
| 7960 | APInt &KnownUndef, APInt &KnownZero, | |||
| 7961 | const SelectionDAG &DAG, unsigned Depth, | |||
| 7962 | bool ResolveKnownElts) { | |||
| 7963 | EVT VT = Op.getValueType(); | |||
| 7964 | if (!VT.isSimple() || !VT.isVector()) | |||
| 7965 | return false; | |||
| 7966 | ||||
| 7967 | if (getTargetShuffleAndZeroables(Op, Mask, Inputs, KnownUndef, KnownZero)) { | |||
| 7968 | if (ResolveKnownElts) | |||
| 7969 | resolveTargetShuffleFromZeroables(Mask, KnownUndef, KnownZero); | |||
| 7970 | return true; | |||
| 7971 | } | |||
| 7972 | if (getFauxShuffleMask(Op, DemandedElts, Mask, Inputs, DAG, Depth, | |||
| 7973 | ResolveKnownElts)) { | |||
| 7974 | resolveZeroablesFromTargetShuffle(Mask, KnownUndef, KnownZero); | |||
| 7975 | return true; | |||
| 7976 | } | |||
| 7977 | return false; | |||
| 7978 | } | |||
| 7979 | ||||
| 7980 | static bool getTargetShuffleInputs(SDValue Op, SmallVectorImpl<SDValue> &Inputs, | |||
| 7981 | SmallVectorImpl<int> &Mask, | |||
| 7982 | const SelectionDAG &DAG, unsigned Depth = 0, | |||
| 7983 | bool ResolveKnownElts = true) { | |||
| 7984 | EVT VT = Op.getValueType(); | |||
| 7985 | if (!VT.isSimple() || !VT.isVector()) | |||
| 7986 | return false; | |||
| 7987 | ||||
| 7988 | APInt KnownUndef, KnownZero; | |||
| 7989 | unsigned NumElts = Op.getValueType().getVectorNumElements(); | |||
| 7990 | APInt DemandedElts = APInt::getAllOnesValue(NumElts); | |||
| 7991 | return getTargetShuffleInputs(Op, DemandedElts, Inputs, Mask, KnownUndef, | |||
| 7992 | KnownZero, DAG, Depth, ResolveKnownElts); | |||
| 7993 | } | |||
| 7994 | ||||
| 7995 | // Attempt to create a scalar/subvector broadcast from the base MemSDNode. | |||
| 7996 | static SDValue getBROADCAST_LOAD(unsigned Opcode, const SDLoc &DL, EVT VT, | |||
| 7997 | EVT MemVT, MemSDNode *Mem, unsigned Offset, | |||
| 7998 | SelectionDAG &DAG) { | |||
| 7999 | assert((Opcode == X86ISD::VBROADCAST_LOAD ||((void)0) | |||
| 8000 | Opcode == X86ISD::SUBV_BROADCAST_LOAD) &&((void)0) | |||
| 8001 | "Unknown broadcast load type")((void)0); | |||
| 8002 | ||||
| 8003 | // Ensure this is a simple (non-atomic, non-voltile), temporal read memop. | |||
| 8004 | if (!Mem || !Mem->readMem() || !Mem->isSimple() || Mem->isNonTemporal()) | |||
| 8005 | return SDValue(); | |||
| 8006 | ||||
| 8007 | SDValue Ptr = | |||
| 8008 | DAG.getMemBasePlusOffset(Mem->getBasePtr(), TypeSize::Fixed(Offset), DL); | |||
| 8009 | SDVTList Tys = DAG.getVTList(VT, MVT::Other); | |||
| 8010 | SDValue Ops[] = {Mem->getChain(), Ptr}; | |||
| 8011 | SDValue BcstLd = DAG.getMemIntrinsicNode( | |||
| 8012 | Opcode, DL, Tys, Ops, MemVT, | |||
| 8013 | DAG.getMachineFunction().getMachineMemOperand( | |||
| 8014 | Mem->getMemOperand(), Offset, MemVT.getStoreSize())); | |||
| 8015 | DAG.makeEquivalentMemoryOrdering(SDValue(Mem, 1), BcstLd.getValue(1)); | |||
| 8016 | return BcstLd; | |||
| 8017 | } | |||
| 8018 | ||||
| 8019 | /// Returns the scalar element that will make up the i'th | |||
| 8020 | /// element of the result of the vector shuffle. | |||
| 8021 | static SDValue getShuffleScalarElt(SDValue Op, unsigned Index, | |||
| 8022 | SelectionDAG &DAG, unsigned Depth) { | |||
| 8023 | if (Depth >= SelectionDAG::MaxRecursionDepth) | |||
| 8024 | return SDValue(); // Limit search depth. | |||
| 8025 | ||||
| 8026 | EVT VT = Op.getValueType(); | |||
| 8027 | unsigned Opcode = Op.getOpcode(); | |||
| 8028 | unsigned NumElems = VT.getVectorNumElements(); | |||
| 8029 | ||||
| 8030 | // Recurse into ISD::VECTOR_SHUFFLE node to find scalars. | |||
| 8031 | if (auto *SV = dyn_cast<ShuffleVectorSDNode>(Op)) { | |||
| 8032 | int Elt = SV->getMaskElt(Index); | |||
| 8033 | ||||
| 8034 | if (Elt < 0) | |||
| 8035 | return DAG.getUNDEF(VT.getVectorElementType()); | |||
| 8036 | ||||
| 8037 | SDValue Src = (Elt < (int)NumElems) ? SV->getOperand(0) : SV->getOperand(1); | |||
| 8038 | return getShuffleScalarElt(Src, Elt % NumElems, DAG, Depth + 1); | |||
| 8039 | } | |||
| 8040 | ||||
| 8041 | // Recurse into target specific vector shuffles to find scalars. | |||
| 8042 | if (isTargetShuffle(Opcode)) { | |||
| 8043 | MVT ShufVT = VT.getSimpleVT(); | |||
| 8044 | MVT ShufSVT = ShufVT.getVectorElementType(); | |||
| 8045 | int NumElems = (int)ShufVT.getVectorNumElements(); | |||
| 8046 | SmallVector<int, 16> ShuffleMask; | |||
| 8047 | SmallVector<SDValue, 16> ShuffleOps; | |||
| 8048 | if (!getTargetShuffleMask(Op.getNode(), ShufVT, true, ShuffleOps, | |||
| 8049 | ShuffleMask)) | |||
| 8050 | return SDValue(); | |||
| 8051 | ||||
| 8052 | int Elt = ShuffleMask[Index]; | |||
| 8053 | if (Elt == SM_SentinelZero) | |||
| 8054 | return ShufSVT.isInteger() ? DAG.getConstant(0, SDLoc(Op), ShufSVT) | |||
| 8055 | : DAG.getConstantFP(+0.0, SDLoc(Op), ShufSVT); | |||
| 8056 | if (Elt == SM_SentinelUndef) | |||
| 8057 | return DAG.getUNDEF(ShufSVT); | |||
| 8058 | ||||
| 8059 | assert(0 <= Elt && Elt < (2 * NumElems) && "Shuffle index out of range")((void)0); | |||
| 8060 | SDValue Src = (Elt < NumElems) ? ShuffleOps[0] : ShuffleOps[1]; | |||
| 8061 | return getShuffleScalarElt(Src, Elt % NumElems, DAG, Depth + 1); | |||
| 8062 | } | |||
| 8063 | ||||
| 8064 | // Recurse into insert_subvector base/sub vector to find scalars. | |||
| 8065 | if (Opcode == ISD::INSERT_SUBVECTOR) { | |||
| 8066 | SDValue Vec = Op.getOperand(0); | |||
| 8067 | SDValue Sub = Op.getOperand(1); | |||
| 8068 | uint64_t SubIdx = Op.getConstantOperandVal(2); | |||
| 8069 | unsigned NumSubElts = Sub.getValueType().getVectorNumElements(); | |||
| 8070 | ||||
| 8071 | if (SubIdx <= Index && Index < (SubIdx + NumSubElts)) | |||
| 8072 | return getShuffleScalarElt(Sub, Index - SubIdx, DAG, Depth + 1); | |||
| 8073 | return getShuffleScalarElt(Vec, Index, DAG, Depth + 1); | |||
| 8074 | } | |||
| 8075 | ||||
| 8076 | // Recurse into concat_vectors sub vector to find scalars. | |||
| 8077 | if (Opcode == ISD::CONCAT_VECTORS) { | |||
| 8078 | EVT SubVT = Op.getOperand(0).getValueType(); | |||
| 8079 | unsigned NumSubElts = SubVT.getVectorNumElements(); | |||
| 8080 | uint64_t SubIdx = Index / NumSubElts; | |||
| 8081 | uint64_t SubElt = Index % NumSubElts; | |||
| 8082 | return getShuffleScalarElt(Op.getOperand(SubIdx), SubElt, DAG, Depth + 1); | |||
| 8083 | } | |||
| 8084 | ||||
| 8085 | // Recurse into extract_subvector src vector to find scalars. | |||
| 8086 | if (Opcode == ISD::EXTRACT_SUBVECTOR) { | |||
| 8087 | SDValue Src = Op.getOperand(0); | |||
| 8088 | uint64_t SrcIdx = Op.getConstantOperandVal(1); | |||
| 8089 | return getShuffleScalarElt(Src, Index + SrcIdx, DAG, Depth + 1); | |||
| 8090 | } | |||
| 8091 | ||||
| 8092 | // We only peek through bitcasts of the same vector width. | |||
| 8093 | if (Opcode == ISD::BITCAST) { | |||
| 8094 | SDValue Src = Op.getOperand(0); | |||
| 8095 | EVT SrcVT = Src.getValueType(); | |||
| 8096 | if (SrcVT.isVector() && SrcVT.getVectorNumElements() == NumElems) | |||
| 8097 | return getShuffleScalarElt(Src, Index, DAG, Depth + 1); | |||
| 8098 | return SDValue(); | |||
| 8099 | } | |||
| 8100 | ||||
| 8101 | // Actual nodes that may contain scalar elements | |||
| 8102 | ||||
| 8103 | // For insert_vector_elt - either return the index matching scalar or recurse | |||
| 8104 | // into the base vector. | |||
| 8105 | if (Opcode == ISD::INSERT_VECTOR_ELT && | |||
| 8106 | isa<ConstantSDNode>(Op.getOperand(2))) { | |||
| 8107 | if (Op.getConstantOperandAPInt(2) == Index) | |||
| 8108 | return Op.getOperand(1); | |||
| 8109 | return getShuffleScalarElt(Op.getOperand(0), Index, DAG, Depth + 1); | |||
| 8110 | } | |||
| 8111 | ||||
| 8112 | if (Opcode == ISD::SCALAR_TO_VECTOR) | |||
| 8113 | return (Index == 0) ? Op.getOperand(0) | |||
| 8114 | : DAG.getUNDEF(VT.getVectorElementType()); | |||
| 8115 | ||||
| 8116 | if (Opcode == ISD::BUILD_VECTOR) | |||
| 8117 | return Op.getOperand(Index); | |||
| 8118 | ||||
| 8119 | return SDValue(); | |||
| 8120 | } | |||
| 8121 | ||||
| 8122 | // Use PINSRB/PINSRW/PINSRD to create a build vector. | |||
| 8123 | static SDValue LowerBuildVectorAsInsert(SDValue Op, const APInt &NonZeroMask, | |||
| 8124 | unsigned NumNonZero, unsigned NumZero, | |||
| 8125 | SelectionDAG &DAG, | |||
| 8126 | const X86Subtarget &Subtarget) { | |||
| 8127 | MVT VT = Op.getSimpleValueType(); | |||
| 8128 | unsigned NumElts = VT.getVectorNumElements(); | |||
| 8129 | assert(((VT == MVT::v8i16 && Subtarget.hasSSE2()) ||((void)0) | |||
| 8130 | ((VT == MVT::v16i8 || VT == MVT::v4i32) && Subtarget.hasSSE41())) &&((void)0) | |||
| 8131 | "Illegal vector insertion")((void)0); | |||
| 8132 | ||||
| 8133 | SDLoc dl(Op); | |||
| 8134 | SDValue V; | |||
| 8135 | bool First = true; | |||
| 8136 | ||||
| 8137 | for (unsigned i = 0; i < NumElts; ++i) { | |||
| 8138 | bool IsNonZero = NonZeroMask[i]; | |||
| 8139 | if (!IsNonZero) | |||
| 8140 | continue; | |||
| 8141 | ||||
| 8142 | // If the build vector contains zeros or our first insertion is not the | |||
| 8143 | // first index then insert into zero vector to break any register | |||
| 8144 | // dependency else use SCALAR_TO_VECTOR. | |||
| 8145 | if (First) { | |||
| 8146 | First = false; | |||
| 8147 | if (NumZero || 0 != i) | |||
| 8148 | V = getZeroVector(VT, Subtarget, DAG, dl); | |||
| 8149 | else { | |||
| 8150 | assert(0 == i && "Expected insertion into zero-index")((void)0); | |||
| 8151 | V = DAG.getAnyExtOrTrunc(Op.getOperand(i), dl, MVT::i32); | |||
| 8152 | V = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, V); | |||
| 8153 | V = DAG.getBitcast(VT, V); | |||
| 8154 | continue; | |||
| 8155 | } | |||
| 8156 | } | |||
| 8157 | V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, V, Op.getOperand(i), | |||
| 8158 | DAG.getIntPtrConstant(i, dl)); | |||
| 8159 | } | |||
| 8160 | ||||
| 8161 | return V; | |||
| 8162 | } | |||
| 8163 | ||||
| 8164 | /// Custom lower build_vector of v16i8. | |||
| 8165 | static SDValue LowerBuildVectorv16i8(SDValue Op, const APInt &NonZeroMask, | |||
| 8166 | unsigned NumNonZero, unsigned NumZero, | |||
| 8167 | SelectionDAG &DAG, | |||
| 8168 | const X86Subtarget &Subtarget) { | |||
| 8169 | if (NumNonZero > 8 && !Subtarget.hasSSE41()) | |||
| 8170 | return SDValue(); | |||
| 8171 | ||||
| 8172 | // SSE4.1 - use PINSRB to insert each byte directly. | |||
| 8173 | if (Subtarget.hasSSE41()) | |||
| 8174 | return LowerBuildVectorAsInsert(Op, NonZeroMask, NumNonZero, NumZero, DAG, | |||
| 8175 | Subtarget); | |||
| 8176 | ||||
| 8177 | SDLoc dl(Op); | |||
| 8178 | SDValue V; | |||
| 8179 | ||||
| 8180 | // Pre-SSE4.1 - merge byte pairs and insert with PINSRW. | |||
| 8181 | for (unsigned i = 0; i < 16; i += 2) { | |||
| 8182 | bool ThisIsNonZero = NonZeroMask[i]; | |||
| 8183 | bool NextIsNonZero = NonZeroMask[i + 1]; | |||
| 8184 | if (!ThisIsNonZero && !NextIsNonZero) | |||
| 8185 | continue; | |||
| 8186 | ||||
| 8187 | // FIXME: Investigate combining the first 4 bytes as a i32 instead. | |||
| 8188 | SDValue Elt; | |||
| 8189 | if (ThisIsNonZero) { | |||
| 8190 | if (NumZero || NextIsNonZero) | |||
| 8191 | Elt = DAG.getZExtOrTrunc(Op.getOperand(i), dl, MVT::i32); | |||
| 8192 | else | |||
| 8193 | Elt = DAG.getAnyExtOrTrunc(Op.getOperand(i), dl, MVT::i32); | |||
| 8194 | } | |||
| 8195 | ||||
| 8196 | if (NextIsNonZero) { | |||
| 8197 | SDValue NextElt = Op.getOperand(i + 1); | |||
| 8198 | if (i == 0 && NumZero) | |||
| 8199 | NextElt = DAG.getZExtOrTrunc(NextElt, dl, MVT::i32); | |||
| 8200 | else | |||
| 8201 | NextElt = DAG.getAnyExtOrTrunc(NextElt, dl, MVT::i32); | |||
| 8202 | NextElt = DAG.getNode(ISD::SHL, dl, MVT::i32, NextElt, | |||
| 8203 | DAG.getConstant(8, dl, MVT::i8)); | |||
| 8204 | if (ThisIsNonZero) | |||
| 8205 | Elt = DAG.getNode(ISD::OR, dl, MVT::i32, NextElt, Elt); | |||
| 8206 | else | |||
| 8207 | Elt = NextElt; | |||
| 8208 | } | |||
| 8209 | ||||
| 8210 | // If our first insertion is not the first index or zeros are needed, then | |||
| 8211 | // insert into zero vector. Otherwise, use SCALAR_TO_VECTOR (leaves high | |||
| 8212 | // elements undefined). | |||
| 8213 | if (!V) { | |||
| 8214 | if (i != 0 || NumZero) | |||
| 8215 | V = getZeroVector(MVT::v8i16, Subtarget, DAG, dl); | |||
| 8216 | else { | |||
| 8217 | V = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, Elt); | |||
| 8218 | V = DAG.getBitcast(MVT::v8i16, V); | |||
| 8219 | continue; | |||
| 8220 | } | |||
| 8221 | } | |||
| 8222 | Elt = DAG.getNode(ISD::TRUNCATE, dl, MVT::i16, Elt); | |||
| 8223 | V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, Elt, | |||
| 8224 | DAG.getIntPtrConstant(i / 2, dl)); | |||
| 8225 | } | |||
| 8226 | ||||
| 8227 | return DAG.getBitcast(MVT::v16i8, V); | |||
| 8228 | } | |||
| 8229 | ||||
| 8230 | /// Custom lower build_vector of v8i16. | |||
| 8231 | static SDValue LowerBuildVectorv8i16(SDValue Op, const APInt &NonZeroMask, | |||
| 8232 | unsigned NumNonZero, unsigned NumZero, | |||
| 8233 | SelectionDAG &DAG, | |||
| 8234 | const X86Subtarget &Subtarget) { | |||
| 8235 | if (NumNonZero > 4 && !Subtarget.hasSSE41()) | |||
| 8236 | return SDValue(); | |||
| 8237 | ||||
| 8238 | // Use PINSRW to insert each byte directly. | |||
| 8239 | return LowerBuildVectorAsInsert(Op, NonZeroMask, NumNonZero, NumZero, DAG, | |||
| 8240 | Subtarget); | |||
| 8241 | } | |||
| 8242 | ||||
| 8243 | /// Custom lower build_vector of v4i32 or v4f32. | |||
| 8244 | static SDValue LowerBuildVectorv4x32(SDValue Op, SelectionDAG &DAG, | |||
| 8245 | const X86Subtarget &Subtarget) { | |||
| 8246 | // If this is a splat of a pair of elements, use MOVDDUP (unless the target | |||
| 8247 | // has XOP; in that case defer lowering to potentially use VPERMIL2PS). | |||
| 8248 | // Because we're creating a less complicated build vector here, we may enable | |||
| 8249 | // further folding of the MOVDDUP via shuffle transforms. | |||
| 8250 | if (Subtarget.hasSSE3() && !Subtarget.hasXOP() && | |||
| 8251 | Op.getOperand(0) == Op.getOperand(2) && | |||
| 8252 | Op.getOperand(1) == Op.getOperand(3) && | |||
| 8253 | Op.getOperand(0) != Op.getOperand(1)) { | |||
| 8254 | SDLoc DL(Op); | |||
| 8255 | MVT VT = Op.getSimpleValueType(); | |||
| 8256 | MVT EltVT = VT.getVectorElementType(); | |||
| 8257 | // Create a new build vector with the first 2 elements followed by undef | |||
| 8258 | // padding, bitcast to v2f64, duplicate, and bitcast back. | |||
| 8259 | SDValue Ops[4] = { Op.getOperand(0), Op.getOperand(1), | |||
| 8260 | DAG.getUNDEF(EltVT), DAG.getUNDEF(EltVT) }; | |||
| 8261 | SDValue NewBV = DAG.getBitcast(MVT::v2f64, DAG.getBuildVector(VT, DL, Ops)); | |||
| 8262 | SDValue Dup = DAG.getNode(X86ISD::MOVDDUP, DL, MVT::v2f64, NewBV); | |||
| 8263 | return DAG.getBitcast(VT, Dup); | |||
| 8264 | } | |||
| 8265 | ||||
| 8266 | // Find all zeroable elements. | |||
| 8267 | std::bitset<4> Zeroable, Undefs; | |||
| 8268 | for (int i = 0; i < 4; ++i) { | |||
| 8269 | SDValue Elt = Op.getOperand(i); | |||
| 8270 | Undefs[i] = Elt.isUndef(); | |||
| 8271 | Zeroable[i] = (Elt.isUndef() || X86::isZeroNode(Elt)); | |||
| 8272 | } | |||
| 8273 | assert(Zeroable.size() - Zeroable.count() > 1 &&((void)0) | |||
| 8274 | "We expect at least two non-zero elements!")((void)0); | |||
| 8275 | ||||
| 8276 | // We only know how to deal with build_vector nodes where elements are either | |||
| 8277 | // zeroable or extract_vector_elt with constant index. | |||
| 8278 | SDValue FirstNonZero; | |||
| 8279 | unsigned FirstNonZeroIdx; | |||
| 8280 | for (unsigned i = 0; i < 4; ++i) { | |||
| 8281 | if (Zeroable[i]) | |||
| 8282 | continue; | |||
| 8283 | SDValue Elt = Op.getOperand(i); | |||
| 8284 | if (Elt.getOpcode() != ISD::EXTRACT_VECTOR_ELT || | |||
| 8285 | !isa<ConstantSDNode>(Elt.getOperand(1))) | |||
| 8286 | return SDValue(); | |||
| 8287 | // Make sure that this node is extracting from a 128-bit vector. | |||
| 8288 | MVT VT = Elt.getOperand(0).getSimpleValueType(); | |||
| 8289 | if (!VT.is128BitVector()) | |||
| 8290 | return SDValue(); | |||
| 8291 | if (!FirstNonZero.getNode()) { | |||
| 8292 | FirstNonZero = Elt; | |||
| 8293 | FirstNonZeroIdx = i; | |||
| 8294 | } | |||
| 8295 | } | |||
| 8296 | ||||
| 8297 | assert(FirstNonZero.getNode() && "Unexpected build vector of all zeros!")((void)0); | |||
| 8298 | SDValue V1 = FirstNonZero.getOperand(0); | |||
| 8299 | MVT VT = V1.getSimpleValueType(); | |||
| 8300 | ||||
| 8301 | // See if this build_vector can be lowered as a blend with zero. | |||
| 8302 | SDValue Elt; | |||
| 8303 | unsigned EltMaskIdx, EltIdx; | |||
| 8304 | int Mask[4]; | |||
| 8305 | for (EltIdx = 0; EltIdx < 4; ++EltIdx) { | |||
| 8306 | if (Zeroable[EltIdx]) { | |||
| 8307 | // The zero vector will be on the right hand side. | |||
| 8308 | Mask[EltIdx] = EltIdx+4; | |||
| 8309 | continue; | |||
| 8310 | } | |||
| 8311 | ||||
| 8312 | Elt = Op->getOperand(EltIdx); | |||
| 8313 | // By construction, Elt is a EXTRACT_VECTOR_ELT with constant index. | |||
| 8314 | EltMaskIdx = Elt.getConstantOperandVal(1); | |||
| 8315 | if (Elt.getOperand(0) != V1 || EltMaskIdx != EltIdx) | |||
| 8316 | break; | |||
| 8317 | Mask[EltIdx] = EltIdx; | |||
| 8318 | } | |||
| 8319 | ||||
| 8320 | if (EltIdx == 4) { | |||
| 8321 | // Let the shuffle legalizer deal with blend operations. | |||
| 8322 | SDValue VZeroOrUndef = (Zeroable == Undefs) | |||
| 8323 | ? DAG.getUNDEF(VT) | |||
| 8324 | : getZeroVector(VT, Subtarget, DAG, SDLoc(Op)); | |||
| 8325 | if (V1.getSimpleValueType() != VT) | |||
| 8326 | V1 = DAG.getBitcast(VT, V1); | |||
| 8327 | return DAG.getVectorShuffle(VT, SDLoc(V1), V1, VZeroOrUndef, Mask); | |||
| 8328 | } | |||
| 8329 | ||||
| 8330 | // See if we can lower this build_vector to a INSERTPS. | |||
| 8331 | if (!Subtarget.hasSSE41()) | |||
| 8332 | return SDValue(); | |||
| 8333 | ||||
| 8334 | SDValue V2 = Elt.getOperand(0); | |||
| 8335 | if (Elt == FirstNonZero && EltIdx == FirstNonZeroIdx) | |||
| 8336 | V1 = SDValue(); | |||
| 8337 | ||||
| 8338 | bool CanFold = true; | |||
| 8339 | for (unsigned i = EltIdx + 1; i < 4 && CanFold; ++i) { | |||
| 8340 | if (Zeroable[i]) | |||
| 8341 | continue; | |||
| 8342 | ||||
| 8343 | SDValue Current = Op->getOperand(i); | |||
| 8344 | SDValue SrcVector = Current->getOperand(0); | |||
| 8345 | if (!V1.getNode()) | |||
| 8346 | V1 = SrcVector; | |||
| 8347 | CanFold = (SrcVector == V1) && (Current.getConstantOperandAPInt(1) == i); | |||
| 8348 | } | |||
| 8349 | ||||
| 8350 | if (!CanFold) | |||
| 8351 | return SDValue(); | |||
| 8352 | ||||
| 8353 | assert(V1.getNode() && "Expected at least two non-zero elements!")((void)0); | |||
| 8354 | if (V1.getSimpleValueType() != MVT::v4f32) | |||
| 8355 | V1 = DAG.getBitcast(MVT::v4f32, V1); | |||
| 8356 | if (V2.getSimpleValueType() != MVT::v4f32) | |||
| 8357 | V2 = DAG.getBitcast(MVT::v4f32, V2); | |||
| 8358 | ||||
| 8359 | // Ok, we can emit an INSERTPS instruction. | |||
| 8360 | unsigned ZMask = Zeroable.to_ulong(); | |||
| 8361 | ||||
| 8362 | unsigned InsertPSMask = EltMaskIdx << 6 | EltIdx << 4 | ZMask; | |||
| 8363 | assert((InsertPSMask & ~0xFFu) == 0 && "Invalid mask!")((void)0); | |||
| 8364 | SDLoc DL(Op); | |||
| 8365 | SDValue Result = DAG.getNode(X86ISD::INSERTPS, DL, MVT::v4f32, V1, V2, | |||
| 8366 | DAG.getIntPtrConstant(InsertPSMask, DL, true)); | |||
| 8367 | return DAG.getBitcast(VT, Result); | |||
| 8368 | } | |||
| 8369 | ||||
| 8370 | /// Return a vector logical shift node. | |||
| 8371 | static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp, unsigned NumBits, | |||
| 8372 | SelectionDAG &DAG, const TargetLowering &TLI, | |||
| 8373 | const SDLoc &dl) { | |||
| 8374 | assert(VT.is128BitVector() && "Unknown type for VShift")((void)0); | |||
| 8375 | MVT ShVT = MVT::v16i8; | |||
| 8376 | unsigned Opc = isLeft ? X86ISD::VSHLDQ : X86ISD::VSRLDQ; | |||
| 8377 | SrcOp = DAG.getBitcast(ShVT, SrcOp); | |||
| 8378 | assert(NumBits % 8 == 0 && "Only support byte sized shifts")((void)0); | |||
| 8379 | SDValue ShiftVal = DAG.getTargetConstant(NumBits / 8, dl, MVT::i8); | |||
| 8380 | return DAG.getBitcast(VT, DAG.getNode(Opc, dl, ShVT, SrcOp, ShiftVal)); | |||
| 8381 | } | |||
| 8382 | ||||
| 8383 | static SDValue LowerAsSplatVectorLoad(SDValue SrcOp, MVT VT, const SDLoc &dl, | |||
| 8384 | SelectionDAG &DAG) { | |||
| 8385 | ||||
| 8386 | // Check if the scalar load can be widened into a vector load. And if | |||
| 8387 | // the address is "base + cst" see if the cst can be "absorbed" into | |||
| 8388 | // the shuffle mask. | |||
| 8389 | if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) { | |||
| 8390 | SDValue Ptr = LD->getBasePtr(); | |||
| 8391 | if (!ISD::isNormalLoad(LD) || !LD->isSimple()) | |||
| 8392 | return SDValue(); | |||
| 8393 | EVT PVT = LD->getValueType(0); | |||
| 8394 | if (PVT != MVT::i32 && PVT != MVT::f32) | |||
| 8395 | return SDValue(); | |||
| 8396 | ||||
| 8397 | int FI = -1; | |||
| 8398 | int64_t Offset = 0; | |||
| 8399 | if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) { | |||
| 8400 | FI = FINode->getIndex(); | |||
| 8401 | Offset = 0; | |||
| 8402 | } else if (DAG.isBaseWithConstantOffset(Ptr) && | |||
| 8403 | isa<FrameIndexSDNode>(Ptr.getOperand(0))) { | |||
| 8404 | FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex(); | |||
| 8405 | Offset = Ptr.getConstantOperandVal(1); | |||
| 8406 | Ptr = Ptr.getOperand(0); | |||
| 8407 | } else { | |||
| 8408 | return SDValue(); | |||
| 8409 | } | |||
| 8410 | ||||
| 8411 | // FIXME: 256-bit vector instructions don't require a strict alignment, | |||
| 8412 | // improve this code to support it better. | |||
| 8413 | Align RequiredAlign(VT.getSizeInBits() / 8); | |||
| 8414 | SDValue Chain = LD->getChain(); | |||
| 8415 | // Make sure the stack object alignment is at least 16 or 32. | |||
| 8416 | MachineFrameInfo &MFI = DAG.getMachineFunction().getFrameInfo(); | |||
| 8417 | MaybeAlign InferredAlign = DAG.InferPtrAlign(Ptr); | |||
| 8418 | if (!InferredAlign || *InferredAlign < RequiredAlign) { | |||
| 8419 | if (MFI.isFixedObjectIndex(FI)) { | |||
| 8420 | // Can't change the alignment. FIXME: It's possible to compute | |||
| 8421 | // the exact stack offset and reference FI + adjust offset instead. | |||
| 8422 | // If someone *really* cares about this. That's the way to implement it. | |||
| 8423 | return SDValue(); | |||
| 8424 | } else { | |||
| 8425 | MFI.setObjectAlignment(FI, RequiredAlign); | |||
| 8426 | } | |||
| 8427 | } | |||
| 8428 | ||||
| 8429 | // (Offset % 16 or 32) must be multiple of 4. Then address is then | |||
| 8430 | // Ptr + (Offset & ~15). | |||
| 8431 | if (Offset < 0) | |||
| 8432 | return SDValue(); | |||
| 8433 | if ((Offset % RequiredAlign.value()) & 3) | |||
| 8434 | return SDValue(); | |||
| 8435 | int64_t StartOffset = Offset & ~int64_t(RequiredAlign.value() - 1); | |||
| 8436 | if (StartOffset) { | |||
| 8437 | SDLoc DL(Ptr); | |||
| 8438 | Ptr = DAG.getNode(ISD::ADD, DL, Ptr.getValueType(), Ptr, | |||
| 8439 | DAG.getConstant(StartOffset, DL, Ptr.getValueType())); | |||
| 8440 | } | |||
| 8441 | ||||
| 8442 | int EltNo = (Offset - StartOffset) >> 2; | |||
| 8443 | unsigned NumElems = VT.getVectorNumElements(); | |||
| 8444 | ||||
| 8445 | EVT NVT = EVT::getVectorVT(*DAG.getContext(), PVT, NumElems); | |||
| 8446 | SDValue V1 = DAG.getLoad(NVT, dl, Chain, Ptr, | |||
| 8447 | LD->getPointerInfo().getWithOffset(StartOffset)); | |||
| 8448 | ||||
| 8449 | SmallVector<int, 8> Mask(NumElems, EltNo); | |||
| 8450 | ||||
| 8451 | return DAG.getVectorShuffle(NVT, dl, V1, DAG.getUNDEF(NVT), Mask); | |||
| 8452 | } | |||
| 8453 | ||||
| 8454 | return SDValue(); | |||
| 8455 | } | |||
| 8456 | ||||
| 8457 | // Recurse to find a LoadSDNode source and the accumulated ByteOffest. | |||
| 8458 | static bool findEltLoadSrc(SDValue Elt, LoadSDNode *&Ld, int64_t &ByteOffset) { | |||
| 8459 | if (ISD::isNON_EXTLoad(Elt.getNode())) { | |||
| 8460 | auto *BaseLd = cast<LoadSDNode>(Elt); | |||
| 8461 | if (!BaseLd->isSimple()) | |||
| 8462 | return false; | |||
| 8463 | Ld = BaseLd; | |||
| 8464 | ByteOffset = 0; | |||
| 8465 | return true; | |||
| 8466 | } | |||
| 8467 | ||||
| 8468 | switch (Elt.getOpcode()) { | |||
| 8469 | case ISD::BITCAST: | |||
| 8470 | case ISD::TRUNCATE: | |||
| 8471 | case ISD::SCALAR_TO_VECTOR: | |||
| 8472 | return findEltLoadSrc(Elt.getOperand(0), Ld, ByteOffset); | |||
| 8473 | case ISD::SRL: | |||
| 8474 | if (auto *IdxC = dyn_cast<ConstantSDNode>(Elt.getOperand(1))) { | |||
| 8475 | uint64_t Idx = IdxC->getZExtValue(); | |||
| 8476 | if ((Idx % 8) == 0 && findEltLoadSrc(Elt.getOperand(0), Ld, ByteOffset)) { | |||
| 8477 | ByteOffset += Idx / 8; | |||
| 8478 | return true; | |||
| 8479 | } | |||
| 8480 | } | |||
| 8481 | break; | |||
| 8482 | case ISD::EXTRACT_VECTOR_ELT: | |||
| 8483 | if (auto *IdxC = dyn_cast<ConstantSDNode>(Elt.getOperand(1))) { | |||
| 8484 | SDValue Src = Elt.getOperand(0); | |||
| 8485 | unsigned SrcSizeInBits = Src.getScalarValueSizeInBits(); | |||
| 8486 | unsigned DstSizeInBits = Elt.getScalarValueSizeInBits(); | |||
| 8487 | if (DstSizeInBits == SrcSizeInBits && (SrcSizeInBits % 8) == 0 && | |||
| 8488 | findEltLoadSrc(Src, Ld, ByteOffset)) { | |||
| 8489 | uint64_t Idx = IdxC->getZExtValue(); | |||
| 8490 | ByteOffset += Idx * (SrcSizeInBits / 8); | |||
| 8491 | return true; | |||
| 8492 | } | |||
| 8493 | } | |||
| 8494 | break; | |||
| 8495 | } | |||
| 8496 | ||||
| 8497 | return false; | |||
| 8498 | } | |||
| 8499 | ||||
| 8500 | /// Given the initializing elements 'Elts' of a vector of type 'VT', see if the | |||
| 8501 | /// elements can be replaced by a single large load which has the same value as | |||
| 8502 | /// a build_vector or insert_subvector whose loaded operands are 'Elts'. | |||
| 8503 | /// | |||
| 8504 | /// Example: <load i32 *a, load i32 *a+4, zero, undef> -> zextload a | |||
| 8505 | static SDValue EltsFromConsecutiveLoads(EVT VT, ArrayRef<SDValue> Elts, | |||
| 8506 | const SDLoc &DL, SelectionDAG &DAG, | |||
| 8507 | const X86Subtarget &Subtarget, | |||
| 8508 | bool IsAfterLegalize) { | |||
| 8509 | if ((VT.getScalarSizeInBits() % 8) != 0) | |||
| 8510 | return SDValue(); | |||
| 8511 | ||||
| 8512 | unsigned NumElems = Elts.size(); | |||
| 8513 | ||||
| 8514 | int LastLoadedElt = -1; | |||
| 8515 | APInt LoadMask = APInt::getNullValue(NumElems); | |||
| 8516 | APInt ZeroMask = APInt::getNullValue(NumElems); | |||
| 8517 | APInt UndefMask = APInt::getNullValue(NumElems); | |||
| 8518 | ||||
| 8519 | SmallVector<LoadSDNode*, 8> Loads(NumElems, nullptr); | |||
| 8520 | SmallVector<int64_t, 8> ByteOffsets(NumElems, 0); | |||
| 8521 | ||||
| 8522 | // For each element in the initializer, see if we've found a load, zero or an | |||
| 8523 | // undef. | |||
| 8524 | for (unsigned i = 0; i < NumElems; ++i) { | |||
| 8525 | SDValue Elt = peekThroughBitcasts(Elts[i]); | |||
| 8526 | if (!Elt.getNode()) | |||
| 8527 | return SDValue(); | |||
| 8528 | if (Elt.isUndef()) { | |||
| 8529 | UndefMask.setBit(i); | |||
| 8530 | continue; | |||
| 8531 | } | |||
| 8532 | if (X86::isZeroNode(Elt) || ISD::isBuildVectorAllZeros(Elt.getNode())) { | |||
| 8533 | ZeroMask.setBit(i); | |||
| 8534 | continue; | |||
| 8535 | } | |||
| 8536 | ||||
| 8537 | // Each loaded element must be the correct fractional portion of the | |||
| 8538 | // requested vector load. | |||
| 8539 | unsigned EltSizeInBits = Elt.getValueSizeInBits(); | |||
| 8540 | if ((NumElems * EltSizeInBits) != VT.getSizeInBits()) | |||
| 8541 | return SDValue(); | |||
| 8542 | ||||
| 8543 | if (!findEltLoadSrc(Elt, Loads[i], ByteOffsets[i]) || ByteOffsets[i] < 0) | |||
| 8544 | return SDValue(); | |||
| 8545 | unsigned LoadSizeInBits = Loads[i]->getValueSizeInBits(0); | |||
| 8546 | if (((ByteOffsets[i] * 8) + EltSizeInBits) > LoadSizeInBits) | |||
| 8547 | return SDValue(); | |||
| 8548 | ||||
| 8549 | LoadMask.setBit(i); | |||
| 8550 | LastLoadedElt = i; | |||
| 8551 | } | |||
| 8552 | assert((ZeroMask.countPopulation() + UndefMask.countPopulation() +((void)0) | |||
| 8553 | LoadMask.countPopulation()) == NumElems &&((void)0) | |||
| 8554 | "Incomplete element masks")((void)0); | |||
| 8555 | ||||
| 8556 | // Handle Special Cases - all undef or undef/zero. | |||
| 8557 | if (UndefMask.countPopulation() == NumElems) | |||
| 8558 | return DAG.getUNDEF(VT); | |||
| 8559 | if ((ZeroMask.countPopulation() + UndefMask.countPopulation()) == NumElems) | |||
| 8560 | return VT.isInteger() ? DAG.getConstant(0, DL, VT) | |||
| 8561 | : DAG.getConstantFP(0.0, DL, VT); | |||
| 8562 | ||||
| 8563 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
| 8564 | int FirstLoadedElt = LoadMask.countTrailingZeros(); | |||
| 8565 | SDValue EltBase = peekThroughBitcasts(Elts[FirstLoadedElt]); | |||
| 8566 | EVT EltBaseVT = EltBase.getValueType(); | |||
| 8567 | assert(EltBaseVT.getSizeInBits() == EltBaseVT.getStoreSizeInBits() &&((void)0) | |||
| 8568 | "Register/Memory size mismatch")((void)0); | |||
| 8569 | LoadSDNode *LDBase = Loads[FirstLoadedElt]; | |||
| 8570 | assert(LDBase && "Did not find base load for merging consecutive loads")((void)0); | |||
| 8571 | unsigned BaseSizeInBits = EltBaseVT.getStoreSizeInBits(); | |||
| 8572 | unsigned BaseSizeInBytes = BaseSizeInBits / 8; | |||
| 8573 | int NumLoadedElts = (1 + LastLoadedElt - FirstLoadedElt); | |||
| 8574 | int LoadSizeInBits = NumLoadedElts * BaseSizeInBits; | |||
| 8575 | assert((BaseSizeInBits % 8) == 0 && "Sub-byte element loads detected")((void)0); | |||
| 8576 | ||||
| 8577 | // TODO: Support offsetting the base load. | |||
| 8578 | if (ByteOffsets[FirstLoadedElt] != 0) | |||
| 8579 | return SDValue(); | |||
| 8580 | ||||
| 8581 | // Check to see if the element's load is consecutive to the base load | |||
| 8582 | // or offset from a previous (already checked) load. | |||
| 8583 | auto CheckConsecutiveLoad = [&](LoadSDNode *Base, int EltIdx) { | |||
| 8584 | LoadSDNode *Ld = Loads[EltIdx]; | |||
| 8585 | int64_t ByteOffset = ByteOffsets[EltIdx]; | |||
| 8586 | if (ByteOffset && (ByteOffset % BaseSizeInBytes) == 0) { | |||
| 8587 | int64_t BaseIdx = EltIdx - (ByteOffset / BaseSizeInBytes); | |||
| 8588 | return (0 <= BaseIdx && BaseIdx < (int)NumElems && LoadMask[BaseIdx] && | |||
| 8589 | Loads[BaseIdx] == Ld && ByteOffsets[BaseIdx] == 0); | |||
| 8590 | } | |||
| 8591 | return DAG.areNonVolatileConsecutiveLoads(Ld, Base, BaseSizeInBytes, | |||
| 8592 | EltIdx - FirstLoadedElt); | |||
| 8593 | }; | |||
| 8594 | ||||
| 8595 | // Consecutive loads can contain UNDEFS but not ZERO elements. | |||
| 8596 | // Consecutive loads with UNDEFs and ZEROs elements require a | |||
| 8597 | // an additional shuffle stage to clear the ZERO elements. | |||
| 8598 | bool IsConsecutiveLoad = true; | |||
| 8599 | bool IsConsecutiveLoadWithZeros = true; | |||
| 8600 | for (int i = FirstLoadedElt + 1; i <= LastLoadedElt; ++i) { | |||
| 8601 | if (LoadMask[i]) { | |||
| 8602 | if (!CheckConsecutiveLoad(LDBase, i)) { | |||
| 8603 | IsConsecutiveLoad = false; | |||
| 8604 | IsConsecutiveLoadWithZeros = false; | |||
| 8605 | break; | |||
| 8606 | } | |||
| 8607 | } else if (ZeroMask[i]) { | |||
| 8608 | IsConsecutiveLoad = false; | |||
| 8609 | } | |||
| 8610 | } | |||
| 8611 | ||||
| 8612 | auto CreateLoad = [&DAG, &DL, &Loads](EVT VT, LoadSDNode *LDBase) { | |||
| 8613 | auto MMOFlags = LDBase->getMemOperand()->getFlags(); | |||
| 8614 | assert(LDBase->isSimple() &&((void)0) | |||
| 8615 | "Cannot merge volatile or atomic loads.")((void)0); | |||
| 8616 | SDValue NewLd = | |||
| 8617 | DAG.getLoad(VT, DL, LDBase->getChain(), LDBase->getBasePtr(), | |||
| 8618 | LDBase->getPointerInfo(), LDBase->getOriginalAlign(), | |||
| 8619 | MMOFlags); | |||
| 8620 | for (auto *LD : Loads) | |||
| 8621 | if (LD) | |||
| 8622 | DAG.makeEquivalentMemoryOrdering(LD, NewLd); | |||
| 8623 | return NewLd; | |||
| 8624 | }; | |||
| 8625 | ||||
| 8626 | // Check if the base load is entirely dereferenceable. | |||
| 8627 | bool IsDereferenceable = LDBase->getPointerInfo().isDereferenceable( | |||
| 8628 | VT.getSizeInBits() / 8, *DAG.getContext(), DAG.getDataLayout()); | |||
| 8629 | ||||
| 8630 | // LOAD - all consecutive load/undefs (must start/end with a load or be | |||
| 8631 | // entirely dereferenceable). If we have found an entire vector of loads and | |||
| 8632 | // undefs, then return a large load of the entire vector width starting at the | |||
| 8633 | // base pointer. If the vector contains zeros, then attempt to shuffle those | |||
| 8634 | // elements. | |||
| 8635 | if (FirstLoadedElt == 0 && | |||
| 8636 | (NumLoadedElts == (int)NumElems || IsDereferenceable) && | |||
| 8637 | (IsConsecutiveLoad || IsConsecutiveLoadWithZeros)) { | |||
| 8638 | if (IsAfterLegalize && !TLI.isOperationLegal(ISD::LOAD, VT)) | |||
| 8639 | return SDValue(); | |||
| 8640 | ||||
| 8641 | // Don't create 256-bit non-temporal aligned loads without AVX2 as these | |||
| 8642 | // will lower to regular temporal loads and use the cache. | |||
| 8643 | if (LDBase->isNonTemporal() && LDBase->getAlignment() >= 32 && | |||
| 8644 | VT.is256BitVector() && !Subtarget.hasInt256()) | |||
| 8645 | return SDValue(); | |||
| 8646 | ||||
| 8647 | if (NumElems == 1) | |||
| 8648 | return DAG.getBitcast(VT, Elts[FirstLoadedElt]); | |||
| 8649 | ||||
| 8650 | if (!ZeroMask) | |||
| 8651 | return CreateLoad(VT, LDBase); | |||
| 8652 | ||||
| 8653 | // IsConsecutiveLoadWithZeros - we need to create a shuffle of the loaded | |||
| 8654 | // vector and a zero vector to clear out the zero elements. | |||
| 8655 | if (!IsAfterLegalize && VT.isVector()) { | |||
| 8656 | unsigned NumMaskElts = VT.getVectorNumElements(); | |||
| 8657 | if ((NumMaskElts % NumElems) == 0) { | |||
| 8658 | unsigned Scale = NumMaskElts / NumElems; | |||
| 8659 | SmallVector<int, 4> ClearMask(NumMaskElts, -1); | |||
| 8660 | for (unsigned i = 0; i < NumElems; ++i) { | |||
| 8661 | if (UndefMask[i]) | |||
| 8662 | continue; | |||
| 8663 | int Offset = ZeroMask[i] ? NumMaskElts : 0; | |||
| 8664 | for (unsigned j = 0; j != Scale; ++j) | |||
| 8665 | ClearMask[(i * Scale) + j] = (i * Scale) + j + Offset; | |||
| 8666 | } | |||
| 8667 | SDValue V = CreateLoad(VT, LDBase); | |||
| 8668 | SDValue Z = VT.isInteger() ? DAG.getConstant(0, DL, VT) | |||
| 8669 | : DAG.getConstantFP(0.0, DL, VT); | |||
| 8670 | return DAG.getVectorShuffle(VT, DL, V, Z, ClearMask); | |||
| 8671 | } | |||
| 8672 | } | |||
| 8673 | } | |||
| 8674 | ||||
| 8675 | // If the upper half of a ymm/zmm load is undef then just load the lower half. | |||
| 8676 | if (VT.is256BitVector() || VT.is512BitVector()) { | |||
| 8677 | unsigned HalfNumElems = NumElems / 2; | |||
| 8678 | if (UndefMask.extractBits(HalfNumElems, HalfNumElems).isAllOnesValue()) { | |||
| 8679 | EVT HalfVT = | |||
| 8680 | EVT::getVectorVT(*DAG.getContext(), VT.getScalarType(), HalfNumElems); | |||
| 8681 | SDValue HalfLD = | |||
| 8682 | EltsFromConsecutiveLoads(HalfVT, Elts.drop_back(HalfNumElems), DL, | |||
| 8683 | DAG, Subtarget, IsAfterLegalize); | |||
| 8684 | if (HalfLD) | |||
| 8685 | return DAG.getNode(ISD::INSERT_SUBVECTOR, DL, VT, DAG.getUNDEF(VT), | |||
| 8686 | HalfLD, DAG.getIntPtrConstant(0, DL)); | |||
| 8687 | } | |||
| 8688 | } | |||
| 8689 | ||||
| 8690 | // VZEXT_LOAD - consecutive 32/64-bit load/undefs followed by zeros/undefs. | |||
| 8691 | if (IsConsecutiveLoad && FirstLoadedElt == 0 && | |||
| 8692 | (LoadSizeInBits == 32 || LoadSizeInBits == 64) && | |||
| 8693 | ((VT.is128BitVector() || VT.is256BitVector() || VT.is512BitVector()))) { | |||
| 8694 | MVT VecSVT = VT.isFloatingPoint() ? MVT::getFloatingPointVT(LoadSizeInBits) | |||
| 8695 | : MVT::getIntegerVT(LoadSizeInBits); | |||
| 8696 | MVT VecVT = MVT::getVectorVT(VecSVT, VT.getSizeInBits() / LoadSizeInBits); | |||
| 8697 | // Allow v4f32 on SSE1 only targets. | |||
| 8698 | // FIXME: Add more isel patterns so we can just use VT directly. | |||
| 8699 | if (!Subtarget.hasSSE2() && VT == MVT::v4f32) | |||
| 8700 | VecVT = MVT::v4f32; | |||
| 8701 | if (TLI.isTypeLegal(VecVT)) { | |||
| 8702 | SDVTList Tys = DAG.getVTList(VecVT, MVT::Other); | |||
| 8703 | SDValue Ops[] = { LDBase->getChain(), LDBase->getBasePtr() }; | |||
| 8704 | SDValue ResNode = DAG.getMemIntrinsicNode( | |||
| 8705 | X86ISD::VZEXT_LOAD, DL, Tys, Ops, VecSVT, LDBase->getPointerInfo(), | |||
| 8706 | LDBase->getOriginalAlign(), MachineMemOperand::MOLoad); | |||
| 8707 | for (auto *LD : Loads) | |||
| 8708 | if (LD) | |||
| 8709 | DAG.makeEquivalentMemoryOrdering(LD, ResNode); | |||
| 8710 | return DAG.getBitcast(VT, ResNode); | |||
| 8711 | } | |||
| 8712 | } | |||
| 8713 | ||||
| 8714 | // BROADCAST - match the smallest possible repetition pattern, load that | |||
| 8715 | // scalar/subvector element and then broadcast to the entire vector. | |||
| 8716 | if (ZeroMask.isNullValue() && isPowerOf2_32(NumElems) && Subtarget.hasAVX() && | |||
| 8717 | (VT.is128BitVector() || VT.is256BitVector() || VT.is512BitVector())) { | |||
| 8718 | for (unsigned SubElems = 1; SubElems < NumElems; SubElems *= 2) { | |||
| 8719 | unsigned RepeatSize = SubElems * BaseSizeInBits; | |||
| 8720 | unsigned ScalarSize = std::min(RepeatSize, 64u); | |||
| 8721 | if (!Subtarget.hasAVX2() && ScalarSize < 32) | |||
| 8722 | continue; | |||
| 8723 | ||||
| 8724 | // Don't attempt a 1:N subvector broadcast - it should be caught by | |||
| 8725 | // combineConcatVectorOps, else will cause infinite loops. | |||
| 8726 | if (RepeatSize > ScalarSize && SubElems == 1) | |||
| 8727 | continue; | |||
| 8728 | ||||
| 8729 | bool Match = true; | |||
| 8730 | SmallVector<SDValue, 8> RepeatedLoads(SubElems, DAG.getUNDEF(EltBaseVT)); | |||
| 8731 | for (unsigned i = 0; i != NumElems && Match; ++i) { | |||
| 8732 | if (!LoadMask[i]) | |||
| 8733 | continue; | |||
| 8734 | SDValue Elt = peekThroughBitcasts(Elts[i]); | |||
| 8735 | if (RepeatedLoads[i % SubElems].isUndef()) | |||
| 8736 | RepeatedLoads[i % SubElems] = Elt; | |||
| 8737 | else | |||
| 8738 | Match &= (RepeatedLoads[i % SubElems] == Elt); | |||
| 8739 | } | |||
| 8740 | ||||
| 8741 | // We must have loads at both ends of the repetition. | |||
| 8742 | Match &= !RepeatedLoads.front().isUndef(); | |||
| 8743 | Match &= !RepeatedLoads.back().isUndef(); | |||
| 8744 | if (!Match) | |||
| 8745 | continue; | |||
| 8746 | ||||
| 8747 | EVT RepeatVT = | |||
| 8748 | VT.isInteger() && (RepeatSize != 64 || TLI.isTypeLegal(MVT::i64)) | |||
| 8749 | ? EVT::getIntegerVT(*DAG.getContext(), ScalarSize) | |||
| 8750 | : EVT::getFloatingPointVT(ScalarSize); | |||
| 8751 | if (RepeatSize > ScalarSize) | |||
| 8752 | RepeatVT = EVT::getVectorVT(*DAG.getContext(), RepeatVT, | |||
| 8753 | RepeatSize / ScalarSize); | |||
| 8754 | EVT BroadcastVT = | |||
| 8755 | EVT::getVectorVT(*DAG.getContext(), RepeatVT.getScalarType(), | |||
| 8756 | VT.getSizeInBits() / ScalarSize); | |||
| 8757 | if (TLI.isTypeLegal(BroadcastVT)) { | |||
| 8758 | if (SDValue RepeatLoad = EltsFromConsecutiveLoads( | |||
| 8759 | RepeatVT, RepeatedLoads, DL, DAG, Subtarget, IsAfterLegalize)) { | |||
| 8760 | SDValue Broadcast = RepeatLoad; | |||
| 8761 | if (RepeatSize > ScalarSize) { | |||
| 8762 | while (Broadcast.getValueSizeInBits() < VT.getSizeInBits()) | |||
| 8763 | Broadcast = concatSubVectors(Broadcast, Broadcast, DAG, DL); | |||
| 8764 | } else { | |||
| 8765 | Broadcast = | |||
| 8766 | DAG.getNode(X86ISD::VBROADCAST, DL, BroadcastVT, RepeatLoad); | |||
| 8767 | } | |||
| 8768 | return DAG.getBitcast(VT, Broadcast); | |||
| 8769 | } | |||
| 8770 | } | |||
| 8771 | } | |||
| 8772 | } | |||
| 8773 | ||||
| 8774 | return SDValue(); | |||
| 8775 | } | |||
| 8776 | ||||
| 8777 | // Combine a vector ops (shuffles etc.) that is equal to build_vector load1, | |||
| 8778 | // load2, load3, load4, <0, 1, 2, 3> into a vector load if the load addresses | |||
| 8779 | // are consecutive, non-overlapping, and in the right order. | |||
| 8780 | static SDValue combineToConsecutiveLoads(EVT VT, SDValue Op, const SDLoc &DL, | |||
| 8781 | SelectionDAG &DAG, | |||
| 8782 | const X86Subtarget &Subtarget, | |||
| 8783 | bool IsAfterLegalize) { | |||
| 8784 | SmallVector<SDValue, 64> Elts; | |||
| 8785 | for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i) { | |||
| 8786 | if (SDValue Elt = getShuffleScalarElt(Op, i, DAG, 0)) { | |||
| 8787 | Elts.push_back(Elt); | |||
| 8788 | continue; | |||
| 8789 | } | |||
| 8790 | return SDValue(); | |||
| 8791 | } | |||
| 8792 | assert(Elts.size() == VT.getVectorNumElements())((void)0); | |||
| 8793 | return EltsFromConsecutiveLoads(VT, Elts, DL, DAG, Subtarget, | |||
| 8794 | IsAfterLegalize); | |||
| 8795 | } | |||
| 8796 | ||||
| 8797 | static Constant *getConstantVector(MVT VT, const APInt &SplatValue, | |||
| 8798 | unsigned SplatBitSize, LLVMContext &C) { | |||
| 8799 | unsigned ScalarSize = VT.getScalarSizeInBits(); | |||
| 8800 | unsigned NumElm = SplatBitSize / ScalarSize; | |||
| 8801 | ||||
| 8802 | SmallVector<Constant *, 32> ConstantVec; | |||
| 8803 | for (unsigned i = 0; i < NumElm; i++) { | |||
| 8804 | APInt Val = SplatValue.extractBits(ScalarSize, ScalarSize * i); | |||
| 8805 | Constant *Const; | |||
| 8806 | if (VT.isFloatingPoint()) { | |||
| 8807 | if (ScalarSize == 32) { | |||
| 8808 | Const = ConstantFP::get(C, APFloat(APFloat::IEEEsingle(), Val)); | |||
| 8809 | } else { | |||
| 8810 | assert(ScalarSize == 64 && "Unsupported floating point scalar size")((void)0); | |||
| 8811 | Const = ConstantFP::get(C, APFloat(APFloat::IEEEdouble(), Val)); | |||
| 8812 | } | |||
| 8813 | } else | |||
| 8814 | Const = Constant::getIntegerValue(Type::getIntNTy(C, ScalarSize), Val); | |||
| 8815 | ConstantVec.push_back(Const); | |||
| 8816 | } | |||
| 8817 | return ConstantVector::get(ArrayRef<Constant *>(ConstantVec)); | |||
| 8818 | } | |||
| 8819 | ||||
| 8820 | static bool isFoldableUseOfShuffle(SDNode *N) { | |||
| 8821 | for (auto *U : N->uses()) { | |||
| 8822 | unsigned Opc = U->getOpcode(); | |||
| 8823 | // VPERMV/VPERMV3 shuffles can never fold their index operands. | |||
| 8824 | if (Opc == X86ISD::VPERMV && U->getOperand(0).getNode() == N) | |||
| 8825 | return false; | |||
| 8826 | if (Opc == X86ISD::VPERMV3 && U->getOperand(1).getNode() == N) | |||
| 8827 | return false; | |||
| 8828 | if (isTargetShuffle(Opc)) | |||
| 8829 | return true; | |||
| 8830 | if (Opc == ISD::BITCAST) // Ignore bitcasts | |||
| 8831 | return isFoldableUseOfShuffle(U); | |||
| 8832 | if (N->hasOneUse()) | |||
| 8833 | return true; | |||
| 8834 | } | |||
| 8835 | return false; | |||
| 8836 | } | |||
| 8837 | ||||
| 8838 | /// Attempt to use the vbroadcast instruction to generate a splat value | |||
| 8839 | /// from a splat BUILD_VECTOR which uses: | |||
| 8840 | /// a. A single scalar load, or a constant. | |||
| 8841 | /// b. Repeated pattern of constants (e.g. <0,1,0,1> or <0,1,2,3,0,1,2,3>). | |||
| 8842 | /// | |||
| 8843 | /// The VBROADCAST node is returned when a pattern is found, | |||
| 8844 | /// or SDValue() otherwise. | |||
| 8845 | static SDValue lowerBuildVectorAsBroadcast(BuildVectorSDNode *BVOp, | |||
| 8846 | const X86Subtarget &Subtarget, | |||
| 8847 | SelectionDAG &DAG) { | |||
| 8848 | // VBROADCAST requires AVX. | |||
| 8849 | // TODO: Splats could be generated for non-AVX CPUs using SSE | |||
| 8850 | // instructions, but there's less potential gain for only 128-bit vectors. | |||
| 8851 | if (!Subtarget.hasAVX()) | |||
| 8852 | return SDValue(); | |||
| 8853 | ||||
| 8854 | MVT VT = BVOp->getSimpleValueType(0); | |||
| 8855 | unsigned NumElts = VT.getVectorNumElements(); | |||
| 8856 | SDLoc dl(BVOp); | |||
| 8857 | ||||
| 8858 | assert((VT.is128BitVector() || VT.is256BitVector() || VT.is512BitVector()) &&((void)0) | |||
| 8859 | "Unsupported vector type for broadcast.")((void)0); | |||
| 8860 | ||||
| 8861 | // See if the build vector is a repeating sequence of scalars (inc. splat). | |||
| 8862 | SDValue Ld; | |||
| 8863 | BitVector UndefElements; | |||
| 8864 | SmallVector<SDValue, 16> Sequence; | |||
| 8865 | if (BVOp->getRepeatedSequence(Sequence, &UndefElements)) { | |||
| 8866 | assert((NumElts % Sequence.size()) == 0 && "Sequence doesn't fit.")((void)0); | |||
| 8867 | if (Sequence.size() == 1) | |||
| 8868 | Ld = Sequence[0]; | |||
| 8869 | } | |||
| 8870 | ||||
| 8871 | // Attempt to use VBROADCASTM | |||
| 8872 | // From this pattern: | |||
| 8873 | // a. t0 = (zext_i64 (bitcast_i8 v2i1 X)) | |||
| 8874 | // b. t1 = (build_vector t0 t0) | |||
| 8875 | // | |||
| 8876 | // Create (VBROADCASTM v2i1 X) | |||
| 8877 | if (!Sequence.empty() && Subtarget.hasCDI()) { | |||
| 8878 | // If not a splat, are the upper sequence values zeroable? | |||
| 8879 | unsigned SeqLen = Sequence.size(); | |||
| 8880 | bool UpperZeroOrUndef = | |||
| 8881 | SeqLen == 1 || | |||
| 8882 | llvm::all_of(makeArrayRef(Sequence).drop_front(), [](SDValue V) { | |||
| 8883 | return !V || V.isUndef() || isNullConstant(V); | |||
| 8884 | }); | |||
| 8885 | SDValue Op0 = Sequence[0]; | |||
| 8886 | if (UpperZeroOrUndef && ((Op0.getOpcode() == ISD::BITCAST) || | |||
| 8887 | (Op0.getOpcode() == ISD::ZERO_EXTEND && | |||
| 8888 | Op0.getOperand(0).getOpcode() == ISD::BITCAST))) { | |||
| 8889 | SDValue BOperand = Op0.getOpcode() == ISD::BITCAST | |||
| 8890 | ? Op0.getOperand(0) | |||
| 8891 | : Op0.getOperand(0).getOperand(0); | |||
| 8892 | MVT MaskVT = BOperand.getSimpleValueType(); | |||
| 8893 | MVT EltType = MVT::getIntegerVT(VT.getScalarSizeInBits() * SeqLen); | |||
| 8894 | if ((EltType == MVT::i64 && MaskVT == MVT::v8i1) || // for broadcastmb2q | |||
| 8895 | (EltType == MVT::i32 && MaskVT == MVT::v16i1)) { // for broadcastmw2d | |||
| 8896 | MVT BcstVT = MVT::getVectorVT(EltType, NumElts / SeqLen); | |||
| 8897 | if (!VT.is512BitVector() && !Subtarget.hasVLX()) { | |||
| 8898 | unsigned Scale = 512 / VT.getSizeInBits(); | |||
| 8899 | BcstVT = MVT::getVectorVT(EltType, Scale * (NumElts / SeqLen)); | |||
| 8900 | } | |||
| 8901 | SDValue Bcst = DAG.getNode(X86ISD::VBROADCASTM, dl, BcstVT, BOperand); | |||
| 8902 | if (BcstVT.getSizeInBits() != VT.getSizeInBits()) | |||
| 8903 | Bcst = extractSubVector(Bcst, 0, DAG, dl, VT.getSizeInBits()); | |||
| 8904 | return DAG.getBitcast(VT, Bcst); | |||
| 8905 | } | |||
| 8906 | } | |||
| 8907 | } | |||
| 8908 | ||||
| 8909 | unsigned NumUndefElts = UndefElements.count(); | |||
| 8910 | if (!Ld || (NumElts - NumUndefElts) <= 1) { | |||
| 8911 | APInt SplatValue, Undef; | |||
| 8912 | unsigned SplatBitSize; | |||
| 8913 | bool HasUndef; | |||
| 8914 | // Check if this is a repeated constant pattern suitable for broadcasting. | |||
| 8915 | if (BVOp->isConstantSplat(SplatValue, Undef, SplatBitSize, HasUndef) && | |||
| 8916 | SplatBitSize > VT.getScalarSizeInBits() && | |||
| 8917 | SplatBitSize < VT.getSizeInBits()) { | |||
| 8918 | // Avoid replacing with broadcast when it's a use of a shuffle | |||
| 8919 | // instruction to preserve the present custom lowering of shuffles. | |||
| 8920 | if (isFoldableUseOfShuffle(BVOp)) | |||
| 8921 | return SDValue(); | |||
| 8922 | // replace BUILD_VECTOR with broadcast of the repeated constants. | |||
| 8923 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
| 8924 | LLVMContext *Ctx = DAG.getContext(); | |||
| 8925 | MVT PVT = TLI.getPointerTy(DAG.getDataLayout()); | |||
| 8926 | if (Subtarget.hasAVX()) { | |||
| 8927 | if (SplatBitSize == 32 || SplatBitSize == 64 || | |||
| 8928 | (SplatBitSize < 32 && Subtarget.hasAVX2())) { | |||
| 8929 | // Splatted value can fit in one INTEGER constant in constant pool. | |||
| 8930 | // Load the constant and broadcast it. | |||
| 8931 | MVT CVT = MVT::getIntegerVT(SplatBitSize); | |||
| 8932 | Type *ScalarTy = Type::getIntNTy(*Ctx, SplatBitSize); | |||
| 8933 | Constant *C = Constant::getIntegerValue(ScalarTy, SplatValue); | |||
| 8934 | SDValue CP = DAG.getConstantPool(C, PVT); | |||
| 8935 | unsigned Repeat = VT.getSizeInBits() / SplatBitSize; | |||
| 8936 | ||||
| 8937 | Align Alignment = cast<ConstantPoolSDNode>(CP)->getAlign(); | |||
| 8938 | SDVTList Tys = | |||
| 8939 | DAG.getVTList(MVT::getVectorVT(CVT, Repeat), MVT::Other); | |||
| 8940 | SDValue Ops[] = {DAG.getEntryNode(), CP}; | |||
| 8941 | MachinePointerInfo MPI = | |||
| 8942 | MachinePointerInfo::getConstantPool(DAG.getMachineFunction()); | |||
| 8943 | SDValue Brdcst = DAG.getMemIntrinsicNode( | |||
| 8944 | X86ISD::VBROADCAST_LOAD, dl, Tys, Ops, CVT, MPI, Alignment, | |||
| 8945 | MachineMemOperand::MOLoad); | |||
| 8946 | return DAG.getBitcast(VT, Brdcst); | |||
| 8947 | } | |||
| 8948 | if (SplatBitSize > 64) { | |||
| 8949 | // Load the vector of constants and broadcast it. | |||
| 8950 | Constant *VecC = getConstantVector(VT, SplatValue, SplatBitSize, | |||
| 8951 | *Ctx); | |||
| 8952 | SDValue VCP = DAG.getConstantPool(VecC, PVT); | |||
| 8953 | unsigned NumElm = SplatBitSize / VT.getScalarSizeInBits(); | |||
| 8954 | MVT VVT = MVT::getVectorVT(VT.getScalarType(), NumElm); | |||
| 8955 | Align Alignment = cast<ConstantPoolSDNode>(VCP)->getAlign(); | |||
| 8956 | SDVTList Tys = DAG.getVTList(VT, MVT::Other); | |||
| 8957 | SDValue Ops[] = {DAG.getEntryNode(), VCP}; | |||
| 8958 | MachinePointerInfo MPI = | |||
| 8959 | MachinePointerInfo::getConstantPool(DAG.getMachineFunction()); | |||
| 8960 | return DAG.getMemIntrinsicNode( | |||
| 8961 | X86ISD::SUBV_BROADCAST_LOAD, dl, Tys, Ops, VVT, MPI, Alignment, | |||
| 8962 | MachineMemOperand::MOLoad); | |||
| 8963 | } | |||
| 8964 | } | |||
| 8965 | } | |||
| 8966 | ||||
| 8967 | // If we are moving a scalar into a vector (Ld must be set and all elements | |||
| 8968 | // but 1 are undef) and that operation is not obviously supported by | |||
| 8969 | // vmovd/vmovq/vmovss/vmovsd, then keep trying to form a broadcast. | |||
| 8970 | // That's better than general shuffling and may eliminate a load to GPR and | |||
| 8971 | // move from scalar to vector register. | |||
| 8972 | if (!Ld || NumElts - NumUndefElts != 1) | |||
| 8973 | return SDValue(); | |||
| 8974 | unsigned ScalarSize = Ld.getValueSizeInBits(); | |||
| 8975 | if (!(UndefElements[0] || (ScalarSize != 32 && ScalarSize != 64))) | |||
| 8976 | return SDValue(); | |||
| 8977 | } | |||
| 8978 | ||||
| 8979 | bool ConstSplatVal = | |||
| 8980 | (Ld.getOpcode() == ISD::Constant || Ld.getOpcode() == ISD::ConstantFP); | |||
| 8981 | bool IsLoad = ISD::isNormalLoad(Ld.getNode()); | |||
| 8982 | ||||
| 8983 | // TODO: Handle broadcasts of non-constant sequences. | |||
| 8984 | ||||
| 8985 | // Make sure that all of the users of a non-constant load are from the | |||
| 8986 | // BUILD_VECTOR node. | |||
| 8987 | // FIXME: Is the use count needed for non-constant, non-load case? | |||
| 8988 | if (!ConstSplatVal && !IsLoad && !BVOp->isOnlyUserOf(Ld.getNode())) | |||
| 8989 | return SDValue(); | |||
| 8990 | ||||
| 8991 | unsigned ScalarSize = Ld.getValueSizeInBits(); | |||
| 8992 | bool IsGE256 = (VT.getSizeInBits() >= 256); | |||
| 8993 | ||||
| 8994 | // When optimizing for size, generate up to 5 extra bytes for a broadcast | |||
| 8995 | // instruction to save 8 or more bytes of constant pool data. | |||
| 8996 | // TODO: If multiple splats are generated to load the same constant, | |||
| 8997 | // it may be detrimental to overall size. There needs to be a way to detect | |||
| 8998 | // that condition to know if this is truly a size win. | |||
| 8999 | bool OptForSize = DAG.shouldOptForSize(); | |||
| 9000 | ||||
| 9001 | // Handle broadcasting a single constant scalar from the constant pool | |||
| 9002 | // into a vector. | |||
| 9003 | // On Sandybridge (no AVX2), it is still better to load a constant vector | |||
| 9004 | // from the constant pool and not to broadcast it from a scalar. | |||
| 9005 | // But override that restriction when optimizing for size. | |||
| 9006 | // TODO: Check if splatting is recommended for other AVX-capable CPUs. | |||
| 9007 | if (ConstSplatVal && (Subtarget.hasAVX2() || OptForSize)) { | |||
| 9008 | EVT CVT = Ld.getValueType(); | |||
| 9009 | assert(!CVT.isVector() && "Must not broadcast a vector type")((void)0); | |||
| 9010 | ||||
| 9011 | // Splat f32, i32, v4f64, v4i64 in all cases with AVX2. | |||
| 9012 | // For size optimization, also splat v2f64 and v2i64, and for size opt | |||
| 9013 | // with AVX2, also splat i8 and i16. | |||
| 9014 | // With pattern matching, the VBROADCAST node may become a VMOVDDUP. | |||
| 9015 | if (ScalarSize == 32 || (IsGE256 && ScalarSize == 64) || | |||
| 9016 | (OptForSize && (ScalarSize == 64 || Subtarget.hasAVX2()))) { | |||
| 9017 | const Constant *C = nullptr; | |||
| 9018 | if (ConstantSDNode *CI = dyn_cast<ConstantSDNode>(Ld)) | |||
| 9019 | C = CI->getConstantIntValue(); | |||
| 9020 | else if (ConstantFPSDNode *CF = dyn_cast<ConstantFPSDNode>(Ld)) | |||
| 9021 | C = CF->getConstantFPValue(); | |||
| 9022 | ||||
| 9023 | assert(C && "Invalid constant type")((void)0); | |||
| 9024 | ||||
| 9025 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
| 9026 | SDValue CP = | |||
| 9027 | DAG.getConstantPool(C, TLI.getPointerTy(DAG.getDataLayout())); | |||
| 9028 | Align Alignment = cast<ConstantPoolSDNode>(CP)->getAlign(); | |||
| 9029 | ||||
| 9030 | SDVTList Tys = DAG.getVTList(VT, MVT::Other); | |||
| 9031 | SDValue Ops[] = {DAG.getEntryNode(), CP}; | |||
| 9032 | MachinePointerInfo MPI = | |||
| 9033 | MachinePointerInfo::getConstantPool(DAG.getMachineFunction()); | |||
| 9034 | return DAG.getMemIntrinsicNode(X86ISD::VBROADCAST_LOAD, dl, Tys, Ops, CVT, | |||
| 9035 | MPI, Alignment, MachineMemOperand::MOLoad); | |||
| 9036 | } | |||
| 9037 | } | |||
| 9038 | ||||
| 9039 | // Handle AVX2 in-register broadcasts. | |||
| 9040 | if (!IsLoad && Subtarget.hasInt256() && | |||
| 9041 | (ScalarSize == 32 || (IsGE256 && ScalarSize == 64))) | |||
| 9042 | return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld); | |||
| 9043 | ||||
| 9044 | // The scalar source must be a normal load. | |||
| 9045 | if (!IsLoad) | |||
| 9046 | return SDValue(); | |||
| 9047 | ||||
| 9048 | // Make sure the non-chain result is only used by this build vector. | |||
| 9049 | if (!Ld->hasNUsesOfValue(NumElts - NumUndefElts, 0)) | |||
| 9050 | return SDValue(); | |||
| 9051 | ||||
| 9052 | if (ScalarSize == 32 || (IsGE256 && ScalarSize == 64) || | |||
| 9053 | (Subtarget.hasVLX() && ScalarSize == 64)) { | |||
| 9054 | auto *LN = cast<LoadSDNode>(Ld); | |||
| 9055 | SDVTList Tys = DAG.getVTList(VT, MVT::Other); | |||
| 9056 | SDValue Ops[] = {LN->getChain(), LN->getBasePtr()}; | |||
| 9057 | SDValue BCast = | |||
| 9058 | DAG.getMemIntrinsicNode(X86ISD::VBROADCAST_LOAD, dl, Tys, Ops, | |||
| 9059 | LN->getMemoryVT(), LN->getMemOperand()); | |||
| 9060 | DAG.ReplaceAllUsesOfValueWith(SDValue(LN, 1), BCast.getValue(1)); | |||
| 9061 | return BCast; | |||
| 9062 | } | |||
| 9063 | ||||
| 9064 | // The integer check is needed for the 64-bit into 128-bit so it doesn't match | |||
| 9065 | // double since there is no vbroadcastsd xmm | |||
| 9066 | if (Subtarget.hasInt256() && Ld.getValueType().isInteger() && | |||
| 9067 | (ScalarSize == 8 || ScalarSize == 16 || ScalarSize == 64)) { | |||
| 9068 | auto *LN = cast<LoadSDNode>(Ld); | |||
| 9069 | SDVTList Tys = DAG.getVTList(VT, MVT::Other); | |||
| 9070 | SDValue Ops[] = {LN->getChain(), LN->getBasePtr()}; | |||
| 9071 | SDValue BCast = | |||
| 9072 | DAG.getMemIntrinsicNode(X86ISD::VBROADCAST_LOAD, dl, Tys, Ops, | |||
| 9073 | LN->getMemoryVT(), LN->getMemOperand()); | |||
| 9074 | DAG.ReplaceAllUsesOfValueWith(SDValue(LN, 1), BCast.getValue(1)); | |||
| 9075 | return BCast; | |||
| 9076 | } | |||
| 9077 | ||||
| 9078 | // Unsupported broadcast. | |||
| 9079 | return SDValue(); | |||
| 9080 | } | |||
| 9081 | ||||
| 9082 | /// For an EXTRACT_VECTOR_ELT with a constant index return the real | |||
| 9083 | /// underlying vector and index. | |||
| 9084 | /// | |||
| 9085 | /// Modifies \p ExtractedFromVec to the real vector and returns the real | |||
| 9086 | /// index. | |||
| 9087 | static int getUnderlyingExtractedFromVec(SDValue &ExtractedFromVec, | |||
| 9088 | SDValue ExtIdx) { | |||
| 9089 | int Idx = cast<ConstantSDNode>(ExtIdx)->getZExtValue(); | |||
| 9090 | if (!isa<ShuffleVectorSDNode>(ExtractedFromVec)) | |||
| 9091 | return Idx; | |||
| 9092 | ||||
| 9093 | // For 256-bit vectors, LowerEXTRACT_VECTOR_ELT_SSE4 may have already | |||
| 9094 | // lowered this: | |||
| 9095 | // (extract_vector_elt (v8f32 %1), Constant<6>) | |||
| 9096 | // to: | |||
| 9097 | // (extract_vector_elt (vector_shuffle<2,u,u,u> | |||
| 9098 | // (extract_subvector (v8f32 %0), Constant<4>), | |||
| 9099 | // undef) | |||
| 9100 | // Constant<0>) | |||
| 9101 | // In this case the vector is the extract_subvector expression and the index | |||
| 9102 | // is 2, as specified by the shuffle. | |||
| 9103 | ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(ExtractedFromVec); | |||
| 9104 | SDValue ShuffleVec = SVOp->getOperand(0); | |||
| 9105 | MVT ShuffleVecVT = ShuffleVec.getSimpleValueType(); | |||
| 9106 | assert(ShuffleVecVT.getVectorElementType() ==((void)0) | |||
| 9107 | ExtractedFromVec.getSimpleValueType().getVectorElementType())((void)0); | |||
| 9108 | ||||
| 9109 | int ShuffleIdx = SVOp->getMaskElt(Idx); | |||
| 9110 | if (isUndefOrInRange(ShuffleIdx, 0, ShuffleVecVT.getVectorNumElements())) { | |||
| 9111 | ExtractedFromVec = ShuffleVec; | |||
| 9112 | return ShuffleIdx; | |||
| 9113 | } | |||
| 9114 | return Idx; | |||
| 9115 | } | |||
| 9116 | ||||
| 9117 | static SDValue buildFromShuffleMostly(SDValue Op, SelectionDAG &DAG) { | |||
| 9118 | MVT VT = Op.getSimpleValueType(); | |||
| 9119 | ||||
| 9120 | // Skip if insert_vec_elt is not supported. | |||
| 9121 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
| 9122 | if (!TLI.isOperationLegalOrCustom(ISD::INSERT_VECTOR_ELT, VT)) | |||
| 9123 | return SDValue(); | |||
| 9124 | ||||
| 9125 | SDLoc DL(Op); | |||
| 9126 | unsigned NumElems = Op.getNumOperands(); | |||
| 9127 | ||||
| 9128 | SDValue VecIn1; | |||
| 9129 | SDValue VecIn2; | |||
| 9130 | SmallVector<unsigned, 4> InsertIndices; | |||
| 9131 | SmallVector<int, 8> Mask(NumElems, -1); | |||
| 9132 | ||||
| 9133 | for (unsigned i = 0; i != NumElems; ++i) { | |||
| 9134 | unsigned Opc = Op.getOperand(i).getOpcode(); | |||
| 9135 | ||||
| 9136 | if (Opc == ISD::UNDEF) | |||
| 9137 | continue; | |||
| 9138 | ||||
| 9139 | if (Opc != ISD::EXTRACT_VECTOR_ELT) { | |||
| 9140 | // Quit if more than 1 elements need inserting. | |||
| 9141 | if (InsertIndices.size() > 1) | |||
| 9142 | return SDValue(); | |||
| 9143 | ||||
| 9144 | InsertIndices.push_back(i); | |||
| 9145 | continue; | |||
| 9146 | } | |||
| 9147 | ||||
| 9148 | SDValue ExtractedFromVec = Op.getOperand(i).getOperand(0); | |||
| 9149 | SDValue ExtIdx = Op.getOperand(i).getOperand(1); | |||
| 9150 | ||||
| 9151 | // Quit if non-constant index. | |||
| 9152 | if (!isa<ConstantSDNode>(ExtIdx)) | |||
| 9153 | return SDValue(); | |||
| 9154 | int Idx = getUnderlyingExtractedFromVec(ExtractedFromVec, ExtIdx); | |||
| 9155 | ||||
| 9156 | // Quit if extracted from vector of different type. | |||
| 9157 | if (ExtractedFromVec.getValueType() != VT) | |||
| 9158 | return SDValue(); | |||
| 9159 | ||||
| 9160 | if (!VecIn1.getNode()) | |||
| 9161 | VecIn1 = ExtractedFromVec; | |||
| 9162 | else if (VecIn1 != ExtractedFromVec) { | |||
| 9163 | if (!VecIn2.getNode()) | |||
| 9164 | VecIn2 = ExtractedFromVec; | |||
| 9165 | else if (VecIn2 != ExtractedFromVec) | |||
| 9166 | // Quit if more than 2 vectors to shuffle | |||
| 9167 | return SDValue(); | |||
| 9168 | } | |||
| 9169 | ||||
| 9170 | if (ExtractedFromVec == VecIn1) | |||
| 9171 | Mask[i] = Idx; | |||
| 9172 | else if (ExtractedFromVec == VecIn2) | |||
| 9173 | Mask[i] = Idx + NumElems; | |||
| 9174 | } | |||
| 9175 | ||||
| 9176 | if (!VecIn1.getNode()) | |||
| 9177 | return SDValue(); | |||
| 9178 | ||||
| 9179 | VecIn2 = VecIn2.getNode() ? VecIn2 : DAG.getUNDEF(VT); | |||
| 9180 | SDValue NV = DAG.getVectorShuffle(VT, DL, VecIn1, VecIn2, Mask); | |||
| 9181 | ||||
| 9182 | for (unsigned Idx : InsertIndices) | |||
| 9183 | NV = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, VT, NV, Op.getOperand(Idx), | |||
| 9184 | DAG.getIntPtrConstant(Idx, DL)); | |||
| 9185 | ||||
| 9186 | return NV; | |||
| 9187 | } | |||
| 9188 | ||||
| 9189 | // Lower BUILD_VECTOR operation for v8i1 and v16i1 types. | |||
| 9190 | static SDValue LowerBUILD_VECTORvXi1(SDValue Op, SelectionDAG &DAG, | |||
| 9191 | const X86Subtarget &Subtarget) { | |||
| 9192 | ||||
| 9193 | MVT VT = Op.getSimpleValueType(); | |||
| 9194 | assert((VT.getVectorElementType() == MVT::i1) &&((void)0) | |||
| 9195 | "Unexpected type in LowerBUILD_VECTORvXi1!")((void)0); | |||
| 9196 | ||||
| 9197 | SDLoc dl(Op); | |||
| 9198 | if (ISD::isBuildVectorAllZeros(Op.getNode()) || | |||
| 9199 | ISD::isBuildVectorAllOnes(Op.getNode())) | |||
| 9200 | return Op; | |||
| 9201 | ||||
| 9202 | uint64_t Immediate = 0; | |||
| 9203 | SmallVector<unsigned, 16> NonConstIdx; | |||
| 9204 | bool IsSplat = true; | |||
| 9205 | bool HasConstElts = false; | |||
| 9206 | int SplatIdx = -1; | |||
| 9207 | for (unsigned idx = 0, e = Op.getNumOperands(); idx < e; ++idx) { | |||
| 9208 | SDValue In = Op.getOperand(idx); | |||
| 9209 | if (In.isUndef()) | |||
| 9210 | continue; | |||
| 9211 | if (auto *InC = dyn_cast<ConstantSDNode>(In)) { | |||
| 9212 | Immediate |= (InC->getZExtValue() & 0x1) << idx; | |||
| 9213 | HasConstElts = true; | |||
| 9214 | } else { | |||
| 9215 | NonConstIdx.push_back(idx); | |||
| 9216 | } | |||
| 9217 | if (SplatIdx < 0) | |||
| 9218 | SplatIdx = idx; | |||
| 9219 | else if (In != Op.getOperand(SplatIdx)) | |||
| 9220 | IsSplat = false; | |||
| 9221 | } | |||
| 9222 | ||||
| 9223 | // for splat use " (select i1 splat_elt, all-ones, all-zeroes)" | |||
| 9224 | if (IsSplat) { | |||
| 9225 | // The build_vector allows the scalar element to be larger than the vector | |||
| 9226 | // element type. We need to mask it to use as a condition unless we know | |||
| 9227 | // the upper bits are zero. | |||
| 9228 | // FIXME: Use computeKnownBits instead of checking specific opcode? | |||
| 9229 | SDValue Cond = Op.getOperand(SplatIdx); | |||
| 9230 | assert(Cond.getValueType() == MVT::i8 && "Unexpected VT!")((void)0); | |||
| 9231 | if (Cond.getOpcode() != ISD::SETCC) | |||
| 9232 | Cond = DAG.getNode(ISD::AND, dl, MVT::i8, Cond, | |||
| 9233 | DAG.getConstant(1, dl, MVT::i8)); | |||
| 9234 | ||||
| 9235 | // Perform the select in the scalar domain so we can use cmov. | |||
| 9236 | if (VT == MVT::v64i1 && !Subtarget.is64Bit()) { | |||
| 9237 | SDValue Select = DAG.getSelect(dl, MVT::i32, Cond, | |||
| 9238 | DAG.getAllOnesConstant(dl, MVT::i32), | |||
| 9239 | DAG.getConstant(0, dl, MVT::i32)); | |||
| 9240 | Select = DAG.getBitcast(MVT::v32i1, Select); | |||
| 9241 | return DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v64i1, Select, Select); | |||
| 9242 | } else { | |||
| 9243 | MVT ImmVT = MVT::getIntegerVT(std::max((unsigned)VT.getSizeInBits(), 8U)); | |||
| 9244 | SDValue Select = DAG.getSelect(dl, ImmVT, Cond, | |||
| 9245 | DAG.getAllOnesConstant(dl, ImmVT), | |||
| 9246 | DAG.getConstant(0, dl, ImmVT)); | |||
| 9247 | MVT VecVT = VT.getSizeInBits() >= 8 ? VT : MVT::v8i1; | |||
| 9248 | Select = DAG.getBitcast(VecVT, Select); | |||
| 9249 | return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, Select, | |||
| 9250 | DAG.getIntPtrConstant(0, dl)); | |||
| 9251 | } | |||
| 9252 | } | |||
| 9253 | ||||
| 9254 | // insert elements one by one | |||
| 9255 | SDValue DstVec; | |||
| 9256 | if (HasConstElts) { | |||
| 9257 | if (VT == MVT::v64i1 && !Subtarget.is64Bit()) { | |||
| 9258 | SDValue ImmL = DAG.getConstant(Lo_32(Immediate), dl, MVT::i32); | |||
| 9259 | SDValue ImmH = DAG.getConstant(Hi_32(Immediate), dl, MVT::i32); | |||
| 9260 | ImmL = DAG.getBitcast(MVT::v32i1, ImmL); | |||
| 9261 | ImmH = DAG.getBitcast(MVT::v32i1, ImmH); | |||
| 9262 | DstVec = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v64i1, ImmL, ImmH); | |||
| 9263 | } else { | |||
| 9264 | MVT ImmVT = MVT::getIntegerVT(std::max((unsigned)VT.getSizeInBits(), 8U)); | |||
| 9265 | SDValue Imm = DAG.getConstant(Immediate, dl, ImmVT); | |||
| 9266 | MVT VecVT = VT.getSizeInBits() >= 8 ? VT : MVT::v8i1; | |||
| 9267 | DstVec = DAG.getBitcast(VecVT, Imm); | |||
| 9268 | DstVec = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, DstVec, | |||
| 9269 | DAG.getIntPtrConstant(0, dl)); | |||
| 9270 | } | |||
| 9271 | } else | |||
| 9272 | DstVec = DAG.getUNDEF(VT); | |||
| 9273 | ||||
| 9274 | for (unsigned i = 0, e = NonConstIdx.size(); i != e; ++i) { | |||
| 9275 | unsigned InsertIdx = NonConstIdx[i]; | |||
| 9276 | DstVec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, DstVec, | |||
| 9277 | Op.getOperand(InsertIdx), | |||
| 9278 | DAG.getIntPtrConstant(InsertIdx, dl)); | |||
| 9279 | } | |||
| 9280 | return DstVec; | |||
| 9281 | } | |||
| 9282 | ||||
| 9283 | LLVM_ATTRIBUTE_UNUSED__attribute__((__unused__)) static bool isHorizOp(unsigned Opcode) { | |||
| 9284 | switch (Opcode) { | |||
| 9285 | case X86ISD::PACKSS: | |||
| 9286 | case X86ISD::PACKUS: | |||
| 9287 | case X86ISD::FHADD: | |||
| 9288 | case X86ISD::FHSUB: | |||
| 9289 | case X86ISD::HADD: | |||
| 9290 | case X86ISD::HSUB: | |||
| 9291 | return true; | |||
| 9292 | } | |||
| 9293 | return false; | |||
| 9294 | } | |||
| 9295 | ||||
| 9296 | /// This is a helper function of LowerToHorizontalOp(). | |||
| 9297 | /// This function checks that the build_vector \p N in input implements a | |||
| 9298 | /// 128-bit partial horizontal operation on a 256-bit vector, but that operation | |||
| 9299 | /// may not match the layout of an x86 256-bit horizontal instruction. | |||
| 9300 | /// In other words, if this returns true, then some extraction/insertion will | |||
| 9301 | /// be required to produce a valid horizontal instruction. | |||
| 9302 | /// | |||
| 9303 | /// Parameter \p Opcode defines the kind of horizontal operation to match. | |||
| 9304 | /// For example, if \p Opcode is equal to ISD::ADD, then this function | |||
| 9305 | /// checks if \p N implements a horizontal arithmetic add; if instead \p Opcode | |||
| 9306 | /// is equal to ISD::SUB, then this function checks if this is a horizontal | |||
| 9307 | /// arithmetic sub. | |||
| 9308 | /// | |||
| 9309 | /// This function only analyzes elements of \p N whose indices are | |||
| 9310 | /// in range [BaseIdx, LastIdx). | |||
| 9311 | /// | |||
| 9312 | /// TODO: This function was originally used to match both real and fake partial | |||
| 9313 | /// horizontal operations, but the index-matching logic is incorrect for that. | |||
| 9314 | /// See the corrected implementation in isHopBuildVector(). Can we reduce this | |||
| 9315 | /// code because it is only used for partial h-op matching now? | |||
| 9316 | static bool isHorizontalBinOpPart(const BuildVectorSDNode *N, unsigned Opcode, | |||
| 9317 | SelectionDAG &DAG, | |||
| 9318 | unsigned BaseIdx, unsigned LastIdx, | |||
| 9319 | SDValue &V0, SDValue &V1) { | |||
| 9320 | EVT VT = N->getValueType(0); | |||
| 9321 | assert(VT.is256BitVector() && "Only use for matching partial 256-bit h-ops")((void)0); | |||
| 9322 | assert(BaseIdx * 2 <= LastIdx && "Invalid Indices in input!")((void)0); | |||
| 9323 | assert(VT.isVector() && VT.getVectorNumElements() >= LastIdx &&((void)0) | |||
| 9324 | "Invalid Vector in input!")((void)0); | |||
| 9325 | ||||
| 9326 | bool IsCommutable = (Opcode == ISD::ADD || Opcode == ISD::FADD); | |||
| 9327 | bool CanFold = true; | |||
| 9328 | unsigned ExpectedVExtractIdx = BaseIdx; | |||
| 9329 | unsigned NumElts = LastIdx - BaseIdx; | |||
| 9330 | V0 = DAG.getUNDEF(VT); | |||
| 9331 | V1 = DAG.getUNDEF(VT); | |||
| 9332 | ||||
| 9333 | // Check if N implements a horizontal binop. | |||
| 9334 | for (unsigned i = 0, e = NumElts; i != e && CanFold; ++i) { | |||
| 9335 | SDValue Op = N->getOperand(i + BaseIdx); | |||
| 9336 | ||||
| 9337 | // Skip UNDEFs. | |||
| 9338 | if (Op->isUndef()) { | |||
| 9339 | // Update the expected vector extract index. | |||
| 9340 | if (i * 2 == NumElts) | |||
| 9341 | ExpectedVExtractIdx = BaseIdx; | |||
| 9342 | ExpectedVExtractIdx += 2; | |||
| 9343 | continue; | |||
| 9344 | } | |||
| 9345 | ||||
| 9346 | CanFold = Op->getOpcode() == Opcode && Op->hasOneUse(); | |||
| 9347 | ||||
| 9348 | if (!CanFold) | |||
| 9349 | break; | |||
| 9350 | ||||
| 9351 | SDValue Op0 = Op.getOperand(0); | |||
| 9352 | SDValue Op1 = Op.getOperand(1); | |||
| 9353 | ||||
| 9354 | // Try to match the following pattern: | |||
| 9355 | // (BINOP (extract_vector_elt A, I), (extract_vector_elt A, I+1)) | |||
| 9356 | CanFold = (Op0.getOpcode() == ISD::EXTRACT_VECTOR_ELT && | |||
| 9357 | Op1.getOpcode() == ISD::EXTRACT_VECTOR_ELT && | |||
| 9358 | Op0.getOperand(0) == Op1.getOperand(0) && | |||
| 9359 | isa<ConstantSDNode>(Op0.getOperand(1)) && | |||
| 9360 | isa<ConstantSDNode>(Op1.getOperand(1))); | |||
| 9361 | if (!CanFold) | |||
| 9362 | break; | |||
| 9363 | ||||
| 9364 | unsigned I0 = Op0.getConstantOperandVal(1); | |||
| 9365 | unsigned I1 = Op1.getConstantOperandVal(1); | |||
| 9366 | ||||
| 9367 | if (i * 2 < NumElts) { | |||
| 9368 | if (V0.isUndef()) { | |||
| 9369 | V0 = Op0.getOperand(0); | |||
| 9370 | if (V0.getValueType() != VT) | |||
| 9371 | return false; | |||
| 9372 | } | |||
| 9373 | } else { | |||
| 9374 | if (V1.isUndef()) { | |||
| 9375 | V1 = Op0.getOperand(0); | |||
| 9376 | if (V1.getValueType() != VT) | |||
| 9377 | return false; | |||
| 9378 | } | |||
| 9379 | if (i * 2 == NumElts) | |||
| 9380 | ExpectedVExtractIdx = BaseIdx; | |||
| 9381 | } | |||
| 9382 | ||||
| 9383 | SDValue Expected = (i * 2 < NumElts) ? V0 : V1; | |||
| 9384 | if (I0 == ExpectedVExtractIdx) | |||
| 9385 | CanFold = I1 == I0 + 1 && Op0.getOperand(0) == Expected; | |||
| 9386 | else if (IsCommutable && I1 == ExpectedVExtractIdx) { | |||
| 9387 | // Try to match the following dag sequence: | |||
| 9388 | // (BINOP (extract_vector_elt A, I+1), (extract_vector_elt A, I)) | |||
| 9389 | CanFold = I0 == I1 + 1 && Op1.getOperand(0) == Expected; | |||
| 9390 | } else | |||
| 9391 | CanFold = false; | |||
| 9392 | ||||
| 9393 | ExpectedVExtractIdx += 2; | |||
| 9394 | } | |||
| 9395 | ||||
| 9396 | return CanFold; | |||
| 9397 | } | |||
| 9398 | ||||
| 9399 | /// Emit a sequence of two 128-bit horizontal add/sub followed by | |||
| 9400 | /// a concat_vector. | |||
| 9401 | /// | |||
| 9402 | /// This is a helper function of LowerToHorizontalOp(). | |||
| 9403 | /// This function expects two 256-bit vectors called V0 and V1. | |||
| 9404 | /// At first, each vector is split into two separate 128-bit vectors. | |||
| 9405 | /// Then, the resulting 128-bit vectors are used to implement two | |||
| 9406 | /// horizontal binary operations. | |||
| 9407 | /// | |||
| 9408 | /// The kind of horizontal binary operation is defined by \p X86Opcode. | |||
| 9409 | /// | |||
| 9410 | /// \p Mode specifies how the 128-bit parts of V0 and V1 are passed in input to | |||
| 9411 | /// the two new horizontal binop. | |||
| 9412 | /// When Mode is set, the first horizontal binop dag node would take as input | |||
| 9413 | /// the lower 128-bit of V0 and the upper 128-bit of V0. The second | |||
| 9414 | /// horizontal binop dag node would take as input the lower 128-bit of V1 | |||
| 9415 | /// and the upper 128-bit of V1. | |||
| 9416 | /// Example: | |||
| 9417 | /// HADD V0_LO, V0_HI | |||
| 9418 | /// HADD V1_LO, V1_HI | |||
| 9419 | /// | |||
| 9420 | /// Otherwise, the first horizontal binop dag node takes as input the lower | |||
| 9421 | /// 128-bit of V0 and the lower 128-bit of V1, and the second horizontal binop | |||
| 9422 | /// dag node takes the upper 128-bit of V0 and the upper 128-bit of V1. | |||
| 9423 | /// Example: | |||
| 9424 | /// HADD V0_LO, V1_LO | |||
| 9425 | /// HADD V0_HI, V1_HI | |||
| 9426 | /// | |||
| 9427 | /// If \p isUndefLO is set, then the algorithm propagates UNDEF to the lower | |||
| 9428 | /// 128-bits of the result. If \p isUndefHI is set, then UNDEF is propagated to | |||
| 9429 | /// the upper 128-bits of the result. | |||
| 9430 | static SDValue ExpandHorizontalBinOp(const SDValue &V0, const SDValue &V1, | |||
| 9431 | const SDLoc &DL, SelectionDAG &DAG, | |||
| 9432 | unsigned X86Opcode, bool Mode, | |||
| 9433 | bool isUndefLO, bool isUndefHI) { | |||
| 9434 | MVT VT = V0.getSimpleValueType(); | |||
| 9435 | assert(VT.is256BitVector() && VT == V1.getSimpleValueType() &&((void)0) | |||
| 9436 | "Invalid nodes in input!")((void)0); | |||
| 9437 | ||||
| 9438 | unsigned NumElts = VT.getVectorNumElements(); | |||
| 9439 | SDValue V0_LO = extract128BitVector(V0, 0, DAG, DL); | |||
| 9440 | SDValue V0_HI = extract128BitVector(V0, NumElts/2, DAG, DL); | |||
| 9441 | SDValue V1_LO = extract128BitVector(V1, 0, DAG, DL); | |||
| 9442 | SDValue V1_HI = extract128BitVector(V1, NumElts/2, DAG, DL); | |||
| 9443 | MVT NewVT = V0_LO.getSimpleValueType(); | |||
| 9444 | ||||
| 9445 | SDValue LO = DAG.getUNDEF(NewVT); | |||
| 9446 | SDValue HI = DAG.getUNDEF(NewVT); | |||
| 9447 | ||||
| 9448 | if (Mode) { | |||
| 9449 | // Don't emit a horizontal binop if the result is expected to be UNDEF. | |||
| 9450 | if (!isUndefLO && !V0->isUndef()) | |||
| 9451 | LO = DAG.getNode(X86Opcode, DL, NewVT, V0_LO, V0_HI); | |||
| 9452 | if (!isUndefHI && !V1->isUndef()) | |||
| 9453 | HI = DAG.getNode(X86Opcode, DL, NewVT, V1_LO, V1_HI); | |||
| 9454 | } else { | |||
| 9455 | // Don't emit a horizontal binop if the result is expected to be UNDEF. | |||
| 9456 | if (!isUndefLO && (!V0_LO->isUndef() || !V1_LO->isUndef())) | |||
| 9457 | LO = DAG.getNode(X86Opcode, DL, NewVT, V0_LO, V1_LO); | |||
| 9458 | ||||
| 9459 | if (!isUndefHI && (!V0_HI->isUndef() || !V1_HI->isUndef())) | |||
| 9460 | HI = DAG.getNode(X86Opcode, DL, NewVT, V0_HI, V1_HI); | |||
| 9461 | } | |||
| 9462 | ||||
| 9463 | return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LO, HI); | |||
| 9464 | } | |||
| 9465 | ||||
| 9466 | /// Returns true iff \p BV builds a vector with the result equivalent to | |||
| 9467 | /// the result of ADDSUB/SUBADD operation. | |||
| 9468 | /// If true is returned then the operands of ADDSUB = Opnd0 +- Opnd1 | |||
| 9469 | /// (SUBADD = Opnd0 -+ Opnd1) operation are written to the parameters | |||
| 9470 | /// \p Opnd0 and \p Opnd1. | |||
| 9471 | static bool isAddSubOrSubAdd(const BuildVectorSDNode *BV, | |||
| 9472 | const X86Subtarget &Subtarget, SelectionDAG &DAG, | |||
| 9473 | SDValue &Opnd0, SDValue &Opnd1, | |||
| 9474 | unsigned &NumExtracts, | |||
| 9475 | bool &IsSubAdd) { | |||
| 9476 | ||||
| 9477 | MVT VT = BV->getSimpleValueType(0); | |||
| 9478 | if (!Subtarget.hasSSE3() || !VT.isFloatingPoint()) | |||
| 9479 | return false; | |||
| 9480 | ||||
| 9481 | unsigned NumElts = VT.getVectorNumElements(); | |||
| 9482 | SDValue InVec0 = DAG.getUNDEF(VT); | |||
| 9483 | SDValue InVec1 = DAG.getUNDEF(VT); | |||
| 9484 | ||||
| 9485 | NumExtracts = 0; | |||
| 9486 | ||||
| 9487 | // Odd-numbered elements in the input build vector are obtained from | |||
| 9488 | // adding/subtracting two integer/float elements. | |||
| 9489 | // Even-numbered elements in the input build vector are obtained from | |||
| 9490 | // subtracting/adding two integer/float elements. | |||
| 9491 | unsigned Opc[2] = {0, 0}; | |||
| 9492 | for (unsigned i = 0, e = NumElts; i != e; ++i) { | |||
| 9493 | SDValue Op = BV->getOperand(i); | |||
| 9494 | ||||
| 9495 | // Skip 'undef' values. | |||
| 9496 | unsigned Opcode = Op.getOpcode(); | |||
| 9497 | if (Opcode == ISD::UNDEF) | |||
| 9498 | continue; | |||
| 9499 | ||||
| 9500 | // Early exit if we found an unexpected opcode. | |||
| 9501 | if (Opcode != ISD::FADD && Opcode != ISD::FSUB) | |||
| 9502 | return false; | |||
| 9503 | ||||
| 9504 | SDValue Op0 = Op.getOperand(0); | |||
| 9505 | SDValue Op1 = Op.getOperand(1); | |||
| 9506 | ||||
| 9507 | // Try to match the following pattern: | |||
| 9508 | // (BINOP (extract_vector_elt A, i), (extract_vector_elt B, i)) | |||
| 9509 | // Early exit if we cannot match that sequence. | |||
| 9510 | if (Op0.getOpcode() != ISD::EXTRACT_VECTOR_ELT || | |||
| 9511 | Op1.getOpcode() != ISD::EXTRACT_VECTOR_ELT || | |||
| 9512 | !isa<ConstantSDNode>(Op0.getOperand(1)) || | |||
| 9513 | Op0.getOperand(1) != Op1.getOperand(1)) | |||
| 9514 | return false; | |||
| 9515 | ||||
| 9516 | unsigned I0 = Op0.getConstantOperandVal(1); | |||
| 9517 | if (I0 != i) | |||
| 9518 | return false; | |||
| 9519 | ||||
| 9520 | // We found a valid add/sub node, make sure its the same opcode as previous | |||
| 9521 | // elements for this parity. | |||
| 9522 | if (Opc[i % 2] != 0 && Opc[i % 2] != Opcode) | |||
| 9523 | return false; | |||
| 9524 | Opc[i % 2] = Opcode; | |||
| 9525 | ||||
| 9526 | // Update InVec0 and InVec1. | |||
| 9527 | if (InVec0.isUndef()) { | |||
| 9528 | InVec0 = Op0.getOperand(0); | |||
| 9529 | if (InVec0.getSimpleValueType() != VT) | |||
| 9530 | return false; | |||
| 9531 | } | |||
| 9532 | if (InVec1.isUndef()) { | |||
| 9533 | InVec1 = Op1.getOperand(0); | |||
| 9534 | if (InVec1.getSimpleValueType() != VT) | |||
| 9535 | return false; | |||
| 9536 | } | |||
| 9537 | ||||
| 9538 | // Make sure that operands in input to each add/sub node always | |||
| 9539 | // come from a same pair of vectors. | |||
| 9540 | if (InVec0 != Op0.getOperand(0)) { | |||
| 9541 | if (Opcode == ISD::FSUB) | |||
| 9542 | return false; | |||
| 9543 | ||||
| 9544 | // FADD is commutable. Try to commute the operands | |||
| 9545 | // and then test again. | |||
| 9546 | std::swap(Op0, Op1); | |||
| 9547 | if (InVec0 != Op0.getOperand(0)) | |||
| 9548 | return false; | |||
| 9549 | } | |||
| 9550 | ||||
| 9551 | if (InVec1 != Op1.getOperand(0)) | |||
| 9552 | return false; | |||
| 9553 | ||||
| 9554 | // Increment the number of extractions done. | |||
| 9555 | ++NumExtracts; | |||
| 9556 | } | |||
| 9557 | ||||
| 9558 | // Ensure we have found an opcode for both parities and that they are | |||
| 9559 | // different. Don't try to fold this build_vector into an ADDSUB/SUBADD if the | |||
| 9560 | // inputs are undef. | |||
| 9561 | if (!Opc[0] || !Opc[1] || Opc[0] == Opc[1] || | |||
| 9562 | InVec0.isUndef() || InVec1.isUndef()) | |||
| 9563 | return false; | |||
| 9564 | ||||
| 9565 | IsSubAdd = Opc[0] == ISD::FADD; | |||
| 9566 | ||||
| 9567 | Opnd0 = InVec0; | |||
| 9568 | Opnd1 = InVec1; | |||
| 9569 | return true; | |||
| 9570 | } | |||
| 9571 | ||||
| 9572 | /// Returns true if is possible to fold MUL and an idiom that has already been | |||
| 9573 | /// recognized as ADDSUB/SUBADD(\p Opnd0, \p Opnd1) into | |||
| 9574 | /// FMADDSUB/FMSUBADD(x, y, \p Opnd1). If (and only if) true is returned, the | |||
| 9575 | /// operands of FMADDSUB/FMSUBADD are written to parameters \p Opnd0, \p Opnd1, \p Opnd2. | |||
| 9576 | /// | |||
| 9577 | /// Prior to calling this function it should be known that there is some | |||
| 9578 | /// SDNode that potentially can be replaced with an X86ISD::ADDSUB operation | |||
| 9579 | /// using \p Opnd0 and \p Opnd1 as operands. Also, this method is called | |||
| 9580 | /// before replacement of such SDNode with ADDSUB operation. Thus the number | |||
| 9581 | /// of \p Opnd0 uses is expected to be equal to 2. | |||
| 9582 | /// For example, this function may be called for the following IR: | |||
| 9583 | /// %AB = fmul fast <2 x double> %A, %B | |||
| 9584 | /// %Sub = fsub fast <2 x double> %AB, %C | |||
| 9585 | /// %Add = fadd fast <2 x double> %AB, %C | |||
| 9586 | /// %Addsub = shufflevector <2 x double> %Sub, <2 x double> %Add, | |||
| 9587 | /// <2 x i32> <i32 0, i32 3> | |||
| 9588 | /// There is a def for %Addsub here, which potentially can be replaced by | |||
| 9589 | /// X86ISD::ADDSUB operation: | |||
| 9590 | /// %Addsub = X86ISD::ADDSUB %AB, %C | |||
| 9591 | /// and such ADDSUB can further be replaced with FMADDSUB: | |||
| 9592 | /// %Addsub = FMADDSUB %A, %B, %C. | |||
| 9593 | /// | |||
| 9594 | /// The main reason why this method is called before the replacement of the | |||
| 9595 | /// recognized ADDSUB idiom with ADDSUB operation is that such replacement | |||
| 9596 | /// is illegal sometimes. E.g. 512-bit ADDSUB is not available, while 512-bit | |||
| 9597 | /// FMADDSUB is. | |||
| 9598 | static bool isFMAddSubOrFMSubAdd(const X86Subtarget &Subtarget, | |||
| 9599 | SelectionDAG &DAG, | |||
| 9600 | SDValue &Opnd0, SDValue &Opnd1, SDValue &Opnd2, | |||
| 9601 | unsigned ExpectedUses) { | |||
| 9602 | if (Opnd0.getOpcode() != ISD::FMUL || | |||
| 9603 | !Opnd0->hasNUsesOfValue(ExpectedUses, 0) || !Subtarget.hasAnyFMA()) | |||
| 9604 | return false; | |||
| 9605 | ||||
| 9606 | // FIXME: These checks must match the similar ones in | |||
| 9607 | // DAGCombiner::visitFADDForFMACombine. It would be good to have one | |||
| 9608 | // function that would answer if it is Ok to fuse MUL + ADD to FMADD | |||
| 9609 | // or MUL + ADDSUB to FMADDSUB. | |||
| 9610 | const TargetOptions &Options = DAG.getTarget().Options; | |||
| 9611 | bool AllowFusion = | |||
| 9612 | (Options.AllowFPOpFusion == FPOpFusion::Fast || Options.UnsafeFPMath); | |||
| 9613 | if (!AllowFusion) | |||
| 9614 | return false; | |||
| 9615 | ||||
| 9616 | Opnd2 = Opnd1; | |||
| 9617 | Opnd1 = Opnd0.getOperand(1); | |||
| 9618 | Opnd0 = Opnd0.getOperand(0); | |||
| 9619 | ||||
| 9620 | return true; | |||
| 9621 | } | |||
| 9622 | ||||
| 9623 | /// Try to fold a build_vector that performs an 'addsub' or 'fmaddsub' or | |||
| 9624 | /// 'fsubadd' operation accordingly to X86ISD::ADDSUB or X86ISD::FMADDSUB or | |||
| 9625 | /// X86ISD::FMSUBADD node. | |||
| 9626 | static SDValue lowerToAddSubOrFMAddSub(const BuildVectorSDNode *BV, | |||
| 9627 | const X86Subtarget &Subtarget, | |||
| 9628 | SelectionDAG &DAG) { | |||
| 9629 | SDValue Opnd0, Opnd1; | |||
| 9630 | unsigned NumExtracts; | |||
| 9631 | bool IsSubAdd; | |||
| 9632 | if (!isAddSubOrSubAdd(BV, Subtarget, DAG, Opnd0, Opnd1, NumExtracts, | |||
| 9633 | IsSubAdd)) | |||
| 9634 | return SDValue(); | |||
| 9635 | ||||
| 9636 | MVT VT = BV->getSimpleValueType(0); | |||
| 9637 | SDLoc DL(BV); | |||
| 9638 | ||||
| 9639 | // Try to generate X86ISD::FMADDSUB node here. | |||
| 9640 | SDValue Opnd2; | |||
| 9641 | if (isFMAddSubOrFMSubAdd(Subtarget, DAG, Opnd0, Opnd1, Opnd2, NumExtracts)) { | |||
| 9642 | unsigned Opc = IsSubAdd ? X86ISD::FMSUBADD : X86ISD::FMADDSUB; | |||
| 9643 | return DAG.getNode(Opc, DL, VT, Opnd0, Opnd1, Opnd2); | |||
| 9644 | } | |||
| 9645 | ||||
| 9646 | // We only support ADDSUB. | |||
| 9647 | if (IsSubAdd) | |||
| 9648 | return SDValue(); | |||
| 9649 | ||||
| 9650 | // Do not generate X86ISD::ADDSUB node for 512-bit types even though | |||
| 9651 | // the ADDSUB idiom has been successfully recognized. There are no known | |||
| 9652 | // X86 targets with 512-bit ADDSUB instructions! | |||
| 9653 | // 512-bit ADDSUB idiom recognition was needed only as part of FMADDSUB idiom | |||
| 9654 | // recognition. | |||
| 9655 | if (VT.is512BitVector()) | |||
| 9656 | return SDValue(); | |||
| 9657 | ||||
| 9658 | return DAG.getNode(X86ISD::ADDSUB, DL, VT, Opnd0, Opnd1); | |||
| 9659 | } | |||
| 9660 | ||||
| 9661 | static bool isHopBuildVector(const BuildVectorSDNode *BV, SelectionDAG &DAG, | |||
| 9662 | unsigned &HOpcode, SDValue &V0, SDValue &V1) { | |||
| 9663 | // Initialize outputs to known values. | |||
| 9664 | MVT VT = BV->getSimpleValueType(0); | |||
| 9665 | HOpcode = ISD::DELETED_NODE; | |||
| 9666 | V0 = DAG.getUNDEF(VT); | |||
| 9667 | V1 = DAG.getUNDEF(VT); | |||
| 9668 | ||||
| 9669 | // x86 256-bit horizontal ops are defined in a non-obvious way. Each 128-bit | |||
| 9670 | // half of the result is calculated independently from the 128-bit halves of | |||
| 9671 | // the inputs, so that makes the index-checking logic below more complicated. | |||
| 9672 | unsigned NumElts = VT.getVectorNumElements(); | |||
| 9673 | unsigned GenericOpcode = ISD::DELETED_NODE; | |||
| 9674 | unsigned Num128BitChunks = VT.is256BitVector() ? 2 : 1; | |||
| 9675 | unsigned NumEltsIn128Bits = NumElts / Num128BitChunks; | |||
| 9676 | unsigned NumEltsIn64Bits = NumEltsIn128Bits / 2; | |||
| 9677 | for (unsigned i = 0; i != Num128BitChunks; ++i) { | |||
| 9678 | for (unsigned j = 0; j != NumEltsIn128Bits; ++j) { | |||
| 9679 | // Ignore undef elements. | |||
| 9680 | SDValue Op = BV->getOperand(i * NumEltsIn128Bits + j); | |||
| 9681 | if (Op.isUndef()) | |||
| 9682 | continue; | |||
| 9683 | ||||
| 9684 | // If there's an opcode mismatch, we're done. | |||
| 9685 | if (HOpcode != ISD::DELETED_NODE && Op.getOpcode() != GenericOpcode) | |||
| 9686 | return false; | |||
| 9687 | ||||
| 9688 | // Initialize horizontal opcode. | |||
| 9689 | if (HOpcode == ISD::DELETED_NODE) { | |||
| 9690 | GenericOpcode = Op.getOpcode(); | |||
| 9691 | switch (GenericOpcode) { | |||
| 9692 | case ISD::ADD: HOpcode = X86ISD::HADD; break; | |||
| 9693 | case ISD::SUB: HOpcode = X86ISD::HSUB; break; | |||
| 9694 | case ISD::FADD: HOpcode = X86ISD::FHADD; break; | |||
| 9695 | case ISD::FSUB: HOpcode = X86ISD::FHSUB; break; | |||
| 9696 | default: return false; | |||
| 9697 | } | |||
| 9698 | } | |||
| 9699 | ||||
| 9700 | SDValue Op0 = Op.getOperand(0); | |||
| 9701 | SDValue Op1 = Op.getOperand(1); | |||
| 9702 | if (Op0.getOpcode() != ISD::EXTRACT_VECTOR_ELT || | |||
| 9703 | Op1.getOpcode() != ISD::EXTRACT_VECTOR_ELT || | |||
| 9704 | Op0.getOperand(0) != Op1.getOperand(0) || | |||
| 9705 | !isa<ConstantSDNode>(Op0.getOperand(1)) || | |||
| 9706 | !isa<ConstantSDNode>(Op1.getOperand(1)) || !Op.hasOneUse()) | |||
| 9707 | return false; | |||
| 9708 | ||||
| 9709 | // The source vector is chosen based on which 64-bit half of the | |||
| 9710 | // destination vector is being calculated. | |||
| 9711 | if (j < NumEltsIn64Bits) { | |||
| 9712 | if (V0.isUndef()) | |||
| 9713 | V0 = Op0.getOperand(0); | |||
| 9714 | } else { | |||
| 9715 | if (V1.isUndef()) | |||
| 9716 | V1 = Op0.getOperand(0); | |||
| 9717 | } | |||
| 9718 | ||||
| 9719 | SDValue SourceVec = (j < NumEltsIn64Bits) ? V0 : V1; | |||
| 9720 | if (SourceVec != Op0.getOperand(0)) | |||
| 9721 | return false; | |||
| 9722 | ||||
| 9723 | // op (extract_vector_elt A, I), (extract_vector_elt A, I+1) | |||
| 9724 | unsigned ExtIndex0 = Op0.getConstantOperandVal(1); | |||
| 9725 | unsigned ExtIndex1 = Op1.getConstantOperandVal(1); | |||
| 9726 | unsigned ExpectedIndex = i * NumEltsIn128Bits + | |||
| 9727 | (j % NumEltsIn64Bits) * 2; | |||
| 9728 | if (ExpectedIndex == ExtIndex0 && ExtIndex1 == ExtIndex0 + 1) | |||
| 9729 | continue; | |||
| 9730 | ||||
| 9731 | // If this is not a commutative op, this does not match. | |||
| 9732 | if (GenericOpcode != ISD::ADD && GenericOpcode != ISD::FADD) | |||
| 9733 | return false; | |||
| 9734 | ||||
| 9735 | // Addition is commutative, so try swapping the extract indexes. | |||
| 9736 | // op (extract_vector_elt A, I+1), (extract_vector_elt A, I) | |||
| 9737 | if (ExpectedIndex == ExtIndex1 && ExtIndex0 == ExtIndex1 + 1) | |||
| 9738 | continue; | |||
| 9739 | ||||
| 9740 | // Extract indexes do not match horizontal requirement. | |||
| 9741 | return false; | |||
| 9742 | } | |||
| 9743 | } | |||
| 9744 | // We matched. Opcode and operands are returned by reference as arguments. | |||
| 9745 | return true; | |||
| 9746 | } | |||
| 9747 | ||||
| 9748 | static SDValue getHopForBuildVector(const BuildVectorSDNode *BV, | |||
| 9749 | SelectionDAG &DAG, unsigned HOpcode, | |||
| 9750 | SDValue V0, SDValue V1) { | |||
| 9751 | // If either input vector is not the same size as the build vector, | |||
| 9752 | // extract/insert the low bits to the correct size. | |||
| 9753 | // This is free (examples: zmm --> xmm, xmm --> ymm). | |||
| 9754 | MVT VT = BV->getSimpleValueType(0); | |||
| 9755 | unsigned Width = VT.getSizeInBits(); | |||
| 9756 | if (V0.getValueSizeInBits() > Width) | |||
| 9757 | V0 = extractSubVector(V0, 0, DAG, SDLoc(BV), Width); | |||
| 9758 | else if (V0.getValueSizeInBits() < Width) | |||
| 9759 | V0 = insertSubVector(DAG.getUNDEF(VT), V0, 0, DAG, SDLoc(BV), Width); | |||
| 9760 | ||||
| 9761 | if (V1.getValueSizeInBits() > Width) | |||
| 9762 | V1 = extractSubVector(V1, 0, DAG, SDLoc(BV), Width); | |||
| 9763 | else if (V1.getValueSizeInBits() < Width) | |||
| 9764 | V1 = insertSubVector(DAG.getUNDEF(VT), V1, 0, DAG, SDLoc(BV), Width); | |||
| 9765 | ||||
| 9766 | unsigned NumElts = VT.getVectorNumElements(); | |||
| 9767 | APInt DemandedElts = APInt::getAllOnesValue(NumElts); | |||
| 9768 | for (unsigned i = 0; i != NumElts; ++i) | |||
| 9769 | if (BV->getOperand(i).isUndef()) | |||
| 9770 | DemandedElts.clearBit(i); | |||
| 9771 | ||||
| 9772 | // If we don't need the upper xmm, then perform as a xmm hop. | |||
| 9773 | unsigned HalfNumElts = NumElts / 2; | |||
| 9774 | if (VT.is256BitVector() && DemandedElts.lshr(HalfNumElts) == 0) { | |||
| 9775 | MVT HalfVT = VT.getHalfNumVectorElementsVT(); | |||
| 9776 | V0 = extractSubVector(V0, 0, DAG, SDLoc(BV), 128); | |||
| 9777 | V1 = extractSubVector(V1, 0, DAG, SDLoc(BV), 128); | |||
| 9778 | SDValue Half = DAG.getNode(HOpcode, SDLoc(BV), HalfVT, V0, V1); | |||
| 9779 | return insertSubVector(DAG.getUNDEF(VT), Half, 0, DAG, SDLoc(BV), 256); | |||
| 9780 | } | |||
| 9781 | ||||
| 9782 | return DAG.getNode(HOpcode, SDLoc(BV), VT, V0, V1); | |||
| 9783 | } | |||
| 9784 | ||||
| 9785 | /// Lower BUILD_VECTOR to a horizontal add/sub operation if possible. | |||
| 9786 | static SDValue LowerToHorizontalOp(const BuildVectorSDNode *BV, | |||
| 9787 | const X86Subtarget &Subtarget, | |||
| 9788 | SelectionDAG &DAG) { | |||
| 9789 | // We need at least 2 non-undef elements to make this worthwhile by default. | |||
| 9790 | unsigned NumNonUndefs = | |||
| 9791 | count_if(BV->op_values(), [](SDValue V) { return !V.isUndef(); }); | |||
| 9792 | if (NumNonUndefs < 2) | |||
| 9793 | return SDValue(); | |||
| 9794 | ||||
| 9795 | // There are 4 sets of horizontal math operations distinguished by type: | |||
| 9796 | // int/FP at 128-bit/256-bit. Each type was introduced with a different | |||
| 9797 | // subtarget feature. Try to match those "native" patterns first. | |||
| 9798 | MVT VT = BV->getSimpleValueType(0); | |||
| 9799 | if (((VT == MVT::v4f32 || VT == MVT::v2f64) && Subtarget.hasSSE3()) || | |||
| 9800 | ((VT == MVT::v8i16 || VT == MVT::v4i32) && Subtarget.hasSSSE3()) || | |||
| 9801 | ((VT == MVT::v8f32 || VT == MVT::v4f64) && Subtarget.hasAVX()) || | |||
| 9802 | ((VT == MVT::v16i16 || VT == MVT::v8i32) && Subtarget.hasAVX2())) { | |||
| 9803 | unsigned HOpcode; | |||
| 9804 | SDValue V0, V1; | |||
| 9805 | if (isHopBuildVector(BV, DAG, HOpcode, V0, V1)) | |||
| 9806 | return getHopForBuildVector(BV, DAG, HOpcode, V0, V1); | |||
| 9807 | } | |||
| 9808 | ||||
| 9809 | // Try harder to match 256-bit ops by using extract/concat. | |||
| 9810 | if (!Subtarget.hasAVX() || !VT.is256BitVector()) | |||
| 9811 | return SDValue(); | |||
| 9812 | ||||
| 9813 | // Count the number of UNDEF operands in the build_vector in input. | |||
| 9814 | unsigned NumElts = VT.getVectorNumElements(); | |||
| 9815 | unsigned Half = NumElts / 2; | |||
| 9816 | unsigned NumUndefsLO = 0; | |||
| 9817 | unsigned NumUndefsHI = 0; | |||
| 9818 | for (unsigned i = 0, e = Half; i != e; ++i) | |||
| 9819 | if (BV->getOperand(i)->isUndef()) | |||
| 9820 | NumUndefsLO++; | |||
| 9821 | ||||
| 9822 | for (unsigned i = Half, e = NumElts; i != e; ++i) | |||
| 9823 | if (BV->getOperand(i)->isUndef()) | |||
| 9824 | NumUndefsHI++; | |||
| 9825 | ||||
| 9826 | SDLoc DL(BV); | |||
| 9827 | SDValue InVec0, InVec1; | |||
| 9828 | if (VT == MVT::v8i32 || VT == MVT::v16i16) { | |||
| 9829 | SDValue InVec2, InVec3; | |||
| 9830 | unsigned X86Opcode; | |||
| 9831 | bool CanFold = true; | |||
| 9832 | ||||
| 9833 | if (isHorizontalBinOpPart(BV, ISD::ADD, DAG, 0, Half, InVec0, InVec1) && | |||
| 9834 | isHorizontalBinOpPart(BV, ISD::ADD, DAG, Half, NumElts, InVec2, | |||
| 9835 | InVec3) && | |||
| 9836 | ((InVec0.isUndef() || InVec2.isUndef()) || InVec0 == InVec2) && | |||
| 9837 | ((InVec1.isUndef() || InVec3.isUndef()) || InVec1 == InVec3)) | |||
| 9838 | X86Opcode = X86ISD::HADD; | |||
| 9839 | else if (isHorizontalBinOpPart(BV, ISD::SUB, DAG, 0, Half, InVec0, | |||
| 9840 | InVec1) && | |||
| 9841 | isHorizontalBinOpPart(BV, ISD::SUB, DAG, Half, NumElts, InVec2, | |||
| 9842 | InVec3) && | |||
| 9843 | ((InVec0.isUndef() || InVec2.isUndef()) || InVec0 == InVec2) && | |||
| 9844 | ((InVec1.isUndef() || InVec3.isUndef()) || InVec1 == InVec3)) | |||
| 9845 | X86Opcode = X86ISD::HSUB; | |||
| 9846 | else | |||
| 9847 | CanFold = false; | |||
| 9848 | ||||
| 9849 | if (CanFold) { | |||
| 9850 | // Do not try to expand this build_vector into a pair of horizontal | |||
| 9851 | // add/sub if we can emit a pair of scalar add/sub. | |||
| 9852 | if (NumUndefsLO + 1 == Half || NumUndefsHI + 1 == Half) | |||
| 9853 | return SDValue(); | |||
| 9854 | ||||
| 9855 | // Convert this build_vector into a pair of horizontal binops followed by | |||
| 9856 | // a concat vector. We must adjust the outputs from the partial horizontal | |||
| 9857 | // matching calls above to account for undefined vector halves. | |||
| 9858 | SDValue V0 = InVec0.isUndef() ? InVec2 : InVec0; | |||
| 9859 | SDValue V1 = InVec1.isUndef() ? InVec3 : InVec1; | |||
| 9860 | assert((!V0.isUndef() || !V1.isUndef()) && "Horizontal-op of undefs?")((void)0); | |||
| 9861 | bool isUndefLO = NumUndefsLO == Half; | |||
| 9862 | bool isUndefHI = NumUndefsHI == Half; | |||
| 9863 | return ExpandHorizontalBinOp(V0, V1, DL, DAG, X86Opcode, false, isUndefLO, | |||
| 9864 | isUndefHI); | |||
| 9865 | } | |||
| 9866 | } | |||
| 9867 | ||||
| 9868 | if (VT == MVT::v8f32 || VT == MVT::v4f64 || VT == MVT::v8i32 || | |||
| 9869 | VT == MVT::v16i16) { | |||
| 9870 | unsigned X86Opcode; | |||
| 9871 | if (isHorizontalBinOpPart(BV, ISD::ADD, DAG, 0, NumElts, InVec0, InVec1)) | |||
| 9872 | X86Opcode = X86ISD::HADD; | |||
| 9873 | else if (isHorizontalBinOpPart(BV, ISD::SUB, DAG, 0, NumElts, InVec0, | |||
| 9874 | InVec1)) | |||
| 9875 | X86Opcode = X86ISD::HSUB; | |||
| 9876 | else if (isHorizontalBinOpPart(BV, ISD::FADD, DAG, 0, NumElts, InVec0, | |||
| 9877 | InVec1)) | |||
| 9878 | X86Opcode = X86ISD::FHADD; | |||
| 9879 | else if (isHorizontalBinOpPart(BV, ISD::FSUB, DAG, 0, NumElts, InVec0, | |||
| 9880 | InVec1)) | |||
| 9881 | X86Opcode = X86ISD::FHSUB; | |||
| 9882 | else | |||
| 9883 | return SDValue(); | |||
| 9884 | ||||
| 9885 | // Don't try to expand this build_vector into a pair of horizontal add/sub | |||
| 9886 | // if we can simply emit a pair of scalar add/sub. | |||
| 9887 | if (NumUndefsLO + 1 == Half || NumUndefsHI + 1 == Half) | |||
| 9888 | return SDValue(); | |||
| 9889 | ||||
| 9890 | // Convert this build_vector into two horizontal add/sub followed by | |||
| 9891 | // a concat vector. | |||
| 9892 | bool isUndefLO = NumUndefsLO == Half; | |||
| 9893 | bool isUndefHI = NumUndefsHI == Half; | |||
| 9894 | return ExpandHorizontalBinOp(InVec0, InVec1, DL, DAG, X86Opcode, true, | |||
| 9895 | isUndefLO, isUndefHI); | |||
| 9896 | } | |||
| 9897 | ||||
| 9898 | return SDValue(); | |||
| 9899 | } | |||
| 9900 | ||||
| 9901 | static SDValue LowerShift(SDValue Op, const X86Subtarget &Subtarget, | |||
| 9902 | SelectionDAG &DAG); | |||
| 9903 | ||||
| 9904 | /// If a BUILD_VECTOR's source elements all apply the same bit operation and | |||
| 9905 | /// one of their operands is constant, lower to a pair of BUILD_VECTOR and | |||
| 9906 | /// just apply the bit to the vectors. | |||
| 9907 | /// NOTE: Its not in our interest to start make a general purpose vectorizer | |||
| 9908 | /// from this, but enough scalar bit operations are created from the later | |||
| 9909 | /// legalization + scalarization stages to need basic support. | |||
| 9910 | static SDValue lowerBuildVectorToBitOp(BuildVectorSDNode *Op, | |||
| 9911 | const X86Subtarget &Subtarget, | |||
| 9912 | SelectionDAG &DAG) { | |||
| 9913 | SDLoc DL(Op); | |||
| 9914 | MVT VT = Op->getSimpleValueType(0); | |||
| 9915 | unsigned NumElems = VT.getVectorNumElements(); | |||
| 9916 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
| 9917 | ||||
| 9918 | // Check that all elements have the same opcode. | |||
| 9919 | // TODO: Should we allow UNDEFS and if so how many? | |||
| 9920 | unsigned Opcode = Op->getOperand(0).getOpcode(); | |||
| 9921 | for (unsigned i = 1; i < NumElems; ++i) | |||
| 9922 | if (Opcode != Op->getOperand(i).getOpcode()) | |||
| 9923 | return SDValue(); | |||
| 9924 | ||||
| 9925 | // TODO: We may be able to add support for other Ops (ADD/SUB + shifts). | |||
| 9926 | bool IsShift = false; | |||
| 9927 | switch (Opcode) { | |||
| 9928 | default: | |||
| 9929 | return SDValue(); | |||
| 9930 | case ISD::SHL: | |||
| 9931 | case ISD::SRL: | |||
| 9932 | case ISD::SRA: | |||
| 9933 | IsShift = true; | |||
| 9934 | break; | |||
| 9935 | case ISD::AND: | |||
| 9936 | case ISD::XOR: | |||
| 9937 | case ISD::OR: | |||
| 9938 | // Don't do this if the buildvector is a splat - we'd replace one | |||
| 9939 | // constant with an entire vector. | |||
| 9940 | if (Op->getSplatValue()) | |||
| 9941 | return SDValue(); | |||
| 9942 | if (!TLI.isOperationLegalOrPromote(Opcode, VT)) | |||
| 9943 | return SDValue(); | |||
| 9944 | break; | |||
| 9945 | } | |||
| 9946 | ||||
| 9947 | SmallVector<SDValue, 4> LHSElts, RHSElts; | |||
| 9948 | for (SDValue Elt : Op->ops()) { | |||
| 9949 | SDValue LHS = Elt.getOperand(0); | |||
| 9950 | SDValue RHS = Elt.getOperand(1); | |||
| 9951 | ||||
| 9952 | // We expect the canonicalized RHS operand to be the constant. | |||
| 9953 | if (!isa<ConstantSDNode>(RHS)) | |||
| 9954 | return SDValue(); | |||
| 9955 | ||||
| 9956 | // Extend shift amounts. | |||
| 9957 | if (RHS.getValueSizeInBits() != VT.getScalarSizeInBits()) { | |||
| 9958 | if (!IsShift) | |||
| 9959 | return SDValue(); | |||
| 9960 | RHS = DAG.getZExtOrTrunc(RHS, DL, VT.getScalarType()); | |||
| 9961 | } | |||
| 9962 | ||||
| 9963 | LHSElts.push_back(LHS); | |||
| 9964 | RHSElts.push_back(RHS); | |||
| 9965 | } | |||
| 9966 | ||||
| 9967 | // Limit to shifts by uniform immediates. | |||
| 9968 | // TODO: Only accept vXi8/vXi64 special cases? | |||
| 9969 | // TODO: Permit non-uniform XOP/AVX2/MULLO cases? | |||
| 9970 | if (IsShift && any_of(RHSElts, [&](SDValue V) { return RHSElts[0] != V; })) | |||
| 9971 | return SDValue(); | |||
| 9972 | ||||
| 9973 | SDValue LHS = DAG.getBuildVector(VT, DL, LHSElts); | |||
| 9974 | SDValue RHS = DAG.getBuildVector(VT, DL, RHSElts); | |||
| 9975 | SDValue Res = DAG.getNode(Opcode, DL, VT, LHS, RHS); | |||
| 9976 | ||||
| 9977 | if (!IsShift) | |||
| 9978 | return Res; | |||
| 9979 | ||||
| 9980 | // Immediately lower the shift to ensure the constant build vector doesn't | |||
| 9981 | // get converted to a constant pool before the shift is lowered. | |||
| 9982 | return LowerShift(Res, Subtarget, DAG); | |||
| 9983 | } | |||
| 9984 | ||||
| 9985 | /// Create a vector constant without a load. SSE/AVX provide the bare minimum | |||
| 9986 | /// functionality to do this, so it's all zeros, all ones, or some derivation | |||
| 9987 | /// that is cheap to calculate. | |||
| 9988 | static SDValue materializeVectorConstant(SDValue Op, SelectionDAG &DAG, | |||
| 9989 | const X86Subtarget &Subtarget) { | |||
| 9990 | SDLoc DL(Op); | |||
| 9991 | MVT VT = Op.getSimpleValueType(); | |||
| 9992 | ||||
| 9993 | // Vectors containing all zeros can be matched by pxor and xorps. | |||
| 9994 | if (ISD::isBuildVectorAllZeros(Op.getNode())) | |||
| 9995 | return Op; | |||
| 9996 | ||||
| 9997 | // Vectors containing all ones can be matched by pcmpeqd on 128-bit width | |||
| 9998 | // vectors or broken into v4i32 operations on 256-bit vectors. AVX2 can use | |||
| 9999 | // vpcmpeqd on 256-bit vectors. | |||
| 10000 | if (Subtarget.hasSSE2() && ISD::isBuildVectorAllOnes(Op.getNode())) { | |||
| 10001 | if (VT == MVT::v4i32 || VT == MVT::v8i32 || VT == MVT::v16i32) | |||
| 10002 | return Op; | |||
| 10003 | ||||
| 10004 | return getOnesVector(VT, DAG, DL); | |||
| 10005 | } | |||
| 10006 | ||||
| 10007 | return SDValue(); | |||
| 10008 | } | |||
| 10009 | ||||
| 10010 | /// Look for opportunities to create a VPERMV/VPERMILPV/PSHUFB variable permute | |||
| 10011 | /// from a vector of source values and a vector of extraction indices. | |||
| 10012 | /// The vectors might be manipulated to match the type of the permute op. | |||
| 10013 | static SDValue createVariablePermute(MVT VT, SDValue SrcVec, SDValue IndicesVec, | |||
| 10014 | SDLoc &DL, SelectionDAG &DAG, | |||
| 10015 | const X86Subtarget &Subtarget) { | |||
| 10016 | MVT ShuffleVT = VT; | |||
| 10017 | EVT IndicesVT = EVT(VT).changeVectorElementTypeToInteger(); | |||
| 10018 | unsigned NumElts = VT.getVectorNumElements(); | |||
| 10019 | unsigned SizeInBits = VT.getSizeInBits(); | |||
| 10020 | ||||
| 10021 | // Adjust IndicesVec to match VT size. | |||
| 10022 | assert(IndicesVec.getValueType().getVectorNumElements() >= NumElts &&((void)0) | |||
| 10023 | "Illegal variable permute mask size")((void)0); | |||
| 10024 | if (IndicesVec.getValueType().getVectorNumElements() > NumElts) { | |||
| 10025 | // Narrow/widen the indices vector to the correct size. | |||
| 10026 | if (IndicesVec.getValueSizeInBits() > SizeInBits) | |||
| 10027 | IndicesVec = extractSubVector(IndicesVec, 0, DAG, SDLoc(IndicesVec), | |||
| 10028 | NumElts * VT.getScalarSizeInBits()); | |||
| 10029 | else if (IndicesVec.getValueSizeInBits() < SizeInBits) | |||
| 10030 | IndicesVec = widenSubVector(IndicesVec, false, Subtarget, DAG, | |||
| 10031 | SDLoc(IndicesVec), SizeInBits); | |||
| 10032 | // Zero-extend the index elements within the vector. | |||
| 10033 | if (IndicesVec.getValueType().getVectorNumElements() > NumElts) | |||
| 10034 | IndicesVec = DAG.getNode(ISD::ZERO_EXTEND_VECTOR_INREG, SDLoc(IndicesVec), | |||
| 10035 | IndicesVT, IndicesVec); | |||
| 10036 | } | |||
| 10037 | IndicesVec = DAG.getZExtOrTrunc(IndicesVec, SDLoc(IndicesVec), IndicesVT); | |||
| 10038 | ||||
| 10039 | // Handle SrcVec that don't match VT type. | |||
| 10040 | if (SrcVec.getValueSizeInBits() != SizeInBits) { | |||
| 10041 | if ((SrcVec.getValueSizeInBits() % SizeInBits) == 0) { | |||
| 10042 | // Handle larger SrcVec by treating it as a larger permute. | |||
| 10043 | unsigned Scale = SrcVec.getValueSizeInBits() / SizeInBits; | |||
| 10044 | VT = MVT::getVectorVT(VT.getScalarType(), Scale * NumElts); | |||
| 10045 | IndicesVT = EVT(VT).changeVectorElementTypeToInteger(); | |||
| 10046 | IndicesVec = widenSubVector(IndicesVT.getSimpleVT(), IndicesVec, false, | |||
| 10047 | Subtarget, DAG, SDLoc(IndicesVec)); | |||
| 10048 | SDValue NewSrcVec = | |||
| 10049 | createVariablePermute(VT, SrcVec, IndicesVec, DL, DAG, Subtarget); | |||
| 10050 | if (NewSrcVec) | |||
| 10051 | return extractSubVector(NewSrcVec, 0, DAG, DL, SizeInBits); | |||
| 10052 | return SDValue(); | |||
| 10053 | } else if (SrcVec.getValueSizeInBits() < SizeInBits) { | |||
| 10054 | // Widen smaller SrcVec to match VT. | |||
| 10055 | SrcVec = widenSubVector(VT, SrcVec, false, Subtarget, DAG, SDLoc(SrcVec)); | |||
| 10056 | } else | |||
| 10057 | return SDValue(); | |||
| 10058 | } | |||
| 10059 | ||||
| 10060 | auto ScaleIndices = [&DAG](SDValue Idx, uint64_t Scale) { | |||
| 10061 | assert(isPowerOf2_64(Scale) && "Illegal variable permute shuffle scale")((void)0); | |||
| 10062 | EVT SrcVT = Idx.getValueType(); | |||
| 10063 | unsigned NumDstBits = SrcVT.getScalarSizeInBits() / Scale; | |||
| 10064 | uint64_t IndexScale = 0; | |||
| 10065 | uint64_t IndexOffset = 0; | |||
| 10066 | ||||
| 10067 | // If we're scaling a smaller permute op, then we need to repeat the | |||
| 10068 | // indices, scaling and offsetting them as well. | |||
| 10069 | // e.g. v4i32 -> v16i8 (Scale = 4) | |||
| 10070 | // IndexScale = v4i32 Splat(4 << 24 | 4 << 16 | 4 << 8 | 4) | |||
| 10071 | // IndexOffset = v4i32 Splat(3 << 24 | 2 << 16 | 1 << 8 | 0) | |||
| 10072 | for (uint64_t i = 0; i != Scale; ++i) { | |||
| 10073 | IndexScale |= Scale << (i * NumDstBits); | |||
| 10074 | IndexOffset |= i << (i * NumDstBits); | |||
| 10075 | } | |||
| 10076 | ||||
| 10077 | Idx = DAG.getNode(ISD::MUL, SDLoc(Idx), SrcVT, Idx, | |||
| 10078 | DAG.getConstant(IndexScale, SDLoc(Idx), SrcVT)); | |||
| 10079 | Idx = DAG.getNode(ISD::ADD, SDLoc(Idx), SrcVT, Idx, | |||
| 10080 | DAG.getConstant(IndexOffset, SDLoc(Idx), SrcVT)); | |||
| 10081 | return Idx; | |||
| 10082 | }; | |||
| 10083 | ||||
| 10084 | unsigned Opcode = 0; | |||
| 10085 | switch (VT.SimpleTy) { | |||
| 10086 | default: | |||
| 10087 | break; | |||
| 10088 | case MVT::v16i8: | |||
| 10089 | if (Subtarget.hasSSSE3()) | |||
| 10090 | Opcode = X86ISD::PSHUFB; | |||
| 10091 | break; | |||
| 10092 | case MVT::v8i16: | |||
| 10093 | if (Subtarget.hasVLX() && Subtarget.hasBWI()) | |||
| 10094 | Opcode = X86ISD::VPERMV; | |||
| 10095 | else if (Subtarget.hasSSSE3()) { | |||
| 10096 | Opcode = X86ISD::PSHUFB; | |||
| 10097 | ShuffleVT = MVT::v16i8; | |||
| 10098 | } | |||
| 10099 | break; | |||
| 10100 | case MVT::v4f32: | |||
| 10101 | case MVT::v4i32: | |||
| 10102 | if (Subtarget.hasAVX()) { | |||
| 10103 | Opcode = X86ISD::VPERMILPV; | |||
| 10104 | ShuffleVT = MVT::v4f32; | |||
| 10105 | } else if (Subtarget.hasSSSE3()) { | |||
| 10106 | Opcode = X86ISD::PSHUFB; | |||
| 10107 | ShuffleVT = MVT::v16i8; | |||
| 10108 | } | |||
| 10109 | break; | |||
| 10110 | case MVT::v2f64: | |||
| 10111 | case MVT::v2i64: | |||
| 10112 | if (Subtarget.hasAVX()) { | |||
| 10113 | // VPERMILPD selects using bit#1 of the index vector, so scale IndicesVec. | |||
| 10114 | IndicesVec = DAG.getNode(ISD::ADD, DL, IndicesVT, IndicesVec, IndicesVec); | |||
| 10115 | Opcode = X86ISD::VPERMILPV; | |||
| 10116 | ShuffleVT = MVT::v2f64; | |||
| 10117 | } else if (Subtarget.hasSSE41()) { | |||
| 10118 | // SSE41 can compare v2i64 - select between indices 0 and 1. | |||
| 10119 | return DAG.getSelectCC( | |||
| 10120 | DL, IndicesVec, | |||
| 10121 | getZeroVector(IndicesVT.getSimpleVT(), Subtarget, DAG, DL), | |||
| 10122 | DAG.getVectorShuffle(VT, DL, SrcVec, SrcVec, {0, 0}), | |||
| 10123 | DAG.getVectorShuffle(VT, DL, SrcVec, SrcVec, {1, 1}), | |||
| 10124 | ISD::CondCode::SETEQ); | |||
| 10125 | } | |||
| 10126 | break; | |||
| 10127 | case MVT::v32i8: | |||
| 10128 | if (Subtarget.hasVLX() && Subtarget.hasVBMI()) | |||
| 10129 | Opcode = X86ISD::VPERMV; | |||
| 10130 | else if (Subtarget.hasXOP()) { | |||
| 10131 | SDValue LoSrc = extract128BitVector(SrcVec, 0, DAG, DL); | |||
| 10132 | SDValue HiSrc = extract128BitVector(SrcVec, 16, DAG, DL); | |||
| 10133 | SDValue LoIdx = extract128BitVector(IndicesVec, 0, DAG, DL); | |||
| 10134 | SDValue HiIdx = extract128BitVector(IndicesVec, 16, DAG, DL); | |||
| 10135 | return DAG.getNode( | |||
| 10136 | ISD::CONCAT_VECTORS, DL, VT, | |||
| 10137 | DAG.getNode(X86ISD::VPPERM, DL, MVT::v16i8, LoSrc, HiSrc, LoIdx), | |||
| 10138 | DAG.getNode(X86ISD::VPPERM, DL, MVT::v16i8, LoSrc, HiSrc, HiIdx)); | |||
| 10139 | } else if (Subtarget.hasAVX()) { | |||
| 10140 | SDValue Lo = extract128BitVector(SrcVec, 0, DAG, DL); | |||
| 10141 | SDValue Hi = extract128BitVector(SrcVec, 16, DAG, DL); | |||
| 10142 | SDValue LoLo = DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Lo, Lo); | |||
| 10143 | SDValue HiHi = DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Hi, Hi); | |||
| 10144 | auto PSHUFBBuilder = [](SelectionDAG &DAG, const SDLoc &DL, | |||
| 10145 | ArrayRef<SDValue> Ops) { | |||
| 10146 | // Permute Lo and Hi and then select based on index range. | |||
| 10147 | // This works as SHUFB uses bits[3:0] to permute elements and we don't | |||
| 10148 | // care about the bit[7] as its just an index vector. | |||
| 10149 | SDValue Idx = Ops[2]; | |||
| 10150 | EVT VT = Idx.getValueType(); | |||
| 10151 | return DAG.getSelectCC(DL, Idx, DAG.getConstant(15, DL, VT), | |||
| 10152 | DAG.getNode(X86ISD::PSHUFB, DL, VT, Ops[1], Idx), | |||
| 10153 | DAG.getNode(X86ISD::PSHUFB, DL, VT, Ops[0], Idx), | |||
| 10154 | ISD::CondCode::SETGT); | |||
| 10155 | }; | |||
| 10156 | SDValue Ops[] = {LoLo, HiHi, IndicesVec}; | |||
| 10157 | return SplitOpsAndApply(DAG, Subtarget, DL, MVT::v32i8, Ops, | |||
| 10158 | PSHUFBBuilder); | |||
| 10159 | } | |||
| 10160 | break; | |||
| 10161 | case MVT::v16i16: | |||
| 10162 | if (Subtarget.hasVLX() && Subtarget.hasBWI()) | |||
| 10163 | Opcode = X86ISD::VPERMV; | |||
| 10164 | else if (Subtarget.hasAVX()) { | |||
| 10165 | // Scale to v32i8 and perform as v32i8. | |||
| 10166 | IndicesVec = ScaleIndices(IndicesVec, 2); | |||
| 10167 | return DAG.getBitcast( | |||
| 10168 | VT, createVariablePermute( | |||
| 10169 | MVT::v32i8, DAG.getBitcast(MVT::v32i8, SrcVec), | |||
| 10170 | DAG.getBitcast(MVT::v32i8, IndicesVec), DL, DAG, Subtarget)); | |||
| 10171 | } | |||
| 10172 | break; | |||
| 10173 | case MVT::v8f32: | |||
| 10174 | case MVT::v8i32: | |||
| 10175 | if (Subtarget.hasAVX2()) | |||
| 10176 | Opcode = X86ISD::VPERMV; | |||
| 10177 | else if (Subtarget.hasAVX()) { | |||
| 10178 | SrcVec = DAG.getBitcast(MVT::v8f32, SrcVec); | |||
| 10179 | SDValue LoLo = DAG.getVectorShuffle(MVT::v8f32, DL, SrcVec, SrcVec, | |||
| 10180 | {0, 1, 2, 3, 0, 1, 2, 3}); | |||
| 10181 | SDValue HiHi = DAG.getVectorShuffle(MVT::v8f32, DL, SrcVec, SrcVec, | |||
| 10182 | {4, 5, 6, 7, 4, 5, 6, 7}); | |||
| 10183 | if (Subtarget.hasXOP()) | |||
| 10184 | return DAG.getBitcast( | |||
| 10185 | VT, DAG.getNode(X86ISD::VPERMIL2, DL, MVT::v8f32, LoLo, HiHi, | |||
| 10186 | IndicesVec, DAG.getTargetConstant(0, DL, MVT::i8))); | |||
| 10187 | // Permute Lo and Hi and then select based on index range. | |||
| 10188 | // This works as VPERMILPS only uses index bits[0:1] to permute elements. | |||
| 10189 | SDValue Res = DAG.getSelectCC( | |||
| 10190 | DL, IndicesVec, DAG.getConstant(3, DL, MVT::v8i32), | |||
| 10191 | DAG.getNode(X86ISD::VPERMILPV, DL, MVT::v8f32, HiHi, IndicesVec), | |||
| 10192 | DAG.getNode(X86ISD::VPERMILPV, DL, MVT::v8f32, LoLo, IndicesVec), | |||
| 10193 | ISD::CondCode::SETGT); | |||
| 10194 | return DAG.getBitcast(VT, Res); | |||
| 10195 | } | |||
| 10196 | break; | |||
| 10197 | case MVT::v4i64: | |||
| 10198 | case MVT::v4f64: | |||
| 10199 | if (Subtarget.hasAVX512()) { | |||
| 10200 | if (!Subtarget.hasVLX()) { | |||
| 10201 | MVT WidenSrcVT = MVT::getVectorVT(VT.getScalarType(), 8); | |||
| 10202 | SrcVec = widenSubVector(WidenSrcVT, SrcVec, false, Subtarget, DAG, | |||
| 10203 | SDLoc(SrcVec)); | |||
| 10204 | IndicesVec = widenSubVector(MVT::v8i64, IndicesVec, false, Subtarget, | |||
| 10205 | DAG, SDLoc(IndicesVec)); | |||
| 10206 | SDValue Res = createVariablePermute(WidenSrcVT, SrcVec, IndicesVec, DL, | |||
| 10207 | DAG, Subtarget); | |||
| 10208 | return extract256BitVector(Res, 0, DAG, DL); | |||
| 10209 | } | |||
| 10210 | Opcode = X86ISD::VPERMV; | |||
| 10211 | } else if (Subtarget.hasAVX()) { | |||
| 10212 | SrcVec = DAG.getBitcast(MVT::v4f64, SrcVec); | |||
| 10213 | SDValue LoLo = | |||
| 10214 | DAG.getVectorShuffle(MVT::v4f64, DL, SrcVec, SrcVec, {0, 1, 0, 1}); | |||
| 10215 | SDValue HiHi = | |||
| 10216 | DAG.getVectorShuffle(MVT::v4f64, DL, SrcVec, SrcVec, {2, 3, 2, 3}); | |||
| 10217 | // VPERMIL2PD selects with bit#1 of the index vector, so scale IndicesVec. | |||
| 10218 | IndicesVec = DAG.getNode(ISD::ADD, DL, IndicesVT, IndicesVec, IndicesVec); | |||
| 10219 | if (Subtarget.hasXOP()) | |||
| 10220 | return DAG.getBitcast( | |||
| 10221 | VT, DAG.getNode(X86ISD::VPERMIL2, DL, MVT::v4f64, LoLo, HiHi, | |||
| 10222 | IndicesVec, DAG.getTargetConstant(0, DL, MVT::i8))); | |||
| 10223 | // Permute Lo and Hi and then select based on index range. | |||
| 10224 | // This works as VPERMILPD only uses index bit[1] to permute elements. | |||
| 10225 | SDValue Res = DAG.getSelectCC( | |||
| 10226 | DL, IndicesVec, DAG.getConstant(2, DL, MVT::v4i64), | |||
| 10227 | DAG.getNode(X86ISD::VPERMILPV, DL, MVT::v4f64, HiHi, IndicesVec), | |||
| 10228 | DAG.getNode(X86ISD::VPERMILPV, DL, MVT::v4f64, LoLo, IndicesVec), | |||
| 10229 | ISD::CondCode::SETGT); | |||
| 10230 | return DAG.getBitcast(VT, Res); | |||
| 10231 | } | |||
| 10232 | break; | |||
| 10233 | case MVT::v64i8: | |||
| 10234 | if (Subtarget.hasVBMI()) | |||
| 10235 | Opcode = X86ISD::VPERMV; | |||
| 10236 | break; | |||
| 10237 | case MVT::v32i16: | |||
| 10238 | if (Subtarget.hasBWI()) | |||
| 10239 | Opcode = X86ISD::VPERMV; | |||
| 10240 | break; | |||
| 10241 | case MVT::v16f32: | |||
| 10242 | case MVT::v16i32: | |||
| 10243 | case MVT::v8f64: | |||
| 10244 | case MVT::v8i64: | |||
| 10245 | if (Subtarget.hasAVX512()) | |||
| 10246 | Opcode = X86ISD::VPERMV; | |||
| 10247 | break; | |||
| 10248 | } | |||
| 10249 | if (!Opcode) | |||
| 10250 | return SDValue(); | |||
| 10251 | ||||
| 10252 | assert((VT.getSizeInBits() == ShuffleVT.getSizeInBits()) &&((void)0) | |||
| 10253 | (VT.getScalarSizeInBits() % ShuffleVT.getScalarSizeInBits()) == 0 &&((void)0) | |||
| 10254 | "Illegal variable permute shuffle type")((void)0); | |||
| 10255 | ||||
| 10256 | uint64_t Scale = VT.getScalarSizeInBits() / ShuffleVT.getScalarSizeInBits(); | |||
| 10257 | if (Scale > 1) | |||
| 10258 | IndicesVec = ScaleIndices(IndicesVec, Scale); | |||
| 10259 | ||||
| 10260 | EVT ShuffleIdxVT = EVT(ShuffleVT).changeVectorElementTypeToInteger(); | |||
| 10261 | IndicesVec = DAG.getBitcast(ShuffleIdxVT, IndicesVec); | |||
| 10262 | ||||
| 10263 | SrcVec = DAG.getBitcast(ShuffleVT, SrcVec); | |||
| 10264 | SDValue Res = Opcode == X86ISD::VPERMV | |||
| 10265 | ? DAG.getNode(Opcode, DL, ShuffleVT, IndicesVec, SrcVec) | |||
| 10266 | : DAG.getNode(Opcode, DL, ShuffleVT, SrcVec, IndicesVec); | |||
| 10267 | return DAG.getBitcast(VT, Res); | |||
| 10268 | } | |||
| 10269 | ||||
| 10270 | // Tries to lower a BUILD_VECTOR composed of extract-extract chains that can be | |||
| 10271 | // reasoned to be a permutation of a vector by indices in a non-constant vector. | |||
| 10272 | // (build_vector (extract_elt V, (extract_elt I, 0)), | |||
| 10273 | // (extract_elt V, (extract_elt I, 1)), | |||
| 10274 | // ... | |||
| 10275 | // -> | |||
| 10276 | // (vpermv I, V) | |||
| 10277 | // | |||
| 10278 | // TODO: Handle undefs | |||
| 10279 | // TODO: Utilize pshufb and zero mask blending to support more efficient | |||
| 10280 | // construction of vectors with constant-0 elements. | |||
| 10281 | static SDValue | |||
| 10282 | LowerBUILD_VECTORAsVariablePermute(SDValue V, SelectionDAG &DAG, | |||
| 10283 | const X86Subtarget &Subtarget) { | |||
| 10284 | SDValue SrcVec, IndicesVec; | |||
| 10285 | // Check for a match of the permute source vector and permute index elements. | |||
| 10286 | // This is done by checking that the i-th build_vector operand is of the form: | |||
| 10287 | // (extract_elt SrcVec, (extract_elt IndicesVec, i)). | |||
| 10288 | for (unsigned Idx = 0, E = V.getNumOperands(); Idx != E; ++Idx) { | |||
| 10289 | SDValue Op = V.getOperand(Idx); | |||
| 10290 | if (Op.getOpcode() != ISD::EXTRACT_VECTOR_ELT) | |||
| 10291 | return SDValue(); | |||
| 10292 | ||||
| 10293 | // If this is the first extract encountered in V, set the source vector, | |||
| 10294 | // otherwise verify the extract is from the previously defined source | |||
| 10295 | // vector. | |||
| 10296 | if (!SrcVec) | |||
| 10297 | SrcVec = Op.getOperand(0); | |||
| 10298 | else if (SrcVec != Op.getOperand(0)) | |||
| 10299 | return SDValue(); | |||
| 10300 | SDValue ExtractedIndex = Op->getOperand(1); | |||
| 10301 | // Peek through extends. | |||
| 10302 | if (ExtractedIndex.getOpcode() == ISD::ZERO_EXTEND || | |||
| 10303 | ExtractedIndex.getOpcode() == ISD::SIGN_EXTEND) | |||
| 10304 | ExtractedIndex = ExtractedIndex.getOperand(0); | |||
| 10305 | if (ExtractedIndex.getOpcode() != ISD::EXTRACT_VECTOR_ELT) | |||
| 10306 | return SDValue(); | |||
| 10307 | ||||
| 10308 | // If this is the first extract from the index vector candidate, set the | |||
| 10309 | // indices vector, otherwise verify the extract is from the previously | |||
| 10310 | // defined indices vector. | |||
| 10311 | if (!IndicesVec) | |||
| 10312 | IndicesVec = ExtractedIndex.getOperand(0); | |||
| 10313 | else if (IndicesVec != ExtractedIndex.getOperand(0)) | |||
| 10314 | return SDValue(); | |||
| 10315 | ||||
| 10316 | auto *PermIdx = dyn_cast<ConstantSDNode>(ExtractedIndex.getOperand(1)); | |||
| 10317 | if (!PermIdx || PermIdx->getAPIntValue() != Idx) | |||
| 10318 | return SDValue(); | |||
| 10319 | } | |||
| 10320 | ||||
| 10321 | SDLoc DL(V); | |||
| 10322 | MVT VT = V.getSimpleValueType(); | |||
| 10323 | return createVariablePermute(VT, SrcVec, IndicesVec, DL, DAG, Subtarget); | |||
| 10324 | } | |||
| 10325 | ||||
| 10326 | SDValue | |||
| 10327 | X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const { | |||
| 10328 | SDLoc dl(Op); | |||
| 10329 | ||||
| 10330 | MVT VT = Op.getSimpleValueType(); | |||
| 10331 | MVT EltVT = VT.getVectorElementType(); | |||
| 10332 | unsigned NumElems = Op.getNumOperands(); | |||
| 10333 | ||||
| 10334 | // Generate vectors for predicate vectors. | |||
| 10335 | if (VT.getVectorElementType() == MVT::i1 && Subtarget.hasAVX512()) | |||
| ||||
| 10336 | return LowerBUILD_VECTORvXi1(Op, DAG, Subtarget); | |||
| 10337 | ||||
| 10338 | if (SDValue VectorConstant = materializeVectorConstant(Op, DAG, Subtarget)) | |||
| 10339 | return VectorConstant; | |||
| 10340 | ||||
| 10341 | unsigned EVTBits = EltVT.getSizeInBits(); | |||
| 10342 | APInt UndefMask = APInt::getNullValue(NumElems); | |||
| 10343 | APInt ZeroMask = APInt::getNullValue(NumElems); | |||
| 10344 | APInt NonZeroMask = APInt::getNullValue(NumElems); | |||
| 10345 | bool IsAllConstants = true; | |||
| 10346 | SmallSet<SDValue, 8> Values; | |||
| 10347 | unsigned NumConstants = NumElems; | |||
| 10348 | for (unsigned i = 0; i < NumElems; ++i) { | |||
| 10349 | SDValue Elt = Op.getOperand(i); | |||
| 10350 | if (Elt.isUndef()) { | |||
| 10351 | UndefMask.setBit(i); | |||
| 10352 | continue; | |||
| 10353 | } | |||
| 10354 | Values.insert(Elt); | |||
| 10355 | if (!isa<ConstantSDNode>(Elt) && !isa<ConstantFPSDNode>(Elt)) { | |||
| 10356 | IsAllConstants = false; | |||
| 10357 | NumConstants--; | |||
| 10358 | } | |||
| 10359 | if (X86::isZeroNode(Elt)) { | |||
| 10360 | ZeroMask.setBit(i); | |||
| 10361 | } else { | |||
| 10362 | NonZeroMask.setBit(i); | |||
| 10363 | } | |||
| 10364 | } | |||
| 10365 | ||||
| 10366 | // All undef vector. Return an UNDEF. All zero vectors were handled above. | |||
| 10367 | if (NonZeroMask == 0) { | |||
| 10368 | assert(UndefMask.isAllOnesValue() && "Fully undef mask expected")((void)0); | |||
| 10369 | return DAG.getUNDEF(VT); | |||
| 10370 | } | |||
| 10371 | ||||
| 10372 | BuildVectorSDNode *BV = cast<BuildVectorSDNode>(Op.getNode()); | |||
| 10373 | ||||
| 10374 | // If the upper elts of a ymm/zmm are undef/zero then we might be better off | |||
| 10375 | // lowering to a smaller build vector and padding with undef/zero. | |||
| 10376 | if ((VT.is256BitVector() || VT.is512BitVector()) && | |||
| 10377 | !isFoldableUseOfShuffle(BV)) { | |||
| 10378 | unsigned UpperElems = NumElems / 2; | |||
| 10379 | APInt UndefOrZeroMask = UndefMask | ZeroMask; | |||
| 10380 | unsigned NumUpperUndefsOrZeros = UndefOrZeroMask.countLeadingOnes(); | |||
| 10381 | if (NumUpperUndefsOrZeros >= UpperElems) { | |||
| 10382 | if (VT.is512BitVector() && | |||
| 10383 | NumUpperUndefsOrZeros >= (NumElems - (NumElems / 4))) | |||
| 10384 | UpperElems = NumElems - (NumElems / 4); | |||
| 10385 | bool UndefUpper = UndefMask.countLeadingOnes() >= UpperElems; | |||
| 10386 | MVT LowerVT = MVT::getVectorVT(EltVT, NumElems - UpperElems); | |||
| 10387 | SDValue NewBV = | |||
| 10388 | DAG.getBuildVector(LowerVT, dl, Op->ops().drop_back(UpperElems)); | |||
| 10389 | return widenSubVector(VT, NewBV, !UndefUpper, Subtarget, DAG, dl); | |||
| 10390 | } | |||
| 10391 | } | |||
| 10392 | ||||
| 10393 | if (SDValue AddSub = lowerToAddSubOrFMAddSub(BV, Subtarget, DAG)) | |||
| 10394 | return AddSub; | |||
| 10395 | if (SDValue HorizontalOp = LowerToHorizontalOp(BV, Subtarget, DAG)) | |||
| 10396 | return HorizontalOp; | |||
| 10397 | if (SDValue Broadcast = lowerBuildVectorAsBroadcast(BV, Subtarget, DAG)) | |||
| 10398 | return Broadcast; | |||
| 10399 | if (SDValue BitOp = lowerBuildVectorToBitOp(BV, Subtarget, DAG)) | |||
| 10400 | return BitOp; | |||
| 10401 | ||||
| 10402 | unsigned NumZero = ZeroMask.countPopulation(); | |||
| 10403 | unsigned NumNonZero = NonZeroMask.countPopulation(); | |||
| 10404 | ||||
| 10405 | // If we are inserting one variable into a vector of non-zero constants, try | |||
| 10406 | // to avoid loading each constant element as a scalar. Load the constants as a | |||
| 10407 | // vector and then insert the variable scalar element. If insertion is not | |||
| 10408 | // supported, fall back to a shuffle to get the scalar blended with the | |||
| 10409 | // constants. Insertion into a zero vector is handled as a special-case | |||
| 10410 | // somewhere below here. | |||
| 10411 | if (NumConstants == NumElems - 1 && NumNonZero != 1 && | |||
| 10412 | (isOperationLegalOrCustom(ISD::INSERT_VECTOR_ELT, VT) || | |||
| 10413 | isOperationLegalOrCustom(ISD::VECTOR_SHUFFLE, VT))) { | |||
| 10414 | // Create an all-constant vector. The variable element in the old | |||
| 10415 | // build vector is replaced by undef in the constant vector. Save the | |||
| 10416 | // variable scalar element and its index for use in the insertelement. | |||
| 10417 | LLVMContext &Context = *DAG.getContext(); | |||
| 10418 | Type *EltType = Op.getValueType().getScalarType().getTypeForEVT(Context); | |||
| 10419 | SmallVector<Constant *, 16> ConstVecOps(NumElems, UndefValue::get(EltType)); | |||
| 10420 | SDValue VarElt; | |||
| 10421 | SDValue InsIndex; | |||
| 10422 | for (unsigned i = 0; i != NumElems; ++i) { | |||
| 10423 | SDValue Elt = Op.getOperand(i); | |||
| 10424 | if (auto *C = dyn_cast<ConstantSDNode>(Elt)) | |||
| 10425 | ConstVecOps[i] = ConstantInt::get(Context, C->getAPIntValue()); | |||
| 10426 | else if (auto *C = dyn_cast<ConstantFPSDNode>(Elt)) | |||
| 10427 | ConstVecOps[i] = ConstantFP::get(Context, C->getValueAPF()); | |||
| 10428 | else if (!Elt.isUndef()) { | |||
| 10429 | assert(!VarElt.getNode() && !InsIndex.getNode() &&((void)0) | |||
| 10430 | "Expected one variable element in this vector")((void)0); | |||
| 10431 | VarElt = Elt; | |||
| 10432 | InsIndex = DAG.getVectorIdxConstant(i, dl); | |||
| 10433 | } | |||
| 10434 | } | |||
| 10435 | Constant *CV = ConstantVector::get(ConstVecOps); | |||
| 10436 | SDValue DAGConstVec = DAG.getConstantPool(CV, VT); | |||
| 10437 | ||||
| 10438 | // The constants we just created may not be legal (eg, floating point). We | |||
| 10439 | // must lower the vector right here because we can not guarantee that we'll | |||
| 10440 | // legalize it before loading it. This is also why we could not just create | |||
| 10441 | // a new build vector here. If the build vector contains illegal constants, | |||
| 10442 | // it could get split back up into a series of insert elements. | |||
| 10443 | // TODO: Improve this by using shorter loads with broadcast/VZEXT_LOAD. | |||
| 10444 | SDValue LegalDAGConstVec = LowerConstantPool(DAGConstVec, DAG); | |||
| 10445 | MachineFunction &MF = DAG.getMachineFunction(); | |||
| 10446 | MachinePointerInfo MPI = MachinePointerInfo::getConstantPool(MF); | |||
| 10447 | SDValue Ld = DAG.getLoad(VT, dl, DAG.getEntryNode(), LegalDAGConstVec, MPI); | |||
| 10448 | unsigned InsertC = cast<ConstantSDNode>(InsIndex)->getZExtValue(); | |||
| 10449 | unsigned NumEltsInLow128Bits = 128 / VT.getScalarSizeInBits(); | |||
| 10450 | if (InsertC < NumEltsInLow128Bits) | |||
| 10451 | return DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Ld, VarElt, InsIndex); | |||
| 10452 | ||||
| 10453 | // There's no good way to insert into the high elements of a >128-bit | |||
| 10454 | // vector, so use shuffles to avoid an extract/insert sequence. | |||
| 10455 | assert(VT.getSizeInBits() > 128 && "Invalid insertion index?")((void)0); | |||
| 10456 | assert(Subtarget.hasAVX() && "Must have AVX with >16-byte vector")((void)0); | |||
| 10457 | SmallVector<int, 8> ShuffleMask; | |||
| 10458 | unsigned NumElts = VT.getVectorNumElements(); | |||
| 10459 | for (unsigned i = 0; i != NumElts; ++i) | |||
| 10460 | ShuffleMask.push_back(i == InsertC ? NumElts : i); | |||
| 10461 | SDValue S2V = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, VarElt); | |||
| 10462 | return DAG.getVectorShuffle(VT, dl, Ld, S2V, ShuffleMask); | |||
| 10463 | } | |||
| 10464 | ||||
| 10465 | // Special case for single non-zero, non-undef, element. | |||
| 10466 | if (NumNonZero == 1) { | |||
| 10467 | unsigned Idx = NonZeroMask.countTrailingZeros(); | |||
| 10468 | SDValue Item = Op.getOperand(Idx); | |||
| 10469 | ||||
| 10470 | // If we have a constant or non-constant insertion into the low element of | |||
| 10471 | // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into | |||
| 10472 | // the rest of the elements. This will be matched as movd/movq/movss/movsd | |||
| 10473 | // depending on what the source datatype is. | |||
| 10474 | if (Idx == 0) { | |||
| 10475 | if (NumZero == 0) | |||
| 10476 | return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item); | |||
| 10477 | ||||
| 10478 | if (EltVT == MVT::i32 || EltVT == MVT::f32 || EltVT == MVT::f64 || | |||
| 10479 | (EltVT == MVT::i64 && Subtarget.is64Bit())) { | |||
| 10480 | assert((VT.is128BitVector() || VT.is256BitVector() ||((void)0) | |||
| 10481 | VT.is512BitVector()) &&((void)0) | |||
| 10482 | "Expected an SSE value type!")((void)0); | |||
| 10483 | Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item); | |||
| 10484 | // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector. | |||
| 10485 | return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG); | |||
| 10486 | } | |||
| 10487 | ||||
| 10488 | // We can't directly insert an i8 or i16 into a vector, so zero extend | |||
| 10489 | // it to i32 first. | |||
| 10490 | if (EltVT == MVT::i16 || EltVT == MVT::i8) { | |||
| 10491 | Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item); | |||
| 10492 | MVT ShufVT = MVT::getVectorVT(MVT::i32, VT.getSizeInBits() / 32); | |||
| 10493 | Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, ShufVT, Item); | |||
| 10494 | Item = getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG); | |||
| 10495 | return DAG.getBitcast(VT, Item); | |||
| 10496 | } | |||
| 10497 | } | |||
| 10498 | ||||
| 10499 | // Is it a vector logical left shift? | |||
| 10500 | if (NumElems == 2 && Idx == 1 && | |||
| 10501 | X86::isZeroNode(Op.getOperand(0)) && | |||
| 10502 | !X86::isZeroNode(Op.getOperand(1))) { | |||
| 10503 | unsigned NumBits = VT.getSizeInBits(); | |||
| 10504 | return getVShift(true, VT, | |||
| 10505 | DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, | |||
| 10506 | VT, Op.getOperand(1)), | |||
| 10507 | NumBits/2, DAG, *this, dl); | |||
| 10508 | } | |||
| 10509 | ||||
| 10510 | if (IsAllConstants) // Otherwise, it's better to do a constpool load. | |||
| 10511 | return SDValue(); | |||
| 10512 | ||||
| 10513 | // Otherwise, if this is a vector with i32 or f32 elements, and the element | |||
| 10514 | // is a non-constant being inserted into an element other than the low one, | |||
| 10515 | // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka | |||
| 10516 | // movd/movss) to move this into the low element, then shuffle it into | |||
| 10517 | // place. | |||
| 10518 | if (EVTBits == 32) { | |||
| 10519 | Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item); | |||
| 10520 | return getShuffleVectorZeroOrUndef(Item, Idx, NumZero > 0, Subtarget, DAG); | |||
| 10521 | } | |||
| 10522 | } | |||
| 10523 | ||||
| 10524 | // Splat is obviously ok. Let legalizer expand it to a shuffle. | |||
| 10525 | if (Values.size() == 1) { | |||
| 10526 | if (EVTBits == 32) { | |||
| 10527 | // Instead of a shuffle like this: | |||
| 10528 | // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0> | |||
| 10529 | // Check if it's possible to issue this instead. | |||
| 10530 | // shuffle (vload ptr)), undef, <1, 1, 1, 1> | |||
| 10531 | unsigned Idx = NonZeroMask.countTrailingZeros(); | |||
| 10532 | SDValue Item = Op.getOperand(Idx); | |||
| 10533 | if (Op.getNode()->isOnlyUserOf(Item.getNode())) | |||
| 10534 | return LowerAsSplatVectorLoad(Item, VT, dl, DAG); | |||
| 10535 | } | |||
| 10536 | return SDValue(); | |||
| 10537 | } | |||
| 10538 | ||||
| 10539 | // A vector full of immediates; various special cases are already | |||
| 10540 | // handled, so this is best done with a single constant-pool load. | |||
| 10541 | if (IsAllConstants) | |||
| 10542 | return SDValue(); | |||
| 10543 | ||||
| 10544 | if (SDValue V = LowerBUILD_VECTORAsVariablePermute(Op, DAG, Subtarget)) | |||
| 10545 | return V; | |||
| 10546 | ||||
| 10547 | // See if we can use a vector load to get all of the elements. | |||
| 10548 | { | |||
| 10549 | SmallVector<SDValue, 64> Ops(Op->op_begin(), Op->op_begin() + NumElems); | |||
| 10550 | if (SDValue LD = | |||
| 10551 | EltsFromConsecutiveLoads(VT, Ops, dl, DAG, Subtarget, false)) | |||
| 10552 | return LD; | |||
| 10553 | } | |||
| 10554 | ||||
| 10555 | // If this is a splat of pairs of 32-bit elements, we can use a narrower | |||
| 10556 | // build_vector and broadcast it. | |||
| 10557 | // TODO: We could probably generalize this more. | |||
| 10558 | if (Subtarget.hasAVX2() && EVTBits == 32 && Values.size() == 2) { | |||
| 10559 | SDValue Ops[4] = { Op.getOperand(0), Op.getOperand(1), | |||
| 10560 | DAG.getUNDEF(EltVT), DAG.getUNDEF(EltVT) }; | |||
| 10561 | auto CanSplat = [](SDValue Op, unsigned NumElems, ArrayRef<SDValue> Ops) { | |||
| 10562 | // Make sure all the even/odd operands match. | |||
| 10563 | for (unsigned i = 2; i != NumElems; ++i) | |||
| 10564 | if (Ops[i % 2] != Op.getOperand(i)) | |||
| 10565 | return false; | |||
| 10566 | return true; | |||
| 10567 | }; | |||
| 10568 | if (CanSplat(Op, NumElems, Ops)) { | |||
| 10569 | MVT WideEltVT = VT.isFloatingPoint() ? MVT::f64 : MVT::i64; | |||
| 10570 | MVT NarrowVT = MVT::getVectorVT(EltVT, 4); | |||
| 10571 | // Create a new build vector and cast to v2i64/v2f64. | |||
| 10572 | SDValue NewBV = DAG.getBitcast(MVT::getVectorVT(WideEltVT, 2), | |||
| 10573 | DAG.getBuildVector(NarrowVT, dl, Ops)); | |||
| 10574 | // Broadcast from v2i64/v2f64 and cast to final VT. | |||
| 10575 | MVT BcastVT = MVT::getVectorVT(WideEltVT, NumElems / 2); | |||
| 10576 | return DAG.getBitcast(VT, DAG.getNode(X86ISD::VBROADCAST, dl, BcastVT, | |||
| 10577 | NewBV)); | |||
| 10578 | } | |||
| 10579 | } | |||
| 10580 | ||||
| 10581 | // For AVX-length vectors, build the individual 128-bit pieces and use | |||
| 10582 | // shuffles to put them in place. | |||
| 10583 | if (VT.getSizeInBits() > 128) { | |||
| 10584 | MVT HVT = MVT::getVectorVT(EltVT, NumElems / 2); | |||
| 10585 | ||||
| 10586 | // Build both the lower and upper subvector. | |||
| 10587 | SDValue Lower = | |||
| 10588 | DAG.getBuildVector(HVT, dl, Op->ops().slice(0, NumElems / 2)); | |||
| 10589 | SDValue Upper = DAG.getBuildVector( | |||
| 10590 | HVT, dl, Op->ops().slice(NumElems / 2, NumElems /2)); | |||
| 10591 | ||||
| 10592 | // Recreate the wider vector with the lower and upper part. | |||
| 10593 | return concatSubVectors(Lower, Upper, DAG, dl); | |||
| 10594 | } | |||
| 10595 | ||||
| 10596 | // Let legalizer expand 2-wide build_vectors. | |||
| 10597 | if (EVTBits == 64) { | |||
| 10598 | if (NumNonZero == 1) { | |||
| 10599 | // One half is zero or undef. | |||
| 10600 | unsigned Idx = NonZeroMask.countTrailingZeros(); | |||
| 10601 | SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, | |||
| 10602 | Op.getOperand(Idx)); | |||
| 10603 | return getShuffleVectorZeroOrUndef(V2, Idx, true, Subtarget, DAG); | |||
| 10604 | } | |||
| 10605 | return SDValue(); | |||
| 10606 | } | |||
| 10607 | ||||
| 10608 | // If element VT is < 32 bits, convert it to inserts into a zero vector. | |||
| 10609 | if (EVTBits == 8 && NumElems == 16) | |||
| 10610 | if (SDValue V = LowerBuildVectorv16i8(Op, NonZeroMask, NumNonZero, NumZero, | |||
| 10611 | DAG, Subtarget)) | |||
| 10612 | return V; | |||
| 10613 | ||||
| 10614 | if (EVTBits == 16 && NumElems == 8) | |||
| 10615 | if (SDValue V = LowerBuildVectorv8i16(Op, NonZeroMask, NumNonZero, NumZero, | |||
| 10616 | DAG, Subtarget)) | |||
| 10617 | return V; | |||
| 10618 | ||||
| 10619 | // If element VT is == 32 bits and has 4 elems, try to generate an INSERTPS | |||
| 10620 | if (EVTBits == 32 && NumElems == 4) | |||
| 10621 | if (SDValue V = LowerBuildVectorv4x32(Op, DAG, Subtarget)) | |||
| 10622 | return V; | |||
| 10623 | ||||
| 10624 | // If element VT is == 32 bits, turn it into a number of shuffles. | |||
| 10625 | if (NumElems == 4 && NumZero > 0) { | |||
| 10626 | SmallVector<SDValue, 8> Ops(NumElems); | |||
| 10627 | for (unsigned i = 0; i < 4; ++i) { | |||
| 10628 | bool isZero = !NonZeroMask[i]; | |||
| 10629 | if (isZero) | |||
| 10630 | Ops[i] = getZeroVector(VT, Subtarget, DAG, dl); | |||
| 10631 | else | |||
| 10632 | Ops[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i)); | |||
| 10633 | } | |||
| 10634 | ||||
| 10635 | for (unsigned i = 0; i < 2; ++i) { | |||
| 10636 | switch (NonZeroMask.extractBitsAsZExtValue(2, i * 2)) { | |||
| 10637 | default: llvm_unreachable("Unexpected NonZero count")__builtin_unreachable(); | |||
| 10638 | case 0: | |||
| 10639 | Ops[i] = Ops[i*2]; // Must be a zero vector. | |||
| 10640 | break; | |||
| 10641 | case 1: | |||
| 10642 | Ops[i] = getMOVL(DAG, dl, VT, Ops[i*2+1], Ops[i*2]); | |||
| 10643 | break; | |||
| 10644 | case 2: | |||
| 10645 | Ops[i] = getMOVL(DAG, dl, VT, Ops[i*2], Ops[i*2+1]); | |||
| 10646 | break; | |||
| 10647 | case 3: | |||
| 10648 | Ops[i] = getUnpackl(DAG, dl, VT, Ops[i*2], Ops[i*2+1]); | |||
| 10649 | break; | |||
| 10650 | } | |||
| 10651 | } | |||
| 10652 | ||||
| 10653 | bool Reverse1 = NonZeroMask.extractBitsAsZExtValue(2, 0) == 2; | |||
| 10654 | bool Reverse2 = NonZeroMask.extractBitsAsZExtValue(2, 2) == 2; | |||
| 10655 | int MaskVec[] = { | |||
| 10656 | Reverse1 ? 1 : 0, | |||
| 10657 | Reverse1 ? 0 : 1, | |||
| 10658 | static_cast<int>(Reverse2 ? NumElems+1 : NumElems), | |||
| 10659 | static_cast<int>(Reverse2 ? NumElems : NumElems+1) | |||
| 10660 | }; | |||
| 10661 | return DAG.getVectorShuffle(VT, dl, Ops[0], Ops[1], MaskVec); | |||
| 10662 | } | |||
| 10663 | ||||
| 10664 | assert(Values.size() > 1 && "Expected non-undef and non-splat vector")((void)0); | |||
| 10665 | ||||
| 10666 | // Check for a build vector from mostly shuffle plus few inserting. | |||
| 10667 | if (SDValue Sh = buildFromShuffleMostly(Op, DAG)) | |||
| 10668 | return Sh; | |||
| 10669 | ||||
| 10670 | // For SSE 4.1, use insertps to put the high elements into the low element. | |||
| 10671 | if (Subtarget.hasSSE41()) { | |||
| 10672 | SDValue Result; | |||
| 10673 | if (!Op.getOperand(0).isUndef()) | |||
| 10674 | Result = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(0)); | |||
| 10675 | else | |||
| 10676 | Result = DAG.getUNDEF(VT); | |||
| 10677 | ||||
| 10678 | for (unsigned i = 1; i < NumElems; ++i) { | |||
| 10679 | if (Op.getOperand(i).isUndef()) continue; | |||
| 10680 | Result = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Result, | |||
| 10681 | Op.getOperand(i), DAG.getIntPtrConstant(i, dl)); | |||
| 10682 | } | |||
| 10683 | return Result; | |||
| 10684 | } | |||
| 10685 | ||||
| 10686 | // Otherwise, expand into a number of unpckl*, start by extending each of | |||
| 10687 | // our (non-undef) elements to the full vector width with the element in the | |||
| 10688 | // bottom slot of the vector (which generates no code for SSE). | |||
| 10689 | SmallVector<SDValue, 8> Ops(NumElems); | |||
| 10690 | for (unsigned i = 0; i < NumElems; ++i) { | |||
| 10691 | if (!Op.getOperand(i).isUndef()) | |||
| 10692 | Ops[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i)); | |||
| 10693 | else | |||
| 10694 | Ops[i] = DAG.getUNDEF(VT); | |||
| 10695 | } | |||
| 10696 | ||||
| 10697 | // Next, we iteratively mix elements, e.g. for v4f32: | |||
| 10698 | // Step 1: unpcklps 0, 1 ==> X: <?, ?, 1, 0> | |||
| 10699 | // : unpcklps 2, 3 ==> Y: <?, ?, 3, 2> | |||
| 10700 | // Step 2: unpcklpd X, Y ==> <3, 2, 1, 0> | |||
| 10701 | for (unsigned Scale = 1; Scale < NumElems; Scale *= 2) { | |||
| 10702 | // Generate scaled UNPCKL shuffle mask. | |||
| 10703 | SmallVector<int, 16> Mask; | |||
| 10704 | for(unsigned i = 0; i != Scale; ++i) | |||
| 10705 | Mask.push_back(i); | |||
| 10706 | for (unsigned i = 0; i != Scale; ++i) | |||
| 10707 | Mask.push_back(NumElems+i); | |||
| 10708 | Mask.append(NumElems - Mask.size(), SM_SentinelUndef); | |||
| 10709 | ||||
| 10710 | for (unsigned i = 0, e = NumElems / (2 * Scale); i != e; ++i) | |||
| 10711 | Ops[i] = DAG.getVectorShuffle(VT, dl, Ops[2*i], Ops[(2*i)+1], Mask); | |||
| 10712 | } | |||
| 10713 | return Ops[0]; | |||
| 10714 | } | |||
| 10715 | ||||
| 10716 | // 256-bit AVX can use the vinsertf128 instruction | |||
| 10717 | // to create 256-bit vectors from two other 128-bit ones. | |||
| 10718 | // TODO: Detect subvector broadcast here instead of DAG combine? | |||
| 10719 | static SDValue LowerAVXCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG, | |||
| 10720 | const X86Subtarget &Subtarget) { | |||
| 10721 | SDLoc dl(Op); | |||
| 10722 | MVT ResVT = Op.getSimpleValueType(); | |||
| 10723 | ||||
| 10724 | assert((ResVT.is256BitVector() ||((void)0) | |||
| 10725 | ResVT.is512BitVector()) && "Value type must be 256-/512-bit wide")((void)0); | |||
| 10726 | ||||
| 10727 | unsigned NumOperands = Op.getNumOperands(); | |||
| 10728 | unsigned NumZero = 0; | |||
| 10729 | unsigned NumNonZero = 0; | |||
| 10730 | unsigned NonZeros = 0; | |||
| 10731 | for (unsigned i = 0; i != NumOperands; ++i) { | |||
| 10732 | SDValue SubVec = Op.getOperand(i); | |||
| 10733 | if (SubVec.isUndef()) | |||
| 10734 | continue; | |||
| 10735 | if (ISD::isBuildVectorAllZeros(SubVec.getNode())) | |||
| 10736 | ++NumZero; | |||
| 10737 | else { | |||
| 10738 | assert(i < sizeof(NonZeros) * CHAR_BIT)((void)0); // Ensure the shift is in range. | |||
| 10739 | NonZeros |= 1 << i; | |||
| 10740 | ++NumNonZero; | |||
| 10741 | } | |||
| 10742 | } | |||
| 10743 | ||||
| 10744 | // If we have more than 2 non-zeros, build each half separately. | |||
| 10745 | if (NumNonZero > 2) { | |||
| 10746 | MVT HalfVT = ResVT.getHalfNumVectorElementsVT(); | |||
| 10747 | ArrayRef<SDUse> Ops = Op->ops(); | |||
| 10748 | SDValue Lo = DAG.getNode(ISD::CONCAT_VECTORS, dl, HalfVT, | |||
| 10749 | Ops.slice(0, NumOperands/2)); | |||
| 10750 | SDValue Hi = DAG.getNode(ISD::CONCAT_VECTORS, dl, HalfVT, | |||
| 10751 | Ops.slice(NumOperands/2)); | |||
| 10752 | return DAG.getNode(ISD::CONCAT_VECTORS, dl, ResVT, Lo, Hi); | |||
| 10753 | } | |||
| 10754 | ||||
| 10755 | // Otherwise, build it up through insert_subvectors. | |||
| 10756 | SDValue Vec = NumZero ? getZeroVector(ResVT, Subtarget, DAG, dl) | |||
| 10757 | : DAG.getUNDEF(ResVT); | |||
| 10758 | ||||
| 10759 | MVT SubVT = Op.getOperand(0).getSimpleValueType(); | |||
| 10760 | unsigned NumSubElems = SubVT.getVectorNumElements(); | |||
| 10761 | for (unsigned i = 0; i != NumOperands; ++i) { | |||
| 10762 | if ((NonZeros & (1 << i)) == 0) | |||
| 10763 | continue; | |||
| 10764 | ||||
| 10765 | Vec = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, Vec, | |||
| 10766 | Op.getOperand(i), | |||
| 10767 | DAG.getIntPtrConstant(i * NumSubElems, dl)); | |||
| 10768 | } | |||
| 10769 | ||||
| 10770 | return Vec; | |||
| 10771 | } | |||
| 10772 | ||||
| 10773 | // Returns true if the given node is a type promotion (by concatenating i1 | |||
| 10774 | // zeros) of the result of a node that already zeros all upper bits of | |||
| 10775 | // k-register. | |||
| 10776 | // TODO: Merge this with LowerAVXCONCAT_VECTORS? | |||
| 10777 | static SDValue LowerCONCAT_VECTORSvXi1(SDValue Op, | |||
| 10778 | const X86Subtarget &Subtarget, | |||
| 10779 | SelectionDAG & DAG) { | |||
| 10780 | SDLoc dl(Op); | |||
| 10781 | MVT ResVT = Op.getSimpleValueType(); | |||
| 10782 | unsigned NumOperands = Op.getNumOperands(); | |||
| 10783 | ||||
| 10784 | assert(NumOperands > 1 && isPowerOf2_32(NumOperands) &&((void)0) | |||
| 10785 | "Unexpected number of operands in CONCAT_VECTORS")((void)0); | |||
| 10786 | ||||
| 10787 | uint64_t Zeros = 0; | |||
| 10788 | uint64_t NonZeros = 0; | |||
| 10789 | for (unsigned i = 0; i != NumOperands; ++i) { | |||
| 10790 | SDValue SubVec = Op.getOperand(i); | |||
| 10791 | if (SubVec.isUndef()) | |||
| 10792 | continue; | |||
| 10793 | assert(i < sizeof(NonZeros) * CHAR_BIT)((void)0); // Ensure the shift is in range. | |||
| 10794 | if (ISD::isBuildVectorAllZeros(SubVec.getNode())) | |||
| 10795 | Zeros |= (uint64_t)1 << i; | |||
| 10796 | else | |||
| 10797 | NonZeros |= (uint64_t)1 << i; | |||
| 10798 | } | |||
| 10799 | ||||
| 10800 | unsigned NumElems = ResVT.getVectorNumElements(); | |||
| 10801 | ||||
| 10802 | // If we are inserting non-zero vector and there are zeros in LSBs and undef | |||
| 10803 | // in the MSBs we need to emit a KSHIFTL. The generic lowering to | |||
| 10804 | // insert_subvector will give us two kshifts. | |||
| 10805 | if (isPowerOf2_64(NonZeros) && Zeros != 0 && NonZeros > Zeros && | |||
| 10806 | Log2_64(NonZeros) != NumOperands - 1) { | |||
| 10807 | MVT ShiftVT = ResVT; | |||
| 10808 | if ((!Subtarget.hasDQI() && NumElems == 8) || NumElems < 8) | |||
| 10809 | ShiftVT = Subtarget.hasDQI() ? MVT::v8i1 : MVT::v16i1; | |||
| 10810 | unsigned Idx = Log2_64(NonZeros); | |||
| 10811 | SDValue SubVec = Op.getOperand(Idx); | |||
| 10812 | unsigned SubVecNumElts = SubVec.getSimpleValueType().getVectorNumElements(); | |||
| 10813 | SubVec = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ShiftVT, | |||
| 10814 | DAG.getUNDEF(ShiftVT), SubVec, | |||
| 10815 | DAG.getIntPtrConstant(0, dl)); | |||
| 10816 | Op = DAG.getNode(X86ISD::KSHIFTL, dl, ShiftVT, SubVec, | |||
| 10817 | DAG.getTargetConstant(Idx * SubVecNumElts, dl, MVT::i8)); | |||
| 10818 | return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, ResVT, Op, | |||
| 10819 | DAG.getIntPtrConstant(0, dl)); | |||
| 10820 | } | |||
| 10821 | ||||
| 10822 | // If there are zero or one non-zeros we can handle this very simply. | |||
| 10823 | if (NonZeros == 0 || isPowerOf2_64(NonZeros)) { | |||
| 10824 | SDValue Vec = Zeros ? DAG.getConstant(0, dl, ResVT) : DAG.getUNDEF(ResVT); | |||
| 10825 | if (!NonZeros) | |||
| 10826 | return Vec; | |||
| 10827 | unsigned Idx = Log2_64(NonZeros); | |||
| 10828 | SDValue SubVec = Op.getOperand(Idx); | |||
| 10829 | unsigned SubVecNumElts = SubVec.getSimpleValueType().getVectorNumElements(); | |||
| 10830 | return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, Vec, SubVec, | |||
| 10831 | DAG.getIntPtrConstant(Idx * SubVecNumElts, dl)); | |||
| 10832 | } | |||
| 10833 | ||||
| 10834 | if (NumOperands > 2) { | |||
| 10835 | MVT HalfVT = ResVT.getHalfNumVectorElementsVT(); | |||
| 10836 | ArrayRef<SDUse> Ops = Op->ops(); | |||
| 10837 | SDValue Lo = DAG.getNode(ISD::CONCAT_VECTORS, dl, HalfVT, | |||
| 10838 | Ops.slice(0, NumOperands/2)); | |||
| 10839 | SDValue Hi = DAG.getNode(ISD::CONCAT_VECTORS, dl, HalfVT, | |||
| 10840 | Ops.slice(NumOperands/2)); | |||
| 10841 | return DAG.getNode(ISD::CONCAT_VECTORS, dl, ResVT, Lo, Hi); | |||
| 10842 | } | |||
| 10843 | ||||
| 10844 | assert(countPopulation(NonZeros) == 2 && "Simple cases not handled?")((void)0); | |||
| 10845 | ||||
| 10846 | if (ResVT.getVectorNumElements() >= 16) | |||
| 10847 | return Op; // The operation is legal with KUNPCK | |||
| 10848 | ||||
| 10849 | SDValue Vec = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, | |||
| 10850 | DAG.getUNDEF(ResVT), Op.getOperand(0), | |||
| 10851 | DAG.getIntPtrConstant(0, dl)); | |||
| 10852 | return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, Vec, Op.getOperand(1), | |||
| 10853 | DAG.getIntPtrConstant(NumElems/2, dl)); | |||
| 10854 | } | |||
| 10855 | ||||
| 10856 | static SDValue LowerCONCAT_VECTORS(SDValue Op, | |||
| 10857 | const X86Subtarget &Subtarget, | |||
| 10858 | SelectionDAG &DAG) { | |||
| 10859 | MVT VT = Op.getSimpleValueType(); | |||
| 10860 | if (VT.getVectorElementType() == MVT::i1) | |||
| 10861 | return LowerCONCAT_VECTORSvXi1(Op, Subtarget, DAG); | |||
| 10862 | ||||
| 10863 | assert((VT.is256BitVector() && Op.getNumOperands() == 2) ||((void)0) | |||
| 10864 | (VT.is512BitVector() && (Op.getNumOperands() == 2 ||((void)0) | |||
| 10865 | Op.getNumOperands() == 4)))((void)0); | |||
| 10866 | ||||
| 10867 | // AVX can use the vinsertf128 instruction to create 256-bit vectors | |||
| 10868 | // from two other 128-bit ones. | |||
| 10869 | ||||
| 10870 | // 512-bit vector may contain 2 256-bit vectors or 4 128-bit vectors | |||
| 10871 | return LowerAVXCONCAT_VECTORS(Op, DAG, Subtarget); | |||
| 10872 | } | |||
| 10873 | ||||
| 10874 | //===----------------------------------------------------------------------===// | |||
| 10875 | // Vector shuffle lowering | |||
| 10876 | // | |||
| 10877 | // This is an experimental code path for lowering vector shuffles on x86. It is | |||
| 10878 | // designed to handle arbitrary vector shuffles and blends, gracefully | |||
| 10879 | // degrading performance as necessary. It works hard to recognize idiomatic | |||
| 10880 | // shuffles and lower them to optimal instruction patterns without leaving | |||
| 10881 | // a framework that allows reasonably efficient handling of all vector shuffle | |||
| 10882 | // patterns. | |||
| 10883 | //===----------------------------------------------------------------------===// | |||
| 10884 | ||||
| 10885 | /// Tiny helper function to identify a no-op mask. | |||
| 10886 | /// | |||
| 10887 | /// This is a somewhat boring predicate function. It checks whether the mask | |||
| 10888 | /// array input, which is assumed to be a single-input shuffle mask of the kind | |||
| 10889 | /// used by the X86 shuffle instructions (not a fully general | |||
| 10890 | /// ShuffleVectorSDNode mask) requires any shuffles to occur. Both undef and an | |||
| 10891 | /// in-place shuffle are 'no-op's. | |||
| 10892 | static bool isNoopShuffleMask(ArrayRef<int> Mask) { | |||
| 10893 | for (int i = 0, Size = Mask.size(); i < Size; ++i) { | |||
| 10894 | assert(Mask[i] >= -1 && "Out of bound mask element!")((void)0); | |||
| 10895 | if (Mask[i] >= 0 && Mask[i] != i) | |||
| 10896 | return false; | |||
| 10897 | } | |||
| 10898 | return true; | |||
| 10899 | } | |||
| 10900 | ||||
| 10901 | /// Test whether there are elements crossing LaneSizeInBits lanes in this | |||
| 10902 | /// shuffle mask. | |||
| 10903 | /// | |||
| 10904 | /// X86 divides up its shuffles into in-lane and cross-lane shuffle operations | |||
| 10905 | /// and we routinely test for these. | |||
| 10906 | static bool isLaneCrossingShuffleMask(unsigned LaneSizeInBits, | |||
| 10907 | unsigned ScalarSizeInBits, | |||
| 10908 | ArrayRef<int> Mask) { | |||
| 10909 | assert(LaneSizeInBits && ScalarSizeInBits &&((void)0) | |||
| 10910 | (LaneSizeInBits % ScalarSizeInBits) == 0 &&((void)0) | |||
| 10911 | "Illegal shuffle lane size")((void)0); | |||
| 10912 | int LaneSize = LaneSizeInBits / ScalarSizeInBits; | |||
| 10913 | int Size = Mask.size(); | |||
| 10914 | for (int i = 0; i < Size; ++i) | |||
| 10915 | if (Mask[i] >= 0 && (Mask[i] % Size) / LaneSize != i / LaneSize) | |||
| 10916 | return true; | |||
| 10917 | return false; | |||
| 10918 | } | |||
| 10919 | ||||
| 10920 | /// Test whether there are elements crossing 128-bit lanes in this | |||
| 10921 | /// shuffle mask. | |||
| 10922 | static bool is128BitLaneCrossingShuffleMask(MVT VT, ArrayRef<int> Mask) { | |||
| 10923 | return isLaneCrossingShuffleMask(128, VT.getScalarSizeInBits(), Mask); | |||
| 10924 | } | |||
| 10925 | ||||
| 10926 | /// Test whether elements in each LaneSizeInBits lane in this shuffle mask come | |||
| 10927 | /// from multiple lanes - this is different to isLaneCrossingShuffleMask to | |||
| 10928 | /// better support 'repeated mask + lane permute' style shuffles. | |||
| 10929 | static bool isMultiLaneShuffleMask(unsigned LaneSizeInBits, | |||
| 10930 | unsigned ScalarSizeInBits, | |||
| 10931 | ArrayRef<int> Mask) { | |||
| 10932 | assert(LaneSizeInBits && ScalarSizeInBits &&((void)0) | |||
| 10933 | (LaneSizeInBits % ScalarSizeInBits) == 0 &&((void)0) | |||
| 10934 | "Illegal shuffle lane size")((void)0); | |||
| 10935 | int NumElts = Mask.size(); | |||
| 10936 | int NumEltsPerLane = LaneSizeInBits / ScalarSizeInBits; | |||
| 10937 | int NumLanes = NumElts / NumEltsPerLane; | |||
| 10938 | if (NumLanes > 1) { | |||
| 10939 | for (int i = 0; i != NumLanes; ++i) { | |||
| 10940 | int SrcLane = -1; | |||
| 10941 | for (int j = 0; j != NumEltsPerLane; ++j) { | |||
| 10942 | int M = Mask[(i * NumEltsPerLane) + j]; | |||
| 10943 | if (M < 0) | |||
| 10944 | continue; | |||
| 10945 | int Lane = (M % NumElts) / NumEltsPerLane; | |||
| 10946 | if (SrcLane >= 0 && SrcLane != Lane) | |||
| 10947 | return true; | |||
| 10948 | SrcLane = Lane; | |||
| 10949 | } | |||
| 10950 | } | |||
| 10951 | } | |||
| 10952 | return false; | |||
| 10953 | } | |||
| 10954 | ||||
| 10955 | /// Test whether a shuffle mask is equivalent within each sub-lane. | |||
| 10956 | /// | |||
| 10957 | /// This checks a shuffle mask to see if it is performing the same | |||
| 10958 | /// lane-relative shuffle in each sub-lane. This trivially implies | |||
| 10959 | /// that it is also not lane-crossing. It may however involve a blend from the | |||
| 10960 | /// same lane of a second vector. | |||
| 10961 | /// | |||
| 10962 | /// The specific repeated shuffle mask is populated in \p RepeatedMask, as it is | |||
| 10963 | /// non-trivial to compute in the face of undef lanes. The representation is | |||
| 10964 | /// suitable for use with existing 128-bit shuffles as entries from the second | |||
| 10965 | /// vector have been remapped to [LaneSize, 2*LaneSize). | |||
| 10966 | static bool isRepeatedShuffleMask(unsigned LaneSizeInBits, MVT VT, | |||
| 10967 | ArrayRef<int> Mask, | |||
| 10968 | SmallVectorImpl<int> &RepeatedMask) { | |||
| 10969 | auto LaneSize = LaneSizeInBits / VT.getScalarSizeInBits(); | |||
| 10970 | RepeatedMask.assign(LaneSize, -1); | |||
| 10971 | int Size = Mask.size(); | |||
| 10972 | for (int i = 0; i < Size; ++i) { | |||
| 10973 | assert(Mask[i] == SM_SentinelUndef || Mask[i] >= 0)((void)0); | |||
| 10974 | if (Mask[i] < 0) | |||
| 10975 | continue; | |||
| 10976 | if ((Mask[i] % Size) / LaneSize != i / LaneSize) | |||
| 10977 | // This entry crosses lanes, so there is no way to model this shuffle. | |||
| 10978 | return false; | |||
| 10979 | ||||
| 10980 | // Ok, handle the in-lane shuffles by detecting if and when they repeat. | |||
| 10981 | // Adjust second vector indices to start at LaneSize instead of Size. | |||
| 10982 | int LocalM = Mask[i] < Size ? Mask[i] % LaneSize | |||
| 10983 | : Mask[i] % LaneSize + LaneSize; | |||
| 10984 | if (RepeatedMask[i % LaneSize] < 0) | |||
| 10985 | // This is the first non-undef entry in this slot of a 128-bit lane. | |||
| 10986 | RepeatedMask[i % LaneSize] = LocalM; | |||
| 10987 | else if (RepeatedMask[i % LaneSize] != LocalM) | |||
| 10988 | // Found a mismatch with the repeated mask. | |||
| 10989 | return false; | |||
| 10990 | } | |||
| 10991 | return true; | |||
| 10992 | } | |||
| 10993 | ||||
| 10994 | /// Test whether a shuffle mask is equivalent within each 128-bit lane. | |||
| 10995 | static bool | |||
| 10996 | is128BitLaneRepeatedShuffleMask(MVT VT, ArrayRef<int> Mask, | |||
| 10997 | SmallVectorImpl<int> &RepeatedMask) { | |||
| 10998 | return isRepeatedShuffleMask(128, VT, Mask, RepeatedMask); | |||
| 10999 | } | |||
| 11000 | ||||
| 11001 | static bool | |||
| 11002 | is128BitLaneRepeatedShuffleMask(MVT VT, ArrayRef<int> Mask) { | |||
| 11003 | SmallVector<int, 32> RepeatedMask; | |||
| 11004 | return isRepeatedShuffleMask(128, VT, Mask, RepeatedMask); | |||
| 11005 | } | |||
| 11006 | ||||
| 11007 | /// Test whether a shuffle mask is equivalent within each 256-bit lane. | |||
| 11008 | static bool | |||
| 11009 | is256BitLaneRepeatedShuffleMask(MVT VT, ArrayRef<int> Mask, | |||
| 11010 | SmallVectorImpl<int> &RepeatedMask) { | |||
| 11011 | return isRepeatedShuffleMask(256, VT, Mask, RepeatedMask); | |||
| 11012 | } | |||
| 11013 | ||||
| 11014 | /// Test whether a target shuffle mask is equivalent within each sub-lane. | |||
| 11015 | /// Unlike isRepeatedShuffleMask we must respect SM_SentinelZero. | |||
| 11016 | static bool isRepeatedTargetShuffleMask(unsigned LaneSizeInBits, | |||
| 11017 | unsigned EltSizeInBits, | |||
| 11018 | ArrayRef<int> Mask, | |||
| 11019 | SmallVectorImpl<int> &RepeatedMask) { | |||
| 11020 | int LaneSize = LaneSizeInBits / EltSizeInBits; | |||
| 11021 | RepeatedMask.assign(LaneSize, SM_SentinelUndef); | |||
| 11022 | int Size = Mask.size(); | |||
| 11023 | for (int i = 0; i < Size; ++i) { | |||
| 11024 | assert(isUndefOrZero(Mask[i]) || (Mask[i] >= 0))((void)0); | |||
| 11025 | if (Mask[i] == SM_SentinelUndef) | |||
| 11026 | continue; | |||
| 11027 | if (Mask[i] == SM_SentinelZero) { | |||
| 11028 | if (!isUndefOrZero(RepeatedMask[i % LaneSize])) | |||
| 11029 | return false; | |||
| 11030 | RepeatedMask[i % LaneSize] = SM_SentinelZero; | |||
| 11031 | continue; | |||
| 11032 | } | |||
| 11033 | if ((Mask[i] % Size) / LaneSize != i / LaneSize) | |||
| 11034 | // This entry crosses lanes, so there is no way to model this shuffle. | |||
| 11035 | return false; | |||
| 11036 | ||||
| 11037 | // Handle the in-lane shuffles by detecting if and when they repeat. Adjust | |||
| 11038 | // later vector indices to start at multiples of LaneSize instead of Size. | |||
| 11039 | int LaneM = Mask[i] / Size; | |||
| 11040 | int LocalM = (Mask[i] % LaneSize) + (LaneM * LaneSize); | |||
| 11041 | if (RepeatedMask[i % LaneSize] == SM_SentinelUndef) | |||
| 11042 | // This is the first non-undef entry in this slot of a 128-bit lane. | |||
| 11043 | RepeatedMask[i % LaneSize] = LocalM; | |||
| 11044 | else if (RepeatedMask[i % LaneSize] != LocalM) | |||
| 11045 | // Found a mismatch with the repeated mask. | |||
| 11046 | return false; | |||
| 11047 | } | |||
| 11048 | return true; | |||
| 11049 | } | |||
| 11050 | ||||
| 11051 | /// Test whether a target shuffle mask is equivalent within each sub-lane. | |||
| 11052 | /// Unlike isRepeatedShuffleMask we must respect SM_SentinelZero. | |||
| 11053 | static bool isRepeatedTargetShuffleMask(unsigned LaneSizeInBits, MVT VT, | |||
| 11054 | ArrayRef<int> Mask, | |||
| 11055 | SmallVectorImpl<int> &RepeatedMask) { | |||
| 11056 | return isRepeatedTargetShuffleMask(LaneSizeInBits, VT.getScalarSizeInBits(), | |||
| 11057 | Mask, RepeatedMask); | |||
| 11058 | } | |||
| 11059 | ||||
| 11060 | /// Checks whether the vector elements referenced by two shuffle masks are | |||
| 11061 | /// equivalent. | |||
| 11062 | static bool IsElementEquivalent(int MaskSize, SDValue Op, SDValue ExpectedOp, | |||
| 11063 | int Idx, int ExpectedIdx) { | |||
| 11064 | assert(0 <= Idx && Idx < MaskSize && 0 <= ExpectedIdx &&((void)0) | |||
| 11065 | ExpectedIdx < MaskSize && "Out of range element index")((void)0); | |||
| 11066 | if (!Op || !ExpectedOp || Op.getOpcode() != ExpectedOp.getOpcode()) | |||
| 11067 | return false; | |||
| 11068 | ||||
| 11069 | switch (Op.getOpcode()) { | |||
| 11070 | case ISD::BUILD_VECTOR: | |||
| 11071 | // If the values are build vectors, we can look through them to find | |||
| 11072 | // equivalent inputs that make the shuffles equivalent. | |||
| 11073 | // TODO: Handle MaskSize != Op.getNumOperands()? | |||
| 11074 | if (MaskSize == (int)Op.getNumOperands() && | |||
| 11075 | MaskSize == (int)ExpectedOp.getNumOperands()) | |||
| 11076 | return Op.getOperand(Idx) == ExpectedOp.getOperand(ExpectedIdx); | |||
| 11077 | break; | |||
| 11078 | case X86ISD::VBROADCAST: | |||
| 11079 | case X86ISD::VBROADCAST_LOAD: | |||
| 11080 | // TODO: Handle MaskSize != Op.getValueType().getVectorNumElements()? | |||
| 11081 | return (Op == ExpectedOp && | |||
| 11082 | (int)Op.getValueType().getVectorNumElements() == MaskSize); | |||
| 11083 | case X86ISD::HADD: | |||
| 11084 | case X86ISD::HSUB: | |||
| 11085 | case X86ISD::FHADD: | |||
| 11086 | case X86ISD::FHSUB: | |||
| 11087 | case X86ISD::PACKSS: | |||
| 11088 | case X86ISD::PACKUS: | |||
| 11089 | // HOP(X,X) can refer to the elt from the lower/upper half of a lane. | |||
| 11090 | // TODO: Handle MaskSize != NumElts? | |||
| 11091 | // TODO: Handle HOP(X,Y) vs HOP(Y,X) equivalence cases. | |||
| 11092 | if (Op == ExpectedOp && Op.getOperand(0) == Op.getOperand(1)) { | |||
| 11093 | MVT VT = Op.getSimpleValueType(); | |||
| 11094 | int NumElts = VT.getVectorNumElements(); | |||
| 11095 | if (MaskSize == NumElts) { | |||
| 11096 | int NumLanes = VT.getSizeInBits() / 128; | |||
| 11097 | int NumEltsPerLane = NumElts / NumLanes; | |||
| 11098 | int NumHalfEltsPerLane = NumEltsPerLane / 2; | |||
| 11099 | bool SameLane = | |||
| 11100 | (Idx / NumEltsPerLane) == (ExpectedIdx / NumEltsPerLane); | |||
| 11101 | bool SameElt = | |||
| 11102 | (Idx % NumHalfEltsPerLane) == (ExpectedIdx % NumHalfEltsPerLane); | |||
| 11103 | return SameLane && SameElt; | |||
| 11104 | } | |||
| 11105 | } | |||
| 11106 | break; | |||
| 11107 | } | |||
| 11108 | ||||
| 11109 | return false; | |||
| 11110 | } | |||
| 11111 | ||||
| 11112 | /// Checks whether a shuffle mask is equivalent to an explicit list of | |||
| 11113 | /// arguments. | |||
| 11114 | /// | |||
| 11115 | /// This is a fast way to test a shuffle mask against a fixed pattern: | |||
| 11116 | /// | |||
| 11117 | /// if (isShuffleEquivalent(Mask, 3, 2, {1, 0})) { ... } | |||
| 11118 | /// | |||
| 11119 | /// It returns true if the mask is exactly as wide as the argument list, and | |||
| 11120 | /// each element of the mask is either -1 (signifying undef) or the value given | |||
| 11121 | /// in the argument. | |||
| 11122 | static bool isShuffleEquivalent(ArrayRef<int> Mask, ArrayRef<int> ExpectedMask, | |||
| 11123 | SDValue V1 = SDValue(), | |||
| 11124 | SDValue V2 = SDValue()) { | |||
| 11125 | int Size = Mask.size(); | |||
| 11126 | if (Size != (int)ExpectedMask.size()) | |||
| 11127 | return false; | |||
| 11128 | ||||
| 11129 | for (int i = 0; i < Size; ++i) { | |||
| 11130 | assert(Mask[i] >= -1 && "Out of bound mask element!")((void)0); | |||
| 11131 | int MaskIdx = Mask[i]; | |||
| 11132 | int ExpectedIdx = ExpectedMask[i]; | |||
| 11133 | if (0 <= MaskIdx && MaskIdx != ExpectedIdx) { | |||
| 11134 | SDValue MaskV = MaskIdx < Size ? V1 : V2; | |||
| 11135 | SDValue ExpectedV = ExpectedIdx < Size ? V1 : V2; | |||
| 11136 | MaskIdx = MaskIdx < Size ? MaskIdx : (MaskIdx - Size); | |||
| 11137 | ExpectedIdx = ExpectedIdx < Size ? ExpectedIdx : (ExpectedIdx - Size); | |||
| 11138 | if (!IsElementEquivalent(Size, MaskV, ExpectedV, MaskIdx, ExpectedIdx)) | |||
| 11139 | return false; | |||
| 11140 | } | |||
| 11141 | } | |||
| 11142 | return true; | |||
| 11143 | } | |||
| 11144 | ||||
| 11145 | /// Checks whether a target shuffle mask is equivalent to an explicit pattern. | |||
| 11146 | /// | |||
| 11147 | /// The masks must be exactly the same width. | |||
| 11148 | /// | |||
| 11149 | /// If an element in Mask matches SM_SentinelUndef (-1) then the corresponding | |||
| 11150 | /// value in ExpectedMask is always accepted. Otherwise the indices must match. | |||
| 11151 | /// | |||
| 11152 | /// SM_SentinelZero is accepted as a valid negative index but must match in | |||
| 11153 | /// both. | |||
| 11154 | static bool isTargetShuffleEquivalent(MVT VT, ArrayRef<int> Mask, | |||
| 11155 | ArrayRef<int> ExpectedMask, | |||
| 11156 | SDValue V1 = SDValue(), | |||
| 11157 | SDValue V2 = SDValue()) { | |||
| 11158 | int Size = Mask.size(); | |||
| 11159 | if (Size != (int)ExpectedMask.size()) | |||
| 11160 | return false; | |||
| 11161 | assert(isUndefOrZeroOrInRange(ExpectedMask, 0, 2 * Size) &&((void)0) | |||
| 11162 | "Illegal target shuffle mask")((void)0); | |||
| 11163 | ||||
| 11164 | // Check for out-of-range target shuffle mask indices. | |||
| 11165 | if (!isUndefOrZeroOrInRange(Mask, 0, 2 * Size)) | |||
| 11166 | return false; | |||
| 11167 | ||||
| 11168 | // Don't use V1/V2 if they're not the same size as the shuffle mask type. | |||
| 11169 | if (V1 && V1.getValueSizeInBits() != VT.getSizeInBits()) | |||
| 11170 | V1 = SDValue(); | |||
| 11171 | if (V2 && V2.getValueSizeInBits() != VT.getSizeInBits()) | |||
| 11172 | V2 = SDValue(); | |||
| 11173 | ||||
| 11174 | for (int i = 0; i < Size; ++i) { | |||
| 11175 | int MaskIdx = Mask[i]; | |||
| 11176 | int ExpectedIdx = ExpectedMask[i]; | |||
| 11177 | if (MaskIdx == SM_SentinelUndef || MaskIdx == ExpectedIdx) | |||
| 11178 | continue; | |||
| 11179 | if (0 <= MaskIdx && 0 <= ExpectedIdx) { | |||
| 11180 | SDValue MaskV = MaskIdx < Size ? V1 : V2; | |||
| 11181 | SDValue ExpectedV = ExpectedIdx < Size ? V1 : V2; | |||
| 11182 | MaskIdx = MaskIdx < Size ? MaskIdx : (MaskIdx - Size); | |||
| 11183 | ExpectedIdx = ExpectedIdx < Size ? ExpectedIdx : (ExpectedIdx - Size); | |||
| 11184 | if (IsElementEquivalent(Size, MaskV, ExpectedV, MaskIdx, ExpectedIdx)) | |||
| 11185 | continue; | |||
| 11186 | } | |||
| 11187 | // TODO - handle SM_Sentinel equivalences. | |||
| 11188 | return false; | |||
| 11189 | } | |||
| 11190 | return true; | |||
| 11191 | } | |||
| 11192 | ||||
| 11193 | // Attempt to create a shuffle mask from a VSELECT condition mask. | |||
| 11194 | static bool createShuffleMaskFromVSELECT(SmallVectorImpl<int> &Mask, | |||
| 11195 | SDValue Cond) { | |||
| 11196 | EVT CondVT = Cond.getValueType(); | |||
| 11197 | unsigned EltSizeInBits = CondVT.getScalarSizeInBits(); | |||
| 11198 | unsigned NumElts = CondVT.getVectorNumElements(); | |||
| 11199 | ||||
| 11200 | APInt UndefElts; | |||
| 11201 | SmallVector<APInt, 32> EltBits; | |||
| 11202 | if (!getTargetConstantBitsFromNode(Cond, EltSizeInBits, UndefElts, EltBits, | |||
| 11203 | true, false)) | |||
| 11204 | return false; | |||
| 11205 | ||||
| 11206 | Mask.resize(NumElts, SM_SentinelUndef); | |||
| 11207 | ||||
| 11208 | for (int i = 0; i != (int)NumElts; ++i) { | |||
| 11209 | Mask[i] = i; | |||
| 11210 | // Arbitrarily choose from the 2nd operand if the select condition element | |||
| 11211 | // is undef. | |||
| 11212 | // TODO: Can we do better by matching patterns such as even/odd? | |||
| 11213 | if (UndefElts[i] || EltBits[i].isNullValue()) | |||
| 11214 | Mask[i] += NumElts; | |||
| 11215 | } | |||
| 11216 | ||||
| 11217 | return true; | |||
| 11218 | } | |||
| 11219 | ||||
| 11220 | // Check if the shuffle mask is suitable for the AVX vpunpcklwd or vpunpckhwd | |||
| 11221 | // instructions. | |||
| 11222 | static bool isUnpackWdShuffleMask(ArrayRef<int> Mask, MVT VT) { | |||
| 11223 | if (VT != MVT::v8i32 && VT != MVT::v8f32) | |||
| 11224 | return false; | |||
| 11225 | ||||
| 11226 | SmallVector<int, 8> Unpcklwd; | |||
| 11227 | createUnpackShuffleMask(MVT::v8i16, Unpcklwd, /* Lo = */ true, | |||
| 11228 | /* Unary = */ false); | |||
| 11229 | SmallVector<int, 8> Unpckhwd; | |||
| 11230 | createUnpackShuffleMask(MVT::v8i16, Unpckhwd, /* Lo = */ false, | |||
| 11231 | /* Unary = */ false); | |||
| 11232 | bool IsUnpackwdMask = (isTargetShuffleEquivalent(VT, Mask, Unpcklwd) || | |||
| 11233 | isTargetShuffleEquivalent(VT, Mask, Unpckhwd)); | |||
| 11234 | return IsUnpackwdMask; | |||
| 11235 | } | |||
| 11236 | ||||
| 11237 | static bool is128BitUnpackShuffleMask(ArrayRef<int> Mask) { | |||
| 11238 | // Create 128-bit vector type based on mask size. | |||
| 11239 | MVT EltVT = MVT::getIntegerVT(128 / Mask.size()); | |||
| 11240 | MVT VT = MVT::getVectorVT(EltVT, Mask.size()); | |||
| 11241 | ||||
| 11242 | // We can't assume a canonical shuffle mask, so try the commuted version too. | |||
| 11243 | SmallVector<int, 4> CommutedMask(Mask.begin(), Mask.end()); | |||
| 11244 | ShuffleVectorSDNode::commuteMask(CommutedMask); | |||
| 11245 | ||||
| 11246 | // Match any of unary/binary or low/high. | |||
| 11247 | for (unsigned i = 0; i != 4; ++i) { | |||
| 11248 | SmallVector<int, 16> UnpackMask; | |||
| 11249 | createUnpackShuffleMask(VT, UnpackMask, (i >> 1) % 2, i % 2); | |||
| 11250 | if (isTargetShuffleEquivalent(VT, Mask, UnpackMask) || | |||
| 11251 | isTargetShuffleEquivalent(VT, CommutedMask, UnpackMask)) | |||
| 11252 | return true; | |||
| 11253 | } | |||
| 11254 | return false; | |||
| 11255 | } | |||
| 11256 | ||||
| 11257 | /// Return true if a shuffle mask chooses elements identically in its top and | |||
| 11258 | /// bottom halves. For example, any splat mask has the same top and bottom | |||
| 11259 | /// halves. If an element is undefined in only one half of the mask, the halves | |||
| 11260 | /// are not considered identical. | |||
| 11261 | static bool hasIdenticalHalvesShuffleMask(ArrayRef<int> Mask) { | |||
| 11262 | assert(Mask.size() % 2 == 0 && "Expecting even number of elements in mask")((void)0); | |||
| 11263 | unsigned HalfSize = Mask.size() / 2; | |||
| 11264 | for (unsigned i = 0; i != HalfSize; ++i) { | |||
| 11265 | if (Mask[i] != Mask[i + HalfSize]) | |||
| 11266 | return false; | |||
| 11267 | } | |||
| 11268 | return true; | |||
| 11269 | } | |||
| 11270 | ||||
| 11271 | /// Get a 4-lane 8-bit shuffle immediate for a mask. | |||
| 11272 | /// | |||
| 11273 | /// This helper function produces an 8-bit shuffle immediate corresponding to | |||
| 11274 | /// the ubiquitous shuffle encoding scheme used in x86 instructions for | |||
| 11275 | /// shuffling 4 lanes. It can be used with most of the PSHUF instructions for | |||
| 11276 | /// example. | |||
| 11277 | /// | |||
| 11278 | /// NB: We rely heavily on "undef" masks preserving the input lane. | |||
| 11279 | static unsigned getV4X86ShuffleImm(ArrayRef<int> Mask) { | |||
| 11280 | assert(Mask.size() == 4 && "Only 4-lane shuffle masks")((void)0); | |||
| 11281 | assert(Mask[0] >= -1 && Mask[0] < 4 && "Out of bound mask element!")((void)0); | |||
| 11282 | assert(Mask[1] >= -1 && Mask[1] < 4 && "Out of bound mask element!")((void)0); | |||
| 11283 | assert(Mask[2] >= -1 && Mask[2] < 4 && "Out of bound mask element!")((void)0); | |||
| 11284 | assert(Mask[3] >= -1 && Mask[3] < 4 && "Out of bound mask element!")((void)0); | |||
| 11285 | ||||
| 11286 | // If the mask only uses one non-undef element, then fully 'splat' it to | |||
| 11287 | // improve later broadcast matching. | |||
| 11288 | int FirstIndex = find_if(Mask, [](int M) { return M >= 0; }) - Mask.begin(); | |||
| 11289 | assert(0 <= FirstIndex && FirstIndex < 4 && "All undef shuffle mask")((void)0); | |||
| 11290 | ||||
| 11291 | int FirstElt = Mask[FirstIndex]; | |||
| 11292 | if (all_of(Mask, [FirstElt](int M) { return M < 0 || M == FirstElt; })) | |||
| 11293 | return (FirstElt << 6) | (FirstElt << 4) | (FirstElt << 2) | FirstElt; | |||
| 11294 | ||||
| 11295 | unsigned Imm = 0; | |||
| 11296 | Imm |= (Mask[0] < 0 ? 0 : Mask[0]) << 0; | |||
| 11297 | Imm |= (Mask[1] < 0 ? 1 : Mask[1]) << 2; | |||
| 11298 | Imm |= (Mask[2] < 0 ? 2 : Mask[2]) << 4; | |||
| 11299 | Imm |= (Mask[3] < 0 ? 3 : Mask[3]) << 6; | |||
| 11300 | return Imm; | |||
| 11301 | } | |||
| 11302 | ||||
| 11303 | static SDValue getV4X86ShuffleImm8ForMask(ArrayRef<int> Mask, const SDLoc &DL, | |||
| 11304 | SelectionDAG &DAG) { | |||
| 11305 | return DAG.getTargetConstant(getV4X86ShuffleImm(Mask), DL, MVT::i8); | |||
| 11306 | } | |||
| 11307 | ||||
| 11308 | // The Shuffle result is as follow: | |||
| 11309 | // 0*a[0]0*a[1]...0*a[n] , n >=0 where a[] elements in a ascending order. | |||
| 11310 | // Each Zeroable's element correspond to a particular Mask's element. | |||
| 11311 | // As described in computeZeroableShuffleElements function. | |||
| 11312 | // | |||
| 11313 | // The function looks for a sub-mask that the nonzero elements are in | |||
| 11314 | // increasing order. If such sub-mask exist. The function returns true. | |||
| 11315 | static bool isNonZeroElementsInOrder(const APInt &Zeroable, | |||
| 11316 | ArrayRef<int> Mask, const EVT &VectorType, | |||
| 11317 | bool &IsZeroSideLeft) { | |||
| 11318 | int NextElement = -1; | |||
| 11319 | // Check if the Mask's nonzero elements are in increasing order. | |||
| 11320 | for (int i = 0, e = Mask.size(); i < e; i++) { | |||
| 11321 | // Checks if the mask's zeros elements are built from only zeros. | |||
| 11322 | assert(Mask[i] >= -1 && "Out of bound mask element!")((void)0); | |||
| 11323 | if (Mask[i] < 0) | |||
| 11324 | return false; | |||
| 11325 | if (Zeroable[i]) | |||
| 11326 | continue; | |||
| 11327 | // Find the lowest non zero element | |||
| 11328 | if (NextElement < 0) { | |||
| 11329 | NextElement = Mask[i] != 0 ? VectorType.getVectorNumElements() : 0; | |||
| 11330 | IsZeroSideLeft = NextElement != 0; | |||
| 11331 | } | |||
| 11332 | // Exit if the mask's non zero elements are not in increasing order. | |||
| 11333 | if (NextElement != Mask[i]) | |||
| 11334 | return false; | |||
| 11335 | NextElement++; | |||
| 11336 | } | |||
| 11337 | return true; | |||
| 11338 | } | |||
| 11339 | ||||
| 11340 | /// Try to lower a shuffle with a single PSHUFB of V1 or V2. | |||
| 11341 | static SDValue lowerShuffleWithPSHUFB(const SDLoc &DL, MVT VT, | |||
| 11342 | ArrayRef<int> Mask, SDValue V1, | |||
| 11343 | SDValue V2, const APInt &Zeroable, | |||
| 11344 | const X86Subtarget &Subtarget, | |||
| 11345 | SelectionDAG &DAG) { | |||
| 11346 | int Size = Mask.size(); | |||
| 11347 | int LaneSize = 128 / VT.getScalarSizeInBits(); | |||
| 11348 | const int NumBytes = VT.getSizeInBits() / 8; | |||
| 11349 | const int NumEltBytes = VT.getScalarSizeInBits() / 8; | |||
| 11350 | ||||
| 11351 | assert((Subtarget.hasSSSE3() && VT.is128BitVector()) ||((void)0) | |||
| 11352 | (Subtarget.hasAVX2() && VT.is256BitVector()) ||((void)0) | |||
| 11353 | (Subtarget.hasBWI() && VT.is512BitVector()))((void)0); | |||
| 11354 | ||||
| 11355 | SmallVector<SDValue, 64> PSHUFBMask(NumBytes); | |||
| 11356 | // Sign bit set in i8 mask means zero element. | |||
| 11357 | SDValue ZeroMask = DAG.getConstant(0x80, DL, MVT::i8); | |||
| 11358 | ||||
| 11359 | SDValue V; | |||
| 11360 | for (int i = 0; i < NumBytes; ++i) { | |||
| 11361 | int M = Mask[i / NumEltBytes]; | |||
| 11362 | if (M < 0) { | |||
| 11363 | PSHUFBMask[i] = DAG.getUNDEF(MVT::i8); | |||
| 11364 | continue; | |||
| 11365 | } | |||
| 11366 | if (Zeroable[i / NumEltBytes]) { | |||
| 11367 | PSHUFBMask[i] = ZeroMask; | |||
| 11368 | continue; | |||
| 11369 | } | |||
| 11370 | ||||
| 11371 | // We can only use a single input of V1 or V2. | |||
| 11372 | SDValue SrcV = (M >= Size ? V2 : V1); | |||
| 11373 | if (V && V != SrcV) | |||
| 11374 | return SDValue(); | |||
| 11375 | V = SrcV; | |||
| 11376 | M %= Size; | |||
| 11377 | ||||
| 11378 | // PSHUFB can't cross lanes, ensure this doesn't happen. | |||
| 11379 | if ((M / LaneSize) != ((i / NumEltBytes) / LaneSize)) | |||
| 11380 | return SDValue(); | |||
| 11381 | ||||
| 11382 | M = M % LaneSize; | |||
| 11383 | M = M * NumEltBytes + (i % NumEltBytes); | |||
| 11384 | PSHUFBMask[i] = DAG.getConstant(M, DL, MVT::i8); | |||
| 11385 | } | |||
| 11386 | assert(V && "Failed to find a source input")((void)0); | |||
| 11387 | ||||
| 11388 | MVT I8VT = MVT::getVectorVT(MVT::i8, NumBytes); | |||
| 11389 | return DAG.getBitcast( | |||
| 11390 | VT, DAG.getNode(X86ISD::PSHUFB, DL, I8VT, DAG.getBitcast(I8VT, V), | |||
| 11391 | DAG.getBuildVector(I8VT, DL, PSHUFBMask))); | |||
| 11392 | } | |||
| 11393 | ||||
| 11394 | static SDValue getMaskNode(SDValue Mask, MVT MaskVT, | |||
| 11395 | const X86Subtarget &Subtarget, SelectionDAG &DAG, | |||
| 11396 | const SDLoc &dl); | |||
| 11397 | ||||
| 11398 | // X86 has dedicated shuffle that can be lowered to VEXPAND | |||
| 11399 | static SDValue lowerShuffleToEXPAND(const SDLoc &DL, MVT VT, | |||
| 11400 | const APInt &Zeroable, | |||
| 11401 | ArrayRef<int> Mask, SDValue &V1, | |||
| 11402 | SDValue &V2, SelectionDAG &DAG, | |||
| 11403 | const X86Subtarget &Subtarget) { | |||
| 11404 | bool IsLeftZeroSide = true; | |||
| 11405 | if (!isNonZeroElementsInOrder(Zeroable, Mask, V1.getValueType(), | |||
| 11406 | IsLeftZeroSide)) | |||
| 11407 | return SDValue(); | |||
| 11408 | unsigned VEXPANDMask = (~Zeroable).getZExtValue(); | |||
| 11409 | MVT IntegerType = | |||
| 11410 | MVT::getIntegerVT(std::max((int)VT.getVectorNumElements(), 8)); | |||
| 11411 | SDValue MaskNode = DAG.getConstant(VEXPANDMask, DL, IntegerType); | |||
| 11412 | unsigned NumElts = VT.getVectorNumElements(); | |||
| 11413 | assert((NumElts == 4 || NumElts == 8 || NumElts == 16) &&((void)0) | |||
| 11414 | "Unexpected number of vector elements")((void)0); | |||
| 11415 | SDValue VMask = getMaskNode(MaskNode, MVT::getVectorVT(MVT::i1, NumElts), | |||
| 11416 | Subtarget, DAG, DL); | |||
| 11417 | SDValue ZeroVector = getZeroVector(VT, Subtarget, DAG, DL); | |||
| 11418 | SDValue ExpandedVector = IsLeftZeroSide ? V2 : V1; | |||
| 11419 | return DAG.getNode(X86ISD::EXPAND, DL, VT, ExpandedVector, ZeroVector, VMask); | |||
| 11420 | } | |||
| 11421 | ||||
| 11422 | static bool matchShuffleWithUNPCK(MVT VT, SDValue &V1, SDValue &V2, | |||
| 11423 | unsigned &UnpackOpcode, bool IsUnary, | |||
| 11424 | ArrayRef<int> TargetMask, const SDLoc &DL, | |||
| 11425 | SelectionDAG &DAG, | |||
| 11426 | const X86Subtarget &Subtarget) { | |||
| 11427 | int NumElts = VT.getVectorNumElements(); | |||
| 11428 | ||||
| 11429 | bool Undef1 = true, Undef2 = true, Zero1 = true, Zero2 = true; | |||
| 11430 | for (int i = 0; i != NumElts; i += 2) { | |||
| 11431 | int M1 = TargetMask[i + 0]; | |||
| 11432 | int M2 = TargetMask[i + 1]; | |||
| 11433 | Undef1 &= (SM_SentinelUndef == M1); | |||
| 11434 | Undef2 &= (SM_SentinelUndef == M2); | |||
| 11435 | Zero1 &= isUndefOrZero(M1); | |||
| 11436 | Zero2 &= isUndefOrZero(M2); | |||
| 11437 | } | |||
| 11438 | assert(!((Undef1 || Zero1) && (Undef2 || Zero2)) &&((void)0) | |||
| 11439 | "Zeroable shuffle detected")((void)0); | |||
| 11440 | ||||
| 11441 | // Attempt to match the target mask against the unpack lo/hi mask patterns. | |||
| 11442 | SmallVector<int, 64> Unpckl, Unpckh; | |||
| 11443 | createUnpackShuffleMask(VT, Unpckl, /* Lo = */ true, IsUnary); | |||
| 11444 | if (isTargetShuffleEquivalent(VT, TargetMask, Unpckl, V1, | |||
| 11445 | (IsUnary ? V1 : V2))) { | |||
| 11446 | UnpackOpcode = X86ISD::UNPCKL; | |||
| 11447 | V2 = (Undef2 ? DAG.getUNDEF(VT) : (IsUnary ? V1 : V2)); | |||
| 11448 | V1 = (Undef1 ? DAG.getUNDEF(VT) : V1); | |||
| 11449 | return true; | |||
| 11450 | } | |||
| 11451 | ||||
| 11452 | createUnpackShuffleMask(VT, Unpckh, /* Lo = */ false, IsUnary); | |||
| 11453 | if (isTargetShuffleEquivalent(VT, TargetMask, Unpckh, V1, | |||
| 11454 | (IsUnary ? V1 : V2))) { | |||
| 11455 | UnpackOpcode = X86ISD::UNPCKH; | |||
| 11456 | V2 = (Undef2 ? DAG.getUNDEF(VT) : (IsUnary ? V1 : V2)); | |||
| 11457 | V1 = (Undef1 ? DAG.getUNDEF(VT) : V1); | |||
| 11458 | return true; | |||
| 11459 | } | |||
| 11460 | ||||
| 11461 | // If an unary shuffle, attempt to match as an unpack lo/hi with zero. | |||
| 11462 | if (IsUnary && (Zero1 || Zero2)) { | |||
| 11463 | // Don't bother if we can blend instead. | |||
| 11464 | if ((Subtarget.hasSSE41() || VT == MVT::v2i64 || VT == MVT::v2f64) && | |||
| 11465 | isSequentialOrUndefOrZeroInRange(TargetMask, 0, NumElts, 0)) | |||
| 11466 | return false; | |||
| 11467 | ||||
| 11468 | bool MatchLo = true, MatchHi = true; | |||
| 11469 | for (int i = 0; (i != NumElts) && (MatchLo || MatchHi); ++i) { | |||
| 11470 | int M = TargetMask[i]; | |||
| 11471 | ||||
| 11472 | // Ignore if the input is known to be zero or the index is undef. | |||
| 11473 | if ((((i & 1) == 0) && Zero1) || (((i & 1) == 1) && Zero2) || | |||
| 11474 | (M == SM_SentinelUndef)) | |||
| 11475 | continue; | |||
| 11476 | ||||
| 11477 | MatchLo &= (M == Unpckl[i]); | |||
| 11478 | MatchHi &= (M == Unpckh[i]); | |||
| 11479 | } | |||
| 11480 | ||||
| 11481 | if (MatchLo || MatchHi) { | |||
| 11482 | UnpackOpcode = MatchLo ? X86ISD::UNPCKL : X86ISD::UNPCKH; | |||
| 11483 | V2 = Zero2 ? getZeroVector(VT, Subtarget, DAG, DL) : V1; | |||
| 11484 | V1 = Zero1 ? getZeroVector(VT, Subtarget, DAG, DL) : V1; | |||
| 11485 | return true; | |||
| 11486 | } | |||
| 11487 | } | |||
| 11488 | ||||
| 11489 | // If a binary shuffle, commute and try again. | |||
| 11490 | if (!IsUnary) { | |||
| 11491 | ShuffleVectorSDNode::commuteMask(Unpckl); | |||
| 11492 | if (isTargetShuffleEquivalent(VT, TargetMask, Unpckl)) { | |||
| 11493 | UnpackOpcode = X86ISD::UNPCKL; | |||
| 11494 | std::swap(V1, V2); | |||
| 11495 | return true; | |||
| 11496 | } | |||
| 11497 | ||||
| 11498 | ShuffleVectorSDNode::commuteMask(Unpckh); | |||
| 11499 | if (isTargetShuffleEquivalent(VT, TargetMask, Unpckh)) { | |||
| 11500 | UnpackOpcode = X86ISD::UNPCKH; | |||
| 11501 | std::swap(V1, V2); | |||
| 11502 | return true; | |||
| 11503 | } | |||
| 11504 | } | |||
| 11505 | ||||
| 11506 | return false; | |||
| 11507 | } | |||
| 11508 | ||||
| 11509 | // X86 has dedicated unpack instructions that can handle specific blend | |||
| 11510 | // operations: UNPCKH and UNPCKL. | |||
| 11511 | static SDValue lowerShuffleWithUNPCK(const SDLoc &DL, MVT VT, | |||
| 11512 | ArrayRef<int> Mask, SDValue V1, SDValue V2, | |||
| 11513 | SelectionDAG &DAG) { | |||
| 11514 | SmallVector<int, 8> Unpckl; | |||
| 11515 | createUnpackShuffleMask(VT, Unpckl, /* Lo = */ true, /* Unary = */ false); | |||
| 11516 | if (isShuffleEquivalent(Mask, Unpckl, V1, V2)) | |||
| 11517 | return DAG.getNode(X86ISD::UNPCKL, DL, VT, V1, V2); | |||
| 11518 | ||||
| 11519 | SmallVector<int, 8> Unpckh; | |||
| 11520 | createUnpackShuffleMask(VT, Unpckh, /* Lo = */ false, /* Unary = */ false); | |||
| 11521 | if (isShuffleEquivalent(Mask, Unpckh, V1, V2)) | |||
| 11522 | return DAG.getNode(X86ISD::UNPCKH, DL, VT, V1, V2); | |||
| 11523 | ||||
| 11524 | // Commute and try again. | |||
| 11525 | ShuffleVectorSDNode::commuteMask(Unpckl); | |||
| 11526 | if (isShuffleEquivalent(Mask, Unpckl, V1, V2)) | |||
| 11527 | return DAG.getNode(X86ISD::UNPCKL, DL, VT, V2, V1); | |||
| 11528 | ||||
| 11529 | ShuffleVectorSDNode::commuteMask(Unpckh); | |||
| 11530 | if (isShuffleEquivalent(Mask, Unpckh, V1, V2)) | |||
| 11531 | return DAG.getNode(X86ISD::UNPCKH, DL, VT, V2, V1); | |||
| 11532 | ||||
| 11533 | return SDValue(); | |||
| 11534 | } | |||
| 11535 | ||||
| 11536 | /// Check if the mask can be mapped to a preliminary shuffle (vperm 64-bit) | |||
| 11537 | /// followed by unpack 256-bit. | |||
| 11538 | static SDValue lowerShuffleWithUNPCK256(const SDLoc &DL, MVT VT, | |||
| 11539 | ArrayRef<int> Mask, SDValue V1, | |||
| 11540 | SDValue V2, SelectionDAG &DAG) { | |||
| 11541 | SmallVector<int, 32> Unpckl, Unpckh; | |||
| 11542 | createSplat2ShuffleMask(VT, Unpckl, /* Lo */ true); | |||
| 11543 | createSplat2ShuffleMask(VT, Unpckh, /* Lo */ false); | |||
| 11544 | ||||
| 11545 | unsigned UnpackOpcode; | |||
| 11546 | if (isShuffleEquivalent(Mask, Unpckl, V1, V2)) | |||
| 11547 | UnpackOpcode = X86ISD::UNPCKL; | |||
| 11548 | else if (isShuffleEquivalent(Mask, Unpckh, V1, V2)) | |||
| 11549 | UnpackOpcode = X86ISD::UNPCKH; | |||
| 11550 | else | |||
| 11551 | return SDValue(); | |||
| 11552 | ||||
| 11553 | // This is a "natural" unpack operation (rather than the 128-bit sectored | |||
| 11554 | // operation implemented by AVX). We need to rearrange 64-bit chunks of the | |||
| 11555 | // input in order to use the x86 instruction. | |||
| 11556 | V1 = DAG.getVectorShuffle(MVT::v4f64, DL, DAG.getBitcast(MVT::v4f64, V1), | |||
| 11557 | DAG.getUNDEF(MVT::v4f64), {0, 2, 1, 3}); | |||
| 11558 | V1 = DAG.getBitcast(VT, V1); | |||
| 11559 | return DAG.getNode(UnpackOpcode, DL, VT, V1, V1); | |||
| 11560 | } | |||
| 11561 | ||||
| 11562 | // Check if the mask can be mapped to a TRUNCATE or VTRUNC, truncating the | |||
| 11563 | // source into the lower elements and zeroing the upper elements. | |||
| 11564 | static bool matchShuffleAsVTRUNC(MVT &SrcVT, MVT &DstVT, MVT VT, | |||
| 11565 | ArrayRef<int> Mask, const APInt &Zeroable, | |||
| 11566 | const X86Subtarget &Subtarget) { | |||
| 11567 | if (!VT.is512BitVector() && !Subtarget.hasVLX()) | |||
| 11568 | return false; | |||
| 11569 | ||||
| 11570 | unsigned NumElts = Mask.size(); | |||
| 11571 | unsigned EltSizeInBits = VT.getScalarSizeInBits(); | |||
| 11572 | unsigned MaxScale = 64 / EltSizeInBits; | |||
| 11573 | ||||
| 11574 | for (unsigned Scale = 2; Scale <= MaxScale; Scale += Scale) { | |||
| 11575 | unsigned SrcEltBits = EltSizeInBits * Scale; | |||
| 11576 | if (SrcEltBits < 32 && !Subtarget.hasBWI()) | |||
| 11577 | continue; | |||
| 11578 | unsigned NumSrcElts = NumElts / Scale; | |||
| 11579 | if (!isSequentialOrUndefInRange(Mask, 0, NumSrcElts, 0, Scale)) | |||
| 11580 | continue; | |||
| 11581 | unsigned UpperElts = NumElts - NumSrcElts; | |||
| 11582 | if (!Zeroable.extractBits(UpperElts, NumSrcElts).isAllOnesValue()) | |||
| 11583 | continue; | |||
| 11584 | SrcVT = MVT::getIntegerVT(EltSizeInBits * Scale); | |||
| 11585 | SrcVT = MVT::getVectorVT(SrcVT, NumSrcElts); | |||
| 11586 | DstVT = MVT::getIntegerVT(EltSizeInBits); | |||
| 11587 | if ((NumSrcElts * EltSizeInBits) >= 128) { | |||
| 11588 | // ISD::TRUNCATE | |||
| 11589 | DstVT = MVT::getVectorVT(DstVT, NumSrcElts); | |||
| 11590 | } else { | |||
| 11591 | // X86ISD::VTRUNC | |||
| 11592 | DstVT = MVT::getVectorVT(DstVT, 128 / EltSizeInBits); | |||
| 11593 | } | |||
| 11594 | return true; | |||
| 11595 | } | |||
| 11596 | ||||
| 11597 | return false; | |||
| 11598 | } | |||
| 11599 | ||||
| 11600 | // Helper to create TRUNCATE/VTRUNC nodes, optionally with zero/undef upper | |||
| 11601 | // element padding to the final DstVT. | |||
| 11602 | static SDValue getAVX512TruncNode(const SDLoc &DL, MVT DstVT, SDValue Src, | |||
| 11603 | const X86Subtarget &Subtarget, | |||
| 11604 | SelectionDAG &DAG, bool ZeroUppers) { | |||
| 11605 | MVT SrcVT = Src.getSimpleValueType(); | |||
| 11606 | MVT DstSVT = DstVT.getScalarType(); | |||
| 11607 | unsigned NumDstElts = DstVT.getVectorNumElements(); | |||
| 11608 | unsigned NumSrcElts = SrcVT.getVectorNumElements(); | |||
| 11609 | unsigned DstEltSizeInBits = DstVT.getScalarSizeInBits(); | |||
| 11610 | ||||
| 11611 | if (!DAG.getTargetLoweringInfo().isTypeLegal(SrcVT)) | |||
| 11612 | return SDValue(); | |||
| 11613 | ||||
| 11614 | // Perform a direct ISD::TRUNCATE if possible. | |||
| 11615 | if (NumSrcElts == NumDstElts) | |||
| 11616 | return DAG.getNode(ISD::TRUNCATE, DL, DstVT, Src); | |||
| 11617 | ||||
| 11618 | if (NumSrcElts > NumDstElts) { | |||
| 11619 | MVT TruncVT = MVT::getVectorVT(DstSVT, NumSrcElts); | |||
| 11620 | SDValue Trunc = DAG.getNode(ISD::TRUNCATE, DL, TruncVT, Src); | |||
| 11621 | return extractSubVector(Trunc, 0, DAG, DL, DstVT.getSizeInBits()); | |||
| 11622 | } | |||
| 11623 | ||||
| 11624 | if ((NumSrcElts * DstEltSizeInBits) >= 128) { | |||
| 11625 | MVT TruncVT = MVT::getVectorVT(DstSVT, NumSrcElts); | |||
| 11626 | SDValue Trunc = DAG.getNode(ISD::TRUNCATE, DL, TruncVT, Src); | |||
| 11627 | return widenSubVector(Trunc, ZeroUppers, Subtarget, DAG, DL, | |||
| 11628 | DstVT.getSizeInBits()); | |||
| 11629 | } | |||
| 11630 | ||||
| 11631 | // Non-VLX targets must truncate from a 512-bit type, so we need to | |||
| 11632 | // widen, truncate and then possibly extract the original subvector. | |||
| 11633 | if (!Subtarget.hasVLX() && !SrcVT.is512BitVector()) { | |||
| 11634 | SDValue NewSrc = widenSubVector(Src, ZeroUppers, Subtarget, DAG, DL, 512); | |||
| 11635 | return getAVX512TruncNode(DL, DstVT, NewSrc, Subtarget, DAG, ZeroUppers); | |||
| 11636 | } | |||
| 11637 | ||||
| 11638 | // Fallback to a X86ISD::VTRUNC, padding if necessary. | |||
| 11639 | MVT TruncVT = MVT::getVectorVT(DstSVT, 128 / DstEltSizeInBits); | |||
| 11640 | SDValue Trunc = DAG.getNode(X86ISD::VTRUNC, DL, TruncVT, Src); | |||
| 11641 | if (DstVT != TruncVT) | |||
| 11642 | Trunc = widenSubVector(Trunc, ZeroUppers, Subtarget, DAG, DL, | |||
| 11643 | DstVT.getSizeInBits()); | |||
| 11644 | return Trunc; | |||
| 11645 | } | |||
| 11646 | ||||
| 11647 | // Try to lower trunc+vector_shuffle to a vpmovdb or a vpmovdw instruction. | |||
| 11648 | // | |||
| 11649 | // An example is the following: | |||
| 11650 | // | |||
| 11651 | // t0: ch = EntryToken | |||
| 11652 | // t2: v4i64,ch = CopyFromReg t0, Register:v4i64 %0 | |||
| 11653 | // t25: v4i32 = truncate t2 | |||
| 11654 | // t41: v8i16 = bitcast t25 | |||
| 11655 | // t21: v8i16 = BUILD_VECTOR undef:i16, undef:i16, undef:i16, undef:i16, | |||
| 11656 | // Constant:i16<0>, Constant:i16<0>, Constant:i16<0>, Constant:i16<0> | |||
| 11657 | // t51: v8i16 = vector_shuffle<0,2,4,6,12,13,14,15> t41, t21 | |||
| 11658 | // t18: v2i64 = bitcast t51 | |||
| 11659 | // | |||
| 11660 | // One can just use a single vpmovdw instruction, without avx512vl we need to | |||
| 11661 | // use the zmm variant and extract the lower subvector, padding with zeroes. | |||
| 11662 | // TODO: Merge with lowerShuffleAsVTRUNC. | |||
| 11663 | static SDValue lowerShuffleWithVPMOV(const SDLoc &DL, MVT VT, SDValue V1, | |||
| 11664 | SDValue V2, ArrayRef<int> Mask, | |||
| 11665 | const APInt &Zeroable, | |||
| 11666 | const X86Subtarget &Subtarget, | |||
| 11667 | SelectionDAG &DAG) { | |||
| 11668 | assert((VT == MVT::v16i8 || VT == MVT::v8i16) && "Unexpected VTRUNC type")((void)0); | |||
| 11669 | if (!Subtarget.hasAVX512()) | |||
| 11670 | return SDValue(); | |||
| 11671 | ||||
| 11672 | unsigned NumElts = VT.getVectorNumElements(); | |||
| 11673 | unsigned EltSizeInBits = VT.getScalarSizeInBits(); | |||
| 11674 | unsigned MaxScale = 64 / EltSizeInBits; | |||
| 11675 | for (unsigned Scale = 2; Scale <= MaxScale; Scale += Scale) { | |||
| 11676 | unsigned NumSrcElts = NumElts / Scale; | |||
| 11677 | unsigned UpperElts = NumElts - NumSrcElts; | |||
| 11678 | if (!isSequentialOrUndefInRange(Mask, 0, NumSrcElts, 0, Scale) || | |||
| 11679 | !Zeroable.extractBits(UpperElts, NumSrcElts).isAllOnesValue()) | |||
| 11680 | continue; | |||
| 11681 | ||||
| 11682 | SDValue Src = V1; | |||
| 11683 | if (!Src.hasOneUse()) | |||
| 11684 | return SDValue(); | |||
| 11685 | ||||
| 11686 | Src = peekThroughOneUseBitcasts(Src); | |||
| 11687 | if (Src.getOpcode() != ISD::TRUNCATE || | |||
| 11688 | Src.getScalarValueSizeInBits() != (EltSizeInBits * Scale)) | |||
| 11689 | return SDValue(); | |||
| 11690 | Src = Src.getOperand(0); | |||
| 11691 | ||||
| 11692 | // VPMOVWB is only available with avx512bw. | |||
| 11693 | MVT SrcVT = Src.getSimpleValueType(); | |||
| 11694 | if (SrcVT.getVectorElementType() == MVT::i16 && VT == MVT::v16i8 && | |||
| 11695 | !Subtarget.hasBWI()) | |||
| 11696 | return SDValue(); | |||
| 11697 | ||||
| 11698 | bool UndefUppers = isUndefInRange(Mask, NumSrcElts, UpperElts); | |||
| 11699 | return getAVX512TruncNode(DL, VT, Src, Subtarget, DAG, !UndefUppers); | |||
| 11700 | } | |||
| 11701 | ||||
| 11702 | return SDValue(); | |||
| 11703 | } | |||
| 11704 | ||||
| 11705 | // Attempt to match binary shuffle patterns as a truncate. | |||
| 11706 | static SDValue lowerShuffleAsVTRUNC(const SDLoc &DL, MVT VT, SDValue V1, | |||
| 11707 | SDValue V2, ArrayRef<int> Mask, | |||
| 11708 | const APInt &Zeroable, | |||
| 11709 | const X86Subtarget &Subtarget, | |||
| 11710 | SelectionDAG &DAG) { | |||
| 11711 | assert((VT.is128BitVector() || VT.is256BitVector()) &&((void)0) | |||
| 11712 | "Unexpected VTRUNC type")((void)0); | |||
| 11713 | if (!Subtarget.hasAVX512()) | |||
| 11714 | return SDValue(); | |||
| 11715 | ||||
| 11716 | unsigned NumElts = VT.getVectorNumElements(); | |||
| 11717 | unsigned EltSizeInBits = VT.getScalarSizeInBits(); | |||
| 11718 | unsigned MaxScale = 64 / EltSizeInBits; | |||
| 11719 | for (unsigned Scale = 2; Scale <= MaxScale; Scale += Scale) { | |||
| 11720 | // TODO: Support non-BWI VPMOVWB truncations? | |||
| 11721 | unsigned SrcEltBits = EltSizeInBits * Scale; | |||
| 11722 | if (SrcEltBits < 32 && !Subtarget.hasBWI()) | |||
| 11723 | continue; | |||
| 11724 | ||||
| 11725 | // Match shuffle <0,Scale,2*Scale,..,undef_or_zero,undef_or_zero,...> | |||
| 11726 | // Bail if the V2 elements are undef. | |||
| 11727 | unsigned NumHalfSrcElts = NumElts / Scale; | |||
| 11728 | unsigned NumSrcElts = 2 * NumHalfSrcElts; | |||
| 11729 | if (!isSequentialOrUndefInRange(Mask, 0, NumSrcElts, 0, Scale) || | |||
| 11730 | isUndefInRange(Mask, NumHalfSrcElts, NumHalfSrcElts)) | |||
| 11731 | continue; | |||
| 11732 | ||||
| 11733 | // The elements beyond the truncation must be undef/zero. | |||
| 11734 | unsigned UpperElts = NumElts - NumSrcElts; | |||
| 11735 | if (UpperElts > 0 && | |||
| 11736 | !Zeroable.extractBits(UpperElts, NumSrcElts).isAllOnesValue()) | |||
| 11737 | continue; | |||
| 11738 | bool UndefUppers = | |||
| 11739 | UpperElts > 0 && isUndefInRange(Mask, NumSrcElts, UpperElts); | |||
| 11740 | ||||
| 11741 | // As we're using both sources then we need to concat them together | |||
| 11742 | // and truncate from the double-sized src. | |||
| 11743 | MVT ConcatVT = MVT::getVectorVT(VT.getScalarType(), NumElts * 2); | |||
| 11744 | SDValue Src = DAG.getNode(ISD::CONCAT_VECTORS, DL, ConcatVT, V1, V2); | |||
| 11745 | ||||
| 11746 | MVT SrcSVT = MVT::getIntegerVT(SrcEltBits); | |||
| 11747 | MVT SrcVT = MVT::getVectorVT(SrcSVT, NumSrcElts); | |||
| 11748 | Src = DAG.getBitcast(SrcVT, Src); | |||
| 11749 | return getAVX512TruncNode(DL, VT, Src, Subtarget, DAG, !UndefUppers); | |||
| 11750 | } | |||
| 11751 | ||||
| 11752 | return SDValue(); | |||
| 11753 | } | |||
| 11754 | ||||
| 11755 | /// Check whether a compaction lowering can be done by dropping even | |||
| 11756 | /// elements and compute how many times even elements must be dropped. | |||
| 11757 | /// | |||
| 11758 | /// This handles shuffles which take every Nth element where N is a power of | |||
| 11759 | /// two. Example shuffle masks: | |||
| 11760 | /// | |||
| 11761 | /// N = 1: 0, 2, 4, 6, 8, 10, 12, 14, 0, 2, 4, 6, 8, 10, 12, 14 | |||
| 11762 | /// N = 1: 0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30 | |||
| 11763 | /// N = 2: 0, 4, 8, 12, 0, 4, 8, 12, 0, 4, 8, 12, 0, 4, 8, 12 | |||
| 11764 | /// N = 2: 0, 4, 8, 12, 16, 20, 24, 28, 0, 4, 8, 12, 16, 20, 24, 28 | |||
| 11765 | /// N = 3: 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8 | |||
| 11766 | /// N = 3: 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24 | |||
| 11767 | /// | |||
| 11768 | /// Any of these lanes can of course be undef. | |||
| 11769 | /// | |||
| 11770 | /// This routine only supports N <= 3. | |||
| 11771 | /// FIXME: Evaluate whether either AVX or AVX-512 have any opportunities here | |||
| 11772 | /// for larger N. | |||
| 11773 | /// | |||
| 11774 | /// \returns N above, or the number of times even elements must be dropped if | |||
| 11775 | /// there is such a number. Otherwise returns zero. | |||
| 11776 | static int canLowerByDroppingEvenElements(ArrayRef<int> Mask, | |||
| 11777 | bool IsSingleInput) { | |||
| 11778 | // The modulus for the shuffle vector entries is based on whether this is | |||
| 11779 | // a single input or not. | |||
| 11780 | int ShuffleModulus = Mask.size() * (IsSingleInput ? 1 : 2); | |||
| 11781 | assert(isPowerOf2_32((uint32_t)ShuffleModulus) &&((void)0) | |||
| 11782 | "We should only be called with masks with a power-of-2 size!")((void)0); | |||
| 11783 | ||||
| 11784 | uint64_t ModMask = (uint64_t)ShuffleModulus - 1; | |||
| 11785 | ||||
| 11786 | // We track whether the input is viable for all power-of-2 strides 2^1, 2^2, | |||
| 11787 | // and 2^3 simultaneously. This is because we may have ambiguity with | |||
| 11788 | // partially undef inputs. | |||
| 11789 | bool ViableForN[3] = {true, true, true}; | |||
| 11790 | ||||
| 11791 | for (int i = 0, e = Mask.size(); i < e; ++i) { | |||
| 11792 | // Ignore undef lanes, we'll optimistically collapse them to the pattern we | |||
| 11793 | // want. | |||
| 11794 | if (Mask[i] < 0) | |||
| 11795 | continue; | |||
| 11796 | ||||
| 11797 | bool IsAnyViable = false; | |||
| 11798 | for (unsigned j = 0; j != array_lengthof(ViableForN); ++j) | |||
| 11799 | if (ViableForN[j]) { | |||
| 11800 | uint64_t N = j + 1; | |||
| 11801 | ||||
| 11802 | // The shuffle mask must be equal to (i * 2^N) % M. | |||
| 11803 | if ((uint64_t)Mask[i] == (((uint64_t)i << N) & ModMask)) | |||
| 11804 | IsAnyViable = true; | |||
| 11805 | else | |||
| 11806 | ViableForN[j] = false; | |||
| 11807 | } | |||
| 11808 | // Early exit if we exhaust the possible powers of two. | |||
| 11809 | if (!IsAnyViable) | |||
| 11810 | break; | |||
| 11811 | } | |||
| 11812 | ||||
| 11813 | for (unsigned j = 0; j != array_lengthof(ViableForN); ++j) | |||
| 11814 | if (ViableForN[j]) | |||
| 11815 | return j + 1; | |||
| 11816 | ||||
| 11817 | // Return 0 as there is no viable power of two. | |||
| 11818 | return 0; | |||
| 11819 | } | |||
| 11820 | ||||
| 11821 | // X86 has dedicated pack instructions that can handle specific truncation | |||
| 11822 | // operations: PACKSS and PACKUS. | |||
| 11823 | // Checks for compaction shuffle masks if MaxStages > 1. | |||
| 11824 | // TODO: Add support for matching multiple PACKSS/PACKUS stages. | |||
| 11825 | static bool matchShuffleWithPACK(MVT VT, MVT &SrcVT, SDValue &V1, SDValue &V2, | |||
| 11826 | unsigned &PackOpcode, ArrayRef<int> TargetMask, | |||
| 11827 | const SelectionDAG &DAG, | |||
| 11828 | const X86Subtarget &Subtarget, | |||
| 11829 | unsigned MaxStages = 1) { | |||
| 11830 | unsigned NumElts = VT.getVectorNumElements(); | |||
| 11831 | unsigned BitSize = VT.getScalarSizeInBits(); | |||
| 11832 | assert(0 < MaxStages && MaxStages <= 3 && (BitSize << MaxStages) <= 64 &&((void)0) | |||
| 11833 | "Illegal maximum compaction")((void)0); | |||
| 11834 | ||||
| 11835 | auto MatchPACK = [&](SDValue N1, SDValue N2, MVT PackVT) { | |||
| 11836 | unsigned NumSrcBits = PackVT.getScalarSizeInBits(); | |||
| 11837 | unsigned NumPackedBits = NumSrcBits - BitSize; | |||
| 11838 | N1 = peekThroughBitcasts(N1); | |||
| 11839 | N2 = peekThroughBitcasts(N2); | |||
| 11840 | unsigned NumBits1 = N1.getScalarValueSizeInBits(); | |||
| 11841 | unsigned NumBits2 = N2.getScalarValueSizeInBits(); | |||
| 11842 | bool IsZero1 = llvm::isNullOrNullSplat(N1, /*AllowUndefs*/ false); | |||
| 11843 | bool IsZero2 = llvm::isNullOrNullSplat(N2, /*AllowUndefs*/ false); | |||
| 11844 | if ((!N1.isUndef() && !IsZero1 && NumBits1 != NumSrcBits) || | |||
| 11845 | (!N2.isUndef() && !IsZero2 && NumBits2 != NumSrcBits)) | |||
| 11846 | return false; | |||
| 11847 | if (Subtarget.hasSSE41() || BitSize == 8) { | |||
| 11848 | APInt ZeroMask = APInt::getHighBitsSet(NumSrcBits, NumPackedBits); | |||
| 11849 | if ((N1.isUndef() || IsZero1 || DAG.MaskedValueIsZero(N1, ZeroMask)) && | |||
| 11850 | (N2.isUndef() || IsZero2 || DAG.MaskedValueIsZero(N2, ZeroMask))) { | |||
| 11851 | V1 = N1; | |||
| 11852 | V2 = N2; | |||
| 11853 | SrcVT = PackVT; | |||
| 11854 | PackOpcode = X86ISD::PACKUS; | |||
| 11855 | return true; | |||
| 11856 | } | |||
| 11857 | } | |||
| 11858 | bool IsAllOnes1 = llvm::isAllOnesOrAllOnesSplat(N1, /*AllowUndefs*/ false); | |||
| 11859 | bool IsAllOnes2 = llvm::isAllOnesOrAllOnesSplat(N2, /*AllowUndefs*/ false); | |||
| 11860 | if ((N1.isUndef() || IsZero1 || IsAllOnes1 || | |||
| 11861 | DAG.ComputeNumSignBits(N1) > NumPackedBits) && | |||
| 11862 | (N2.isUndef() || IsZero2 || IsAllOnes2 || | |||
| 11863 | DAG.ComputeNumSignBits(N2) > NumPackedBits)) { | |||
| 11864 | V1 = N1; | |||
| 11865 | V2 = N2; | |||
| 11866 | SrcVT = PackVT; | |||
| 11867 | PackOpcode = X86ISD::PACKSS; | |||
| 11868 | return true; | |||
| 11869 | } | |||
| 11870 | return false; | |||
| 11871 | }; | |||
| 11872 | ||||
| 11873 | // Attempt to match against wider and wider compaction patterns. | |||
| 11874 | for (unsigned NumStages = 1; NumStages <= MaxStages; ++NumStages) { | |||
| 11875 | MVT PackSVT = MVT::getIntegerVT(BitSize << NumStages); | |||
| 11876 | MVT PackVT = MVT::getVectorVT(PackSVT, NumElts >> NumStages); | |||
| 11877 | ||||
| 11878 | // Try binary shuffle. | |||
| 11879 | SmallVector<int, 32> BinaryMask; | |||
| 11880 | createPackShuffleMask(VT, BinaryMask, false, NumStages); | |||
| 11881 | if (isTargetShuffleEquivalent(VT, TargetMask, BinaryMask, V1, V2)) | |||
| 11882 | if (MatchPACK(V1, V2, PackVT)) | |||
| 11883 | return true; | |||
| 11884 | ||||
| 11885 | // Try unary shuffle. | |||
| 11886 | SmallVector<int, 32> UnaryMask; | |||
| 11887 | createPackShuffleMask(VT, UnaryMask, true, NumStages); | |||
| 11888 | if (isTargetShuffleEquivalent(VT, TargetMask, UnaryMask, V1)) | |||
| 11889 | if (MatchPACK(V1, V1, PackVT)) | |||
| 11890 | return true; | |||
| 11891 | } | |||
| 11892 | ||||
| 11893 | return false; | |||
| 11894 | } | |||
| 11895 | ||||
| 11896 | static SDValue lowerShuffleWithPACK(const SDLoc &DL, MVT VT, ArrayRef<int> Mask, | |||
| 11897 | SDValue V1, SDValue V2, SelectionDAG &DAG, | |||
| 11898 | const X86Subtarget &Subtarget) { | |||
| 11899 | MVT PackVT; | |||
| 11900 | unsigned PackOpcode; | |||
| 11901 | unsigned SizeBits = VT.getSizeInBits(); | |||
| 11902 | unsigned EltBits = VT.getScalarSizeInBits(); | |||
| 11903 | unsigned MaxStages = Log2_32(64 / EltBits); | |||
| 11904 | if (!matchShuffleWithPACK(VT, PackVT, V1, V2, PackOpcode, Mask, DAG, | |||
| 11905 | Subtarget, MaxStages)) | |||
| 11906 | return SDValue(); | |||
| 11907 | ||||
| 11908 | unsigned CurrentEltBits = PackVT.getScalarSizeInBits(); | |||
| 11909 | unsigned NumStages = Log2_32(CurrentEltBits / EltBits); | |||
| 11910 | ||||
| 11911 | // Don't lower multi-stage packs on AVX512, truncation is better. | |||
| 11912 | if (NumStages != 1 && SizeBits == 128 && Subtarget.hasVLX()) | |||
| 11913 | return SDValue(); | |||
| 11914 | ||||
| 11915 | // Pack to the largest type possible: | |||
| 11916 | // vXi64/vXi32 -> PACK*SDW and vXi16 -> PACK*SWB. | |||
| 11917 | unsigned MaxPackBits = 16; | |||
| 11918 | if (CurrentEltBits > 16 && | |||
| 11919 | (PackOpcode == X86ISD::PACKSS || Subtarget.hasSSE41())) | |||
| 11920 | MaxPackBits = 32; | |||
| 11921 | ||||
| 11922 | // Repeatedly pack down to the target size. | |||
| 11923 | SDValue Res; | |||
| 11924 | for (unsigned i = 0; i != NumStages; ++i) { | |||
| 11925 | unsigned SrcEltBits = std::min(MaxPackBits, CurrentEltBits); | |||
| 11926 | unsigned NumSrcElts = SizeBits / SrcEltBits; | |||
| 11927 | MVT SrcSVT = MVT::getIntegerVT(SrcEltBits); | |||
| 11928 | MVT DstSVT = MVT::getIntegerVT(SrcEltBits / 2); | |||
| 11929 | MVT SrcVT = MVT::getVectorVT(SrcSVT, NumSrcElts); | |||
| 11930 | MVT DstVT = MVT::getVectorVT(DstSVT, NumSrcElts * 2); | |||
| 11931 | Res = DAG.getNode(PackOpcode, DL, DstVT, DAG.getBitcast(SrcVT, V1), | |||
| 11932 | DAG.getBitcast(SrcVT, V2)); | |||
| 11933 | V1 = V2 = Res; | |||
| 11934 | CurrentEltBits /= 2; | |||
| 11935 | } | |||
| 11936 | assert(Res && Res.getValueType() == VT &&((void)0) | |||
| 11937 | "Failed to lower compaction shuffle")((void)0); | |||
| 11938 | return Res; | |||
| 11939 | } | |||
| 11940 | ||||
| 11941 | /// Try to emit a bitmask instruction for a shuffle. | |||
| 11942 | /// | |||
| 11943 | /// This handles cases where we can model a blend exactly as a bitmask due to | |||
| 11944 | /// one of the inputs being zeroable. | |||
| 11945 | static SDValue lowerShuffleAsBitMask(const SDLoc &DL, MVT VT, SDValue V1, | |||
| 11946 | SDValue V2, ArrayRef<int> Mask, | |||
| 11947 | const APInt &Zeroable, | |||
| 11948 | const X86Subtarget &Subtarget, | |||
| 11949 | SelectionDAG &DAG) { | |||
| 11950 | MVT MaskVT = VT; | |||
| 11951 | MVT EltVT = VT.getVectorElementType(); | |||
| 11952 | SDValue Zero, AllOnes; | |||
| 11953 | // Use f64 if i64 isn't legal. | |||
| 11954 | if (EltVT == MVT::i64 && !Subtarget.is64Bit()) { | |||
| 11955 | EltVT = MVT::f64; | |||
| 11956 | MaskVT = MVT::getVectorVT(EltVT, Mask.size()); | |||
| 11957 | } | |||
| 11958 | ||||
| 11959 | MVT LogicVT = VT; | |||
| 11960 | if (EltVT == MVT::f32 || EltVT == MVT::f64) { | |||
| 11961 | Zero = DAG.getConstantFP(0.0, DL, EltVT); | |||
| 11962 | APFloat AllOnesValue = APFloat::getAllOnesValue( | |||
| 11963 | SelectionDAG::EVTToAPFloatSemantics(EltVT), EltVT.getSizeInBits()); | |||
| 11964 | AllOnes = DAG.getConstantFP(AllOnesValue, DL, EltVT); | |||
| 11965 | LogicVT = | |||
| 11966 | MVT::getVectorVT(EltVT == MVT::f64 ? MVT::i64 : MVT::i32, Mask.size()); | |||
| 11967 | } else { | |||
| 11968 | Zero = DAG.getConstant(0, DL, EltVT); | |||
| 11969 | AllOnes = DAG.getAllOnesConstant(DL, EltVT); | |||
| 11970 | } | |||
| 11971 | ||||
| 11972 | SmallVector<SDValue, 16> VMaskOps(Mask.size(), Zero); | |||
| 11973 | SDValue V; | |||
| 11974 | for (int i = 0, Size = Mask.size(); i < Size; ++i) { | |||
| 11975 | if (Zeroable[i]) | |||
| 11976 | continue; | |||
| 11977 | if (Mask[i] % Size != i) | |||
| 11978 | return SDValue(); // Not a blend. | |||
| 11979 | if (!V) | |||
| 11980 | V = Mask[i] < Size ? V1 : V2; | |||
| 11981 | else if (V != (Mask[i] < Size ? V1 : V2)) | |||
| 11982 | return SDValue(); // Can only let one input through the mask. | |||
| 11983 | ||||
| 11984 | VMaskOps[i] = AllOnes; | |||
| 11985 | } | |||
| 11986 | if (!V) | |||
| 11987 | return SDValue(); // No non-zeroable elements! | |||
| 11988 | ||||
| 11989 | SDValue VMask = DAG.getBuildVector(MaskVT, DL, VMaskOps); | |||
| 11990 | VMask = DAG.getBitcast(LogicVT, VMask); | |||
| 11991 | V = DAG.getBitcast(LogicVT, V); | |||
| 11992 | SDValue And = DAG.getNode(ISD::AND, DL, LogicVT, V, VMask); | |||
| 11993 | return DAG.getBitcast(VT, And); | |||
| 11994 | } | |||
| 11995 | ||||
| 11996 | /// Try to emit a blend instruction for a shuffle using bit math. | |||
| 11997 | /// | |||
| 11998 | /// This is used as a fallback approach when first class blend instructions are | |||
| 11999 | /// unavailable. Currently it is only suitable for integer vectors, but could | |||
| 12000 | /// be generalized for floating point vectors if desirable. | |||
| 12001 | static SDValue lowerShuffleAsBitBlend(const SDLoc &DL, MVT VT, SDValue V1, | |||
| 12002 | SDValue V2, ArrayRef<int> Mask, | |||
| 12003 | SelectionDAG &DAG) { | |||
| 12004 | assert(VT.isInteger() && "Only supports integer vector types!")((void)0); | |||
| 12005 | MVT EltVT = VT.getVectorElementType(); | |||
| 12006 | SDValue Zero = DAG.getConstant(0, DL, EltVT); | |||
| 12007 | SDValue AllOnes = DAG.getAllOnesConstant(DL, EltVT); | |||
| 12008 | SmallVector<SDValue, 16> MaskOps; | |||
| 12009 | for (int i = 0, Size = Mask.size(); i < Size; ++i) { | |||
| 12010 | if (Mask[i] >= 0 && Mask[i] != i && Mask[i] != i + Size) | |||
| 12011 | return SDValue(); // Shuffled input! | |||
| 12012 | MaskOps.push_back(Mask[i] < Size ? AllOnes : Zero); | |||
| 12013 | } | |||
| 12014 | ||||
| 12015 | SDValue V1Mask = DAG.getBuildVector(VT, DL, MaskOps); | |||
| 12016 | V1 = DAG.getNode(ISD::AND, DL, VT, V1, V1Mask); | |||
| 12017 | V2 = DAG.getNode(X86ISD::ANDNP, DL, VT, V1Mask, V2); | |||
| 12018 | return DAG.getNode(ISD::OR, DL, VT, V1, V2); | |||
| 12019 | } | |||
| 12020 | ||||
| 12021 | static SDValue getVectorMaskingNode(SDValue Op, SDValue Mask, | |||
| 12022 | SDValue PreservedSrc, | |||
| 12023 | const X86Subtarget &Subtarget, | |||
| 12024 | SelectionDAG &DAG); | |||
| 12025 | ||||
| 12026 | static bool matchShuffleAsBlend(SDValue V1, SDValue V2, | |||
| 12027 | MutableArrayRef<int> Mask, | |||
| 12028 | const APInt &Zeroable, bool &ForceV1Zero, | |||
| 12029 | bool &ForceV2Zero, uint64_t &BlendMask) { | |||
| 12030 | bool V1IsZeroOrUndef = | |||
| 12031 | V1.isUndef() || ISD::isBuildVectorAllZeros(V1.getNode()); | |||
| 12032 | bool V2IsZeroOrUndef = | |||
| 12033 | V2.isUndef() || ISD::isBuildVectorAllZeros(V2.getNode()); | |||
| 12034 | ||||
| 12035 | BlendMask = 0; | |||
| 12036 | ForceV1Zero = false, ForceV2Zero = false; | |||
| 12037 | assert(Mask.size() <= 64 && "Shuffle mask too big for blend mask")((void)0); | |||
| 12038 | ||||
| 12039 | // Attempt to generate the binary blend mask. If an input is zero then | |||
| 12040 | // we can use any lane. | |||
| 12041 | for (int i = 0, Size = Mask.size(); i < Size; ++i) { | |||
| 12042 | int M = Mask[i]; | |||
| 12043 | if (M == SM_SentinelUndef) | |||
| 12044 | continue; | |||
| 12045 | if (M == i) | |||
| 12046 | continue; | |||
| 12047 | if (M == i + Size) { | |||
| 12048 | BlendMask |= 1ull << i; | |||
| 12049 | continue; | |||
| 12050 | } | |||
| 12051 | if (Zeroable[i]) { | |||
| 12052 | if (V1IsZeroOrUndef) { | |||
| 12053 | ForceV1Zero = true; | |||
| 12054 | Mask[i] = i; | |||
| 12055 | continue; | |||
| 12056 | } | |||
| 12057 | if (V2IsZeroOrUndef) { | |||
| 12058 | ForceV2Zero = true; | |||
| 12059 | BlendMask |= 1ull << i; | |||
| 12060 | Mask[i] = i + Size; | |||
| 12061 | continue; | |||
| 12062 | } | |||
| 12063 | } | |||
| 12064 | return false; | |||
| 12065 | } | |||
| 12066 | return true; | |||
| 12067 | } | |||
| 12068 | ||||
| 12069 | static uint64_t scaleVectorShuffleBlendMask(uint64_t BlendMask, int Size, | |||
| 12070 | int Scale) { | |||
| 12071 | uint64_t ScaledMask = 0; | |||
| 12072 | for (int i = 0; i != Size; ++i) | |||
| 12073 | if (BlendMask & (1ull << i)) | |||
| 12074 | ScaledMask |= ((1ull << Scale) - 1) << (i * Scale); | |||
| 12075 | return ScaledMask; | |||
| 12076 | } | |||
| 12077 | ||||
| 12078 | /// Try to emit a blend instruction for a shuffle. | |||
| 12079 | /// | |||
| 12080 | /// This doesn't do any checks for the availability of instructions for blending | |||
| 12081 | /// these values. It relies on the availability of the X86ISD::BLENDI pattern to | |||
| 12082 | /// be matched in the backend with the type given. What it does check for is | |||
| 12083 | /// that the shuffle mask is a blend, or convertible into a blend with zero. | |||
| 12084 | static SDValue lowerShuffleAsBlend(const SDLoc &DL, MVT VT, SDValue V1, | |||
| 12085 | SDValue V2, ArrayRef<int> Original, | |||
| 12086 | const APInt &Zeroable, | |||
| 12087 | const X86Subtarget &Subtarget, | |||
| 12088 | SelectionDAG &DAG) { | |||
| 12089 | uint64_t BlendMask = 0; | |||
| 12090 | bool ForceV1Zero = false, ForceV2Zero = false; | |||
| 12091 | SmallVector<int, 64> Mask(Original.begin(), Original.end()); | |||
| 12092 | if (!matchShuffleAsBlend(V1, V2, Mask, Zeroable, ForceV1Zero, ForceV2Zero, | |||
| 12093 | BlendMask)) | |||
| 12094 | return SDValue(); | |||
| 12095 | ||||
| 12096 | // Create a REAL zero vector - ISD::isBuildVectorAllZeros allows UNDEFs. | |||
| 12097 | if (ForceV1Zero) | |||
| 12098 | V1 = getZeroVector(VT, Subtarget, DAG, DL); | |||
| 12099 | if (ForceV2Zero) | |||
| 12100 | V2 = getZeroVector(VT, Subtarget, DAG, DL); | |||
| 12101 | ||||
| 12102 | switch (VT.SimpleTy) { | |||
| 12103 | case MVT::v4i64: | |||
| 12104 | case MVT::v8i32: | |||
| 12105 | assert(Subtarget.hasAVX2() && "256-bit integer blends require AVX2!")((void)0); | |||
| 12106 | LLVM_FALLTHROUGH[[gnu::fallthrough]]; | |||
| 12107 | case MVT::v4f64: | |||
| 12108 | case MVT::v8f32: | |||
| 12109 | assert(Subtarget.hasAVX() && "256-bit float blends require AVX!")((void)0); | |||
| 12110 | LLVM_FALLTHROUGH[[gnu::fallthrough]]; | |||
| 12111 | case MVT::v2f64: | |||
| 12112 | case MVT::v2i64: | |||
| 12113 | case MVT::v4f32: | |||
| 12114 | case MVT::v4i32: | |||
| 12115 | case MVT::v8i16: | |||
| 12116 | assert(Subtarget.hasSSE41() && "128-bit blends require SSE41!")((void)0); | |||
| 12117 | return DAG.getNode(X86ISD::BLENDI, DL, VT, V1, V2, | |||
| 12118 | DAG.getTargetConstant(BlendMask, DL, MVT::i8)); | |||
| 12119 | case MVT::v16i16: { | |||
| 12120 | assert(Subtarget.hasAVX2() && "v16i16 blends require AVX2!")((void)0); | |||
| 12121 | SmallVector<int, 8> RepeatedMask; | |||
| 12122 | if (is128BitLaneRepeatedShuffleMask(MVT::v16i16, Mask, RepeatedMask)) { | |||
| 12123 | // We can lower these with PBLENDW which is mirrored across 128-bit lanes. | |||
| 12124 | assert(RepeatedMask.size() == 8 && "Repeated mask size doesn't match!")((void)0); | |||
| 12125 | BlendMask = 0; | |||
| 12126 | for (int i = 0; i < 8; ++i) | |||
| 12127 | if (RepeatedMask[i] >= 8) | |||
| 12128 | BlendMask |= 1ull << i; | |||
| 12129 | return DAG.getNode(X86ISD::BLENDI, DL, MVT::v16i16, V1, V2, | |||
| 12130 | DAG.getTargetConstant(BlendMask, DL, MVT::i8)); | |||
| 12131 | } | |||
| 12132 | // Use PBLENDW for lower/upper lanes and then blend lanes. | |||
| 12133 | // TODO - we should allow 2 PBLENDW here and leave shuffle combine to | |||
| 12134 | // merge to VSELECT where useful. | |||
| 12135 | uint64_t LoMask = BlendMask & 0xFF; | |||
| 12136 | uint64_t HiMask = (BlendMask >> 8) & 0xFF; | |||
| 12137 | if (LoMask == 0 || LoMask == 255 || HiMask == 0 || HiMask == 255) { | |||
| 12138 | SDValue Lo = DAG.getNode(X86ISD::BLENDI, DL, MVT::v16i16, V1, V2, | |||
| 12139 | DAG.getTargetConstant(LoMask, DL, MVT::i8)); | |||
| 12140 | SDValue Hi = DAG.getNode(X86ISD::BLENDI, DL, MVT::v16i16, V1, V2, | |||
| 12141 | DAG.getTargetConstant(HiMask, DL, MVT::i8)); | |||
| 12142 | return DAG.getVectorShuffle( | |||
| 12143 | MVT::v16i16, DL, Lo, Hi, | |||
| 12144 | {0, 1, 2, 3, 4, 5, 6, 7, 24, 25, 26, 27, 28, 29, 30, 31}); | |||
| 12145 | } | |||
| 12146 | LLVM_FALLTHROUGH[[gnu::fallthrough]]; | |||
| 12147 | } | |||
| 12148 | case MVT::v32i8: | |||
| 12149 | assert(Subtarget.hasAVX2() && "256-bit byte-blends require AVX2!")((void)0); | |||
| 12150 | LLVM_FALLTHROUGH[[gnu::fallthrough]]; | |||
| 12151 | case MVT::v16i8: { | |||
| 12152 | assert(Subtarget.hasSSE41() && "128-bit byte-blends require SSE41!")((void)0); | |||
| 12153 | ||||
| 12154 | // Attempt to lower to a bitmask if we can. VPAND is faster than VPBLENDVB. | |||
| 12155 | if (SDValue Masked = lowerShuffleAsBitMask(DL, VT, V1, V2, Mask, Zeroable, | |||
| 12156 | Subtarget, DAG)) | |||
| 12157 | return Masked; | |||
| 12158 | ||||
| 12159 | if (Subtarget.hasBWI() && Subtarget.hasVLX()) { | |||
| 12160 | MVT IntegerType = | |||
| 12161 | MVT::getIntegerVT(std::max((int)VT.getVectorNumElements(), 8)); | |||
| 12162 | SDValue MaskNode = DAG.getConstant(BlendMask, DL, IntegerType); | |||
| 12163 | return getVectorMaskingNode(V2, MaskNode, V1, Subtarget, DAG); | |||
| 12164 | } | |||
| 12165 | ||||
| 12166 | // If we have VPTERNLOG, we can use that as a bit blend. | |||
| 12167 | if (Subtarget.hasVLX()) | |||
| 12168 | if (SDValue BitBlend = | |||
| 12169 | lowerShuffleAsBitBlend(DL, VT, V1, V2, Mask, DAG)) | |||
| 12170 | return BitBlend; | |||
| 12171 | ||||
| 12172 | // Scale the blend by the number of bytes per element. | |||
| 12173 | int Scale = VT.getScalarSizeInBits() / 8; | |||
| 12174 | ||||
| 12175 | // This form of blend is always done on bytes. Compute the byte vector | |||
| 12176 | // type. | |||
| 12177 | MVT BlendVT = MVT::getVectorVT(MVT::i8, VT.getSizeInBits() / 8); | |||
| 12178 | ||||
| 12179 | // x86 allows load folding with blendvb from the 2nd source operand. But | |||
| 12180 | // we are still using LLVM select here (see comment below), so that's V1. | |||
| 12181 | // If V2 can be load-folded and V1 cannot be load-folded, then commute to | |||
| 12182 | // allow that load-folding possibility. | |||
| 12183 | if (!ISD::isNormalLoad(V1.getNode()) && ISD::isNormalLoad(V2.getNode())) { | |||
| 12184 | ShuffleVectorSDNode::commuteMask(Mask); | |||
| 12185 | std::swap(V1, V2); | |||
| 12186 | } | |||
| 12187 | ||||
| 12188 | // Compute the VSELECT mask. Note that VSELECT is really confusing in the | |||
| 12189 | // mix of LLVM's code generator and the x86 backend. We tell the code | |||
| 12190 | // generator that boolean values in the elements of an x86 vector register | |||
| 12191 | // are -1 for true and 0 for false. We then use the LLVM semantics of 'true' | |||
| 12192 | // mapping a select to operand #1, and 'false' mapping to operand #2. The | |||
| 12193 | // reality in x86 is that vector masks (pre-AVX-512) use only the high bit | |||
| 12194 | // of the element (the remaining are ignored) and 0 in that high bit would | |||
| 12195 | // mean operand #1 while 1 in the high bit would mean operand #2. So while | |||
| 12196 | // the LLVM model for boolean values in vector elements gets the relevant | |||
| 12197 | // bit set, it is set backwards and over constrained relative to x86's | |||
| 12198 | // actual model. | |||
| 12199 | SmallVector<SDValue, 32> VSELECTMask; | |||
| 12200 | for (int i = 0, Size = Mask.size(); i < Size; ++i) | |||
| 12201 | for (int j = 0; j < Scale; ++j) | |||
| 12202 | VSELECTMask.push_back( | |||
| 12203 | Mask[i] < 0 ? DAG.getUNDEF(MVT::i8) | |||
| 12204 | : DAG.getConstant(Mask[i] < Size ? -1 : 0, DL, | |||
| 12205 | MVT::i8)); | |||
| 12206 | ||||
| 12207 | V1 = DAG.getBitcast(BlendVT, V1); | |||
| 12208 | V2 = DAG.getBitcast(BlendVT, V2); | |||
| 12209 | return DAG.getBitcast( | |||
| 12210 | VT, | |||
| 12211 | DAG.getSelect(DL, BlendVT, DAG.getBuildVector(BlendVT, DL, VSELECTMask), | |||
| 12212 | V1, V2)); | |||
| 12213 | } | |||
| 12214 | case MVT::v16f32: | |||
| 12215 | case MVT::v8f64: | |||
| 12216 | case MVT::v8i64: | |||
| 12217 | case MVT::v16i32: | |||
| 12218 | case MVT::v32i16: | |||
| 12219 | case MVT::v64i8: { | |||
| 12220 | // Attempt to lower to a bitmask if we can. Only if not optimizing for size. | |||
| 12221 | bool OptForSize = DAG.shouldOptForSize(); | |||
| 12222 | if (!OptForSize) { | |||
| 12223 | if (SDValue Masked = lowerShuffleAsBitMask(DL, VT, V1, V2, Mask, Zeroable, | |||
| 12224 | Subtarget, DAG)) | |||
| 12225 | return Masked; | |||
| 12226 | } | |||
| 12227 | ||||
| 12228 | // Otherwise load an immediate into a GPR, cast to k-register, and use a | |||
| 12229 | // masked move. | |||
| 12230 | MVT IntegerType = | |||
| 12231 | MVT::getIntegerVT(std::max((int)VT.getVectorNumElements(), 8)); | |||
| 12232 | SDValue MaskNode = DAG.getConstant(BlendMask, DL, IntegerType); | |||
| 12233 | return getVectorMaskingNode(V2, MaskNode, V1, Subtarget, DAG); | |||
| 12234 | } | |||
| 12235 | default: | |||
| 12236 | llvm_unreachable("Not a supported integer vector type!")__builtin_unreachable(); | |||
| 12237 | } | |||
| 12238 | } | |||
| 12239 | ||||
| 12240 | /// Try to lower as a blend of elements from two inputs followed by | |||
| 12241 | /// a single-input permutation. | |||
| 12242 | /// | |||
| 12243 | /// This matches the pattern where we can blend elements from two inputs and | |||
| 12244 | /// then reduce the shuffle to a single-input permutation. | |||
| 12245 | static SDValue lowerShuffleAsBlendAndPermute(const SDLoc &DL, MVT VT, | |||
| 12246 | SDValue V1, SDValue V2, | |||
| 12247 | ArrayRef<int> Mask, | |||
| 12248 | SelectionDAG &DAG, | |||
| 12249 | bool ImmBlends = false) { | |||
| 12250 | // We build up the blend mask while checking whether a blend is a viable way | |||
| 12251 | // to reduce the shuffle. | |||
| 12252 | SmallVector<int, 32> BlendMask(Mask.size(), -1); | |||
| 12253 | SmallVector<int, 32> PermuteMask(Mask.size(), -1); | |||
| 12254 | ||||
| 12255 | for (int i = 0, Size = Mask.size(); i < Size; ++i) { | |||
| 12256 | if (Mask[i] < 0) | |||
| 12257 | continue; | |||
| 12258 | ||||
| 12259 | assert(Mask[i] < Size * 2 && "Shuffle input is out of bounds.")((void)0); | |||
| 12260 | ||||
| 12261 | if (BlendMask[Mask[i] % Size] < 0) | |||
| 12262 | BlendMask[Mask[i] % Size] = Mask[i]; | |||
| 12263 | else if (BlendMask[Mask[i] % Size] != Mask[i]) | |||
| 12264 | return SDValue(); // Can't blend in the needed input! | |||
| 12265 | ||||
| 12266 | PermuteMask[i] = Mask[i] % Size; | |||
| 12267 | } | |||
| 12268 | ||||
| 12269 | // If only immediate blends, then bail if the blend mask can't be widened to | |||
| 12270 | // i16. | |||
| 12271 | unsigned EltSize = VT.getScalarSizeInBits(); | |||
| 12272 | if (ImmBlends && EltSize == 8 && !canWidenShuffleElements(BlendMask)) | |||
| 12273 | return SDValue(); | |||
| 12274 | ||||
| 12275 | SDValue V = DAG.getVectorShuffle(VT, DL, V1, V2, BlendMask); | |||
| 12276 | return DAG.getVectorShuffle(VT, DL, V, DAG.getUNDEF(VT), PermuteMask); | |||
| 12277 | } | |||
| 12278 | ||||
| 12279 | /// Try to lower as an unpack of elements from two inputs followed by | |||
| 12280 | /// a single-input permutation. | |||
| 12281 | /// | |||
| 12282 | /// This matches the pattern where we can unpack elements from two inputs and | |||
| 12283 | /// then reduce the shuffle to a single-input (wider) permutation. | |||
| 12284 | static SDValue lowerShuffleAsUNPCKAndPermute(const SDLoc &DL, MVT VT, | |||
| 12285 | SDValue V1, SDValue V2, | |||
| 12286 | ArrayRef<int> Mask, | |||
| 12287 | SelectionDAG &DAG) { | |||
| 12288 | int NumElts = Mask.size(); | |||
| 12289 | int NumLanes = VT.getSizeInBits() / 128; | |||
| 12290 | int NumLaneElts = NumElts / NumLanes; | |||
| 12291 | int NumHalfLaneElts = NumLaneElts / 2; | |||
| 12292 | ||||
| 12293 | bool MatchLo = true, MatchHi = true; | |||
| 12294 | SDValue Ops[2] = {DAG.getUNDEF(VT), DAG.getUNDEF(VT)}; | |||
| 12295 | ||||
| 12296 | // Determine UNPCKL/UNPCKH type and operand order. | |||
| 12297 | for (int Lane = 0; Lane != NumElts; Lane += NumLaneElts) { | |||
| 12298 | for (int Elt = 0; Elt != NumLaneElts; ++Elt) { | |||
| 12299 | int M = Mask[Lane + Elt]; | |||
| 12300 | if (M < 0) | |||
| 12301 | continue; | |||
| 12302 | ||||
| 12303 | SDValue &Op = Ops[Elt & 1]; | |||
| 12304 | if (M < NumElts && (Op.isUndef() || Op == V1)) | |||
| 12305 | Op = V1; | |||
| 12306 | else if (NumElts <= M && (Op.isUndef() || Op == V2)) | |||
| 12307 | Op = V2; | |||
| 12308 | else | |||
| 12309 | return SDValue(); | |||
| 12310 | ||||
| 12311 | int Lo = Lane, Mid = Lane + NumHalfLaneElts, Hi = Lane + NumLaneElts; | |||
| 12312 | MatchLo &= isUndefOrInRange(M, Lo, Mid) || | |||
| 12313 | isUndefOrInRange(M, NumElts + Lo, NumElts + Mid); | |||
| 12314 | MatchHi &= isUndefOrInRange(M, Mid, Hi) || | |||
| 12315 | isUndefOrInRange(M, NumElts + Mid, NumElts + Hi); | |||
| 12316 | if (!MatchLo && !MatchHi) | |||
| 12317 | return SDValue(); | |||
| 12318 | } | |||
| 12319 | } | |||
| 12320 | assert((MatchLo ^ MatchHi) && "Failed to match UNPCKLO/UNPCKHI")((void)0); | |||
| 12321 | ||||
| 12322 | // Now check that each pair of elts come from the same unpack pair | |||
| 12323 | // and set the permute mask based on each pair. | |||
| 12324 | // TODO - Investigate cases where we permute individual elements. | |||
| 12325 | SmallVector<int, 32> PermuteMask(NumElts, -1); | |||
| 12326 | for (int Lane = 0; Lane != NumElts; Lane += NumLaneElts) { | |||
| 12327 | for (int Elt = 0; Elt != NumLaneElts; Elt += 2) { | |||
| 12328 | int M0 = Mask[Lane + Elt + 0]; | |||
| 12329 | int M1 = Mask[Lane + Elt + 1]; | |||
| 12330 | if (0 <= M0 && 0 <= M1 && | |||
| 12331 | (M0 % NumHalfLaneElts) != (M1 % NumHalfLaneElts)) | |||
| 12332 | return SDValue(); | |||
| 12333 | if (0 <= M0) | |||
| 12334 | PermuteMask[Lane + Elt + 0] = Lane + (2 * (M0 % NumHalfLaneElts)); | |||
| 12335 | if (0 <= M1) | |||
| 12336 | PermuteMask[Lane + Elt + 1] = Lane + (2 * (M1 % NumHalfLaneElts)) + 1; | |||
| 12337 | } | |||
| 12338 | } | |||
| 12339 | ||||
| 12340 | unsigned UnpckOp = MatchLo ? X86ISD::UNPCKL : X86ISD::UNPCKH; | |||
| 12341 | SDValue Unpck = DAG.getNode(UnpckOp, DL, VT, Ops); | |||
| 12342 | return DAG.getVectorShuffle(VT, DL, Unpck, DAG.getUNDEF(VT), PermuteMask); | |||
| 12343 | } | |||
| 12344 | ||||
| 12345 | /// Helper to form a PALIGNR-based rotate+permute, merging 2 inputs and then | |||
| 12346 | /// permuting the elements of the result in place. | |||
| 12347 | static SDValue lowerShuffleAsByteRotateAndPermute( | |||
| 12348 | const SDLoc &DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask, | |||
| 12349 | const X86Subtarget &Subtarget, SelectionDAG &DAG) { | |||
| 12350 | if ((VT.is128BitVector() && !Subtarget.hasSSSE3()) || | |||
| 12351 | (VT.is256BitVector() && !Subtarget.hasAVX2()) || | |||
| 12352 | (VT.is512BitVector() && !Subtarget.hasBWI())) | |||
| 12353 | return SDValue(); | |||
| 12354 | ||||
| 12355 | // We don't currently support lane crossing permutes. | |||
| 12356 | if (is128BitLaneCrossingShuffleMask(VT, Mask)) | |||
| 12357 | return SDValue(); | |||
| 12358 | ||||
| 12359 | int Scale = VT.getScalarSizeInBits() / 8; | |||
| 12360 | int NumLanes = VT.getSizeInBits() / 128; | |||
| 12361 | int NumElts = VT.getVectorNumElements(); | |||
| 12362 | int NumEltsPerLane = NumElts / NumLanes; | |||
| 12363 | ||||
| 12364 | // Determine range of mask elts. | |||
| 12365 | bool Blend1 = true; | |||
| 12366 | bool Blend2 = true; | |||
| 12367 | std::pair<int, int> Range1 = std::make_pair(INT_MAX2147483647, INT_MIN(-2147483647 -1)); | |||
| 12368 | std::pair<int, int> Range2 = std::make_pair(INT_MAX2147483647, INT_MIN(-2147483647 -1)); | |||
| 12369 | for (int Lane = 0; Lane != NumElts; Lane += NumEltsPerLane) { | |||
| 12370 | for (int Elt = 0; Elt != NumEltsPerLane; ++Elt) { | |||
| 12371 | int M = Mask[Lane + Elt]; | |||
| 12372 | if (M < 0) | |||
| 12373 | continue; | |||
| 12374 | if (M < NumElts) { | |||
| 12375 | Blend1 &= (M == (Lane + Elt)); | |||
| 12376 | assert(Lane <= M && M < (Lane + NumEltsPerLane) && "Out of range mask")((void)0); | |||
| 12377 | M = M % NumEltsPerLane; | |||
| 12378 | Range1.first = std::min(Range1.first, M); | |||
| 12379 | Range1.second = std::max(Range1.second, M); | |||
| 12380 | } else { | |||
| 12381 | M -= NumElts; | |||
| 12382 | Blend2 &= (M == (Lane + Elt)); | |||
| 12383 | assert(Lane <= M && M < (Lane + NumEltsPerLane) && "Out of range mask")((void)0); | |||
| 12384 | M = M % NumEltsPerLane; | |||
| 12385 | Range2.first = std::min(Range2.first, M); | |||
| 12386 | Range2.second = std::max(Range2.second, M); | |||
| 12387 | } | |||
| 12388 | } | |||
| 12389 | } | |||
| 12390 | ||||
| 12391 | // Bail if we don't need both elements. | |||
| 12392 | // TODO - it might be worth doing this for unary shuffles if the permute | |||
| 12393 | // can be widened. | |||
| 12394 | if (!(0 <= Range1.first && Range1.second < NumEltsPerLane) || | |||
| 12395 | !(0 <= Range2.first && Range2.second < NumEltsPerLane)) | |||
| 12396 | return SDValue(); | |||
| 12397 | ||||
| 12398 | if (VT.getSizeInBits() > 128 && (Blend1 || Blend2)) | |||
| 12399 | return SDValue(); | |||
| 12400 | ||||
| 12401 | // Rotate the 2 ops so we can access both ranges, then permute the result. | |||
| 12402 | auto RotateAndPermute = [&](SDValue Lo, SDValue Hi, int RotAmt, int Ofs) { | |||
| 12403 | MVT ByteVT = MVT::getVectorVT(MVT::i8, VT.getSizeInBits() / 8); | |||
| 12404 | SDValue Rotate = DAG.getBitcast( | |||
| 12405 | VT, DAG.getNode(X86ISD::PALIGNR, DL, ByteVT, DAG.getBitcast(ByteVT, Hi), | |||
| 12406 | DAG.getBitcast(ByteVT, Lo), | |||
| 12407 | DAG.getTargetConstant(Scale * RotAmt, DL, MVT::i8))); | |||
| 12408 | SmallVector<int, 64> PermMask(NumElts, SM_SentinelUndef); | |||
| 12409 | for (int Lane = 0; Lane != NumElts; Lane += NumEltsPerLane) { | |||
| 12410 | for (int Elt = 0; Elt != NumEltsPerLane; ++Elt) { | |||
| 12411 | int M = Mask[Lane + Elt]; | |||
| 12412 | if (M < 0) | |||
| 12413 | continue; | |||
| 12414 | if (M < NumElts) | |||
| 12415 | PermMask[Lane + Elt] = Lane + ((M + Ofs - RotAmt) % NumEltsPerLane); | |||
| 12416 | else | |||
| 12417 | PermMask[Lane + Elt] = Lane + ((M - Ofs - RotAmt) % NumEltsPerLane); | |||
| 12418 | } | |||
| 12419 | } | |||
| 12420 | return DAG.getVectorShuffle(VT, DL, Rotate, DAG.getUNDEF(VT), PermMask); | |||
| 12421 | }; | |||
| 12422 | ||||
| 12423 | // Check if the ranges are small enough to rotate from either direction. | |||
| 12424 | if (Range2.second < Range1.first) | |||
| 12425 | return RotateAndPermute(V1, V2, Range1.first, 0); | |||
| 12426 | if (Range1.second < Range2.first) | |||
| 12427 | return RotateAndPermute(V2, V1, Range2.first, NumElts); | |||
| 12428 | return SDValue(); | |||
| 12429 | } | |||
| 12430 | ||||
| 12431 | /// Generic routine to decompose a shuffle and blend into independent | |||
| 12432 | /// blends and permutes. | |||
| 12433 | /// | |||
| 12434 | /// This matches the extremely common pattern for handling combined | |||
| 12435 | /// shuffle+blend operations on newer X86 ISAs where we have very fast blend | |||
| 12436 | /// operations. It will try to pick the best arrangement of shuffles and | |||
| 12437 | /// blends. For vXi8/vXi16 shuffles we may use unpack instead of blend. | |||
| 12438 | static SDValue lowerShuffleAsDecomposedShuffleMerge( | |||
| 12439 | const SDLoc &DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask, | |||
| 12440 | const X86Subtarget &Subtarget, SelectionDAG &DAG) { | |||
| 12441 | int NumElts = Mask.size(); | |||
| 12442 | int NumLanes = VT.getSizeInBits() / 128; | |||
| 12443 | int NumEltsPerLane = NumElts / NumLanes; | |||
| 12444 | ||||
| 12445 | // Shuffle the input elements into the desired positions in V1 and V2 and | |||
| 12446 | // unpack/blend them together. | |||
| 12447 | bool IsAlternating = true; | |||
| 12448 | SmallVector<int, 32> V1Mask(NumElts, -1); | |||
| 12449 | SmallVector<int, 32> V2Mask(NumElts, -1); | |||
| 12450 | SmallVector<int, 32> FinalMask(NumElts, -1); | |||
| 12451 | for (int i = 0; i < NumElts; ++i) { | |||
| 12452 | int M = Mask[i]; | |||
| 12453 | if (M >= 0 && M < NumElts) { | |||
| 12454 | V1Mask[i] = M; | |||
| 12455 | FinalMask[i] = i; | |||
| 12456 | IsAlternating &= (i & 1) == 0; | |||
| 12457 | } else if (M >= NumElts) { | |||
| 12458 | V2Mask[i] = M - NumElts; | |||
| 12459 | FinalMask[i] = i + NumElts; | |||
| 12460 | IsAlternating &= (i & 1) == 1; | |||
| 12461 | } | |||
| 12462 | } | |||
| 12463 | ||||
| 12464 | // Try to lower with the simpler initial blend/unpack/rotate strategies unless | |||
| 12465 | // one of the input shuffles would be a no-op. We prefer to shuffle inputs as | |||
| 12466 | // the shuffle may be able to fold with a load or other benefit. However, when | |||
| 12467 | // we'll have to do 2x as many shuffles in order to achieve this, a 2-input | |||
| 12468 | // pre-shuffle first is a better strategy. | |||
| 12469 | if (!isNoopShuffleMask(V1Mask) && !isNoopShuffleMask(V2Mask)) { | |||
| 12470 | // Only prefer immediate blends to unpack/rotate. | |||
| 12471 | if (SDValue BlendPerm = lowerShuffleAsBlendAndPermute(DL, VT, V1, V2, Mask, | |||
| 12472 | DAG, true)) | |||
| 12473 | return BlendPerm; | |||
| 12474 | if (SDValue UnpackPerm = lowerShuffleAsUNPCKAndPermute(DL, VT, V1, V2, Mask, | |||
| 12475 | DAG)) | |||
| 12476 | return UnpackPerm; | |||
| 12477 | if (SDValue RotatePerm = lowerShuffleAsByteRotateAndPermute( | |||
| 12478 | DL, VT, V1, V2, Mask, Subtarget, DAG)) | |||
| 12479 | return RotatePerm; | |||
| 12480 | // Unpack/rotate failed - try again with variable blends. | |||
| 12481 | if (SDValue BlendPerm = lowerShuffleAsBlendAndPermute(DL, VT, V1, V2, Mask, | |||
| 12482 | DAG)) | |||
| 12483 | return BlendPerm; | |||
| 12484 | } | |||
| 12485 | ||||
| 12486 | // If the final mask is an alternating blend of vXi8/vXi16, convert to an | |||
| 12487 | // UNPCKL(SHUFFLE, SHUFFLE) pattern. | |||
| 12488 | // TODO: It doesn't have to be alternating - but each lane mustn't have more | |||
| 12489 | // than half the elements coming from each source. | |||
| 12490 | if (IsAlternating && VT.getScalarSizeInBits() < 32) { | |||
| 12491 | V1Mask.assign(NumElts, -1); | |||
| 12492 | V2Mask.assign(NumElts, -1); | |||
| 12493 | FinalMask.assign(NumElts, -1); | |||
| 12494 | for (int i = 0; i != NumElts; i += NumEltsPerLane) | |||
| 12495 | for (int j = 0; j != NumEltsPerLane; ++j) { | |||
| 12496 | int M = Mask[i + j]; | |||
| 12497 | if (M >= 0 && M < NumElts) { | |||
| 12498 | V1Mask[i + (j / 2)] = M; | |||
| 12499 | FinalMask[i + j] = i + (j / 2); | |||
| 12500 | } else if (M >= NumElts) { | |||
| 12501 | V2Mask[i + (j / 2)] = M - NumElts; | |||
| 12502 | FinalMask[i + j] = i + (j / 2) + NumElts; | |||
| 12503 | } | |||
| 12504 | } | |||
| 12505 | } | |||
| 12506 | ||||
| 12507 | V1 = DAG.getVectorShuffle(VT, DL, V1, DAG.getUNDEF(VT), V1Mask); | |||
| 12508 | V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Mask); | |||
| 12509 | return DAG.getVectorShuffle(VT, DL, V1, V2, FinalMask); | |||
| 12510 | } | |||
| 12511 | ||||
| 12512 | /// Try to lower a vector shuffle as a bit rotation. | |||
| 12513 | /// | |||
| 12514 | /// Look for a repeated rotation pattern in each sub group. | |||
| 12515 | /// Returns a ISD::ROTL element rotation amount or -1 if failed. | |||
| 12516 | static int matchShuffleAsBitRotate(ArrayRef<int> Mask, int NumSubElts) { | |||
| 12517 | int NumElts = Mask.size(); | |||
| 12518 | assert((NumElts % NumSubElts) == 0 && "Illegal shuffle mask")((void)0); | |||
| 12519 | ||||
| 12520 | int RotateAmt = -1; | |||
| 12521 | for (int i = 0; i != NumElts; i += NumSubElts) { | |||
| 12522 | for (int j = 0; j != NumSubElts; ++j) { | |||
| 12523 | int M = Mask[i + j]; | |||
| 12524 | if (M < 0) | |||
| 12525 | continue; | |||
| 12526 | if (!isInRange(M, i, i + NumSubElts)) | |||
| 12527 | return -1; | |||
| 12528 | int Offset = (NumSubElts - (M - (i + j))) % NumSubElts; | |||
| 12529 | if (0 <= RotateAmt && Offset != RotateAmt) | |||
| 12530 | return -1; | |||
| 12531 | RotateAmt = Offset; | |||
| 12532 | } | |||
| 12533 | } | |||
| 12534 | return RotateAmt; | |||
| 12535 | } | |||
| 12536 | ||||
| 12537 | static int matchShuffleAsBitRotate(MVT &RotateVT, int EltSizeInBits, | |||
| 12538 | const X86Subtarget &Subtarget, | |||
| 12539 | ArrayRef<int> Mask) { | |||
| 12540 | assert(!isNoopShuffleMask(Mask) && "We shouldn't lower no-op shuffles!")((void)0); | |||
| 12541 | assert(EltSizeInBits < 64 && "Can't rotate 64-bit integers")((void)0); | |||
| 12542 | ||||
| 12543 | // AVX512 only has vXi32/vXi64 rotates, so limit the rotation sub group size. | |||
| 12544 | int MinSubElts = Subtarget.hasAVX512() ? std::max(32 / EltSizeInBits, 2) : 2; | |||
| 12545 | int MaxSubElts = 64 / EltSizeInBits; | |||
| 12546 | for (int NumSubElts = MinSubElts; NumSubElts <= MaxSubElts; NumSubElts *= 2) { | |||
| 12547 | int RotateAmt = matchShuffleAsBitRotate(Mask, NumSubElts); | |||
| 12548 | if (RotateAmt < 0) | |||
| 12549 | continue; | |||
| 12550 | ||||
| 12551 | int NumElts = Mask.size(); | |||
| 12552 | MVT RotateSVT = MVT::getIntegerVT(EltSizeInBits * NumSubElts); | |||
| 12553 | RotateVT = MVT::getVectorVT(RotateSVT, NumElts / NumSubElts); | |||
| 12554 | return RotateAmt * EltSizeInBits; | |||
| 12555 | } | |||
| 12556 | ||||
| 12557 | return -1; | |||
| 12558 | } | |||
| 12559 | ||||
| 12560 | /// Lower shuffle using X86ISD::VROTLI rotations. | |||
| 12561 | static SDValue lowerShuffleAsBitRotate(const SDLoc &DL, MVT VT, SDValue V1, | |||
| 12562 | ArrayRef<int> Mask, | |||
| 12563 | const X86Subtarget &Subtarget, | |||
| 12564 | SelectionDAG &DAG) { | |||
| 12565 | // Only XOP + AVX512 targets have bit rotation instructions. | |||
| 12566 | // If we at least have SSSE3 (PSHUFB) then we shouldn't attempt to use this. | |||
| 12567 | bool IsLegal = | |||
| 12568 | (VT.is128BitVector() && Subtarget.hasXOP()) || Subtarget.hasAVX512(); | |||
| 12569 | if (!IsLegal && Subtarget.hasSSE3()) | |||
| 12570 | return SDValue(); | |||
| 12571 | ||||
| 12572 | MVT RotateVT; | |||
| 12573 | int RotateAmt = matchShuffleAsBitRotate(RotateVT, VT.getScalarSizeInBits(), | |||
| 12574 | Subtarget, Mask); | |||
| 12575 | if (RotateAmt < 0) | |||
| 12576 | return SDValue(); | |||
| 12577 | ||||
| 12578 | // For pre-SSSE3 targets, if we are shuffling vXi8 elts then ISD::ROTL, | |||
| 12579 | // expanded to OR(SRL,SHL), will be more efficient, but if they can | |||
| 12580 | // widen to vXi16 or more then existing lowering should will be better. | |||
| 12581 | if (!IsLegal) { | |||
| 12582 | if ((RotateAmt % 16) == 0) | |||
| 12583 | return SDValue(); | |||
| 12584 | // TODO: Use getTargetVShiftByConstNode. | |||
| 12585 | unsigned ShlAmt = RotateAmt; | |||
| 12586 | unsigned SrlAmt = RotateVT.getScalarSizeInBits() - RotateAmt; | |||
| 12587 | V1 = DAG.getBitcast(RotateVT, V1); | |||
| 12588 | SDValue SHL = DAG.getNode(X86ISD::VSHLI, DL, RotateVT, V1, | |||
| 12589 | DAG.getTargetConstant(ShlAmt, DL, MVT::i8)); | |||
| 12590 | SDValue SRL = DAG.getNode(X86ISD::VSRLI, DL, RotateVT, V1, | |||
| 12591 | DAG.getTargetConstant(SrlAmt, DL, MVT::i8)); | |||
| 12592 | SDValue Rot = DAG.getNode(ISD::OR, DL, RotateVT, SHL, SRL); | |||
| 12593 | return DAG.getBitcast(VT, Rot); | |||
| 12594 | } | |||
| 12595 | ||||
| 12596 | SDValue Rot = | |||
| 12597 | DAG.getNode(X86ISD::VROTLI, DL, RotateVT, DAG.getBitcast(RotateVT, V1), | |||
| 12598 | DAG.getTargetConstant(RotateAmt, DL, MVT::i8)); | |||
| 12599 | return DAG.getBitcast(VT, Rot); | |||
| 12600 | } | |||
| 12601 | ||||
| 12602 | /// Try to match a vector shuffle as an element rotation. | |||
| 12603 | /// | |||
| 12604 | /// This is used for support PALIGNR for SSSE3 or VALIGND/Q for AVX512. | |||
| 12605 | static int matchShuffleAsElementRotate(SDValue &V1, SDValue &V2, | |||
| 12606 | ArrayRef<int> Mask) { | |||
| 12607 | int NumElts = Mask.size(); | |||
| 12608 | ||||
| 12609 | // We need to detect various ways of spelling a rotation: | |||
| 12610 | // [11, 12, 13, 14, 15, 0, 1, 2] | |||
| 12611 | // [-1, 12, 13, 14, -1, -1, 1, -1] | |||
| 12612 | // [-1, -1, -1, -1, -1, -1, 1, 2] | |||
| 12613 | // [ 3, 4, 5, 6, 7, 8, 9, 10] | |||
| 12614 | // [-1, 4, 5, 6, -1, -1, 9, -1] | |||
| 12615 | // [-1, 4, 5, 6, -1, -1, -1, -1] | |||
| 12616 | int Rotation = 0; | |||
| 12617 | SDValue Lo, Hi; | |||
| 12618 | for (int i = 0; i < NumElts; ++i) { | |||
| 12619 | int M = Mask[i]; | |||
| 12620 | assert((M == SM_SentinelUndef || (0 <= M && M < (2*NumElts))) &&((void)0) | |||
| 12621 | "Unexpected mask index.")((void)0); | |||
| 12622 | if (M < 0) | |||
| 12623 | continue; | |||
| 12624 | ||||
| 12625 | // Determine where a rotated vector would have started. | |||
| 12626 | int StartIdx = i - (M % NumElts); | |||
| 12627 | if (StartIdx == 0) | |||
| 12628 | // The identity rotation isn't interesting, stop. | |||
| 12629 | return -1; | |||
| 12630 | ||||
| 12631 | // If we found the tail of a vector the rotation must be the missing | |||
| 12632 | // front. If we found the head of a vector, it must be how much of the | |||
| 12633 | // head. | |||
| 12634 | int CandidateRotation = StartIdx < 0 ? -StartIdx : NumElts - StartIdx; | |||
| 12635 | ||||
| 12636 | if (Rotation == 0) | |||
| 12637 | Rotation = CandidateRotation; | |||
| 12638 | else if (Rotation != CandidateRotation) | |||
| 12639 | // The rotations don't match, so we can't match this mask. | |||
| 12640 | return -1; | |||
| 12641 | ||||
| 12642 | // Compute which value this mask is pointing at. | |||
| 12643 | SDValue MaskV = M < NumElts ? V1 : V2; | |||
| 12644 | ||||
| 12645 | // Compute which of the two target values this index should be assigned | |||
| 12646 | // to. This reflects whether the high elements are remaining or the low | |||
| 12647 | // elements are remaining. | |||
| 12648 | SDValue &TargetV = StartIdx < 0 ? Hi : Lo; | |||
| 12649 | ||||
| 12650 | // Either set up this value if we've not encountered it before, or check | |||
| 12651 | // that it remains consistent. | |||
| 12652 | if (!TargetV) | |||
| 12653 | TargetV = MaskV; | |||
| 12654 | else if (TargetV != MaskV) | |||
| 12655 | // This may be a rotation, but it pulls from the inputs in some | |||
| 12656 | // unsupported interleaving. | |||
| 12657 | return -1; | |||
| 12658 | } | |||
| 12659 | ||||
| 12660 | // Check that we successfully analyzed the mask, and normalize the results. | |||
| 12661 | assert(Rotation != 0 && "Failed to locate a viable rotation!")((void)0); | |||
| 12662 | assert((Lo || Hi) && "Failed to find a rotated input vector!")((void)0); | |||
| 12663 | if (!Lo) | |||
| 12664 | Lo = Hi; | |||
| 12665 | else if (!Hi) | |||
| 12666 | Hi = Lo; | |||
| 12667 | ||||
| 12668 | V1 = Lo; | |||
| 12669 | V2 = Hi; | |||
| 12670 | ||||
| 12671 | return Rotation; | |||
| 12672 | } | |||
| 12673 | ||||
| 12674 | /// Try to lower a vector shuffle as a byte rotation. | |||
| 12675 | /// | |||
| 12676 | /// SSSE3 has a generic PALIGNR instruction in x86 that will do an arbitrary | |||
| 12677 | /// byte-rotation of the concatenation of two vectors; pre-SSSE3 can use | |||
| 12678 | /// a PSRLDQ/PSLLDQ/POR pattern to get a similar effect. This routine will | |||
| 12679 | /// try to generically lower a vector shuffle through such an pattern. It | |||
| 12680 | /// does not check for the profitability of lowering either as PALIGNR or | |||
| 12681 | /// PSRLDQ/PSLLDQ/POR, only whether the mask is valid to lower in that form. | |||
| 12682 | /// This matches shuffle vectors that look like: | |||
| 12683 | /// | |||
| 12684 | /// v8i16 [11, 12, 13, 14, 15, 0, 1, 2] | |||
| 12685 | /// | |||
| 12686 | /// Essentially it concatenates V1 and V2, shifts right by some number of | |||
| 12687 | /// elements, and takes the low elements as the result. Note that while this is | |||
| 12688 | /// specified as a *right shift* because x86 is little-endian, it is a *left | |||
| 12689 | /// rotate* of the vector lanes. | |||
| 12690 | static int matchShuffleAsByteRotate(MVT VT, SDValue &V1, SDValue &V2, | |||
| 12691 | ArrayRef<int> Mask) { | |||
| 12692 | // Don't accept any shuffles with zero elements. | |||
| 12693 | if (isAnyZero(Mask)) | |||
| 12694 | return -1; | |||
| 12695 | ||||
| 12696 | // PALIGNR works on 128-bit lanes. | |||
| 12697 | SmallVector<int, 16> RepeatedMask; | |||
| 12698 | if (!is128BitLaneRepeatedShuffleMask(VT, Mask, RepeatedMask)) | |||
| 12699 | return -1; | |||
| 12700 | ||||
| 12701 | int Rotation = matchShuffleAsElementRotate(V1, V2, RepeatedMask); | |||
| 12702 | if (Rotation <= 0) | |||
| 12703 | return -1; | |||
| 12704 | ||||
| 12705 | // PALIGNR rotates bytes, so we need to scale the | |||
| 12706 | // rotation based on how many bytes are in the vector lane. | |||
| 12707 | int NumElts = RepeatedMask.size(); | |||
| 12708 | int Scale = 16 / NumElts; | |||
| 12709 | return Rotation * Scale; | |||
| 12710 | } | |||
| 12711 | ||||
| 12712 | static SDValue lowerShuffleAsByteRotate(const SDLoc &DL, MVT VT, SDValue V1, | |||
| 12713 | SDValue V2, ArrayRef<int> Mask, | |||
| 12714 | const X86Subtarget &Subtarget, | |||
| 12715 | SelectionDAG &DAG) { | |||
| 12716 | assert(!isNoopShuffleMask(Mask) && "We shouldn't lower no-op shuffles!")((void)0); | |||
| 12717 | ||||
| 12718 | SDValue Lo = V1, Hi = V2; | |||
| 12719 | int ByteRotation = matchShuffleAsByteRotate(VT, Lo, Hi, Mask); | |||
| 12720 | if (ByteRotation <= 0) | |||
| 12721 | return SDValue(); | |||
| 12722 | ||||
| 12723 | // Cast the inputs to i8 vector of correct length to match PALIGNR or | |||
| 12724 | // PSLLDQ/PSRLDQ. | |||
| 12725 | MVT ByteVT = MVT::getVectorVT(MVT::i8, VT.getSizeInBits() / 8); | |||
| 12726 | Lo = DAG.getBitcast(ByteVT, Lo); | |||
| 12727 | Hi = DAG.getBitcast(ByteVT, Hi); | |||
| 12728 | ||||
| 12729 | // SSSE3 targets can use the palignr instruction. | |||
| 12730 | if (Subtarget.hasSSSE3()) { | |||
| 12731 | assert((!VT.is512BitVector() || Subtarget.hasBWI()) &&((void)0) | |||
| 12732 | "512-bit PALIGNR requires BWI instructions")((void)0); | |||
| 12733 | return DAG.getBitcast( | |||
| 12734 | VT, DAG.getNode(X86ISD::PALIGNR, DL, ByteVT, Lo, Hi, | |||
| 12735 | DAG.getTargetConstant(ByteRotation, DL, MVT::i8))); | |||
| 12736 | } | |||
| 12737 | ||||
| 12738 | assert(VT.is128BitVector() &&((void)0) | |||
| 12739 | "Rotate-based lowering only supports 128-bit lowering!")((void)0); | |||
| 12740 | assert(Mask.size() <= 16 &&((void)0) | |||
| 12741 | "Can shuffle at most 16 bytes in a 128-bit vector!")((void)0); | |||
| 12742 | assert(ByteVT == MVT::v16i8 &&((void)0) | |||
| 12743 | "SSE2 rotate lowering only needed for v16i8!")((void)0); | |||
| 12744 | ||||
| 12745 | // Default SSE2 implementation | |||
| 12746 | int LoByteShift = 16 - ByteRotation; | |||
| 12747 | int HiByteShift = ByteRotation; | |||
| 12748 | ||||
| 12749 | SDValue LoShift = | |||
| 12750 | DAG.getNode(X86ISD::VSHLDQ, DL, MVT::v16i8, Lo, | |||
| 12751 | DAG.getTargetConstant(LoByteShift, DL, MVT::i8)); | |||
| 12752 | SDValue HiShift = | |||
| 12753 | DAG.getNode(X86ISD::VSRLDQ, DL, MVT::v16i8, Hi, | |||
| 12754 | DAG.getTargetConstant(HiByteShift, DL, MVT::i8)); | |||
| 12755 | return DAG.getBitcast(VT, | |||
| 12756 | DAG.getNode(ISD::OR, DL, MVT::v16i8, LoShift, HiShift)); | |||
| 12757 | } | |||
| 12758 | ||||
| 12759 | /// Try to lower a vector shuffle as a dword/qword rotation. | |||
| 12760 | /// | |||
| 12761 | /// AVX512 has a VALIGND/VALIGNQ instructions that will do an arbitrary | |||
| 12762 | /// rotation of the concatenation of two vectors; This routine will | |||
| 12763 | /// try to generically lower a vector shuffle through such an pattern. | |||
| 12764 | /// | |||
| 12765 | /// Essentially it concatenates V1 and V2, shifts right by some number of | |||
| 12766 | /// elements, and takes the low elements as the result. Note that while this is | |||
| 12767 | /// specified as a *right shift* because x86 is little-endian, it is a *left | |||
| 12768 | /// rotate* of the vector lanes. | |||
| 12769 | static SDValue lowerShuffleAsVALIGN(const SDLoc &DL, MVT VT, SDValue V1, | |||
| 12770 | SDValue V2, ArrayRef<int> Mask, | |||
| 12771 | const X86Subtarget &Subtarget, | |||
| 12772 | SelectionDAG &DAG) { | |||
| 12773 | assert((VT.getScalarType() == MVT::i32 || VT.getScalarType() == MVT::i64) &&((void)0) | |||
| 12774 | "Only 32-bit and 64-bit elements are supported!")((void)0); | |||
| 12775 | ||||
| 12776 | // 128/256-bit vectors are only supported with VLX. | |||
| 12777 | assert((Subtarget.hasVLX() || (!VT.is128BitVector() && !VT.is256BitVector()))((void)0) | |||
| 12778 | && "VLX required for 128/256-bit vectors")((void)0); | |||
| 12779 | ||||
| 12780 | SDValue Lo = V1, Hi = V2; | |||
| 12781 | int Rotation = matchShuffleAsElementRotate(Lo, Hi, Mask); | |||
| 12782 | if (Rotation <= 0) | |||
| 12783 | return SDValue(); | |||
| 12784 | ||||
| 12785 | return DAG.getNode(X86ISD::VALIGN, DL, VT, Lo, Hi, | |||
| 12786 | DAG.getTargetConstant(Rotation, DL, MVT::i8)); | |||
| 12787 | } | |||
| 12788 | ||||
| 12789 | /// Try to lower a vector shuffle as a byte shift sequence. | |||
| 12790 | static SDValue lowerShuffleAsByteShiftMask(const SDLoc &DL, MVT VT, SDValue V1, | |||
| 12791 | SDValue V2, ArrayRef<int> Mask, | |||
| 12792 | const APInt &Zeroable, | |||
| 12793 | const X86Subtarget &Subtarget, | |||
| 12794 | SelectionDAG &DAG) { | |||
| 12795 | assert(!isNoopShuffleMask(Mask) && "We shouldn't lower no-op shuffles!")((void)0); | |||
| 12796 | assert(VT.is128BitVector() && "Only 128-bit vectors supported")((void)0); | |||
| 12797 | ||||
| 12798 | // We need a shuffle that has zeros at one/both ends and a sequential | |||
| 12799 | // shuffle from one source within. | |||
| 12800 | unsigned ZeroLo = Zeroable.countTrailingOnes(); | |||
| 12801 | unsigned ZeroHi = Zeroable.countLeadingOnes(); | |||
| 12802 | if (!ZeroLo && !ZeroHi) | |||
| 12803 | return SDValue(); | |||
| 12804 | ||||
| 12805 | unsigned NumElts = Mask.size(); | |||
| 12806 | unsigned Len = NumElts - (ZeroLo + ZeroHi); | |||
| 12807 | if (!isSequentialOrUndefInRange(Mask, ZeroLo, Len, Mask[ZeroLo])) | |||
| 12808 | return SDValue(); | |||
| 12809 | ||||
| 12810 | unsigned Scale = VT.getScalarSizeInBits() / 8; | |||
| 12811 | ArrayRef<int> StubMask = Mask.slice(ZeroLo, Len); | |||
| 12812 | if (!isUndefOrInRange(StubMask, 0, NumElts) && | |||
| 12813 | !isUndefOrInRange(StubMask, NumElts, 2 * NumElts)) | |||
| 12814 | return SDValue(); | |||
| 12815 | ||||
| 12816 | SDValue Res = Mask[ZeroLo] < (int)NumElts ? V1 : V2; | |||
| 12817 | Res = DAG.getBitcast(MVT::v16i8, Res); | |||
| 12818 | ||||
| 12819 | // Use VSHLDQ/VSRLDQ ops to zero the ends of a vector and leave an | |||
| 12820 | // inner sequential set of elements, possibly offset: | |||
| 12821 | // 01234567 --> zzzzzz01 --> 1zzzzzzz | |||
| 12822 | // 01234567 --> 4567zzzz --> zzzzz456 | |||
| 12823 | // 01234567 --> z0123456 --> 3456zzzz --> zz3456zz | |||
| 12824 | if (ZeroLo == 0) { | |||
| 12825 | unsigned Shift = (NumElts - 1) - (Mask[ZeroLo + Len - 1] % NumElts); | |||
| 12826 | Res = DAG.getNode(X86ISD::VSHLDQ, DL, MVT::v16i8, Res, | |||
| 12827 | DAG.getTargetConstant(Scale * Shift, DL, MVT::i8)); | |||
| 12828 | Res = DAG.getNode(X86ISD::VSRLDQ, DL, MVT::v16i8, Res, | |||
| 12829 | DAG.getTargetConstant(Scale * ZeroHi, DL, MVT::i8)); | |||
| 12830 | } else if (ZeroHi == 0) { | |||
| 12831 | unsigned Shift = Mask[ZeroLo] % NumElts; | |||
| 12832 | Res = DAG.getNode(X86ISD::VSRLDQ, DL, MVT::v16i8, Res, | |||
| 12833 | DAG.getTargetConstant(Scale * Shift, DL, MVT::i8)); | |||
| 12834 | Res = DAG.getNode(X86ISD::VSHLDQ, DL, MVT::v16i8, Res, | |||
| 12835 | DAG.getTargetConstant(Scale * ZeroLo, DL, MVT::i8)); | |||
| 12836 | } else if (!Subtarget.hasSSSE3()) { | |||
| 12837 | // If we don't have PSHUFB then its worth avoiding an AND constant mask | |||
| 12838 | // by performing 3 byte shifts. Shuffle combining can kick in above that. | |||
| 12839 | // TODO: There may be some cases where VSH{LR}DQ+PAND is still better. | |||
| 12840 | unsigned Shift = (NumElts - 1) - (Mask[ZeroLo + Len - 1] % NumElts); | |||
| 12841 | Res = DAG.getNode(X86ISD::VSHLDQ, DL, MVT::v16i8, Res, | |||
| 12842 | DAG.getTargetConstant(Scale * Shift, DL, MVT::i8)); | |||
| 12843 | Shift += Mask[ZeroLo] % NumElts; | |||
| 12844 | Res = DAG.getNode(X86ISD::VSRLDQ, DL, MVT::v16i8, Res, | |||
| 12845 | DAG.getTargetConstant(Scale * Shift, DL, MVT::i8)); | |||
| 12846 | Res = DAG.getNode(X86ISD::VSHLDQ, DL, MVT::v16i8, Res, | |||
| 12847 | DAG.getTargetConstant(Scale * ZeroLo, DL, MVT::i8)); | |||
| 12848 | } else | |||
| 12849 | return SDValue(); | |||
| 12850 | ||||
| 12851 | return DAG.getBitcast(VT, Res); | |||
| 12852 | } | |||
| 12853 | ||||
| 12854 | /// Try to lower a vector shuffle as a bit shift (shifts in zeros). | |||
| 12855 | /// | |||
| 12856 | /// Attempts to match a shuffle mask against the PSLL(W/D/Q/DQ) and | |||
| 12857 | /// PSRL(W/D/Q/DQ) SSE2 and AVX2 logical bit-shift instructions. The function | |||
| 12858 | /// matches elements from one of the input vectors shuffled to the left or | |||
| 12859 | /// right with zeroable elements 'shifted in'. It handles both the strictly | |||
| 12860 | /// bit-wise element shifts and the byte shift across an entire 128-bit double | |||
| 12861 | /// quad word lane. | |||
| 12862 | /// | |||
| 12863 | /// PSHL : (little-endian) left bit shift. | |||
| 12864 | /// [ zz, 0, zz, 2 ] | |||
| 12865 | /// [ -1, 4, zz, -1 ] | |||
| 12866 | /// PSRL : (little-endian) right bit shift. | |||
| 12867 | /// [ 1, zz, 3, zz] | |||
| 12868 | /// [ -1, -1, 7, zz] | |||
| 12869 | /// PSLLDQ : (little-endian) left byte shift | |||
| 12870 | /// [ zz, 0, 1, 2, 3, 4, 5, 6] | |||
| 12871 | /// [ zz, zz, -1, -1, 2, 3, 4, -1] | |||
| 12872 | /// [ zz, zz, zz, zz, zz, zz, -1, 1] | |||
| 12873 | /// PSRLDQ : (little-endian) right byte shift | |||
| 12874 | /// [ 5, 6, 7, zz, zz, zz, zz, zz] | |||
| 12875 | /// [ -1, 5, 6, 7, zz, zz, zz, zz] | |||
| 12876 | /// [ 1, 2, -1, -1, -1, -1, zz, zz] | |||
| 12877 | static int matchShuffleAsShift(MVT &ShiftVT, unsigned &Opcode, | |||
| 12878 | unsigned ScalarSizeInBits, ArrayRef<int> Mask, | |||
| 12879 | int MaskOffset, const APInt &Zeroable, | |||
| 12880 | const X86Subtarget &Subtarget) { | |||
| 12881 | int Size = Mask.size(); | |||
| 12882 | unsigned SizeInBits = Size * ScalarSizeInBits; | |||
| 12883 | ||||
| 12884 | auto CheckZeros = [&](int Shift, int Scale, bool Left) { | |||
| 12885 | for (int i = 0; i < Size; i += Scale) | |||
| 12886 | for (int j = 0; j < Shift; ++j) | |||
| 12887 | if (!Zeroable[i + j + (Left ? 0 : (Scale - Shift))]) | |||
| 12888 | return false; | |||
| 12889 | ||||
| 12890 | return true; | |||
| 12891 | }; | |||
| 12892 | ||||
| 12893 | auto MatchShift = [&](int Shift, int Scale, bool Left) { | |||
| 12894 | for (int i = 0; i != Size; i += Scale) { | |||
| 12895 | unsigned Pos = Left ? i + Shift : i; | |||
| 12896 | unsigned Low = Left ? i : i + Shift; | |||
| 12897 | unsigned Len = Scale - Shift; | |||
| 12898 | if (!isSequentialOrUndefInRange(Mask, Pos, Len, Low + MaskOffset)) | |||
| 12899 | return -1; | |||
| 12900 | } | |||
| 12901 | ||||
| 12902 | int ShiftEltBits = ScalarSizeInBits * Scale; | |||
| 12903 | bool ByteShift = ShiftEltBits > 64; | |||
| 12904 | Opcode = Left ? (ByteShift ? X86ISD::VSHLDQ : X86ISD::VSHLI) | |||
| 12905 | : (ByteShift ? X86ISD::VSRLDQ : X86ISD::VSRLI); | |||
| 12906 | int ShiftAmt = Shift * ScalarSizeInBits / (ByteShift ? 8 : 1); | |||
| 12907 | ||||
| 12908 | // Normalize the scale for byte shifts to still produce an i64 element | |||
| 12909 | // type. | |||
| 12910 | Scale = ByteShift ? Scale / 2 : Scale; | |||
| 12911 | ||||
| 12912 | // We need to round trip through the appropriate type for the shift. | |||
| 12913 | MVT ShiftSVT = MVT::getIntegerVT(ScalarSizeInBits * Scale); | |||
| 12914 | ShiftVT = ByteShift ? MVT::getVectorVT(MVT::i8, SizeInBits / 8) | |||
| 12915 | : MVT::getVectorVT(ShiftSVT, Size / Scale); | |||
| 12916 | return (int)ShiftAmt; | |||
| 12917 | }; | |||
| 12918 | ||||
| 12919 | // SSE/AVX supports logical shifts up to 64-bit integers - so we can just | |||
| 12920 | // keep doubling the size of the integer elements up to that. We can | |||
| 12921 | // then shift the elements of the integer vector by whole multiples of | |||
| 12922 | // their width within the elements of the larger integer vector. Test each | |||
| 12923 | // multiple to see if we can find a match with the moved element indices | |||
| 12924 | // and that the shifted in elements are all zeroable. | |||
| 12925 | unsigned MaxWidth = ((SizeInBits == 512) && !Subtarget.hasBWI() ? 64 : 128); | |||
| 12926 | for (int Scale = 2; Scale * ScalarSizeInBits <= MaxWidth; Scale *= 2) | |||
| 12927 | for (int Shift = 1; Shift != Scale; ++Shift) | |||
| 12928 | for (bool Left : {true, false}) | |||
| 12929 | if (CheckZeros(Shift, Scale, Left)) { | |||
| 12930 | int ShiftAmt = MatchShift(Shift, Scale, Left); | |||
| 12931 | if (0 < ShiftAmt) | |||
| 12932 | return ShiftAmt; | |||
| 12933 | } | |||
| 12934 | ||||
| 12935 | // no match | |||
| 12936 | return -1; | |||
| 12937 | } | |||
| 12938 | ||||
| 12939 | static SDValue lowerShuffleAsShift(const SDLoc &DL, MVT VT, SDValue V1, | |||
| 12940 | SDValue V2, ArrayRef<int> Mask, | |||
| 12941 | const APInt &Zeroable, | |||
| 12942 | const X86Subtarget &Subtarget, | |||
| 12943 | SelectionDAG &DAG) { | |||
| 12944 | int Size = Mask.size(); | |||
| 12945 | assert(Size == (int)VT.getVectorNumElements() && "Unexpected mask size")((void)0); | |||
| 12946 | ||||
| 12947 | MVT ShiftVT; | |||
| 12948 | SDValue V = V1; | |||
| 12949 | unsigned Opcode; | |||
| 12950 | ||||
| 12951 | // Try to match shuffle against V1 shift. | |||
| 12952 | int ShiftAmt = matchShuffleAsShift(ShiftVT, Opcode, VT.getScalarSizeInBits(), | |||
| 12953 | Mask, 0, Zeroable, Subtarget); | |||
| 12954 | ||||
| 12955 | // If V1 failed, try to match shuffle against V2 shift. | |||
| 12956 | if (ShiftAmt < 0) { | |||
| 12957 | ShiftAmt = matchShuffleAsShift(ShiftVT, Opcode, VT.getScalarSizeInBits(), | |||
| 12958 | Mask, Size, Zeroable, Subtarget); | |||
| 12959 | V = V2; | |||
| 12960 | } | |||
| 12961 | ||||
| 12962 | if (ShiftAmt < 0) | |||
| 12963 | return SDValue(); | |||
| 12964 | ||||
| 12965 | assert(DAG.getTargetLoweringInfo().isTypeLegal(ShiftVT) &&((void)0) | |||
| 12966 | "Illegal integer vector type")((void)0); | |||
| 12967 | V = DAG.getBitcast(ShiftVT, V); | |||
| 12968 | V = DAG.getNode(Opcode, DL, ShiftVT, V, | |||
| 12969 | DAG.getTargetConstant(ShiftAmt, DL, MVT::i8)); | |||
| 12970 | return DAG.getBitcast(VT, V); | |||
| 12971 | } | |||
| 12972 | ||||
| 12973 | // EXTRQ: Extract Len elements from lower half of source, starting at Idx. | |||
| 12974 | // Remainder of lower half result is zero and upper half is all undef. | |||
| 12975 | static bool matchShuffleAsEXTRQ(MVT VT, SDValue &V1, SDValue &V2, | |||
| 12976 | ArrayRef<int> Mask, uint64_t &BitLen, | |||
| 12977 | uint64_t &BitIdx, const APInt &Zeroable) { | |||
| 12978 | int Size = Mask.size(); | |||
| 12979 | int HalfSize = Size / 2; | |||
| 12980 | assert(Size == (int)VT.getVectorNumElements() && "Unexpected mask size")((void)0); | |||
| 12981 | assert(!Zeroable.isAllOnesValue() && "Fully zeroable shuffle mask")((void)0); | |||
| 12982 | ||||
| 12983 | // Upper half must be undefined. | |||
| 12984 | if (!isUndefUpperHalf(Mask)) | |||
| 12985 | return false; | |||
| 12986 | ||||
| 12987 | // Determine the extraction length from the part of the | |||
| 12988 | // lower half that isn't zeroable. | |||
| 12989 | int Len = HalfSize; | |||
| 12990 | for (; Len > 0; --Len) | |||
| 12991 | if (!Zeroable[Len - 1]) | |||
| 12992 | break; | |||
| 12993 | assert(Len > 0 && "Zeroable shuffle mask")((void)0); | |||
| 12994 | ||||
| 12995 | // Attempt to match first Len sequential elements from the lower half. | |||
| 12996 | SDValue Src; | |||
| 12997 | int Idx = -1; | |||
| 12998 | for (int i = 0; i != Len; ++i) { | |||
| 12999 | int M = Mask[i]; | |||
| 13000 | if (M == SM_SentinelUndef) | |||
| 13001 | continue; | |||
| 13002 | SDValue &V = (M < Size ? V1 : V2); | |||
| 13003 | M = M % Size; | |||
| 13004 | ||||
| 13005 | // The extracted elements must start at a valid index and all mask | |||
| 13006 | // elements must be in the lower half. | |||
| 13007 | if (i > M || M >= HalfSize) | |||
| 13008 | return false; | |||
| 13009 | ||||
| 13010 | if (Idx < 0 || (Src == V && Idx == (M - i))) { | |||
| 13011 | Src = V; | |||
| 13012 | Idx = M - i; | |||
| 13013 | continue; | |||
| 13014 | } | |||
| 13015 | return false; | |||
| 13016 | } | |||
| 13017 | ||||
| 13018 | if (!Src || Idx < 0) | |||
| 13019 | return false; | |||
| 13020 | ||||
| 13021 | assert((Idx + Len) <= HalfSize && "Illegal extraction mask")((void)0); | |||
| 13022 | BitLen = (Len * VT.getScalarSizeInBits()) & 0x3f; | |||
| 13023 | BitIdx = (Idx * VT.getScalarSizeInBits()) & 0x3f; | |||
| 13024 | V1 = Src; | |||
| 13025 | return true; | |||
| 13026 | } | |||
| 13027 | ||||
| 13028 | // INSERTQ: Extract lowest Len elements from lower half of second source and | |||
| 13029 | // insert over first source, starting at Idx. | |||
| 13030 | // { A[0], .., A[Idx-1], B[0], .., B[Len-1], A[Idx+Len], .., UNDEF, ... } | |||
| 13031 | static bool matchShuffleAsINSERTQ(MVT VT, SDValue &V1, SDValue &V2, | |||
| 13032 | ArrayRef<int> Mask, uint64_t &BitLen, | |||
| 13033 | uint64_t &BitIdx) { | |||
| 13034 | int Size = Mask.size(); | |||
| 13035 | int HalfSize = Size / 2; | |||
| 13036 | assert(Size == (int)VT.getVectorNumElements() && "Unexpected mask size")((void)0); | |||
| 13037 | ||||
| 13038 | // Upper half must be undefined. | |||
| 13039 | if (!isUndefUpperHalf(Mask)) | |||
| 13040 | return false; | |||
| 13041 | ||||
| 13042 | for (int Idx = 0; Idx != HalfSize; ++Idx) { | |||
| 13043 | SDValue Base; | |||
| 13044 | ||||
| 13045 | // Attempt to match first source from mask before insertion point. | |||
| 13046 | if (isUndefInRange(Mask, 0, Idx)) { | |||
| 13047 | /* EMPTY */ | |||
| 13048 | } else if (isSequentialOrUndefInRange(Mask, 0, Idx, 0)) { | |||
| 13049 | Base = V1; | |||
| 13050 | } else if (isSequentialOrUndefInRange(Mask, 0, Idx, Size)) { | |||
| 13051 | Base = V2; | |||
| 13052 | } else { | |||
| 13053 | continue; | |||
| 13054 | } | |||
| 13055 | ||||
| 13056 | // Extend the extraction length looking to match both the insertion of | |||
| 13057 | // the second source and the remaining elements of the first. | |||
| 13058 | for (int Hi = Idx + 1; Hi <= HalfSize; ++Hi) { | |||
| 13059 | SDValue Insert; | |||
| 13060 | int Len = Hi - Idx; | |||
| 13061 | ||||
| 13062 | // Match insertion. | |||
| 13063 | if (isSequentialOrUndefInRange(Mask, Idx, Len, 0)) { | |||
| 13064 | Insert = V1; | |||
| 13065 | } else if (isSequentialOrUndefInRange(Mask, Idx, Len, Size)) { | |||
| 13066 | Insert = V2; | |||
| 13067 | } else { | |||
| 13068 | continue; | |||
| 13069 | } | |||
| 13070 | ||||
| 13071 | // Match the remaining elements of the lower half. | |||
| 13072 | if (isUndefInRange(Mask, Hi, HalfSize - Hi)) { | |||
| 13073 | /* EMPTY */ | |||
| 13074 | } else if ((!Base || (Base == V1)) && | |||
| 13075 | isSequentialOrUndefInRange(Mask, Hi, HalfSize - Hi, Hi)) { | |||
| 13076 | Base = V1; | |||
| 13077 | } else if ((!Base || (Base == V2)) && | |||
| 13078 | isSequentialOrUndefInRange(Mask, Hi, HalfSize - Hi, | |||
| 13079 | Size + Hi)) { | |||
| 13080 | Base = V2; | |||
| 13081 | } else { | |||
| 13082 | continue; | |||
| 13083 | } | |||
| 13084 | ||||
| 13085 | BitLen = (Len * VT.getScalarSizeInBits()) & 0x3f; | |||
| 13086 | BitIdx = (Idx * VT.getScalarSizeInBits()) & 0x3f; | |||
| 13087 | V1 = Base; | |||
| 13088 | V2 = Insert; | |||
| 13089 | return true; | |||
| 13090 | } | |||
| 13091 | } | |||
| 13092 | ||||
| 13093 | return false; | |||
| 13094 | } | |||
| 13095 | ||||
| 13096 | /// Try to lower a vector shuffle using SSE4a EXTRQ/INSERTQ. | |||
| 13097 | static SDValue lowerShuffleWithSSE4A(const SDLoc &DL, MVT VT, SDValue V1, | |||
| 13098 | SDValue V2, ArrayRef<int> Mask, | |||
| 13099 | const APInt &Zeroable, SelectionDAG &DAG) { | |||
| 13100 | uint64_t BitLen, BitIdx; | |||
| 13101 | if (matchShuffleAsEXTRQ(VT, V1, V2, Mask, BitLen, BitIdx, Zeroable)) | |||
| 13102 | return DAG.getNode(X86ISD::EXTRQI, DL, VT, V1, | |||
| 13103 | DAG.getTargetConstant(BitLen, DL, MVT::i8), | |||
| 13104 | DAG.getTargetConstant(BitIdx, DL, MVT::i8)); | |||
| 13105 | ||||
| 13106 | if (matchShuffleAsINSERTQ(VT, V1, V2, Mask, BitLen, BitIdx)) | |||
| 13107 | return DAG.getNode(X86ISD::INSERTQI, DL, VT, V1 ? V1 : DAG.getUNDEF(VT), | |||
| 13108 | V2 ? V2 : DAG.getUNDEF(VT), | |||
| 13109 | DAG.getTargetConstant(BitLen, DL, MVT::i8), | |||
| 13110 | DAG.getTargetConstant(BitIdx, DL, MVT::i8)); | |||
| 13111 | ||||
| 13112 | return SDValue(); | |||
| 13113 | } | |||
| 13114 | ||||
| 13115 | /// Lower a vector shuffle as a zero or any extension. | |||
| 13116 | /// | |||
| 13117 | /// Given a specific number of elements, element bit width, and extension | |||
| 13118 | /// stride, produce either a zero or any extension based on the available | |||
| 13119 | /// features of the subtarget. The extended elements are consecutive and | |||
| 13120 | /// begin and can start from an offsetted element index in the input; to | |||
| 13121 | /// avoid excess shuffling the offset must either being in the bottom lane | |||
| 13122 | /// or at the start of a higher lane. All extended elements must be from | |||
| 13123 | /// the same lane. | |||
| 13124 | static SDValue lowerShuffleAsSpecificZeroOrAnyExtend( | |||
| 13125 | const SDLoc &DL, MVT VT, int Scale, int Offset, bool AnyExt, SDValue InputV, | |||
| 13126 | ArrayRef<int> Mask, const X86Subtarget &Subtarget, SelectionDAG &DAG) { | |||
| 13127 | assert(Scale > 1 && "Need a scale to extend.")((void)0); | |||
| 13128 | int EltBits = VT.getScalarSizeInBits(); | |||
| 13129 | int NumElements = VT.getVectorNumElements(); | |||
| 13130 | int NumEltsPerLane = 128 / EltBits; | |||
| 13131 | int OffsetLane = Offset / NumEltsPerLane; | |||
| 13132 | assert((EltBits == 8 || EltBits == 16 || EltBits == 32) &&((void)0) | |||
| 13133 | "Only 8, 16, and 32 bit elements can be extended.")((void)0); | |||
| 13134 | assert(Scale * EltBits <= 64 && "Cannot zero extend past 64 bits.")((void)0); | |||
| 13135 | assert(0 <= Offset && "Extension offset must be positive.")((void)0); | |||
| 13136 | assert((Offset < NumEltsPerLane || Offset % NumEltsPerLane == 0) &&((void)0) | |||
| 13137 | "Extension offset must be in the first lane or start an upper lane.")((void)0); | |||
| 13138 | ||||
| 13139 | // Check that an index is in same lane as the base offset. | |||
| 13140 | auto SafeOffset = [&](int Idx) { | |||
| 13141 | return OffsetLane == (Idx / NumEltsPerLane); | |||
| 13142 | }; | |||
| 13143 | ||||
| 13144 | // Shift along an input so that the offset base moves to the first element. | |||
| 13145 | auto ShuffleOffset = [&](SDValue V) { | |||
| 13146 | if (!Offset) | |||
| 13147 | return V; | |||
| 13148 | ||||
| 13149 | SmallVector<int, 8> ShMask((unsigned)NumElements, -1); | |||
| 13150 | for (int i = 0; i * Scale < NumElements; ++i) { | |||
| 13151 | int SrcIdx = i + Offset; | |||
| 13152 | ShMask[i] = SafeOffset(SrcIdx) ? SrcIdx : -1; | |||
| 13153 | } | |||
| 13154 | return DAG.getVectorShuffle(VT, DL, V, DAG.getUNDEF(VT), ShMask); | |||
| 13155 | }; | |||
| 13156 | ||||
| 13157 | // Found a valid a/zext mask! Try various lowering strategies based on the | |||
| 13158 | // input type and available ISA extensions. | |||
| 13159 | if (Subtarget.hasSSE41()) { | |||
| 13160 | // Not worth offsetting 128-bit vectors if scale == 2, a pattern using | |||
| 13161 | // PUNPCK will catch this in a later shuffle match. | |||
| 13162 | if (Offset && Scale == 2 && VT.is128BitVector()) | |||
| 13163 | return SDValue(); | |||
| 13164 | MVT ExtVT = MVT::getVectorVT(MVT::getIntegerVT(EltBits * Scale), | |||
| 13165 | NumElements / Scale); | |||
| 13166 | InputV = ShuffleOffset(InputV); | |||
| 13167 | InputV = getEXTEND_VECTOR_INREG(AnyExt ? ISD::ANY_EXTEND : ISD::ZERO_EXTEND, | |||
| 13168 | DL, ExtVT, InputV, DAG); | |||
| 13169 | return DAG.getBitcast(VT, InputV); | |||
| 13170 | } | |||
| 13171 | ||||
| 13172 | assert(VT.is128BitVector() && "Only 128-bit vectors can be extended.")((void)0); | |||
| 13173 | ||||
| 13174 | // For any extends we can cheat for larger element sizes and use shuffle | |||
| 13175 | // instructions that can fold with a load and/or copy. | |||
| 13176 | if (AnyExt && EltBits == 32) { | |||
| 13177 | int PSHUFDMask[4] = {Offset, -1, SafeOffset(Offset + 1) ? Offset + 1 : -1, | |||
| 13178 | -1}; | |||
| 13179 | return DAG.getBitcast( | |||
| 13180 | VT, DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, | |||
| 13181 | DAG.getBitcast(MVT::v4i32, InputV), | |||
| 13182 | getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG))); | |||
| 13183 | } | |||
| 13184 | if (AnyExt && EltBits == 16 && Scale > 2) { | |||
| 13185 | int PSHUFDMask[4] = {Offset / 2, -1, | |||
| 13186 | SafeOffset(Offset + 1) ? (Offset + 1) / 2 : -1, -1}; | |||
| 13187 | InputV = DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, | |||
| 13188 | DAG.getBitcast(MVT::v4i32, InputV), | |||
| 13189 | getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)); | |||
| 13190 | int PSHUFWMask[4] = {1, -1, -1, -1}; | |||
| 13191 | unsigned OddEvenOp = (Offset & 1) ? X86ISD::PSHUFLW : X86ISD::PSHUFHW; | |||
| 13192 | return DAG.getBitcast( | |||
| 13193 | VT, DAG.getNode(OddEvenOp, DL, MVT::v8i16, | |||
| 13194 | DAG.getBitcast(MVT::v8i16, InputV), | |||
| 13195 | getV4X86ShuffleImm8ForMask(PSHUFWMask, DL, DAG))); | |||
| 13196 | } | |||
| 13197 | ||||
| 13198 | // The SSE4A EXTRQ instruction can efficiently extend the first 2 lanes | |||
| 13199 | // to 64-bits. | |||
| 13200 | if ((Scale * EltBits) == 64 && EltBits < 32 && Subtarget.hasSSE4A()) { | |||
| 13201 | assert(NumElements == (int)Mask.size() && "Unexpected shuffle mask size!")((void)0); | |||
| 13202 | assert(VT.is128BitVector() && "Unexpected vector width!")((void)0); | |||
| 13203 | ||||
| 13204 | int LoIdx = Offset * EltBits; | |||
| 13205 | SDValue Lo = DAG.getBitcast( | |||
| 13206 | MVT::v2i64, DAG.getNode(X86ISD::EXTRQI, DL, VT, InputV, | |||
| 13207 | DAG.getTargetConstant(EltBits, DL, MVT::i8), | |||
| 13208 | DAG.getTargetConstant(LoIdx, DL, MVT::i8))); | |||
| 13209 | ||||
| 13210 | if (isUndefUpperHalf(Mask) || !SafeOffset(Offset + 1)) | |||
| 13211 | return DAG.getBitcast(VT, Lo); | |||
| 13212 | ||||
| 13213 | int HiIdx = (Offset + 1) * EltBits; | |||
| 13214 | SDValue Hi = DAG.getBitcast( | |||
| 13215 | MVT::v2i64, DAG.getNode(X86ISD::EXTRQI, DL, VT, InputV, | |||
| 13216 | DAG.getTargetConstant(EltBits, DL, MVT::i8), | |||
| 13217 | DAG.getTargetConstant(HiIdx, DL, MVT::i8))); | |||
| 13218 | return DAG.getBitcast(VT, | |||
| 13219 | DAG.getNode(X86ISD::UNPCKL, DL, MVT::v2i64, Lo, Hi)); | |||
| 13220 | } | |||
| 13221 | ||||
| 13222 | // If this would require more than 2 unpack instructions to expand, use | |||
| 13223 | // pshufb when available. We can only use more than 2 unpack instructions | |||
| 13224 | // when zero extending i8 elements which also makes it easier to use pshufb. | |||
| 13225 | if (Scale > 4 && EltBits == 8 && Subtarget.hasSSSE3()) { | |||
| 13226 | assert(NumElements == 16 && "Unexpected byte vector width!")((void)0); | |||
| 13227 | SDValue PSHUFBMask[16]; | |||
| 13228 | for (int i = 0; i < 16; ++i) { | |||
| 13229 | int Idx = Offset + (i / Scale); | |||
| 13230 | if ((i % Scale == 0 && SafeOffset(Idx))) { | |||
| 13231 | PSHUFBMask[i] = DAG.getConstant(Idx, DL, MVT::i8); | |||
| 13232 | continue; | |||
| 13233 | } | |||
| 13234 | PSHUFBMask[i] = | |||
| 13235 | AnyExt ? DAG.getUNDEF(MVT::i8) : DAG.getConstant(0x80, DL, MVT::i8); | |||
| 13236 | } | |||
| 13237 | InputV = DAG.getBitcast(MVT::v16i8, InputV); | |||
| 13238 | return DAG.getBitcast( | |||
| 13239 | VT, DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8, InputV, | |||
| 13240 | DAG.getBuildVector(MVT::v16i8, DL, PSHUFBMask))); | |||
| 13241 | } | |||
| 13242 | ||||
| 13243 | // If we are extending from an offset, ensure we start on a boundary that | |||
| 13244 | // we can unpack from. | |||
| 13245 | int AlignToUnpack = Offset % (NumElements / Scale); | |||
| 13246 | if (AlignToUnpack) { | |||
| 13247 | SmallVector<int, 8> ShMask((unsigned)NumElements, -1); | |||
| 13248 | for (int i = AlignToUnpack; i < NumElements; ++i) | |||
| 13249 | ShMask[i - AlignToUnpack] = i; | |||
| 13250 | InputV = DAG.getVectorShuffle(VT, DL, InputV, DAG.getUNDEF(VT), ShMask); | |||
| 13251 | Offset -= AlignToUnpack; | |||
| 13252 | } | |||
| 13253 | ||||
| 13254 | // Otherwise emit a sequence of unpacks. | |||
| 13255 | do { | |||
| 13256 | unsigned UnpackLoHi = X86ISD::UNPCKL; | |||
| 13257 | if (Offset >= (NumElements / 2)) { | |||
| 13258 | UnpackLoHi = X86ISD::UNPCKH; | |||
| 13259 | Offset -= (NumElements / 2); | |||
| 13260 | } | |||
| 13261 | ||||
| 13262 | MVT InputVT = MVT::getVectorVT(MVT::getIntegerVT(EltBits), NumElements); | |||
| 13263 | SDValue Ext = AnyExt ? DAG.getUNDEF(InputVT) | |||
| 13264 | : getZeroVector(InputVT, Subtarget, DAG, DL); | |||
| 13265 | InputV = DAG.getBitcast(InputVT, InputV); | |||
| 13266 | InputV = DAG.getNode(UnpackLoHi, DL, InputVT, InputV, Ext); | |||
| 13267 | Scale /= 2; | |||
| 13268 | EltBits *= 2; | |||
| 13269 | NumElements /= 2; | |||
| 13270 | } while (Scale > 1); | |||
| 13271 | return DAG.getBitcast(VT, InputV); | |||
| 13272 | } | |||
| 13273 | ||||
| 13274 | /// Try to lower a vector shuffle as a zero extension on any microarch. | |||
| 13275 | /// | |||
| 13276 | /// This routine will try to do everything in its power to cleverly lower | |||
| 13277 | /// a shuffle which happens to match the pattern of a zero extend. It doesn't | |||
| 13278 | /// check for the profitability of this lowering, it tries to aggressively | |||
| 13279 | /// match this pattern. It will use all of the micro-architectural details it | |||
| 13280 | /// can to emit an efficient lowering. It handles both blends with all-zero | |||
| 13281 | /// inputs to explicitly zero-extend and undef-lanes (sometimes undef due to | |||
| 13282 | /// masking out later). | |||
| 13283 | /// | |||
| 13284 | /// The reason we have dedicated lowering for zext-style shuffles is that they | |||
| 13285 | /// are both incredibly common and often quite performance sensitive. | |||
| 13286 | static SDValue lowerShuffleAsZeroOrAnyExtend( | |||
| 13287 | const SDLoc &DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask, | |||
| 13288 | const APInt &Zeroable, const X86Subtarget &Subtarget, | |||
| 13289 | SelectionDAG &DAG) { | |||
| 13290 | int Bits = VT.getSizeInBits(); | |||
| 13291 | int NumLanes = Bits / 128; | |||
| 13292 | int NumElements = VT.getVectorNumElements(); | |||
| 13293 | int NumEltsPerLane = NumElements / NumLanes; | |||
| 13294 | assert(VT.getScalarSizeInBits() <= 32 &&((void)0) | |||
| 13295 | "Exceeds 32-bit integer zero extension limit")((void)0); | |||
| 13296 | assert((int)Mask.size() == NumElements && "Unexpected shuffle mask size")((void)0); | |||
| 13297 | ||||
| 13298 | // Define a helper function to check a particular ext-scale and lower to it if | |||
| 13299 | // valid. | |||
| 13300 | auto Lower = [&](int Scale) -> SDValue { | |||
| 13301 | SDValue InputV; | |||
| 13302 | bool AnyExt = true; | |||
| 13303 | int Offset = 0; | |||
| 13304 | int Matches = 0; | |||
| 13305 | for (int i = 0; i < NumElements; ++i) { | |||
| 13306 | int M = Mask[i]; | |||
| 13307 | if (M < 0) | |||
| 13308 | continue; // Valid anywhere but doesn't tell us anything. | |||
| 13309 | if (i % Scale != 0) { | |||
| 13310 | // Each of the extended elements need to be zeroable. | |||
| 13311 | if (!Zeroable[i]) | |||
| 13312 | return SDValue(); | |||
| 13313 | ||||
| 13314 | // We no longer are in the anyext case. | |||
| 13315 | AnyExt = false; | |||
| 13316 | continue; | |||
| 13317 | } | |||
| 13318 | ||||
| 13319 | // Each of the base elements needs to be consecutive indices into the | |||
| 13320 | // same input vector. | |||
| 13321 | SDValue V = M < NumElements ? V1 : V2; | |||
| 13322 | M = M % NumElements; | |||
| 13323 | if (!InputV) { | |||
| 13324 | InputV = V; | |||
| 13325 | Offset = M - (i / Scale); | |||
| 13326 | } else if (InputV != V) | |||
| 13327 | return SDValue(); // Flip-flopping inputs. | |||
| 13328 | ||||
| 13329 | // Offset must start in the lowest 128-bit lane or at the start of an | |||
| 13330 | // upper lane. | |||
| 13331 | // FIXME: Is it ever worth allowing a negative base offset? | |||
| 13332 | if (!((0 <= Offset && Offset < NumEltsPerLane) || | |||
| 13333 | (Offset % NumEltsPerLane) == 0)) | |||
| 13334 | return SDValue(); | |||
| 13335 | ||||
| 13336 | // If we are offsetting, all referenced entries must come from the same | |||
| 13337 | // lane. | |||
| 13338 | if (Offset && (Offset / NumEltsPerLane) != (M / NumEltsPerLane)) | |||
| 13339 | return SDValue(); | |||
| 13340 | ||||
| 13341 | if ((M % NumElements) != (Offset + (i / Scale))) | |||
| 13342 | return SDValue(); // Non-consecutive strided elements. | |||
| 13343 | Matches++; | |||
| 13344 | } | |||
| 13345 | ||||
| 13346 | // If we fail to find an input, we have a zero-shuffle which should always | |||
| 13347 | // have already been handled. | |||
| 13348 | // FIXME: Maybe handle this here in case during blending we end up with one? | |||
| 13349 | if (!InputV) | |||
| 13350 | return SDValue(); | |||
| 13351 | ||||
| 13352 | // If we are offsetting, don't extend if we only match a single input, we | |||
| 13353 | // can always do better by using a basic PSHUF or PUNPCK. | |||
| 13354 | if (Offset != 0 && Matches < 2) | |||
| 13355 | return SDValue(); | |||
| 13356 | ||||
| 13357 | return lowerShuffleAsSpecificZeroOrAnyExtend(DL, VT, Scale, Offset, AnyExt, | |||
| 13358 | InputV, Mask, Subtarget, DAG); | |||
| 13359 | }; | |||
| 13360 | ||||
| 13361 | // The widest scale possible for extending is to a 64-bit integer. | |||
| 13362 | assert(Bits % 64 == 0 &&((void)0) | |||
| 13363 | "The number of bits in a vector must be divisible by 64 on x86!")((void)0); | |||
| 13364 | int NumExtElements = Bits / 64; | |||
| 13365 | ||||
| 13366 | // Each iteration, try extending the elements half as much, but into twice as | |||
| 13367 | // many elements. | |||
| 13368 | for (; NumExtElements < NumElements; NumExtElements *= 2) { | |||
| 13369 | assert(NumElements % NumExtElements == 0 &&((void)0) | |||
| 13370 | "The input vector size must be divisible by the extended size.")((void)0); | |||
| 13371 | if (SDValue V = Lower(NumElements / NumExtElements)) | |||
| 13372 | return V; | |||
| 13373 | } | |||
| 13374 | ||||
| 13375 | // General extends failed, but 128-bit vectors may be able to use MOVQ. | |||
| 13376 | if (Bits != 128) | |||
| 13377 | return SDValue(); | |||
| 13378 | ||||
| 13379 | // Returns one of the source operands if the shuffle can be reduced to a | |||
| 13380 | // MOVQ, copying the lower 64-bits and zero-extending to the upper 64-bits. | |||
| 13381 | auto CanZExtLowHalf = [&]() { | |||
| 13382 | for (int i = NumElements / 2; i != NumElements; ++i) | |||
| 13383 | if (!Zeroable[i]) | |||
| 13384 | return SDValue(); | |||
| 13385 | if (isSequentialOrUndefInRange(Mask, 0, NumElements / 2, 0)) | |||
| 13386 | return V1; | |||
| 13387 | if (isSequentialOrUndefInRange(Mask, 0, NumElements / 2, NumElements)) | |||
| 13388 | return V2; | |||
| 13389 | return SDValue(); | |||
| 13390 | }; | |||
| 13391 | ||||
| 13392 | if (SDValue V = CanZExtLowHalf()) { | |||
| 13393 | V = DAG.getBitcast(MVT::v2i64, V); | |||
| 13394 | V = DAG.getNode(X86ISD::VZEXT_MOVL, DL, MVT::v2i64, V); | |||
| 13395 | return DAG.getBitcast(VT, V); | |||
| 13396 | } | |||
| 13397 | ||||
| 13398 | // No viable ext lowering found. | |||
| 13399 | return SDValue(); | |||
| 13400 | } | |||
| 13401 | ||||
| 13402 | /// Try to get a scalar value for a specific element of a vector. | |||
| 13403 | /// | |||
| 13404 | /// Looks through BUILD_VECTOR and SCALAR_TO_VECTOR nodes to find a scalar. | |||
| 13405 | static SDValue getScalarValueForVectorElement(SDValue V, int Idx, | |||
| 13406 | SelectionDAG &DAG) { | |||
| 13407 | MVT VT = V.getSimpleValueType(); | |||
| 13408 | MVT EltVT = VT.getVectorElementType(); | |||
| 13409 | V = peekThroughBitcasts(V); | |||
| 13410 | ||||
| 13411 | // If the bitcasts shift the element size, we can't extract an equivalent | |||
| 13412 | // element from it. | |||
| 13413 | MVT NewVT = V.getSimpleValueType(); | |||
| 13414 | if (!NewVT.isVector() || NewVT.getScalarSizeInBits() != VT.getScalarSizeInBits()) | |||
| 13415 | return SDValue(); | |||
| 13416 | ||||
| 13417 | if (V.getOpcode() == ISD::BUILD_VECTOR || | |||
| 13418 | (Idx == 0 && V.getOpcode() == ISD::SCALAR_TO_VECTOR)) { | |||
| 13419 | // Ensure the scalar operand is the same size as the destination. | |||
| 13420 | // FIXME: Add support for scalar truncation where possible. | |||
| 13421 | SDValue S = V.getOperand(Idx); | |||
| 13422 | if (EltVT.getSizeInBits() == S.getSimpleValueType().getSizeInBits()) | |||
| 13423 | return DAG.getBitcast(EltVT, S); | |||
| 13424 | } | |||
| 13425 | ||||
| 13426 | return SDValue(); | |||
| 13427 | } | |||
| 13428 | ||||
| 13429 | /// Helper to test for a load that can be folded with x86 shuffles. | |||
| 13430 | /// | |||
| 13431 | /// This is particularly important because the set of instructions varies | |||
| 13432 | /// significantly based on whether the operand is a load or not. | |||
| 13433 | static bool isShuffleFoldableLoad(SDValue V) { | |||
| 13434 | V = peekThroughBitcasts(V); | |||
| 13435 | return ISD::isNON_EXTLoad(V.getNode()); | |||
| 13436 | } | |||
| 13437 | ||||
| 13438 | /// Try to lower insertion of a single element into a zero vector. | |||
| 13439 | /// | |||
| 13440 | /// This is a common pattern that we have especially efficient patterns to lower | |||
| 13441 | /// across all subtarget feature sets. | |||
| 13442 | static SDValue lowerShuffleAsElementInsertion( | |||
| 13443 | const SDLoc &DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask, | |||
| 13444 | const APInt &Zeroable, const X86Subtarget &Subtarget, | |||
| 13445 | SelectionDAG &DAG) { | |||
| 13446 | MVT ExtVT = VT; | |||
| 13447 | MVT EltVT = VT.getVectorElementType(); | |||
| 13448 | ||||
| 13449 | int V2Index = | |||
| 13450 | find_if(Mask, [&Mask](int M) { return M >= (int)Mask.size(); }) - | |||
| 13451 | Mask.begin(); | |||
| 13452 | bool IsV1Zeroable = true; | |||
| 13453 | for (int i = 0, Size = Mask.size(); i < Size; ++i) | |||
| 13454 | if (i != V2Index && !Zeroable[i]) { | |||
| 13455 | IsV1Zeroable = false; | |||
| 13456 | break; | |||
| 13457 | } | |||
| 13458 | ||||
| 13459 | // Check for a single input from a SCALAR_TO_VECTOR node. | |||
| 13460 | // FIXME: All of this should be canonicalized into INSERT_VECTOR_ELT and | |||
| 13461 | // all the smarts here sunk into that routine. However, the current | |||
| 13462 | // lowering of BUILD_VECTOR makes that nearly impossible until the old | |||
| 13463 | // vector shuffle lowering is dead. | |||
| 13464 | SDValue V2S = getScalarValueForVectorElement(V2, Mask[V2Index] - Mask.size(), | |||
| 13465 | DAG); | |||
| 13466 | if (V2S && DAG.getTargetLoweringInfo().isTypeLegal(V2S.getValueType())) { | |||
| 13467 | // We need to zext the scalar if it is smaller than an i32. | |||
| 13468 | V2S = DAG.getBitcast(EltVT, V2S); | |||
| 13469 | if (EltVT == MVT::i8 || EltVT == MVT::i16) { | |||
| 13470 | // Using zext to expand a narrow element won't work for non-zero | |||
| 13471 | // insertions. | |||
| 13472 | if (!IsV1Zeroable) | |||
| 13473 | return SDValue(); | |||
| 13474 | ||||
| 13475 | // Zero-extend directly to i32. | |||
| 13476 | ExtVT = MVT::getVectorVT(MVT::i32, ExtVT.getSizeInBits() / 32); | |||
| 13477 | V2S = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, V2S); | |||
| 13478 | } | |||
| 13479 | V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, ExtVT, V2S); | |||
| 13480 | } else if (Mask[V2Index] != (int)Mask.size() || EltVT == MVT::i8 || | |||
| 13481 | EltVT == MVT::i16) { | |||
| 13482 | // Either not inserting from the low element of the input or the input | |||
| 13483 | // element size is too small to use VZEXT_MOVL to clear the high bits. | |||
| 13484 | return SDValue(); | |||
| 13485 | } | |||
| 13486 | ||||
| 13487 | if (!IsV1Zeroable) { | |||
| 13488 | // If V1 can't be treated as a zero vector we have fewer options to lower | |||
| 13489 | // this. We can't support integer vectors or non-zero targets cheaply, and | |||
| 13490 | // the V1 elements can't be permuted in any way. | |||
| 13491 | assert(VT == ExtVT && "Cannot change extended type when non-zeroable!")((void)0); | |||
| 13492 | if (!VT.isFloatingPoint() || V2Index != 0) | |||
| 13493 | return SDValue(); | |||
| 13494 | SmallVector<int, 8> V1Mask(Mask.begin(), Mask.end()); | |||
| 13495 | V1Mask[V2Index] = -1; | |||
| 13496 | if (!isNoopShuffleMask(V1Mask)) | |||
| 13497 | return SDValue(); | |||
| 13498 | if (!VT.is128BitVector()) | |||
| 13499 | return SDValue(); | |||
| 13500 | ||||
| 13501 | // Otherwise, use MOVSD or MOVSS. | |||
| 13502 | assert((EltVT == MVT::f32 || EltVT == MVT::f64) &&((void)0) | |||
| 13503 | "Only two types of floating point element types to handle!")((void)0); | |||
| 13504 | return DAG.getNode(EltVT == MVT::f32 ? X86ISD::MOVSS : X86ISD::MOVSD, DL, | |||
| 13505 | ExtVT, V1, V2); | |||
| 13506 | } | |||
| 13507 | ||||
| 13508 | // This lowering only works for the low element with floating point vectors. | |||
| 13509 | if (VT.isFloatingPoint() && V2Index != 0) | |||
| 13510 | return SDValue(); | |||
| 13511 | ||||
| 13512 | V2 = DAG.getNode(X86ISD::VZEXT_MOVL, DL, ExtVT, V2); | |||
| 13513 | if (ExtVT != VT) | |||
| 13514 | V2 = DAG.getBitcast(VT, V2); | |||
| 13515 | ||||
| 13516 | if (V2Index != 0) { | |||
| 13517 | // If we have 4 or fewer lanes we can cheaply shuffle the element into | |||
| 13518 | // the desired position. Otherwise it is more efficient to do a vector | |||
| 13519 | // shift left. We know that we can do a vector shift left because all | |||
| 13520 | // the inputs are zero. | |||
| 13521 | if (VT.isFloatingPoint() || VT.getVectorNumElements() <= 4) { | |||
| 13522 | SmallVector<int, 4> V2Shuffle(Mask.size(), 1); | |||
| 13523 | V2Shuffle[V2Index] = 0; | |||
| 13524 | V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Shuffle); | |||
| 13525 | } else { | |||
| 13526 | V2 = DAG.getBitcast(MVT::v16i8, V2); | |||
| 13527 | V2 = DAG.getNode(X86ISD::VSHLDQ, DL, MVT::v16i8, V2, | |||
| 13528 | DAG.getTargetConstant( | |||
| 13529 | V2Index * EltVT.getSizeInBits() / 8, DL, MVT::i8)); | |||
| 13530 | V2 = DAG.getBitcast(VT, V2); | |||
| 13531 | } | |||
| 13532 | } | |||
| 13533 | return V2; | |||
| 13534 | } | |||
| 13535 | ||||
| 13536 | /// Try to lower broadcast of a single - truncated - integer element, | |||
| 13537 | /// coming from a scalar_to_vector/build_vector node \p V0 with larger elements. | |||
| 13538 | /// | |||
| 13539 | /// This assumes we have AVX2. | |||
| 13540 | static SDValue lowerShuffleAsTruncBroadcast(const SDLoc &DL, MVT VT, SDValue V0, | |||
| 13541 | int BroadcastIdx, | |||
| 13542 | const X86Subtarget &Subtarget, | |||
| 13543 | SelectionDAG &DAG) { | |||
| 13544 | assert(Subtarget.hasAVX2() &&((void)0) | |||
| 13545 | "We can only lower integer broadcasts with AVX2!")((void)0); | |||
| 13546 | ||||
| 13547 | MVT EltVT = VT.getVectorElementType(); | |||
| 13548 | MVT V0VT = V0.getSimpleValueType(); | |||
| 13549 | ||||
| 13550 | assert(VT.isInteger() && "Unexpected non-integer trunc broadcast!")((void)0); | |||
| 13551 | assert(V0VT.isVector() && "Unexpected non-vector vector-sized value!")((void)0); | |||
| 13552 | ||||
| 13553 | MVT V0EltVT = V0VT.getVectorElementType(); | |||
| 13554 | if (!V0EltVT.isInteger()) | |||
| 13555 | return SDValue(); | |||
| 13556 | ||||
| 13557 | const unsigned EltSize = EltVT.getSizeInBits(); | |||
| 13558 | const unsigned V0EltSize = V0EltVT.getSizeInBits(); | |||
| 13559 | ||||
| 13560 | // This is only a truncation if the original element type is larger. | |||
| 13561 | if (V0EltSize <= EltSize) | |||
| 13562 | return SDValue(); | |||
| 13563 | ||||
| 13564 | assert(((V0EltSize % EltSize) == 0) &&((void)0) | |||
| 13565 | "Scalar type sizes must all be powers of 2 on x86!")((void)0); | |||
| 13566 | ||||
| 13567 | const unsigned V0Opc = V0.getOpcode(); | |||
| 13568 | const unsigned Scale = V0EltSize / EltSize; | |||
| 13569 | const unsigned V0BroadcastIdx = BroadcastIdx / Scale; | |||
| 13570 | ||||
| 13571 | if ((V0Opc != ISD::SCALAR_TO_VECTOR || V0BroadcastIdx != 0) && | |||
| 13572 | V0Opc != ISD::BUILD_VECTOR) | |||
| 13573 | return SDValue(); | |||
| 13574 | ||||
| 13575 | SDValue Scalar = V0.getOperand(V0BroadcastIdx); | |||
| 13576 | ||||
| 13577 | // If we're extracting non-least-significant bits, shift so we can truncate. | |||
| 13578 | // Hopefully, we can fold away the trunc/srl/load into the broadcast. | |||
| 13579 | // Even if we can't (and !isShuffleFoldableLoad(Scalar)), prefer | |||
| 13580 | // vpbroadcast+vmovd+shr to vpshufb(m)+vmovd. | |||
| 13581 | if (const int OffsetIdx = BroadcastIdx % Scale) | |||
| 13582 | Scalar = DAG.getNode(ISD::SRL, DL, Scalar.getValueType(), Scalar, | |||
| 13583 | DAG.getConstant(OffsetIdx * EltSize, DL, MVT::i8)); | |||
| 13584 | ||||
| 13585 | return DAG.getNode(X86ISD::VBROADCAST, DL, VT, | |||
| 13586 | DAG.getNode(ISD::TRUNCATE, DL, EltVT, Scalar)); | |||
| 13587 | } | |||
| 13588 | ||||
| 13589 | /// Test whether this can be lowered with a single SHUFPS instruction. | |||
| 13590 | /// | |||
| 13591 | /// This is used to disable more specialized lowerings when the shufps lowering | |||
| 13592 | /// will happen to be efficient. | |||
| 13593 | static bool isSingleSHUFPSMask(ArrayRef<int> Mask) { | |||
| 13594 | // This routine only handles 128-bit shufps. | |||
| 13595 | assert(Mask.size() == 4 && "Unsupported mask size!")((void)0); | |||
| 13596 | assert(Mask[0] >= -1 && Mask[0] < 8 && "Out of bound mask element!")((void)0); | |||
| 13597 | assert(Mask[1] >= -1 && Mask[1] < 8 && "Out of bound mask element!")((void)0); | |||
| 13598 | assert(Mask[2] >= -1 && Mask[2] < 8 && "Out of bound mask element!")((void)0); | |||
| 13599 | assert(Mask[3] >= -1 && Mask[3] < 8 && "Out of bound mask element!")((void)0); | |||
| 13600 | ||||
| 13601 | // To lower with a single SHUFPS we need to have the low half and high half | |||
| 13602 | // each requiring a single input. | |||
| 13603 | if (Mask[0] >= 0 && Mask[1] >= 0 && (Mask[0] < 4) != (Mask[1] < 4)) | |||
| 13604 | return false; | |||
| 13605 | if (Mask[2] >= 0 && Mask[3] >= 0 && (Mask[2] < 4) != (Mask[3] < 4)) | |||
| 13606 | return false; | |||
| 13607 | ||||
| 13608 | return true; | |||
| 13609 | } | |||
| 13610 | ||||
| 13611 | /// If we are extracting two 128-bit halves of a vector and shuffling the | |||
| 13612 | /// result, match that to a 256-bit AVX2 vperm* instruction to avoid a | |||
| 13613 | /// multi-shuffle lowering. | |||
| 13614 | static SDValue lowerShuffleOfExtractsAsVperm(const SDLoc &DL, SDValue N0, | |||
| 13615 | SDValue N1, ArrayRef<int> Mask, | |||
| 13616 | SelectionDAG &DAG) { | |||
| 13617 | MVT VT = N0.getSimpleValueType(); | |||
| 13618 | assert((VT.is128BitVector() &&((void)0) | |||
| 13619 | (VT.getScalarSizeInBits() == 32 || VT.getScalarSizeInBits() == 64)) &&((void)0) | |||
| 13620 | "VPERM* family of shuffles requires 32-bit or 64-bit elements")((void)0); | |||
| 13621 | ||||
| 13622 | // Check that both sources are extracts of the same source vector. | |||
| 13623 | if (!N0.hasOneUse() || !N1.hasOneUse() || | |||
| 13624 | N0.getOpcode() != ISD::EXTRACT_SUBVECTOR || | |||
| 13625 | N1.getOpcode() != ISD::EXTRACT_SUBVECTOR || | |||
| 13626 | N0.getOperand(0) != N1.getOperand(0)) | |||
| 13627 | return SDValue(); | |||
| 13628 | ||||
| 13629 | SDValue WideVec = N0.getOperand(0); | |||
| 13630 | MVT WideVT = WideVec.getSimpleValueType(); | |||
| 13631 | if (!WideVT.is256BitVector()) | |||
| 13632 | return SDValue(); | |||
| 13633 | ||||
| 13634 | // Match extracts of each half of the wide source vector. Commute the shuffle | |||
| 13635 | // if the extract of the low half is N1. | |||
| 13636 | unsigned NumElts = VT.getVectorNumElements(); | |||
| 13637 | SmallVector<int, 4> NewMask(Mask.begin(), Mask.end()); | |||
| 13638 | const APInt &ExtIndex0 = N0.getConstantOperandAPInt(1); | |||
| 13639 | const APInt &ExtIndex1 = N1.getConstantOperandAPInt(1); | |||
| 13640 | if (ExtIndex1 == 0 && ExtIndex0 == NumElts) | |||
| 13641 | ShuffleVectorSDNode::commuteMask(NewMask); | |||
| 13642 | else if (ExtIndex0 != 0 || ExtIndex1 != NumElts) | |||
| 13643 | return SDValue(); | |||
| 13644 | ||||
| 13645 | // Final bailout: if the mask is simple, we are better off using an extract | |||
| 13646 | // and a simple narrow shuffle. Prefer extract+unpack(h/l)ps to vpermps | |||
| 13647 | // because that avoids a constant load from memory. | |||
| 13648 | if (NumElts == 4 && | |||
| 13649 | (isSingleSHUFPSMask(NewMask) || is128BitUnpackShuffleMask(NewMask))) | |||
| 13650 | return SDValue(); | |||
| 13651 | ||||
| 13652 | // Extend the shuffle mask with undef elements. | |||
| 13653 | NewMask.append(NumElts, -1); | |||
| 13654 | ||||
| 13655 | // shuf (extract X, 0), (extract X, 4), M --> extract (shuf X, undef, M'), 0 | |||
| 13656 | SDValue Shuf = DAG.getVectorShuffle(WideVT, DL, WideVec, DAG.getUNDEF(WideVT), | |||
| 13657 | NewMask); | |||
| 13658 | // This is free: ymm -> xmm. | |||
| 13659 | return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, Shuf, | |||
| 13660 | DAG.getIntPtrConstant(0, DL)); | |||
| 13661 | } | |||
| 13662 | ||||
| 13663 | /// Try to lower broadcast of a single element. | |||
| 13664 | /// | |||
| 13665 | /// For convenience, this code also bundles all of the subtarget feature set | |||
| 13666 | /// filtering. While a little annoying to re-dispatch on type here, there isn't | |||
| 13667 | /// a convenient way to factor it out. | |||
| 13668 | static SDValue lowerShuffleAsBroadcast(const SDLoc &DL, MVT VT, SDValue V1, | |||
| 13669 | SDValue V2, ArrayRef<int> Mask, | |||
| 13670 | const X86Subtarget &Subtarget, | |||
| 13671 | SelectionDAG &DAG) { | |||
| 13672 | if (!((Subtarget.hasSSE3() && VT == MVT::v2f64) || | |||
| 13673 | (Subtarget.hasAVX() && VT.isFloatingPoint()) || | |||
| 13674 | (Subtarget.hasAVX2() && VT.isInteger()))) | |||
| 13675 | return SDValue(); | |||
| 13676 | ||||
| 13677 | // With MOVDDUP (v2f64) we can broadcast from a register or a load, otherwise | |||
| 13678 | // we can only broadcast from a register with AVX2. | |||
| 13679 | unsigned NumEltBits = VT.getScalarSizeInBits(); | |||
| 13680 | unsigned Opcode = (VT == MVT::v2f64 && !Subtarget.hasAVX2()) | |||
| 13681 | ? X86ISD::MOVDDUP | |||
| 13682 | : X86ISD::VBROADCAST; | |||
| 13683 | bool BroadcastFromReg = (Opcode == X86ISD::MOVDDUP) || Subtarget.hasAVX2(); | |||
| 13684 | ||||
| 13685 | // Check that the mask is a broadcast. | |||
| 13686 | int BroadcastIdx = getSplatIndex(Mask); | |||
| 13687 | if (BroadcastIdx < 0) | |||
| 13688 | return SDValue(); | |||
| 13689 | assert(BroadcastIdx < (int)Mask.size() && "We only expect to be called with "((void)0) | |||
| 13690 | "a sorted mask where the broadcast "((void)0) | |||
| 13691 | "comes from V1.")((void)0); | |||
| 13692 | ||||
| 13693 | // Go up the chain of (vector) values to find a scalar load that we can | |||
| 13694 | // combine with the broadcast. | |||
| 13695 | // TODO: Combine this logic with findEltLoadSrc() used by | |||
| 13696 | // EltsFromConsecutiveLoads(). | |||
| 13697 | int BitOffset = BroadcastIdx * NumEltBits; | |||
| 13698 | SDValue V = V1; | |||
| 13699 | for (;;) { | |||
| 13700 | switch (V.getOpcode()) { | |||
| 13701 | case ISD::BITCAST: { | |||
| 13702 | V = V.getOperand(0); | |||
| 13703 | continue; | |||
| 13704 | } | |||
| 13705 | case ISD::CONCAT_VECTORS: { | |||
| 13706 | int OpBitWidth = V.getOperand(0).getValueSizeInBits(); | |||
| 13707 | int OpIdx = BitOffset / OpBitWidth; | |||
| 13708 | V = V.getOperand(OpIdx); | |||
| 13709 | BitOffset %= OpBitWidth; | |||
| 13710 | continue; | |||
| 13711 | } | |||
| 13712 | case ISD::EXTRACT_SUBVECTOR: { | |||
| 13713 | // The extraction index adds to the existing offset. | |||
| 13714 | unsigned EltBitWidth = V.getScalarValueSizeInBits(); | |||
| 13715 | unsigned Idx = V.getConstantOperandVal(1); | |||
| 13716 | unsigned BeginOffset = Idx * EltBitWidth; | |||
| 13717 | BitOffset += BeginOffset; | |||
| 13718 | V = V.getOperand(0); | |||
| 13719 | continue; | |||
| 13720 | } | |||
| 13721 | case ISD::INSERT_SUBVECTOR: { | |||
| 13722 | SDValue VOuter = V.getOperand(0), VInner = V.getOperand(1); | |||
| 13723 | int EltBitWidth = VOuter.getScalarValueSizeInBits(); | |||
| 13724 | int Idx = (int)V.getConstantOperandVal(2); | |||
| 13725 | int NumSubElts = (int)VInner.getSimpleValueType().getVectorNumElements(); | |||
| 13726 | int BeginOffset = Idx * EltBitWidth; | |||
| 13727 | int EndOffset = BeginOffset + NumSubElts * EltBitWidth; | |||
| 13728 | if (BeginOffset <= BitOffset && BitOffset < EndOffset) { | |||
| 13729 | BitOffset -= BeginOffset; | |||
| 13730 | V = VInner; | |||
| 13731 | } else { | |||
| 13732 | V = VOuter; | |||
| 13733 | } | |||
| 13734 | continue; | |||
| 13735 | } | |||
| 13736 | } | |||
| 13737 | break; | |||
| 13738 | } | |||
| 13739 | assert((BitOffset % NumEltBits) == 0 && "Illegal bit-offset")((void)0); | |||
| 13740 | BroadcastIdx = BitOffset / NumEltBits; | |||
| 13741 | ||||
| 13742 | // Do we need to bitcast the source to retrieve the original broadcast index? | |||
| 13743 | bool BitCastSrc = V.getScalarValueSizeInBits() != NumEltBits; | |||
| 13744 | ||||
| 13745 | // Check if this is a broadcast of a scalar. We special case lowering | |||
| 13746 | // for scalars so that we can more effectively fold with loads. | |||
| 13747 | // If the original value has a larger element type than the shuffle, the | |||
| 13748 | // broadcast element is in essence truncated. Make that explicit to ease | |||
| 13749 | // folding. | |||
| 13750 | if (BitCastSrc && VT.isInteger()) | |||
| 13751 | if (SDValue TruncBroadcast = lowerShuffleAsTruncBroadcast( | |||
| 13752 | DL, VT, V, BroadcastIdx, Subtarget, DAG)) | |||
| 13753 | return TruncBroadcast; | |||
| 13754 | ||||
| 13755 | // Also check the simpler case, where we can directly reuse the scalar. | |||
| 13756 | if (!BitCastSrc && | |||
| 13757 | ((V.getOpcode() == ISD::BUILD_VECTOR && V.hasOneUse()) || | |||
| 13758 | (V.getOpcode() == ISD::SCALAR_TO_VECTOR && BroadcastIdx == 0))) { | |||
| 13759 | V = V.getOperand(BroadcastIdx); | |||
| 13760 | ||||
| 13761 | // If we can't broadcast from a register, check that the input is a load. | |||
| 13762 | if (!BroadcastFromReg && !isShuffleFoldableLoad(V)) | |||
| 13763 | return SDValue(); | |||
| 13764 | } else if (ISD::isNormalLoad(V.getNode()) && | |||
| 13765 | cast<LoadSDNode>(V)->isSimple()) { | |||
| 13766 | // We do not check for one-use of the vector load because a broadcast load | |||
| 13767 | // is expected to be a win for code size, register pressure, and possibly | |||
| 13768 | // uops even if the original vector load is not eliminated. | |||
| 13769 | ||||
| 13770 | // Reduce the vector load and shuffle to a broadcasted scalar load. | |||
| 13771 | LoadSDNode *Ld = cast<LoadSDNode>(V); | |||
| 13772 | SDValue BaseAddr = Ld->getOperand(1); | |||
| 13773 | MVT SVT = VT.getScalarType(); | |||
| 13774 | unsigned Offset = BroadcastIdx * SVT.getStoreSize(); | |||
| 13775 | assert((int)(Offset * 8) == BitOffset && "Unexpected bit-offset")((void)0); | |||
| 13776 | SDValue NewAddr = | |||
| 13777 | DAG.getMemBasePlusOffset(BaseAddr, TypeSize::Fixed(Offset), DL); | |||
| 13778 | ||||
| 13779 | // Directly form VBROADCAST_LOAD if we're using VBROADCAST opcode rather | |||
| 13780 | // than MOVDDUP. | |||
| 13781 | // FIXME: Should we add VBROADCAST_LOAD isel patterns for pre-AVX? | |||
| 13782 | if (Opcode == X86ISD::VBROADCAST) { | |||
| 13783 | SDVTList Tys = DAG.getVTList(VT, MVT::Other); | |||
| 13784 | SDValue Ops[] = {Ld->getChain(), NewAddr}; | |||
| 13785 | V = DAG.getMemIntrinsicNode( | |||
| 13786 | X86ISD::VBROADCAST_LOAD, DL, Tys, Ops, SVT, | |||
| 13787 | DAG.getMachineFunction().getMachineMemOperand( | |||
| 13788 | Ld->getMemOperand(), Offset, SVT.getStoreSize())); | |||
| 13789 | DAG.makeEquivalentMemoryOrdering(Ld, V); | |||
| 13790 | return DAG.getBitcast(VT, V); | |||
| 13791 | } | |||
| 13792 | assert(SVT == MVT::f64 && "Unexpected VT!")((void)0); | |||
| 13793 | V = DAG.getLoad(SVT, DL, Ld->getChain(), NewAddr, | |||
| 13794 | DAG.getMachineFunction().getMachineMemOperand( | |||
| 13795 | Ld->getMemOperand(), Offset, SVT.getStoreSize())); | |||
| 13796 | DAG.makeEquivalentMemoryOrdering(Ld, V); | |||
| 13797 | } else if (!BroadcastFromReg) { | |||
| 13798 | // We can't broadcast from a vector register. | |||
| 13799 | return SDValue(); | |||
| 13800 | } else if (BitOffset != 0) { | |||
| 13801 | // We can only broadcast from the zero-element of a vector register, | |||
| 13802 | // but it can be advantageous to broadcast from the zero-element of a | |||
| 13803 | // subvector. | |||
| 13804 | if (!VT.is256BitVector() && !VT.is512BitVector()) | |||
| 13805 | return SDValue(); | |||
| 13806 | ||||
| 13807 | // VPERMQ/VPERMPD can perform the cross-lane shuffle directly. | |||
| 13808 | if (VT == MVT::v4f64 || VT == MVT::v4i64) | |||
| 13809 | return SDValue(); | |||
| 13810 | ||||
| 13811 | // Only broadcast the zero-element of a 128-bit subvector. | |||
| 13812 | if ((BitOffset % 128) != 0) | |||
| 13813 | return SDValue(); | |||
| 13814 | ||||
| 13815 | assert((BitOffset % V.getScalarValueSizeInBits()) == 0 &&((void)0) | |||
| 13816 | "Unexpected bit-offset")((void)0); | |||
| 13817 | assert((V.getValueSizeInBits() == 256 || V.getValueSizeInBits() == 512) &&((void)0) | |||
| 13818 | "Unexpected vector size")((void)0); | |||
| 13819 | unsigned ExtractIdx = BitOffset / V.getScalarValueSizeInBits(); | |||
| 13820 | V = extract128BitVector(V, ExtractIdx, DAG, DL); | |||
| 13821 | } | |||
| 13822 | ||||
| 13823 | // On AVX we can use VBROADCAST directly for scalar sources. | |||
| 13824 | if (Opcode == X86ISD::MOVDDUP && !V.getValueType().isVector()) { | |||
| 13825 | V = DAG.getBitcast(MVT::f64, V); | |||
| 13826 | if (Subtarget.hasAVX()) { | |||
| 13827 | V = DAG.getNode(X86ISD::VBROADCAST, DL, MVT::v2f64, V); | |||
| 13828 | return DAG.getBitcast(VT, V); | |||
| 13829 | } | |||
| 13830 | V = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v2f64, V); | |||
| 13831 | } | |||
| 13832 | ||||
| 13833 | // If this is a scalar, do the broadcast on this type and bitcast. | |||
| 13834 | if (!V.getValueType().isVector()) { | |||
| 13835 | assert(V.getScalarValueSizeInBits() == NumEltBits &&((void)0) | |||
| 13836 | "Unexpected scalar size")((void)0); | |||
| 13837 | MVT BroadcastVT = MVT::getVectorVT(V.getSimpleValueType(), | |||
| 13838 | VT.getVectorNumElements()); | |||
| 13839 | return DAG.getBitcast(VT, DAG.getNode(Opcode, DL, BroadcastVT, V)); | |||
| 13840 | } | |||
| 13841 | ||||
| 13842 | // We only support broadcasting from 128-bit vectors to minimize the | |||
| 13843 | // number of patterns we need to deal with in isel. So extract down to | |||
| 13844 | // 128-bits, removing as many bitcasts as possible. | |||
| 13845 | if (V.getValueSizeInBits() > 128) | |||
| 13846 | V = extract128BitVector(peekThroughBitcasts(V), 0, DAG, DL); | |||
| 13847 | ||||
| 13848 | // Otherwise cast V to a vector with the same element type as VT, but | |||
| 13849 | // possibly narrower than VT. Then perform the broadcast. | |||
| 13850 | unsigned NumSrcElts = V.getValueSizeInBits() / NumEltBits; | |||
| 13851 | MVT CastVT = MVT::getVectorVT(VT.getVectorElementType(), NumSrcElts); | |||
| 13852 | return DAG.getNode(Opcode, DL, VT, DAG.getBitcast(CastVT, V)); | |||
| 13853 | } | |||
| 13854 | ||||
| 13855 | // Check for whether we can use INSERTPS to perform the shuffle. We only use | |||
| 13856 | // INSERTPS when the V1 elements are already in the correct locations | |||
| 13857 | // because otherwise we can just always use two SHUFPS instructions which | |||
| 13858 | // are much smaller to encode than a SHUFPS and an INSERTPS. We can also | |||
| 13859 | // perform INSERTPS if a single V1 element is out of place and all V2 | |||
| 13860 | // elements are zeroable. | |||
| 13861 | static bool matchShuffleAsInsertPS(SDValue &V1, SDValue &V2, | |||
| 13862 | unsigned &InsertPSMask, | |||
| 13863 | const APInt &Zeroable, | |||
| 13864 | ArrayRef<int> Mask, SelectionDAG &DAG) { | |||
| 13865 | assert(V1.getSimpleValueType().is128BitVector() && "Bad operand type!")((void)0); | |||
| 13866 | assert(V2.getSimpleValueType().is128BitVector() && "Bad operand type!")((void)0); | |||
| 13867 | assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!")((void)0); | |||
| 13868 | ||||
| 13869 | // Attempt to match INSERTPS with one element from VA or VB being | |||
| 13870 | // inserted into VA (or undef). If successful, V1, V2 and InsertPSMask | |||
| 13871 | // are updated. | |||
| 13872 | auto matchAsInsertPS = [&](SDValue VA, SDValue VB, | |||
| 13873 | ArrayRef<int> CandidateMask) { | |||
| 13874 | unsigned ZMask = 0; | |||
| 13875 | int VADstIndex = -1; | |||
| 13876 | int VBDstIndex = -1; | |||
| 13877 | bool VAUsedInPlace = false; | |||
| 13878 | ||||
| 13879 | for (int i = 0; i < 4; ++i) { | |||
| 13880 | // Synthesize a zero mask from the zeroable elements (includes undefs). | |||
| 13881 | if (Zeroable[i]) { | |||
| 13882 | ZMask |= 1 << i; | |||
| 13883 | continue; | |||
| 13884 | } | |||
| 13885 | ||||
| 13886 | // Flag if we use any VA inputs in place. | |||
| 13887 | if (i == CandidateMask[i]) { | |||
| 13888 | VAUsedInPlace = true; | |||
| 13889 | continue; | |||
| 13890 | } | |||
| 13891 | ||||
| 13892 | // We can only insert a single non-zeroable element. | |||
| 13893 | if (VADstIndex >= 0 || VBDstIndex >= 0) | |||
| 13894 | return false; | |||
| 13895 | ||||
| 13896 | if (CandidateMask[i] < 4) { | |||
| 13897 | // VA input out of place for insertion. | |||
| 13898 | VADstIndex = i; | |||
| 13899 | } else { | |||
| 13900 | // VB input for insertion. | |||
| 13901 | VBDstIndex = i; | |||
| 13902 | } | |||
| 13903 | } | |||
| 13904 | ||||
| 13905 | // Don't bother if we have no (non-zeroable) element for insertion. | |||
| 13906 | if (VADstIndex < 0 && VBDstIndex < 0) | |||
| 13907 | return false; | |||
| 13908 | ||||
| 13909 | // Determine element insertion src/dst indices. The src index is from the | |||
| 13910 | // start of the inserted vector, not the start of the concatenated vector. | |||
| 13911 | unsigned VBSrcIndex = 0; | |||
| 13912 | if (VADstIndex >= 0) { | |||
| 13913 | // If we have a VA input out of place, we use VA as the V2 element | |||
| 13914 | // insertion and don't use the original V2 at all. | |||
| 13915 | VBSrcIndex = CandidateMask[VADstIndex]; | |||
| 13916 | VBDstIndex = VADstIndex; | |||
| 13917 | VB = VA; | |||
| 13918 | } else { | |||
| 13919 | VBSrcIndex = CandidateMask[VBDstIndex] - 4; | |||
| 13920 | } | |||
| 13921 | ||||
| 13922 | // If no V1 inputs are used in place, then the result is created only from | |||
| 13923 | // the zero mask and the V2 insertion - so remove V1 dependency. | |||
| 13924 | if (!VAUsedInPlace) | |||
| 13925 | VA = DAG.getUNDEF(MVT::v4f32); | |||
| 13926 | ||||
| 13927 | // Update V1, V2 and InsertPSMask accordingly. | |||
| 13928 | V1 = VA; | |||
| 13929 | V2 = VB; | |||
| 13930 | ||||
| 13931 | // Insert the V2 element into the desired position. | |||
| 13932 | InsertPSMask = VBSrcIndex << 6 | VBDstIndex << 4 | ZMask; | |||
| 13933 | assert((InsertPSMask & ~0xFFu) == 0 && "Invalid mask!")((void)0); | |||
| 13934 | return true; | |||
| 13935 | }; | |||
| 13936 | ||||
| 13937 | if (matchAsInsertPS(V1, V2, Mask)) | |||
| 13938 | return true; | |||
| 13939 | ||||
| 13940 | // Commute and try again. | |||
| 13941 | SmallVector<int, 4> CommutedMask(Mask.begin(), Mask.end()); | |||
| 13942 | ShuffleVectorSDNode::commuteMask(CommutedMask); | |||
| 13943 | if (matchAsInsertPS(V2, V1, CommutedMask)) | |||
| 13944 | return true; | |||
| 13945 | ||||
| 13946 | return false; | |||
| 13947 | } | |||
| 13948 | ||||
| 13949 | static SDValue lowerShuffleAsInsertPS(const SDLoc &DL, SDValue V1, SDValue V2, | |||
| 13950 | ArrayRef<int> Mask, const APInt &Zeroable, | |||
| 13951 | SelectionDAG &DAG) { | |||
| 13952 | assert(V1.getSimpleValueType() == MVT::v4f32 && "Bad operand type!")((void)0); | |||
| 13953 | assert(V2.getSimpleValueType() == MVT::v4f32 && "Bad operand type!")((void)0); | |||
| 13954 | ||||
| 13955 | // Attempt to match the insertps pattern. | |||
| 13956 | unsigned InsertPSMask = 0; | |||
| 13957 | if (!matchShuffleAsInsertPS(V1, V2, InsertPSMask, Zeroable, Mask, DAG)) | |||
| 13958 | return SDValue(); | |||
| 13959 | ||||
| 13960 | // Insert the V2 element into the desired position. | |||
| 13961 | return DAG.getNode(X86ISD::INSERTPS, DL, MVT::v4f32, V1, V2, | |||
| 13962 | DAG.getTargetConstant(InsertPSMask, DL, MVT::i8)); | |||
| 13963 | } | |||
| 13964 | ||||
| 13965 | /// Try to lower a shuffle as a permute of the inputs followed by an | |||
| 13966 | /// UNPCK instruction. | |||
| 13967 | /// | |||
| 13968 | /// This specifically targets cases where we end up with alternating between | |||
| 13969 | /// the two inputs, and so can permute them into something that feeds a single | |||
| 13970 | /// UNPCK instruction. Note that this routine only targets integer vectors | |||
| 13971 | /// because for floating point vectors we have a generalized SHUFPS lowering | |||
| 13972 | /// strategy that handles everything that doesn't *exactly* match an unpack, | |||
| 13973 | /// making this clever lowering unnecessary. | |||
| 13974 | static SDValue lowerShuffleAsPermuteAndUnpack( | |||
| 13975 | const SDLoc &DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask, | |||
| 13976 | const X86Subtarget &Subtarget, SelectionDAG &DAG) { | |||
| 13977 | assert(!VT.isFloatingPoint() &&((void)0) | |||
| 13978 | "This routine only supports integer vectors.")((void)0); | |||
| 13979 | assert(VT.is128BitVector() &&((void)0) | |||
| 13980 | "This routine only works on 128-bit vectors.")((void)0); | |||
| 13981 | assert(!V2.isUndef() &&((void)0) | |||
| 13982 | "This routine should only be used when blending two inputs.")((void)0); | |||
| 13983 | assert(Mask.size() >= 2 && "Single element masks are invalid.")((void)0); | |||
| 13984 | ||||
| 13985 | int Size = Mask.size(); | |||
| 13986 | ||||
| 13987 | int NumLoInputs = | |||
| 13988 | count_if(Mask, [Size](int M) { return M >= 0 && M % Size < Size / 2; }); | |||
| 13989 | int NumHiInputs = | |||
| 13990 | count_if(Mask, [Size](int M) { return M % Size >= Size / 2; }); | |||
| 13991 | ||||
| 13992 | bool UnpackLo = NumLoInputs >= NumHiInputs; | |||
| 13993 | ||||
| 13994 | auto TryUnpack = [&](int ScalarSize, int Scale) { | |||
| 13995 | SmallVector<int, 16> V1Mask((unsigned)Size, -1); | |||
| 13996 | SmallVector<int, 16> V2Mask((unsigned)Size, -1); | |||
| 13997 | ||||
| 13998 | for (int i = 0; i < Size; ++i) { | |||
| 13999 | if (Mask[i] < 0) | |||
| 14000 | continue; | |||
| 14001 | ||||
| 14002 | // Each element of the unpack contains Scale elements from this mask. | |||
| 14003 | int UnpackIdx = i / Scale; | |||
| 14004 | ||||
| 14005 | // We only handle the case where V1 feeds the first slots of the unpack. | |||
| 14006 | // We rely on canonicalization to ensure this is the case. | |||
| 14007 | if ((UnpackIdx % 2 == 0) != (Mask[i] < Size)) | |||
| 14008 | return SDValue(); | |||
| 14009 | ||||
| 14010 | // Setup the mask for this input. The indexing is tricky as we have to | |||
| 14011 | // handle the unpack stride. | |||
| 14012 | SmallVectorImpl<int> &VMask = (UnpackIdx % 2 == 0) ? V1Mask : V2Mask; | |||
| 14013 | VMask[(UnpackIdx / 2) * Scale + i % Scale + (UnpackLo ? 0 : Size / 2)] = | |||
| 14014 | Mask[i] % Size; | |||
| 14015 | } | |||
| 14016 | ||||
| 14017 | // If we will have to shuffle both inputs to use the unpack, check whether | |||
| 14018 | // we can just unpack first and shuffle the result. If so, skip this unpack. | |||
| 14019 | if ((NumLoInputs == 0 || NumHiInputs == 0) && !isNoopShuffleMask(V1Mask) && | |||
| 14020 | !isNoopShuffleMask(V2Mask)) | |||
| 14021 | return SDValue(); | |||
| 14022 | ||||
| 14023 | // Shuffle the inputs into place. | |||
| 14024 | V1 = DAG.getVectorShuffle(VT, DL, V1, DAG.getUNDEF(VT), V1Mask); | |||
| 14025 | V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Mask); | |||
| 14026 | ||||
| 14027 | // Cast the inputs to the type we will use to unpack them. | |||
| 14028 | MVT UnpackVT = MVT::getVectorVT(MVT::getIntegerVT(ScalarSize), Size / Scale); | |||
| 14029 | V1 = DAG.getBitcast(UnpackVT, V1); | |||
| 14030 | V2 = DAG.getBitcast(UnpackVT, V2); | |||
| 14031 | ||||
| 14032 | // Unpack the inputs and cast the result back to the desired type. | |||
| 14033 | return DAG.getBitcast( | |||
| 14034 | VT, DAG.getNode(UnpackLo ? X86ISD::UNPCKL : X86ISD::UNPCKH, DL, | |||
| 14035 | UnpackVT, V1, V2)); | |||
| 14036 | }; | |||
| 14037 | ||||
| 14038 | // We try each unpack from the largest to the smallest to try and find one | |||
| 14039 | // that fits this mask. | |||
| 14040 | int OrigScalarSize = VT.getScalarSizeInBits(); | |||
| 14041 | for (int ScalarSize = 64; ScalarSize >= OrigScalarSize; ScalarSize /= 2) | |||
| 14042 | if (SDValue Unpack = TryUnpack(ScalarSize, ScalarSize / OrigScalarSize)) | |||
| 14043 | return Unpack; | |||
| 14044 | ||||
| 14045 | // If we're shuffling with a zero vector then we're better off not doing | |||
| 14046 | // VECTOR_SHUFFLE(UNPCK()) as we lose track of those zero elements. | |||
| 14047 | if (ISD::isBuildVectorAllZeros(V1.getNode()) || | |||
| 14048 | ISD::isBuildVectorAllZeros(V2.getNode())) | |||
| 14049 | return SDValue(); | |||
| 14050 | ||||
| 14051 | // If none of the unpack-rooted lowerings worked (or were profitable) try an | |||
| 14052 | // initial unpack. | |||
| 14053 | if (NumLoInputs == 0 || NumHiInputs == 0) { | |||
| 14054 | assert((NumLoInputs > 0 || NumHiInputs > 0) &&((void)0) | |||
| 14055 | "We have to have *some* inputs!")((void)0); | |||
| 14056 | int HalfOffset = NumLoInputs == 0 ? Size / 2 : 0; | |||
| 14057 | ||||
| 14058 | // FIXME: We could consider the total complexity of the permute of each | |||
| 14059 | // possible unpacking. Or at the least we should consider how many | |||
| 14060 | // half-crossings are created. | |||
| 14061 | // FIXME: We could consider commuting the unpacks. | |||
| 14062 | ||||
| 14063 | SmallVector<int, 32> PermMask((unsigned)Size, -1); | |||
| 14064 | for (int i = 0; i < Size; ++i) { | |||
| 14065 | if (Mask[i] < 0) | |||
| 14066 | continue; | |||
| 14067 | ||||
| 14068 | assert(Mask[i] % Size >= HalfOffset && "Found input from wrong half!")((void)0); | |||
| 14069 | ||||
| 14070 | PermMask[i] = | |||
| 14071 | 2 * ((Mask[i] % Size) - HalfOffset) + (Mask[i] < Size ? 0 : 1); | |||
| 14072 | } | |||
| 14073 | return DAG.getVectorShuffle( | |||
| 14074 | VT, DL, DAG.getNode(NumLoInputs == 0 ? X86ISD::UNPCKH : X86ISD::UNPCKL, | |||
| 14075 | DL, VT, V1, V2), | |||
| 14076 | DAG.getUNDEF(VT), PermMask); | |||
| 14077 | } | |||
| 14078 | ||||
| 14079 | return SDValue(); | |||
| 14080 | } | |||
| 14081 | ||||
| 14082 | /// Handle lowering of 2-lane 64-bit floating point shuffles. | |||
| 14083 | /// | |||
| 14084 | /// This is the basis function for the 2-lane 64-bit shuffles as we have full | |||
| 14085 | /// support for floating point shuffles but not integer shuffles. These | |||
| 14086 | /// instructions will incur a domain crossing penalty on some chips though so | |||
| 14087 | /// it is better to avoid lowering through this for integer vectors where | |||
| 14088 | /// possible. | |||
| 14089 | static SDValue lowerV2F64Shuffle(const SDLoc &DL, ArrayRef<int> Mask, | |||
| 14090 | const APInt &Zeroable, SDValue V1, SDValue V2, | |||
| 14091 | const X86Subtarget &Subtarget, | |||
| 14092 | SelectionDAG &DAG) { | |||
| 14093 | assert(V1.getSimpleValueType() == MVT::v2f64 && "Bad operand type!")((void)0); | |||
| 14094 | assert(V2.getSimpleValueType() == MVT::v2f64 && "Bad operand type!")((void)0); | |||
| 14095 | assert(Mask.size() == 2 && "Unexpected mask size for v2 shuffle!")((void)0); | |||
| 14096 | ||||
| 14097 | if (V2.isUndef()) { | |||
| 14098 | // Check for being able to broadcast a single element. | |||
| 14099 | if (SDValue Broadcast = lowerShuffleAsBroadcast(DL, MVT::v2f64, V1, V2, | |||
| 14100 | Mask, Subtarget, DAG)) | |||
| 14101 | return Broadcast; | |||
| 14102 | ||||
| 14103 | // Straight shuffle of a single input vector. Simulate this by using the | |||
| 14104 | // single input as both of the "inputs" to this instruction.. | |||
| 14105 | unsigned SHUFPDMask = (Mask[0] == 1) | ((Mask[1] == 1) << 1); | |||
| 14106 | ||||
| 14107 | if (Subtarget.hasAVX()) { | |||
| 14108 | // If we have AVX, we can use VPERMILPS which will allow folding a load | |||
| 14109 | // into the shuffle. | |||
| 14110 | return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v2f64, V1, | |||
| 14111 | DAG.getTargetConstant(SHUFPDMask, DL, MVT::i8)); | |||
| 14112 | } | |||
| 14113 | ||||
| 14114 | return DAG.getNode( | |||
| 14115 | X86ISD::SHUFP, DL, MVT::v2f64, | |||
| 14116 | Mask[0] == SM_SentinelUndef ? DAG.getUNDEF(MVT::v2f64) : V1, | |||
| 14117 | Mask[1] == SM_SentinelUndef ? DAG.getUNDEF(MVT::v2f64) : V1, | |||
| 14118 | DAG.getTargetConstant(SHUFPDMask, DL, MVT::i8)); | |||
| 14119 | } | |||
| 14120 | assert(Mask[0] >= 0 && "No undef lanes in multi-input v2 shuffles!")((void)0); | |||
| 14121 | assert(Mask[1] >= 0 && "No undef lanes in multi-input v2 shuffles!")((void)0); | |||
| 14122 | assert(Mask[0] < 2 && "We sort V1 to be the first input.")((void)0); | |||
| 14123 | assert(Mask[1] >= 2 && "We sort V2 to be the second input.")((void)0); | |||
| 14124 | ||||
| 14125 | if (Subtarget.hasAVX2()) | |||
| 14126 | if (SDValue Extract = lowerShuffleOfExtractsAsVperm(DL, V1, V2, Mask, DAG)) | |||
| 14127 | return Extract; | |||
| 14128 | ||||
| 14129 | // When loading a scalar and then shuffling it into a vector we can often do | |||
| 14130 | // the insertion cheaply. | |||
| 14131 | if (SDValue Insertion = lowerShuffleAsElementInsertion( | |||
| 14132 | DL, MVT::v2f64, V1, V2, Mask, Zeroable, Subtarget, DAG)) | |||
| 14133 | return Insertion; | |||
| 14134 | // Try inverting the insertion since for v2 masks it is easy to do and we | |||
| 14135 | // can't reliably sort the mask one way or the other. | |||
| 14136 | int InverseMask[2] = {Mask[0] < 0 ? -1 : (Mask[0] ^ 2), | |||
| 14137 | Mask[1] < 0 ? -1 : (Mask[1] ^ 2)}; | |||
| 14138 | if (SDValue Insertion = lowerShuffleAsElementInsertion( | |||
| 14139 | DL, MVT::v2f64, V2, V1, InverseMask, Zeroable, Subtarget, DAG)) | |||
| 14140 | return Insertion; | |||
| 14141 | ||||
| 14142 | // Try to use one of the special instruction patterns to handle two common | |||
| 14143 | // blend patterns if a zero-blend above didn't work. | |||
| 14144 | if (isShuffleEquivalent(Mask, {0, 3}, V1, V2) || | |||
| 14145 | isShuffleEquivalent(Mask, {1, 3}, V1, V2)) | |||
| 14146 | if (SDValue V1S = getScalarValueForVectorElement(V1, Mask[0], DAG)) | |||
| 14147 | // We can either use a special instruction to load over the low double or | |||
| 14148 | // to move just the low double. | |||
| 14149 | return DAG.getNode( | |||
| 14150 | X86ISD::MOVSD, DL, MVT::v2f64, V2, | |||
| 14151 | DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v2f64, V1S)); | |||
| 14152 | ||||
| 14153 | if (Subtarget.hasSSE41()) | |||
| 14154 | if (SDValue Blend = lowerShuffleAsBlend(DL, MVT::v2f64, V1, V2, Mask, | |||
| 14155 | Zeroable, Subtarget, DAG)) | |||
| 14156 | return Blend; | |||
| 14157 | ||||
| 14158 | // Use dedicated unpack instructions for masks that match their pattern. | |||
| 14159 | if (SDValue V = lowerShuffleWithUNPCK(DL, MVT::v2f64, Mask, V1, V2, DAG)) | |||
| 14160 | return V; | |||
| 14161 | ||||
| 14162 | unsigned SHUFPDMask = (Mask[0] == 1) | (((Mask[1] - 2) == 1) << 1); | |||
| 14163 | return DAG.getNode(X86ISD::SHUFP, DL, MVT::v2f64, V1, V2, | |||
| 14164 | DAG.getTargetConstant(SHUFPDMask, DL, MVT::i8)); | |||
| 14165 | } | |||
| 14166 | ||||
| 14167 | /// Handle lowering of 2-lane 64-bit integer shuffles. | |||
| 14168 | /// | |||
| 14169 | /// Tries to lower a 2-lane 64-bit shuffle using shuffle operations provided by | |||
| 14170 | /// the integer unit to minimize domain crossing penalties. However, for blends | |||
| 14171 | /// it falls back to the floating point shuffle operation with appropriate bit | |||
| 14172 | /// casting. | |||
| 14173 | static SDValue lowerV2I64Shuffle(const SDLoc &DL, ArrayRef<int> Mask, | |||
| 14174 | const APInt &Zeroable, SDValue V1, SDValue V2, | |||
| 14175 | const X86Subtarget &Subtarget, | |||
| 14176 | SelectionDAG &DAG) { | |||
| 14177 | assert(V1.getSimpleValueType() == MVT::v2i64 && "Bad operand type!")((void)0); | |||
| 14178 | assert(V2.getSimpleValueType() == MVT::v2i64 && "Bad operand type!")((void)0); | |||
| 14179 | assert(Mask.size() == 2 && "Unexpected mask size for v2 shuffle!")((void)0); | |||
| 14180 | ||||
| 14181 | if (V2.isUndef()) { | |||
| 14182 | // Check for being able to broadcast a single element. | |||
| 14183 | if (SDValue Broadcast = lowerShuffleAsBroadcast(DL, MVT::v2i64, V1, V2, | |||
| 14184 | Mask, Subtarget, DAG)) | |||
| 14185 | return Broadcast; | |||
| 14186 | ||||
| 14187 | // Straight shuffle of a single input vector. For everything from SSE2 | |||
| 14188 | // onward this has a single fast instruction with no scary immediates. | |||
| 14189 | // We have to map the mask as it is actually a v4i32 shuffle instruction. | |||
| 14190 | V1 = DAG.getBitcast(MVT::v4i32, V1); | |||
| 14191 | int WidenedMask[4] = {Mask[0] < 0 ? -1 : (Mask[0] * 2), | |||
| 14192 | Mask[0] < 0 ? -1 : ((Mask[0] * 2) + 1), | |||
| 14193 | Mask[1] < 0 ? -1 : (Mask[1] * 2), | |||
| 14194 | Mask[1] < 0 ? -1 : ((Mask[1] * 2) + 1)}; | |||
| 14195 | return DAG.getBitcast( | |||
| 14196 | MVT::v2i64, | |||
| 14197 | DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, V1, | |||
| 14198 | getV4X86ShuffleImm8ForMask(WidenedMask, DL, DAG))); | |||
| 14199 | } | |||
| 14200 | assert(Mask[0] != -1 && "No undef lanes in multi-input v2 shuffles!")((void)0); | |||
| 14201 | assert(Mask[1] != -1 && "No undef lanes in multi-input v2 shuffles!")((void)0); | |||
| 14202 | assert(Mask[0] < 2 && "We sort V1 to be the first input.")((void)0); | |||
| 14203 | assert(Mask[1] >= 2 && "We sort V2 to be the second input.")((void)0); | |||
| 14204 | ||||
| 14205 | if (Subtarget.hasAVX2()) | |||
| 14206 | if (SDValue Extract = lowerShuffleOfExtractsAsVperm(DL, V1, V2, Mask, DAG)) | |||
| 14207 | return Extract; | |||
| 14208 | ||||
| 14209 | // Try to use shift instructions. | |||
| 14210 | if (SDValue Shift = lowerShuffleAsShift(DL, MVT::v2i64, V1, V2, Mask, | |||
| 14211 | Zeroable, Subtarget, DAG)) | |||
| 14212 | return Shift; | |||
| 14213 | ||||
| 14214 | // When loading a scalar and then shuffling it into a vector we can often do | |||
| 14215 | // the insertion cheaply. | |||
| 14216 | if (SDValue Insertion = lowerShuffleAsElementInsertion( | |||
| 14217 | DL, MVT::v2i64, V1, V2, Mask, Zeroable, Subtarget, DAG)) | |||
| 14218 | return Insertion; | |||
| 14219 | // Try inverting the insertion since for v2 masks it is easy to do and we | |||
| 14220 | // can't reliably sort the mask one way or the other. | |||
| 14221 | int InverseMask[2] = {Mask[0] ^ 2, Mask[1] ^ 2}; | |||
| 14222 | if (SDValue Insertion = lowerShuffleAsElementInsertion( | |||
| 14223 | DL, MVT::v2i64, V2, V1, InverseMask, Zeroable, Subtarget, DAG)) | |||
| 14224 | return Insertion; | |||
| 14225 | ||||
| 14226 | // We have different paths for blend lowering, but they all must use the | |||
| 14227 | // *exact* same predicate. | |||
| 14228 | bool IsBlendSupported = Subtarget.hasSSE41(); | |||
| 14229 | if (IsBlendSupported) | |||
| 14230 | if (SDValue Blend = lowerShuffleAsBlend(DL, MVT::v2i64, V1, V2, Mask, | |||
| 14231 | Zeroable, Subtarget, DAG)) | |||
| 14232 | return Blend; | |||
| 14233 | ||||
| 14234 | // Use dedicated unpack instructions for masks that match their pattern. | |||
| 14235 | if (SDValue V = lowerShuffleWithUNPCK(DL, MVT::v2i64, Mask, V1, V2, DAG)) | |||
| 14236 | return V; | |||
| 14237 | ||||
| 14238 | // Try to use byte rotation instructions. | |||
| 14239 | // Its more profitable for pre-SSSE3 to use shuffles/unpacks. | |||
| 14240 | if (Subtarget.hasSSSE3()) { | |||
| 14241 | if (Subtarget.hasVLX()) | |||
| 14242 | if (SDValue Rotate = lowerShuffleAsVALIGN(DL, MVT::v2i64, V1, V2, Mask, | |||
| 14243 | Subtarget, DAG)) | |||
| 14244 | return Rotate; | |||
| 14245 | ||||
| 14246 | if (SDValue Rotate = lowerShuffleAsByteRotate(DL, MVT::v2i64, V1, V2, Mask, | |||
| 14247 | Subtarget, DAG)) | |||
| 14248 | return Rotate; | |||
| 14249 | } | |||
| 14250 | ||||
| 14251 | // If we have direct support for blends, we should lower by decomposing into | |||
| 14252 | // a permute. That will be faster than the domain cross. | |||
| 14253 | if (IsBlendSupported) | |||
| 14254 | return lowerShuffleAsDecomposedShuffleMerge(DL, MVT::v2i64, V1, V2, Mask, | |||
| 14255 | Subtarget, DAG); | |||
| 14256 | ||||
| 14257 | // We implement this with SHUFPD which is pretty lame because it will likely | |||
| 14258 | // incur 2 cycles of stall for integer vectors on Nehalem and older chips. | |||
| 14259 | // However, all the alternatives are still more cycles and newer chips don't | |||
| 14260 | // have this problem. It would be really nice if x86 had better shuffles here. | |||
| 14261 | V1 = DAG.getBitcast(MVT::v2f64, V1); | |||
| 14262 | V2 = DAG.getBitcast(MVT::v2f64, V2); | |||
| 14263 | return DAG.getBitcast(MVT::v2i64, | |||
| 14264 | DAG.getVectorShuffle(MVT::v2f64, DL, V1, V2, Mask)); | |||
| 14265 | } | |||
| 14266 | ||||
| 14267 | /// Lower a vector shuffle using the SHUFPS instruction. | |||
| 14268 | /// | |||
| 14269 | /// This is a helper routine dedicated to lowering vector shuffles using SHUFPS. | |||
| 14270 | /// It makes no assumptions about whether this is the *best* lowering, it simply | |||
| 14271 | /// uses it. | |||
| 14272 | static SDValue lowerShuffleWithSHUFPS(const SDLoc &DL, MVT VT, | |||
| 14273 | ArrayRef<int> Mask, SDValue V1, | |||
| 14274 | SDValue V2, SelectionDAG &DAG) { | |||
| 14275 | SDValue LowV = V1, HighV = V2; | |||
| 14276 | SmallVector<int, 4> NewMask(Mask.begin(), Mask.end()); | |||
| 14277 | int NumV2Elements = count_if(Mask, [](int M) { return M >= 4; }); | |||
| 14278 | ||||
| 14279 | if (NumV2Elements == 1) { | |||
| 14280 | int V2Index = find_if(Mask, [](int M) { return M >= 4; }) - Mask.begin(); | |||
| 14281 | ||||
| 14282 | // Compute the index adjacent to V2Index and in the same half by toggling | |||
| 14283 | // the low bit. | |||
| 14284 | int V2AdjIndex = V2Index ^ 1; | |||
| 14285 | ||||
| 14286 | if (Mask[V2AdjIndex] < 0) { | |||
| 14287 | // Handles all the cases where we have a single V2 element and an undef. | |||
| 14288 | // This will only ever happen in the high lanes because we commute the | |||
| 14289 | // vector otherwise. | |||
| 14290 | if (V2Index < 2) | |||
| 14291 | std::swap(LowV, HighV); | |||
| 14292 | NewMask[V2Index] -= 4; | |||
| 14293 | } else { | |||
| 14294 | // Handle the case where the V2 element ends up adjacent to a V1 element. | |||
| 14295 | // To make this work, blend them together as the first step. | |||
| 14296 | int V1Index = V2AdjIndex; | |||
| 14297 | int BlendMask[4] = {Mask[V2Index] - 4, 0, Mask[V1Index], 0}; | |||
| 14298 | V2 = DAG.getNode(X86ISD::SHUFP, DL, VT, V2, V1, | |||
| 14299 | getV4X86ShuffleImm8ForMask(BlendMask, DL, DAG)); | |||
| 14300 | ||||
| 14301 | // Now proceed to reconstruct the final blend as we have the necessary | |||
| 14302 | // high or low half formed. | |||
| 14303 | if (V2Index < 2) { | |||
| 14304 | LowV = V2; | |||
| 14305 | HighV = V1; | |||
| 14306 | } else { | |||
| 14307 | HighV = V2; | |||
| 14308 | } | |||
| 14309 | NewMask[V1Index] = 2; // We put the V1 element in V2[2]. | |||
| 14310 | NewMask[V2Index] = 0; // We shifted the V2 element into V2[0]. | |||
| 14311 | } | |||
| 14312 | } else if (NumV2Elements == 2) { | |||
| 14313 | if (Mask[0] < 4 && Mask[1] < 4) { | |||
| 14314 | // Handle the easy case where we have V1 in the low lanes and V2 in the | |||
| 14315 | // high lanes. | |||
| 14316 | NewMask[2] -= 4; | |||
| 14317 | NewMask[3] -= 4; | |||
| 14318 | } else if (Mask[2] < 4 && Mask[3] < 4) { | |||
| 14319 | // We also handle the reversed case because this utility may get called | |||
| 14320 | // when we detect a SHUFPS pattern but can't easily commute the shuffle to | |||
| 14321 | // arrange things in the right direction. | |||
| 14322 | NewMask[0] -= 4; | |||
| 14323 | NewMask[1] -= 4; | |||
| 14324 | HighV = V1; | |||
| 14325 | LowV = V2; | |||
| 14326 | } else { | |||
| 14327 | // We have a mixture of V1 and V2 in both low and high lanes. Rather than | |||
| 14328 | // trying to place elements directly, just blend them and set up the final | |||
| 14329 | // shuffle to place them. | |||
| 14330 | ||||
| 14331 | // The first two blend mask elements are for V1, the second two are for | |||
| 14332 | // V2. | |||
| 14333 | int BlendMask[4] = {Mask[0] < 4 ? Mask[0] : Mask[1], | |||
| 14334 | Mask[2] < 4 ? Mask[2] : Mask[3], | |||
| 14335 | (Mask[0] >= 4 ? Mask[0] : Mask[1]) - 4, | |||
| 14336 | (Mask[2] >= 4 ? Mask[2] : Mask[3]) - 4}; | |||
| 14337 | V1 = DAG.getNode(X86ISD::SHUFP, DL, VT, V1, V2, | |||
| 14338 | getV4X86ShuffleImm8ForMask(BlendMask, DL, DAG)); | |||
| 14339 | ||||
| 14340 | // Now we do a normal shuffle of V1 by giving V1 as both operands to | |||
| 14341 | // a blend. | |||
| 14342 | LowV = HighV = V1; | |||
| 14343 | NewMask[0] = Mask[0] < 4 ? 0 : 2; | |||
| 14344 | NewMask[1] = Mask[0] < 4 ? 2 : 0; | |||
| 14345 | NewMask[2] = Mask[2] < 4 ? 1 : 3; | |||
| 14346 | NewMask[3] = Mask[2] < 4 ? 3 : 1; | |||
| 14347 | } | |||
| 14348 | } else if (NumV2Elements == 3) { | |||
| 14349 | // Ideally canonicalizeShuffleMaskWithCommute should have caught this, but | |||
| 14350 | // we can get here due to other paths (e.g repeated mask matching) that we | |||
| 14351 | // don't want to do another round of lowerVECTOR_SHUFFLE. | |||
| 14352 | ShuffleVectorSDNode::commuteMask(NewMask); | |||
| 14353 | return lowerShuffleWithSHUFPS(DL, VT, NewMask, V2, V1, DAG); | |||
| 14354 | } | |||
| 14355 | return DAG.getNode(X86ISD::SHUFP, DL, VT, LowV, HighV, | |||
| 14356 | getV4X86ShuffleImm8ForMask(NewMask, DL, DAG)); | |||
| 14357 | } | |||
| 14358 | ||||
| 14359 | /// Lower 4-lane 32-bit floating point shuffles. | |||
| 14360 | /// | |||
| 14361 | /// Uses instructions exclusively from the floating point unit to minimize | |||
| 14362 | /// domain crossing penalties, as these are sufficient to implement all v4f32 | |||
| 14363 | /// shuffles. | |||
| 14364 | static SDValue lowerV4F32Shuffle(const SDLoc &DL, ArrayRef<int> Mask, | |||
| 14365 | const APInt &Zeroable, SDValue V1, SDValue V2, | |||
| 14366 | const X86Subtarget &Subtarget, | |||
| 14367 | SelectionDAG &DAG) { | |||
| 14368 | assert(V1.getSimpleValueType() == MVT::v4f32 && "Bad operand type!")((void)0); | |||
| 14369 | assert(V2.getSimpleValueType() == MVT::v4f32 && "Bad operand type!")((void)0); | |||
| 14370 | assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!")((void)0); | |||
| 14371 | ||||
| 14372 | int NumV2Elements = count_if(Mask, [](int M) { return M >= 4; }); | |||
| 14373 | ||||
| 14374 | if (NumV2Elements == 0) { | |||
| 14375 | // Check for being able to broadcast a single element. | |||
| 14376 | if (SDValue Broadcast = lowerShuffleAsBroadcast(DL, MVT::v4f32, V1, V2, | |||
| 14377 | Mask, Subtarget, DAG)) | |||
| 14378 | return Broadcast; | |||
| 14379 | ||||
| 14380 | // Use even/odd duplicate instructions for masks that match their pattern. | |||
| 14381 | if (Subtarget.hasSSE3()) { | |||
| 14382 | if (isShuffleEquivalent(Mask, {0, 0, 2, 2}, V1, V2)) | |||
| 14383 | return DAG.getNode(X86ISD::MOVSLDUP, DL, MVT::v4f32, V1); | |||
| 14384 | if (isShuffleEquivalent(Mask, {1, 1, 3, 3}, V1, V2)) | |||
| 14385 | return DAG.getNode(X86ISD::MOVSHDUP, DL, MVT::v4f32, V1); | |||
| 14386 | } | |||
| 14387 | ||||
| 14388 | if (Subtarget.hasAVX()) { | |||
| 14389 | // If we have AVX, we can use VPERMILPS which will allow folding a load | |||
| 14390 | // into the shuffle. | |||
| 14391 | return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v4f32, V1, | |||
| 14392 | getV4X86ShuffleImm8ForMask(Mask, DL, DAG)); | |||
| 14393 | } | |||
| 14394 | ||||
| 14395 | // Use MOVLHPS/MOVHLPS to simulate unary shuffles. These are only valid | |||
| 14396 | // in SSE1 because otherwise they are widened to v2f64 and never get here. | |||
| 14397 | if (!Subtarget.hasSSE2()) { | |||
| 14398 | if (isShuffleEquivalent(Mask, {0, 1, 0, 1}, V1, V2)) | |||
| 14399 | return DAG.getNode(X86ISD::MOVLHPS, DL, MVT::v4f32, V1, V1); | |||
| 14400 | if (isShuffleEquivalent(Mask, {2, 3, 2, 3}, V1, V2)) | |||
| 14401 | return DAG.getNode(X86ISD::MOVHLPS, DL, MVT::v4f32, V1, V1); | |||
| 14402 | } | |||
| 14403 | ||||
| 14404 | // Otherwise, use a straight shuffle of a single input vector. We pass the | |||
| 14405 | // input vector to both operands to simulate this with a SHUFPS. | |||
| 14406 | return DAG.getNode(X86ISD::SHUFP, DL, MVT::v4f32, V1, V1, | |||
| 14407 | getV4X86ShuffleImm8ForMask(Mask, DL, DAG)); | |||
| 14408 | } | |||
| 14409 | ||||
| 14410 | if (Subtarget.hasAVX2()) | |||
| 14411 | if (SDValue Extract = lowerShuffleOfExtractsAsVperm(DL, V1, V2, Mask, DAG)) | |||
| 14412 | return Extract; | |||
| 14413 | ||||
| 14414 | // There are special ways we can lower some single-element blends. However, we | |||
| 14415 | // have custom ways we can lower more complex single-element blends below that | |||
| 14416 | // we defer to if both this and BLENDPS fail to match, so restrict this to | |||
| 14417 | // when the V2 input is targeting element 0 of the mask -- that is the fast | |||
| 14418 | // case here. | |||
| 14419 | if (NumV2Elements == 1 && Mask[0] >= 4) | |||
| 14420 | if (SDValue V = lowerShuffleAsElementInsertion( | |||
| 14421 | DL, MVT::v4f32, V1, V2, Mask, Zeroable, Subtarget, DAG)) | |||
| 14422 | return V; | |||
| 14423 | ||||
| 14424 | if (Subtarget.hasSSE41()) { | |||
| 14425 | if (SDValue Blend = lowerShuffleAsBlend(DL, MVT::v4f32, V1, V2, Mask, | |||
| 14426 | Zeroable, Subtarget, DAG)) | |||
| 14427 | return Blend; | |||
| 14428 | ||||
| 14429 | // Use INSERTPS if we can complete the shuffle efficiently. | |||
| 14430 | if (SDValue V = lowerShuffleAsInsertPS(DL, V1, V2, Mask, Zeroable, DAG)) | |||
| 14431 | return V; | |||
| 14432 | ||||
| 14433 | if (!isSingleSHUFPSMask(Mask)) | |||
| 14434 | if (SDValue BlendPerm = lowerShuffleAsBlendAndPermute(DL, MVT::v4f32, V1, | |||
| 14435 | V2, Mask, DAG)) | |||
| 14436 | return BlendPerm; | |||
| 14437 | } | |||
| 14438 | ||||
| 14439 | // Use low/high mov instructions. These are only valid in SSE1 because | |||
| 14440 | // otherwise they are widened to v2f64 and never get here. | |||
| 14441 | if (!Subtarget.hasSSE2()) { | |||
| 14442 | if (isShuffleEquivalent(Mask, {0, 1, 4, 5}, V1, V2)) | |||
| 14443 | return DAG.getNode(X86ISD::MOVLHPS, DL, MVT::v4f32, V1, V2); | |||
| 14444 | if (isShuffleEquivalent(Mask, {2, 3, 6, 7}, V1, V2)) | |||
| 14445 | return DAG.getNode(X86ISD::MOVHLPS, DL, MVT::v4f32, V2, V1); | |||
| 14446 | } | |||
| 14447 | ||||
| 14448 | // Use dedicated unpack instructions for masks that match their pattern. | |||
| 14449 | if (SDValue V = lowerShuffleWithUNPCK(DL, MVT::v4f32, Mask, V1, V2, DAG)) | |||
| 14450 | return V; | |||
| 14451 | ||||
| 14452 | // Otherwise fall back to a SHUFPS lowering strategy. | |||
| 14453 | return lowerShuffleWithSHUFPS(DL, MVT::v4f32, Mask, V1, V2, DAG); | |||
| 14454 | } | |||
| 14455 | ||||
| 14456 | /// Lower 4-lane i32 vector shuffles. | |||
| 14457 | /// | |||
| 14458 | /// We try to handle these with integer-domain shuffles where we can, but for | |||
| 14459 | /// blends we use the floating point domain blend instructions. | |||
| 14460 | static SDValue lowerV4I32Shuffle(const SDLoc &DL, ArrayRef<int> Mask, | |||
| 14461 | const APInt &Zeroable, SDValue V1, SDValue V2, | |||
| 14462 | const X86Subtarget &Subtarget, | |||
| 14463 | SelectionDAG &DAG) { | |||
| 14464 | assert(V1.getSimpleValueType() == MVT::v4i32 && "Bad operand type!")((void)0); | |||
| 14465 | assert(V2.getSimpleValueType() == MVT::v4i32 && "Bad operand type!")((void)0); | |||
| 14466 | assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!")((void)0); | |||
| 14467 | ||||
| 14468 | // Whenever we can lower this as a zext, that instruction is strictly faster | |||
| 14469 | // than any alternative. It also allows us to fold memory operands into the | |||
| 14470 | // shuffle in many cases. | |||
| 14471 | if (SDValue ZExt = lowerShuffleAsZeroOrAnyExtend(DL, MVT::v4i32, V1, V2, Mask, | |||
| 14472 | Zeroable, Subtarget, DAG)) | |||
| 14473 | return ZExt; | |||
| 14474 | ||||
| 14475 | int NumV2Elements = count_if(Mask, [](int M) { return M >= 4; }); | |||
| 14476 | ||||
| 14477 | if (NumV2Elements == 0) { | |||
| 14478 | // Try to use broadcast unless the mask only has one non-undef element. | |||
| 14479 | if (count_if(Mask, [](int M) { return M >= 0 && M < 4; }) > 1) { | |||
| 14480 | if (SDValue Broadcast = lowerShuffleAsBroadcast(DL, MVT::v4i32, V1, V2, | |||
| 14481 | Mask, Subtarget, DAG)) | |||
| 14482 | return Broadcast; | |||
| 14483 | } | |||
| 14484 | ||||
| 14485 | // Straight shuffle of a single input vector. For everything from SSE2 | |||
| 14486 | // onward this has a single fast instruction with no scary immediates. | |||
| 14487 | // We coerce the shuffle pattern to be compatible with UNPCK instructions | |||
| 14488 | // but we aren't actually going to use the UNPCK instruction because doing | |||
| 14489 | // so prevents folding a load into this instruction or making a copy. | |||
| 14490 | const int UnpackLoMask[] = {0, 0, 1, 1}; | |||
| 14491 | const int UnpackHiMask[] = {2, 2, 3, 3}; | |||
| 14492 | if (isShuffleEquivalent(Mask, {0, 0, 1, 1}, V1, V2)) | |||
| 14493 | Mask = UnpackLoMask; | |||
| 14494 | else if (isShuffleEquivalent(Mask, {2, 2, 3, 3}, V1, V2)) | |||
| 14495 | Mask = UnpackHiMask; | |||
| 14496 | ||||
| 14497 | return DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, V1, | |||
| 14498 | getV4X86ShuffleImm8ForMask(Mask, DL, DAG)); | |||
| 14499 | } | |||
| 14500 | ||||
| 14501 | if (Subtarget.hasAVX2()) | |||
| 14502 | if (SDValue Extract = lowerShuffleOfExtractsAsVperm(DL, V1, V2, Mask, DAG)) | |||
| 14503 | return Extract; | |||
| 14504 | ||||
| 14505 | // Try to use shift instructions. | |||
| 14506 | if (SDValue Shift = lowerShuffleAsShift(DL, MVT::v4i32, V1, V2, Mask, | |||
| 14507 | Zeroable, Subtarget, DAG)) | |||
| 14508 | return Shift; | |||
| 14509 | ||||
| 14510 | // There are special ways we can lower some single-element blends. | |||
| 14511 | if (NumV2Elements == 1) | |||
| 14512 | if (SDValue V = lowerShuffleAsElementInsertion( | |||
| 14513 | DL, MVT::v4i32, V1, V2, Mask, Zeroable, Subtarget, DAG)) | |||
| 14514 | return V; | |||
| 14515 | ||||
| 14516 | // We have different paths for blend lowering, but they all must use the | |||
| 14517 | // *exact* same predicate. | |||
| 14518 | bool IsBlendSupported = Subtarget.hasSSE41(); | |||
| 14519 | if (IsBlendSupported) | |||
| 14520 | if (SDValue Blend = lowerShuffleAsBlend(DL, MVT::v4i32, V1, V2, Mask, | |||
| 14521 | Zeroable, Subtarget, DAG)) | |||
| 14522 | return Blend; | |||
| 14523 | ||||
| 14524 | if (SDValue Masked = lowerShuffleAsBitMask(DL, MVT::v4i32, V1, V2, Mask, | |||
| 14525 | Zeroable, Subtarget, DAG)) | |||
| 14526 | return Masked; | |||
| 14527 | ||||
| 14528 | // Use dedicated unpack instructions for masks that match their pattern. | |||
| 14529 | if (SDValue V = lowerShuffleWithUNPCK(DL, MVT::v4i32, Mask, V1, V2, DAG)) | |||
| 14530 | return V; | |||
| 14531 | ||||
| 14532 | // Try to use byte rotation instructions. | |||
| 14533 | // Its more profitable for pre-SSSE3 to use shuffles/unpacks. | |||
| 14534 | if (Subtarget.hasSSSE3()) { | |||
| 14535 | if (Subtarget.hasVLX()) | |||
| 14536 | if (SDValue Rotate = lowerShuffleAsVALIGN(DL, MVT::v4i32, V1, V2, Mask, | |||
| 14537 | Subtarget, DAG)) | |||
| 14538 | return Rotate; | |||
| 14539 | ||||
| 14540 | if (SDValue Rotate = lowerShuffleAsByteRotate(DL, MVT::v4i32, V1, V2, Mask, | |||
| 14541 | Subtarget, DAG)) | |||
| 14542 | return Rotate; | |||
| 14543 | } | |||
| 14544 | ||||
| 14545 | // Assume that a single SHUFPS is faster than an alternative sequence of | |||
| 14546 | // multiple instructions (even if the CPU has a domain penalty). | |||
| 14547 | // If some CPU is harmed by the domain switch, we can fix it in a later pass. | |||
| 14548 | if (!isSingleSHUFPSMask(Mask)) { | |||
| 14549 | // If we have direct support for blends, we should lower by decomposing into | |||
| 14550 | // a permute. That will be faster than the domain cross. | |||
| 14551 | if (IsBlendSupported) | |||
| 14552 | return lowerShuffleAsDecomposedShuffleMerge(DL, MVT::v4i32, V1, V2, Mask, | |||
| 14553 | Subtarget, DAG); | |||
| 14554 | ||||
| 14555 | // Try to lower by permuting the inputs into an unpack instruction. | |||
| 14556 | if (SDValue Unpack = lowerShuffleAsPermuteAndUnpack(DL, MVT::v4i32, V1, V2, | |||
| 14557 | Mask, Subtarget, DAG)) | |||
| 14558 | return Unpack; | |||
| 14559 | } | |||
| 14560 | ||||
| 14561 | // We implement this with SHUFPS because it can blend from two vectors. | |||
| 14562 | // Because we're going to eventually use SHUFPS, we use SHUFPS even to build | |||
| 14563 | // up the inputs, bypassing domain shift penalties that we would incur if we | |||
| 14564 | // directly used PSHUFD on Nehalem and older. For newer chips, this isn't | |||
| 14565 | // relevant. | |||
| 14566 | SDValue CastV1 = DAG.getBitcast(MVT::v4f32, V1); | |||
| 14567 | SDValue CastV2 = DAG.getBitcast(MVT::v4f32, V2); | |||
| 14568 | SDValue ShufPS = DAG.getVectorShuffle(MVT::v4f32, DL, CastV1, CastV2, Mask); | |||
| 14569 | return DAG.getBitcast(MVT::v4i32, ShufPS); | |||
| 14570 | } | |||
| 14571 | ||||
| 14572 | /// Lowering of single-input v8i16 shuffles is the cornerstone of SSE2 | |||
| 14573 | /// shuffle lowering, and the most complex part. | |||
| 14574 | /// | |||
| 14575 | /// The lowering strategy is to try to form pairs of input lanes which are | |||
| 14576 | /// targeted at the same half of the final vector, and then use a dword shuffle | |||
| 14577 | /// to place them onto the right half, and finally unpack the paired lanes into | |||
| 14578 | /// their final position. | |||
| 14579 | /// | |||
| 14580 | /// The exact breakdown of how to form these dword pairs and align them on the | |||
| 14581 | /// correct sides is really tricky. See the comments within the function for | |||
| 14582 | /// more of the details. | |||
| 14583 | /// | |||
| 14584 | /// This code also handles repeated 128-bit lanes of v8i16 shuffles, but each | |||
| 14585 | /// lane must shuffle the *exact* same way. In fact, you must pass a v8 Mask to | |||
| 14586 | /// this routine for it to work correctly. To shuffle a 256-bit or 512-bit i16 | |||
| 14587 | /// vector, form the analogous 128-bit 8-element Mask. | |||
| 14588 | static SDValue lowerV8I16GeneralSingleInputShuffle( | |||
| 14589 | const SDLoc &DL, MVT VT, SDValue V, MutableArrayRef<int> Mask, | |||
| 14590 | const X86Subtarget &Subtarget, SelectionDAG &DAG) { | |||
| 14591 | assert(VT.getVectorElementType() == MVT::i16 && "Bad input type!")((void)0); | |||
| 14592 | MVT PSHUFDVT = MVT::getVectorVT(MVT::i32, VT.getVectorNumElements() / 2); | |||
| 14593 | ||||
| 14594 | assert(Mask.size() == 8 && "Shuffle mask length doesn't match!")((void)0); | |||
| 14595 | MutableArrayRef<int> LoMask = Mask.slice(0, 4); | |||
| 14596 | MutableArrayRef<int> HiMask = Mask.slice(4, 4); | |||
| 14597 | ||||
| 14598 | // Attempt to directly match PSHUFLW or PSHUFHW. | |||
| 14599 | if (isUndefOrInRange(LoMask, 0, 4) && | |||
| 14600 | isSequentialOrUndefInRange(HiMask, 0, 4, 4)) { | |||
| 14601 | return DAG.getNode(X86ISD::PSHUFLW, DL, VT, V, | |||
| 14602 | getV4X86ShuffleImm8ForMask(LoMask, DL, DAG)); | |||
| 14603 | } | |||
| 14604 | if (isUndefOrInRange(HiMask, 4, 8) && | |||
| 14605 | isSequentialOrUndefInRange(LoMask, 0, 4, 0)) { | |||
| 14606 | for (int i = 0; i != 4; ++i) | |||
| 14607 | HiMask[i] = (HiMask[i] < 0 ? HiMask[i] : (HiMask[i] - 4)); | |||
| 14608 | return DAG.getNode(X86ISD::PSHUFHW, DL, VT, V, | |||
| 14609 | getV4X86ShuffleImm8ForMask(HiMask, DL, DAG)); | |||
| 14610 | } | |||
| 14611 | ||||
| 14612 | SmallVector<int, 4> LoInputs; | |||
| 14613 | copy_if(LoMask, std::back_inserter(LoInputs), [](int M) { return M >= 0; }); | |||
| 14614 | array_pod_sort(LoInputs.begin(), LoInputs.end()); | |||
| 14615 | LoInputs.erase(std::unique(LoInputs.begin(), LoInputs.end()), LoInputs.end()); | |||
| 14616 | SmallVector<int, 4> HiInputs; | |||
| 14617 | copy_if(HiMask, std::back_inserter(HiInputs), [](int M) { return M >= 0; }); | |||
| 14618 | array_pod_sort(HiInputs.begin(), HiInputs.end()); | |||
| 14619 | HiInputs.erase(std::unique(HiInputs.begin(), HiInputs.end()), HiInputs.end()); | |||
| 14620 | int NumLToL = llvm::lower_bound(LoInputs, 4) - LoInputs.begin(); | |||
| 14621 | int NumHToL = LoInputs.size() - NumLToL; | |||
| 14622 | int NumLToH = llvm::lower_bound(HiInputs, 4) - HiInputs.begin(); | |||
| 14623 | int NumHToH = HiInputs.size() - NumLToH; | |||
| 14624 | MutableArrayRef<int> LToLInputs(LoInputs.data(), NumLToL); | |||
| 14625 | MutableArrayRef<int> LToHInputs(HiInputs.data(), NumLToH); | |||
| 14626 | MutableArrayRef<int> HToLInputs(LoInputs.data() + NumLToL, NumHToL); | |||
| 14627 | MutableArrayRef<int> HToHInputs(HiInputs.data() + NumLToH, NumHToH); | |||
| 14628 | ||||
| 14629 | // If we are shuffling values from one half - check how many different DWORD | |||
| 14630 | // pairs we need to create. If only 1 or 2 then we can perform this as a | |||
| 14631 | // PSHUFLW/PSHUFHW + PSHUFD instead of the PSHUFD+PSHUFLW+PSHUFHW chain below. | |||
| 14632 | auto ShuffleDWordPairs = [&](ArrayRef<int> PSHUFHalfMask, | |||
| 14633 | ArrayRef<int> PSHUFDMask, unsigned ShufWOp) { | |||
| 14634 | V = DAG.getNode(ShufWOp, DL, VT, V, | |||
| 14635 | getV4X86ShuffleImm8ForMask(PSHUFHalfMask, DL, DAG)); | |||
| 14636 | V = DAG.getBitcast(PSHUFDVT, V); | |||
| 14637 | V = DAG.getNode(X86ISD::PSHUFD, DL, PSHUFDVT, V, | |||
| 14638 | getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)); | |||
| 14639 | return DAG.getBitcast(VT, V); | |||
| 14640 | }; | |||
| 14641 | ||||
| 14642 | if ((NumHToL + NumHToH) == 0 || (NumLToL + NumLToH) == 0) { | |||
| 14643 | int PSHUFDMask[4] = { -1, -1, -1, -1 }; | |||
| 14644 | SmallVector<std::pair<int, int>, 4> DWordPairs; | |||
| 14645 | int DOffset = ((NumHToL + NumHToH) == 0 ? 0 : 2); | |||
| 14646 | ||||
| 14647 | // Collect the different DWORD pairs. | |||
| 14648 | for (int DWord = 0; DWord != 4; ++DWord) { | |||
| 14649 | int M0 = Mask[2 * DWord + 0]; | |||
| 14650 | int M1 = Mask[2 * DWord + 1]; | |||
| 14651 | M0 = (M0 >= 0 ? M0 % 4 : M0); | |||
| 14652 | M1 = (M1 >= 0 ? M1 % 4 : M1); | |||
| 14653 | if (M0 < 0 && M1 < 0) | |||
| 14654 | continue; | |||
| 14655 | ||||
| 14656 | bool Match = false; | |||
| 14657 | for (int j = 0, e = DWordPairs.size(); j < e; ++j) { | |||
| 14658 | auto &DWordPair = DWordPairs[j]; | |||
| 14659 | if ((M0 < 0 || isUndefOrEqual(DWordPair.first, M0)) && | |||
| 14660 | (M1 < 0 || isUndefOrEqual(DWordPair.second, M1))) { | |||
| 14661 | DWordPair.first = (M0 >= 0 ? M0 : DWordPair.first); | |||
| 14662 | DWordPair.second = (M1 >= 0 ? M1 : DWordPair.second); | |||
| 14663 | PSHUFDMask[DWord] = DOffset + j; | |||
| 14664 | Match = true; | |||
| 14665 | break; | |||
| 14666 | } | |||
| 14667 | } | |||
| 14668 | if (!Match) { | |||
| 14669 | PSHUFDMask[DWord] = DOffset + DWordPairs.size(); | |||
| 14670 | DWordPairs.push_back(std::make_pair(M0, M1)); | |||
| 14671 | } | |||
| 14672 | } | |||
| 14673 | ||||
| 14674 | if (DWordPairs.size() <= 2) { | |||
| 14675 | DWordPairs.resize(2, std::make_pair(-1, -1)); | |||
| 14676 | int PSHUFHalfMask[4] = {DWordPairs[0].first, DWordPairs[0].second, | |||
| 14677 | DWordPairs[1].first, DWordPairs[1].second}; | |||
| 14678 | if ((NumHToL + NumHToH) == 0) | |||
| 14679 | return ShuffleDWordPairs(PSHUFHalfMask, PSHUFDMask, X86ISD::PSHUFLW); | |||
| 14680 | if ((NumLToL + NumLToH) == 0) | |||
| 14681 | return ShuffleDWordPairs(PSHUFHalfMask, PSHUFDMask, X86ISD::PSHUFHW); | |||
| 14682 | } | |||
| 14683 | } | |||
| 14684 | ||||
| 14685 | // Simplify the 1-into-3 and 3-into-1 cases with a single pshufd. For all | |||
| 14686 | // such inputs we can swap two of the dwords across the half mark and end up | |||
| 14687 | // with <=2 inputs to each half in each half. Once there, we can fall through | |||
| 14688 | // to the generic code below. For example: | |||
| 14689 | // | |||
| 14690 | // Input: [a, b, c, d, e, f, g, h] -PSHUFD[0,2,1,3]-> [a, b, e, f, c, d, g, h] | |||
| 14691 | // Mask: [0, 1, 2, 7, 4, 5, 6, 3] -----------------> [0, 1, 4, 7, 2, 3, 6, 5] | |||
| 14692 | // | |||
| 14693 | // However in some very rare cases we have a 1-into-3 or 3-into-1 on one half | |||
| 14694 | // and an existing 2-into-2 on the other half. In this case we may have to | |||
| 14695 | // pre-shuffle the 2-into-2 half to avoid turning it into a 3-into-1 or | |||
| 14696 | // 1-into-3 which could cause us to cycle endlessly fixing each side in turn. | |||
| 14697 | // Fortunately, we don't have to handle anything but a 2-into-2 pattern | |||
| 14698 | // because any other situation (including a 3-into-1 or 1-into-3 in the other | |||
| 14699 | // half than the one we target for fixing) will be fixed when we re-enter this | |||
| 14700 | // path. We will also combine away any sequence of PSHUFD instructions that | |||
| 14701 | // result into a single instruction. Here is an example of the tricky case: | |||
| 14702 | // | |||
| 14703 | // Input: [a, b, c, d, e, f, g, h] -PSHUFD[0,2,1,3]-> [a, b, e, f, c, d, g, h] | |||
| 14704 | // Mask: [3, 7, 1, 0, 2, 7, 3, 5] -THIS-IS-BAD!!!!-> [5, 7, 1, 0, 4, 7, 5, 3] | |||
| 14705 | // | |||
| 14706 | // This now has a 1-into-3 in the high half! Instead, we do two shuffles: | |||
| 14707 | // | |||
| 14708 | // Input: [a, b, c, d, e, f, g, h] PSHUFHW[0,2,1,3]-> [a, b, c, d, e, g, f, h] | |||
| 14709 | // Mask: [3, 7, 1, 0, 2, 7, 3, 5] -----------------> [3, 7, 1, 0, 2, 7, 3, 6] | |||
| 14710 | // | |||
| 14711 | // Input: [a, b, c, d, e, g, f, h] -PSHUFD[0,2,1,3]-> [a, b, e, g, c, d, f, h] | |||
| 14712 | // Mask: [3, 7, 1, 0, 2, 7, 3, 6] -----------------> [5, 7, 1, 0, 4, 7, 5, 6] | |||
| 14713 | // | |||
| 14714 | // The result is fine to be handled by the generic logic. | |||
| 14715 | auto balanceSides = [&](ArrayRef<int> AToAInputs, ArrayRef<int> BToAInputs, | |||
| 14716 | ArrayRef<int> BToBInputs, ArrayRef<int> AToBInputs, | |||
| 14717 | int AOffset, int BOffset) { | |||
| 14718 | assert((AToAInputs.size() == 3 || AToAInputs.size() == 1) &&((void)0) | |||
| 14719 | "Must call this with A having 3 or 1 inputs from the A half.")((void)0); | |||
| 14720 | assert((BToAInputs.size() == 1 || BToAInputs.size() == 3) &&((void)0) | |||
| 14721 | "Must call this with B having 1 or 3 inputs from the B half.")((void)0); | |||
| 14722 | assert(AToAInputs.size() + BToAInputs.size() == 4 &&((void)0) | |||
| 14723 | "Must call this with either 3:1 or 1:3 inputs (summing to 4).")((void)0); | |||
| 14724 | ||||
| 14725 | bool ThreeAInputs = AToAInputs.size() == 3; | |||
| 14726 | ||||
| 14727 | // Compute the index of dword with only one word among the three inputs in | |||
| 14728 | // a half by taking the sum of the half with three inputs and subtracting | |||
| 14729 | // the sum of the actual three inputs. The difference is the remaining | |||
| 14730 | // slot. | |||
| 14731 | int ADWord = 0, BDWord = 0; | |||
| 14732 | int &TripleDWord = ThreeAInputs ? ADWord : BDWord; | |||
| 14733 | int &OneInputDWord = ThreeAInputs ? BDWord : ADWord; | |||
| 14734 | int TripleInputOffset = ThreeAInputs ? AOffset : BOffset; | |||
| 14735 | ArrayRef<int> TripleInputs = ThreeAInputs ? AToAInputs : BToAInputs; | |||
| 14736 | int OneInput = ThreeAInputs ? BToAInputs[0] : AToAInputs[0]; | |||
| 14737 | int TripleInputSum = 0 + 1 + 2 + 3 + (4 * TripleInputOffset); | |||
| 14738 | int TripleNonInputIdx = | |||
| 14739 | TripleInputSum - std::accumulate(TripleInputs.begin(), TripleInputs.end(), 0); | |||
| 14740 | TripleDWord = TripleNonInputIdx / 2; | |||
| 14741 | ||||
| 14742 | // We use xor with one to compute the adjacent DWord to whichever one the | |||
| 14743 | // OneInput is in. | |||
| 14744 | OneInputDWord = (OneInput / 2) ^ 1; | |||
| 14745 | ||||
| 14746 | // Check for one tricky case: We're fixing a 3<-1 or a 1<-3 shuffle for AToA | |||
| 14747 | // and BToA inputs. If there is also such a problem with the BToB and AToB | |||
| 14748 | // inputs, we don't try to fix it necessarily -- we'll recurse and see it in | |||
| 14749 | // the next pass. However, if we have a 2<-2 in the BToB and AToB inputs, it | |||
| 14750 | // is essential that we don't *create* a 3<-1 as then we might oscillate. | |||
| 14751 | if (BToBInputs.size() == 2 && AToBInputs.size() == 2) { | |||
| 14752 | // Compute how many inputs will be flipped by swapping these DWords. We | |||
| 14753 | // need | |||
| 14754 | // to balance this to ensure we don't form a 3-1 shuffle in the other | |||
| 14755 | // half. | |||
| 14756 | int NumFlippedAToBInputs = | |||
| 14757 | std::count(AToBInputs.begin(), AToBInputs.end(), 2 * ADWord) + | |||
| 14758 | std::count(AToBInputs.begin(), AToBInputs.end(), 2 * ADWord + 1); | |||
| 14759 | int NumFlippedBToBInputs = | |||
| 14760 | std::count(BToBInputs.begin(), BToBInputs.end(), 2 * BDWord) + | |||
| 14761 | std::count(BToBInputs.begin(), BToBInputs.end(), 2 * BDWord + 1); | |||
| 14762 | if ((NumFlippedAToBInputs == 1 && | |||
| 14763 | (NumFlippedBToBInputs == 0 || NumFlippedBToBInputs == 2)) || | |||
| 14764 | (NumFlippedBToBInputs == 1 && | |||
| 14765 | (NumFlippedAToBInputs == 0 || NumFlippedAToBInputs == 2))) { | |||
| 14766 | // We choose whether to fix the A half or B half based on whether that | |||
| 14767 | // half has zero flipped inputs. At zero, we may not be able to fix it | |||
| 14768 | // with that half. We also bias towards fixing the B half because that | |||
| 14769 | // will more commonly be the high half, and we have to bias one way. | |||
| 14770 | auto FixFlippedInputs = [&V, &DL, &Mask, &DAG](int PinnedIdx, int DWord, | |||
| 14771 | ArrayRef<int> Inputs) { | |||
| 14772 | int FixIdx = PinnedIdx ^ 1; // The adjacent slot to the pinned slot. | |||
| 14773 | bool IsFixIdxInput = is_contained(Inputs, PinnedIdx ^ 1); | |||
| 14774 | // Determine whether the free index is in the flipped dword or the | |||
| 14775 | // unflipped dword based on where the pinned index is. We use this bit | |||
| 14776 | // in an xor to conditionally select the adjacent dword. | |||
| 14777 | int FixFreeIdx = 2 * (DWord ^ (PinnedIdx / 2 == DWord)); | |||
| 14778 | bool IsFixFreeIdxInput = is_contained(Inputs, FixFreeIdx); | |||
| 14779 | if (IsFixIdxInput == IsFixFreeIdxInput) | |||
| 14780 | FixFreeIdx += 1; | |||
| 14781 | IsFixFreeIdxInput = is_contained(Inputs, FixFreeIdx); | |||
| 14782 | assert(IsFixIdxInput != IsFixFreeIdxInput &&((void)0) | |||
| 14783 | "We need to be changing the number of flipped inputs!")((void)0); | |||
| 14784 | int PSHUFHalfMask[] = {0, 1, 2, 3}; | |||
| 14785 | std::swap(PSHUFHalfMask[FixFreeIdx % 4], PSHUFHalfMask[FixIdx % 4]); | |||
| 14786 | V = DAG.getNode( | |||
| 14787 | FixIdx < 4 ? X86ISD::PSHUFLW : X86ISD::PSHUFHW, DL, | |||
| 14788 | MVT::getVectorVT(MVT::i16, V.getValueSizeInBits() / 16), V, | |||
| 14789 | getV4X86ShuffleImm8ForMask(PSHUFHalfMask, DL, DAG)); | |||
| 14790 | ||||
| 14791 | for (int &M : Mask) | |||
| 14792 | if (M >= 0 && M == FixIdx) | |||
| 14793 | M = FixFreeIdx; | |||
| 14794 | else if (M >= 0 && M == FixFreeIdx) | |||
| 14795 | M = FixIdx; | |||
| 14796 | }; | |||
| 14797 | if (NumFlippedBToBInputs != 0) { | |||
| 14798 | int BPinnedIdx = | |||
| 14799 | BToAInputs.size() == 3 ? TripleNonInputIdx : OneInput; | |||
| 14800 | FixFlippedInputs(BPinnedIdx, BDWord, BToBInputs); | |||
| 14801 | } else { | |||
| 14802 | assert(NumFlippedAToBInputs != 0 && "Impossible given predicates!")((void)0); | |||
| 14803 | int APinnedIdx = ThreeAInputs ? TripleNonInputIdx : OneInput; | |||
| 14804 | FixFlippedInputs(APinnedIdx, ADWord, AToBInputs); | |||
| 14805 | } | |||
| 14806 | } | |||
| 14807 | } | |||
| 14808 | ||||
| 14809 | int PSHUFDMask[] = {0, 1, 2, 3}; | |||
| 14810 | PSHUFDMask[ADWord] = BDWord; | |||
| 14811 | PSHUFDMask[BDWord] = ADWord; | |||
| 14812 | V = DAG.getBitcast( | |||
| 14813 | VT, | |||
| 14814 | DAG.getNode(X86ISD::PSHUFD, DL, PSHUFDVT, DAG.getBitcast(PSHUFDVT, V), | |||
| 14815 | getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG))); | |||
| 14816 | ||||
| 14817 | // Adjust the mask to match the new locations of A and B. | |||
| 14818 | for (int &M : Mask) | |||
| 14819 | if (M >= 0 && M/2 == ADWord) | |||
| 14820 | M = 2 * BDWord + M % 2; | |||
| 14821 | else if (M >= 0 && M/2 == BDWord) | |||
| 14822 | M = 2 * ADWord + M % 2; | |||
| 14823 | ||||
| 14824 | // Recurse back into this routine to re-compute state now that this isn't | |||
| 14825 | // a 3 and 1 problem. | |||
| 14826 | return lowerV8I16GeneralSingleInputShuffle(DL, VT, V, Mask, Subtarget, DAG); | |||
| 14827 | }; | |||
| 14828 | if ((NumLToL == 3 && NumHToL == 1) || (NumLToL == 1 && NumHToL == 3)) | |||
| 14829 | return balanceSides(LToLInputs, HToLInputs, HToHInputs, LToHInputs, 0, 4); | |||
| 14830 | if ((NumHToH == 3 && NumLToH == 1) || (NumHToH == 1 && NumLToH == 3)) | |||
| 14831 | return balanceSides(HToHInputs, LToHInputs, LToLInputs, HToLInputs, 4, 0); | |||
| 14832 | ||||
| 14833 | // At this point there are at most two inputs to the low and high halves from | |||
| 14834 | // each half. That means the inputs can always be grouped into dwords and | |||
| 14835 | // those dwords can then be moved to the correct half with a dword shuffle. | |||
| 14836 | // We use at most one low and one high word shuffle to collect these paired | |||
| 14837 | // inputs into dwords, and finally a dword shuffle to place them. | |||
| 14838 | int PSHUFLMask[4] = {-1, -1, -1, -1}; | |||
| 14839 | int PSHUFHMask[4] = {-1, -1, -1, -1}; | |||
| 14840 | int PSHUFDMask[4] = {-1, -1, -1, -1}; | |||
| 14841 | ||||
| 14842 | // First fix the masks for all the inputs that are staying in their | |||
| 14843 | // original halves. This will then dictate the targets of the cross-half | |||
| 14844 | // shuffles. | |||
| 14845 | auto fixInPlaceInputs = | |||
| 14846 | [&PSHUFDMask](ArrayRef<int> InPlaceInputs, ArrayRef<int> IncomingInputs, | |||
| 14847 | MutableArrayRef<int> SourceHalfMask, | |||
| 14848 | MutableArrayRef<int> HalfMask, int HalfOffset) { | |||
| 14849 | if (InPlaceInputs.empty()) | |||
| 14850 | return; | |||
| 14851 | if (InPlaceInputs.size() == 1) { | |||
| 14852 | SourceHalfMask[InPlaceInputs[0] - HalfOffset] = | |||
| 14853 | InPlaceInputs[0] - HalfOffset; | |||
| 14854 | PSHUFDMask[InPlaceInputs[0] / 2] = InPlaceInputs[0] / 2; | |||
| 14855 | return; | |||
| 14856 | } | |||
| 14857 | if (IncomingInputs.empty()) { | |||
| 14858 | // Just fix all of the in place inputs. | |||
| 14859 | for (int Input : InPlaceInputs) { | |||
| 14860 | SourceHalfMask[Input - HalfOffset] = Input - HalfOffset; | |||
| 14861 | PSHUFDMask[Input / 2] = Input / 2; | |||
| 14862 | } | |||
| 14863 | return; | |||
| 14864 | } | |||
| 14865 | ||||
| 14866 | assert(InPlaceInputs.size() == 2 && "Cannot handle 3 or 4 inputs!")((void)0); | |||
| 14867 | SourceHalfMask[InPlaceInputs[0] - HalfOffset] = | |||
| 14868 | InPlaceInputs[0] - HalfOffset; | |||
| 14869 | // Put the second input next to the first so that they are packed into | |||
| 14870 | // a dword. We find the adjacent index by toggling the low bit. | |||
| 14871 | int AdjIndex = InPlaceInputs[0] ^ 1; | |||
| 14872 | SourceHalfMask[AdjIndex - HalfOffset] = InPlaceInputs[1] - HalfOffset; | |||
| 14873 | std::replace(HalfMask.begin(), HalfMask.end(), InPlaceInputs[1], AdjIndex); | |||
| 14874 | PSHUFDMask[AdjIndex / 2] = AdjIndex / 2; | |||
| 14875 | }; | |||
| 14876 | fixInPlaceInputs(LToLInputs, HToLInputs, PSHUFLMask, LoMask, 0); | |||
| 14877 | fixInPlaceInputs(HToHInputs, LToHInputs, PSHUFHMask, HiMask, 4); | |||
| 14878 | ||||
| 14879 | // Now gather the cross-half inputs and place them into a free dword of | |||
| 14880 | // their target half. | |||
| 14881 | // FIXME: This operation could almost certainly be simplified dramatically to | |||
| 14882 | // look more like the 3-1 fixing operation. | |||
| 14883 | auto moveInputsToRightHalf = [&PSHUFDMask]( | |||
| 14884 | MutableArrayRef<int> IncomingInputs, ArrayRef<int> ExistingInputs, | |||
| 14885 | MutableArrayRef<int> SourceHalfMask, MutableArrayRef<int> HalfMask, | |||
| 14886 | MutableArrayRef<int> FinalSourceHalfMask, int SourceOffset, | |||
| 14887 | int DestOffset) { | |||
| 14888 | auto isWordClobbered = [](ArrayRef<int> SourceHalfMask, int Word) { | |||
| 14889 | return SourceHalfMask[Word] >= 0 && SourceHalfMask[Word] != Word; | |||
| 14890 | }; | |||
| 14891 | auto isDWordClobbered = [&isWordClobbered](ArrayRef<int> SourceHalfMask, | |||
| 14892 | int Word) { | |||
| 14893 | int LowWord = Word & ~1; | |||
| 14894 | int HighWord = Word | 1; | |||
| 14895 | return isWordClobbered(SourceHalfMask, LowWord) || | |||
| 14896 | isWordClobbered(SourceHalfMask, HighWord); | |||
| 14897 | }; | |||
| 14898 | ||||
| 14899 | if (IncomingInputs.empty()) | |||
| 14900 | return; | |||
| 14901 | ||||
| 14902 | if (ExistingInputs.empty()) { | |||
| 14903 | // Map any dwords with inputs from them into the right half. | |||
| 14904 | for (int Input : IncomingInputs) { | |||
| 14905 | // If the source half mask maps over the inputs, turn those into | |||
| 14906 | // swaps and use the swapped lane. | |||
| 14907 | if (isWordClobbered(SourceHalfMask, Input - SourceOffset)) { | |||
| 14908 | if (SourceHalfMask[SourceHalfMask[Input - SourceOffset]] < 0) { | |||
| 14909 | SourceHalfMask[SourceHalfMask[Input - SourceOffset]] = | |||
| 14910 | Input - SourceOffset; | |||
| 14911 | // We have to swap the uses in our half mask in one sweep. | |||
| 14912 | for (int &M : HalfMask) | |||
| 14913 | if (M == SourceHalfMask[Input - SourceOffset] + SourceOffset) | |||
| 14914 | M = Input; | |||
| 14915 | else if (M == Input) | |||
| 14916 | M = SourceHalfMask[Input - SourceOffset] + SourceOffset; | |||
| 14917 | } else { | |||
| 14918 | assert(SourceHalfMask[SourceHalfMask[Input - SourceOffset]] ==((void)0) | |||
| 14919 | Input - SourceOffset &&((void)0) | |||
| 14920 | "Previous placement doesn't match!")((void)0); | |||
| 14921 | } | |||
| 14922 | // Note that this correctly re-maps both when we do a swap and when | |||
| 14923 | // we observe the other side of the swap above. We rely on that to | |||
| 14924 | // avoid swapping the members of the input list directly. | |||
| 14925 | Input = SourceHalfMask[Input - SourceOffset] + SourceOffset; | |||
| 14926 | } | |||
| 14927 | ||||
| 14928 | // Map the input's dword into the correct half. | |||
| 14929 | if (PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] < 0) | |||
| 14930 | PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] = Input / 2; | |||
| 14931 | else | |||
| 14932 | assert(PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] ==((void)0) | |||
| 14933 | Input / 2 &&((void)0) | |||
| 14934 | "Previous placement doesn't match!")((void)0); | |||
| 14935 | } | |||
| 14936 | ||||
| 14937 | // And just directly shift any other-half mask elements to be same-half | |||
| 14938 | // as we will have mirrored the dword containing the element into the | |||
| 14939 | // same position within that half. | |||
| 14940 | for (int &M : HalfMask) | |||
| 14941 | if (M >= SourceOffset && M < SourceOffset + 4) { | |||
| 14942 | M = M - SourceOffset + DestOffset; | |||
| 14943 | assert(M >= 0 && "This should never wrap below zero!")((void)0); | |||
| 14944 | } | |||
| 14945 | return; | |||
| 14946 | } | |||
| 14947 | ||||
| 14948 | // Ensure we have the input in a viable dword of its current half. This | |||
| 14949 | // is particularly tricky because the original position may be clobbered | |||
| 14950 | // by inputs being moved and *staying* in that half. | |||
| 14951 | if (IncomingInputs.size() == 1) { | |||
| 14952 | if (isWordClobbered(SourceHalfMask, IncomingInputs[0] - SourceOffset)) { | |||
| 14953 | int InputFixed = find(SourceHalfMask, -1) - std::begin(SourceHalfMask) + | |||
| 14954 | SourceOffset; | |||
| 14955 | SourceHalfMask[InputFixed - SourceOffset] = | |||
| 14956 | IncomingInputs[0] - SourceOffset; | |||
| 14957 | std::replace(HalfMask.begin(), HalfMask.end(), IncomingInputs[0], | |||
| 14958 | InputFixed); | |||
| 14959 | IncomingInputs[0] = InputFixed; | |||
| 14960 | } | |||
| 14961 | } else if (IncomingInputs.size() == 2) { | |||
| 14962 | if (IncomingInputs[0] / 2 != IncomingInputs[1] / 2 || | |||
| 14963 | isDWordClobbered(SourceHalfMask, IncomingInputs[0] - SourceOffset)) { | |||
| 14964 | // We have two non-adjacent or clobbered inputs we need to extract from | |||
| 14965 | // the source half. To do this, we need to map them into some adjacent | |||
| 14966 | // dword slot in the source mask. | |||
| 14967 | int InputsFixed[2] = {IncomingInputs[0] - SourceOffset, | |||
| 14968 | IncomingInputs[1] - SourceOffset}; | |||
| 14969 | ||||
| 14970 | // If there is a free slot in the source half mask adjacent to one of | |||
| 14971 | // the inputs, place the other input in it. We use (Index XOR 1) to | |||
| 14972 | // compute an adjacent index. | |||
| 14973 | if (!isWordClobbered(SourceHalfMask, InputsFixed[0]) && | |||
| 14974 | SourceHalfMask[InputsFixed[0] ^ 1] < 0) { | |||
| 14975 | SourceHalfMask[InputsFixed[0]] = InputsFixed[0]; | |||
| 14976 | SourceHalfMask[InputsFixed[0] ^ 1] = InputsFixed[1]; | |||
| 14977 | InputsFixed[1] = InputsFixed[0] ^ 1; | |||
| 14978 | } else if (!isWordClobbered(SourceHalfMask, InputsFixed[1]) && | |||
| 14979 | SourceHalfMask[InputsFixed[1] ^ 1] < 0) { | |||
| 14980 | SourceHalfMask[InputsFixed[1]] = InputsFixed[1]; | |||
| 14981 | SourceHalfMask[InputsFixed[1] ^ 1] = InputsFixed[0]; | |||
| 14982 | InputsFixed[0] = InputsFixed[1] ^ 1; | |||
| 14983 | } else if (SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1)] < 0 && | |||
| 14984 | SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1) + 1] < 0) { | |||
| 14985 | // The two inputs are in the same DWord but it is clobbered and the | |||
| 14986 | // adjacent DWord isn't used at all. Move both inputs to the free | |||
| 14987 | // slot. | |||
| 14988 | SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1)] = InputsFixed[0]; | |||
| 14989 | SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1) + 1] = InputsFixed[1]; | |||
| 14990 | InputsFixed[0] = 2 * ((InputsFixed[0] / 2) ^ 1); | |||
| 14991 | InputsFixed[1] = 2 * ((InputsFixed[0] / 2) ^ 1) + 1; | |||
| 14992 | } else { | |||
| 14993 | // The only way we hit this point is if there is no clobbering | |||
| 14994 | // (because there are no off-half inputs to this half) and there is no | |||
| 14995 | // free slot adjacent to one of the inputs. In this case, we have to | |||
| 14996 | // swap an input with a non-input. | |||
| 14997 | for (int i = 0; i < 4; ++i) | |||
| 14998 | assert((SourceHalfMask[i] < 0 || SourceHalfMask[i] == i) &&((void)0) | |||
| 14999 | "We can't handle any clobbers here!")((void)0); | |||
| 15000 | assert(InputsFixed[1] != (InputsFixed[0] ^ 1) &&((void)0) | |||
| 15001 | "Cannot have adjacent inputs here!")((void)0); | |||
| 15002 | ||||
| 15003 | SourceHalfMask[InputsFixed[0] ^ 1] = InputsFixed[1]; | |||
| 15004 | SourceHalfMask[InputsFixed[1]] = InputsFixed[0] ^ 1; | |||
| 15005 | ||||
| 15006 | // We also have to update the final source mask in this case because | |||
| 15007 | // it may need to undo the above swap. | |||
| 15008 | for (int &M : FinalSourceHalfMask) | |||
| 15009 | if (M == (InputsFixed[0] ^ 1) + SourceOffset) | |||
| 15010 | M = InputsFixed[1] + SourceOffset; | |||
| 15011 | else if (M == InputsFixed[1] + SourceOffset) | |||
| 15012 | M = (InputsFixed[0] ^ 1) + SourceOffset; | |||
| 15013 | ||||
| 15014 | InputsFixed[1] = InputsFixed[0] ^ 1; | |||
| 15015 | } | |||
| 15016 | ||||
| 15017 | // Point everything at the fixed inputs. | |||
| 15018 | for (int &M : HalfMask) | |||
| 15019 | if (M == IncomingInputs[0]) | |||
| 15020 | M = InputsFixed[0] + SourceOffset; | |||
| 15021 | else if (M == IncomingInputs[1]) | |||
| 15022 | M = InputsFixed[1] + SourceOffset; | |||
| 15023 | ||||
| 15024 | IncomingInputs[0] = InputsFixed[0] + SourceOffset; | |||
| 15025 | IncomingInputs[1] = InputsFixed[1] + SourceOffset; | |||
| 15026 | } | |||
| 15027 | } else { | |||
| 15028 | llvm_unreachable("Unhandled input size!")__builtin_unreachable(); | |||
| 15029 | } | |||
| 15030 | ||||
| 15031 | // Now hoist the DWord down to the right half. | |||
| 15032 | int FreeDWord = (PSHUFDMask[DestOffset / 2] < 0 ? 0 : 1) + DestOffset / 2; | |||
| 15033 | assert(PSHUFDMask[FreeDWord] < 0 && "DWord not free")((void)0); | |||
| 15034 | PSHUFDMask[FreeDWord] = IncomingInputs[0] / 2; | |||
| 15035 | for (int &M : HalfMask) | |||
| 15036 | for (int Input : IncomingInputs) | |||
| 15037 | if (M == Input) | |||
| 15038 | M = FreeDWord * 2 + Input % 2; | |||
| 15039 | }; | |||
| 15040 | moveInputsToRightHalf(HToLInputs, LToLInputs, PSHUFHMask, LoMask, HiMask, | |||
| 15041 | /*SourceOffset*/ 4, /*DestOffset*/ 0); | |||
| 15042 | moveInputsToRightHalf(LToHInputs, HToHInputs, PSHUFLMask, HiMask, LoMask, | |||
| 15043 | /*SourceOffset*/ 0, /*DestOffset*/ 4); | |||
| 15044 | ||||
| 15045 | // Now enact all the shuffles we've computed to move the inputs into their | |||
| 15046 | // target half. | |||
| 15047 | if (!isNoopShuffleMask(PSHUFLMask)) | |||
| 15048 | V = DAG.getNode(X86ISD::PSHUFLW, DL, VT, V, | |||
| 15049 | getV4X86ShuffleImm8ForMask(PSHUFLMask, DL, DAG)); | |||
| 15050 | if (!isNoopShuffleMask(PSHUFHMask)) | |||
| 15051 | V = DAG.getNode(X86ISD::PSHUFHW, DL, VT, V, | |||
| 15052 | getV4X86ShuffleImm8ForMask(PSHUFHMask, DL, DAG)); | |||
| 15053 | if (!isNoopShuffleMask(PSHUFDMask)) | |||
| 15054 | V = DAG.getBitcast( | |||
| 15055 | VT, | |||
| 15056 | DAG.getNode(X86ISD::PSHUFD, DL, PSHUFDVT, DAG.getBitcast(PSHUFDVT, V), | |||
| 15057 | getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG))); | |||
| 15058 | ||||
| 15059 | // At this point, each half should contain all its inputs, and we can then | |||
| 15060 | // just shuffle them into their final position. | |||
| 15061 | assert(count_if(LoMask, [](int M) { return M >= 4; }) == 0 &&((void)0) | |||
| 15062 | "Failed to lift all the high half inputs to the low mask!")((void)0); | |||
| 15063 | assert(count_if(HiMask, [](int M) { return M >= 0 && M < 4; }) == 0 &&((void)0) | |||
| 15064 | "Failed to lift all the low half inputs to the high mask!")((void)0); | |||
| 15065 | ||||
| 15066 | // Do a half shuffle for the low mask. | |||
| 15067 | if (!isNoopShuffleMask(LoMask)) | |||
| 15068 | V = DAG.getNode(X86ISD::PSHUFLW, DL, VT, V, | |||
| 15069 | getV4X86ShuffleImm8ForMask(LoMask, DL, DAG)); | |||
| 15070 | ||||
| 15071 | // Do a half shuffle with the high mask after shifting its values down. | |||
| 15072 | for (int &M : HiMask) | |||
| 15073 | if (M >= 0) | |||
| 15074 | M -= 4; | |||
| 15075 | if (!isNoopShuffleMask(HiMask)) | |||
| 15076 | V = DAG.getNode(X86ISD::PSHUFHW, DL, VT, V, | |||
| 15077 | getV4X86ShuffleImm8ForMask(HiMask, DL, DAG)); | |||
| 15078 | ||||
| 15079 | return V; | |||
| 15080 | } | |||
| 15081 | ||||
| 15082 | /// Helper to form a PSHUFB-based shuffle+blend, opportunistically avoiding the | |||
| 15083 | /// blend if only one input is used. | |||
| 15084 | static SDValue lowerShuffleAsBlendOfPSHUFBs( | |||
| 15085 | const SDLoc &DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask, | |||
| 15086 | const APInt &Zeroable, SelectionDAG &DAG, bool &V1InUse, bool &V2InUse) { | |||
| 15087 | assert(!is128BitLaneCrossingShuffleMask(VT, Mask) &&((void)0) | |||
| 15088 | "Lane crossing shuffle masks not supported")((void)0); | |||
| 15089 | ||||
| 15090 | int NumBytes = VT.getSizeInBits() / 8; | |||
| 15091 | int Size = Mask.size(); | |||
| 15092 | int Scale = NumBytes / Size; | |||
| 15093 | ||||
| 15094 | SmallVector<SDValue, 64> V1Mask(NumBytes, DAG.getUNDEF(MVT::i8)); | |||
| 15095 | SmallVector<SDValue, 64> V2Mask(NumBytes, DAG.getUNDEF(MVT::i8)); | |||
| 15096 | V1InUse = false; | |||
| 15097 | V2InUse = false; | |||
| 15098 | ||||
| 15099 | for (int i = 0; i < NumBytes; ++i) { | |||
| 15100 | int M = Mask[i / Scale]; | |||
| 15101 | if (M < 0) | |||
| 15102 | continue; | |||
| 15103 | ||||
| 15104 | const int ZeroMask = 0x80; | |||
| 15105 | int V1Idx = M < Size ? M * Scale + i % Scale : ZeroMask; | |||
| 15106 | int V2Idx = M < Size ? ZeroMask : (M - Size) * Scale + i % Scale; | |||
| 15107 | if (Zeroable[i / Scale]) | |||
| 15108 | V1Idx = V2Idx = ZeroMask; | |||
| 15109 | ||||
| 15110 | V1Mask[i] = DAG.getConstant(V1Idx, DL, MVT::i8); | |||
| 15111 | V2Mask[i] = DAG.getConstant(V2Idx, DL, MVT::i8); | |||
| 15112 | V1InUse |= (ZeroMask != V1Idx); | |||
| 15113 | V2InUse |= (ZeroMask != V2Idx); | |||
| 15114 | } | |||
| 15115 | ||||
| 15116 | MVT ShufVT = MVT::getVectorVT(MVT::i8, NumBytes); | |||
| 15117 | if (V1InUse) | |||
| 15118 | V1 = DAG.getNode(X86ISD::PSHUFB, DL, ShufVT, DAG.getBitcast(ShufVT, V1), | |||
| 15119 | DAG.getBuildVector(ShufVT, DL, V1Mask)); | |||
| 15120 | if (V2InUse) | |||
| 15121 | V2 = DAG.getNode(X86ISD::PSHUFB, DL, ShufVT, DAG.getBitcast(ShufVT, V2), | |||
| 15122 | DAG.getBuildVector(ShufVT, DL, V2Mask)); | |||
| 15123 | ||||
| 15124 | // If we need shuffled inputs from both, blend the two. | |||
| 15125 | SDValue V; | |||
| 15126 | if (V1InUse && V2InUse) | |||
| 15127 | V = DAG.getNode(ISD::OR, DL, ShufVT, V1, V2); | |||
| 15128 | else | |||
| 15129 | V = V1InUse ? V1 : V2; | |||
| 15130 | ||||
| 15131 | // Cast the result back to the correct type. | |||
| 15132 | return DAG.getBitcast(VT, V); | |||
| 15133 | } | |||
| 15134 | ||||
| 15135 | /// Generic lowering of 8-lane i16 shuffles. | |||
| 15136 | /// | |||
| 15137 | /// This handles both single-input shuffles and combined shuffle/blends with | |||
| 15138 | /// two inputs. The single input shuffles are immediately delegated to | |||
| 15139 | /// a dedicated lowering routine. | |||
| 15140 | /// | |||
| 15141 | /// The blends are lowered in one of three fundamental ways. If there are few | |||
| 15142 | /// enough inputs, it delegates to a basic UNPCK-based strategy. If the shuffle | |||
| 15143 | /// of the input is significantly cheaper when lowered as an interleaving of | |||
| 15144 | /// the two inputs, try to interleave them. Otherwise, blend the low and high | |||
| 15145 | /// halves of the inputs separately (making them have relatively few inputs) | |||
| 15146 | /// and then concatenate them. | |||
| 15147 | static SDValue lowerV8I16Shuffle(const SDLoc &DL, ArrayRef<int> Mask, | |||
| 15148 | const APInt &Zeroable, SDValue V1, SDValue V2, | |||
| 15149 | const X86Subtarget &Subtarget, | |||
| 15150 | SelectionDAG &DAG) { | |||
| 15151 | assert(V1.getSimpleValueType() == MVT::v8i16 && "Bad operand type!")((void)0); | |||
| 15152 | assert(V2.getSimpleValueType() == MVT::v8i16 && "Bad operand type!")((void)0); | |||
| 15153 | assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!")((void)0); | |||
| 15154 | ||||
| 15155 | // Whenever we can lower this as a zext, that instruction is strictly faster | |||
| 15156 | // than any alternative. | |||
| 15157 | if (SDValue ZExt = lowerShuffleAsZeroOrAnyExtend(DL, MVT::v8i16, V1, V2, Mask, | |||
| 15158 | Zeroable, Subtarget, DAG)) | |||
| 15159 | return ZExt; | |||
| 15160 | ||||
| 15161 | // Try to use lower using a truncation. | |||
| 15162 | if (SDValue V = lowerShuffleWithVPMOV(DL, MVT::v8i16, V1, V2, Mask, Zeroable, | |||
| 15163 | Subtarget, DAG)) | |||
| 15164 | return V; | |||
| 15165 | ||||
| 15166 | int NumV2Inputs = count_if(Mask, [](int M) { return M >= 8; }); | |||
| 15167 | ||||
| 15168 | if (NumV2Inputs == 0) { | |||
| 15169 | // Try to use shift instructions. | |||
| 15170 | if (SDValue Shift = lowerShuffleAsShift(DL, MVT::v8i16, V1, V1, Mask, | |||
| 15171 | Zeroable, Subtarget, DAG)) | |||
| 15172 | return Shift; | |||
| 15173 | ||||
| 15174 | // Check for being able to broadcast a single element. | |||
| 15175 | if (SDValue Broadcast = lowerShuffleAsBroadcast(DL, MVT::v8i16, V1, V2, | |||
| 15176 | Mask, Subtarget, DAG)) | |||
| 15177 | return Broadcast; | |||
| 15178 | ||||
| 15179 | // Try to use bit rotation instructions. | |||
| 15180 | if (SDValue Rotate = lowerShuffleAsBitRotate(DL, MVT::v8i16, V1, Mask, | |||
| 15181 | Subtarget, DAG)) | |||
| 15182 | return Rotate; | |||
| 15183 | ||||
| 15184 | // Use dedicated unpack instructions for masks that match their pattern. | |||
| 15185 | if (SDValue V = lowerShuffleWithUNPCK(DL, MVT::v8i16, Mask, V1, V2, DAG)) | |||
| 15186 | return V; | |||
| 15187 | ||||
| 15188 | // Use dedicated pack instructions for masks that match their pattern. | |||
| 15189 | if (SDValue V = lowerShuffleWithPACK(DL, MVT::v8i16, Mask, V1, V2, DAG, | |||
| 15190 | Subtarget)) | |||
| 15191 | return V; | |||
| 15192 | ||||
| 15193 | // Try to use byte rotation instructions. | |||
| 15194 | if (SDValue Rotate = lowerShuffleAsByteRotate(DL, MVT::v8i16, V1, V1, Mask, | |||
| 15195 | Subtarget, DAG)) | |||
| 15196 | return Rotate; | |||
| 15197 | ||||
| 15198 | // Make a copy of the mask so it can be modified. | |||
| 15199 | SmallVector<int, 8> MutableMask(Mask.begin(), Mask.end()); | |||
| 15200 | return lowerV8I16GeneralSingleInputShuffle(DL, MVT::v8i16, V1, MutableMask, | |||
| 15201 | Subtarget, DAG); | |||
| 15202 | } | |||
| 15203 | ||||
| 15204 | assert(llvm::any_of(Mask, [](int M) { return M >= 0 && M < 8; }) &&((void)0) | |||
| 15205 | "All single-input shuffles should be canonicalized to be V1-input "((void)0) | |||
| 15206 | "shuffles.")((void)0); | |||
| 15207 | ||||
| 15208 | // Try to use shift instructions. | |||
| 15209 | if (SDValue Shift = lowerShuffleAsShift(DL, MVT::v8i16, V1, V2, Mask, | |||
| 15210 | Zeroable, Subtarget, DAG)) | |||
| 15211 | return Shift; | |||
| 15212 | ||||
| 15213 | // See if we can use SSE4A Extraction / Insertion. | |||
| 15214 | if (Subtarget.hasSSE4A()) | |||
| 15215 | if (SDValue V = lowerShuffleWithSSE4A(DL, MVT::v8i16, V1, V2, Mask, | |||
| 15216 | Zeroable, DAG)) | |||
| 15217 | return V; | |||
| 15218 | ||||
| 15219 | // There are special ways we can lower some single-element blends. | |||
| 15220 | if (NumV2Inputs == 1) | |||
| 15221 | if (SDValue V = lowerShuffleAsElementInsertion( | |||
| 15222 | DL, MVT::v8i16, V1, V2, Mask, Zeroable, Subtarget, DAG)) | |||
| 15223 | return V; | |||
| 15224 | ||||
| 15225 | // We have different paths for blend lowering, but they all must use the | |||
| 15226 | // *exact* same predicate. | |||
| 15227 | bool IsBlendSupported = Subtarget.hasSSE41(); | |||
| 15228 | if (IsBlendSupported) | |||
| 15229 | if (SDValue Blend = lowerShuffleAsBlend(DL, MVT::v8i16, V1, V2, Mask, | |||
| 15230 | Zeroable, Subtarget, DAG)) | |||
| 15231 | return Blend; | |||
| 15232 | ||||
| 15233 | if (SDValue Masked = lowerShuffleAsBitMask(DL, MVT::v8i16, V1, V2, Mask, | |||
| 15234 | Zeroable, Subtarget, DAG)) | |||
| 15235 | return Masked; | |||
| 15236 | ||||
| 15237 | // Use dedicated unpack instructions for masks that match their pattern. | |||
| 15238 | if (SDValue V = lowerShuffleWithUNPCK(DL, MVT::v8i16, Mask, V1, V2, DAG)) | |||
| 15239 | return V; | |||
| 15240 | ||||
| 15241 | // Use dedicated pack instructions for masks that match their pattern. | |||
| 15242 | if (SDValue V = lowerShuffleWithPACK(DL, MVT::v8i16, Mask, V1, V2, DAG, | |||
| 15243 | Subtarget)) | |||
| 15244 | return V; | |||
| 15245 | ||||
| 15246 | // Try to use lower using a truncation. | |||
| 15247 | if (SDValue V = lowerShuffleAsVTRUNC(DL, MVT::v8i16, V1, V2, Mask, Zeroable, | |||
| 15248 | Subtarget, DAG)) | |||
| 15249 | return V; | |||
| 15250 | ||||
| 15251 | // Try to use byte rotation instructions. | |||
| 15252 | if (SDValue Rotate = lowerShuffleAsByteRotate(DL, MVT::v8i16, V1, V2, Mask, | |||
| 15253 | Subtarget, DAG)) | |||
| 15254 | return Rotate; | |||
| 15255 | ||||
| 15256 | if (SDValue BitBlend = | |||
| 15257 | lowerShuffleAsBitBlend(DL, MVT::v8i16, V1, V2, Mask, DAG)) | |||
| 15258 | return BitBlend; | |||
| 15259 | ||||
| 15260 | // Try to use byte shift instructions to mask. | |||
| 15261 | if (SDValue V = lowerShuffleAsByteShiftMask(DL, MVT::v8i16, V1, V2, Mask, | |||
| 15262 | Zeroable, Subtarget, DAG)) | |||
| 15263 | return V; | |||
| 15264 | ||||
| 15265 | // Attempt to lower using compaction, SSE41 is necessary for PACKUSDW. | |||
| 15266 | // We could use SIGN_EXTEND_INREG+PACKSSDW for older targets but this seems to | |||
| 15267 | // be slower than a PSHUFLW+PSHUFHW+PSHUFD chain. | |||
| 15268 | int NumEvenDrops = canLowerByDroppingEvenElements(Mask, false); | |||
| 15269 | if ((NumEvenDrops == 1 || NumEvenDrops == 2) && Subtarget.hasSSE41() && | |||
| 15270 | !Subtarget.hasVLX()) { | |||
| 15271 | SmallVector<SDValue, 8> DWordClearOps(4, DAG.getConstant(0, DL, MVT::i32)); | |||
| 15272 | for (unsigned i = 0; i != 4; i += 1 << (NumEvenDrops - 1)) | |||
| 15273 | DWordClearOps[i] = DAG.getConstant(0xFFFF, DL, MVT::i32); | |||
| 15274 | SDValue DWordClearMask = DAG.getBuildVector(MVT::v4i32, DL, DWordClearOps); | |||
| 15275 | V1 = DAG.getNode(ISD::AND, DL, MVT::v4i32, DAG.getBitcast(MVT::v4i32, V1), | |||
| 15276 | DWordClearMask); | |||
| 15277 | V2 = DAG.getNode(ISD::AND, DL, MVT::v4i32, DAG.getBitcast(MVT::v4i32, V2), | |||
| 15278 | DWordClearMask); | |||
| 15279 | // Now pack things back together. | |||
| 15280 | SDValue Result = DAG.getNode(X86ISD::PACKUS, DL, MVT::v8i16, V1, V2); | |||
| 15281 | if (NumEvenDrops == 2) { | |||
| 15282 | Result = DAG.getBitcast(MVT::v4i32, Result); | |||
| 15283 | Result = DAG.getNode(X86ISD::PACKUS, DL, MVT::v8i16, Result, Result); | |||
| 15284 | } | |||
| 15285 | return Result; | |||
| 15286 | } | |||
| 15287 | ||||
| 15288 | // Try to lower by permuting the inputs into an unpack instruction. | |||
| 15289 | if (SDValue Unpack = lowerShuffleAsPermuteAndUnpack(DL, MVT::v8i16, V1, V2, | |||
| 15290 | Mask, Subtarget, DAG)) | |||
| 15291 | return Unpack; | |||
| 15292 | ||||
| 15293 | // If we can't directly blend but can use PSHUFB, that will be better as it | |||
| 15294 | // can both shuffle and set up the inefficient blend. | |||
| 15295 | if (!IsBlendSupported && Subtarget.hasSSSE3()) { | |||
| 15296 | bool V1InUse, V2InUse; | |||
| 15297 | return lowerShuffleAsBlendOfPSHUFBs(DL, MVT::v8i16, V1, V2, Mask, | |||
| 15298 | Zeroable, DAG, V1InUse, V2InUse); | |||
| 15299 | } | |||
| 15300 | ||||
| 15301 | // We can always bit-blend if we have to so the fallback strategy is to | |||
| 15302 | // decompose into single-input permutes and blends/unpacks. | |||
| 15303 | return lowerShuffleAsDecomposedShuffleMerge(DL, MVT::v8i16, V1, V2, | |||
| 15304 | Mask, Subtarget, DAG); | |||
| 15305 | } | |||
| 15306 | ||||
| 15307 | // Lowers unary/binary shuffle as VPERMV/VPERMV3, for non-VLX targets, | |||
| 15308 | // sub-512-bit shuffles are padded to 512-bits for the shuffle and then | |||
| 15309 | // the active subvector is extracted. | |||
| 15310 | static SDValue lowerShuffleWithPERMV(const SDLoc &DL, MVT VT, | |||
| 15311 | ArrayRef<int> Mask, SDValue V1, SDValue V2, | |||
| 15312 | const X86Subtarget &Subtarget, | |||
| 15313 | SelectionDAG &DAG) { | |||
| 15314 | MVT MaskVT = VT.changeTypeToInteger(); | |||
| 15315 | SDValue MaskNode; | |||
| 15316 | MVT ShuffleVT = VT; | |||
| 15317 | if (!VT.is512BitVector() && !Subtarget.hasVLX()) { | |||
| 15318 | V1 = widenSubVector(V1, false, Subtarget, DAG, DL, 512); | |||
| 15319 | V2 = widenSubVector(V2, false, Subtarget, DAG, DL, 512); | |||
| 15320 | ShuffleVT = V1.getSimpleValueType(); | |||
| 15321 | ||||
| 15322 | // Adjust mask to correct indices for the second input. | |||
| 15323 | int NumElts = VT.getVectorNumElements(); | |||
| 15324 | unsigned Scale = 512 / VT.getSizeInBits(); | |||
| 15325 | SmallVector<int, 32> AdjustedMask(Mask.begin(), Mask.end()); | |||
| 15326 | for (int &M : AdjustedMask) | |||
| 15327 | if (NumElts <= M) | |||
| 15328 | M += (Scale - 1) * NumElts; | |||
| 15329 | MaskNode = getConstVector(AdjustedMask, MaskVT, DAG, DL, true); | |||
| 15330 | MaskNode = widenSubVector(MaskNode, false, Subtarget, DAG, DL, 512); | |||
| 15331 | } else { | |||
| 15332 | MaskNode = getConstVector(Mask, MaskVT, DAG, DL, true); | |||
| 15333 | } | |||
| 15334 | ||||
| 15335 | SDValue Result; | |||
| 15336 | if (V2.isUndef()) | |||
| 15337 | Result = DAG.getNode(X86ISD::VPERMV, DL, ShuffleVT, MaskNode, V1); | |||
| 15338 | else | |||
| 15339 | Result = DAG.getNode(X86ISD::VPERMV3, DL, ShuffleVT, V1, MaskNode, V2); | |||
| 15340 | ||||
| 15341 | if (VT != ShuffleVT) | |||
| 15342 | Result = extractSubVector(Result, 0, DAG, DL, VT.getSizeInBits()); | |||
| 15343 | ||||
| 15344 | return Result; | |||
| 15345 | } | |||
| 15346 | ||||
| 15347 | /// Generic lowering of v16i8 shuffles. | |||
| 15348 | /// | |||
| 15349 | /// This is a hybrid strategy to lower v16i8 vectors. It first attempts to | |||
| 15350 | /// detect any complexity reducing interleaving. If that doesn't help, it uses | |||
| 15351 | /// UNPCK to spread the i8 elements across two i16-element vectors, and uses | |||
| 15352 | /// the existing lowering for v8i16 blends on each half, finally PACK-ing them | |||
| 15353 | /// back together. | |||
| 15354 | static SDValue lowerV16I8Shuffle(const SDLoc &DL, ArrayRef<int> Mask, | |||
| 15355 | const APInt &Zeroable, SDValue V1, SDValue V2, | |||
| 15356 | const X86Subtarget &Subtarget, | |||
| 15357 | SelectionDAG &DAG) { | |||
| 15358 | assert(V1.getSimpleValueType() == MVT::v16i8 && "Bad operand type!")((void)0); | |||
| 15359 | assert(V2.getSimpleValueType() == MVT::v16i8 && "Bad operand type!")((void)0); | |||
| 15360 | assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!")((void)0); | |||
| 15361 | ||||
| 15362 | // Try to use shift instructions. | |||
| 15363 | if (SDValue Shift = lowerShuffleAsShift(DL, MVT::v16i8, V1, V2, Mask, | |||
| 15364 | Zeroable, Subtarget, DAG)) | |||
| 15365 | return Shift; | |||
| 15366 | ||||
| 15367 | // Try to use byte rotation instructions. | |||
| 15368 | if (SDValue Rotate = lowerShuffleAsByteRotate(DL, MVT::v16i8, V1, V2, Mask, | |||
| 15369 | Subtarget, DAG)) | |||
| 15370 | return Rotate; | |||
| 15371 | ||||
| 15372 | // Use dedicated pack instructions for masks that match their pattern. | |||
| 15373 | if (SDValue V = lowerShuffleWithPACK(DL, MVT::v16i8, Mask, V1, V2, DAG, | |||
| 15374 | Subtarget)) | |||
| 15375 | return V; | |||
| 15376 | ||||
| 15377 | // Try to use a zext lowering. | |||
| 15378 | if (SDValue ZExt = lowerShuffleAsZeroOrAnyExtend(DL, MVT::v16i8, V1, V2, Mask, | |||
| 15379 | Zeroable, Subtarget, DAG)) | |||
| 15380 | return ZExt; | |||
| 15381 | ||||
| 15382 | // Try to use lower using a truncation. | |||
| 15383 | if (SDValue V = lowerShuffleWithVPMOV(DL, MVT::v16i8, V1, V2, Mask, Zeroable, | |||
| 15384 | Subtarget, DAG)) | |||
| 15385 | return V; | |||
| 15386 | ||||
| 15387 | if (SDValue V = lowerShuffleAsVTRUNC(DL, MVT::v16i8, V1, V2, Mask, Zeroable, | |||
| 15388 | Subtarget, DAG)) | |||
| 15389 | return V; | |||
| 15390 | ||||
| 15391 | // See if we can use SSE4A Extraction / Insertion. | |||
| 15392 | if (Subtarget.hasSSE4A()) | |||
| 15393 | if (SDValue V = lowerShuffleWithSSE4A(DL, MVT::v16i8, V1, V2, Mask, | |||
| 15394 | Zeroable, DAG)) | |||
| 15395 | return V; | |||
| 15396 | ||||
| 15397 | int NumV2Elements = count_if(Mask, [](int M) { return M >= 16; }); | |||
| 15398 | ||||
| 15399 | // For single-input shuffles, there are some nicer lowering tricks we can use. | |||
| 15400 | if (NumV2Elements == 0) { | |||
| 15401 | // Check for being able to broadcast a single element. | |||
| 15402 | if (SDValue Broadcast = lowerShuffleAsBroadcast(DL, MVT::v16i8, V1, V2, | |||
| 15403 | Mask, Subtarget, DAG)) | |||
| 15404 | return Broadcast; | |||
| 15405 | ||||
| 15406 | // Try to use bit rotation instructions. | |||
| 15407 | if (SDValue Rotate = lowerShuffleAsBitRotate(DL, MVT::v16i8, V1, Mask, | |||
| 15408 | Subtarget, DAG)) | |||
| 15409 | return Rotate; | |||
| 15410 | ||||
| 15411 | if (SDValue V = lowerShuffleWithUNPCK(DL, MVT::v16i8, Mask, V1, V2, DAG)) | |||
| 15412 | return V; | |||
| 15413 | ||||
| 15414 | // Check whether we can widen this to an i16 shuffle by duplicating bytes. | |||
| 15415 | // Notably, this handles splat and partial-splat shuffles more efficiently. | |||
| 15416 | // However, it only makes sense if the pre-duplication shuffle simplifies | |||
| 15417 | // things significantly. Currently, this means we need to be able to | |||
| 15418 | // express the pre-duplication shuffle as an i16 shuffle. | |||
| 15419 | // | |||
| 15420 | // FIXME: We should check for other patterns which can be widened into an | |||
| 15421 | // i16 shuffle as well. | |||
| 15422 | auto canWidenViaDuplication = [](ArrayRef<int> Mask) { | |||
| 15423 | for (int i = 0; i < 16; i += 2) | |||
| 15424 | if (Mask[i] >= 0 && Mask[i + 1] >= 0 && Mask[i] != Mask[i + 1]) | |||
| 15425 | return false; | |||
| 15426 | ||||
| 15427 | return true; | |||
| 15428 | }; | |||
| 15429 | auto tryToWidenViaDuplication = [&]() -> SDValue { | |||
| 15430 | if (!canWidenViaDuplication(Mask)) | |||
| 15431 | return SDValue(); | |||
| 15432 | SmallVector<int, 4> LoInputs; | |||
| 15433 | copy_if(Mask, std::back_inserter(LoInputs), | |||
| 15434 | [](int M) { return M >= 0 && M < 8; }); | |||
| 15435 | array_pod_sort(LoInputs.begin(), LoInputs.end()); | |||
| 15436 | LoInputs.erase(std::unique(LoInputs.begin(), LoInputs.end()), | |||
| 15437 | LoInputs.end()); | |||
| 15438 | SmallVector<int, 4> HiInputs; | |||
| 15439 | copy_if(Mask, std::back_inserter(HiInputs), [](int M) { return M >= 8; }); | |||
| 15440 | array_pod_sort(HiInputs.begin(), HiInputs.end()); | |||
| 15441 | HiInputs.erase(std::unique(HiInputs.begin(), HiInputs.end()), | |||
| 15442 | HiInputs.end()); | |||
| 15443 | ||||
| 15444 | bool TargetLo = LoInputs.size() >= HiInputs.size(); | |||
| 15445 | ArrayRef<int> InPlaceInputs = TargetLo ? LoInputs : HiInputs; | |||
| 15446 | ArrayRef<int> MovingInputs = TargetLo ? HiInputs : LoInputs; | |||
| 15447 | ||||
| 15448 | int PreDupI16Shuffle[] = {-1, -1, -1, -1, -1, -1, -1, -1}; | |||
| 15449 | SmallDenseMap<int, int, 8> LaneMap; | |||
| 15450 | for (int I : InPlaceInputs) { | |||
| 15451 | PreDupI16Shuffle[I/2] = I/2; | |||
| 15452 | LaneMap[I] = I; | |||
| 15453 | } | |||
| 15454 | int j = TargetLo ? 0 : 4, je = j + 4; | |||
| 15455 | for (int i = 0, ie = MovingInputs.size(); i < ie; ++i) { | |||
| 15456 | // Check if j is already a shuffle of this input. This happens when | |||
| 15457 | // there are two adjacent bytes after we move the low one. | |||
| 15458 | if (PreDupI16Shuffle[j] != MovingInputs[i] / 2) { | |||
| 15459 | // If we haven't yet mapped the input, search for a slot into which | |||
| 15460 | // we can map it. | |||
| 15461 | while (j < je && PreDupI16Shuffle[j] >= 0) | |||
| 15462 | ++j; | |||
| 15463 | ||||
| 15464 | if (j == je) | |||
| 15465 | // We can't place the inputs into a single half with a simple i16 shuffle, so bail. | |||
| 15466 | return SDValue(); | |||
| 15467 | ||||
| 15468 | // Map this input with the i16 shuffle. | |||
| 15469 | PreDupI16Shuffle[j] = MovingInputs[i] / 2; | |||
| 15470 | } | |||
| 15471 | ||||
| 15472 | // Update the lane map based on the mapping we ended up with. | |||
| 15473 | LaneMap[MovingInputs[i]] = 2 * j + MovingInputs[i] % 2; | |||
| 15474 | } | |||
| 15475 | V1 = DAG.getBitcast( | |||
| 15476 | MVT::v16i8, | |||
| 15477 | DAG.getVectorShuffle(MVT::v8i16, DL, DAG.getBitcast(MVT::v8i16, V1), | |||
| 15478 | DAG.getUNDEF(MVT::v8i16), PreDupI16Shuffle)); | |||
| 15479 | ||||
| 15480 | // Unpack the bytes to form the i16s that will be shuffled into place. | |||
| 15481 | bool EvenInUse = false, OddInUse = false; | |||
| 15482 | for (int i = 0; i < 16; i += 2) { | |||
| 15483 | EvenInUse |= (Mask[i + 0] >= 0); | |||
| 15484 | OddInUse |= (Mask[i + 1] >= 0); | |||
| 15485 | if (EvenInUse && OddInUse) | |||
| 15486 | break; | |||
| 15487 | } | |||
| 15488 | V1 = DAG.getNode(TargetLo ? X86ISD::UNPCKL : X86ISD::UNPCKH, DL, | |||
| 15489 | MVT::v16i8, EvenInUse ? V1 : DAG.getUNDEF(MVT::v16i8), | |||
| 15490 | OddInUse ? V1 : DAG.getUNDEF(MVT::v16i8)); | |||
| 15491 | ||||
| 15492 | int PostDupI16Shuffle[8] = {-1, -1, -1, -1, -1, -1, -1, -1}; | |||
| 15493 | for (int i = 0; i < 16; ++i) | |||
| 15494 | if (Mask[i] >= 0) { | |||
| 15495 | int MappedMask = LaneMap[Mask[i]] - (TargetLo ? 0 : 8); | |||
| 15496 | assert(MappedMask < 8 && "Invalid v8 shuffle mask!")((void)0); | |||
| 15497 | if (PostDupI16Shuffle[i / 2] < 0) | |||
| 15498 | PostDupI16Shuffle[i / 2] = MappedMask; | |||
| 15499 | else | |||
| 15500 | assert(PostDupI16Shuffle[i / 2] == MappedMask &&((void)0) | |||
| 15501 | "Conflicting entries in the original shuffle!")((void)0); | |||
| 15502 | } | |||
| 15503 | return DAG.getBitcast( | |||
| 15504 | MVT::v16i8, | |||
| 15505 | DAG.getVectorShuffle(MVT::v8i16, DL, DAG.getBitcast(MVT::v8i16, V1), | |||
| 15506 | DAG.getUNDEF(MVT::v8i16), PostDupI16Shuffle)); | |||
| 15507 | }; | |||
| 15508 | if (SDValue V = tryToWidenViaDuplication()) | |||
| 15509 | return V; | |||
| 15510 | } | |||
| 15511 | ||||
| 15512 | if (SDValue Masked = lowerShuffleAsBitMask(DL, MVT::v16i8, V1, V2, Mask, | |||
| 15513 | Zeroable, Subtarget, DAG)) | |||
| 15514 | return Masked; | |||
| 15515 | ||||
| 15516 | // Use dedicated unpack instructions for masks that match their pattern. | |||
| 15517 | if (SDValue V = lowerShuffleWithUNPCK(DL, MVT::v16i8, Mask, V1, V2, DAG)) | |||
| 15518 | return V; | |||
| 15519 | ||||
| 15520 | // Try to use byte shift instructions to mask. | |||
| 15521 | if (SDValue V = lowerShuffleAsByteShiftMask(DL, MVT::v16i8, V1, V2, Mask, | |||
| 15522 | Zeroable, Subtarget, DAG)) | |||
| 15523 | return V; | |||
| 15524 | ||||
| 15525 | // Check for compaction patterns. | |||
| 15526 | bool IsSingleInput = V2.isUndef(); | |||
| 15527 | int NumEvenDrops = canLowerByDroppingEvenElements(Mask, IsSingleInput); | |||
| 15528 | ||||
| 15529 | // Check for SSSE3 which lets us lower all v16i8 shuffles much more directly | |||
| 15530 | // with PSHUFB. It is important to do this before we attempt to generate any | |||
| 15531 | // blends but after all of the single-input lowerings. If the single input | |||
| 15532 | // lowerings can find an instruction sequence that is faster than a PSHUFB, we | |||
| 15533 | // want to preserve that and we can DAG combine any longer sequences into | |||
| 15534 | // a PSHUFB in the end. But once we start blending from multiple inputs, | |||
| 15535 | // the complexity of DAG combining bad patterns back into PSHUFB is too high, | |||
| 15536 | // and there are *very* few patterns that would actually be faster than the | |||
| 15537 | // PSHUFB approach because of its ability to zero lanes. | |||
| 15538 | // | |||
| 15539 | // If the mask is a binary compaction, we can more efficiently perform this | |||
| 15540 | // as a PACKUS(AND(),AND()) - which is quicker than UNPACK(PSHUFB(),PSHUFB()). | |||
| 15541 | // | |||
| 15542 | // FIXME: The only exceptions to the above are blends which are exact | |||
| 15543 | // interleavings with direct instructions supporting them. We currently don't | |||
| 15544 | // handle those well here. | |||
| 15545 | if (Subtarget.hasSSSE3() && (IsSingleInput || NumEvenDrops != 1)) { | |||
| 15546 | bool V1InUse = false; | |||
| 15547 | bool V2InUse = false; | |||
| 15548 | ||||
| 15549 | SDValue PSHUFB = lowerShuffleAsBlendOfPSHUFBs( | |||
| 15550 | DL, MVT::v16i8, V1, V2, Mask, Zeroable, DAG, V1InUse, V2InUse); | |||
| 15551 | ||||
| 15552 | // If both V1 and V2 are in use and we can use a direct blend or an unpack, | |||
| 15553 | // do so. This avoids using them to handle blends-with-zero which is | |||
| 15554 | // important as a single pshufb is significantly faster for that. | |||
| 15555 | if (V1InUse && V2InUse) { | |||
| 15556 | if (Subtarget.hasSSE41()) | |||
| 15557 | if (SDValue Blend = lowerShuffleAsBlend(DL, MVT::v16i8, V1, V2, Mask, | |||
| 15558 | Zeroable, Subtarget, DAG)) | |||
| 15559 | return Blend; | |||
| 15560 | ||||
| 15561 | // We can use an unpack to do the blending rather than an or in some | |||
| 15562 | // cases. Even though the or may be (very minorly) more efficient, we | |||
| 15563 | // preference this lowering because there are common cases where part of | |||
| 15564 | // the complexity of the shuffles goes away when we do the final blend as | |||
| 15565 | // an unpack. | |||
| 15566 | // FIXME: It might be worth trying to detect if the unpack-feeding | |||
| 15567 | // shuffles will both be pshufb, in which case we shouldn't bother with | |||
| 15568 | // this. | |||
| 15569 | if (SDValue Unpack = lowerShuffleAsPermuteAndUnpack( | |||
| 15570 | DL, MVT::v16i8, V1, V2, Mask, Subtarget, DAG)) | |||
| 15571 | return Unpack; | |||
| 15572 | ||||
| 15573 | // AVX512VBMI can lower to VPERMB (non-VLX will pad to v64i8). | |||
| 15574 | if (Subtarget.hasVBMI()) | |||
| 15575 | return lowerShuffleWithPERMV(DL, MVT::v16i8, Mask, V1, V2, Subtarget, | |||
| 15576 | DAG); | |||
| 15577 | ||||
| 15578 | // If we have XOP we can use one VPPERM instead of multiple PSHUFBs. | |||
| 15579 | if (Subtarget.hasXOP()) { | |||
| 15580 | SDValue MaskNode = getConstVector(Mask, MVT::v16i8, DAG, DL, true); | |||
| 15581 | return DAG.getNode(X86ISD::VPPERM, DL, MVT::v16i8, V1, V2, MaskNode); | |||
| 15582 | } | |||
| 15583 | ||||
| 15584 | // Use PALIGNR+Permute if possible - permute might become PSHUFB but the | |||
| 15585 | // PALIGNR will be cheaper than the second PSHUFB+OR. | |||
| 15586 | if (SDValue V = lowerShuffleAsByteRotateAndPermute( | |||
| 15587 | DL, MVT::v16i8, V1, V2, Mask, Subtarget, DAG)) | |||
| 15588 | return V; | |||
| 15589 | } | |||
| 15590 | ||||
| 15591 | return PSHUFB; | |||
| 15592 | } | |||
| 15593 | ||||
| 15594 | // There are special ways we can lower some single-element blends. | |||
| 15595 | if (NumV2Elements == 1) | |||
| 15596 | if (SDValue V = lowerShuffleAsElementInsertion( | |||
| 15597 | DL, MVT::v16i8, V1, V2, Mask, Zeroable, Subtarget, DAG)) | |||
| 15598 | return V; | |||
| 15599 | ||||
| 15600 | if (SDValue Blend = lowerShuffleAsBitBlend(DL, MVT::v16i8, V1, V2, Mask, DAG)) | |||
| 15601 | return Blend; | |||
| 15602 | ||||
| 15603 | // Check whether a compaction lowering can be done. This handles shuffles | |||
| 15604 | // which take every Nth element for some even N. See the helper function for | |||
| 15605 | // details. | |||
| 15606 | // | |||
| 15607 | // We special case these as they can be particularly efficiently handled with | |||
| 15608 | // the PACKUSB instruction on x86 and they show up in common patterns of | |||
| 15609 | // rearranging bytes to truncate wide elements. | |||
| 15610 | if (NumEvenDrops) { | |||
| 15611 | // NumEvenDrops is the power of two stride of the elements. Another way of | |||
| 15612 | // thinking about it is that we need to drop the even elements this many | |||
| 15613 | // times to get the original input. | |||
| 15614 | ||||
| 15615 | // First we need to zero all the dropped bytes. | |||
| 15616 | assert(NumEvenDrops <= 3 &&((void)0) | |||
| 15617 | "No support for dropping even elements more than 3 times.")((void)0); | |||
| 15618 | SmallVector<SDValue, 8> WordClearOps(8, DAG.getConstant(0, DL, MVT::i16)); | |||
| 15619 | for (unsigned i = 0; i != 8; i += 1 << (NumEvenDrops - 1)) | |||
| 15620 | WordClearOps[i] = DAG.getConstant(0xFF, DL, MVT::i16); | |||
| 15621 | SDValue WordClearMask = DAG.getBuildVector(MVT::v8i16, DL, WordClearOps); | |||
| 15622 | V1 = DAG.getNode(ISD::AND, DL, MVT::v8i16, DAG.getBitcast(MVT::v8i16, V1), | |||
| 15623 | WordClearMask); | |||
| 15624 | if (!IsSingleInput) | |||
| 15625 | V2 = DAG.getNode(ISD::AND, DL, MVT::v8i16, DAG.getBitcast(MVT::v8i16, V2), | |||
| 15626 | WordClearMask); | |||
| 15627 | ||||
| 15628 | // Now pack things back together. | |||
| 15629 | SDValue Result = DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, V1, | |||
| 15630 | IsSingleInput ? V1 : V2); | |||
| 15631 | for (int i = 1; i < NumEvenDrops; ++i) { | |||
| 15632 | Result = DAG.getBitcast(MVT::v8i16, Result); | |||
| 15633 | Result = DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, Result, Result); | |||
| 15634 | } | |||
| 15635 | return Result; | |||
| 15636 | } | |||
| 15637 | ||||
| 15638 | // Handle multi-input cases by blending/unpacking single-input shuffles. | |||
| 15639 | if (NumV2Elements > 0) | |||
| 15640 | return lowerShuffleAsDecomposedShuffleMerge(DL, MVT::v16i8, V1, V2, Mask, | |||
| 15641 | Subtarget, DAG); | |||
| 15642 | ||||
| 15643 | // The fallback path for single-input shuffles widens this into two v8i16 | |||
| 15644 | // vectors with unpacks, shuffles those, and then pulls them back together | |||
| 15645 | // with a pack. | |||
| 15646 | SDValue V = V1; | |||
| 15647 | ||||
| 15648 | std::array<int, 8> LoBlendMask = {{-1, -1, -1, -1, -1, -1, -1, -1}}; | |||
| 15649 | std::array<int, 8> HiBlendMask = {{-1, -1, -1, -1, -1, -1, -1, -1}}; | |||
| 15650 | for (int i = 0; i < 16; ++i) | |||
| 15651 | if (Mask[i] >= 0) | |||
| 15652 | (i < 8 ? LoBlendMask[i] : HiBlendMask[i % 8]) = Mask[i]; | |||
| 15653 | ||||
| 15654 | SDValue VLoHalf, VHiHalf; | |||
| 15655 | // Check if any of the odd lanes in the v16i8 are used. If not, we can mask | |||
| 15656 | // them out and avoid using UNPCK{L,H} to extract the elements of V as | |||
| 15657 | // i16s. | |||
| 15658 | if (none_of(LoBlendMask, [](int M) { return M >= 0 && M % 2 == 1; }) && | |||
| 15659 | none_of(HiBlendMask, [](int M) { return M >= 0 && M % 2 == 1; })) { | |||
| 15660 | // Use a mask to drop the high bytes. | |||
| 15661 | VLoHalf = DAG.getBitcast(MVT::v8i16, V); | |||
| 15662 | VLoHalf = DAG.getNode(ISD::AND, DL, MVT::v8i16, VLoHalf, | |||
| 15663 | DAG.getConstant(0x00FF, DL, MVT::v8i16)); | |||
| 15664 | ||||
| 15665 | // This will be a single vector shuffle instead of a blend so nuke VHiHalf. | |||
| 15666 | VHiHalf = DAG.getUNDEF(MVT::v8i16); | |||
| 15667 | ||||
| 15668 | // Squash the masks to point directly into VLoHalf. | |||
| 15669 | for (int &M : LoBlendMask) | |||
| 15670 | if (M >= 0) | |||
| 15671 | M /= 2; | |||
| 15672 | for (int &M : HiBlendMask) | |||
| 15673 | if (M >= 0) | |||
| 15674 | M /= 2; | |||
| 15675 | } else { | |||
| 15676 | // Otherwise just unpack the low half of V into VLoHalf and the high half into | |||
| 15677 | // VHiHalf so that we can blend them as i16s. | |||
| 15678 | SDValue Zero = getZeroVector(MVT::v16i8, Subtarget, DAG, DL); | |||
| 15679 | ||||
| 15680 | VLoHalf = DAG.getBitcast( | |||
| 15681 | MVT::v8i16, DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i8, V, Zero)); | |||
| 15682 | VHiHalf = DAG.getBitcast( | |||
| 15683 | MVT::v8i16, DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i8, V, Zero)); | |||
| 15684 | } | |||
| 15685 | ||||
| 15686 | SDValue LoV = DAG.getVectorShuffle(MVT::v8i16, DL, VLoHalf, VHiHalf, LoBlendMask); | |||
| 15687 | SDValue HiV = DAG.getVectorShuffle(MVT::v8i16, DL, VLoHalf, VHiHalf, HiBlendMask); | |||
| 15688 | ||||
| 15689 | return DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, LoV, HiV); | |||
| 15690 | } | |||
| 15691 | ||||
| 15692 | /// Dispatching routine to lower various 128-bit x86 vector shuffles. | |||
| 15693 | /// | |||
| 15694 | /// This routine breaks down the specific type of 128-bit shuffle and | |||
| 15695 | /// dispatches to the lowering routines accordingly. | |||
| 15696 | static SDValue lower128BitShuffle(const SDLoc &DL, ArrayRef<int> Mask, | |||
| 15697 | MVT VT, SDValue V1, SDValue V2, | |||
| 15698 | const APInt &Zeroable, | |||
| 15699 | const X86Subtarget &Subtarget, | |||
| 15700 | SelectionDAG &DAG) { | |||
| 15701 | switch (VT.SimpleTy) { | |||
| 15702 | case MVT::v2i64: | |||
| 15703 | return lowerV2I64Shuffle(DL, Mask, Zeroable, V1, V2, Subtarget, DAG); | |||
| 15704 | case MVT::v2f64: | |||
| 15705 | return lowerV2F64Shuffle(DL, Mask, Zeroable, V1, V2, Subtarget, DAG); | |||
| 15706 | case MVT::v4i32: | |||
| 15707 | return lowerV4I32Shuffle(DL, Mask, Zeroable, V1, V2, Subtarget, DAG); | |||
| 15708 | case MVT::v4f32: | |||
| 15709 | return lowerV4F32Shuffle(DL, Mask, Zeroable, V1, V2, Subtarget, DAG); | |||
| 15710 | case MVT::v8i16: | |||
| 15711 | return lowerV8I16Shuffle(DL, Mask, Zeroable, V1, V2, Subtarget, DAG); | |||
| 15712 | case MVT::v16i8: | |||
| 15713 | return lowerV16I8Shuffle(DL, Mask, Zeroable, V1, V2, Subtarget, DAG); | |||
| 15714 | ||||
| 15715 | default: | |||
| 15716 | llvm_unreachable("Unimplemented!")__builtin_unreachable(); | |||
| 15717 | } | |||
| 15718 | } | |||
| 15719 | ||||
| 15720 | /// Generic routine to split vector shuffle into half-sized shuffles. | |||
| 15721 | /// | |||
| 15722 | /// This routine just extracts two subvectors, shuffles them independently, and | |||
| 15723 | /// then concatenates them back together. This should work effectively with all | |||
| 15724 | /// AVX vector shuffle types. | |||
| 15725 | static SDValue splitAndLowerShuffle(const SDLoc &DL, MVT VT, SDValue V1, | |||
| 15726 | SDValue V2, ArrayRef<int> Mask, | |||
| 15727 | SelectionDAG &DAG) { | |||
| 15728 | assert(VT.getSizeInBits() >= 256 &&((void)0) | |||
| 15729 | "Only for 256-bit or wider vector shuffles!")((void)0); | |||
| 15730 | assert(V1.getSimpleValueType() == VT && "Bad operand type!")((void)0); | |||
| 15731 | assert(V2.getSimpleValueType() == VT && "Bad operand type!")((void)0); | |||
| 15732 | ||||
| 15733 | ArrayRef<int> LoMask = Mask.slice(0, Mask.size() / 2); | |||
| 15734 | ArrayRef<int> HiMask = Mask.slice(Mask.size() / 2); | |||
| 15735 | ||||
| 15736 | int NumElements = VT.getVectorNumElements(); | |||
| 15737 | int SplitNumElements = NumElements / 2; | |||
| 15738 | MVT ScalarVT = VT.getVectorElementType(); | |||
| 15739 | MVT SplitVT = MVT::getVectorVT(ScalarVT, SplitNumElements); | |||
| 15740 | ||||
| 15741 | // Use splitVector/extractSubVector so that split build-vectors just build two | |||
| 15742 | // narrower build vectors. This helps shuffling with splats and zeros. | |||
| 15743 | auto SplitVector = [&](SDValue V) { | |||
| 15744 | SDValue LoV, HiV; | |||
| 15745 | std::tie(LoV, HiV) = splitVector(peekThroughBitcasts(V), DAG, DL); | |||
| 15746 | return std::make_pair(DAG.getBitcast(SplitVT, LoV), | |||
| 15747 | DAG.getBitcast(SplitVT, HiV)); | |||
| 15748 | }; | |||
| 15749 | ||||
| 15750 | SDValue LoV1, HiV1, LoV2, HiV2; | |||
| 15751 | std::tie(LoV1, HiV1) = SplitVector(V1); | |||
| 15752 | std::tie(LoV2, HiV2) = SplitVector(V2); | |||
| 15753 | ||||
| 15754 | // Now create two 4-way blends of these half-width vectors. | |||
| 15755 | auto HalfBlend = [&](ArrayRef<int> HalfMask) { | |||
| 15756 | bool UseLoV1 = false, UseHiV1 = false, UseLoV2 = false, UseHiV2 = false; | |||
| 15757 | SmallVector<int, 32> V1BlendMask((unsigned)SplitNumElements, -1); | |||
| 15758 | SmallVector<int, 32> V2BlendMask((unsigned)SplitNumElements, -1); | |||
| 15759 | SmallVector<int, 32> BlendMask((unsigned)SplitNumElements, -1); | |||
| 15760 | for (int i = 0; i < SplitNumElements; ++i) { | |||
| 15761 | int M = HalfMask[i]; | |||
| 15762 | if (M >= NumElements) { | |||
| 15763 | if (M >= NumElements + SplitNumElements) | |||
| 15764 | UseHiV2 = true; | |||
| 15765 | else | |||
| 15766 | UseLoV2 = true; | |||
| 15767 | V2BlendMask[i] = M - NumElements; | |||
| 15768 | BlendMask[i] = SplitNumElements + i; | |||
| 15769 | } else if (M >= 0) { | |||
| 15770 | if (M >= SplitNumElements) | |||
| 15771 | UseHiV1 = true; | |||
| 15772 | else | |||
| 15773 | UseLoV1 = true; | |||
| 15774 | V1BlendMask[i] = M; | |||
| 15775 | BlendMask[i] = i; | |||
| 15776 | } | |||
| 15777 | } | |||
| 15778 | ||||
| 15779 | // Because the lowering happens after all combining takes place, we need to | |||
| 15780 | // manually combine these blend masks as much as possible so that we create | |||
| 15781 | // a minimal number of high-level vector shuffle nodes. | |||
| 15782 | ||||
| 15783 | // First try just blending the halves of V1 or V2. | |||
| 15784 | if (!UseLoV1 && !UseHiV1 && !UseLoV2 && !UseHiV2) | |||
| 15785 | return DAG.getUNDEF(SplitVT); | |||
| 15786 | if (!UseLoV2 && !UseHiV2) | |||
| 15787 | return DAG.getVectorShuffle(SplitVT, DL, LoV1, HiV1, V1BlendMask); | |||
| 15788 | if (!UseLoV1 && !UseHiV1) | |||
| 15789 | return DAG.getVectorShuffle(SplitVT, DL, LoV2, HiV2, V2BlendMask); | |||
| 15790 | ||||
| 15791 | SDValue V1Blend, V2Blend; | |||
| 15792 | if (UseLoV1 && UseHiV1) { | |||
| 15793 | V1Blend = | |||
| 15794 | DAG.getVectorShuffle(SplitVT, DL, LoV1, HiV1, V1BlendMask); | |||
| 15795 | } else { | |||
| 15796 | // We only use half of V1 so map the usage down into the final blend mask. | |||
| 15797 | V1Blend = UseLoV1 ? LoV1 : HiV1; | |||
| 15798 | for (int i = 0; i < SplitNumElements; ++i) | |||
| 15799 | if (BlendMask[i] >= 0 && BlendMask[i] < SplitNumElements) | |||
| 15800 | BlendMask[i] = V1BlendMask[i] - (UseLoV1 ? 0 : SplitNumElements); | |||
| 15801 | } | |||
| 15802 | if (UseLoV2 && UseHiV2) { | |||
| 15803 | V2Blend = | |||
| 15804 | DAG.getVectorShuffle(SplitVT, DL, LoV2, HiV2, V2BlendMask); | |||
| 15805 | } else { | |||
| 15806 | // We only use half of V2 so map the usage down into the final blend mask. | |||
| 15807 | V2Blend = UseLoV2 ? LoV2 : HiV2; | |||
| 15808 | for (int i = 0; i < SplitNumElements; ++i) | |||
| 15809 | if (BlendMask[i] >= SplitNumElements) | |||
| 15810 | BlendMask[i] = V2BlendMask[i] + (UseLoV2 ? SplitNumElements : 0); | |||
| 15811 | } | |||
| 15812 | return DAG.getVectorShuffle(SplitVT, DL, V1Blend, V2Blend, BlendMask); | |||
| 15813 | }; | |||
| 15814 | SDValue Lo = HalfBlend(LoMask); | |||
| 15815 | SDValue Hi = HalfBlend(HiMask); | |||
| 15816 | return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Lo, Hi); | |||
| 15817 | } | |||
| 15818 | ||||
| 15819 | /// Either split a vector in halves or decompose the shuffles and the | |||
| 15820 | /// blend/unpack. | |||
| 15821 | /// | |||
| 15822 | /// This is provided as a good fallback for many lowerings of non-single-input | |||
| 15823 | /// shuffles with more than one 128-bit lane. In those cases, we want to select | |||
| 15824 | /// between splitting the shuffle into 128-bit components and stitching those | |||
| 15825 | /// back together vs. extracting the single-input shuffles and blending those | |||
| 15826 | /// results. | |||
| 15827 | static SDValue lowerShuffleAsSplitOrBlend(const SDLoc &DL, MVT VT, SDValue V1, | |||
| 15828 | SDValue V2, ArrayRef<int> Mask, | |||
| 15829 | const X86Subtarget &Subtarget, | |||
| 15830 | SelectionDAG &DAG) { | |||
| 15831 | assert(!V2.isUndef() && "This routine must not be used to lower single-input "((void)0) | |||
| 15832 | "shuffles as it could then recurse on itself.")((void)0); | |||
| 15833 | int Size = Mask.size(); | |||
| 15834 | ||||
| 15835 | // If this can be modeled as a broadcast of two elements followed by a blend, | |||
| 15836 | // prefer that lowering. This is especially important because broadcasts can | |||
| 15837 | // often fold with memory operands. | |||
| 15838 | auto DoBothBroadcast = [&] { | |||
| 15839 | int V1BroadcastIdx = -1, V2BroadcastIdx = -1; | |||
| 15840 | for (int M : Mask) | |||
| 15841 | if (M >= Size) { | |||
| 15842 | if (V2BroadcastIdx < 0) | |||
| 15843 | V2BroadcastIdx = M - Size; | |||
| 15844 | else if (M - Size != V2BroadcastIdx) | |||
| 15845 | return false; | |||
| 15846 | } else if (M >= 0) { | |||
| 15847 | if (V1BroadcastIdx < 0) | |||
| 15848 | V1BroadcastIdx = M; | |||
| 15849 | else if (M != V1BroadcastIdx) | |||
| 15850 | return false; | |||
| 15851 | } | |||
| 15852 | return true; | |||
| 15853 | }; | |||
| 15854 | if (DoBothBroadcast()) | |||
| 15855 | return lowerShuffleAsDecomposedShuffleMerge(DL, VT, V1, V2, Mask, Subtarget, | |||
| 15856 | DAG); | |||
| 15857 | ||||
| 15858 | // If the inputs all stem from a single 128-bit lane of each input, then we | |||
| 15859 | // split them rather than blending because the split will decompose to | |||
| 15860 | // unusually few instructions. | |||
| 15861 | int LaneCount = VT.getSizeInBits() / 128; | |||
| 15862 | int LaneSize = Size / LaneCount; | |||
| 15863 | SmallBitVector LaneInputs[2]; | |||
| 15864 | LaneInputs[0].resize(LaneCount, false); | |||
| 15865 | LaneInputs[1].resize(LaneCount, false); | |||
| 15866 | for (int i = 0; i < Size; ++i) | |||
| 15867 | if (Mask[i] >= 0) | |||
| 15868 | LaneInputs[Mask[i] / Size][(Mask[i] % Size) / LaneSize] = true; | |||
| 15869 | if (LaneInputs[0].count() <= 1 && LaneInputs[1].count() <= 1) | |||
| 15870 | return splitAndLowerShuffle(DL, VT, V1, V2, Mask, DAG); | |||
| 15871 | ||||
| 15872 | // Otherwise, just fall back to decomposed shuffles and a blend/unpack. This | |||
| 15873 | // requires that the decomposed single-input shuffles don't end up here. | |||
| 15874 | return lowerShuffleAsDecomposedShuffleMerge(DL, VT, V1, V2, Mask, Subtarget, | |||
| 15875 | DAG); | |||
| 15876 | } | |||
| 15877 | ||||
| 15878 | // Lower as SHUFPD(VPERM2F128(V1, V2), VPERM2F128(V1, V2)). | |||
| 15879 | // TODO: Extend to support v8f32 (+ 512-bit shuffles). | |||
| 15880 | static SDValue lowerShuffleAsLanePermuteAndSHUFP(const SDLoc &DL, MVT VT, | |||
| 15881 | SDValue V1, SDValue V2, | |||
| 15882 | ArrayRef<int> Mask, | |||
| 15883 | SelectionDAG &DAG) { | |||
| 15884 | assert(VT == MVT::v4f64 && "Only for v4f64 shuffles")((void)0); | |||
| 15885 | ||||
| 15886 | int LHSMask[4] = {-1, -1, -1, -1}; | |||
| 15887 | int RHSMask[4] = {-1, -1, -1, -1}; | |||
| 15888 | unsigned SHUFPMask = 0; | |||
| 15889 | ||||
| 15890 | // As SHUFPD uses a single LHS/RHS element per lane, we can always | |||
| 15891 | // perform the shuffle once the lanes have been shuffled in place. | |||
| 15892 | for (int i = 0; i != 4; ++i) { | |||
| 15893 | int M = Mask[i]; | |||
| 15894 | if (M < 0) | |||
| 15895 | continue; | |||
| 15896 | int LaneBase = i & ~1; | |||
| 15897 | auto &LaneMask = (i & 1) ? RHSMask : LHSMask; | |||
| 15898 | LaneMask[LaneBase + (M & 1)] = M; | |||
| 15899 | SHUFPMask |= (M & 1) << i; | |||
| 15900 | } | |||
| 15901 | ||||
| 15902 | SDValue LHS = DAG.getVectorShuffle(VT, DL, V1, V2, LHSMask); | |||
| 15903 | SDValue RHS = DAG.getVectorShuffle(VT, DL, V1, V2, RHSMask); | |||
| 15904 | return DAG.getNode(X86ISD::SHUFP, DL, VT, LHS, RHS, | |||
| 15905 | DAG.getTargetConstant(SHUFPMask, DL, MVT::i8)); | |||
| 15906 | } | |||
| 15907 | ||||
| 15908 | /// Lower a vector shuffle crossing multiple 128-bit lanes as | |||
| 15909 | /// a lane permutation followed by a per-lane permutation. | |||
| 15910 | /// | |||
| 15911 | /// This is mainly for cases where we can have non-repeating permutes | |||
| 15912 | /// in each lane. | |||
| 15913 | /// | |||
| 15914 | /// TODO: This is very similar to lowerShuffleAsLanePermuteAndRepeatedMask, | |||
| 15915 | /// we should investigate merging them. | |||
| 15916 | static SDValue lowerShuffleAsLanePermuteAndPermute( | |||
| 15917 | const SDLoc &DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask, | |||
| 15918 | SelectionDAG &DAG, const X86Subtarget &Subtarget) { | |||
| 15919 | int NumElts = VT.getVectorNumElements(); | |||
| 15920 | int NumLanes = VT.getSizeInBits() / 128; | |||
| 15921 | int NumEltsPerLane = NumElts / NumLanes; | |||
| 15922 | bool CanUseSublanes = Subtarget.hasAVX2() && V2.isUndef(); | |||
| 15923 | ||||
| 15924 | /// Attempts to find a sublane permute with the given size | |||
| 15925 | /// that gets all elements into their target lanes. | |||
| 15926 | /// | |||
| 15927 | /// If successful, fills CrossLaneMask and InLaneMask and returns true. | |||
| 15928 | /// If unsuccessful, returns false and may overwrite InLaneMask. | |||
| 15929 | auto getSublanePermute = [&](int NumSublanes) -> SDValue { | |||
| 15930 | int NumSublanesPerLane = NumSublanes / NumLanes; | |||
| 15931 | int NumEltsPerSublane = NumElts / NumSublanes; | |||
| 15932 | ||||
| 15933 | SmallVector<int, 16> CrossLaneMask; | |||
| 15934 | SmallVector<int, 16> InLaneMask(NumElts, SM_SentinelUndef); | |||
| 15935 | // CrossLaneMask but one entry == one sublane. | |||
| 15936 | SmallVector<int, 16> CrossLaneMaskLarge(NumSublanes, SM_SentinelUndef); | |||
| 15937 | ||||
| 15938 | for (int i = 0; i != NumElts; ++i) { | |||
| 15939 | int M = Mask[i]; | |||
| 15940 | if (M < 0) | |||
| 15941 | continue; | |||
| 15942 | ||||
| 15943 | int SrcSublane = M / NumEltsPerSublane; | |||
| 15944 | int DstLane = i / NumEltsPerLane; | |||
| 15945 | ||||
| 15946 | // We only need to get the elements into the right lane, not sublane. | |||
| 15947 | // So search all sublanes that make up the destination lane. | |||
| 15948 | bool Found = false; | |||
| 15949 | int DstSubStart = DstLane * NumSublanesPerLane; | |||
| 15950 | int DstSubEnd = DstSubStart + NumSublanesPerLane; | |||
| 15951 | for (int DstSublane = DstSubStart; DstSublane < DstSubEnd; ++DstSublane) { | |||
| 15952 | if (!isUndefOrEqual(CrossLaneMaskLarge[DstSublane], SrcSublane)) | |||
| 15953 | continue; | |||
| 15954 | ||||
| 15955 | Found = true; | |||
| 15956 | CrossLaneMaskLarge[DstSublane] = SrcSublane; | |||
| 15957 | int DstSublaneOffset = DstSublane * NumEltsPerSublane; | |||
| 15958 | InLaneMask[i] = DstSublaneOffset + M % NumEltsPerSublane; | |||
| 15959 | break; | |||
| 15960 | } | |||
| 15961 | if (!Found) | |||
| 15962 | return SDValue(); | |||
| 15963 | } | |||
| 15964 | ||||
| 15965 | // Fill CrossLaneMask using CrossLaneMaskLarge. | |||
| 15966 | narrowShuffleMaskElts(NumEltsPerSublane, CrossLaneMaskLarge, CrossLaneMask); | |||
| 15967 | ||||
| 15968 | if (!CanUseSublanes) { | |||
| 15969 | // If we're only shuffling a single lowest lane and the rest are identity | |||
| 15970 | // then don't bother. | |||
| 15971 | // TODO - isShuffleMaskInputInPlace could be extended to something like | |||
| 15972 | // this. | |||
| 15973 | int NumIdentityLanes = 0; | |||
| 15974 | bool OnlyShuffleLowestLane = true; | |||
| 15975 | for (int i = 0; i != NumLanes; ++i) { | |||
| 15976 | int LaneOffset = i * NumEltsPerLane; | |||
| 15977 | if (isSequentialOrUndefInRange(InLaneMask, LaneOffset, NumEltsPerLane, | |||
| 15978 | i * NumEltsPerLane)) | |||
| 15979 | NumIdentityLanes++; | |||
| 15980 | else if (CrossLaneMask[LaneOffset] != 0) | |||
| 15981 | OnlyShuffleLowestLane = false; | |||
| 15982 | } | |||
| 15983 | if (OnlyShuffleLowestLane && NumIdentityLanes == (NumLanes - 1)) | |||
| 15984 | return SDValue(); | |||
| 15985 | } | |||
| 15986 | ||||
| 15987 | SDValue CrossLane = DAG.getVectorShuffle(VT, DL, V1, V2, CrossLaneMask); | |||
| 15988 | return DAG.getVectorShuffle(VT, DL, CrossLane, DAG.getUNDEF(VT), | |||
| 15989 | InLaneMask); | |||
| 15990 | }; | |||
| 15991 | ||||
| 15992 | // First attempt a solution with full lanes. | |||
| 15993 | if (SDValue V = getSublanePermute(/*NumSublanes=*/NumLanes)) | |||
| 15994 | return V; | |||
| 15995 | ||||
| 15996 | // The rest of the solutions use sublanes. | |||
| 15997 | if (!CanUseSublanes) | |||
| 15998 | return SDValue(); | |||
| 15999 | ||||
| 16000 | // Then attempt a solution with 64-bit sublanes (vpermq). | |||
| 16001 | if (SDValue V = getSublanePermute(/*NumSublanes=*/NumLanes * 2)) | |||
| 16002 | return V; | |||
| 16003 | ||||
| 16004 | // If that doesn't work and we have fast variable cross-lane shuffle, | |||
| 16005 | // attempt 32-bit sublanes (vpermd). | |||
| 16006 | if (!Subtarget.hasFastVariableCrossLaneShuffle()) | |||
| 16007 | return SDValue(); | |||
| 16008 | ||||
| 16009 | return getSublanePermute(/*NumSublanes=*/NumLanes * 4); | |||
| 16010 | } | |||
| 16011 | ||||
| 16012 | /// Lower a vector shuffle crossing multiple 128-bit lanes by shuffling one | |||
| 16013 | /// source with a lane permutation. | |||
| 16014 | /// | |||
| 16015 | /// This lowering strategy results in four instructions in the worst case for a | |||
| 16016 | /// single-input cross lane shuffle which is lower than any other fully general | |||
| 16017 | /// cross-lane shuffle strategy I'm aware of. Special cases for each particular | |||
| 16018 | /// shuffle pattern should be handled prior to trying this lowering. | |||
| 16019 | static SDValue lowerShuffleAsLanePermuteAndShuffle( | |||
| 16020 | const SDLoc &DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask, | |||
| 16021 | SelectionDAG &DAG, const X86Subtarget &Subtarget) { | |||
| 16022 | // FIXME: This should probably be generalized for 512-bit vectors as well. | |||
| 16023 | assert(VT.is256BitVector() && "Only for 256-bit vector shuffles!")((void)0); | |||
| 16024 | int Size = Mask.size(); | |||
| 16025 | int LaneSize = Size / 2; | |||
| 16026 | ||||
| 16027 | // Fold to SHUFPD(VPERM2F128(V1, V2), VPERM2F128(V1, V2)). | |||
| 16028 | // Only do this if the elements aren't all from the lower lane, | |||
| 16029 | // otherwise we're (probably) better off doing a split. | |||
| 16030 | if (VT == MVT::v4f64 && | |||
| 16031 | !all_of(Mask, [LaneSize](int M) { return M < LaneSize; })) | |||
| 16032 | if (SDValue V = | |||
| 16033 | lowerShuffleAsLanePermuteAndSHUFP(DL, VT, V1, V2, Mask, DAG)) | |||
| 16034 | return V; | |||
| 16035 | ||||
| 16036 | // If there are only inputs from one 128-bit lane, splitting will in fact be | |||
| 16037 | // less expensive. The flags track whether the given lane contains an element | |||
| 16038 | // that crosses to another lane. | |||
| 16039 | if (!Subtarget.hasAVX2()) { | |||
| 16040 | bool LaneCrossing[2] = {false, false}; | |||
| 16041 | for (int i = 0; i < Size; ++i) | |||
| 16042 | if (Mask[i] >= 0 && ((Mask[i] % Size) / LaneSize) != (i / LaneSize)) | |||
| 16043 | LaneCrossing[(Mask[i] % Size) / LaneSize] = true; | |||
| 16044 | if (!LaneCrossing[0] || !LaneCrossing[1]) | |||
| 16045 | return splitAndLowerShuffle(DL, VT, V1, V2, Mask, DAG); | |||
| 16046 | } else { | |||
| 16047 | bool LaneUsed[2] = {false, false}; | |||
| 16048 | for (int i = 0; i < Size; ++i) | |||
| 16049 | if (Mask[i] >= 0) | |||
| 16050 | LaneUsed[(Mask[i] % Size) / LaneSize] = true; | |||
| 16051 | if (!LaneUsed[0] || !LaneUsed[1]) | |||
| 16052 | return splitAndLowerShuffle(DL, VT, V1, V2, Mask, DAG); | |||
| 16053 | } | |||
| 16054 | ||||
| 16055 | // TODO - we could support shuffling V2 in the Flipped input. | |||
| 16056 | assert(V2.isUndef() &&((void)0) | |||
| 16057 | "This last part of this routine only works on single input shuffles")((void)0); | |||
| 16058 | ||||
| 16059 | SmallVector<int, 32> InLaneMask(Mask.begin(), Mask.end()); | |||
| 16060 | for (int i = 0; i < Size; ++i) { | |||
| 16061 | int &M = InLaneMask[i]; | |||
| 16062 | if (M < 0) | |||
| 16063 | continue; | |||
| 16064 | if (((M % Size) / LaneSize) != (i / LaneSize)) | |||
| 16065 | M = (M % LaneSize) + ((i / LaneSize) * LaneSize) + Size; | |||
| 16066 | } | |||
| 16067 | assert(!is128BitLaneCrossingShuffleMask(VT, InLaneMask) &&((void)0) | |||
| 16068 | "In-lane shuffle mask expected")((void)0); | |||
| 16069 | ||||
| 16070 | // Flip the lanes, and shuffle the results which should now be in-lane. | |||
| 16071 | MVT PVT = VT.isFloatingPoint() ? MVT::v4f64 : MVT::v4i64; | |||
| 16072 | SDValue Flipped = DAG.getBitcast(PVT, V1); | |||
| 16073 | Flipped = | |||
| 16074 | DAG.getVectorShuffle(PVT, DL, Flipped, DAG.getUNDEF(PVT), {2, 3, 0, 1}); | |||
| 16075 | Flipped = DAG.getBitcast(VT, Flipped); | |||
| 16076 | return DAG.getVectorShuffle(VT, DL, V1, Flipped, InLaneMask); | |||
| 16077 | } | |||
| 16078 | ||||
| 16079 | /// Handle lowering 2-lane 128-bit shuffles. | |||
| 16080 | static SDValue lowerV2X128Shuffle(const SDLoc &DL, MVT VT, SDValue V1, | |||
| 16081 | SDValue V2, ArrayRef<int> Mask, | |||
| 16082 | const APInt &Zeroable, | |||
| 16083 | const X86Subtarget &Subtarget, | |||
| 16084 | SelectionDAG &DAG) { | |||
| 16085 | if (V2.isUndef()) { | |||
| 16086 | // Attempt to match VBROADCAST*128 subvector broadcast load. | |||
| 16087 | bool SplatLo = isShuffleEquivalent(Mask, {0, 1, 0, 1}, V1); | |||
| 16088 | bool SplatHi = isShuffleEquivalent(Mask, {2, 3, 2, 3}, V1); | |||
| 16089 | if ((SplatLo || SplatHi) && !Subtarget.hasAVX512() && V1.hasOneUse() && | |||
| 16090 | MayFoldLoad(peekThroughOneUseBitcasts(V1))) { | |||
| 16091 | auto *Ld = cast<LoadSDNode>(peekThroughOneUseBitcasts(V1)); | |||
| 16092 | if (!Ld->isNonTemporal()) { | |||
| 16093 | MVT MemVT = VT.getHalfNumVectorElementsVT(); | |||
| 16094 | unsigned Ofs = SplatLo ? 0 : MemVT.getStoreSize(); | |||
| 16095 | SDVTList Tys = DAG.getVTList(VT, MVT::Other); | |||
| 16096 | SDValue Ptr = DAG.getMemBasePlusOffset(Ld->getBasePtr(), | |||
| 16097 | TypeSize::Fixed(Ofs), DL); | |||
| 16098 | SDValue Ops[] = {Ld->getChain(), Ptr}; | |||
| 16099 | SDValue BcastLd = DAG.getMemIntrinsicNode( | |||
| 16100 | X86ISD::SUBV_BROADCAST_LOAD, DL, Tys, Ops, MemVT, | |||
| 16101 | DAG.getMachineFunction().getMachineMemOperand( | |||
| 16102 | Ld->getMemOperand(), Ofs, MemVT.getStoreSize())); | |||
| 16103 | DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), BcastLd.getValue(1)); | |||
| 16104 | return BcastLd; | |||
| 16105 | } | |||
| 16106 | } | |||
| 16107 | ||||
| 16108 | // With AVX2, use VPERMQ/VPERMPD for unary shuffles to allow memory folding. | |||
| 16109 | if (Subtarget.hasAVX2()) | |||
| 16110 | return SDValue(); | |||
| 16111 | } | |||
| 16112 | ||||
| 16113 | bool V2IsZero = !V2.isUndef() && ISD::isBuildVectorAllZeros(V2.getNode()); | |||
| 16114 | ||||
| 16115 | SmallVector<int, 4> WidenedMask; | |||
| 16116 | if (!canWidenShuffleElements(Mask, Zeroable, V2IsZero, WidenedMask)) | |||
| 16117 | return SDValue(); | |||
| 16118 | ||||
| 16119 | bool IsLowZero = (Zeroable & 0x3) == 0x3; | |||
| 16120 | bool IsHighZero = (Zeroable & 0xc) == 0xc; | |||
| 16121 | ||||
| 16122 | // Try to use an insert into a zero vector. | |||
| 16123 | if (WidenedMask[0] == 0 && IsHighZero) { | |||
| 16124 | MVT SubVT = MVT::getVectorVT(VT.getVectorElementType(), 2); | |||
| 16125 | SDValue LoV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT, V1, | |||
| 16126 | DAG.getIntPtrConstant(0, DL)); | |||
| 16127 | return DAG.getNode(ISD::INSERT_SUBVECTOR, DL, VT, | |||
| 16128 | getZeroVector(VT, Subtarget, DAG, DL), LoV, | |||
| 16129 | DAG.getIntPtrConstant(0, DL)); | |||
| 16130 | } | |||
| 16131 | ||||
| 16132 | // TODO: If minimizing size and one of the inputs is a zero vector and the | |||
| 16133 | // the zero vector has only one use, we could use a VPERM2X128 to save the | |||
| 16134 | // instruction bytes needed to explicitly generate the zero vector. | |||
| 16135 | ||||
| 16136 | // Blends are faster and handle all the non-lane-crossing cases. | |||
| 16137 | if (SDValue Blend = lowerShuffleAsBlend(DL, VT, V1, V2, Mask, Zeroable, | |||
| 16138 | Subtarget, DAG)) | |||
| 16139 | return Blend; | |||
| 16140 | ||||
| 16141 | // If either input operand is a zero vector, use VPERM2X128 because its mask | |||
| 16142 | // allows us to replace the zero input with an implicit zero. | |||
| 16143 | if (!IsLowZero && !IsHighZero) { | |||
| 16144 | // Check for patterns which can be matched with a single insert of a 128-bit | |||
| 16145 | // subvector. | |||
| 16146 | bool OnlyUsesV1 = isShuffleEquivalent(Mask, {0, 1, 0, 1}, V1, V2); | |||
| 16147 | if (OnlyUsesV1 || isShuffleEquivalent(Mask, {0, 1, 4, 5}, V1, V2)) { | |||
| 16148 | ||||
| 16149 | // With AVX1, use vperm2f128 (below) to allow load folding. Otherwise, | |||
| 16150 | // this will likely become vinsertf128 which can't fold a 256-bit memop. | |||
| 16151 | if (!isa<LoadSDNode>(peekThroughBitcasts(V1))) { | |||
| 16152 | MVT SubVT = MVT::getVectorVT(VT.getVectorElementType(), 2); | |||
| 16153 | SDValue SubVec = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT, | |||
| 16154 | OnlyUsesV1 ? V1 : V2, | |||
| 16155 | DAG.getIntPtrConstant(0, DL)); | |||
| 16156 | return DAG.getNode(ISD::INSERT_SUBVECTOR, DL, VT, V1, SubVec, | |||
| 16157 | DAG.getIntPtrConstant(2, DL)); | |||
| 16158 | } | |||
| 16159 | } | |||
| 16160 | ||||
| 16161 | // Try to use SHUF128 if possible. | |||
| 16162 | if (Subtarget.hasVLX()) { | |||
| 16163 | if (WidenedMask[0] < 2 && WidenedMask[1] >= 2) { | |||
| 16164 | unsigned PermMask = ((WidenedMask[0] % 2) << 0) | | |||
| 16165 | ((WidenedMask[1] % 2) << 1); | |||
| 16166 | return DAG.getNode(X86ISD::SHUF128, DL, VT, V1, V2, | |||
| 16167 | DAG.getTargetConstant(PermMask, DL, MVT::i8)); | |||
| 16168 | } | |||
| 16169 | } | |||
| 16170 | } | |||
| 16171 | ||||
| 16172 | // Otherwise form a 128-bit permutation. After accounting for undefs, | |||
| 16173 | // convert the 64-bit shuffle mask selection values into 128-bit | |||
| 16174 | // selection bits by dividing the indexes by 2 and shifting into positions | |||
| 16175 | // defined by a vperm2*128 instruction's immediate control byte. | |||
| 16176 | ||||
| 16177 | // The immediate permute control byte looks like this: | |||
| 16178 | // [1:0] - select 128 bits from sources for low half of destination | |||
| 16179 | // [2] - ignore | |||
| 16180 | // [3] - zero low half of destination | |||
| 16181 | // [5:4] - select 128 bits from sources for high half of destination | |||
| 16182 | // [6] - ignore | |||
| 16183 | // [7] - zero high half of destination | |||
| 16184 | ||||
| 16185 | assert((WidenedMask[0] >= 0 || IsLowZero) &&((void)0) | |||
| 16186 | (WidenedMask[1] >= 0 || IsHighZero) && "Undef half?")((void)0); | |||
| 16187 | ||||
| 16188 | unsigned PermMask = 0; | |||
| 16189 | PermMask |= IsLowZero ? 0x08 : (WidenedMask[0] << 0); | |||
| 16190 | PermMask |= IsHighZero ? 0x80 : (WidenedMask[1] << 4); | |||
| 16191 | ||||
| 16192 | // Check the immediate mask and replace unused sources with undef. | |||
| 16193 | if ((PermMask & 0x0a) != 0x00 && (PermMask & 0xa0) != 0x00) | |||
| 16194 | V1 = DAG.getUNDEF(VT); | |||
| 16195 | if ((PermMask & 0x0a) != 0x02 && (PermMask & 0xa0) != 0x20) | |||
| 16196 | V2 = DAG.getUNDEF(VT); | |||
| 16197 | ||||
| 16198 | return DAG.getNode(X86ISD::VPERM2X128, DL, VT, V1, V2, | |||
| 16199 | DAG.getTargetConstant(PermMask, DL, MVT::i8)); | |||
| 16200 | } | |||
| 16201 | ||||
| 16202 | /// Lower a vector shuffle by first fixing the 128-bit lanes and then | |||
| 16203 | /// shuffling each lane. | |||
| 16204 | /// | |||
| 16205 | /// This attempts to create a repeated lane shuffle where each lane uses one | |||
| 16206 | /// or two of the lanes of the inputs. The lanes of the input vectors are | |||
| 16207 | /// shuffled in one or two independent shuffles to get the lanes into the | |||
| 16208 | /// position needed by the final shuffle. | |||
| 16209 | static SDValue lowerShuffleAsLanePermuteAndRepeatedMask( | |||
| 16210 | const SDLoc &DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask, | |||
| 16211 | const X86Subtarget &Subtarget, SelectionDAG &DAG) { | |||
| 16212 | assert(!V2.isUndef() && "This is only useful with multiple inputs.")((void)0); | |||
| 16213 | ||||
| 16214 | if (is128BitLaneRepeatedShuffleMask(VT, Mask)) | |||
| 16215 | return SDValue(); | |||
| 16216 | ||||
| 16217 | int NumElts = Mask.size(); | |||
| 16218 | int NumLanes = VT.getSizeInBits() / 128; | |||
| 16219 | int NumLaneElts = 128 / VT.getScalarSizeInBits(); | |||
| 16220 | SmallVector<int, 16> RepeatMask(NumLaneElts, -1); | |||
| 16221 | SmallVector<std::array<int, 2>, 2> LaneSrcs(NumLanes, {{-1, -1}}); | |||
| 16222 | ||||
| 16223 | // First pass will try to fill in the RepeatMask from lanes that need two | |||
| 16224 | // sources. | |||
| 16225 | for (int Lane = 0; Lane != NumLanes; ++Lane) { | |||
| 16226 | int Srcs[2] = {-1, -1}; | |||
| 16227 | SmallVector<int, 16> InLaneMask(NumLaneElts, -1); | |||
| 16228 | for (int i = 0; i != NumLaneElts; ++i) { | |||
| 16229 | int M = Mask[(Lane * NumLaneElts) + i]; | |||
| 16230 | if (M < 0) | |||
| 16231 | continue; | |||
| 16232 | // Determine which of the possible input lanes (NumLanes from each source) | |||
| 16233 | // this element comes from. Assign that as one of the sources for this | |||
| 16234 | // lane. We can assign up to 2 sources for this lane. If we run out | |||
| 16235 | // sources we can't do anything. | |||
| 16236 | int LaneSrc = M / NumLaneElts; | |||
| 16237 | int Src; | |||
| 16238 | if (Srcs[0] < 0 || Srcs[0] == LaneSrc) | |||
| 16239 | Src = 0; | |||
| 16240 | else if (Srcs[1] < 0 || Srcs[1] == LaneSrc) | |||
| 16241 | Src = 1; | |||
| 16242 | else | |||
| 16243 | return SDValue(); | |||
| 16244 | ||||
| 16245 | Srcs[Src] = LaneSrc; | |||
| 16246 | InLaneMask[i] = (M % NumLaneElts) + Src * NumElts; | |||
| 16247 | } | |||
| 16248 | ||||
| 16249 | // If this lane has two sources, see if it fits with the repeat mask so far. | |||
| 16250 | if (Srcs[1] < 0) | |||
| 16251 | continue; | |||
| 16252 | ||||
| 16253 | LaneSrcs[Lane][0] = Srcs[0]; | |||
| 16254 | LaneSrcs[Lane][1] = Srcs[1]; | |||
| 16255 | ||||
| 16256 | auto MatchMasks = [](ArrayRef<int> M1, ArrayRef<int> M2) { | |||
| 16257 | assert(M1.size() == M2.size() && "Unexpected mask size")((void)0); | |||
| 16258 | for (int i = 0, e = M1.size(); i != e; ++i) | |||
| 16259 | if (M1[i] >= 0 && M2[i] >= 0 && M1[i] != M2[i]) | |||
| 16260 | return false; | |||
| 16261 | return true; | |||
| 16262 | }; | |||
| 16263 | ||||
| 16264 | auto MergeMasks = [](ArrayRef<int> Mask, MutableArrayRef<int> MergedMask) { | |||
| 16265 | assert(Mask.size() == MergedMask.size() && "Unexpected mask size")((void)0); | |||
| 16266 | for (int i = 0, e = MergedMask.size(); i != e; ++i) { | |||
| 16267 | int M = Mask[i]; | |||
| 16268 | if (M < 0) | |||
| 16269 | continue; | |||
| 16270 | assert((MergedMask[i] < 0 || MergedMask[i] == M) &&((void)0) | |||
| 16271 | "Unexpected mask element")((void)0); | |||
| 16272 | MergedMask[i] = M; | |||
| 16273 | } | |||
| 16274 | }; | |||
| 16275 | ||||
| 16276 | if (MatchMasks(InLaneMask, RepeatMask)) { | |||
| 16277 | // Merge this lane mask into the final repeat mask. | |||
| 16278 | MergeMasks(InLaneMask, RepeatMask); | |||
| 16279 | continue; | |||
| 16280 | } | |||
| 16281 | ||||
| 16282 | // Didn't find a match. Swap the operands and try again. | |||
| 16283 | std::swap(LaneSrcs[Lane][0], LaneSrcs[Lane][1]); | |||
| 16284 | ShuffleVectorSDNode::commuteMask(InLaneMask); | |||
| 16285 | ||||
| 16286 | if (MatchMasks(InLaneMask, RepeatMask)) { | |||
| 16287 | // Merge this lane mask into the final repeat mask. | |||
| 16288 | MergeMasks(InLaneMask, RepeatMask); | |||
| 16289 | continue; | |||
| 16290 | } | |||
| 16291 | ||||
| 16292 | // Couldn't find a match with the operands in either order. | |||
| 16293 | return SDValue(); | |||
| 16294 | } | |||
| 16295 | ||||
| 16296 | // Now handle any lanes with only one source. | |||
| 16297 | for (int Lane = 0; Lane != NumLanes; ++Lane) { | |||
| 16298 | // If this lane has already been processed, skip it. | |||
| 16299 | if (LaneSrcs[Lane][0] >= 0) | |||
| 16300 | continue; | |||
| 16301 | ||||
| 16302 | for (int i = 0; i != NumLaneElts; ++i) { | |||
| 16303 | int M = Mask[(Lane * NumLaneElts) + i]; | |||
| 16304 | if (M < 0) | |||
| 16305 | continue; | |||
| 16306 | ||||
| 16307 | // If RepeatMask isn't defined yet we can define it ourself. | |||
| 16308 | if (RepeatMask[i] < 0) | |||
| 16309 | RepeatMask[i] = M % NumLaneElts; | |||
| 16310 | ||||
| 16311 | if (RepeatMask[i] < NumElts) { | |||
| 16312 | if (RepeatMask[i] != M % NumLaneElts) | |||
| 16313 | return SDValue(); | |||
| 16314 | LaneSrcs[Lane][0] = M / NumLaneElts; | |||
| 16315 | } else { | |||
| 16316 | if (RepeatMask[i] != ((M % NumLaneElts) + NumElts)) | |||
| 16317 | return SDValue(); | |||
| 16318 | LaneSrcs[Lane][1] = M / NumLaneElts; | |||
| 16319 | } | |||
| 16320 | } | |||
| 16321 | ||||
| 16322 | if (LaneSrcs[Lane][0] < 0 && LaneSrcs[Lane][1] < 0) | |||
| 16323 | return SDValue(); | |||
| 16324 | } | |||
| 16325 | ||||
| 16326 | SmallVector<int, 16> NewMask(NumElts, -1); | |||
| 16327 | for (int Lane = 0; Lane != NumLanes; ++Lane) { | |||
| 16328 | int Src = LaneSrcs[Lane][0]; | |||
| 16329 | for (int i = 0; i != NumLaneElts; ++i) { | |||
| 16330 | int M = -1; | |||
| 16331 | if (Src >= 0) | |||
| 16332 | M = Src * NumLaneElts + i; | |||
| 16333 | NewMask[Lane * NumLaneElts + i] = M; | |||
| 16334 | } | |||
| 16335 | } | |||
| 16336 | SDValue NewV1 = DAG.getVectorShuffle(VT, DL, V1, V2, NewMask); | |||
| 16337 | // Ensure we didn't get back the shuffle we started with. | |||
| 16338 | // FIXME: This is a hack to make up for some splat handling code in | |||
| 16339 | // getVectorShuffle. | |||
| 16340 | if (isa<ShuffleVectorSDNode>(NewV1) && | |||
| 16341 | cast<ShuffleVectorSDNode>(NewV1)->getMask() == Mask) | |||
| 16342 | return SDValue(); | |||
| 16343 | ||||
| 16344 | for (int Lane = 0; Lane != NumLanes; ++Lane) { | |||
| 16345 | int Src = LaneSrcs[Lane][1]; | |||
| 16346 | for (int i = 0; i != NumLaneElts; ++i) { | |||
| 16347 | int M = -1; | |||
| 16348 | if (Src >= 0) | |||
| 16349 | M = Src * NumLaneElts + i; | |||
| 16350 | NewMask[Lane * NumLaneElts + i] = M; | |||
| 16351 | } | |||
| 16352 | } | |||
| 16353 | SDValue NewV2 = DAG.getVectorShuffle(VT, DL, V1, V2, NewMask); | |||
| 16354 | // Ensure we didn't get back the shuffle we started with. | |||
| 16355 | // FIXME: This is a hack to make up for some splat handling code in | |||
| 16356 | // getVectorShuffle. | |||
| 16357 | if (isa<ShuffleVectorSDNode>(NewV2) && | |||
| 16358 | cast<ShuffleVectorSDNode>(NewV2)->getMask() == Mask) | |||
| 16359 | return SDValue(); | |||
| 16360 | ||||
| 16361 | for (int i = 0; i != NumElts; ++i) { | |||
| 16362 | NewMask[i] = RepeatMask[i % NumLaneElts]; | |||
| 16363 | if (NewMask[i] < 0) | |||
| 16364 | continue; | |||
| 16365 | ||||
| 16366 | NewMask[i] += (i / NumLaneElts) * NumLaneElts; | |||
| 16367 | } | |||
| 16368 | return DAG.getVectorShuffle(VT, DL, NewV1, NewV2, NewMask); | |||
| 16369 | } | |||
| 16370 | ||||
| 16371 | /// If the input shuffle mask results in a vector that is undefined in all upper | |||
| 16372 | /// or lower half elements and that mask accesses only 2 halves of the | |||
| 16373 | /// shuffle's operands, return true. A mask of half the width with mask indexes | |||
| 16374 | /// adjusted to access the extracted halves of the original shuffle operands is | |||
| 16375 | /// returned in HalfMask. HalfIdx1 and HalfIdx2 return whether the upper or | |||
| 16376 | /// lower half of each input operand is accessed. | |||
| 16377 | static bool | |||
| 16378 | getHalfShuffleMask(ArrayRef<int> Mask, MutableArrayRef<int> HalfMask, | |||
| 16379 | int &HalfIdx1, int &HalfIdx2) { | |||
| 16380 | assert((Mask.size() == HalfMask.size() * 2) &&((void)0) | |||
| 16381 | "Expected input mask to be twice as long as output")((void)0); | |||
| 16382 | ||||
| 16383 | // Exactly one half of the result must be undef to allow narrowing. | |||
| 16384 | bool UndefLower = isUndefLowerHalf(Mask); | |||
| 16385 | bool UndefUpper = isUndefUpperHalf(Mask); | |||
| 16386 | if (UndefLower == UndefUpper) | |||
| 16387 | return false; | |||
| 16388 | ||||
| 16389 | unsigned HalfNumElts = HalfMask.size(); | |||
| 16390 | unsigned MaskIndexOffset = UndefLower ? HalfNumElts : 0; | |||
| 16391 | HalfIdx1 = -1; | |||
| 16392 | HalfIdx2 = -1; | |||
| 16393 | for (unsigned i = 0; i != HalfNumElts; ++i) { | |||
| 16394 | int M = Mask[i + MaskIndexOffset]; | |||
| 16395 | if (M < 0) { | |||
| 16396 | HalfMask[i] = M; | |||
| 16397 | continue; | |||
| 16398 | } | |||
| 16399 | ||||
| 16400 | // Determine which of the 4 half vectors this element is from. | |||
| 16401 | // i.e. 0 = Lower V1, 1 = Upper V1, 2 = Lower V2, 3 = Upper V2. | |||
| 16402 | int HalfIdx = M / HalfNumElts; | |||
| 16403 | ||||
| 16404 | // Determine the element index into its half vector source. | |||
| 16405 | int HalfElt = M % HalfNumElts; | |||
| 16406 | ||||
| 16407 | // We can shuffle with up to 2 half vectors, set the new 'half' | |||
| 16408 | // shuffle mask accordingly. | |||
| 16409 | if (HalfIdx1 < 0 || HalfIdx1 == HalfIdx) { | |||
| 16410 | HalfMask[i] = HalfElt; | |||
| 16411 | HalfIdx1 = HalfIdx; | |||
| 16412 | continue; | |||
| 16413 | } | |||
| 16414 | if (HalfIdx2 < 0 || HalfIdx2 == HalfIdx) { | |||
| 16415 | HalfMask[i] = HalfElt + HalfNumElts; | |||
| 16416 | HalfIdx2 = HalfIdx; | |||
| 16417 | continue; | |||
| 16418 | } | |||
| 16419 | ||||
| 16420 | // Too many half vectors referenced. | |||
| 16421 | return false; | |||
| 16422 | } | |||
| 16423 | ||||
| 16424 | return true; | |||
| 16425 | } | |||
| 16426 | ||||
| 16427 | /// Given the output values from getHalfShuffleMask(), create a half width | |||
| 16428 | /// shuffle of extracted vectors followed by an insert back to full width. | |||
| 16429 | static SDValue getShuffleHalfVectors(const SDLoc &DL, SDValue V1, SDValue V2, | |||
| 16430 | ArrayRef<int> HalfMask, int HalfIdx1, | |||
| 16431 | int HalfIdx2, bool UndefLower, | |||
| 16432 | SelectionDAG &DAG, bool UseConcat = false) { | |||
| 16433 | assert(V1.getValueType() == V2.getValueType() && "Different sized vectors?")((void)0); | |||
| 16434 | assert(V1.getValueType().isSimple() && "Expecting only simple types")((void)0); | |||
| 16435 | ||||
| 16436 | MVT VT = V1.getSimpleValueType(); | |||
| 16437 | MVT HalfVT = VT.getHalfNumVectorElementsVT(); | |||
| 16438 | unsigned HalfNumElts = HalfVT.getVectorNumElements(); | |||
| 16439 | ||||
| 16440 | auto getHalfVector = [&](int HalfIdx) { | |||
| 16441 | if (HalfIdx < 0) | |||
| 16442 | return DAG.getUNDEF(HalfVT); | |||
| 16443 | SDValue V = (HalfIdx < 2 ? V1 : V2); | |||
| 16444 | HalfIdx = (HalfIdx % 2) * HalfNumElts; | |||
| 16445 | return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, HalfVT, V, | |||
| 16446 | DAG.getIntPtrConstant(HalfIdx, DL)); | |||
| 16447 | }; | |||
| 16448 | ||||
| 16449 | // ins undef, (shuf (ext V1, HalfIdx1), (ext V2, HalfIdx2), HalfMask), Offset | |||
| 16450 | SDValue Half1 = getHalfVector(HalfIdx1); | |||
| 16451 | SDValue Half2 = getHalfVector(HalfIdx2); | |||
| 16452 | SDValue V = DAG.getVectorShuffle(HalfVT, DL, Half1, Half2, HalfMask); | |||
| 16453 | if (UseConcat) { | |||
| 16454 | SDValue Op0 = V; | |||
| 16455 | SDValue Op1 = DAG.getUNDEF(HalfVT); | |||
| 16456 | if (UndefLower) | |||
| 16457 | std::swap(Op0, Op1); | |||
| 16458 | return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Op0, Op1); | |||
| 16459 | } | |||
| 16460 | ||||
| 16461 | unsigned Offset = UndefLower ? HalfNumElts : 0; | |||
| 16462 | return DAG.getNode(ISD::INSERT_SUBVECTOR, DL, VT, DAG.getUNDEF(VT), V, | |||
| 16463 | DAG.getIntPtrConstant(Offset, DL)); | |||
| 16464 | } | |||
| 16465 | ||||
| 16466 | /// Lower shuffles where an entire half of a 256 or 512-bit vector is UNDEF. | |||
| 16467 | /// This allows for fast cases such as subvector extraction/insertion | |||
| 16468 | /// or shuffling smaller vector types which can lower more efficiently. | |||
| 16469 | static SDValue lowerShuffleWithUndefHalf(const SDLoc &DL, MVT VT, SDValue V1, | |||
| 16470 | SDValue V2, ArrayRef<int> Mask, | |||
| 16471 | const X86Subtarget &Subtarget, | |||
| 16472 | SelectionDAG &DAG) { | |||
| 16473 | assert((VT.is256BitVector() || VT.is512BitVector()) &&((void)0) | |||
| 16474 | "Expected 256-bit or 512-bit vector")((void)0); | |||
| 16475 | ||||
| 16476 | bool UndefLower = isUndefLowerHalf(Mask); | |||
| 16477 | if (!UndefLower && !isUndefUpperHalf(Mask)) | |||
| 16478 | return SDValue(); | |||
| 16479 | ||||
| 16480 | assert((!UndefLower || !isUndefUpperHalf(Mask)) &&((void)0) | |||
| 16481 | "Completely undef shuffle mask should have been simplified already")((void)0); | |||
| 16482 | ||||
| 16483 | // Upper half is undef and lower half is whole upper subvector. | |||
| 16484 | // e.g. vector_shuffle <4, 5, 6, 7, u, u, u, u> or <2, 3, u, u> | |||
| 16485 | MVT HalfVT = VT.getHalfNumVectorElementsVT(); | |||
| 16486 | unsigned HalfNumElts = HalfVT.getVectorNumElements(); | |||
| 16487 | if (!UndefLower && | |||
| 16488 | isSequentialOrUndefInRange(Mask, 0, HalfNumElts, HalfNumElts)) { | |||
| 16489 | SDValue Hi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, HalfVT, V1, | |||
| 16490 | DAG.getIntPtrConstant(HalfNumElts, DL)); | |||
| 16491 | return DAG.getNode(ISD::INSERT_SUBVECTOR, DL, VT, DAG.getUNDEF(VT), Hi, | |||
| 16492 | DAG.getIntPtrConstant(0, DL)); | |||
| 16493 | } | |||
| 16494 | ||||
| 16495 | // Lower half is undef and upper half is whole lower subvector. | |||
| 16496 | // e.g. vector_shuffle <u, u, u, u, 0, 1, 2, 3> or <u, u, 0, 1> | |||
| 16497 | if (UndefLower && | |||
| 16498 | isSequentialOrUndefInRange(Mask, HalfNumElts, HalfNumElts, 0)) { | |||
| 16499 | SDValue Hi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, HalfVT, V1, | |||
| 16500 | DAG.getIntPtrConstant(0, DL)); | |||
| 16501 | return DAG.getNode(ISD::INSERT_SUBVECTOR, DL, VT, DAG.getUNDEF(VT), Hi, | |||
| 16502 | DAG.getIntPtrConstant(HalfNumElts, DL)); | |||
| 16503 | } | |||
| 16504 | ||||
| 16505 | int HalfIdx1, HalfIdx2; | |||
| 16506 | SmallVector<int, 8> HalfMask(HalfNumElts); | |||
| 16507 | if (!getHalfShuffleMask(Mask, HalfMask, HalfIdx1, HalfIdx2)) | |||
| 16508 | return SDValue(); | |||
| 16509 | ||||
| 16510 | assert(HalfMask.size() == HalfNumElts && "Unexpected shuffle mask length")((void)0); | |||
| 16511 | ||||
| 16512 | // Only shuffle the halves of the inputs when useful. | |||
| 16513 | unsigned NumLowerHalves = | |||
| 16514 | (HalfIdx1 == 0 || HalfIdx1 == 2) + (HalfIdx2 == 0 || HalfIdx2 == 2); | |||
| 16515 | unsigned NumUpperHalves = | |||
| 16516 | (HalfIdx1 == 1 || HalfIdx1 == 3) + (HalfIdx2 == 1 || HalfIdx2 == 3); | |||
| 16517 | assert(NumLowerHalves + NumUpperHalves <= 2 && "Only 1 or 2 halves allowed")((void)0); | |||
| 16518 | ||||
| 16519 | // Determine the larger pattern of undef/halves, then decide if it's worth | |||
| 16520 | // splitting the shuffle based on subtarget capabilities and types. | |||
| 16521 | unsigned EltWidth = VT.getVectorElementType().getSizeInBits(); | |||
| 16522 | if (!UndefLower) { | |||
| 16523 | // XXXXuuuu: no insert is needed. | |||
| 16524 | // Always extract lowers when setting lower - these are all free subreg ops. | |||
| 16525 | if (NumUpperHalves == 0) | |||
| 16526 | return getShuffleHalfVectors(DL, V1, V2, HalfMask, HalfIdx1, HalfIdx2, | |||
| 16527 | UndefLower, DAG); | |||
| 16528 | ||||
| 16529 | if (NumUpperHalves == 1) { | |||
| 16530 | // AVX2 has efficient 32/64-bit element cross-lane shuffles. | |||
| 16531 | if (Subtarget.hasAVX2()) { | |||
| 16532 | // extract128 + vunpckhps/vshufps, is better than vblend + vpermps. | |||
| 16533 | if (EltWidth == 32 && NumLowerHalves && HalfVT.is128BitVector() && | |||
| 16534 | !is128BitUnpackShuffleMask(HalfMask) && | |||
| 16535 | (!isSingleSHUFPSMask(HalfMask) || | |||
| 16536 | Subtarget.hasFastVariableCrossLaneShuffle())) | |||
| 16537 | return SDValue(); | |||
| 16538 | // If this is a unary shuffle (assume that the 2nd operand is | |||
| 16539 | // canonicalized to undef), then we can use vpermpd. Otherwise, we | |||
| 16540 | // are better off extracting the upper half of 1 operand and using a | |||
| 16541 | // narrow shuffle. | |||
| 16542 | if (EltWidth == 64 && V2.isUndef()) | |||
| 16543 | return SDValue(); | |||
| 16544 | } | |||
| 16545 | // AVX512 has efficient cross-lane shuffles for all legal 512-bit types. | |||
| 16546 | if (Subtarget.hasAVX512() && VT.is512BitVector()) | |||
| 16547 | return SDValue(); | |||
| 16548 | // Extract + narrow shuffle is better than the wide alternative. | |||
| 16549 | return getShuffleHalfVectors(DL, V1, V2, HalfMask, HalfIdx1, HalfIdx2, | |||
| 16550 | UndefLower, DAG); | |||
| 16551 | } | |||
| 16552 | ||||
| 16553 | // Don't extract both uppers, instead shuffle and then extract. | |||
| 16554 | assert(NumUpperHalves == 2 && "Half vector count went wrong")((void)0); | |||
| 16555 | return SDValue(); | |||
| 16556 | } | |||
| 16557 | ||||
| 16558 | // UndefLower - uuuuXXXX: an insert to high half is required if we split this. | |||
| 16559 | if (NumUpperHalves == 0) { | |||
| 16560 | // AVX2 has efficient 64-bit element cross-lane shuffles. | |||
| 16561 | // TODO: Refine to account for unary shuffle, splat, and other masks? | |||
| 16562 | if (Subtarget.hasAVX2() && EltWidth == 64) | |||
| 16563 | return SDValue(); | |||
| 16564 | // AVX512 has efficient cross-lane shuffles for all legal 512-bit types. | |||
| 16565 | if (Subtarget.hasAVX512() && VT.is512BitVector()) | |||
| 16566 | return SDValue(); | |||
| 16567 | // Narrow shuffle + insert is better than the wide alternative. | |||
| 16568 | return getShuffleHalfVectors(DL, V1, V2, HalfMask, HalfIdx1, HalfIdx2, | |||
| 16569 | UndefLower, DAG); | |||
| 16570 | } | |||
| 16571 | ||||
| 16572 | // NumUpperHalves != 0: don't bother with extract, shuffle, and then insert. | |||
| 16573 | return SDValue(); | |||
| 16574 | } | |||
| 16575 | ||||
| 16576 | /// Test whether the specified input (0 or 1) is in-place blended by the | |||
| 16577 | /// given mask. | |||
| 16578 | /// | |||
| 16579 | /// This returns true if the elements from a particular input are already in the | |||
| 16580 | /// slot required by the given mask and require no permutation. | |||
| 16581 | static bool isShuffleMaskInputInPlace(int Input, ArrayRef<int> Mask) { | |||
| 16582 | assert((Input == 0 || Input == 1) && "Only two inputs to shuffles.")((void)0); | |||
| 16583 | int Size = Mask.size(); | |||
| 16584 | for (int i = 0; i < Size; ++i) | |||
| 16585 | if (Mask[i] >= 0 && Mask[i] / Size == Input && Mask[i] % Size != i) | |||
| 16586 | return false; | |||
| 16587 | ||||
| 16588 | return true; | |||
| 16589 | } | |||
| 16590 | ||||
| 16591 | /// Handle case where shuffle sources are coming from the same 128-bit lane and | |||
| 16592 | /// every lane can be represented as the same repeating mask - allowing us to | |||
| 16593 | /// shuffle the sources with the repeating shuffle and then permute the result | |||
| 16594 | /// to the destination lanes. | |||
| 16595 | static SDValue lowerShuffleAsRepeatedMaskAndLanePermute( | |||
| 16596 | const SDLoc &DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask, | |||
| 16597 | const X86Subtarget &Subtarget, SelectionDAG &DAG) { | |||
| 16598 | int NumElts = VT.getVectorNumElements(); | |||
| 16599 | int NumLanes = VT.getSizeInBits() / 128; | |||
| 16600 | int NumLaneElts = NumElts / NumLanes; | |||
| 16601 | ||||
| 16602 | // On AVX2 we may be able to just shuffle the lowest elements and then | |||
| 16603 | // broadcast the result. | |||
| 16604 | if (Subtarget.hasAVX2()) { | |||
| 16605 | for (unsigned BroadcastSize : {16, 32, 64}) { | |||
| 16606 | if (BroadcastSize <= VT.getScalarSizeInBits()) | |||
| 16607 | continue; | |||
| 16608 | int NumBroadcastElts = BroadcastSize / VT.getScalarSizeInBits(); | |||
| 16609 | ||||
| 16610 | // Attempt to match a repeating pattern every NumBroadcastElts, | |||
| 16611 | // accounting for UNDEFs but only references the lowest 128-bit | |||
| 16612 | // lane of the inputs. | |||
| 16613 | auto FindRepeatingBroadcastMask = [&](SmallVectorImpl<int> &RepeatMask) { | |||
| 16614 | for (int i = 0; i != NumElts; i += NumBroadcastElts) | |||
| 16615 | for (int j = 0; j != NumBroadcastElts; ++j) { | |||
| 16616 | int M = Mask[i + j]; | |||
| 16617 | if (M < 0) | |||
| 16618 | continue; | |||
| 16619 | int &R = RepeatMask[j]; | |||
| 16620 | if (0 != ((M % NumElts) / NumLaneElts)) | |||
| 16621 | return false; | |||
| 16622 | if (0 <= R && R != M) | |||
| 16623 | return false; | |||
| 16624 | R = M; | |||
| 16625 | } | |||
| 16626 | return true; | |||
| 16627 | }; | |||
| 16628 | ||||
| 16629 | SmallVector<int, 8> RepeatMask((unsigned)NumElts, -1); | |||
| 16630 | if (!FindRepeatingBroadcastMask(RepeatMask)) | |||
| 16631 | continue; | |||
| 16632 | ||||
| 16633 | // Shuffle the (lowest) repeated elements in place for broadcast. | |||
| 16634 | SDValue RepeatShuf = DAG.getVectorShuffle(VT, DL, V1, V2, RepeatMask); | |||
| 16635 | ||||
| 16636 | // Shuffle the actual broadcast. | |||
| 16637 | SmallVector<int, 8> BroadcastMask((unsigned)NumElts, -1); | |||
| 16638 | for (int i = 0; i != NumElts; i += NumBroadcastElts) | |||
| 16639 | for (int j = 0; j != NumBroadcastElts; ++j) | |||
| 16640 | BroadcastMask[i + j] = j; | |||
| 16641 | return DAG.getVectorShuffle(VT, DL, RepeatShuf, DAG.getUNDEF(VT), | |||
| 16642 | BroadcastMask); | |||
| 16643 | } | |||
| 16644 | } | |||
| 16645 | ||||
| 16646 | // Bail if the shuffle mask doesn't cross 128-bit lanes. | |||
| 16647 | if (!is128BitLaneCrossingShuffleMask(VT, Mask)) | |||
| 16648 | return SDValue(); | |||
| 16649 | ||||
| 16650 | // Bail if we already have a repeated lane shuffle mask. | |||
| 16651 | SmallVector<int, 8> RepeatedShuffleMask; | |||
| 16652 | if (is128BitLaneRepeatedShuffleMask(VT, Mask, RepeatedShuffleMask)) | |||
| 16653 | return SDValue(); | |||
| 16654 | ||||
| 16655 | // On AVX2 targets we can permute 256-bit vectors as 64-bit sub-lanes | |||
| 16656 | // (with PERMQ/PERMPD), otherwise we can only permute whole 128-bit lanes. | |||
| 16657 | int SubLaneScale = Subtarget.hasAVX2() && VT.is256BitVector() ? 2 : 1; | |||
| 16658 | int NumSubLanes = NumLanes * SubLaneScale; | |||
| 16659 | int NumSubLaneElts = NumLaneElts / SubLaneScale; | |||
| 16660 | ||||
| 16661 | // Check that all the sources are coming from the same lane and see if we can | |||
| 16662 | // form a repeating shuffle mask (local to each sub-lane). At the same time, | |||
| 16663 | // determine the source sub-lane for each destination sub-lane. | |||
| 16664 | int TopSrcSubLane = -1; | |||
| 16665 | SmallVector<int, 8> Dst2SrcSubLanes((unsigned)NumSubLanes, -1); | |||
| 16666 | SmallVector<int, 8> RepeatedSubLaneMasks[2] = { | |||
| 16667 | SmallVector<int, 8>((unsigned)NumSubLaneElts, SM_SentinelUndef), | |||
| 16668 | SmallVector<int, 8>((unsigned)NumSubLaneElts, SM_SentinelUndef)}; | |||
| 16669 | ||||
| 16670 | for (int DstSubLane = 0; DstSubLane != NumSubLanes; ++DstSubLane) { | |||
| 16671 | // Extract the sub-lane mask, check that it all comes from the same lane | |||
| 16672 | // and normalize the mask entries to come from the first lane. | |||
| 16673 | int SrcLane = -1; | |||
| 16674 | SmallVector<int, 8> SubLaneMask((unsigned)NumSubLaneElts, -1); | |||
| 16675 | for (int Elt = 0; Elt != NumSubLaneElts; ++Elt) { | |||
| 16676 | int M = Mask[(DstSubLane * NumSubLaneElts) + Elt]; | |||
| 16677 | if (M < 0) | |||
| 16678 | continue; | |||
| 16679 | int Lane = (M % NumElts) / NumLaneElts; | |||
| 16680 | if ((0 <= SrcLane) && (SrcLane != Lane)) | |||
| 16681 | return SDValue(); | |||
| 16682 | SrcLane = Lane; | |||
| 16683 | int LocalM = (M % NumLaneElts) + (M < NumElts ? 0 : NumElts); | |||
| 16684 | SubLaneMask[Elt] = LocalM; | |||
| 16685 | } | |||
| 16686 | ||||
| 16687 | // Whole sub-lane is UNDEF. | |||
| 16688 | if (SrcLane < 0) | |||
| 16689 | continue; | |||
| 16690 | ||||
| 16691 | // Attempt to match against the candidate repeated sub-lane masks. | |||
| 16692 | for (int SubLane = 0; SubLane != SubLaneScale; ++SubLane) { | |||
| 16693 | auto MatchMasks = [NumSubLaneElts](ArrayRef<int> M1, ArrayRef<int> M2) { | |||
| 16694 | for (int i = 0; i != NumSubLaneElts; ++i) { | |||
| 16695 | if (M1[i] < 0 || M2[i] < 0) | |||
| 16696 | continue; | |||
| 16697 | if (M1[i] != M2[i]) | |||
| 16698 | return false; | |||
| 16699 | } | |||
| 16700 | return true; | |||
| 16701 | }; | |||
| 16702 | ||||
| 16703 | auto &RepeatedSubLaneMask = RepeatedSubLaneMasks[SubLane]; | |||
| 16704 | if (!MatchMasks(SubLaneMask, RepeatedSubLaneMask)) | |||
| 16705 | continue; | |||
| 16706 | ||||
| 16707 | // Merge the sub-lane mask into the matching repeated sub-lane mask. | |||
| 16708 | for (int i = 0; i != NumSubLaneElts; ++i) { | |||
| 16709 | int M = SubLaneMask[i]; | |||
| 16710 | if (M < 0) | |||
| 16711 | continue; | |||
| 16712 | assert((RepeatedSubLaneMask[i] < 0 || RepeatedSubLaneMask[i] == M) &&((void)0) | |||
| 16713 | "Unexpected mask element")((void)0); | |||
| 16714 | RepeatedSubLaneMask[i] = M; | |||
| 16715 | } | |||
| 16716 | ||||
| 16717 | // Track the top most source sub-lane - by setting the remaining to UNDEF | |||
| 16718 | // we can greatly simplify shuffle matching. | |||
| 16719 | int SrcSubLane = (SrcLane * SubLaneScale) + SubLane; | |||
| 16720 | TopSrcSubLane = std::max(TopSrcSubLane, SrcSubLane); | |||
| 16721 | Dst2SrcSubLanes[DstSubLane] = SrcSubLane; | |||
| 16722 | break; | |||
| 16723 | } | |||
| 16724 | ||||
| 16725 | // Bail if we failed to find a matching repeated sub-lane mask. | |||
| 16726 | if (Dst2SrcSubLanes[DstSubLane] < 0) | |||
| 16727 | return SDValue(); | |||
| 16728 | } | |||
| 16729 | assert(0 <= TopSrcSubLane && TopSrcSubLane < NumSubLanes &&((void)0) | |||
| 16730 | "Unexpected source lane")((void)0); | |||
| 16731 | ||||
| 16732 | // Create a repeating shuffle mask for the entire vector. | |||
| 16733 | SmallVector<int, 8> RepeatedMask((unsigned)NumElts, -1); | |||
| 16734 | for (int SubLane = 0; SubLane <= TopSrcSubLane; ++SubLane) { | |||
| 16735 | int Lane = SubLane / SubLaneScale; | |||
| 16736 | auto &RepeatedSubLaneMask = RepeatedSubLaneMasks[SubLane % SubLaneScale]; | |||
| 16737 | for (int Elt = 0; Elt != NumSubLaneElts; ++Elt) { | |||
| 16738 | int M = RepeatedSubLaneMask[Elt]; | |||
| 16739 | if (M < 0) | |||
| 16740 | continue; | |||
| 16741 | int Idx = (SubLane * NumSubLaneElts) + Elt; | |||
| 16742 | RepeatedMask[Idx] = M + (Lane * NumLaneElts); | |||
| 16743 | } | |||
| 16744 | } | |||
| 16745 | SDValue RepeatedShuffle = DAG.getVectorShuffle(VT, DL, V1, V2, RepeatedMask); | |||
| 16746 | ||||
| 16747 | // Shuffle each source sub-lane to its destination. | |||
| 16748 | SmallVector<int, 8> SubLaneMask((unsigned)NumElts, -1); | |||
| 16749 | for (int i = 0; i != NumElts; i += NumSubLaneElts) { | |||
| 16750 | int SrcSubLane = Dst2SrcSubLanes[i / NumSubLaneElts]; | |||
| 16751 | if (SrcSubLane < 0) | |||
| 16752 | continue; | |||
| 16753 | for (int j = 0; j != NumSubLaneElts; ++j) | |||
| 16754 | SubLaneMask[i + j] = j + (SrcSubLane * NumSubLaneElts); | |||
| 16755 | } | |||
| 16756 | ||||
| 16757 | return DAG.getVectorShuffle(VT, DL, RepeatedShuffle, DAG.getUNDEF(VT), | |||
| 16758 | SubLaneMask); | |||
| 16759 | } | |||
| 16760 | ||||
| 16761 | static bool matchShuffleWithSHUFPD(MVT VT, SDValue &V1, SDValue &V2, | |||
| 16762 | bool &ForceV1Zero, bool &ForceV2Zero, | |||
| 16763 | unsigned &ShuffleImm, ArrayRef<int> Mask, | |||
| 16764 | const APInt &Zeroable) { | |||
| 16765 | int NumElts = VT.getVectorNumElements(); | |||
| 16766 | assert(VT.getScalarSizeInBits() == 64 &&((void)0) | |||
| 16767 | (NumElts == 2 || NumElts == 4 || NumElts == 8) &&((void)0) | |||
| 16768 | "Unexpected data type for VSHUFPD")((void)0); | |||
| 16769 | assert(isUndefOrZeroOrInRange(Mask, 0, 2 * NumElts) &&((void)0) | |||
| 16770 | "Illegal shuffle mask")((void)0); | |||
| 16771 | ||||
| 16772 | bool ZeroLane[2] = { true, true }; | |||
| 16773 | for (int i = 0; i < NumElts; ++i) | |||
| 16774 | ZeroLane[i & 1] &= Zeroable[i]; | |||
| 16775 | ||||
| 16776 | // Mask for V8F64: 0/1, 8/9, 2/3, 10/11, 4/5, .. | |||
| 16777 | // Mask for V4F64; 0/1, 4/5, 2/3, 6/7.. | |||
| 16778 | ShuffleImm = 0; | |||
| 16779 | bool ShufpdMask = true; | |||
| 16780 | bool CommutableMask = true; | |||
| 16781 | for (int i = 0; i < NumElts; ++i) { | |||
| 16782 | if (Mask[i] == SM_SentinelUndef || ZeroLane[i & 1]) | |||
| 16783 | continue; | |||
| 16784 | if (Mask[i] < 0) | |||
| 16785 | return false; | |||
| 16786 | int Val = (i & 6) + NumElts * (i & 1); | |||
| 16787 | int CommutVal = (i & 0xe) + NumElts * ((i & 1) ^ 1); | |||
| 16788 | if (Mask[i] < Val || Mask[i] > Val + 1) | |||
| 16789 | ShufpdMask = false; | |||
| 16790 | if (Mask[i] < CommutVal || Mask[i] > CommutVal + 1) | |||
| 16791 | CommutableMask = false; | |||
| 16792 | ShuffleImm |= (Mask[i] % 2) << i; | |||
| 16793 | } | |||
| 16794 | ||||
| 16795 | if (!ShufpdMask && !CommutableMask) | |||
| 16796 | return false; | |||
| 16797 | ||||
| 16798 | if (!ShufpdMask && CommutableMask) | |||
| 16799 | std::swap(V1, V2); | |||
| 16800 | ||||
| 16801 | ForceV1Zero = ZeroLane[0]; | |||
| 16802 | ForceV2Zero = ZeroLane[1]; | |||
| 16803 | return true; | |||
| 16804 | } | |||
| 16805 | ||||
| 16806 | static SDValue lowerShuffleWithSHUFPD(const SDLoc &DL, MVT VT, SDValue V1, | |||
| 16807 | SDValue V2, ArrayRef<int> Mask, | |||
| 16808 | const APInt &Zeroable, | |||
| 16809 | const X86Subtarget &Subtarget, | |||
| 16810 | SelectionDAG &DAG) { | |||
| 16811 | assert((VT == MVT::v2f64 || VT == MVT::v4f64 || VT == MVT::v8f64) &&((void)0) | |||
| 16812 | "Unexpected data type for VSHUFPD")((void)0); | |||
| 16813 | ||||
| 16814 | unsigned Immediate = 0; | |||
| 16815 | bool ForceV1Zero = false, ForceV2Zero = false; | |||
| 16816 | if (!matchShuffleWithSHUFPD(VT, V1, V2, ForceV1Zero, ForceV2Zero, Immediate, | |||
| 16817 | Mask, Zeroable)) | |||
| 16818 | return SDValue(); | |||
| 16819 | ||||
| 16820 | // Create a REAL zero vector - ISD::isBuildVectorAllZeros allows UNDEFs. | |||
| 16821 | if (ForceV1Zero) | |||
| 16822 | V1 = getZeroVector(VT, Subtarget, DAG, DL); | |||
| 16823 | if (ForceV2Zero) | |||
| 16824 | V2 = getZeroVector(VT, Subtarget, DAG, DL); | |||
| 16825 | ||||
| 16826 | return DAG.getNode(X86ISD::SHUFP, DL, VT, V1, V2, | |||
| 16827 | DAG.getTargetConstant(Immediate, DL, MVT::i8)); | |||
| 16828 | } | |||
| 16829 | ||||
| 16830 | // Look for {0, 8, 16, 24, 32, 40, 48, 56 } in the first 8 elements. Followed | |||
| 16831 | // by zeroable elements in the remaining 24 elements. Turn this into two | |||
| 16832 | // vmovqb instructions shuffled together. | |||
| 16833 | static SDValue lowerShuffleAsVTRUNCAndUnpack(const SDLoc &DL, MVT VT, | |||
| 16834 | SDValue V1, SDValue V2, | |||
| 16835 | ArrayRef<int> Mask, | |||
| 16836 | const APInt &Zeroable, | |||
| 16837 | SelectionDAG &DAG) { | |||
| 16838 | assert(VT == MVT::v32i8 && "Unexpected type!")((void)0); | |||
| 16839 | ||||
| 16840 | // The first 8 indices should be every 8th element. | |||
| 16841 | if (!isSequentialOrUndefInRange(Mask, 0, 8, 0, 8)) | |||
| 16842 | return SDValue(); | |||
| 16843 | ||||
| 16844 | // Remaining elements need to be zeroable. | |||
| 16845 | if (Zeroable.countLeadingOnes() < (Mask.size() - 8)) | |||
| 16846 | return SDValue(); | |||
| 16847 | ||||
| 16848 | V1 = DAG.getBitcast(MVT::v4i64, V1); | |||
| 16849 | V2 = DAG.getBitcast(MVT::v4i64, V2); | |||
| 16850 | ||||
| 16851 | V1 = DAG.getNode(X86ISD::VTRUNC, DL, MVT::v16i8, V1); | |||
| 16852 | V2 = DAG.getNode(X86ISD::VTRUNC, DL, MVT::v16i8, V2); | |||
| 16853 | ||||
| 16854 | // The VTRUNCs will put 0s in the upper 12 bytes. Use them to put zeroes in | |||
| 16855 | // the upper bits of the result using an unpckldq. | |||
| 16856 | SDValue Unpack = DAG.getVectorShuffle(MVT::v16i8, DL, V1, V2, | |||
| 16857 | { 0, 1, 2, 3, 16, 17, 18, 19, | |||
| 16858 | 4, 5, 6, 7, 20, 21, 22, 23 }); | |||
| 16859 | // Insert the unpckldq into a zero vector to widen to v32i8. | |||
| 16860 | return DAG.getNode(ISD::INSERT_SUBVECTOR, DL, MVT::v32i8, | |||
| 16861 | DAG.getConstant(0, DL, MVT::v32i8), Unpack, | |||
| 16862 | DAG.getIntPtrConstant(0, DL)); | |||
| 16863 | } | |||
| 16864 | ||||
| 16865 | ||||
| 16866 | /// Handle lowering of 4-lane 64-bit floating point shuffles. | |||
| 16867 | /// | |||
| 16868 | /// Also ends up handling lowering of 4-lane 64-bit integer shuffles when AVX2 | |||
| 16869 | /// isn't available. | |||
| 16870 | static SDValue lowerV4F64Shuffle(const SDLoc &DL, ArrayRef<int> Mask, | |||
| 16871 | const APInt &Zeroable, SDValue V1, SDValue V2, | |||
| 16872 | const X86Subtarget &Subtarget, | |||
| 16873 | SelectionDAG &DAG) { | |||
| 16874 | assert(V1.getSimpleValueType() == MVT::v4f64 && "Bad operand type!")((void)0); | |||
| 16875 | assert(V2.getSimpleValueType() == MVT::v4f64 && "Bad operand type!")((void)0); | |||
| 16876 | assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!")((void)0); | |||
| 16877 | ||||
| 16878 | if (SDValue V = lowerV2X128Shuffle(DL, MVT::v4f64, V1, V2, Mask, Zeroable, | |||
| 16879 | Subtarget, DAG)) | |||
| 16880 | return V; | |||
| 16881 | ||||
| 16882 | if (V2.isUndef()) { | |||
| 16883 | // Check for being able to broadcast a single element. | |||
| 16884 | if (SDValue Broadcast = lowerShuffleAsBroadcast(DL, MVT::v4f64, V1, V2, | |||
| 16885 | Mask, Subtarget, DAG)) | |||
| 16886 | return Broadcast; | |||
| 16887 | ||||
| 16888 | // Use low duplicate instructions for masks that match their pattern. | |||
| 16889 | if (isShuffleEquivalent(Mask, {0, 0, 2, 2}, V1, V2)) | |||
| 16890 | return DAG.getNode(X86ISD::MOVDDUP, DL, MVT::v4f64, V1); | |||
| 16891 | ||||
| 16892 | if (!is128BitLaneCrossingShuffleMask(MVT::v4f64, Mask)) { | |||
| 16893 | // Non-half-crossing single input shuffles can be lowered with an | |||
| 16894 | // interleaved permutation. | |||
| 16895 | unsigned VPERMILPMask = (Mask[0] == 1) | ((Mask[1] == 1) << 1) | | |||
| 16896 | ((Mask[2] == 3) << 2) | ((Mask[3] == 3) << 3); | |||
| 16897 | return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v4f64, V1, | |||
| 16898 | DAG.getTargetConstant(VPERMILPMask, DL, MVT::i8)); | |||
| 16899 | } | |||
| 16900 | ||||
| 16901 | // With AVX2 we have direct support for this permutation. | |||
| 16902 | if (Subtarget.hasAVX2()) | |||
| 16903 | return DAG.getNode(X86ISD::VPERMI, DL, MVT::v4f64, V1, | |||
| 16904 | getV4X86ShuffleImm8ForMask(Mask, DL, DAG)); | |||
| 16905 | ||||
| 16906 | // Try to create an in-lane repeating shuffle mask and then shuffle the | |||
| 16907 | // results into the target lanes. | |||
| 16908 | if (SDValue V = lowerShuffleAsRepeatedMaskAndLanePermute( | |||
| 16909 | DL, MVT::v4f64, V1, V2, Mask, Subtarget, DAG)) | |||
| 16910 | return V; | |||
| 16911 | ||||
| 16912 | // Try to permute the lanes and then use a per-lane permute. | |||
| 16913 | if (SDValue V = lowerShuffleAsLanePermuteAndPermute(DL, MVT::v4f64, V1, V2, | |||
| 16914 | Mask, DAG, Subtarget)) | |||
| 16915 | return V; | |||
| 16916 | ||||
| 16917 | // Otherwise, fall back. | |||
| 16918 | return lowerShuffleAsLanePermuteAndShuffle(DL, MVT::v4f64, V1, V2, Mask, | |||
| 16919 | DAG, Subtarget); | |||
| 16920 | } | |||
| 16921 | ||||
| 16922 | // Use dedicated unpack instructions for masks that match their pattern. | |||
| 16923 | if (SDValue V = lowerShuffleWithUNPCK(DL, MVT::v4f64, Mask, V1, V2, DAG)) | |||
| 16924 | return V; | |||
| 16925 | ||||
| 16926 | if (SDValue Blend = lowerShuffleAsBlend(DL, MVT::v4f64, V1, V2, Mask, | |||
| 16927 | Zeroable, Subtarget, DAG)) | |||
| 16928 | return Blend; | |||
| 16929 | ||||
| 16930 | // Check if the blend happens to exactly fit that of SHUFPD. | |||
| 16931 | if (SDValue Op = lowerShuffleWithSHUFPD(DL, MVT::v4f64, V1, V2, Mask, | |||
| 16932 | Zeroable, Subtarget, DAG)) | |||
| 16933 | return Op; | |||
| 16934 | ||||
| 16935 | // If we have lane crossing shuffles AND they don't all come from the lower | |||
| 16936 | // lane elements, lower to SHUFPD(VPERM2F128(V1, V2), VPERM2F128(V1, V2)). | |||
| 16937 | // TODO: Handle BUILD_VECTOR sources which getVectorShuffle currently | |||
| 16938 | // canonicalize to a blend of splat which isn't necessary for this combine. | |||
| 16939 | if (is128BitLaneCrossingShuffleMask(MVT::v4f64, Mask) && | |||
| 16940 | !all_of(Mask, [](int M) { return M < 2 || (4 <= M && M < 6); }) && | |||
| 16941 | (V1.getOpcode() != ISD::BUILD_VECTOR) && | |||
| 16942 | (V2.getOpcode() != ISD::BUILD_VECTOR)) | |||
| 16943 | if (SDValue Op = lowerShuffleAsLanePermuteAndSHUFP(DL, MVT::v4f64, V1, V2, | |||
| 16944 | Mask, DAG)) | |||
| 16945 | return Op; | |||
| 16946 | ||||
| 16947 | // If we have one input in place, then we can permute the other input and | |||
| 16948 | // blend the result. | |||
| 16949 | if (isShuffleMaskInputInPlace(0, Mask) || isShuffleMaskInputInPlace(1, Mask)) | |||
| 16950 | return lowerShuffleAsDecomposedShuffleMerge(DL, MVT::v4f64, V1, V2, Mask, | |||
| 16951 | Subtarget, DAG); | |||
| 16952 | ||||
| 16953 | // Try to create an in-lane repeating shuffle mask and then shuffle the | |||
| 16954 | // results into the target lanes. | |||
| 16955 | if (SDValue V = lowerShuffleAsRepeatedMaskAndLanePermute( | |||
| 16956 | DL, MVT::v4f64, V1, V2, Mask, Subtarget, DAG)) | |||
| 16957 | return V; | |||
| 16958 | ||||
| 16959 | // Try to simplify this by merging 128-bit lanes to enable a lane-based | |||
| 16960 | // shuffle. However, if we have AVX2 and either inputs are already in place, | |||
| 16961 | // we will be able to shuffle even across lanes the other input in a single | |||
| 16962 | // instruction so skip this pattern. | |||
| 16963 | if (!(Subtarget.hasAVX2() && (isShuffleMaskInputInPlace(0, Mask) || | |||
| 16964 | isShuffleMaskInputInPlace(1, Mask)))) | |||
| 16965 | if (SDValue V = lowerShuffleAsLanePermuteAndRepeatedMask( | |||
| 16966 | DL, MVT::v4f64, V1, V2, Mask, Subtarget, DAG)) | |||
| 16967 | return V; | |||
| 16968 | ||||
| 16969 | // If we have VLX support, we can use VEXPAND. | |||
| 16970 | if (Subtarget.hasVLX()) | |||
| 16971 | if (SDValue V = lowerShuffleToEXPAND(DL, MVT::v4f64, Zeroable, Mask, V1, V2, | |||
| 16972 | DAG, Subtarget)) | |||
| 16973 | return V; | |||
| 16974 | ||||
| 16975 | // If we have AVX2 then we always want to lower with a blend because an v4 we | |||
| 16976 | // can fully permute the elements. | |||
| 16977 | if (Subtarget.hasAVX2()) | |||
| 16978 | return lowerShuffleAsDecomposedShuffleMerge(DL, MVT::v4f64, V1, V2, Mask, | |||
| 16979 | Subtarget, DAG); | |||
| 16980 | ||||
| 16981 | // Otherwise fall back on generic lowering. | |||
| 16982 | return lowerShuffleAsSplitOrBlend(DL, MVT::v4f64, V1, V2, Mask, | |||
| 16983 | Subtarget, DAG); | |||
| 16984 | } | |||
| 16985 | ||||
| 16986 | /// Handle lowering of 4-lane 64-bit integer shuffles. | |||
| 16987 | /// | |||
| 16988 | /// This routine is only called when we have AVX2 and thus a reasonable | |||
| 16989 | /// instruction set for v4i64 shuffling.. | |||
| 16990 | static SDValue lowerV4I64Shuffle(const SDLoc &DL, ArrayRef<int> Mask, | |||
| 16991 | const APInt &Zeroable, SDValue V1, SDValue V2, | |||
| 16992 | const X86Subtarget &Subtarget, | |||
| 16993 | SelectionDAG &DAG) { | |||
| 16994 | assert(V1.getSimpleValueType() == MVT::v4i64 && "Bad operand type!")((void)0); | |||
| 16995 | assert(V2.getSimpleValueType() == MVT::v4i64 && "Bad operand type!")((void)0); | |||
| 16996 | assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!")((void)0); | |||
| 16997 | assert(Subtarget.hasAVX2() && "We can only lower v4i64 with AVX2!")((void)0); | |||
| 16998 | ||||
| 16999 | if (SDValue V = lowerV2X128Shuffle(DL, MVT::v4i64, V1, V2, Mask, Zeroable, | |||
| 17000 | Subtarget, DAG)) | |||
| 17001 | return V; | |||
| 17002 | ||||
| 17003 | if (SDValue Blend = lowerShuffleAsBlend(DL, MVT::v4i64, V1, V2, Mask, | |||
| 17004 | Zeroable, Subtarget, DAG)) | |||
| 17005 | return Blend; | |||
| 17006 | ||||
| 17007 | // Check for being able to broadcast a single element. | |||
| 17008 | if (SDValue Broadcast = lowerShuffleAsBroadcast(DL, MVT::v4i64, V1, V2, Mask, | |||
| 17009 | Subtarget, DAG)) | |||
| 17010 | return Broadcast; | |||
| 17011 | ||||
| 17012 | if (V2.isUndef()) { | |||
| 17013 | // When the shuffle is mirrored between the 128-bit lanes of the unit, we | |||
| 17014 | // can use lower latency instructions that will operate on both lanes. | |||
| 17015 | SmallVector<int, 2> RepeatedMask; | |||
| 17016 | if (is128BitLaneRepeatedShuffleMask(MVT::v4i64, Mask, RepeatedMask)) { | |||
| 17017 | SmallVector<int, 4> PSHUFDMask; | |||
| 17018 | narrowShuffleMaskElts(2, RepeatedMask, PSHUFDMask); | |||
| 17019 | return DAG.getBitcast( | |||
| 17020 | MVT::v4i64, | |||
| 17021 | DAG.getNode(X86ISD::PSHUFD, DL, MVT::v8i32, | |||
| 17022 | DAG.getBitcast(MVT::v8i32, V1), | |||
| 17023 | getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG))); | |||
| 17024 | } | |||
| 17025 | ||||
| 17026 | // AVX2 provides a direct instruction for permuting a single input across | |||
| 17027 | // lanes. | |||
| 17028 | return DAG.getNode(X86ISD::VPERMI, DL, MVT::v4i64, V1, | |||
| 17029 | getV4X86ShuffleImm8ForMask(Mask, DL, DAG)); | |||
| 17030 | } | |||
| 17031 | ||||
| 17032 | // Try to use shift instructions. | |||
| 17033 | if (SDValue Shift = lowerShuffleAsShift(DL, MVT::v4i64, V1, V2, Mask, | |||
| 17034 | Zeroable, Subtarget, DAG)) | |||
| 17035 | return Shift; | |||
| 17036 | ||||
| 17037 | // If we have VLX support, we can use VALIGN or VEXPAND. | |||
| 17038 | if (Subtarget.hasVLX()) { | |||
| 17039 | if (SDValue Rotate = lowerShuffleAsVALIGN(DL, MVT::v4i64, V1, V2, Mask, | |||
| 17040 | Subtarget, DAG)) | |||
| 17041 | return Rotate; | |||
| 17042 | ||||
| 17043 | if (SDValue V = lowerShuffleToEXPAND(DL, MVT::v4i64, Zeroable, Mask, V1, V2, | |||
| 17044 | DAG, Subtarget)) | |||
| 17045 | return V; | |||
| 17046 | } | |||
| 17047 | ||||
| 17048 | // Try to use PALIGNR. | |||
| 17049 | if (SDValue Rotate = lowerShuffleAsByteRotate(DL, MVT::v4i64, V1, V2, Mask, | |||
| 17050 | Subtarget, DAG)) | |||
| 17051 | return Rotate; | |||
| 17052 | ||||
| 17053 | // Use dedicated unpack instructions for masks that match their pattern. | |||
| 17054 | if (SDValue V = lowerShuffleWithUNPCK(DL, MVT::v4i64, Mask, V1, V2, DAG)) | |||
| 17055 | return V; | |||
| 17056 | ||||
| 17057 | // If we have one input in place, then we can permute the other input and | |||
| 17058 | // blend the result. | |||
| 17059 | if (isShuffleMaskInputInPlace(0, Mask) || isShuffleMaskInputInPlace(1, Mask)) | |||
| 17060 | return lowerShuffleAsDecomposedShuffleMerge(DL, MVT::v4i64, V1, V2, Mask, | |||
| 17061 | Subtarget, DAG); | |||
| 17062 | ||||
| 17063 | // Try to create an in-lane repeating shuffle mask and then shuffle the | |||
| 17064 | // results into the target lanes. | |||
| 17065 | if (SDValue V = lowerShuffleAsRepeatedMaskAndLanePermute( | |||
| 17066 | DL, MVT::v4i64, V1, V2, Mask, Subtarget, DAG)) | |||
| 17067 | return V; | |||
| 17068 | ||||
| 17069 | // Try to simplify this by merging 128-bit lanes to enable a lane-based | |||
| 17070 | // shuffle. However, if we have AVX2 and either inputs are already in place, | |||
| 17071 | // we will be able to shuffle even across lanes the other input in a single | |||
| 17072 | // instruction so skip this pattern. | |||
| 17073 | if (!isShuffleMaskInputInPlace(0, Mask) && | |||
| 17074 | !isShuffleMaskInputInPlace(1, Mask)) | |||
| 17075 | if (SDValue Result = lowerShuffleAsLanePermuteAndRepeatedMask( | |||
| 17076 | DL, MVT::v4i64, V1, V2, Mask, Subtarget, DAG)) | |||
| 17077 | return Result; | |||
| 17078 | ||||
| 17079 | // Otherwise fall back on generic blend lowering. | |||
| 17080 | return lowerShuffleAsDecomposedShuffleMerge(DL, MVT::v4i64, V1, V2, Mask, | |||
| 17081 | Subtarget, DAG); | |||
| 17082 | } | |||
| 17083 | ||||
| 17084 | /// Handle lowering of 8-lane 32-bit floating point shuffles. | |||
| 17085 | /// | |||
| 17086 | /// Also ends up handling lowering of 8-lane 32-bit integer shuffles when AVX2 | |||
| 17087 | /// isn't available. | |||
| 17088 | static SDValue lowerV8F32Shuffle(const SDLoc &DL, ArrayRef<int> Mask, | |||
| 17089 | const APInt &Zeroable, SDValue V1, SDValue V2, | |||
| 17090 | const X86Subtarget &Subtarget, | |||
| 17091 | SelectionDAG &DAG) { | |||
| 17092 | assert(V1.getSimpleValueType() == MVT::v8f32 && "Bad operand type!")((void)0); | |||
| 17093 | assert(V2.getSimpleValueType() == MVT::v8f32 && "Bad operand type!")((void)0); | |||
| 17094 | assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!")((void)0); | |||
| 17095 | ||||
| 17096 | if (SDValue Blend = lowerShuffleAsBlend(DL, MVT::v8f32, V1, V2, Mask, | |||
| 17097 | Zeroable, Subtarget, DAG)) | |||
| 17098 | return Blend; | |||
| 17099 | ||||
| 17100 | // Check for being able to broadcast a single element. | |||
| 17101 | if (SDValue Broadcast = lowerShuffleAsBroadcast(DL, MVT::v8f32, V1, V2, Mask, | |||
| 17102 | Subtarget, DAG)) | |||
| 17103 | return Broadcast; | |||
| 17104 | ||||
| 17105 | // If the shuffle mask is repeated in each 128-bit lane, we have many more | |||
| 17106 | // options to efficiently lower the shuffle. | |||
| 17107 | SmallVector<int, 4> RepeatedMask; | |||
| 17108 | if (is128BitLaneRepeatedShuffleMask(MVT::v8f32, Mask, RepeatedMask)) { | |||
| 17109 | assert(RepeatedMask.size() == 4 &&((void)0) | |||
| 17110 | "Repeated masks must be half the mask width!")((void)0); | |||
| 17111 | ||||
| 17112 | // Use even/odd duplicate instructions for masks that match their pattern. | |||
| 17113 | if (isShuffleEquivalent(RepeatedMask, {0, 0, 2, 2}, V1, V2)) | |||
| 17114 | return DAG.getNode(X86ISD::MOVSLDUP, DL, MVT::v8f32, V1); | |||
| 17115 | if (isShuffleEquivalent(RepeatedMask, {1, 1, 3, 3}, V1, V2)) | |||
| 17116 | return DAG.getNode(X86ISD::MOVSHDUP, DL, MVT::v8f32, V1); | |||
| 17117 | ||||
| 17118 | if (V2.isUndef()) | |||
| 17119 | return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v8f32, V1, | |||
| 17120 | getV4X86ShuffleImm8ForMask(RepeatedMask, DL, DAG)); | |||
| 17121 | ||||
| 17122 | // Use dedicated unpack instructions for masks that match their pattern. | |||
| 17123 | if (SDValue V = lowerShuffleWithUNPCK(DL, MVT::v8f32, Mask, V1, V2, DAG)) | |||
| 17124 | return V; | |||
| 17125 | ||||
| 17126 | // Otherwise, fall back to a SHUFPS sequence. Here it is important that we | |||
| 17127 | // have already handled any direct blends. | |||
| 17128 | return lowerShuffleWithSHUFPS(DL, MVT::v8f32, RepeatedMask, V1, V2, DAG); | |||
| 17129 | } | |||
| 17130 | ||||
| 17131 | // Try to create an in-lane repeating shuffle mask and then shuffle the | |||
| 17132 | // results into the target lanes. | |||
| 17133 | if (SDValue V = lowerShuffleAsRepeatedMaskAndLanePermute( | |||
| 17134 | DL, MVT::v8f32, V1, V2, Mask, Subtarget, DAG)) | |||
| 17135 | return V; | |||
| 17136 | ||||
| 17137 | // If we have a single input shuffle with different shuffle patterns in the | |||
| 17138 | // two 128-bit lanes use the variable mask to VPERMILPS. | |||
| 17139 | if (V2.isUndef()) { | |||
| 17140 | if (!is128BitLaneCrossingShuffleMask(MVT::v8f32, Mask)) { | |||
| 17141 | SDValue VPermMask = getConstVector(Mask, MVT::v8i32, DAG, DL, true); | |||
| 17142 | return DAG.getNode(X86ISD::VPERMILPV, DL, MVT::v8f32, V1, VPermMask); | |||
| 17143 | } | |||
| 17144 | if (Subtarget.hasAVX2()) { | |||
| 17145 | SDValue VPermMask = getConstVector(Mask, MVT::v8i32, DAG, DL, true); | |||
| 17146 | return DAG.getNode(X86ISD::VPERMV, DL, MVT::v8f32, VPermMask, V1); | |||
| 17147 | } | |||
| 17148 | // Otherwise, fall back. | |||
| 17149 | return lowerShuffleAsLanePermuteAndShuffle(DL, MVT::v8f32, V1, V2, Mask, | |||
| 17150 | DAG, Subtarget); | |||
| 17151 | } | |||
| 17152 | ||||
| 17153 | // Try to simplify this by merging 128-bit lanes to enable a lane-based | |||
| 17154 | // shuffle. | |||
| 17155 | if (SDValue Result = lowerShuffleAsLanePermuteAndRepeatedMask( | |||
| 17156 | DL, MVT::v8f32, V1, V2, Mask, Subtarget, DAG)) | |||
| 17157 | return Result; | |||
| 17158 | ||||
| 17159 | // If we have VLX support, we can use VEXPAND. | |||
| 17160 | if (Subtarget.hasVLX()) | |||
| 17161 | if (SDValue V = lowerShuffleToEXPAND(DL, MVT::v8f32, Zeroable, Mask, V1, V2, | |||
| 17162 | DAG, Subtarget)) | |||
| 17163 | return V; | |||
| 17164 | ||||
| 17165 | // For non-AVX512 if the Mask is of 16bit elements in lane then try to split | |||
| 17166 | // since after split we get a more efficient code using vpunpcklwd and | |||
| 17167 | // vpunpckhwd instrs than vblend. | |||
| 17168 | if (!Subtarget.hasAVX512() && isUnpackWdShuffleMask(Mask, MVT::v8f32)) | |||
| 17169 | return lowerShuffleAsSplitOrBlend(DL, MVT::v8f32, V1, V2, Mask, Subtarget, | |||
| 17170 | DAG); | |||
| 17171 | ||||
| 17172 | // If we have AVX2 then we always want to lower with a blend because at v8 we | |||
| 17173 | // can fully permute the elements. | |||
| 17174 | if (Subtarget.hasAVX2()) | |||
| 17175 | return lowerShuffleAsDecomposedShuffleMerge(DL, MVT::v8f32, V1, V2, Mask, | |||
| 17176 | Subtarget, DAG); | |||
| 17177 | ||||
| 17178 | // Otherwise fall back on generic lowering. | |||
| 17179 | return lowerShuffleAsSplitOrBlend(DL, MVT::v8f32, V1, V2, Mask, | |||
| 17180 | Subtarget, DAG); | |||
| 17181 | } | |||
| 17182 | ||||
| 17183 | /// Handle lowering of 8-lane 32-bit integer shuffles. | |||
| 17184 | /// | |||
| 17185 | /// This routine is only called when we have AVX2 and thus a reasonable | |||
| 17186 | /// instruction set for v8i32 shuffling.. | |||
| 17187 | static SDValue lowerV8I32Shuffle(const SDLoc &DL, ArrayRef<int> Mask, | |||
| 17188 | const APInt &Zeroable, SDValue V1, SDValue V2, | |||
| 17189 | const X86Subtarget &Subtarget, | |||
| 17190 | SelectionDAG &DAG) { | |||
| 17191 | assert(V1.getSimpleValueType() == MVT::v8i32 && "Bad operand type!")((void)0); | |||
| 17192 | assert(V2.getSimpleValueType() == MVT::v8i32 && "Bad operand type!")((void)0); | |||
| 17193 | assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!")((void)0); | |||
| 17194 | assert(Subtarget.hasAVX2() && "We can only lower v8i32 with AVX2!")((void)0); | |||
| 17195 | ||||
| 17196 | // Whenever we can lower this as a zext, that instruction is strictly faster | |||
| 17197 | // than any alternative. It also allows us to fold memory operands into the | |||
| 17198 | // shuffle in many cases. | |||
| 17199 | if (SDValue ZExt = lowerShuffleAsZeroOrAnyExtend(DL, MVT::v8i32, V1, V2, Mask, | |||
| 17200 | Zeroable, Subtarget, DAG)) | |||
| 17201 | return ZExt; | |||
| 17202 | ||||
| 17203 | // For non-AVX512 if the Mask is of 16bit elements in lane then try to split | |||
| 17204 | // since after split we get a more efficient code than vblend by using | |||
| 17205 | // vpunpcklwd and vpunpckhwd instrs. | |||
| 17206 | if (isUnpackWdShuffleMask(Mask, MVT::v8i32) && !V2.isUndef() && | |||
| 17207 | !Subtarget.hasAVX512()) | |||
| 17208 | return lowerShuffleAsSplitOrBlend(DL, MVT::v8i32, V1, V2, Mask, Subtarget, | |||
| 17209 | DAG); | |||
| 17210 | ||||
| 17211 | if (SDValue Blend = lowerShuffleAsBlend(DL, MVT::v8i32, V1, V2, Mask, | |||
| 17212 | Zeroable, Subtarget, DAG)) | |||
| 17213 | return Blend; | |||
| 17214 | ||||
| 17215 | // Check for being able to broadcast a single element. | |||
| 17216 | if (SDValue Broadcast = lowerShuffleAsBroadcast(DL, MVT::v8i32, V1, V2, Mask, | |||
| 17217 | Subtarget, DAG)) | |||
| 17218 | return Broadcast; | |||
| 17219 | ||||
| 17220 | // If the shuffle mask is repeated in each 128-bit lane we can use more | |||
| 17221 | // efficient instructions that mirror the shuffles across the two 128-bit | |||
| 17222 | // lanes. | |||
| 17223 | SmallVector<int, 4> RepeatedMask; | |||
| 17224 | bool Is128BitLaneRepeatedShuffle = | |||
| 17225 | is128BitLaneRepeatedShuffleMask(MVT::v8i32, Mask, RepeatedMask); | |||
| 17226 | if (Is128BitLaneRepeatedShuffle) { | |||
| 17227 | assert(RepeatedMask.size() == 4 && "Unexpected repeated mask size!")((void)0); | |||
| 17228 | if (V2.isUndef()) | |||
| 17229 | return DAG.getNode(X86ISD::PSHUFD, DL, MVT::v8i32, V1, | |||
| 17230 | getV4X86ShuffleImm8ForMask(RepeatedMask, DL, DAG)); | |||
| 17231 | ||||
| 17232 | // Use dedicated unpack instructions for masks that match their pattern. | |||
| 17233 | if (SDValue V = lowerShuffleWithUNPCK(DL, MVT::v8i32, Mask, V1, V2, DAG)) | |||
| 17234 | return V; | |||
| 17235 | } | |||
| 17236 | ||||
| 17237 | // Try to use shift instructions. | |||
| 17238 | if (SDValue Shift = lowerShuffleAsShift(DL, MVT::v8i32, V1, V2, Mask, | |||
| 17239 | Zeroable, Subtarget, DAG)) | |||
| 17240 | return Shift; | |||
| 17241 | ||||
| 17242 | // If we have VLX support, we can use VALIGN or EXPAND. | |||
| 17243 | if (Subtarget.hasVLX()) { | |||
| 17244 | if (SDValue Rotate = lowerShuffleAsVALIGN(DL, MVT::v8i32, V1, V2, Mask, | |||
| 17245 | Subtarget, DAG)) | |||
| 17246 | return Rotate; | |||
| 17247 | ||||
| 17248 | if (SDValue V = lowerShuffleToEXPAND(DL, MVT::v8i32, Zeroable, Mask, V1, V2, | |||
| 17249 | DAG, Subtarget)) | |||
| 17250 | return V; | |||
| 17251 | } | |||
| 17252 | ||||
| 17253 | // Try to use byte rotation instructions. | |||
| 17254 | if (SDValue Rotate = lowerShuffleAsByteRotate(DL, MVT::v8i32, V1, V2, Mask, | |||
| 17255 | Subtarget, DAG)) | |||
| 17256 | return Rotate; | |||
| 17257 | ||||
| 17258 | // Try to create an in-lane repeating shuffle mask and then shuffle the | |||
| 17259 | // results into the target lanes. | |||
| 17260 | if (SDValue V = lowerShuffleAsRepeatedMaskAndLanePermute( | |||
| 17261 | DL, MVT::v8i32, V1, V2, Mask, Subtarget, DAG)) | |||
| 17262 | return V; | |||
| 17263 | ||||
| 17264 | if (V2.isUndef()) { | |||
| 17265 | // Try to produce a fixed cross-128-bit lane permute followed by unpack | |||
| 17266 | // because that should be faster than the variable permute alternatives. | |||
| 17267 | if (SDValue V = lowerShuffleWithUNPCK256(DL, MVT::v8i32, Mask, V1, V2, DAG)) | |||
| 17268 | return V; | |||
| 17269 | ||||
| 17270 | // If the shuffle patterns aren't repeated but it's a single input, directly | |||
| 17271 | // generate a cross-lane VPERMD instruction. | |||
| 17272 | SDValue VPermMask = getConstVector(Mask, MVT::v8i32, DAG, DL, true); | |||
| 17273 | return DAG.getNode(X86ISD::VPERMV, DL, MVT::v8i32, VPermMask, V1); | |||
| 17274 | } | |||
| 17275 | ||||
| 17276 | // Assume that a single SHUFPS is faster than an alternative sequence of | |||
| 17277 | // multiple instructions (even if the CPU has a domain penalty). | |||
| 17278 | // If some CPU is harmed by the domain switch, we can fix it in a later pass. | |||
| 17279 | if (Is128BitLaneRepeatedShuffle && isSingleSHUFPSMask(RepeatedMask)) { | |||
| 17280 | SDValue CastV1 = DAG.getBitcast(MVT::v8f32, V1); | |||
| 17281 | SDValue CastV2 = DAG.getBitcast(MVT::v8f32, V2); | |||
| 17282 | SDValue ShufPS = lowerShuffleWithSHUFPS(DL, MVT::v8f32, RepeatedMask, | |||
| 17283 | CastV1, CastV2, DAG); | |||
| 17284 | return DAG.getBitcast(MVT::v8i32, ShufPS); | |||
| 17285 | } | |||
| 17286 | ||||
| 17287 | // Try to simplify this by merging 128-bit lanes to enable a lane-based | |||
| 17288 | // shuffle. | |||
| 17289 | if (SDValue Result = lowerShuffleAsLanePermuteAndRepeatedMask( | |||
| 17290 | DL, MVT::v8i32, V1, V2, Mask, Subtarget, DAG)) | |||
| 17291 | return Result; | |||
| 17292 | ||||
| 17293 | // Otherwise fall back on generic blend lowering. | |||
| 17294 | return lowerShuffleAsDecomposedShuffleMerge(DL, MVT::v8i32, V1, V2, Mask, | |||
| 17295 | Subtarget, DAG); | |||
| 17296 | } | |||
| 17297 | ||||
| 17298 | /// Handle lowering of 16-lane 16-bit integer shuffles. | |||
| 17299 | /// | |||
| 17300 | /// This routine is only called when we have AVX2 and thus a reasonable | |||
| 17301 | /// instruction set for v16i16 shuffling.. | |||
| 17302 | static SDValue lowerV16I16Shuffle(const SDLoc &DL, ArrayRef<int> Mask, | |||
| 17303 | const APInt &Zeroable, SDValue V1, SDValue V2, | |||
| 17304 | const X86Subtarget &Subtarget, | |||
| 17305 | SelectionDAG &DAG) { | |||
| 17306 | assert(V1.getSimpleValueType() == MVT::v16i16 && "Bad operand type!")((void)0); | |||
| 17307 | assert(V2.getSimpleValueType() == MVT::v16i16 && "Bad operand type!")((void)0); | |||
| 17308 | assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!")((void)0); | |||
| 17309 | assert(Subtarget.hasAVX2() && "We can only lower v16i16 with AVX2!")((void)0); | |||
| 17310 | ||||
| 17311 | // Whenever we can lower this as a zext, that instruction is strictly faster | |||
| 17312 | // than any alternative. It also allows us to fold memory operands into the | |||
| 17313 | // shuffle in many cases. | |||
| 17314 | if (SDValue ZExt = lowerShuffleAsZeroOrAnyExtend( | |||
| 17315 | DL, MVT::v16i16, V1, V2, Mask, Zeroable, Subtarget, DAG)) | |||
| 17316 | return ZExt; | |||
| 17317 | ||||
| 17318 | // Check for being able to broadcast a single element. | |||
| 17319 | if (SDValue Broadcast = lowerShuffleAsBroadcast(DL, MVT::v16i16, V1, V2, Mask, | |||
| 17320 | Subtarget, DAG)) | |||
| 17321 | return Broadcast; | |||
| 17322 | ||||
| 17323 | if (SDValue Blend = lowerShuffleAsBlend(DL, MVT::v16i16, V1, V2, Mask, | |||
| 17324 | Zeroable, Subtarget, DAG)) | |||
| 17325 | return Blend; | |||
| 17326 | ||||
| 17327 | // Use dedicated unpack instructions for masks that match their pattern. | |||
| 17328 | if (SDValue V = lowerShuffleWithUNPCK(DL, MVT::v16i16, Mask, V1, V2, DAG)) | |||
| 17329 | return V; | |||
| 17330 | ||||
| 17331 | // Use dedicated pack instructions for masks that match their pattern. | |||
| 17332 | if (SDValue V = lowerShuffleWithPACK(DL, MVT::v16i16, Mask, V1, V2, DAG, | |||
| 17333 | Subtarget)) | |||
| 17334 | return V; | |||
| 17335 | ||||
| 17336 | // Try to use lower using a truncation. | |||
| 17337 | if (SDValue V = lowerShuffleAsVTRUNC(DL, MVT::v16i16, V1, V2, Mask, Zeroable, | |||
| 17338 | Subtarget, DAG)) | |||
| 17339 | return V; | |||
| 17340 | ||||
| 17341 | // Try to use shift instructions. | |||
| 17342 | if (SDValue Shift = lowerShuffleAsShift(DL, MVT::v16i16, V1, V2, Mask, | |||
| 17343 | Zeroable, Subtarget, DAG)) | |||
| 17344 | return Shift; | |||
| 17345 | ||||
| 17346 | // Try to use byte rotation instructions. | |||
| 17347 | if (SDValue Rotate = lowerShuffleAsByteRotate(DL, MVT::v16i16, V1, V2, Mask, | |||
| 17348 | Subtarget, DAG)) | |||
| 17349 | return Rotate; | |||
| 17350 | ||||
| 17351 | // Try to create an in-lane repeating shuffle mask and then shuffle the | |||
| 17352 | // results into the target lanes. | |||
| 17353 | if (SDValue V = lowerShuffleAsRepeatedMaskAndLanePermute( | |||
| 17354 | DL, MVT::v16i16, V1, V2, Mask, Subtarget, DAG)) | |||
| 17355 | return V; | |||
| 17356 | ||||
| 17357 | if (V2.isUndef()) { | |||
| 17358 | // Try to use bit rotation instructions. | |||
| 17359 | if (SDValue Rotate = | |||
| 17360 | lowerShuffleAsBitRotate(DL, MVT::v16i16, V1, Mask, Subtarget, DAG)) | |||
| 17361 | return Rotate; | |||
| 17362 | ||||
| 17363 | // Try to produce a fixed cross-128-bit lane permute followed by unpack | |||
| 17364 | // because that should be faster than the variable permute alternatives. | |||
| 17365 | if (SDValue V = lowerShuffleWithUNPCK256(DL, MVT::v16i16, Mask, V1, V2, DAG)) | |||
| 17366 | return V; | |||
| 17367 | ||||
| 17368 | // There are no generalized cross-lane shuffle operations available on i16 | |||
| 17369 | // element types. | |||
| 17370 | if (is128BitLaneCrossingShuffleMask(MVT::v16i16, Mask)) { | |||
| 17371 | if (SDValue V = lowerShuffleAsLanePermuteAndPermute( | |||
| 17372 | DL, MVT::v16i16, V1, V2, Mask, DAG, Subtarget)) | |||
| 17373 | return V; | |||
| 17374 | ||||
| 17375 | return lowerShuffleAsLanePermuteAndShuffle(DL, MVT::v16i16, V1, V2, Mask, | |||
| 17376 | DAG, Subtarget); | |||
| 17377 | } | |||
| 17378 | ||||
| 17379 | SmallVector<int, 8> RepeatedMask; | |||
| 17380 | if (is128BitLaneRepeatedShuffleMask(MVT::v16i16, Mask, RepeatedMask)) { | |||
| 17381 | // As this is a single-input shuffle, the repeated mask should be | |||
| 17382 | // a strictly valid v8i16 mask that we can pass through to the v8i16 | |||
| 17383 | // lowering to handle even the v16 case. | |||
| 17384 | return lowerV8I16GeneralSingleInputShuffle( | |||
| 17385 | DL, MVT::v16i16, V1, RepeatedMask, Subtarget, DAG); | |||
| 17386 | } | |||
| 17387 | } | |||
| 17388 | ||||
| 17389 | if (SDValue PSHUFB = lowerShuffleWithPSHUFB(DL, MVT::v16i16, Mask, V1, V2, | |||
| 17390 | Zeroable, Subtarget, DAG)) | |||
| 17391 | return PSHUFB; | |||
| 17392 | ||||
| 17393 | // AVX512BW can lower to VPERMW (non-VLX will pad to v32i16). | |||
| 17394 | if (Subtarget.hasBWI()) | |||
| 17395 | return lowerShuffleWithPERMV(DL, MVT::v16i16, Mask, V1, V2, Subtarget, DAG); | |||
| 17396 | ||||
| 17397 | // Try to simplify this by merging 128-bit lanes to enable a lane-based | |||
| 17398 | // shuffle. | |||
| 17399 | if (SDValue Result = lowerShuffleAsLanePermuteAndRepeatedMask( | |||
| 17400 | DL, MVT::v16i16, V1, V2, Mask, Subtarget, DAG)) | |||
| 17401 | return Result; | |||
| 17402 | ||||
| 17403 | // Try to permute the lanes and then use a per-lane permute. | |||
| 17404 | if (SDValue V = lowerShuffleAsLanePermuteAndPermute( | |||
| 17405 | DL, MVT::v16i16, V1, V2, Mask, DAG, Subtarget)) | |||
| 17406 | return V; | |||
| 17407 | ||||
| 17408 | // Otherwise fall back on generic lowering. | |||
| 17409 | return lowerShuffleAsSplitOrBlend(DL, MVT::v16i16, V1, V2, Mask, | |||
| 17410 | Subtarget, DAG); | |||
| 17411 | } | |||
| 17412 | ||||
| 17413 | /// Handle lowering of 32-lane 8-bit integer shuffles. | |||
| 17414 | /// | |||
| 17415 | /// This routine is only called when we have AVX2 and thus a reasonable | |||
| 17416 | /// instruction set for v32i8 shuffling.. | |||
| 17417 | static SDValue lowerV32I8Shuffle(const SDLoc &DL, ArrayRef<int> Mask, | |||
| 17418 | const APInt &Zeroable, SDValue V1, SDValue V2, | |||
| 17419 | const X86Subtarget &Subtarget, | |||
| 17420 | SelectionDAG &DAG) { | |||
| 17421 | assert(V1.getSimpleValueType() == MVT::v32i8 && "Bad operand type!")((void)0); | |||
| 17422 | assert(V2.getSimpleValueType() == MVT::v32i8 && "Bad operand type!")((void)0); | |||
| 17423 | assert(Mask.size() == 32 && "Unexpected mask size for v32 shuffle!")((void)0); | |||
| 17424 | assert(Subtarget.hasAVX2() && "We can only lower v32i8 with AVX2!")((void)0); | |||
| 17425 | ||||
| 17426 | // Whenever we can lower this as a zext, that instruction is strictly faster | |||
| 17427 | // than any alternative. It also allows us to fold memory operands into the | |||
| 17428 | // shuffle in many cases. | |||
| 17429 | if (SDValue ZExt = lowerShuffleAsZeroOrAnyExtend(DL, MVT::v32i8, V1, V2, Mask, | |||
| 17430 | Zeroable, Subtarget, DAG)) | |||
| 17431 | return ZExt; | |||
| 17432 | ||||
| 17433 | // Check for being able to broadcast a single element. | |||
| 17434 | if (SDValue Broadcast = lowerShuffleAsBroadcast(DL, MVT::v32i8, V1, V2, Mask, | |||
| 17435 | Subtarget, DAG)) | |||
| 17436 | return Broadcast; | |||
| 17437 | ||||
| 17438 | if (SDValue Blend = lowerShuffleAsBlend(DL, MVT::v32i8, V1, V2, Mask, | |||
| 17439 | Zeroable, Subtarget, DAG)) | |||
| 17440 | return Blend; | |||
| 17441 | ||||
| 17442 | // Use dedicated unpack instructions for masks that match their pattern. | |||
| 17443 | if (SDValue V = lowerShuffleWithUNPCK(DL, MVT::v32i8, Mask, V1, V2, DAG)) | |||
| 17444 | return V; | |||
| 17445 | ||||
| 17446 | // Use dedicated pack instructions for masks that match their pattern. | |||
| 17447 | if (SDValue V = lowerShuffleWithPACK(DL, MVT::v32i8, Mask, V1, V2, DAG, | |||
| 17448 | Subtarget)) | |||
| 17449 | return V; | |||
| 17450 | ||||
| 17451 | // Try to use lower using a truncation. | |||
| 17452 | if (SDValue V = lowerShuffleAsVTRUNC(DL, MVT::v32i8, V1, V2, Mask, Zeroable, | |||
| 17453 | Subtarget, DAG)) | |||
| 17454 | return V; | |||
| 17455 | ||||
| 17456 | // Try to use shift instructions. | |||
| 17457 | if (SDValue Shift = lowerShuffleAsShift(DL, MVT::v32i8, V1, V2, Mask, | |||
| 17458 | Zeroable, Subtarget, DAG)) | |||
| 17459 | return Shift; | |||
| 17460 | ||||
| 17461 | // Try to use byte rotation instructions. | |||
| 17462 | if (SDValue Rotate = lowerShuffleAsByteRotate(DL, MVT::v32i8, V1, V2, Mask, | |||
| 17463 | Subtarget, DAG)) | |||
| 17464 | return Rotate; | |||
| 17465 | ||||
| 17466 | // Try to use bit rotation instructions. | |||
| 17467 | if (V2.isUndef()) | |||
| 17468 | if (SDValue Rotate = | |||
| 17469 | lowerShuffleAsBitRotate(DL, MVT::v32i8, V1, Mask, Subtarget, DAG)) | |||
| 17470 | return Rotate; | |||
| 17471 | ||||
| 17472 | // Try to create an in-lane repeating shuffle mask and then shuffle the | |||
| 17473 | // results into the target lanes. | |||
| 17474 | if (SDValue V = lowerShuffleAsRepeatedMaskAndLanePermute( | |||
| 17475 | DL, MVT::v32i8, V1, V2, Mask, Subtarget, DAG)) | |||
| 17476 | return V; | |||
| 17477 | ||||
| 17478 | // There are no generalized cross-lane shuffle operations available on i8 | |||
| 17479 | // element types. | |||
| 17480 | if (V2.isUndef() && is128BitLaneCrossingShuffleMask(MVT::v32i8, Mask)) { | |||
| 17481 | // Try to produce a fixed cross-128-bit lane permute followed by unpack | |||
| 17482 | // because that should be faster than the variable permute alternatives. | |||
| 17483 | if (SDValue V = lowerShuffleWithUNPCK256(DL, MVT::v32i8, Mask, V1, V2, DAG)) | |||
| 17484 | return V; | |||
| 17485 | ||||
| 17486 | if (SDValue V = lowerShuffleAsLanePermuteAndPermute( | |||
| 17487 | DL, MVT::v32i8, V1, V2, Mask, DAG, Subtarget)) | |||
| 17488 | return V; | |||
| 17489 | ||||
| 17490 | return lowerShuffleAsLanePermuteAndShuffle(DL, MVT::v32i8, V1, V2, Mask, | |||
| 17491 | DAG, Subtarget); | |||
| 17492 | } | |||
| 17493 | ||||
| 17494 | if (SDValue PSHUFB = lowerShuffleWithPSHUFB(DL, MVT::v32i8, Mask, V1, V2, | |||
| 17495 | Zeroable, Subtarget, DAG)) | |||
| 17496 | return PSHUFB; | |||
| 17497 | ||||
| 17498 | // AVX512VBMI can lower to VPERMB (non-VLX will pad to v64i8). | |||
| 17499 | if (Subtarget.hasVBMI()) | |||
| 17500 | return lowerShuffleWithPERMV(DL, MVT::v32i8, Mask, V1, V2, Subtarget, DAG); | |||
| 17501 | ||||
| 17502 | // Try to simplify this by merging 128-bit lanes to enable a lane-based | |||
| 17503 | // shuffle. | |||
| 17504 | if (SDValue Result = lowerShuffleAsLanePermuteAndRepeatedMask( | |||
| 17505 | DL, MVT::v32i8, V1, V2, Mask, Subtarget, DAG)) | |||
| 17506 | return Result; | |||
| 17507 | ||||
| 17508 | // Try to permute the lanes and then use a per-lane permute. | |||
| 17509 | if (SDValue V = lowerShuffleAsLanePermuteAndPermute( | |||
| 17510 | DL, MVT::v32i8, V1, V2, Mask, DAG, Subtarget)) | |||
| 17511 | return V; | |||
| 17512 | ||||
| 17513 | // Look for {0, 8, 16, 24, 32, 40, 48, 56 } in the first 8 elements. Followed | |||
| 17514 | // by zeroable elements in the remaining 24 elements. Turn this into two | |||
| 17515 | // vmovqb instructions shuffled together. | |||
| 17516 | if (Subtarget.hasVLX()) | |||
| 17517 | if (SDValue V = lowerShuffleAsVTRUNCAndUnpack(DL, MVT::v32i8, V1, V2, | |||
| 17518 | Mask, Zeroable, DAG)) | |||
| 17519 | return V; | |||
| 17520 | ||||
| 17521 | // Otherwise fall back on generic lowering. | |||
| 17522 | return lowerShuffleAsSplitOrBlend(DL, MVT::v32i8, V1, V2, Mask, | |||
| 17523 | Subtarget, DAG); | |||
| 17524 | } | |||
| 17525 | ||||
| 17526 | /// High-level routine to lower various 256-bit x86 vector shuffles. | |||
| 17527 | /// | |||
| 17528 | /// This routine either breaks down the specific type of a 256-bit x86 vector | |||
| 17529 | /// shuffle or splits it into two 128-bit shuffles and fuses the results back | |||
| 17530 | /// together based on the available instructions. | |||
| 17531 | static SDValue lower256BitShuffle(const SDLoc &DL, ArrayRef<int> Mask, MVT VT, | |||
| 17532 | SDValue V1, SDValue V2, const APInt &Zeroable, | |||
| 17533 | const X86Subtarget &Subtarget, | |||
| 17534 | SelectionDAG &DAG) { | |||
| 17535 | // If we have a single input to the zero element, insert that into V1 if we | |||
| 17536 | // can do so cheaply. | |||
| 17537 | int NumElts = VT.getVectorNumElements(); | |||
| 17538 | int NumV2Elements = count_if(Mask, [NumElts](int M) { return M >= NumElts; }); | |||
| 17539 | ||||
| 17540 | if (NumV2Elements == 1 && Mask[0] >= NumElts) | |||
| 17541 | if (SDValue Insertion = lowerShuffleAsElementInsertion( | |||
| 17542 | DL, VT, V1, V2, Mask, Zeroable, Subtarget, DAG)) | |||
| 17543 | return Insertion; | |||
| 17544 | ||||
| 17545 | // Handle special cases where the lower or upper half is UNDEF. | |||
| 17546 | if (SDValue V = | |||
| 17547 | lowerShuffleWithUndefHalf(DL, VT, V1, V2, Mask, Subtarget, DAG)) | |||
| 17548 | return V; | |||
| 17549 | ||||
| 17550 | // There is a really nice hard cut-over between AVX1 and AVX2 that means we | |||
| 17551 | // can check for those subtargets here and avoid much of the subtarget | |||
| 17552 | // querying in the per-vector-type lowering routines. With AVX1 we have | |||
| 17553 | // essentially *zero* ability to manipulate a 256-bit vector with integer | |||
| 17554 | // types. Since we'll use floating point types there eventually, just | |||
| 17555 | // immediately cast everything to a float and operate entirely in that domain. | |||
| 17556 | if (VT.isInteger() && !Subtarget.hasAVX2()) { | |||
| 17557 | int ElementBits = VT.getScalarSizeInBits(); | |||
| 17558 | if (ElementBits < 32) { | |||
| 17559 | // No floating point type available, if we can't use the bit operations | |||
| 17560 | // for masking/blending then decompose into 128-bit vectors. | |||
| 17561 | if (SDValue V = lowerShuffleAsBitMask(DL, VT, V1, V2, Mask, Zeroable, | |||
| 17562 | Subtarget, DAG)) | |||
| 17563 | return V; | |||
| 17564 | if (SDValue V = lowerShuffleAsBitBlend(DL, VT, V1, V2, Mask, DAG)) | |||
| 17565 | return V; | |||
| 17566 | return splitAndLowerShuffle(DL, VT, V1, V2, Mask, DAG); | |||
| 17567 | } | |||
| 17568 | ||||
| 17569 | MVT FpVT = MVT::getVectorVT(MVT::getFloatingPointVT(ElementBits), | |||
| 17570 | VT.getVectorNumElements()); | |||
| 17571 | V1 = DAG.getBitcast(FpVT, V1); | |||
| 17572 | V2 = DAG.getBitcast(FpVT, V2); | |||
| 17573 | return DAG.getBitcast(VT, DAG.getVectorShuffle(FpVT, DL, V1, V2, Mask)); | |||
| 17574 | } | |||
| 17575 | ||||
| 17576 | switch (VT.SimpleTy) { | |||
| 17577 | case MVT::v4f64: | |||
| 17578 | return lowerV4F64Shuffle(DL, Mask, Zeroable, V1, V2, Subtarget, DAG); | |||
| 17579 | case MVT::v4i64: | |||
| 17580 | return lowerV4I64Shuffle(DL, Mask, Zeroable, V1, V2, Subtarget, DAG); | |||
| 17581 | case MVT::v8f32: | |||
| 17582 | return lowerV8F32Shuffle(DL, Mask, Zeroable, V1, V2, Subtarget, DAG); | |||
| 17583 | case MVT::v8i32: | |||
| 17584 | return lowerV8I32Shuffle(DL, Mask, Zeroable, V1, V2, Subtarget, DAG); | |||
| 17585 | case MVT::v16i16: | |||
| 17586 | return lowerV16I16Shuffle(DL, Mask, Zeroable, V1, V2, Subtarget, DAG); | |||
| 17587 | case MVT::v32i8: | |||
| 17588 | return lowerV32I8Shuffle(DL, Mask, Zeroable, V1, V2, Subtarget, DAG); | |||
| 17589 | ||||
| 17590 | default: | |||
| 17591 | llvm_unreachable("Not a valid 256-bit x86 vector type!")__builtin_unreachable(); | |||
| 17592 | } | |||
| 17593 | } | |||
| 17594 | ||||
| 17595 | /// Try to lower a vector shuffle as a 128-bit shuffles. | |||
| 17596 | static SDValue lowerV4X128Shuffle(const SDLoc &DL, MVT VT, ArrayRef<int> Mask, | |||
| 17597 | const APInt &Zeroable, SDValue V1, SDValue V2, | |||
| 17598 | const X86Subtarget &Subtarget, | |||
| 17599 | SelectionDAG &DAG) { | |||
| 17600 | assert(VT.getScalarSizeInBits() == 64 &&((void)0) | |||
| 17601 | "Unexpected element type size for 128bit shuffle.")((void)0); | |||
| 17602 | ||||
| 17603 | // To handle 256 bit vector requires VLX and most probably | |||
| 17604 | // function lowerV2X128VectorShuffle() is better solution. | |||
| 17605 | assert(VT.is512BitVector() && "Unexpected vector size for 512bit shuffle.")((void)0); | |||
| 17606 | ||||
| 17607 | // TODO - use Zeroable like we do for lowerV2X128VectorShuffle? | |||
| 17608 | SmallVector<int, 4> Widened128Mask; | |||
| 17609 | if (!canWidenShuffleElements(Mask, Widened128Mask)) | |||
| 17610 | return SDValue(); | |||
| 17611 | assert(Widened128Mask.size() == 4 && "Shuffle widening mismatch")((void)0); | |||
| 17612 | ||||
| 17613 | // Try to use an insert into a zero vector. | |||
| 17614 | if (Widened128Mask[0] == 0 && (Zeroable & 0xf0) == 0xf0 && | |||
| 17615 | (Widened128Mask[1] == 1 || (Zeroable & 0x0c) == 0x0c)) { | |||
| 17616 | unsigned NumElts = ((Zeroable & 0x0c) == 0x0c) ? 2 : 4; | |||
| 17617 | MVT SubVT = MVT::getVectorVT(VT.getVectorElementType(), NumElts); | |||
| 17618 | SDValue LoV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT, V1, | |||
| 17619 | DAG.getIntPtrConstant(0, DL)); | |||
| 17620 | return DAG.getNode(ISD::INSERT_SUBVECTOR, DL, VT, | |||
| 17621 | getZeroVector(VT, Subtarget, DAG, DL), LoV, | |||
| 17622 | DAG.getIntPtrConstant(0, DL)); | |||
| 17623 | } | |||
| 17624 | ||||
| 17625 | // Check for patterns which can be matched with a single insert of a 256-bit | |||
| 17626 | // subvector. | |||
| 17627 | bool OnlyUsesV1 = isShuffleEquivalent(Mask, {0, 1, 2, 3, 0, 1, 2, 3}, V1, V2); | |||
| 17628 | if (OnlyUsesV1 || | |||
| 17629 | isShuffleEquivalent(Mask, {0, 1, 2, 3, 8, 9, 10, 11}, V1, V2)) { | |||
| 17630 | MVT SubVT = MVT::getVectorVT(VT.getVectorElementType(), 4); | |||
| 17631 | SDValue SubVec = | |||
| 17632 | DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT, OnlyUsesV1 ? V1 : V2, | |||
| 17633 | DAG.getIntPtrConstant(0, DL)); | |||
| 17634 | return DAG.getNode(ISD::INSERT_SUBVECTOR, DL, VT, V1, SubVec, | |||
| 17635 | DAG.getIntPtrConstant(4, DL)); | |||
| 17636 | } | |||
| 17637 | ||||
| 17638 | // See if this is an insertion of the lower 128-bits of V2 into V1. | |||
| 17639 | bool IsInsert = true; | |||
| 17640 | int V2Index = -1; | |||
| 17641 | for (int i = 0; i < 4; ++i) { | |||
| 17642 | assert(Widened128Mask[i] >= -1 && "Illegal shuffle sentinel value")((void)0); | |||
| 17643 | if (Widened128Mask[i] < 0) | |||
| 17644 | continue; | |||
| 17645 | ||||
| 17646 | // Make sure all V1 subvectors are in place. | |||
| 17647 | if (Widened128Mask[i] < 4) { | |||
| 17648 | if (Widened128Mask[i] != i) { | |||
| 17649 | IsInsert = false; | |||
| 17650 | break; | |||
| 17651 | } | |||
| 17652 | } else { | |||
| 17653 | // Make sure we only have a single V2 index and its the lowest 128-bits. | |||
| 17654 | if (V2Index >= 0 || Widened128Mask[i] != 4) { | |||
| 17655 | IsInsert = false; | |||
| 17656 | break; | |||
| 17657 | } | |||
| 17658 | V2Index = i; | |||
| 17659 | } | |||
| 17660 | } | |||
| 17661 | if (IsInsert && V2Index >= 0) { | |||
| 17662 | MVT SubVT = MVT::getVectorVT(VT.getVectorElementType(), 2); | |||
| 17663 | SDValue Subvec = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT, V2, | |||
| 17664 | DAG.getIntPtrConstant(0, DL)); | |||
| 17665 | return insert128BitVector(V1, Subvec, V2Index * 2, DAG, DL); | |||
| 17666 | } | |||
| 17667 | ||||
| 17668 | // See if we can widen to a 256-bit lane shuffle, we're going to lose 128-lane | |||
| 17669 | // UNDEF info by lowering to X86ISD::SHUF128 anyway, so by widening where | |||
| 17670 | // possible we at least ensure the lanes stay sequential to help later | |||
| 17671 | // combines. | |||
| 17672 | SmallVector<int, 2> Widened256Mask; | |||
| 17673 | if (canWidenShuffleElements(Widened128Mask, Widened256Mask)) { | |||
| 17674 | Widened128Mask.clear(); | |||
| 17675 | narrowShuffleMaskElts(2, Widened256Mask, Widened128Mask); | |||
| 17676 | } | |||
| 17677 | ||||
| 17678 | // Try to lower to vshuf64x2/vshuf32x4. | |||
| 17679 | SDValue Ops[2] = {DAG.getUNDEF(VT), DAG.getUNDEF(VT)}; | |||
| 17680 | unsigned PermMask = 0; | |||
| 17681 | // Insure elements came from the same Op. | |||
| 17682 | for (int i = 0; i < 4; ++i) { | |||
| 17683 | assert(Widened128Mask[i] >= -1 && "Illegal shuffle sentinel value")((void)0); | |||
| 17684 | if (Widened128Mask[i] < 0) | |||
| 17685 | continue; | |||
| 17686 | ||||
| 17687 | SDValue Op = Widened128Mask[i] >= 4 ? V2 : V1; | |||
| 17688 | unsigned OpIndex = i / 2; | |||
| 17689 | if (Ops[OpIndex].isUndef()) | |||
| 17690 | Ops[OpIndex] = Op; | |||
| 17691 | else if (Ops[OpIndex] != Op) | |||
| 17692 | return SDValue(); | |||
| 17693 | ||||
| 17694 | // Convert the 128-bit shuffle mask selection values into 128-bit selection | |||
| 17695 | // bits defined by a vshuf64x2 instruction's immediate control byte. | |||
| 17696 | PermMask |= (Widened128Mask[i] % 4) << (i * 2); | |||
| 17697 | } | |||
| 17698 | ||||
| 17699 | return DAG.getNode(X86ISD::SHUF128, DL, VT, Ops[0], Ops[1], | |||
| 17700 | DAG.getTargetConstant(PermMask, DL, MVT::i8)); | |||
| 17701 | } | |||
| 17702 | ||||
| 17703 | /// Handle lowering of 8-lane 64-bit floating point shuffles. | |||
| 17704 | static SDValue lowerV8F64Shuffle(const SDLoc &DL, ArrayRef<int> Mask, | |||
| 17705 | const APInt &Zeroable, SDValue V1, SDValue V2, | |||
| 17706 | const X86Subtarget &Subtarget, | |||
| 17707 | SelectionDAG &DAG) { | |||
| 17708 | assert(V1.getSimpleValueType() == MVT::v8f64 && "Bad operand type!")((void)0); | |||
| 17709 | assert(V2.getSimpleValueType() == MVT::v8f64 && "Bad operand type!")((void)0); | |||
| 17710 | assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!")((void)0); | |||
| 17711 | ||||
| 17712 | if (V2.isUndef()) { | |||
| 17713 | // Use low duplicate instructions for masks that match their pattern. | |||
| 17714 | if (isShuffleEquivalent(Mask, {0, 0, 2, 2, 4, 4, 6, 6}, V1, V2)) | |||
| 17715 | return DAG.getNode(X86ISD::MOVDDUP, DL, MVT::v8f64, V1); | |||
| 17716 | ||||
| 17717 | if (!is128BitLaneCrossingShuffleMask(MVT::v8f64, Mask)) { | |||
| 17718 | // Non-half-crossing single input shuffles can be lowered with an | |||
| 17719 | // interleaved permutation. | |||
| 17720 | unsigned VPERMILPMask = (Mask[0] == 1) | ((Mask[1] == 1) << 1) | | |||
| 17721 | ((Mask[2] == 3) << 2) | ((Mask[3] == 3) << 3) | | |||
| 17722 | ((Mask[4] == 5) << 4) | ((Mask[5] == 5) << 5) | | |||
| 17723 | ((Mask[6] == 7) << 6) | ((Mask[7] == 7) << 7); | |||
| 17724 | return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v8f64, V1, | |||
| 17725 | DAG.getTargetConstant(VPERMILPMask, DL, MVT::i8)); | |||
| 17726 | } | |||
| 17727 | ||||
| 17728 | SmallVector<int, 4> RepeatedMask; | |||
| 17729 | if (is256BitLaneRepeatedShuffleMask(MVT::v8f64, Mask, RepeatedMask)) | |||
| 17730 | return DAG.getNode(X86ISD::VPERMI, DL, MVT::v8f64, V1, | |||
| 17731 | getV4X86ShuffleImm8ForMask(RepeatedMask, DL, DAG)); | |||
| 17732 | } | |||
| 17733 | ||||
| 17734 | if (SDValue Shuf128 = lowerV4X128Shuffle(DL, MVT::v8f64, Mask, Zeroable, V1, | |||
| 17735 | V2, Subtarget, DAG)) | |||
| 17736 | return Shuf128; | |||
| 17737 | ||||
| 17738 | if (SDValue Unpck = lowerShuffleWithUNPCK(DL, MVT::v8f64, Mask, V1, V2, DAG)) | |||
| 17739 | return Unpck; | |||
| 17740 | ||||
| 17741 | // Check if the blend happens to exactly fit that of SHUFPD. | |||
| 17742 | if (SDValue Op = lowerShuffleWithSHUFPD(DL, MVT::v8f64, V1, V2, Mask, | |||
| 17743 | Zeroable, Subtarget, DAG)) | |||
| 17744 | return Op; | |||
| 17745 | ||||
| 17746 | if (SDValue V = lowerShuffleToEXPAND(DL, MVT::v8f64, Zeroable, Mask, V1, V2, | |||
| 17747 | DAG, Subtarget)) | |||
| 17748 | return V; | |||
| 17749 | ||||
| 17750 | if (SDValue Blend = lowerShuffleAsBlend(DL, MVT::v8f64, V1, V2, Mask, | |||
| 17751 | Zeroable, Subtarget, DAG)) | |||
| 17752 | return Blend; | |||
| 17753 | ||||
| 17754 | return lowerShuffleWithPERMV(DL, MVT::v8f64, Mask, V1, V2, Subtarget, DAG); | |||
| 17755 | } | |||
| 17756 | ||||
| 17757 | /// Handle lowering of 16-lane 32-bit floating point shuffles. | |||
| 17758 | static SDValue lowerV16F32Shuffle(const SDLoc &DL, ArrayRef<int> Mask, | |||
| 17759 | const APInt &Zeroable, SDValue V1, SDValue V2, | |||
| 17760 | const X86Subtarget &Subtarget, | |||
| 17761 | SelectionDAG &DAG) { | |||
| 17762 | assert(V1.getSimpleValueType() == MVT::v16f32 && "Bad operand type!")((void)0); | |||
| 17763 | assert(V2.getSimpleValueType() == MVT::v16f32 && "Bad operand type!")((void)0); | |||
| 17764 | assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!")((void)0); | |||
| 17765 | ||||
| 17766 | // If the shuffle mask is repeated in each 128-bit lane, we have many more | |||
| 17767 | // options to efficiently lower the shuffle. | |||
| 17768 | SmallVector<int, 4> RepeatedMask; | |||
| 17769 | if (is128BitLaneRepeatedShuffleMask(MVT::v16f32, Mask, RepeatedMask)) { | |||
| 17770 | assert(RepeatedMask.size() == 4 && "Unexpected repeated mask size!")((void)0); | |||
| 17771 | ||||
| 17772 | // Use even/odd duplicate instructions for masks that match their pattern. | |||
| 17773 | if (isShuffleEquivalent(RepeatedMask, {0, 0, 2, 2}, V1, V2)) | |||
| 17774 | return DAG.getNode(X86ISD::MOVSLDUP, DL, MVT::v16f32, V1); | |||
| 17775 | if (isShuffleEquivalent(RepeatedMask, {1, 1, 3, 3}, V1, V2)) | |||
| 17776 | return DAG.getNode(X86ISD::MOVSHDUP, DL, MVT::v16f32, V1); | |||
| 17777 | ||||
| 17778 | if (V2.isUndef()) | |||
| 17779 | return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v16f32, V1, | |||
| 17780 | getV4X86ShuffleImm8ForMask(RepeatedMask, DL, DAG)); | |||
| 17781 | ||||
| 17782 | // Use dedicated unpack instructions for masks that match their pattern. | |||
| 17783 | if (SDValue V = lowerShuffleWithUNPCK(DL, MVT::v16f32, Mask, V1, V2, DAG)) | |||
| 17784 | return V; | |||
| 17785 | ||||
| 17786 | if (SDValue Blend = lowerShuffleAsBlend(DL, MVT::v16f32, V1, V2, Mask, | |||
| 17787 | Zeroable, Subtarget, DAG)) | |||
| 17788 | return Blend; | |||
| 17789 | ||||
| 17790 | // Otherwise, fall back to a SHUFPS sequence. | |||
| 17791 | return lowerShuffleWithSHUFPS(DL, MVT::v16f32, RepeatedMask, V1, V2, DAG); | |||
| 17792 | } | |||
| 17793 | ||||
| 17794 | // Try to create an in-lane repeating shuffle mask and then shuffle the | |||
| 17795 | // results into the target lanes. | |||
| 17796 | if (SDValue V = lowerShuffleAsRepeatedMaskAndLanePermute( | |||
| 17797 | DL, MVT::v16f32, V1, V2, Mask, Subtarget, DAG)) | |||
| 17798 | return V; | |||
| 17799 | ||||
| 17800 | // If we have a single input shuffle with different shuffle patterns in the | |||
| 17801 | // 128-bit lanes and don't lane cross, use variable mask VPERMILPS. | |||
| 17802 | if (V2.isUndef() && | |||
| 17803 | !is128BitLaneCrossingShuffleMask(MVT::v16f32, Mask)) { | |||
| 17804 | SDValue VPermMask = getConstVector(Mask, MVT::v16i32, DAG, DL, true); | |||
| 17805 | return DAG.getNode(X86ISD::VPERMILPV, DL, MVT::v16f32, V1, VPermMask); | |||
| 17806 | } | |||
| 17807 | ||||
| 17808 | // If we have AVX512F support, we can use VEXPAND. | |||
| 17809 | if (SDValue V = lowerShuffleToEXPAND(DL, MVT::v16f32, Zeroable, Mask, | |||
| 17810 | V1, V2, DAG, Subtarget)) | |||
| 17811 | return V; | |||
| 17812 | ||||
| 17813 | return lowerShuffleWithPERMV(DL, MVT::v16f32, Mask, V1, V2, Subtarget, DAG); | |||
| 17814 | } | |||
| 17815 | ||||
| 17816 | /// Handle lowering of 8-lane 64-bit integer shuffles. | |||
| 17817 | static SDValue lowerV8I64Shuffle(const SDLoc &DL, ArrayRef<int> Mask, | |||
| 17818 | const APInt &Zeroable, SDValue V1, SDValue V2, | |||
| 17819 | const X86Subtarget &Subtarget, | |||
| 17820 | SelectionDAG &DAG) { | |||
| 17821 | assert(V1.getSimpleValueType() == MVT::v8i64 && "Bad operand type!")((void)0); | |||
| 17822 | assert(V2.getSimpleValueType() == MVT::v8i64 && "Bad operand type!")((void)0); | |||
| 17823 | assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!")((void)0); | |||
| 17824 | ||||
| 17825 | if (V2.isUndef()) { | |||
| 17826 | // When the shuffle is mirrored between the 128-bit lanes of the unit, we | |||
| 17827 | // can use lower latency instructions that will operate on all four | |||
| 17828 | // 128-bit lanes. | |||
| 17829 | SmallVector<int, 2> Repeated128Mask; | |||
| 17830 | if (is128BitLaneRepeatedShuffleMask(MVT::v8i64, Mask, Repeated128Mask)) { | |||
| 17831 | SmallVector<int, 4> PSHUFDMask; | |||
| 17832 | narrowShuffleMaskElts(2, Repeated128Mask, PSHUFDMask); | |||
| 17833 | return DAG.getBitcast( | |||
| 17834 | MVT::v8i64, | |||
| 17835 | DAG.getNode(X86ISD::PSHUFD, DL, MVT::v16i32, | |||
| 17836 | DAG.getBitcast(MVT::v16i32, V1), | |||
| 17837 | getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG))); | |||
| 17838 | } | |||
| 17839 | ||||
| 17840 | SmallVector<int, 4> Repeated256Mask; | |||
| 17841 | if (is256BitLaneRepeatedShuffleMask(MVT::v8i64, Mask, Repeated256Mask)) | |||
| 17842 | return DAG.getNode(X86ISD::VPERMI, DL, MVT::v8i64, V1, | |||
| 17843 | getV4X86ShuffleImm8ForMask(Repeated256Mask, DL, DAG)); | |||
| 17844 | } | |||
| 17845 | ||||
| 17846 | if (SDValue Shuf128 = lowerV4X128Shuffle(DL, MVT::v8i64, Mask, Zeroable, V1, | |||
| 17847 | V2, Subtarget, DAG)) | |||
| 17848 | return Shuf128; | |||
| 17849 | ||||
| 17850 | // Try to use shift instructions. | |||
| 17851 | if (SDValue Shift = lowerShuffleAsShift(DL, MVT::v8i64, V1, V2, Mask, | |||
| 17852 | Zeroable, Subtarget, DAG)) | |||
| 17853 | return Shift; | |||
| 17854 | ||||
| 17855 | // Try to use VALIGN. | |||
| 17856 | if (SDValue Rotate = lowerShuffleAsVALIGN(DL, MVT::v8i64, V1, V2, Mask, | |||
| 17857 | Subtarget, DAG)) | |||
| 17858 | return Rotate; | |||
| 17859 | ||||
| 17860 | // Try to use PALIGNR. | |||
| 17861 | if (Subtarget.hasBWI()) | |||
| 17862 | if (SDValue Rotate = lowerShuffleAsByteRotate(DL, MVT::v8i64, V1, V2, Mask, | |||
| 17863 | Subtarget, DAG)) | |||
| 17864 | return Rotate; | |||
| 17865 | ||||
| 17866 | if (SDValue Unpck = lowerShuffleWithUNPCK(DL, MVT::v8i64, Mask, V1, V2, DAG)) | |||
| 17867 | return Unpck; | |||
| 17868 | ||||
| 17869 | // If we have AVX512F support, we can use VEXPAND. | |||
| 17870 | if (SDValue V = lowerShuffleToEXPAND(DL, MVT::v8i64, Zeroable, Mask, V1, V2, | |||
| 17871 | DAG, Subtarget)) | |||
| 17872 | return V; | |||
| 17873 | ||||
| 17874 | if (SDValue Blend = lowerShuffleAsBlend(DL, MVT::v8i64, V1, V2, Mask, | |||
| 17875 | Zeroable, Subtarget, DAG)) | |||
| 17876 | return Blend; | |||
| 17877 | ||||
| 17878 | return lowerShuffleWithPERMV(DL, MVT::v8i64, Mask, V1, V2, Subtarget, DAG); | |||
| 17879 | } | |||
| 17880 | ||||
| 17881 | /// Handle lowering of 16-lane 32-bit integer shuffles. | |||
| 17882 | static SDValue lowerV16I32Shuffle(const SDLoc &DL, ArrayRef<int> Mask, | |||
| 17883 | const APInt &Zeroable, SDValue V1, SDValue V2, | |||
| 17884 | const X86Subtarget &Subtarget, | |||
| 17885 | SelectionDAG &DAG) { | |||
| 17886 | assert(V1.getSimpleValueType() == MVT::v16i32 && "Bad operand type!")((void)0); | |||
| 17887 | assert(V2.getSimpleValueType() == MVT::v16i32 && "Bad operand type!")((void)0); | |||
| 17888 | assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!")((void)0); | |||
| 17889 | ||||
| 17890 | // Whenever we can lower this as a zext, that instruction is strictly faster | |||
| 17891 | // than any alternative. It also allows us to fold memory operands into the | |||
| 17892 | // shuffle in many cases. | |||
| 17893 | if (SDValue ZExt = lowerShuffleAsZeroOrAnyExtend( | |||
| 17894 | DL, MVT::v16i32, V1, V2, Mask, Zeroable, Subtarget, DAG)) | |||
| 17895 | return ZExt; | |||
| 17896 | ||||
| 17897 | // If the shuffle mask is repeated in each 128-bit lane we can use more | |||
| 17898 | // efficient instructions that mirror the shuffles across the four 128-bit | |||
| 17899 | // lanes. | |||
| 17900 | SmallVector<int, 4> RepeatedMask; | |||
| 17901 | bool Is128BitLaneRepeatedShuffle = | |||
| 17902 | is128BitLaneRepeatedShuffleMask(MVT::v16i32, Mask, RepeatedMask); | |||
| 17903 | if (Is128BitLaneRepeatedShuffle) { | |||
| 17904 | assert(RepeatedMask.size() == 4 && "Unexpected repeated mask size!")((void)0); | |||
| 17905 | if (V2.isUndef()) | |||
| 17906 | return DAG.getNode(X86ISD::PSHUFD, DL, MVT::v16i32, V1, | |||
| 17907 | getV4X86ShuffleImm8ForMask(RepeatedMask, DL, DAG)); | |||
| 17908 | ||||
| 17909 | // Use dedicated unpack instructions for masks that match their pattern. | |||
| 17910 | if (SDValue V = lowerShuffleWithUNPCK(DL, MVT::v16i32, Mask, V1, V2, DAG)) | |||
| 17911 | return V; | |||
| 17912 | } | |||
| 17913 | ||||
| 17914 | // Try to use shift instructions. | |||
| 17915 | if (SDValue Shift = lowerShuffleAsShift(DL, MVT::v16i32, V1, V2, Mask, | |||
| 17916 | Zeroable, Subtarget, DAG)) | |||
| 17917 | return Shift; | |||
| 17918 | ||||
| 17919 | // Try to use VALIGN. | |||
| 17920 | if (SDValue Rotate = lowerShuffleAsVALIGN(DL, MVT::v16i32, V1, V2, Mask, | |||
| 17921 | Subtarget, DAG)) | |||
| 17922 | return Rotate; | |||
| 17923 | ||||
| 17924 | // Try to use byte rotation instructions. | |||
| 17925 | if (Subtarget.hasBWI()) | |||
| 17926 | if (SDValue Rotate = lowerShuffleAsByteRotate(DL, MVT::v16i32, V1, V2, Mask, | |||
| 17927 | Subtarget, DAG)) | |||
| 17928 | return Rotate; | |||
| 17929 | ||||
| 17930 | // Assume that a single SHUFPS is faster than using a permv shuffle. | |||
| 17931 | // If some CPU is harmed by the domain switch, we can fix it in a later pass. | |||
| 17932 | if (Is128BitLaneRepeatedShuffle && isSingleSHUFPSMask(RepeatedMask)) { | |||
| 17933 | SDValue CastV1 = DAG.getBitcast(MVT::v16f32, V1); | |||
| 17934 | SDValue CastV2 = DAG.getBitcast(MVT::v16f32, V2); | |||
| 17935 | SDValue ShufPS = lowerShuffleWithSHUFPS(DL, MVT::v16f32, RepeatedMask, | |||
| 17936 | CastV1, CastV2, DAG); | |||
| 17937 | return DAG.getBitcast(MVT::v16i32, ShufPS); | |||
| 17938 | } | |||
| 17939 | ||||
| 17940 | // Try to create an in-lane repeating shuffle mask and then shuffle the | |||
| 17941 | // results into the target lanes. | |||
| 17942 | if (SDValue V = lowerShuffleAsRepeatedMaskAndLanePermute( | |||
| 17943 | DL, MVT::v16i32, V1, V2, Mask, Subtarget, DAG)) | |||
| 17944 | return V; | |||
| 17945 | ||||
| 17946 | // If we have AVX512F support, we can use VEXPAND. | |||
| 17947 | if (SDValue V = lowerShuffleToEXPAND(DL, MVT::v16i32, Zeroable, Mask, V1, V2, | |||
| 17948 | DAG, Subtarget)) | |||
| 17949 | return V; | |||
| 17950 | ||||
| 17951 | if (SDValue Blend = lowerShuffleAsBlend(DL, MVT::v16i32, V1, V2, Mask, | |||
| 17952 | Zeroable, Subtarget, DAG)) | |||
| 17953 | return Blend; | |||
| 17954 | ||||
| 17955 | return lowerShuffleWithPERMV(DL, MVT::v16i32, Mask, V1, V2, Subtarget, DAG); | |||
| 17956 | } | |||
| 17957 | ||||
| 17958 | /// Handle lowering of 32-lane 16-bit integer shuffles. | |||
| 17959 | static SDValue lowerV32I16Shuffle(const SDLoc &DL, ArrayRef<int> Mask, | |||
| 17960 | const APInt &Zeroable, SDValue V1, SDValue V2, | |||
| 17961 | const X86Subtarget &Subtarget, | |||
| 17962 | SelectionDAG &DAG) { | |||
| 17963 | assert(V1.getSimpleValueType() == MVT::v32i16 && "Bad operand type!")((void)0); | |||
| 17964 | assert(V2.getSimpleValueType() == MVT::v32i16 && "Bad operand type!")((void)0); | |||
| 17965 | assert(Mask.size() == 32 && "Unexpected mask size for v32 shuffle!")((void)0); | |||
| 17966 | assert(Subtarget.hasBWI() && "We can only lower v32i16 with AVX-512-BWI!")((void)0); | |||
| 17967 | ||||
| 17968 | // Whenever we can lower this as a zext, that instruction is strictly faster | |||
| 17969 | // than any alternative. It also allows us to fold memory operands into the | |||
| 17970 | // shuffle in many cases. | |||
| 17971 | if (SDValue ZExt = lowerShuffleAsZeroOrAnyExtend( | |||
| 17972 | DL, MVT::v32i16, V1, V2, Mask, Zeroable, Subtarget, DAG)) | |||
| 17973 | return ZExt; | |||
| 17974 | ||||
| 17975 | // Use dedicated unpack instructions for masks that match their pattern. | |||
| 17976 | if (SDValue V = lowerShuffleWithUNPCK(DL, MVT::v32i16, Mask, V1, V2, DAG)) | |||
| 17977 | return V; | |||
| 17978 | ||||
| 17979 | // Use dedicated pack instructions for masks that match their pattern. | |||
| 17980 | if (SDValue V = | |||
| 17981 | lowerShuffleWithPACK(DL, MVT::v32i16, Mask, V1, V2, DAG, Subtarget)) | |||
| 17982 | return V; | |||
| 17983 | ||||
| 17984 | // Try to use shift instructions. | |||
| 17985 | if (SDValue Shift = lowerShuffleAsShift(DL, MVT::v32i16, V1, V2, Mask, | |||
| 17986 | Zeroable, Subtarget, DAG)) | |||
| 17987 | return Shift; | |||
| 17988 | ||||
| 17989 | // Try to use byte rotation instructions. | |||
| 17990 | if (SDValue Rotate = lowerShuffleAsByteRotate(DL, MVT::v32i16, V1, V2, Mask, | |||
| 17991 | Subtarget, DAG)) | |||
| 17992 | return Rotate; | |||
| 17993 | ||||
| 17994 | if (V2.isUndef()) { | |||
| 17995 | // Try to use bit rotation instructions. | |||
| 17996 | if (SDValue Rotate = | |||
| 17997 | lowerShuffleAsBitRotate(DL, MVT::v32i16, V1, Mask, Subtarget, DAG)) | |||
| 17998 | return Rotate; | |||
| 17999 | ||||
| 18000 | SmallVector<int, 8> RepeatedMask; | |||
| 18001 | if (is128BitLaneRepeatedShuffleMask(MVT::v32i16, Mask, RepeatedMask)) { | |||
| 18002 | // As this is a single-input shuffle, the repeated mask should be | |||
| 18003 | // a strictly valid v8i16 mask that we can pass through to the v8i16 | |||
| 18004 | // lowering to handle even the v32 case. | |||
| 18005 | return lowerV8I16GeneralSingleInputShuffle(DL, MVT::v32i16, V1, | |||
| 18006 | RepeatedMask, Subtarget, DAG); | |||
| 18007 | } | |||
| 18008 | } | |||
| 18009 | ||||
| 18010 | if (SDValue Blend = lowerShuffleAsBlend(DL, MVT::v32i16, V1, V2, Mask, | |||
| 18011 | Zeroable, Subtarget, DAG)) | |||
| 18012 | return Blend; | |||
| 18013 | ||||
| 18014 | if (SDValue PSHUFB = lowerShuffleWithPSHUFB(DL, MVT::v32i16, Mask, V1, V2, | |||
| 18015 | Zeroable, Subtarget, DAG)) | |||
| 18016 | return PSHUFB; | |||
| 18017 | ||||
| 18018 | return lowerShuffleWithPERMV(DL, MVT::v32i16, Mask, V1, V2, Subtarget, DAG); | |||
| 18019 | } | |||
| 18020 | ||||
| 18021 | /// Handle lowering of 64-lane 8-bit integer shuffles. | |||
| 18022 | static SDValue lowerV64I8Shuffle(const SDLoc &DL, ArrayRef<int> Mask, | |||
| 18023 | const APInt &Zeroable, SDValue V1, SDValue V2, | |||
| 18024 | const X86Subtarget &Subtarget, | |||
| 18025 | SelectionDAG &DAG) { | |||
| 18026 | assert(V1.getSimpleValueType() == MVT::v64i8 && "Bad operand type!")((void)0); | |||
| 18027 | assert(V2.getSimpleValueType() == MVT::v64i8 && "Bad operand type!")((void)0); | |||
| 18028 | assert(Mask.size() == 64 && "Unexpected mask size for v64 shuffle!")((void)0); | |||
| 18029 | assert(Subtarget.hasBWI() && "We can only lower v64i8 with AVX-512-BWI!")((void)0); | |||
| 18030 | ||||
| 18031 | // Whenever we can lower this as a zext, that instruction is strictly faster | |||
| 18032 | // than any alternative. It also allows us to fold memory operands into the | |||
| 18033 | // shuffle in many cases. | |||
| 18034 | if (SDValue ZExt = lowerShuffleAsZeroOrAnyExtend( | |||
| 18035 | DL, MVT::v64i8, V1, V2, Mask, Zeroable, Subtarget, DAG)) | |||
| 18036 | return ZExt; | |||
| 18037 | ||||
| 18038 | // Use dedicated unpack instructions for masks that match their pattern. | |||
| 18039 | if (SDValue V = lowerShuffleWithUNPCK(DL, MVT::v64i8, Mask, V1, V2, DAG)) | |||
| 18040 | return V; | |||
| 18041 | ||||
| 18042 | // Use dedicated pack instructions for masks that match their pattern. | |||
| 18043 | if (SDValue V = lowerShuffleWithPACK(DL, MVT::v64i8, Mask, V1, V2, DAG, | |||
| 18044 | Subtarget)) | |||
| 18045 | return V; | |||
| 18046 | ||||
| 18047 | // Try to use shift instructions. | |||
| 18048 | if (SDValue Shift = lowerShuffleAsShift(DL, MVT::v64i8, V1, V2, Mask, | |||
| 18049 | Zeroable, Subtarget, DAG)) | |||
| 18050 | return Shift; | |||
| 18051 | ||||
| 18052 | // Try to use byte rotation instructions. | |||
| 18053 | if (SDValue Rotate = lowerShuffleAsByteRotate(DL, MVT::v64i8, V1, V2, Mask, | |||
| 18054 | Subtarget, DAG)) | |||
| 18055 | return Rotate; | |||
| 18056 | ||||
| 18057 | // Try to use bit rotation instructions. | |||
| 18058 | if (V2.isUndef()) | |||
| 18059 | if (SDValue Rotate = | |||
| 18060 | lowerShuffleAsBitRotate(DL, MVT::v64i8, V1, Mask, Subtarget, DAG)) | |||
| 18061 | return Rotate; | |||
| 18062 | ||||
| 18063 | // Lower as AND if possible. | |||
| 18064 | if (SDValue Masked = lowerShuffleAsBitMask(DL, MVT::v64i8, V1, V2, Mask, | |||
| 18065 | Zeroable, Subtarget, DAG)) | |||
| 18066 | return Masked; | |||
| 18067 | ||||
| 18068 | if (SDValue PSHUFB = lowerShuffleWithPSHUFB(DL, MVT::v64i8, Mask, V1, V2, | |||
| 18069 | Zeroable, Subtarget, DAG)) | |||
| 18070 | return PSHUFB; | |||
| 18071 | ||||
| 18072 | // VBMI can use VPERMV/VPERMV3 byte shuffles. | |||
| 18073 | if (Subtarget.hasVBMI()) | |||
| 18074 | return lowerShuffleWithPERMV(DL, MVT::v64i8, Mask, V1, V2, Subtarget, DAG); | |||
| 18075 | ||||
| 18076 | // Try to create an in-lane repeating shuffle mask and then shuffle the | |||
| 18077 | // results into the target lanes. | |||
| 18078 | if (SDValue V = lowerShuffleAsRepeatedMaskAndLanePermute( | |||
| 18079 | DL, MVT::v64i8, V1, V2, Mask, Subtarget, DAG)) | |||
| 18080 | return V; | |||
| 18081 | ||||
| 18082 | if (SDValue Blend = lowerShuffleAsBlend(DL, MVT::v64i8, V1, V2, Mask, | |||
| 18083 | Zeroable, Subtarget, DAG)) | |||
| 18084 | return Blend; | |||
| 18085 | ||||
| 18086 | // Try to simplify this by merging 128-bit lanes to enable a lane-based | |||
| 18087 | // shuffle. | |||
| 18088 | if (!V2.isUndef()) | |||
| 18089 | if (SDValue Result = lowerShuffleAsLanePermuteAndRepeatedMask( | |||
| 18090 | DL, MVT::v64i8, V1, V2, Mask, Subtarget, DAG)) | |||
| 18091 | return Result; | |||
| 18092 | ||||
| 18093 | // FIXME: Implement direct support for this type! | |||
| 18094 | return splitAndLowerShuffle(DL, MVT::v64i8, V1, V2, Mask, DAG); | |||
| 18095 | } | |||
| 18096 | ||||
| 18097 | /// High-level routine to lower various 512-bit x86 vector shuffles. | |||
| 18098 | /// | |||
| 18099 | /// This routine either breaks down the specific type of a 512-bit x86 vector | |||
| 18100 | /// shuffle or splits it into two 256-bit shuffles and fuses the results back | |||
| 18101 | /// together based on the available instructions. | |||
| 18102 | static SDValue lower512BitShuffle(const SDLoc &DL, ArrayRef<int> Mask, | |||
| 18103 | MVT VT, SDValue V1, SDValue V2, | |||
| 18104 | const APInt &Zeroable, | |||
| 18105 | const X86Subtarget &Subtarget, | |||
| 18106 | SelectionDAG &DAG) { | |||
| 18107 | assert(Subtarget.hasAVX512() &&((void)0) | |||
| 18108 | "Cannot lower 512-bit vectors w/ basic ISA!")((void)0); | |||
| 18109 | ||||
| 18110 | // If we have a single input to the zero element, insert that into V1 if we | |||
| 18111 | // can do so cheaply. | |||
| 18112 | int NumElts = Mask.size(); | |||
| 18113 | int NumV2Elements = count_if(Mask, [NumElts](int M) { return M >= NumElts; }); | |||
| 18114 | ||||
| 18115 | if (NumV2Elements == 1 && Mask[0] >= NumElts) | |||
| 18116 | if (SDValue Insertion = lowerShuffleAsElementInsertion( | |||
| 18117 | DL, VT, V1, V2, Mask, Zeroable, Subtarget, DAG)) | |||
| 18118 | return Insertion; | |||
| 18119 | ||||
| 18120 | // Handle special cases where the lower or upper half is UNDEF. | |||
| 18121 | if (SDValue V = | |||
| 18122 | lowerShuffleWithUndefHalf(DL, VT, V1, V2, Mask, Subtarget, DAG)) | |||
| 18123 | return V; | |||
| 18124 | ||||
| 18125 | // Check for being able to broadcast a single element. | |||
| 18126 | if (SDValue Broadcast = lowerShuffleAsBroadcast(DL, VT, V1, V2, Mask, | |||
| 18127 | Subtarget, DAG)) | |||
| 18128 | return Broadcast; | |||
| 18129 | ||||
| 18130 | if ((VT == MVT::v32i16 || VT == MVT::v64i8) && !Subtarget.hasBWI()) { | |||
| 18131 | // Try using bit ops for masking and blending before falling back to | |||
| 18132 | // splitting. | |||
| 18133 | if (SDValue V = lowerShuffleAsBitMask(DL, VT, V1, V2, Mask, Zeroable, | |||
| 18134 | Subtarget, DAG)) | |||
| 18135 | return V; | |||
| 18136 | if (SDValue V = lowerShuffleAsBitBlend(DL, VT, V1, V2, Mask, DAG)) | |||
| 18137 | return V; | |||
| 18138 | ||||
| 18139 | return splitAndLowerShuffle(DL, VT, V1, V2, Mask, DAG); | |||
| 18140 | } | |||
| 18141 | ||||
| 18142 | // Dispatch to each element type for lowering. If we don't have support for | |||
| 18143 | // specific element type shuffles at 512 bits, immediately split them and | |||
| 18144 | // lower them. Each lowering routine of a given type is allowed to assume that | |||
| 18145 | // the requisite ISA extensions for that element type are available. | |||
| 18146 | switch (VT.SimpleTy) { | |||
| 18147 | case MVT::v8f64: | |||
| 18148 | return lowerV8F64Shuffle(DL, Mask, Zeroable, V1, V2, Subtarget, DAG); | |||
| 18149 | case MVT::v16f32: | |||
| 18150 | return lowerV16F32Shuffle(DL, Mask, Zeroable, V1, V2, Subtarget, DAG); | |||
| 18151 | case MVT::v8i64: | |||
| 18152 | return lowerV8I64Shuffle(DL, Mask, Zeroable, V1, V2, Subtarget, DAG); | |||
| 18153 | case MVT::v16i32: | |||
| 18154 | return lowerV16I32Shuffle(DL, Mask, Zeroable, V1, V2, Subtarget, DAG); | |||
| 18155 | case MVT::v32i16: | |||
| 18156 | return lowerV32I16Shuffle(DL, Mask, Zeroable, V1, V2, Subtarget, DAG); | |||
| 18157 | case MVT::v64i8: | |||
| 18158 | return lowerV64I8Shuffle(DL, Mask, Zeroable, V1, V2, Subtarget, DAG); | |||
| 18159 | ||||
| 18160 | default: | |||
| 18161 | llvm_unreachable("Not a valid 512-bit x86 vector type!")__builtin_unreachable(); | |||
| 18162 | } | |||
| 18163 | } | |||
| 18164 | ||||
| 18165 | static SDValue lower1BitShuffleAsKSHIFTR(const SDLoc &DL, ArrayRef<int> Mask, | |||
| 18166 | MVT VT, SDValue V1, SDValue V2, | |||
| 18167 | const X86Subtarget &Subtarget, | |||
| 18168 | SelectionDAG &DAG) { | |||
| 18169 | // Shuffle should be unary. | |||
| 18170 | if (!V2.isUndef()) | |||
| 18171 | return SDValue(); | |||
| 18172 | ||||
| 18173 | int ShiftAmt = -1; | |||
| 18174 | int NumElts = Mask.size(); | |||
| 18175 | for (int i = 0; i != NumElts; ++i) { | |||
| 18176 | int M = Mask[i]; | |||
| 18177 | assert((M == SM_SentinelUndef || (0 <= M && M < NumElts)) &&((void)0) | |||
| 18178 | "Unexpected mask index.")((void)0); | |||
| 18179 | if (M < 0) | |||
| 18180 | continue; | |||
| 18181 | ||||
| 18182 | // The first non-undef element determines our shift amount. | |||
| 18183 | if (ShiftAmt < 0) { | |||
| 18184 | ShiftAmt = M - i; | |||
| 18185 | // Need to be shifting right. | |||
| 18186 | if (ShiftAmt <= 0) | |||
| 18187 | return SDValue(); | |||
| 18188 | } | |||
| 18189 | // All non-undef elements must shift by the same amount. | |||
| 18190 | if (ShiftAmt != M - i) | |||
| 18191 | return SDValue(); | |||
| 18192 | } | |||
| 18193 | assert(ShiftAmt >= 0 && "All undef?")((void)0); | |||
| 18194 | ||||
| 18195 | // Great we found a shift right. | |||
| 18196 | MVT WideVT = VT; | |||
| 18197 | if ((!Subtarget.hasDQI() && NumElts == 8) || NumElts < 8) | |||
| 18198 | WideVT = Subtarget.hasDQI() ? MVT::v8i1 : MVT::v16i1; | |||
| 18199 | SDValue Res = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, WideVT, | |||
| 18200 | DAG.getUNDEF(WideVT), V1, | |||
| 18201 | DAG.getIntPtrConstant(0, DL)); | |||
| 18202 | Res = DAG.getNode(X86ISD::KSHIFTR, DL, WideVT, Res, | |||
| 18203 | DAG.getTargetConstant(ShiftAmt, DL, MVT::i8)); | |||
| 18204 | return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, Res, | |||
| 18205 | DAG.getIntPtrConstant(0, DL)); | |||
| 18206 | } | |||
| 18207 | ||||
| 18208 | // Determine if this shuffle can be implemented with a KSHIFT instruction. | |||
| 18209 | // Returns the shift amount if possible or -1 if not. This is a simplified | |||
| 18210 | // version of matchShuffleAsShift. | |||
| 18211 | static int match1BitShuffleAsKSHIFT(unsigned &Opcode, ArrayRef<int> Mask, | |||
| 18212 | int MaskOffset, const APInt &Zeroable) { | |||
| 18213 | int Size = Mask.size(); | |||
| 18214 | ||||
| 18215 | auto CheckZeros = [&](int Shift, bool Left) { | |||
| 18216 | for (int j = 0; j < Shift; ++j) | |||
| 18217 | if (!Zeroable[j + (Left ? 0 : (Size - Shift))]) | |||
| 18218 | return false; | |||
| 18219 | ||||
| 18220 | return true; | |||
| 18221 | }; | |||
| 18222 | ||||
| 18223 | auto MatchShift = [&](int Shift, bool Left) { | |||
| 18224 | unsigned Pos = Left ? Shift : 0; | |||
| 18225 | unsigned Low = Left ? 0 : Shift; | |||
| 18226 | unsigned Len = Size - Shift; | |||
| 18227 | return isSequentialOrUndefInRange(Mask, Pos, Len, Low + MaskOffset); | |||
| 18228 | }; | |||
| 18229 | ||||
| 18230 | for (int Shift = 1; Shift != Size; ++Shift) | |||
| 18231 | for (bool Left : {true, false}) | |||
| 18232 | if (CheckZeros(Shift, Left) && MatchShift(Shift, Left)) { | |||
| 18233 | Opcode = Left ? X86ISD::KSHIFTL : X86ISD::KSHIFTR; | |||
| 18234 | return Shift; | |||
| 18235 | } | |||
| 18236 | ||||
| 18237 | return -1; | |||
| 18238 | } | |||
| 18239 | ||||
| 18240 | ||||
| 18241 | // Lower vXi1 vector shuffles. | |||
| 18242 | // There is no a dedicated instruction on AVX-512 that shuffles the masks. | |||
| 18243 | // The only way to shuffle bits is to sign-extend the mask vector to SIMD | |||
| 18244 | // vector, shuffle and then truncate it back. | |||
| 18245 | static SDValue lower1BitShuffle(const SDLoc &DL, ArrayRef<int> Mask, | |||
| 18246 | MVT VT, SDValue V1, SDValue V2, | |||
| 18247 | const APInt &Zeroable, | |||
| 18248 | const X86Subtarget &Subtarget, | |||
| 18249 | SelectionDAG &DAG) { | |||
| 18250 | assert(Subtarget.hasAVX512() &&((void)0) | |||
| 18251 | "Cannot lower 512-bit vectors w/o basic ISA!")((void)0); | |||
| 18252 | ||||
| 18253 | int NumElts = Mask.size(); | |||
| 18254 | ||||
| 18255 | // Try to recognize shuffles that are just padding a subvector with zeros. | |||
| 18256 | int SubvecElts = 0; | |||
| 18257 | int Src = -1; | |||
| 18258 | for (int i = 0; i != NumElts; ++i) { | |||
| 18259 | if (Mask[i] >= 0) { | |||
| 18260 | // Grab the source from the first valid mask. All subsequent elements need | |||
| 18261 | // to use this same source. | |||
| 18262 | if (Src < 0) | |||
| 18263 | Src = Mask[i] / NumElts; | |||
| 18264 | if (Src != (Mask[i] / NumElts) || (Mask[i] % NumElts) != i) | |||
| 18265 | break; | |||
| 18266 | } | |||
| 18267 | ||||
| 18268 | ++SubvecElts; | |||
| 18269 | } | |||
| 18270 | assert(SubvecElts != NumElts && "Identity shuffle?")((void)0); | |||
| 18271 | ||||
| 18272 | // Clip to a power 2. | |||
| 18273 | SubvecElts = PowerOf2Floor(SubvecElts); | |||
| 18274 | ||||
| 18275 | // Make sure the number of zeroable bits in the top at least covers the bits | |||
| 18276 | // not covered by the subvector. | |||
| 18277 | if ((int)Zeroable.countLeadingOnes() >= (NumElts - SubvecElts)) { | |||
| 18278 | assert(Src >= 0 && "Expected a source!")((void)0); | |||
| 18279 | MVT ExtractVT = MVT::getVectorVT(MVT::i1, SubvecElts); | |||
| 18280 | SDValue Extract = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, ExtractVT, | |||
| 18281 | Src == 0 ? V1 : V2, | |||
| 18282 | DAG.getIntPtrConstant(0, DL)); | |||
| 18283 | return DAG.getNode(ISD::INSERT_SUBVECTOR, DL, VT, | |||
| 18284 | DAG.getConstant(0, DL, VT), | |||
| 18285 | Extract, DAG.getIntPtrConstant(0, DL)); | |||
| 18286 | } | |||
| 18287 | ||||
| 18288 | // Try a simple shift right with undef elements. Later we'll try with zeros. | |||
| 18289 | if (SDValue Shift = lower1BitShuffleAsKSHIFTR(DL, Mask, VT, V1, V2, Subtarget, | |||
| 18290 | DAG)) | |||
| 18291 | return Shift; | |||
| 18292 | ||||
| 18293 | // Try to match KSHIFTs. | |||
| 18294 | unsigned Offset = 0; | |||
| 18295 | for (SDValue V : { V1, V2 }) { | |||
| 18296 | unsigned Opcode; | |||
| 18297 | int ShiftAmt = match1BitShuffleAsKSHIFT(Opcode, Mask, Offset, Zeroable); | |||
| 18298 | if (ShiftAmt >= 0) { | |||
| 18299 | MVT WideVT = VT; | |||
| 18300 | if ((!Subtarget.hasDQI() && NumElts == 8) || NumElts < 8) | |||
| 18301 | WideVT = Subtarget.hasDQI() ? MVT::v8i1 : MVT::v16i1; | |||
| 18302 | SDValue Res = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, WideVT, | |||
| 18303 | DAG.getUNDEF(WideVT), V, | |||
| 18304 | DAG.getIntPtrConstant(0, DL)); | |||
| 18305 | // Widened right shifts need two shifts to ensure we shift in zeroes. | |||
| 18306 | if (Opcode == X86ISD::KSHIFTR && WideVT != VT) { | |||
| 18307 | int WideElts = WideVT.getVectorNumElements(); | |||
| 18308 | // Shift left to put the original vector in the MSBs of the new size. | |||
| 18309 | Res = DAG.getNode(X86ISD::KSHIFTL, DL, WideVT, Res, | |||
| 18310 | DAG.getTargetConstant(WideElts - NumElts, DL, MVT::i8)); | |||
| 18311 | // Increase the shift amount to account for the left shift. | |||
| 18312 | ShiftAmt += WideElts - NumElts; | |||
| 18313 | } | |||
| 18314 | ||||
| 18315 | Res = DAG.getNode(Opcode, DL, WideVT, Res, | |||
| 18316 | DAG.getTargetConstant(ShiftAmt, DL, MVT::i8)); | |||
| 18317 | return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, Res, | |||
| 18318 | DAG.getIntPtrConstant(0, DL)); | |||
| 18319 | } | |||
| 18320 | Offset += NumElts; // Increment for next iteration. | |||
| 18321 | } | |||
| 18322 | ||||
| 18323 | ||||
| 18324 | ||||
| 18325 | MVT ExtVT; | |||
| 18326 | switch (VT.SimpleTy) { | |||
| 18327 | default: | |||
| 18328 | llvm_unreachable("Expected a vector of i1 elements")__builtin_unreachable(); | |||
| 18329 | case MVT::v2i1: | |||
| 18330 | ExtVT = MVT::v2i64; | |||
| 18331 | break; | |||
| 18332 | case MVT::v4i1: | |||
| 18333 | ExtVT = MVT::v4i32; | |||
| 18334 | break; | |||
| 18335 | case MVT::v8i1: | |||
| 18336 | // Take 512-bit type, more shuffles on KNL. If we have VLX use a 256-bit | |||
| 18337 | // shuffle. | |||
| 18338 | ExtVT = Subtarget.hasVLX() ? MVT::v8i32 : MVT::v8i64; | |||
| 18339 | break; | |||
| 18340 | case MVT::v16i1: | |||
| 18341 | // Take 512-bit type, unless we are avoiding 512-bit types and have the | |||
| 18342 | // 256-bit operation available. | |||
| 18343 | ExtVT = Subtarget.canExtendTo512DQ() ? MVT::v16i32 : MVT::v16i16; | |||
| 18344 | break; | |||
| 18345 | case MVT::v32i1: | |||
| 18346 | // Take 512-bit type, unless we are avoiding 512-bit types and have the | |||
| 18347 | // 256-bit operation available. | |||
| 18348 | assert(Subtarget.hasBWI() && "Expected AVX512BW support")((void)0); | |||
| 18349 | ExtVT = Subtarget.canExtendTo512BW() ? MVT::v32i16 : MVT::v32i8; | |||
| 18350 | break; | |||
| 18351 | case MVT::v64i1: | |||
| 18352 | // Fall back to scalarization. FIXME: We can do better if the shuffle | |||
| 18353 | // can be partitioned cleanly. | |||
| 18354 | if (!Subtarget.useBWIRegs()) | |||
| 18355 | return SDValue(); | |||
| 18356 | ExtVT = MVT::v64i8; | |||
| 18357 | break; | |||
| 18358 | } | |||
| 18359 | ||||
| 18360 | V1 = DAG.getNode(ISD::SIGN_EXTEND, DL, ExtVT, V1); | |||
| 18361 | V2 = DAG.getNode(ISD::SIGN_EXTEND, DL, ExtVT, V2); | |||
| 18362 | ||||
| 18363 | SDValue Shuffle = DAG.getVectorShuffle(ExtVT, DL, V1, V2, Mask); | |||
| 18364 | // i1 was sign extended we can use X86ISD::CVT2MASK. | |||
| 18365 | int NumElems = VT.getVectorNumElements(); | |||
| 18366 | if ((Subtarget.hasBWI() && (NumElems >= 32)) || | |||
| 18367 | (Subtarget.hasDQI() && (NumElems < 32))) | |||
| 18368 | return DAG.getSetCC(DL, VT, DAG.getConstant(0, DL, ExtVT), | |||
| 18369 | Shuffle, ISD::SETGT); | |||
| 18370 | ||||
| 18371 | return DAG.getNode(ISD::TRUNCATE, DL, VT, Shuffle); | |||
| 18372 | } | |||
| 18373 | ||||
| 18374 | /// Helper function that returns true if the shuffle mask should be | |||
| 18375 | /// commuted to improve canonicalization. | |||
| 18376 | static bool canonicalizeShuffleMaskWithCommute(ArrayRef<int> Mask) { | |||
| 18377 | int NumElements = Mask.size(); | |||
| 18378 | ||||
| 18379 | int NumV1Elements = 0, NumV2Elements = 0; | |||
| 18380 | for (int M : Mask) | |||
| 18381 | if (M < 0) | |||
| 18382 | continue; | |||
| 18383 | else if (M < NumElements) | |||
| 18384 | ++NumV1Elements; | |||
| 18385 | else | |||
| 18386 | ++NumV2Elements; | |||
| 18387 | ||||
| 18388 | // Commute the shuffle as needed such that more elements come from V1 than | |||
| 18389 | // V2. This allows us to match the shuffle pattern strictly on how many | |||
| 18390 | // elements come from V1 without handling the symmetric cases. | |||
| 18391 | if (NumV2Elements > NumV1Elements) | |||
| 18392 | return true; | |||
| 18393 | ||||
| 18394 | assert(NumV1Elements > 0 && "No V1 indices")((void)0); | |||
| 18395 | ||||
| 18396 | if (NumV2Elements == 0) | |||
| 18397 | return false; | |||
| 18398 | ||||
| 18399 | // When the number of V1 and V2 elements are the same, try to minimize the | |||
| 18400 | // number of uses of V2 in the low half of the vector. When that is tied, | |||
| 18401 | // ensure that the sum of indices for V1 is equal to or lower than the sum | |||
| 18402 | // indices for V2. When those are equal, try to ensure that the number of odd | |||
| 18403 | // indices for V1 is lower than the number of odd indices for V2. | |||
| 18404 | if (NumV1Elements == NumV2Elements) { | |||
| 18405 | int LowV1Elements = 0, LowV2Elements = 0; | |||
| 18406 | for (int M : Mask.slice(0, NumElements / 2)) | |||
| 18407 | if (M >= NumElements) | |||
| 18408 | ++LowV2Elements; | |||
| 18409 | else if (M >= 0) | |||
| 18410 | ++LowV1Elements; | |||
| 18411 | if (LowV2Elements > LowV1Elements) | |||
| 18412 | return true; | |||
| 18413 | if (LowV2Elements == LowV1Elements) { | |||
| 18414 | int SumV1Indices = 0, SumV2Indices = 0; | |||
| 18415 | for (int i = 0, Size = Mask.size(); i < Size; ++i) | |||
| 18416 | if (Mask[i] >= NumElements) | |||
| 18417 | SumV2Indices += i; | |||
| 18418 | else if (Mask[i] >= 0) | |||
| 18419 | SumV1Indices += i; | |||
| 18420 | if (SumV2Indices < SumV1Indices) | |||
| 18421 | return true; | |||
| 18422 | if (SumV2Indices == SumV1Indices) { | |||
| 18423 | int NumV1OddIndices = 0, NumV2OddIndices = 0; | |||
| 18424 | for (int i = 0, Size = Mask.size(); i < Size; ++i) | |||
| 18425 | if (Mask[i] >= NumElements) | |||
| 18426 | NumV2OddIndices += i % 2; | |||
| 18427 | else if (Mask[i] >= 0) | |||
| 18428 | NumV1OddIndices += i % 2; | |||
| 18429 | if (NumV2OddIndices < NumV1OddIndices) | |||
| 18430 | return true; | |||
| 18431 | } | |||
| 18432 | } | |||
| 18433 | } | |||
| 18434 | ||||
| 18435 | return false; | |||
| 18436 | } | |||
| 18437 | ||||
| 18438 | /// Top-level lowering for x86 vector shuffles. | |||
| 18439 | /// | |||
| 18440 | /// This handles decomposition, canonicalization, and lowering of all x86 | |||
| 18441 | /// vector shuffles. Most of the specific lowering strategies are encapsulated | |||
| 18442 | /// above in helper routines. The canonicalization attempts to widen shuffles | |||
| 18443 | /// to involve fewer lanes of wider elements, consolidate symmetric patterns | |||
| 18444 | /// s.t. only one of the two inputs needs to be tested, etc. | |||
| 18445 | static SDValue lowerVECTOR_SHUFFLE(SDValue Op, const X86Subtarget &Subtarget, | |||
| 18446 | SelectionDAG &DAG) { | |||
| 18447 | ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op); | |||
| 18448 | ArrayRef<int> OrigMask = SVOp->getMask(); | |||
| 18449 | SDValue V1 = Op.getOperand(0); | |||
| 18450 | SDValue V2 = Op.getOperand(1); | |||
| 18451 | MVT VT = Op.getSimpleValueType(); | |||
| 18452 | int NumElements = VT.getVectorNumElements(); | |||
| 18453 | SDLoc DL(Op); | |||
| 18454 | bool Is1BitVector = (VT.getVectorElementType() == MVT::i1); | |||
| 18455 | ||||
| 18456 | assert((VT.getSizeInBits() != 64 || Is1BitVector) &&((void)0) | |||
| 18457 | "Can't lower MMX shuffles")((void)0); | |||
| 18458 | ||||
| 18459 | bool V1IsUndef = V1.isUndef(); | |||
| 18460 | bool V2IsUndef = V2.isUndef(); | |||
| 18461 | if (V1IsUndef && V2IsUndef) | |||
| 18462 | return DAG.getUNDEF(VT); | |||
| 18463 | ||||
| 18464 | // When we create a shuffle node we put the UNDEF node to second operand, | |||
| 18465 | // but in some cases the first operand may be transformed to UNDEF. | |||
| 18466 | // In this case we should just commute the node. | |||
| 18467 | if (V1IsUndef) | |||
| 18468 | return DAG.getCommutedVectorShuffle(*SVOp); | |||
| 18469 | ||||
| 18470 | // Check for non-undef masks pointing at an undef vector and make the masks | |||
| 18471 | // undef as well. This makes it easier to match the shuffle based solely on | |||
| 18472 | // the mask. | |||
| 18473 | if (V2IsUndef && | |||
| 18474 | any_of(OrigMask, [NumElements](int M) { return M >= NumElements; })) { | |||
| 18475 | SmallVector<int, 8> NewMask(OrigMask.begin(), OrigMask.end()); | |||
| 18476 | for (int &M : NewMask) | |||
| 18477 | if (M >= NumElements) | |||
| 18478 | M = -1; | |||
| 18479 | return DAG.getVectorShuffle(VT, DL, V1, V2, NewMask); | |||
| 18480 | } | |||
| 18481 | ||||
| 18482 | // Check for illegal shuffle mask element index values. | |||
| 18483 | int MaskUpperLimit = OrigMask.size() * (V2IsUndef ? 1 : 2); | |||
| 18484 | (void)MaskUpperLimit; | |||
| 18485 | assert(llvm::all_of(OrigMask,((void)0) | |||
| 18486 | [&](int M) { return -1 <= M && M < MaskUpperLimit; }) &&((void)0) | |||
| 18487 | "Out of bounds shuffle index")((void)0); | |||
| 18488 | ||||
| 18489 | // We actually see shuffles that are entirely re-arrangements of a set of | |||
| 18490 | // zero inputs. This mostly happens while decomposing complex shuffles into | |||
| 18491 | // simple ones. Directly lower these as a buildvector of zeros. | |||
| 18492 | APInt KnownUndef, KnownZero; | |||
| 18493 | computeZeroableShuffleElements(OrigMask, V1, V2, KnownUndef, KnownZero); | |||
| 18494 | ||||
| 18495 | APInt Zeroable = KnownUndef | KnownZero; | |||
| 18496 | if (Zeroable.isAllOnesValue()) | |||
| 18497 | return getZeroVector(VT, Subtarget, DAG, DL); | |||
| 18498 | ||||
| 18499 | bool V2IsZero = !V2IsUndef && ISD::isBuildVectorAllZeros(V2.getNode()); | |||
| 18500 | ||||
| 18501 | // Try to collapse shuffles into using a vector type with fewer elements but | |||
| 18502 | // wider element types. We cap this to not form integers or floating point | |||
| 18503 | // elements wider than 64 bits. It does not seem beneficial to form i128 | |||
| 18504 | // integers to handle flipping the low and high halves of AVX 256-bit vectors. | |||
| 18505 | SmallVector<int, 16> WidenedMask; | |||
| 18506 | if (VT.getScalarSizeInBits() < 64 && !Is1BitVector && | |||
| 18507 | canWidenShuffleElements(OrigMask, Zeroable, V2IsZero, WidenedMask)) { | |||
| 18508 | // Shuffle mask widening should not interfere with a broadcast opportunity | |||
| 18509 | // by obfuscating the operands with bitcasts. | |||
| 18510 | // TODO: Avoid lowering directly from this top-level function: make this | |||
| 18511 | // a query (canLowerAsBroadcast) and defer lowering to the type-based calls. | |||
| 18512 | if (SDValue Broadcast = lowerShuffleAsBroadcast(DL, VT, V1, V2, OrigMask, | |||
| 18513 | Subtarget, DAG)) | |||
| 18514 | return Broadcast; | |||
| 18515 | ||||
| 18516 | MVT NewEltVT = VT.isFloatingPoint() | |||
| 18517 | ? MVT::getFloatingPointVT(VT.getScalarSizeInBits() * 2) | |||
| 18518 | : MVT::getIntegerVT(VT.getScalarSizeInBits() * 2); | |||
| 18519 | int NewNumElts = NumElements / 2; | |||
| 18520 | MVT NewVT = MVT::getVectorVT(NewEltVT, NewNumElts); | |||
| 18521 | // Make sure that the new vector type is legal. For example, v2f64 isn't | |||
| 18522 | // legal on SSE1. | |||
| 18523 | if (DAG.getTargetLoweringInfo().isTypeLegal(NewVT)) { | |||
| 18524 | if (V2IsZero) { | |||
| 18525 | // Modify the new Mask to take all zeros from the all-zero vector. | |||
| 18526 | // Choose indices that are blend-friendly. | |||
| 18527 | bool UsedZeroVector = false; | |||
| 18528 | assert(is_contained(WidenedMask, SM_SentinelZero) &&((void)0) | |||
| 18529 | "V2's non-undef elements are used?!")((void)0); | |||
| 18530 | for (int i = 0; i != NewNumElts; ++i) | |||
| 18531 | if (WidenedMask[i] == SM_SentinelZero) { | |||
| 18532 | WidenedMask[i] = i + NewNumElts; | |||
| 18533 | UsedZeroVector = true; | |||
| 18534 | } | |||
| 18535 | // Ensure all elements of V2 are zero - isBuildVectorAllZeros permits | |||
| 18536 | // some elements to be undef. | |||
| 18537 | if (UsedZeroVector) | |||
| 18538 | V2 = getZeroVector(NewVT, Subtarget, DAG, DL); | |||
| 18539 | } | |||
| 18540 | V1 = DAG.getBitcast(NewVT, V1); | |||
| 18541 | V2 = DAG.getBitcast(NewVT, V2); | |||
| 18542 | return DAG.getBitcast( | |||
| 18543 | VT, DAG.getVectorShuffle(NewVT, DL, V1, V2, WidenedMask)); | |||
| 18544 | } | |||
| 18545 | } | |||
| 18546 | ||||
| 18547 | // Commute the shuffle if it will improve canonicalization. | |||
| 18548 | SmallVector<int, 64> Mask(OrigMask.begin(), OrigMask.end()); | |||
| 18549 | if (canonicalizeShuffleMaskWithCommute(Mask)) { | |||
| 18550 | ShuffleVectorSDNode::commuteMask(Mask); | |||
| 18551 | std::swap(V1, V2); | |||
| 18552 | } | |||
| 18553 | ||||
| 18554 | // For each vector width, delegate to a specialized lowering routine. | |||
| 18555 | if (VT.is128BitVector()) | |||
| 18556 | return lower128BitShuffle(DL, Mask, VT, V1, V2, Zeroable, Subtarget, DAG); | |||
| 18557 | ||||
| 18558 | if (VT.is256BitVector()) | |||
| 18559 | return lower256BitShuffle(DL, Mask, VT, V1, V2, Zeroable, Subtarget, DAG); | |||
| 18560 | ||||
| 18561 | if (VT.is512BitVector()) | |||
| 18562 | return lower512BitShuffle(DL, Mask, VT, V1, V2, Zeroable, Subtarget, DAG); | |||
| 18563 | ||||
| 18564 | if (Is1BitVector) | |||
| 18565 | return lower1BitShuffle(DL, Mask, VT, V1, V2, Zeroable, Subtarget, DAG); | |||
| 18566 | ||||
| 18567 | llvm_unreachable("Unimplemented!")__builtin_unreachable(); | |||
| 18568 | } | |||
| 18569 | ||||
| 18570 | /// Try to lower a VSELECT instruction to a vector shuffle. | |||
| 18571 | static SDValue lowerVSELECTtoVectorShuffle(SDValue Op, | |||
| 18572 | const X86Subtarget &Subtarget, | |||
| 18573 | SelectionDAG &DAG) { | |||
| 18574 | SDValue Cond = Op.getOperand(0); | |||
| 18575 | SDValue LHS = Op.getOperand(1); | |||
| 18576 | SDValue RHS = Op.getOperand(2); | |||
| 18577 | MVT VT = Op.getSimpleValueType(); | |||
| 18578 | ||||
| 18579 | // Only non-legal VSELECTs reach this lowering, convert those into generic | |||
| 18580 | // shuffles and re-use the shuffle lowering path for blends. | |||
| 18581 | if (ISD::isBuildVectorOfConstantSDNodes(Cond.getNode())) { | |||
| 18582 | SmallVector<int, 32> Mask; | |||
| 18583 | if (createShuffleMaskFromVSELECT(Mask, Cond)) | |||
| 18584 | return DAG.getVectorShuffle(VT, SDLoc(Op), LHS, RHS, Mask); | |||
| 18585 | } | |||
| 18586 | ||||
| 18587 | return SDValue(); | |||
| 18588 | } | |||
| 18589 | ||||
| 18590 | SDValue X86TargetLowering::LowerVSELECT(SDValue Op, SelectionDAG &DAG) const { | |||
| 18591 | SDValue Cond = Op.getOperand(0); | |||
| 18592 | SDValue LHS = Op.getOperand(1); | |||
| 18593 | SDValue RHS = Op.getOperand(2); | |||
| 18594 | ||||
| 18595 | // A vselect where all conditions and data are constants can be optimized into | |||
| 18596 | // a single vector load by SelectionDAGLegalize::ExpandBUILD_VECTOR(). | |||
| 18597 | if (ISD::isBuildVectorOfConstantSDNodes(Cond.getNode()) && | |||
| 18598 | ISD::isBuildVectorOfConstantSDNodes(LHS.getNode()) && | |||
| 18599 | ISD::isBuildVectorOfConstantSDNodes(RHS.getNode())) | |||
| 18600 | return SDValue(); | |||
| 18601 | ||||
| 18602 | // Try to lower this to a blend-style vector shuffle. This can handle all | |||
| 18603 | // constant condition cases. | |||
| 18604 | if (SDValue BlendOp = lowerVSELECTtoVectorShuffle(Op, Subtarget, DAG)) | |||
| 18605 | return BlendOp; | |||
| 18606 | ||||
| 18607 | // If this VSELECT has a vector if i1 as a mask, it will be directly matched | |||
| 18608 | // with patterns on the mask registers on AVX-512. | |||
| 18609 | MVT CondVT = Cond.getSimpleValueType(); | |||
| 18610 | unsigned CondEltSize = Cond.getScalarValueSizeInBits(); | |||
| 18611 | if (CondEltSize == 1) | |||
| 18612 | return Op; | |||
| 18613 | ||||
| 18614 | // Variable blends are only legal from SSE4.1 onward. | |||
| 18615 | if (!Subtarget.hasSSE41()) | |||
| 18616 | return SDValue(); | |||
| 18617 | ||||
| 18618 | SDLoc dl(Op); | |||
| 18619 | MVT VT = Op.getSimpleValueType(); | |||
| 18620 | unsigned EltSize = VT.getScalarSizeInBits(); | |||
| 18621 | unsigned NumElts = VT.getVectorNumElements(); | |||
| 18622 | ||||
| 18623 | // Expand v32i16/v64i8 without BWI. | |||
| 18624 | if ((VT == MVT::v32i16 || VT == MVT::v64i8) && !Subtarget.hasBWI()) | |||
| 18625 | return SDValue(); | |||
| 18626 | ||||
| 18627 | // If the VSELECT is on a 512-bit type, we have to convert a non-i1 condition | |||
| 18628 | // into an i1 condition so that we can use the mask-based 512-bit blend | |||
| 18629 | // instructions. | |||
| 18630 | if (VT.getSizeInBits() == 512) { | |||
| 18631 | // Build a mask by testing the condition against zero. | |||
| 18632 | MVT MaskVT = MVT::getVectorVT(MVT::i1, NumElts); | |||
| 18633 | SDValue Mask = DAG.getSetCC(dl, MaskVT, Cond, | |||
| 18634 | DAG.getConstant(0, dl, CondVT), | |||
| 18635 | ISD::SETNE); | |||
| 18636 | // Now return a new VSELECT using the mask. | |||
| 18637 | return DAG.getSelect(dl, VT, Mask, LHS, RHS); | |||
| 18638 | } | |||
| 18639 | ||||
| 18640 | // SEXT/TRUNC cases where the mask doesn't match the destination size. | |||
| 18641 | if (CondEltSize != EltSize) { | |||
| 18642 | // If we don't have a sign splat, rely on the expansion. | |||
| 18643 | if (CondEltSize != DAG.ComputeNumSignBits(Cond)) | |||
| 18644 | return SDValue(); | |||
| 18645 | ||||
| 18646 | MVT NewCondSVT = MVT::getIntegerVT(EltSize); | |||
| 18647 | MVT NewCondVT = MVT::getVectorVT(NewCondSVT, NumElts); | |||
| 18648 | Cond = DAG.getSExtOrTrunc(Cond, dl, NewCondVT); | |||
| 18649 | return DAG.getNode(ISD::VSELECT, dl, VT, Cond, LHS, RHS); | |||
| 18650 | } | |||
| 18651 | ||||
| 18652 | // Only some types will be legal on some subtargets. If we can emit a legal | |||
| 18653 | // VSELECT-matching blend, return Op, and but if we need to expand, return | |||
| 18654 | // a null value. | |||
| 18655 | switch (VT.SimpleTy) { | |||
| 18656 | default: | |||
| 18657 | // Most of the vector types have blends past SSE4.1. | |||
| 18658 | return Op; | |||
| 18659 | ||||
| 18660 | case MVT::v32i8: | |||
| 18661 | // The byte blends for AVX vectors were introduced only in AVX2. | |||
| 18662 | if (Subtarget.hasAVX2()) | |||
| 18663 | return Op; | |||
| 18664 | ||||
| 18665 | return SDValue(); | |||
| 18666 | ||||
| 18667 | case MVT::v8i16: | |||
| 18668 | case MVT::v16i16: { | |||
| 18669 | // Bitcast everything to the vXi8 type and use a vXi8 vselect. | |||
| 18670 | MVT CastVT = MVT::getVectorVT(MVT::i8, NumElts * 2); | |||
| 18671 | Cond = DAG.getBitcast(CastVT, Cond); | |||
| 18672 | LHS = DAG.getBitcast(CastVT, LHS); | |||
| 18673 | RHS = DAG.getBitcast(CastVT, RHS); | |||
| 18674 | SDValue Select = DAG.getNode(ISD::VSELECT, dl, CastVT, Cond, LHS, RHS); | |||
| 18675 | return DAG.getBitcast(VT, Select); | |||
| 18676 | } | |||
| 18677 | } | |||
| 18678 | } | |||
| 18679 | ||||
| 18680 | static SDValue LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG) { | |||
| 18681 | MVT VT = Op.getSimpleValueType(); | |||
| 18682 | SDValue Vec = Op.getOperand(0); | |||
| 18683 | SDValue Idx = Op.getOperand(1); | |||
| 18684 | assert(isa<ConstantSDNode>(Idx) && "Constant index expected")((void)0); | |||
| 18685 | SDLoc dl(Op); | |||
| 18686 | ||||
| 18687 | if (!Vec.getSimpleValueType().is128BitVector()) | |||
| 18688 | return SDValue(); | |||
| 18689 | ||||
| 18690 | if (VT.getSizeInBits() == 8) { | |||
| 18691 | // If IdxVal is 0, it's cheaper to do a move instead of a pextrb, unless | |||
| 18692 | // we're going to zero extend the register or fold the store. | |||
| 18693 | if (llvm::isNullConstant(Idx) && !MayFoldIntoZeroExtend(Op) && | |||
| 18694 | !MayFoldIntoStore(Op)) | |||
| 18695 | return DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, | |||
| 18696 | DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32, | |||
| 18697 | DAG.getBitcast(MVT::v4i32, Vec), Idx)); | |||
| 18698 | ||||
| 18699 | unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue(); | |||
| 18700 | SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32, Vec, | |||
| 18701 | DAG.getTargetConstant(IdxVal, dl, MVT::i8)); | |||
| 18702 | return DAG.getNode(ISD::TRUNCATE, dl, VT, Extract); | |||
| 18703 | } | |||
| 18704 | ||||
| 18705 | if (VT == MVT::f32) { | |||
| 18706 | // EXTRACTPS outputs to a GPR32 register which will require a movd to copy | |||
| 18707 | // the result back to FR32 register. It's only worth matching if the | |||
| 18708 | // result has a single use which is a store or a bitcast to i32. And in | |||
| 18709 | // the case of a store, it's not worth it if the index is a constant 0, | |||
| 18710 | // because a MOVSSmr can be used instead, which is smaller and faster. | |||
| 18711 | if (!Op.hasOneUse()) | |||
| 18712 | return SDValue(); | |||
| 18713 | SDNode *User = *Op.getNode()->use_begin(); | |||
| 18714 | if ((User->getOpcode() != ISD::STORE || isNullConstant(Idx)) && | |||
| 18715 | (User->getOpcode() != ISD::BITCAST || | |||
| 18716 | User->getValueType(0) != MVT::i32)) | |||
| 18717 | return SDValue(); | |||
| 18718 | SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32, | |||
| 18719 | DAG.getBitcast(MVT::v4i32, Vec), Idx); | |||
| 18720 | return DAG.getBitcast(MVT::f32, Extract); | |||
| 18721 | } | |||
| 18722 | ||||
| 18723 | if (VT == MVT::i32 || VT == MVT::i64) | |||
| 18724 | return Op; | |||
| 18725 | ||||
| 18726 | return SDValue(); | |||
| 18727 | } | |||
| 18728 | ||||
| 18729 | /// Extract one bit from mask vector, like v16i1 or v8i1. | |||
| 18730 | /// AVX-512 feature. | |||
| 18731 | static SDValue ExtractBitFromMaskVector(SDValue Op, SelectionDAG &DAG, | |||
| 18732 | const X86Subtarget &Subtarget) { | |||
| 18733 | SDValue Vec = Op.getOperand(0); | |||
| 18734 | SDLoc dl(Vec); | |||
| 18735 | MVT VecVT = Vec.getSimpleValueType(); | |||
| 18736 | SDValue Idx = Op.getOperand(1); | |||
| 18737 | auto* IdxC = dyn_cast<ConstantSDNode>(Idx); | |||
| 18738 | MVT EltVT = Op.getSimpleValueType(); | |||
| 18739 | ||||
| 18740 | assert((VecVT.getVectorNumElements() <= 16 || Subtarget.hasBWI()) &&((void)0) | |||
| 18741 | "Unexpected vector type in ExtractBitFromMaskVector")((void)0); | |||
| 18742 | ||||
| 18743 | // variable index can't be handled in mask registers, | |||
| 18744 | // extend vector to VR512/128 | |||
| 18745 | if (!IdxC) { | |||
| 18746 | unsigned NumElts = VecVT.getVectorNumElements(); | |||
| 18747 | // Extending v8i1/v16i1 to 512-bit get better performance on KNL | |||
| 18748 | // than extending to 128/256bit. | |||
| 18749 | MVT ExtEltVT = (NumElts <= 8) ? MVT::getIntegerVT(128 / NumElts) : MVT::i8; | |||
| 18750 | MVT ExtVecVT = MVT::getVectorVT(ExtEltVT, NumElts); | |||
| 18751 | SDValue Ext = DAG.getNode(ISD::SIGN_EXTEND, dl, ExtVecVT, Vec); | |||
| 18752 | SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ExtEltVT, Ext, Idx); | |||
| 18753 | return DAG.getNode(ISD::TRUNCATE, dl, EltVT, Elt); | |||
| 18754 | } | |||
| 18755 | ||||
| 18756 | unsigned IdxVal = IdxC->getZExtValue(); | |||
| 18757 | if (IdxVal == 0) // the operation is legal | |||
| 18758 | return Op; | |||
| 18759 | ||||
| 18760 | // Extend to natively supported kshift. | |||
| 18761 | unsigned NumElems = VecVT.getVectorNumElements(); | |||
| 18762 | MVT WideVecVT = VecVT; | |||
| 18763 | if ((!Subtarget.hasDQI() && NumElems == 8) || NumElems < 8) { | |||
| 18764 | WideVecVT = Subtarget.hasDQI() ? MVT::v8i1 : MVT::v16i1; | |||
| 18765 | Vec = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, WideVecVT, | |||
| 18766 | DAG.getUNDEF(WideVecVT), Vec, | |||
| 18767 | DAG.getIntPtrConstant(0, dl)); | |||
| 18768 | } | |||
| 18769 | ||||
| 18770 | // Use kshiftr instruction to move to the lower element. | |||
| 18771 | Vec = DAG.getNode(X86ISD::KSHIFTR, dl, WideVecVT, Vec, | |||
| 18772 | DAG.getTargetConstant(IdxVal, dl, MVT::i8)); | |||
| 18773 | ||||
| 18774 | return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(), Vec, | |||
| 18775 | DAG.getIntPtrConstant(0, dl)); | |||
| 18776 | } | |||
| 18777 | ||||
| 18778 | SDValue | |||
| 18779 | X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op, | |||
| 18780 | SelectionDAG &DAG) const { | |||
| 18781 | SDLoc dl(Op); | |||
| 18782 | SDValue Vec = Op.getOperand(0); | |||
| 18783 | MVT VecVT = Vec.getSimpleValueType(); | |||
| 18784 | SDValue Idx = Op.getOperand(1); | |||
| 18785 | auto* IdxC = dyn_cast<ConstantSDNode>(Idx); | |||
| 18786 | ||||
| 18787 | if (VecVT.getVectorElementType() == MVT::i1) | |||
| 18788 | return ExtractBitFromMaskVector(Op, DAG, Subtarget); | |||
| 18789 | ||||
| 18790 | if (!IdxC) { | |||
| 18791 | // Its more profitable to go through memory (1 cycles throughput) | |||
| 18792 | // than using VMOVD + VPERMV/PSHUFB sequence ( 2/3 cycles throughput) | |||
| 18793 | // IACA tool was used to get performance estimation | |||
| 18794 | // (https://software.intel.com/en-us/articles/intel-architecture-code-analyzer) | |||
| 18795 | // | |||
| 18796 | // example : extractelement <16 x i8> %a, i32 %i | |||
| 18797 | // | |||
| 18798 | // Block Throughput: 3.00 Cycles | |||
| 18799 | // Throughput Bottleneck: Port5 | |||
| 18800 | // | |||
| 18801 | // | Num Of | Ports pressure in cycles | | | |||
| 18802 | // | Uops | 0 - DV | 5 | 6 | 7 | | | |||
| 18803 | // --------------------------------------------- | |||
| 18804 | // | 1 | | 1.0 | | | CP | vmovd xmm1, edi | |||
| 18805 | // | 1 | | 1.0 | | | CP | vpshufb xmm0, xmm0, xmm1 | |||
| 18806 | // | 2 | 1.0 | 1.0 | | | CP | vpextrb eax, xmm0, 0x0 | |||
| 18807 | // Total Num Of Uops: 4 | |||
| 18808 | // | |||
| 18809 | // | |||
| 18810 | // Block Throughput: 1.00 Cycles | |||
| 18811 | // Throughput Bottleneck: PORT2_AGU, PORT3_AGU, Port4 | |||
| 18812 | // | |||
| 18813 | // | | Ports pressure in cycles | | | |||
| 18814 | // |Uops| 1 | 2 - D |3 - D | 4 | 5 | | | |||
| 18815 | // --------------------------------------------------------- | |||
| 18816 | // |2^ | | 0.5 | 0.5 |1.0| |CP| vmovaps xmmword ptr [rsp-0x18], xmm0 | |||
| 18817 | // |1 |0.5| | | |0.5| | lea rax, ptr [rsp-0x18] | |||
| 18818 | // |1 | |0.5, 0.5|0.5, 0.5| | |CP| mov al, byte ptr [rdi+rax*1] | |||
| 18819 | // Total Num Of Uops: 4 | |||
| 18820 | ||||
| 18821 | return SDValue(); | |||
| 18822 | } | |||
| 18823 | ||||
| 18824 | unsigned IdxVal = IdxC->getZExtValue(); | |||
| 18825 | ||||
| 18826 | // If this is a 256-bit vector result, first extract the 128-bit vector and | |||
| 18827 | // then extract the element from the 128-bit vector. | |||
| 18828 | if (VecVT.is256BitVector() || VecVT.is512BitVector()) { | |||
| 18829 | // Get the 128-bit vector. | |||
| 18830 | Vec = extract128BitVector(Vec, IdxVal, DAG, dl); | |||
| 18831 | MVT EltVT = VecVT.getVectorElementType(); | |||
| 18832 | ||||
| 18833 | unsigned ElemsPerChunk = 128 / EltVT.getSizeInBits(); | |||
| 18834 | assert(isPowerOf2_32(ElemsPerChunk) && "Elements per chunk not power of 2")((void)0); | |||
| 18835 | ||||
| 18836 | // Find IdxVal modulo ElemsPerChunk. Since ElemsPerChunk is a power of 2 | |||
| 18837 | // this can be done with a mask. | |||
| 18838 | IdxVal &= ElemsPerChunk - 1; | |||
| 18839 | return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(), Vec, | |||
| 18840 | DAG.getIntPtrConstant(IdxVal, dl)); | |||
| 18841 | } | |||
| 18842 | ||||
| 18843 | assert(VecVT.is128BitVector() && "Unexpected vector length")((void)0); | |||
| 18844 | ||||
| 18845 | MVT VT = Op.getSimpleValueType(); | |||
| 18846 | ||||
| 18847 | if (VT.getSizeInBits() == 16) { | |||
| 18848 | // If IdxVal is 0, it's cheaper to do a move instead of a pextrw, unless | |||
| 18849 | // we're going to zero extend the register or fold the store (SSE41 only). | |||
| 18850 | if (IdxVal == 0 && !MayFoldIntoZeroExtend(Op) && | |||
| 18851 | !(Subtarget.hasSSE41() && MayFoldIntoStore(Op))) | |||
| 18852 | return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16, | |||
| 18853 | DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32, | |||
| 18854 | DAG.getBitcast(MVT::v4i32, Vec), Idx)); | |||
| 18855 | ||||
| 18856 | SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32, Vec, | |||
| 18857 | DAG.getTargetConstant(IdxVal, dl, MVT::i8)); | |||
| 18858 | return DAG.getNode(ISD::TRUNCATE, dl, VT, Extract); | |||
| 18859 | } | |||
| 18860 | ||||
| 18861 | if (Subtarget.hasSSE41()) | |||
| 18862 | if (SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG)) | |||
| 18863 | return Res; | |||
| 18864 | ||||
| 18865 | // TODO: We only extract a single element from v16i8, we can probably afford | |||
| 18866 | // to be more aggressive here before using the default approach of spilling to | |||
| 18867 | // stack. | |||
| 18868 | if (VT.getSizeInBits() == 8 && Op->isOnlyUserOf(Vec.getNode())) { | |||
| 18869 | // Extract either the lowest i32 or any i16, and extract the sub-byte. | |||
| 18870 | int DWordIdx = IdxVal / 4; | |||
| 18871 | if (DWordIdx == 0) { | |||
| 18872 | SDValue Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32, | |||
| 18873 | DAG.getBitcast(MVT::v4i32, Vec), | |||
| 18874 | DAG.getIntPtrConstant(DWordIdx, dl)); | |||
| 18875 | int ShiftVal = (IdxVal % 4) * 8; | |||
| 18876 | if (ShiftVal != 0) | |||
| 18877 | Res = DAG.getNode(ISD::SRL, dl, MVT::i32, Res, | |||
| 18878 | DAG.getConstant(ShiftVal, dl, MVT::i8)); | |||
| 18879 | return DAG.getNode(ISD::TRUNCATE, dl, VT, Res); | |||
| 18880 | } | |||
| 18881 | ||||
| 18882 | int WordIdx = IdxVal / 2; | |||
| 18883 | SDValue Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, | |||
| 18884 | DAG.getBitcast(MVT::v8i16, Vec), | |||
| 18885 | DAG.getIntPtrConstant(WordIdx, dl)); | |||
| 18886 | int ShiftVal = (IdxVal % 2) * 8; | |||
| 18887 | if (ShiftVal != 0) | |||
| 18888 | Res = DAG.getNode(ISD::SRL, dl, MVT::i16, Res, | |||
| 18889 | DAG.getConstant(ShiftVal, dl, MVT::i8)); | |||
| 18890 | return DAG.getNode(ISD::TRUNCATE, dl, VT, Res); | |||
| 18891 | } | |||
| 18892 | ||||
| 18893 | if (VT.getSizeInBits() == 32) { | |||
| 18894 | if (IdxVal == 0) | |||
| 18895 | return Op; | |||
| 18896 | ||||
| 18897 | // SHUFPS the element to the lowest double word, then movss. | |||
| 18898 | int Mask[4] = { static_cast<int>(IdxVal), -1, -1, -1 }; | |||
| 18899 | Vec = DAG.getVectorShuffle(VecVT, dl, Vec, DAG.getUNDEF(VecVT), Mask); | |||
| 18900 | return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec, | |||
| 18901 | DAG.getIntPtrConstant(0, dl)); | |||
| 18902 | } | |||
| 18903 | ||||
| 18904 | if (VT.getSizeInBits() == 64) { | |||
| 18905 | // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b | |||
| 18906 | // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught | |||
| 18907 | // to match extract_elt for f64. | |||
| 18908 | if (IdxVal == 0) | |||
| 18909 | return Op; | |||
| 18910 | ||||
| 18911 | // UNPCKHPD the element to the lowest double word, then movsd. | |||
| 18912 | // Note if the lower 64 bits of the result of the UNPCKHPD is then stored | |||
| 18913 | // to a f64mem, the whole operation is folded into a single MOVHPDmr. | |||
| 18914 | int Mask[2] = { 1, -1 }; | |||
| 18915 | Vec = DAG.getVectorShuffle(VecVT, dl, Vec, DAG.getUNDEF(VecVT), Mask); | |||
| 18916 | return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec, | |||
| 18917 | DAG.getIntPtrConstant(0, dl)); | |||
| 18918 | } | |||
| 18919 | ||||
| 18920 | return SDValue(); | |||
| 18921 | } | |||
| 18922 | ||||
| 18923 | /// Insert one bit to mask vector, like v16i1 or v8i1. | |||
| 18924 | /// AVX-512 feature. | |||
| 18925 | static SDValue InsertBitToMaskVector(SDValue Op, SelectionDAG &DAG, | |||
| 18926 | const X86Subtarget &Subtarget) { | |||
| 18927 | SDLoc dl(Op); | |||
| 18928 | SDValue Vec = Op.getOperand(0); | |||
| 18929 | SDValue Elt = Op.getOperand(1); | |||
| 18930 | SDValue Idx = Op.getOperand(2); | |||
| 18931 | MVT VecVT = Vec.getSimpleValueType(); | |||
| 18932 | ||||
| 18933 | if (!isa<ConstantSDNode>(Idx)) { | |||
| 18934 | // Non constant index. Extend source and destination, | |||
| 18935 | // insert element and then truncate the result. | |||
| 18936 | unsigned NumElts = VecVT.getVectorNumElements(); | |||
| 18937 | MVT ExtEltVT = (NumElts <= 8) ? MVT::getIntegerVT(128 / NumElts) : MVT::i8; | |||
| 18938 | MVT ExtVecVT = MVT::getVectorVT(ExtEltVT, NumElts); | |||
| 18939 | SDValue ExtOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ExtVecVT, | |||
| 18940 | DAG.getNode(ISD::SIGN_EXTEND, dl, ExtVecVT, Vec), | |||
| 18941 | DAG.getNode(ISD::SIGN_EXTEND, dl, ExtEltVT, Elt), Idx); | |||
| 18942 | return DAG.getNode(ISD::TRUNCATE, dl, VecVT, ExtOp); | |||
| 18943 | } | |||
| 18944 | ||||
| 18945 | // Copy into a k-register, extract to v1i1 and insert_subvector. | |||
| 18946 | SDValue EltInVec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i1, Elt); | |||
| 18947 | return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, VecVT, Vec, EltInVec, Idx); | |||
| 18948 | } | |||
| 18949 | ||||
| 18950 | SDValue X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, | |||
| 18951 | SelectionDAG &DAG) const { | |||
| 18952 | MVT VT = Op.getSimpleValueType(); | |||
| 18953 | MVT EltVT = VT.getVectorElementType(); | |||
| 18954 | unsigned NumElts = VT.getVectorNumElements(); | |||
| 18955 | unsigned EltSizeInBits = EltVT.getScalarSizeInBits(); | |||
| 18956 | ||||
| 18957 | if (EltVT == MVT::i1) | |||
| 18958 | return InsertBitToMaskVector(Op, DAG, Subtarget); | |||
| 18959 | ||||
| 18960 | SDLoc dl(Op); | |||
| 18961 | SDValue N0 = Op.getOperand(0); | |||
| 18962 | SDValue N1 = Op.getOperand(1); | |||
| 18963 | SDValue N2 = Op.getOperand(2); | |||
| 18964 | auto *N2C = dyn_cast<ConstantSDNode>(N2); | |||
| 18965 | ||||
| 18966 | if (!N2C) { | |||
| 18967 | // Variable insertion indices, usually we're better off spilling to stack, | |||
| 18968 | // but AVX512 can use a variable compare+select by comparing against all | |||
| 18969 | // possible vector indices, and FP insertion has less gpr->simd traffic. | |||
| 18970 | if (!(Subtarget.hasBWI() || | |||
| 18971 | (Subtarget.hasAVX512() && EltSizeInBits >= 32) || | |||
| 18972 | (Subtarget.hasSSE41() && VT.isFloatingPoint()))) | |||
| 18973 | return SDValue(); | |||
| 18974 | ||||
| 18975 | MVT IdxSVT = MVT::getIntegerVT(EltSizeInBits); | |||
| 18976 | MVT IdxVT = MVT::getVectorVT(IdxSVT, NumElts); | |||
| 18977 | if (!isTypeLegal(IdxSVT) || !isTypeLegal(IdxVT)) | |||
| 18978 | return SDValue(); | |||
| 18979 | ||||
| 18980 | SDValue IdxExt = DAG.getZExtOrTrunc(N2, dl, IdxSVT); | |||
| 18981 | SDValue IdxSplat = DAG.getSplatBuildVector(IdxVT, dl, IdxExt); | |||
| 18982 | SDValue EltSplat = DAG.getSplatBuildVector(VT, dl, N1); | |||
| 18983 | ||||
| 18984 | SmallVector<SDValue, 16> RawIndices; | |||
| 18985 | for (unsigned I = 0; I != NumElts; ++I) | |||
| 18986 | RawIndices.push_back(DAG.getConstant(I, dl, IdxSVT)); | |||
| 18987 | SDValue Indices = DAG.getBuildVector(IdxVT, dl, RawIndices); | |||
| 18988 | ||||
| 18989 | // inselt N0, N1, N2 --> select (SplatN2 == {0,1,2...}) ? SplatN1 : N0. | |||
| 18990 | return DAG.getSelectCC(dl, IdxSplat, Indices, EltSplat, N0, | |||
| 18991 | ISD::CondCode::SETEQ); | |||
| 18992 | } | |||
| 18993 | ||||
| 18994 | if (N2C->getAPIntValue().uge(NumElts)) | |||
| 18995 | return SDValue(); | |||
| 18996 | uint64_t IdxVal = N2C->getZExtValue(); | |||
| 18997 | ||||
| 18998 | bool IsZeroElt = X86::isZeroNode(N1); | |||
| 18999 | bool IsAllOnesElt = VT.isInteger() && llvm::isAllOnesConstant(N1); | |||
| 19000 | ||||
| 19001 | // If we are inserting a element, see if we can do this more efficiently with | |||
| 19002 | // a blend shuffle with a rematerializable vector than a costly integer | |||
| 19003 | // insertion. | |||
| 19004 | if ((IsZeroElt || IsAllOnesElt) && Subtarget.hasSSE41() && | |||
| 19005 | (16 <= EltSizeInBits || (IsZeroElt && !VT.is128BitVector()))) { | |||
| 19006 | SmallVector<int, 8> BlendMask; | |||
| 19007 | for (unsigned i = 0; i != NumElts; ++i) | |||
| 19008 | BlendMask.push_back(i == IdxVal ? i + NumElts : i); | |||
| 19009 | SDValue CstVector = IsZeroElt ? getZeroVector(VT, Subtarget, DAG, dl) | |||
| 19010 | : getOnesVector(VT, DAG, dl); | |||
| 19011 | return DAG.getVectorShuffle(VT, dl, N0, CstVector, BlendMask); | |||
| 19012 | } | |||
| 19013 | ||||
| 19014 | // If the vector is wider than 128 bits, extract the 128-bit subvector, insert | |||
| 19015 | // into that, and then insert the subvector back into the result. | |||
| 19016 | if (VT.is256BitVector() || VT.is512BitVector()) { | |||
| 19017 | // With a 256-bit vector, we can insert into the zero element efficiently | |||
| 19018 | // using a blend if we have AVX or AVX2 and the right data type. | |||
| 19019 | if (VT.is256BitVector() && IdxVal == 0) { | |||
| 19020 | // TODO: It is worthwhile to cast integer to floating point and back | |||
| 19021 | // and incur a domain crossing penalty if that's what we'll end up | |||
| 19022 | // doing anyway after extracting to a 128-bit vector. | |||
| 19023 | if ((Subtarget.hasAVX() && (EltVT == MVT::f64 || EltVT == MVT::f32)) || | |||
| 19024 | (Subtarget.hasAVX2() && EltVT == MVT::i32)) { | |||
| 19025 | SDValue N1Vec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, N1); | |||
| 19026 | return DAG.getNode(X86ISD::BLENDI, dl, VT, N0, N1Vec, | |||
| 19027 | DAG.getTargetConstant(1, dl, MVT::i8)); | |||
| 19028 | } | |||
| 19029 | } | |||
| 19030 | ||||
| 19031 | // Get the desired 128-bit vector chunk. | |||
| 19032 | SDValue V = extract128BitVector(N0, IdxVal, DAG, dl); | |||
| 19033 | ||||
| 19034 | // Insert the element into the desired chunk. | |||
| 19035 | unsigned NumEltsIn128 = 128 / EltSizeInBits; | |||
| 19036 | assert(isPowerOf2_32(NumEltsIn128))((void)0); | |||
| 19037 | // Since NumEltsIn128 is a power of 2 we can use mask instead of modulo. | |||
| 19038 | unsigned IdxIn128 = IdxVal & (NumEltsIn128 - 1); | |||
| 19039 | ||||
| 19040 | V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, V.getValueType(), V, N1, | |||
| 19041 | DAG.getIntPtrConstant(IdxIn128, dl)); | |||
| 19042 | ||||
| 19043 | // Insert the changed part back into the bigger vector | |||
| 19044 | return insert128BitVector(N0, V, IdxVal, DAG, dl); | |||
| 19045 | } | |||
| 19046 | assert(VT.is128BitVector() && "Only 128-bit vector types should be left!")((void)0); | |||
| 19047 | ||||
| 19048 | // This will be just movd/movq/movss/movsd. | |||
| 19049 | if (IdxVal == 0 && ISD::isBuildVectorAllZeros(N0.getNode())) { | |||
| 19050 | if (EltVT == MVT::i32 || EltVT == MVT::f32 || EltVT == MVT::f64 || | |||
| 19051 | EltVT == MVT::i64) { | |||
| 19052 | N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, N1); | |||
| 19053 | return getShuffleVectorZeroOrUndef(N1, 0, true, Subtarget, DAG); | |||
| 19054 | } | |||
| 19055 | ||||
| 19056 | // We can't directly insert an i8 or i16 into a vector, so zero extend | |||
| 19057 | // it to i32 first. | |||
| 19058 | if (EltVT == MVT::i16 || EltVT == MVT::i8) { | |||
| 19059 | N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, N1); | |||
| 19060 | MVT ShufVT = MVT::getVectorVT(MVT::i32, VT.getSizeInBits() / 32); | |||
| 19061 | N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, ShufVT, N1); | |||
| 19062 | N1 = getShuffleVectorZeroOrUndef(N1, 0, true, Subtarget, DAG); | |||
| 19063 | return DAG.getBitcast(VT, N1); | |||
| 19064 | } | |||
| 19065 | } | |||
| 19066 | ||||
| 19067 | // Transform it so it match pinsr{b,w} which expects a GR32 as its second | |||
| 19068 | // argument. SSE41 required for pinsrb. | |||
| 19069 | if (VT == MVT::v8i16 || (VT == MVT::v16i8 && Subtarget.hasSSE41())) { | |||
| 19070 | unsigned Opc; | |||
| 19071 | if (VT == MVT::v8i16) { | |||
| 19072 | assert(Subtarget.hasSSE2() && "SSE2 required for PINSRW")((void)0); | |||
| 19073 | Opc = X86ISD::PINSRW; | |||
| 19074 | } else { | |||
| 19075 | assert(VT == MVT::v16i8 && "PINSRB requires v16i8 vector")((void)0); | |||
| 19076 | assert(Subtarget.hasSSE41() && "SSE41 required for PINSRB")((void)0); | |||
| 19077 | Opc = X86ISD::PINSRB; | |||
| 19078 | } | |||
| 19079 | ||||
| 19080 | assert(N1.getValueType() != MVT::i32 && "Unexpected VT")((void)0); | |||
| 19081 | N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1); | |||
| 19082 | N2 = DAG.getTargetConstant(IdxVal, dl, MVT::i8); | |||
| 19083 | return DAG.getNode(Opc, dl, VT, N0, N1, N2); | |||
| 19084 | } | |||
| 19085 | ||||
| 19086 | if (Subtarget.hasSSE41()) { | |||
| 19087 | if (EltVT == MVT::f32) { | |||
| 19088 | // Bits [7:6] of the constant are the source select. This will always be | |||
| 19089 | // zero here. The DAG Combiner may combine an extract_elt index into | |||
| 19090 | // these bits. For example (insert (extract, 3), 2) could be matched by | |||
| 19091 | // putting the '3' into bits [7:6] of X86ISD::INSERTPS. | |||
| 19092 | // Bits [5:4] of the constant are the destination select. This is the | |||
| 19093 | // value of the incoming immediate. | |||
| 19094 | // Bits [3:0] of the constant are the zero mask. The DAG Combiner may | |||
| 19095 | // combine either bitwise AND or insert of float 0.0 to set these bits. | |||
| 19096 | ||||
| 19097 | bool MinSize = DAG.getMachineFunction().getFunction().hasMinSize(); | |||
| 19098 | if (IdxVal == 0 && (!MinSize || !MayFoldLoad(N1))) { | |||
| 19099 | // If this is an insertion of 32-bits into the low 32-bits of | |||
| 19100 | // a vector, we prefer to generate a blend with immediate rather | |||
| 19101 | // than an insertps. Blends are simpler operations in hardware and so | |||
| 19102 | // will always have equal or better performance than insertps. | |||
| 19103 | // But if optimizing for size and there's a load folding opportunity, | |||
| 19104 | // generate insertps because blendps does not have a 32-bit memory | |||
| 19105 | // operand form. | |||
| 19106 | N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1); | |||
| 19107 | return DAG.getNode(X86ISD::BLENDI, dl, VT, N0, N1, | |||
| 19108 | DAG.getTargetConstant(1, dl, MVT::i8)); | |||
| 19109 | } | |||
| 19110 | // Create this as a scalar to vector.. | |||
| 19111 | N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1); | |||
| 19112 | return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, | |||
| 19113 | DAG.getTargetConstant(IdxVal << 4, dl, MVT::i8)); | |||
| 19114 | } | |||
| 19115 | ||||
| 19116 | // PINSR* works with constant index. | |||
| 19117 | if (EltVT == MVT::i32 || EltVT == MVT::i64) | |||
| 19118 | return Op; | |||
| 19119 | } | |||
| 19120 | ||||
| 19121 | return SDValue(); | |||
| 19122 | } | |||
| 19123 | ||||
| 19124 | static SDValue LowerSCALAR_TO_VECTOR(SDValue Op, const X86Subtarget &Subtarget, | |||
| 19125 | SelectionDAG &DAG) { | |||
| 19126 | SDLoc dl(Op); | |||
| 19127 | MVT OpVT = Op.getSimpleValueType(); | |||
| 19128 | ||||
| 19129 | // It's always cheaper to replace a xor+movd with xorps and simplifies further | |||
| 19130 | // combines. | |||
| 19131 | if (X86::isZeroNode(Op.getOperand(0))) | |||
| 19132 | return getZeroVector(OpVT, Subtarget, DAG, dl); | |||
| 19133 | ||||
| 19134 | // If this is a 256-bit vector result, first insert into a 128-bit | |||
| 19135 | // vector and then insert into the 256-bit vector. | |||
| 19136 | if (!OpVT.is128BitVector()) { | |||
| 19137 | // Insert into a 128-bit vector. | |||
| 19138 | unsigned SizeFactor = OpVT.getSizeInBits() / 128; | |||
| 19139 | MVT VT128 = MVT::getVectorVT(OpVT.getVectorElementType(), | |||
| 19140 | OpVT.getVectorNumElements() / SizeFactor); | |||
| 19141 | ||||
| 19142 | Op = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT128, Op.getOperand(0)); | |||
| 19143 | ||||
| 19144 | // Insert the 128-bit vector. | |||
| 19145 | return insert128BitVector(DAG.getUNDEF(OpVT), Op, 0, DAG, dl); | |||
| 19146 | } | |||
| 19147 | assert(OpVT.is128BitVector() && OpVT.isInteger() && OpVT != MVT::v2i64 &&((void)0) | |||
| 19148 | "Expected an SSE type!")((void)0); | |||
| 19149 | ||||
| 19150 | // Pass through a v4i32 SCALAR_TO_VECTOR as that's what we use in tblgen. | |||
| 19151 | if (OpVT == MVT::v4i32) | |||
| 19152 | return Op; | |||
| 19153 | ||||
| 19154 | SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0)); | |||
| 19155 | return DAG.getBitcast( | |||
| 19156 | OpVT, DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, AnyExt)); | |||
| 19157 | } | |||
| 19158 | ||||
| 19159 | // Lower a node with an INSERT_SUBVECTOR opcode. This may result in a | |||
| 19160 | // simple superregister reference or explicit instructions to insert | |||
| 19161 | // the upper bits of a vector. | |||
| 19162 | static SDValue LowerINSERT_SUBVECTOR(SDValue Op, const X86Subtarget &Subtarget, | |||
| 19163 | SelectionDAG &DAG) { | |||
| 19164 | assert(Op.getSimpleValueType().getVectorElementType() == MVT::i1)((void)0); | |||
| 19165 | ||||
| 19166 | return insert1BitVector(Op, DAG, Subtarget); | |||
| 19167 | } | |||
| 19168 | ||||
| 19169 | static SDValue LowerEXTRACT_SUBVECTOR(SDValue Op, const X86Subtarget &Subtarget, | |||
| 19170 | SelectionDAG &DAG) { | |||
| 19171 | assert(Op.getSimpleValueType().getVectorElementType() == MVT::i1 &&((void)0) | |||
| 19172 | "Only vXi1 extract_subvectors need custom lowering")((void)0); | |||
| 19173 | ||||
| 19174 | SDLoc dl(Op); | |||
| 19175 | SDValue Vec = Op.getOperand(0); | |||
| 19176 | uint64_t IdxVal = Op.getConstantOperandVal(1); | |||
| 19177 | ||||
| 19178 | if (IdxVal == 0) // the operation is legal | |||
| 19179 | return Op; | |||
| 19180 | ||||
| 19181 | MVT VecVT = Vec.getSimpleValueType(); | |||
| 19182 | unsigned NumElems = VecVT.getVectorNumElements(); | |||
| 19183 | ||||
| 19184 | // Extend to natively supported kshift. | |||
| 19185 | MVT WideVecVT = VecVT; | |||
| 19186 | if ((!Subtarget.hasDQI() && NumElems == 8) || NumElems < 8) { | |||
| 19187 | WideVecVT = Subtarget.hasDQI() ? MVT::v8i1 : MVT::v16i1; | |||
| 19188 | Vec = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, WideVecVT, | |||
| 19189 | DAG.getUNDEF(WideVecVT), Vec, | |||
| 19190 | DAG.getIntPtrConstant(0, dl)); | |||
| 19191 | } | |||
| 19192 | ||||
| 19193 | // Shift to the LSB. | |||
| 19194 | Vec = DAG.getNode(X86ISD::KSHIFTR, dl, WideVecVT, Vec, | |||
| 19195 | DAG.getTargetConstant(IdxVal, dl, MVT::i8)); | |||
| 19196 | ||||
| 19197 | return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, Op.getValueType(), Vec, | |||
| 19198 | DAG.getIntPtrConstant(0, dl)); | |||
| 19199 | } | |||
| 19200 | ||||
| 19201 | // Returns the appropriate wrapper opcode for a global reference. | |||
| 19202 | unsigned X86TargetLowering::getGlobalWrapperKind( | |||
| 19203 | const GlobalValue *GV, const unsigned char OpFlags) const { | |||
| 19204 | // References to absolute symbols are never PC-relative. | |||
| 19205 | if (GV && GV->isAbsoluteSymbolRef()) | |||
| 19206 | return X86ISD::Wrapper; | |||
| 19207 | ||||
| 19208 | CodeModel::Model M = getTargetMachine().getCodeModel(); | |||
| 19209 | if (Subtarget.isPICStyleRIPRel() && | |||
| 19210 | (M == CodeModel::Small || M == CodeModel::Kernel)) | |||
| 19211 | return X86ISD::WrapperRIP; | |||
| 19212 | ||||
| 19213 | // GOTPCREL references must always use RIP. | |||
| 19214 | if (OpFlags == X86II::MO_GOTPCREL) | |||
| 19215 | return X86ISD::WrapperRIP; | |||
| 19216 | ||||
| 19217 | return X86ISD::Wrapper; | |||
| 19218 | } | |||
| 19219 | ||||
| 19220 | // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as | |||
| 19221 | // their target counterpart wrapped in the X86ISD::Wrapper node. Suppose N is | |||
| 19222 | // one of the above mentioned nodes. It has to be wrapped because otherwise | |||
| 19223 | // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only | |||
| 19224 | // be used to form addressing mode. These wrapped nodes will be selected | |||
| 19225 | // into MOV32ri. | |||
| 19226 | SDValue | |||
| 19227 | X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) const { | |||
| 19228 | ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op); | |||
| 19229 | ||||
| 19230 | // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the | |||
| 19231 | // global base reg. | |||
| 19232 | unsigned char OpFlag = Subtarget.classifyLocalReference(nullptr); | |||
| 19233 | ||||
| 19234 | auto PtrVT = getPointerTy(DAG.getDataLayout()); | |||
| 19235 | SDValue Result = DAG.getTargetConstantPool( | |||
| 19236 | CP->getConstVal(), PtrVT, CP->getAlign(), CP->getOffset(), OpFlag); | |||
| 19237 | SDLoc DL(CP); | |||
| 19238 | Result = DAG.getNode(getGlobalWrapperKind(), DL, PtrVT, Result); | |||
| 19239 | // With PIC, the address is actually $g + Offset. | |||
| 19240 | if (OpFlag) { | |||
| 19241 | Result = | |||
| 19242 | DAG.getNode(ISD::ADD, DL, PtrVT, | |||
| 19243 | DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), Result); | |||
| 19244 | } | |||
| 19245 | ||||
| 19246 | return Result; | |||
| 19247 | } | |||
| 19248 | ||||
| 19249 | SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const { | |||
| 19250 | JumpTableSDNode *JT = cast<JumpTableSDNode>(Op); | |||
| 19251 | ||||
| 19252 | // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the | |||
| 19253 | // global base reg. | |||
| 19254 | unsigned char OpFlag = Subtarget.classifyLocalReference(nullptr); | |||
| 19255 | ||||
| 19256 | auto PtrVT = getPointerTy(DAG.getDataLayout()); | |||
| 19257 | SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, OpFlag); | |||
| 19258 | SDLoc DL(JT); | |||
| 19259 | Result = DAG.getNode(getGlobalWrapperKind(), DL, PtrVT, Result); | |||
| 19260 | ||||
| 19261 | // With PIC, the address is actually $g + Offset. | |||
| 19262 | if (OpFlag) | |||
| 19263 | Result = | |||
| 19264 | DAG.getNode(ISD::ADD, DL, PtrVT, | |||
| 19265 | DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), Result); | |||
| 19266 | ||||
| 19267 | return Result; | |||
| 19268 | } | |||
| 19269 | ||||
| 19270 | SDValue X86TargetLowering::LowerExternalSymbol(SDValue Op, | |||
| 19271 | SelectionDAG &DAG) const { | |||
| 19272 | return LowerGlobalOrExternal(Op, DAG, /*ForCall=*/false); | |||
| 19273 | } | |||
| 19274 | ||||
| 19275 | SDValue | |||
| 19276 | X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const { | |||
| 19277 | // Create the TargetBlockAddressAddress node. | |||
| 19278 | unsigned char OpFlags = | |||
| 19279 | Subtarget.classifyBlockAddressReference(); | |||
| 19280 | const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress(); | |||
| 19281 | int64_t Offset = cast<BlockAddressSDNode>(Op)->getOffset(); | |||
| 19282 | SDLoc dl(Op); | |||
| 19283 | auto PtrVT = getPointerTy(DAG.getDataLayout()); | |||
| 19284 | SDValue Result = DAG.getTargetBlockAddress(BA, PtrVT, Offset, OpFlags); | |||
| 19285 | Result = DAG.getNode(getGlobalWrapperKind(), dl, PtrVT, Result); | |||
| 19286 | ||||
| 19287 | // With PIC, the address is actually $g + Offset. | |||
| 19288 | if (isGlobalRelativeToPICBase(OpFlags)) { | |||
| 19289 | Result = DAG.getNode(ISD::ADD, dl, PtrVT, | |||
| 19290 | DAG.getNode(X86ISD::GlobalBaseReg, dl, PtrVT), Result); | |||
| 19291 | } | |||
| 19292 | ||||
| 19293 | return Result; | |||
| 19294 | } | |||
| 19295 | ||||
| 19296 | /// Creates target global address or external symbol nodes for calls or | |||
| 19297 | /// other uses. | |||
| 19298 | SDValue X86TargetLowering::LowerGlobalOrExternal(SDValue Op, SelectionDAG &DAG, | |||
| 19299 | bool ForCall) const { | |||
| 19300 | // Unpack the global address or external symbol. | |||
| 19301 | const SDLoc &dl = SDLoc(Op); | |||
| 19302 | const GlobalValue *GV = nullptr; | |||
| 19303 | int64_t Offset = 0; | |||
| 19304 | const char *ExternalSym = nullptr; | |||
| 19305 | if (const auto *G = dyn_cast<GlobalAddressSDNode>(Op)) { | |||
| 19306 | GV = G->getGlobal(); | |||
| 19307 | Offset = G->getOffset(); | |||
| 19308 | } else { | |||
| 19309 | const auto *ES = cast<ExternalSymbolSDNode>(Op); | |||
| 19310 | ExternalSym = ES->getSymbol(); | |||
| 19311 | } | |||
| 19312 | ||||
| 19313 | // Calculate some flags for address lowering. | |||
| 19314 | const Module &Mod = *DAG.getMachineFunction().getFunction().getParent(); | |||
| 19315 | unsigned char OpFlags; | |||
| 19316 | if (ForCall) | |||
| 19317 | OpFlags = Subtarget.classifyGlobalFunctionReference(GV, Mod); | |||
| 19318 | else | |||
| 19319 | OpFlags = Subtarget.classifyGlobalReference(GV, Mod); | |||
| 19320 | bool HasPICReg = isGlobalRelativeToPICBase(OpFlags); | |||
| 19321 | bool NeedsLoad = isGlobalStubReference(OpFlags); | |||
| 19322 | ||||
| 19323 | CodeModel::Model M = DAG.getTarget().getCodeModel(); | |||
| 19324 | auto PtrVT = getPointerTy(DAG.getDataLayout()); | |||
| 19325 | SDValue Result; | |||
| 19326 | ||||
| 19327 | if (GV) { | |||
| 19328 | // Create a target global address if this is a global. If possible, fold the | |||
| 19329 | // offset into the global address reference. Otherwise, ADD it on later. | |||
| 19330 | // Suppress the folding if Offset is negative: movl foo-1, %eax is not | |||
| 19331 | // allowed because if the address of foo is 0, the ELF R_X86_64_32 | |||
| 19332 | // relocation will compute to a negative value, which is invalid. | |||
| 19333 | int64_t GlobalOffset = 0; | |||
| 19334 | if (OpFlags == X86II::MO_NO_FLAG && Offset >= 0 && | |||
| 19335 | X86::isOffsetSuitableForCodeModel(Offset, M, true)) { | |||
| 19336 | std::swap(GlobalOffset, Offset); | |||
| 19337 | } | |||
| 19338 | Result = DAG.getTargetGlobalAddress(GV, dl, PtrVT, GlobalOffset, OpFlags); | |||
| 19339 | } else { | |||
| 19340 | // If this is not a global address, this must be an external symbol. | |||
| 19341 | Result = DAG.getTargetExternalSymbol(ExternalSym, PtrVT, OpFlags); | |||
| 19342 | } | |||
| 19343 | ||||
| 19344 | // If this is a direct call, avoid the wrapper if we don't need to do any | |||
| 19345 | // loads or adds. This allows SDAG ISel to match direct calls. | |||
| 19346 | if (ForCall && !NeedsLoad && !HasPICReg && Offset == 0) | |||
| 19347 | return Result; | |||
| 19348 | ||||
| 19349 | Result = DAG.getNode(getGlobalWrapperKind(GV, OpFlags), dl, PtrVT, Result); | |||
| 19350 | ||||
| 19351 | // With PIC, the address is actually $g + Offset. | |||
| 19352 | if (HasPICReg) { | |||
| 19353 | Result = DAG.getNode(ISD::ADD, dl, PtrVT, | |||
| 19354 | DAG.getNode(X86ISD::GlobalBaseReg, dl, PtrVT), Result); | |||
| 19355 | } | |||
| 19356 | ||||
| 19357 | // For globals that require a load from a stub to get the address, emit the | |||
| 19358 | // load. | |||
| 19359 | if (NeedsLoad) | |||
| 19360 | Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result, | |||
| 19361 | MachinePointerInfo::getGOT(DAG.getMachineFunction())); | |||
| 19362 | ||||
| 19363 | // If there was a non-zero offset that we didn't fold, create an explicit | |||
| 19364 | // addition for it. | |||
| 19365 | if (Offset != 0) | |||
| 19366 | Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, | |||
| 19367 | DAG.getConstant(Offset, dl, PtrVT)); | |||
| 19368 | ||||
| 19369 | return Result; | |||
| 19370 | } | |||
| 19371 | ||||
| 19372 | SDValue | |||
| 19373 | X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const { | |||
| 19374 | return LowerGlobalOrExternal(Op, DAG, /*ForCall=*/false); | |||
| 19375 | } | |||
| 19376 | ||||
| 19377 | static SDValue | |||
| 19378 | GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA, | |||
| 19379 | SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg, | |||
| 19380 | unsigned char OperandFlags, bool LocalDynamic = false) { | |||
| 19381 | MachineFrameInfo &MFI = DAG.getMachineFunction().getFrameInfo(); | |||
| 19382 | SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue); | |||
| 19383 | SDLoc dl(GA); | |||
| 19384 | SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl, | |||
| 19385 | GA->getValueType(0), | |||
| 19386 | GA->getOffset(), | |||
| 19387 | OperandFlags); | |||
| 19388 | ||||
| 19389 | X86ISD::NodeType CallType = LocalDynamic ? X86ISD::TLSBASEADDR | |||
| 19390 | : X86ISD::TLSADDR; | |||
| 19391 | ||||
| 19392 | if (InFlag) { | |||
| 19393 | SDValue Ops[] = { Chain, TGA, *InFlag }; | |||
| 19394 | Chain = DAG.getNode(CallType, dl, NodeTys, Ops); | |||
| 19395 | } else { | |||
| 19396 | SDValue Ops[] = { Chain, TGA }; | |||
| 19397 | Chain = DAG.getNode(CallType, dl, NodeTys, Ops); | |||
| 19398 | } | |||
| 19399 | ||||
| 19400 | // TLSADDR will be codegen'ed as call. Inform MFI that function has calls. | |||
| 19401 | MFI.setAdjustsStack(true); | |||
| 19402 | MFI.setHasCalls(true); | |||
| 19403 | ||||
| 19404 | SDValue Flag = Chain.getValue(1); | |||
| 19405 | return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag); | |||
| 19406 | } | |||
| 19407 | ||||
| 19408 | // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit | |||
| 19409 | static SDValue | |||
| 19410 | LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG, | |||
| 19411 | const EVT PtrVT) { | |||
| 19412 | SDValue InFlag; | |||
| 19413 | SDLoc dl(GA); // ? function entry point might be better | |||
| 19414 | SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX, | |||
| 19415 | DAG.getNode(X86ISD::GlobalBaseReg, | |||
| 19416 | SDLoc(), PtrVT), InFlag); | |||
| 19417 | InFlag = Chain.getValue(1); | |||
| 19418 | ||||
| 19419 | return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD); | |||
| 19420 | } | |||
| 19421 | ||||
| 19422 | // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit LP64 | |||
| 19423 | static SDValue | |||
| 19424 | LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG, | |||
| 19425 | const EVT PtrVT) { | |||
| 19426 | return GetTLSADDR(DAG, DAG.getEntryNode(), GA, nullptr, PtrVT, | |||
| 19427 | X86::RAX, X86II::MO_TLSGD); | |||
| 19428 | } | |||
| 19429 | ||||
| 19430 | // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit ILP32 | |||
| 19431 | static SDValue | |||
| 19432 | LowerToTLSGeneralDynamicModelX32(GlobalAddressSDNode *GA, SelectionDAG &DAG, | |||
| 19433 | const EVT PtrVT) { | |||
| 19434 | return GetTLSADDR(DAG, DAG.getEntryNode(), GA, nullptr, PtrVT, | |||
| 19435 | X86::EAX, X86II::MO_TLSGD); | |||
| 19436 | } | |||
| 19437 | ||||
| 19438 | static SDValue LowerToTLSLocalDynamicModel(GlobalAddressSDNode *GA, | |||
| 19439 | SelectionDAG &DAG, const EVT PtrVT, | |||
| 19440 | bool Is64Bit, bool Is64BitLP64) { | |||
| 19441 | SDLoc dl(GA); | |||
| 19442 | ||||
| 19443 | // Get the start address of the TLS block for this module. | |||
| 19444 | X86MachineFunctionInfo *MFI = DAG.getMachineFunction() | |||
| 19445 | .getInfo<X86MachineFunctionInfo>(); | |||
| 19446 | MFI->incNumLocalDynamicTLSAccesses(); | |||
| 19447 | ||||
| 19448 | SDValue Base; | |||
| 19449 | if (Is64Bit) { | |||
| 19450 | unsigned ReturnReg = Is64BitLP64 ? X86::RAX : X86::EAX; | |||
| 19451 | Base = GetTLSADDR(DAG, DAG.getEntryNode(), GA, nullptr, PtrVT, ReturnReg, | |||
| 19452 | X86II::MO_TLSLD, /*LocalDynamic=*/true); | |||
| 19453 | } else { | |||
| 19454 | SDValue InFlag; | |||
| 19455 | SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX, | |||
| 19456 | DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), InFlag); | |||
| 19457 | InFlag = Chain.getValue(1); | |||
| 19458 | Base = GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, | |||
| 19459 | X86II::MO_TLSLDM, /*LocalDynamic=*/true); | |||
| 19460 | } | |||
| 19461 | ||||
| 19462 | // Note: the CleanupLocalDynamicTLSPass will remove redundant computations | |||
| 19463 | // of Base. | |||
| 19464 | ||||
| 19465 | // Build x@dtpoff. | |||
| 19466 | unsigned char OperandFlags = X86II::MO_DTPOFF; | |||
| 19467 | unsigned WrapperKind = X86ISD::Wrapper; | |||
| 19468 | SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl, | |||
| 19469 | GA->getValueType(0), | |||
| 19470 | GA->getOffset(), OperandFlags); | |||
| 19471 | SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA); | |||
| 19472 | ||||
| 19473 | // Add x@dtpoff with the base. | |||
| 19474 | return DAG.getNode(ISD::ADD, dl, PtrVT, Offset, Base); | |||
| 19475 | } | |||
| 19476 | ||||
| 19477 | // Lower ISD::GlobalTLSAddress using the "initial exec" or "local exec" model. | |||
| 19478 | static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG, | |||
| 19479 | const EVT PtrVT, TLSModel::Model model, | |||
| 19480 | bool is64Bit, bool isPIC) { | |||
| 19481 | SDLoc dl(GA); | |||
| 19482 | ||||
| 19483 | // Get the Thread Pointer, which is %gs:0 (32-bit) or %fs:0 (64-bit). | |||
| 19484 | Value *Ptr = Constant::getNullValue(Type::getInt8PtrTy(*DAG.getContext(), | |||
| 19485 | is64Bit ? 257 : 256)); | |||
| 19486 | ||||
| 19487 | SDValue ThreadPointer = | |||
| 19488 | DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), DAG.getIntPtrConstant(0, dl), | |||
| 19489 | MachinePointerInfo(Ptr)); | |||
| 19490 | ||||
| 19491 | unsigned char OperandFlags = 0; | |||
| 19492 | // Most TLS accesses are not RIP relative, even on x86-64. One exception is | |||
| 19493 | // initialexec. | |||
| 19494 | unsigned WrapperKind = X86ISD::Wrapper; | |||
| 19495 | if (model == TLSModel::LocalExec) { | |||
| 19496 | OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF; | |||
| 19497 | } else if (model == TLSModel::InitialExec) { | |||
| 19498 | if (is64Bit) { | |||
| 19499 | OperandFlags = X86II::MO_GOTTPOFF; | |||
| 19500 | WrapperKind = X86ISD::WrapperRIP; | |||
| 19501 | } else { | |||
| 19502 | OperandFlags = isPIC ? X86II::MO_GOTNTPOFF : X86II::MO_INDNTPOFF; | |||
| 19503 | } | |||
| 19504 | } else { | |||
| 19505 | llvm_unreachable("Unexpected model")__builtin_unreachable(); | |||
| 19506 | } | |||
| 19507 | ||||
| 19508 | // emit "addl x@ntpoff,%eax" (local exec) | |||
| 19509 | // or "addl x@indntpoff,%eax" (initial exec) | |||
| 19510 | // or "addl x@gotntpoff(%ebx) ,%eax" (initial exec, 32-bit pic) | |||
| 19511 | SDValue TGA = | |||
| 19512 | DAG.getTargetGlobalAddress(GA->getGlobal(), dl, GA->getValueType(0), | |||
| 19513 | GA->getOffset(), OperandFlags); | |||
| 19514 | SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA); | |||
| 19515 | ||||
| 19516 | if (model == TLSModel::InitialExec) { | |||
| 19517 | if (isPIC && !is64Bit) { | |||
| 19518 | Offset = DAG.getNode(ISD::ADD, dl, PtrVT, | |||
| 19519 | DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), | |||
| 19520 | Offset); | |||
| 19521 | } | |||
| 19522 | ||||
| 19523 | Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset, | |||
| 19524 | MachinePointerInfo::getGOT(DAG.getMachineFunction())); | |||
| 19525 | } | |||
| 19526 | ||||
| 19527 | // The address of the thread local variable is the add of the thread | |||
| 19528 | // pointer with the offset of the variable. | |||
| 19529 | return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset); | |||
| 19530 | } | |||
| 19531 | ||||
| 19532 | SDValue | |||
| 19533 | X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const { | |||
| 19534 | ||||
| 19535 | GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op); | |||
| 19536 | ||||
| 19537 | if (DAG.getTarget().useEmulatedTLS()) | |||
| 19538 | return LowerToTLSEmulatedModel(GA, DAG); | |||
| 19539 | ||||
| 19540 | const GlobalValue *GV = GA->getGlobal(); | |||
| 19541 | auto PtrVT = getPointerTy(DAG.getDataLayout()); | |||
| 19542 | bool PositionIndependent = isPositionIndependent(); | |||
| 19543 | ||||
| 19544 | if (Subtarget.isTargetELF()) { | |||
| 19545 | TLSModel::Model model = DAG.getTarget().getTLSModel(GV); | |||
| 19546 | switch (model) { | |||
| 19547 | case TLSModel::GeneralDynamic: | |||
| 19548 | if (Subtarget.is64Bit()) { | |||
| 19549 | if (Subtarget.isTarget64BitLP64()) | |||
| 19550 | return LowerToTLSGeneralDynamicModel64(GA, DAG, PtrVT); | |||
| 19551 | return LowerToTLSGeneralDynamicModelX32(GA, DAG, PtrVT); | |||
| 19552 | } | |||
| 19553 | return LowerToTLSGeneralDynamicModel32(GA, DAG, PtrVT); | |||
| 19554 | case TLSModel::LocalDynamic: | |||
| 19555 | return LowerToTLSLocalDynamicModel(GA, DAG, PtrVT, Subtarget.is64Bit(), | |||
| 19556 | Subtarget.isTarget64BitLP64()); | |||
| 19557 | case TLSModel::InitialExec: | |||
| 19558 | case TLSModel::LocalExec: | |||
| 19559 | return LowerToTLSExecModel(GA, DAG, PtrVT, model, Subtarget.is64Bit(), | |||
| 19560 | PositionIndependent); | |||
| 19561 | } | |||
| 19562 | llvm_unreachable("Unknown TLS model.")__builtin_unreachable(); | |||
| 19563 | } | |||
| 19564 | ||||
| 19565 | if (Subtarget.isTargetDarwin()) { | |||
| 19566 | // Darwin only has one model of TLS. Lower to that. | |||
| 19567 | unsigned char OpFlag = 0; | |||
| 19568 | unsigned WrapperKind = Subtarget.isPICStyleRIPRel() ? | |||
| 19569 | X86ISD::WrapperRIP : X86ISD::Wrapper; | |||
| 19570 | ||||
| 19571 | // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the | |||
| 19572 | // global base reg. | |||
| 19573 | bool PIC32 = PositionIndependent && !Subtarget.is64Bit(); | |||
| 19574 | if (PIC32) | |||
| 19575 | OpFlag = X86II::MO_TLVP_PIC_BASE; | |||
| 19576 | else | |||
| 19577 | OpFlag = X86II::MO_TLVP; | |||
| 19578 | SDLoc DL(Op); | |||
| 19579 | SDValue Result = DAG.getTargetGlobalAddress(GA->getGlobal(), DL, | |||
| 19580 | GA->getValueType(0), | |||
| 19581 | GA->getOffset(), OpFlag); | |||
| 19582 | SDValue Offset = DAG.getNode(WrapperKind, DL, PtrVT, Result); | |||
| 19583 | ||||
| 19584 | // With PIC32, the address is actually $g + Offset. | |||
| 19585 | if (PIC32) | |||
| 19586 | Offset = DAG.getNode(ISD::ADD, DL, PtrVT, | |||
| 19587 | DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), | |||
| 19588 | Offset); | |||
| 19589 | ||||
| 19590 | // Lowering the machine isd will make sure everything is in the right | |||
| 19591 | // location. | |||
| 19592 | SDValue Chain = DAG.getEntryNode(); | |||
| 19593 | SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue); | |||
| 19594 | Chain = DAG.getCALLSEQ_START(Chain, 0, 0, DL); | |||
| 19595 | SDValue Args[] = { Chain, Offset }; | |||
| 19596 | Chain = DAG.getNode(X86ISD::TLSCALL, DL, NodeTys, Args); | |||
| 19597 | Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, DL, true), | |||
| 19598 | DAG.getIntPtrConstant(0, DL, true), | |||
| 19599 | Chain.getValue(1), DL); | |||
| 19600 | ||||
| 19601 | // TLSCALL will be codegen'ed as call. Inform MFI that function has calls. | |||
| 19602 | MachineFrameInfo &MFI = DAG.getMachineFunction().getFrameInfo(); | |||
| 19603 | MFI.setAdjustsStack(true); | |||
| 19604 | ||||
| 19605 | // And our return value (tls address) is in the standard call return value | |||
| 19606 | // location. | |||
| 19607 | unsigned Reg = Subtarget.is64Bit() ? X86::RAX : X86::EAX; | |||
| 19608 | return DAG.getCopyFromReg(Chain, DL, Reg, PtrVT, Chain.getValue(1)); | |||
| 19609 | } | |||
| 19610 | ||||
| 19611 | if (Subtarget.isOSWindows()) { | |||
| 19612 | // Just use the implicit TLS architecture | |||
| 19613 | // Need to generate something similar to: | |||
| 19614 | // mov rdx, qword [gs:abs 58H]; Load pointer to ThreadLocalStorage | |||
| 19615 | // ; from TEB | |||
| 19616 | // mov ecx, dword [rel _tls_index]: Load index (from C runtime) | |||
| 19617 | // mov rcx, qword [rdx+rcx*8] | |||
| 19618 | // mov eax, .tls$:tlsvar | |||
| 19619 | // [rax+rcx] contains the address | |||
| 19620 | // Windows 64bit: gs:0x58 | |||
| 19621 | // Windows 32bit: fs:__tls_array | |||
| 19622 | ||||
| 19623 | SDLoc dl(GA); | |||
| 19624 | SDValue Chain = DAG.getEntryNode(); | |||
| 19625 | ||||
| 19626 | // Get the Thread Pointer, which is %fs:__tls_array (32-bit) or | |||
| 19627 | // %gs:0x58 (64-bit). On MinGW, __tls_array is not available, so directly | |||
| 19628 | // use its literal value of 0x2C. | |||
| 19629 | Value *Ptr = Constant::getNullValue(Subtarget.is64Bit() | |||
| 19630 | ? Type::getInt8PtrTy(*DAG.getContext(), | |||
| 19631 | 256) | |||
| 19632 | : Type::getInt32PtrTy(*DAG.getContext(), | |||
| 19633 | 257)); | |||
| 19634 | ||||
| 19635 | SDValue TlsArray = Subtarget.is64Bit() | |||
| 19636 | ? DAG.getIntPtrConstant(0x58, dl) | |||
| 19637 | : (Subtarget.isTargetWindowsGNU() | |||
| 19638 | ? DAG.getIntPtrConstant(0x2C, dl) | |||
| 19639 | : DAG.getExternalSymbol("_tls_array", PtrVT)); | |||
| 19640 | ||||
| 19641 | SDValue ThreadPointer = | |||
| 19642 | DAG.getLoad(PtrVT, dl, Chain, TlsArray, MachinePointerInfo(Ptr)); | |||
| 19643 | ||||
| 19644 | SDValue res; | |||
| 19645 | if (GV->getThreadLocalMode() == GlobalVariable::LocalExecTLSModel) { | |||
| 19646 | res = ThreadPointer; | |||
| 19647 | } else { | |||
| 19648 | // Load the _tls_index variable | |||
| 19649 | SDValue IDX = DAG.getExternalSymbol("_tls_index", PtrVT); | |||
| 19650 | if (Subtarget.is64Bit()) | |||
| 19651 | IDX = DAG.getExtLoad(ISD::ZEXTLOAD, dl, PtrVT, Chain, IDX, | |||
| 19652 | MachinePointerInfo(), MVT::i32); | |||
| 19653 | else | |||
| 19654 | IDX = DAG.getLoad(PtrVT, dl, Chain, IDX, MachinePointerInfo()); | |||
| 19655 | ||||
| 19656 | const DataLayout &DL = DAG.getDataLayout(); | |||
| 19657 | SDValue Scale = | |||
| 19658 | DAG.getConstant(Log2_64_Ceil(DL.getPointerSize()), dl, MVT::i8); | |||
| 19659 | IDX = DAG.getNode(ISD::SHL, dl, PtrVT, IDX, Scale); | |||
| 19660 | ||||
| 19661 | res = DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, IDX); | |||
| 19662 | } | |||
| 19663 | ||||
| 19664 | res = DAG.getLoad(PtrVT, dl, Chain, res, MachinePointerInfo()); | |||
| 19665 | ||||
| 19666 | // Get the offset of start of .tls section | |||
| 19667 | SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl, | |||
| 19668 | GA->getValueType(0), | |||
| 19669 | GA->getOffset(), X86II::MO_SECREL); | |||
| 19670 | SDValue Offset = DAG.getNode(X86ISD::Wrapper, dl, PtrVT, TGA); | |||
| 19671 | ||||
| 19672 | // The address of the thread local variable is the add of the thread | |||
| 19673 | // pointer with the offset of the variable. | |||
| 19674 | return DAG.getNode(ISD::ADD, dl, PtrVT, res, Offset); | |||
| 19675 | } | |||
| 19676 | ||||
| 19677 | llvm_unreachable("TLS not implemented for this target.")__builtin_unreachable(); | |||
| 19678 | } | |||
| 19679 | ||||
| 19680 | /// Lower SRA_PARTS and friends, which return two i32 values | |||
| 19681 | /// and take a 2 x i32 value to shift plus a shift amount. | |||
| 19682 | /// TODO: Can this be moved to general expansion code? | |||
| 19683 | static SDValue LowerShiftParts(SDValue Op, SelectionDAG &DAG) { | |||
| 19684 | SDValue Lo, Hi; | |||
| 19685 | DAG.getTargetLoweringInfo().expandShiftParts(Op.getNode(), Lo, Hi, DAG); | |||
| 19686 | return DAG.getMergeValues({Lo, Hi}, SDLoc(Op)); | |||
| 19687 | } | |||
| 19688 | ||||
| 19689 | static SDValue LowerFunnelShift(SDValue Op, const X86Subtarget &Subtarget, | |||
| 19690 | SelectionDAG &DAG) { | |||
| 19691 | MVT VT = Op.getSimpleValueType(); | |||
| 19692 | assert((Op.getOpcode() == ISD::FSHL || Op.getOpcode() == ISD::FSHR) &&((void)0) | |||
| 19693 | "Unexpected funnel shift opcode!")((void)0); | |||
| 19694 | ||||
| 19695 | SDLoc DL(Op); | |||
| 19696 | SDValue Op0 = Op.getOperand(0); | |||
| 19697 | SDValue Op1 = Op.getOperand(1); | |||
| 19698 | SDValue Amt = Op.getOperand(2); | |||
| 19699 | ||||
| 19700 | bool IsFSHR = Op.getOpcode() == ISD::FSHR; | |||
| 19701 | ||||
| 19702 | if (VT.isVector()) { | |||
| 19703 | assert(Subtarget.hasVBMI2() && "Expected VBMI2")((void)0); | |||
| 19704 | ||||
| 19705 | if (IsFSHR) | |||
| 19706 | std::swap(Op0, Op1); | |||
| 19707 | ||||
| 19708 | // With AVX512, but not VLX we need to widen to get a 512-bit result type. | |||
| 19709 | if (!Subtarget.hasVLX() && !VT.is512BitVector()) { | |||
| 19710 | Op0 = widenSubVector(Op0, false, Subtarget, DAG, DL, 512); | |||
| 19711 | Op1 = widenSubVector(Op1, false, Subtarget, DAG, DL, 512); | |||
| 19712 | } | |||
| 19713 | ||||
| 19714 | SDValue Funnel; | |||
| 19715 | APInt APIntShiftAmt; | |||
| 19716 | MVT ResultVT = Op0.getSimpleValueType(); | |||
| 19717 | if (X86::isConstantSplat(Amt, APIntShiftAmt)) { | |||
| 19718 | uint64_t ShiftAmt = APIntShiftAmt.urem(VT.getScalarSizeInBits()); | |||
| 19719 | Funnel = | |||
| 19720 | DAG.getNode(IsFSHR ? X86ISD::VSHRD : X86ISD::VSHLD, DL, ResultVT, Op0, | |||
| 19721 | Op1, DAG.getTargetConstant(ShiftAmt, DL, MVT::i8)); | |||
| 19722 | } else { | |||
| 19723 | if (!Subtarget.hasVLX() && !VT.is512BitVector()) | |||
| 19724 | Amt = widenSubVector(Amt, false, Subtarget, DAG, DL, 512); | |||
| 19725 | Funnel = DAG.getNode(IsFSHR ? X86ISD::VSHRDV : X86ISD::VSHLDV, DL, | |||
| 19726 | ResultVT, Op0, Op1, Amt); | |||
| 19727 | } | |||
| 19728 | if (!Subtarget.hasVLX() && !VT.is512BitVector()) | |||
| 19729 | Funnel = extractSubVector(Funnel, 0, DAG, DL, VT.getSizeInBits()); | |||
| 19730 | return Funnel; | |||
| 19731 | } | |||
| 19732 | assert(((void)0) | |||
| 19733 | (VT == MVT::i8 || VT == MVT::i16 || VT == MVT::i32 || VT == MVT::i64) &&((void)0) | |||
| 19734 | "Unexpected funnel shift type!")((void)0); | |||
| 19735 | ||||
| 19736 | // Expand slow SHLD/SHRD cases if we are not optimizing for size. | |||
| 19737 | bool OptForSize = DAG.shouldOptForSize(); | |||
| 19738 | bool ExpandFunnel = !OptForSize && Subtarget.isSHLDSlow(); | |||
| 19739 | ||||
| 19740 | // fshl(x,y,z) -> (((aext(x) << bw) | zext(y)) << (z & (bw-1))) >> bw. | |||
| 19741 | // fshr(x,y,z) -> (((aext(x) << bw) | zext(y)) >> (z & (bw-1))). | |||
| 19742 | if ((VT == MVT::i8 || (ExpandFunnel && VT == MVT::i16)) && | |||
| 19743 | !isa<ConstantSDNode>(Amt)) { | |||
| 19744 | unsigned EltSizeInBits = VT.getScalarSizeInBits(); | |||
| 19745 | SDValue Mask = DAG.getConstant(EltSizeInBits - 1, DL, Amt.getValueType()); | |||
| 19746 | SDValue HiShift = DAG.getConstant(EltSizeInBits, DL, Amt.getValueType()); | |||
| 19747 | Op0 = DAG.getAnyExtOrTrunc(Op0, DL, MVT::i32); | |||
| 19748 | Op1 = DAG.getZExtOrTrunc(Op1, DL, MVT::i32); | |||
| 19749 | Amt = DAG.getNode(ISD::AND, DL, Amt.getValueType(), Amt, Mask); | |||
| 19750 | SDValue Res = DAG.getNode(ISD::SHL, DL, MVT::i32, Op0, HiShift); | |||
| 19751 | Res = DAG.getNode(ISD::OR, DL, MVT::i32, Res, Op1); | |||
| 19752 | if (IsFSHR) { | |||
| 19753 | Res = DAG.getNode(ISD::SRL, DL, MVT::i32, Res, Amt); | |||
| 19754 | } else { | |||
| 19755 | Res = DAG.getNode(ISD::SHL, DL, MVT::i32, Res, Amt); | |||
| 19756 | Res = DAG.getNode(ISD::SRL, DL, MVT::i32, Res, HiShift); | |||
| 19757 | } | |||
| 19758 | return DAG.getZExtOrTrunc(Res, DL, VT); | |||
| 19759 | } | |||
| 19760 | ||||
| 19761 | if (VT == MVT::i8 || ExpandFunnel) | |||
| 19762 | return SDValue(); | |||
| 19763 | ||||
| 19764 | // i16 needs to modulo the shift amount, but i32/i64 have implicit modulo. | |||
| 19765 | if (VT == MVT::i16) { | |||
| 19766 | Amt = DAG.getNode(ISD::AND, DL, Amt.getValueType(), Amt, | |||
| 19767 | DAG.getConstant(15, DL, Amt.getValueType())); | |||
| 19768 | unsigned FSHOp = (IsFSHR ? X86ISD::FSHR : X86ISD::FSHL); | |||
| 19769 | return DAG.getNode(FSHOp, DL, VT, Op0, Op1, Amt); | |||
| 19770 | } | |||
| 19771 | ||||
| 19772 | return Op; | |||
| 19773 | } | |||
| 19774 | ||||
| 19775 | // Try to use a packed vector operation to handle i64 on 32-bit targets when | |||
| 19776 | // AVX512DQ is enabled. | |||
| 19777 | static SDValue LowerI64IntToFP_AVX512DQ(SDValue Op, SelectionDAG &DAG, | |||
| 19778 | const X86Subtarget &Subtarget) { | |||
| 19779 | assert((Op.getOpcode() == ISD::SINT_TO_FP ||((void)0) | |||
| 19780 | Op.getOpcode() == ISD::STRICT_SINT_TO_FP ||((void)0) | |||
| 19781 | Op.getOpcode() == ISD::STRICT_UINT_TO_FP ||((void)0) | |||
| 19782 | Op.getOpcode() == ISD::UINT_TO_FP) &&((void)0) | |||
| 19783 | "Unexpected opcode!")((void)0); | |||
| 19784 | bool IsStrict = Op->isStrictFPOpcode(); | |||
| 19785 | unsigned OpNo = IsStrict ? 1 : 0; | |||
| 19786 | SDValue Src = Op.getOperand(OpNo); | |||
| 19787 | MVT SrcVT = Src.getSimpleValueType(); | |||
| 19788 | MVT VT = Op.getSimpleValueType(); | |||
| 19789 | ||||
| 19790 | if (!Subtarget.hasDQI() || SrcVT != MVT::i64 || Subtarget.is64Bit() || | |||
| 19791 | (VT != MVT::f32 && VT != MVT::f64)) | |||
| 19792 | return SDValue(); | |||
| 19793 | ||||
| 19794 | // Pack the i64 into a vector, do the operation and extract. | |||
| 19795 | ||||
| 19796 | // Using 256-bit to ensure result is 128-bits for f32 case. | |||
| 19797 | unsigned NumElts = Subtarget.hasVLX() ? 4 : 8; | |||
| 19798 | MVT VecInVT = MVT::getVectorVT(MVT::i64, NumElts); | |||
| 19799 | MVT VecVT = MVT::getVectorVT(VT, NumElts); | |||
| 19800 | ||||
| 19801 | SDLoc dl(Op); | |||
| 19802 | SDValue InVec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecInVT, Src); | |||
| 19803 | if (IsStrict) { | |||
| 19804 | SDValue CvtVec = DAG.getNode(Op.getOpcode(), dl, {VecVT, MVT::Other}, | |||
| 19805 | {Op.getOperand(0), InVec}); | |||
| 19806 | SDValue Chain = CvtVec.getValue(1); | |||
| 19807 | SDValue Value = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, CvtVec, | |||
| 19808 | DAG.getIntPtrConstant(0, dl)); | |||
| 19809 | return DAG.getMergeValues({Value, Chain}, dl); | |||
| 19810 | } | |||
| 19811 | ||||
| 19812 | SDValue CvtVec = DAG.getNode(Op.getOpcode(), dl, VecVT, InVec); | |||
| 19813 | ||||
| 19814 | return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, CvtVec, | |||
| 19815 | DAG.getIntPtrConstant(0, dl)); | |||
| 19816 | } | |||
| 19817 | ||||
| 19818 | static bool useVectorCast(unsigned Opcode, MVT FromVT, MVT ToVT, | |||
| 19819 | const X86Subtarget &Subtarget) { | |||
| 19820 | switch (Opcode) { | |||
| 19821 | case ISD::SINT_TO_FP: | |||
| 19822 | // TODO: Handle wider types with AVX/AVX512. | |||
| 19823 | if (!Subtarget.hasSSE2() || FromVT != MVT::v4i32) | |||
| 19824 | return false; | |||
| 19825 | // CVTDQ2PS or (V)CVTDQ2PD | |||
| 19826 | return ToVT == MVT::v4f32 || (Subtarget.hasAVX() && ToVT == MVT::v4f64); | |||
| 19827 | ||||
| 19828 | case ISD::UINT_TO_FP: | |||
| 19829 | // TODO: Handle wider types and i64 elements. | |||
| 19830 | if (!Subtarget.hasAVX512() || FromVT != MVT::v4i32) | |||
| 19831 | return false; | |||
| 19832 | // VCVTUDQ2PS or VCVTUDQ2PD | |||
| 19833 | return ToVT == MVT::v4f32 || ToVT == MVT::v4f64; | |||
| 19834 | ||||
| 19835 | default: | |||
| 19836 | return false; | |||
| 19837 | } | |||
| 19838 | } | |||
| 19839 | ||||
| 19840 | /// Given a scalar cast operation that is extracted from a vector, try to | |||
| 19841 | /// vectorize the cast op followed by extraction. This will avoid an expensive | |||
| 19842 | /// round-trip between XMM and GPR. | |||
| 19843 | static SDValue vectorizeExtractedCast(SDValue Cast, SelectionDAG &DAG, | |||
| 19844 | const X86Subtarget &Subtarget) { | |||
| 19845 | // TODO: This could be enhanced to handle smaller integer types by peeking | |||
| 19846 | // through an extend. | |||
| 19847 | SDValue Extract = Cast.getOperand(0); | |||
| 19848 | MVT DestVT = Cast.getSimpleValueType(); | |||
| 19849 | if (Extract.getOpcode() != ISD::EXTRACT_VECTOR_ELT || | |||
| 19850 | !isa<ConstantSDNode>(Extract.getOperand(1))) | |||
| 19851 | return SDValue(); | |||
| 19852 | ||||
| 19853 | // See if we have a 128-bit vector cast op for this type of cast. | |||
| 19854 | SDValue VecOp = Extract.getOperand(0); | |||
| 19855 | MVT FromVT = VecOp.getSimpleValueType(); | |||
| 19856 | unsigned NumEltsInXMM = 128 / FromVT.getScalarSizeInBits(); | |||
| 19857 | MVT Vec128VT = MVT::getVectorVT(FromVT.getScalarType(), NumEltsInXMM); | |||
| 19858 | MVT ToVT = MVT::getVectorVT(DestVT, NumEltsInXMM); | |||
| 19859 | if (!useVectorCast(Cast.getOpcode(), Vec128VT, ToVT, Subtarget)) | |||
| 19860 | return SDValue(); | |||
| 19861 | ||||
| 19862 | // If we are extracting from a non-zero element, first shuffle the source | |||
| 19863 | // vector to allow extracting from element zero. | |||
| 19864 | SDLoc DL(Cast); | |||
| 19865 | if (!isNullConstant(Extract.getOperand(1))) { | |||
| 19866 | SmallVector<int, 16> Mask(FromVT.getVectorNumElements(), -1); | |||
| 19867 | Mask[0] = Extract.getConstantOperandVal(1); | |||
| 19868 | VecOp = DAG.getVectorShuffle(FromVT, DL, VecOp, DAG.getUNDEF(FromVT), Mask); | |||
| 19869 | } | |||
| 19870 | // If the source vector is wider than 128-bits, extract the low part. Do not | |||
| 19871 | // create an unnecessarily wide vector cast op. | |||
| 19872 | if (FromVT != Vec128VT) | |||
| 19873 | VecOp = extract128BitVector(VecOp, 0, DAG, DL); | |||
| 19874 | ||||
| 19875 | // cast (extelt V, 0) --> extelt (cast (extract_subv V)), 0 | |||
| 19876 | // cast (extelt V, C) --> extelt (cast (extract_subv (shuffle V, [C...]))), 0 | |||
| 19877 | SDValue VCast = DAG.getNode(Cast.getOpcode(), DL, ToVT, VecOp); | |||
| 19878 | return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, DestVT, VCast, | |||
| 19879 | DAG.getIntPtrConstant(0, DL)); | |||
| 19880 | } | |||
| 19881 | ||||
| 19882 | /// Given a scalar cast to FP with a cast to integer operand (almost an ftrunc), | |||
| 19883 | /// try to vectorize the cast ops. This will avoid an expensive round-trip | |||
| 19884 | /// between XMM and GPR. | |||
| 19885 | static SDValue lowerFPToIntToFP(SDValue CastToFP, SelectionDAG &DAG, | |||
| 19886 | const X86Subtarget &Subtarget) { | |||
| 19887 | // TODO: Allow FP_TO_UINT. | |||
| 19888 | SDValue CastToInt = CastToFP.getOperand(0); | |||
| 19889 | MVT VT = CastToFP.getSimpleValueType(); | |||
| 19890 | if (CastToInt.getOpcode() != ISD::FP_TO_SINT || VT.isVector()) | |||
| 19891 | return SDValue(); | |||
| 19892 | ||||
| 19893 | MVT IntVT = CastToInt.getSimpleValueType(); | |||
| 19894 | SDValue X = CastToInt.getOperand(0); | |||
| 19895 | MVT SrcVT = X.getSimpleValueType(); | |||
| 19896 | if (SrcVT != MVT::f32 && SrcVT != MVT::f64) | |||
| 19897 | return SDValue(); | |||
| 19898 | ||||
| 19899 | // See if we have 128-bit vector cast instructions for this type of cast. | |||
| 19900 | // We need cvttps2dq/cvttpd2dq and cvtdq2ps/cvtdq2pd. | |||
| 19901 | if (!Subtarget.hasSSE2() || (VT != MVT::f32 && VT != MVT::f64) || | |||
| 19902 | IntVT != MVT::i32) | |||
| 19903 | return SDValue(); | |||
| 19904 | ||||
| 19905 | unsigned SrcSize = SrcVT.getSizeInBits(); | |||
| 19906 | unsigned IntSize = IntVT.getSizeInBits(); | |||
| 19907 | unsigned VTSize = VT.getSizeInBits(); | |||
| 19908 | MVT VecSrcVT = MVT::getVectorVT(SrcVT, 128 / SrcSize); | |||
| 19909 | MVT VecIntVT = MVT::getVectorVT(IntVT, 128 / IntSize); | |||
| 19910 | MVT VecVT = MVT::getVectorVT(VT, 128 / VTSize); | |||
| 19911 | ||||
| 19912 | // We need target-specific opcodes if this is v2f64 -> v4i32 -> v2f64. | |||
| 19913 | unsigned ToIntOpcode = | |||
| 19914 | SrcSize != IntSize ? X86ISD::CVTTP2SI : (unsigned)ISD::FP_TO_SINT; | |||
| 19915 | unsigned ToFPOpcode = | |||
| 19916 | IntSize != VTSize ? X86ISD::CVTSI2P : (unsigned)ISD::SINT_TO_FP; | |||
| 19917 | ||||
| 19918 | // sint_to_fp (fp_to_sint X) --> extelt (sint_to_fp (fp_to_sint (s2v X))), 0 | |||
| 19919 | // | |||
| 19920 | // We are not defining the high elements (for example, zero them) because | |||
| 19921 | // that could nullify any performance advantage that we hoped to gain from | |||
| 19922 | // this vector op hack. We do not expect any adverse effects (like denorm | |||
| 19923 | // penalties) with cast ops. | |||
| 19924 | SDLoc DL(CastToFP); | |||
| 19925 | SDValue ZeroIdx = DAG.getIntPtrConstant(0, DL); | |||
| 19926 | SDValue VecX = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecSrcVT, X); | |||
| 19927 | SDValue VCastToInt = DAG.getNode(ToIntOpcode, DL, VecIntVT, VecX); | |||
| 19928 | SDValue VCastToFP = DAG.getNode(ToFPOpcode, DL, VecVT, VCastToInt); | |||
| 19929 | return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT, VCastToFP, ZeroIdx); | |||
| 19930 | } | |||
| 19931 | ||||
| 19932 | static SDValue lowerINT_TO_FP_vXi64(SDValue Op, SelectionDAG &DAG, | |||
| 19933 | const X86Subtarget &Subtarget) { | |||
| 19934 | SDLoc DL(Op); | |||
| 19935 | bool IsStrict = Op->isStrictFPOpcode(); | |||
| 19936 | MVT VT = Op->getSimpleValueType(0); | |||
| 19937 | SDValue Src = Op->getOperand(IsStrict ? 1 : 0); | |||
| 19938 | ||||
| 19939 | if (Subtarget.hasDQI()) { | |||
| 19940 | assert(!Subtarget.hasVLX() && "Unexpected features")((void)0); | |||
| 19941 | ||||
| 19942 | assert((Src.getSimpleValueType() == MVT::v2i64 ||((void)0) | |||
| 19943 | Src.getSimpleValueType() == MVT::v4i64) &&((void)0) | |||
| 19944 | "Unsupported custom type")((void)0); | |||
| 19945 | ||||
| 19946 | // With AVX512DQ, but not VLX we need to widen to get a 512-bit result type. | |||
| 19947 | assert((VT == MVT::v4f32 || VT == MVT::v2f64 || VT == MVT::v4f64) &&((void)0) | |||
| 19948 | "Unexpected VT!")((void)0); | |||
| 19949 | MVT WideVT = VT == MVT::v4f32 ? MVT::v8f32 : MVT::v8f64; | |||
| 19950 | ||||
| 19951 | // Need to concat with zero vector for strict fp to avoid spurious | |||
| 19952 | // exceptions. | |||
| 19953 | SDValue Tmp = IsStrict ? DAG.getConstant(0, DL, MVT::v8i64) | |||
| 19954 | : DAG.getUNDEF(MVT::v8i64); | |||
| 19955 | Src = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, MVT::v8i64, Tmp, Src, | |||
| 19956 | DAG.getIntPtrConstant(0, DL)); | |||
| 19957 | SDValue Res, Chain; | |||
| 19958 | if (IsStrict) { | |||
| 19959 | Res = DAG.getNode(Op.getOpcode(), DL, {WideVT, MVT::Other}, | |||
| 19960 | {Op->getOperand(0), Src}); | |||
| 19961 | Chain = Res.getValue(1); | |||
| 19962 | } else { | |||
| 19963 | Res = DAG.getNode(Op.getOpcode(), DL, WideVT, Src); | |||
| 19964 | } | |||
| 19965 | ||||
| 19966 | Res = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, Res, | |||
| 19967 | DAG.getIntPtrConstant(0, DL)); | |||
| 19968 | ||||
| 19969 | if (IsStrict) | |||
| 19970 | return DAG.getMergeValues({Res, Chain}, DL); | |||
| 19971 | return Res; | |||
| 19972 | } | |||
| 19973 | ||||
| 19974 | bool IsSigned = Op->getOpcode() == ISD::SINT_TO_FP || | |||
| 19975 | Op->getOpcode() == ISD::STRICT_SINT_TO_FP; | |||
| 19976 | if (VT != MVT::v4f32 || IsSigned) | |||
| 19977 | return SDValue(); | |||
| 19978 | ||||
| 19979 | SDValue Zero = DAG.getConstant(0, DL, MVT::v4i64); | |||
| 19980 | SDValue One = DAG.getConstant(1, DL, MVT::v4i64); | |||
| 19981 | SDValue Sign = DAG.getNode(ISD::OR, DL, MVT::v4i64, | |||
| 19982 | DAG.getNode(ISD::SRL, DL, MVT::v4i64, Src, One), | |||
| 19983 | DAG.getNode(ISD::AND, DL, MVT::v4i64, Src, One)); | |||
| 19984 | SDValue IsNeg = DAG.getSetCC(DL, MVT::v4i64, Src, Zero, ISD::SETLT); | |||
| 19985 | SDValue SignSrc = DAG.getSelect(DL, MVT::v4i64, IsNeg, Sign, Src); | |||
| 19986 | SmallVector<SDValue, 4> SignCvts(4); | |||
| 19987 | SmallVector<SDValue, 4> Chains(4); | |||
| 19988 | for (int i = 0; i != 4; ++i) { | |||
| 19989 | SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i64, SignSrc, | |||
| 19990 | DAG.getIntPtrConstant(i, DL)); | |||
| 19991 | if (IsStrict) { | |||
| 19992 | SignCvts[i] = | |||
| 19993 | DAG.getNode(ISD::STRICT_SINT_TO_FP, DL, {MVT::f32, MVT::Other}, | |||
| 19994 | {Op.getOperand(0), Elt}); | |||
| 19995 | Chains[i] = SignCvts[i].getValue(1); | |||
| 19996 | } else { | |||
| 19997 | SignCvts[i] = DAG.getNode(ISD::SINT_TO_FP, DL, MVT::f32, Elt); | |||
| 19998 | } | |||
| 19999 | } | |||
| 20000 | SDValue SignCvt = DAG.getBuildVector(VT, DL, SignCvts); | |||
| 20001 | ||||
| 20002 | SDValue Slow, Chain; | |||
| 20003 | if (IsStrict) { | |||
| 20004 | Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains); | |||
| 20005 | Slow = DAG.getNode(ISD::STRICT_FADD, DL, {MVT::v4f32, MVT::Other}, | |||
| 20006 | {Chain, SignCvt, SignCvt}); | |||
| 20007 | Chain = Slow.getValue(1); | |||
| 20008 | } else { | |||
| 20009 | Slow = DAG.getNode(ISD::FADD, DL, MVT::v4f32, SignCvt, SignCvt); | |||
| 20010 | } | |||
| 20011 | ||||
| 20012 | IsNeg = DAG.getNode(ISD::TRUNCATE, DL, MVT::v4i32, IsNeg); | |||
| 20013 | SDValue Cvt = DAG.getSelect(DL, MVT::v4f32, IsNeg, Slow, SignCvt); | |||
| 20014 | ||||
| 20015 | if (IsStrict) | |||
| 20016 | return DAG.getMergeValues({Cvt, Chain}, DL); | |||
| 20017 | ||||
| 20018 | return Cvt; | |||
| 20019 | } | |||
| 20020 | ||||
| 20021 | SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op, | |||
| 20022 | SelectionDAG &DAG) const { | |||
| 20023 | bool IsStrict = Op->isStrictFPOpcode(); | |||
| 20024 | unsigned OpNo = IsStrict ? 1 : 0; | |||
| 20025 | SDValue Src = Op.getOperand(OpNo); | |||
| 20026 | SDValue Chain = IsStrict ? Op->getOperand(0) : DAG.getEntryNode(); | |||
| 20027 | MVT SrcVT = Src.getSimpleValueType(); | |||
| 20028 | MVT VT = Op.getSimpleValueType(); | |||
| 20029 | SDLoc dl(Op); | |||
| 20030 | ||||
| 20031 | if (SDValue Extract = vectorizeExtractedCast(Op, DAG, Subtarget)) | |||
| 20032 | return Extract; | |||
| 20033 | ||||
| 20034 | if (SDValue R = lowerFPToIntToFP(Op, DAG, Subtarget)) | |||
| 20035 | return R; | |||
| 20036 | ||||
| 20037 | if (SrcVT.isVector()) { | |||
| 20038 | if (SrcVT == MVT::v2i32 && VT == MVT::v2f64) { | |||
| 20039 | // Note: Since v2f64 is a legal type. We don't need to zero extend the | |||
| 20040 | // source for strict FP. | |||
| 20041 | if (IsStrict) | |||
| 20042 | return DAG.getNode( | |||
| 20043 | X86ISD::STRICT_CVTSI2P, dl, {VT, MVT::Other}, | |||
| 20044 | {Chain, DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4i32, Src, | |||
| 20045 | DAG.getUNDEF(SrcVT))}); | |||
| 20046 | return DAG.getNode(X86ISD::CVTSI2P, dl, VT, | |||
| 20047 | DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4i32, Src, | |||
| 20048 | DAG.getUNDEF(SrcVT))); | |||
| 20049 | } | |||
| 20050 | if (SrcVT == MVT::v2i64 || SrcVT == MVT::v4i64) | |||
| 20051 | return lowerINT_TO_FP_vXi64(Op, DAG, Subtarget); | |||
| 20052 | ||||
| 20053 | return SDValue(); | |||
| 20054 | } | |||
| 20055 | ||||
| 20056 | assert(SrcVT <= MVT::i64 && SrcVT >= MVT::i16 &&((void)0) | |||
| 20057 | "Unknown SINT_TO_FP to lower!")((void)0); | |||
| 20058 | ||||
| 20059 | bool UseSSEReg = isScalarFPTypeInSSEReg(VT); | |||
| 20060 | ||||
| 20061 | // These are really Legal; return the operand so the caller accepts it as | |||
| 20062 | // Legal. | |||
| 20063 | if (SrcVT == MVT::i32 && UseSSEReg) | |||
| 20064 | return Op; | |||
| 20065 | if (SrcVT == MVT::i64 && UseSSEReg && Subtarget.is64Bit()) | |||
| 20066 | return Op; | |||
| 20067 | ||||
| 20068 | if (SDValue V = LowerI64IntToFP_AVX512DQ(Op, DAG, Subtarget)) | |||
| 20069 | return V; | |||
| 20070 | ||||
| 20071 | // SSE doesn't have an i16 conversion so we need to promote. | |||
| 20072 | if (SrcVT == MVT::i16 && (UseSSEReg || VT == MVT::f128)) { | |||
| 20073 | SDValue Ext = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::i32, Src); | |||
| 20074 | if (IsStrict) | |||
| 20075 | return DAG.getNode(ISD::STRICT_SINT_TO_FP, dl, {VT, MVT::Other}, | |||
| 20076 | {Chain, Ext}); | |||
| 20077 | ||||
| 20078 | return DAG.getNode(ISD::SINT_TO_FP, dl, VT, Ext); | |||
| 20079 | } | |||
| 20080 | ||||
| 20081 | if (VT == MVT::f128) | |||
| 20082 | return SDValue(); | |||
| 20083 | ||||
| 20084 | SDValue ValueToStore = Src; | |||
| 20085 | if (SrcVT == MVT::i64 && Subtarget.hasSSE2() && !Subtarget.is64Bit()) | |||
| 20086 | // Bitcasting to f64 here allows us to do a single 64-bit store from | |||
| 20087 | // an SSE register, avoiding the store forwarding penalty that would come | |||
| 20088 | // with two 32-bit stores. | |||
| 20089 | ValueToStore = DAG.getBitcast(MVT::f64, ValueToStore); | |||
| 20090 | ||||
| 20091 | unsigned Size = SrcVT.getStoreSize(); | |||
| 20092 | Align Alignment(Size); | |||
| 20093 | MachineFunction &MF = DAG.getMachineFunction(); | |||
| 20094 | auto PtrVT = getPointerTy(MF.getDataLayout()); | |||
| 20095 | int SSFI = MF.getFrameInfo().CreateStackObject(Size, Alignment, false); | |||
| 20096 | MachinePointerInfo MPI = | |||
| 20097 | MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI); | |||
| 20098 | SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT); | |||
| 20099 | Chain = DAG.getStore(Chain, dl, ValueToStore, StackSlot, MPI, Alignment); | |||
| 20100 | std::pair<SDValue, SDValue> Tmp = | |||
| 20101 | BuildFILD(VT, SrcVT, dl, Chain, StackSlot, MPI, Alignment, DAG); | |||
| 20102 | ||||
| 20103 | if (IsStrict) | |||
| 20104 | return DAG.getMergeValues({Tmp.first, Tmp.second}, dl); | |||
| 20105 | ||||
| 20106 | return Tmp.first; | |||
| 20107 | } | |||
| 20108 | ||||
| 20109 | std::pair<SDValue, SDValue> X86TargetLowering::BuildFILD( | |||
| 20110 | EVT DstVT, EVT SrcVT, const SDLoc &DL, SDValue Chain, SDValue Pointer, | |||
| 20111 | MachinePointerInfo PtrInfo, Align Alignment, SelectionDAG &DAG) const { | |||
| 20112 | // Build the FILD | |||
| 20113 | SDVTList Tys; | |||
| 20114 | bool useSSE = isScalarFPTypeInSSEReg(DstVT); | |||
| 20115 | if (useSSE) | |||
| 20116 | Tys = DAG.getVTList(MVT::f80, MVT::Other); | |||
| 20117 | else | |||
| 20118 | Tys = DAG.getVTList(DstVT, MVT::Other); | |||
| 20119 | ||||
| 20120 | SDValue FILDOps[] = {Chain, Pointer}; | |||
| 20121 | SDValue Result = | |||
| 20122 | DAG.getMemIntrinsicNode(X86ISD::FILD, DL, Tys, FILDOps, SrcVT, PtrInfo, | |||
| 20123 | Alignment, MachineMemOperand::MOLoad); | |||
| 20124 | Chain = Result.getValue(1); | |||
| 20125 | ||||
| 20126 | if (useSSE) { | |||
| 20127 | MachineFunction &MF = DAG.getMachineFunction(); | |||
| 20128 | unsigned SSFISize = DstVT.getStoreSize(); | |||
| 20129 | int SSFI = | |||
| 20130 | MF.getFrameInfo().CreateStackObject(SSFISize, Align(SSFISize), false); | |||
| 20131 | auto PtrVT = getPointerTy(MF.getDataLayout()); | |||
| 20132 | SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT); | |||
| 20133 | Tys = DAG.getVTList(MVT::Other); | |||
| 20134 | SDValue FSTOps[] = {Chain, Result, StackSlot}; | |||
| 20135 | MachineMemOperand *StoreMMO = DAG.getMachineFunction().getMachineMemOperand( | |||
| 20136 | MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI), | |||
| 20137 | MachineMemOperand::MOStore, SSFISize, Align(SSFISize)); | |||
| 20138 | ||||
| 20139 | Chain = | |||
| 20140 | DAG.getMemIntrinsicNode(X86ISD::FST, DL, Tys, FSTOps, DstVT, StoreMMO); | |||
| 20141 | Result = DAG.getLoad( | |||
| 20142 | DstVT, DL, Chain, StackSlot, | |||
| 20143 | MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI)); | |||
| 20144 | Chain = Result.getValue(1); | |||
| 20145 | } | |||
| 20146 | ||||
| 20147 | return { Result, Chain }; | |||
| 20148 | } | |||
| 20149 | ||||
| 20150 | /// Horizontal vector math instructions may be slower than normal math with | |||
| 20151 | /// shuffles. Limit horizontal op codegen based on size/speed trade-offs, uarch | |||
| 20152 | /// implementation, and likely shuffle complexity of the alternate sequence. | |||
| 20153 | static bool shouldUseHorizontalOp(bool IsSingleSource, SelectionDAG &DAG, | |||
| 20154 | const X86Subtarget &Subtarget) { | |||
| 20155 | bool IsOptimizingSize = DAG.shouldOptForSize(); | |||
| 20156 | bool HasFastHOps = Subtarget.hasFastHorizontalOps(); | |||
| 20157 | return !IsSingleSource || IsOptimizingSize || HasFastHOps; | |||
| 20158 | } | |||
| 20159 | ||||
| 20160 | /// 64-bit unsigned integer to double expansion. | |||
| 20161 | static SDValue LowerUINT_TO_FP_i64(SDValue Op, SelectionDAG &DAG, | |||
| 20162 | const X86Subtarget &Subtarget) { | |||
| 20163 | // We can't use this algorithm for strict fp. It produces -0.0 instead of +0.0 | |||
| 20164 | // when converting 0 when rounding toward negative infinity. Caller will | |||
| 20165 | // fall back to Expand for when i64 or is legal or use FILD in 32-bit mode. | |||
| 20166 | assert(!Op->isStrictFPOpcode() && "Expected non-strict uint_to_fp!")((void)0); | |||
| 20167 | // This algorithm is not obvious. Here it is what we're trying to output: | |||
| 20168 | /* | |||
| 20169 | movq %rax, %xmm0 | |||
| 20170 | punpckldq (c0), %xmm0 // c0: (uint4){ 0x43300000U, 0x45300000U, 0U, 0U } | |||
| 20171 | subpd (c1), %xmm0 // c1: (double2){ 0x1.0p52, 0x1.0p52 * 0x1.0p32 } | |||
| 20172 | #ifdef __SSE3__ | |||
| 20173 | haddpd %xmm0, %xmm0 | |||
| 20174 | #else | |||
| 20175 | pshufd $0x4e, %xmm0, %xmm1 | |||
| 20176 | addpd %xmm1, %xmm0 | |||
| 20177 | #endif | |||
| 20178 | */ | |||
| 20179 | ||||
| 20180 | SDLoc dl(Op); | |||
| 20181 | LLVMContext *Context = DAG.getContext(); | |||
| 20182 | ||||
| 20183 | // Build some magic constants. | |||
| 20184 | static const uint32_t CV0[] = { 0x43300000, 0x45300000, 0, 0 }; | |||
| 20185 | Constant *C0 = ConstantDataVector::get(*Context, CV0); | |||
| 20186 | auto PtrVT = DAG.getTargetLoweringInfo().getPointerTy(DAG.getDataLayout()); | |||
| 20187 | SDValue CPIdx0 = DAG.getConstantPool(C0, PtrVT, Align(16)); | |||
| 20188 | ||||
| 20189 | SmallVector<Constant*,2> CV1; | |||
| 20190 | CV1.push_back( | |||
| 20191 | ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble(), | |||
| 20192 | APInt(64, 0x4330000000000000ULL)))); | |||
| 20193 | CV1.push_back( | |||
| 20194 | ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble(), | |||
| 20195 | APInt(64, 0x4530000000000000ULL)))); | |||
| 20196 | Constant *C1 = ConstantVector::get(CV1); | |||
| 20197 | SDValue CPIdx1 = DAG.getConstantPool(C1, PtrVT, Align(16)); | |||
| 20198 | ||||
| 20199 | // Load the 64-bit value into an XMM register. | |||
| 20200 | SDValue XR1 = | |||
| 20201 | DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Op.getOperand(0)); | |||
| 20202 | SDValue CLod0 = DAG.getLoad( | |||
| 20203 | MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0, | |||
| 20204 | MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), Align(16)); | |||
| 20205 | SDValue Unpck1 = | |||
| 20206 | getUnpackl(DAG, dl, MVT::v4i32, DAG.getBitcast(MVT::v4i32, XR1), CLod0); | |||
| 20207 | ||||
| 20208 | SDValue CLod1 = DAG.getLoad( | |||
| 20209 | MVT::v2f64, dl, CLod0.getValue(1), CPIdx1, | |||
| 20210 | MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), Align(16)); | |||
| 20211 | SDValue XR2F = DAG.getBitcast(MVT::v2f64, Unpck1); | |||
| 20212 | // TODO: Are there any fast-math-flags to propagate here? | |||
| 20213 | SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1); | |||
| 20214 | SDValue Result; | |||
| 20215 | ||||
| 20216 | if (Subtarget.hasSSE3() && | |||
| 20217 | shouldUseHorizontalOp(true, DAG, Subtarget)) { | |||
| 20218 | Result = DAG.getNode(X86ISD::FHADD, dl, MVT::v2f64, Sub, Sub); | |||
| 20219 | } else { | |||
| 20220 | SDValue Shuffle = DAG.getVectorShuffle(MVT::v2f64, dl, Sub, Sub, {1,-1}); | |||
| 20221 | Result = DAG.getNode(ISD::FADD, dl, MVT::v2f64, Shuffle, Sub); | |||
| 20222 | } | |||
| 20223 | Result = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Result, | |||
| 20224 | DAG.getIntPtrConstant(0, dl)); | |||
| 20225 | return Result; | |||
| 20226 | } | |||
| 20227 | ||||
| 20228 | /// 32-bit unsigned integer to float expansion. | |||
| 20229 | static SDValue LowerUINT_TO_FP_i32(SDValue Op, SelectionDAG &DAG, | |||
| 20230 | const X86Subtarget &Subtarget) { | |||
| 20231 | unsigned OpNo = Op.getNode()->isStrictFPOpcode() ? 1 : 0; | |||
| 20232 | SDLoc dl(Op); | |||
| 20233 | // FP constant to bias correct the final result. | |||
| 20234 | SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL), dl, | |||
| 20235 | MVT::f64); | |||
| 20236 | ||||
| 20237 | // Load the 32-bit value into an XMM register. | |||
| 20238 | SDValue Load = | |||
| 20239 | DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, Op.getOperand(OpNo)); | |||
| 20240 | ||||
| 20241 | // Zero out the upper parts of the register. | |||
| 20242 | Load = getShuffleVectorZeroOrUndef(Load, 0, true, Subtarget, DAG); | |||
| 20243 | ||||
| 20244 | // Or the load with the bias. | |||
| 20245 | SDValue Or = DAG.getNode( | |||
| 20246 | ISD::OR, dl, MVT::v2i64, | |||
| 20247 | DAG.getBitcast(MVT::v2i64, Load), | |||
| 20248 | DAG.getBitcast(MVT::v2i64, | |||
| 20249 | DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, Bias))); | |||
| 20250 | Or = | |||
| 20251 | DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, | |||
| 20252 | DAG.getBitcast(MVT::v2f64, Or), DAG.getIntPtrConstant(0, dl)); | |||
| 20253 | ||||
| 20254 | if (Op.getNode()->isStrictFPOpcode()) { | |||
| 20255 | // Subtract the bias. | |||
| 20256 | // TODO: Are there any fast-math-flags to propagate here? | |||
| 20257 | SDValue Chain = Op.getOperand(0); | |||
| 20258 | SDValue Sub = DAG.getNode(ISD::STRICT_FSUB, dl, {MVT::f64, MVT::Other}, | |||
| 20259 | {Chain, Or, Bias}); | |||
| 20260 | ||||
| 20261 | if (Op.getValueType() == Sub.getValueType()) | |||
| 20262 | return Sub; | |||
| 20263 | ||||
| 20264 | // Handle final rounding. | |||
| 20265 | std::pair<SDValue, SDValue> ResultPair = DAG.getStrictFPExtendOrRound( | |||
| 20266 | Sub, Sub.getValue(1), dl, Op.getSimpleValueType()); | |||
| 20267 | ||||
| 20268 | return DAG.getMergeValues({ResultPair.first, ResultPair.second}, dl); | |||
| 20269 | } | |||
| 20270 | ||||
| 20271 | // Subtract the bias. | |||
| 20272 | // TODO: Are there any fast-math-flags to propagate here? | |||
| 20273 | SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias); | |||
| 20274 | ||||
| 20275 | // Handle final rounding. | |||
| 20276 | return DAG.getFPExtendOrRound(Sub, dl, Op.getSimpleValueType()); | |||
| 20277 | } | |||
| 20278 | ||||
| 20279 | static SDValue lowerUINT_TO_FP_v2i32(SDValue Op, SelectionDAG &DAG, | |||
| 20280 | const X86Subtarget &Subtarget, | |||
| 20281 | const SDLoc &DL) { | |||
| 20282 | if (Op.getSimpleValueType() != MVT::v2f64) | |||
| 20283 | return SDValue(); | |||
| 20284 | ||||
| 20285 | bool IsStrict = Op->isStrictFPOpcode(); | |||
| 20286 | ||||
| 20287 | SDValue N0 = Op.getOperand(IsStrict ? 1 : 0); | |||
| 20288 | assert(N0.getSimpleValueType() == MVT::v2i32 && "Unexpected input type")((void)0); | |||
| 20289 | ||||
| 20290 | if (Subtarget.hasAVX512()) { | |||
| 20291 | if (!Subtarget.hasVLX()) { | |||
| 20292 | // Let generic type legalization widen this. | |||
| 20293 | if (!IsStrict) | |||
| 20294 | return SDValue(); | |||
| 20295 | // Otherwise pad the integer input with 0s and widen the operation. | |||
| 20296 | N0 = DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v4i32, N0, | |||
| 20297 | DAG.getConstant(0, DL, MVT::v2i32)); | |||
| 20298 | SDValue Res = DAG.getNode(Op->getOpcode(), DL, {MVT::v4f64, MVT::Other}, | |||
| 20299 | {Op.getOperand(0), N0}); | |||
| 20300 | SDValue Chain = Res.getValue(1); | |||
| 20301 | Res = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2f64, Res, | |||
| 20302 | DAG.getIntPtrConstant(0, DL)); | |||
| 20303 | return DAG.getMergeValues({Res, Chain}, DL); | |||
| 20304 | } | |||
| 20305 | ||||
| 20306 | // Legalize to v4i32 type. | |||
| 20307 | N0 = DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v4i32, N0, | |||
| 20308 | DAG.getUNDEF(MVT::v2i32)); | |||
| 20309 | if (IsStrict) | |||
| 20310 | return DAG.getNode(X86ISD::STRICT_CVTUI2P, DL, {MVT::v2f64, MVT::Other}, | |||
| 20311 | {Op.getOperand(0), N0}); | |||
| 20312 | return DAG.getNode(X86ISD::CVTUI2P, DL, MVT::v2f64, N0); | |||
| 20313 | } | |||
| 20314 | ||||
| 20315 | // Zero extend to 2i64, OR with the floating point representation of 2^52. | |||
| 20316 | // This gives us the floating point equivalent of 2^52 + the i32 integer | |||
| 20317 | // since double has 52-bits of mantissa. Then subtract 2^52 in floating | |||
| 20318 | // point leaving just our i32 integers in double format. | |||
| 20319 | SDValue ZExtIn = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v2i64, N0); | |||
| 20320 | SDValue VBias = | |||
| 20321 | DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL), DL, MVT::v2f64); | |||
| 20322 | SDValue Or = DAG.getNode(ISD::OR, DL, MVT::v2i64, ZExtIn, | |||
| 20323 | DAG.getBitcast(MVT::v2i64, VBias)); | |||
| 20324 | Or = DAG.getBitcast(MVT::v2f64, Or); | |||
| 20325 | ||||
| 20326 | if (IsStrict) | |||
| 20327 | return DAG.getNode(ISD::STRICT_FSUB, DL, {MVT::v2f64, MVT::Other}, | |||
| 20328 | {Op.getOperand(0), Or, VBias}); | |||
| 20329 | return DAG.getNode(ISD::FSUB, DL, MVT::v2f64, Or, VBias); | |||
| 20330 | } | |||
| 20331 | ||||
| 20332 | static SDValue lowerUINT_TO_FP_vXi32(SDValue Op, SelectionDAG &DAG, | |||
| 20333 | const X86Subtarget &Subtarget) { | |||
| 20334 | SDLoc DL(Op); | |||
| 20335 | bool IsStrict = Op->isStrictFPOpcode(); | |||
| 20336 | SDValue V = Op->getOperand(IsStrict ? 1 : 0); | |||
| 20337 | MVT VecIntVT = V.getSimpleValueType(); | |||
| 20338 | assert((VecIntVT == MVT::v4i32 || VecIntVT == MVT::v8i32) &&((void)0) | |||
| 20339 | "Unsupported custom type")((void)0); | |||
| 20340 | ||||
| 20341 | if (Subtarget.hasAVX512()) { | |||
| 20342 | // With AVX512, but not VLX we need to widen to get a 512-bit result type. | |||
| 20343 | assert(!Subtarget.hasVLX() && "Unexpected features")((void)0); | |||
| 20344 | MVT VT = Op->getSimpleValueType(0); | |||
| 20345 | ||||
| 20346 | // v8i32->v8f64 is legal with AVX512 so just return it. | |||
| 20347 | if (VT == MVT::v8f64) | |||
| 20348 | return Op; | |||
| 20349 | ||||
| 20350 | assert((VT == MVT::v4f32 || VT == MVT::v8f32 || VT == MVT::v4f64) &&((void)0) | |||
| 20351 | "Unexpected VT!")((void)0); | |||
| 20352 | MVT WideVT = VT == MVT::v4f64 ? MVT::v8f64 : MVT::v16f32; | |||
| 20353 | MVT WideIntVT = VT == MVT::v4f64 ? MVT::v8i32 : MVT::v16i32; | |||
| 20354 | // Need to concat with zero vector for strict fp to avoid spurious | |||
| 20355 | // exceptions. | |||
| 20356 | SDValue Tmp = | |||
| 20357 | IsStrict ? DAG.getConstant(0, DL, WideIntVT) : DAG.getUNDEF(WideIntVT); | |||
| 20358 | V = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, WideIntVT, Tmp, V, | |||
| 20359 | DAG.getIntPtrConstant(0, DL)); | |||
| 20360 | SDValue Res, Chain; | |||
| 20361 | if (IsStrict) { | |||
| 20362 | Res = DAG.getNode(ISD::STRICT_UINT_TO_FP, DL, {WideVT, MVT::Other}, | |||
| 20363 | {Op->getOperand(0), V}); | |||
| 20364 | Chain = Res.getValue(1); | |||
| 20365 | } else { | |||
| 20366 | Res = DAG.getNode(ISD::UINT_TO_FP, DL, WideVT, V); | |||
| 20367 | } | |||
| 20368 | ||||
| 20369 | Res = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, Res, | |||
| 20370 | DAG.getIntPtrConstant(0, DL)); | |||
| 20371 | ||||
| 20372 | if (IsStrict) | |||
| 20373 | return DAG.getMergeValues({Res, Chain}, DL); | |||
| 20374 | return Res; | |||
| 20375 | } | |||
| 20376 | ||||
| 20377 | if (Subtarget.hasAVX() && VecIntVT == MVT::v4i32 && | |||
| 20378 | Op->getSimpleValueType(0) == MVT::v4f64) { | |||
| 20379 | SDValue ZExtIn = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v4i64, V); | |||
| 20380 | Constant *Bias = ConstantFP::get( | |||
| 20381 | *DAG.getContext(), | |||
| 20382 | APFloat(APFloat::IEEEdouble(), APInt(64, 0x4330000000000000ULL))); | |||
| 20383 | auto PtrVT = DAG.getTargetLoweringInfo().getPointerTy(DAG.getDataLayout()); | |||
| 20384 | SDValue CPIdx = DAG.getConstantPool(Bias, PtrVT, Align(8)); | |||
| 20385 | SDVTList Tys = DAG.getVTList(MVT::v4f64, MVT::Other); | |||
| 20386 | SDValue Ops[] = {DAG.getEntryNode(), CPIdx}; | |||
| 20387 | SDValue VBias = DAG.getMemIntrinsicNode( | |||
| 20388 | X86ISD::VBROADCAST_LOAD, DL, Tys, Ops, MVT::f64, | |||
| 20389 | MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), Align(8), | |||
| 20390 | MachineMemOperand::MOLoad); | |||
| 20391 | ||||
| 20392 | SDValue Or = DAG.getNode(ISD::OR, DL, MVT::v4i64, ZExtIn, | |||
| 20393 | DAG.getBitcast(MVT::v4i64, VBias)); | |||
| 20394 | Or = DAG.getBitcast(MVT::v4f64, Or); | |||
| 20395 | ||||
| 20396 | if (IsStrict) | |||
| 20397 | return DAG.getNode(ISD::STRICT_FSUB, DL, {MVT::v4f64, MVT::Other}, | |||
| 20398 | {Op.getOperand(0), Or, VBias}); | |||
| 20399 | return DAG.getNode(ISD::FSUB, DL, MVT::v4f64, Or, VBias); | |||
| 20400 | } | |||
| 20401 | ||||
| 20402 | // The algorithm is the following: | |||
| 20403 | // #ifdef __SSE4_1__ | |||
| 20404 | // uint4 lo = _mm_blend_epi16( v, (uint4) 0x4b000000, 0xaa); | |||
| 20405 | // uint4 hi = _mm_blend_epi16( _mm_srli_epi32(v,16), | |||
| 20406 | // (uint4) 0x53000000, 0xaa); | |||
| 20407 | // #else | |||
| 20408 | // uint4 lo = (v & (uint4) 0xffff) | (uint4) 0x4b000000; | |||
| 20409 | // uint4 hi = (v >> 16) | (uint4) 0x53000000; | |||
| 20410 | // #endif | |||
| 20411 | // float4 fhi = (float4) hi - (0x1.0p39f + 0x1.0p23f); | |||
| 20412 | // return (float4) lo + fhi; | |||
| 20413 | ||||
| 20414 | bool Is128 = VecIntVT == MVT::v4i32; | |||
| 20415 | MVT VecFloatVT = Is128 ? MVT::v4f32 : MVT::v8f32; | |||
| 20416 | // If we convert to something else than the supported type, e.g., to v4f64, | |||
| 20417 | // abort early. | |||
| 20418 | if (VecFloatVT != Op->getSimpleValueType(0)) | |||
| 20419 | return SDValue(); | |||
| 20420 | ||||
| 20421 | // In the #idef/#else code, we have in common: | |||
| 20422 | // - The vector of constants: | |||
| 20423 | // -- 0x4b000000 | |||
| 20424 | // -- 0x53000000 | |||
| 20425 | // - A shift: | |||
| 20426 | // -- v >> 16 | |||
| 20427 | ||||
| 20428 | // Create the splat vector for 0x4b000000. | |||
| 20429 | SDValue VecCstLow = DAG.getConstant(0x4b000000, DL, VecIntVT); | |||
| 20430 | // Create the splat vector for 0x53000000. | |||
| 20431 | SDValue VecCstHigh = DAG.getConstant(0x53000000, DL, VecIntVT); | |||
| 20432 | ||||
| 20433 | // Create the right shift. | |||
| 20434 | SDValue VecCstShift = DAG.getConstant(16, DL, VecIntVT); | |||
| 20435 | SDValue HighShift = DAG.getNode(ISD::SRL, DL, VecIntVT, V, VecCstShift); | |||
| 20436 | ||||
| 20437 | SDValue Low, High; | |||
| 20438 | if (Subtarget.hasSSE41()) { | |||
| 20439 | MVT VecI16VT = Is128 ? MVT::v8i16 : MVT::v16i16; | |||
| 20440 | // uint4 lo = _mm_blend_epi16( v, (uint4) 0x4b000000, 0xaa); | |||
| 20441 | SDValue VecCstLowBitcast = DAG.getBitcast(VecI16VT, VecCstLow); | |||
| 20442 | SDValue VecBitcast = DAG.getBitcast(VecI16VT, V); | |||
| 20443 | // Low will be bitcasted right away, so do not bother bitcasting back to its | |||
| 20444 | // original type. | |||
| 20445 | Low = DAG.getNode(X86ISD::BLENDI, DL, VecI16VT, VecBitcast, | |||
| 20446 | VecCstLowBitcast, DAG.getTargetConstant(0xaa, DL, MVT::i8)); | |||
| 20447 | // uint4 hi = _mm_blend_epi16( _mm_srli_epi32(v,16), | |||
| 20448 | // (uint4) 0x53000000, 0xaa); | |||
| 20449 | SDValue VecCstHighBitcast = DAG.getBitcast(VecI16VT, VecCstHigh); | |||
| 20450 | SDValue VecShiftBitcast = DAG.getBitcast(VecI16VT, HighShift); | |||
| 20451 | // High will be bitcasted right away, so do not bother bitcasting back to | |||
| 20452 | // its original type. | |||
| 20453 | High = DAG.getNode(X86ISD::BLENDI, DL, VecI16VT, VecShiftBitcast, | |||
| 20454 | VecCstHighBitcast, DAG.getTargetConstant(0xaa, DL, MVT::i8)); | |||
| 20455 | } else { | |||
| 20456 | SDValue VecCstMask = DAG.getConstant(0xffff, DL, VecIntVT); | |||
| 20457 | // uint4 lo = (v & (uint4) 0xffff) | (uint4) 0x4b000000; | |||
| 20458 | SDValue LowAnd = DAG.getNode(ISD::AND, DL, VecIntVT, V, VecCstMask); | |||
| 20459 | Low = DAG.getNode(ISD::OR, DL, VecIntVT, LowAnd, VecCstLow); | |||
| 20460 | ||||
| 20461 | // uint4 hi = (v >> 16) | (uint4) 0x53000000; | |||
| 20462 | High = DAG.getNode(ISD::OR, DL, VecIntVT, HighShift, VecCstHigh); | |||
| 20463 | } | |||
| 20464 | ||||
| 20465 | // Create the vector constant for (0x1.0p39f + 0x1.0p23f). | |||
| 20466 | SDValue VecCstFSub = DAG.getConstantFP( | |||
| 20467 | APFloat(APFloat::IEEEsingle(), APInt(32, 0x53000080)), DL, VecFloatVT); | |||
| 20468 | ||||
| 20469 | // float4 fhi = (float4) hi - (0x1.0p39f + 0x1.0p23f); | |||
| 20470 | // NOTE: By using fsub of a positive constant instead of fadd of a negative | |||
| 20471 | // constant, we avoid reassociation in MachineCombiner when unsafe-fp-math is | |||
| 20472 | // enabled. See PR24512. | |||
| 20473 | SDValue HighBitcast = DAG.getBitcast(VecFloatVT, High); | |||
| 20474 | // TODO: Are there any fast-math-flags to propagate here? | |||
| 20475 | // (float4) lo; | |||
| 20476 | SDValue LowBitcast = DAG.getBitcast(VecFloatVT, Low); | |||
| 20477 | // return (float4) lo + fhi; | |||
| 20478 | if (IsStrict) { | |||
| 20479 | SDValue FHigh = DAG.getNode(ISD::STRICT_FSUB, DL, {VecFloatVT, MVT::Other}, | |||
| 20480 | {Op.getOperand(0), HighBitcast, VecCstFSub}); | |||
| 20481 | return DAG.getNode(ISD::STRICT_FADD, DL, {VecFloatVT, MVT::Other}, | |||
| 20482 | {FHigh.getValue(1), LowBitcast, FHigh}); | |||
| 20483 | } | |||
| 20484 | ||||
| 20485 | SDValue FHigh = | |||
| 20486 | DAG.getNode(ISD::FSUB, DL, VecFloatVT, HighBitcast, VecCstFSub); | |||
| 20487 | return DAG.getNode(ISD::FADD, DL, VecFloatVT, LowBitcast, FHigh); | |||
| 20488 | } | |||
| 20489 | ||||
| 20490 | static SDValue lowerUINT_TO_FP_vec(SDValue Op, SelectionDAG &DAG, | |||
| 20491 | const X86Subtarget &Subtarget) { | |||
| 20492 | unsigned OpNo = Op.getNode()->isStrictFPOpcode() ? 1 : 0; | |||
| 20493 | SDValue N0 = Op.getOperand(OpNo); | |||
| 20494 | MVT SrcVT = N0.getSimpleValueType(); | |||
| 20495 | SDLoc dl(Op); | |||
| 20496 | ||||
| 20497 | switch (SrcVT.SimpleTy) { | |||
| 20498 | default: | |||
| 20499 | llvm_unreachable("Custom UINT_TO_FP is not supported!")__builtin_unreachable(); | |||
| 20500 | case MVT::v2i32: | |||
| 20501 | return lowerUINT_TO_FP_v2i32(Op, DAG, Subtarget, dl); | |||
| 20502 | case MVT::v4i32: | |||
| 20503 | case MVT::v8i32: | |||
| 20504 | return lowerUINT_TO_FP_vXi32(Op, DAG, Subtarget); | |||
| 20505 | case MVT::v2i64: | |||
| 20506 | case MVT::v4i64: | |||
| 20507 | return lowerINT_TO_FP_vXi64(Op, DAG, Subtarget); | |||
| 20508 | } | |||
| 20509 | } | |||
| 20510 | ||||
| 20511 | SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op, | |||
| 20512 | SelectionDAG &DAG) const { | |||
| 20513 | bool IsStrict = Op->isStrictFPOpcode(); | |||
| 20514 | unsigned OpNo = IsStrict ? 1 : 0; | |||
| 20515 | SDValue Src = Op.getOperand(OpNo); | |||
| 20516 | SDLoc dl(Op); | |||
| 20517 | auto PtrVT = getPointerTy(DAG.getDataLayout()); | |||
| 20518 | MVT SrcVT = Src.getSimpleValueType(); | |||
| 20519 | MVT DstVT = Op->getSimpleValueType(0); | |||
| 20520 | SDValue Chain = IsStrict ? Op.getOperand(0) : DAG.getEntryNode(); | |||
| 20521 | ||||
| 20522 | if (DstVT == MVT::f128) | |||
| 20523 | return SDValue(); | |||
| 20524 | ||||
| 20525 | if (DstVT.isVector()) | |||
| 20526 | return lowerUINT_TO_FP_vec(Op, DAG, Subtarget); | |||
| 20527 | ||||
| 20528 | if (SDValue Extract = vectorizeExtractedCast(Op, DAG, Subtarget)) | |||
| 20529 | return Extract; | |||
| 20530 | ||||
| 20531 | if (Subtarget.hasAVX512() && isScalarFPTypeInSSEReg(DstVT) && | |||
| 20532 | (SrcVT == MVT::i32 || (SrcVT == MVT::i64 && Subtarget.is64Bit()))) { | |||
| 20533 | // Conversions from unsigned i32 to f32/f64 are legal, | |||
| 20534 | // using VCVTUSI2SS/SD. Same for i64 in 64-bit mode. | |||
| 20535 | return Op; | |||
| 20536 | } | |||
| 20537 | ||||
| 20538 | // Promote i32 to i64 and use a signed conversion on 64-bit targets. | |||
| 20539 | if (SrcVT == MVT::i32 && Subtarget.is64Bit()) { | |||
| 20540 | Src = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i64, Src); | |||
| 20541 | if (IsStrict) | |||
| 20542 | return DAG.getNode(ISD::STRICT_SINT_TO_FP, dl, {DstVT, MVT::Other}, | |||
| 20543 | {Chain, Src}); | |||
| 20544 | return DAG.getNode(ISD::SINT_TO_FP, dl, DstVT, Src); | |||
| 20545 | } | |||
| 20546 | ||||
| 20547 | if (SDValue V = LowerI64IntToFP_AVX512DQ(Op, DAG, Subtarget)) | |||
| 20548 | return V; | |||
| 20549 | ||||
| 20550 | // The transform for i64->f64 isn't correct for 0 when rounding to negative | |||
| 20551 | // infinity. It produces -0.0, so disable under strictfp. | |||
| 20552 | if (SrcVT == MVT::i64 && DstVT == MVT::f64 && X86ScalarSSEf64 && !IsStrict) | |||
| 20553 | return LowerUINT_TO_FP_i64(Op, DAG, Subtarget); | |||
| 20554 | if (SrcVT == MVT::i32 && X86ScalarSSEf64 && DstVT != MVT::f80) | |||
| 20555 | return LowerUINT_TO_FP_i32(Op, DAG, Subtarget); | |||
| 20556 | if (Subtarget.is64Bit() && SrcVT == MVT::i64 && | |||
| 20557 | (DstVT == MVT::f32 || DstVT == MVT::f64)) | |||
| 20558 | return SDValue(); | |||
| 20559 | ||||
| 20560 | // Make a 64-bit buffer, and use it to build an FILD. | |||
| 20561 | SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64, 8); | |||
| 20562 | int SSFI = cast<FrameIndexSDNode>(StackSlot)->getIndex(); | |||
| 20563 | Align SlotAlign(8); | |||
| 20564 | MachinePointerInfo MPI = | |||
| 20565 | MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI); | |||
| 20566 | if (SrcVT == MVT::i32) { | |||
| 20567 | SDValue OffsetSlot = | |||
| 20568 | DAG.getMemBasePlusOffset(StackSlot, TypeSize::Fixed(4), dl); | |||
| 20569 | SDValue Store1 = DAG.getStore(Chain, dl, Src, StackSlot, MPI, SlotAlign); | |||
| 20570 | SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, dl, MVT::i32), | |||
| 20571 | OffsetSlot, MPI.getWithOffset(4), SlotAlign); | |||
| 20572 | std::pair<SDValue, SDValue> Tmp = | |||
| 20573 | BuildFILD(DstVT, MVT::i64, dl, Store2, StackSlot, MPI, SlotAlign, DAG); | |||
| 20574 | if (IsStrict) | |||
| 20575 | return DAG.getMergeValues({Tmp.first, Tmp.second}, dl); | |||
| 20576 | ||||
| 20577 | return Tmp.first; | |||
| 20578 | } | |||
| 20579 | ||||
| 20580 | assert(SrcVT == MVT::i64 && "Unexpected type in UINT_TO_FP")((void)0); | |||
| 20581 | SDValue ValueToStore = Src; | |||
| 20582 | if (isScalarFPTypeInSSEReg(Op.getValueType()) && !Subtarget.is64Bit()) { | |||
| 20583 | // Bitcasting to f64 here allows us to do a single 64-bit store from | |||
| 20584 | // an SSE register, avoiding the store forwarding penalty that would come | |||
| 20585 | // with two 32-bit stores. | |||
| 20586 | ValueToStore = DAG.getBitcast(MVT::f64, ValueToStore); | |||
| 20587 | } | |||
| 20588 | SDValue Store = | |||
| 20589 | DAG.getStore(Chain, dl, ValueToStore, StackSlot, MPI, SlotAlign); | |||
| 20590 | // For i64 source, we need to add the appropriate power of 2 if the input | |||
| 20591 | // was negative. We must be careful to do the computation in x87 extended | |||
| 20592 | // precision, not in SSE. | |||
| 20593 | SDVTList Tys = DAG.getVTList(MVT::f80, MVT::Other); | |||
| 20594 | SDValue Ops[] = { Store, StackSlot }; | |||
| 20595 | SDValue Fild = | |||
| 20596 | DAG.getMemIntrinsicNode(X86ISD::FILD, dl, Tys, Ops, MVT::i64, MPI, | |||
| 20597 | SlotAlign, MachineMemOperand::MOLoad); | |||
| 20598 | Chain = Fild.getValue(1); | |||
| 20599 | ||||
| 20600 | ||||
| 20601 | // Check whether the sign bit is set. | |||
| 20602 | SDValue SignSet = DAG.getSetCC( | |||
| 20603 | dl, getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), MVT::i64), | |||
| 20604 | Op.getOperand(OpNo), DAG.getConstant(0, dl, MVT::i64), ISD::SETLT); | |||
| 20605 | ||||
| 20606 | // Build a 64 bit pair (FF, 0) in the constant pool, with FF in the hi bits. | |||
| 20607 | APInt FF(64, 0x5F80000000000000ULL); | |||
| 20608 | SDValue FudgePtr = DAG.getConstantPool( | |||
| 20609 | ConstantInt::get(*DAG.getContext(), FF), PtrVT); | |||
| 20610 | Align CPAlignment = cast<ConstantPoolSDNode>(FudgePtr)->getAlign(); | |||
| 20611 | ||||
| 20612 | // Get a pointer to FF if the sign bit was set, or to 0 otherwise. | |||
| 20613 | SDValue Zero = DAG.getIntPtrConstant(0, dl); | |||
| 20614 | SDValue Four = DAG.getIntPtrConstant(4, dl); | |||
| 20615 | SDValue Offset = DAG.getSelect(dl, Zero.getValueType(), SignSet, Four, Zero); | |||
| 20616 | FudgePtr = DAG.getNode(ISD::ADD, dl, PtrVT, FudgePtr, Offset); | |||
| 20617 | ||||
| 20618 | // Load the value out, extending it from f32 to f80. | |||
| 20619 | SDValue Fudge = DAG.getExtLoad( | |||
| 20620 | ISD::EXTLOAD, dl, MVT::f80, Chain, FudgePtr, | |||
| 20621 | MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), MVT::f32, | |||
| 20622 | CPAlignment); | |||
| 20623 | Chain = Fudge.getValue(1); | |||
| 20624 | // Extend everything to 80 bits to force it to be done on x87. | |||
| 20625 | // TODO: Are there any fast-math-flags to propagate here? | |||
| 20626 | if (IsStrict) { | |||
| 20627 | SDValue Add = DAG.getNode(ISD::STRICT_FADD, dl, {MVT::f80, MVT::Other}, | |||
| 20628 | {Chain, Fild, Fudge}); | |||
| 20629 | // STRICT_FP_ROUND can't handle equal types. | |||
| 20630 | if (DstVT == MVT::f80) | |||
| 20631 | return Add; | |||
| 20632 | return DAG.getNode(ISD::STRICT_FP_ROUND, dl, {DstVT, MVT::Other}, | |||
| 20633 | {Add.getValue(1), Add, DAG.getIntPtrConstant(0, dl)}); | |||
| 20634 | } | |||
| 20635 | SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::f80, Fild, Fudge); | |||
| 20636 | return DAG.getNode(ISD::FP_ROUND, dl, DstVT, Add, | |||
| 20637 | DAG.getIntPtrConstant(0, dl)); | |||
| 20638 | } | |||
| 20639 | ||||
| 20640 | // If the given FP_TO_SINT (IsSigned) or FP_TO_UINT (!IsSigned) operation | |||
| 20641 | // is legal, or has an fp128 or f16 source (which needs to be promoted to f32), | |||
| 20642 | // just return an SDValue(). | |||
| 20643 | // Otherwise it is assumed to be a conversion from one of f32, f64 or f80 | |||
| 20644 | // to i16, i32 or i64, and we lower it to a legal sequence and return the | |||
| 20645 | // result. | |||
| 20646 | SDValue | |||
| 20647 | X86TargetLowering::FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG, | |||
| 20648 | bool IsSigned, SDValue &Chain) const { | |||
| 20649 | bool IsStrict = Op->isStrictFPOpcode(); | |||
| 20650 | SDLoc DL(Op); | |||
| 20651 | ||||
| 20652 | EVT DstTy = Op.getValueType(); | |||
| 20653 | SDValue Value = Op.getOperand(IsStrict ? 1 : 0); | |||
| 20654 | EVT TheVT = Value.getValueType(); | |||
| 20655 | auto PtrVT = getPointerTy(DAG.getDataLayout()); | |||
| 20656 | ||||
| 20657 | if (TheVT != MVT::f32 && TheVT != MVT::f64 && TheVT != MVT::f80) { | |||
| 20658 | // f16 must be promoted before using the lowering in this routine. | |||
| 20659 | // fp128 does not use this lowering. | |||
| 20660 | return SDValue(); | |||
| 20661 | } | |||
| 20662 | ||||
| 20663 | // If using FIST to compute an unsigned i64, we'll need some fixup | |||
| 20664 | // to handle values above the maximum signed i64. A FIST is always | |||
| 20665 | // used for the 32-bit subtarget, but also for f80 on a 64-bit target. | |||
| 20666 | bool UnsignedFixup = !IsSigned && DstTy == MVT::i64; | |||
| 20667 | ||||
| 20668 | // FIXME: This does not generate an invalid exception if the input does not | |||
| 20669 | // fit in i32. PR44019 | |||
| 20670 | if (!IsSigned && DstTy != MVT::i64) { | |||
| 20671 | // Replace the fp-to-uint32 operation with an fp-to-sint64 FIST. | |||
| 20672 | // The low 32 bits of the fist result will have the correct uint32 result. | |||
| 20673 | assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT")((void)0); | |||
| 20674 | DstTy = MVT::i64; | |||
| 20675 | } | |||
| 20676 | ||||
| 20677 | assert(DstTy.getSimpleVT() <= MVT::i64 &&((void)0) | |||
| 20678 | DstTy.getSimpleVT() >= MVT::i16 &&((void)0) | |||
| 20679 | "Unknown FP_TO_INT to lower!")((void)0); | |||
| 20680 | ||||
| 20681 | // We lower FP->int64 into FISTP64 followed by a load from a temporary | |||
| 20682 | // stack slot. | |||
| 20683 | MachineFunction &MF = DAG.getMachineFunction(); | |||
| 20684 | unsigned MemSize = DstTy.getStoreSize(); | |||
| 20685 | int SSFI = | |||
| 20686 | MF.getFrameInfo().CreateStackObject(MemSize, Align(MemSize), false); | |||
| 20687 | SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT); | |||
| 20688 | ||||
| 20689 | Chain = IsStrict ? Op.getOperand(0) : DAG.getEntryNode(); | |||
| 20690 | ||||
| 20691 | SDValue Adjust; // 0x0 or 0x80000000, for result sign bit adjustment. | |||
| 20692 | ||||
| 20693 | if (UnsignedFixup) { | |||
| 20694 | // | |||
| 20695 | // Conversion to unsigned i64 is implemented with a select, | |||
| 20696 | // depending on whether the source value fits in the range | |||
| 20697 | // of a signed i64. Let Thresh be the FP equivalent of | |||
| 20698 | // 0x8000000000000000ULL. | |||
| 20699 | // | |||
| 20700 | // Adjust = (Value >= Thresh) ? 0x80000000 : 0; | |||
| 20701 | // FltOfs = (Value >= Thresh) ? 0x80000000 : 0; | |||
| 20702 | // FistSrc = (Value - FltOfs); | |||
| 20703 | // Fist-to-mem64 FistSrc | |||
| 20704 | // Add 0 or 0x800...0ULL to the 64-bit result, which is equivalent | |||
| 20705 | // to XOR'ing the high 32 bits with Adjust. | |||
| 20706 | // | |||
| 20707 | // Being a power of 2, Thresh is exactly representable in all FP formats. | |||
| 20708 | // For X87 we'd like to use the smallest FP type for this constant, but | |||
| 20709 | // for DAG type consistency we have to match the FP operand type. | |||
| 20710 | ||||
| 20711 | APFloat Thresh(APFloat::IEEEsingle(), APInt(32, 0x5f000000)); | |||
| 20712 | LLVM_ATTRIBUTE_UNUSED__attribute__((__unused__)) APFloat::opStatus Status = APFloat::opOK; | |||
| 20713 | bool LosesInfo = false; | |||
| 20714 | if (TheVT == MVT::f64) | |||
| 20715 | // The rounding mode is irrelevant as the conversion should be exact. | |||
| 20716 | Status = Thresh.convert(APFloat::IEEEdouble(), APFloat::rmNearestTiesToEven, | |||
| 20717 | &LosesInfo); | |||
| 20718 | else if (TheVT == MVT::f80) | |||
| 20719 | Status = Thresh.convert(APFloat::x87DoubleExtended(), | |||
| 20720 | APFloat::rmNearestTiesToEven, &LosesInfo); | |||
| 20721 | ||||
| 20722 | assert(Status == APFloat::opOK && !LosesInfo &&((void)0) | |||
| 20723 | "FP conversion should have been exact")((void)0); | |||
| 20724 | ||||
| 20725 | SDValue ThreshVal = DAG.getConstantFP(Thresh, DL, TheVT); | |||
| 20726 | ||||
| 20727 | EVT ResVT = getSetCCResultType(DAG.getDataLayout(), | |||
| 20728 | *DAG.getContext(), TheVT); | |||
| 20729 | SDValue Cmp; | |||
| 20730 | if (IsStrict) { | |||
| 20731 | Cmp = DAG.getSetCC(DL, ResVT, Value, ThreshVal, ISD::SETGE, Chain, | |||
| 20732 | /*IsSignaling*/ true); | |||
| 20733 | Chain = Cmp.getValue(1); | |||
| 20734 | } else { | |||
| 20735 | Cmp = DAG.getSetCC(DL, ResVT, Value, ThreshVal, ISD::SETGE); | |||
| 20736 | } | |||
| 20737 | ||||
| 20738 | // Our preferred lowering of | |||
| 20739 | // | |||
| 20740 | // (Value >= Thresh) ? 0x8000000000000000ULL : 0 | |||
| 20741 | // | |||
| 20742 | // is | |||
| 20743 | // | |||
| 20744 | // (Value >= Thresh) << 63 | |||
| 20745 | // | |||
| 20746 | // but since we can get here after LegalOperations, DAGCombine might do the | |||
| 20747 | // wrong thing if we create a select. So, directly create the preferred | |||
| 20748 | // version. | |||
| 20749 | SDValue Zext = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i64, Cmp); | |||
| 20750 | SDValue Const63 = DAG.getConstant(63, DL, MVT::i8); | |||
| 20751 | Adjust = DAG.getNode(ISD::SHL, DL, MVT::i64, Zext, Const63); | |||
| 20752 | ||||
| 20753 | SDValue FltOfs = DAG.getSelect(DL, TheVT, Cmp, ThreshVal, | |||
| 20754 | DAG.getConstantFP(0.0, DL, TheVT)); | |||
| 20755 | ||||
| 20756 | if (IsStrict) { | |||
| 20757 | Value = DAG.getNode(ISD::STRICT_FSUB, DL, { TheVT, MVT::Other}, | |||
| 20758 | { Chain, Value, FltOfs }); | |||
| 20759 | Chain = Value.getValue(1); | |||
| 20760 | } else | |||
| 20761 | Value = DAG.getNode(ISD::FSUB, DL, TheVT, Value, FltOfs); | |||
| 20762 | } | |||
| 20763 | ||||
| 20764 | MachinePointerInfo MPI = MachinePointerInfo::getFixedStack(MF, SSFI); | |||
| 20765 | ||||
| 20766 | // FIXME This causes a redundant load/store if the SSE-class value is already | |||
| 20767 | // in memory, such as if it is on the callstack. | |||
| 20768 | if (isScalarFPTypeInSSEReg(TheVT)) { | |||
| 20769 | assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!")((void)0); | |||
| 20770 | Chain = DAG.getStore(Chain, DL, Value, StackSlot, MPI); | |||
| 20771 | SDVTList Tys = DAG.getVTList(MVT::f80, MVT::Other); | |||
| 20772 | SDValue Ops[] = { Chain, StackSlot }; | |||
| 20773 | ||||
| 20774 | unsigned FLDSize = TheVT.getStoreSize(); | |||
| 20775 | assert(FLDSize <= MemSize && "Stack slot not big enough")((void)0); | |||
| 20776 | MachineMemOperand *MMO = MF.getMachineMemOperand( | |||
| 20777 | MPI, MachineMemOperand::MOLoad, FLDSize, Align(FLDSize)); | |||
| 20778 | Value = DAG.getMemIntrinsicNode(X86ISD::FLD, DL, Tys, Ops, TheVT, MMO); | |||
| 20779 | Chain = Value.getValue(1); | |||
| 20780 | } | |||
| 20781 | ||||
| 20782 | // Build the FP_TO_INT*_IN_MEM | |||
| 20783 | MachineMemOperand *MMO = MF.getMachineMemOperand( | |||
| 20784 | MPI, MachineMemOperand::MOStore, MemSize, Align(MemSize)); | |||
| 20785 | SDValue Ops[] = { Chain, Value, StackSlot }; | |||
| 20786 | SDValue FIST = DAG.getMemIntrinsicNode(X86ISD::FP_TO_INT_IN_MEM, DL, | |||
| 20787 | DAG.getVTList(MVT::Other), | |||
| 20788 | Ops, DstTy, MMO); | |||
| 20789 | ||||
| 20790 | SDValue Res = DAG.getLoad(Op.getValueType(), SDLoc(Op), FIST, StackSlot, MPI); | |||
| 20791 | Chain = Res.getValue(1); | |||
| 20792 | ||||
| 20793 | // If we need an unsigned fixup, XOR the result with adjust. | |||
| 20794 | if (UnsignedFixup) | |||
| 20795 | Res = DAG.getNode(ISD::XOR, DL, MVT::i64, Res, Adjust); | |||
| 20796 | ||||
| 20797 | return Res; | |||
| 20798 | } | |||
| 20799 | ||||
| 20800 | static SDValue LowerAVXExtend(SDValue Op, SelectionDAG &DAG, | |||
| 20801 | const X86Subtarget &Subtarget) { | |||
| 20802 | MVT VT = Op.getSimpleValueType(); | |||
| 20803 | SDValue In = Op.getOperand(0); | |||
| 20804 | MVT InVT = In.getSimpleValueType(); | |||
| 20805 | SDLoc dl(Op); | |||
| 20806 | unsigned Opc = Op.getOpcode(); | |||
| 20807 | ||||
| 20808 | assert(VT.isVector() && InVT.isVector() && "Expected vector type")((void)0); | |||
| 20809 | assert((Opc == ISD::ANY_EXTEND || Opc == ISD::ZERO_EXTEND) &&((void)0) | |||
| 20810 | "Unexpected extension opcode")((void)0); | |||
| 20811 | assert(VT.getVectorNumElements() == InVT.getVectorNumElements() &&((void)0) | |||
| 20812 | "Expected same number of elements")((void)0); | |||
| 20813 | assert((VT.getVectorElementType() == MVT::i16 ||((void)0) | |||
| 20814 | VT.getVectorElementType() == MVT::i32 ||((void)0) | |||
| 20815 | VT.getVectorElementType() == MVT::i64) &&((void)0) | |||
| 20816 | "Unexpected element type")((void)0); | |||
| 20817 | assert((InVT.getVectorElementType() == MVT::i8 ||((void)0) | |||
| 20818 | InVT.getVectorElementType() == MVT::i16 ||((void)0) | |||
| 20819 | InVT.getVectorElementType() == MVT::i32) &&((void)0) | |||
| 20820 | "Unexpected element type")((void)0); | |||
| 20821 | ||||
| 20822 | unsigned ExtendInVecOpc = getOpcode_EXTEND_VECTOR_INREG(Opc); | |||
| 20823 | ||||
| 20824 | if (VT == MVT::v32i16 && !Subtarget.hasBWI()) { | |||
| 20825 | assert(InVT == MVT::v32i8 && "Unexpected VT!")((void)0); | |||
| 20826 | return splitVectorIntUnary(Op, DAG); | |||
| 20827 | } | |||
| 20828 | ||||
| 20829 | if (Subtarget.hasInt256()) | |||
| 20830 | return Op; | |||
| 20831 | ||||
| 20832 | // Optimize vectors in AVX mode: | |||
| 20833 | // | |||
| 20834 | // v8i16 -> v8i32 | |||
| 20835 | // Use vpmovzwd for 4 lower elements v8i16 -> v4i32. | |||
| 20836 | // Use vpunpckhwd for 4 upper elements v8i16 -> v4i32. | |||
| 20837 | // Concat upper and lower parts. | |||
| 20838 | // | |||
| 20839 | // v4i32 -> v4i64 | |||
| 20840 | // Use vpmovzdq for 4 lower elements v4i32 -> v2i64. | |||
| 20841 | // Use vpunpckhdq for 4 upper elements v4i32 -> v2i64. | |||
| 20842 | // Concat upper and lower parts. | |||
| 20843 | // | |||
| 20844 | MVT HalfVT = VT.getHalfNumVectorElementsVT(); | |||
| 20845 | SDValue OpLo = DAG.getNode(ExtendInVecOpc, dl, HalfVT, In); | |||
| 20846 | ||||
| 20847 | // Short-circuit if we can determine that each 128-bit half is the same value. | |||
| 20848 | // Otherwise, this is difficult to match and optimize. | |||
| 20849 | if (auto *Shuf = dyn_cast<ShuffleVectorSDNode>(In)) | |||
| 20850 | if (hasIdenticalHalvesShuffleMask(Shuf->getMask())) | |||
| 20851 | return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpLo); | |||
| 20852 | ||||
| 20853 | SDValue ZeroVec = DAG.getConstant(0, dl, InVT); | |||
| 20854 | SDValue Undef = DAG.getUNDEF(InVT); | |||
| 20855 | bool NeedZero = Opc == ISD::ZERO_EXTEND; | |||
| 20856 | SDValue OpHi = getUnpackh(DAG, dl, InVT, In, NeedZero ? ZeroVec : Undef); | |||
| 20857 | OpHi = DAG.getBitcast(HalfVT, OpHi); | |||
| 20858 | ||||
| 20859 | return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi); | |||
| 20860 | } | |||
| 20861 | ||||
| 20862 | // Helper to split and extend a v16i1 mask to v16i8 or v16i16. | |||
| 20863 | static SDValue SplitAndExtendv16i1(unsigned ExtOpc, MVT VT, SDValue In, | |||
| 20864 | const SDLoc &dl, SelectionDAG &DAG) { | |||
| 20865 | assert((VT == MVT::v16i8 || VT == MVT::v16i16) && "Unexpected VT.")((void)0); | |||
| 20866 | SDValue Lo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v8i1, In, | |||
| 20867 | DAG.getIntPtrConstant(0, dl)); | |||
| 20868 | SDValue Hi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v8i1, In, | |||
| 20869 | DAG.getIntPtrConstant(8, dl)); | |||
| 20870 | Lo = DAG.getNode(ExtOpc, dl, MVT::v8i16, Lo); | |||
| 20871 | Hi = DAG.getNode(ExtOpc, dl, MVT::v8i16, Hi); | |||
| 20872 | SDValue Res = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v16i16, Lo, Hi); | |||
| 20873 | return DAG.getNode(ISD::TRUNCATE, dl, VT, Res); | |||
| 20874 | } | |||
| 20875 | ||||
| 20876 | static SDValue LowerZERO_EXTEND_Mask(SDValue Op, | |||
| 20877 | const X86Subtarget &Subtarget, | |||
| 20878 | SelectionDAG &DAG) { | |||
| 20879 | MVT VT = Op->getSimpleValueType(0); | |||
| 20880 | SDValue In = Op->getOperand(0); | |||
| 20881 | MVT InVT = In.getSimpleValueType(); | |||
| 20882 | assert(InVT.getVectorElementType() == MVT::i1 && "Unexpected input type!")((void)0); | |||
| 20883 | SDLoc DL(Op); | |||
| 20884 | unsigned NumElts = VT.getVectorNumElements(); | |||
| 20885 | ||||
| 20886 | // For all vectors, but vXi8 we can just emit a sign_extend and a shift. This | |||
| 20887 | // avoids a constant pool load. | |||
| 20888 | if (VT.getVectorElementType() != MVT::i8) { | |||
| 20889 | SDValue Extend = DAG.getNode(ISD::SIGN_EXTEND, DL, VT, In); | |||
| 20890 | return DAG.getNode(ISD::SRL, DL, VT, Extend, | |||
| 20891 | DAG.getConstant(VT.getScalarSizeInBits() - 1, DL, VT)); | |||
| 20892 | } | |||
| 20893 | ||||
| 20894 | // Extend VT if BWI is not supported. | |||
| 20895 | MVT ExtVT = VT; | |||
| 20896 | if (!Subtarget.hasBWI()) { | |||
| 20897 | // If v16i32 is to be avoided, we'll need to split and concatenate. | |||
| 20898 | if (NumElts == 16 && !Subtarget.canExtendTo512DQ()) | |||
| 20899 | return SplitAndExtendv16i1(ISD::ZERO_EXTEND, VT, In, DL, DAG); | |||
| 20900 | ||||
| 20901 | ExtVT = MVT::getVectorVT(MVT::i32, NumElts); | |||
| 20902 | } | |||
| 20903 | ||||
| 20904 | // Widen to 512-bits if VLX is not supported. | |||
| 20905 | MVT WideVT = ExtVT; | |||
| 20906 | if (!ExtVT.is512BitVector() && !Subtarget.hasVLX()) { | |||
| 20907 | NumElts *= 512 / ExtVT.getSizeInBits(); | |||
| 20908 | InVT = MVT::getVectorVT(MVT::i1, NumElts); | |||
| 20909 | In = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, InVT, DAG.getUNDEF(InVT), | |||
| 20910 | In, DAG.getIntPtrConstant(0, DL)); | |||
| 20911 | WideVT = MVT::getVectorVT(ExtVT.getVectorElementType(), | |||
| 20912 | NumElts); | |||
| 20913 | } | |||
| 20914 | ||||
| 20915 | SDValue One = DAG.getConstant(1, DL, WideVT); | |||
| 20916 | SDValue Zero = DAG.getConstant(0, DL, WideVT); | |||
| 20917 | ||||
| 20918 | SDValue SelectedVal = DAG.getSelect(DL, WideVT, In, One, Zero); | |||
| 20919 | ||||
| 20920 | // Truncate if we had to extend above. | |||
| 20921 | if (VT != ExtVT) { | |||
| 20922 | WideVT = MVT::getVectorVT(MVT::i8, NumElts); | |||
| 20923 | SelectedVal = DAG.getNode(ISD::TRUNCATE, DL, WideVT, SelectedVal); | |||
| 20924 | } | |||
| 20925 | ||||
| 20926 | // Extract back to 128/256-bit if we widened. | |||
| 20927 | if (WideVT != VT) | |||
| 20928 | SelectedVal = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, SelectedVal, | |||
| 20929 | DAG.getIntPtrConstant(0, DL)); | |||
| 20930 | ||||
| 20931 | return SelectedVal; | |||
| 20932 | } | |||
| 20933 | ||||
| 20934 | static SDValue LowerZERO_EXTEND(SDValue Op, const X86Subtarget &Subtarget, | |||
| 20935 | SelectionDAG &DAG) { | |||
| 20936 | SDValue In = Op.getOperand(0); | |||
| 20937 | MVT SVT = In.getSimpleValueType(); | |||
| 20938 | ||||
| 20939 | if (SVT.getVectorElementType() == MVT::i1) | |||
| 20940 | return LowerZERO_EXTEND_Mask(Op, Subtarget, DAG); | |||
| 20941 | ||||
| 20942 | assert(Subtarget.hasAVX() && "Expected AVX support")((void)0); | |||
| 20943 | return LowerAVXExtend(Op, DAG, Subtarget); | |||
| 20944 | } | |||
| 20945 | ||||
| 20946 | /// Helper to recursively truncate vector elements in half with PACKSS/PACKUS. | |||
| 20947 | /// It makes use of the fact that vectors with enough leading sign/zero bits | |||
| 20948 | /// prevent the PACKSS/PACKUS from saturating the results. | |||
| 20949 | /// AVX2 (Int256) sub-targets require extra shuffling as the PACK*S operates | |||
| 20950 | /// within each 128-bit lane. | |||
| 20951 | static SDValue truncateVectorWithPACK(unsigned Opcode, EVT DstVT, SDValue In, | |||
| 20952 | const SDLoc &DL, SelectionDAG &DAG, | |||
| 20953 | const X86Subtarget &Subtarget) { | |||
| 20954 | assert((Opcode == X86ISD::PACKSS || Opcode == X86ISD::PACKUS) &&((void)0) | |||
| 20955 | "Unexpected PACK opcode")((void)0); | |||
| 20956 | assert(DstVT.isVector() && "VT not a vector?")((void)0); | |||
| 20957 | ||||
| 20958 | // Requires SSE2 for PACKSS (SSE41 PACKUSDW is handled below). | |||
| 20959 | if (!Subtarget.hasSSE2()) | |||
| 20960 | return SDValue(); | |||
| 20961 | ||||
| 20962 | EVT SrcVT = In.getValueType(); | |||
| 20963 | ||||
| 20964 | // No truncation required, we might get here due to recursive calls. | |||
| 20965 | if (SrcVT == DstVT) | |||
| 20966 | return In; | |||
| 20967 | ||||
| 20968 | // We only support vector truncation to 64bits or greater from a | |||
| 20969 | // 128bits or greater source. | |||
| 20970 | unsigned DstSizeInBits = DstVT.getSizeInBits(); | |||
| 20971 | unsigned SrcSizeInBits = SrcVT.getSizeInBits(); | |||
| 20972 | if ((DstSizeInBits % 64) != 0 || (SrcSizeInBits % 128) != 0) | |||
| 20973 | return SDValue(); | |||
| 20974 | ||||
| 20975 | unsigned NumElems = SrcVT.getVectorNumElements(); | |||
| 20976 | if (!isPowerOf2_32(NumElems)) | |||
| 20977 | return SDValue(); | |||
| 20978 | ||||
| 20979 | LLVMContext &Ctx = *DAG.getContext(); | |||
| 20980 | assert(DstVT.getVectorNumElements() == NumElems && "Illegal truncation")((void)0); | |||
| 20981 | assert(SrcSizeInBits > DstSizeInBits && "Illegal truncation")((void)0); | |||
| 20982 | ||||
| 20983 | EVT PackedSVT = EVT::getIntegerVT(Ctx, SrcVT.getScalarSizeInBits() / 2); | |||
| 20984 | ||||
| 20985 | // Pack to the largest type possible: | |||
| 20986 | // vXi64/vXi32 -> PACK*SDW and vXi16 -> PACK*SWB. | |||
| 20987 | EVT InVT = MVT::i16, OutVT = MVT::i8; | |||
| 20988 | if (SrcVT.getScalarSizeInBits() > 16 && | |||
| 20989 | (Opcode == X86ISD::PACKSS || Subtarget.hasSSE41())) { | |||
| 20990 | InVT = MVT::i32; | |||
| 20991 | OutVT = MVT::i16; | |||
| 20992 | } | |||
| 20993 | ||||
| 20994 | // 128bit -> 64bit truncate - PACK 128-bit src in the lower subvector. | |||
| 20995 | if (SrcVT.is128BitVector()) { | |||
| 20996 | InVT = EVT::getVectorVT(Ctx, InVT, 128 / InVT.getSizeInBits()); | |||
| 20997 | OutVT = EVT::getVectorVT(Ctx, OutVT, 128 / OutVT.getSizeInBits()); | |||
| 20998 | In = DAG.getBitcast(InVT, In); | |||
| 20999 | SDValue Res = DAG.getNode(Opcode, DL, OutVT, In, DAG.getUNDEF(InVT)); | |||
| 21000 | Res = extractSubVector(Res, 0, DAG, DL, 64); | |||
| 21001 | return DAG.getBitcast(DstVT, Res); | |||
| 21002 | } | |||
| 21003 | ||||
| 21004 | // Split lower/upper subvectors. | |||
| 21005 | SDValue Lo, Hi; | |||
| 21006 | std::tie(Lo, Hi) = splitVector(In, DAG, DL); | |||
| 21007 | ||||
| 21008 | unsigned SubSizeInBits = SrcSizeInBits / 2; | |||
| 21009 | InVT = EVT::getVectorVT(Ctx, InVT, SubSizeInBits / InVT.getSizeInBits()); | |||
| 21010 | OutVT = EVT::getVectorVT(Ctx, OutVT, SubSizeInBits / OutVT.getSizeInBits()); | |||
| 21011 | ||||
| 21012 | // 256bit -> 128bit truncate - PACK lower/upper 128-bit subvectors. | |||
| 21013 | if (SrcVT.is256BitVector() && DstVT.is128BitVector()) { | |||
| 21014 | Lo = DAG.getBitcast(InVT, Lo); | |||
| 21015 | Hi = DAG.getBitcast(InVT, Hi); | |||
| 21016 | SDValue Res = DAG.getNode(Opcode, DL, OutVT, Lo, Hi); | |||
| 21017 | return DAG.getBitcast(DstVT, Res); | |||
| 21018 | } | |||
| 21019 | ||||
| 21020 | // AVX2: 512bit -> 256bit truncate - PACK lower/upper 256-bit subvectors. | |||
| 21021 | // AVX2: 512bit -> 128bit truncate - PACK(PACK, PACK). | |||
| 21022 | if (SrcVT.is512BitVector() && Subtarget.hasInt256()) { | |||
| 21023 | Lo = DAG.getBitcast(InVT, Lo); | |||
| 21024 | Hi = DAG.getBitcast(InVT, Hi); | |||
| 21025 | SDValue Res = DAG.getNode(Opcode, DL, OutVT, Lo, Hi); | |||
| 21026 | ||||
| 21027 | // 256-bit PACK(ARG0, ARG1) leaves us with ((LO0,LO1),(HI0,HI1)), | |||
| 21028 | // so we need to shuffle to get ((LO0,HI0),(LO1,HI1)). | |||
| 21029 | // Scale shuffle mask to avoid bitcasts and help ComputeNumSignBits. | |||
| 21030 | SmallVector<int, 64> Mask; | |||
| 21031 | int Scale = 64 / OutVT.getScalarSizeInBits(); | |||
| 21032 | narrowShuffleMaskElts(Scale, { 0, 2, 1, 3 }, Mask); | |||
| 21033 | Res = DAG.getVectorShuffle(OutVT, DL, Res, Res, Mask); | |||
| 21034 | ||||
| 21035 | if (DstVT.is256BitVector()) | |||
| 21036 | return DAG.getBitcast(DstVT, Res); | |||
| 21037 | ||||
| 21038 | // If 512bit -> 128bit truncate another stage. | |||
| 21039 | EVT PackedVT = EVT::getVectorVT(Ctx, PackedSVT, NumElems); | |||
| 21040 | Res = DAG.getBitcast(PackedVT, Res); | |||
| 21041 | return truncateVectorWithPACK(Opcode, DstVT, Res, DL, DAG, Subtarget); | |||
| 21042 | } | |||
| 21043 | ||||
| 21044 | // Recursively pack lower/upper subvectors, concat result and pack again. | |||
| 21045 | assert(SrcSizeInBits >= 256 && "Expected 256-bit vector or greater")((void)0); | |||
| 21046 | EVT PackedVT = EVT::getVectorVT(Ctx, PackedSVT, NumElems / 2); | |||
| 21047 | Lo = truncateVectorWithPACK(Opcode, PackedVT, Lo, DL, DAG, Subtarget); | |||
| 21048 | Hi = truncateVectorWithPACK(Opcode, PackedVT, Hi, DL, DAG, Subtarget); | |||
| 21049 | ||||
| 21050 | PackedVT = EVT::getVectorVT(Ctx, PackedSVT, NumElems); | |||
| 21051 | SDValue Res = DAG.getNode(ISD::CONCAT_VECTORS, DL, PackedVT, Lo, Hi); | |||
| 21052 | return truncateVectorWithPACK(Opcode, DstVT, Res, DL, DAG, Subtarget); | |||
| 21053 | } | |||
| 21054 | ||||
| 21055 | static SDValue LowerTruncateVecI1(SDValue Op, SelectionDAG &DAG, | |||
| 21056 | const X86Subtarget &Subtarget) { | |||
| 21057 | ||||
| 21058 | SDLoc DL(Op); | |||
| 21059 | MVT VT = Op.getSimpleValueType(); | |||
| 21060 | SDValue In = Op.getOperand(0); | |||
| 21061 | MVT InVT = In.getSimpleValueType(); | |||
| 21062 | ||||
| 21063 | assert(VT.getVectorElementType() == MVT::i1 && "Unexpected vector type.")((void)0); | |||
| 21064 | ||||
| 21065 | // Shift LSB to MSB and use VPMOVB/W2M or TESTD/Q. | |||
| 21066 | unsigned ShiftInx = InVT.getScalarSizeInBits() - 1; | |||
| 21067 | if (InVT.getScalarSizeInBits() <= 16) { | |||
| 21068 | if (Subtarget.hasBWI()) { | |||
| 21069 | // legal, will go to VPMOVB2M, VPMOVW2M | |||
| 21070 | if (DAG.ComputeNumSignBits(In) < InVT.getScalarSizeInBits()) { | |||
| 21071 | // We need to shift to get the lsb into sign position. | |||
| 21072 | // Shift packed bytes not supported natively, bitcast to word | |||
| 21073 | MVT ExtVT = MVT::getVectorVT(MVT::i16, InVT.getSizeInBits()/16); | |||
| 21074 | In = DAG.getNode(ISD::SHL, DL, ExtVT, | |||
| 21075 | DAG.getBitcast(ExtVT, In), | |||
| 21076 | DAG.getConstant(ShiftInx, DL, ExtVT)); | |||
| 21077 | In = DAG.getBitcast(InVT, In); | |||
| 21078 | } | |||
| 21079 | return DAG.getSetCC(DL, VT, DAG.getConstant(0, DL, InVT), | |||
| 21080 | In, ISD::SETGT); | |||
| 21081 | } | |||
| 21082 | // Use TESTD/Q, extended vector to packed dword/qword. | |||
| 21083 | assert((InVT.is256BitVector() || InVT.is128BitVector()) &&((void)0) | |||
| 21084 | "Unexpected vector type.")((void)0); | |||
| 21085 | unsigned NumElts = InVT.getVectorNumElements(); | |||
| 21086 | assert((NumElts == 8 || NumElts == 16) && "Unexpected number of elements")((void)0); | |||
| 21087 | // We need to change to a wider element type that we have support for. | |||
| 21088 | // For 8 element vectors this is easy, we either extend to v8i32 or v8i64. | |||
| 21089 | // For 16 element vectors we extend to v16i32 unless we are explicitly | |||
| 21090 | // trying to avoid 512-bit vectors. If we are avoiding 512-bit vectors | |||
| 21091 | // we need to split into two 8 element vectors which we can extend to v8i32, | |||
| 21092 | // truncate and concat the results. There's an additional complication if | |||
| 21093 | // the original type is v16i8. In that case we can't split the v16i8 | |||
| 21094 | // directly, so we need to shuffle high elements to low and use | |||
| 21095 | // sign_extend_vector_inreg. | |||
| 21096 | if (NumElts == 16 && !Subtarget.canExtendTo512DQ()) { | |||
| 21097 | SDValue Lo, Hi; | |||
| 21098 | if (InVT == MVT::v16i8) { | |||
| 21099 | Lo = DAG.getNode(ISD::SIGN_EXTEND_VECTOR_INREG, DL, MVT::v8i32, In); | |||
| 21100 | Hi = DAG.getVectorShuffle( | |||
| 21101 | InVT, DL, In, In, | |||
| 21102 | {8, 9, 10, 11, 12, 13, 14, 15, -1, -1, -1, -1, -1, -1, -1, -1}); | |||
| 21103 | Hi = DAG.getNode(ISD::SIGN_EXTEND_VECTOR_INREG, DL, MVT::v8i32, Hi); | |||
| 21104 | } else { | |||
| 21105 | assert(InVT == MVT::v16i16 && "Unexpected VT!")((void)0); | |||
| 21106 | Lo = extract128BitVector(In, 0, DAG, DL); | |||
| 21107 | Hi = extract128BitVector(In, 8, DAG, DL); | |||
| 21108 | } | |||
| 21109 | // We're split now, just emit two truncates and a concat. The two | |||
| 21110 | // truncates will trigger legalization to come back to this function. | |||
| 21111 | Lo = DAG.getNode(ISD::TRUNCATE, DL, MVT::v8i1, Lo); | |||
| 21112 | Hi = DAG.getNode(ISD::TRUNCATE, DL, MVT::v8i1, Hi); | |||
| 21113 | return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Lo, Hi); | |||
| 21114 | } | |||
| 21115 | // We either have 8 elements or we're allowed to use 512-bit vectors. | |||
| 21116 | // If we have VLX, we want to use the narrowest vector that can get the | |||
| 21117 | // job done so we use vXi32. | |||
| 21118 | MVT EltVT = Subtarget.hasVLX() ? MVT::i32 : MVT::getIntegerVT(512/NumElts); | |||
| 21119 | MVT ExtVT = MVT::getVectorVT(EltVT, NumElts); | |||
| 21120 | In = DAG.getNode(ISD::SIGN_EXTEND, DL, ExtVT, In); | |||
| 21121 | InVT = ExtVT; | |||
| 21122 | ShiftInx = InVT.getScalarSizeInBits() - 1; | |||
| 21123 | } | |||
| 21124 | ||||
| 21125 | if (DAG.ComputeNumSignBits(In) < InVT.getScalarSizeInBits()) { | |||
| 21126 | // We need to shift to get the lsb into sign position. | |||
| 21127 | In = DAG.getNode(ISD::SHL, DL, InVT, In, | |||
| 21128 | DAG.getConstant(ShiftInx, DL, InVT)); | |||
| 21129 | } | |||
| 21130 | // If we have DQI, emit a pattern that will be iseled as vpmovq2m/vpmovd2m. | |||
| 21131 | if (Subtarget.hasDQI()) | |||
| 21132 | return DAG.getSetCC(DL, VT, DAG.getConstant(0, DL, InVT), In, ISD::SETGT); | |||
| 21133 | return DAG.getSetCC(DL, VT, In, DAG.getConstant(0, DL, InVT), ISD::SETNE); | |||
| 21134 | } | |||
| 21135 | ||||
| 21136 | SDValue X86TargetLowering::LowerTRUNCATE(SDValue Op, SelectionDAG &DAG) const { | |||
| 21137 | SDLoc DL(Op); | |||
| 21138 | MVT VT = Op.getSimpleValueType(); | |||
| 21139 | SDValue In = Op.getOperand(0); | |||
| 21140 | MVT InVT = In.getSimpleValueType(); | |||
| 21141 | unsigned InNumEltBits = InVT.getScalarSizeInBits(); | |||
| 21142 | ||||
| 21143 | assert(VT.getVectorNumElements() == InVT.getVectorNumElements() &&((void)0) | |||
| 21144 | "Invalid TRUNCATE operation")((void)0); | |||
| 21145 | ||||
| 21146 | // If we're called by the type legalizer, handle a few cases. | |||
| 21147 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
| 21148 | if (!TLI.isTypeLegal(InVT)) { | |||
| 21149 | if ((InVT == MVT::v8i64 || InVT == MVT::v16i32 || InVT == MVT::v16i64) && | |||
| 21150 | VT.is128BitVector()) { | |||
| 21151 | assert((InVT == MVT::v16i64 || Subtarget.hasVLX()) &&((void)0) | |||
| 21152 | "Unexpected subtarget!")((void)0); | |||
| 21153 | // The default behavior is to truncate one step, concatenate, and then | |||
| 21154 | // truncate the remainder. We'd rather produce two 64-bit results and | |||
| 21155 | // concatenate those. | |||
| 21156 | SDValue Lo, Hi; | |||
| 21157 | std::tie(Lo, Hi) = DAG.SplitVector(In, DL); | |||
| 21158 | ||||
| 21159 | EVT LoVT, HiVT; | |||
| 21160 | std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(VT); | |||
| 21161 | ||||
| 21162 | Lo = DAG.getNode(ISD::TRUNCATE, DL, LoVT, Lo); | |||
| 21163 | Hi = DAG.getNode(ISD::TRUNCATE, DL, HiVT, Hi); | |||
| 21164 | return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Lo, Hi); | |||
| 21165 | } | |||
| 21166 | ||||
| 21167 | // Otherwise let default legalization handle it. | |||
| 21168 | return SDValue(); | |||
| 21169 | } | |||
| 21170 | ||||
| 21171 | if (VT.getVectorElementType() == MVT::i1) | |||
| 21172 | return LowerTruncateVecI1(Op, DAG, Subtarget); | |||
| 21173 | ||||
| 21174 | // vpmovqb/w/d, vpmovdb/w, vpmovwb | |||
| 21175 | if (Subtarget.hasAVX512()) { | |||
| 21176 | if (InVT == MVT::v32i16 && !Subtarget.hasBWI()) { | |||
| 21177 | assert(VT == MVT::v32i8 && "Unexpected VT!")((void)0); | |||
| 21178 | return splitVectorIntUnary(Op, DAG); | |||
| 21179 | } | |||
| 21180 | ||||
| 21181 | // word to byte only under BWI. Otherwise we have to promoted to v16i32 | |||
| 21182 | // and then truncate that. But we should only do that if we haven't been | |||
| 21183 | // asked to avoid 512-bit vectors. The actual promotion to v16i32 will be | |||
| 21184 | // handled by isel patterns. | |||
| 21185 | if (InVT != MVT::v16i16 || Subtarget.hasBWI() || | |||
| 21186 | Subtarget.canExtendTo512DQ()) | |||
| 21187 | return Op; | |||
| 21188 | } | |||
| 21189 | ||||
| 21190 | unsigned NumPackedSignBits = std::min<unsigned>(VT.getScalarSizeInBits(), 16); | |||
| 21191 | unsigned NumPackedZeroBits = Subtarget.hasSSE41() ? NumPackedSignBits : 8; | |||
| 21192 | ||||
| 21193 | // Truncate with PACKUS if we are truncating a vector with leading zero bits | |||
| 21194 | // that extend all the way to the packed/truncated value. | |||
| 21195 | // Pre-SSE41 we can only use PACKUSWB. | |||
| 21196 | KnownBits Known = DAG.computeKnownBits(In); | |||
| 21197 | if ((InNumEltBits - NumPackedZeroBits) <= Known.countMinLeadingZeros()) | |||
| 21198 | if (SDValue V = | |||
| 21199 | truncateVectorWithPACK(X86ISD::PACKUS, VT, In, DL, DAG, Subtarget)) | |||
| 21200 | return V; | |||
| 21201 | ||||
| 21202 | // Truncate with PACKSS if we are truncating a vector with sign-bits that | |||
| 21203 | // extend all the way to the packed/truncated value. | |||
| 21204 | if ((InNumEltBits - NumPackedSignBits) < DAG.ComputeNumSignBits(In)) | |||
| 21205 | if (SDValue V = | |||
| 21206 | truncateVectorWithPACK(X86ISD::PACKSS, VT, In, DL, DAG, Subtarget)) | |||
| 21207 | return V; | |||
| 21208 | ||||
| 21209 | // Handle truncation of V256 to V128 using shuffles. | |||
| 21210 | assert(VT.is128BitVector() && InVT.is256BitVector() && "Unexpected types!")((void)0); | |||
| 21211 | ||||
| 21212 | if ((VT == MVT::v4i32) && (InVT == MVT::v4i64)) { | |||
| 21213 | In = DAG.getBitcast(MVT::v8i32, In); | |||
| 21214 | ||||
| 21215 | // On AVX2, v4i64 -> v4i32 becomes VPERMD. | |||
| 21216 | if (Subtarget.hasInt256()) { | |||
| 21217 | static const int ShufMask[] = {0, 2, 4, 6, -1, -1, -1, -1}; | |||
| 21218 | In = DAG.getVectorShuffle(MVT::v8i32, DL, In, In, ShufMask); | |||
| 21219 | return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, In, | |||
| 21220 | DAG.getIntPtrConstant(0, DL)); | |||
| 21221 | } | |||
| 21222 | ||||
| 21223 | SDValue OpLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i32, In, | |||
| 21224 | DAG.getIntPtrConstant(0, DL)); | |||
| 21225 | SDValue OpHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i32, In, | |||
| 21226 | DAG.getIntPtrConstant(4, DL)); | |||
| 21227 | static const int ShufMask[] = {0, 2, 4, 6}; | |||
| 21228 | return DAG.getVectorShuffle(VT, DL, OpLo, OpHi, ShufMask); | |||
| 21229 | } | |||
| 21230 | ||||
| 21231 | if ((VT == MVT::v8i16) && (InVT == MVT::v8i32)) { | |||
| 21232 | In = DAG.getBitcast(MVT::v32i8, In); | |||
| 21233 | ||||
| 21234 | // On AVX2, v8i32 -> v8i16 becomes PSHUFB. | |||
| 21235 | if (Subtarget.hasInt256()) { | |||
| 21236 | // The PSHUFB mask: | |||
| 21237 | static const int ShufMask1[] = { 0, 1, 4, 5, 8, 9, 12, 13, | |||
| 21238 | -1, -1, -1, -1, -1, -1, -1, -1, | |||
| 21239 | 16, 17, 20, 21, 24, 25, 28, 29, | |||
| 21240 | -1, -1, -1, -1, -1, -1, -1, -1 }; | |||
| 21241 | In = DAG.getVectorShuffle(MVT::v32i8, DL, In, In, ShufMask1); | |||
| 21242 | In = DAG.getBitcast(MVT::v4i64, In); | |||
| 21243 | ||||
| 21244 | static const int ShufMask2[] = {0, 2, -1, -1}; | |||
| 21245 | In = DAG.getVectorShuffle(MVT::v4i64, DL, In, In, ShufMask2); | |||
| 21246 | return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v8i16, | |||
| 21247 | DAG.getBitcast(MVT::v16i16, In), | |||
| 21248 | DAG.getIntPtrConstant(0, DL)); | |||
| 21249 | } | |||
| 21250 | ||||
| 21251 | SDValue OpLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v16i8, In, | |||
| 21252 | DAG.getIntPtrConstant(0, DL)); | |||
| 21253 | SDValue OpHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v16i8, In, | |||
| 21254 | DAG.getIntPtrConstant(16, DL)); | |||
| 21255 | ||||
| 21256 | // The PSHUFB mask: | |||
| 21257 | static const int ShufMask1[] = {0, 1, 4, 5, 8, 9, 12, 13, | |||
| 21258 | -1, -1, -1, -1, -1, -1, -1, -1}; | |||
| 21259 | ||||
| 21260 | OpLo = DAG.getVectorShuffle(MVT::v16i8, DL, OpLo, OpLo, ShufMask1); | |||
| 21261 | OpHi = DAG.getVectorShuffle(MVT::v16i8, DL, OpHi, OpHi, ShufMask1); | |||
| 21262 | ||||
| 21263 | OpLo = DAG.getBitcast(MVT::v4i32, OpLo); | |||
| 21264 | OpHi = DAG.getBitcast(MVT::v4i32, OpHi); | |||
| 21265 | ||||
| 21266 | // The MOVLHPS Mask: | |||
| 21267 | static const int ShufMask2[] = {0, 1, 4, 5}; | |||
| 21268 | SDValue res = DAG.getVectorShuffle(MVT::v4i32, DL, OpLo, OpHi, ShufMask2); | |||
| 21269 | return DAG.getBitcast(MVT::v8i16, res); | |||
| 21270 | } | |||
| 21271 | ||||
| 21272 | if (VT == MVT::v16i8 && InVT == MVT::v16i16) { | |||
| 21273 | // Use an AND to zero uppper bits for PACKUS. | |||
| 21274 | In = DAG.getNode(ISD::AND, DL, InVT, In, DAG.getConstant(255, DL, InVT)); | |||
| 21275 | ||||
| 21276 | SDValue InLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v8i16, In, | |||
| 21277 | DAG.getIntPtrConstant(0, DL)); | |||
| 21278 | SDValue InHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v8i16, In, | |||
| 21279 | DAG.getIntPtrConstant(8, DL)); | |||
| 21280 | return DAG.getNode(X86ISD::PACKUS, DL, VT, InLo, InHi); | |||
| 21281 | } | |||
| 21282 | ||||
| 21283 | llvm_unreachable("All 256->128 cases should have been handled above!")__builtin_unreachable(); | |||
| 21284 | } | |||
| 21285 | ||||
| 21286 | // We can leverage the specific way the "cvttps2dq/cvttpd2dq" instruction | |||
| 21287 | // behaves on out of range inputs to generate optimized conversions. | |||
| 21288 | static SDValue expandFP_TO_UINT_SSE(MVT VT, SDValue Src, const SDLoc &dl, | |||
| 21289 | SelectionDAG &DAG, | |||
| 21290 | const X86Subtarget &Subtarget) { | |||
| 21291 | MVT SrcVT = Src.getSimpleValueType(); | |||
| 21292 | unsigned DstBits = VT.getScalarSizeInBits(); | |||
| 21293 | assert(DstBits == 32 && "expandFP_TO_UINT_SSE - only vXi32 supported")((void)0); | |||
| 21294 | ||||
| 21295 | // Calculate the converted result for values in the range 0 to | |||
| 21296 | // 2^31-1 ("Small") and from 2^31 to 2^32-1 ("Big"). | |||
| 21297 | SDValue Small = DAG.getNode(X86ISD::CVTTP2SI, dl, VT, Src); | |||
| 21298 | SDValue Big = | |||
| 21299 | DAG.getNode(X86ISD::CVTTP2SI, dl, VT, | |||
| 21300 | DAG.getNode(ISD::FSUB, dl, SrcVT, Src, | |||
| 21301 | DAG.getConstantFP(2147483648.0f, dl, SrcVT))); | |||
| 21302 | ||||
| 21303 | // The "CVTTP2SI" instruction conveniently sets the sign bit if | |||
| 21304 | // and only if the value was out of range. So we can use that | |||
| 21305 | // as our indicator that we rather use "Big" instead of "Small". | |||
| 21306 | // | |||
| 21307 | // Use "Small" if "IsOverflown" has all bits cleared | |||
| 21308 | // and "0x80000000 | Big" if all bits in "IsOverflown" are set. | |||
| 21309 | ||||
| 21310 | // AVX1 can't use the signsplat masking for 256-bit vectors - we have to | |||
| 21311 | // use the slightly slower blendv select instead. | |||
| 21312 | if (VT == MVT::v8i32 && !Subtarget.hasAVX2()) { | |||
| 21313 | SDValue Overflow = DAG.getNode(ISD::OR, dl, VT, Small, Big); | |||
| 21314 | return DAG.getNode(X86ISD::BLENDV, dl, VT, Small, Overflow, Small); | |||
| 21315 | } | |||
| 21316 | ||||
| 21317 | SDValue IsOverflown = | |||
| 21318 | DAG.getNode(X86ISD::VSRAI, dl, VT, Small, | |||
| 21319 | DAG.getTargetConstant(DstBits - 1, dl, MVT::i8)); | |||
| 21320 | return DAG.getNode(ISD::OR, dl, VT, Small, | |||
| 21321 | DAG.getNode(ISD::AND, dl, VT, Big, IsOverflown)); | |||
| 21322 | } | |||
| 21323 | ||||
| 21324 | SDValue X86TargetLowering::LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) const { | |||
| 21325 | bool IsStrict = Op->isStrictFPOpcode(); | |||
| 21326 | bool IsSigned = Op.getOpcode() == ISD::FP_TO_SINT || | |||
| 21327 | Op.getOpcode() == ISD::STRICT_FP_TO_SINT; | |||
| 21328 | MVT VT = Op->getSimpleValueType(0); | |||
| 21329 | SDValue Src = Op.getOperand(IsStrict ? 1 : 0); | |||
| 21330 | MVT SrcVT = Src.getSimpleValueType(); | |||
| 21331 | SDLoc dl(Op); | |||
| 21332 | ||||
| 21333 | if (VT.isVector()) { | |||
| 21334 | if (VT == MVT::v2i1 && SrcVT == MVT::v2f64) { | |||
| 21335 | MVT ResVT = MVT::v4i32; | |||
| 21336 | MVT TruncVT = MVT::v4i1; | |||
| 21337 | unsigned Opc; | |||
| 21338 | if (IsStrict) | |||
| 21339 | Opc = IsSigned ? X86ISD::STRICT_CVTTP2SI : X86ISD::STRICT_CVTTP2UI; | |||
| 21340 | else | |||
| 21341 | Opc = IsSigned ? X86ISD::CVTTP2SI : X86ISD::CVTTP2UI; | |||
| 21342 | ||||
| 21343 | if (!IsSigned && !Subtarget.hasVLX()) { | |||
| 21344 | assert(Subtarget.useAVX512Regs() && "Unexpected features!")((void)0); | |||
| 21345 | // Widen to 512-bits. | |||
| 21346 | ResVT = MVT::v8i32; | |||
| 21347 | TruncVT = MVT::v8i1; | |||
| 21348 | Opc = Op.getOpcode(); | |||
| 21349 | // Need to concat with zero vector for strict fp to avoid spurious | |||
| 21350 | // exceptions. | |||
| 21351 | // TODO: Should we just do this for non-strict as well? | |||
| 21352 | SDValue Tmp = IsStrict ? DAG.getConstantFP(0.0, dl, MVT::v8f64) | |||
| 21353 | : DAG.getUNDEF(MVT::v8f64); | |||
| 21354 | Src = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, MVT::v8f64, Tmp, Src, | |||
| 21355 | DAG.getIntPtrConstant(0, dl)); | |||
| 21356 | } | |||
| 21357 | SDValue Res, Chain; | |||
| 21358 | if (IsStrict) { | |||
| 21359 | Res = | |||
| 21360 | DAG.getNode(Opc, dl, {ResVT, MVT::Other}, {Op->getOperand(0), Src}); | |||
| 21361 | Chain = Res.getValue(1); | |||
| 21362 | } else { | |||
| 21363 | Res = DAG.getNode(Opc, dl, ResVT, Src); | |||
| 21364 | } | |||
| 21365 | ||||
| 21366 | Res = DAG.getNode(ISD::TRUNCATE, dl, TruncVT, Res); | |||
| 21367 | Res = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v2i1, Res, | |||
| 21368 | DAG.getIntPtrConstant(0, dl)); | |||
| 21369 | if (IsStrict) | |||
| 21370 | return DAG.getMergeValues({Res, Chain}, dl); | |||
| 21371 | return Res; | |||
| 21372 | } | |||
| 21373 | ||||
| 21374 | // v8f64->v8i32 is legal, but we need v8i32 to be custom for v8f32. | |||
| 21375 | if (VT == MVT::v8i32 && SrcVT == MVT::v8f64) { | |||
| 21376 | assert(!IsSigned && "Expected unsigned conversion!")((void)0); | |||
| 21377 | assert(Subtarget.useAVX512Regs() && "Requires avx512f")((void)0); | |||
| 21378 | return Op; | |||
| 21379 | } | |||
| 21380 | ||||
| 21381 | // Widen vXi32 fp_to_uint with avx512f to 512-bit source. | |||
| 21382 | if ((VT == MVT::v4i32 || VT == MVT::v8i32) && | |||
| 21383 | (SrcVT == MVT::v4f64 || SrcVT == MVT::v4f32 || SrcVT == MVT::v8f32) && | |||
| 21384 | Subtarget.useAVX512Regs()) { | |||
| 21385 | assert(!IsSigned && "Expected unsigned conversion!")((void)0); | |||
| 21386 | assert(!Subtarget.hasVLX() && "Unexpected features!")((void)0); | |||
| 21387 | MVT WideVT = SrcVT == MVT::v4f64 ? MVT::v8f64 : MVT::v16f32; | |||
| 21388 | MVT ResVT = SrcVT == MVT::v4f64 ? MVT::v8i32 : MVT::v16i32; | |||
| 21389 | // Need to concat with zero vector for strict fp to avoid spurious | |||
| 21390 | // exceptions. | |||
| 21391 | // TODO: Should we just do this for non-strict as well? | |||
| 21392 | SDValue Tmp = | |||
| 21393 | IsStrict ? DAG.getConstantFP(0.0, dl, WideVT) : DAG.getUNDEF(WideVT); | |||
| 21394 | Src = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, WideVT, Tmp, Src, | |||
| 21395 | DAG.getIntPtrConstant(0, dl)); | |||
| 21396 | ||||
| 21397 | SDValue Res, Chain; | |||
| 21398 | if (IsStrict) { | |||
| 21399 | Res = DAG.getNode(ISD::STRICT_FP_TO_UINT, dl, {ResVT, MVT::Other}, | |||
| 21400 | {Op->getOperand(0), Src}); | |||
| 21401 | Chain = Res.getValue(1); | |||
| 21402 | } else { | |||
| 21403 | Res = DAG.getNode(ISD::FP_TO_UINT, dl, ResVT, Src); | |||
| 21404 | } | |||
| 21405 | ||||
| 21406 | Res = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, Res, | |||
| 21407 | DAG.getIntPtrConstant(0, dl)); | |||
| 21408 | ||||
| 21409 | if (IsStrict) | |||
| 21410 | return DAG.getMergeValues({Res, Chain}, dl); | |||
| 21411 | return Res; | |||
| 21412 | } | |||
| 21413 | ||||
| 21414 | // Widen vXi64 fp_to_uint/fp_to_sint with avx512dq to 512-bit source. | |||
| 21415 | if ((VT == MVT::v2i64 || VT == MVT::v4i64) && | |||
| 21416 | (SrcVT == MVT::v2f64 || SrcVT == MVT::v4f64 || SrcVT == MVT::v4f32) && | |||
| 21417 | Subtarget.useAVX512Regs() && Subtarget.hasDQI()) { | |||
| 21418 | assert(!Subtarget.hasVLX() && "Unexpected features!")((void)0); | |||
| 21419 | MVT WideVT = SrcVT == MVT::v4f32 ? MVT::v8f32 : MVT::v8f64; | |||
| 21420 | // Need to concat with zero vector for strict fp to avoid spurious | |||
| 21421 | // exceptions. | |||
| 21422 | // TODO: Should we just do this for non-strict as well? | |||
| 21423 | SDValue Tmp = | |||
| 21424 | IsStrict ? DAG.getConstantFP(0.0, dl, WideVT) : DAG.getUNDEF(WideVT); | |||
| 21425 | Src = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, WideVT, Tmp, Src, | |||
| 21426 | DAG.getIntPtrConstant(0, dl)); | |||
| 21427 | ||||
| 21428 | SDValue Res, Chain; | |||
| 21429 | if (IsStrict) { | |||
| 21430 | Res = DAG.getNode(Op.getOpcode(), dl, {MVT::v8i64, MVT::Other}, | |||
| 21431 | {Op->getOperand(0), Src}); | |||
| 21432 | Chain = Res.getValue(1); | |||
| 21433 | } else { | |||
| 21434 | Res = DAG.getNode(Op.getOpcode(), dl, MVT::v8i64, Src); | |||
| 21435 | } | |||
| 21436 | ||||
| 21437 | Res = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, Res, | |||
| 21438 | DAG.getIntPtrConstant(0, dl)); | |||
| 21439 | ||||
| 21440 | if (IsStrict) | |||
| 21441 | return DAG.getMergeValues({Res, Chain}, dl); | |||
| 21442 | return Res; | |||
| 21443 | } | |||
| 21444 | ||||
| 21445 | if (VT == MVT::v2i64 && SrcVT == MVT::v2f32) { | |||
| 21446 | if (!Subtarget.hasVLX()) { | |||
| 21447 | // Non-strict nodes without VLX can we widened to v4f32->v4i64 by type | |||
| 21448 | // legalizer and then widened again by vector op legalization. | |||
| 21449 | if (!IsStrict) | |||
| 21450 | return SDValue(); | |||
| 21451 | ||||
| 21452 | SDValue Zero = DAG.getConstantFP(0.0, dl, MVT::v2f32); | |||
| 21453 | SDValue Tmp = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8f32, | |||
| 21454 | {Src, Zero, Zero, Zero}); | |||
| 21455 | Tmp = DAG.getNode(Op.getOpcode(), dl, {MVT::v8i64, MVT::Other}, | |||
| 21456 | {Op->getOperand(0), Tmp}); | |||
| 21457 | SDValue Chain = Tmp.getValue(1); | |||
| 21458 | Tmp = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v2i64, Tmp, | |||
| 21459 | DAG.getIntPtrConstant(0, dl)); | |||
| 21460 | return DAG.getMergeValues({Tmp, Chain}, dl); | |||
| 21461 | } | |||
| 21462 | ||||
| 21463 | assert(Subtarget.hasDQI() && Subtarget.hasVLX() && "Requires AVX512DQVL")((void)0); | |||
| 21464 | SDValue Tmp = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4f32, Src, | |||
| 21465 | DAG.getUNDEF(MVT::v2f32)); | |||
| 21466 | if (IsStrict) { | |||
| 21467 | unsigned Opc = IsSigned ? X86ISD::STRICT_CVTTP2SI | |||
| 21468 | : X86ISD::STRICT_CVTTP2UI; | |||
| 21469 | return DAG.getNode(Opc, dl, {VT, MVT::Other}, {Op->getOperand(0), Tmp}); | |||
| 21470 | } | |||
| 21471 | unsigned Opc = IsSigned ? X86ISD::CVTTP2SI : X86ISD::CVTTP2UI; | |||
| 21472 | return DAG.getNode(Opc, dl, VT, Tmp); | |||
| 21473 | } | |||
| 21474 | ||||
| 21475 | // Generate optimized instructions for pre AVX512 unsigned conversions from | |||
| 21476 | // vXf32 to vXi32. | |||
| 21477 | if ((VT == MVT::v4i32 && SrcVT == MVT::v4f32) || | |||
| 21478 | (VT == MVT::v4i32 && SrcVT == MVT::v4f64) || | |||
| 21479 | (VT == MVT::v8i32 && SrcVT == MVT::v8f32)) { | |||
| 21480 | assert(!IsSigned && "Expected unsigned conversion!")((void)0); | |||
| 21481 | return expandFP_TO_UINT_SSE(VT, Src, dl, DAG, Subtarget); | |||
| 21482 | } | |||
| 21483 | ||||
| 21484 | return SDValue(); | |||
| 21485 | } | |||
| 21486 | ||||
| 21487 | assert(!VT.isVector())((void)0); | |||
| 21488 | ||||
| 21489 | bool UseSSEReg = isScalarFPTypeInSSEReg(SrcVT); | |||
| 21490 | ||||
| 21491 | if (!IsSigned && UseSSEReg) { | |||
| 21492 | // Conversions from f32/f64 with AVX512 should be legal. | |||
| 21493 | if (Subtarget.hasAVX512()) | |||
| 21494 | return Op; | |||
| 21495 | ||||
| 21496 | // We can leverage the specific way the "cvttss2si/cvttsd2si" instruction | |||
| 21497 | // behaves on out of range inputs to generate optimized conversions. | |||
| 21498 | if (!IsStrict && ((VT == MVT::i32 && !Subtarget.is64Bit()) || | |||
| 21499 | (VT == MVT::i64 && Subtarget.is64Bit()))) { | |||
| 21500 | unsigned DstBits = VT.getScalarSizeInBits(); | |||
| 21501 | APInt UIntLimit = APInt::getSignMask(DstBits); | |||
| 21502 | SDValue FloatOffset = DAG.getNode(ISD::UINT_TO_FP, dl, SrcVT, | |||
| 21503 | DAG.getConstant(UIntLimit, dl, VT)); | |||
| 21504 | MVT SrcVecVT = MVT::getVectorVT(SrcVT, 128 / SrcVT.getScalarSizeInBits()); | |||
| 21505 | ||||
| 21506 | // Calculate the converted result for values in the range: | |||
| 21507 | // (i32) 0 to 2^31-1 ("Small") and from 2^31 to 2^32-1 ("Big"). | |||
| 21508 | // (i64) 0 to 2^63-1 ("Small") and from 2^63 to 2^64-1 ("Big"). | |||
| 21509 | SDValue Small = | |||
| 21510 | DAG.getNode(X86ISD::CVTTS2SI, dl, VT, | |||
| 21511 | DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, SrcVecVT, Src)); | |||
| 21512 | SDValue Big = DAG.getNode( | |||
| 21513 | X86ISD::CVTTS2SI, dl, VT, | |||
| 21514 | DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, SrcVecVT, | |||
| 21515 | DAG.getNode(ISD::FSUB, dl, SrcVT, Src, FloatOffset))); | |||
| 21516 | ||||
| 21517 | // The "CVTTS2SI" instruction conveniently sets the sign bit if | |||
| 21518 | // and only if the value was out of range. So we can use that | |||
| 21519 | // as our indicator that we rather use "Big" instead of "Small". | |||
| 21520 | // | |||
| 21521 | // Use "Small" if "IsOverflown" has all bits cleared | |||
| 21522 | // and "0x80000000 | Big" if all bits in "IsOverflown" are set. | |||
| 21523 | SDValue IsOverflown = DAG.getNode( | |||
| 21524 | ISD::SRA, dl, VT, Small, DAG.getConstant(DstBits - 1, dl, MVT::i8)); | |||
| 21525 | return DAG.getNode(ISD::OR, dl, VT, Small, | |||
| 21526 | DAG.getNode(ISD::AND, dl, VT, Big, IsOverflown)); | |||
| 21527 | } | |||
| 21528 | ||||
| 21529 | // Use default expansion for i64. | |||
| 21530 | if (VT == MVT::i64) | |||
| 21531 | return SDValue(); | |||
| 21532 | ||||
| 21533 | assert(VT == MVT::i32 && "Unexpected VT!")((void)0); | |||
| 21534 | ||||
| 21535 | // Promote i32 to i64 and use a signed operation on 64-bit targets. | |||
| 21536 | // FIXME: This does not generate an invalid exception if the input does not | |||
| 21537 | // fit in i32. PR44019 | |||
| 21538 | if (Subtarget.is64Bit()) { | |||
| 21539 | SDValue Res, Chain; | |||
| 21540 | if (IsStrict) { | |||
| 21541 | Res = DAG.getNode(ISD::STRICT_FP_TO_SINT, dl, { MVT::i64, MVT::Other}, | |||
| 21542 | { Op.getOperand(0), Src }); | |||
| 21543 | Chain = Res.getValue(1); | |||
| 21544 | } else | |||
| 21545 | Res = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::i64, Src); | |||
| 21546 | ||||
| 21547 | Res = DAG.getNode(ISD::TRUNCATE, dl, VT, Res); | |||
| 21548 | if (IsStrict) | |||
| 21549 | return DAG.getMergeValues({ Res, Chain }, dl); | |||
| 21550 | return Res; | |||
| 21551 | } | |||
| 21552 | ||||
| 21553 | // Use default expansion for SSE1/2 targets without SSE3. With SSE3 we can | |||
| 21554 | // use fisttp which will be handled later. | |||
| 21555 | if (!Subtarget.hasSSE3()) | |||
| 21556 | return SDValue(); | |||
| 21557 | } | |||
| 21558 | ||||
| 21559 | // Promote i16 to i32 if we can use a SSE operation or the type is f128. | |||
| 21560 | // FIXME: This does not generate an invalid exception if the input does not | |||
| 21561 | // fit in i16. PR44019 | |||
| 21562 | if (VT == MVT::i16 && (UseSSEReg || SrcVT == MVT::f128)) { | |||
| 21563 | assert(IsSigned && "Expected i16 FP_TO_UINT to have been promoted!")((void)0); | |||
| 21564 | SDValue Res, Chain; | |||
| 21565 | if (IsStrict) { | |||
| 21566 | Res = DAG.getNode(ISD::STRICT_FP_TO_SINT, dl, { MVT::i32, MVT::Other}, | |||
| 21567 | { Op.getOperand(0), Src }); | |||
| 21568 | Chain = Res.getValue(1); | |||
| 21569 | } else | |||
| 21570 | Res = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::i32, Src); | |||
| 21571 | ||||
| 21572 | Res = DAG.getNode(ISD::TRUNCATE, dl, VT, Res); | |||
| 21573 | if (IsStrict) | |||
| 21574 | return DAG.getMergeValues({ Res, Chain }, dl); | |||
| 21575 | return Res; | |||
| 21576 | } | |||
| 21577 | ||||
| 21578 | // If this is a FP_TO_SINT using SSEReg we're done. | |||
| 21579 | if (UseSSEReg && IsSigned) | |||
| 21580 | return Op; | |||
| 21581 | ||||
| 21582 | // fp128 needs to use a libcall. | |||
| 21583 | if (SrcVT == MVT::f128) { | |||
| 21584 | RTLIB::Libcall LC; | |||
| 21585 | if (IsSigned) | |||
| 21586 | LC = RTLIB::getFPTOSINT(SrcVT, VT); | |||
| 21587 | else | |||
| 21588 | LC = RTLIB::getFPTOUINT(SrcVT, VT); | |||
| 21589 | ||||
| 21590 | SDValue Chain = IsStrict ? Op.getOperand(0) : SDValue(); | |||
| 21591 | MakeLibCallOptions CallOptions; | |||
| 21592 | std::pair<SDValue, SDValue> Tmp = makeLibCall(DAG, LC, VT, Src, CallOptions, | |||
| 21593 | SDLoc(Op), Chain); | |||
| 21594 | ||||
| 21595 | if (IsStrict) | |||
| 21596 | return DAG.getMergeValues({ Tmp.first, Tmp.second }, dl); | |||
| 21597 | ||||
| 21598 | return Tmp.first; | |||
| 21599 | } | |||
| 21600 | ||||
| 21601 | // Fall back to X87. | |||
| 21602 | SDValue Chain; | |||
| 21603 | if (SDValue V = FP_TO_INTHelper(Op, DAG, IsSigned, Chain)) { | |||
| 21604 | if (IsStrict) | |||
| 21605 | return DAG.getMergeValues({V, Chain}, dl); | |||
| 21606 | return V; | |||
| 21607 | } | |||
| 21608 | ||||
| 21609 | llvm_unreachable("Expected FP_TO_INTHelper to handle all remaining cases.")__builtin_unreachable(); | |||
| 21610 | } | |||
| 21611 | ||||
| 21612 | SDValue X86TargetLowering::LowerLRINT_LLRINT(SDValue Op, | |||
| 21613 | SelectionDAG &DAG) const { | |||
| 21614 | SDValue Src = Op.getOperand(0); | |||
| 21615 | MVT SrcVT = Src.getSimpleValueType(); | |||
| 21616 | ||||
| 21617 | // If the source is in an SSE register, the node is Legal. | |||
| 21618 | if (isScalarFPTypeInSSEReg(SrcVT)) | |||
| 21619 | return Op; | |||
| 21620 | ||||
| 21621 | return LRINT_LLRINTHelper(Op.getNode(), DAG); | |||
| 21622 | } | |||
| 21623 | ||||
| 21624 | SDValue X86TargetLowering::LRINT_LLRINTHelper(SDNode *N, | |||
| 21625 | SelectionDAG &DAG) const { | |||
| 21626 | EVT DstVT = N->getValueType(0); | |||
| 21627 | SDValue Src = N->getOperand(0); | |||
| 21628 | EVT SrcVT = Src.getValueType(); | |||
| 21629 | ||||
| 21630 | if (SrcVT != MVT::f32 && SrcVT != MVT::f64 && SrcVT != MVT::f80) { | |||
| 21631 | // f16 must be promoted before using the lowering in this routine. | |||
| 21632 | // fp128 does not use this lowering. | |||
| 21633 | return SDValue(); | |||
| 21634 | } | |||
| 21635 | ||||
| 21636 | SDLoc DL(N); | |||
| 21637 | SDValue Chain = DAG.getEntryNode(); | |||
| 21638 | ||||
| 21639 | bool UseSSE = isScalarFPTypeInSSEReg(SrcVT); | |||
| 21640 | ||||
| 21641 | // If we're converting from SSE, the stack slot needs to hold both types. | |||
| 21642 | // Otherwise it only needs to hold the DstVT. | |||
| 21643 | EVT OtherVT = UseSSE ? SrcVT : DstVT; | |||
| 21644 | SDValue StackPtr = DAG.CreateStackTemporary(DstVT, OtherVT); | |||
| 21645 | int SPFI = cast<FrameIndexSDNode>(StackPtr.getNode())->getIndex(); | |||
| 21646 | MachinePointerInfo MPI = | |||
| 21647 | MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SPFI); | |||
| 21648 | ||||
| 21649 | if (UseSSE) { | |||
| 21650 | assert(DstVT == MVT::i64 && "Invalid LRINT/LLRINT to lower!")((void)0); | |||
| 21651 | Chain = DAG.getStore(Chain, DL, Src, StackPtr, MPI); | |||
| 21652 | SDVTList Tys = DAG.getVTList(MVT::f80, MVT::Other); | |||
| 21653 | SDValue Ops[] = { Chain, StackPtr }; | |||
| 21654 | ||||
| 21655 | Src = DAG.getMemIntrinsicNode(X86ISD::FLD, DL, Tys, Ops, SrcVT, MPI, | |||
| 21656 | /*Align*/ None, MachineMemOperand::MOLoad); | |||
| 21657 | Chain = Src.getValue(1); | |||
| 21658 | } | |||
| 21659 | ||||
| 21660 | SDValue StoreOps[] = { Chain, Src, StackPtr }; | |||
| 21661 | Chain = DAG.getMemIntrinsicNode(X86ISD::FIST, DL, DAG.getVTList(MVT::Other), | |||
| 21662 | StoreOps, DstVT, MPI, /*Align*/ None, | |||
| 21663 | MachineMemOperand::MOStore); | |||
| 21664 | ||||
| 21665 | return DAG.getLoad(DstVT, DL, Chain, StackPtr, MPI); | |||
| 21666 | } | |||
| 21667 | ||||
| 21668 | SDValue | |||
| 21669 | X86TargetLowering::LowerFP_TO_INT_SAT(SDValue Op, SelectionDAG &DAG) const { | |||
| 21670 | // This is based on the TargetLowering::expandFP_TO_INT_SAT implementation, | |||
| 21671 | // but making use of X86 specifics to produce better instruction sequences. | |||
| 21672 | SDNode *Node = Op.getNode(); | |||
| 21673 | bool IsSigned = Node->getOpcode() == ISD::FP_TO_SINT_SAT; | |||
| 21674 | unsigned FpToIntOpcode = IsSigned ? ISD::FP_TO_SINT : ISD::FP_TO_UINT; | |||
| 21675 | SDLoc dl(SDValue(Node, 0)); | |||
| 21676 | SDValue Src = Node->getOperand(0); | |||
| 21677 | ||||
| 21678 | // There are three types involved here: SrcVT is the source floating point | |||
| 21679 | // type, DstVT is the type of the result, and TmpVT is the result of the | |||
| 21680 | // intermediate FP_TO_*INT operation we'll use (which may be a promotion of | |||
| 21681 | // DstVT). | |||
| 21682 | EVT SrcVT = Src.getValueType(); | |||
| 21683 | EVT DstVT = Node->getValueType(0); | |||
| 21684 | EVT TmpVT = DstVT; | |||
| 21685 | ||||
| 21686 | // This code is only for floats and doubles. Fall back to generic code for | |||
| 21687 | // anything else. | |||
| 21688 | if (!isScalarFPTypeInSSEReg(SrcVT)) | |||
| 21689 | return SDValue(); | |||
| 21690 | ||||
| 21691 | EVT SatVT = cast<VTSDNode>(Node->getOperand(1))->getVT(); | |||
| 21692 | unsigned SatWidth = SatVT.getScalarSizeInBits(); | |||
| 21693 | unsigned DstWidth = DstVT.getScalarSizeInBits(); | |||
| 21694 | unsigned TmpWidth = TmpVT.getScalarSizeInBits(); | |||
| 21695 | assert(SatWidth <= DstWidth && SatWidth <= TmpWidth &&((void)0) | |||
| 21696 | "Expected saturation width smaller than result width")((void)0); | |||
| 21697 | ||||
| 21698 | // Promote result of FP_TO_*INT to at least 32 bits. | |||
| 21699 | if (TmpWidth < 32) { | |||
| 21700 | TmpVT = MVT::i32; | |||
| 21701 | TmpWidth = 32; | |||
| 21702 | } | |||
| 21703 | ||||
| 21704 | // Promote conversions to unsigned 32-bit to 64-bit, because it will allow | |||
| 21705 | // us to use a native signed conversion instead. | |||
| 21706 | if (SatWidth == 32 && !IsSigned && Subtarget.is64Bit()) { | |||
| 21707 | TmpVT = MVT::i64; | |||
| 21708 | TmpWidth = 64; | |||
| 21709 | } | |||
| 21710 | ||||
| 21711 | // If the saturation width is smaller than the size of the temporary result, | |||
| 21712 | // we can always use signed conversion, which is native. | |||
| 21713 | if (SatWidth < TmpWidth) | |||
| 21714 | FpToIntOpcode = ISD::FP_TO_SINT; | |||
| 21715 | ||||
| 21716 | // Determine minimum and maximum integer values and their corresponding | |||
| 21717 | // floating-point values. | |||
| 21718 | APInt MinInt, MaxInt; | |||
| 21719 | if (IsSigned) { | |||
| 21720 | MinInt = APInt::getSignedMinValue(SatWidth).sextOrSelf(DstWidth); | |||
| 21721 | MaxInt = APInt::getSignedMaxValue(SatWidth).sextOrSelf(DstWidth); | |||
| 21722 | } else { | |||
| 21723 | MinInt = APInt::getMinValue(SatWidth).zextOrSelf(DstWidth); | |||
| 21724 | MaxInt = APInt::getMaxValue(SatWidth).zextOrSelf(DstWidth); | |||
| 21725 | } | |||
| 21726 | ||||
| 21727 | APFloat MinFloat(DAG.EVTToAPFloatSemantics(SrcVT)); | |||
| 21728 | APFloat MaxFloat(DAG.EVTToAPFloatSemantics(SrcVT)); | |||
| 21729 | ||||
| 21730 | APFloat::opStatus MinStatus = MinFloat.convertFromAPInt( | |||
| 21731 | MinInt, IsSigned, APFloat::rmTowardZero); | |||
| 21732 | APFloat::opStatus MaxStatus = MaxFloat.convertFromAPInt( | |||
| 21733 | MaxInt, IsSigned, APFloat::rmTowardZero); | |||
| 21734 | bool AreExactFloatBounds = !(MinStatus & APFloat::opStatus::opInexact) | |||
| 21735 | && !(MaxStatus & APFloat::opStatus::opInexact); | |||
| 21736 | ||||
| 21737 | SDValue MinFloatNode = DAG.getConstantFP(MinFloat, dl, SrcVT); | |||
| 21738 | SDValue MaxFloatNode = DAG.getConstantFP(MaxFloat, dl, SrcVT); | |||
| 21739 | ||||
| 21740 | // If the integer bounds are exactly representable as floats, emit a | |||
| 21741 | // min+max+fptoi sequence. Otherwise use comparisons and selects. | |||
| 21742 | if (AreExactFloatBounds) { | |||
| 21743 | if (DstVT != TmpVT) { | |||
| 21744 | // Clamp by MinFloat from below. If Src is NaN, propagate NaN. | |||
| 21745 | SDValue MinClamped = DAG.getNode( | |||
| 21746 | X86ISD::FMAX, dl, SrcVT, MinFloatNode, Src); | |||
| 21747 | // Clamp by MaxFloat from above. If Src is NaN, propagate NaN. | |||
| 21748 | SDValue BothClamped = DAG.getNode( | |||
| 21749 | X86ISD::FMIN, dl, SrcVT, MaxFloatNode, MinClamped); | |||
| 21750 | // Convert clamped value to integer. | |||
| 21751 | SDValue FpToInt = DAG.getNode(FpToIntOpcode, dl, TmpVT, BothClamped); | |||
| 21752 | ||||
| 21753 | // NaN will become INDVAL, with the top bit set and the rest zero. | |||
| 21754 | // Truncation will discard the top bit, resulting in zero. | |||
| 21755 | return DAG.getNode(ISD::TRUNCATE, dl, DstVT, FpToInt); | |||
| 21756 | } | |||
| 21757 | ||||
| 21758 | // Clamp by MinFloat from below. If Src is NaN, the result is MinFloat. | |||
| 21759 | SDValue MinClamped = DAG.getNode( | |||
| 21760 | X86ISD::FMAX, dl, SrcVT, Src, MinFloatNode); | |||
| 21761 | // Clamp by MaxFloat from above. NaN cannot occur. | |||
| 21762 | SDValue BothClamped = DAG.getNode( | |||
| 21763 | X86ISD::FMINC, dl, SrcVT, MinClamped, MaxFloatNode); | |||
| 21764 | // Convert clamped value to integer. | |||
| 21765 | SDValue FpToInt = DAG.getNode(FpToIntOpcode, dl, DstVT, BothClamped); | |||
| 21766 | ||||
| 21767 | if (!IsSigned) { | |||
| 21768 | // In the unsigned case we're done, because we mapped NaN to MinFloat, | |||
| 21769 | // which is zero. | |||
| 21770 | return FpToInt; | |||
| 21771 | } | |||
| 21772 | ||||
| 21773 | // Otherwise, select zero if Src is NaN. | |||
| 21774 | SDValue ZeroInt = DAG.getConstant(0, dl, DstVT); | |||
| 21775 | return DAG.getSelectCC( | |||
| 21776 | dl, Src, Src, ZeroInt, FpToInt, ISD::CondCode::SETUO); | |||
| 21777 | } | |||
| 21778 | ||||
| 21779 | SDValue MinIntNode = DAG.getConstant(MinInt, dl, DstVT); | |||
| 21780 | SDValue MaxIntNode = DAG.getConstant(MaxInt, dl, DstVT); | |||
| 21781 | ||||
| 21782 | // Result of direct conversion, which may be selected away. | |||
| 21783 | SDValue FpToInt = DAG.getNode(FpToIntOpcode, dl, TmpVT, Src); | |||
| 21784 | ||||
| 21785 | if (DstVT != TmpVT) { | |||
| 21786 | // NaN will become INDVAL, with the top bit set and the rest zero. | |||
| 21787 | // Truncation will discard the top bit, resulting in zero. | |||
| 21788 | FpToInt = DAG.getNode(ISD::TRUNCATE, dl, DstVT, FpToInt); | |||
| 21789 | } | |||
| 21790 | ||||
| 21791 | SDValue Select = FpToInt; | |||
| 21792 | // For signed conversions where we saturate to the same size as the | |||
| 21793 | // result type of the fptoi instructions, INDVAL coincides with integer | |||
| 21794 | // minimum, so we don't need to explicitly check it. | |||
| 21795 | if (!IsSigned || SatWidth != TmpVT.getScalarSizeInBits()) { | |||
| 21796 | // If Src ULT MinFloat, select MinInt. In particular, this also selects | |||
| 21797 | // MinInt if Src is NaN. | |||
| 21798 | Select = DAG.getSelectCC( | |||
| 21799 | dl, Src, MinFloatNode, MinIntNode, Select, ISD::CondCode::SETULT); | |||
| 21800 | } | |||
| 21801 | ||||
| 21802 | // If Src OGT MaxFloat, select MaxInt. | |||
| 21803 | Select = DAG.getSelectCC( | |||
| 21804 | dl, Src, MaxFloatNode, MaxIntNode, Select, ISD::CondCode::SETOGT); | |||
| 21805 | ||||
| 21806 | // In the unsigned case we are done, because we mapped NaN to MinInt, which | |||
| 21807 | // is already zero. The promoted case was already handled above. | |||
| 21808 | if (!IsSigned || DstVT != TmpVT) { | |||
| 21809 | return Select; | |||
| 21810 | } | |||
| 21811 | ||||
| 21812 | // Otherwise, select 0 if Src is NaN. | |||
| 21813 | SDValue ZeroInt = DAG.getConstant(0, dl, DstVT); | |||
| 21814 | return DAG.getSelectCC( | |||
| 21815 | dl, Src, Src, ZeroInt, Select, ISD::CondCode::SETUO); | |||
| 21816 | } | |||
| 21817 | ||||
| 21818 | SDValue X86TargetLowering::LowerFP_EXTEND(SDValue Op, SelectionDAG &DAG) const { | |||
| 21819 | bool IsStrict = Op->isStrictFPOpcode(); | |||
| 21820 | ||||
| 21821 | SDLoc DL(Op); | |||
| 21822 | MVT VT = Op.getSimpleValueType(); | |||
| 21823 | SDValue In = Op.getOperand(IsStrict ? 1 : 0); | |||
| 21824 | MVT SVT = In.getSimpleValueType(); | |||
| 21825 | ||||
| 21826 | if (VT == MVT::f128) | |||
| 21827 | return SDValue(); | |||
| 21828 | ||||
| 21829 | assert(SVT == MVT::v2f32 && "Only customize MVT::v2f32 type legalization!")((void)0); | |||
| 21830 | ||||
| 21831 | SDValue Res = | |||
| 21832 | DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v4f32, In, DAG.getUNDEF(SVT)); | |||
| 21833 | if (IsStrict) | |||
| 21834 | return DAG.getNode(X86ISD::STRICT_VFPEXT, DL, {VT, MVT::Other}, | |||
| 21835 | {Op->getOperand(0), Res}); | |||
| 21836 | return DAG.getNode(X86ISD::VFPEXT, DL, VT, Res); | |||
| 21837 | } | |||
| 21838 | ||||
| 21839 | SDValue X86TargetLowering::LowerFP_ROUND(SDValue Op, SelectionDAG &DAG) const { | |||
| 21840 | bool IsStrict = Op->isStrictFPOpcode(); | |||
| 21841 | SDValue In = Op.getOperand(IsStrict ? 1 : 0); | |||
| 21842 | // It's legal except when f128 is involved | |||
| 21843 | if (In.getSimpleValueType() != MVT::f128) | |||
| 21844 | return Op; | |||
| 21845 | ||||
| 21846 | return SDValue(); | |||
| 21847 | } | |||
| 21848 | ||||
| 21849 | static SDValue LowerFP16_TO_FP(SDValue Op, SelectionDAG &DAG) { | |||
| 21850 | bool IsStrict = Op->isStrictFPOpcode(); | |||
| 21851 | SDValue Src = Op.getOperand(IsStrict ? 1 : 0); | |||
| 21852 | assert(Src.getValueType() == MVT::i16 && Op.getValueType() == MVT::f32 &&((void)0) | |||
| 21853 | "Unexpected VT!")((void)0); | |||
| 21854 | ||||
| 21855 | SDLoc dl(Op); | |||
| 21856 | SDValue Res = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, | |||
| 21857 | DAG.getConstant(0, dl, MVT::v8i16), Src, | |||
| 21858 | DAG.getIntPtrConstant(0, dl)); | |||
| 21859 | ||||
| 21860 | SDValue Chain; | |||
| 21861 | if (IsStrict) { | |||
| 21862 | Res = DAG.getNode(X86ISD::STRICT_CVTPH2PS, dl, {MVT::v4f32, MVT::Other}, | |||
| 21863 | {Op.getOperand(0), Res}); | |||
| 21864 | Chain = Res.getValue(1); | |||
| 21865 | } else { | |||
| 21866 | Res = DAG.getNode(X86ISD::CVTPH2PS, dl, MVT::v4f32, Res); | |||
| 21867 | } | |||
| 21868 | ||||
| 21869 | Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, Res, | |||
| 21870 | DAG.getIntPtrConstant(0, dl)); | |||
| 21871 | ||||
| 21872 | if (IsStrict) | |||
| 21873 | return DAG.getMergeValues({Res, Chain}, dl); | |||
| 21874 | ||||
| 21875 | return Res; | |||
| 21876 | } | |||
| 21877 | ||||
| 21878 | static SDValue LowerFP_TO_FP16(SDValue Op, SelectionDAG &DAG) { | |||
| 21879 | bool IsStrict = Op->isStrictFPOpcode(); | |||
| 21880 | SDValue Src = Op.getOperand(IsStrict ? 1 : 0); | |||
| 21881 | assert(Src.getValueType() == MVT::f32 && Op.getValueType() == MVT::i16 &&((void)0) | |||
| 21882 | "Unexpected VT!")((void)0); | |||
| 21883 | ||||
| 21884 | SDLoc dl(Op); | |||
| 21885 | SDValue Res, Chain; | |||
| 21886 | if (IsStrict) { | |||
| 21887 | Res = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v4f32, | |||
| 21888 | DAG.getConstantFP(0, dl, MVT::v4f32), Src, | |||
| 21889 | DAG.getIntPtrConstant(0, dl)); | |||
| 21890 | Res = DAG.getNode( | |||
| 21891 | X86ISD::STRICT_CVTPS2PH, dl, {MVT::v8i16, MVT::Other}, | |||
| 21892 | {Op.getOperand(0), Res, DAG.getTargetConstant(4, dl, MVT::i32)}); | |||
| 21893 | Chain = Res.getValue(1); | |||
| 21894 | } else { | |||
| 21895 | // FIXME: Should we use zeros for upper elements for non-strict? | |||
| 21896 | Res = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, Src); | |||
| 21897 | Res = DAG.getNode(X86ISD::CVTPS2PH, dl, MVT::v8i16, Res, | |||
| 21898 | DAG.getTargetConstant(4, dl, MVT::i32)); | |||
| 21899 | } | |||
| 21900 | ||||
| 21901 | Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Res, | |||
| 21902 | DAG.getIntPtrConstant(0, dl)); | |||
| 21903 | ||||
| 21904 | if (IsStrict) | |||
| 21905 | return DAG.getMergeValues({Res, Chain}, dl); | |||
| 21906 | ||||
| 21907 | return Res; | |||
| 21908 | } | |||
| 21909 | ||||
| 21910 | /// Depending on uarch and/or optimizing for size, we might prefer to use a | |||
| 21911 | /// vector operation in place of the typical scalar operation. | |||
| 21912 | static SDValue lowerAddSubToHorizontalOp(SDValue Op, SelectionDAG &DAG, | |||
| 21913 | const X86Subtarget &Subtarget) { | |||
| 21914 | // If both operands have other uses, this is probably not profitable. | |||
| 21915 | SDValue LHS = Op.getOperand(0); | |||
| 21916 | SDValue RHS = Op.getOperand(1); | |||
| 21917 | if (!LHS.hasOneUse() && !RHS.hasOneUse()) | |||
| 21918 | return Op; | |||
| 21919 | ||||
| 21920 | // FP horizontal add/sub were added with SSE3. Integer with SSSE3. | |||
| 21921 | bool IsFP = Op.getSimpleValueType().isFloatingPoint(); | |||
| 21922 | if (IsFP && !Subtarget.hasSSE3()) | |||
| 21923 | return Op; | |||
| 21924 | if (!IsFP && !Subtarget.hasSSSE3()) | |||
| 21925 | return Op; | |||
| 21926 | ||||
| 21927 | // Extract from a common vector. | |||
| 21928 | if (LHS.getOpcode() != ISD::EXTRACT_VECTOR_ELT || | |||
| 21929 | RHS.getOpcode() != ISD::EXTRACT_VECTOR_ELT || | |||
| 21930 | LHS.getOperand(0) != RHS.getOperand(0) || | |||
| 21931 | !isa<ConstantSDNode>(LHS.getOperand(1)) || | |||
| 21932 | !isa<ConstantSDNode>(RHS.getOperand(1)) || | |||
| 21933 | !shouldUseHorizontalOp(true, DAG, Subtarget)) | |||
| 21934 | return Op; | |||
| 21935 | ||||
| 21936 | // Allow commuted 'hadd' ops. | |||
| 21937 | // TODO: Allow commuted (f)sub by negating the result of (F)HSUB? | |||
| 21938 | unsigned HOpcode; | |||
| 21939 | switch (Op.getOpcode()) { | |||
| 21940 | case ISD::ADD: HOpcode = X86ISD::HADD; break; | |||
| 21941 | case ISD::SUB: HOpcode = X86ISD::HSUB; break; | |||
| 21942 | case ISD::FADD: HOpcode = X86ISD::FHADD; break; | |||
| 21943 | case ISD::FSUB: HOpcode = X86ISD::FHSUB; break; | |||
| 21944 | default: | |||
| 21945 | llvm_unreachable("Trying to lower unsupported opcode to horizontal op")__builtin_unreachable(); | |||
| 21946 | } | |||
| 21947 | unsigned LExtIndex = LHS.getConstantOperandVal(1); | |||
| 21948 | unsigned RExtIndex = RHS.getConstantOperandVal(1); | |||
| 21949 | if ((LExtIndex & 1) == 1 && (RExtIndex & 1) == 0 && | |||
| 21950 | (HOpcode == X86ISD::HADD || HOpcode == X86ISD::FHADD)) | |||
| 21951 | std::swap(LExtIndex, RExtIndex); | |||
| 21952 | ||||
| 21953 | if ((LExtIndex & 1) != 0 || RExtIndex != (LExtIndex + 1)) | |||
| 21954 | return Op; | |||
| 21955 | ||||
| 21956 | SDValue X = LHS.getOperand(0); | |||
| 21957 | EVT VecVT = X.getValueType(); | |||
| 21958 | unsigned BitWidth = VecVT.getSizeInBits(); | |||
| 21959 | unsigned NumLanes = BitWidth / 128; | |||
| 21960 | unsigned NumEltsPerLane = VecVT.getVectorNumElements() / NumLanes; | |||
| 21961 | assert((BitWidth == 128 || BitWidth == 256 || BitWidth == 512) &&((void)0) | |||
| 21962 | "Not expecting illegal vector widths here")((void)0); | |||
| 21963 | ||||
| 21964 | // Creating a 256-bit horizontal op would be wasteful, and there is no 512-bit | |||
| 21965 | // equivalent, so extract the 256/512-bit source op to 128-bit if we can. | |||
| 21966 | SDLoc DL(Op); | |||
| 21967 | if (BitWidth == 256 || BitWidth == 512) { | |||
| 21968 | unsigned LaneIdx = LExtIndex / NumEltsPerLane; | |||
| 21969 | X = extract128BitVector(X, LaneIdx * NumEltsPerLane, DAG, DL); | |||
| 21970 | LExtIndex %= NumEltsPerLane; | |||
| 21971 | } | |||
| 21972 | ||||
| 21973 | // add (extractelt (X, 0), extractelt (X, 1)) --> extractelt (hadd X, X), 0 | |||
| 21974 | // add (extractelt (X, 1), extractelt (X, 0)) --> extractelt (hadd X, X), 0 | |||
| 21975 | // add (extractelt (X, 2), extractelt (X, 3)) --> extractelt (hadd X, X), 1 | |||
| 21976 | // sub (extractelt (X, 0), extractelt (X, 1)) --> extractelt (hsub X, X), 0 | |||
| 21977 | SDValue HOp = DAG.getNode(HOpcode, DL, X.getValueType(), X, X); | |||
| 21978 | return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, Op.getSimpleValueType(), HOp, | |||
| 21979 | DAG.getIntPtrConstant(LExtIndex / 2, DL)); | |||
| 21980 | } | |||
| 21981 | ||||
| 21982 | /// Depending on uarch and/or optimizing for size, we might prefer to use a | |||
| 21983 | /// vector operation in place of the typical scalar operation. | |||
| 21984 | SDValue X86TargetLowering::lowerFaddFsub(SDValue Op, SelectionDAG &DAG) const { | |||
| 21985 | assert((Op.getValueType() == MVT::f32 || Op.getValueType() == MVT::f64) &&((void)0) | |||
| 21986 | "Only expecting float/double")((void)0); | |||
| 21987 | return lowerAddSubToHorizontalOp(Op, DAG, Subtarget); | |||
| 21988 | } | |||
| 21989 | ||||
| 21990 | /// ISD::FROUND is defined to round to nearest with ties rounding away from 0. | |||
| 21991 | /// This mode isn't supported in hardware on X86. But as long as we aren't | |||
| 21992 | /// compiling with trapping math, we can emulate this with | |||
| 21993 | /// floor(X + copysign(nextafter(0.5, 0.0), X)). | |||
| 21994 | static SDValue LowerFROUND(SDValue Op, SelectionDAG &DAG) { | |||
| 21995 | SDValue N0 = Op.getOperand(0); | |||
| 21996 | SDLoc dl(Op); | |||
| 21997 | MVT VT = Op.getSimpleValueType(); | |||
| 21998 | ||||
| 21999 | // N0 += copysign(nextafter(0.5, 0.0), N0) | |||
| 22000 | const fltSemantics &Sem = SelectionDAG::EVTToAPFloatSemantics(VT); | |||
| 22001 | bool Ignored; | |||
| 22002 | APFloat Point5Pred = APFloat(0.5f); | |||
| 22003 | Point5Pred.convert(Sem, APFloat::rmNearestTiesToEven, &Ignored); | |||
| 22004 | Point5Pred.next(/*nextDown*/true); | |||
| 22005 | ||||
| 22006 | SDValue Adder = DAG.getNode(ISD::FCOPYSIGN, dl, VT, | |||
| 22007 | DAG.getConstantFP(Point5Pred, dl, VT), N0); | |||
| 22008 | N0 = DAG.getNode(ISD::FADD, dl, VT, N0, Adder); | |||
| 22009 | ||||
| 22010 | // Truncate the result to remove fraction. | |||
| 22011 | return DAG.getNode(ISD::FTRUNC, dl, VT, N0); | |||
| 22012 | } | |||
| 22013 | ||||
| 22014 | /// The only differences between FABS and FNEG are the mask and the logic op. | |||
| 22015 | /// FNEG also has a folding opportunity for FNEG(FABS(x)). | |||
| 22016 | static SDValue LowerFABSorFNEG(SDValue Op, SelectionDAG &DAG) { | |||
| 22017 | assert((Op.getOpcode() == ISD::FABS || Op.getOpcode() == ISD::FNEG) &&((void)0) | |||
| 22018 | "Wrong opcode for lowering FABS or FNEG.")((void)0); | |||
| 22019 | ||||
| 22020 | bool IsFABS = (Op.getOpcode() == ISD::FABS); | |||
| 22021 | ||||
| 22022 | // If this is a FABS and it has an FNEG user, bail out to fold the combination | |||
| 22023 | // into an FNABS. We'll lower the FABS after that if it is still in use. | |||
| 22024 | if (IsFABS) | |||
| 22025 | for (SDNode *User : Op->uses()) | |||
| 22026 | if (User->getOpcode() == ISD::FNEG) | |||
| 22027 | return Op; | |||
| 22028 | ||||
| 22029 | SDLoc dl(Op); | |||
| 22030 | MVT VT = Op.getSimpleValueType(); | |||
| 22031 | ||||
| 22032 | bool IsF128 = (VT == MVT::f128); | |||
| 22033 | assert((VT == MVT::f64 || VT == MVT::f32 || VT == MVT::f128 ||((void)0) | |||
| 22034 | VT == MVT::v2f64 || VT == MVT::v4f64 || VT == MVT::v4f32 ||((void)0) | |||
| 22035 | VT == MVT::v8f32 || VT == MVT::v8f64 || VT == MVT::v16f32) &&((void)0) | |||
| 22036 | "Unexpected type in LowerFABSorFNEG")((void)0); | |||
| 22037 | ||||
| 22038 | // FIXME: Use function attribute "OptimizeForSize" and/or CodeGenOpt::Level to | |||
| 22039 | // decide if we should generate a 16-byte constant mask when we only need 4 or | |||
| 22040 | // 8 bytes for the scalar case. | |||
| 22041 | ||||
| 22042 | // There are no scalar bitwise logical SSE/AVX instructions, so we | |||
| 22043 | // generate a 16-byte vector constant and logic op even for the scalar case. | |||
| 22044 | // Using a 16-byte mask allows folding the load of the mask with | |||
| 22045 | // the logic op, so it can save (~4 bytes) on code size. | |||
| 22046 | bool IsFakeVector = !VT.isVector() && !IsF128; | |||
| 22047 | MVT LogicVT = VT; | |||
| 22048 | if (IsFakeVector) | |||
| 22049 | LogicVT = (VT == MVT::f64) ? MVT::v2f64 : MVT::v4f32; | |||
| 22050 | ||||
| 22051 | unsigned EltBits = VT.getScalarSizeInBits(); | |||
| 22052 | // For FABS, mask is 0x7f...; for FNEG, mask is 0x80... | |||
| 22053 | APInt MaskElt = IsFABS ? APInt::getSignedMaxValue(EltBits) : | |||
| 22054 | APInt::getSignMask(EltBits); | |||
| 22055 | const fltSemantics &Sem = SelectionDAG::EVTToAPFloatSemantics(VT); | |||
| 22056 | SDValue Mask = DAG.getConstantFP(APFloat(Sem, MaskElt), dl, LogicVT); | |||
| 22057 | ||||
| 22058 | SDValue Op0 = Op.getOperand(0); | |||
| 22059 | bool IsFNABS = !IsFABS && (Op0.getOpcode() == ISD::FABS); | |||
| 22060 | unsigned LogicOp = IsFABS ? X86ISD::FAND : | |||
| 22061 | IsFNABS ? X86ISD::FOR : | |||
| 22062 | X86ISD::FXOR; | |||
| 22063 | SDValue Operand = IsFNABS ? Op0.getOperand(0) : Op0; | |||
| 22064 | ||||
| 22065 | if (VT.isVector() || IsF128) | |||
| 22066 | return DAG.getNode(LogicOp, dl, LogicVT, Operand, Mask); | |||
| 22067 | ||||
| 22068 | // For the scalar case extend to a 128-bit vector, perform the logic op, | |||
| 22069 | // and extract the scalar result back out. | |||
| 22070 | Operand = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LogicVT, Operand); | |||
| 22071 | SDValue LogicNode = DAG.getNode(LogicOp, dl, LogicVT, Operand, Mask); | |||
| 22072 | return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, LogicNode, | |||
| 22073 | DAG.getIntPtrConstant(0, dl)); | |||
| 22074 | } | |||
| 22075 | ||||
| 22076 | static SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) { | |||
| 22077 | SDValue Mag = Op.getOperand(0); | |||
| 22078 | SDValue Sign = Op.getOperand(1); | |||
| 22079 | SDLoc dl(Op); | |||
| 22080 | ||||
| 22081 | // If the sign operand is smaller, extend it first. | |||
| 22082 | MVT VT = Op.getSimpleValueType(); | |||
| 22083 | if (Sign.getSimpleValueType().bitsLT(VT)) | |||
| 22084 | Sign = DAG.getNode(ISD::FP_EXTEND, dl, VT, Sign); | |||
| 22085 | ||||
| 22086 | // And if it is bigger, shrink it first. | |||
| 22087 | if (Sign.getSimpleValueType().bitsGT(VT)) | |||
| 22088 | Sign = | |||
| 22089 | DAG.getNode(ISD::FP_ROUND, dl, VT, Sign, DAG.getIntPtrConstant(0, dl)); | |||
| 22090 | ||||
| 22091 | // At this point the operands and the result should have the same | |||
| 22092 | // type, and that won't be f80 since that is not custom lowered. | |||
| 22093 | bool IsF128 = (VT == MVT::f128); | |||
| 22094 | assert((VT == MVT::f64 || VT == MVT::f32 || VT == MVT::f128 ||((void)0) | |||
| 22095 | VT == MVT::v2f64 || VT == MVT::v4f64 || VT == MVT::v4f32 ||((void)0) | |||
| 22096 | VT == MVT::v8f32 || VT == MVT::v8f64 || VT == MVT::v16f32) &&((void)0) | |||
| 22097 | "Unexpected type in LowerFCOPYSIGN")((void)0); | |||
| 22098 | ||||
| 22099 | const fltSemantics &Sem = SelectionDAG::EVTToAPFloatSemantics(VT); | |||
| 22100 | ||||
| 22101 | // Perform all scalar logic operations as 16-byte vectors because there are no | |||
| 22102 | // scalar FP logic instructions in SSE. | |||
| 22103 | // TODO: This isn't necessary. If we used scalar types, we might avoid some | |||
| 22104 | // unnecessary splats, but we might miss load folding opportunities. Should | |||
| 22105 | // this decision be based on OptimizeForSize? | |||
| 22106 | bool IsFakeVector = !VT.isVector() && !IsF128; | |||
| 22107 | MVT LogicVT = VT; | |||
| 22108 | if (IsFakeVector) | |||
| 22109 | LogicVT = (VT == MVT::f64) ? MVT::v2f64 : MVT::v4f32; | |||
| 22110 | ||||
| 22111 | // The mask constants are automatically splatted for vector types. | |||
| 22112 | unsigned EltSizeInBits = VT.getScalarSizeInBits(); | |||
| 22113 | SDValue SignMask = DAG.getConstantFP( | |||
| 22114 | APFloat(Sem, APInt::getSignMask(EltSizeInBits)), dl, LogicVT); | |||
| 22115 | SDValue MagMask = DAG.getConstantFP( | |||
| 22116 | APFloat(Sem, APInt::getSignedMaxValue(EltSizeInBits)), dl, LogicVT); | |||
| 22117 | ||||
| 22118 | // First, clear all bits but the sign bit from the second operand (sign). | |||
| 22119 | if (IsFakeVector) | |||
| 22120 | Sign = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LogicVT, Sign); | |||
| 22121 | SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, LogicVT, Sign, SignMask); | |||
| 22122 | ||||
| 22123 | // Next, clear the sign bit from the first operand (magnitude). | |||
| 22124 | // TODO: If we had general constant folding for FP logic ops, this check | |||
| 22125 | // wouldn't be necessary. | |||
| 22126 | SDValue MagBits; | |||
| 22127 | if (ConstantFPSDNode *Op0CN = isConstOrConstSplatFP(Mag)) { | |||
| 22128 | APFloat APF = Op0CN->getValueAPF(); | |||
| 22129 | APF.clearSign(); | |||
| 22130 | MagBits = DAG.getConstantFP(APF, dl, LogicVT); | |||
| 22131 | } else { | |||
| 22132 | // If the magnitude operand wasn't a constant, we need to AND out the sign. | |||
| 22133 | if (IsFakeVector) | |||
| 22134 | Mag = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LogicVT, Mag); | |||
| 22135 | MagBits = DAG.getNode(X86ISD::FAND, dl, LogicVT, Mag, MagMask); | |||
| 22136 | } | |||
| 22137 | ||||
| 22138 | // OR the magnitude value with the sign bit. | |||
| 22139 | SDValue Or = DAG.getNode(X86ISD::FOR, dl, LogicVT, MagBits, SignBit); | |||
| 22140 | return !IsFakeVector ? Or : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Or, | |||
| 22141 | DAG.getIntPtrConstant(0, dl)); | |||
| 22142 | } | |||
| 22143 | ||||
| 22144 | static SDValue LowerFGETSIGN(SDValue Op, SelectionDAG &DAG) { | |||
| 22145 | SDValue N0 = Op.getOperand(0); | |||
| 22146 | SDLoc dl(Op); | |||
| 22147 | MVT VT = Op.getSimpleValueType(); | |||
| 22148 | ||||
| 22149 | MVT OpVT = N0.getSimpleValueType(); | |||
| 22150 | assert((OpVT == MVT::f32 || OpVT == MVT::f64) &&((void)0) | |||
| 22151 | "Unexpected type for FGETSIGN")((void)0); | |||
| 22152 | ||||
| 22153 | // Lower ISD::FGETSIGN to (AND (X86ISD::MOVMSK ...) 1). | |||
| 22154 | MVT VecVT = (OpVT == MVT::f32 ? MVT::v4f32 : MVT::v2f64); | |||
| 22155 | SDValue Res = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, N0); | |||
| 22156 | Res = DAG.getNode(X86ISD::MOVMSK, dl, MVT::i32, Res); | |||
| 22157 | Res = DAG.getZExtOrTrunc(Res, dl, VT); | |||
| 22158 | Res = DAG.getNode(ISD::AND, dl, VT, Res, DAG.getConstant(1, dl, VT)); | |||
| 22159 | return Res; | |||
| 22160 | } | |||
| 22161 | ||||
| 22162 | /// Helper for creating a X86ISD::SETCC node. | |||
| 22163 | static SDValue getSETCC(X86::CondCode Cond, SDValue EFLAGS, const SDLoc &dl, | |||
| 22164 | SelectionDAG &DAG) { | |||
| 22165 | return DAG.getNode(X86ISD::SETCC, dl, MVT::i8, | |||
| 22166 | DAG.getTargetConstant(Cond, dl, MVT::i8), EFLAGS); | |||
| 22167 | } | |||
| 22168 | ||||
| 22169 | /// Helper for matching OR(EXTRACTELT(X,0),OR(EXTRACTELT(X,1),...)) | |||
| 22170 | /// style scalarized (associative) reduction patterns. Partial reductions | |||
| 22171 | /// are supported when the pointer SrcMask is non-null. | |||
| 22172 | /// TODO - move this to SelectionDAG? | |||
| 22173 | static bool matchScalarReduction(SDValue Op, ISD::NodeType BinOp, | |||
| 22174 | SmallVectorImpl<SDValue> &SrcOps, | |||
| 22175 | SmallVectorImpl<APInt> *SrcMask = nullptr) { | |||
| 22176 | SmallVector<SDValue, 8> Opnds; | |||
| 22177 | DenseMap<SDValue, APInt> SrcOpMap; | |||
| 22178 | EVT VT = MVT::Other; | |||
| 22179 | ||||
| 22180 | // Recognize a special case where a vector is casted into wide integer to | |||
| 22181 | // test all 0s. | |||
| 22182 | assert(Op.getOpcode() == unsigned(BinOp) &&((void)0) | |||
| 22183 | "Unexpected bit reduction opcode")((void)0); | |||
| 22184 | Opnds.push_back(Op.getOperand(0)); | |||
| 22185 | Opnds.push_back(Op.getOperand(1)); | |||
| 22186 | ||||
| 22187 | for (unsigned Slot = 0, e = Opnds.size(); Slot < e; ++Slot) { | |||
| 22188 | SmallVectorImpl<SDValue>::const_iterator I = Opnds.begin() + Slot; | |||
| 22189 | // BFS traverse all BinOp operands. | |||
| 22190 | if (I->getOpcode() == unsigned(BinOp)) { | |||
| 22191 | Opnds.push_back(I->getOperand(0)); | |||
| 22192 | Opnds.push_back(I->getOperand(1)); | |||
| 22193 | // Re-evaluate the number of nodes to be traversed. | |||
| 22194 | e += 2; // 2 more nodes (LHS and RHS) are pushed. | |||
| 22195 | continue; | |||
| 22196 | } | |||
| 22197 | ||||
| 22198 | // Quit if a non-EXTRACT_VECTOR_ELT | |||
| 22199 | if (I->getOpcode() != ISD::EXTRACT_VECTOR_ELT) | |||
| 22200 | return false; | |||
| 22201 | ||||
| 22202 | // Quit if without a constant index. | |||
| 22203 | auto *Idx = dyn_cast<ConstantSDNode>(I->getOperand(1)); | |||
| 22204 | if (!Idx) | |||
| 22205 | return false; | |||
| 22206 | ||||
| 22207 | SDValue Src = I->getOperand(0); | |||
| 22208 | DenseMap<SDValue, APInt>::iterator M = SrcOpMap.find(Src); | |||
| 22209 | if (M == SrcOpMap.end()) { | |||
| 22210 | VT = Src.getValueType(); | |||
| 22211 | // Quit if not the same type. | |||
| 22212 | if (!SrcOpMap.empty() && VT != SrcOpMap.begin()->first.getValueType()) | |||
| 22213 | return false; | |||
| 22214 | unsigned NumElts = VT.getVectorNumElements(); | |||
| 22215 | APInt EltCount = APInt::getNullValue(NumElts); | |||
| 22216 | M = SrcOpMap.insert(std::make_pair(Src, EltCount)).first; | |||
| 22217 | SrcOps.push_back(Src); | |||
| 22218 | } | |||
| 22219 | ||||
| 22220 | // Quit if element already used. | |||
| 22221 | unsigned CIdx = Idx->getZExtValue(); | |||
| 22222 | if (M->second[CIdx]) | |||
| 22223 | return false; | |||
| 22224 | M->second.setBit(CIdx); | |||
| 22225 | } | |||
| 22226 | ||||
| 22227 | if (SrcMask) { | |||
| 22228 | // Collect the source partial masks. | |||
| 22229 | for (SDValue &SrcOp : SrcOps) | |||
| 22230 | SrcMask->push_back(SrcOpMap[SrcOp]); | |||
| 22231 | } else { | |||
| 22232 | // Quit if not all elements are used. | |||
| 22233 | for (const auto &I : SrcOpMap) | |||
| 22234 | if (!I.second.isAllOnesValue()) | |||
| 22235 | return false; | |||
| 22236 | } | |||
| 22237 | ||||
| 22238 | return true; | |||
| 22239 | } | |||
| 22240 | ||||
| 22241 | // Helper function for comparing all bits of a vector against zero. | |||
| 22242 | static SDValue LowerVectorAllZero(const SDLoc &DL, SDValue V, ISD::CondCode CC, | |||
| 22243 | const APInt &Mask, | |||
| 22244 | const X86Subtarget &Subtarget, | |||
| 22245 | SelectionDAG &DAG, X86::CondCode &X86CC) { | |||
| 22246 | EVT VT = V.getValueType(); | |||
| 22247 | unsigned ScalarSize = VT.getScalarSizeInBits(); | |||
| 22248 | if (Mask.getBitWidth() != ScalarSize) { | |||
| 22249 | assert(ScalarSize == 1 && "Element Mask vs Vector bitwidth mismatch")((void)0); | |||
| 22250 | return SDValue(); | |||
| 22251 | } | |||
| 22252 | ||||
| 22253 | assert((CC == ISD::SETEQ || CC == ISD::SETNE) && "Unsupported ISD::CondCode")((void)0); | |||
| 22254 | X86CC = (CC == ISD::SETEQ ? X86::COND_E : X86::COND_NE); | |||
| 22255 | ||||
| 22256 | auto MaskBits = [&](SDValue Src) { | |||
| 22257 | if (Mask.isAllOnesValue()) | |||
| 22258 | return Src; | |||
| 22259 | EVT SrcVT = Src.getValueType(); | |||
| 22260 | SDValue MaskValue = DAG.getConstant(Mask, DL, SrcVT); | |||
| 22261 | return DAG.getNode(ISD::AND, DL, SrcVT, Src, MaskValue); | |||
| 22262 | }; | |||
| 22263 | ||||
| 22264 | // For sub-128-bit vector, cast to (legal) integer and compare with zero. | |||
| 22265 | if (VT.getSizeInBits() < 128) { | |||
| 22266 | EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), VT.getSizeInBits()); | |||
| 22267 | if (!DAG.getTargetLoweringInfo().isTypeLegal(IntVT)) | |||
| 22268 | return SDValue(); | |||
| 22269 | return DAG.getNode(X86ISD::CMP, DL, MVT::i32, | |||
| 22270 | DAG.getBitcast(IntVT, MaskBits(V)), | |||
| 22271 | DAG.getConstant(0, DL, IntVT)); | |||
| 22272 | } | |||
| 22273 | ||||
| 22274 | // Quit if not splittable to 128/256-bit vector. | |||
| 22275 | if (!isPowerOf2_32(VT.getSizeInBits())) | |||
| 22276 | return SDValue(); | |||
| 22277 | ||||
| 22278 | // Split down to 128/256-bit vector. | |||
| 22279 | unsigned TestSize = Subtarget.hasAVX() ? 256 : 128; | |||
| 22280 | while (VT.getSizeInBits() > TestSize) { | |||
| 22281 | auto Split = DAG.SplitVector(V, DL); | |||
| 22282 | VT = Split.first.getValueType(); | |||
| 22283 | V = DAG.getNode(ISD::OR, DL, VT, Split.first, Split.second); | |||
| 22284 | } | |||
| 22285 | ||||
| 22286 | bool UsePTEST = Subtarget.hasSSE41(); | |||
| 22287 | if (UsePTEST) { | |||
| 22288 | MVT TestVT = VT.is128BitVector() ? MVT::v2i64 : MVT::v4i64; | |||
| 22289 | V = DAG.getBitcast(TestVT, MaskBits(V)); | |||
| 22290 | return DAG.getNode(X86ISD::PTEST, DL, MVT::i32, V, V); | |||
| 22291 | } | |||
| 22292 | ||||
| 22293 | // Without PTEST, a masked v2i64 or-reduction is not faster than | |||
| 22294 | // scalarization. | |||
| 22295 | if (!Mask.isAllOnesValue() && VT.getScalarSizeInBits() > 32) | |||
| 22296 | return SDValue(); | |||
| 22297 | ||||
| 22298 | V = DAG.getBitcast(MVT::v16i8, MaskBits(V)); | |||
| 22299 | V = DAG.getNode(X86ISD::PCMPEQ, DL, MVT::v16i8, V, | |||
| 22300 | getZeroVector(MVT::v16i8, Subtarget, DAG, DL)); | |||
| 22301 | V = DAG.getNode(X86ISD::MOVMSK, DL, MVT::i32, V); | |||
| 22302 | return DAG.getNode(X86ISD::CMP, DL, MVT::i32, V, | |||
| 22303 | DAG.getConstant(0xFFFF, DL, MVT::i32)); | |||
| 22304 | } | |||
| 22305 | ||||
| 22306 | // Check whether an OR'd reduction tree is PTEST-able, or if we can fallback to | |||
| 22307 | // CMP(MOVMSK(PCMPEQB(X,0))). | |||
| 22308 | static SDValue MatchVectorAllZeroTest(SDValue Op, ISD::CondCode CC, | |||
| 22309 | const SDLoc &DL, | |||
| 22310 | const X86Subtarget &Subtarget, | |||
| 22311 | SelectionDAG &DAG, SDValue &X86CC) { | |||
| 22312 | assert((CC == ISD::SETEQ || CC == ISD::SETNE) && "Unsupported ISD::CondCode")((void)0); | |||
| 22313 | ||||
| 22314 | if (!Subtarget.hasSSE2() || !Op->hasOneUse()) | |||
| 22315 | return SDValue(); | |||
| 22316 | ||||
| 22317 | // Check whether we're masking/truncating an OR-reduction result, in which | |||
| 22318 | // case track the masked bits. | |||
| 22319 | APInt Mask = APInt::getAllOnesValue(Op.getScalarValueSizeInBits()); | |||
| 22320 | switch (Op.getOpcode()) { | |||
| 22321 | case ISD::TRUNCATE: { | |||
| 22322 | SDValue Src = Op.getOperand(0); | |||
| 22323 | Mask = APInt::getLowBitsSet(Src.getScalarValueSizeInBits(), | |||
| 22324 | Op.getScalarValueSizeInBits()); | |||
| 22325 | Op = Src; | |||
| 22326 | break; | |||
| 22327 | } | |||
| 22328 | case ISD::AND: { | |||
| 22329 | if (auto *Cst = dyn_cast<ConstantSDNode>(Op.getOperand(1))) { | |||
| 22330 | Mask = Cst->getAPIntValue(); | |||
| 22331 | Op = Op.getOperand(0); | |||
| 22332 | } | |||
| 22333 | break; | |||
| 22334 | } | |||
| 22335 | } | |||
| 22336 | ||||
| 22337 | SmallVector<SDValue, 8> VecIns; | |||
| 22338 | if (Op.getOpcode() == ISD::OR && matchScalarReduction(Op, ISD::OR, VecIns)) { | |||
| 22339 | EVT VT = VecIns[0].getValueType(); | |||
| 22340 | assert(llvm::all_of(VecIns,((void)0) | |||
| 22341 | [VT](SDValue V) { return VT == V.getValueType(); }) &&((void)0) | |||
| 22342 | "Reduction source vector mismatch")((void)0); | |||
| 22343 | ||||
| 22344 | // Quit if less than 128-bits or not splittable to 128/256-bit vector. | |||
| 22345 | if (VT.getSizeInBits() < 128 || !isPowerOf2_32(VT.getSizeInBits())) | |||
| 22346 | return SDValue(); | |||
| 22347 | ||||
| 22348 | // If more than one full vector is evaluated, OR them first before PTEST. | |||
| 22349 | for (unsigned Slot = 0, e = VecIns.size(); e - Slot > 1; | |||
| 22350 | Slot += 2, e += 1) { | |||
| 22351 | // Each iteration will OR 2 nodes and append the result until there is | |||
| 22352 | // only 1 node left, i.e. the final OR'd value of all vectors. | |||
| 22353 | SDValue LHS = VecIns[Slot]; | |||
| 22354 | SDValue RHS = VecIns[Slot + 1]; | |||
| 22355 | VecIns.push_back(DAG.getNode(ISD::OR, DL, VT, LHS, RHS)); | |||
| 22356 | } | |||
| 22357 | ||||
| 22358 | X86::CondCode CCode; | |||
| 22359 | if (SDValue V = LowerVectorAllZero(DL, VecIns.back(), CC, Mask, Subtarget, | |||
| 22360 | DAG, CCode)) { | |||
| 22361 | X86CC = DAG.getTargetConstant(CCode, DL, MVT::i8); | |||
| 22362 | return V; | |||
| 22363 | } | |||
| 22364 | } | |||
| 22365 | ||||
| 22366 | if (Op.getOpcode() == ISD::EXTRACT_VECTOR_ELT) { | |||
| 22367 | ISD::NodeType BinOp; | |||
| 22368 | if (SDValue Match = | |||
| 22369 | DAG.matchBinOpReduction(Op.getNode(), BinOp, {ISD::OR})) { | |||
| 22370 | X86::CondCode CCode; | |||
| 22371 | if (SDValue V = | |||
| 22372 | LowerVectorAllZero(DL, Match, CC, Mask, Subtarget, DAG, CCode)) { | |||
| 22373 | X86CC = DAG.getTargetConstant(CCode, DL, MVT::i8); | |||
| 22374 | return V; | |||
| 22375 | } | |||
| 22376 | } | |||
| 22377 | } | |||
| 22378 | ||||
| 22379 | return SDValue(); | |||
| 22380 | } | |||
| 22381 | ||||
| 22382 | /// return true if \c Op has a use that doesn't just read flags. | |||
| 22383 | static bool hasNonFlagsUse(SDValue Op) { | |||
| 22384 | for (SDNode::use_iterator UI = Op->use_begin(), UE = Op->use_end(); UI != UE; | |||
| 22385 | ++UI) { | |||
| 22386 | SDNode *User = *UI; | |||
| 22387 | unsigned UOpNo = UI.getOperandNo(); | |||
| 22388 | if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) { | |||
| 22389 | // Look pass truncate. | |||
| 22390 | UOpNo = User->use_begin().getOperandNo(); | |||
| 22391 | User = *User->use_begin(); | |||
| 22392 | } | |||
| 22393 | ||||
| 22394 | if (User->getOpcode() != ISD::BRCOND && User->getOpcode() != ISD::SETCC && | |||
| 22395 | !(User->getOpcode() == ISD::SELECT && UOpNo == 0)) | |||
| 22396 | return true; | |||
| 22397 | } | |||
| 22398 | return false; | |||
| 22399 | } | |||
| 22400 | ||||
| 22401 | // Transform to an x86-specific ALU node with flags if there is a chance of | |||
| 22402 | // using an RMW op or only the flags are used. Otherwise, leave | |||
| 22403 | // the node alone and emit a 'cmp' or 'test' instruction. | |||
| 22404 | static bool isProfitableToUseFlagOp(SDValue Op) { | |||
| 22405 | for (SDNode *U : Op->uses()) | |||
| 22406 | if (U->getOpcode() != ISD::CopyToReg && | |||
| 22407 | U->getOpcode() != ISD::SETCC && | |||
| 22408 | U->getOpcode() != ISD::STORE) | |||
| 22409 | return false; | |||
| 22410 | ||||
| 22411 | return true; | |||
| 22412 | } | |||
| 22413 | ||||
| 22414 | /// Emit nodes that will be selected as "test Op0,Op0", or something | |||
| 22415 | /// equivalent. | |||
| 22416 | static SDValue EmitTest(SDValue Op, unsigned X86CC, const SDLoc &dl, | |||
| 22417 | SelectionDAG &DAG, const X86Subtarget &Subtarget) { | |||
| 22418 | // CF and OF aren't always set the way we want. Determine which | |||
| 22419 | // of these we need. | |||
| 22420 | bool NeedCF = false; | |||
| 22421 | bool NeedOF = false; | |||
| 22422 | switch (X86CC) { | |||
| 22423 | default: break; | |||
| 22424 | case X86::COND_A: case X86::COND_AE: | |||
| 22425 | case X86::COND_B: case X86::COND_BE: | |||
| 22426 | NeedCF = true; | |||
| 22427 | break; | |||
| 22428 | case X86::COND_G: case X86::COND_GE: | |||
| 22429 | case X86::COND_L: case X86::COND_LE: | |||
| 22430 | case X86::COND_O: case X86::COND_NO: { | |||
| 22431 | // Check if we really need to set the | |||
| 22432 | // Overflow flag. If NoSignedWrap is present | |||
| 22433 | // that is not actually needed. | |||
| 22434 | switch (Op->getOpcode()) { | |||
| 22435 | case ISD::ADD: | |||
| 22436 | case ISD::SUB: | |||
| 22437 | case ISD::MUL: | |||
| 22438 | case ISD::SHL: | |||
| 22439 | if (Op.getNode()->getFlags().hasNoSignedWrap()) | |||
| 22440 | break; | |||
| 22441 | LLVM_FALLTHROUGH[[gnu::fallthrough]]; | |||
| 22442 | default: | |||
| 22443 | NeedOF = true; | |||
| 22444 | break; | |||
| 22445 | } | |||
| 22446 | break; | |||
| 22447 | } | |||
| 22448 | } | |||
| 22449 | // See if we can use the EFLAGS value from the operand instead of | |||
| 22450 | // doing a separate TEST. TEST always sets OF and CF to 0, so unless | |||
| 22451 | // we prove that the arithmetic won't overflow, we can't use OF or CF. | |||
| 22452 | if (Op.getResNo() != 0 || NeedOF || NeedCF) { | |||
| 22453 | // Emit a CMP with 0, which is the TEST pattern. | |||
| 22454 | return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op, | |||
| 22455 | DAG.getConstant(0, dl, Op.getValueType())); | |||
| 22456 | } | |||
| 22457 | unsigned Opcode = 0; | |||
| 22458 | unsigned NumOperands = 0; | |||
| 22459 | ||||
| 22460 | SDValue ArithOp = Op; | |||
| 22461 | ||||
| 22462 | // NOTICE: In the code below we use ArithOp to hold the arithmetic operation | |||
| 22463 | // which may be the result of a CAST. We use the variable 'Op', which is the | |||
| 22464 | // non-casted variable when we check for possible users. | |||
| 22465 | switch (ArithOp.getOpcode()) { | |||
| 22466 | case ISD::AND: | |||
| 22467 | // If the primary 'and' result isn't used, don't bother using X86ISD::AND, | |||
| 22468 | // because a TEST instruction will be better. | |||
| 22469 | if (!hasNonFlagsUse(Op)) | |||
| 22470 | break; | |||
| 22471 | ||||
| 22472 | LLVM_FALLTHROUGH[[gnu::fallthrough]]; | |||
| 22473 | case ISD::ADD: | |||
| 22474 | case ISD::SUB: | |||
| 22475 | case ISD::OR: | |||
| 22476 | case ISD::XOR: | |||
| 22477 | if (!isProfitableToUseFlagOp(Op)) | |||
| 22478 | break; | |||
| 22479 | ||||
| 22480 | // Otherwise use a regular EFLAGS-setting instruction. | |||
| 22481 | switch (ArithOp.getOpcode()) { | |||
| 22482 | default: llvm_unreachable("unexpected operator!")__builtin_unreachable(); | |||
| 22483 | case ISD::ADD: Opcode = X86ISD::ADD; break; | |||
| 22484 | case ISD::SUB: Opcode = X86ISD::SUB; break; | |||
| 22485 | case ISD::XOR: Opcode = X86ISD::XOR; break; | |||
| 22486 | case ISD::AND: Opcode = X86ISD::AND; break; | |||
| 22487 | case ISD::OR: Opcode = X86ISD::OR; break; | |||
| 22488 | } | |||
| 22489 | ||||
| 22490 | NumOperands = 2; | |||
| 22491 | break; | |||
| 22492 | case X86ISD::ADD: | |||
| 22493 | case X86ISD::SUB: | |||
| 22494 | case X86ISD::OR: | |||
| 22495 | case X86ISD::XOR: | |||
| 22496 | case X86ISD::AND: | |||
| 22497 | return SDValue(Op.getNode(), 1); | |||
| 22498 | case ISD::SSUBO: | |||
| 22499 | case ISD::USUBO: { | |||
| 22500 | // /USUBO/SSUBO will become a X86ISD::SUB and we can use its Z flag. | |||
| 22501 | SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32); | |||
| 22502 | return DAG.getNode(X86ISD::SUB, dl, VTs, Op->getOperand(0), | |||
| 22503 | Op->getOperand(1)).getValue(1); | |||
| 22504 | } | |||
| 22505 | default: | |||
| 22506 | break; | |||
| 22507 | } | |||
| 22508 | ||||
| 22509 | if (Opcode == 0) { | |||
| 22510 | // Emit a CMP with 0, which is the TEST pattern. | |||
| 22511 | return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op, | |||
| 22512 | DAG.getConstant(0, dl, Op.getValueType())); | |||
| 22513 | } | |||
| 22514 | SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32); | |||
| 22515 | SmallVector<SDValue, 4> Ops(Op->op_begin(), Op->op_begin() + NumOperands); | |||
| 22516 | ||||
| 22517 | SDValue New = DAG.getNode(Opcode, dl, VTs, Ops); | |||
| 22518 | DAG.ReplaceAllUsesOfValueWith(SDValue(Op.getNode(), 0), New); | |||
| 22519 | return SDValue(New.getNode(), 1); | |||
| 22520 | } | |||
| 22521 | ||||
| 22522 | /// Emit nodes that will be selected as "cmp Op0,Op1", or something | |||
| 22523 | /// equivalent. | |||
| 22524 | static SDValue EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC, | |||
| 22525 | const SDLoc &dl, SelectionDAG &DAG, | |||
| 22526 | const X86Subtarget &Subtarget) { | |||
| 22527 | if (isNullConstant(Op1)) | |||
| 22528 | return EmitTest(Op0, X86CC, dl, DAG, Subtarget); | |||
| 22529 | ||||
| 22530 | EVT CmpVT = Op0.getValueType(); | |||
| 22531 | ||||
| 22532 | assert((CmpVT == MVT::i8 || CmpVT == MVT::i16 ||((void)0) | |||
| 22533 | CmpVT == MVT::i32 || CmpVT == MVT::i64) && "Unexpected VT!")((void)0); | |||
| 22534 | ||||
| 22535 | // Only promote the compare up to I32 if it is a 16 bit operation | |||
| 22536 | // with an immediate. 16 bit immediates are to be avoided. | |||
| 22537 | if (CmpVT == MVT::i16 && !Subtarget.isAtom() && | |||
| 22538 | !DAG.getMachineFunction().getFunction().hasMinSize()) { | |||
| 22539 | ConstantSDNode *COp0 = dyn_cast<ConstantSDNode>(Op0); | |||
| 22540 | ConstantSDNode *COp1 = dyn_cast<ConstantSDNode>(Op1); | |||
| 22541 | // Don't do this if the immediate can fit in 8-bits. | |||
| 22542 | if ((COp0 && !COp0->getAPIntValue().isSignedIntN(8)) || | |||
| 22543 | (COp1 && !COp1->getAPIntValue().isSignedIntN(8))) { | |||
| 22544 | unsigned ExtendOp = | |||
| 22545 | isX86CCSigned(X86CC) ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND; | |||
| 22546 | if (X86CC == X86::COND_E || X86CC == X86::COND_NE) { | |||
| 22547 | // For equality comparisons try to use SIGN_EXTEND if the input was | |||
| 22548 | // truncate from something with enough sign bits. | |||
| 22549 | if (Op0.getOpcode() == ISD::TRUNCATE) { | |||
| 22550 | SDValue In = Op0.getOperand(0); | |||
| 22551 | unsigned EffBits = | |||
| 22552 | In.getScalarValueSizeInBits() - DAG.ComputeNumSignBits(In) + 1; | |||
| 22553 | if (EffBits <= 16) | |||
| 22554 | ExtendOp = ISD::SIGN_EXTEND; | |||
| 22555 | } else if (Op1.getOpcode() == ISD::TRUNCATE) { | |||
| 22556 | SDValue In = Op1.getOperand(0); | |||
| 22557 | unsigned EffBits = | |||
| 22558 | In.getScalarValueSizeInBits() - DAG.ComputeNumSignBits(In) + 1; | |||
| 22559 | if (EffBits <= 16) | |||
| 22560 | ExtendOp = ISD::SIGN_EXTEND; | |||
| 22561 | } | |||
| 22562 | } | |||
| 22563 | ||||
| 22564 | CmpVT = MVT::i32; | |||
| 22565 | Op0 = DAG.getNode(ExtendOp, dl, CmpVT, Op0); | |||
| 22566 | Op1 = DAG.getNode(ExtendOp, dl, CmpVT, Op1); | |||
| 22567 | } | |||
| 22568 | } | |||
| 22569 | ||||
| 22570 | // Try to shrink i64 compares if the input has enough zero bits. | |||
| 22571 | // FIXME: Do this for non-constant compares for constant on LHS? | |||
| 22572 | if (CmpVT == MVT::i64 && isa<ConstantSDNode>(Op1) && !isX86CCSigned(X86CC) && | |||
| 22573 | Op0.hasOneUse() && // Hacky way to not break CSE opportunities with sub. | |||
| 22574 | cast<ConstantSDNode>(Op1)->getAPIntValue().getActiveBits() <= 32 && | |||
| 22575 | DAG.MaskedValueIsZero(Op0, APInt::getHighBitsSet(64, 32))) { | |||
| 22576 | CmpVT = MVT::i32; | |||
| 22577 | Op0 = DAG.getNode(ISD::TRUNCATE, dl, CmpVT, Op0); | |||
| 22578 | Op1 = DAG.getNode(ISD::TRUNCATE, dl, CmpVT, Op1); | |||
| 22579 | } | |||
| 22580 | ||||
| 22581 | // 0-x == y --> x+y == 0 | |||
| 22582 | // 0-x != y --> x+y != 0 | |||
| 22583 | if (Op0.getOpcode() == ISD::SUB && isNullConstant(Op0.getOperand(0)) && | |||
| 22584 | Op0.hasOneUse() && (X86CC == X86::COND_E || X86CC == X86::COND_NE)) { | |||
| 22585 | SDVTList VTs = DAG.getVTList(CmpVT, MVT::i32); | |||
| 22586 | SDValue Add = DAG.getNode(X86ISD::ADD, dl, VTs, Op0.getOperand(1), Op1); | |||
| 22587 | return Add.getValue(1); | |||
| 22588 | } | |||
| 22589 | ||||
| 22590 | // x == 0-y --> x+y == 0 | |||
| 22591 | // x != 0-y --> x+y != 0 | |||
| 22592 | if (Op1.getOpcode() == ISD::SUB && isNullConstant(Op1.getOperand(0)) && | |||
| 22593 | Op1.hasOneUse() && (X86CC == X86::COND_E || X86CC == X86::COND_NE)) { | |||
| 22594 | SDVTList VTs = DAG.getVTList(CmpVT, MVT::i32); | |||
| 22595 | SDValue Add = DAG.getNode(X86ISD::ADD, dl, VTs, Op0, Op1.getOperand(1)); | |||
| 22596 | return Add.getValue(1); | |||
| 22597 | } | |||
| 22598 | ||||
| 22599 | // Use SUB instead of CMP to enable CSE between SUB and CMP. | |||
| 22600 | SDVTList VTs = DAG.getVTList(CmpVT, MVT::i32); | |||
| 22601 | SDValue Sub = DAG.getNode(X86ISD::SUB, dl, VTs, Op0, Op1); | |||
| 22602 | return Sub.getValue(1); | |||
| 22603 | } | |||
| 22604 | ||||
| 22605 | /// Check if replacement of SQRT with RSQRT should be disabled. | |||
| 22606 | bool X86TargetLowering::isFsqrtCheap(SDValue Op, SelectionDAG &DAG) const { | |||
| 22607 | EVT VT = Op.getValueType(); | |||
| 22608 | ||||
| 22609 | // We never want to use both SQRT and RSQRT instructions for the same input. | |||
| 22610 | if (DAG.getNodeIfExists(X86ISD::FRSQRT, DAG.getVTList(VT), Op)) | |||
| 22611 | return false; | |||
| 22612 | ||||
| 22613 | if (VT.isVector()) | |||
| 22614 | return Subtarget.hasFastVectorFSQRT(); | |||
| 22615 | return Subtarget.hasFastScalarFSQRT(); | |||
| 22616 | } | |||
| 22617 | ||||
| 22618 | /// The minimum architected relative accuracy is 2^-12. We need one | |||
| 22619 | /// Newton-Raphson step to have a good float result (24 bits of precision). | |||
| 22620 | SDValue X86TargetLowering::getSqrtEstimate(SDValue Op, | |||
| 22621 | SelectionDAG &DAG, int Enabled, | |||
| 22622 | int &RefinementSteps, | |||
| 22623 | bool &UseOneConstNR, | |||
| 22624 | bool Reciprocal) const { | |||
| 22625 | EVT VT = Op.getValueType(); | |||
| 22626 | ||||
| 22627 | // SSE1 has rsqrtss and rsqrtps. AVX adds a 256-bit variant for rsqrtps. | |||
| 22628 | // It is likely not profitable to do this for f64 because a double-precision | |||
| 22629 | // rsqrt estimate with refinement on x86 prior to FMA requires at least 16 | |||
| 22630 | // instructions: convert to single, rsqrtss, convert back to double, refine | |||
| 22631 | // (3 steps = at least 13 insts). If an 'rsqrtsd' variant was added to the ISA | |||
| 22632 | // along with FMA, this could be a throughput win. | |||
| 22633 | // TODO: SQRT requires SSE2 to prevent the introduction of an illegal v4i32 | |||
| 22634 | // after legalize types. | |||
| 22635 | if ((VT == MVT::f32 && Subtarget.hasSSE1()) || | |||
| 22636 | (VT == MVT::v4f32 && Subtarget.hasSSE1() && Reciprocal) || | |||
| 22637 | (VT == MVT::v4f32 && Subtarget.hasSSE2() && !Reciprocal) || | |||
| 22638 | (VT == MVT::v8f32 && Subtarget.hasAVX()) || | |||
| 22639 | (VT == MVT::v16f32 && Subtarget.useAVX512Regs())) { | |||
| 22640 | if (RefinementSteps == ReciprocalEstimate::Unspecified) | |||
| 22641 | RefinementSteps = 1; | |||
| 22642 | ||||
| 22643 | UseOneConstNR = false; | |||
| 22644 | // There is no FSQRT for 512-bits, but there is RSQRT14. | |||
| 22645 | unsigned Opcode = VT == MVT::v16f32 ? X86ISD::RSQRT14 : X86ISD::FRSQRT; | |||
| 22646 | return DAG.getNode(Opcode, SDLoc(Op), VT, Op); | |||
| 22647 | } | |||
| 22648 | return SDValue(); | |||
| 22649 | } | |||
| 22650 | ||||
| 22651 | /// The minimum architected relative accuracy is 2^-12. We need one | |||
| 22652 | /// Newton-Raphson step to have a good float result (24 bits of precision). | |||
| 22653 | SDValue X86TargetLowering::getRecipEstimate(SDValue Op, SelectionDAG &DAG, | |||
| 22654 | int Enabled, | |||
| 22655 | int &RefinementSteps) const { | |||
| 22656 | EVT VT = Op.getValueType(); | |||
| 22657 | ||||
| 22658 | // SSE1 has rcpss and rcpps. AVX adds a 256-bit variant for rcpps. | |||
| 22659 | // It is likely not profitable to do this for f64 because a double-precision | |||
| 22660 | // reciprocal estimate with refinement on x86 prior to FMA requires | |||
| 22661 | // 15 instructions: convert to single, rcpss, convert back to double, refine | |||
| 22662 | // (3 steps = 12 insts). If an 'rcpsd' variant was added to the ISA | |||
| 22663 | // along with FMA, this could be a throughput win. | |||
| 22664 | ||||
| 22665 | if ((VT == MVT::f32 && Subtarget.hasSSE1()) || | |||
| 22666 | (VT == MVT::v4f32 && Subtarget.hasSSE1()) || | |||
| 22667 | (VT == MVT::v8f32 && Subtarget.hasAVX()) || | |||
| 22668 | (VT == MVT::v16f32 && Subtarget.useAVX512Regs())) { | |||
| 22669 | // Enable estimate codegen with 1 refinement step for vector division. | |||
| 22670 | // Scalar division estimates are disabled because they break too much | |||
| 22671 | // real-world code. These defaults are intended to match GCC behavior. | |||
| 22672 | if (VT == MVT::f32 && Enabled == ReciprocalEstimate::Unspecified) | |||
| 22673 | return SDValue(); | |||
| 22674 | ||||
| 22675 | if (RefinementSteps == ReciprocalEstimate::Unspecified) | |||
| 22676 | RefinementSteps = 1; | |||
| 22677 | ||||
| 22678 | // There is no FSQRT for 512-bits, but there is RCP14. | |||
| 22679 | unsigned Opcode = VT == MVT::v16f32 ? X86ISD::RCP14 : X86ISD::FRCP; | |||
| 22680 | return DAG.getNode(Opcode, SDLoc(Op), VT, Op); | |||
| 22681 | } | |||
| 22682 | return SDValue(); | |||
| 22683 | } | |||
| 22684 | ||||
| 22685 | /// If we have at least two divisions that use the same divisor, convert to | |||
| 22686 | /// multiplication by a reciprocal. This may need to be adjusted for a given | |||
| 22687 | /// CPU if a division's cost is not at least twice the cost of a multiplication. | |||
| 22688 | /// This is because we still need one division to calculate the reciprocal and | |||
| 22689 | /// then we need two multiplies by that reciprocal as replacements for the | |||
| 22690 | /// original divisions. | |||
| 22691 | unsigned X86TargetLowering::combineRepeatedFPDivisors() const { | |||
| 22692 | return 2; | |||
| 22693 | } | |||
| 22694 | ||||
| 22695 | SDValue | |||
| 22696 | X86TargetLowering::BuildSDIVPow2(SDNode *N, const APInt &Divisor, | |||
| 22697 | SelectionDAG &DAG, | |||
| 22698 | SmallVectorImpl<SDNode *> &Created) const { | |||
| 22699 | AttributeList Attr = DAG.getMachineFunction().getFunction().getAttributes(); | |||
| 22700 | if (isIntDivCheap(N->getValueType(0), Attr)) | |||
| 22701 | return SDValue(N,0); // Lower SDIV as SDIV | |||
| 22702 | ||||
| 22703 | assert((Divisor.isPowerOf2() || (-Divisor).isPowerOf2()) &&((void)0) | |||
| 22704 | "Unexpected divisor!")((void)0); | |||
| 22705 | ||||
| 22706 | // Only perform this transform if CMOV is supported otherwise the select | |||
| 22707 | // below will become a branch. | |||
| 22708 | if (!Subtarget.hasCMov()) | |||
| 22709 | return SDValue(); | |||
| 22710 | ||||
| 22711 | // fold (sdiv X, pow2) | |||
| 22712 | EVT VT = N->getValueType(0); | |||
| 22713 | // FIXME: Support i8. | |||
| 22714 | if (VT != MVT::i16 && VT != MVT::i32 && | |||
| 22715 | !(Subtarget.is64Bit() && VT == MVT::i64)) | |||
| 22716 | return SDValue(); | |||
| 22717 | ||||
| 22718 | unsigned Lg2 = Divisor.countTrailingZeros(); | |||
| 22719 | ||||
| 22720 | // If the divisor is 2 or -2, the default expansion is better. | |||
| 22721 | if (Lg2 == 1) | |||
| 22722 | return SDValue(); | |||
| 22723 | ||||
| 22724 | SDLoc DL(N); | |||
| 22725 | SDValue N0 = N->getOperand(0); | |||
| 22726 | SDValue Zero = DAG.getConstant(0, DL, VT); | |||
| 22727 | APInt Lg2Mask = APInt::getLowBitsSet(VT.getSizeInBits(), Lg2); | |||
| 22728 | SDValue Pow2MinusOne = DAG.getConstant(Lg2Mask, DL, VT); | |||
| 22729 | ||||
| 22730 | // If N0 is negative, we need to add (Pow2 - 1) to it before shifting right. | |||
| 22731 | SDValue Cmp = DAG.getSetCC(DL, MVT::i8, N0, Zero, ISD::SETLT); | |||
| 22732 | SDValue Add = DAG.getNode(ISD::ADD, DL, VT, N0, Pow2MinusOne); | |||
| 22733 | SDValue CMov = DAG.getNode(ISD::SELECT, DL, VT, Cmp, Add, N0); | |||
| 22734 | ||||
| 22735 | Created.push_back(Cmp.getNode()); | |||
| 22736 | Created.push_back(Add.getNode()); | |||
| 22737 | Created.push_back(CMov.getNode()); | |||
| 22738 | ||||
| 22739 | // Divide by pow2. | |||
| 22740 | SDValue SRA = | |||
| 22741 | DAG.getNode(ISD::SRA, DL, VT, CMov, DAG.getConstant(Lg2, DL, MVT::i8)); | |||
| 22742 | ||||
| 22743 | // If we're dividing by a positive value, we're done. Otherwise, we must | |||
| 22744 | // negate the result. | |||
| 22745 | if (Divisor.isNonNegative()) | |||
| 22746 | return SRA; | |||
| 22747 | ||||
| 22748 | Created.push_back(SRA.getNode()); | |||
| 22749 | return DAG.getNode(ISD::SUB, DL, VT, Zero, SRA); | |||
| 22750 | } | |||
| 22751 | ||||
| 22752 | /// Result of 'and' is compared against zero. Change to a BT node if possible. | |||
| 22753 | /// Returns the BT node and the condition code needed to use it. | |||
| 22754 | static SDValue LowerAndToBT(SDValue And, ISD::CondCode CC, | |||
| 22755 | const SDLoc &dl, SelectionDAG &DAG, | |||
| 22756 | SDValue &X86CC) { | |||
| 22757 | assert(And.getOpcode() == ISD::AND && "Expected AND node!")((void)0); | |||
| 22758 | SDValue Op0 = And.getOperand(0); | |||
| 22759 | SDValue Op1 = And.getOperand(1); | |||
| 22760 | if (Op0.getOpcode() == ISD::TRUNCATE) | |||
| 22761 | Op0 = Op0.getOperand(0); | |||
| 22762 | if (Op1.getOpcode() == ISD::TRUNCATE) | |||
| 22763 | Op1 = Op1.getOperand(0); | |||
| 22764 | ||||
| 22765 | SDValue Src, BitNo; | |||
| 22766 | if (Op1.getOpcode() == ISD::SHL) | |||
| 22767 | std::swap(Op0, Op1); | |||
| 22768 | if (Op0.getOpcode() == ISD::SHL) { | |||
| 22769 | if (isOneConstant(Op0.getOperand(0))) { | |||
| 22770 | // If we looked past a truncate, check that it's only truncating away | |||
| 22771 | // known zeros. | |||
| 22772 | unsigned BitWidth = Op0.getValueSizeInBits(); | |||
| 22773 | unsigned AndBitWidth = And.getValueSizeInBits(); | |||
| 22774 | if (BitWidth > AndBitWidth) { | |||
| 22775 | KnownBits Known = DAG.computeKnownBits(Op0); | |||
| 22776 | if (Known.countMinLeadingZeros() < BitWidth - AndBitWidth) | |||
| 22777 | return SDValue(); | |||
| 22778 | } | |||
| 22779 | Src = Op1; | |||
| 22780 | BitNo = Op0.getOperand(1); | |||
| 22781 | } | |||
| 22782 | } else if (Op1.getOpcode() == ISD::Constant) { | |||
| 22783 | ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op1); | |||
| 22784 | uint64_t AndRHSVal = AndRHS->getZExtValue(); | |||
| 22785 | SDValue AndLHS = Op0; | |||
| 22786 | ||||
| 22787 | if (AndRHSVal == 1 && AndLHS.getOpcode() == ISD::SRL) { | |||
| 22788 | Src = AndLHS.getOperand(0); | |||
| 22789 | BitNo = AndLHS.getOperand(1); | |||
| 22790 | } else { | |||
| 22791 | // Use BT if the immediate can't be encoded in a TEST instruction or we | |||
| 22792 | // are optimizing for size and the immedaite won't fit in a byte. | |||
| 22793 | bool OptForSize = DAG.shouldOptForSize(); | |||
| 22794 | if ((!isUInt<32>(AndRHSVal) || (OptForSize && !isUInt<8>(AndRHSVal))) && | |||
| 22795 | isPowerOf2_64(AndRHSVal)) { | |||
| 22796 | Src = AndLHS; | |||
| 22797 | BitNo = DAG.getConstant(Log2_64_Ceil(AndRHSVal), dl, | |||
| 22798 | Src.getValueType()); | |||
| 22799 | } | |||
| 22800 | } | |||
| 22801 | } | |||
| 22802 | ||||
| 22803 | // No patterns found, give up. | |||
| 22804 | if (!Src.getNode()) | |||
| 22805 | return SDValue(); | |||
| 22806 | ||||
| 22807 | // If Src is i8, promote it to i32 with any_extend. There is no i8 BT | |||
| 22808 | // instruction. Since the shift amount is in-range-or-undefined, we know | |||
| 22809 | // that doing a bittest on the i32 value is ok. We extend to i32 because | |||
| 22810 | // the encoding for the i16 version is larger than the i32 version. | |||
| 22811 | // Also promote i16 to i32 for performance / code size reason. | |||
| 22812 | if (Src.getValueType() == MVT::i8 || Src.getValueType() == MVT::i16) | |||
| 22813 | Src = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Src); | |||
| 22814 | ||||
| 22815 | // See if we can use the 32-bit instruction instead of the 64-bit one for a | |||
| 22816 | // shorter encoding. Since the former takes the modulo 32 of BitNo and the | |||
| 22817 | // latter takes the modulo 64, this is only valid if the 5th bit of BitNo is | |||
| 22818 | // known to be zero. | |||
| 22819 | if (Src.getValueType() == MVT::i64 && | |||
| 22820 | DAG.MaskedValueIsZero(BitNo, APInt(BitNo.getValueSizeInBits(), 32))) | |||
| 22821 | Src = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Src); | |||
| 22822 | ||||
| 22823 | // If the operand types disagree, extend the shift amount to match. Since | |||
| 22824 | // BT ignores high bits (like shifts) we can use anyextend. | |||
| 22825 | if (Src.getValueType() != BitNo.getValueType()) | |||
| 22826 | BitNo = DAG.getNode(ISD::ANY_EXTEND, dl, Src.getValueType(), BitNo); | |||
| 22827 | ||||
| 22828 | X86CC = DAG.getTargetConstant(CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B, | |||
| 22829 | dl, MVT::i8); | |||
| 22830 | return DAG.getNode(X86ISD::BT, dl, MVT::i32, Src, BitNo); | |||
| 22831 | } | |||
| 22832 | ||||
| 22833 | /// Turns an ISD::CondCode into a value suitable for SSE floating-point mask | |||
| 22834 | /// CMPs. | |||
| 22835 | static unsigned translateX86FSETCC(ISD::CondCode SetCCOpcode, SDValue &Op0, | |||
| 22836 | SDValue &Op1, bool &IsAlwaysSignaling) { | |||
| 22837 | unsigned SSECC; | |||
| 22838 | bool Swap = false; | |||
| 22839 | ||||
| 22840 | // SSE Condition code mapping: | |||
| 22841 | // 0 - EQ | |||
| 22842 | // 1 - LT | |||
| 22843 | // 2 - LE | |||
| 22844 | // 3 - UNORD | |||
| 22845 | // 4 - NEQ | |||
| 22846 | // 5 - NLT | |||
| 22847 | // 6 - NLE | |||
| 22848 | // 7 - ORD | |||
| 22849 | switch (SetCCOpcode) { | |||
| 22850 | default: llvm_unreachable("Unexpected SETCC condition")__builtin_unreachable(); | |||
| 22851 | case ISD::SETOEQ: | |||
| 22852 | case ISD::SETEQ: SSECC = 0; break; | |||
| 22853 | case ISD::SETOGT: | |||
| 22854 | case ISD::SETGT: Swap = true; LLVM_FALLTHROUGH[[gnu::fallthrough]]; | |||
| 22855 | case ISD::SETLT: | |||
| 22856 | case ISD::SETOLT: SSECC = 1; break; | |||
| 22857 | case ISD::SETOGE: | |||
| 22858 | case ISD::SETGE: Swap = true; LLVM_FALLTHROUGH[[gnu::fallthrough]]; | |||
| 22859 | case ISD::SETLE: | |||
| 22860 | case ISD::SETOLE: SSECC = 2; break; | |||
| 22861 | case ISD::SETUO: SSECC = 3; break; | |||
| 22862 | case ISD::SETUNE: | |||
| 22863 | case ISD::SETNE: SSECC = 4; break; | |||
| 22864 | case ISD::SETULE: Swap = true; LLVM_FALLTHROUGH[[gnu::fallthrough]]; | |||
| 22865 | case ISD::SETUGE: SSECC = 5; break; | |||
| 22866 | case ISD::SETULT: Swap = true; LLVM_FALLTHROUGH[[gnu::fallthrough]]; | |||
| 22867 | case ISD::SETUGT: SSECC = 6; break; | |||
| 22868 | case ISD::SETO: SSECC = 7; break; | |||
| 22869 | case ISD::SETUEQ: SSECC = 8; break; | |||
| 22870 | case ISD::SETONE: SSECC = 12; break; | |||
| 22871 | } | |||
| 22872 | if (Swap) | |||
| 22873 | std::swap(Op0, Op1); | |||
| 22874 | ||||
| 22875 | switch (SetCCOpcode) { | |||
| 22876 | default: | |||
| 22877 | IsAlwaysSignaling = true; | |||
| 22878 | break; | |||
| 22879 | case ISD::SETEQ: | |||
| 22880 | case ISD::SETOEQ: | |||
| 22881 | case ISD::SETUEQ: | |||
| 22882 | case ISD::SETNE: | |||
| 22883 | case ISD::SETONE: | |||
| 22884 | case ISD::SETUNE: | |||
| 22885 | case ISD::SETO: | |||
| 22886 | case ISD::SETUO: | |||
| 22887 | IsAlwaysSignaling = false; | |||
| 22888 | break; | |||
| 22889 | } | |||
| 22890 | ||||
| 22891 | return SSECC; | |||
| 22892 | } | |||
| 22893 | ||||
| 22894 | /// Break a VSETCC 256-bit integer VSETCC into two new 128 ones and then | |||
| 22895 | /// concatenate the result back. | |||
| 22896 | static SDValue splitIntVSETCC(EVT VT, SDValue LHS, SDValue RHS, | |||
| 22897 | ISD::CondCode Cond, SelectionDAG &DAG, | |||
| 22898 | const SDLoc &dl) { | |||
| 22899 | assert(VT.isInteger() && VT == LHS.getValueType() &&((void)0) | |||
| 22900 | VT == RHS.getValueType() && "Unsupported VTs!")((void)0); | |||
| 22901 | ||||
| 22902 | SDValue CC = DAG.getCondCode(Cond); | |||
| 22903 | ||||
| 22904 | // Extract the LHS Lo/Hi vectors | |||
| 22905 | SDValue LHS1, LHS2; | |||
| 22906 | std::tie(LHS1, LHS2) = splitVector(LHS, DAG, dl); | |||
| 22907 | ||||
| 22908 | // Extract the RHS Lo/Hi vectors | |||
| 22909 | SDValue RHS1, RHS2; | |||
| 22910 | std::tie(RHS1, RHS2) = splitVector(RHS, DAG, dl); | |||
| 22911 | ||||
| 22912 | // Issue the operation on the smaller types and concatenate the result back | |||
| 22913 | EVT LoVT, HiVT; | |||
| 22914 | std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(VT); | |||
| 22915 | return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, | |||
| 22916 | DAG.getNode(ISD::SETCC, dl, LoVT, LHS1, RHS1, CC), | |||
| 22917 | DAG.getNode(ISD::SETCC, dl, HiVT, LHS2, RHS2, CC)); | |||
| 22918 | } | |||
| 22919 | ||||
| 22920 | static SDValue LowerIntVSETCC_AVX512(SDValue Op, SelectionDAG &DAG) { | |||
| 22921 | ||||
| 22922 | SDValue Op0 = Op.getOperand(0); | |||
| 22923 | SDValue Op1 = Op.getOperand(1); | |||
| 22924 | SDValue CC = Op.getOperand(2); | |||
| 22925 | MVT VT = Op.getSimpleValueType(); | |||
| 22926 | SDLoc dl(Op); | |||
| 22927 | ||||
| 22928 | assert(VT.getVectorElementType() == MVT::i1 &&((void)0) | |||
| 22929 | "Cannot set masked compare for this operation")((void)0); | |||
| 22930 | ||||
| 22931 | ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get(); | |||
| 22932 | ||||
| 22933 | // Prefer SETGT over SETLT. | |||
| 22934 | if (SetCCOpcode == ISD::SETLT) { | |||
| 22935 | SetCCOpcode = ISD::getSetCCSwappedOperands(SetCCOpcode); | |||
| 22936 | std::swap(Op0, Op1); | |||
| 22937 | } | |||
| 22938 | ||||
| 22939 | return DAG.getSetCC(dl, VT, Op0, Op1, SetCCOpcode); | |||
| 22940 | } | |||
| 22941 | ||||
| 22942 | /// Given a buildvector constant, return a new vector constant with each element | |||
| 22943 | /// incremented or decremented. If incrementing or decrementing would result in | |||
| 22944 | /// unsigned overflow or underflow or this is not a simple vector constant, | |||
| 22945 | /// return an empty value. | |||
| 22946 | static SDValue incDecVectorConstant(SDValue V, SelectionDAG &DAG, bool IsInc) { | |||
| 22947 | auto *BV = dyn_cast<BuildVectorSDNode>(V.getNode()); | |||
| 22948 | if (!BV) | |||
| 22949 | return SDValue(); | |||
| 22950 | ||||
| 22951 | MVT VT = V.getSimpleValueType(); | |||
| 22952 | MVT EltVT = VT.getVectorElementType(); | |||
| 22953 | unsigned NumElts = VT.getVectorNumElements(); | |||
| 22954 | SmallVector<SDValue, 8> NewVecC; | |||
| 22955 | SDLoc DL(V); | |||
| 22956 | for (unsigned i = 0; i < NumElts; ++i) { | |||
| 22957 | auto *Elt = dyn_cast<ConstantSDNode>(BV->getOperand(i)); | |||
| 22958 | if (!Elt || Elt->isOpaque() || Elt->getSimpleValueType(0) != EltVT) | |||
| 22959 | return SDValue(); | |||
| 22960 | ||||
| 22961 | // Avoid overflow/underflow. | |||
| 22962 | const APInt &EltC = Elt->getAPIntValue(); | |||
| 22963 | if ((IsInc && EltC.isMaxValue()) || (!IsInc && EltC.isNullValue())) | |||
| 22964 | return SDValue(); | |||
| 22965 | ||||
| 22966 | NewVecC.push_back(DAG.getConstant(EltC + (IsInc ? 1 : -1), DL, EltVT)); | |||
| 22967 | } | |||
| 22968 | ||||
| 22969 | return DAG.getBuildVector(VT, DL, NewVecC); | |||
| 22970 | } | |||
| 22971 | ||||
| 22972 | /// As another special case, use PSUBUS[BW] when it's profitable. E.g. for | |||
| 22973 | /// Op0 u<= Op1: | |||
| 22974 | /// t = psubus Op0, Op1 | |||
| 22975 | /// pcmpeq t, <0..0> | |||
| 22976 | static SDValue LowerVSETCCWithSUBUS(SDValue Op0, SDValue Op1, MVT VT, | |||
| 22977 | ISD::CondCode Cond, const SDLoc &dl, | |||
| 22978 | const X86Subtarget &Subtarget, | |||
| 22979 | SelectionDAG &DAG) { | |||
| 22980 | if (!Subtarget.hasSSE2()) | |||
| 22981 | return SDValue(); | |||
| 22982 | ||||
| 22983 | MVT VET = VT.getVectorElementType(); | |||
| 22984 | if (VET != MVT::i8 && VET != MVT::i16) | |||
| 22985 | return SDValue(); | |||
| 22986 | ||||
| 22987 | switch (Cond) { | |||
| 22988 | default: | |||
| 22989 | return SDValue(); | |||
| 22990 | case ISD::SETULT: { | |||
| 22991 | // If the comparison is against a constant we can turn this into a | |||
| 22992 | // setule. With psubus, setule does not require a swap. This is | |||
| 22993 | // beneficial because the constant in the register is no longer | |||
| 22994 | // destructed as the destination so it can be hoisted out of a loop. | |||
| 22995 | // Only do this pre-AVX since vpcmp* is no longer destructive. | |||
| 22996 | if (Subtarget.hasAVX()) | |||
| 22997 | return SDValue(); | |||
| 22998 | SDValue ULEOp1 = incDecVectorConstant(Op1, DAG, /*IsInc*/false); | |||
| 22999 | if (!ULEOp1) | |||
| 23000 | return SDValue(); | |||
| 23001 | Op1 = ULEOp1; | |||
| 23002 | break; | |||
| 23003 | } | |||
| 23004 | case ISD::SETUGT: { | |||
| 23005 | // If the comparison is against a constant, we can turn this into a setuge. | |||
| 23006 | // This is beneficial because materializing a constant 0 for the PCMPEQ is | |||
| 23007 | // probably cheaper than XOR+PCMPGT using 2 different vector constants: | |||
| 23008 | // cmpgt (xor X, SignMaskC) CmpC --> cmpeq (usubsat (CmpC+1), X), 0 | |||
| 23009 | SDValue UGEOp1 = incDecVectorConstant(Op1, DAG, /*IsInc*/true); | |||
| 23010 | if (!UGEOp1) | |||
| 23011 | return SDValue(); | |||
| 23012 | Op1 = Op0; | |||
| 23013 | Op0 = UGEOp1; | |||
| 23014 | break; | |||
| 23015 | } | |||
| 23016 | // Psubus is better than flip-sign because it requires no inversion. | |||
| 23017 | case ISD::SETUGE: | |||
| 23018 | std::swap(Op0, Op1); | |||
| 23019 | break; | |||
| 23020 | case ISD::SETULE: | |||
| 23021 | break; | |||
| 23022 | } | |||
| 23023 | ||||
| 23024 | SDValue Result = DAG.getNode(ISD::USUBSAT, dl, VT, Op0, Op1); | |||
| 23025 | return DAG.getNode(X86ISD::PCMPEQ, dl, VT, Result, | |||
| 23026 | DAG.getConstant(0, dl, VT)); | |||
| 23027 | } | |||
| 23028 | ||||
| 23029 | static SDValue LowerVSETCC(SDValue Op, const X86Subtarget &Subtarget, | |||
| 23030 | SelectionDAG &DAG) { | |||
| 23031 | bool IsStrict = Op.getOpcode() == ISD::STRICT_FSETCC || | |||
| 23032 | Op.getOpcode() == ISD::STRICT_FSETCCS; | |||
| 23033 | SDValue Op0 = Op.getOperand(IsStrict ? 1 : 0); | |||
| 23034 | SDValue Op1 = Op.getOperand(IsStrict ? 2 : 1); | |||
| 23035 | SDValue CC = Op.getOperand(IsStrict ? 3 : 2); | |||
| 23036 | MVT VT = Op->getSimpleValueType(0); | |||
| 23037 | ISD::CondCode Cond = cast<CondCodeSDNode>(CC)->get(); | |||
| 23038 | bool isFP = Op1.getSimpleValueType().isFloatingPoint(); | |||
| 23039 | SDLoc dl(Op); | |||
| 23040 | ||||
| 23041 | if (isFP) { | |||
| 23042 | #ifndef NDEBUG1 | |||
| 23043 | MVT EltVT = Op0.getSimpleValueType().getVectorElementType(); | |||
| 23044 | assert(EltVT == MVT::f32 || EltVT == MVT::f64)((void)0); | |||
| 23045 | #endif | |||
| 23046 | ||||
| 23047 | bool IsSignaling = Op.getOpcode() == ISD::STRICT_FSETCCS; | |||
| 23048 | SDValue Chain = IsStrict ? Op.getOperand(0) : SDValue(); | |||
| 23049 | ||||
| 23050 | // If we have a strict compare with a vXi1 result and the input is 128/256 | |||
| 23051 | // bits we can't use a masked compare unless we have VLX. If we use a wider | |||
| 23052 | // compare like we do for non-strict, we might trigger spurious exceptions | |||
| 23053 | // from the upper elements. Instead emit a AVX compare and convert to mask. | |||
| 23054 | unsigned Opc; | |||
| 23055 | if (Subtarget.hasAVX512() && VT.getVectorElementType() == MVT::i1 && | |||
| 23056 | (!IsStrict || Subtarget.hasVLX() || | |||
| 23057 | Op0.getSimpleValueType().is512BitVector())) { | |||
| 23058 | assert(VT.getVectorNumElements() <= 16)((void)0); | |||
| 23059 | Opc = IsStrict ? X86ISD::STRICT_CMPM : X86ISD::CMPM; | |||
| 23060 | } else { | |||
| 23061 | Opc = IsStrict ? X86ISD::STRICT_CMPP : X86ISD::CMPP; | |||
| 23062 | // The SSE/AVX packed FP comparison nodes are defined with a | |||
| 23063 | // floating-point vector result that matches the operand type. This allows | |||
| 23064 | // them to work with an SSE1 target (integer vector types are not legal). | |||
| 23065 | VT = Op0.getSimpleValueType(); | |||
| 23066 | } | |||
| 23067 | ||||
| 23068 | SDValue Cmp; | |||
| 23069 | bool IsAlwaysSignaling; | |||
| 23070 | unsigned SSECC = translateX86FSETCC(Cond, Op0, Op1, IsAlwaysSignaling); | |||
| 23071 | if (!Subtarget.hasAVX()) { | |||
| 23072 | // TODO: We could use following steps to handle a quiet compare with | |||
| 23073 | // signaling encodings. | |||
| 23074 | // 1. Get ordered masks from a quiet ISD::SETO | |||
| 23075 | // 2. Use the masks to mask potential unordered elements in operand A, B | |||
| 23076 | // 3. Get the compare results of masked A, B | |||
| 23077 | // 4. Calculating final result using the mask and result from 3 | |||
| 23078 | // But currently, we just fall back to scalar operations. | |||
| 23079 | if (IsStrict && IsAlwaysSignaling && !IsSignaling) | |||
| 23080 | return SDValue(); | |||
| 23081 | ||||
| 23082 | // Insert an extra signaling instruction to raise exception. | |||
| 23083 | if (IsStrict && !IsAlwaysSignaling && IsSignaling) { | |||
| 23084 | SDValue SignalCmp = DAG.getNode( | |||
| 23085 | Opc, dl, {VT, MVT::Other}, | |||
| 23086 | {Chain, Op0, Op1, DAG.getTargetConstant(1, dl, MVT::i8)}); // LT_OS | |||
| 23087 | // FIXME: It seems we need to update the flags of all new strict nodes. | |||
| 23088 | // Otherwise, mayRaiseFPException in MI will return false due to | |||
| 23089 | // NoFPExcept = false by default. However, I didn't find it in other | |||
| 23090 | // patches. | |||
| 23091 | SignalCmp->setFlags(Op->getFlags()); | |||
| 23092 | Chain = SignalCmp.getValue(1); | |||
| 23093 | } | |||
| 23094 | ||||
| 23095 | // In the two cases not handled by SSE compare predicates (SETUEQ/SETONE), | |||
| 23096 | // emit two comparisons and a logic op to tie them together. | |||
| 23097 | if (SSECC >= 8) { | |||
| 23098 | // LLVM predicate is SETUEQ or SETONE. | |||
| 23099 | unsigned CC0, CC1; | |||
| 23100 | unsigned CombineOpc; | |||
| 23101 | if (Cond == ISD::SETUEQ) { | |||
| 23102 | CC0 = 3; // UNORD | |||
| 23103 | CC1 = 0; // EQ | |||
| 23104 | CombineOpc = X86ISD::FOR; | |||
| 23105 | } else { | |||
| 23106 | assert(Cond == ISD::SETONE)((void)0); | |||
| 23107 | CC0 = 7; // ORD | |||
| 23108 | CC1 = 4; // NEQ | |||
| 23109 | CombineOpc = X86ISD::FAND; | |||
| 23110 | } | |||
| 23111 | ||||
| 23112 | SDValue Cmp0, Cmp1; | |||
| 23113 | if (IsStrict) { | |||
| 23114 | Cmp0 = DAG.getNode( | |||
| 23115 | Opc, dl, {VT, MVT::Other}, | |||
| 23116 | {Chain, Op0, Op1, DAG.getTargetConstant(CC0, dl, MVT::i8)}); | |||
| 23117 | Cmp1 = DAG.getNode( | |||
| 23118 | Opc, dl, {VT, MVT::Other}, | |||
| 23119 | {Chain, Op0, Op1, DAG.getTargetConstant(CC1, dl, MVT::i8)}); | |||
| 23120 | Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Cmp0.getValue(1), | |||
| 23121 | Cmp1.getValue(1)); | |||
| 23122 | } else { | |||
| 23123 | Cmp0 = DAG.getNode( | |||
| 23124 | Opc, dl, VT, Op0, Op1, DAG.getTargetConstant(CC0, dl, MVT::i8)); | |||
| 23125 | Cmp1 = DAG.getNode( | |||
| 23126 | Opc, dl, VT, Op0, Op1, DAG.getTargetConstant(CC1, dl, MVT::i8)); | |||
| 23127 | } | |||
| 23128 | Cmp = DAG.getNode(CombineOpc, dl, VT, Cmp0, Cmp1); | |||
| 23129 | } else { | |||
| 23130 | if (IsStrict) { | |||
| 23131 | Cmp = DAG.getNode( | |||
| 23132 | Opc, dl, {VT, MVT::Other}, | |||
| 23133 | {Chain, Op0, Op1, DAG.getTargetConstant(SSECC, dl, MVT::i8)}); | |||
| 23134 | Chain = Cmp.getValue(1); | |||
| 23135 | } else | |||
| 23136 | Cmp = DAG.getNode( | |||
| 23137 | Opc, dl, VT, Op0, Op1, DAG.getTargetConstant(SSECC, dl, MVT::i8)); | |||
| 23138 | } | |||
| 23139 | } else { | |||
| 23140 | // Handle all other FP comparisons here. | |||
| 23141 | if (IsStrict) { | |||
| 23142 | // Make a flip on already signaling CCs before setting bit 4 of AVX CC. | |||
| 23143 | SSECC |= (IsAlwaysSignaling ^ IsSignaling) << 4; | |||
| 23144 | Cmp = DAG.getNode( | |||
| 23145 | Opc, dl, {VT, MVT::Other}, | |||
| 23146 | {Chain, Op0, Op1, DAG.getTargetConstant(SSECC, dl, MVT::i8)}); | |||
| 23147 | Chain = Cmp.getValue(1); | |||
| 23148 | } else | |||
| 23149 | Cmp = DAG.getNode( | |||
| 23150 | Opc, dl, VT, Op0, Op1, DAG.getTargetConstant(SSECC, dl, MVT::i8)); | |||
| 23151 | } | |||
| 23152 | ||||
| 23153 | if (VT.getFixedSizeInBits() > | |||
| 23154 | Op.getSimpleValueType().getFixedSizeInBits()) { | |||
| 23155 | // We emitted a compare with an XMM/YMM result. Finish converting to a | |||
| 23156 | // mask register using a vptestm. | |||
| 23157 | EVT CastVT = EVT(VT).changeVectorElementTypeToInteger(); | |||
| 23158 | Cmp = DAG.getBitcast(CastVT, Cmp); | |||
| 23159 | Cmp = DAG.getSetCC(dl, Op.getSimpleValueType(), Cmp, | |||
| 23160 | DAG.getConstant(0, dl, CastVT), ISD::SETNE); | |||
| 23161 | } else { | |||
| 23162 | // If this is SSE/AVX CMPP, bitcast the result back to integer to match | |||
| 23163 | // the result type of SETCC. The bitcast is expected to be optimized | |||
| 23164 | // away during combining/isel. | |||
| 23165 | Cmp = DAG.getBitcast(Op.getSimpleValueType(), Cmp); | |||
| 23166 | } | |||
| 23167 | ||||
| 23168 | if (IsStrict) | |||
| 23169 | return DAG.getMergeValues({Cmp, Chain}, dl); | |||
| 23170 | ||||
| 23171 | return Cmp; | |||
| 23172 | } | |||
| 23173 | ||||
| 23174 | assert(!IsStrict && "Strict SETCC only handles FP operands.")((void)0); | |||
| 23175 | ||||
| 23176 | MVT VTOp0 = Op0.getSimpleValueType(); | |||
| 23177 | (void)VTOp0; | |||
| 23178 | assert(VTOp0 == Op1.getSimpleValueType() &&((void)0) | |||
| 23179 | "Expected operands with same type!")((void)0); | |||
| 23180 | assert(VT.getVectorNumElements() == VTOp0.getVectorNumElements() &&((void)0) | |||
| 23181 | "Invalid number of packed elements for source and destination!")((void)0); | |||
| 23182 | ||||
| 23183 | // The non-AVX512 code below works under the assumption that source and | |||
| 23184 | // destination types are the same. | |||
| 23185 | assert((Subtarget.hasAVX512() || (VT == VTOp0)) &&((void)0) | |||
| 23186 | "Value types for source and destination must be the same!")((void)0); | |||
| 23187 | ||||
| 23188 | // The result is boolean, but operands are int/float | |||
| 23189 | if (VT.getVectorElementType() == MVT::i1) { | |||
| 23190 | // In AVX-512 architecture setcc returns mask with i1 elements, | |||
| 23191 | // But there is no compare instruction for i8 and i16 elements in KNL. | |||
| 23192 | assert((VTOp0.getScalarSizeInBits() >= 32 || Subtarget.hasBWI()) &&((void)0) | |||
| 23193 | "Unexpected operand type")((void)0); | |||
| 23194 | return LowerIntVSETCC_AVX512(Op, DAG); | |||
| 23195 | } | |||
| 23196 | ||||
| 23197 | // Lower using XOP integer comparisons. | |||
| 23198 | if (VT.is128BitVector() && Subtarget.hasXOP()) { | |||
| 23199 | // Translate compare code to XOP PCOM compare mode. | |||
| 23200 | unsigned CmpMode = 0; | |||
| 23201 | switch (Cond) { | |||
| 23202 | default: llvm_unreachable("Unexpected SETCC condition")__builtin_unreachable(); | |||
| 23203 | case ISD::SETULT: | |||
| 23204 | case ISD::SETLT: CmpMode = 0x00; break; | |||
| 23205 | case ISD::SETULE: | |||
| 23206 | case ISD::SETLE: CmpMode = 0x01; break; | |||
| 23207 | case ISD::SETUGT: | |||
| 23208 | case ISD::SETGT: CmpMode = 0x02; break; | |||
| 23209 | case ISD::SETUGE: | |||
| 23210 | case ISD::SETGE: CmpMode = 0x03; break; | |||
| 23211 | case ISD::SETEQ: CmpMode = 0x04; break; | |||
| 23212 | case ISD::SETNE: CmpMode = 0x05; break; | |||
| 23213 | } | |||
| 23214 | ||||
| 23215 | // Are we comparing unsigned or signed integers? | |||
| 23216 | unsigned Opc = | |||
| 23217 | ISD::isUnsignedIntSetCC(Cond) ? X86ISD::VPCOMU : X86ISD::VPCOM; | |||
| 23218 | ||||
| 23219 | return DAG.getNode(Opc, dl, VT, Op0, Op1, | |||
| 23220 | DAG.getTargetConstant(CmpMode, dl, MVT::i8)); | |||
| 23221 | } | |||
| 23222 | ||||
| 23223 | // (X & Y) != 0 --> (X & Y) == Y iff Y is power-of-2. | |||
| 23224 | // Revert part of the simplifySetCCWithAnd combine, to avoid an invert. | |||
| 23225 | if (Cond == ISD::SETNE && ISD::isBuildVectorAllZeros(Op1.getNode())) { | |||
| 23226 | SDValue BC0 = peekThroughBitcasts(Op0); | |||
| 23227 | if (BC0.getOpcode() == ISD::AND) { | |||
| 23228 | APInt UndefElts; | |||
| 23229 | SmallVector<APInt, 64> EltBits; | |||
| 23230 | if (getTargetConstantBitsFromNode(BC0.getOperand(1), | |||
| 23231 | VT.getScalarSizeInBits(), UndefElts, | |||
| 23232 | EltBits, false, false)) { | |||
| 23233 | if (llvm::all_of(EltBits, [](APInt &V) { return V.isPowerOf2(); })) { | |||
| 23234 | Cond = ISD::SETEQ; | |||
| 23235 | Op1 = DAG.getBitcast(VT, BC0.getOperand(1)); | |||
| 23236 | } | |||
| 23237 | } | |||
| 23238 | } | |||
| 23239 | } | |||
| 23240 | ||||
| 23241 | // ICMP_EQ(AND(X,C),C) -> SRA(SHL(X,LOG2(C)),BW-1) iff C is power-of-2. | |||
| 23242 | if (Cond == ISD::SETEQ && Op0.getOpcode() == ISD::AND && | |||
| 23243 | Op0.getOperand(1) == Op1 && Op0.hasOneUse()) { | |||
| 23244 | ConstantSDNode *C1 = isConstOrConstSplat(Op1); | |||
| 23245 | if (C1 && C1->getAPIntValue().isPowerOf2()) { | |||
| 23246 | unsigned BitWidth = VT.getScalarSizeInBits(); | |||
| 23247 | unsigned ShiftAmt = BitWidth - C1->getAPIntValue().logBase2() - 1; | |||
| 23248 | ||||
| 23249 | SDValue Result = Op0.getOperand(0); | |||
| 23250 | Result = DAG.getNode(ISD::SHL, dl, VT, Result, | |||
| 23251 | DAG.getConstant(ShiftAmt, dl, VT)); | |||
| 23252 | Result = DAG.getNode(ISD::SRA, dl, VT, Result, | |||
| 23253 | DAG.getConstant(BitWidth - 1, dl, VT)); | |||
| 23254 | return Result; | |||
| 23255 | } | |||
| 23256 | } | |||
| 23257 | ||||
| 23258 | // Break 256-bit integer vector compare into smaller ones. | |||
| 23259 | if (VT.is256BitVector() && !Subtarget.hasInt256()) | |||
| 23260 | return splitIntVSETCC(VT, Op0, Op1, Cond, DAG, dl); | |||
| 23261 | ||||
| 23262 | if (VT == MVT::v32i16 || VT == MVT::v64i8) { | |||
| 23263 | assert(!Subtarget.hasBWI() && "Unexpected VT with AVX512BW!")((void)0); | |||
| 23264 | return splitIntVSETCC(VT, Op0, Op1, Cond, DAG, dl); | |||
| 23265 | } | |||
| 23266 | ||||
| 23267 | // If we have a limit constant, try to form PCMPGT (signed cmp) to avoid | |||
| 23268 | // not-of-PCMPEQ: | |||
| 23269 | // X != INT_MIN --> X >s INT_MIN | |||
| 23270 | // X != INT_MAX --> X <s INT_MAX --> INT_MAX >s X | |||
| 23271 | // +X != 0 --> +X >s 0 | |||
| 23272 | APInt ConstValue; | |||
| 23273 | if (Cond == ISD::SETNE && | |||
| 23274 | ISD::isConstantSplatVector(Op1.getNode(), ConstValue)) { | |||
| 23275 | if (ConstValue.isMinSignedValue()) | |||
| 23276 | Cond = ISD::SETGT; | |||
| 23277 | else if (ConstValue.isMaxSignedValue()) | |||
| 23278 | Cond = ISD::SETLT; | |||
| 23279 | else if (ConstValue.isNullValue() && DAG.SignBitIsZero(Op0)) | |||
| 23280 | Cond = ISD::SETGT; | |||
| 23281 | } | |||
| 23282 | ||||
| 23283 | // If both operands are known non-negative, then an unsigned compare is the | |||
| 23284 | // same as a signed compare and there's no need to flip signbits. | |||
| 23285 | // TODO: We could check for more general simplifications here since we're | |||
| 23286 | // computing known bits. | |||
| 23287 | bool FlipSigns = ISD::isUnsignedIntSetCC(Cond) && | |||
| 23288 | !(DAG.SignBitIsZero(Op0) && DAG.SignBitIsZero(Op1)); | |||
| 23289 | ||||
| 23290 | // Special case: Use min/max operations for unsigned compares. | |||
| 23291 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
| 23292 | if (ISD::isUnsignedIntSetCC(Cond) && | |||
| 23293 | (FlipSigns || ISD::isTrueWhenEqual(Cond)) && | |||
| 23294 | TLI.isOperationLegal(ISD::UMIN, VT)) { | |||
| 23295 | // If we have a constant operand, increment/decrement it and change the | |||
| 23296 | // condition to avoid an invert. | |||
| 23297 | if (Cond == ISD::SETUGT) { | |||
| 23298 | // X > C --> X >= (C+1) --> X == umax(X, C+1) | |||
| 23299 | if (SDValue UGTOp1 = incDecVectorConstant(Op1, DAG, /*IsInc*/true)) { | |||
| 23300 | Op1 = UGTOp1; | |||
| 23301 | Cond = ISD::SETUGE; | |||
| 23302 | } | |||
| 23303 | } | |||
| 23304 | if (Cond == ISD::SETULT) { | |||
| 23305 | // X < C --> X <= (C-1) --> X == umin(X, C-1) | |||
| 23306 | if (SDValue ULTOp1 = incDecVectorConstant(Op1, DAG, /*IsInc*/false)) { | |||
| 23307 | Op1 = ULTOp1; | |||
| 23308 | Cond = ISD::SETULE; | |||
| 23309 | } | |||
| 23310 | } | |||
| 23311 | bool Invert = false; | |||
| 23312 | unsigned Opc; | |||
| 23313 | switch (Cond) { | |||
| 23314 | default: llvm_unreachable("Unexpected condition code")__builtin_unreachable(); | |||
| 23315 | case ISD::SETUGT: Invert = true; LLVM_FALLTHROUGH[[gnu::fallthrough]]; | |||
| 23316 | case ISD::SETULE: Opc = ISD::UMIN; break; | |||
| 23317 | case ISD::SETULT: Invert = true; LLVM_FALLTHROUGH[[gnu::fallthrough]]; | |||
| 23318 | case ISD::SETUGE: Opc = ISD::UMAX; break; | |||
| 23319 | } | |||
| 23320 | ||||
| 23321 | SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1); | |||
| 23322 | Result = DAG.getNode(X86ISD::PCMPEQ, dl, VT, Op0, Result); | |||
| 23323 | ||||
| 23324 | // If the logical-not of the result is required, perform that now. | |||
| 23325 | if (Invert) | |||
| 23326 | Result = DAG.getNOT(dl, Result, VT); | |||
| 23327 | ||||
| 23328 | return Result; | |||
| 23329 | } | |||
| 23330 | ||||
| 23331 | // Try to use SUBUS and PCMPEQ. | |||
| 23332 | if (FlipSigns) | |||
| 23333 | if (SDValue V = | |||
| 23334 | LowerVSETCCWithSUBUS(Op0, Op1, VT, Cond, dl, Subtarget, DAG)) | |||
| 23335 | return V; | |||
| 23336 | ||||
| 23337 | // We are handling one of the integer comparisons here. Since SSE only has | |||
| 23338 | // GT and EQ comparisons for integer, swapping operands and multiple | |||
| 23339 | // operations may be required for some comparisons. | |||
| 23340 | unsigned Opc = (Cond == ISD::SETEQ || Cond == ISD::SETNE) ? X86ISD::PCMPEQ | |||
| 23341 | : X86ISD::PCMPGT; | |||
| 23342 | bool Swap = Cond == ISD::SETLT || Cond == ISD::SETULT || | |||
| 23343 | Cond == ISD::SETGE || Cond == ISD::SETUGE; | |||
| 23344 | bool Invert = Cond == ISD::SETNE || | |||
| 23345 | (Cond != ISD::SETEQ && ISD::isTrueWhenEqual(Cond)); | |||
| 23346 | ||||
| 23347 | if (Swap) | |||
| 23348 | std::swap(Op0, Op1); | |||
| 23349 | ||||
| 23350 | // Check that the operation in question is available (most are plain SSE2, | |||
| 23351 | // but PCMPGTQ and PCMPEQQ have different requirements). | |||
| 23352 | if (VT == MVT::v2i64) { | |||
| 23353 | if (Opc == X86ISD::PCMPGT && !Subtarget.hasSSE42()) { | |||
| 23354 | assert(Subtarget.hasSSE2() && "Don't know how to lower!")((void)0); | |||
| 23355 | ||||
| 23356 | // Special case for sign bit test. We can use a v4i32 PCMPGT and shuffle | |||
| 23357 | // the odd elements over the even elements. | |||
| 23358 | if (!FlipSigns && !Invert && ISD::isBuildVectorAllZeros(Op0.getNode())) { | |||
| 23359 | Op0 = DAG.getConstant(0, dl, MVT::v4i32); | |||
| 23360 | Op1 = DAG.getBitcast(MVT::v4i32, Op1); | |||
| 23361 | ||||
| 23362 | SDValue GT = DAG.getNode(X86ISD::PCMPGT, dl, MVT::v4i32, Op0, Op1); | |||
| 23363 | static const int MaskHi[] = { 1, 1, 3, 3 }; | |||
| 23364 | SDValue Result = DAG.getVectorShuffle(MVT::v4i32, dl, GT, GT, MaskHi); | |||
| 23365 | ||||
| 23366 | return DAG.getBitcast(VT, Result); | |||
| 23367 | } | |||
| 23368 | ||||
| 23369 | if (!FlipSigns && !Invert && ISD::isBuildVectorAllOnes(Op1.getNode())) { | |||
| 23370 | Op0 = DAG.getBitcast(MVT::v4i32, Op0); | |||
| 23371 | Op1 = DAG.getConstant(-1, dl, MVT::v4i32); | |||
| 23372 | ||||
| 23373 | SDValue GT = DAG.getNode(X86ISD::PCMPGT, dl, MVT::v4i32, Op0, Op1); | |||
| 23374 | static const int MaskHi[] = { 1, 1, 3, 3 }; | |||
| 23375 | SDValue Result = DAG.getVectorShuffle(MVT::v4i32, dl, GT, GT, MaskHi); | |||
| 23376 | ||||
| 23377 | return DAG.getBitcast(VT, Result); | |||
| 23378 | } | |||
| 23379 | ||||
| 23380 | // Since SSE has no unsigned integer comparisons, we need to flip the sign | |||
| 23381 | // bits of the inputs before performing those operations. The lower | |||
| 23382 | // compare is always unsigned. | |||
| 23383 | SDValue SB; | |||
| 23384 | if (FlipSigns) { | |||
| 23385 | SB = DAG.getConstant(0x8000000080000000ULL, dl, MVT::v2i64); | |||
| 23386 | } else { | |||
| 23387 | SB = DAG.getConstant(0x0000000080000000ULL, dl, MVT::v2i64); | |||
| 23388 | } | |||
| 23389 | Op0 = DAG.getNode(ISD::XOR, dl, MVT::v2i64, Op0, SB); | |||
| 23390 | Op1 = DAG.getNode(ISD::XOR, dl, MVT::v2i64, Op1, SB); | |||
| 23391 | ||||
| 23392 | // Cast everything to the right type. | |||
| 23393 | Op0 = DAG.getBitcast(MVT::v4i32, Op0); | |||
| 23394 | Op1 = DAG.getBitcast(MVT::v4i32, Op1); | |||
| 23395 | ||||
| 23396 | // Emulate PCMPGTQ with (hi1 > hi2) | ((hi1 == hi2) & (lo1 > lo2)) | |||
| 23397 | SDValue GT = DAG.getNode(X86ISD::PCMPGT, dl, MVT::v4i32, Op0, Op1); | |||
| 23398 | SDValue EQ = DAG.getNode(X86ISD::PCMPEQ, dl, MVT::v4i32, Op0, Op1); | |||
| 23399 | ||||
| 23400 | // Create masks for only the low parts/high parts of the 64 bit integers. | |||
| 23401 | static const int MaskHi[] = { 1, 1, 3, 3 }; | |||
| 23402 | static const int MaskLo[] = { 0, 0, 2, 2 }; | |||
| 23403 | SDValue EQHi = DAG.getVectorShuffle(MVT::v4i32, dl, EQ, EQ, MaskHi); | |||
| 23404 | SDValue GTLo = DAG.getVectorShuffle(MVT::v4i32, dl, GT, GT, MaskLo); | |||
| 23405 | SDValue GTHi = DAG.getVectorShuffle(MVT::v4i32, dl, GT, GT, MaskHi); | |||
| 23406 | ||||
| 23407 | SDValue Result = DAG.getNode(ISD::AND, dl, MVT::v4i32, EQHi, GTLo); | |||
| 23408 | Result = DAG.getNode(ISD::OR, dl, MVT::v4i32, Result, GTHi); | |||
| 23409 | ||||
| 23410 | if (Invert) | |||
| 23411 | Result = DAG.getNOT(dl, Result, MVT::v4i32); | |||
| 23412 | ||||
| 23413 | return DAG.getBitcast(VT, Result); | |||
| 23414 | } | |||
| 23415 | ||||
| 23416 | if (Opc == X86ISD::PCMPEQ && !Subtarget.hasSSE41()) { | |||
| 23417 | // If pcmpeqq is missing but pcmpeqd is available synthesize pcmpeqq with | |||
| 23418 | // pcmpeqd + pshufd + pand. | |||
| 23419 | assert(Subtarget.hasSSE2() && !FlipSigns && "Don't know how to lower!")((void)0); | |||
| 23420 | ||||
| 23421 | // First cast everything to the right type. | |||
| 23422 | Op0 = DAG.getBitcast(MVT::v4i32, Op0); | |||
| 23423 | Op1 = DAG.getBitcast(MVT::v4i32, Op1); | |||
| 23424 | ||||
| 23425 | // Do the compare. | |||
| 23426 | SDValue Result = DAG.getNode(Opc, dl, MVT::v4i32, Op0, Op1); | |||
| 23427 | ||||
| 23428 | // Make sure the lower and upper halves are both all-ones. | |||
| 23429 | static const int Mask[] = { 1, 0, 3, 2 }; | |||
| 23430 | SDValue Shuf = DAG.getVectorShuffle(MVT::v4i32, dl, Result, Result, Mask); | |||
| 23431 | Result = DAG.getNode(ISD::AND, dl, MVT::v4i32, Result, Shuf); | |||
| 23432 | ||||
| 23433 | if (Invert) | |||
| 23434 | Result = DAG.getNOT(dl, Result, MVT::v4i32); | |||
| 23435 | ||||
| 23436 | return DAG.getBitcast(VT, Result); | |||
| 23437 | } | |||
| 23438 | } | |||
| 23439 | ||||
| 23440 | // Since SSE has no unsigned integer comparisons, we need to flip the sign | |||
| 23441 | // bits of the inputs before performing those operations. | |||
| 23442 | if (FlipSigns) { | |||
| 23443 | MVT EltVT = VT.getVectorElementType(); | |||
| 23444 | SDValue SM = DAG.getConstant(APInt::getSignMask(EltVT.getSizeInBits()), dl, | |||
| 23445 | VT); | |||
| 23446 | Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SM); | |||
| 23447 | Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SM); | |||
| 23448 | } | |||
| 23449 | ||||
| 23450 | SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1); | |||
| 23451 | ||||
| 23452 | // If the logical-not of the result is required, perform that now. | |||
| 23453 | if (Invert) | |||
| 23454 | Result = DAG.getNOT(dl, Result, VT); | |||
| 23455 | ||||
| 23456 | return Result; | |||
| 23457 | } | |||
| 23458 | ||||
| 23459 | // Try to select this as a KORTEST+SETCC or KTEST+SETCC if possible. | |||
| 23460 | static SDValue EmitAVX512Test(SDValue Op0, SDValue Op1, ISD::CondCode CC, | |||
| 23461 | const SDLoc &dl, SelectionDAG &DAG, | |||
| 23462 | const X86Subtarget &Subtarget, | |||
| 23463 | SDValue &X86CC) { | |||
| 23464 | // Only support equality comparisons. | |||
| 23465 | if (CC != ISD::SETEQ && CC != ISD::SETNE) | |||
| 23466 | return SDValue(); | |||
| 23467 | ||||
| 23468 | // Must be a bitcast from vXi1. | |||
| 23469 | if (Op0.getOpcode() != ISD::BITCAST) | |||
| 23470 | return SDValue(); | |||
| 23471 | ||||
| 23472 | Op0 = Op0.getOperand(0); | |||
| 23473 | MVT VT = Op0.getSimpleValueType(); | |||
| 23474 | if (!(Subtarget.hasAVX512() && VT == MVT::v16i1) && | |||
| 23475 | !(Subtarget.hasDQI() && VT == MVT::v8i1) && | |||
| 23476 | !(Subtarget.hasBWI() && (VT == MVT::v32i1 || VT == MVT::v64i1))) | |||
| 23477 | return SDValue(); | |||
| 23478 | ||||
| 23479 | X86::CondCode X86Cond; | |||
| 23480 | if (isNullConstant(Op1)) { | |||
| 23481 | X86Cond = CC == ISD::SETEQ ? X86::COND_E : X86::COND_NE; | |||
| 23482 | } else if (isAllOnesConstant(Op1)) { | |||
| 23483 | // C flag is set for all ones. | |||
| 23484 | X86Cond = CC == ISD::SETEQ ? X86::COND_B : X86::COND_AE; | |||
| 23485 | } else | |||
| 23486 | return SDValue(); | |||
| 23487 | ||||
| 23488 | // If the input is an AND, we can combine it's operands into the KTEST. | |||
| 23489 | bool KTestable = false; | |||
| 23490 | if (Subtarget.hasDQI() && (VT == MVT::v8i1 || VT == MVT::v16i1)) | |||
| 23491 | KTestable = true; | |||
| 23492 | if (Subtarget.hasBWI() && (VT == MVT::v32i1 || VT == MVT::v64i1)) | |||
| 23493 | KTestable = true; | |||
| 23494 | if (!isNullConstant(Op1)) | |||
| 23495 | KTestable = false; | |||
| 23496 | if (KTestable && Op0.getOpcode() == ISD::AND && Op0.hasOneUse()) { | |||
| 23497 | SDValue LHS = Op0.getOperand(0); | |||
| 23498 | SDValue RHS = Op0.getOperand(1); | |||
| 23499 | X86CC = DAG.getTargetConstant(X86Cond, dl, MVT::i8); | |||
| 23500 | return DAG.getNode(X86ISD::KTEST, dl, MVT::i32, LHS, RHS); | |||
| 23501 | } | |||
| 23502 | ||||
| 23503 | // If the input is an OR, we can combine it's operands into the KORTEST. | |||
| 23504 | SDValue LHS = Op0; | |||
| 23505 | SDValue RHS = Op0; | |||
| 23506 | if (Op0.getOpcode() == ISD::OR && Op0.hasOneUse()) { | |||
| 23507 | LHS = Op0.getOperand(0); | |||
| 23508 | RHS = Op0.getOperand(1); | |||
| 23509 | } | |||
| 23510 | ||||
| 23511 | X86CC = DAG.getTargetConstant(X86Cond, dl, MVT::i8); | |||
| 23512 | return DAG.getNode(X86ISD::KORTEST, dl, MVT::i32, LHS, RHS); | |||
| 23513 | } | |||
| 23514 | ||||
| 23515 | /// Emit flags for the given setcc condition and operands. Also returns the | |||
| 23516 | /// corresponding X86 condition code constant in X86CC. | |||
| 23517 | SDValue X86TargetLowering::emitFlagsForSetcc(SDValue Op0, SDValue Op1, | |||
| 23518 | ISD::CondCode CC, const SDLoc &dl, | |||
| 23519 | SelectionDAG &DAG, | |||
| 23520 | SDValue &X86CC) const { | |||
| 23521 | // Optimize to BT if possible. | |||
| 23522 | // Lower (X & (1 << N)) == 0 to BT(X, N). | |||
| 23523 | // Lower ((X >>u N) & 1) != 0 to BT(X, N). | |||
| 23524 | // Lower ((X >>s N) & 1) != 0 to BT(X, N). | |||
| 23525 | if (Op0.getOpcode() == ISD::AND && Op0.hasOneUse() && isNullConstant(Op1) && | |||
| 23526 | (CC == ISD::SETEQ || CC == ISD::SETNE)) { | |||
| 23527 | if (SDValue BT = LowerAndToBT(Op0, CC, dl, DAG, X86CC)) | |||
| 23528 | return BT; | |||
| 23529 | } | |||
| 23530 | ||||
| 23531 | // Try to use PTEST/PMOVMSKB for a tree ORs equality compared with 0. | |||
| 23532 | // TODO: We could do AND tree with all 1s as well by using the C flag. | |||
| 23533 | if (isNullConstant(Op1) && (CC == ISD::SETEQ || CC == ISD::SETNE)) | |||
| 23534 | if (SDValue CmpZ = | |||
| 23535 | MatchVectorAllZeroTest(Op0, CC, dl, Subtarget, DAG, X86CC)) | |||
| 23536 | return CmpZ; | |||
| 23537 | ||||
| 23538 | // Try to lower using KORTEST or KTEST. | |||
| 23539 | if (SDValue Test = EmitAVX512Test(Op0, Op1, CC, dl, DAG, Subtarget, X86CC)) | |||
| 23540 | return Test; | |||
| 23541 | ||||
| 23542 | // Look for X == 0, X == 1, X != 0, or X != 1. We can simplify some forms of | |||
| 23543 | // these. | |||
| 23544 | if ((isOneConstant(Op1) || isNullConstant(Op1)) && | |||
| 23545 | (CC == ISD::SETEQ || CC == ISD::SETNE)) { | |||
| 23546 | // If the input is a setcc, then reuse the input setcc or use a new one with | |||
| 23547 | // the inverted condition. | |||
| 23548 | if (Op0.getOpcode() == X86ISD::SETCC) { | |||
| 23549 | bool Invert = (CC == ISD::SETNE) ^ isNullConstant(Op1); | |||
| 23550 | ||||
| 23551 | X86CC = Op0.getOperand(0); | |||
| 23552 | if (Invert) { | |||
| 23553 | X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0); | |||
| 23554 | CCode = X86::GetOppositeBranchCondition(CCode); | |||
| 23555 | X86CC = DAG.getTargetConstant(CCode, dl, MVT::i8); | |||
| 23556 | } | |||
| 23557 | ||||
| 23558 | return Op0.getOperand(1); | |||
| 23559 | } | |||
| 23560 | } | |||
| 23561 | ||||
| 23562 | // Try to use the carry flag from the add in place of an separate CMP for: | |||
| 23563 | // (seteq (add X, -1), -1). Similar for setne. | |||
| 23564 | if (isAllOnesConstant(Op1) && Op0.getOpcode() == ISD::ADD && | |||
| 23565 | Op0.getOperand(1) == Op1 && (CC == ISD::SETEQ || CC == ISD::SETNE)) { | |||
| 23566 | if (isProfitableToUseFlagOp(Op0)) { | |||
| 23567 | SDVTList VTs = DAG.getVTList(Op0.getValueType(), MVT::i32); | |||
| 23568 | ||||
| 23569 | SDValue New = DAG.getNode(X86ISD::ADD, dl, VTs, Op0.getOperand(0), | |||
| 23570 | Op0.getOperand(1)); | |||
| 23571 | DAG.ReplaceAllUsesOfValueWith(SDValue(Op0.getNode(), 0), New); | |||
| 23572 | X86::CondCode CCode = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B; | |||
| 23573 | X86CC = DAG.getTargetConstant(CCode, dl, MVT::i8); | |||
| 23574 | return SDValue(New.getNode(), 1); | |||
| 23575 | } | |||
| 23576 | } | |||
| 23577 | ||||
| 23578 | X86::CondCode CondCode = | |||
| 23579 | TranslateX86CC(CC, dl, /*IsFP*/ false, Op0, Op1, DAG); | |||
| 23580 | assert(CondCode != X86::COND_INVALID && "Unexpected condition code!")((void)0); | |||
| 23581 | ||||
| 23582 | SDValue EFLAGS = EmitCmp(Op0, Op1, CondCode, dl, DAG, Subtarget); | |||
| 23583 | X86CC = DAG.getTargetConstant(CondCode, dl, MVT::i8); | |||
| 23584 | return EFLAGS; | |||
| 23585 | } | |||
| 23586 | ||||
| 23587 | SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const { | |||
| 23588 | ||||
| 23589 | bool IsStrict = Op.getOpcode() == ISD::STRICT_FSETCC || | |||
| 23590 | Op.getOpcode() == ISD::STRICT_FSETCCS; | |||
| 23591 | MVT VT = Op->getSimpleValueType(0); | |||
| 23592 | ||||
| 23593 | if (VT.isVector()) return LowerVSETCC(Op, Subtarget, DAG); | |||
| 23594 | ||||
| 23595 | assert(VT == MVT::i8 && "SetCC type must be 8-bit integer")((void)0); | |||
| 23596 | SDValue Chain = IsStrict ? Op.getOperand(0) : SDValue(); | |||
| 23597 | SDValue Op0 = Op.getOperand(IsStrict ? 1 : 0); | |||
| 23598 | SDValue Op1 = Op.getOperand(IsStrict ? 2 : 1); | |||
| 23599 | SDLoc dl(Op); | |||
| 23600 | ISD::CondCode CC = | |||
| 23601 | cast<CondCodeSDNode>(Op.getOperand(IsStrict ? 3 : 2))->get(); | |||
| 23602 | ||||
| 23603 | // Handle f128 first, since one possible outcome is a normal integer | |||
| 23604 | // comparison which gets handled by emitFlagsForSetcc. | |||
| 23605 | if (Op0.getValueType() == MVT::f128) { | |||
| 23606 | softenSetCCOperands(DAG, MVT::f128, Op0, Op1, CC, dl, Op0, Op1, Chain, | |||
| 23607 | Op.getOpcode() == ISD::STRICT_FSETCCS); | |||
| 23608 | ||||
| 23609 | // If softenSetCCOperands returned a scalar, use it. | |||
| 23610 | if (!Op1.getNode()) { | |||
| 23611 | assert(Op0.getValueType() == Op.getValueType() &&((void)0) | |||
| 23612 | "Unexpected setcc expansion!")((void)0); | |||
| 23613 | if (IsStrict) | |||
| 23614 | return DAG.getMergeValues({Op0, Chain}, dl); | |||
| 23615 | return Op0; | |||
| 23616 | } | |||
| 23617 | } | |||
| 23618 | ||||
| 23619 | if (Op0.getSimpleValueType().isInteger()) { | |||
| 23620 | // Attempt to canonicalize SGT/UGT -> SGE/UGE compares with constant which | |||
| 23621 | // reduces the number of EFLAGs bit reads (the GE conditions don't read ZF), | |||
| 23622 | // this may translate to less uops depending on uarch implementation. The | |||
| 23623 | // equivalent for SLE/ULE -> SLT/ULT isn't likely to happen as we already | |||
| 23624 | // canonicalize to that CondCode. | |||
| 23625 | // NOTE: Only do this if incrementing the constant doesn't increase the bit | |||
| 23626 | // encoding size - so it must either already be a i8 or i32 immediate, or it | |||
| 23627 | // shrinks down to that. We don't do this for any i64's to avoid additional | |||
| 23628 | // constant materializations. | |||
| 23629 | // TODO: Can we move this to TranslateX86CC to handle jumps/branches too? | |||
| 23630 | if (auto *Op1C = dyn_cast<ConstantSDNode>(Op1)) { | |||
| 23631 | const APInt &Op1Val = Op1C->getAPIntValue(); | |||
| 23632 | if (!Op1Val.isNullValue()) { | |||
| 23633 | // Ensure the constant+1 doesn't overflow. | |||
| 23634 | if ((CC == ISD::CondCode::SETGT && !Op1Val.isMaxSignedValue()) || | |||
| 23635 | (CC == ISD::CondCode::SETUGT && !Op1Val.isMaxValue())) { | |||
| 23636 | APInt Op1ValPlusOne = Op1Val + 1; | |||
| 23637 | if (Op1ValPlusOne.isSignedIntN(32) && | |||
| 23638 | (!Op1Val.isSignedIntN(8) || Op1ValPlusOne.isSignedIntN(8))) { | |||
| 23639 | Op1 = DAG.getConstant(Op1ValPlusOne, dl, Op0.getValueType()); | |||
| 23640 | CC = CC == ISD::CondCode::SETGT ? ISD::CondCode::SETGE | |||
| 23641 | : ISD::CondCode::SETUGE; | |||
| 23642 | } | |||
| 23643 | } | |||
| 23644 | } | |||
| 23645 | } | |||
| 23646 | ||||
| 23647 | SDValue X86CC; | |||
| 23648 | SDValue EFLAGS = emitFlagsForSetcc(Op0, Op1, CC, dl, DAG, X86CC); | |||
| 23649 | SDValue Res = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, X86CC, EFLAGS); | |||
| 23650 | return IsStrict ? DAG.getMergeValues({Res, Chain}, dl) : Res; | |||
| 23651 | } | |||
| 23652 | ||||
| 23653 | // Handle floating point. | |||
| 23654 | X86::CondCode CondCode = TranslateX86CC(CC, dl, /*IsFP*/ true, Op0, Op1, DAG); | |||
| 23655 | if (CondCode == X86::COND_INVALID) | |||
| 23656 | return SDValue(); | |||
| 23657 | ||||
| 23658 | SDValue EFLAGS; | |||
| 23659 | if (IsStrict) { | |||
| 23660 | bool IsSignaling = Op.getOpcode() == ISD::STRICT_FSETCCS; | |||
| 23661 | EFLAGS = | |||
| 23662 | DAG.getNode(IsSignaling ? X86ISD::STRICT_FCMPS : X86ISD::STRICT_FCMP, | |||
| 23663 | dl, {MVT::i32, MVT::Other}, {Chain, Op0, Op1}); | |||
| 23664 | Chain = EFLAGS.getValue(1); | |||
| 23665 | } else { | |||
| 23666 | EFLAGS = DAG.getNode(X86ISD::FCMP, dl, MVT::i32, Op0, Op1); | |||
| 23667 | } | |||
| 23668 | ||||
| 23669 | SDValue X86CC = DAG.getTargetConstant(CondCode, dl, MVT::i8); | |||
| 23670 | SDValue Res = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, X86CC, EFLAGS); | |||
| 23671 | return IsStrict ? DAG.getMergeValues({Res, Chain}, dl) : Res; | |||
| 23672 | } | |||
| 23673 | ||||
| 23674 | SDValue X86TargetLowering::LowerSETCCCARRY(SDValue Op, SelectionDAG &DAG) const { | |||
| 23675 | SDValue LHS = Op.getOperand(0); | |||
| 23676 | SDValue RHS = Op.getOperand(1); | |||
| 23677 | SDValue Carry = Op.getOperand(2); | |||
| 23678 | SDValue Cond = Op.getOperand(3); | |||
| 23679 | SDLoc DL(Op); | |||
| 23680 | ||||
| 23681 | assert(LHS.getSimpleValueType().isInteger() && "SETCCCARRY is integer only.")((void)0); | |||
| 23682 | X86::CondCode CC = TranslateIntegerX86CC(cast<CondCodeSDNode>(Cond)->get()); | |||
| 23683 | ||||
| 23684 | // Recreate the carry if needed. | |||
| 23685 | EVT CarryVT = Carry.getValueType(); | |||
| 23686 | Carry = DAG.getNode(X86ISD::ADD, DL, DAG.getVTList(CarryVT, MVT::i32), | |||
| 23687 | Carry, DAG.getAllOnesConstant(DL, CarryVT)); | |||
| 23688 | ||||
| 23689 | SDVTList VTs = DAG.getVTList(LHS.getValueType(), MVT::i32); | |||
| 23690 | SDValue Cmp = DAG.getNode(X86ISD::SBB, DL, VTs, LHS, RHS, Carry.getValue(1)); | |||
| 23691 | return getSETCC(CC, Cmp.getValue(1), DL, DAG); | |||
| 23692 | } | |||
| 23693 | ||||
| 23694 | // This function returns three things: the arithmetic computation itself | |||
| 23695 | // (Value), an EFLAGS result (Overflow), and a condition code (Cond). The | |||
| 23696 | // flag and the condition code define the case in which the arithmetic | |||
| 23697 | // computation overflows. | |||
| 23698 | static std::pair<SDValue, SDValue> | |||
| 23699 | getX86XALUOOp(X86::CondCode &Cond, SDValue Op, SelectionDAG &DAG) { | |||
| 23700 | assert(Op.getResNo() == 0 && "Unexpected result number!")((void)0); | |||
| 23701 | SDValue Value, Overflow; | |||
| 23702 | SDValue LHS = Op.getOperand(0); | |||
| 23703 | SDValue RHS = Op.getOperand(1); | |||
| 23704 | unsigned BaseOp = 0; | |||
| 23705 | SDLoc DL(Op); | |||
| 23706 | switch (Op.getOpcode()) { | |||
| 23707 | default: llvm_unreachable("Unknown ovf instruction!")__builtin_unreachable(); | |||
| 23708 | case ISD::SADDO: | |||
| 23709 | BaseOp = X86ISD::ADD; | |||
| 23710 | Cond = X86::COND_O; | |||
| 23711 | break; | |||
| 23712 | case ISD::UADDO: | |||
| 23713 | BaseOp = X86ISD::ADD; | |||
| 23714 | Cond = isOneConstant(RHS) ? X86::COND_E : X86::COND_B; | |||
| 23715 | break; | |||
| 23716 | case ISD::SSUBO: | |||
| 23717 | BaseOp = X86ISD::SUB; | |||
| 23718 | Cond = X86::COND_O; | |||
| 23719 | break; | |||
| 23720 | case ISD::USUBO: | |||
| 23721 | BaseOp = X86ISD::SUB; | |||
| 23722 | Cond = X86::COND_B; | |||
| 23723 | break; | |||
| 23724 | case ISD::SMULO: | |||
| 23725 | BaseOp = X86ISD::SMUL; | |||
| 23726 | Cond = X86::COND_O; | |||
| 23727 | break; | |||
| 23728 | case ISD::UMULO: | |||
| 23729 | BaseOp = X86ISD::UMUL; | |||
| 23730 | Cond = X86::COND_O; | |||
| 23731 | break; | |||
| 23732 | } | |||
| 23733 | ||||
| 23734 | if (BaseOp) { | |||
| 23735 | // Also sets EFLAGS. | |||
| 23736 | SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32); | |||
| 23737 | Value = DAG.getNode(BaseOp, DL, VTs, LHS, RHS); | |||
| 23738 | Overflow = Value.getValue(1); | |||
| 23739 | } | |||
| 23740 | ||||
| 23741 | return std::make_pair(Value, Overflow); | |||
| 23742 | } | |||
| 23743 | ||||
| 23744 | static SDValue LowerXALUO(SDValue Op, SelectionDAG &DAG) { | |||
| 23745 | // Lower the "add/sub/mul with overflow" instruction into a regular ins plus | |||
| 23746 | // a "setcc" instruction that checks the overflow flag. The "brcond" lowering | |||
| 23747 | // looks for this combo and may remove the "setcc" instruction if the "setcc" | |||
| 23748 | // has only one use. | |||
| 23749 | SDLoc DL(Op); | |||
| 23750 | X86::CondCode Cond; | |||
| 23751 | SDValue Value, Overflow; | |||
| 23752 | std::tie(Value, Overflow) = getX86XALUOOp(Cond, Op, DAG); | |||
| 23753 | ||||
| 23754 | SDValue SetCC = getSETCC(Cond, Overflow, DL, DAG); | |||
| 23755 | assert(Op->getValueType(1) == MVT::i8 && "Unexpected VT!")((void)0); | |||
| 23756 | return DAG.getNode(ISD::MERGE_VALUES, DL, Op->getVTList(), Value, SetCC); | |||
| 23757 | } | |||
| 23758 | ||||
| 23759 | /// Return true if opcode is a X86 logical comparison. | |||
| 23760 | static bool isX86LogicalCmp(SDValue Op) { | |||
| 23761 | unsigned Opc = Op.getOpcode(); | |||
| 23762 | if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI || | |||
| 23763 | Opc == X86ISD::FCMP) | |||
| 23764 | return true; | |||
| 23765 | if (Op.getResNo() == 1 && | |||
| 23766 | (Opc == X86ISD::ADD || Opc == X86ISD::SUB || Opc == X86ISD::ADC || | |||
| 23767 | Opc == X86ISD::SBB || Opc == X86ISD::SMUL || Opc == X86ISD::UMUL || | |||
| 23768 | Opc == X86ISD::OR || Opc == X86ISD::XOR || Opc == X86ISD::AND)) | |||
| 23769 | return true; | |||
| 23770 | ||||
| 23771 | return false; | |||
| 23772 | } | |||
| 23773 | ||||
| 23774 | static bool isTruncWithZeroHighBitsInput(SDValue V, SelectionDAG &DAG) { | |||
| 23775 | if (V.getOpcode() != ISD::TRUNCATE) | |||
| 23776 | return false; | |||
| 23777 | ||||
| 23778 | SDValue VOp0 = V.getOperand(0); | |||
| 23779 | unsigned InBits = VOp0.getValueSizeInBits(); | |||
| 23780 | unsigned Bits = V.getValueSizeInBits(); | |||
| 23781 | return DAG.MaskedValueIsZero(VOp0, APInt::getHighBitsSet(InBits,InBits-Bits)); | |||
| 23782 | } | |||
| 23783 | ||||
| 23784 | SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const { | |||
| 23785 | bool AddTest = true; | |||
| 23786 | SDValue Cond = Op.getOperand(0); | |||
| 23787 | SDValue Op1 = Op.getOperand(1); | |||
| 23788 | SDValue Op2 = Op.getOperand(2); | |||
| 23789 | SDLoc DL(Op); | |||
| 23790 | MVT VT = Op1.getSimpleValueType(); | |||
| 23791 | SDValue CC; | |||
| 23792 | ||||
| 23793 | // Lower FP selects into a CMP/AND/ANDN/OR sequence when the necessary SSE ops | |||
| 23794 | // are available or VBLENDV if AVX is available. | |||
| 23795 | // Otherwise FP cmovs get lowered into a less efficient branch sequence later. | |||
| 23796 | if (Cond.getOpcode() == ISD::SETCC && isScalarFPTypeInSSEReg(VT) && | |||
| 23797 | VT == Cond.getOperand(0).getSimpleValueType() && Cond->hasOneUse()) { | |||
| 23798 | SDValue CondOp0 = Cond.getOperand(0), CondOp1 = Cond.getOperand(1); | |||
| 23799 | bool IsAlwaysSignaling; | |||
| 23800 | unsigned SSECC = | |||
| 23801 | translateX86FSETCC(cast<CondCodeSDNode>(Cond.getOperand(2))->get(), | |||
| 23802 | CondOp0, CondOp1, IsAlwaysSignaling); | |||
| 23803 | ||||
| 23804 | if (Subtarget.hasAVX512()) { | |||
| 23805 | SDValue Cmp = | |||
| 23806 | DAG.getNode(X86ISD::FSETCCM, DL, MVT::v1i1, CondOp0, CondOp1, | |||
| 23807 | DAG.getTargetConstant(SSECC, DL, MVT::i8)); | |||
| 23808 | assert(!VT.isVector() && "Not a scalar type?")((void)0); | |||
| 23809 | return DAG.getNode(X86ISD::SELECTS, DL, VT, Cmp, Op1, Op2); | |||
| 23810 | } | |||
| 23811 | ||||
| 23812 | if (SSECC < 8 || Subtarget.hasAVX()) { | |||
| 23813 | SDValue Cmp = DAG.getNode(X86ISD::FSETCC, DL, VT, CondOp0, CondOp1, | |||
| 23814 | DAG.getTargetConstant(SSECC, DL, MVT::i8)); | |||
| 23815 | ||||
| 23816 | // If we have AVX, we can use a variable vector select (VBLENDV) instead | |||
| 23817 | // of 3 logic instructions for size savings and potentially speed. | |||
| 23818 | // Unfortunately, there is no scalar form of VBLENDV. | |||
| 23819 | ||||
| 23820 | // If either operand is a +0.0 constant, don't try this. We can expect to | |||
| 23821 | // optimize away at least one of the logic instructions later in that | |||
| 23822 | // case, so that sequence would be faster than a variable blend. | |||
| 23823 | ||||
| 23824 | // BLENDV was introduced with SSE 4.1, but the 2 register form implicitly | |||
| 23825 | // uses XMM0 as the selection register. That may need just as many | |||
| 23826 | // instructions as the AND/ANDN/OR sequence due to register moves, so | |||
| 23827 | // don't bother. | |||
| 23828 | if (Subtarget.hasAVX() && !isNullFPConstant(Op1) && | |||
| 23829 | !isNullFPConstant(Op2)) { | |||
| 23830 | // Convert to vectors, do a VSELECT, and convert back to scalar. | |||
| 23831 | // All of the conversions should be optimized away. | |||
| 23832 | MVT VecVT = VT == MVT::f32 ? MVT::v4f32 : MVT::v2f64; | |||
| 23833 | SDValue VOp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, Op1); | |||
| 23834 | SDValue VOp2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, Op2); | |||
| 23835 | SDValue VCmp = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, Cmp); | |||
| 23836 | ||||
| 23837 | MVT VCmpVT = VT == MVT::f32 ? MVT::v4i32 : MVT::v2i64; | |||
| 23838 | VCmp = DAG.getBitcast(VCmpVT, VCmp); | |||
| 23839 | ||||
| 23840 | SDValue VSel = DAG.getSelect(DL, VecVT, VCmp, VOp1, VOp2); | |||
| 23841 | ||||
| 23842 | return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT, | |||
| 23843 | VSel, DAG.getIntPtrConstant(0, DL)); | |||
| 23844 | } | |||
| 23845 | SDValue AndN = DAG.getNode(X86ISD::FANDN, DL, VT, Cmp, Op2); | |||
| 23846 | SDValue And = DAG.getNode(X86ISD::FAND, DL, VT, Cmp, Op1); | |||
| 23847 | return DAG.getNode(X86ISD::FOR, DL, VT, AndN, And); | |||
| 23848 | } | |||
| 23849 | } | |||
| 23850 | ||||
| 23851 | // AVX512 fallback is to lower selects of scalar floats to masked moves. | |||
| 23852 | if (isScalarFPTypeInSSEReg(VT) && Subtarget.hasAVX512()) { | |||
| 23853 | SDValue Cmp = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v1i1, Cond); | |||
| 23854 | return DAG.getNode(X86ISD::SELECTS, DL, VT, Cmp, Op1, Op2); | |||
| 23855 | } | |||
| 23856 | ||||
| 23857 | if (Cond.getOpcode() == ISD::SETCC) { | |||
| 23858 | if (SDValue NewCond = LowerSETCC(Cond, DAG)) { | |||
| 23859 | Cond = NewCond; | |||
| 23860 | // If the condition was updated, it's possible that the operands of the | |||
| 23861 | // select were also updated (for example, EmitTest has a RAUW). Refresh | |||
| 23862 | // the local references to the select operands in case they got stale. | |||
| 23863 | Op1 = Op.getOperand(1); | |||
| 23864 | Op2 = Op.getOperand(2); | |||
| 23865 | } | |||
| 23866 | } | |||
| 23867 | ||||
| 23868 | // (select (x == 0), -1, y) -> (sign_bit (x - 1)) | y | |||
| 23869 | // (select (x == 0), y, -1) -> ~(sign_bit (x - 1)) | y | |||
| 23870 | // (select (x != 0), y, -1) -> (sign_bit (x - 1)) | y | |||
| 23871 | // (select (x != 0), -1, y) -> ~(sign_bit (x - 1)) | y | |||
| 23872 | // (select (and (x , 0x1) == 0), y, (z ^ y) ) -> (-(and (x , 0x1)) & z ) ^ y | |||
| 23873 | // (select (and (x , 0x1) == 0), y, (z | y) ) -> (-(and (x , 0x1)) & z ) | y | |||
| 23874 | if (Cond.getOpcode() == X86ISD::SETCC && | |||
| 23875 | Cond.getOperand(1).getOpcode() == X86ISD::CMP && | |||
| 23876 | isNullConstant(Cond.getOperand(1).getOperand(1))) { | |||
| 23877 | SDValue Cmp = Cond.getOperand(1); | |||
| 23878 | SDValue CmpOp0 = Cmp.getOperand(0); | |||
| 23879 | unsigned CondCode = Cond.getConstantOperandVal(0); | |||
| 23880 | ||||
| 23881 | // Special handling for __builtin_ffs(X) - 1 pattern which looks like | |||
| 23882 | // (select (seteq X, 0), -1, (cttz_zero_undef X)). Disable the special | |||
| 23883 | // handle to keep the CMP with 0. This should be removed by | |||
| 23884 | // optimizeCompareInst by using the flags from the BSR/TZCNT used for the | |||
| 23885 | // cttz_zero_undef. | |||
| 23886 | auto MatchFFSMinus1 = [&](SDValue Op1, SDValue Op2) { | |||
| 23887 | return (Op1.getOpcode() == ISD::CTTZ_ZERO_UNDEF && Op1.hasOneUse() && | |||
| 23888 | Op1.getOperand(0) == CmpOp0 && isAllOnesConstant(Op2)); | |||
| 23889 | }; | |||
| 23890 | if (Subtarget.hasCMov() && (VT == MVT::i32 || VT == MVT::i64) && | |||
| 23891 | ((CondCode == X86::COND_NE && MatchFFSMinus1(Op1, Op2)) || | |||
| 23892 | (CondCode == X86::COND_E && MatchFFSMinus1(Op2, Op1)))) { | |||
| 23893 | // Keep Cmp. | |||
| 23894 | } else if ((isAllOnesConstant(Op1) || isAllOnesConstant(Op2)) && | |||
| 23895 | (CondCode == X86::COND_E || CondCode == X86::COND_NE)) { | |||
| 23896 | SDValue Y = isAllOnesConstant(Op2) ? Op1 : Op2; | |||
| 23897 | ||||
| 23898 | SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32); | |||
| 23899 | SDVTList CmpVTs = DAG.getVTList(CmpOp0.getValueType(), MVT::i32); | |||
| 23900 | ||||
| 23901 | // Apply further optimizations for special cases | |||
| 23902 | // (select (x != 0), -1, 0) -> neg & sbb | |||
| 23903 | // (select (x == 0), 0, -1) -> neg & sbb | |||
| 23904 | if (isNullConstant(Y) && | |||
| 23905 | (isAllOnesConstant(Op1) == (CondCode == X86::COND_NE))) { | |||
| 23906 | SDValue Zero = DAG.getConstant(0, DL, CmpOp0.getValueType()); | |||
| 23907 | SDValue Neg = DAG.getNode(X86ISD::SUB, DL, CmpVTs, Zero, CmpOp0); | |||
| 23908 | Zero = DAG.getConstant(0, DL, Op.getValueType()); | |||
| 23909 | return DAG.getNode(X86ISD::SBB, DL, VTs, Zero, Zero, Neg.getValue(1)); | |||
| 23910 | } | |||
| 23911 | ||||
| 23912 | Cmp = DAG.getNode(X86ISD::SUB, DL, CmpVTs, | |||
| 23913 | CmpOp0, DAG.getConstant(1, DL, CmpOp0.getValueType())); | |||
| 23914 | ||||
| 23915 | SDValue Zero = DAG.getConstant(0, DL, Op.getValueType()); | |||
| 23916 | SDValue Res = // Res = 0 or -1. | |||
| 23917 | DAG.getNode(X86ISD::SBB, DL, VTs, Zero, Zero, Cmp.getValue(1)); | |||
| 23918 | ||||
| 23919 | if (isAllOnesConstant(Op1) != (CondCode == X86::COND_E)) | |||
| 23920 | Res = DAG.getNOT(DL, Res, Res.getValueType()); | |||
| 23921 | ||||
| 23922 | return DAG.getNode(ISD::OR, DL, Res.getValueType(), Res, Y); | |||
| 23923 | } else if (!Subtarget.hasCMov() && CondCode == X86::COND_E && | |||
| 23924 | Cmp.getOperand(0).getOpcode() == ISD::AND && | |||
| 23925 | isOneConstant(Cmp.getOperand(0).getOperand(1))) { | |||
| 23926 | SDValue Src1, Src2; | |||
| 23927 | // true if Op2 is XOR or OR operator and one of its operands | |||
| 23928 | // is equal to Op1 | |||
| 23929 | // ( a , a op b) || ( b , a op b) | |||
| 23930 | auto isOrXorPattern = [&]() { | |||
| 23931 | if ((Op2.getOpcode() == ISD::XOR || Op2.getOpcode() == ISD::OR) && | |||
| 23932 | (Op2.getOperand(0) == Op1 || Op2.getOperand(1) == Op1)) { | |||
| 23933 | Src1 = | |||
| 23934 | Op2.getOperand(0) == Op1 ? Op2.getOperand(1) : Op2.getOperand(0); | |||
| 23935 | Src2 = Op1; | |||
| 23936 | return true; | |||
| 23937 | } | |||
| 23938 | return false; | |||
| 23939 | }; | |||
| 23940 | ||||
| 23941 | if (isOrXorPattern()) { | |||
| 23942 | SDValue Neg; | |||
| 23943 | unsigned int CmpSz = CmpOp0.getSimpleValueType().getSizeInBits(); | |||
| 23944 | // we need mask of all zeros or ones with same size of the other | |||
| 23945 | // operands. | |||
| 23946 | if (CmpSz > VT.getSizeInBits()) | |||
| 23947 | Neg = DAG.getNode(ISD::TRUNCATE, DL, VT, CmpOp0); | |||
| 23948 | else if (CmpSz < VT.getSizeInBits()) | |||
| 23949 | Neg = DAG.getNode(ISD::AND, DL, VT, | |||
| 23950 | DAG.getNode(ISD::ANY_EXTEND, DL, VT, CmpOp0.getOperand(0)), | |||
| 23951 | DAG.getConstant(1, DL, VT)); | |||
| 23952 | else | |||
| 23953 | Neg = CmpOp0; | |||
| 23954 | SDValue Mask = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, DL, VT), | |||
| 23955 | Neg); // -(and (x, 0x1)) | |||
| 23956 | SDValue And = DAG.getNode(ISD::AND, DL, VT, Mask, Src1); // Mask & z | |||
| 23957 | return DAG.getNode(Op2.getOpcode(), DL, VT, And, Src2); // And Op y | |||
| 23958 | } | |||
| 23959 | } | |||
| 23960 | } | |||
| 23961 | ||||
| 23962 | // Look past (and (setcc_carry (cmp ...)), 1). | |||
| 23963 | if (Cond.getOpcode() == ISD::AND && | |||
| 23964 | Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY && | |||
| 23965 | isOneConstant(Cond.getOperand(1))) | |||
| 23966 | Cond = Cond.getOperand(0); | |||
| 23967 | ||||
| 23968 | // If condition flag is set by a X86ISD::CMP, then use it as the condition | |||
| 23969 | // setting operand in place of the X86ISD::SETCC. | |||
| 23970 | unsigned CondOpcode = Cond.getOpcode(); | |||
| 23971 | if (CondOpcode == X86ISD::SETCC || | |||
| 23972 | CondOpcode == X86ISD::SETCC_CARRY) { | |||
| 23973 | CC = Cond.getOperand(0); | |||
| 23974 | ||||
| 23975 | SDValue Cmp = Cond.getOperand(1); | |||
| 23976 | bool IllegalFPCMov = false; | |||
| 23977 | if (VT.isFloatingPoint() && !VT.isVector() && | |||
| 23978 | !isScalarFPTypeInSSEReg(VT) && Subtarget.hasCMov()) // FPStack? | |||
| 23979 | IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue()); | |||
| 23980 | ||||
| 23981 | if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) || | |||
| 23982 | Cmp.getOpcode() == X86ISD::BT) { // FIXME | |||
| 23983 | Cond = Cmp; | |||
| 23984 | AddTest = false; | |||
| 23985 | } | |||
| 23986 | } else if (CondOpcode == ISD::USUBO || CondOpcode == ISD::SSUBO || | |||
| 23987 | CondOpcode == ISD::UADDO || CondOpcode == ISD::SADDO || | |||
| 23988 | CondOpcode == ISD::UMULO || CondOpcode == ISD::SMULO) { | |||
| 23989 | SDValue Value; | |||
| 23990 | X86::CondCode X86Cond; | |||
| 23991 | std::tie(Value, Cond) = getX86XALUOOp(X86Cond, Cond.getValue(0), DAG); | |||
| 23992 | ||||
| 23993 | CC = DAG.getTargetConstant(X86Cond, DL, MVT::i8); | |||
| 23994 | AddTest = false; | |||
| 23995 | } | |||
| 23996 | ||||
| 23997 | if (AddTest) { | |||
| 23998 | // Look past the truncate if the high bits are known zero. | |||
| 23999 | if (isTruncWithZeroHighBitsInput(Cond, DAG)) | |||
| 24000 | Cond = Cond.getOperand(0); | |||
| 24001 | ||||
| 24002 | // We know the result of AND is compared against zero. Try to match | |||
| 24003 | // it to BT. | |||
| 24004 | if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) { | |||
| 24005 | SDValue BTCC; | |||
| 24006 | if (SDValue BT = LowerAndToBT(Cond, ISD::SETNE, DL, DAG, BTCC)) { | |||
| 24007 | CC = BTCC; | |||
| 24008 | Cond = BT; | |||
| 24009 | AddTest = false; | |||
| 24010 | } | |||
| 24011 | } | |||
| 24012 | } | |||
| 24013 | ||||
| 24014 | if (AddTest) { | |||
| 24015 | CC = DAG.getTargetConstant(X86::COND_NE, DL, MVT::i8); | |||
| 24016 | Cond = EmitTest(Cond, X86::COND_NE, DL, DAG, Subtarget); | |||
| 24017 | } | |||
| 24018 | ||||
| 24019 | // a < b ? -1 : 0 -> RES = ~setcc_carry | |||
| 24020 | // a < b ? 0 : -1 -> RES = setcc_carry | |||
| 24021 | // a >= b ? -1 : 0 -> RES = setcc_carry | |||
| 24022 | // a >= b ? 0 : -1 -> RES = ~setcc_carry | |||
| 24023 | if (Cond.getOpcode() == X86ISD::SUB) { | |||
| 24024 | unsigned CondCode = cast<ConstantSDNode>(CC)->getZExtValue(); | |||
| 24025 | ||||
| 24026 | if ((CondCode == X86::COND_AE || CondCode == X86::COND_B) && | |||
| 24027 | (isAllOnesConstant(Op1) || isAllOnesConstant(Op2)) && | |||
| 24028 | (isNullConstant(Op1) || isNullConstant(Op2))) { | |||
| 24029 | SDValue Res = | |||
| 24030 | DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(), | |||
| 24031 | DAG.getTargetConstant(X86::COND_B, DL, MVT::i8), Cond); | |||
| 24032 | if (isAllOnesConstant(Op1) != (CondCode == X86::COND_B)) | |||
| 24033 | return DAG.getNOT(DL, Res, Res.getValueType()); | |||
| 24034 | return Res; | |||
| 24035 | } | |||
| 24036 | } | |||
| 24037 | ||||
| 24038 | // X86 doesn't have an i8 cmov. If both operands are the result of a truncate | |||
| 24039 | // widen the cmov and push the truncate through. This avoids introducing a new | |||
| 24040 | // branch during isel and doesn't add any extensions. | |||
| 24041 | if (Op.getValueType() == MVT::i8 && | |||
| 24042 | Op1.getOpcode() == ISD::TRUNCATE && Op2.getOpcode() == ISD::TRUNCATE) { | |||
| 24043 | SDValue T1 = Op1.getOperand(0), T2 = Op2.getOperand(0); | |||
| 24044 | if (T1.getValueType() == T2.getValueType() && | |||
| 24045 | // Exclude CopyFromReg to avoid partial register stalls. | |||
| 24046 | T1.getOpcode() != ISD::CopyFromReg && T2.getOpcode()!=ISD::CopyFromReg){ | |||
| 24047 | SDValue Cmov = DAG.getNode(X86ISD::CMOV, DL, T1.getValueType(), T2, T1, | |||
| 24048 | CC, Cond); | |||
| 24049 | return DAG.getNode(ISD::TRUNCATE, DL, Op.getValueType(), Cmov); | |||
| 24050 | } | |||
| 24051 | } | |||
| 24052 | ||||
| 24053 | // Or finally, promote i8 cmovs if we have CMOV, | |||
| 24054 | // or i16 cmovs if it won't prevent folding a load. | |||
| 24055 | // FIXME: we should not limit promotion of i8 case to only when the CMOV is | |||
| 24056 | // legal, but EmitLoweredSelect() can not deal with these extensions | |||
| 24057 | // being inserted between two CMOV's. (in i16 case too TBN) | |||
| 24058 | // https://bugs.llvm.org/show_bug.cgi?id=40974 | |||
| 24059 | if ((Op.getValueType() == MVT::i8 && Subtarget.hasCMov()) || | |||
| 24060 | (Op.getValueType() == MVT::i16 && !MayFoldLoad(Op1) && | |||
| 24061 | !MayFoldLoad(Op2))) { | |||
| 24062 | Op1 = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, Op1); | |||
| 24063 | Op2 = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, Op2); | |||
| 24064 | SDValue Ops[] = { Op2, Op1, CC, Cond }; | |||
| 24065 | SDValue Cmov = DAG.getNode(X86ISD::CMOV, DL, MVT::i32, Ops); | |||
| 24066 | return DAG.getNode(ISD::TRUNCATE, DL, Op.getValueType(), Cmov); | |||
| 24067 | } | |||
| 24068 | ||||
| 24069 | // X86ISD::CMOV means set the result (which is operand 1) to the RHS if | |||
| 24070 | // condition is true. | |||
| 24071 | SDValue Ops[] = { Op2, Op1, CC, Cond }; | |||
| 24072 | return DAG.getNode(X86ISD::CMOV, DL, Op.getValueType(), Ops); | |||
| 24073 | } | |||
| 24074 | ||||
| 24075 | static SDValue LowerSIGN_EXTEND_Mask(SDValue Op, | |||
| 24076 | const X86Subtarget &Subtarget, | |||
| 24077 | SelectionDAG &DAG) { | |||
| 24078 | MVT VT = Op->getSimpleValueType(0); | |||
| 24079 | SDValue In = Op->getOperand(0); | |||
| 24080 | MVT InVT = In.getSimpleValueType(); | |||
| 24081 | assert(InVT.getVectorElementType() == MVT::i1 && "Unexpected input type!")((void)0); | |||
| 24082 | MVT VTElt = VT.getVectorElementType(); | |||
| 24083 | SDLoc dl(Op); | |||
| 24084 | ||||
| 24085 | unsigned NumElts = VT.getVectorNumElements(); | |||
| 24086 | ||||
| 24087 | // Extend VT if the scalar type is i8/i16 and BWI is not supported. | |||
| 24088 | MVT ExtVT = VT; | |||
| 24089 | if (!Subtarget.hasBWI() && VTElt.getSizeInBits() <= 16) { | |||
| 24090 | // If v16i32 is to be avoided, we'll need to split and concatenate. | |||
| 24091 | if (NumElts == 16 && !Subtarget.canExtendTo512DQ()) | |||
| 24092 | return SplitAndExtendv16i1(Op.getOpcode(), VT, In, dl, DAG); | |||
| 24093 | ||||
| 24094 | ExtVT = MVT::getVectorVT(MVT::i32, NumElts); | |||
| 24095 | } | |||
| 24096 | ||||
| 24097 | // Widen to 512-bits if VLX is not supported. | |||
| 24098 | MVT WideVT = ExtVT; | |||
| 24099 | if (!ExtVT.is512BitVector() && !Subtarget.hasVLX()) { | |||
| 24100 | NumElts *= 512 / ExtVT.getSizeInBits(); | |||
| 24101 | InVT = MVT::getVectorVT(MVT::i1, NumElts); | |||
| 24102 | In = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, InVT, DAG.getUNDEF(InVT), | |||
| 24103 | In, DAG.getIntPtrConstant(0, dl)); | |||
| 24104 | WideVT = MVT::getVectorVT(ExtVT.getVectorElementType(), NumElts); | |||
| 24105 | } | |||
| 24106 | ||||
| 24107 | SDValue V; | |||
| 24108 | MVT WideEltVT = WideVT.getVectorElementType(); | |||
| 24109 | if ((Subtarget.hasDQI() && WideEltVT.getSizeInBits() >= 32) || | |||
| 24110 | (Subtarget.hasBWI() && WideEltVT.getSizeInBits() <= 16)) { | |||
| 24111 | V = DAG.getNode(Op.getOpcode(), dl, WideVT, In); | |||
| 24112 | } else { | |||
| 24113 | SDValue NegOne = DAG.getConstant(-1, dl, WideVT); | |||
| 24114 | SDValue Zero = DAG.getConstant(0, dl, WideVT); | |||
| 24115 | V = DAG.getSelect(dl, WideVT, In, NegOne, Zero); | |||
| 24116 | } | |||
| 24117 | ||||
| 24118 | // Truncate if we had to extend i16/i8 above. | |||
| 24119 | if (VT != ExtVT) { | |||
| 24120 | WideVT = MVT::getVectorVT(VTElt, NumElts); | |||
| 24121 | V = DAG.getNode(ISD::TRUNCATE, dl, WideVT, V); | |||
| 24122 | } | |||
| 24123 | ||||
| 24124 | // Extract back to 128/256-bit if we widened. | |||
| 24125 | if (WideVT != VT) | |||
| 24126 | V = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, V, | |||
| 24127 | DAG.getIntPtrConstant(0, dl)); | |||
| 24128 | ||||
| 24129 | return V; | |||
| 24130 | } | |||
| 24131 | ||||
| 24132 | static SDValue LowerANY_EXTEND(SDValue Op, const X86Subtarget &Subtarget, | |||
| 24133 | SelectionDAG &DAG) { | |||
| 24134 | SDValue In = Op->getOperand(0); | |||
| 24135 | MVT InVT = In.getSimpleValueType(); | |||
| 24136 | ||||
| 24137 | if (InVT.getVectorElementType() == MVT::i1) | |||
| 24138 | return LowerSIGN_EXTEND_Mask(Op, Subtarget, DAG); | |||
| 24139 | ||||
| 24140 | assert(Subtarget.hasAVX() && "Expected AVX support")((void)0); | |||
| 24141 | return LowerAVXExtend(Op, DAG, Subtarget); | |||
| 24142 | } | |||
| 24143 | ||||
| 24144 | // Lowering for SIGN_EXTEND_VECTOR_INREG and ZERO_EXTEND_VECTOR_INREG. | |||
| 24145 | // For sign extend this needs to handle all vector sizes and SSE4.1 and | |||
| 24146 | // non-SSE4.1 targets. For zero extend this should only handle inputs of | |||
| 24147 | // MVT::v64i8 when BWI is not supported, but AVX512 is. | |||
| 24148 | static SDValue LowerEXTEND_VECTOR_INREG(SDValue Op, | |||
| 24149 | const X86Subtarget &Subtarget, | |||
| 24150 | SelectionDAG &DAG) { | |||
| 24151 | SDValue In = Op->getOperand(0); | |||
| 24152 | MVT VT = Op->getSimpleValueType(0); | |||
| 24153 | MVT InVT = In.getSimpleValueType(); | |||
| 24154 | ||||
| 24155 | MVT SVT = VT.getVectorElementType(); | |||
| 24156 | MVT InSVT = InVT.getVectorElementType(); | |||
| 24157 | assert(SVT.getFixedSizeInBits() > InSVT.getFixedSizeInBits())((void)0); | |||
| 24158 | ||||
| 24159 | if (SVT != MVT::i64 && SVT != MVT::i32 && SVT != MVT::i16) | |||
| 24160 | return SDValue(); | |||
| 24161 | if (InSVT != MVT::i32 && InSVT != MVT::i16 && InSVT != MVT::i8) | |||
| 24162 | return SDValue(); | |||
| 24163 | if (!(VT.is128BitVector() && Subtarget.hasSSE2()) && | |||
| 24164 | !(VT.is256BitVector() && Subtarget.hasAVX()) && | |||
| 24165 | !(VT.is512BitVector() && Subtarget.hasAVX512())) | |||
| 24166 | return SDValue(); | |||
| 24167 | ||||
| 24168 | SDLoc dl(Op); | |||
| 24169 | unsigned Opc = Op.getOpcode(); | |||
| 24170 | unsigned NumElts = VT.getVectorNumElements(); | |||
| 24171 | ||||
| 24172 | // For 256-bit vectors, we only need the lower (128-bit) half of the input. | |||
| 24173 | // For 512-bit vectors, we need 128-bits or 256-bits. | |||
| 24174 | if (InVT.getSizeInBits() > 128) { | |||
| 24175 | // Input needs to be at least the same number of elements as output, and | |||
| 24176 | // at least 128-bits. | |||
| 24177 | int InSize = InSVT.getSizeInBits() * NumElts; | |||
| 24178 | In = extractSubVector(In, 0, DAG, dl, std::max(InSize, 128)); | |||
| 24179 | InVT = In.getSimpleValueType(); | |||
| 24180 | } | |||
| 24181 | ||||
| 24182 | // SSE41 targets can use the pmov[sz]x* instructions directly for 128-bit results, | |||
| 24183 | // so are legal and shouldn't occur here. AVX2/AVX512 pmovsx* instructions still | |||
| 24184 | // need to be handled here for 256/512-bit results. | |||
| 24185 | if (Subtarget.hasInt256()) { | |||
| 24186 | assert(VT.getSizeInBits() > 128 && "Unexpected 128-bit vector extension")((void)0); | |||
| 24187 | ||||
| 24188 | if (InVT.getVectorNumElements() != NumElts) | |||
| 24189 | return DAG.getNode(Op.getOpcode(), dl, VT, In); | |||
| 24190 | ||||
| 24191 | // FIXME: Apparently we create inreg operations that could be regular | |||
| 24192 | // extends. | |||
| 24193 | unsigned ExtOpc = | |||
| 24194 | Opc == ISD::SIGN_EXTEND_VECTOR_INREG ? ISD::SIGN_EXTEND | |||
| 24195 | : ISD::ZERO_EXTEND; | |||
| 24196 | return DAG.getNode(ExtOpc, dl, VT, In); | |||
| 24197 | } | |||
| 24198 | ||||
| 24199 | // pre-AVX2 256-bit extensions need to be split into 128-bit instructions. | |||
| 24200 | if (Subtarget.hasAVX()) { | |||
| 24201 | assert(VT.is256BitVector() && "256-bit vector expected")((void)0); | |||
| 24202 | MVT HalfVT = VT.getHalfNumVectorElementsVT(); | |||
| 24203 | int HalfNumElts = HalfVT.getVectorNumElements(); | |||
| 24204 | ||||
| 24205 | unsigned NumSrcElts = InVT.getVectorNumElements(); | |||
| 24206 | SmallVector<int, 16> HiMask(NumSrcElts, SM_SentinelUndef); | |||
| 24207 | for (int i = 0; i != HalfNumElts; ++i) | |||
| 24208 | HiMask[i] = HalfNumElts + i; | |||
| 24209 | ||||
| 24210 | SDValue Lo = DAG.getNode(Opc, dl, HalfVT, In); | |||
| 24211 | SDValue Hi = DAG.getVectorShuffle(InVT, dl, In, DAG.getUNDEF(InVT), HiMask); | |||
| 24212 | Hi = DAG.getNode(Opc, dl, HalfVT, Hi); | |||
| 24213 | return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, Lo, Hi); | |||
| 24214 | } | |||
| 24215 | ||||
| 24216 | // We should only get here for sign extend. | |||
| 24217 | assert(Opc == ISD::SIGN_EXTEND_VECTOR_INREG && "Unexpected opcode!")((void)0); | |||
| 24218 | assert(VT.is128BitVector() && InVT.is128BitVector() && "Unexpected VTs")((void)0); | |||
| 24219 | ||||
| 24220 | // pre-SSE41 targets unpack lower lanes and then sign-extend using SRAI. | |||
| 24221 | SDValue Curr = In; | |||
| 24222 | SDValue SignExt = Curr; | |||
| 24223 | ||||
| 24224 | // As SRAI is only available on i16/i32 types, we expand only up to i32 | |||
| 24225 | // and handle i64 separately. | |||
| 24226 | if (InVT != MVT::v4i32) { | |||
| 24227 | MVT DestVT = VT == MVT::v2i64 ? MVT::v4i32 : VT; | |||
| 24228 | ||||
| 24229 | unsigned DestWidth = DestVT.getScalarSizeInBits(); | |||
| 24230 | unsigned Scale = DestWidth / InSVT.getSizeInBits(); | |||
| 24231 | ||||
| 24232 | unsigned InNumElts = InVT.getVectorNumElements(); | |||
| 24233 | unsigned DestElts = DestVT.getVectorNumElements(); | |||
| 24234 | ||||
| 24235 | // Build a shuffle mask that takes each input element and places it in the | |||
| 24236 | // MSBs of the new element size. | |||
| 24237 | SmallVector<int, 16> Mask(InNumElts, SM_SentinelUndef); | |||
| 24238 | for (unsigned i = 0; i != DestElts; ++i) | |||
| 24239 | Mask[i * Scale + (Scale - 1)] = i; | |||
| 24240 | ||||
| 24241 | Curr = DAG.getVectorShuffle(InVT, dl, In, In, Mask); | |||
| 24242 | Curr = DAG.getBitcast(DestVT, Curr); | |||
| 24243 | ||||
| 24244 | unsigned SignExtShift = DestWidth - InSVT.getSizeInBits(); | |||
| 24245 | SignExt = DAG.getNode(X86ISD::VSRAI, dl, DestVT, Curr, | |||
| 24246 | DAG.getTargetConstant(SignExtShift, dl, MVT::i8)); | |||
| 24247 | } | |||
| 24248 | ||||
| 24249 | if (VT == MVT::v2i64) { | |||
| 24250 | assert(Curr.getValueType() == MVT::v4i32 && "Unexpected input VT")((void)0); | |||
| 24251 | SDValue Zero = DAG.getConstant(0, dl, MVT::v4i32); | |||
| 24252 | SDValue Sign = DAG.getSetCC(dl, MVT::v4i32, Zero, Curr, ISD::SETGT); | |||
| 24253 | SignExt = DAG.getVectorShuffle(MVT::v4i32, dl, SignExt, Sign, {0, 4, 1, 5}); | |||
| 24254 | SignExt = DAG.getBitcast(VT, SignExt); | |||
| 24255 | } | |||
| 24256 | ||||
| 24257 | return SignExt; | |||
| 24258 | } | |||
| 24259 | ||||
| 24260 | static SDValue LowerSIGN_EXTEND(SDValue Op, const X86Subtarget &Subtarget, | |||
| 24261 | SelectionDAG &DAG) { | |||
| 24262 | MVT VT = Op->getSimpleValueType(0); | |||
| 24263 | SDValue In = Op->getOperand(0); | |||
| 24264 | MVT InVT = In.getSimpleValueType(); | |||
| 24265 | SDLoc dl(Op); | |||
| 24266 | ||||
| 24267 | if (InVT.getVectorElementType() == MVT::i1) | |||
| 24268 | return LowerSIGN_EXTEND_Mask(Op, Subtarget, DAG); | |||
| 24269 | ||||
| 24270 | assert(VT.isVector() && InVT.isVector() && "Expected vector type")((void)0); | |||
| 24271 | assert(VT.getVectorNumElements() == InVT.getVectorNumElements() &&((void)0) | |||
| 24272 | "Expected same number of elements")((void)0); | |||
| 24273 | assert((VT.getVectorElementType() == MVT::i16 ||((void)0) | |||
| 24274 | VT.getVectorElementType() == MVT::i32 ||((void)0) | |||
| 24275 | VT.getVectorElementType() == MVT::i64) &&((void)0) | |||
| 24276 | "Unexpected element type")((void)0); | |||
| 24277 | assert((InVT.getVectorElementType() == MVT::i8 ||((void)0) | |||
| 24278 | InVT.getVectorElementType() == MVT::i16 ||((void)0) | |||
| 24279 | InVT.getVectorElementType() == MVT::i32) &&((void)0) | |||
| 24280 | "Unexpected element type")((void)0); | |||
| 24281 | ||||
| 24282 | if (VT == MVT::v32i16 && !Subtarget.hasBWI()) { | |||
| 24283 | assert(InVT == MVT::v32i8 && "Unexpected VT!")((void)0); | |||
| 24284 | return splitVectorIntUnary(Op, DAG); | |||
| 24285 | } | |||
| 24286 | ||||
| 24287 | if (Subtarget.hasInt256()) | |||
| 24288 | return Op; | |||
| 24289 | ||||
| 24290 | // Optimize vectors in AVX mode | |||
| 24291 | // Sign extend v8i16 to v8i32 and | |||
| 24292 | // v4i32 to v4i64 | |||
| 24293 | // | |||
| 24294 | // Divide input vector into two parts | |||
| 24295 | // for v4i32 the high shuffle mask will be {2, 3, -1, -1} | |||
| 24296 | // use vpmovsx instruction to extend v4i32 -> v2i64; v8i16 -> v4i32 | |||
| 24297 | // concat the vectors to original VT | |||
| 24298 | MVT HalfVT = VT.getHalfNumVectorElementsVT(); | |||
| 24299 | SDValue OpLo = DAG.getNode(ISD::SIGN_EXTEND_VECTOR_INREG, dl, HalfVT, In); | |||
| 24300 | ||||
| 24301 | unsigned NumElems = InVT.getVectorNumElements(); | |||
| 24302 | SmallVector<int,8> ShufMask(NumElems, -1); | |||
| 24303 | for (unsigned i = 0; i != NumElems/2; ++i) | |||
| 24304 | ShufMask[i] = i + NumElems/2; | |||
| 24305 | ||||
| 24306 | SDValue OpHi = DAG.getVectorShuffle(InVT, dl, In, In, ShufMask); | |||
| 24307 | OpHi = DAG.getNode(ISD::SIGN_EXTEND_VECTOR_INREG, dl, HalfVT, OpHi); | |||
| 24308 | ||||
| 24309 | return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi); | |||
| 24310 | } | |||
| 24311 | ||||
| 24312 | /// Change a vector store into a pair of half-size vector stores. | |||
| 24313 | static SDValue splitVectorStore(StoreSDNode *Store, SelectionDAG &DAG) { | |||
| 24314 | SDValue StoredVal = Store->getValue(); | |||
| 24315 | assert((StoredVal.getValueType().is256BitVector() ||((void)0) | |||
| 24316 | StoredVal.getValueType().is512BitVector()) &&((void)0) | |||
| 24317 | "Expecting 256/512-bit op")((void)0); | |||
| 24318 | ||||
| 24319 | // Splitting volatile memory ops is not allowed unless the operation was not | |||
| 24320 | // legal to begin with. Assume the input store is legal (this transform is | |||
| 24321 | // only used for targets with AVX). Note: It is possible that we have an | |||
| 24322 | // illegal type like v2i128, and so we could allow splitting a volatile store | |||
| 24323 | // in that case if that is important. | |||
| 24324 | if (!Store->isSimple()) | |||
| 24325 | return SDValue(); | |||
| 24326 | ||||
| 24327 | SDLoc DL(Store); | |||
| 24328 | SDValue Value0, Value1; | |||
| 24329 | std::tie(Value0, Value1) = splitVector(StoredVal, DAG, DL); | |||
| 24330 | unsigned HalfOffset = Value0.getValueType().getStoreSize(); | |||
| 24331 | SDValue Ptr0 = Store->getBasePtr(); | |||
| 24332 | SDValue Ptr1 = | |||
| 24333 | DAG.getMemBasePlusOffset(Ptr0, TypeSize::Fixed(HalfOffset), DL); | |||
| 24334 | SDValue Ch0 = | |||
| 24335 | DAG.getStore(Store->getChain(), DL, Value0, Ptr0, Store->getPointerInfo(), | |||
| 24336 | Store->getOriginalAlign(), | |||
| 24337 | Store->getMemOperand()->getFlags()); | |||
| 24338 | SDValue Ch1 = DAG.getStore(Store->getChain(), DL, Value1, Ptr1, | |||
| 24339 | Store->getPointerInfo().getWithOffset(HalfOffset), | |||
| 24340 | Store->getOriginalAlign(), | |||
| 24341 | Store->getMemOperand()->getFlags()); | |||
| 24342 | return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Ch0, Ch1); | |||
| 24343 | } | |||
| 24344 | ||||
| 24345 | /// Scalarize a vector store, bitcasting to TargetVT to determine the scalar | |||
| 24346 | /// type. | |||
| 24347 | static SDValue scalarizeVectorStore(StoreSDNode *Store, MVT StoreVT, | |||
| 24348 | SelectionDAG &DAG) { | |||
| 24349 | SDValue StoredVal = Store->getValue(); | |||
| 24350 | assert(StoreVT.is128BitVector() &&((void)0) | |||
| 24351 | StoredVal.getValueType().is128BitVector() && "Expecting 128-bit op")((void)0); | |||
| 24352 | StoredVal = DAG.getBitcast(StoreVT, StoredVal); | |||
| 24353 | ||||
| 24354 | // Splitting volatile memory ops is not allowed unless the operation was not | |||
| 24355 | // legal to begin with. We are assuming the input op is legal (this transform | |||
| 24356 | // is only used for targets with AVX). | |||
| 24357 | if (!Store->isSimple()) | |||
| 24358 | return SDValue(); | |||
| 24359 | ||||
| 24360 | MVT StoreSVT = StoreVT.getScalarType(); | |||
| 24361 | unsigned NumElems = StoreVT.getVectorNumElements(); | |||
| 24362 | unsigned ScalarSize = StoreSVT.getStoreSize(); | |||
| 24363 | ||||
| 24364 | SDLoc DL(Store); | |||
| 24365 | SmallVector<SDValue, 4> Stores; | |||
| 24366 | for (unsigned i = 0; i != NumElems; ++i) { | |||
| 24367 | unsigned Offset = i * ScalarSize; | |||
| 24368 | SDValue Ptr = DAG.getMemBasePlusOffset(Store->getBasePtr(), | |||
| 24369 | TypeSize::Fixed(Offset), DL); | |||
| 24370 | SDValue Scl = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, StoreSVT, StoredVal, | |||
| 24371 | DAG.getIntPtrConstant(i, DL)); | |||
| 24372 | SDValue Ch = DAG.getStore(Store->getChain(), DL, Scl, Ptr, | |||
| 24373 | Store->getPointerInfo().getWithOffset(Offset), | |||
| 24374 | Store->getOriginalAlign(), | |||
| 24375 | Store->getMemOperand()->getFlags()); | |||
| 24376 | Stores.push_back(Ch); | |||
| 24377 | } | |||
| 24378 | return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Stores); | |||
| 24379 | } | |||
| 24380 | ||||
| 24381 | static SDValue LowerStore(SDValue Op, const X86Subtarget &Subtarget, | |||
| 24382 | SelectionDAG &DAG) { | |||
| 24383 | StoreSDNode *St = cast<StoreSDNode>(Op.getNode()); | |||
| 24384 | SDLoc dl(St); | |||
| 24385 | SDValue StoredVal = St->getValue(); | |||
| 24386 | ||||
| 24387 | // Without AVX512DQ, we need to use a scalar type for v2i1/v4i1/v8i1 stores. | |||
| 24388 | if (StoredVal.getValueType().isVector() && | |||
| 24389 | StoredVal.getValueType().getVectorElementType() == MVT::i1) { | |||
| 24390 | unsigned NumElts = StoredVal.getValueType().getVectorNumElements(); | |||
| 24391 | assert(NumElts <= 8 && "Unexpected VT")((void)0); | |||
| 24392 | assert(!St->isTruncatingStore() && "Expected non-truncating store")((void)0); | |||
| 24393 | assert(Subtarget.hasAVX512() && !Subtarget.hasDQI() &&((void)0) | |||
| 24394 | "Expected AVX512F without AVX512DQI")((void)0); | |||
| 24395 | ||||
| 24396 | // We must pad with zeros to ensure we store zeroes to any unused bits. | |||
| 24397 | StoredVal = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, MVT::v16i1, | |||
| 24398 | DAG.getUNDEF(MVT::v16i1), StoredVal, | |||
| 24399 | DAG.getIntPtrConstant(0, dl)); | |||
| 24400 | StoredVal = DAG.getBitcast(MVT::i16, StoredVal); | |||
| 24401 | StoredVal = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, StoredVal); | |||
| 24402 | // Make sure we store zeros in the extra bits. | |||
| 24403 | if (NumElts < 8) | |||
| 24404 | StoredVal = DAG.getZeroExtendInReg( | |||
| 24405 | StoredVal, dl, EVT::getIntegerVT(*DAG.getContext(), NumElts)); | |||
| 24406 | ||||
| 24407 | return DAG.getStore(St->getChain(), dl, StoredVal, St->getBasePtr(), | |||
| 24408 | St->getPointerInfo(), St->getOriginalAlign(), | |||
| 24409 | St->getMemOperand()->getFlags()); | |||
| 24410 | } | |||
| 24411 | ||||
| 24412 | if (St->isTruncatingStore()) | |||
| 24413 | return SDValue(); | |||
| 24414 | ||||
| 24415 | // If this is a 256-bit store of concatenated ops, we are better off splitting | |||
| 24416 | // that store into two 128-bit stores. This avoids spurious use of 256-bit ops | |||
| 24417 | // and each half can execute independently. Some cores would split the op into | |||
| 24418 | // halves anyway, so the concat (vinsertf128) is purely an extra op. | |||
| 24419 | MVT StoreVT = StoredVal.getSimpleValueType(); | |||
| 24420 | if (StoreVT.is256BitVector() || | |||
| 24421 | ((StoreVT == MVT::v32i16 || StoreVT == MVT::v64i8) && | |||
| 24422 | !Subtarget.hasBWI())) { | |||
| 24423 | SmallVector<SDValue, 4> CatOps; | |||
| 24424 | if (StoredVal.hasOneUse() && collectConcatOps(StoredVal.getNode(), CatOps)) | |||
| 24425 | return splitVectorStore(St, DAG); | |||
| 24426 | return SDValue(); | |||
| 24427 | } | |||
| 24428 | ||||
| 24429 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
| 24430 | assert(StoreVT.isVector() && StoreVT.getSizeInBits() == 64 &&((void)0) | |||
| 24431 | "Unexpected VT")((void)0); | |||
| 24432 | assert(TLI.getTypeAction(*DAG.getContext(), StoreVT) ==((void)0) | |||
| 24433 | TargetLowering::TypeWidenVector && "Unexpected type action!")((void)0); | |||
| 24434 | ||||
| 24435 | EVT WideVT = TLI.getTypeToTransformTo(*DAG.getContext(), StoreVT); | |||
| 24436 | StoredVal = DAG.getNode(ISD::CONCAT_VECTORS, dl, WideVT, StoredVal, | |||
| 24437 | DAG.getUNDEF(StoreVT)); | |||
| 24438 | ||||
| 24439 | if (Subtarget.hasSSE2()) { | |||
| 24440 | // Widen the vector, cast to a v2x64 type, extract the single 64-bit element | |||
| 24441 | // and store it. | |||
| 24442 | MVT StVT = Subtarget.is64Bit() && StoreVT.isInteger() ? MVT::i64 : MVT::f64; | |||
| 24443 | MVT CastVT = MVT::getVectorVT(StVT, 2); | |||
| 24444 | StoredVal = DAG.getBitcast(CastVT, StoredVal); | |||
| 24445 | StoredVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, StVT, StoredVal, | |||
| 24446 | DAG.getIntPtrConstant(0, dl)); | |||
| 24447 | ||||
| 24448 | return DAG.getStore(St->getChain(), dl, StoredVal, St->getBasePtr(), | |||
| 24449 | St->getPointerInfo(), St->getOriginalAlign(), | |||
| 24450 | St->getMemOperand()->getFlags()); | |||
| 24451 | } | |||
| 24452 | assert(Subtarget.hasSSE1() && "Expected SSE")((void)0); | |||
| 24453 | SDVTList Tys = DAG.getVTList(MVT::Other); | |||
| 24454 | SDValue Ops[] = {St->getChain(), StoredVal, St->getBasePtr()}; | |||
| 24455 | return DAG.getMemIntrinsicNode(X86ISD::VEXTRACT_STORE, dl, Tys, Ops, MVT::i64, | |||
| 24456 | St->getMemOperand()); | |||
| 24457 | } | |||
| 24458 | ||||
| 24459 | // Lower vector extended loads using a shuffle. If SSSE3 is not available we | |||
| 24460 | // may emit an illegal shuffle but the expansion is still better than scalar | |||
| 24461 | // code. We generate sext/sext_invec for SEXTLOADs if it's available, otherwise | |||
| 24462 | // we'll emit a shuffle and a arithmetic shift. | |||
| 24463 | // FIXME: Is the expansion actually better than scalar code? It doesn't seem so. | |||
| 24464 | // TODO: It is possible to support ZExt by zeroing the undef values during | |||
| 24465 | // the shuffle phase or after the shuffle. | |||
| 24466 | static SDValue LowerLoad(SDValue Op, const X86Subtarget &Subtarget, | |||
| 24467 | SelectionDAG &DAG) { | |||
| 24468 | MVT RegVT = Op.getSimpleValueType(); | |||
| 24469 | assert(RegVT.isVector() && "We only custom lower vector loads.")((void)0); | |||
| 24470 | assert(RegVT.isInteger() &&((void)0) | |||
| 24471 | "We only custom lower integer vector loads.")((void)0); | |||
| 24472 | ||||
| 24473 | LoadSDNode *Ld = cast<LoadSDNode>(Op.getNode()); | |||
| 24474 | SDLoc dl(Ld); | |||
| 24475 | ||||
| 24476 | // Without AVX512DQ, we need to use a scalar type for v2i1/v4i1/v8i1 loads. | |||
| 24477 | if (RegVT.getVectorElementType() == MVT::i1) { | |||
| 24478 | assert(EVT(RegVT) == Ld->getMemoryVT() && "Expected non-extending load")((void)0); | |||
| 24479 | assert(RegVT.getVectorNumElements() <= 8 && "Unexpected VT")((void)0); | |||
| 24480 | assert(Subtarget.hasAVX512() && !Subtarget.hasDQI() &&((void)0) | |||
| 24481 | "Expected AVX512F without AVX512DQI")((void)0); | |||
| 24482 | ||||
| 24483 | SDValue NewLd = DAG.getLoad(MVT::i8, dl, Ld->getChain(), Ld->getBasePtr(), | |||
| 24484 | Ld->getPointerInfo(), Ld->getOriginalAlign(), | |||
| 24485 | Ld->getMemOperand()->getFlags()); | |||
| 24486 | ||||
| 24487 | // Replace chain users with the new chain. | |||
| 24488 | assert(NewLd->getNumValues() == 2 && "Loads must carry a chain!")((void)0); | |||
| 24489 | ||||
| 24490 | SDValue Val = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i16, NewLd); | |||
| 24491 | Val = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, RegVT, | |||
| 24492 | DAG.getBitcast(MVT::v16i1, Val), | |||
| 24493 | DAG.getIntPtrConstant(0, dl)); | |||
| 24494 | return DAG.getMergeValues({Val, NewLd.getValue(1)}, dl); | |||
| 24495 | } | |||
| 24496 | ||||
| 24497 | return SDValue(); | |||
| 24498 | } | |||
| 24499 | ||||
| 24500 | /// Return true if node is an ISD::AND or ISD::OR of two X86ISD::SETCC nodes | |||
| 24501 | /// each of which has no other use apart from the AND / OR. | |||
| 24502 | static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) { | |||
| 24503 | Opc = Op.getOpcode(); | |||
| 24504 | if (Opc != ISD::OR && Opc != ISD::AND) | |||
| 24505 | return false; | |||
| 24506 | return (Op.getOperand(0).getOpcode() == X86ISD::SETCC && | |||
| 24507 | Op.getOperand(0).hasOneUse() && | |||
| 24508 | Op.getOperand(1).getOpcode() == X86ISD::SETCC && | |||
| 24509 | Op.getOperand(1).hasOneUse()); | |||
| 24510 | } | |||
| 24511 | ||||
| 24512 | SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const { | |||
| 24513 | SDValue Chain = Op.getOperand(0); | |||
| 24514 | SDValue Cond = Op.getOperand(1); | |||
| 24515 | SDValue Dest = Op.getOperand(2); | |||
| 24516 | SDLoc dl(Op); | |||
| 24517 | ||||
| 24518 | if (Cond.getOpcode() == ISD::SETCC && | |||
| 24519 | Cond.getOperand(0).getValueType() != MVT::f128) { | |||
| 24520 | SDValue LHS = Cond.getOperand(0); | |||
| 24521 | SDValue RHS = Cond.getOperand(1); | |||
| 24522 | ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get(); | |||
| 24523 | ||||
| 24524 | // Special case for | |||
| 24525 | // setcc([su]{add,sub,mul}o == 0) | |||
| 24526 | // setcc([su]{add,sub,mul}o != 1) | |||
| 24527 | if (ISD::isOverflowIntrOpRes(LHS) && | |||
| 24528 | (CC == ISD::SETEQ || CC == ISD::SETNE) && | |||
| 24529 | (isNullConstant(RHS) || isOneConstant(RHS))) { | |||
| 24530 | SDValue Value, Overflow; | |||
| 24531 | X86::CondCode X86Cond; | |||
| 24532 | std::tie(Value, Overflow) = getX86XALUOOp(X86Cond, LHS.getValue(0), DAG); | |||
| 24533 | ||||
| 24534 | if ((CC == ISD::SETEQ) == isNullConstant(RHS)) | |||
| 24535 | X86Cond = X86::GetOppositeBranchCondition(X86Cond); | |||
| 24536 | ||||
| 24537 | SDValue CCVal = DAG.getTargetConstant(X86Cond, dl, MVT::i8); | |||
| 24538 | return DAG.getNode(X86ISD::BRCOND, dl, MVT::Other, Chain, Dest, CCVal, | |||
| 24539 | Overflow); | |||
| 24540 | } | |||
| 24541 | ||||
| 24542 | if (LHS.getSimpleValueType().isInteger()) { | |||
| 24543 | SDValue CCVal; | |||
| 24544 | SDValue EFLAGS = emitFlagsForSetcc(LHS, RHS, CC, SDLoc(Cond), DAG, CCVal); | |||
| 24545 | return DAG.getNode(X86ISD::BRCOND, dl, MVT::Other, Chain, Dest, CCVal, | |||
| 24546 | EFLAGS); | |||
| 24547 | } | |||
| 24548 | ||||
| 24549 | if (CC == ISD::SETOEQ) { | |||
| 24550 | // For FCMP_OEQ, we can emit | |||
| 24551 | // two branches instead of an explicit AND instruction with a | |||
| 24552 | // separate test. However, we only do this if this block doesn't | |||
| 24553 | // have a fall-through edge, because this requires an explicit | |||
| 24554 | // jmp when the condition is false. | |||
| 24555 | if (Op.getNode()->hasOneUse()) { | |||
| 24556 | SDNode *User = *Op.getNode()->use_begin(); | |||
| 24557 | // Look for an unconditional branch following this conditional branch. | |||
| 24558 | // We need this because we need to reverse the successors in order | |||
| 24559 | // to implement FCMP_OEQ. | |||
| 24560 | if (User->getOpcode() == ISD::BR) { | |||
| 24561 | SDValue FalseBB = User->getOperand(1); | |||
| 24562 | SDNode *NewBR = | |||
| 24563 | DAG.UpdateNodeOperands(User, User->getOperand(0), Dest); | |||
| 24564 | assert(NewBR == User)((void)0); | |||
| 24565 | (void)NewBR; | |||
| 24566 | Dest = FalseBB; | |||
| 24567 | ||||
| 24568 | SDValue Cmp = | |||
| 24569 | DAG.getNode(X86ISD::FCMP, SDLoc(Cond), MVT::i32, LHS, RHS); | |||
| 24570 | SDValue CCVal = DAG.getTargetConstant(X86::COND_NE, dl, MVT::i8); | |||
| 24571 | Chain = DAG.getNode(X86ISD::BRCOND, dl, MVT::Other, Chain, Dest, | |||
| 24572 | CCVal, Cmp); | |||
| 24573 | CCVal = DAG.getTargetConstant(X86::COND_P, dl, MVT::i8); | |||
| 24574 | return DAG.getNode(X86ISD::BRCOND, dl, MVT::Other, Chain, Dest, CCVal, | |||
| 24575 | Cmp); | |||
| 24576 | } | |||
| 24577 | } | |||
| 24578 | } else if (CC == ISD::SETUNE) { | |||
| 24579 | // For FCMP_UNE, we can emit | |||
| 24580 | // two branches instead of an explicit OR instruction with a | |||
| 24581 | // separate test. | |||
| 24582 | SDValue Cmp = DAG.getNode(X86ISD::FCMP, SDLoc(Cond), MVT::i32, LHS, RHS); | |||
| 24583 | SDValue CCVal = DAG.getTargetConstant(X86::COND_NE, dl, MVT::i8); | |||
| 24584 | Chain = | |||
| 24585 | DAG.getNode(X86ISD::BRCOND, dl, MVT::Other, Chain, Dest, CCVal, Cmp); | |||
| 24586 | CCVal = DAG.getTargetConstant(X86::COND_P, dl, MVT::i8); | |||
| 24587 | return DAG.getNode(X86ISD::BRCOND, dl, MVT::Other, Chain, Dest, CCVal, | |||
| 24588 | Cmp); | |||
| 24589 | } else { | |||
| 24590 | X86::CondCode X86Cond = | |||
| 24591 | TranslateX86CC(CC, dl, /*IsFP*/ true, LHS, RHS, DAG); | |||
| 24592 | SDValue Cmp = DAG.getNode(X86ISD::FCMP, SDLoc(Cond), MVT::i32, LHS, RHS); | |||
| 24593 | SDValue CCVal = DAG.getTargetConstant(X86Cond, dl, MVT::i8); | |||
| 24594 | return DAG.getNode(X86ISD::BRCOND, dl, MVT::Other, Chain, Dest, CCVal, | |||
| 24595 | Cmp); | |||
| 24596 | } | |||
| 24597 | } | |||
| 24598 | ||||
| 24599 | if (ISD::isOverflowIntrOpRes(Cond)) { | |||
| 24600 | SDValue Value, Overflow; | |||
| 24601 | X86::CondCode X86Cond; | |||
| 24602 | std::tie(Value, Overflow) = getX86XALUOOp(X86Cond, Cond.getValue(0), DAG); | |||
| 24603 | ||||
| 24604 | SDValue CCVal = DAG.getTargetConstant(X86Cond, dl, MVT::i8); | |||
| 24605 | return DAG.getNode(X86ISD::BRCOND, dl, MVT::Other, Chain, Dest, CCVal, | |||
| 24606 | Overflow); | |||
| 24607 | } | |||
| 24608 | ||||
| 24609 | // Look past the truncate if the high bits are known zero. | |||
| 24610 | if (isTruncWithZeroHighBitsInput(Cond, DAG)) | |||
| 24611 | Cond = Cond.getOperand(0); | |||
| 24612 | ||||
| 24613 | EVT CondVT = Cond.getValueType(); | |||
| 24614 | ||||
| 24615 | // Add an AND with 1 if we don't already have one. | |||
| 24616 | if (!(Cond.getOpcode() == ISD::AND && isOneConstant(Cond.getOperand(1)))) | |||
| 24617 | Cond = | |||
| 24618 | DAG.getNode(ISD::AND, dl, CondVT, Cond, DAG.getConstant(1, dl, CondVT)); | |||
| 24619 | ||||
| 24620 | SDValue LHS = Cond; | |||
| 24621 | SDValue RHS = DAG.getConstant(0, dl, CondVT); | |||
| 24622 | ||||
| 24623 | SDValue CCVal; | |||
| 24624 | SDValue EFLAGS = emitFlagsForSetcc(LHS, RHS, ISD::SETNE, dl, DAG, CCVal); | |||
| 24625 | return DAG.getNode(X86ISD::BRCOND, dl, MVT::Other, Chain, Dest, CCVal, | |||
| 24626 | EFLAGS); | |||
| 24627 | } | |||
| 24628 | ||||
| 24629 | // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets. | |||
| 24630 | // Calls to _alloca are needed to probe the stack when allocating more than 4k | |||
| 24631 | // bytes in one go. Touching the stack at 4K increments is necessary to ensure | |||
| 24632 | // that the guard pages used by the OS virtual memory manager are allocated in | |||
| 24633 | // correct sequence. | |||
| 24634 | SDValue | |||
| 24635 | X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op, | |||
| 24636 | SelectionDAG &DAG) const { | |||
| 24637 | MachineFunction &MF = DAG.getMachineFunction(); | |||
| 24638 | bool SplitStack = MF.shouldSplitStack(); | |||
| 24639 | bool EmitStackProbeCall = hasStackProbeSymbol(MF); | |||
| 24640 | bool Lower = (Subtarget.isOSWindows() && !Subtarget.isTargetMachO()) || | |||
| 24641 | SplitStack || EmitStackProbeCall; | |||
| 24642 | SDLoc dl(Op); | |||
| 24643 | ||||
| 24644 | // Get the inputs. | |||
| 24645 | SDNode *Node = Op.getNode(); | |||
| 24646 | SDValue Chain = Op.getOperand(0); | |||
| 24647 | SDValue Size = Op.getOperand(1); | |||
| 24648 | MaybeAlign Alignment(Op.getConstantOperandVal(2)); | |||
| 24649 | EVT VT = Node->getValueType(0); | |||
| 24650 | ||||
| 24651 | // Chain the dynamic stack allocation so that it doesn't modify the stack | |||
| 24652 | // pointer when other instructions are using the stack. | |||
| 24653 | Chain = DAG.getCALLSEQ_START(Chain, 0, 0, dl); | |||
| 24654 | ||||
| 24655 | bool Is64Bit = Subtarget.is64Bit(); | |||
| 24656 | MVT SPTy = getPointerTy(DAG.getDataLayout()); | |||
| 24657 | ||||
| 24658 | SDValue Result; | |||
| 24659 | if (!Lower) { | |||
| 24660 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
| 24661 | Register SPReg = TLI.getStackPointerRegisterToSaveRestore(); | |||
| 24662 | assert(SPReg && "Target cannot require DYNAMIC_STACKALLOC expansion and"((void)0) | |||
| 24663 | " not tell us which reg is the stack pointer!")((void)0); | |||
| 24664 | ||||
| 24665 | const TargetFrameLowering &TFI = *Subtarget.getFrameLowering(); | |||
| 24666 | const Align StackAlign = TFI.getStackAlign(); | |||
| 24667 | if (hasInlineStackProbe(MF)) { | |||
| 24668 | MachineRegisterInfo &MRI = MF.getRegInfo(); | |||
| 24669 | ||||
| 24670 | const TargetRegisterClass *AddrRegClass = getRegClassFor(SPTy); | |||
| 24671 | Register Vreg = MRI.createVirtualRegister(AddrRegClass); | |||
| 24672 | Chain = DAG.getCopyToReg(Chain, dl, Vreg, Size); | |||
| 24673 | Result = DAG.getNode(X86ISD::PROBED_ALLOCA, dl, SPTy, Chain, | |||
| 24674 | DAG.getRegister(Vreg, SPTy)); | |||
| 24675 | } else { | |||
| 24676 | SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, VT); | |||
| 24677 | Chain = SP.getValue(1); | |||
| 24678 | Result = DAG.getNode(ISD::SUB, dl, VT, SP, Size); // Value | |||
| 24679 | } | |||
| 24680 | if (Alignment && *Alignment > StackAlign) | |||
| 24681 | Result = | |||
| 24682 | DAG.getNode(ISD::AND, dl, VT, Result, | |||
| 24683 | DAG.getConstant(~(Alignment->value() - 1ULL), dl, VT)); | |||
| 24684 | Chain = DAG.getCopyToReg(Chain, dl, SPReg, Result); // Output chain | |||
| 24685 | } else if (SplitStack) { | |||
| 24686 | MachineRegisterInfo &MRI = MF.getRegInfo(); | |||
| 24687 | ||||
| 24688 | if (Is64Bit) { | |||
| 24689 | // The 64 bit implementation of segmented stacks needs to clobber both r10 | |||
| 24690 | // r11. This makes it impossible to use it along with nested parameters. | |||
| 24691 | const Function &F = MF.getFunction(); | |||
| 24692 | for (const auto &A : F.args()) { | |||
| 24693 | if (A.hasNestAttr()) | |||
| 24694 | report_fatal_error("Cannot use segmented stacks with functions that " | |||
| 24695 | "have nested arguments."); | |||
| 24696 | } | |||
| 24697 | } | |||
| 24698 | ||||
| 24699 | const TargetRegisterClass *AddrRegClass = getRegClassFor(SPTy); | |||
| 24700 | Register Vreg = MRI.createVirtualRegister(AddrRegClass); | |||
| 24701 | Chain = DAG.getCopyToReg(Chain, dl, Vreg, Size); | |||
| 24702 | Result = DAG.getNode(X86ISD::SEG_ALLOCA, dl, SPTy, Chain, | |||
| 24703 | DAG.getRegister(Vreg, SPTy)); | |||
| 24704 | } else { | |||
| 24705 | SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue); | |||
| 24706 | Chain = DAG.getNode(X86ISD::WIN_ALLOCA, dl, NodeTys, Chain, Size); | |||
| 24707 | MF.getInfo<X86MachineFunctionInfo>()->setHasWinAlloca(true); | |||
| 24708 | ||||
| 24709 | const X86RegisterInfo *RegInfo = Subtarget.getRegisterInfo(); | |||
| 24710 | Register SPReg = RegInfo->getStackRegister(); | |||
| 24711 | SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, SPTy); | |||
| 24712 | Chain = SP.getValue(1); | |||
| 24713 | ||||
| 24714 | if (Alignment) { | |||
| 24715 | SP = DAG.getNode(ISD::AND, dl, VT, SP.getValue(0), | |||
| 24716 | DAG.getConstant(~(Alignment->value() - 1ULL), dl, VT)); | |||
| 24717 | Chain = DAG.getCopyToReg(Chain, dl, SPReg, SP); | |||
| 24718 | } | |||
| 24719 | ||||
| 24720 | Result = SP; | |||
| 24721 | } | |||
| 24722 | ||||
| 24723 | Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, dl, true), | |||
| 24724 | DAG.getIntPtrConstant(0, dl, true), SDValue(), dl); | |||
| 24725 | ||||
| 24726 | SDValue Ops[2] = {Result, Chain}; | |||
| 24727 | return DAG.getMergeValues(Ops, dl); | |||
| 24728 | } | |||
| 24729 | ||||
| 24730 | SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const { | |||
| 24731 | MachineFunction &MF = DAG.getMachineFunction(); | |||
| 24732 | auto PtrVT = getPointerTy(MF.getDataLayout()); | |||
| 24733 | X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>(); | |||
| 24734 | ||||
| 24735 | const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue(); | |||
| 24736 | SDLoc DL(Op); | |||
| 24737 | ||||
| 24738 | if (!Subtarget.is64Bit() || | |||
| 24739 | Subtarget.isCallingConvWin64(MF.getFunction().getCallingConv())) { | |||
| 24740 | // vastart just stores the address of the VarArgsFrameIndex slot into the | |||
| 24741 | // memory location argument. | |||
| 24742 | SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT); | |||
| 24743 | return DAG.getStore(Op.getOperand(0), DL, FR, Op.getOperand(1), | |||
| 24744 | MachinePointerInfo(SV)); | |||
| 24745 | } | |||
| 24746 | ||||
| 24747 | // __va_list_tag: | |||
| 24748 | // gp_offset (0 - 6 * 8) | |||
| 24749 | // fp_offset (48 - 48 + 8 * 16) | |||
| 24750 | // overflow_arg_area (point to parameters coming in memory). | |||
| 24751 | // reg_save_area | |||
| 24752 | SmallVector<SDValue, 8> MemOps; | |||
| 24753 | SDValue FIN = Op.getOperand(1); | |||
| 24754 | // Store gp_offset | |||
| 24755 | SDValue Store = DAG.getStore( | |||
| 24756 | Op.getOperand(0), DL, | |||
| 24757 | DAG.getConstant(FuncInfo->getVarArgsGPOffset(), DL, MVT::i32), FIN, | |||
| 24758 | MachinePointerInfo(SV)); | |||
| 24759 | MemOps.push_back(Store); | |||
| 24760 | ||||
| 24761 | // Store fp_offset | |||
| 24762 | FIN = DAG.getMemBasePlusOffset(FIN, TypeSize::Fixed(4), DL); | |||
| 24763 | Store = DAG.getStore( | |||
| 24764 | Op.getOperand(0), DL, | |||
| 24765 | DAG.getConstant(FuncInfo->getVarArgsFPOffset(), DL, MVT::i32), FIN, | |||
| 24766 | MachinePointerInfo(SV, 4)); | |||
| 24767 | MemOps.push_back(Store); | |||
| 24768 | ||||
| 24769 | // Store ptr to overflow_arg_area | |||
| 24770 | FIN = DAG.getNode(ISD::ADD, DL, PtrVT, FIN, DAG.getIntPtrConstant(4, DL)); | |||
| 24771 | SDValue OVFIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT); | |||
| 24772 | Store = | |||
| 24773 | DAG.getStore(Op.getOperand(0), DL, OVFIN, FIN, MachinePointerInfo(SV, 8)); | |||
| 24774 | MemOps.push_back(Store); | |||
| 24775 | ||||
| 24776 | // Store ptr to reg_save_area. | |||
| 24777 | FIN = DAG.getNode(ISD::ADD, DL, PtrVT, FIN, DAG.getIntPtrConstant( | |||
| 24778 | Subtarget.isTarget64BitLP64() ? 8 : 4, DL)); | |||
| 24779 | SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(), PtrVT); | |||
| 24780 | Store = DAG.getStore( | |||
| 24781 | Op.getOperand(0), DL, RSFIN, FIN, | |||
| 24782 | MachinePointerInfo(SV, Subtarget.isTarget64BitLP64() ? 16 : 12)); | |||
| 24783 | MemOps.push_back(Store); | |||
| 24784 | return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOps); | |||
| 24785 | } | |||
| 24786 | ||||
| 24787 | SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const { | |||
| 24788 | assert(Subtarget.is64Bit() &&((void)0) | |||
| 24789 | "LowerVAARG only handles 64-bit va_arg!")((void)0); | |||
| 24790 | assert(Op.getNumOperands() == 4)((void)0); | |||
| 24791 | ||||
| 24792 | MachineFunction &MF = DAG.getMachineFunction(); | |||
| 24793 | if (Subtarget.isCallingConvWin64(MF.getFunction().getCallingConv())) | |||
| 24794 | // The Win64 ABI uses char* instead of a structure. | |||
| 24795 | return DAG.expandVAArg(Op.getNode()); | |||
| 24796 | ||||
| 24797 | SDValue Chain = Op.getOperand(0); | |||
| 24798 | SDValue SrcPtr = Op.getOperand(1); | |||
| 24799 | const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue(); | |||
| 24800 | unsigned Align = Op.getConstantOperandVal(3); | |||
| 24801 | SDLoc dl(Op); | |||
| 24802 | ||||
| 24803 | EVT ArgVT = Op.getNode()->getValueType(0); | |||
| 24804 | Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext()); | |||
| 24805 | uint32_t ArgSize = DAG.getDataLayout().getTypeAllocSize(ArgTy); | |||
| 24806 | uint8_t ArgMode; | |||
| 24807 | ||||
| 24808 | // Decide which area this value should be read from. | |||
| 24809 | // TODO: Implement the AMD64 ABI in its entirety. This simple | |||
| 24810 | // selection mechanism works only for the basic types. | |||
| 24811 | assert(ArgVT != MVT::f80 && "va_arg for f80 not yet implemented")((void)0); | |||
| 24812 | if (ArgVT.isFloatingPoint() && ArgSize <= 16 /*bytes*/) { | |||
| 24813 | ArgMode = 2; // Argument passed in XMM register. Use fp_offset. | |||
| 24814 | } else { | |||
| 24815 | assert(ArgVT.isInteger() && ArgSize <= 32 /*bytes*/ &&((void)0) | |||
| 24816 | "Unhandled argument type in LowerVAARG")((void)0); | |||
| 24817 | ArgMode = 1; // Argument passed in GPR64 register(s). Use gp_offset. | |||
| 24818 | } | |||
| 24819 | ||||
| 24820 | if (ArgMode == 2) { | |||
| 24821 | // Sanity Check: Make sure using fp_offset makes sense. | |||
| 24822 | assert(!Subtarget.useSoftFloat() &&((void)0) | |||
| 24823 | !(MF.getFunction().hasFnAttribute(Attribute::NoImplicitFloat)) &&((void)0) | |||
| 24824 | Subtarget.hasSSE1())((void)0); | |||
| 24825 | } | |||
| 24826 | ||||
| 24827 | // Insert VAARG node into the DAG | |||
| 24828 | // VAARG returns two values: Variable Argument Address, Chain | |||
| 24829 | SDValue InstOps[] = {Chain, SrcPtr, | |||
| 24830 | DAG.getTargetConstant(ArgSize, dl, MVT::i32), | |||
| 24831 | DAG.getTargetConstant(ArgMode, dl, MVT::i8), | |||
| 24832 | DAG.getTargetConstant(Align, dl, MVT::i32)}; | |||
| 24833 | SDVTList VTs = DAG.getVTList(getPointerTy(DAG.getDataLayout()), MVT::Other); | |||
| 24834 | SDValue VAARG = DAG.getMemIntrinsicNode( | |||
| 24835 | Subtarget.isTarget64BitLP64() ? X86ISD::VAARG_64 : X86ISD::VAARG_X32, dl, | |||
| 24836 | VTs, InstOps, MVT::i64, MachinePointerInfo(SV), | |||
| 24837 | /*Alignment=*/None, | |||
| 24838 | MachineMemOperand::MOLoad | MachineMemOperand::MOStore); | |||
| 24839 | Chain = VAARG.getValue(1); | |||
| 24840 | ||||
| 24841 | // Load the next argument and return it | |||
| 24842 | return DAG.getLoad(ArgVT, dl, Chain, VAARG, MachinePointerInfo()); | |||
| 24843 | } | |||
| 24844 | ||||
| 24845 | static SDValue LowerVACOPY(SDValue Op, const X86Subtarget &Subtarget, | |||
| 24846 | SelectionDAG &DAG) { | |||
| 24847 | // X86-64 va_list is a struct { i32, i32, i8*, i8* }, except on Windows, | |||
| 24848 | // where a va_list is still an i8*. | |||
| 24849 | assert(Subtarget.is64Bit() && "This code only handles 64-bit va_copy!")((void)0); | |||
| 24850 | if (Subtarget.isCallingConvWin64( | |||
| 24851 | DAG.getMachineFunction().getFunction().getCallingConv())) | |||
| 24852 | // Probably a Win64 va_copy. | |||
| 24853 | return DAG.expandVACopy(Op.getNode()); | |||
| 24854 | ||||
| 24855 | SDValue Chain = Op.getOperand(0); | |||
| 24856 | SDValue DstPtr = Op.getOperand(1); | |||
| 24857 | SDValue SrcPtr = Op.getOperand(2); | |||
| 24858 | const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue(); | |||
| 24859 | const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue(); | |||
| 24860 | SDLoc DL(Op); | |||
| 24861 | ||||
| 24862 | return DAG.getMemcpy( | |||
| 24863 | Chain, DL, DstPtr, SrcPtr, | |||
| 24864 | DAG.getIntPtrConstant(Subtarget.isTarget64BitLP64() ? 24 : 16, DL), | |||
| 24865 | Align(Subtarget.isTarget64BitLP64() ? 8 : 4), /*isVolatile*/ false, false, | |||
| 24866 | false, MachinePointerInfo(DstSV), MachinePointerInfo(SrcSV)); | |||
| 24867 | } | |||
| 24868 | ||||
| 24869 | // Helper to get immediate/variable SSE shift opcode from other shift opcodes. | |||
| 24870 | static unsigned getTargetVShiftUniformOpcode(unsigned Opc, bool IsVariable) { | |||
| 24871 | switch (Opc) { | |||
| 24872 | case ISD::SHL: | |||
| 24873 | case X86ISD::VSHL: | |||
| 24874 | case X86ISD::VSHLI: | |||
| 24875 | return IsVariable ? X86ISD::VSHL : X86ISD::VSHLI; | |||
| 24876 | case ISD::SRL: | |||
| 24877 | case X86ISD::VSRL: | |||
| 24878 | case X86ISD::VSRLI: | |||
| 24879 | return IsVariable ? X86ISD::VSRL : X86ISD::VSRLI; | |||
| 24880 | case ISD::SRA: | |||
| 24881 | case X86ISD::VSRA: | |||
| 24882 | case X86ISD::VSRAI: | |||
| 24883 | return IsVariable ? X86ISD::VSRA : X86ISD::VSRAI; | |||
| 24884 | } | |||
| 24885 | llvm_unreachable("Unknown target vector shift node")__builtin_unreachable(); | |||
| 24886 | } | |||
| 24887 | ||||
| 24888 | /// Handle vector element shifts where the shift amount is a constant. | |||
| 24889 | /// Takes immediate version of shift as input. | |||
| 24890 | static SDValue getTargetVShiftByConstNode(unsigned Opc, const SDLoc &dl, MVT VT, | |||
| 24891 | SDValue SrcOp, uint64_t ShiftAmt, | |||
| 24892 | SelectionDAG &DAG) { | |||
| 24893 | MVT ElementType = VT.getVectorElementType(); | |||
| 24894 | ||||
| 24895 | // Bitcast the source vector to the output type, this is mainly necessary for | |||
| 24896 | // vXi8/vXi64 shifts. | |||
| 24897 | if (VT != SrcOp.getSimpleValueType()) | |||
| 24898 | SrcOp = DAG.getBitcast(VT, SrcOp); | |||
| 24899 | ||||
| 24900 | // Fold this packed shift into its first operand if ShiftAmt is 0. | |||
| 24901 | if (ShiftAmt == 0) | |||
| 24902 | return SrcOp; | |||
| 24903 | ||||
| 24904 | // Check for ShiftAmt >= element width | |||
| 24905 | if (ShiftAmt >= ElementType.getSizeInBits()) { | |||
| 24906 | if (Opc == X86ISD::VSRAI) | |||
| 24907 | ShiftAmt = ElementType.getSizeInBits() - 1; | |||
| 24908 | else | |||
| 24909 | return DAG.getConstant(0, dl, VT); | |||
| 24910 | } | |||
| 24911 | ||||
| 24912 | assert((Opc == X86ISD::VSHLI || Opc == X86ISD::VSRLI || Opc == X86ISD::VSRAI)((void)0) | |||
| 24913 | && "Unknown target vector shift-by-constant node")((void)0); | |||
| 24914 | ||||
| 24915 | // Fold this packed vector shift into a build vector if SrcOp is a | |||
| 24916 | // vector of Constants or UNDEFs. | |||
| 24917 | if (ISD::isBuildVectorOfConstantSDNodes(SrcOp.getNode())) { | |||
| 24918 | SmallVector<SDValue, 8> Elts; | |||
| 24919 | unsigned NumElts = SrcOp->getNumOperands(); | |||
| 24920 | ||||
| 24921 | switch (Opc) { | |||
| 24922 | default: llvm_unreachable("Unknown opcode!")__builtin_unreachable(); | |||
| 24923 | case X86ISD::VSHLI: | |||
| 24924 | for (unsigned i = 0; i != NumElts; ++i) { | |||
| 24925 | SDValue CurrentOp = SrcOp->getOperand(i); | |||
| 24926 | if (CurrentOp->isUndef()) { | |||
| 24927 | // Must produce 0s in the correct bits. | |||
| 24928 | Elts.push_back(DAG.getConstant(0, dl, ElementType)); | |||
| 24929 | continue; | |||
| 24930 | } | |||
| 24931 | auto *ND = cast<ConstantSDNode>(CurrentOp); | |||
| 24932 | const APInt &C = ND->getAPIntValue(); | |||
| 24933 | Elts.push_back(DAG.getConstant(C.shl(ShiftAmt), dl, ElementType)); | |||
| 24934 | } | |||
| 24935 | break; | |||
| 24936 | case X86ISD::VSRLI: | |||
| 24937 | for (unsigned i = 0; i != NumElts; ++i) { | |||
| 24938 | SDValue CurrentOp = SrcOp->getOperand(i); | |||
| 24939 | if (CurrentOp->isUndef()) { | |||
| 24940 | // Must produce 0s in the correct bits. | |||
| 24941 | Elts.push_back(DAG.getConstant(0, dl, ElementType)); | |||
| 24942 | continue; | |||
| 24943 | } | |||
| 24944 | auto *ND = cast<ConstantSDNode>(CurrentOp); | |||
| 24945 | const APInt &C = ND->getAPIntValue(); | |||
| 24946 | Elts.push_back(DAG.getConstant(C.lshr(ShiftAmt), dl, ElementType)); | |||
| 24947 | } | |||
| 24948 | break; | |||
| 24949 | case X86ISD::VSRAI: | |||
| 24950 | for (unsigned i = 0; i != NumElts; ++i) { | |||
| 24951 | SDValue CurrentOp = SrcOp->getOperand(i); | |||
| 24952 | if (CurrentOp->isUndef()) { | |||
| 24953 | // All shifted in bits must be the same so use 0. | |||
| 24954 | Elts.push_back(DAG.getConstant(0, dl, ElementType)); | |||
| 24955 | continue; | |||
| 24956 | } | |||
| 24957 | auto *ND = cast<ConstantSDNode>(CurrentOp); | |||
| 24958 | const APInt &C = ND->getAPIntValue(); | |||
| 24959 | Elts.push_back(DAG.getConstant(C.ashr(ShiftAmt), dl, ElementType)); | |||
| 24960 | } | |||
| 24961 | break; | |||
| 24962 | } | |||
| 24963 | ||||
| 24964 | return DAG.getBuildVector(VT, dl, Elts); | |||
| 24965 | } | |||
| 24966 | ||||
| 24967 | return DAG.getNode(Opc, dl, VT, SrcOp, | |||
| 24968 | DAG.getTargetConstant(ShiftAmt, dl, MVT::i8)); | |||
| 24969 | } | |||
| 24970 | ||||
| 24971 | /// Handle vector element shifts where the shift amount may or may not be a | |||
| 24972 | /// constant. Takes immediate version of shift as input. | |||
| 24973 | static SDValue getTargetVShiftNode(unsigned Opc, const SDLoc &dl, MVT VT, | |||
| 24974 | SDValue SrcOp, SDValue ShAmt, | |||
| 24975 | const X86Subtarget &Subtarget, | |||
| 24976 | SelectionDAG &DAG) { | |||
| 24977 | MVT SVT = ShAmt.getSimpleValueType(); | |||
| 24978 | assert((SVT == MVT::i32 || SVT == MVT::i64) && "Unexpected value type!")((void)0); | |||
| 24979 | ||||
| 24980 | // Catch shift-by-constant. | |||
| 24981 | if (ConstantSDNode *CShAmt = dyn_cast<ConstantSDNode>(ShAmt)) | |||
| 24982 | return getTargetVShiftByConstNode(Opc, dl, VT, SrcOp, | |||
| 24983 | CShAmt->getZExtValue(), DAG); | |||
| 24984 | ||||
| 24985 | // Change opcode to non-immediate version. | |||
| 24986 | Opc = getTargetVShiftUniformOpcode(Opc, true); | |||
| 24987 | ||||
| 24988 | // Need to build a vector containing shift amount. | |||
| 24989 | // SSE/AVX packed shifts only use the lower 64-bit of the shift count. | |||
| 24990 | // +====================+============+=======================================+ | |||
| 24991 | // | ShAmt is | HasSSE4.1? | Construct ShAmt vector as | | |||
| 24992 | // +====================+============+=======================================+ | |||
| 24993 | // | i64 | Yes, No | Use ShAmt as lowest elt | | |||
| 24994 | // | i32 | Yes | zero-extend in-reg | | |||
| 24995 | // | (i32 zext(i16/i8)) | Yes | zero-extend in-reg | | |||
| 24996 | // | (i32 zext(i16/i8)) | No | byte-shift-in-reg | | |||
| 24997 | // | i16/i32 | No | v4i32 build_vector(ShAmt, 0, ud, ud)) | | |||
| 24998 | // +====================+============+=======================================+ | |||
| 24999 | ||||
| 25000 | if (SVT == MVT::i64) | |||
| 25001 | ShAmt = DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(ShAmt), MVT::v2i64, ShAmt); | |||
| 25002 | else if (ShAmt.getOpcode() == ISD::ZERO_EXTEND && | |||
| 25003 | ShAmt.getOperand(0).getOpcode() == ISD::EXTRACT_VECTOR_ELT && | |||
| 25004 | (ShAmt.getOperand(0).getSimpleValueType() == MVT::i16 || | |||
| 25005 | ShAmt.getOperand(0).getSimpleValueType() == MVT::i8)) { | |||
| 25006 | ShAmt = ShAmt.getOperand(0); | |||
| 25007 | MVT AmtTy = ShAmt.getSimpleValueType() == MVT::i8 ? MVT::v16i8 : MVT::v8i16; | |||
| 25008 | ShAmt = DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(ShAmt), AmtTy, ShAmt); | |||
| 25009 | if (Subtarget.hasSSE41()) | |||
| 25010 | ShAmt = DAG.getNode(ISD::ZERO_EXTEND_VECTOR_INREG, SDLoc(ShAmt), | |||
| 25011 | MVT::v2i64, ShAmt); | |||
| 25012 | else { | |||
| 25013 | SDValue ByteShift = DAG.getTargetConstant( | |||
| 25014 | (128 - AmtTy.getScalarSizeInBits()) / 8, SDLoc(ShAmt), MVT::i8); | |||
| 25015 | ShAmt = DAG.getBitcast(MVT::v16i8, ShAmt); | |||
| 25016 | ShAmt = DAG.getNode(X86ISD::VSHLDQ, SDLoc(ShAmt), MVT::v16i8, ShAmt, | |||
| 25017 | ByteShift); | |||
| 25018 | ShAmt = DAG.getNode(X86ISD::VSRLDQ, SDLoc(ShAmt), MVT::v16i8, ShAmt, | |||
| 25019 | ByteShift); | |||
| 25020 | } | |||
| 25021 | } else if (Subtarget.hasSSE41() && | |||
| 25022 | ShAmt.getOpcode() == ISD::EXTRACT_VECTOR_ELT) { | |||
| 25023 | ShAmt = DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(ShAmt), MVT::v4i32, ShAmt); | |||
| 25024 | ShAmt = DAG.getNode(ISD::ZERO_EXTEND_VECTOR_INREG, SDLoc(ShAmt), | |||
| 25025 | MVT::v2i64, ShAmt); | |||
| 25026 | } else { | |||
| 25027 | SDValue ShOps[4] = {ShAmt, DAG.getConstant(0, dl, SVT), DAG.getUNDEF(SVT), | |||
| 25028 | DAG.getUNDEF(SVT)}; | |||
| 25029 | ShAmt = DAG.getBuildVector(MVT::v4i32, dl, ShOps); | |||
| 25030 | } | |||
| 25031 | ||||
| 25032 | // The return type has to be a 128-bit type with the same element | |||
| 25033 | // type as the input type. | |||
| 25034 | MVT EltVT = VT.getVectorElementType(); | |||
| 25035 | MVT ShVT = MVT::getVectorVT(EltVT, 128 / EltVT.getSizeInBits()); | |||
| 25036 | ||||
| 25037 | ShAmt = DAG.getBitcast(ShVT, ShAmt); | |||
| 25038 | return DAG.getNode(Opc, dl, VT, SrcOp, ShAmt); | |||
| 25039 | } | |||
| 25040 | ||||
| 25041 | /// Return Mask with the necessary casting or extending | |||
| 25042 | /// for \p Mask according to \p MaskVT when lowering masking intrinsics | |||
| 25043 | static SDValue getMaskNode(SDValue Mask, MVT MaskVT, | |||
| 25044 | const X86Subtarget &Subtarget, SelectionDAG &DAG, | |||
| 25045 | const SDLoc &dl) { | |||
| 25046 | ||||
| 25047 | if (isAllOnesConstant(Mask)) | |||
| 25048 | return DAG.getConstant(1, dl, MaskVT); | |||
| 25049 | if (X86::isZeroNode(Mask)) | |||
| 25050 | return DAG.getConstant(0, dl, MaskVT); | |||
| 25051 | ||||
| 25052 | assert(MaskVT.bitsLE(Mask.getSimpleValueType()) && "Unexpected mask size!")((void)0); | |||
| 25053 | ||||
| 25054 | if (Mask.getSimpleValueType() == MVT::i64 && Subtarget.is32Bit()) { | |||
| 25055 | assert(MaskVT == MVT::v64i1 && "Expected v64i1 mask!")((void)0); | |||
| 25056 | assert(Subtarget.hasBWI() && "Expected AVX512BW target!")((void)0); | |||
| 25057 | // In case 32bit mode, bitcast i64 is illegal, extend/split it. | |||
| 25058 | SDValue Lo, Hi; | |||
| 25059 | Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Mask, | |||
| 25060 | DAG.getConstant(0, dl, MVT::i32)); | |||
| 25061 | Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Mask, | |||
| 25062 | DAG.getConstant(1, dl, MVT::i32)); | |||
| 25063 | ||||
| 25064 | Lo = DAG.getBitcast(MVT::v32i1, Lo); | |||
| 25065 | Hi = DAG.getBitcast(MVT::v32i1, Hi); | |||
| 25066 | ||||
| 25067 | return DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v64i1, Lo, Hi); | |||
| 25068 | } else { | |||
| 25069 | MVT BitcastVT = MVT::getVectorVT(MVT::i1, | |||
| 25070 | Mask.getSimpleValueType().getSizeInBits()); | |||
| 25071 | // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements | |||
| 25072 | // are extracted by EXTRACT_SUBVECTOR. | |||
| 25073 | return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT, | |||
| 25074 | DAG.getBitcast(BitcastVT, Mask), | |||
| 25075 | DAG.getIntPtrConstant(0, dl)); | |||
| 25076 | } | |||
| 25077 | } | |||
| 25078 | ||||
| 25079 | /// Return (and \p Op, \p Mask) for compare instructions or | |||
| 25080 | /// (vselect \p Mask, \p Op, \p PreservedSrc) for others along with the | |||
| 25081 | /// necessary casting or extending for \p Mask when lowering masking intrinsics | |||
| 25082 | static SDValue getVectorMaskingNode(SDValue Op, SDValue Mask, | |||
| 25083 | SDValue PreservedSrc, | |||
| 25084 | const X86Subtarget &Subtarget, | |||
| 25085 | SelectionDAG &DAG) { | |||
| 25086 | MVT VT = Op.getSimpleValueType(); | |||
| 25087 | MVT MaskVT = MVT::getVectorVT(MVT::i1, VT.getVectorNumElements()); | |||
| 25088 | unsigned OpcodeSelect = ISD::VSELECT; | |||
| 25089 | SDLoc dl(Op); | |||
| 25090 | ||||
| 25091 | if (isAllOnesConstant(Mask)) | |||
| 25092 | return Op; | |||
| 25093 | ||||
| 25094 | SDValue VMask = getMaskNode(Mask, MaskVT, Subtarget, DAG, dl); | |||
| 25095 | ||||
| 25096 | if (PreservedSrc.isUndef()) | |||
| 25097 | PreservedSrc = getZeroVector(VT, Subtarget, DAG, dl); | |||
| 25098 | return DAG.getNode(OpcodeSelect, dl, VT, VMask, Op, PreservedSrc); | |||
| 25099 | } | |||
| 25100 | ||||
| 25101 | /// Creates an SDNode for a predicated scalar operation. | |||
| 25102 | /// \returns (X86vselect \p Mask, \p Op, \p PreservedSrc). | |||
| 25103 | /// The mask is coming as MVT::i8 and it should be transformed | |||
| 25104 | /// to MVT::v1i1 while lowering masking intrinsics. | |||
| 25105 | /// The main difference between ScalarMaskingNode and VectorMaskingNode is using | |||
| 25106 | /// "X86select" instead of "vselect". We just can't create the "vselect" node | |||
| 25107 | /// for a scalar instruction. | |||
| 25108 | static SDValue getScalarMaskingNode(SDValue Op, SDValue Mask, | |||
| 25109 | SDValue PreservedSrc, | |||
| 25110 | const X86Subtarget &Subtarget, | |||
| 25111 | SelectionDAG &DAG) { | |||
| 25112 | ||||
| 25113 | if (auto *MaskConst = dyn_cast<ConstantSDNode>(Mask)) | |||
| 25114 | if (MaskConst->getZExtValue() & 0x1) | |||
| 25115 | return Op; | |||
| 25116 | ||||
| 25117 | MVT VT = Op.getSimpleValueType(); | |||
| 25118 | SDLoc dl(Op); | |||
| 25119 | ||||
| 25120 | assert(Mask.getValueType() == MVT::i8 && "Unexpect type")((void)0); | |||
| 25121 | SDValue IMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v1i1, | |||
| 25122 | DAG.getBitcast(MVT::v8i1, Mask), | |||
| 25123 | DAG.getIntPtrConstant(0, dl)); | |||
| 25124 | if (Op.getOpcode() == X86ISD::FSETCCM || | |||
| 25125 | Op.getOpcode() == X86ISD::FSETCCM_SAE || | |||
| 25126 | Op.getOpcode() == X86ISD::VFPCLASSS) | |||
| 25127 | return DAG.getNode(ISD::AND, dl, VT, Op, IMask); | |||
| 25128 | ||||
| 25129 | if (PreservedSrc.isUndef()) | |||
| 25130 | PreservedSrc = getZeroVector(VT, Subtarget, DAG, dl); | |||
| 25131 | return DAG.getNode(X86ISD::SELECTS, dl, VT, IMask, Op, PreservedSrc); | |||
| 25132 | } | |||
| 25133 | ||||
| 25134 | static int getSEHRegistrationNodeSize(const Function *Fn) { | |||
| 25135 | if (!Fn->hasPersonalityFn()) | |||
| 25136 | report_fatal_error( | |||
| 25137 | "querying registration node size for function without personality"); | |||
| 25138 | // The RegNodeSize is 6 32-bit words for SEH and 4 for C++ EH. See | |||
| 25139 | // WinEHStatePass for the full struct definition. | |||
| 25140 | switch (classifyEHPersonality(Fn->getPersonalityFn())) { | |||
| 25141 | case EHPersonality::MSVC_X86SEH: return 24; | |||
| 25142 | case EHPersonality::MSVC_CXX: return 16; | |||
| 25143 | default: break; | |||
| 25144 | } | |||
| 25145 | report_fatal_error( | |||
| 25146 | "can only recover FP for 32-bit MSVC EH personality functions"); | |||
| 25147 | } | |||
| 25148 | ||||
| 25149 | /// When the MSVC runtime transfers control to us, either to an outlined | |||
| 25150 | /// function or when returning to a parent frame after catching an exception, we | |||
| 25151 | /// recover the parent frame pointer by doing arithmetic on the incoming EBP. | |||
| 25152 | /// Here's the math: | |||
| 25153 | /// RegNodeBase = EntryEBP - RegNodeSize | |||
| 25154 | /// ParentFP = RegNodeBase - ParentFrameOffset | |||
| 25155 | /// Subtracting RegNodeSize takes us to the offset of the registration node, and | |||
| 25156 | /// subtracting the offset (negative on x86) takes us back to the parent FP. | |||
| 25157 | static SDValue recoverFramePointer(SelectionDAG &DAG, const Function *Fn, | |||
| 25158 | SDValue EntryEBP) { | |||
| 25159 | MachineFunction &MF = DAG.getMachineFunction(); | |||
| 25160 | SDLoc dl; | |||
| 25161 | ||||
| 25162 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
| 25163 | MVT PtrVT = TLI.getPointerTy(DAG.getDataLayout()); | |||
| 25164 | ||||
| 25165 | // It's possible that the parent function no longer has a personality function | |||
| 25166 | // if the exceptional code was optimized away, in which case we just return | |||
| 25167 | // the incoming EBP. | |||
| 25168 | if (!Fn->hasPersonalityFn()) | |||
| 25169 | return EntryEBP; | |||
| 25170 | ||||
| 25171 | // Get an MCSymbol that will ultimately resolve to the frame offset of the EH | |||
| 25172 | // registration, or the .set_setframe offset. | |||
| 25173 | MCSymbol *OffsetSym = | |||
| 25174 | MF.getMMI().getContext().getOrCreateParentFrameOffsetSymbol( | |||
| 25175 | GlobalValue::dropLLVMManglingEscape(Fn->getName())); | |||
| 25176 | SDValue OffsetSymVal = DAG.getMCSymbol(OffsetSym, PtrVT); | |||
| 25177 | SDValue ParentFrameOffset = | |||
| 25178 | DAG.getNode(ISD::LOCAL_RECOVER, dl, PtrVT, OffsetSymVal); | |||
| 25179 | ||||
| 25180 | // Return EntryEBP + ParentFrameOffset for x64. This adjusts from RSP after | |||
| 25181 | // prologue to RBP in the parent function. | |||
| 25182 | const X86Subtarget &Subtarget = | |||
| 25183 | static_cast<const X86Subtarget &>(DAG.getSubtarget()); | |||
| 25184 | if (Subtarget.is64Bit()) | |||
| 25185 | return DAG.getNode(ISD::ADD, dl, PtrVT, EntryEBP, ParentFrameOffset); | |||
| 25186 | ||||
| 25187 | int RegNodeSize = getSEHRegistrationNodeSize(Fn); | |||
| 25188 | // RegNodeBase = EntryEBP - RegNodeSize | |||
| 25189 | // ParentFP = RegNodeBase - ParentFrameOffset | |||
| 25190 | SDValue RegNodeBase = DAG.getNode(ISD::SUB, dl, PtrVT, EntryEBP, | |||
| 25191 | DAG.getConstant(RegNodeSize, dl, PtrVT)); | |||
| 25192 | return DAG.getNode(ISD::SUB, dl, PtrVT, RegNodeBase, ParentFrameOffset); | |||
| 25193 | } | |||
| 25194 | ||||
| 25195 | SDValue X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, | |||
| 25196 | SelectionDAG &DAG) const { | |||
| 25197 | // Helper to detect if the operand is CUR_DIRECTION rounding mode. | |||
| 25198 | auto isRoundModeCurDirection = [](SDValue Rnd) { | |||
| 25199 | if (auto *C = dyn_cast<ConstantSDNode>(Rnd)) | |||
| 25200 | return C->getAPIntValue() == X86::STATIC_ROUNDING::CUR_DIRECTION; | |||
| 25201 | ||||
| 25202 | return false; | |||
| 25203 | }; | |||
| 25204 | auto isRoundModeSAE = [](SDValue Rnd) { | |||
| 25205 | if (auto *C = dyn_cast<ConstantSDNode>(Rnd)) { | |||
| 25206 | unsigned RC = C->getZExtValue(); | |||
| 25207 | if (RC & X86::STATIC_ROUNDING::NO_EXC) { | |||
| 25208 | // Clear the NO_EXC bit and check remaining bits. | |||
| 25209 | RC ^= X86::STATIC_ROUNDING::NO_EXC; | |||
| 25210 | // As a convenience we allow no other bits or explicitly | |||
| 25211 | // current direction. | |||
| 25212 | return RC == 0 || RC == X86::STATIC_ROUNDING::CUR_DIRECTION; | |||
| 25213 | } | |||
| 25214 | } | |||
| 25215 | ||||
| 25216 | return false; | |||
| 25217 | }; | |||
| 25218 | auto isRoundModeSAEToX = [](SDValue Rnd, unsigned &RC) { | |||
| 25219 | if (auto *C = dyn_cast<ConstantSDNode>(Rnd)) { | |||
| 25220 | RC = C->getZExtValue(); | |||
| 25221 | if (RC & X86::STATIC_ROUNDING::NO_EXC) { | |||
| 25222 | // Clear the NO_EXC bit and check remaining bits. | |||
| 25223 | RC ^= X86::STATIC_ROUNDING::NO_EXC; | |||
| 25224 | return RC == X86::STATIC_ROUNDING::TO_NEAREST_INT || | |||
| 25225 | RC == X86::STATIC_ROUNDING::TO_NEG_INF || | |||
| 25226 | RC == X86::STATIC_ROUNDING::TO_POS_INF || | |||
| 25227 | RC == X86::STATIC_ROUNDING::TO_ZERO; | |||
| 25228 | } | |||
| 25229 | } | |||
| 25230 | ||||
| 25231 | return false; | |||
| 25232 | }; | |||
| 25233 | ||||
| 25234 | SDLoc dl(Op); | |||
| 25235 | unsigned IntNo = Op.getConstantOperandVal(0); | |||
| 25236 | MVT VT = Op.getSimpleValueType(); | |||
| 25237 | const IntrinsicData* IntrData = getIntrinsicWithoutChain(IntNo); | |||
| 25238 | ||||
| 25239 | // Propagate flags from original node to transformed node(s). | |||
| 25240 | SelectionDAG::FlagInserter FlagsInserter(DAG, Op->getFlags()); | |||
| 25241 | ||||
| 25242 | if (IntrData) { | |||
| 25243 | switch(IntrData->Type) { | |||
| 25244 | case INTR_TYPE_1OP: { | |||
| 25245 | // We specify 2 possible opcodes for intrinsics with rounding modes. | |||
| 25246 | // First, we check if the intrinsic may have non-default rounding mode, | |||
| 25247 | // (IntrData->Opc1 != 0), then we check the rounding mode operand. | |||
| 25248 | unsigned IntrWithRoundingModeOpcode = IntrData->Opc1; | |||
| 25249 | if (IntrWithRoundingModeOpcode != 0) { | |||
| 25250 | SDValue Rnd = Op.getOperand(2); | |||
| 25251 | unsigned RC = 0; | |||
| 25252 | if (isRoundModeSAEToX(Rnd, RC)) | |||
| 25253 | return DAG.getNode(IntrWithRoundingModeOpcode, dl, Op.getValueType(), | |||
| 25254 | Op.getOperand(1), | |||
| 25255 | DAG.getTargetConstant(RC, dl, MVT::i32)); | |||
| 25256 | if (!isRoundModeCurDirection(Rnd)) | |||
| 25257 | return SDValue(); | |||
| 25258 | } | |||
| 25259 | return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), | |||
| 25260 | Op.getOperand(1)); | |||
| 25261 | } | |||
| 25262 | case INTR_TYPE_1OP_SAE: { | |||
| 25263 | SDValue Sae = Op.getOperand(2); | |||
| 25264 | ||||
| 25265 | unsigned Opc; | |||
| 25266 | if (isRoundModeCurDirection(Sae)) | |||
| 25267 | Opc = IntrData->Opc0; | |||
| 25268 | else if (isRoundModeSAE(Sae)) | |||
| 25269 | Opc = IntrData->Opc1; | |||
| 25270 | else | |||
| 25271 | return SDValue(); | |||
| 25272 | ||||
| 25273 | return DAG.getNode(Opc, dl, Op.getValueType(), Op.getOperand(1)); | |||
| 25274 | } | |||
| 25275 | case INTR_TYPE_2OP: { | |||
| 25276 | SDValue Src2 = Op.getOperand(2); | |||
| 25277 | ||||
| 25278 | // We specify 2 possible opcodes for intrinsics with rounding modes. | |||
| 25279 | // First, we check if the intrinsic may have non-default rounding mode, | |||
| 25280 | // (IntrData->Opc1 != 0), then we check the rounding mode operand. | |||
| 25281 | unsigned IntrWithRoundingModeOpcode = IntrData->Opc1; | |||
| 25282 | if (IntrWithRoundingModeOpcode != 0) { | |||
| 25283 | SDValue Rnd = Op.getOperand(3); | |||
| 25284 | unsigned RC = 0; | |||
| 25285 | if (isRoundModeSAEToX(Rnd, RC)) | |||
| 25286 | return DAG.getNode(IntrWithRoundingModeOpcode, dl, Op.getValueType(), | |||
| 25287 | Op.getOperand(1), Src2, | |||
| 25288 | DAG.getTargetConstant(RC, dl, MVT::i32)); | |||
| 25289 | if (!isRoundModeCurDirection(Rnd)) | |||
| 25290 | return SDValue(); | |||
| 25291 | } | |||
| 25292 | ||||
| 25293 | return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), | |||
| 25294 | Op.getOperand(1), Src2); | |||
| 25295 | } | |||
| 25296 | case INTR_TYPE_2OP_SAE: { | |||
| 25297 | SDValue Sae = Op.getOperand(3); | |||
| 25298 | ||||
| 25299 | unsigned Opc; | |||
| 25300 | if (isRoundModeCurDirection(Sae)) | |||
| 25301 | Opc = IntrData->Opc0; | |||
| 25302 | else if (isRoundModeSAE(Sae)) | |||
| 25303 | Opc = IntrData->Opc1; | |||
| 25304 | else | |||
| 25305 | return SDValue(); | |||
| 25306 | ||||
| 25307 | return DAG.getNode(Opc, dl, Op.getValueType(), Op.getOperand(1), | |||
| 25308 | Op.getOperand(2)); | |||
| 25309 | } | |||
| 25310 | case INTR_TYPE_3OP: | |||
| 25311 | case INTR_TYPE_3OP_IMM8: { | |||
| 25312 | SDValue Src1 = Op.getOperand(1); | |||
| 25313 | SDValue Src2 = Op.getOperand(2); | |||
| 25314 | SDValue Src3 = Op.getOperand(3); | |||
| 25315 | ||||
| 25316 | if (IntrData->Type == INTR_TYPE_3OP_IMM8 && | |||
| 25317 | Src3.getValueType() != MVT::i8) { | |||
| 25318 | Src3 = DAG.getTargetConstant( | |||
| 25319 | cast<ConstantSDNode>(Src3)->getZExtValue() & 0xff, dl, MVT::i8); | |||
| 25320 | } | |||
| 25321 | ||||
| 25322 | // We specify 2 possible opcodes for intrinsics with rounding modes. | |||
| 25323 | // First, we check if the intrinsic may have non-default rounding mode, | |||
| 25324 | // (IntrData->Opc1 != 0), then we check the rounding mode operand. | |||
| 25325 | unsigned IntrWithRoundingModeOpcode = IntrData->Opc1; | |||
| 25326 | if (IntrWithRoundingModeOpcode != 0) { | |||
| 25327 | SDValue Rnd = Op.getOperand(4); | |||
| 25328 | unsigned RC = 0; | |||
| 25329 | if (isRoundModeSAEToX(Rnd, RC)) | |||
| 25330 | return DAG.getNode(IntrWithRoundingModeOpcode, dl, Op.getValueType(), | |||
| 25331 | Src1, Src2, Src3, | |||
| 25332 | DAG.getTargetConstant(RC, dl, MVT::i32)); | |||
| 25333 | if (!isRoundModeCurDirection(Rnd)) | |||
| 25334 | return SDValue(); | |||
| 25335 | } | |||
| 25336 | ||||
| 25337 | return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), | |||
| 25338 | {Src1, Src2, Src3}); | |||
| 25339 | } | |||
| 25340 | case INTR_TYPE_4OP_IMM8: { | |||
| 25341 | assert(Op.getOperand(4)->getOpcode() == ISD::TargetConstant)((void)0); | |||
| 25342 | SDValue Src4 = Op.getOperand(4); | |||
| 25343 | if (Src4.getValueType() != MVT::i8) { | |||
| 25344 | Src4 = DAG.getTargetConstant( | |||
| 25345 | cast<ConstantSDNode>(Src4)->getZExtValue() & 0xff, dl, MVT::i8); | |||
| 25346 | } | |||
| 25347 | ||||
| 25348 | return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), | |||
| 25349 | Op.getOperand(1), Op.getOperand(2), Op.getOperand(3), | |||
| 25350 | Src4); | |||
| 25351 | } | |||
| 25352 | case INTR_TYPE_1OP_MASK: { | |||
| 25353 | SDValue Src = Op.getOperand(1); | |||
| 25354 | SDValue PassThru = Op.getOperand(2); | |||
| 25355 | SDValue Mask = Op.getOperand(3); | |||
| 25356 | // We add rounding mode to the Node when | |||
| 25357 | // - RC Opcode is specified and | |||
| 25358 | // - RC is not "current direction". | |||
| 25359 | unsigned IntrWithRoundingModeOpcode = IntrData->Opc1; | |||
| 25360 | if (IntrWithRoundingModeOpcode != 0) { | |||
| 25361 | SDValue Rnd = Op.getOperand(4); | |||
| 25362 | unsigned RC = 0; | |||
| 25363 | if (isRoundModeSAEToX(Rnd, RC)) | |||
| 25364 | return getVectorMaskingNode( | |||
| 25365 | DAG.getNode(IntrWithRoundingModeOpcode, dl, Op.getValueType(), | |||
| 25366 | Src, DAG.getTargetConstant(RC, dl, MVT::i32)), | |||
| 25367 | Mask, PassThru, Subtarget, DAG); | |||
| 25368 | if (!isRoundModeCurDirection(Rnd)) | |||
| 25369 | return SDValue(); | |||
| 25370 | } | |||
| 25371 | return getVectorMaskingNode( | |||
| 25372 | DAG.getNode(IntrData->Opc0, dl, VT, Src), Mask, PassThru, | |||
| 25373 | Subtarget, DAG); | |||
| 25374 | } | |||
| 25375 | case INTR_TYPE_1OP_MASK_SAE: { | |||
| 25376 | SDValue Src = Op.getOperand(1); | |||
| 25377 | SDValue PassThru = Op.getOperand(2); | |||
| 25378 | SDValue Mask = Op.getOperand(3); | |||
| 25379 | SDValue Rnd = Op.getOperand(4); | |||
| 25380 | ||||
| 25381 | unsigned Opc; | |||
| 25382 | if (isRoundModeCurDirection(Rnd)) | |||
| 25383 | Opc = IntrData->Opc0; | |||
| 25384 | else if (isRoundModeSAE(Rnd)) | |||
| 25385 | Opc = IntrData->Opc1; | |||
| 25386 | else | |||
| 25387 | return SDValue(); | |||
| 25388 | ||||
| 25389 | return getVectorMaskingNode(DAG.getNode(Opc, dl, VT, Src), Mask, PassThru, | |||
| 25390 | Subtarget, DAG); | |||
| 25391 | } | |||
| 25392 | case INTR_TYPE_SCALAR_MASK: { | |||
| 25393 | SDValue Src1 = Op.getOperand(1); | |||
| 25394 | SDValue Src2 = Op.getOperand(2); | |||
| 25395 | SDValue passThru = Op.getOperand(3); | |||
| 25396 | SDValue Mask = Op.getOperand(4); | |||
| 25397 | unsigned IntrWithRoundingModeOpcode = IntrData->Opc1; | |||
| 25398 | // There are 2 kinds of intrinsics in this group: | |||
| 25399 | // (1) With suppress-all-exceptions (sae) or rounding mode- 6 operands | |||
| 25400 | // (2) With rounding mode and sae - 7 operands. | |||
| 25401 | bool HasRounding = IntrWithRoundingModeOpcode != 0; | |||
| 25402 | if (Op.getNumOperands() == (5U + HasRounding)) { | |||
| 25403 | if (HasRounding) { | |||
| 25404 | SDValue Rnd = Op.getOperand(5); | |||
| 25405 | unsigned RC = 0; | |||
| 25406 | if (isRoundModeSAEToX(Rnd, RC)) | |||
| 25407 | return getScalarMaskingNode( | |||
| 25408 | DAG.getNode(IntrWithRoundingModeOpcode, dl, VT, Src1, Src2, | |||
| 25409 | DAG.getTargetConstant(RC, dl, MVT::i32)), | |||
| 25410 | Mask, passThru, Subtarget, DAG); | |||
| 25411 | if (!isRoundModeCurDirection(Rnd)) | |||
| 25412 | return SDValue(); | |||
| 25413 | } | |||
| 25414 | return getScalarMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src1, | |||
| 25415 | Src2), | |||
| 25416 | Mask, passThru, Subtarget, DAG); | |||
| 25417 | } | |||
| 25418 | ||||
| 25419 | assert(Op.getNumOperands() == (6U + HasRounding) &&((void)0) | |||
| 25420 | "Unexpected intrinsic form")((void)0); | |||
| 25421 | SDValue RoundingMode = Op.getOperand(5); | |||
| 25422 | unsigned Opc = IntrData->Opc0; | |||
| 25423 | if (HasRounding) { | |||
| 25424 | SDValue Sae = Op.getOperand(6); | |||
| 25425 | if (isRoundModeSAE(Sae)) | |||
| 25426 | Opc = IntrWithRoundingModeOpcode; | |||
| 25427 | else if (!isRoundModeCurDirection(Sae)) | |||
| 25428 | return SDValue(); | |||
| 25429 | } | |||
| 25430 | return getScalarMaskingNode(DAG.getNode(Opc, dl, VT, Src1, | |||
| 25431 | Src2, RoundingMode), | |||
| 25432 | Mask, passThru, Subtarget, DAG); | |||
| 25433 | } | |||
| 25434 | case INTR_TYPE_SCALAR_MASK_RND: { | |||
| 25435 | SDValue Src1 = Op.getOperand(1); | |||
| 25436 | SDValue Src2 = Op.getOperand(2); | |||
| 25437 | SDValue passThru = Op.getOperand(3); | |||
| 25438 | SDValue Mask = Op.getOperand(4); | |||
| 25439 | SDValue Rnd = Op.getOperand(5); | |||
| 25440 | ||||
| 25441 | SDValue NewOp; | |||
| 25442 | unsigned RC = 0; | |||
| 25443 | if (isRoundModeCurDirection(Rnd)) | |||
| 25444 | NewOp = DAG.getNode(IntrData->Opc0, dl, VT, Src1, Src2); | |||
| 25445 | else if (isRoundModeSAEToX(Rnd, RC)) | |||
| 25446 | NewOp = DAG.getNode(IntrData->Opc1, dl, VT, Src1, Src2, | |||
| 25447 | DAG.getTargetConstant(RC, dl, MVT::i32)); | |||
| 25448 | else | |||
| 25449 | return SDValue(); | |||
| 25450 | ||||
| 25451 | return getScalarMaskingNode(NewOp, Mask, passThru, Subtarget, DAG); | |||
| 25452 | } | |||
| 25453 | case INTR_TYPE_SCALAR_MASK_SAE: { | |||
| 25454 | SDValue Src1 = Op.getOperand(1); | |||
| 25455 | SDValue Src2 = Op.getOperand(2); | |||
| 25456 | SDValue passThru = Op.getOperand(3); | |||
| 25457 | SDValue Mask = Op.getOperand(4); | |||
| 25458 | SDValue Sae = Op.getOperand(5); | |||
| 25459 | unsigned Opc; | |||
| 25460 | if (isRoundModeCurDirection(Sae)) | |||
| 25461 | Opc = IntrData->Opc0; | |||
| 25462 | else if (isRoundModeSAE(Sae)) | |||
| 25463 | Opc = IntrData->Opc1; | |||
| 25464 | else | |||
| 25465 | return SDValue(); | |||
| 25466 | ||||
| 25467 | return getScalarMaskingNode(DAG.getNode(Opc, dl, VT, Src1, Src2), | |||
| 25468 | Mask, passThru, Subtarget, DAG); | |||
| 25469 | } | |||
| 25470 | case INTR_TYPE_2OP_MASK: { | |||
| 25471 | SDValue Src1 = Op.getOperand(1); | |||
| 25472 | SDValue Src2 = Op.getOperand(2); | |||
| 25473 | SDValue PassThru = Op.getOperand(3); | |||
| 25474 | SDValue Mask = Op.getOperand(4); | |||
| 25475 | SDValue NewOp; | |||
| 25476 | if (IntrData->Opc1 != 0) { | |||
| 25477 | SDValue Rnd = Op.getOperand(5); | |||
| 25478 | unsigned RC = 0; | |||
| 25479 | if (isRoundModeSAEToX(Rnd, RC)) | |||
| 25480 | NewOp = DAG.getNode(IntrData->Opc1, dl, VT, Src1, Src2, | |||
| 25481 | DAG.getTargetConstant(RC, dl, MVT::i32)); | |||
| 25482 | else if (!isRoundModeCurDirection(Rnd)) | |||
| 25483 | return SDValue(); | |||
| 25484 | } | |||
| 25485 | if (!NewOp) | |||
| 25486 | NewOp = DAG.getNode(IntrData->Opc0, dl, VT, Src1, Src2); | |||
| 25487 | return getVectorMaskingNode(NewOp, Mask, PassThru, Subtarget, DAG); | |||
| 25488 | } | |||
| 25489 | case INTR_TYPE_2OP_MASK_SAE: { | |||
| 25490 | SDValue Src1 = Op.getOperand(1); | |||
| 25491 | SDValue Src2 = Op.getOperand(2); | |||
| 25492 | SDValue PassThru = Op.getOperand(3); | |||
| 25493 | SDValue Mask = Op.getOperand(4); | |||
| 25494 | ||||
| 25495 | unsigned Opc = IntrData->Opc0; | |||
| 25496 | if (IntrData->Opc1 != 0) { | |||
| 25497 | SDValue Sae = Op.getOperand(5); | |||
| 25498 | if (isRoundModeSAE(Sae)) | |||
| 25499 | Opc = IntrData->Opc1; | |||
| 25500 | else if (!isRoundModeCurDirection(Sae)) | |||
| 25501 | return SDValue(); | |||
| 25502 | } | |||
| 25503 | ||||
| 25504 | return getVectorMaskingNode(DAG.getNode(Opc, dl, VT, Src1, Src2), | |||
| 25505 | Mask, PassThru, Subtarget, DAG); | |||
| 25506 | } | |||
| 25507 | case INTR_TYPE_3OP_SCALAR_MASK_SAE: { | |||
| 25508 | SDValue Src1 = Op.getOperand(1); | |||
| 25509 | SDValue Src2 = Op.getOperand(2); | |||
| 25510 | SDValue Src3 = Op.getOperand(3); | |||
| 25511 | SDValue PassThru = Op.getOperand(4); | |||
| 25512 | SDValue Mask = Op.getOperand(5); | |||
| 25513 | SDValue Sae = Op.getOperand(6); | |||
| 25514 | unsigned Opc; | |||
| 25515 | if (isRoundModeCurDirection(Sae)) | |||
| 25516 | Opc = IntrData->Opc0; | |||
| 25517 | else if (isRoundModeSAE(Sae)) | |||
| 25518 | Opc = IntrData->Opc1; | |||
| 25519 | else | |||
| 25520 | return SDValue(); | |||
| 25521 | ||||
| 25522 | return getScalarMaskingNode(DAG.getNode(Opc, dl, VT, Src1, Src2, Src3), | |||
| 25523 | Mask, PassThru, Subtarget, DAG); | |||
| 25524 | } | |||
| 25525 | case INTR_TYPE_3OP_MASK_SAE: { | |||
| 25526 | SDValue Src1 = Op.getOperand(1); | |||
| 25527 | SDValue Src2 = Op.getOperand(2); | |||
| 25528 | SDValue Src3 = Op.getOperand(3); | |||
| 25529 | SDValue PassThru = Op.getOperand(4); | |||
| 25530 | SDValue Mask = Op.getOperand(5); | |||
| 25531 | ||||
| 25532 | unsigned Opc = IntrData->Opc0; | |||
| 25533 | if (IntrData->Opc1 != 0) { | |||
| 25534 | SDValue Sae = Op.getOperand(6); | |||
| 25535 | if (isRoundModeSAE(Sae)) | |||
| 25536 | Opc = IntrData->Opc1; | |||
| 25537 | else if (!isRoundModeCurDirection(Sae)) | |||
| 25538 | return SDValue(); | |||
| 25539 | } | |||
| 25540 | return getVectorMaskingNode(DAG.getNode(Opc, dl, VT, Src1, Src2, Src3), | |||
| 25541 | Mask, PassThru, Subtarget, DAG); | |||
| 25542 | } | |||
| 25543 | case BLENDV: { | |||
| 25544 | SDValue Src1 = Op.getOperand(1); | |||
| 25545 | SDValue Src2 = Op.getOperand(2); | |||
| 25546 | SDValue Src3 = Op.getOperand(3); | |||
| 25547 | ||||
| 25548 | EVT MaskVT = Src3.getValueType().changeVectorElementTypeToInteger(); | |||
| 25549 | Src3 = DAG.getBitcast(MaskVT, Src3); | |||
| 25550 | ||||
| 25551 | // Reverse the operands to match VSELECT order. | |||
| 25552 | return DAG.getNode(IntrData->Opc0, dl, VT, Src3, Src2, Src1); | |||
| 25553 | } | |||
| 25554 | case VPERM_2OP : { | |||
| 25555 | SDValue Src1 = Op.getOperand(1); | |||
| 25556 | SDValue Src2 = Op.getOperand(2); | |||
| 25557 | ||||
| 25558 | // Swap Src1 and Src2 in the node creation | |||
| 25559 | return DAG.getNode(IntrData->Opc0, dl, VT,Src2, Src1); | |||
| 25560 | } | |||
| 25561 | case IFMA_OP: | |||
| 25562 | // NOTE: We need to swizzle the operands to pass the multiply operands | |||
| 25563 | // first. | |||
| 25564 | return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), | |||
| 25565 | Op.getOperand(2), Op.getOperand(3), Op.getOperand(1)); | |||
| 25566 | case FPCLASSS: { | |||
| 25567 | SDValue Src1 = Op.getOperand(1); | |||
| 25568 | SDValue Imm = Op.getOperand(2); | |||
| 25569 | SDValue Mask = Op.getOperand(3); | |||
| 25570 | SDValue FPclass = DAG.getNode(IntrData->Opc0, dl, MVT::v1i1, Src1, Imm); | |||
| 25571 | SDValue FPclassMask = getScalarMaskingNode(FPclass, Mask, SDValue(), | |||
| 25572 | Subtarget, DAG); | |||
| 25573 | // Need to fill with zeros to ensure the bitcast will produce zeroes | |||
| 25574 | // for the upper bits. An EXTRACT_ELEMENT here wouldn't guarantee that. | |||
| 25575 | SDValue Ins = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, MVT::v8i1, | |||
| 25576 | DAG.getConstant(0, dl, MVT::v8i1), | |||
| 25577 | FPclassMask, DAG.getIntPtrConstant(0, dl)); | |||
| 25578 | return DAG.getBitcast(MVT::i8, Ins); | |||
| 25579 | } | |||
| 25580 | ||||
| 25581 | case CMP_MASK_CC: { | |||
| 25582 | MVT MaskVT = Op.getSimpleValueType(); | |||
| 25583 | SDValue CC = Op.getOperand(3); | |||
| 25584 | SDValue Mask = Op.getOperand(4); | |||
| 25585 | // We specify 2 possible opcodes for intrinsics with rounding modes. | |||
| 25586 | // First, we check if the intrinsic may have non-default rounding mode, | |||
| 25587 | // (IntrData->Opc1 != 0), then we check the rounding mode operand. | |||
| 25588 | if (IntrData->Opc1 != 0) { | |||
| 25589 | SDValue Sae = Op.getOperand(5); | |||
| 25590 | if (isRoundModeSAE(Sae)) | |||
| 25591 | return DAG.getNode(IntrData->Opc1, dl, MaskVT, Op.getOperand(1), | |||
| 25592 | Op.getOperand(2), CC, Mask, Sae); | |||
| 25593 | if (!isRoundModeCurDirection(Sae)) | |||
| 25594 | return SDValue(); | |||
| 25595 | } | |||
| 25596 | //default rounding mode | |||
| 25597 | return DAG.getNode(IntrData->Opc0, dl, MaskVT, | |||
| 25598 | {Op.getOperand(1), Op.getOperand(2), CC, Mask}); | |||
| 25599 | } | |||
| 25600 | case CMP_MASK_SCALAR_CC: { | |||
| 25601 | SDValue Src1 = Op.getOperand(1); | |||
| 25602 | SDValue Src2 = Op.getOperand(2); | |||
| 25603 | SDValue CC = Op.getOperand(3); | |||
| 25604 | SDValue Mask = Op.getOperand(4); | |||
| 25605 | ||||
| 25606 | SDValue Cmp; | |||
| 25607 | if (IntrData->Opc1 != 0) { | |||
| 25608 | SDValue Sae = Op.getOperand(5); | |||
| 25609 | if (isRoundModeSAE(Sae)) | |||
| 25610 | Cmp = DAG.getNode(IntrData->Opc1, dl, MVT::v1i1, Src1, Src2, CC, Sae); | |||
| 25611 | else if (!isRoundModeCurDirection(Sae)) | |||
| 25612 | return SDValue(); | |||
| 25613 | } | |||
| 25614 | //default rounding mode | |||
| 25615 | if (!Cmp.getNode()) | |||
| 25616 | Cmp = DAG.getNode(IntrData->Opc0, dl, MVT::v1i1, Src1, Src2, CC); | |||
| 25617 | ||||
| 25618 | SDValue CmpMask = getScalarMaskingNode(Cmp, Mask, SDValue(), | |||
| 25619 | Subtarget, DAG); | |||
| 25620 | // Need to fill with zeros to ensure the bitcast will produce zeroes | |||
| 25621 | // for the upper bits. An EXTRACT_ELEMENT here wouldn't guarantee that. | |||
| 25622 | SDValue Ins = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, MVT::v8i1, | |||
| 25623 | DAG.getConstant(0, dl, MVT::v8i1), | |||
| 25624 | CmpMask, DAG.getIntPtrConstant(0, dl)); | |||
| 25625 | return DAG.getBitcast(MVT::i8, Ins); | |||
| 25626 | } | |||
| 25627 | case COMI: { // Comparison intrinsics | |||
| 25628 | ISD::CondCode CC = (ISD::CondCode)IntrData->Opc1; | |||
| 25629 | SDValue LHS = Op.getOperand(1); | |||
| 25630 | SDValue RHS = Op.getOperand(2); | |||
| 25631 | // Some conditions require the operands to be swapped. | |||
| 25632 | if (CC == ISD::SETLT || CC == ISD::SETLE) | |||
| 25633 | std::swap(LHS, RHS); | |||
| 25634 | ||||
| 25635 | SDValue Comi = DAG.getNode(IntrData->Opc0, dl, MVT::i32, LHS, RHS); | |||
| 25636 | SDValue SetCC; | |||
| 25637 | switch (CC) { | |||
| 25638 | case ISD::SETEQ: { // (ZF = 0 and PF = 0) | |||
| 25639 | SetCC = getSETCC(X86::COND_E, Comi, dl, DAG); | |||
| 25640 | SDValue SetNP = getSETCC(X86::COND_NP, Comi, dl, DAG); | |||
| 25641 | SetCC = DAG.getNode(ISD::AND, dl, MVT::i8, SetCC, SetNP); | |||
| 25642 | break; | |||
| 25643 | } | |||
| 25644 | case ISD::SETNE: { // (ZF = 1 or PF = 1) | |||
| 25645 | SetCC = getSETCC(X86::COND_NE, Comi, dl, DAG); | |||
| 25646 | SDValue SetP = getSETCC(X86::COND_P, Comi, dl, DAG); | |||
| 25647 | SetCC = DAG.getNode(ISD::OR, dl, MVT::i8, SetCC, SetP); | |||
| 25648 | break; | |||
| 25649 | } | |||
| 25650 | case ISD::SETGT: // (CF = 0 and ZF = 0) | |||
| 25651 | case ISD::SETLT: { // Condition opposite to GT. Operands swapped above. | |||
| 25652 | SetCC = getSETCC(X86::COND_A, Comi, dl, DAG); | |||
| 25653 | break; | |||
| 25654 | } | |||
| 25655 | case ISD::SETGE: // CF = 0 | |||
| 25656 | case ISD::SETLE: // Condition opposite to GE. Operands swapped above. | |||
| 25657 | SetCC = getSETCC(X86::COND_AE, Comi, dl, DAG); | |||
| 25658 | break; | |||
| 25659 | default: | |||
| 25660 | llvm_unreachable("Unexpected illegal condition!")__builtin_unreachable(); | |||
| 25661 | } | |||
| 25662 | return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC); | |||
| 25663 | } | |||
| 25664 | case COMI_RM: { // Comparison intrinsics with Sae | |||
| 25665 | SDValue LHS = Op.getOperand(1); | |||
| 25666 | SDValue RHS = Op.getOperand(2); | |||
| 25667 | unsigned CondVal = Op.getConstantOperandVal(3); | |||
| 25668 | SDValue Sae = Op.getOperand(4); | |||
| 25669 | ||||
| 25670 | SDValue FCmp; | |||
| 25671 | if (isRoundModeCurDirection(Sae)) | |||
| 25672 | FCmp = DAG.getNode(X86ISD::FSETCCM, dl, MVT::v1i1, LHS, RHS, | |||
| 25673 | DAG.getTargetConstant(CondVal, dl, MVT::i8)); | |||
| 25674 | else if (isRoundModeSAE(Sae)) | |||
| 25675 | FCmp = DAG.getNode(X86ISD::FSETCCM_SAE, dl, MVT::v1i1, LHS, RHS, | |||
| 25676 | DAG.getTargetConstant(CondVal, dl, MVT::i8), Sae); | |||
| 25677 | else | |||
| 25678 | return SDValue(); | |||
| 25679 | // Need to fill with zeros to ensure the bitcast will produce zeroes | |||
| 25680 | // for the upper bits. An EXTRACT_ELEMENT here wouldn't guarantee that. | |||
| 25681 | SDValue Ins = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, MVT::v16i1, | |||
| 25682 | DAG.getConstant(0, dl, MVT::v16i1), | |||
| 25683 | FCmp, DAG.getIntPtrConstant(0, dl)); | |||
| 25684 | return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, | |||
| 25685 | DAG.getBitcast(MVT::i16, Ins)); | |||
| 25686 | } | |||
| 25687 | case VSHIFT: | |||
| 25688 | return getTargetVShiftNode(IntrData->Opc0, dl, Op.getSimpleValueType(), | |||
| 25689 | Op.getOperand(1), Op.getOperand(2), Subtarget, | |||
| 25690 | DAG); | |||
| 25691 | case COMPRESS_EXPAND_IN_REG: { | |||
| 25692 | SDValue Mask = Op.getOperand(3); | |||
| 25693 | SDValue DataToCompress = Op.getOperand(1); | |||
| 25694 | SDValue PassThru = Op.getOperand(2); | |||
| 25695 | if (ISD::isBuildVectorAllOnes(Mask.getNode())) // return data as is | |||
| 25696 | return Op.getOperand(1); | |||
| 25697 | ||||
| 25698 | // Avoid false dependency. | |||
| 25699 | if (PassThru.isUndef()) | |||
| 25700 | PassThru = DAG.getConstant(0, dl, VT); | |||
| 25701 | ||||
| 25702 | return DAG.getNode(IntrData->Opc0, dl, VT, DataToCompress, PassThru, | |||
| 25703 | Mask); | |||
| 25704 | } | |||
| 25705 | case FIXUPIMM: | |||
| 25706 | case FIXUPIMM_MASKZ: { | |||
| 25707 | SDValue Src1 = Op.getOperand(1); | |||
| 25708 | SDValue Src2 = Op.getOperand(2); | |||
| 25709 | SDValue Src3 = Op.getOperand(3); | |||
| 25710 | SDValue Imm = Op.getOperand(4); | |||
| 25711 | SDValue Mask = Op.getOperand(5); | |||
| 25712 | SDValue Passthru = (IntrData->Type == FIXUPIMM) | |||
| 25713 | ? Src1 | |||
| 25714 | : getZeroVector(VT, Subtarget, DAG, dl); | |||
| 25715 | ||||
| 25716 | unsigned Opc = IntrData->Opc0; | |||
| 25717 | if (IntrData->Opc1 != 0) { | |||
| 25718 | SDValue Sae = Op.getOperand(6); | |||
| 25719 | if (isRoundModeSAE(Sae)) | |||
| 25720 | Opc = IntrData->Opc1; | |||
| 25721 | else if (!isRoundModeCurDirection(Sae)) | |||
| 25722 | return SDValue(); | |||
| 25723 | } | |||
| 25724 | ||||
| 25725 | SDValue FixupImm = DAG.getNode(Opc, dl, VT, Src1, Src2, Src3, Imm); | |||
| 25726 | ||||
| 25727 | if (Opc == X86ISD::VFIXUPIMM || Opc == X86ISD::VFIXUPIMM_SAE) | |||
| 25728 | return getVectorMaskingNode(FixupImm, Mask, Passthru, Subtarget, DAG); | |||
| 25729 | ||||
| 25730 | return getScalarMaskingNode(FixupImm, Mask, Passthru, Subtarget, DAG); | |||
| 25731 | } | |||
| 25732 | case ROUNDP: { | |||
| 25733 | assert(IntrData->Opc0 == X86ISD::VRNDSCALE && "Unexpected opcode")((void)0); | |||
| 25734 | // Clear the upper bits of the rounding immediate so that the legacy | |||
| 25735 | // intrinsic can't trigger the scaling behavior of VRNDSCALE. | |||
| 25736 | auto Round = cast<ConstantSDNode>(Op.getOperand(2)); | |||
| 25737 | SDValue RoundingMode = | |||
| 25738 | DAG.getTargetConstant(Round->getZExtValue() & 0xf, dl, MVT::i32); | |||
| 25739 | return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), | |||
| 25740 | Op.getOperand(1), RoundingMode); | |||
| 25741 | } | |||
| 25742 | case ROUNDS: { | |||
| 25743 | assert(IntrData->Opc0 == X86ISD::VRNDSCALES && "Unexpected opcode")((void)0); | |||
| 25744 | // Clear the upper bits of the rounding immediate so that the legacy | |||
| 25745 | // intrinsic can't trigger the scaling behavior of VRNDSCALE. | |||
| 25746 | auto Round = cast<ConstantSDNode>(Op.getOperand(3)); | |||
| 25747 | SDValue RoundingMode = | |||
| 25748 | DAG.getTargetConstant(Round->getZExtValue() & 0xf, dl, MVT::i32); | |||
| 25749 | return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), | |||
| 25750 | Op.getOperand(1), Op.getOperand(2), RoundingMode); | |||
| 25751 | } | |||
| 25752 | case BEXTRI: { | |||
| 25753 | assert(IntrData->Opc0 == X86ISD::BEXTRI && "Unexpected opcode")((void)0); | |||
| 25754 | ||||
| 25755 | uint64_t Imm = Op.getConstantOperandVal(2); | |||
| 25756 | SDValue Control = DAG.getTargetConstant(Imm & 0xffff, dl, | |||
| 25757 | Op.getValueType()); | |||
| 25758 | return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), | |||
| 25759 | Op.getOperand(1), Control); | |||
| 25760 | } | |||
| 25761 | // ADC/ADCX/SBB | |||
| 25762 | case ADX: { | |||
| 25763 | SDVTList CFVTs = DAG.getVTList(Op->getValueType(0), MVT::i32); | |||
| 25764 | SDVTList VTs = DAG.getVTList(Op.getOperand(2).getValueType(), MVT::i32); | |||
| 25765 | ||||
| 25766 | SDValue Res; | |||
| 25767 | // If the carry in is zero, then we should just use ADD/SUB instead of | |||
| 25768 | // ADC/SBB. | |||
| 25769 | if (isNullConstant(Op.getOperand(1))) { | |||
| 25770 | Res = DAG.getNode(IntrData->Opc1, dl, VTs, Op.getOperand(2), | |||
| 25771 | Op.getOperand(3)); | |||
| 25772 | } else { | |||
| 25773 | SDValue GenCF = DAG.getNode(X86ISD::ADD, dl, CFVTs, Op.getOperand(1), | |||
| 25774 | DAG.getConstant(-1, dl, MVT::i8)); | |||
| 25775 | Res = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(2), | |||
| 25776 | Op.getOperand(3), GenCF.getValue(1)); | |||
| 25777 | } | |||
| 25778 | SDValue SetCC = getSETCC(X86::COND_B, Res.getValue(1), dl, DAG); | |||
| 25779 | SDValue Results[] = { SetCC, Res }; | |||
| 25780 | return DAG.getMergeValues(Results, dl); | |||
| 25781 | } | |||
| 25782 | case CVTPD2PS_MASK: | |||
| 25783 | case CVTPD2DQ_MASK: | |||
| 25784 | case CVTQQ2PS_MASK: | |||
| 25785 | case TRUNCATE_TO_REG: { | |||
| 25786 | SDValue Src = Op.getOperand(1); | |||
| 25787 | SDValue PassThru = Op.getOperand(2); | |||
| 25788 | SDValue Mask = Op.getOperand(3); | |||
| 25789 | ||||
| 25790 | if (isAllOnesConstant(Mask)) | |||
| 25791 | return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Src); | |||
| 25792 | ||||
| 25793 | MVT SrcVT = Src.getSimpleValueType(); | |||
| 25794 | MVT MaskVT = MVT::getVectorVT(MVT::i1, SrcVT.getVectorNumElements()); | |||
| 25795 | Mask = getMaskNode(Mask, MaskVT, Subtarget, DAG, dl); | |||
| 25796 | return DAG.getNode(IntrData->Opc1, dl, Op.getValueType(), | |||
| 25797 | {Src, PassThru, Mask}); | |||
| 25798 | } | |||
| 25799 | case CVTPS2PH_MASK: { | |||
| 25800 | SDValue Src = Op.getOperand(1); | |||
| 25801 | SDValue Rnd = Op.getOperand(2); | |||
| 25802 | SDValue PassThru = Op.getOperand(3); | |||
| 25803 | SDValue Mask = Op.getOperand(4); | |||
| 25804 | ||||
| 25805 | if (isAllOnesConstant(Mask)) | |||
| 25806 | return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Src, Rnd); | |||
| 25807 | ||||
| 25808 | MVT SrcVT = Src.getSimpleValueType(); | |||
| 25809 | MVT MaskVT = MVT::getVectorVT(MVT::i1, SrcVT.getVectorNumElements()); | |||
| 25810 | Mask = getMaskNode(Mask, MaskVT, Subtarget, DAG, dl); | |||
| 25811 | return DAG.getNode(IntrData->Opc1, dl, Op.getValueType(), Src, Rnd, | |||
| 25812 | PassThru, Mask); | |||
| 25813 | ||||
| 25814 | } | |||
| 25815 | case CVTNEPS2BF16_MASK: { | |||
| 25816 | SDValue Src = Op.getOperand(1); | |||
| 25817 | SDValue PassThru = Op.getOperand(2); | |||
| 25818 | SDValue Mask = Op.getOperand(3); | |||
| 25819 | ||||
| 25820 | if (ISD::isBuildVectorAllOnes(Mask.getNode())) | |||
| 25821 | return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Src); | |||
| 25822 | ||||
| 25823 | // Break false dependency. | |||
| 25824 | if (PassThru.isUndef()) | |||
| 25825 | PassThru = DAG.getConstant(0, dl, PassThru.getValueType()); | |||
| 25826 | ||||
| 25827 | return DAG.getNode(IntrData->Opc1, dl, Op.getValueType(), Src, PassThru, | |||
| 25828 | Mask); | |||
| 25829 | } | |||
| 25830 | default: | |||
| 25831 | break; | |||
| 25832 | } | |||
| 25833 | } | |||
| 25834 | ||||
| 25835 | switch (IntNo) { | |||
| 25836 | default: return SDValue(); // Don't custom lower most intrinsics. | |||
| 25837 | ||||
| 25838 | // ptest and testp intrinsics. The intrinsic these come from are designed to | |||
| 25839 | // return an integer value, not just an instruction so lower it to the ptest | |||
| 25840 | // or testp pattern and a setcc for the result. | |||
| 25841 | case Intrinsic::x86_avx512_ktestc_b: | |||
| 25842 | case Intrinsic::x86_avx512_ktestc_w: | |||
| 25843 | case Intrinsic::x86_avx512_ktestc_d: | |||
| 25844 | case Intrinsic::x86_avx512_ktestc_q: | |||
| 25845 | case Intrinsic::x86_avx512_ktestz_b: | |||
| 25846 | case Intrinsic::x86_avx512_ktestz_w: | |||
| 25847 | case Intrinsic::x86_avx512_ktestz_d: | |||
| 25848 | case Intrinsic::x86_avx512_ktestz_q: | |||
| 25849 | case Intrinsic::x86_sse41_ptestz: | |||
| 25850 | case Intrinsic::x86_sse41_ptestc: | |||
| 25851 | case Intrinsic::x86_sse41_ptestnzc: | |||
| 25852 | case Intrinsic::x86_avx_ptestz_256: | |||
| 25853 | case Intrinsic::x86_avx_ptestc_256: | |||
| 25854 | case Intrinsic::x86_avx_ptestnzc_256: | |||
| 25855 | case Intrinsic::x86_avx_vtestz_ps: | |||
| 25856 | case Intrinsic::x86_avx_vtestc_ps: | |||
| 25857 | case Intrinsic::x86_avx_vtestnzc_ps: | |||
| 25858 | case Intrinsic::x86_avx_vtestz_pd: | |||
| 25859 | case Intrinsic::x86_avx_vtestc_pd: | |||
| 25860 | case Intrinsic::x86_avx_vtestnzc_pd: | |||
| 25861 | case Intrinsic::x86_avx_vtestz_ps_256: | |||
| 25862 | case Intrinsic::x86_avx_vtestc_ps_256: | |||
| 25863 | case Intrinsic::x86_avx_vtestnzc_ps_256: | |||
| 25864 | case Intrinsic::x86_avx_vtestz_pd_256: | |||
| 25865 | case Intrinsic::x86_avx_vtestc_pd_256: | |||
| 25866 | case Intrinsic::x86_avx_vtestnzc_pd_256: { | |||
| 25867 | unsigned TestOpc = X86ISD::PTEST; | |||
| 25868 | X86::CondCode X86CC; | |||
| 25869 | switch (IntNo) { | |||
| 25870 | default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.")__builtin_unreachable(); | |||
| 25871 | case Intrinsic::x86_avx512_ktestc_b: | |||
| 25872 | case Intrinsic::x86_avx512_ktestc_w: | |||
| 25873 | case Intrinsic::x86_avx512_ktestc_d: | |||
| 25874 | case Intrinsic::x86_avx512_ktestc_q: | |||
| 25875 | // CF = 1 | |||
| 25876 | TestOpc = X86ISD::KTEST; | |||
| 25877 | X86CC = X86::COND_B; | |||
| 25878 | break; | |||
| 25879 | case Intrinsic::x86_avx512_ktestz_b: | |||
| 25880 | case Intrinsic::x86_avx512_ktestz_w: | |||
| 25881 | case Intrinsic::x86_avx512_ktestz_d: | |||
| 25882 | case Intrinsic::x86_avx512_ktestz_q: | |||
| 25883 | TestOpc = X86ISD::KTEST; | |||
| 25884 | X86CC = X86::COND_E; | |||
| 25885 | break; | |||
| 25886 | case Intrinsic::x86_avx_vtestz_ps: | |||
| 25887 | case Intrinsic::x86_avx_vtestz_pd: | |||
| 25888 | case Intrinsic::x86_avx_vtestz_ps_256: | |||
| 25889 | case Intrinsic::x86_avx_vtestz_pd_256: | |||
| 25890 | TestOpc = X86ISD::TESTP; | |||
| 25891 | LLVM_FALLTHROUGH[[gnu::fallthrough]]; | |||
| 25892 | case Intrinsic::x86_sse41_ptestz: | |||
| 25893 | case Intrinsic::x86_avx_ptestz_256: | |||
| 25894 | // ZF = 1 | |||
| 25895 | X86CC = X86::COND_E; | |||
| 25896 | break; | |||
| 25897 | case Intrinsic::x86_avx_vtestc_ps: | |||
| 25898 | case Intrinsic::x86_avx_vtestc_pd: | |||
| 25899 | case Intrinsic::x86_avx_vtestc_ps_256: | |||
| 25900 | case Intrinsic::x86_avx_vtestc_pd_256: | |||
| 25901 | TestOpc = X86ISD::TESTP; | |||
| 25902 | LLVM_FALLTHROUGH[[gnu::fallthrough]]; | |||
| 25903 | case Intrinsic::x86_sse41_ptestc: | |||
| 25904 | case Intrinsic::x86_avx_ptestc_256: | |||
| 25905 | // CF = 1 | |||
| 25906 | X86CC = X86::COND_B; | |||
| 25907 | break; | |||
| 25908 | case Intrinsic::x86_avx_vtestnzc_ps: | |||
| 25909 | case Intrinsic::x86_avx_vtestnzc_pd: | |||
| 25910 | case Intrinsic::x86_avx_vtestnzc_ps_256: | |||
| 25911 | case Intrinsic::x86_avx_vtestnzc_pd_256: | |||
| 25912 | TestOpc = X86ISD::TESTP; | |||
| 25913 | LLVM_FALLTHROUGH[[gnu::fallthrough]]; | |||
| 25914 | case Intrinsic::x86_sse41_ptestnzc: | |||
| 25915 | case Intrinsic::x86_avx_ptestnzc_256: | |||
| 25916 | // ZF and CF = 0 | |||
| 25917 | X86CC = X86::COND_A; | |||
| 25918 | break; | |||
| 25919 | } | |||
| 25920 | ||||
| 25921 | SDValue LHS = Op.getOperand(1); | |||
| 25922 | SDValue RHS = Op.getOperand(2); | |||
| 25923 | SDValue Test = DAG.getNode(TestOpc, dl, MVT::i32, LHS, RHS); | |||
| 25924 | SDValue SetCC = getSETCC(X86CC, Test, dl, DAG); | |||
| 25925 | return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC); | |||
| 25926 | } | |||
| 25927 | ||||
| 25928 | case Intrinsic::x86_sse42_pcmpistria128: | |||
| 25929 | case Intrinsic::x86_sse42_pcmpestria128: | |||
| 25930 | case Intrinsic::x86_sse42_pcmpistric128: | |||
| 25931 | case Intrinsic::x86_sse42_pcmpestric128: | |||
| 25932 | case Intrinsic::x86_sse42_pcmpistrio128: | |||
| 25933 | case Intrinsic::x86_sse42_pcmpestrio128: | |||
| 25934 | case Intrinsic::x86_sse42_pcmpistris128: | |||
| 25935 | case Intrinsic::x86_sse42_pcmpestris128: | |||
| 25936 | case Intrinsic::x86_sse42_pcmpistriz128: | |||
| 25937 | case Intrinsic::x86_sse42_pcmpestriz128: { | |||
| 25938 | unsigned Opcode; | |||
| 25939 | X86::CondCode X86CC; | |||
| 25940 | switch (IntNo) { | |||
| 25941 | default: llvm_unreachable("Impossible intrinsic")__builtin_unreachable(); // Can't reach here. | |||
| 25942 | case Intrinsic::x86_sse42_pcmpistria128: | |||
| 25943 | Opcode = X86ISD::PCMPISTR; | |||
| 25944 | X86CC = X86::COND_A; | |||
| 25945 | break; | |||
| 25946 | case Intrinsic::x86_sse42_pcmpestria128: | |||
| 25947 | Opcode = X86ISD::PCMPESTR; | |||
| 25948 | X86CC = X86::COND_A; | |||
| 25949 | break; | |||
| 25950 | case Intrinsic::x86_sse42_pcmpistric128: | |||
| 25951 | Opcode = X86ISD::PCMPISTR; | |||
| 25952 | X86CC = X86::COND_B; | |||
| 25953 | break; | |||
| 25954 | case Intrinsic::x86_sse42_pcmpestric128: | |||
| 25955 | Opcode = X86ISD::PCMPESTR; | |||
| 25956 | X86CC = X86::COND_B; | |||
| 25957 | break; | |||
| 25958 | case Intrinsic::x86_sse42_pcmpistrio128: | |||
| 25959 | Opcode = X86ISD::PCMPISTR; | |||
| 25960 | X86CC = X86::COND_O; | |||
| 25961 | break; | |||
| 25962 | case Intrinsic::x86_sse42_pcmpestrio128: | |||
| 25963 | Opcode = X86ISD::PCMPESTR; | |||
| 25964 | X86CC = X86::COND_O; | |||
| 25965 | break; | |||
| 25966 | case Intrinsic::x86_sse42_pcmpistris128: | |||
| 25967 | Opcode = X86ISD::PCMPISTR; | |||
| 25968 | X86CC = X86::COND_S; | |||
| 25969 | break; | |||
| 25970 | case Intrinsic::x86_sse42_pcmpestris128: | |||
| 25971 | Opcode = X86ISD::PCMPESTR; | |||
| 25972 | X86CC = X86::COND_S; | |||
| 25973 | break; | |||
| 25974 | case Intrinsic::x86_sse42_pcmpistriz128: | |||
| 25975 | Opcode = X86ISD::PCMPISTR; | |||
| 25976 | X86CC = X86::COND_E; | |||
| 25977 | break; | |||
| 25978 | case Intrinsic::x86_sse42_pcmpestriz128: | |||
| 25979 | Opcode = X86ISD::PCMPESTR; | |||
| 25980 | X86CC = X86::COND_E; | |||
| 25981 | break; | |||
| 25982 | } | |||
| 25983 | SmallVector<SDValue, 5> NewOps(Op->op_begin()+1, Op->op_end()); | |||
| 25984 | SDVTList VTs = DAG.getVTList(MVT::i32, MVT::v16i8, MVT::i32); | |||
| 25985 | SDValue PCMP = DAG.getNode(Opcode, dl, VTs, NewOps).getValue(2); | |||
| 25986 | SDValue SetCC = getSETCC(X86CC, PCMP, dl, DAG); | |||
| 25987 | return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC); | |||
| 25988 | } | |||
| 25989 | ||||
| 25990 | case Intrinsic::x86_sse42_pcmpistri128: | |||
| 25991 | case Intrinsic::x86_sse42_pcmpestri128: { | |||
| 25992 | unsigned Opcode; | |||
| 25993 | if (IntNo == Intrinsic::x86_sse42_pcmpistri128) | |||
| 25994 | Opcode = X86ISD::PCMPISTR; | |||
| 25995 | else | |||
| 25996 | Opcode = X86ISD::PCMPESTR; | |||
| 25997 | ||||
| 25998 | SmallVector<SDValue, 5> NewOps(Op->op_begin()+1, Op->op_end()); | |||
| 25999 | SDVTList VTs = DAG.getVTList(MVT::i32, MVT::v16i8, MVT::i32); | |||
| 26000 | return DAG.getNode(Opcode, dl, VTs, NewOps); | |||
| 26001 | } | |||
| 26002 | ||||
| 26003 | case Intrinsic::x86_sse42_pcmpistrm128: | |||
| 26004 | case Intrinsic::x86_sse42_pcmpestrm128: { | |||
| 26005 | unsigned Opcode; | |||
| 26006 | if (IntNo == Intrinsic::x86_sse42_pcmpistrm128) | |||
| 26007 | Opcode = X86ISD::PCMPISTR; | |||
| 26008 | else | |||
| 26009 | Opcode = X86ISD::PCMPESTR; | |||
| 26010 | ||||
| 26011 | SmallVector<SDValue, 5> NewOps(Op->op_begin()+1, Op->op_end()); | |||
| 26012 | SDVTList VTs = DAG.getVTList(MVT::i32, MVT::v16i8, MVT::i32); | |||
| 26013 | return DAG.getNode(Opcode, dl, VTs, NewOps).getValue(1); | |||
| 26014 | } | |||
| 26015 | ||||
| 26016 | case Intrinsic::eh_sjlj_lsda: { | |||
| 26017 | MachineFunction &MF = DAG.getMachineFunction(); | |||
| 26018 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
| 26019 | MVT PtrVT = TLI.getPointerTy(DAG.getDataLayout()); | |||
| 26020 | auto &Context = MF.getMMI().getContext(); | |||
| 26021 | MCSymbol *S = Context.getOrCreateSymbol(Twine("GCC_except_table") + | |||
| 26022 | Twine(MF.getFunctionNumber())); | |||
| 26023 | return DAG.getNode(getGlobalWrapperKind(), dl, VT, | |||
| 26024 | DAG.getMCSymbol(S, PtrVT)); | |||
| 26025 | } | |||
| 26026 | ||||
| 26027 | case Intrinsic::x86_seh_lsda: { | |||
| 26028 | // Compute the symbol for the LSDA. We know it'll get emitted later. | |||
| 26029 | MachineFunction &MF = DAG.getMachineFunction(); | |||
| 26030 | SDValue Op1 = Op.getOperand(1); | |||
| 26031 | auto *Fn = cast<Function>(cast<GlobalAddressSDNode>(Op1)->getGlobal()); | |||
| 26032 | MCSymbol *LSDASym = MF.getMMI().getContext().getOrCreateLSDASymbol( | |||
| 26033 | GlobalValue::dropLLVMManglingEscape(Fn->getName())); | |||
| 26034 | ||||
| 26035 | // Generate a simple absolute symbol reference. This intrinsic is only | |||
| 26036 | // supported on 32-bit Windows, which isn't PIC. | |||
| 26037 | SDValue Result = DAG.getMCSymbol(LSDASym, VT); | |||
| 26038 | return DAG.getNode(X86ISD::Wrapper, dl, VT, Result); | |||
| 26039 | } | |||
| 26040 | ||||
| 26041 | case Intrinsic::eh_recoverfp: { | |||
| 26042 | SDValue FnOp = Op.getOperand(1); | |||
| 26043 | SDValue IncomingFPOp = Op.getOperand(2); | |||
| 26044 | GlobalAddressSDNode *GSD = dyn_cast<GlobalAddressSDNode>(FnOp); | |||
| 26045 | auto *Fn = dyn_cast_or_null<Function>(GSD ? GSD->getGlobal() : nullptr); | |||
| 26046 | if (!Fn) | |||
| 26047 | report_fatal_error( | |||
| 26048 | "llvm.eh.recoverfp must take a function as the first argument"); | |||
| 26049 | return recoverFramePointer(DAG, Fn, IncomingFPOp); | |||
| 26050 | } | |||
| 26051 | ||||
| 26052 | case Intrinsic::localaddress: { | |||
| 26053 | // Returns one of the stack, base, or frame pointer registers, depending on | |||
| 26054 | // which is used to reference local variables. | |||
| 26055 | MachineFunction &MF = DAG.getMachineFunction(); | |||
| 26056 | const X86RegisterInfo *RegInfo = Subtarget.getRegisterInfo(); | |||
| 26057 | unsigned Reg; | |||
| 26058 | if (RegInfo->hasBasePointer(MF)) | |||
| 26059 | Reg = RegInfo->getBaseRegister(); | |||
| 26060 | else { // Handles the SP or FP case. | |||
| 26061 | bool CantUseFP = RegInfo->hasStackRealignment(MF); | |||
| 26062 | if (CantUseFP) | |||
| 26063 | Reg = RegInfo->getPtrSizedStackRegister(MF); | |||
| 26064 | else | |||
| 26065 | Reg = RegInfo->getPtrSizedFrameRegister(MF); | |||
| 26066 | } | |||
| 26067 | return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT); | |||
| 26068 | } | |||
| 26069 | case Intrinsic::swift_async_context_addr: { | |||
| 26070 | auto &MF = DAG.getMachineFunction(); | |||
| 26071 | auto X86FI = MF.getInfo<X86MachineFunctionInfo>(); | |||
| 26072 | if (Subtarget.is64Bit()) { | |||
| 26073 | MF.getFrameInfo().setFrameAddressIsTaken(true); | |||
| 26074 | X86FI->setHasSwiftAsyncContext(true); | |||
| 26075 | return SDValue( | |||
| 26076 | DAG.getMachineNode( | |||
| 26077 | X86::SUB64ri8, dl, MVT::i64, | |||
| 26078 | DAG.getCopyFromReg(DAG.getEntryNode(), dl, X86::RBP, MVT::i64), | |||
| 26079 | DAG.getTargetConstant(8, dl, MVT::i32)), | |||
| 26080 | 0); | |||
| 26081 | } else { | |||
| 26082 | // 32-bit so no special extended frame, create or reuse an existing stack | |||
| 26083 | // slot. | |||
| 26084 | if (!X86FI->getSwiftAsyncContextFrameIdx()) | |||
| 26085 | X86FI->setSwiftAsyncContextFrameIdx( | |||
| 26086 | MF.getFrameInfo().CreateStackObject(4, Align(4), false)); | |||
| 26087 | return DAG.getFrameIndex(*X86FI->getSwiftAsyncContextFrameIdx(), MVT::i32); | |||
| 26088 | } | |||
| 26089 | } | |||
| 26090 | case Intrinsic::x86_avx512_vp2intersect_q_512: | |||
| 26091 | case Intrinsic::x86_avx512_vp2intersect_q_256: | |||
| 26092 | case Intrinsic::x86_avx512_vp2intersect_q_128: | |||
| 26093 | case Intrinsic::x86_avx512_vp2intersect_d_512: | |||
| 26094 | case Intrinsic::x86_avx512_vp2intersect_d_256: | |||
| 26095 | case Intrinsic::x86_avx512_vp2intersect_d_128: { | |||
| 26096 | MVT MaskVT = Op.getSimpleValueType(); | |||
| 26097 | ||||
| 26098 | SDVTList VTs = DAG.getVTList(MVT::Untyped, MVT::Other); | |||
| 26099 | SDLoc DL(Op); | |||
| 26100 | ||||
| 26101 | SDValue Operation = | |||
| 26102 | DAG.getNode(X86ISD::VP2INTERSECT, DL, VTs, | |||
| 26103 | Op->getOperand(1), Op->getOperand(2)); | |||
| 26104 | ||||
| 26105 | SDValue Result0 = DAG.getTargetExtractSubreg(X86::sub_mask_0, DL, | |||
| 26106 | MaskVT, Operation); | |||
| 26107 | SDValue Result1 = DAG.getTargetExtractSubreg(X86::sub_mask_1, DL, | |||
| 26108 | MaskVT, Operation); | |||
| 26109 | return DAG.getMergeValues({Result0, Result1}, DL); | |||
| 26110 | } | |||
| 26111 | case Intrinsic::x86_mmx_pslli_w: | |||
| 26112 | case Intrinsic::x86_mmx_pslli_d: | |||
| 26113 | case Intrinsic::x86_mmx_pslli_q: | |||
| 26114 | case Intrinsic::x86_mmx_psrli_w: | |||
| 26115 | case Intrinsic::x86_mmx_psrli_d: | |||
| 26116 | case Intrinsic::x86_mmx_psrli_q: | |||
| 26117 | case Intrinsic::x86_mmx_psrai_w: | |||
| 26118 | case Intrinsic::x86_mmx_psrai_d: { | |||
| 26119 | SDLoc DL(Op); | |||
| 26120 | SDValue ShAmt = Op.getOperand(2); | |||
| 26121 | // If the argument is a constant, convert it to a target constant. | |||
| 26122 | if (auto *C = dyn_cast<ConstantSDNode>(ShAmt)) { | |||
| 26123 | // Clamp out of bounds shift amounts since they will otherwise be masked | |||
| 26124 | // to 8-bits which may make it no longer out of bounds. | |||
| 26125 | unsigned ShiftAmount = C->getAPIntValue().getLimitedValue(255); | |||
| 26126 | if (ShiftAmount == 0) | |||
| 26127 | return Op.getOperand(1); | |||
| 26128 | ||||
| 26129 | return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, Op.getValueType(), | |||
| 26130 | Op.getOperand(0), Op.getOperand(1), | |||
| 26131 | DAG.getTargetConstant(ShiftAmount, DL, MVT::i32)); | |||
| 26132 | } | |||
| 26133 | ||||
| 26134 | unsigned NewIntrinsic; | |||
| 26135 | switch (IntNo) { | |||
| 26136 | default: llvm_unreachable("Impossible intrinsic")__builtin_unreachable(); // Can't reach here. | |||
| 26137 | case Intrinsic::x86_mmx_pslli_w: | |||
| 26138 | NewIntrinsic = Intrinsic::x86_mmx_psll_w; | |||
| 26139 | break; | |||
| 26140 | case Intrinsic::x86_mmx_pslli_d: | |||
| 26141 | NewIntrinsic = Intrinsic::x86_mmx_psll_d; | |||
| 26142 | break; | |||
| 26143 | case Intrinsic::x86_mmx_pslli_q: | |||
| 26144 | NewIntrinsic = Intrinsic::x86_mmx_psll_q; | |||
| 26145 | break; | |||
| 26146 | case Intrinsic::x86_mmx_psrli_w: | |||
| 26147 | NewIntrinsic = Intrinsic::x86_mmx_psrl_w; | |||
| 26148 | break; | |||
| 26149 | case Intrinsic::x86_mmx_psrli_d: | |||
| 26150 | NewIntrinsic = Intrinsic::x86_mmx_psrl_d; | |||
| 26151 | break; | |||
| 26152 | case Intrinsic::x86_mmx_psrli_q: | |||
| 26153 | NewIntrinsic = Intrinsic::x86_mmx_psrl_q; | |||
| 26154 | break; | |||
| 26155 | case Intrinsic::x86_mmx_psrai_w: | |||
| 26156 | NewIntrinsic = Intrinsic::x86_mmx_psra_w; | |||
| 26157 | break; | |||
| 26158 | case Intrinsic::x86_mmx_psrai_d: | |||
| 26159 | NewIntrinsic = Intrinsic::x86_mmx_psra_d; | |||
| 26160 | break; | |||
| 26161 | } | |||
| 26162 | ||||
| 26163 | // The vector shift intrinsics with scalars uses 32b shift amounts but | |||
| 26164 | // the sse2/mmx shift instructions reads 64 bits. Copy the 32 bits to an | |||
| 26165 | // MMX register. | |||
| 26166 | ShAmt = DAG.getNode(X86ISD::MMX_MOVW2D, DL, MVT::x86mmx, ShAmt); | |||
| 26167 | return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, Op.getValueType(), | |||
| 26168 | DAG.getTargetConstant(NewIntrinsic, DL, | |||
| 26169 | getPointerTy(DAG.getDataLayout())), | |||
| 26170 | Op.getOperand(1), ShAmt); | |||
| 26171 | } | |||
| 26172 | } | |||
| 26173 | } | |||
| 26174 | ||||
| 26175 | static SDValue getAVX2GatherNode(unsigned Opc, SDValue Op, SelectionDAG &DAG, | |||
| 26176 | SDValue Src, SDValue Mask, SDValue Base, | |||
| 26177 | SDValue Index, SDValue ScaleOp, SDValue Chain, | |||
| 26178 | const X86Subtarget &Subtarget) { | |||
| 26179 | SDLoc dl(Op); | |||
| 26180 | auto *C = dyn_cast<ConstantSDNode>(ScaleOp); | |||
| 26181 | // Scale must be constant. | |||
| 26182 | if (!C) | |||
| 26183 | return SDValue(); | |||
| 26184 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
| 26185 | SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), dl, | |||
| 26186 | TLI.getPointerTy(DAG.getDataLayout())); | |||
| 26187 | EVT MaskVT = Mask.getValueType().changeVectorElementTypeToInteger(); | |||
| 26188 | SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Other); | |||
| 26189 | // If source is undef or we know it won't be used, use a zero vector | |||
| 26190 | // to break register dependency. | |||
| 26191 | // TODO: use undef instead and let BreakFalseDeps deal with it? | |||
| 26192 | if (Src.isUndef() || ISD::isBuildVectorAllOnes(Mask.getNode())) | |||
| 26193 | Src = getZeroVector(Op.getSimpleValueType(), Subtarget, DAG, dl); | |||
| 26194 | ||||
| 26195 | // Cast mask to an integer type. | |||
| 26196 | Mask = DAG.getBitcast(MaskVT, Mask); | |||
| 26197 | ||||
| 26198 | MemIntrinsicSDNode *MemIntr = cast<MemIntrinsicSDNode>(Op); | |||
| 26199 | ||||
| 26200 | SDValue Ops[] = {Chain, Src, Mask, Base, Index, Scale }; | |||
| 26201 | SDValue Res = | |||
| 26202 | DAG.getMemIntrinsicNode(X86ISD::MGATHER, dl, VTs, Ops, | |||
| 26203 | MemIntr->getMemoryVT(), MemIntr->getMemOperand()); | |||
| 26204 | return DAG.getMergeValues({Res, Res.getValue(1)}, dl); | |||
| 26205 | } | |||
| 26206 | ||||
| 26207 | static SDValue getGatherNode(SDValue Op, SelectionDAG &DAG, | |||
| 26208 | SDValue Src, SDValue Mask, SDValue Base, | |||
| 26209 | SDValue Index, SDValue ScaleOp, SDValue Chain, | |||
| 26210 | const X86Subtarget &Subtarget) { | |||
| 26211 | MVT VT = Op.getSimpleValueType(); | |||
| 26212 | SDLoc dl(Op); | |||
| 26213 | auto *C = dyn_cast<ConstantSDNode>(ScaleOp); | |||
| 26214 | // Scale must be constant. | |||
| 26215 | if (!C) | |||
| 26216 | return SDValue(); | |||
| 26217 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
| 26218 | SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), dl, | |||
| 26219 | TLI.getPointerTy(DAG.getDataLayout())); | |||
| 26220 | unsigned MinElts = std::min(Index.getSimpleValueType().getVectorNumElements(), | |||
| 26221 | VT.getVectorNumElements()); | |||
| 26222 | MVT MaskVT = MVT::getVectorVT(MVT::i1, MinElts); | |||
| 26223 | ||||
| 26224 | // We support two versions of the gather intrinsics. One with scalar mask and | |||
| 26225 | // one with vXi1 mask. Convert scalar to vXi1 if necessary. | |||
| 26226 | if (Mask.getValueType() != MaskVT) | |||
| 26227 | Mask = getMaskNode(Mask, MaskVT, Subtarget, DAG, dl); | |||
| 26228 | ||||
| 26229 | SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Other); | |||
| 26230 | // If source is undef or we know it won't be used, use a zero vector | |||
| 26231 | // to break register dependency. | |||
| 26232 | // TODO: use undef instead and let BreakFalseDeps deal with it? | |||
| 26233 | if (Src.isUndef() || ISD::isBuildVectorAllOnes(Mask.getNode())) | |||
| 26234 | Src = getZeroVector(Op.getSimpleValueType(), Subtarget, DAG, dl); | |||
| 26235 | ||||
| 26236 | MemIntrinsicSDNode *MemIntr = cast<MemIntrinsicSDNode>(Op); | |||
| 26237 | ||||
| 26238 | SDValue Ops[] = {Chain, Src, Mask, Base, Index, Scale }; | |||
| 26239 | SDValue Res = | |||
| 26240 | DAG.getMemIntrinsicNode(X86ISD::MGATHER, dl, VTs, Ops, | |||
| 26241 | MemIntr->getMemoryVT(), MemIntr->getMemOperand()); | |||
| 26242 | return DAG.getMergeValues({Res, Res.getValue(1)}, dl); | |||
| 26243 | } | |||
| 26244 | ||||
| 26245 | static SDValue getScatterNode(unsigned Opc, SDValue Op, SelectionDAG &DAG, | |||
| 26246 | SDValue Src, SDValue Mask, SDValue Base, | |||
| 26247 | SDValue Index, SDValue ScaleOp, SDValue Chain, | |||
| 26248 | const X86Subtarget &Subtarget) { | |||
| 26249 | SDLoc dl(Op); | |||
| 26250 | auto *C = dyn_cast<ConstantSDNode>(ScaleOp); | |||
| 26251 | // Scale must be constant. | |||
| 26252 | if (!C) | |||
| 26253 | return SDValue(); | |||
| 26254 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
| 26255 | SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), dl, | |||
| 26256 | TLI.getPointerTy(DAG.getDataLayout())); | |||
| 26257 | unsigned MinElts = std::min(Index.getSimpleValueType().getVectorNumElements(), | |||
| 26258 | Src.getSimpleValueType().getVectorNumElements()); | |||
| 26259 | MVT MaskVT = MVT::getVectorVT(MVT::i1, MinElts); | |||
| 26260 | ||||
| 26261 | // We support two versions of the scatter intrinsics. One with scalar mask and | |||
| 26262 | // one with vXi1 mask. Convert scalar to vXi1 if necessary. | |||
| 26263 | if (Mask.getValueType() != MaskVT) | |||
| 26264 | Mask = getMaskNode(Mask, MaskVT, Subtarget, DAG, dl); | |||
| 26265 | ||||
| 26266 | MemIntrinsicSDNode *MemIntr = cast<MemIntrinsicSDNode>(Op); | |||
| 26267 | ||||
| 26268 | SDVTList VTs = DAG.getVTList(MVT::Other); | |||
| 26269 | SDValue Ops[] = {Chain, Src, Mask, Base, Index, Scale}; | |||
| 26270 | SDValue Res = | |||
| 26271 | DAG.getMemIntrinsicNode(X86ISD::MSCATTER, dl, VTs, Ops, | |||
| 26272 | MemIntr->getMemoryVT(), MemIntr->getMemOperand()); | |||
| 26273 | return Res; | |||
| 26274 | } | |||
| 26275 | ||||
| 26276 | static SDValue getPrefetchNode(unsigned Opc, SDValue Op, SelectionDAG &DAG, | |||
| 26277 | SDValue Mask, SDValue Base, SDValue Index, | |||
| 26278 | SDValue ScaleOp, SDValue Chain, | |||
| 26279 | const X86Subtarget &Subtarget) { | |||
| 26280 | SDLoc dl(Op); | |||
| 26281 | auto *C = dyn_cast<ConstantSDNode>(ScaleOp); | |||
| 26282 | // Scale must be constant. | |||
| 26283 | if (!C) | |||
| 26284 | return SDValue(); | |||
| 26285 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
| 26286 | SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), dl, | |||
| 26287 | TLI.getPointerTy(DAG.getDataLayout())); | |||
| 26288 | SDValue Disp = DAG.getTargetConstant(0, dl, MVT::i32); | |||
| 26289 | SDValue Segment = DAG.getRegister(0, MVT::i32); | |||
| 26290 | MVT MaskVT = | |||
| 26291 | MVT::getVectorVT(MVT::i1, Index.getSimpleValueType().getVectorNumElements()); | |||
| 26292 | SDValue VMask = getMaskNode(Mask, MaskVT, Subtarget, DAG, dl); | |||
| 26293 | SDValue Ops[] = {VMask, Base, Scale, Index, Disp, Segment, Chain}; | |||
| 26294 | SDNode *Res = DAG.getMachineNode(Opc, dl, MVT::Other, Ops); | |||
| 26295 | return SDValue(Res, 0); | |||
| 26296 | } | |||
| 26297 | ||||
| 26298 | /// Handles the lowering of builtin intrinsics with chain that return their | |||
| 26299 | /// value into registers EDX:EAX. | |||
| 26300 | /// If operand ScrReg is a valid register identifier, then operand 2 of N is | |||
| 26301 | /// copied to SrcReg. The assumption is that SrcReg is an implicit input to | |||
| 26302 | /// TargetOpcode. | |||
| 26303 | /// Returns a Glue value which can be used to add extra copy-from-reg if the | |||
| 26304 | /// expanded intrinsics implicitly defines extra registers (i.e. not just | |||
| 26305 | /// EDX:EAX). | |||
| 26306 | static SDValue expandIntrinsicWChainHelper(SDNode *N, const SDLoc &DL, | |||
| 26307 | SelectionDAG &DAG, | |||
| 26308 | unsigned TargetOpcode, | |||
| 26309 | unsigned SrcReg, | |||
| 26310 | const X86Subtarget &Subtarget, | |||
| 26311 | SmallVectorImpl<SDValue> &Results) { | |||
| 26312 | SDValue Chain = N->getOperand(0); | |||
| 26313 | SDValue Glue; | |||
| 26314 | ||||
| 26315 | if (SrcReg) { | |||
| 26316 | assert(N->getNumOperands() == 3 && "Unexpected number of operands!")((void)0); | |||
| 26317 | Chain = DAG.getCopyToReg(Chain, DL, SrcReg, N->getOperand(2), Glue); | |||
| 26318 | Glue = Chain.getValue(1); | |||
| 26319 | } | |||
| 26320 | ||||
| 26321 | SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue); | |||
| 26322 | SDValue N1Ops[] = {Chain, Glue}; | |||
| 26323 | SDNode *N1 = DAG.getMachineNode( | |||
| 26324 | TargetOpcode, DL, Tys, ArrayRef<SDValue>(N1Ops, Glue.getNode() ? 2 : 1)); | |||
| 26325 | Chain = SDValue(N1, 0); | |||
| 26326 | ||||
| 26327 | // Reads the content of XCR and returns it in registers EDX:EAX. | |||
| 26328 | SDValue LO, HI; | |||
| 26329 | if (Subtarget.is64Bit()) { | |||
| 26330 | LO = DAG.getCopyFromReg(Chain, DL, X86::RAX, MVT::i64, SDValue(N1, 1)); | |||
| 26331 | HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::RDX, MVT::i64, | |||
| 26332 | LO.getValue(2)); | |||
| 26333 | } else { | |||
| 26334 | LO = DAG.getCopyFromReg(Chain, DL, X86::EAX, MVT::i32, SDValue(N1, 1)); | |||
| 26335 | HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::EDX, MVT::i32, | |||
| 26336 | LO.getValue(2)); | |||
| 26337 | } | |||
| 26338 | Chain = HI.getValue(1); | |||
| 26339 | Glue = HI.getValue(2); | |||
| 26340 | ||||
| 26341 | if (Subtarget.is64Bit()) { | |||
| 26342 | // Merge the two 32-bit values into a 64-bit one. | |||
| 26343 | SDValue Tmp = DAG.getNode(ISD::SHL, DL, MVT::i64, HI, | |||
| 26344 | DAG.getConstant(32, DL, MVT::i8)); | |||
| 26345 | Results.push_back(DAG.getNode(ISD::OR, DL, MVT::i64, LO, Tmp)); | |||
| 26346 | Results.push_back(Chain); | |||
| 26347 | return Glue; | |||
| 26348 | } | |||
| 26349 | ||||
| 26350 | // Use a buildpair to merge the two 32-bit values into a 64-bit one. | |||
| 26351 | SDValue Ops[] = { LO, HI }; | |||
| 26352 | SDValue Pair = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops); | |||
| 26353 | Results.push_back(Pair); | |||
| 26354 | Results.push_back(Chain); | |||
| 26355 | return Glue; | |||
| 26356 | } | |||
| 26357 | ||||
| 26358 | /// Handles the lowering of builtin intrinsics that read the time stamp counter | |||
| 26359 | /// (x86_rdtsc and x86_rdtscp). This function is also used to custom lower | |||
| 26360 | /// READCYCLECOUNTER nodes. | |||
| 26361 | static void getReadTimeStampCounter(SDNode *N, const SDLoc &DL, unsigned Opcode, | |||
| 26362 | SelectionDAG &DAG, | |||
| 26363 | const X86Subtarget &Subtarget, | |||
| 26364 | SmallVectorImpl<SDValue> &Results) { | |||
| 26365 | // The processor's time-stamp counter (a 64-bit MSR) is stored into the | |||
| 26366 | // EDX:EAX registers. EDX is loaded with the high-order 32 bits of the MSR | |||
| 26367 | // and the EAX register is loaded with the low-order 32 bits. | |||
| 26368 | SDValue Glue = expandIntrinsicWChainHelper(N, DL, DAG, Opcode, | |||
| 26369 | /* NoRegister */0, Subtarget, | |||
| 26370 | Results); | |||
| 26371 | if (Opcode != X86::RDTSCP) | |||
| 26372 | return; | |||
| 26373 | ||||
| 26374 | SDValue Chain = Results[1]; | |||
| 26375 | // Instruction RDTSCP loads the IA32:TSC_AUX_MSR (address C000_0103H) into | |||
| 26376 | // the ECX register. Add 'ecx' explicitly to the chain. | |||
| 26377 | SDValue ecx = DAG.getCopyFromReg(Chain, DL, X86::ECX, MVT::i32, Glue); | |||
| 26378 | Results[1] = ecx; | |||
| 26379 | Results.push_back(ecx.getValue(1)); | |||
| 26380 | } | |||
| 26381 | ||||
| 26382 | static SDValue LowerREADCYCLECOUNTER(SDValue Op, const X86Subtarget &Subtarget, | |||
| 26383 | SelectionDAG &DAG) { | |||
| 26384 | SmallVector<SDValue, 3> Results; | |||
| 26385 | SDLoc DL(Op); | |||
| 26386 | getReadTimeStampCounter(Op.getNode(), DL, X86::RDTSC, DAG, Subtarget, | |||
| 26387 | Results); | |||
| 26388 | return DAG.getMergeValues(Results, DL); | |||
| 26389 | } | |||
| 26390 | ||||
| 26391 | static SDValue MarkEHRegistrationNode(SDValue Op, SelectionDAG &DAG) { | |||
| 26392 | MachineFunction &MF = DAG.getMachineFunction(); | |||
| 26393 | SDValue Chain = Op.getOperand(0); | |||
| 26394 | SDValue RegNode = Op.getOperand(2); | |||
| 26395 | WinEHFuncInfo *EHInfo = MF.getWinEHFuncInfo(); | |||
| 26396 | if (!EHInfo) | |||
| 26397 | report_fatal_error("EH registrations only live in functions using WinEH"); | |||
| 26398 | ||||
| 26399 | // Cast the operand to an alloca, and remember the frame index. | |||
| 26400 | auto *FINode = dyn_cast<FrameIndexSDNode>(RegNode); | |||
| 26401 | if (!FINode) | |||
| 26402 | report_fatal_error("llvm.x86.seh.ehregnode expects a static alloca"); | |||
| 26403 | EHInfo->EHRegNodeFrameIndex = FINode->getIndex(); | |||
| 26404 | ||||
| 26405 | // Return the chain operand without making any DAG nodes. | |||
| 26406 | return Chain; | |||
| 26407 | } | |||
| 26408 | ||||
| 26409 | static SDValue MarkEHGuard(SDValue Op, SelectionDAG &DAG) { | |||
| 26410 | MachineFunction &MF = DAG.getMachineFunction(); | |||
| 26411 | SDValue Chain = Op.getOperand(0); | |||
| 26412 | SDValue EHGuard = Op.getOperand(2); | |||
| 26413 | WinEHFuncInfo *EHInfo = MF.getWinEHFuncInfo(); | |||
| 26414 | if (!EHInfo) | |||
| 26415 | report_fatal_error("EHGuard only live in functions using WinEH"); | |||
| 26416 | ||||
| 26417 | // Cast the operand to an alloca, and remember the frame index. | |||
| 26418 | auto *FINode = dyn_cast<FrameIndexSDNode>(EHGuard); | |||
| 26419 | if (!FINode) | |||
| 26420 | report_fatal_error("llvm.x86.seh.ehguard expects a static alloca"); | |||
| 26421 | EHInfo->EHGuardFrameIndex = FINode->getIndex(); | |||
| 26422 | ||||
| 26423 | // Return the chain operand without making any DAG nodes. | |||
| 26424 | return Chain; | |||
| 26425 | } | |||
| 26426 | ||||
| 26427 | /// Emit Truncating Store with signed or unsigned saturation. | |||
| 26428 | static SDValue | |||
| 26429 | EmitTruncSStore(bool SignedSat, SDValue Chain, const SDLoc &Dl, SDValue Val, | |||
| 26430 | SDValue Ptr, EVT MemVT, MachineMemOperand *MMO, | |||
| 26431 | SelectionDAG &DAG) { | |||
| 26432 | SDVTList VTs = DAG.getVTList(MVT::Other); | |||
| 26433 | SDValue Undef = DAG.getUNDEF(Ptr.getValueType()); | |||
| 26434 | SDValue Ops[] = { Chain, Val, Ptr, Undef }; | |||
| 26435 | unsigned Opc = SignedSat ? X86ISD::VTRUNCSTORES : X86ISD::VTRUNCSTOREUS; | |||
| 26436 | return DAG.getMemIntrinsicNode(Opc, Dl, VTs, Ops, MemVT, MMO); | |||
| 26437 | } | |||
| 26438 | ||||
| 26439 | /// Emit Masked Truncating Store with signed or unsigned saturation. | |||
| 26440 | static SDValue | |||
| 26441 | EmitMaskedTruncSStore(bool SignedSat, SDValue Chain, const SDLoc &Dl, | |||
| 26442 | SDValue Val, SDValue Ptr, SDValue Mask, EVT MemVT, | |||
| 26443 | MachineMemOperand *MMO, SelectionDAG &DAG) { | |||
| 26444 | SDVTList VTs = DAG.getVTList(MVT::Other); | |||
| 26445 | SDValue Ops[] = { Chain, Val, Ptr, Mask }; | |||
| 26446 | unsigned Opc = SignedSat ? X86ISD::VMTRUNCSTORES : X86ISD::VMTRUNCSTOREUS; | |||
| 26447 | return DAG.getMemIntrinsicNode(Opc, Dl, VTs, Ops, MemVT, MMO); | |||
| 26448 | } | |||
| 26449 | ||||
| 26450 | static SDValue LowerINTRINSIC_W_CHAIN(SDValue Op, const X86Subtarget &Subtarget, | |||
| 26451 | SelectionDAG &DAG) { | |||
| 26452 | unsigned IntNo = Op.getConstantOperandVal(1); | |||
| 26453 | const IntrinsicData *IntrData = getIntrinsicWithChain(IntNo); | |||
| 26454 | if (!IntrData) { | |||
| 26455 | switch (IntNo) { | |||
| 26456 | case llvm::Intrinsic::x86_seh_ehregnode: | |||
| 26457 | return MarkEHRegistrationNode(Op, DAG); | |||
| 26458 | case llvm::Intrinsic::x86_seh_ehguard: | |||
| 26459 | return MarkEHGuard(Op, DAG); | |||
| 26460 | case llvm::Intrinsic::x86_rdpkru: { | |||
| 26461 | SDLoc dl(Op); | |||
| 26462 | SDVTList VTs = DAG.getVTList(MVT::i32, MVT::Other); | |||
| 26463 | // Create a RDPKRU node and pass 0 to the ECX parameter. | |||
| 26464 | return DAG.getNode(X86ISD::RDPKRU, dl, VTs, Op.getOperand(0), | |||
| 26465 | DAG.getConstant(0, dl, MVT::i32)); | |||
| 26466 | } | |||
| 26467 | case llvm::Intrinsic::x86_wrpkru: { | |||
| 26468 | SDLoc dl(Op); | |||
| 26469 | // Create a WRPKRU node, pass the input to the EAX parameter, and pass 0 | |||
| 26470 | // to the EDX and ECX parameters. | |||
| 26471 | return DAG.getNode(X86ISD::WRPKRU, dl, MVT::Other, | |||
| 26472 | Op.getOperand(0), Op.getOperand(2), | |||
| 26473 | DAG.getConstant(0, dl, MVT::i32), | |||
| 26474 | DAG.getConstant(0, dl, MVT::i32)); | |||
| 26475 | } | |||
| 26476 | case llvm::Intrinsic::x86_flags_read_u32: | |||
| 26477 | case llvm::Intrinsic::x86_flags_read_u64: | |||
| 26478 | case llvm::Intrinsic::x86_flags_write_u32: | |||
| 26479 | case llvm::Intrinsic::x86_flags_write_u64: { | |||
| 26480 | // We need a frame pointer because this will get lowered to a PUSH/POP | |||
| 26481 | // sequence. | |||
| 26482 | MachineFrameInfo &MFI = DAG.getMachineFunction().getFrameInfo(); | |||
| 26483 | MFI.setHasCopyImplyingStackAdjustment(true); | |||
| 26484 | // Don't do anything here, we will expand these intrinsics out later | |||
| 26485 | // during FinalizeISel in EmitInstrWithCustomInserter. | |||
| 26486 | return Op; | |||
| 26487 | } | |||
| 26488 | case Intrinsic::x86_lwpins32: | |||
| 26489 | case Intrinsic::x86_lwpins64: | |||
| 26490 | case Intrinsic::x86_umwait: | |||
| 26491 | case Intrinsic::x86_tpause: { | |||
| 26492 | SDLoc dl(Op); | |||
| 26493 | SDValue Chain = Op->getOperand(0); | |||
| 26494 | SDVTList VTs = DAG.getVTList(MVT::i32, MVT::Other); | |||
| 26495 | unsigned Opcode; | |||
| 26496 | ||||
| 26497 | switch (IntNo) { | |||
| 26498 | default: llvm_unreachable("Impossible intrinsic")__builtin_unreachable(); | |||
| 26499 | case Intrinsic::x86_umwait: | |||
| 26500 | Opcode = X86ISD::UMWAIT; | |||
| 26501 | break; | |||
| 26502 | case Intrinsic::x86_tpause: | |||
| 26503 | Opcode = X86ISD::TPAUSE; | |||
| 26504 | break; | |||
| 26505 | case Intrinsic::x86_lwpins32: | |||
| 26506 | case Intrinsic::x86_lwpins64: | |||
| 26507 | Opcode = X86ISD::LWPINS; | |||
| 26508 | break; | |||
| 26509 | } | |||
| 26510 | ||||
| 26511 | SDValue Operation = | |||
| 26512 | DAG.getNode(Opcode, dl, VTs, Chain, Op->getOperand(2), | |||
| 26513 | Op->getOperand(3), Op->getOperand(4)); | |||
| 26514 | SDValue SetCC = getSETCC(X86::COND_B, Operation.getValue(0), dl, DAG); | |||
| 26515 | return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(), SetCC, | |||
| 26516 | Operation.getValue(1)); | |||
| 26517 | } | |||
| 26518 | case Intrinsic::x86_enqcmd: | |||
| 26519 | case Intrinsic::x86_enqcmds: { | |||
| 26520 | SDLoc dl(Op); | |||
| 26521 | SDValue Chain = Op.getOperand(0); | |||
| 26522 | SDVTList VTs = DAG.getVTList(MVT::i32, MVT::Other); | |||
| 26523 | unsigned Opcode; | |||
| 26524 | switch (IntNo) { | |||
| 26525 | default: llvm_unreachable("Impossible intrinsic!")__builtin_unreachable(); | |||
| 26526 | case Intrinsic::x86_enqcmd: | |||
| 26527 | Opcode = X86ISD::ENQCMD; | |||
| 26528 | break; | |||
| 26529 | case Intrinsic::x86_enqcmds: | |||
| 26530 | Opcode = X86ISD::ENQCMDS; | |||
| 26531 | break; | |||
| 26532 | } | |||
| 26533 | SDValue Operation = DAG.getNode(Opcode, dl, VTs, Chain, Op.getOperand(2), | |||
| 26534 | Op.getOperand(3)); | |||
| 26535 | SDValue SetCC = getSETCC(X86::COND_E, Operation.getValue(0), dl, DAG); | |||
| 26536 | return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(), SetCC, | |||
| 26537 | Operation.getValue(1)); | |||
| 26538 | } | |||
| 26539 | case Intrinsic::x86_aesenc128kl: | |||
| 26540 | case Intrinsic::x86_aesdec128kl: | |||
| 26541 | case Intrinsic::x86_aesenc256kl: | |||
| 26542 | case Intrinsic::x86_aesdec256kl: { | |||
| 26543 | SDLoc DL(Op); | |||
| 26544 | SDVTList VTs = DAG.getVTList(MVT::v2i64, MVT::i32, MVT::Other); | |||
| 26545 | SDValue Chain = Op.getOperand(0); | |||
| 26546 | unsigned Opcode; | |||
| 26547 | ||||
| 26548 | switch (IntNo) { | |||
| 26549 | default: llvm_unreachable("Impossible intrinsic")__builtin_unreachable(); | |||
| 26550 | case Intrinsic::x86_aesenc128kl: | |||
| 26551 | Opcode = X86ISD::AESENC128KL; | |||
| 26552 | break; | |||
| 26553 | case Intrinsic::x86_aesdec128kl: | |||
| 26554 | Opcode = X86ISD::AESDEC128KL; | |||
| 26555 | break; | |||
| 26556 | case Intrinsic::x86_aesenc256kl: | |||
| 26557 | Opcode = X86ISD::AESENC256KL; | |||
| 26558 | break; | |||
| 26559 | case Intrinsic::x86_aesdec256kl: | |||
| 26560 | Opcode = X86ISD::AESDEC256KL; | |||
| 26561 | break; | |||
| 26562 | } | |||
| 26563 | ||||
| 26564 | MemIntrinsicSDNode *MemIntr = cast<MemIntrinsicSDNode>(Op); | |||
| 26565 | MachineMemOperand *MMO = MemIntr->getMemOperand(); | |||
| 26566 | EVT MemVT = MemIntr->getMemoryVT(); | |||
| 26567 | SDValue Operation = DAG.getMemIntrinsicNode( | |||
| 26568 | Opcode, DL, VTs, {Chain, Op.getOperand(2), Op.getOperand(3)}, MemVT, | |||
| 26569 | MMO); | |||
| 26570 | SDValue ZF = getSETCC(X86::COND_E, Operation.getValue(1), DL, DAG); | |||
| 26571 | ||||
| 26572 | return DAG.getNode(ISD::MERGE_VALUES, DL, Op->getVTList(), | |||
| 26573 | {ZF, Operation.getValue(0), Operation.getValue(2)}); | |||
| 26574 | } | |||
| 26575 | case Intrinsic::x86_aesencwide128kl: | |||
| 26576 | case Intrinsic::x86_aesdecwide128kl: | |||
| 26577 | case Intrinsic::x86_aesencwide256kl: | |||
| 26578 | case Intrinsic::x86_aesdecwide256kl: { | |||
| 26579 | SDLoc DL(Op); | |||
| 26580 | SDVTList VTs = DAG.getVTList( | |||
| 26581 | {MVT::i32, MVT::v2i64, MVT::v2i64, MVT::v2i64, MVT::v2i64, MVT::v2i64, | |||
| 26582 | MVT::v2i64, MVT::v2i64, MVT::v2i64, MVT::Other}); | |||
| 26583 | SDValue Chain = Op.getOperand(0); | |||
| 26584 | unsigned Opcode; | |||
| 26585 | ||||
| 26586 | switch (IntNo) { | |||
| 26587 | default: llvm_unreachable("Impossible intrinsic")__builtin_unreachable(); | |||
| 26588 | case Intrinsic::x86_aesencwide128kl: | |||
| 26589 | Opcode = X86ISD::AESENCWIDE128KL; | |||
| 26590 | break; | |||
| 26591 | case Intrinsic::x86_aesdecwide128kl: | |||
| 26592 | Opcode = X86ISD::AESDECWIDE128KL; | |||
| 26593 | break; | |||
| 26594 | case Intrinsic::x86_aesencwide256kl: | |||
| 26595 | Opcode = X86ISD::AESENCWIDE256KL; | |||
| 26596 | break; | |||
| 26597 | case Intrinsic::x86_aesdecwide256kl: | |||
| 26598 | Opcode = X86ISD::AESDECWIDE256KL; | |||
| 26599 | break; | |||
| 26600 | } | |||
| 26601 | ||||
| 26602 | MemIntrinsicSDNode *MemIntr = cast<MemIntrinsicSDNode>(Op); | |||
| 26603 | MachineMemOperand *MMO = MemIntr->getMemOperand(); | |||
| 26604 | EVT MemVT = MemIntr->getMemoryVT(); | |||
| 26605 | SDValue Operation = DAG.getMemIntrinsicNode( | |||
| 26606 | Opcode, DL, VTs, | |||
| 26607 | {Chain, Op.getOperand(2), Op.getOperand(3), Op.getOperand(4), | |||
| 26608 | Op.getOperand(5), Op.getOperand(6), Op.getOperand(7), | |||
| 26609 | Op.getOperand(8), Op.getOperand(9), Op.getOperand(10)}, | |||
| 26610 | MemVT, MMO); | |||
| 26611 | SDValue ZF = getSETCC(X86::COND_E, Operation.getValue(0), DL, DAG); | |||
| 26612 | ||||
| 26613 | return DAG.getNode(ISD::MERGE_VALUES, DL, Op->getVTList(), | |||
| 26614 | {ZF, Operation.getValue(1), Operation.getValue(2), | |||
| 26615 | Operation.getValue(3), Operation.getValue(4), | |||
| 26616 | Operation.getValue(5), Operation.getValue(6), | |||
| 26617 | Operation.getValue(7), Operation.getValue(8), | |||
| 26618 | Operation.getValue(9)}); | |||
| 26619 | } | |||
| 26620 | case Intrinsic::x86_testui: { | |||
| 26621 | SDLoc dl(Op); | |||
| 26622 | SDValue Chain = Op.getOperand(0); | |||
| 26623 | SDVTList VTs = DAG.getVTList(MVT::i32, MVT::Other); | |||
| 26624 | SDValue Operation = DAG.getNode(X86ISD::TESTUI, dl, VTs, Chain); | |||
| 26625 | SDValue SetCC = getSETCC(X86::COND_B, Operation.getValue(0), dl, DAG); | |||
| 26626 | return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(), SetCC, | |||
| 26627 | Operation.getValue(1)); | |||
| 26628 | } | |||
| 26629 | } | |||
| 26630 | return SDValue(); | |||
| 26631 | } | |||
| 26632 | ||||
| 26633 | SDLoc dl(Op); | |||
| 26634 | switch(IntrData->Type) { | |||
| 26635 | default: llvm_unreachable("Unknown Intrinsic Type")__builtin_unreachable(); | |||
| 26636 | case RDSEED: | |||
| 26637 | case RDRAND: { | |||
| 26638 | // Emit the node with the right value type. | |||
| 26639 | SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::i32, MVT::Other); | |||
| 26640 | SDValue Result = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(0)); | |||
| 26641 | ||||
| 26642 | // If the value returned by RDRAND/RDSEED was valid (CF=1), return 1. | |||
| 26643 | // Otherwise return the value from Rand, which is always 0, casted to i32. | |||
| 26644 | SDValue Ops[] = {DAG.getZExtOrTrunc(Result, dl, Op->getValueType(1)), | |||
| 26645 | DAG.getConstant(1, dl, Op->getValueType(1)), | |||
| 26646 | DAG.getTargetConstant(X86::COND_B, dl, MVT::i8), | |||
| 26647 | SDValue(Result.getNode(), 1)}; | |||
| 26648 | SDValue isValid = DAG.getNode(X86ISD::CMOV, dl, Op->getValueType(1), Ops); | |||
| 26649 | ||||
| 26650 | // Return { result, isValid, chain }. | |||
| 26651 | return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(), Result, isValid, | |||
| 26652 | SDValue(Result.getNode(), 2)); | |||
| 26653 | } | |||
| 26654 | case GATHER_AVX2: { | |||
| 26655 | SDValue Chain = Op.getOperand(0); | |||
| 26656 | SDValue Src = Op.getOperand(2); | |||
| 26657 | SDValue Base = Op.getOperand(3); | |||
| 26658 | SDValue Index = Op.getOperand(4); | |||
| 26659 | SDValue Mask = Op.getOperand(5); | |||
| 26660 | SDValue Scale = Op.getOperand(6); | |||
| 26661 | return getAVX2GatherNode(IntrData->Opc0, Op, DAG, Src, Mask, Base, Index, | |||
| 26662 | Scale, Chain, Subtarget); | |||
| 26663 | } | |||
| 26664 | case GATHER: { | |||
| 26665 | //gather(v1, mask, index, base, scale); | |||
| 26666 | SDValue Chain = Op.getOperand(0); | |||
| 26667 | SDValue Src = Op.getOperand(2); | |||
| 26668 | SDValue Base = Op.getOperand(3); | |||
| 26669 | SDValue Index = Op.getOperand(4); | |||
| 26670 | SDValue Mask = Op.getOperand(5); | |||
| 26671 | SDValue Scale = Op.getOperand(6); | |||
| 26672 | return getGatherNode(Op, DAG, Src, Mask, Base, Index, Scale, | |||
| 26673 | Chain, Subtarget); | |||
| 26674 | } | |||
| 26675 | case SCATTER: { | |||
| 26676 | //scatter(base, mask, index, v1, scale); | |||
| 26677 | SDValue Chain = Op.getOperand(0); | |||
| 26678 | SDValue Base = Op.getOperand(2); | |||
| 26679 | SDValue Mask = Op.getOperand(3); | |||
| 26680 | SDValue Index = Op.getOperand(4); | |||
| 26681 | SDValue Src = Op.getOperand(5); | |||
| 26682 | SDValue Scale = Op.getOperand(6); | |||
| 26683 | return getScatterNode(IntrData->Opc0, Op, DAG, Src, Mask, Base, Index, | |||
| 26684 | Scale, Chain, Subtarget); | |||
| 26685 | } | |||
| 26686 | case PREFETCH: { | |||
| 26687 | const APInt &HintVal = Op.getConstantOperandAPInt(6); | |||
| 26688 | assert((HintVal == 2 || HintVal == 3) &&((void)0) | |||
| 26689 | "Wrong prefetch hint in intrinsic: should be 2 or 3")((void)0); | |||
| 26690 | unsigned Opcode = (HintVal == 2 ? IntrData->Opc1 : IntrData->Opc0); | |||
| 26691 | SDValue Chain = Op.getOperand(0); | |||
| 26692 | SDValue Mask = Op.getOperand(2); | |||
| 26693 | SDValue Index = Op.getOperand(3); | |||
| 26694 | SDValue Base = Op.getOperand(4); | |||
| 26695 | SDValue Scale = Op.getOperand(5); | |||
| 26696 | return getPrefetchNode(Opcode, Op, DAG, Mask, Base, Index, Scale, Chain, | |||
| 26697 | Subtarget); | |||
| 26698 | } | |||
| 26699 | // Read Time Stamp Counter (RDTSC) and Processor ID (RDTSCP). | |||
| 26700 | case RDTSC: { | |||
| 26701 | SmallVector<SDValue, 2> Results; | |||
| 26702 | getReadTimeStampCounter(Op.getNode(), dl, IntrData->Opc0, DAG, Subtarget, | |||
| 26703 | Results); | |||
| 26704 | return DAG.getMergeValues(Results, dl); | |||
| 26705 | } | |||
| 26706 | // Read Performance Monitoring Counters. | |||
| 26707 | case RDPMC: | |||
| 26708 | // GetExtended Control Register. | |||
| 26709 | case XGETBV: { | |||
| 26710 | SmallVector<SDValue, 2> Results; | |||
| 26711 | ||||
| 26712 | // RDPMC uses ECX to select the index of the performance counter to read. | |||
| 26713 | // XGETBV uses ECX to select the index of the XCR register to return. | |||
| 26714 | // The result is stored into registers EDX:EAX. | |||
| 26715 | expandIntrinsicWChainHelper(Op.getNode(), dl, DAG, IntrData->Opc0, X86::ECX, | |||
| 26716 | Subtarget, Results); | |||
| 26717 | return DAG.getMergeValues(Results, dl); | |||
| 26718 | } | |||
| 26719 | // XTEST intrinsics. | |||
| 26720 | case XTEST: { | |||
| 26721 | SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::Other); | |||
| 26722 | SDValue InTrans = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(0)); | |||
| 26723 | ||||
| 26724 | SDValue SetCC = getSETCC(X86::COND_NE, InTrans, dl, DAG); | |||
| 26725 | SDValue Ret = DAG.getNode(ISD::ZERO_EXTEND, dl, Op->getValueType(0), SetCC); | |||
| 26726 | return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(), | |||
| 26727 | Ret, SDValue(InTrans.getNode(), 1)); | |||
| 26728 | } | |||
| 26729 | case TRUNCATE_TO_MEM_VI8: | |||
| 26730 | case TRUNCATE_TO_MEM_VI16: | |||
| 26731 | case TRUNCATE_TO_MEM_VI32: { | |||
| 26732 | SDValue Mask = Op.getOperand(4); | |||
| 26733 | SDValue DataToTruncate = Op.getOperand(3); | |||
| 26734 | SDValue Addr = Op.getOperand(2); | |||
| 26735 | SDValue Chain = Op.getOperand(0); | |||
| 26736 | ||||
| 26737 | MemIntrinsicSDNode *MemIntr = dyn_cast<MemIntrinsicSDNode>(Op); | |||
| 26738 | assert(MemIntr && "Expected MemIntrinsicSDNode!")((void)0); | |||
| 26739 | ||||
| 26740 | EVT MemVT = MemIntr->getMemoryVT(); | |||
| 26741 | ||||
| 26742 | uint16_t TruncationOp = IntrData->Opc0; | |||
| 26743 | switch (TruncationOp) { | |||
| 26744 | case X86ISD::VTRUNC: { | |||
| 26745 | if (isAllOnesConstant(Mask)) // return just a truncate store | |||
| 26746 | return DAG.getTruncStore(Chain, dl, DataToTruncate, Addr, MemVT, | |||
| 26747 | MemIntr->getMemOperand()); | |||
| 26748 | ||||
| 26749 | MVT MaskVT = MVT::getVectorVT(MVT::i1, MemVT.getVectorNumElements()); | |||
| 26750 | SDValue VMask = getMaskNode(Mask, MaskVT, Subtarget, DAG, dl); | |||
| 26751 | SDValue Offset = DAG.getUNDEF(VMask.getValueType()); | |||
| 26752 | ||||
| 26753 | return DAG.getMaskedStore(Chain, dl, DataToTruncate, Addr, Offset, VMask, | |||
| 26754 | MemVT, MemIntr->getMemOperand(), ISD::UNINDEXED, | |||
| 26755 | true /* truncating */); | |||
| 26756 | } | |||
| 26757 | case X86ISD::VTRUNCUS: | |||
| 26758 | case X86ISD::VTRUNCS: { | |||
| 26759 | bool IsSigned = (TruncationOp == X86ISD::VTRUNCS); | |||
| 26760 | if (isAllOnesConstant(Mask)) | |||
| 26761 | return EmitTruncSStore(IsSigned, Chain, dl, DataToTruncate, Addr, MemVT, | |||
| 26762 | MemIntr->getMemOperand(), DAG); | |||
| 26763 | ||||
| 26764 | MVT MaskVT = MVT::getVectorVT(MVT::i1, MemVT.getVectorNumElements()); | |||
| 26765 | SDValue VMask = getMaskNode(Mask, MaskVT, Subtarget, DAG, dl); | |||
| 26766 | ||||
| 26767 | return EmitMaskedTruncSStore(IsSigned, Chain, dl, DataToTruncate, Addr, | |||
| 26768 | VMask, MemVT, MemIntr->getMemOperand(), DAG); | |||
| 26769 | } | |||
| 26770 | default: | |||
| 26771 | llvm_unreachable("Unsupported truncstore intrinsic")__builtin_unreachable(); | |||
| 26772 | } | |||
| 26773 | } | |||
| 26774 | } | |||
| 26775 | } | |||
| 26776 | ||||
| 26777 | SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op, | |||
| 26778 | SelectionDAG &DAG) const { | |||
| 26779 | MachineFrameInfo &MFI = DAG.getMachineFunction().getFrameInfo(); | |||
| 26780 | MFI.setReturnAddressIsTaken(true); | |||
| 26781 | ||||
| 26782 | if (verifyReturnAddressArgumentIsConstant(Op, DAG)) | |||
| 26783 | return SDValue(); | |||
| 26784 | ||||
| 26785 | unsigned Depth = Op.getConstantOperandVal(0); | |||
| 26786 | SDLoc dl(Op); | |||
| 26787 | EVT PtrVT = getPointerTy(DAG.getDataLayout()); | |||
| 26788 | ||||
| 26789 | if (Depth > 0) { | |||
| 26790 | SDValue FrameAddr = LowerFRAMEADDR(Op, DAG); | |||
| 26791 | const X86RegisterInfo *RegInfo = Subtarget.getRegisterInfo(); | |||
| 26792 | SDValue Offset = DAG.getConstant(RegInfo->getSlotSize(), dl, PtrVT); | |||
| 26793 | return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), | |||
| 26794 | DAG.getNode(ISD::ADD, dl, PtrVT, FrameAddr, Offset), | |||
| 26795 | MachinePointerInfo()); | |||
| 26796 | } | |||
| 26797 | ||||
| 26798 | // Just load the return address. | |||
| 26799 | SDValue RetAddrFI = getReturnAddressFrameIndex(DAG); | |||
| 26800 | return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), RetAddrFI, | |||
| 26801 | MachinePointerInfo()); | |||
| 26802 | } | |||
| 26803 | ||||
| 26804 | SDValue X86TargetLowering::LowerADDROFRETURNADDR(SDValue Op, | |||
| 26805 | SelectionDAG &DAG) const { | |||
| 26806 | DAG.getMachineFunction().getFrameInfo().setReturnAddressIsTaken(true); | |||
| 26807 | return getReturnAddressFrameIndex(DAG); | |||
| 26808 | } | |||
| 26809 | ||||
| 26810 | SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const { | |||
| 26811 | MachineFunction &MF = DAG.getMachineFunction(); | |||
| 26812 | MachineFrameInfo &MFI = MF.getFrameInfo(); | |||
| 26813 | X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>(); | |||
| 26814 | const X86RegisterInfo *RegInfo = Subtarget.getRegisterInfo(); | |||
| 26815 | EVT VT = Op.getValueType(); | |||
| 26816 | ||||
| 26817 | MFI.setFrameAddressIsTaken(true); | |||
| 26818 | ||||
| 26819 | if (MF.getTarget().getMCAsmInfo()->usesWindowsCFI()) { | |||
| 26820 | // Depth > 0 makes no sense on targets which use Windows unwind codes. It | |||
| 26821 | // is not possible to crawl up the stack without looking at the unwind codes | |||
| 26822 | // simultaneously. | |||
| 26823 | int FrameAddrIndex = FuncInfo->getFAIndex(); | |||
| 26824 | if (!FrameAddrIndex) { | |||
| 26825 | // Set up a frame object for the return address. | |||
| 26826 | unsigned SlotSize = RegInfo->getSlotSize(); | |||
| 26827 | FrameAddrIndex = MF.getFrameInfo().CreateFixedObject( | |||
| 26828 | SlotSize, /*SPOffset=*/0, /*IsImmutable=*/false); | |||
| 26829 | FuncInfo->setFAIndex(FrameAddrIndex); | |||
| 26830 | } | |||
| 26831 | return DAG.getFrameIndex(FrameAddrIndex, VT); | |||
| 26832 | } | |||
| 26833 | ||||
| 26834 | unsigned FrameReg = | |||
| 26835 | RegInfo->getPtrSizedFrameRegister(DAG.getMachineFunction()); | |||
| 26836 | SDLoc dl(Op); // FIXME probably not meaningful | |||
| 26837 | unsigned Depth = Op.getConstantOperandVal(0); | |||
| 26838 | assert(((FrameReg == X86::RBP && VT == MVT::i64) ||((void)0) | |||
| 26839 | (FrameReg == X86::EBP && VT == MVT::i32)) &&((void)0) | |||
| 26840 | "Invalid Frame Register!")((void)0); | |||
| 26841 | SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT); | |||
| 26842 | while (Depth--) | |||
| 26843 | FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, | |||
| 26844 | MachinePointerInfo()); | |||
| 26845 | return FrameAddr; | |||
| 26846 | } | |||
| 26847 | ||||
| 26848 | // FIXME? Maybe this could be a TableGen attribute on some registers and | |||
| 26849 | // this table could be generated automatically from RegInfo. | |||
| 26850 | Register X86TargetLowering::getRegisterByName(const char* RegName, LLT VT, | |||
| 26851 | const MachineFunction &MF) const { | |||
| 26852 | const TargetFrameLowering &TFI = *Subtarget.getFrameLowering(); | |||
| 26853 | ||||
| 26854 | Register Reg = StringSwitch<unsigned>(RegName) | |||
| 26855 | .Case("esp", X86::ESP) | |||
| 26856 | .Case("rsp", X86::RSP) | |||
| 26857 | .Case("ebp", X86::EBP) | |||
| 26858 | .Case("rbp", X86::RBP) | |||
| 26859 | .Default(0); | |||
| 26860 | ||||
| 26861 | if (Reg == X86::EBP || Reg == X86::RBP) { | |||
| 26862 | if (!TFI.hasFP(MF)) | |||
| 26863 | report_fatal_error("register " + StringRef(RegName) + | |||
| 26864 | " is allocatable: function has no frame pointer"); | |||
| 26865 | #ifndef NDEBUG1 | |||
| 26866 | else { | |||
| 26867 | const X86RegisterInfo *RegInfo = Subtarget.getRegisterInfo(); | |||
| 26868 | Register FrameReg = RegInfo->getPtrSizedFrameRegister(MF); | |||
| 26869 | assert((FrameReg == X86::EBP || FrameReg == X86::RBP) &&((void)0) | |||
| 26870 | "Invalid Frame Register!")((void)0); | |||
| 26871 | } | |||
| 26872 | #endif | |||
| 26873 | } | |||
| 26874 | ||||
| 26875 | if (Reg) | |||
| 26876 | return Reg; | |||
| 26877 | ||||
| 26878 | report_fatal_error("Invalid register name global variable"); | |||
| 26879 | } | |||
| 26880 | ||||
| 26881 | SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op, | |||
| 26882 | SelectionDAG &DAG) const { | |||
| 26883 | const X86RegisterInfo *RegInfo = Subtarget.getRegisterInfo(); | |||
| 26884 | return DAG.getIntPtrConstant(2 * RegInfo->getSlotSize(), SDLoc(Op)); | |||
| 26885 | } | |||
| 26886 | ||||
| 26887 | Register X86TargetLowering::getExceptionPointerRegister( | |||
| 26888 | const Constant *PersonalityFn) const { | |||
| 26889 | if (classifyEHPersonality(PersonalityFn) == EHPersonality::CoreCLR) | |||
| 26890 | return Subtarget.isTarget64BitLP64() ? X86::RDX : X86::EDX; | |||
| 26891 | ||||
| 26892 | return Subtarget.isTarget64BitLP64() ? X86::RAX : X86::EAX; | |||
| 26893 | } | |||
| 26894 | ||||
| 26895 | Register X86TargetLowering::getExceptionSelectorRegister( | |||
| 26896 | const Constant *PersonalityFn) const { | |||
| 26897 | // Funclet personalities don't use selectors (the runtime does the selection). | |||
| 26898 | if (isFuncletEHPersonality(classifyEHPersonality(PersonalityFn))) | |||
| 26899 | return X86::NoRegister; | |||
| 26900 | return Subtarget.isTarget64BitLP64() ? X86::RDX : X86::EDX; | |||
| 26901 | } | |||
| 26902 | ||||
| 26903 | bool X86TargetLowering::needsFixedCatchObjects() const { | |||
| 26904 | return Subtarget.isTargetWin64(); | |||
| 26905 | } | |||
| 26906 | ||||
| 26907 | SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const { | |||
| 26908 | SDValue Chain = Op.getOperand(0); | |||
| 26909 | SDValue Offset = Op.getOperand(1); | |||
| 26910 | SDValue Handler = Op.getOperand(2); | |||
| 26911 | SDLoc dl (Op); | |||
| 26912 | ||||
| 26913 | EVT PtrVT = getPointerTy(DAG.getDataLayout()); | |||
| 26914 | const X86RegisterInfo *RegInfo = Subtarget.getRegisterInfo(); | |||
| 26915 | Register FrameReg = RegInfo->getFrameRegister(DAG.getMachineFunction()); | |||
| 26916 | assert(((FrameReg == X86::RBP && PtrVT == MVT::i64) ||((void)0) | |||
| 26917 | (FrameReg == X86::EBP && PtrVT == MVT::i32)) &&((void)0) | |||
| 26918 | "Invalid Frame Register!")((void)0); | |||
| 26919 | SDValue Frame = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, PtrVT); | |||
| 26920 | Register StoreAddrReg = (PtrVT == MVT::i64) ? X86::RCX : X86::ECX; | |||
| 26921 | ||||
| 26922 | SDValue StoreAddr = DAG.getNode(ISD::ADD, dl, PtrVT, Frame, | |||
| 26923 | DAG.getIntPtrConstant(RegInfo->getSlotSize(), | |||
| 26924 | dl)); | |||
| 26925 | StoreAddr = DAG.getNode(ISD::ADD, dl, PtrVT, StoreAddr, Offset); | |||
| 26926 | Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, MachinePointerInfo()); | |||
| 26927 | Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr); | |||
| 26928 | ||||
| 26929 | return DAG.getNode(X86ISD::EH_RETURN, dl, MVT::Other, Chain, | |||
| 26930 | DAG.getRegister(StoreAddrReg, PtrVT)); | |||
| 26931 | } | |||
| 26932 | ||||
| 26933 | SDValue X86TargetLowering::lowerEH_SJLJ_SETJMP(SDValue Op, | |||
| 26934 | SelectionDAG &DAG) const { | |||
| 26935 | SDLoc DL(Op); | |||
| 26936 | // If the subtarget is not 64bit, we may need the global base reg | |||
| 26937 | // after isel expand pseudo, i.e., after CGBR pass ran. | |||
| 26938 | // Therefore, ask for the GlobalBaseReg now, so that the pass | |||
| 26939 | // inserts the code for us in case we need it. | |||
| 26940 | // Otherwise, we will end up in a situation where we will | |||
| 26941 | // reference a virtual register that is not defined! | |||
| 26942 | if (!Subtarget.is64Bit()) { | |||
| 26943 | const X86InstrInfo *TII = Subtarget.getInstrInfo(); | |||
| 26944 | (void)TII->getGlobalBaseReg(&DAG.getMachineFunction()); | |||
| 26945 | } | |||
| 26946 | return DAG.getNode(X86ISD::EH_SJLJ_SETJMP, DL, | |||
| 26947 | DAG.getVTList(MVT::i32, MVT::Other), | |||
| 26948 | Op.getOperand(0), Op.getOperand(1)); | |||
| 26949 | } | |||
| 26950 | ||||
| 26951 | SDValue X86TargetLowering::lowerEH_SJLJ_LONGJMP(SDValue Op, | |||
| 26952 | SelectionDAG &DAG) const { | |||
| 26953 | SDLoc DL(Op); | |||
| 26954 | return DAG.getNode(X86ISD::EH_SJLJ_LONGJMP, DL, MVT::Other, | |||
| 26955 | Op.getOperand(0), Op.getOperand(1)); | |||
| 26956 | } | |||
| 26957 | ||||
| 26958 | SDValue X86TargetLowering::lowerEH_SJLJ_SETUP_DISPATCH(SDValue Op, | |||
| 26959 | SelectionDAG &DAG) const { | |||
| 26960 | SDLoc DL(Op); | |||
| 26961 | return DAG.getNode(X86ISD::EH_SJLJ_SETUP_DISPATCH, DL, MVT::Other, | |||
| 26962 | Op.getOperand(0)); | |||
| 26963 | } | |||
| 26964 | ||||
| 26965 | static SDValue LowerADJUST_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) { | |||
| 26966 | return Op.getOperand(0); | |||
| 26967 | } | |||
| 26968 | ||||
| 26969 | SDValue X86TargetLowering::LowerINIT_TRAMPOLINE(SDValue Op, | |||
| 26970 | SelectionDAG &DAG) const { | |||
| 26971 | SDValue Root = Op.getOperand(0); | |||
| 26972 | SDValue Trmp = Op.getOperand(1); // trampoline | |||
| 26973 | SDValue FPtr = Op.getOperand(2); // nested function | |||
| 26974 | SDValue Nest = Op.getOperand(3); // 'nest' parameter value | |||
| 26975 | SDLoc dl (Op); | |||
| 26976 | ||||
| 26977 | const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue(); | |||
| 26978 | const TargetRegisterInfo *TRI = Subtarget.getRegisterInfo(); | |||
| 26979 | ||||
| 26980 | if (Subtarget.is64Bit()) { | |||
| 26981 | SDValue OutChains[6]; | |||
| 26982 | ||||
| 26983 | // Large code-model. | |||
| 26984 | const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode. | |||
| 26985 | const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode. | |||
| 26986 | ||||
| 26987 | const unsigned char N86R10 = TRI->getEncodingValue(X86::R10) & 0x7; | |||
| 26988 | const unsigned char N86R11 = TRI->getEncodingValue(X86::R11) & 0x7; | |||
| 26989 | ||||
| 26990 | const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix | |||
| 26991 | ||||
| 26992 | // Load the pointer to the nested function into R11. | |||
| 26993 | unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11 | |||
| 26994 | SDValue Addr = Trmp; | |||
| 26995 | OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, dl, MVT::i16), | |||
| 26996 | Addr, MachinePointerInfo(TrmpAddr)); | |||
| 26997 | ||||
| 26998 | Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp, | |||
| 26999 | DAG.getConstant(2, dl, MVT::i64)); | |||
| 27000 | OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr, | |||
| 27001 | MachinePointerInfo(TrmpAddr, 2), Align(2)); | |||
| 27002 | ||||
| 27003 | // Load the 'nest' parameter value into R10. | |||
| 27004 | // R10 is specified in X86CallingConv.td | |||
| 27005 | OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10 | |||
| 27006 | Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp, | |||
| 27007 | DAG.getConstant(10, dl, MVT::i64)); | |||
| 27008 | OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, dl, MVT::i16), | |||
| 27009 | Addr, MachinePointerInfo(TrmpAddr, 10)); | |||
| 27010 | ||||
| 27011 | Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp, | |||
| 27012 | DAG.getConstant(12, dl, MVT::i64)); | |||
| 27013 | OutChains[3] = DAG.getStore(Root, dl, Nest, Addr, | |||
| 27014 | MachinePointerInfo(TrmpAddr, 12), Align(2)); | |||
| 27015 | ||||
| 27016 | // Jump to the nested function. | |||
| 27017 | OpCode = (JMP64r << 8) | REX_WB; // jmpq *... | |||
| 27018 | Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp, | |||
| 27019 | DAG.getConstant(20, dl, MVT::i64)); | |||
| 27020 | OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, dl, MVT::i16), | |||
| 27021 | Addr, MachinePointerInfo(TrmpAddr, 20)); | |||
| 27022 | ||||
| 27023 | unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11 | |||
| 27024 | Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp, | |||
| 27025 | DAG.getConstant(22, dl, MVT::i64)); | |||
| 27026 | OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, dl, MVT::i8), | |||
| 27027 | Addr, MachinePointerInfo(TrmpAddr, 22)); | |||
| 27028 | ||||
| 27029 | return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains); | |||
| 27030 | } else { | |||
| 27031 | const Function *Func = | |||
| 27032 | cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue()); | |||
| 27033 | CallingConv::ID CC = Func->getCallingConv(); | |||
| 27034 | unsigned NestReg; | |||
| 27035 | ||||
| 27036 | switch (CC) { | |||
| 27037 | default: | |||
| 27038 | llvm_unreachable("Unsupported calling convention")__builtin_unreachable(); | |||
| 27039 | case CallingConv::C: | |||
| 27040 | case CallingConv::X86_StdCall: { | |||
| 27041 | // Pass 'nest' parameter in ECX. | |||
| 27042 | // Must be kept in sync with X86CallingConv.td | |||
| 27043 | NestReg = X86::ECX; | |||
| 27044 | ||||
| 27045 | // Check that ECX wasn't needed by an 'inreg' parameter. | |||
| 27046 | FunctionType *FTy = Func->getFunctionType(); | |||
| 27047 | const AttributeList &Attrs = Func->getAttributes(); | |||
| 27048 | ||||
| 27049 | if (!Attrs.isEmpty() && !Func->isVarArg()) { | |||
| 27050 | unsigned InRegCount = 0; | |||
| 27051 | unsigned Idx = 1; | |||
| 27052 | ||||
| 27053 | for (FunctionType::param_iterator I = FTy->param_begin(), | |||
| 27054 | E = FTy->param_end(); I != E; ++I, ++Idx) | |||
| 27055 | if (Attrs.hasAttribute(Idx, Attribute::InReg)) { | |||
| 27056 | const DataLayout &DL = DAG.getDataLayout(); | |||
| 27057 | // FIXME: should only count parameters that are lowered to integers. | |||
| 27058 | InRegCount += (DL.getTypeSizeInBits(*I) + 31) / 32; | |||
| 27059 | } | |||
| 27060 | ||||
| 27061 | if (InRegCount > 2) { | |||
| 27062 | report_fatal_error("Nest register in use - reduce number of inreg" | |||
| 27063 | " parameters!"); | |||
| 27064 | } | |||
| 27065 | } | |||
| 27066 | break; | |||
| 27067 | } | |||
| 27068 | case CallingConv::X86_FastCall: | |||
| 27069 | case CallingConv::X86_ThisCall: | |||
| 27070 | case CallingConv::Fast: | |||
| 27071 | case CallingConv::Tail: | |||
| 27072 | case CallingConv::SwiftTail: | |||
| 27073 | // Pass 'nest' parameter in EAX. | |||
| 27074 | // Must be kept in sync with X86CallingConv.td | |||
| 27075 | NestReg = X86::EAX; | |||
| 27076 | break; | |||
| 27077 | } | |||
| 27078 | ||||
| 27079 | SDValue OutChains[4]; | |||
| 27080 | SDValue Addr, Disp; | |||
| 27081 | ||||
| 27082 | Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp, | |||
| 27083 | DAG.getConstant(10, dl, MVT::i32)); | |||
| 27084 | Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr); | |||
| 27085 | ||||
| 27086 | // This is storing the opcode for MOV32ri. | |||
| 27087 | const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte. | |||
| 27088 | const unsigned char N86Reg = TRI->getEncodingValue(NestReg) & 0x7; | |||
| 27089 | OutChains[0] = | |||
| 27090 | DAG.getStore(Root, dl, DAG.getConstant(MOV32ri | N86Reg, dl, MVT::i8), | |||
| 27091 | Trmp, MachinePointerInfo(TrmpAddr)); | |||
| 27092 | ||||
| 27093 | Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp, | |||
| 27094 | DAG.getConstant(1, dl, MVT::i32)); | |||
| 27095 | OutChains[1] = DAG.getStore(Root, dl, Nest, Addr, | |||
| 27096 | MachinePointerInfo(TrmpAddr, 1), Align(1)); | |||
| 27097 | ||||
| 27098 | const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode. | |||
| 27099 | Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp, | |||
| 27100 | DAG.getConstant(5, dl, MVT::i32)); | |||
| 27101 | OutChains[2] = | |||
| 27102 | DAG.getStore(Root, dl, DAG.getConstant(JMP, dl, MVT::i8), Addr, | |||
| 27103 | MachinePointerInfo(TrmpAddr, 5), Align(1)); | |||
| 27104 | ||||
| 27105 | Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp, | |||
| 27106 | DAG.getConstant(6, dl, MVT::i32)); | |||
| 27107 | OutChains[3] = DAG.getStore(Root, dl, Disp, Addr, | |||
| 27108 | MachinePointerInfo(TrmpAddr, 6), Align(1)); | |||
| 27109 | ||||
| 27110 | return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains); | |||
| 27111 | } | |||
| 27112 | } | |||
| 27113 | ||||
| 27114 | SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op, | |||
| 27115 | SelectionDAG &DAG) const { | |||
| 27116 | /* | |||
| 27117 | The rounding mode is in bits 11:10 of FPSR, and has the following | |||
| 27118 | settings: | |||
| 27119 | 00 Round to nearest | |||
| 27120 | 01 Round to -inf | |||
| 27121 | 10 Round to +inf | |||
| 27122 | 11 Round to 0 | |||
| 27123 | ||||
| 27124 | FLT_ROUNDS, on the other hand, expects the following: | |||
| 27125 | -1 Undefined | |||
| 27126 | 0 Round to 0 | |||
| 27127 | 1 Round to nearest | |||
| 27128 | 2 Round to +inf | |||
| 27129 | 3 Round to -inf | |||
| 27130 | ||||
| 27131 | To perform the conversion, we use a packed lookup table of the four 2-bit | |||
| 27132 | values that we can index by FPSP[11:10] | |||
| 27133 | 0x2d --> (0b00,10,11,01) --> (0,2,3,1) >> FPSR[11:10] | |||
| 27134 | ||||
| 27135 | (0x2d >> ((FPSR & 0xc00) >> 9)) & 3 | |||
| 27136 | */ | |||
| 27137 | ||||
| 27138 | MachineFunction &MF = DAG.getMachineFunction(); | |||
| 27139 | MVT VT = Op.getSimpleValueType(); | |||
| 27140 | SDLoc DL(Op); | |||
| 27141 | ||||
| 27142 | // Save FP Control Word to stack slot | |||
| 27143 | int SSFI = MF.getFrameInfo().CreateStackObject(2, Align(2), false); | |||
| 27144 | SDValue StackSlot = | |||
| 27145 | DAG.getFrameIndex(SSFI, getPointerTy(DAG.getDataLayout())); | |||
| 27146 | ||||
| 27147 | MachinePointerInfo MPI = MachinePointerInfo::getFixedStack(MF, SSFI); | |||
| 27148 | ||||
| 27149 | SDValue Chain = Op.getOperand(0); | |||
| 27150 | SDValue Ops[] = {Chain, StackSlot}; | |||
| 27151 | Chain = DAG.getMemIntrinsicNode(X86ISD::FNSTCW16m, DL, | |||
| 27152 | DAG.getVTList(MVT::Other), Ops, MVT::i16, MPI, | |||
| 27153 | Align(2), MachineMemOperand::MOStore); | |||
| 27154 | ||||
| 27155 | // Load FP Control Word from stack slot | |||
| 27156 | SDValue CWD = DAG.getLoad(MVT::i16, DL, Chain, StackSlot, MPI, Align(2)); | |||
| 27157 | Chain = CWD.getValue(1); | |||
| 27158 | ||||
| 27159 | // Mask and turn the control bits into a shift for the lookup table. | |||
| 27160 | SDValue Shift = | |||
| 27161 | DAG.getNode(ISD::SRL, DL, MVT::i16, | |||
| 27162 | DAG.getNode(ISD::AND, DL, MVT::i16, | |||
| 27163 | CWD, DAG.getConstant(0xc00, DL, MVT::i16)), | |||
| 27164 | DAG.getConstant(9, DL, MVT::i8)); | |||
| 27165 | Shift = DAG.getNode(ISD::TRUNCATE, DL, MVT::i8, Shift); | |||
| 27166 | ||||
| 27167 | SDValue LUT = DAG.getConstant(0x2d, DL, MVT::i32); | |||
| 27168 | SDValue RetVal = | |||
| 27169 | DAG.getNode(ISD::AND, DL, MVT::i32, | |||
| 27170 | DAG.getNode(ISD::SRL, DL, MVT::i32, LUT, Shift), | |||
| 27171 | DAG.getConstant(3, DL, MVT::i32)); | |||
| 27172 | ||||
| 27173 | RetVal = DAG.getZExtOrTrunc(RetVal, DL, VT); | |||
| 27174 | ||||
| 27175 | return DAG.getMergeValues({RetVal, Chain}, DL); | |||
| 27176 | } | |||
| 27177 | ||||
| 27178 | SDValue X86TargetLowering::LowerSET_ROUNDING(SDValue Op, | |||
| 27179 | SelectionDAG &DAG) const { | |||
| 27180 | MachineFunction &MF = DAG.getMachineFunction(); | |||
| 27181 | SDLoc DL(Op); | |||
| 27182 | SDValue Chain = Op.getNode()->getOperand(0); | |||
| 27183 | ||||
| 27184 | // FP control word may be set only from data in memory. So we need to allocate | |||
| 27185 | // stack space to save/load FP control word. | |||
| 27186 | int OldCWFrameIdx = MF.getFrameInfo().CreateStackObject(4, Align(4), false); | |||
| 27187 | SDValue StackSlot = | |||
| 27188 | DAG.getFrameIndex(OldCWFrameIdx, getPointerTy(DAG.getDataLayout())); | |||
| 27189 | MachinePointerInfo MPI = MachinePointerInfo::getFixedStack(MF, OldCWFrameIdx); | |||
| 27190 | MachineMemOperand *MMO = | |||
| 27191 | MF.getMachineMemOperand(MPI, MachineMemOperand::MOStore, 2, Align(2)); | |||
| 27192 | ||||
| 27193 | // Store FP control word into memory. | |||
| 27194 | SDValue Ops[] = {Chain, StackSlot}; | |||
| 27195 | Chain = DAG.getMemIntrinsicNode( | |||
| 27196 | X86ISD::FNSTCW16m, DL, DAG.getVTList(MVT::Other), Ops, MVT::i16, MMO); | |||
| 27197 | ||||
| 27198 | // Load FP Control Word from stack slot and clear RM field (bits 11:10). | |||
| 27199 | SDValue CWD = DAG.getLoad(MVT::i16, DL, Chain, StackSlot, MPI); | |||
| 27200 | Chain = CWD.getValue(1); | |||
| 27201 | CWD = DAG.getNode(ISD::AND, DL, MVT::i16, CWD.getValue(0), | |||
| 27202 | DAG.getConstant(0xf3ff, DL, MVT::i16)); | |||
| 27203 | ||||
| 27204 | // Calculate new rounding mode. | |||
| 27205 | SDValue NewRM = Op.getNode()->getOperand(1); | |||
| 27206 | SDValue RMBits; | |||
| 27207 | if (auto *CVal = dyn_cast<ConstantSDNode>(NewRM)) { | |||
| 27208 | uint64_t RM = CVal->getZExtValue(); | |||
| 27209 | int FieldVal; | |||
| 27210 | switch (static_cast<RoundingMode>(RM)) { | |||
| 27211 | case RoundingMode::NearestTiesToEven: FieldVal = X86::rmToNearest; break; | |||
| 27212 | case RoundingMode::TowardNegative: FieldVal = X86::rmDownward; break; | |||
| 27213 | case RoundingMode::TowardPositive: FieldVal = X86::rmUpward; break; | |||
| 27214 | case RoundingMode::TowardZero: FieldVal = X86::rmTowardZero; break; | |||
| 27215 | default: | |||
| 27216 | llvm_unreachable("rounding mode is not supported by X86 hardware")__builtin_unreachable(); | |||
| 27217 | } | |||
| 27218 | RMBits = DAG.getConstant(FieldVal, DL, MVT::i16); | |||
| 27219 | } else { | |||
| 27220 | // Need to convert argument into bits of control word: | |||
| 27221 | // 0 Round to 0 -> 11 | |||
| 27222 | // 1 Round to nearest -> 00 | |||
| 27223 | // 2 Round to +inf -> 10 | |||
| 27224 | // 3 Round to -inf -> 01 | |||
| 27225 | // The 2-bit value needs then to be shifted so that it occupies bits 11:10. | |||
| 27226 | // To make the conversion, put all these values into a value 0xc9 and shift | |||
| 27227 | // it left depending on the rounding mode: | |||
| 27228 | // (0xc9 << 4) & 0xc00 = X86::rmTowardZero | |||
| 27229 | // (0xc9 << 6) & 0xc00 = X86::rmToNearest | |||
| 27230 | // ... | |||
| 27231 | // (0xc9 << (2 * NewRM + 4)) & 0xc00 | |||
| 27232 | SDValue ShiftValue = | |||
| 27233 | DAG.getNode(ISD::TRUNCATE, DL, MVT::i8, | |||
| 27234 | DAG.getNode(ISD::ADD, DL, MVT::i32, | |||
| 27235 | DAG.getNode(ISD::SHL, DL, MVT::i32, NewRM, | |||
| 27236 | DAG.getConstant(1, DL, MVT::i8)), | |||
| 27237 | DAG.getConstant(4, DL, MVT::i32))); | |||
| 27238 | SDValue Shifted = | |||
| 27239 | DAG.getNode(ISD::SHL, DL, MVT::i16, DAG.getConstant(0xc9, DL, MVT::i16), | |||
| 27240 | ShiftValue); | |||
| 27241 | RMBits = DAG.getNode(ISD::AND, DL, MVT::i16, Shifted, | |||
| 27242 | DAG.getConstant(0xc00, DL, MVT::i16)); | |||
| 27243 | } | |||
| 27244 | ||||
| 27245 | // Update rounding mode bits and store the new FP Control Word into stack. | |||
| 27246 | CWD = DAG.getNode(ISD::OR, DL, MVT::i16, CWD, RMBits); | |||
| 27247 | Chain = DAG.getStore(Chain, DL, CWD, StackSlot, MPI, /* Alignment = */ 2); | |||
| 27248 | ||||
| 27249 | // Load FP control word from the slot. | |||
| 27250 | SDValue OpsLD[] = {Chain, StackSlot}; | |||
| 27251 | MachineMemOperand *MMOL = | |||
| 27252 | MF.getMachineMemOperand(MPI, MachineMemOperand::MOLoad, 2, Align(2)); | |||
| 27253 | Chain = DAG.getMemIntrinsicNode( | |||
| 27254 | X86ISD::FLDCW16m, DL, DAG.getVTList(MVT::Other), OpsLD, MVT::i16, MMOL); | |||
| 27255 | ||||
| 27256 | // If target supports SSE, set MXCSR as well. Rounding mode is encoded in the | |||
| 27257 | // same way but in bits 14:13. | |||
| 27258 | if (Subtarget.hasSSE1()) { | |||
| 27259 | // Store MXCSR into memory. | |||
| 27260 | Chain = DAG.getNode( | |||
| 27261 | ISD::INTRINSIC_VOID, DL, DAG.getVTList(MVT::Other), Chain, | |||
| 27262 | DAG.getTargetConstant(Intrinsic::x86_sse_stmxcsr, DL, MVT::i32), | |||
| 27263 | StackSlot); | |||
| 27264 | ||||
| 27265 | // Load MXCSR from stack slot and clear RM field (bits 14:13). | |||
| 27266 | SDValue CWD = DAG.getLoad(MVT::i32, DL, Chain, StackSlot, MPI); | |||
| 27267 | Chain = CWD.getValue(1); | |||
| 27268 | CWD = DAG.getNode(ISD::AND, DL, MVT::i32, CWD.getValue(0), | |||
| 27269 | DAG.getConstant(0xffff9fff, DL, MVT::i32)); | |||
| 27270 | ||||
| 27271 | // Shift X87 RM bits from 11:10 to 14:13. | |||
| 27272 | RMBits = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, RMBits); | |||
| 27273 | RMBits = DAG.getNode(ISD::SHL, DL, MVT::i32, RMBits, | |||
| 27274 | DAG.getConstant(3, DL, MVT::i8)); | |||
| 27275 | ||||
| 27276 | // Update rounding mode bits and store the new FP Control Word into stack. | |||
| 27277 | CWD = DAG.getNode(ISD::OR, DL, MVT::i32, CWD, RMBits); | |||
| 27278 | Chain = DAG.getStore(Chain, DL, CWD, StackSlot, MPI, /* Alignment = */ 4); | |||
| 27279 | ||||
| 27280 | // Load MXCSR from the slot. | |||
| 27281 | Chain = DAG.getNode( | |||
| 27282 | ISD::INTRINSIC_VOID, DL, DAG.getVTList(MVT::Other), Chain, | |||
| 27283 | DAG.getTargetConstant(Intrinsic::x86_sse_ldmxcsr, DL, MVT::i32), | |||
| 27284 | StackSlot); | |||
| 27285 | } | |||
| 27286 | ||||
| 27287 | return Chain; | |||
| 27288 | } | |||
| 27289 | ||||
| 27290 | /// Lower a vector CTLZ using native supported vector CTLZ instruction. | |||
| 27291 | // | |||
| 27292 | // i8/i16 vector implemented using dword LZCNT vector instruction | |||
| 27293 | // ( sub(trunc(lzcnt(zext32(x)))) ). In case zext32(x) is illegal, | |||
| 27294 | // split the vector, perform operation on it's Lo a Hi part and | |||
| 27295 | // concatenate the results. | |||
| 27296 | static SDValue LowerVectorCTLZ_AVX512CDI(SDValue Op, SelectionDAG &DAG, | |||
| 27297 | const X86Subtarget &Subtarget) { | |||
| 27298 | assert(Op.getOpcode() == ISD::CTLZ)((void)0); | |||
| 27299 | SDLoc dl(Op); | |||
| 27300 | MVT VT = Op.getSimpleValueType(); | |||
| 27301 | MVT EltVT = VT.getVectorElementType(); | |||
| 27302 | unsigned NumElems = VT.getVectorNumElements(); | |||
| 27303 | ||||
| 27304 | assert((EltVT == MVT::i8 || EltVT == MVT::i16) &&((void)0) | |||
| 27305 | "Unsupported element type")((void)0); | |||
| 27306 | ||||
| 27307 | // Split vector, it's Lo and Hi parts will be handled in next iteration. | |||
| 27308 | if (NumElems > 16 || | |||
| 27309 | (NumElems == 16 && !Subtarget.canExtendTo512DQ())) | |||
| 27310 | return splitVectorIntUnary(Op, DAG); | |||
| 27311 | ||||
| 27312 | MVT NewVT = MVT::getVectorVT(MVT::i32, NumElems); | |||
| 27313 | assert((NewVT.is256BitVector() || NewVT.is512BitVector()) &&((void)0) | |||
| 27314 | "Unsupported value type for operation")((void)0); | |||
| 27315 | ||||
| 27316 | // Use native supported vector instruction vplzcntd. | |||
| 27317 | Op = DAG.getNode(ISD::ZERO_EXTEND, dl, NewVT, Op.getOperand(0)); | |||
| 27318 | SDValue CtlzNode = DAG.getNode(ISD::CTLZ, dl, NewVT, Op); | |||
| 27319 | SDValue TruncNode = DAG.getNode(ISD::TRUNCATE, dl, VT, CtlzNode); | |||
| 27320 | SDValue Delta = DAG.getConstant(32 - EltVT.getSizeInBits(), dl, VT); | |||
| 27321 | ||||
| 27322 | return DAG.getNode(ISD::SUB, dl, VT, TruncNode, Delta); | |||
| 27323 | } | |||
| 27324 | ||||
| 27325 | // Lower CTLZ using a PSHUFB lookup table implementation. | |||
| 27326 | static SDValue LowerVectorCTLZInRegLUT(SDValue Op, const SDLoc &DL, | |||
| 27327 | const X86Subtarget &Subtarget, | |||
| 27328 | SelectionDAG &DAG) { | |||
| 27329 | MVT VT = Op.getSimpleValueType(); | |||
| 27330 | int NumElts = VT.getVectorNumElements(); | |||
| 27331 | int NumBytes = NumElts * (VT.getScalarSizeInBits() / 8); | |||
| 27332 | MVT CurrVT = MVT::getVectorVT(MVT::i8, NumBytes); | |||
| 27333 | ||||
| 27334 | // Per-nibble leading zero PSHUFB lookup table. | |||
| 27335 | const int LUT[16] = {/* 0 */ 4, /* 1 */ 3, /* 2 */ 2, /* 3 */ 2, | |||
| 27336 | /* 4 */ 1, /* 5 */ 1, /* 6 */ 1, /* 7 */ 1, | |||
| 27337 | /* 8 */ 0, /* 9 */ 0, /* a */ 0, /* b */ 0, | |||
| 27338 | /* c */ 0, /* d */ 0, /* e */ 0, /* f */ 0}; | |||
| 27339 | ||||
| 27340 | SmallVector<SDValue, 64> LUTVec; | |||
| 27341 | for (int i = 0; i < NumBytes; ++i) | |||
| 27342 | LUTVec.push_back(DAG.getConstant(LUT[i % 16], DL, MVT::i8)); | |||
| 27343 | SDValue InRegLUT = DAG.getBuildVector(CurrVT, DL, LUTVec); | |||
| 27344 | ||||
| 27345 | // Begin by bitcasting the input to byte vector, then split those bytes | |||
| 27346 | // into lo/hi nibbles and use the PSHUFB LUT to perform CLTZ on each of them. | |||
| 27347 | // If the hi input nibble is zero then we add both results together, otherwise | |||
| 27348 | // we just take the hi result (by masking the lo result to zero before the | |||
| 27349 | // add). | |||
| 27350 | SDValue Op0 = DAG.getBitcast(CurrVT, Op.getOperand(0)); | |||
| 27351 | SDValue Zero = DAG.getConstant(0, DL, CurrVT); | |||
| 27352 | ||||
| 27353 | SDValue NibbleShift = DAG.getConstant(0x4, DL, CurrVT); | |||
| 27354 | SDValue Lo = Op0; | |||
| 27355 | SDValue Hi = DAG.getNode(ISD::SRL, DL, CurrVT, Op0, NibbleShift); | |||
| 27356 | SDValue HiZ; | |||
| 27357 | if (CurrVT.is512BitVector()) { | |||
| 27358 | MVT MaskVT = MVT::getVectorVT(MVT::i1, CurrVT.getVectorNumElements()); | |||
| 27359 | HiZ = DAG.getSetCC(DL, MaskVT, Hi, Zero, ISD::SETEQ); | |||
| 27360 | HiZ = DAG.getNode(ISD::SIGN_EXTEND, DL, CurrVT, HiZ); | |||
| 27361 | } else { | |||
| 27362 | HiZ = DAG.getSetCC(DL, CurrVT, Hi, Zero, ISD::SETEQ); | |||
| 27363 | } | |||
| 27364 | ||||
| 27365 | Lo = DAG.getNode(X86ISD::PSHUFB, DL, CurrVT, InRegLUT, Lo); | |||
| 27366 | Hi = DAG.getNode(X86ISD::PSHUFB, DL, CurrVT, InRegLUT, Hi); | |||
| 27367 | Lo = DAG.getNode(ISD::AND, DL, CurrVT, Lo, HiZ); | |||
| 27368 | SDValue Res = DAG.getNode(ISD::ADD, DL, CurrVT, Lo, Hi); | |||
| 27369 | ||||
| 27370 | // Merge result back from vXi8 back to VT, working on the lo/hi halves | |||
| 27371 | // of the current vector width in the same way we did for the nibbles. | |||
| 27372 | // If the upper half of the input element is zero then add the halves' | |||
| 27373 | // leading zero counts together, otherwise just use the upper half's. | |||
| 27374 | // Double the width of the result until we are at target width. | |||
| 27375 | while (CurrVT != VT) { | |||
| 27376 | int CurrScalarSizeInBits = CurrVT.getScalarSizeInBits(); | |||
| 27377 | int CurrNumElts = CurrVT.getVectorNumElements(); | |||
| 27378 | MVT NextSVT = MVT::getIntegerVT(CurrScalarSizeInBits * 2); | |||
| 27379 | MVT NextVT = MVT::getVectorVT(NextSVT, CurrNumElts / 2); | |||
| 27380 | SDValue Shift = DAG.getConstant(CurrScalarSizeInBits, DL, NextVT); | |||
| 27381 | ||||
| 27382 | // Check if the upper half of the input element is zero. | |||
| 27383 | if (CurrVT.is512BitVector()) { | |||
| 27384 | MVT MaskVT = MVT::getVectorVT(MVT::i1, CurrVT.getVectorNumElements()); | |||
| 27385 | HiZ = DAG.getSetCC(DL, MaskVT, DAG.getBitcast(CurrVT, Op0), | |||
| 27386 | DAG.getBitcast(CurrVT, Zero), ISD::SETEQ); | |||
| 27387 | HiZ = DAG.getNode(ISD::SIGN_EXTEND, DL, CurrVT, HiZ); | |||
| 27388 | } else { | |||
| 27389 | HiZ = DAG.getSetCC(DL, CurrVT, DAG.getBitcast(CurrVT, Op0), | |||
| 27390 | DAG.getBitcast(CurrVT, Zero), ISD::SETEQ); | |||
| 27391 | } | |||
| 27392 | HiZ = DAG.getBitcast(NextVT, HiZ); | |||
| 27393 | ||||
| 27394 | // Move the upper/lower halves to the lower bits as we'll be extending to | |||
| 27395 | // NextVT. Mask the lower result to zero if HiZ is true and add the results | |||
| 27396 | // together. | |||
| 27397 | SDValue ResNext = Res = DAG.getBitcast(NextVT, Res); | |||
| 27398 | SDValue R0 = DAG.getNode(ISD::SRL, DL, NextVT, ResNext, Shift); | |||
| 27399 | SDValue R1 = DAG.getNode(ISD::SRL, DL, NextVT, HiZ, Shift); | |||
| 27400 | R1 = DAG.getNode(ISD::AND, DL, NextVT, ResNext, R1); | |||
| 27401 | Res = DAG.getNode(ISD::ADD, DL, NextVT, R0, R1); | |||
| 27402 | CurrVT = NextVT; | |||
| 27403 | } | |||
| 27404 | ||||
| 27405 | return Res; | |||
| 27406 | } | |||
| 27407 | ||||
| 27408 | static SDValue LowerVectorCTLZ(SDValue Op, const SDLoc &DL, | |||
| 27409 | const X86Subtarget &Subtarget, | |||
| 27410 | SelectionDAG &DAG) { | |||
| 27411 | MVT VT = Op.getSimpleValueType(); | |||
| 27412 | ||||
| 27413 | if (Subtarget.hasCDI() && | |||
| 27414 | // vXi8 vectors need to be promoted to 512-bits for vXi32. | |||
| 27415 | (Subtarget.canExtendTo512DQ() || VT.getVectorElementType() != MVT::i8)) | |||
| 27416 | return LowerVectorCTLZ_AVX512CDI(Op, DAG, Subtarget); | |||
| 27417 | ||||
| 27418 | // Decompose 256-bit ops into smaller 128-bit ops. | |||
| 27419 | if (VT.is256BitVector() && !Subtarget.hasInt256()) | |||
| 27420 | return splitVectorIntUnary(Op, DAG); | |||
| 27421 | ||||
| 27422 | // Decompose 512-bit ops into smaller 256-bit ops. | |||
| 27423 | if (VT.is512BitVector() && !Subtarget.hasBWI()) | |||
| 27424 | return splitVectorIntUnary(Op, DAG); | |||
| 27425 | ||||
| 27426 | assert(Subtarget.hasSSSE3() && "Expected SSSE3 support for PSHUFB")((void)0); | |||
| 27427 | return LowerVectorCTLZInRegLUT(Op, DL, Subtarget, DAG); | |||
| 27428 | } | |||
| 27429 | ||||
| 27430 | static SDValue LowerCTLZ(SDValue Op, const X86Subtarget &Subtarget, | |||
| 27431 | SelectionDAG &DAG) { | |||
| 27432 | MVT VT = Op.getSimpleValueType(); | |||
| 27433 | MVT OpVT = VT; | |||
| 27434 | unsigned NumBits = VT.getSizeInBits(); | |||
| 27435 | SDLoc dl(Op); | |||
| 27436 | unsigned Opc = Op.getOpcode(); | |||
| 27437 | ||||
| 27438 | if (VT.isVector()) | |||
| 27439 | return LowerVectorCTLZ(Op, dl, Subtarget, DAG); | |||
| 27440 | ||||
| 27441 | Op = Op.getOperand(0); | |||
| 27442 | if (VT == MVT::i8) { | |||
| 27443 | // Zero extend to i32 since there is not an i8 bsr. | |||
| 27444 | OpVT = MVT::i32; | |||
| 27445 | Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op); | |||
| 27446 | } | |||
| 27447 | ||||
| 27448 | // Issue a bsr (scan bits in reverse) which also sets EFLAGS. | |||
| 27449 | SDVTList VTs = DAG.getVTList(OpVT, MVT::i32); | |||
| 27450 | Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op); | |||
| 27451 | ||||
| 27452 | if (Opc == ISD::CTLZ) { | |||
| 27453 | // If src is zero (i.e. bsr sets ZF), returns NumBits. | |||
| 27454 | SDValue Ops[] = {Op, DAG.getConstant(NumBits + NumBits - 1, dl, OpVT), | |||
| 27455 | DAG.getTargetConstant(X86::COND_E, dl, MVT::i8), | |||
| 27456 | Op.getValue(1)}; | |||
| 27457 | Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops); | |||
| 27458 | } | |||
| 27459 | ||||
| 27460 | // Finally xor with NumBits-1. | |||
| 27461 | Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, | |||
| 27462 | DAG.getConstant(NumBits - 1, dl, OpVT)); | |||
| 27463 | ||||
| 27464 | if (VT == MVT::i8) | |||
| 27465 | Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op); | |||
| 27466 | return Op; | |||
| 27467 | } | |||
| 27468 | ||||
| 27469 | static SDValue LowerCTTZ(SDValue Op, const X86Subtarget &Subtarget, | |||
| 27470 | SelectionDAG &DAG) { | |||
| 27471 | MVT VT = Op.getSimpleValueType(); | |||
| 27472 | unsigned NumBits = VT.getScalarSizeInBits(); | |||
| 27473 | SDValue N0 = Op.getOperand(0); | |||
| 27474 | SDLoc dl(Op); | |||
| 27475 | ||||
| 27476 | assert(!VT.isVector() && Op.getOpcode() == ISD::CTTZ &&((void)0) | |||
| 27477 | "Only scalar CTTZ requires custom lowering")((void)0); | |||
| 27478 | ||||
| 27479 | // Issue a bsf (scan bits forward) which also sets EFLAGS. | |||
| 27480 | SDVTList VTs = DAG.getVTList(VT, MVT::i32); | |||
| 27481 | Op = DAG.getNode(X86ISD::BSF, dl, VTs, N0); | |||
| 27482 | ||||
| 27483 | // If src is zero (i.e. bsf sets ZF), returns NumBits. | |||
| 27484 | SDValue Ops[] = {Op, DAG.getConstant(NumBits, dl, VT), | |||
| 27485 | DAG.getTargetConstant(X86::COND_E, dl, MVT::i8), | |||
| 27486 | Op.getValue(1)}; | |||
| 27487 | return DAG.getNode(X86ISD::CMOV, dl, VT, Ops); | |||
| 27488 | } | |||
| 27489 | ||||
| 27490 | static SDValue lowerAddSub(SDValue Op, SelectionDAG &DAG, | |||
| 27491 | const X86Subtarget &Subtarget) { | |||
| 27492 | MVT VT = Op.getSimpleValueType(); | |||
| 27493 | if (VT == MVT::i16 || VT == MVT::i32) | |||
| 27494 | return lowerAddSubToHorizontalOp(Op, DAG, Subtarget); | |||
| 27495 | ||||
| 27496 | if (VT == MVT::v32i16 || VT == MVT::v64i8) | |||
| 27497 | return splitVectorIntBinary(Op, DAG); | |||
| 27498 | ||||
| 27499 | assert(Op.getSimpleValueType().is256BitVector() &&((void)0) | |||
| 27500 | Op.getSimpleValueType().isInteger() &&((void)0) | |||
| 27501 | "Only handle AVX 256-bit vector integer operation")((void)0); | |||
| 27502 | return splitVectorIntBinary(Op, DAG); | |||
| 27503 | } | |||
| 27504 | ||||
| 27505 | static SDValue LowerADDSAT_SUBSAT(SDValue Op, SelectionDAG &DAG, | |||
| 27506 | const X86Subtarget &Subtarget) { | |||
| 27507 | MVT VT = Op.getSimpleValueType(); | |||
| 27508 | SDValue X = Op.getOperand(0), Y = Op.getOperand(1); | |||
| 27509 | unsigned Opcode = Op.getOpcode(); | |||
| 27510 | SDLoc DL(Op); | |||
| 27511 | ||||
| 27512 | if (VT == MVT::v32i16 || VT == MVT::v64i8 || | |||
| 27513 | (VT.is256BitVector() && !Subtarget.hasInt256())) { | |||
| 27514 | assert(Op.getSimpleValueType().isInteger() &&((void)0) | |||
| 27515 | "Only handle AVX vector integer operation")((void)0); | |||
| 27516 | return splitVectorIntBinary(Op, DAG); | |||
| 27517 | } | |||
| 27518 | ||||
| 27519 | // Avoid the generic expansion with min/max if we don't have pminu*/pmaxu*. | |||
| 27520 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
| 27521 | EVT SetCCResultType = | |||
| 27522 | TLI.getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT); | |||
| 27523 | ||||
| 27524 | if (Opcode == ISD::USUBSAT && !TLI.isOperationLegal(ISD::UMAX, VT)) { | |||
| 27525 | // usubsat X, Y --> (X >u Y) ? X - Y : 0 | |||
| 27526 | SDValue Sub = DAG.getNode(ISD::SUB, DL, VT, X, Y); | |||
| 27527 | SDValue Cmp = DAG.getSetCC(DL, SetCCResultType, X, Y, ISD::SETUGT); | |||
| 27528 | // TODO: Move this to DAGCombiner? | |||
| 27529 | if (SetCCResultType == VT && | |||
| 27530 | DAG.ComputeNumSignBits(Cmp) == VT.getScalarSizeInBits()) | |||
| 27531 | return DAG.getNode(ISD::AND, DL, VT, Cmp, Sub); | |||
| 27532 | return DAG.getSelect(DL, VT, Cmp, Sub, DAG.getConstant(0, DL, VT)); | |||
| 27533 | } | |||
| 27534 | ||||
| 27535 | // Use default expansion. | |||
| 27536 | return SDValue(); | |||
| 27537 | } | |||
| 27538 | ||||
| 27539 | static SDValue LowerABS(SDValue Op, const X86Subtarget &Subtarget, | |||
| 27540 | SelectionDAG &DAG) { | |||
| 27541 | MVT VT = Op.getSimpleValueType(); | |||
| 27542 | if (VT == MVT::i16 || VT == MVT::i32 || VT == MVT::i64) { | |||
| 27543 | // Since X86 does not have CMOV for 8-bit integer, we don't convert | |||
| 27544 | // 8-bit integer abs to NEG and CMOV. | |||
| 27545 | SDLoc DL(Op); | |||
| 27546 | SDValue N0 = Op.getOperand(0); | |||
| 27547 | SDValue Neg = DAG.getNode(X86ISD::SUB, DL, DAG.getVTList(VT, MVT::i32), | |||
| 27548 | DAG.getConstant(0, DL, VT), N0); | |||
| 27549 | SDValue Ops[] = {N0, Neg, DAG.getTargetConstant(X86::COND_GE, DL, MVT::i8), | |||
| 27550 | SDValue(Neg.getNode(), 1)}; | |||
| 27551 | return DAG.getNode(X86ISD::CMOV, DL, VT, Ops); | |||
| 27552 | } | |||
| 27553 | ||||
| 27554 | // ABS(vXi64 X) --> VPBLENDVPD(X, 0-X, X). | |||
| 27555 | if ((VT == MVT::v2i64 || VT == MVT::v4i64) && Subtarget.hasSSE41()) { | |||
| 27556 | SDLoc DL(Op); | |||
| 27557 | SDValue Src = Op.getOperand(0); | |||
| 27558 | SDValue Sub = | |||
| 27559 | DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, DL, VT), Src); | |||
| 27560 | return DAG.getNode(X86ISD::BLENDV, DL, VT, Src, Sub, Src); | |||
| 27561 | } | |||
| 27562 | ||||
| 27563 | if (VT.is256BitVector() && !Subtarget.hasInt256()) { | |||
| 27564 | assert(VT.isInteger() &&((void)0) | |||
| 27565 | "Only handle AVX 256-bit vector integer operation")((void)0); | |||
| 27566 | return splitVectorIntUnary(Op, DAG); | |||
| 27567 | } | |||
| 27568 | ||||
| 27569 | if ((VT == MVT::v32i16 || VT == MVT::v64i8) && !Subtarget.hasBWI()) | |||
| 27570 | return splitVectorIntUnary(Op, DAG); | |||
| 27571 | ||||
| 27572 | // Default to expand. | |||
| 27573 | return SDValue(); | |||
| 27574 | } | |||
| 27575 | ||||
| 27576 | static SDValue LowerMINMAX(SDValue Op, SelectionDAG &DAG) { | |||
| 27577 | MVT VT = Op.getSimpleValueType(); | |||
| 27578 | ||||
| 27579 | // For AVX1 cases, split to use legal ops (everything but v4i64). | |||
| 27580 | if (VT.getScalarType() != MVT::i64 && VT.is256BitVector()) | |||
| 27581 | return splitVectorIntBinary(Op, DAG); | |||
| 27582 | ||||
| 27583 | if (VT == MVT::v32i16 || VT == MVT::v64i8) | |||
| 27584 | return splitVectorIntBinary(Op, DAG); | |||
| 27585 | ||||
| 27586 | // Default to expand. | |||
| 27587 | return SDValue(); | |||
| 27588 | } | |||
| 27589 | ||||
| 27590 | static SDValue LowerMUL(SDValue Op, const X86Subtarget &Subtarget, | |||
| 27591 | SelectionDAG &DAG) { | |||
| 27592 | SDLoc dl(Op); | |||
| 27593 | MVT VT = Op.getSimpleValueType(); | |||
| 27594 | ||||
| 27595 | // Decompose 256-bit ops into 128-bit ops. | |||
| 27596 | if (VT.is256BitVector() && !Subtarget.hasInt256()) | |||
| 27597 | return splitVectorIntBinary(Op, DAG); | |||
| 27598 | ||||
| 27599 | if ((VT == MVT::v32i16 || VT == MVT::v64i8) && !Subtarget.hasBWI()) | |||
| 27600 | return splitVectorIntBinary(Op, DAG); | |||
| 27601 | ||||
| 27602 | SDValue A = Op.getOperand(0); | |||
| 27603 | SDValue B = Op.getOperand(1); | |||
| 27604 | ||||
| 27605 | // Lower v16i8/v32i8/v64i8 mul as sign-extension to v8i16/v16i16/v32i16 | |||
| 27606 | // vector pairs, multiply and truncate. | |||
| 27607 | if (VT == MVT::v16i8 || VT == MVT::v32i8 || VT == MVT::v64i8) { | |||
| 27608 | unsigned NumElts = VT.getVectorNumElements(); | |||
| 27609 | ||||
| 27610 | if ((VT == MVT::v16i8 && Subtarget.hasInt256()) || | |||
| 27611 | (VT == MVT::v32i8 && Subtarget.canExtendTo512BW())) { | |||
| 27612 | MVT ExVT = MVT::getVectorVT(MVT::i16, VT.getVectorNumElements()); | |||
| 27613 | return DAG.getNode( | |||
| 27614 | ISD::TRUNCATE, dl, VT, | |||
| 27615 | DAG.getNode(ISD::MUL, dl, ExVT, | |||
| 27616 | DAG.getNode(ISD::ANY_EXTEND, dl, ExVT, A), | |||
| 27617 | DAG.getNode(ISD::ANY_EXTEND, dl, ExVT, B))); | |||
| 27618 | } | |||
| 27619 | ||||
| 27620 | MVT ExVT = MVT::getVectorVT(MVT::i16, NumElts / 2); | |||
| 27621 | ||||
| 27622 | // Extract the lo/hi parts to any extend to i16. | |||
| 27623 | // We're going to mask off the low byte of each result element of the | |||
| 27624 | // pmullw, so it doesn't matter what's in the high byte of each 16-bit | |||
| 27625 | // element. | |||
| 27626 | SDValue Undef = DAG.getUNDEF(VT); | |||
| 27627 | SDValue ALo = DAG.getBitcast(ExVT, getUnpackl(DAG, dl, VT, A, Undef)); | |||
| 27628 | SDValue AHi = DAG.getBitcast(ExVT, getUnpackh(DAG, dl, VT, A, Undef)); | |||
| 27629 | ||||
| 27630 | SDValue BLo, BHi; | |||
| 27631 | if (ISD::isBuildVectorOfConstantSDNodes(B.getNode())) { | |||
| 27632 | // If the RHS is a constant, manually unpackl/unpackh. | |||
| 27633 | SmallVector<SDValue, 16> LoOps, HiOps; | |||
| 27634 | for (unsigned i = 0; i != NumElts; i += 16) { | |||
| 27635 | for (unsigned j = 0; j != 8; ++j) { | |||
| 27636 | LoOps.push_back(DAG.getAnyExtOrTrunc(B.getOperand(i + j), dl, | |||
| 27637 | MVT::i16)); | |||
| 27638 | HiOps.push_back(DAG.getAnyExtOrTrunc(B.getOperand(i + j + 8), dl, | |||
| 27639 | MVT::i16)); | |||
| 27640 | } | |||
| 27641 | } | |||
| 27642 | ||||
| 27643 | BLo = DAG.getBuildVector(ExVT, dl, LoOps); | |||
| 27644 | BHi = DAG.getBuildVector(ExVT, dl, HiOps); | |||
| 27645 | } else { | |||
| 27646 | BLo = DAG.getBitcast(ExVT, getUnpackl(DAG, dl, VT, B, Undef)); | |||
| 27647 | BHi = DAG.getBitcast(ExVT, getUnpackh(DAG, dl, VT, B, Undef)); | |||
| 27648 | } | |||
| 27649 | ||||
| 27650 | // Multiply, mask the lower 8bits of the lo/hi results and pack. | |||
| 27651 | SDValue RLo = DAG.getNode(ISD::MUL, dl, ExVT, ALo, BLo); | |||
| 27652 | SDValue RHi = DAG.getNode(ISD::MUL, dl, ExVT, AHi, BHi); | |||
| 27653 | RLo = DAG.getNode(ISD::AND, dl, ExVT, RLo, DAG.getConstant(255, dl, ExVT)); | |||
| 27654 | RHi = DAG.getNode(ISD::AND, dl, ExVT, RHi, DAG.getConstant(255, dl, ExVT)); | |||
| 27655 | return DAG.getNode(X86ISD::PACKUS, dl, VT, RLo, RHi); | |||
| 27656 | } | |||
| 27657 | ||||
| 27658 | // Lower v4i32 mul as 2x shuffle, 2x pmuludq, 2x shuffle. | |||
| 27659 | if (VT == MVT::v4i32) { | |||
| 27660 | assert(Subtarget.hasSSE2() && !Subtarget.hasSSE41() &&((void)0) | |||
| 27661 | "Should not custom lower when pmulld is available!")((void)0); | |||
| 27662 | ||||
| 27663 | // Extract the odd parts. | |||
| 27664 | static const int UnpackMask[] = { 1, -1, 3, -1 }; | |||
| 27665 | SDValue Aodds = DAG.getVectorShuffle(VT, dl, A, A, UnpackMask); | |||
| 27666 | SDValue Bodds = DAG.getVectorShuffle(VT, dl, B, B, UnpackMask); | |||
| 27667 | ||||
| 27668 | // Multiply the even parts. | |||
| 27669 | SDValue Evens = DAG.getNode(X86ISD::PMULUDQ, dl, MVT::v2i64, | |||
| 27670 | DAG.getBitcast(MVT::v2i64, A), | |||
| 27671 | DAG.getBitcast(MVT::v2i64, B)); | |||
| 27672 | // Now multiply odd parts. | |||
| 27673 | SDValue Odds = DAG.getNode(X86ISD::PMULUDQ, dl, MVT::v2i64, | |||
| 27674 | DAG.getBitcast(MVT::v2i64, Aodds), | |||
| 27675 | DAG.getBitcast(MVT::v2i64, Bodds)); | |||
| 27676 | ||||
| 27677 | Evens = DAG.getBitcast(VT, Evens); | |||
| 27678 | Odds = DAG.getBitcast(VT, Odds); | |||
| 27679 | ||||
| 27680 | // Merge the two vectors back together with a shuffle. This expands into 2 | |||
| 27681 | // shuffles. | |||
| 27682 | static const int ShufMask[] = { 0, 4, 2, 6 }; | |||
| 27683 | return DAG.getVectorShuffle(VT, dl, Evens, Odds, ShufMask); | |||
| 27684 | } | |||
| 27685 | ||||
| 27686 | assert((VT == MVT::v2i64 || VT == MVT::v4i64 || VT == MVT::v8i64) &&((void)0) | |||
| 27687 | "Only know how to lower V2I64/V4I64/V8I64 multiply")((void)0); | |||
| 27688 | assert(!Subtarget.hasDQI() && "DQI should use MULLQ")((void)0); | |||
| 27689 | ||||
| 27690 | // Ahi = psrlqi(a, 32); | |||
| 27691 | // Bhi = psrlqi(b, 32); | |||
| 27692 | // | |||
| 27693 | // AloBlo = pmuludq(a, b); | |||
| 27694 | // AloBhi = pmuludq(a, Bhi); | |||
| 27695 | // AhiBlo = pmuludq(Ahi, b); | |||
| 27696 | // | |||
| 27697 | // Hi = psllqi(AloBhi + AhiBlo, 32); | |||
| 27698 | // return AloBlo + Hi; | |||
| 27699 | KnownBits AKnown = DAG.computeKnownBits(A); | |||
| 27700 | KnownBits BKnown = DAG.computeKnownBits(B); | |||
| 27701 | ||||
| 27702 | APInt LowerBitsMask = APInt::getLowBitsSet(64, 32); | |||
| 27703 | bool ALoIsZero = LowerBitsMask.isSubsetOf(AKnown.Zero); | |||
| 27704 | bool BLoIsZero = LowerBitsMask.isSubsetOf(BKnown.Zero); | |||
| 27705 | ||||
| 27706 | APInt UpperBitsMask = APInt::getHighBitsSet(64, 32); | |||
| 27707 | bool AHiIsZero = UpperBitsMask.isSubsetOf(AKnown.Zero); | |||
| 27708 | bool BHiIsZero = UpperBitsMask.isSubsetOf(BKnown.Zero); | |||
| 27709 | ||||
| 27710 | SDValue Zero = DAG.getConstant(0, dl, VT); | |||
| 27711 | ||||
| 27712 | // Only multiply lo/hi halves that aren't known to be zero. | |||
| 27713 | SDValue AloBlo = Zero; | |||
| 27714 | if (!ALoIsZero && !BLoIsZero) | |||
| 27715 | AloBlo = DAG.getNode(X86ISD::PMULUDQ, dl, VT, A, B); | |||
| 27716 | ||||
| 27717 | SDValue AloBhi = Zero; | |||
| 27718 | if (!ALoIsZero && !BHiIsZero) { | |||
| 27719 | SDValue Bhi = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, B, 32, DAG); | |||
| 27720 | AloBhi = DAG.getNode(X86ISD::PMULUDQ, dl, VT, A, Bhi); | |||
| 27721 | } | |||
| 27722 | ||||
| 27723 | SDValue AhiBlo = Zero; | |||
| 27724 | if (!AHiIsZero && !BLoIsZero) { | |||
| 27725 | SDValue Ahi = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, A, 32, DAG); | |||
| 27726 | AhiBlo = DAG.getNode(X86ISD::PMULUDQ, dl, VT, Ahi, B); | |||
| 27727 | } | |||
| 27728 | ||||
| 27729 | SDValue Hi = DAG.getNode(ISD::ADD, dl, VT, AloBhi, AhiBlo); | |||
| 27730 | Hi = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, Hi, 32, DAG); | |||
| 27731 | ||||
| 27732 | return DAG.getNode(ISD::ADD, dl, VT, AloBlo, Hi); | |||
| 27733 | } | |||
| 27734 | ||||
| 27735 | static SDValue LowervXi8MulWithUNPCK(SDValue A, SDValue B, const SDLoc &dl, | |||
| 27736 | MVT VT, bool IsSigned, | |||
| 27737 | const X86Subtarget &Subtarget, | |||
| 27738 | SelectionDAG &DAG, | |||
| 27739 | SDValue *Low = nullptr) { | |||
| 27740 | unsigned NumElts = VT.getVectorNumElements(); | |||
| 27741 | ||||
| 27742 | // For vXi8 we will unpack the low and high half of each 128 bit lane to widen | |||
| 27743 | // to a vXi16 type. Do the multiplies, shift the results and pack the half | |||
| 27744 | // lane results back together. | |||
| 27745 | ||||
| 27746 | // We'll take different approaches for signed and unsigned. | |||
| 27747 | // For unsigned we'll use punpcklbw/punpckhbw to put zero extend the bytes | |||
| 27748 | // and use pmullw to calculate the full 16-bit product. | |||
| 27749 | // For signed we'll use punpcklbw/punpckbw to extend the bytes to words and | |||
| 27750 | // shift them left into the upper byte of each word. This allows us to use | |||
| 27751 | // pmulhw to calculate the full 16-bit product. This trick means we don't | |||
| 27752 | // need to sign extend the bytes to use pmullw. | |||
| 27753 | ||||
| 27754 | MVT ExVT = MVT::getVectorVT(MVT::i16, NumElts / 2); | |||
| 27755 | SDValue Zero = DAG.getConstant(0, dl, VT); | |||
| 27756 | ||||
| 27757 | SDValue ALo, AHi; | |||
| 27758 | if (IsSigned) { | |||
| 27759 | ALo = DAG.getBitcast(ExVT, getUnpackl(DAG, dl, VT, Zero, A)); | |||
| 27760 | AHi = DAG.getBitcast(ExVT, getUnpackh(DAG, dl, VT, Zero, A)); | |||
| 27761 | } else { | |||
| 27762 | ALo = DAG.getBitcast(ExVT, getUnpackl(DAG, dl, VT, A, Zero)); | |||
| 27763 | AHi = DAG.getBitcast(ExVT, getUnpackh(DAG, dl, VT, A, Zero)); | |||
| 27764 | } | |||
| 27765 | ||||
| 27766 | SDValue BLo, BHi; | |||
| 27767 | if (ISD::isBuildVectorOfConstantSDNodes(B.getNode())) { | |||
| 27768 | // If the RHS is a constant, manually unpackl/unpackh and extend. | |||
| 27769 | SmallVector<SDValue, 16> LoOps, HiOps; | |||
| 27770 | for (unsigned i = 0; i != NumElts; i += 16) { | |||
| 27771 | for (unsigned j = 0; j != 8; ++j) { | |||
| 27772 | SDValue LoOp = B.getOperand(i + j); | |||
| 27773 | SDValue HiOp = B.getOperand(i + j + 8); | |||
| 27774 | ||||
| 27775 | if (IsSigned) { | |||
| 27776 | LoOp = DAG.getAnyExtOrTrunc(LoOp, dl, MVT::i16); | |||
| 27777 | HiOp = DAG.getAnyExtOrTrunc(HiOp, dl, MVT::i16); | |||
| 27778 | LoOp = DAG.getNode(ISD::SHL, dl, MVT::i16, LoOp, | |||
| 27779 | DAG.getConstant(8, dl, MVT::i16)); | |||
| 27780 | HiOp = DAG.getNode(ISD::SHL, dl, MVT::i16, HiOp, | |||
| 27781 | DAG.getConstant(8, dl, MVT::i16)); | |||
| 27782 | } else { | |||
| 27783 | LoOp = DAG.getZExtOrTrunc(LoOp, dl, MVT::i16); | |||
| 27784 | HiOp = DAG.getZExtOrTrunc(HiOp, dl, MVT::i16); | |||
| 27785 | } | |||
| 27786 | ||||
| 27787 | LoOps.push_back(LoOp); | |||
| 27788 | HiOps.push_back(HiOp); | |||
| 27789 | } | |||
| 27790 | } | |||
| 27791 | ||||
| 27792 | BLo = DAG.getBuildVector(ExVT, dl, LoOps); | |||
| 27793 | BHi = DAG.getBuildVector(ExVT, dl, HiOps); | |||
| 27794 | } else if (IsSigned) { | |||
| 27795 | BLo = DAG.getBitcast(ExVT, getUnpackl(DAG, dl, VT, Zero, B)); | |||
| 27796 | BHi = DAG.getBitcast(ExVT, getUnpackh(DAG, dl, VT, Zero, B)); | |||
| 27797 | } else { | |||
| 27798 | BLo = DAG.getBitcast(ExVT, getUnpackl(DAG, dl, VT, B, Zero)); | |||
| 27799 | BHi = DAG.getBitcast(ExVT, getUnpackh(DAG, dl, VT, B, Zero)); | |||
| 27800 | } | |||
| 27801 | ||||
| 27802 | // Multiply, lshr the upper 8bits to the lower 8bits of the lo/hi results and | |||
| 27803 | // pack back to vXi8. | |||
| 27804 | unsigned MulOpc = IsSigned ? ISD::MULHS : ISD::MUL; | |||
| 27805 | SDValue RLo = DAG.getNode(MulOpc, dl, ExVT, ALo, BLo); | |||
| 27806 | SDValue RHi = DAG.getNode(MulOpc, dl, ExVT, AHi, BHi); | |||
| 27807 | ||||
| 27808 | if (Low) { | |||
| 27809 | // Mask the lower bits and pack the results to rejoin the halves. | |||
| 27810 | SDValue Mask = DAG.getConstant(255, dl, ExVT); | |||
| 27811 | SDValue LLo = DAG.getNode(ISD::AND, dl, ExVT, RLo, Mask); | |||
| 27812 | SDValue LHi = DAG.getNode(ISD::AND, dl, ExVT, RHi, Mask); | |||
| 27813 | *Low = DAG.getNode(X86ISD::PACKUS, dl, VT, LLo, LHi); | |||
| 27814 | } | |||
| 27815 | ||||
| 27816 | RLo = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, ExVT, RLo, 8, DAG); | |||
| 27817 | RHi = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, ExVT, RHi, 8, DAG); | |||
| 27818 | ||||
| 27819 | // Bitcast back to VT and then pack all the even elements from Lo and Hi. | |||
| 27820 | return DAG.getNode(X86ISD::PACKUS, dl, VT, RLo, RHi); | |||
| 27821 | } | |||
| 27822 | ||||
| 27823 | static SDValue LowerMULH(SDValue Op, const X86Subtarget &Subtarget, | |||
| 27824 | SelectionDAG &DAG) { | |||
| 27825 | SDLoc dl(Op); | |||
| 27826 | MVT VT = Op.getSimpleValueType(); | |||
| 27827 | bool IsSigned = Op->getOpcode() == ISD::MULHS; | |||
| 27828 | unsigned NumElts = VT.getVectorNumElements(); | |||
| 27829 | SDValue A = Op.getOperand(0); | |||
| 27830 | SDValue B = Op.getOperand(1); | |||
| 27831 | ||||
| 27832 | // Decompose 256-bit ops into 128-bit ops. | |||
| 27833 | if (VT.is256BitVector() && !Subtarget.hasInt256()) | |||
| 27834 | return splitVectorIntBinary(Op, DAG); | |||
| 27835 | ||||
| 27836 | if ((VT == MVT::v32i16 || VT == MVT::v64i8) && !Subtarget.hasBWI()) | |||
| 27837 | return splitVectorIntBinary(Op, DAG); | |||
| 27838 | ||||
| 27839 | if (VT == MVT::v4i32 || VT == MVT::v8i32 || VT == MVT::v16i32) { | |||
| 27840 | assert((VT == MVT::v4i32 && Subtarget.hasSSE2()) ||((void)0) | |||
| 27841 | (VT == MVT::v8i32 && Subtarget.hasInt256()) ||((void)0) | |||
| 27842 | (VT == MVT::v16i32 && Subtarget.hasAVX512()))((void)0); | |||
| 27843 | ||||
| 27844 | // PMULxD operations multiply each even value (starting at 0) of LHS with | |||
| 27845 | // the related value of RHS and produce a widen result. | |||
| 27846 | // E.g., PMULUDQ <4 x i32> <a|b|c|d>, <4 x i32> <e|f|g|h> | |||
| 27847 | // => <2 x i64> <ae|cg> | |||
| 27848 | // | |||
| 27849 | // In other word, to have all the results, we need to perform two PMULxD: | |||
| 27850 | // 1. one with the even values. | |||
| 27851 | // 2. one with the odd values. | |||
| 27852 | // To achieve #2, with need to place the odd values at an even position. | |||
| 27853 | // | |||
| 27854 | // Place the odd value at an even position (basically, shift all values 1 | |||
| 27855 | // step to the left): | |||
| 27856 | const int Mask[] = {1, -1, 3, -1, 5, -1, 7, -1, | |||
| 27857 | 9, -1, 11, -1, 13, -1, 15, -1}; | |||
| 27858 | // <a|b|c|d> => <b|undef|d|undef> | |||
| 27859 | SDValue Odd0 = DAG.getVectorShuffle(VT, dl, A, A, | |||
| 27860 | makeArrayRef(&Mask[0], NumElts)); | |||
| 27861 | // <e|f|g|h> => <f|undef|h|undef> | |||
| 27862 | SDValue Odd1 = DAG.getVectorShuffle(VT, dl, B, B, | |||
| 27863 | makeArrayRef(&Mask[0], NumElts)); | |||
| 27864 | ||||
| 27865 | // Emit two multiplies, one for the lower 2 ints and one for the higher 2 | |||
| 27866 | // ints. | |||
| 27867 | MVT MulVT = MVT::getVectorVT(MVT::i64, NumElts / 2); | |||
| 27868 | unsigned Opcode = | |||
| 27869 | (IsSigned && Subtarget.hasSSE41()) ? X86ISD::PMULDQ : X86ISD::PMULUDQ; | |||
| 27870 | // PMULUDQ <4 x i32> <a|b|c|d>, <4 x i32> <e|f|g|h> | |||
| 27871 | // => <2 x i64> <ae|cg> | |||
| 27872 | SDValue Mul1 = DAG.getBitcast(VT, DAG.getNode(Opcode, dl, MulVT, | |||
| 27873 | DAG.getBitcast(MulVT, A), | |||
| 27874 | DAG.getBitcast(MulVT, B))); | |||
| 27875 | // PMULUDQ <4 x i32> <b|undef|d|undef>, <4 x i32> <f|undef|h|undef> | |||
| 27876 | // => <2 x i64> <bf|dh> | |||
| 27877 | SDValue Mul2 = DAG.getBitcast(VT, DAG.getNode(Opcode, dl, MulVT, | |||
| 27878 | DAG.getBitcast(MulVT, Odd0), | |||
| 27879 | DAG.getBitcast(MulVT, Odd1))); | |||
| 27880 | ||||
| 27881 | // Shuffle it back into the right order. | |||
| 27882 | SmallVector<int, 16> ShufMask(NumElts); | |||
| 27883 | for (int i = 0; i != (int)NumElts; ++i) | |||
| 27884 | ShufMask[i] = (i / 2) * 2 + ((i % 2) * NumElts) + 1; | |||
| 27885 | ||||
| 27886 | SDValue Res = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, ShufMask); | |||
| 27887 | ||||
| 27888 | // If we have a signed multiply but no PMULDQ fix up the result of an | |||
| 27889 | // unsigned multiply. | |||
| 27890 | if (IsSigned && !Subtarget.hasSSE41()) { | |||
| 27891 | SDValue Zero = DAG.getConstant(0, dl, VT); | |||
| 27892 | SDValue T1 = DAG.getNode(ISD::AND, dl, VT, | |||
| 27893 | DAG.getSetCC(dl, VT, Zero, A, ISD::SETGT), B); | |||
| 27894 | SDValue T2 = DAG.getNode(ISD::AND, dl, VT, | |||
| 27895 | DAG.getSetCC(dl, VT, Zero, B, ISD::SETGT), A); | |||
| 27896 | ||||
| 27897 | SDValue Fixup = DAG.getNode(ISD::ADD, dl, VT, T1, T2); | |||
| 27898 | Res = DAG.getNode(ISD::SUB, dl, VT, Res, Fixup); | |||
| 27899 | } | |||
| 27900 | ||||
| 27901 | return Res; | |||
| 27902 | } | |||
| 27903 | ||||
| 27904 | // Only i8 vectors should need custom lowering after this. | |||
| 27905 | assert((VT == MVT::v16i8 || (VT == MVT::v32i8 && Subtarget.hasInt256()) ||((void)0) | |||
| 27906 | (VT == MVT::v64i8 && Subtarget.hasBWI())) &&((void)0) | |||
| 27907 | "Unsupported vector type")((void)0); | |||
| 27908 | ||||
| 27909 | // Lower v16i8/v32i8 as extension to v8i16/v16i16 vector pairs, multiply, | |||
| 27910 | // logical shift down the upper half and pack back to i8. | |||
| 27911 | ||||
| 27912 | // With SSE41 we can use sign/zero extend, but for pre-SSE41 we unpack | |||
| 27913 | // and then ashr/lshr the upper bits down to the lower bits before multiply. | |||
| 27914 | ||||
| 27915 | if ((VT == MVT::v16i8 && Subtarget.hasInt256()) || | |||
| 27916 | (VT == MVT::v32i8 && Subtarget.canExtendTo512BW())) { | |||
| 27917 | MVT ExVT = MVT::getVectorVT(MVT::i16, NumElts); | |||
| 27918 | unsigned ExAVX = IsSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND; | |||
| 27919 | SDValue ExA = DAG.getNode(ExAVX, dl, ExVT, A); | |||
| 27920 | SDValue ExB = DAG.getNode(ExAVX, dl, ExVT, B); | |||
| 27921 | SDValue Mul = DAG.getNode(ISD::MUL, dl, ExVT, ExA, ExB); | |||
| 27922 | Mul = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, ExVT, Mul, 8, DAG); | |||
| 27923 | return DAG.getNode(ISD::TRUNCATE, dl, VT, Mul); | |||
| 27924 | } | |||
| 27925 | ||||
| 27926 | return LowervXi8MulWithUNPCK(A, B, dl, VT, IsSigned, Subtarget, DAG); | |||
| 27927 | } | |||
| 27928 | ||||
| 27929 | // Custom lowering for SMULO/UMULO. | |||
| 27930 | static SDValue LowerMULO(SDValue Op, const X86Subtarget &Subtarget, | |||
| 27931 | SelectionDAG &DAG) { | |||
| 27932 | MVT VT = Op.getSimpleValueType(); | |||
| 27933 | ||||
| 27934 | // Scalars defer to LowerXALUO. | |||
| 27935 | if (!VT.isVector()) | |||
| 27936 | return LowerXALUO(Op, DAG); | |||
| 27937 | ||||
| 27938 | SDLoc dl(Op); | |||
| 27939 | bool IsSigned = Op->getOpcode() == ISD::SMULO; | |||
| 27940 | SDValue A = Op.getOperand(0); | |||
| 27941 | SDValue B = Op.getOperand(1); | |||
| 27942 | EVT OvfVT = Op->getValueType(1); | |||
| 27943 | ||||
| 27944 | if ((VT == MVT::v32i8 && !Subtarget.hasInt256()) || | |||
| 27945 | (VT == MVT::v64i8 && !Subtarget.hasBWI())) { | |||
| 27946 | // Extract the LHS Lo/Hi vectors | |||
| 27947 | SDValue LHSLo, LHSHi; | |||
| 27948 | std::tie(LHSLo, LHSHi) = splitVector(A, DAG, dl); | |||
| 27949 | ||||
| 27950 | // Extract the RHS Lo/Hi vectors | |||
| 27951 | SDValue RHSLo, RHSHi; | |||
| 27952 | std::tie(RHSLo, RHSHi) = splitVector(B, DAG, dl); | |||
| 27953 | ||||
| 27954 | EVT LoOvfVT, HiOvfVT; | |||
| 27955 | std::tie(LoOvfVT, HiOvfVT) = DAG.GetSplitDestVTs(OvfVT); | |||
| 27956 | SDVTList LoVTs = DAG.getVTList(LHSLo.getValueType(), LoOvfVT); | |||
| 27957 | SDVTList HiVTs = DAG.getVTList(LHSHi.getValueType(), HiOvfVT); | |||
| 27958 | ||||
| 27959 | // Issue the split operations. | |||
| 27960 | SDValue Lo = DAG.getNode(Op.getOpcode(), dl, LoVTs, LHSLo, RHSLo); | |||
| 27961 | SDValue Hi = DAG.getNode(Op.getOpcode(), dl, HiVTs, LHSHi, RHSHi); | |||
| 27962 | ||||
| 27963 | // Join the separate data results and the overflow results. | |||
| 27964 | SDValue Res = DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, Lo, Hi); | |||
| 27965 | SDValue Ovf = DAG.getNode(ISD::CONCAT_VECTORS, dl, OvfVT, Lo.getValue(1), | |||
| 27966 | Hi.getValue(1)); | |||
| 27967 | ||||
| 27968 | return DAG.getMergeValues({Res, Ovf}, dl); | |||
| 27969 | } | |||
| 27970 | ||||
| 27971 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
| 27972 | EVT SetccVT = | |||
| 27973 | TLI.getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT); | |||
| 27974 | ||||
| 27975 | if ((VT == MVT::v16i8 && Subtarget.hasInt256()) || | |||
| 27976 | (VT == MVT::v32i8 && Subtarget.canExtendTo512BW())) { | |||
| 27977 | unsigned NumElts = VT.getVectorNumElements(); | |||
| 27978 | MVT ExVT = MVT::getVectorVT(MVT::i16, NumElts); | |||
| 27979 | unsigned ExAVX = IsSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND; | |||
| 27980 | SDValue ExA = DAG.getNode(ExAVX, dl, ExVT, A); | |||
| 27981 | SDValue ExB = DAG.getNode(ExAVX, dl, ExVT, B); | |||
| 27982 | SDValue Mul = DAG.getNode(ISD::MUL, dl, ExVT, ExA, ExB); | |||
| 27983 | ||||
| 27984 | SDValue Low = DAG.getNode(ISD::TRUNCATE, dl, VT, Mul); | |||
| 27985 | ||||
| 27986 | SDValue Ovf; | |||
| 27987 | if (IsSigned) { | |||
| 27988 | SDValue High, LowSign; | |||
| 27989 | if (OvfVT.getVectorElementType() == MVT::i1 && | |||
| 27990 | (Subtarget.hasBWI() || Subtarget.canExtendTo512DQ())) { | |||
| 27991 | // Rather the truncating try to do the compare on vXi16 or vXi32. | |||
| 27992 | // Shift the high down filling with sign bits. | |||
| 27993 | High = getTargetVShiftByConstNode(X86ISD::VSRAI, dl, ExVT, Mul, 8, DAG); | |||
| 27994 | // Fill all 16 bits with the sign bit from the low. | |||
| 27995 | LowSign = | |||
| 27996 | getTargetVShiftByConstNode(X86ISD::VSHLI, dl, ExVT, Mul, 8, DAG); | |||
| 27997 | LowSign = getTargetVShiftByConstNode(X86ISD::VSRAI, dl, ExVT, LowSign, | |||
| 27998 | 15, DAG); | |||
| 27999 | SetccVT = OvfVT; | |||
| 28000 | if (!Subtarget.hasBWI()) { | |||
| 28001 | // We can't do a vXi16 compare so sign extend to v16i32. | |||
| 28002 | High = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v16i32, High); | |||
| 28003 | LowSign = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v16i32, LowSign); | |||
| 28004 | } | |||
| 28005 | } else { | |||
| 28006 | // Otherwise do the compare at vXi8. | |||
| 28007 | High = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, ExVT, Mul, 8, DAG); | |||
| 28008 | High = DAG.getNode(ISD::TRUNCATE, dl, VT, High); | |||
| 28009 | LowSign = | |||
| 28010 | DAG.getNode(ISD::SRA, dl, VT, Low, DAG.getConstant(7, dl, VT)); | |||
| 28011 | } | |||
| 28012 | ||||
| 28013 | Ovf = DAG.getSetCC(dl, SetccVT, LowSign, High, ISD::SETNE); | |||
| 28014 | } else { | |||
| 28015 | SDValue High = | |||
| 28016 | getTargetVShiftByConstNode(X86ISD::VSRLI, dl, ExVT, Mul, 8, DAG); | |||
| 28017 | if (OvfVT.getVectorElementType() == MVT::i1 && | |||
| 28018 | (Subtarget.hasBWI() || Subtarget.canExtendTo512DQ())) { | |||
| 28019 | // Rather the truncating try to do the compare on vXi16 or vXi32. | |||
| 28020 | SetccVT = OvfVT; | |||
| 28021 | if (!Subtarget.hasBWI()) { | |||
| 28022 | // We can't do a vXi16 compare so sign extend to v16i32. | |||
| 28023 | High = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v16i32, High); | |||
| 28024 | } | |||
| 28025 | } else { | |||
| 28026 | // Otherwise do the compare at vXi8. | |||
| 28027 | High = DAG.getNode(ISD::TRUNCATE, dl, VT, High); | |||
| 28028 | } | |||
| 28029 | ||||
| 28030 | Ovf = | |||
| 28031 | DAG.getSetCC(dl, SetccVT, High, | |||
| 28032 | DAG.getConstant(0, dl, High.getValueType()), ISD::SETNE); | |||
| 28033 | } | |||
| 28034 | ||||
| 28035 | Ovf = DAG.getSExtOrTrunc(Ovf, dl, OvfVT); | |||
| 28036 | ||||
| 28037 | return DAG.getMergeValues({Low, Ovf}, dl); | |||
| 28038 | } | |||
| 28039 | ||||
| 28040 | SDValue Low; | |||
| 28041 | SDValue High = | |||
| 28042 | LowervXi8MulWithUNPCK(A, B, dl, VT, IsSigned, Subtarget, DAG, &Low); | |||
| 28043 | ||||
| 28044 | SDValue Ovf; | |||
| 28045 | if (IsSigned) { | |||
| 28046 | // SMULO overflows if the high bits don't match the sign of the low. | |||
| 28047 | SDValue LowSign = | |||
| 28048 | DAG.getNode(ISD::SRA, dl, VT, Low, DAG.getConstant(7, dl, VT)); | |||
| 28049 | Ovf = DAG.getSetCC(dl, SetccVT, LowSign, High, ISD::SETNE); | |||
| 28050 | } else { | |||
| 28051 | // UMULO overflows if the high bits are non-zero. | |||
| 28052 | Ovf = | |||
| 28053 | DAG.getSetCC(dl, SetccVT, High, DAG.getConstant(0, dl, VT), ISD::SETNE); | |||
| 28054 | } | |||
| 28055 | ||||
| 28056 | Ovf = DAG.getSExtOrTrunc(Ovf, dl, OvfVT); | |||
| 28057 | ||||
| 28058 | return DAG.getMergeValues({Low, Ovf}, dl); | |||
| 28059 | } | |||
| 28060 | ||||
| 28061 | SDValue X86TargetLowering::LowerWin64_i128OP(SDValue Op, SelectionDAG &DAG) const { | |||
| 28062 | assert(Subtarget.isTargetWin64() && "Unexpected target")((void)0); | |||
| 28063 | EVT VT = Op.getValueType(); | |||
| 28064 | assert(VT.isInteger() && VT.getSizeInBits() == 128 &&((void)0) | |||
| 28065 | "Unexpected return type for lowering")((void)0); | |||
| 28066 | ||||
| 28067 | RTLIB::Libcall LC; | |||
| 28068 | bool isSigned; | |||
| 28069 | switch (Op->getOpcode()) { | |||
| 28070 | default: llvm_unreachable("Unexpected request for libcall!")__builtin_unreachable(); | |||
| 28071 | case ISD::SDIV: isSigned = true; LC = RTLIB::SDIV_I128; break; | |||
| 28072 | case ISD::UDIV: isSigned = false; LC = RTLIB::UDIV_I128; break; | |||
| 28073 | case ISD::SREM: isSigned = true; LC = RTLIB::SREM_I128; break; | |||
| 28074 | case ISD::UREM: isSigned = false; LC = RTLIB::UREM_I128; break; | |||
| 28075 | } | |||
| 28076 | ||||
| 28077 | SDLoc dl(Op); | |||
| 28078 | SDValue InChain = DAG.getEntryNode(); | |||
| 28079 | ||||
| 28080 | TargetLowering::ArgListTy Args; | |||
| 28081 | TargetLowering::ArgListEntry Entry; | |||
| 28082 | for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i) { | |||
| 28083 | EVT ArgVT = Op->getOperand(i).getValueType(); | |||
| 28084 | assert(ArgVT.isInteger() && ArgVT.getSizeInBits() == 128 &&((void)0) | |||
| 28085 | "Unexpected argument type for lowering")((void)0); | |||
| 28086 | SDValue StackPtr = DAG.CreateStackTemporary(ArgVT, 16); | |||
| 28087 | int SPFI = cast<FrameIndexSDNode>(StackPtr.getNode())->getIndex(); | |||
| 28088 | MachinePointerInfo MPI = | |||
| 28089 | MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SPFI); | |||
| 28090 | Entry.Node = StackPtr; | |||
| 28091 | InChain = | |||
| 28092 | DAG.getStore(InChain, dl, Op->getOperand(i), StackPtr, MPI, Align(16)); | |||
| 28093 | Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext()); | |||
| 28094 | Entry.Ty = PointerType::get(ArgTy,0); | |||
| 28095 | Entry.IsSExt = false; | |||
| 28096 | Entry.IsZExt = false; | |||
| 28097 | Args.push_back(Entry); | |||
| 28098 | } | |||
| 28099 | ||||
| 28100 | SDValue Callee = DAG.getExternalSymbol(getLibcallName(LC), | |||
| 28101 | getPointerTy(DAG.getDataLayout())); | |||
| 28102 | ||||
| 28103 | TargetLowering::CallLoweringInfo CLI(DAG); | |||
| 28104 | CLI.setDebugLoc(dl) | |||
| 28105 | .setChain(InChain) | |||
| 28106 | .setLibCallee( | |||
| 28107 | getLibcallCallingConv(LC), | |||
| 28108 | static_cast<EVT>(MVT::v2i64).getTypeForEVT(*DAG.getContext()), Callee, | |||
| 28109 | std::move(Args)) | |||
| 28110 | .setInRegister() | |||
| 28111 | .setSExtResult(isSigned) | |||
| 28112 | .setZExtResult(!isSigned); | |||
| 28113 | ||||
| 28114 | std::pair<SDValue, SDValue> CallInfo = LowerCallTo(CLI); | |||
| 28115 | return DAG.getBitcast(VT, CallInfo.first); | |||
| 28116 | } | |||
| 28117 | ||||
| 28118 | // Return true if the required (according to Opcode) shift-imm form is natively | |||
| 28119 | // supported by the Subtarget | |||
| 28120 | static bool SupportedVectorShiftWithImm(MVT VT, const X86Subtarget &Subtarget, | |||
| 28121 | unsigned Opcode) { | |||
| 28122 | if (VT.getScalarSizeInBits() < 16) | |||
| 28123 | return false; | |||
| 28124 | ||||
| 28125 | if (VT.is512BitVector() && Subtarget.hasAVX512() && | |||
| 28126 | (VT.getScalarSizeInBits() > 16 || Subtarget.hasBWI())) | |||
| 28127 | return true; | |||
| 28128 | ||||
| 28129 | bool LShift = (VT.is128BitVector() && Subtarget.hasSSE2()) || | |||
| 28130 | (VT.is256BitVector() && Subtarget.hasInt256()); | |||
| 28131 | ||||
| 28132 | bool AShift = LShift && (Subtarget.hasAVX512() || | |||
| 28133 | (VT != MVT::v2i64 && VT != MVT::v4i64)); | |||
| 28134 | return (Opcode == ISD::SRA) ? AShift : LShift; | |||
| 28135 | } | |||
| 28136 | ||||
| 28137 | // The shift amount is a variable, but it is the same for all vector lanes. | |||
| 28138 | // These instructions are defined together with shift-immediate. | |||
| 28139 | static | |||
| 28140 | bool SupportedVectorShiftWithBaseAmnt(MVT VT, const X86Subtarget &Subtarget, | |||
| 28141 | unsigned Opcode) { | |||
| 28142 | return SupportedVectorShiftWithImm(VT, Subtarget, Opcode); | |||
| 28143 | } | |||
| 28144 | ||||
| 28145 | // Return true if the required (according to Opcode) variable-shift form is | |||
| 28146 | // natively supported by the Subtarget | |||
| 28147 | static bool SupportedVectorVarShift(MVT VT, const X86Subtarget &Subtarget, | |||
| 28148 | unsigned Opcode) { | |||
| 28149 | ||||
| 28150 | if (!Subtarget.hasInt256() || VT.getScalarSizeInBits() < 16) | |||
| 28151 | return false; | |||
| 28152 | ||||
| 28153 | // vXi16 supported only on AVX-512, BWI | |||
| 28154 | if (VT.getScalarSizeInBits() == 16 && !Subtarget.hasBWI()) | |||
| 28155 | return false; | |||
| 28156 | ||||
| 28157 | if (Subtarget.hasAVX512()) | |||
| 28158 | return true; | |||
| 28159 | ||||
| 28160 | bool LShift = VT.is128BitVector() || VT.is256BitVector(); | |||
| 28161 | bool AShift = LShift && VT != MVT::v2i64 && VT != MVT::v4i64; | |||
| 28162 | return (Opcode == ISD::SRA) ? AShift : LShift; | |||
| 28163 | } | |||
| 28164 | ||||
| 28165 | static SDValue LowerScalarImmediateShift(SDValue Op, SelectionDAG &DAG, | |||
| 28166 | const X86Subtarget &Subtarget) { | |||
| 28167 | MVT VT = Op.getSimpleValueType(); | |||
| 28168 | SDLoc dl(Op); | |||
| 28169 | SDValue R = Op.getOperand(0); | |||
| 28170 | SDValue Amt = Op.getOperand(1); | |||
| 28171 | unsigned X86Opc = getTargetVShiftUniformOpcode(Op.getOpcode(), false); | |||
| 28172 | ||||
| 28173 | auto ArithmeticShiftRight64 = [&](uint64_t ShiftAmt) { | |||
| 28174 | assert((VT == MVT::v2i64 || VT == MVT::v4i64) && "Unexpected SRA type")((void)0); | |||
| 28175 | MVT ExVT = MVT::getVectorVT(MVT::i32, VT.getVectorNumElements() * 2); | |||
| 28176 | SDValue Ex = DAG.getBitcast(ExVT, R); | |||
| 28177 | ||||
| 28178 | // ashr(R, 63) === cmp_slt(R, 0) | |||
| 28179 | if (ShiftAmt == 63 && Subtarget.hasSSE42()) { | |||
| 28180 | assert((VT != MVT::v4i64 || Subtarget.hasInt256()) &&((void)0) | |||
| 28181 | "Unsupported PCMPGT op")((void)0); | |||
| 28182 | return DAG.getNode(X86ISD::PCMPGT, dl, VT, DAG.getConstant(0, dl, VT), R); | |||
| 28183 | } | |||
| 28184 | ||||
| 28185 | if (ShiftAmt >= 32) { | |||
| 28186 | // Splat sign to upper i32 dst, and SRA upper i32 src to lower i32. | |||
| 28187 | SDValue Upper = | |||
| 28188 | getTargetVShiftByConstNode(X86ISD::VSRAI, dl, ExVT, Ex, 31, DAG); | |||
| 28189 | SDValue Lower = getTargetVShiftByConstNode(X86ISD::VSRAI, dl, ExVT, Ex, | |||
| 28190 | ShiftAmt - 32, DAG); | |||
| 28191 | if (VT == MVT::v2i64) | |||
| 28192 | Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower, {5, 1, 7, 3}); | |||
| 28193 | if (VT == MVT::v4i64) | |||
| 28194 | Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower, | |||
| 28195 | {9, 1, 11, 3, 13, 5, 15, 7}); | |||
| 28196 | } else { | |||
| 28197 | // SRA upper i32, SRL whole i64 and select lower i32. | |||
| 28198 | SDValue Upper = getTargetVShiftByConstNode(X86ISD::VSRAI, dl, ExVT, Ex, | |||
| 28199 | ShiftAmt, DAG); | |||
| 28200 | SDValue Lower = | |||
| 28201 | getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, R, ShiftAmt, DAG); | |||
| 28202 | Lower = DAG.getBitcast(ExVT, Lower); | |||
| 28203 | if (VT == MVT::v2i64) | |||
| 28204 | Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower, {4, 1, 6, 3}); | |||
| 28205 | if (VT == MVT::v4i64) | |||
| 28206 | Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower, | |||
| 28207 | {8, 1, 10, 3, 12, 5, 14, 7}); | |||
| 28208 | } | |||
| 28209 | return DAG.getBitcast(VT, Ex); | |||
| 28210 | }; | |||
| 28211 | ||||
| 28212 | // Optimize shl/srl/sra with constant shift amount. | |||
| 28213 | APInt APIntShiftAmt; | |||
| 28214 | if (!X86::isConstantSplat(Amt, APIntShiftAmt)) | |||
| 28215 | return SDValue(); | |||
| 28216 | ||||
| 28217 | // If the shift amount is out of range, return undef. | |||
| 28218 | if (APIntShiftAmt.uge(VT.getScalarSizeInBits())) | |||
| 28219 | return DAG.getUNDEF(VT); | |||
| 28220 | ||||
| 28221 | uint64_t ShiftAmt = APIntShiftAmt.getZExtValue(); | |||
| 28222 | ||||
| 28223 | if (SupportedVectorShiftWithImm(VT, Subtarget, Op.getOpcode())) | |||
| 28224 | return getTargetVShiftByConstNode(X86Opc, dl, VT, R, ShiftAmt, DAG); | |||
| 28225 | ||||
| 28226 | // i64 SRA needs to be performed as partial shifts. | |||
| 28227 | if (((!Subtarget.hasXOP() && VT == MVT::v2i64) || | |||
| 28228 | (Subtarget.hasInt256() && VT == MVT::v4i64)) && | |||
| 28229 | Op.getOpcode() == ISD::SRA) | |||
| 28230 | return ArithmeticShiftRight64(ShiftAmt); | |||
| 28231 | ||||
| 28232 | if (VT == MVT::v16i8 || (Subtarget.hasInt256() && VT == MVT::v32i8) || | |||
| 28233 | (Subtarget.hasBWI() && VT == MVT::v64i8)) { | |||
| 28234 | unsigned NumElts = VT.getVectorNumElements(); | |||
| 28235 | MVT ShiftVT = MVT::getVectorVT(MVT::i16, NumElts / 2); | |||
| 28236 | ||||
| 28237 | // Simple i8 add case | |||
| 28238 | if (Op.getOpcode() == ISD::SHL && ShiftAmt == 1) | |||
| 28239 | return DAG.getNode(ISD::ADD, dl, VT, R, R); | |||
| 28240 | ||||
| 28241 | // ashr(R, 7) === cmp_slt(R, 0) | |||
| 28242 | if (Op.getOpcode() == ISD::SRA && ShiftAmt == 7) { | |||
| 28243 | SDValue Zeros = DAG.getConstant(0, dl, VT); | |||
| 28244 | if (VT.is512BitVector()) { | |||
| 28245 | assert(VT == MVT::v64i8 && "Unexpected element type!")((void)0); | |||
| 28246 | SDValue CMP = DAG.getSetCC(dl, MVT::v64i1, Zeros, R, ISD::SETGT); | |||
| 28247 | return DAG.getNode(ISD::SIGN_EXTEND, dl, VT, CMP); | |||
| 28248 | } | |||
| 28249 | return DAG.getNode(X86ISD::PCMPGT, dl, VT, Zeros, R); | |||
| 28250 | } | |||
| 28251 | ||||
| 28252 | // XOP can shift v16i8 directly instead of as shift v8i16 + mask. | |||
| 28253 | if (VT == MVT::v16i8 && Subtarget.hasXOP()) | |||
| 28254 | return SDValue(); | |||
| 28255 | ||||
| 28256 | if (Op.getOpcode() == ISD::SHL) { | |||
| 28257 | // Make a large shift. | |||
| 28258 | SDValue SHL = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, ShiftVT, R, | |||
| 28259 | ShiftAmt, DAG); | |||
| 28260 | SHL = DAG.getBitcast(VT, SHL); | |||
| 28261 | // Zero out the rightmost bits. | |||
| 28262 | APInt Mask = APInt::getHighBitsSet(8, 8 - ShiftAmt); | |||
| 28263 | return DAG.getNode(ISD::AND, dl, VT, SHL, DAG.getConstant(Mask, dl, VT)); | |||
| 28264 | } | |||
| 28265 | if (Op.getOpcode() == ISD::SRL) { | |||
| 28266 | // Make a large shift. | |||
| 28267 | SDValue SRL = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, ShiftVT, R, | |||
| 28268 | ShiftAmt, DAG); | |||
| 28269 | SRL = DAG.getBitcast(VT, SRL); | |||
| 28270 | // Zero out the leftmost bits. | |||
| 28271 | APInt Mask = APInt::getLowBitsSet(8, 8 - ShiftAmt); | |||
| 28272 | return DAG.getNode(ISD::AND, dl, VT, SRL, DAG.getConstant(Mask, dl, VT)); | |||
| 28273 | } | |||
| 28274 | if (Op.getOpcode() == ISD::SRA) { | |||
| 28275 | // ashr(R, Amt) === sub(xor(lshr(R, Amt), Mask), Mask) | |||
| 28276 | SDValue Res = DAG.getNode(ISD::SRL, dl, VT, R, Amt); | |||
| 28277 | ||||
| 28278 | SDValue Mask = DAG.getConstant(128 >> ShiftAmt, dl, VT); | |||
| 28279 | Res = DAG.getNode(ISD::XOR, dl, VT, Res, Mask); | |||
| 28280 | Res = DAG.getNode(ISD::SUB, dl, VT, Res, Mask); | |||
| 28281 | return Res; | |||
| 28282 | } | |||
| 28283 | llvm_unreachable("Unknown shift opcode.")__builtin_unreachable(); | |||
| 28284 | } | |||
| 28285 | ||||
| 28286 | return SDValue(); | |||
| 28287 | } | |||
| 28288 | ||||
| 28289 | static SDValue LowerScalarVariableShift(SDValue Op, SelectionDAG &DAG, | |||
| 28290 | const X86Subtarget &Subtarget) { | |||
| 28291 | MVT VT = Op.getSimpleValueType(); | |||
| 28292 | SDLoc dl(Op); | |||
| 28293 | SDValue R = Op.getOperand(0); | |||
| 28294 | SDValue Amt = Op.getOperand(1); | |||
| 28295 | unsigned Opcode = Op.getOpcode(); | |||
| 28296 | unsigned X86OpcI = getTargetVShiftUniformOpcode(Opcode, false); | |||
| 28297 | unsigned X86OpcV = getTargetVShiftUniformOpcode(Opcode, true); | |||
| 28298 | ||||
| 28299 | if (SDValue BaseShAmt = DAG.getSplatValue(Amt)) { | |||
| 28300 | if (SupportedVectorShiftWithBaseAmnt(VT, Subtarget, Opcode)) { | |||
| 28301 | MVT EltVT = VT.getVectorElementType(); | |||
| 28302 | assert(EltVT.bitsLE(MVT::i64) && "Unexpected element type!")((void)0); | |||
| 28303 | if (EltVT != MVT::i64 && EltVT.bitsGT(MVT::i32)) | |||
| 28304 | BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i64, BaseShAmt); | |||
| 28305 | else if (EltVT.bitsLT(MVT::i32)) | |||
| 28306 | BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, BaseShAmt); | |||
| 28307 | ||||
| 28308 | return getTargetVShiftNode(X86OpcI, dl, VT, R, BaseShAmt, Subtarget, DAG); | |||
| 28309 | } | |||
| 28310 | ||||
| 28311 | // vXi8 shifts - shift as v8i16 + mask result. | |||
| 28312 | if (((VT == MVT::v16i8 && !Subtarget.canExtendTo512DQ()) || | |||
| 28313 | (VT == MVT::v32i8 && !Subtarget.canExtendTo512BW()) || | |||
| 28314 | VT == MVT::v64i8) && | |||
| 28315 | !Subtarget.hasXOP()) { | |||
| 28316 | unsigned NumElts = VT.getVectorNumElements(); | |||
| 28317 | MVT ExtVT = MVT::getVectorVT(MVT::i16, NumElts / 2); | |||
| 28318 | if (SupportedVectorShiftWithBaseAmnt(ExtVT, Subtarget, Opcode)) { | |||
| 28319 | unsigned LogicalOp = (Opcode == ISD::SHL ? ISD::SHL : ISD::SRL); | |||
| 28320 | unsigned LogicalX86Op = getTargetVShiftUniformOpcode(LogicalOp, false); | |||
| 28321 | BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, BaseShAmt); | |||
| 28322 | ||||
| 28323 | // Create the mask using vXi16 shifts. For shift-rights we need to move | |||
| 28324 | // the upper byte down before splatting the vXi8 mask. | |||
| 28325 | SDValue BitMask = DAG.getConstant(-1, dl, ExtVT); | |||
| 28326 | BitMask = getTargetVShiftNode(LogicalX86Op, dl, ExtVT, BitMask, | |||
| 28327 | BaseShAmt, Subtarget, DAG); | |||
| 28328 | if (Opcode != ISD::SHL) | |||
| 28329 | BitMask = getTargetVShiftByConstNode(LogicalX86Op, dl, ExtVT, BitMask, | |||
| 28330 | 8, DAG); | |||
| 28331 | BitMask = DAG.getBitcast(VT, BitMask); | |||
| 28332 | BitMask = DAG.getVectorShuffle(VT, dl, BitMask, BitMask, | |||
| 28333 | SmallVector<int, 64>(NumElts, 0)); | |||
| 28334 | ||||
| 28335 | SDValue Res = getTargetVShiftNode(LogicalX86Op, dl, ExtVT, | |||
| 28336 | DAG.getBitcast(ExtVT, R), BaseShAmt, | |||
| 28337 | Subtarget, DAG); | |||
| 28338 | Res = DAG.getBitcast(VT, Res); | |||
| 28339 | Res = DAG.getNode(ISD::AND, dl, VT, Res, BitMask); | |||
| 28340 | ||||
| 28341 | if (Opcode == ISD::SRA) { | |||
| 28342 | // ashr(R, Amt) === sub(xor(lshr(R, Amt), SignMask), SignMask) | |||
| 28343 | // SignMask = lshr(SignBit, Amt) - safe to do this with PSRLW. | |||
| 28344 | SDValue SignMask = DAG.getConstant(0x8080, dl, ExtVT); | |||
| 28345 | SignMask = getTargetVShiftNode(LogicalX86Op, dl, ExtVT, SignMask, | |||
| 28346 | BaseShAmt, Subtarget, DAG); | |||
| 28347 | SignMask = DAG.getBitcast(VT, SignMask); | |||
| 28348 | Res = DAG.getNode(ISD::XOR, dl, VT, Res, SignMask); | |||
| 28349 | Res = DAG.getNode(ISD::SUB, dl, VT, Res, SignMask); | |||
| 28350 | } | |||
| 28351 | return Res; | |||
| 28352 | } | |||
| 28353 | } | |||
| 28354 | } | |||
| 28355 | ||||
| 28356 | // Check cases (mainly 32-bit) where i64 is expanded into high and low parts. | |||
| 28357 | if (VT == MVT::v2i64 && Amt.getOpcode() == ISD::BITCAST && | |||
| 28358 | Amt.getOperand(0).getOpcode() == ISD::BUILD_VECTOR) { | |||
| 28359 | Amt = Amt.getOperand(0); | |||
| 28360 | unsigned Ratio = 64 / Amt.getScalarValueSizeInBits(); | |||
| 28361 | std::vector<SDValue> Vals(Ratio); | |||
| 28362 | for (unsigned i = 0; i != Ratio; ++i) | |||
| 28363 | Vals[i] = Amt.getOperand(i); | |||
| 28364 | for (unsigned i = Ratio, e = Amt.getNumOperands(); i != e; i += Ratio) { | |||
| 28365 | for (unsigned j = 0; j != Ratio; ++j) | |||
| 28366 | if (Vals[j] != Amt.getOperand(i + j)) | |||
| 28367 | return SDValue(); | |||
| 28368 | } | |||
| 28369 | ||||
| 28370 | if (SupportedVectorShiftWithBaseAmnt(VT, Subtarget, Op.getOpcode())) | |||
| 28371 | return DAG.getNode(X86OpcV, dl, VT, R, Op.getOperand(1)); | |||
| 28372 | } | |||
| 28373 | return SDValue(); | |||
| 28374 | } | |||
| 28375 | ||||
| 28376 | // Convert a shift/rotate left amount to a multiplication scale factor. | |||
| 28377 | static SDValue convertShiftLeftToScale(SDValue Amt, const SDLoc &dl, | |||
| 28378 | const X86Subtarget &Subtarget, | |||
| 28379 | SelectionDAG &DAG) { | |||
| 28380 | MVT VT = Amt.getSimpleValueType(); | |||
| 28381 | if (!(VT == MVT::v8i16 || VT == MVT::v4i32 || | |||
| 28382 | (Subtarget.hasInt256() && VT == MVT::v16i16) || | |||
| 28383 | (Subtarget.hasVBMI2() && VT == MVT::v32i16) || | |||
| 28384 | (!Subtarget.hasAVX512() && VT == MVT::v16i8))) | |||
| 28385 | return SDValue(); | |||
| 28386 | ||||
| 28387 | if (ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) { | |||
| 28388 | SmallVector<SDValue, 8> Elts; | |||
| 28389 | MVT SVT = VT.getVectorElementType(); | |||
| 28390 | unsigned SVTBits = SVT.getSizeInBits(); | |||
| 28391 | APInt One(SVTBits, 1); | |||
| 28392 | unsigned NumElems = VT.getVectorNumElements(); | |||
| 28393 | ||||
| 28394 | for (unsigned i = 0; i != NumElems; ++i) { | |||
| 28395 | SDValue Op = Amt->getOperand(i); | |||
| 28396 | if (Op->isUndef()) { | |||
| 28397 | Elts.push_back(Op); | |||
| 28398 | continue; | |||
| 28399 | } | |||
| 28400 | ||||
| 28401 | ConstantSDNode *ND = cast<ConstantSDNode>(Op); | |||
| 28402 | APInt C(SVTBits, ND->getZExtValue()); | |||
| 28403 | uint64_t ShAmt = C.getZExtValue(); | |||
| 28404 | if (ShAmt >= SVTBits) { | |||
| 28405 | Elts.push_back(DAG.getUNDEF(SVT)); | |||
| 28406 | continue; | |||
| 28407 | } | |||
| 28408 | Elts.push_back(DAG.getConstant(One.shl(ShAmt), dl, SVT)); | |||
| 28409 | } | |||
| 28410 | return DAG.getBuildVector(VT, dl, Elts); | |||
| 28411 | } | |||
| 28412 | ||||
| 28413 | // If the target doesn't support variable shifts, use either FP conversion | |||
| 28414 | // or integer multiplication to avoid shifting each element individually. | |||
| 28415 | if (VT == MVT::v4i32) { | |||
| 28416 | Amt = DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(23, dl, VT)); | |||
| 28417 | Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, | |||
| 28418 | DAG.getConstant(0x3f800000U, dl, VT)); | |||
| 28419 | Amt = DAG.getBitcast(MVT::v4f32, Amt); | |||
| 28420 | return DAG.getNode(ISD::FP_TO_SINT, dl, VT, Amt); | |||
| 28421 | } | |||
| 28422 | ||||
| 28423 | // AVX2 can more effectively perform this as a zext/trunc to/from v8i32. | |||
| 28424 | if (VT == MVT::v8i16 && !Subtarget.hasAVX2()) { | |||
| 28425 | SDValue Z = DAG.getConstant(0, dl, VT); | |||
| 28426 | SDValue Lo = DAG.getBitcast(MVT::v4i32, getUnpackl(DAG, dl, VT, Amt, Z)); | |||
| 28427 | SDValue Hi = DAG.getBitcast(MVT::v4i32, getUnpackh(DAG, dl, VT, Amt, Z)); | |||
| 28428 | Lo = convertShiftLeftToScale(Lo, dl, Subtarget, DAG); | |||
| 28429 | Hi = convertShiftLeftToScale(Hi, dl, Subtarget, DAG); | |||
| 28430 | if (Subtarget.hasSSE41()) | |||
| 28431 | return DAG.getNode(X86ISD::PACKUS, dl, VT, Lo, Hi); | |||
| 28432 | ||||
| 28433 | return DAG.getVectorShuffle(VT, dl, DAG.getBitcast(VT, Lo), | |||
| 28434 | DAG.getBitcast(VT, Hi), | |||
| 28435 | {0, 2, 4, 6, 8, 10, 12, 14}); | |||
| 28436 | } | |||
| 28437 | ||||
| 28438 | return SDValue(); | |||
| 28439 | } | |||
| 28440 | ||||
| 28441 | static SDValue LowerShift(SDValue Op, const X86Subtarget &Subtarget, | |||
| 28442 | SelectionDAG &DAG) { | |||
| 28443 | MVT VT = Op.getSimpleValueType(); | |||
| 28444 | SDLoc dl(Op); | |||
| 28445 | SDValue R = Op.getOperand(0); | |||
| 28446 | SDValue Amt = Op.getOperand(1); | |||
| 28447 | unsigned EltSizeInBits = VT.getScalarSizeInBits(); | |||
| 28448 | bool ConstantAmt = ISD::isBuildVectorOfConstantSDNodes(Amt.getNode()); | |||
| 28449 | ||||
| 28450 | unsigned Opc = Op.getOpcode(); | |||
| 28451 | unsigned X86OpcV = getTargetVShiftUniformOpcode(Opc, true); | |||
| 28452 | unsigned X86OpcI = getTargetVShiftUniformOpcode(Opc, false); | |||
| 28453 | ||||
| 28454 | assert(VT.isVector() && "Custom lowering only for vector shifts!")((void)0); | |||
| 28455 | assert(Subtarget.hasSSE2() && "Only custom lower when we have SSE2!")((void)0); | |||
| 28456 | ||||
| 28457 | if (SDValue V = LowerScalarImmediateShift(Op, DAG, Subtarget)) | |||
| 28458 | return V; | |||
| 28459 | ||||
| 28460 | if (SDValue V = LowerScalarVariableShift(Op, DAG, Subtarget)) | |||
| 28461 | return V; | |||
| 28462 | ||||
| 28463 | if (SupportedVectorVarShift(VT, Subtarget, Opc)) | |||
| 28464 | return Op; | |||
| 28465 | ||||
| 28466 | // XOP has 128-bit variable logical/arithmetic shifts. | |||
| 28467 | // +ve/-ve Amt = shift left/right. | |||
| 28468 | if (Subtarget.hasXOP() && (VT == MVT::v2i64 || VT == MVT::v4i32 || | |||
| 28469 | VT == MVT::v8i16 || VT == MVT::v16i8)) { | |||
| 28470 | if (Opc == ISD::SRL || Opc == ISD::SRA) { | |||
| 28471 | SDValue Zero = DAG.getConstant(0, dl, VT); | |||
| 28472 | Amt = DAG.getNode(ISD::SUB, dl, VT, Zero, Amt); | |||
| 28473 | } | |||
| 28474 | if (Opc == ISD::SHL || Opc == ISD::SRL) | |||
| 28475 | return DAG.getNode(X86ISD::VPSHL, dl, VT, R, Amt); | |||
| 28476 | if (Opc == ISD::SRA) | |||
| 28477 | return DAG.getNode(X86ISD::VPSHA, dl, VT, R, Amt); | |||
| 28478 | } | |||
| 28479 | ||||
| 28480 | // 2i64 vector logical shifts can efficiently avoid scalarization - do the | |||
| 28481 | // shifts per-lane and then shuffle the partial results back together. | |||
| 28482 | if (VT == MVT::v2i64 && Opc != ISD::SRA) { | |||
| 28483 | // Splat the shift amounts so the scalar shifts above will catch it. | |||
| 28484 | SDValue Amt0 = DAG.getVectorShuffle(VT, dl, Amt, Amt, {0, 0}); | |||
| 28485 | SDValue Amt1 = DAG.getVectorShuffle(VT, dl, Amt, Amt, {1, 1}); | |||
| 28486 | SDValue R0 = DAG.getNode(Opc, dl, VT, R, Amt0); | |||
| 28487 | SDValue R1 = DAG.getNode(Opc, dl, VT, R, Amt1); | |||
| 28488 | return DAG.getVectorShuffle(VT, dl, R0, R1, {0, 3}); | |||
| 28489 | } | |||
| 28490 | ||||
| 28491 | // i64 vector arithmetic shift can be emulated with the transform: | |||
| 28492 | // M = lshr(SIGN_MASK, Amt) | |||
| 28493 | // ashr(R, Amt) === sub(xor(lshr(R, Amt), M), M) | |||
| 28494 | if ((VT == MVT::v2i64 || (VT == MVT::v4i64 && Subtarget.hasInt256())) && | |||
| 28495 | Opc == ISD::SRA) { | |||
| 28496 | SDValue S = DAG.getConstant(APInt::getSignMask(64), dl, VT); | |||
| 28497 | SDValue M = DAG.getNode(ISD::SRL, dl, VT, S, Amt); | |||
| 28498 | R = DAG.getNode(ISD::SRL, dl, VT, R, Amt); | |||
| 28499 | R = DAG.getNode(ISD::XOR, dl, VT, R, M); | |||
| 28500 | R = DAG.getNode(ISD::SUB, dl, VT, R, M); | |||
| 28501 | return R; | |||
| 28502 | } | |||
| 28503 | ||||
| 28504 | // If possible, lower this shift as a sequence of two shifts by | |||
| 28505 | // constant plus a BLENDing shuffle instead of scalarizing it. | |||
| 28506 | // Example: | |||
| 28507 | // (v4i32 (srl A, (build_vector < X, Y, Y, Y>))) | |||
| 28508 | // | |||
| 28509 | // Could be rewritten as: | |||
| 28510 | // (v4i32 (MOVSS (srl A, <Y,Y,Y,Y>), (srl A, <X,X,X,X>))) | |||
| 28511 | // | |||
| 28512 | // The advantage is that the two shifts from the example would be | |||
| 28513 | // lowered as X86ISD::VSRLI nodes in parallel before blending. | |||
| 28514 | if (ConstantAmt && (VT == MVT::v8i16 || VT == MVT::v4i32 || | |||
| 28515 | (VT == MVT::v16i16 && Subtarget.hasInt256()))) { | |||
| 28516 | SDValue Amt1, Amt2; | |||
| 28517 | unsigned NumElts = VT.getVectorNumElements(); | |||
| 28518 | SmallVector<int, 8> ShuffleMask; | |||
| 28519 | for (unsigned i = 0; i != NumElts; ++i) { | |||
| 28520 | SDValue A = Amt->getOperand(i); | |||
| 28521 | if (A.isUndef()) { | |||
| 28522 | ShuffleMask.push_back(SM_SentinelUndef); | |||
| 28523 | continue; | |||
| 28524 | } | |||
| 28525 | if (!Amt1 || Amt1 == A) { | |||
| 28526 | ShuffleMask.push_back(i); | |||
| 28527 | Amt1 = A; | |||
| 28528 | continue; | |||
| 28529 | } | |||
| 28530 | if (!Amt2 || Amt2 == A) { | |||
| 28531 | ShuffleMask.push_back(i + NumElts); | |||
| 28532 | Amt2 = A; | |||
| 28533 | continue; | |||
| 28534 | } | |||
| 28535 | break; | |||
| 28536 | } | |||
| 28537 | ||||
| 28538 | // Only perform this blend if we can perform it without loading a mask. | |||
| 28539 | if (ShuffleMask.size() == NumElts && Amt1 && Amt2 && | |||
| 28540 | (VT != MVT::v16i16 || | |||
| 28541 | is128BitLaneRepeatedShuffleMask(VT, ShuffleMask)) && | |||
| 28542 | (VT == MVT::v4i32 || Subtarget.hasSSE41() || Opc != ISD::SHL || | |||
| 28543 | canWidenShuffleElements(ShuffleMask))) { | |||
| 28544 | auto *Cst1 = dyn_cast<ConstantSDNode>(Amt1); | |||
| 28545 | auto *Cst2 = dyn_cast<ConstantSDNode>(Amt2); | |||
| 28546 | if (Cst1 && Cst2 && Cst1->getAPIntValue().ult(EltSizeInBits) && | |||
| 28547 | Cst2->getAPIntValue().ult(EltSizeInBits)) { | |||
| 28548 | SDValue Shift1 = getTargetVShiftByConstNode(X86OpcI, dl, VT, R, | |||
| 28549 | Cst1->getZExtValue(), DAG); | |||
| 28550 | SDValue Shift2 = getTargetVShiftByConstNode(X86OpcI, dl, VT, R, | |||
| 28551 | Cst2->getZExtValue(), DAG); | |||
| 28552 | return DAG.getVectorShuffle(VT, dl, Shift1, Shift2, ShuffleMask); | |||
| 28553 | } | |||
| 28554 | } | |||
| 28555 | } | |||
| 28556 | ||||
| 28557 | // If possible, lower this packed shift into a vector multiply instead of | |||
| 28558 | // expanding it into a sequence of scalar shifts. | |||
| 28559 | if (Opc == ISD::SHL) | |||
| 28560 | if (SDValue Scale = convertShiftLeftToScale(Amt, dl, Subtarget, DAG)) | |||
| 28561 | return DAG.getNode(ISD::MUL, dl, VT, R, Scale); | |||
| 28562 | ||||
| 28563 | // Constant ISD::SRL can be performed efficiently on vXi16 vectors as we | |||
| 28564 | // can replace with ISD::MULHU, creating scale factor from (NumEltBits - Amt). | |||
| 28565 | if (Opc == ISD::SRL && ConstantAmt && | |||
| 28566 | (VT == MVT::v8i16 || (VT == MVT::v16i16 && Subtarget.hasInt256()))) { | |||
| 28567 | SDValue EltBits = DAG.getConstant(EltSizeInBits, dl, VT); | |||
| 28568 | SDValue RAmt = DAG.getNode(ISD::SUB, dl, VT, EltBits, Amt); | |||
| 28569 | if (SDValue Scale = convertShiftLeftToScale(RAmt, dl, Subtarget, DAG)) { | |||
| 28570 | SDValue Zero = DAG.getConstant(0, dl, VT); | |||
| 28571 | SDValue ZAmt = DAG.getSetCC(dl, VT, Amt, Zero, ISD::SETEQ); | |||
| 28572 | SDValue Res = DAG.getNode(ISD::MULHU, dl, VT, R, Scale); | |||
| 28573 | return DAG.getSelect(dl, VT, ZAmt, R, Res); | |||
| 28574 | } | |||
| 28575 | } | |||
| 28576 | ||||
| 28577 | // Constant ISD::SRA can be performed efficiently on vXi16 vectors as we | |||
| 28578 | // can replace with ISD::MULHS, creating scale factor from (NumEltBits - Amt). | |||
| 28579 | // TODO: Special case handling for shift by 0/1, really we can afford either | |||
| 28580 | // of these cases in pre-SSE41/XOP/AVX512 but not both. | |||
| 28581 | if (Opc == ISD::SRA && ConstantAmt && | |||
| 28582 | (VT == MVT::v8i16 || (VT == MVT::v16i16 && Subtarget.hasInt256())) && | |||
| 28583 | ((Subtarget.hasSSE41() && !Subtarget.hasXOP() && | |||
| 28584 | !Subtarget.hasAVX512()) || | |||
| 28585 | DAG.isKnownNeverZero(Amt))) { | |||
| 28586 | SDValue EltBits = DAG.getConstant(EltSizeInBits, dl, VT); | |||
| 28587 | SDValue RAmt = DAG.getNode(ISD::SUB, dl, VT, EltBits, Amt); | |||
| 28588 | if (SDValue Scale = convertShiftLeftToScale(RAmt, dl, Subtarget, DAG)) { | |||
| 28589 | SDValue Amt0 = | |||
| 28590 | DAG.getSetCC(dl, VT, Amt, DAG.getConstant(0, dl, VT), ISD::SETEQ); | |||
| 28591 | SDValue Amt1 = | |||
| 28592 | DAG.getSetCC(dl, VT, Amt, DAG.getConstant(1, dl, VT), ISD::SETEQ); | |||
| 28593 | SDValue Sra1 = | |||
| 28594 | getTargetVShiftByConstNode(X86ISD::VSRAI, dl, VT, R, 1, DAG); | |||
| 28595 | SDValue Res = DAG.getNode(ISD::MULHS, dl, VT, R, Scale); | |||
| 28596 | Res = DAG.getSelect(dl, VT, Amt0, R, Res); | |||
| 28597 | return DAG.getSelect(dl, VT, Amt1, Sra1, Res); | |||
| 28598 | } | |||
| 28599 | } | |||
| 28600 | ||||
| 28601 | // v4i32 Non Uniform Shifts. | |||
| 28602 | // If the shift amount is constant we can shift each lane using the SSE2 | |||
| 28603 | // immediate shifts, else we need to zero-extend each lane to the lower i64 | |||
| 28604 | // and shift using the SSE2 variable shifts. | |||
| 28605 | // The separate results can then be blended together. | |||
| 28606 | if (VT == MVT::v4i32) { | |||
| 28607 | SDValue Amt0, Amt1, Amt2, Amt3; | |||
| 28608 | if (ConstantAmt) { | |||
| 28609 | Amt0 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {0, 0, 0, 0}); | |||
| 28610 | Amt1 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {1, 1, 1, 1}); | |||
| 28611 | Amt2 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {2, 2, 2, 2}); | |||
| 28612 | Amt3 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {3, 3, 3, 3}); | |||
| 28613 | } else { | |||
| 28614 | // The SSE2 shifts use the lower i64 as the same shift amount for | |||
| 28615 | // all lanes and the upper i64 is ignored. On AVX we're better off | |||
| 28616 | // just zero-extending, but for SSE just duplicating the top 16-bits is | |||
| 28617 | // cheaper and has the same effect for out of range values. | |||
| 28618 | if (Subtarget.hasAVX()) { | |||
| 28619 | SDValue Z = DAG.getConstant(0, dl, VT); | |||
| 28620 | Amt0 = DAG.getVectorShuffle(VT, dl, Amt, Z, {0, 4, -1, -1}); | |||
| 28621 | Amt1 = DAG.getVectorShuffle(VT, dl, Amt, Z, {1, 5, -1, -1}); | |||
| 28622 | Amt2 = DAG.getVectorShuffle(VT, dl, Amt, Z, {2, 6, -1, -1}); | |||
| 28623 | Amt3 = DAG.getVectorShuffle(VT, dl, Amt, Z, {3, 7, -1, -1}); | |||
| 28624 | } else { | |||
| 28625 | SDValue Amt01 = DAG.getBitcast(MVT::v8i16, Amt); | |||
| 28626 | SDValue Amt23 = DAG.getVectorShuffle(MVT::v8i16, dl, Amt01, Amt01, | |||
| 28627 | {4, 5, 6, 7, -1, -1, -1, -1}); | |||
| 28628 | Amt0 = DAG.getVectorShuffle(MVT::v8i16, dl, Amt01, Amt01, | |||
| 28629 | {0, 1, 1, 1, -1, -1, -1, -1}); | |||
| 28630 | Amt1 = DAG.getVectorShuffle(MVT::v8i16, dl, Amt01, Amt01, | |||
| 28631 | {2, 3, 3, 3, -1, -1, -1, -1}); | |||
| 28632 | Amt2 = DAG.getVectorShuffle(MVT::v8i16, dl, Amt23, Amt23, | |||
| 28633 | {0, 1, 1, 1, -1, -1, -1, -1}); | |||
| 28634 | Amt3 = DAG.getVectorShuffle(MVT::v8i16, dl, Amt23, Amt23, | |||
| 28635 | {2, 3, 3, 3, -1, -1, -1, -1}); | |||
| 28636 | } | |||
| 28637 | } | |||
| 28638 | ||||
| 28639 | unsigned ShOpc = ConstantAmt ? Opc : X86OpcV; | |||
| 28640 | SDValue R0 = DAG.getNode(ShOpc, dl, VT, R, DAG.getBitcast(VT, Amt0)); | |||
| 28641 | SDValue R1 = DAG.getNode(ShOpc, dl, VT, R, DAG.getBitcast(VT, Amt1)); | |||
| 28642 | SDValue R2 = DAG.getNode(ShOpc, dl, VT, R, DAG.getBitcast(VT, Amt2)); | |||
| 28643 | SDValue R3 = DAG.getNode(ShOpc, dl, VT, R, DAG.getBitcast(VT, Amt3)); | |||
| 28644 | ||||
| 28645 | // Merge the shifted lane results optimally with/without PBLENDW. | |||
| 28646 | // TODO - ideally shuffle combining would handle this. | |||
| 28647 | if (Subtarget.hasSSE41()) { | |||
| 28648 | SDValue R02 = DAG.getVectorShuffle(VT, dl, R0, R2, {0, -1, 6, -1}); | |||
| 28649 | SDValue R13 = DAG.getVectorShuffle(VT, dl, R1, R3, {-1, 1, -1, 7}); | |||
| 28650 | return DAG.getVectorShuffle(VT, dl, R02, R13, {0, 5, 2, 7}); | |||
| 28651 | } | |||
| 28652 | SDValue R01 = DAG.getVectorShuffle(VT, dl, R0, R1, {0, -1, -1, 5}); | |||
| 28653 | SDValue R23 = DAG.getVectorShuffle(VT, dl, R2, R3, {2, -1, -1, 7}); | |||
| 28654 | return DAG.getVectorShuffle(VT, dl, R01, R23, {0, 3, 4, 7}); | |||
| 28655 | } | |||
| 28656 | ||||
| 28657 | // It's worth extending once and using the vXi16/vXi32 shifts for smaller | |||
| 28658 | // types, but without AVX512 the extra overheads to get from vXi8 to vXi32 | |||
| 28659 | // make the existing SSE solution better. | |||
| 28660 | // NOTE: We honor prefered vector width before promoting to 512-bits. | |||
| 28661 | if ((Subtarget.hasInt256() && VT == MVT::v8i16) || | |||
| 28662 | (Subtarget.canExtendTo512DQ() && VT == MVT::v16i16) || | |||
| 28663 | (Subtarget.canExtendTo512DQ() && VT == MVT::v16i8) || | |||
| 28664 | (Subtarget.canExtendTo512BW() && VT == MVT::v32i8) || | |||
| 28665 | (Subtarget.hasBWI() && Subtarget.hasVLX() && VT == MVT::v16i8)) { | |||
| 28666 | assert((!Subtarget.hasBWI() || VT == MVT::v32i8 || VT == MVT::v16i8) &&((void)0) | |||
| 28667 | "Unexpected vector type")((void)0); | |||
| 28668 | MVT EvtSVT = Subtarget.hasBWI() ? MVT::i16 : MVT::i32; | |||
| 28669 | MVT ExtVT = MVT::getVectorVT(EvtSVT, VT.getVectorNumElements()); | |||
| 28670 | unsigned ExtOpc = Opc == ISD::SRA ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND; | |||
| 28671 | R = DAG.getNode(ExtOpc, dl, ExtVT, R); | |||
| 28672 | Amt = DAG.getNode(ISD::ZERO_EXTEND, dl, ExtVT, Amt); | |||
| 28673 | return DAG.getNode(ISD::TRUNCATE, dl, VT, | |||
| 28674 | DAG.getNode(Opc, dl, ExtVT, R, Amt)); | |||
| 28675 | } | |||
| 28676 | ||||
| 28677 | // Constant ISD::SRA/SRL can be performed efficiently on vXi8 vectors as we | |||
| 28678 | // extend to vXi16 to perform a MUL scale effectively as a MUL_LOHI. | |||
| 28679 | if (ConstantAmt && (Opc == ISD::SRA || Opc == ISD::SRL) && | |||
| 28680 | (VT == MVT::v16i8 || (VT == MVT::v32i8 && Subtarget.hasInt256()) || | |||
| 28681 | (VT == MVT::v64i8 && Subtarget.hasBWI())) && | |||
| 28682 | !Subtarget.hasXOP()) { | |||
| 28683 | int NumElts = VT.getVectorNumElements(); | |||
| 28684 | SDValue Cst8 = DAG.getTargetConstant(8, dl, MVT::i8); | |||
| 28685 | ||||
| 28686 | // Extend constant shift amount to vXi16 (it doesn't matter if the type | |||
| 28687 | // isn't legal). | |||
| 28688 | MVT ExVT = MVT::getVectorVT(MVT::i16, NumElts); | |||
| 28689 | Amt = DAG.getZExtOrTrunc(Amt, dl, ExVT); | |||
| 28690 | Amt = DAG.getNode(ISD::SUB, dl, ExVT, DAG.getConstant(8, dl, ExVT), Amt); | |||
| 28691 | Amt = DAG.getNode(ISD::SHL, dl, ExVT, DAG.getConstant(1, dl, ExVT), Amt); | |||
| 28692 | assert(ISD::isBuildVectorOfConstantSDNodes(Amt.getNode()) &&((void)0) | |||
| 28693 | "Constant build vector expected")((void)0); | |||
| 28694 | ||||
| 28695 | if (VT == MVT::v16i8 && Subtarget.hasInt256()) { | |||
| 28696 | R = Opc == ISD::SRA ? DAG.getSExtOrTrunc(R, dl, ExVT) | |||
| 28697 | : DAG.getZExtOrTrunc(R, dl, ExVT); | |||
| 28698 | R = DAG.getNode(ISD::MUL, dl, ExVT, R, Amt); | |||
| 28699 | R = DAG.getNode(X86ISD::VSRLI, dl, ExVT, R, Cst8); | |||
| 28700 | return DAG.getZExtOrTrunc(R, dl, VT); | |||
| 28701 | } | |||
| 28702 | ||||
| 28703 | SmallVector<SDValue, 16> LoAmt, HiAmt; | |||
| 28704 | for (int i = 0; i != NumElts; i += 16) { | |||
| 28705 | for (int j = 0; j != 8; ++j) { | |||
| 28706 | LoAmt.push_back(Amt.getOperand(i + j)); | |||
| 28707 | HiAmt.push_back(Amt.getOperand(i + j + 8)); | |||
| 28708 | } | |||
| 28709 | } | |||
| 28710 | ||||
| 28711 | MVT VT16 = MVT::getVectorVT(MVT::i16, NumElts / 2); | |||
| 28712 | SDValue LoA = DAG.getBuildVector(VT16, dl, LoAmt); | |||
| 28713 | SDValue HiA = DAG.getBuildVector(VT16, dl, HiAmt); | |||
| 28714 | ||||
| 28715 | SDValue LoR = DAG.getBitcast(VT16, getUnpackl(DAG, dl, VT, R, R)); | |||
| 28716 | SDValue HiR = DAG.getBitcast(VT16, getUnpackh(DAG, dl, VT, R, R)); | |||
| 28717 | LoR = DAG.getNode(X86OpcI, dl, VT16, LoR, Cst8); | |||
| 28718 | HiR = DAG.getNode(X86OpcI, dl, VT16, HiR, Cst8); | |||
| 28719 | LoR = DAG.getNode(ISD::MUL, dl, VT16, LoR, LoA); | |||
| 28720 | HiR = DAG.getNode(ISD::MUL, dl, VT16, HiR, HiA); | |||
| 28721 | LoR = DAG.getNode(X86ISD::VSRLI, dl, VT16, LoR, Cst8); | |||
| 28722 | HiR = DAG.getNode(X86ISD::VSRLI, dl, VT16, HiR, Cst8); | |||
| 28723 | return DAG.getNode(X86ISD::PACKUS, dl, VT, LoR, HiR); | |||
| 28724 | } | |||
| 28725 | ||||
| 28726 | if (VT == MVT::v16i8 || | |||
| 28727 | (VT == MVT::v32i8 && Subtarget.hasInt256() && !Subtarget.hasXOP()) || | |||
| 28728 | (VT == MVT::v64i8 && Subtarget.hasBWI())) { | |||
| 28729 | MVT ExtVT = MVT::getVectorVT(MVT::i16, VT.getVectorNumElements() / 2); | |||
| 28730 | ||||
| 28731 | auto SignBitSelect = [&](MVT SelVT, SDValue Sel, SDValue V0, SDValue V1) { | |||
| 28732 | if (VT.is512BitVector()) { | |||
| 28733 | // On AVX512BW targets we make use of the fact that VSELECT lowers | |||
| 28734 | // to a masked blend which selects bytes based just on the sign bit | |||
| 28735 | // extracted to a mask. | |||
| 28736 | MVT MaskVT = MVT::getVectorVT(MVT::i1, VT.getVectorNumElements()); | |||
| 28737 | V0 = DAG.getBitcast(VT, V0); | |||
| 28738 | V1 = DAG.getBitcast(VT, V1); | |||
| 28739 | Sel = DAG.getBitcast(VT, Sel); | |||
| 28740 | Sel = DAG.getSetCC(dl, MaskVT, DAG.getConstant(0, dl, VT), Sel, | |||
| 28741 | ISD::SETGT); | |||
| 28742 | return DAG.getBitcast(SelVT, DAG.getSelect(dl, VT, Sel, V0, V1)); | |||
| 28743 | } else if (Subtarget.hasSSE41()) { | |||
| 28744 | // On SSE41 targets we can use PBLENDVB which selects bytes based just | |||
| 28745 | // on the sign bit. | |||
| 28746 | V0 = DAG.getBitcast(VT, V0); | |||
| 28747 | V1 = DAG.getBitcast(VT, V1); | |||
| 28748 | Sel = DAG.getBitcast(VT, Sel); | |||
| 28749 | return DAG.getBitcast(SelVT, | |||
| 28750 | DAG.getNode(X86ISD::BLENDV, dl, VT, Sel, V0, V1)); | |||
| 28751 | } | |||
| 28752 | // On pre-SSE41 targets we test for the sign bit by comparing to | |||
| 28753 | // zero - a negative value will set all bits of the lanes to true | |||
| 28754 | // and VSELECT uses that in its OR(AND(V0,C),AND(V1,~C)) lowering. | |||
| 28755 | SDValue Z = DAG.getConstant(0, dl, SelVT); | |||
| 28756 | SDValue C = DAG.getNode(X86ISD::PCMPGT, dl, SelVT, Z, Sel); | |||
| 28757 | return DAG.getSelect(dl, SelVT, C, V0, V1); | |||
| 28758 | }; | |||
| 28759 | ||||
| 28760 | // Turn 'a' into a mask suitable for VSELECT: a = a << 5; | |||
| 28761 | // We can safely do this using i16 shifts as we're only interested in | |||
| 28762 | // the 3 lower bits of each byte. | |||
| 28763 | Amt = DAG.getBitcast(ExtVT, Amt); | |||
| 28764 | Amt = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, ExtVT, Amt, 5, DAG); | |||
| 28765 | Amt = DAG.getBitcast(VT, Amt); | |||
| 28766 | ||||
| 28767 | if (Opc == ISD::SHL || Opc == ISD::SRL) { | |||
| 28768 | // r = VSELECT(r, shift(r, 4), a); | |||
| 28769 | SDValue M = DAG.getNode(Opc, dl, VT, R, DAG.getConstant(4, dl, VT)); | |||
| 28770 | R = SignBitSelect(VT, Amt, M, R); | |||
| 28771 | ||||
| 28772 | // a += a | |||
| 28773 | Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt); | |||
| 28774 | ||||
| 28775 | // r = VSELECT(r, shift(r, 2), a); | |||
| 28776 | M = DAG.getNode(Opc, dl, VT, R, DAG.getConstant(2, dl, VT)); | |||
| 28777 | R = SignBitSelect(VT, Amt, M, R); | |||
| 28778 | ||||
| 28779 | // a += a | |||
| 28780 | Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt); | |||
| 28781 | ||||
| 28782 | // return VSELECT(r, shift(r, 1), a); | |||
| 28783 | M = DAG.getNode(Opc, dl, VT, R, DAG.getConstant(1, dl, VT)); | |||
| 28784 | R = SignBitSelect(VT, Amt, M, R); | |||
| 28785 | return R; | |||
| 28786 | } | |||
| 28787 | ||||
| 28788 | if (Opc == ISD::SRA) { | |||
| 28789 | // For SRA we need to unpack each byte to the higher byte of a i16 vector | |||
| 28790 | // so we can correctly sign extend. We don't care what happens to the | |||
| 28791 | // lower byte. | |||
| 28792 | SDValue ALo = getUnpackl(DAG, dl, VT, DAG.getUNDEF(VT), Amt); | |||
| 28793 | SDValue AHi = getUnpackh(DAG, dl, VT, DAG.getUNDEF(VT), Amt); | |||
| 28794 | SDValue RLo = getUnpackl(DAG, dl, VT, DAG.getUNDEF(VT), R); | |||
| 28795 | SDValue RHi = getUnpackh(DAG, dl, VT, DAG.getUNDEF(VT), R); | |||
| 28796 | ALo = DAG.getBitcast(ExtVT, ALo); | |||
| 28797 | AHi = DAG.getBitcast(ExtVT, AHi); | |||
| 28798 | RLo = DAG.getBitcast(ExtVT, RLo); | |||
| 28799 | RHi = DAG.getBitcast(ExtVT, RHi); | |||
| 28800 | ||||
| 28801 | // r = VSELECT(r, shift(r, 4), a); | |||
| 28802 | SDValue MLo = getTargetVShiftByConstNode(X86OpcI, dl, ExtVT, RLo, 4, DAG); | |||
| 28803 | SDValue MHi = getTargetVShiftByConstNode(X86OpcI, dl, ExtVT, RHi, 4, DAG); | |||
| 28804 | RLo = SignBitSelect(ExtVT, ALo, MLo, RLo); | |||
| 28805 | RHi = SignBitSelect(ExtVT, AHi, MHi, RHi); | |||
| 28806 | ||||
| 28807 | // a += a | |||
| 28808 | ALo = DAG.getNode(ISD::ADD, dl, ExtVT, ALo, ALo); | |||
| 28809 | AHi = DAG.getNode(ISD::ADD, dl, ExtVT, AHi, AHi); | |||
| 28810 | ||||
| 28811 | // r = VSELECT(r, shift(r, 2), a); | |||
| 28812 | MLo = getTargetVShiftByConstNode(X86OpcI, dl, ExtVT, RLo, 2, DAG); | |||
| 28813 | MHi = getTargetVShiftByConstNode(X86OpcI, dl, ExtVT, RHi, 2, DAG); | |||
| 28814 | RLo = SignBitSelect(ExtVT, ALo, MLo, RLo); | |||
| 28815 | RHi = SignBitSelect(ExtVT, AHi, MHi, RHi); | |||
| 28816 | ||||
| 28817 | // a += a | |||
| 28818 | ALo = DAG.getNode(ISD::ADD, dl, ExtVT, ALo, ALo); | |||
| 28819 | AHi = DAG.getNode(ISD::ADD, dl, ExtVT, AHi, AHi); | |||
| 28820 | ||||
| 28821 | // r = VSELECT(r, shift(r, 1), a); | |||
| 28822 | MLo = getTargetVShiftByConstNode(X86OpcI, dl, ExtVT, RLo, 1, DAG); | |||
| 28823 | MHi = getTargetVShiftByConstNode(X86OpcI, dl, ExtVT, RHi, 1, DAG); | |||
| 28824 | RLo = SignBitSelect(ExtVT, ALo, MLo, RLo); | |||
| 28825 | RHi = SignBitSelect(ExtVT, AHi, MHi, RHi); | |||
| 28826 | ||||
| 28827 | // Logical shift the result back to the lower byte, leaving a zero upper | |||
| 28828 | // byte meaning that we can safely pack with PACKUSWB. | |||
| 28829 | RLo = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, ExtVT, RLo, 8, DAG); | |||
| 28830 | RHi = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, ExtVT, RHi, 8, DAG); | |||
| 28831 | return DAG.getNode(X86ISD::PACKUS, dl, VT, RLo, RHi); | |||
| 28832 | } | |||
| 28833 | } | |||
| 28834 | ||||
| 28835 | if (Subtarget.hasInt256() && !Subtarget.hasXOP() && VT == MVT::v16i16) { | |||
| 28836 | MVT ExtVT = MVT::v8i32; | |||
| 28837 | SDValue Z = DAG.getConstant(0, dl, VT); | |||
| 28838 | SDValue ALo = getUnpackl(DAG, dl, VT, Amt, Z); | |||
| 28839 | SDValue AHi = getUnpackh(DAG, dl, VT, Amt, Z); | |||
| 28840 | SDValue RLo = getUnpackl(DAG, dl, VT, Z, R); | |||
| 28841 | SDValue RHi = getUnpackh(DAG, dl, VT, Z, R); | |||
| 28842 | ALo = DAG.getBitcast(ExtVT, ALo); | |||
| 28843 | AHi = DAG.getBitcast(ExtVT, AHi); | |||
| 28844 | RLo = DAG.getBitcast(ExtVT, RLo); | |||
| 28845 | RHi = DAG.getBitcast(ExtVT, RHi); | |||
| 28846 | SDValue Lo = DAG.getNode(Opc, dl, ExtVT, RLo, ALo); | |||
| 28847 | SDValue Hi = DAG.getNode(Opc, dl, ExtVT, RHi, AHi); | |||
| 28848 | Lo = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, ExtVT, Lo, 16, DAG); | |||
| 28849 | Hi = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, ExtVT, Hi, 16, DAG); | |||
| 28850 | return DAG.getNode(X86ISD::PACKUS, dl, VT, Lo, Hi); | |||
| 28851 | } | |||
| 28852 | ||||
| 28853 | if (VT == MVT::v8i16) { | |||
| 28854 | // If we have a constant shift amount, the non-SSE41 path is best as | |||
| 28855 | // avoiding bitcasts make it easier to constant fold and reduce to PBLENDW. | |||
| 28856 | bool UseSSE41 = Subtarget.hasSSE41() && | |||
| 28857 | !ISD::isBuildVectorOfConstantSDNodes(Amt.getNode()); | |||
| 28858 | ||||
| 28859 | auto SignBitSelect = [&](SDValue Sel, SDValue V0, SDValue V1) { | |||
| 28860 | // On SSE41 targets we can use PBLENDVB which selects bytes based just on | |||
| 28861 | // the sign bit. | |||
| 28862 | if (UseSSE41) { | |||
| 28863 | MVT ExtVT = MVT::getVectorVT(MVT::i8, VT.getVectorNumElements() * 2); | |||
| 28864 | V0 = DAG.getBitcast(ExtVT, V0); | |||
| 28865 | V1 = DAG.getBitcast(ExtVT, V1); | |||
| 28866 | Sel = DAG.getBitcast(ExtVT, Sel); | |||
| 28867 | return DAG.getBitcast( | |||
| 28868 | VT, DAG.getNode(X86ISD::BLENDV, dl, ExtVT, Sel, V0, V1)); | |||
| 28869 | } | |||
| 28870 | // On pre-SSE41 targets we splat the sign bit - a negative value will | |||
| 28871 | // set all bits of the lanes to true and VSELECT uses that in | |||
| 28872 | // its OR(AND(V0,C),AND(V1,~C)) lowering. | |||
| 28873 | SDValue C = | |||
| 28874 | getTargetVShiftByConstNode(X86ISD::VSRAI, dl, VT, Sel, 15, DAG); | |||
| 28875 | return DAG.getSelect(dl, VT, C, V0, V1); | |||
| 28876 | }; | |||
| 28877 | ||||
| 28878 | // Turn 'a' into a mask suitable for VSELECT: a = a << 12; | |||
| 28879 | if (UseSSE41) { | |||
| 28880 | // On SSE41 targets we need to replicate the shift mask in both | |||
| 28881 | // bytes for PBLENDVB. | |||
| 28882 | Amt = DAG.getNode( | |||
| 28883 | ISD::OR, dl, VT, | |||
| 28884 | getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, Amt, 4, DAG), | |||
| 28885 | getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, Amt, 12, DAG)); | |||
| 28886 | } else { | |||
| 28887 | Amt = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, Amt, 12, DAG); | |||
| 28888 | } | |||
| 28889 | ||||
| 28890 | // r = VSELECT(r, shift(r, 8), a); | |||
| 28891 | SDValue M = getTargetVShiftByConstNode(X86OpcI, dl, VT, R, 8, DAG); | |||
| 28892 | R = SignBitSelect(Amt, M, R); | |||
| 28893 | ||||
| 28894 | // a += a | |||
| 28895 | Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt); | |||
| 28896 | ||||
| 28897 | // r = VSELECT(r, shift(r, 4), a); | |||
| 28898 | M = getTargetVShiftByConstNode(X86OpcI, dl, VT, R, 4, DAG); | |||
| 28899 | R = SignBitSelect(Amt, M, R); | |||
| 28900 | ||||
| 28901 | // a += a | |||
| 28902 | Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt); | |||
| 28903 | ||||
| 28904 | // r = VSELECT(r, shift(r, 2), a); | |||
| 28905 | M = getTargetVShiftByConstNode(X86OpcI, dl, VT, R, 2, DAG); | |||
| 28906 | R = SignBitSelect(Amt, M, R); | |||
| 28907 | ||||
| 28908 | // a += a | |||
| 28909 | Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt); | |||
| 28910 | ||||
| 28911 | // return VSELECT(r, shift(r, 1), a); | |||
| 28912 | M = getTargetVShiftByConstNode(X86OpcI, dl, VT, R, 1, DAG); | |||
| 28913 | R = SignBitSelect(Amt, M, R); | |||
| 28914 | return R; | |||
| 28915 | } | |||
| 28916 | ||||
| 28917 | // Decompose 256-bit shifts into 128-bit shifts. | |||
| 28918 | if (VT.is256BitVector()) | |||
| 28919 | return splitVectorIntBinary(Op, DAG); | |||
| 28920 | ||||
| 28921 | if (VT == MVT::v32i16 || VT == MVT::v64i8) | |||
| 28922 | return splitVectorIntBinary(Op, DAG); | |||
| 28923 | ||||
| 28924 | return SDValue(); | |||
| 28925 | } | |||
| 28926 | ||||
| 28927 | static SDValue LowerRotate(SDValue Op, const X86Subtarget &Subtarget, | |||
| 28928 | SelectionDAG &DAG) { | |||
| 28929 | MVT VT = Op.getSimpleValueType(); | |||
| 28930 | assert(VT.isVector() && "Custom lowering only for vector rotates!")((void)0); | |||
| 28931 | ||||
| 28932 | SDLoc DL(Op); | |||
| 28933 | SDValue R = Op.getOperand(0); | |||
| 28934 | SDValue Amt = Op.getOperand(1); | |||
| 28935 | unsigned Opcode = Op.getOpcode(); | |||
| 28936 | unsigned EltSizeInBits = VT.getScalarSizeInBits(); | |||
| 28937 | int NumElts = VT.getVectorNumElements(); | |||
| 28938 | ||||
| 28939 | // Check for constant splat rotation amount. | |||
| 28940 | APInt CstSplatValue; | |||
| 28941 | bool IsCstSplat = X86::isConstantSplat(Amt, CstSplatValue); | |||
| 28942 | ||||
| 28943 | // Check for splat rotate by zero. | |||
| 28944 | if (IsCstSplat && CstSplatValue.urem(EltSizeInBits) == 0) | |||
| 28945 | return R; | |||
| 28946 | ||||
| 28947 | // AVX512 implicitly uses modulo rotation amounts. | |||
| 28948 | if (Subtarget.hasAVX512() && 32 <= EltSizeInBits) { | |||
| 28949 | // Attempt to rotate by immediate. | |||
| 28950 | if (IsCstSplat) { | |||
| 28951 | unsigned RotOpc = (Opcode == ISD::ROTL ? X86ISD::VROTLI : X86ISD::VROTRI); | |||
| 28952 | uint64_t RotAmt = CstSplatValue.urem(EltSizeInBits); | |||
| 28953 | return DAG.getNode(RotOpc, DL, VT, R, | |||
| 28954 | DAG.getTargetConstant(RotAmt, DL, MVT::i8)); | |||
| 28955 | } | |||
| 28956 | ||||
| 28957 | // Else, fall-back on VPROLV/VPRORV. | |||
| 28958 | return Op; | |||
| 28959 | } | |||
| 28960 | ||||
| 28961 | // AVX512 VBMI2 vXi16 - lower to funnel shifts. | |||
| 28962 | if (Subtarget.hasVBMI2() && 16 == EltSizeInBits) { | |||
| 28963 | unsigned FunnelOpc = (Opcode == ISD::ROTL ? ISD::FSHL : ISD::FSHR); | |||
| 28964 | return DAG.getNode(FunnelOpc, DL, VT, R, R, Amt); | |||
| 28965 | } | |||
| 28966 | ||||
| 28967 | assert((Opcode == ISD::ROTL) && "Only ROTL supported")((void)0); | |||
| 28968 | ||||
| 28969 | // XOP has 128-bit vector variable + immediate rotates. | |||
| 28970 | // +ve/-ve Amt = rotate left/right - just need to handle ISD::ROTL. | |||
| 28971 | // XOP implicitly uses modulo rotation amounts. | |||
| 28972 | if (Subtarget.hasXOP()) { | |||
| 28973 | if (VT.is256BitVector()) | |||
| 28974 | return splitVectorIntBinary(Op, DAG); | |||
| 28975 | assert(VT.is128BitVector() && "Only rotate 128-bit vectors!")((void)0); | |||
| 28976 | ||||
| 28977 | // Attempt to rotate by immediate. | |||
| 28978 | if (IsCstSplat) { | |||
| 28979 | uint64_t RotAmt = CstSplatValue.urem(EltSizeInBits); | |||
| 28980 | return DAG.getNode(X86ISD::VROTLI, DL, VT, R, | |||
| 28981 | DAG.getTargetConstant(RotAmt, DL, MVT::i8)); | |||
| 28982 | } | |||
| 28983 | ||||
| 28984 | // Use general rotate by variable (per-element). | |||
| 28985 | return Op; | |||
| 28986 | } | |||
| 28987 | ||||
| 28988 | // Split 256-bit integers on pre-AVX2 targets. | |||
| 28989 | if (VT.is256BitVector() && !Subtarget.hasAVX2()) | |||
| 28990 | return splitVectorIntBinary(Op, DAG); | |||
| 28991 | ||||
| 28992 | assert((VT == MVT::v4i32 || VT == MVT::v8i16 || VT == MVT::v16i8 ||((void)0) | |||
| 28993 | ((VT == MVT::v8i32 || VT == MVT::v16i16 || VT == MVT::v32i8 ||((void)0) | |||
| 28994 | VT == MVT::v32i16) &&((void)0) | |||
| 28995 | Subtarget.hasAVX2())) &&((void)0) | |||
| 28996 | "Only vXi32/vXi16/vXi8 vector rotates supported")((void)0); | |||
| 28997 | ||||
| 28998 | // Rotate by an uniform constant - expand back to shifts. | |||
| 28999 | if (IsCstSplat) | |||
| 29000 | return SDValue(); | |||
| 29001 | ||||
| 29002 | bool IsSplatAmt = DAG.isSplatValue(Amt); | |||
| 29003 | ||||
| 29004 | // v16i8/v32i8: Split rotation into rot4/rot2/rot1 stages and select by | |||
| 29005 | // the amount bit. | |||
| 29006 | if (EltSizeInBits == 8 && !IsSplatAmt) { | |||
| 29007 | if (ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) | |||
| 29008 | return SDValue(); | |||
| 29009 | ||||
| 29010 | // We don't need ModuloAmt here as we just peek at individual bits. | |||
| 29011 | MVT ExtVT = MVT::getVectorVT(MVT::i16, NumElts / 2); | |||
| 29012 | ||||
| 29013 | auto SignBitSelect = [&](MVT SelVT, SDValue Sel, SDValue V0, SDValue V1) { | |||
| 29014 | if (Subtarget.hasSSE41()) { | |||
| 29015 | // On SSE41 targets we can use PBLENDVB which selects bytes based just | |||
| 29016 | // on the sign bit. | |||
| 29017 | V0 = DAG.getBitcast(VT, V0); | |||
| 29018 | V1 = DAG.getBitcast(VT, V1); | |||
| 29019 | Sel = DAG.getBitcast(VT, Sel); | |||
| 29020 | return DAG.getBitcast(SelVT, | |||
| 29021 | DAG.getNode(X86ISD::BLENDV, DL, VT, Sel, V0, V1)); | |||
| 29022 | } | |||
| 29023 | // On pre-SSE41 targets we test for the sign bit by comparing to | |||
| 29024 | // zero - a negative value will set all bits of the lanes to true | |||
| 29025 | // and VSELECT uses that in its OR(AND(V0,C),AND(V1,~C)) lowering. | |||
| 29026 | SDValue Z = DAG.getConstant(0, DL, SelVT); | |||
| 29027 | SDValue C = DAG.getNode(X86ISD::PCMPGT, DL, SelVT, Z, Sel); | |||
| 29028 | return DAG.getSelect(DL, SelVT, C, V0, V1); | |||
| 29029 | }; | |||
| 29030 | ||||
| 29031 | // Turn 'a' into a mask suitable for VSELECT: a = a << 5; | |||
| 29032 | // We can safely do this using i16 shifts as we're only interested in | |||
| 29033 | // the 3 lower bits of each byte. | |||
| 29034 | Amt = DAG.getBitcast(ExtVT, Amt); | |||
| 29035 | Amt = DAG.getNode(ISD::SHL, DL, ExtVT, Amt, DAG.getConstant(5, DL, ExtVT)); | |||
| 29036 | Amt = DAG.getBitcast(VT, Amt); | |||
| 29037 | ||||
| 29038 | // r = VSELECT(r, rot(r, 4), a); | |||
| 29039 | SDValue M; | |||
| 29040 | M = DAG.getNode( | |||
| 29041 | ISD::OR, DL, VT, | |||
| 29042 | DAG.getNode(ISD::SHL, DL, VT, R, DAG.getConstant(4, DL, VT)), | |||
| 29043 | DAG.getNode(ISD::SRL, DL, VT, R, DAG.getConstant(4, DL, VT))); | |||
| 29044 | R = SignBitSelect(VT, Amt, M, R); | |||
| 29045 | ||||
| 29046 | // a += a | |||
| 29047 | Amt = DAG.getNode(ISD::ADD, DL, VT, Amt, Amt); | |||
| 29048 | ||||
| 29049 | // r = VSELECT(r, rot(r, 2), a); | |||
| 29050 | M = DAG.getNode( | |||
| 29051 | ISD::OR, DL, VT, | |||
| 29052 | DAG.getNode(ISD::SHL, DL, VT, R, DAG.getConstant(2, DL, VT)), | |||
| 29053 | DAG.getNode(ISD::SRL, DL, VT, R, DAG.getConstant(6, DL, VT))); | |||
| 29054 | R = SignBitSelect(VT, Amt, M, R); | |||
| 29055 | ||||
| 29056 | // a += a | |||
| 29057 | Amt = DAG.getNode(ISD::ADD, DL, VT, Amt, Amt); | |||
| 29058 | ||||
| 29059 | // return VSELECT(r, rot(r, 1), a); | |||
| 29060 | M = DAG.getNode( | |||
| 29061 | ISD::OR, DL, VT, | |||
| 29062 | DAG.getNode(ISD::SHL, DL, VT, R, DAG.getConstant(1, DL, VT)), | |||
| 29063 | DAG.getNode(ISD::SRL, DL, VT, R, DAG.getConstant(7, DL, VT))); | |||
| 29064 | return SignBitSelect(VT, Amt, M, R); | |||
| 29065 | } | |||
| 29066 | ||||
| 29067 | // ISD::ROT* uses modulo rotate amounts. | |||
| 29068 | if (SDValue BaseRotAmt = DAG.getSplatValue(Amt)) { | |||
| 29069 | // If the amount is a splat, perform the modulo BEFORE the splat, | |||
| 29070 | // this helps LowerScalarVariableShift to remove the splat later. | |||
| 29071 | Amt = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VT, BaseRotAmt); | |||
| 29072 | Amt = DAG.getNode(ISD::AND, DL, VT, Amt, | |||
| 29073 | DAG.getConstant(EltSizeInBits - 1, DL, VT)); | |||
| 29074 | Amt = DAG.getVectorShuffle(VT, DL, Amt, DAG.getUNDEF(VT), | |||
| 29075 | SmallVector<int>(NumElts, 0)); | |||
| 29076 | } else { | |||
| 29077 | Amt = DAG.getNode(ISD::AND, DL, VT, Amt, | |||
| 29078 | DAG.getConstant(EltSizeInBits - 1, DL, VT)); | |||
| 29079 | } | |||
| 29080 | ||||
| 29081 | bool ConstantAmt = ISD::isBuildVectorOfConstantSDNodes(Amt.getNode()); | |||
| 29082 | bool LegalVarShifts = SupportedVectorVarShift(VT, Subtarget, ISD::SHL) && | |||
| 29083 | SupportedVectorVarShift(VT, Subtarget, ISD::SRL); | |||
| 29084 | ||||
| 29085 | // Fallback for splats + all supported variable shifts. | |||
| 29086 | // Fallback for non-constants AVX2 vXi16 as well. | |||
| 29087 | if (IsSplatAmt || LegalVarShifts || (Subtarget.hasAVX2() && !ConstantAmt)) { | |||
| 29088 | SDValue AmtR = DAG.getConstant(EltSizeInBits, DL, VT); | |||
| 29089 | AmtR = DAG.getNode(ISD::SUB, DL, VT, AmtR, Amt); | |||
| 29090 | SDValue SHL = DAG.getNode(ISD::SHL, DL, VT, R, Amt); | |||
| 29091 | SDValue SRL = DAG.getNode(ISD::SRL, DL, VT, R, AmtR); | |||
| 29092 | return DAG.getNode(ISD::OR, DL, VT, SHL, SRL); | |||
| 29093 | } | |||
| 29094 | ||||
| 29095 | // As with shifts, convert the rotation amount to a multiplication factor. | |||
| 29096 | SDValue Scale = convertShiftLeftToScale(Amt, DL, Subtarget, DAG); | |||
| 29097 | assert(Scale && "Failed to convert ROTL amount to scale")((void)0); | |||
| 29098 | ||||
| 29099 | // v8i16/v16i16: perform unsigned multiply hi/lo and OR the results. | |||
| 29100 | if (EltSizeInBits == 16) { | |||
| 29101 | SDValue Lo = DAG.getNode(ISD::MUL, DL, VT, R, Scale); | |||
| 29102 | SDValue Hi = DAG.getNode(ISD::MULHU, DL, VT, R, Scale); | |||
| 29103 | return DAG.getNode(ISD::OR, DL, VT, Lo, Hi); | |||
| 29104 | } | |||
| 29105 | ||||
| 29106 | // v4i32: make use of the PMULUDQ instruction to multiply 2 lanes of v4i32 | |||
| 29107 | // to v2i64 results at a time. The upper 32-bits contain the wrapped bits | |||
| 29108 | // that can then be OR'd with the lower 32-bits. | |||
| 29109 | assert(VT == MVT::v4i32 && "Only v4i32 vector rotate expected")((void)0); | |||
| 29110 | static const int OddMask[] = {1, -1, 3, -1}; | |||
| 29111 | SDValue R13 = DAG.getVectorShuffle(VT, DL, R, R, OddMask); | |||
| 29112 | SDValue Scale13 = DAG.getVectorShuffle(VT, DL, Scale, Scale, OddMask); | |||
| 29113 | ||||
| 29114 | SDValue Res02 = DAG.getNode(X86ISD::PMULUDQ, DL, MVT::v2i64, | |||
| 29115 | DAG.getBitcast(MVT::v2i64, R), | |||
| 29116 | DAG.getBitcast(MVT::v2i64, Scale)); | |||
| 29117 | SDValue Res13 = DAG.getNode(X86ISD::PMULUDQ, DL, MVT::v2i64, | |||
| 29118 | DAG.getBitcast(MVT::v2i64, R13), | |||
| 29119 | DAG.getBitcast(MVT::v2i64, Scale13)); | |||
| 29120 | Res02 = DAG.getBitcast(VT, Res02); | |||
| 29121 | Res13 = DAG.getBitcast(VT, Res13); | |||
| 29122 | ||||
| 29123 | return DAG.getNode(ISD::OR, DL, VT, | |||
| 29124 | DAG.getVectorShuffle(VT, DL, Res02, Res13, {0, 4, 2, 6}), | |||
| 29125 | DAG.getVectorShuffle(VT, DL, Res02, Res13, {1, 5, 3, 7})); | |||
| 29126 | } | |||
| 29127 | ||||
| 29128 | /// Returns true if the operand type is exactly twice the native width, and | |||
| 29129 | /// the corresponding cmpxchg8b or cmpxchg16b instruction is available. | |||
| 29130 | /// Used to know whether to use cmpxchg8/16b when expanding atomic operations | |||
| 29131 | /// (otherwise we leave them alone to become __sync_fetch_and_... calls). | |||
| 29132 | bool X86TargetLowering::needsCmpXchgNb(Type *MemType) const { | |||
| 29133 | unsigned OpWidth = MemType->getPrimitiveSizeInBits(); | |||
| 29134 | ||||
| 29135 | if (OpWidth == 64) | |||
| 29136 | return Subtarget.hasCmpxchg8b() && !Subtarget.is64Bit(); | |||
| 29137 | if (OpWidth == 128) | |||
| 29138 | return Subtarget.hasCmpxchg16b(); | |||
| 29139 | ||||
| 29140 | return false; | |||
| 29141 | } | |||
| 29142 | ||||
| 29143 | bool X86TargetLowering::shouldExpandAtomicStoreInIR(StoreInst *SI) const { | |||
| 29144 | Type *MemType = SI->getValueOperand()->getType(); | |||
| 29145 | ||||
| 29146 | bool NoImplicitFloatOps = | |||
| 29147 | SI->getFunction()->hasFnAttribute(Attribute::NoImplicitFloat); | |||
| 29148 | if (MemType->getPrimitiveSizeInBits() == 64 && !Subtarget.is64Bit() && | |||
| 29149 | !Subtarget.useSoftFloat() && !NoImplicitFloatOps && | |||
| 29150 | (Subtarget.hasSSE1() || Subtarget.hasX87())) | |||
| 29151 | return false; | |||
| 29152 | ||||
| 29153 | return needsCmpXchgNb(MemType); | |||
| 29154 | } | |||
| 29155 | ||||
| 29156 | // Note: this turns large loads into lock cmpxchg8b/16b. | |||
| 29157 | // TODO: In 32-bit mode, use MOVLPS when SSE1 is available? | |||
| 29158 | TargetLowering::AtomicExpansionKind | |||
| 29159 | X86TargetLowering::shouldExpandAtomicLoadInIR(LoadInst *LI) const { | |||
| 29160 | Type *MemType = LI->getType(); | |||
| 29161 | ||||
| 29162 | // If this a 64 bit atomic load on a 32-bit target and SSE2 is enabled, we | |||
| 29163 | // can use movq to do the load. If we have X87 we can load into an 80-bit | |||
| 29164 | // X87 register and store it to a stack temporary. | |||
| 29165 | bool NoImplicitFloatOps = | |||
| 29166 | LI->getFunction()->hasFnAttribute(Attribute::NoImplicitFloat); | |||
| 29167 | if (MemType->getPrimitiveSizeInBits() == 64 && !Subtarget.is64Bit() && | |||
| 29168 | !Subtarget.useSoftFloat() && !NoImplicitFloatOps && | |||
| 29169 | (Subtarget.hasSSE1() || Subtarget.hasX87())) | |||
| 29170 | return AtomicExpansionKind::None; | |||
| 29171 | ||||
| 29172 | return needsCmpXchgNb(MemType) ? AtomicExpansionKind::CmpXChg | |||
| 29173 | : AtomicExpansionKind::None; | |||
| 29174 | } | |||
| 29175 | ||||
| 29176 | TargetLowering::AtomicExpansionKind | |||
| 29177 | X86TargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const { | |||
| 29178 | unsigned NativeWidth = Subtarget.is64Bit() ? 64 : 32; | |||
| 29179 | Type *MemType = AI->getType(); | |||
| 29180 | ||||
| 29181 | // If the operand is too big, we must see if cmpxchg8/16b is available | |||
| 29182 | // and default to library calls otherwise. | |||
| 29183 | if (MemType->getPrimitiveSizeInBits() > NativeWidth) { | |||
| 29184 | return needsCmpXchgNb(MemType) ? AtomicExpansionKind::CmpXChg | |||
| 29185 | : AtomicExpansionKind::None; | |||
| 29186 | } | |||
| 29187 | ||||
| 29188 | AtomicRMWInst::BinOp Op = AI->getOperation(); | |||
| 29189 | switch (Op) { | |||
| 29190 | default: | |||
| 29191 | llvm_unreachable("Unknown atomic operation")__builtin_unreachable(); | |||
| 29192 | case AtomicRMWInst::Xchg: | |||
| 29193 | case AtomicRMWInst::Add: | |||
| 29194 | case AtomicRMWInst::Sub: | |||
| 29195 | // It's better to use xadd, xsub or xchg for these in all cases. | |||
| 29196 | return AtomicExpansionKind::None; | |||
| 29197 | case AtomicRMWInst::Or: | |||
| 29198 | case AtomicRMWInst::And: | |||
| 29199 | case AtomicRMWInst::Xor: | |||
| 29200 | // If the atomicrmw's result isn't actually used, we can just add a "lock" | |||
| 29201 | // prefix to a normal instruction for these operations. | |||
| 29202 | return !AI->use_empty() ? AtomicExpansionKind::CmpXChg | |||
| 29203 | : AtomicExpansionKind::None; | |||
| 29204 | case AtomicRMWInst::Nand: | |||
| 29205 | case AtomicRMWInst::Max: | |||
| 29206 | case AtomicRMWInst::Min: | |||
| 29207 | case AtomicRMWInst::UMax: | |||
| 29208 | case AtomicRMWInst::UMin: | |||
| 29209 | case AtomicRMWInst::FAdd: | |||
| 29210 | case AtomicRMWInst::FSub: | |||
| 29211 | // These always require a non-trivial set of data operations on x86. We must | |||
| 29212 | // use a cmpxchg loop. | |||
| 29213 | return AtomicExpansionKind::CmpXChg; | |||
| 29214 | } | |||
| 29215 | } | |||
| 29216 | ||||
| 29217 | LoadInst * | |||
| 29218 | X86TargetLowering::lowerIdempotentRMWIntoFencedLoad(AtomicRMWInst *AI) const { | |||
| 29219 | unsigned NativeWidth = Subtarget.is64Bit() ? 64 : 32; | |||
| 29220 | Type *MemType = AI->getType(); | |||
| 29221 | // Accesses larger than the native width are turned into cmpxchg/libcalls, so | |||
| 29222 | // there is no benefit in turning such RMWs into loads, and it is actually | |||
| 29223 | // harmful as it introduces a mfence. | |||
| 29224 | if (MemType->getPrimitiveSizeInBits() > NativeWidth) | |||
| 29225 | return nullptr; | |||
| 29226 | ||||
| 29227 | // If this is a canonical idempotent atomicrmw w/no uses, we have a better | |||
| 29228 | // lowering available in lowerAtomicArith. | |||
| 29229 | // TODO: push more cases through this path. | |||
| 29230 | if (auto *C = dyn_cast<ConstantInt>(AI->getValOperand())) | |||
| 29231 | if (AI->getOperation() == AtomicRMWInst::Or && C->isZero() && | |||
| 29232 | AI->use_empty()) | |||
| 29233 | return nullptr; | |||
| 29234 | ||||
| 29235 | IRBuilder<> Builder(AI); | |||
| 29236 | Module *M = Builder.GetInsertBlock()->getParent()->getParent(); | |||
| 29237 | auto SSID = AI->getSyncScopeID(); | |||
| 29238 | // We must restrict the ordering to avoid generating loads with Release or | |||
| 29239 | // ReleaseAcquire orderings. | |||
| 29240 | auto Order = AtomicCmpXchgInst::getStrongestFailureOrdering(AI->getOrdering()); | |||
| 29241 | ||||
| 29242 | // Before the load we need a fence. Here is an example lifted from | |||
| 29243 | // http://www.hpl.hp.com/techreports/2012/HPL-2012-68.pdf showing why a fence | |||
| 29244 | // is required: | |||
| 29245 | // Thread 0: | |||
| 29246 | // x.store(1, relaxed); | |||
| 29247 | // r1 = y.fetch_add(0, release); | |||
| 29248 | // Thread 1: | |||
| 29249 | // y.fetch_add(42, acquire); | |||
| 29250 | // r2 = x.load(relaxed); | |||
| 29251 | // r1 = r2 = 0 is impossible, but becomes possible if the idempotent rmw is | |||
| 29252 | // lowered to just a load without a fence. A mfence flushes the store buffer, | |||
| 29253 | // making the optimization clearly correct. | |||
| 29254 | // FIXME: it is required if isReleaseOrStronger(Order) but it is not clear | |||
| 29255 | // otherwise, we might be able to be more aggressive on relaxed idempotent | |||
| 29256 | // rmw. In practice, they do not look useful, so we don't try to be | |||
| 29257 | // especially clever. | |||
| 29258 | if (SSID == SyncScope::SingleThread) | |||
| 29259 | // FIXME: we could just insert an X86ISD::MEMBARRIER here, except we are at | |||
| 29260 | // the IR level, so we must wrap it in an intrinsic. | |||
| 29261 | return nullptr; | |||
| 29262 | ||||
| 29263 | if (!Subtarget.hasMFence()) | |||
| 29264 | // FIXME: it might make sense to use a locked operation here but on a | |||
| 29265 | // different cache-line to prevent cache-line bouncing. In practice it | |||
| 29266 | // is probably a small win, and x86 processors without mfence are rare | |||
| 29267 | // enough that we do not bother. | |||
| 29268 | return nullptr; | |||
| 29269 | ||||
| 29270 | Function *MFence = | |||
| 29271 | llvm::Intrinsic::getDeclaration(M, Intrinsic::x86_sse2_mfence); | |||
| 29272 | Builder.CreateCall(MFence, {}); | |||
| 29273 | ||||
| 29274 | // Finally we can emit the atomic load. | |||
| 29275 | LoadInst *Loaded = Builder.CreateAlignedLoad( | |||
| 29276 | AI->getType(), AI->getPointerOperand(), AI->getAlign()); | |||
| 29277 | Loaded->setAtomic(Order, SSID); | |||
| 29278 | AI->replaceAllUsesWith(Loaded); | |||
| 29279 | AI->eraseFromParent(); | |||
| 29280 | return Loaded; | |||
| 29281 | } | |||
| 29282 | ||||
| 29283 | bool X86TargetLowering::lowerAtomicStoreAsStoreSDNode(const StoreInst &SI) const { | |||
| 29284 | if (!SI.isUnordered()) | |||
| 29285 | return false; | |||
| 29286 | return ExperimentalUnorderedISEL; | |||
| 29287 | } | |||
| 29288 | bool X86TargetLowering::lowerAtomicLoadAsLoadSDNode(const LoadInst &LI) const { | |||
| 29289 | if (!LI.isUnordered()) | |||
| 29290 | return false; | |||
| 29291 | return ExperimentalUnorderedISEL; | |||
| 29292 | } | |||
| 29293 | ||||
| 29294 | ||||
| 29295 | /// Emit a locked operation on a stack location which does not change any | |||
| 29296 | /// memory location, but does involve a lock prefix. Location is chosen to be | |||
| 29297 | /// a) very likely accessed only by a single thread to minimize cache traffic, | |||
| 29298 | /// and b) definitely dereferenceable. Returns the new Chain result. | |||
| 29299 | static SDValue emitLockedStackOp(SelectionDAG &DAG, | |||
| 29300 | const X86Subtarget &Subtarget, SDValue Chain, | |||
| 29301 | const SDLoc &DL) { | |||
| 29302 | // Implementation notes: | |||
| 29303 | // 1) LOCK prefix creates a full read/write reordering barrier for memory | |||
| 29304 | // operations issued by the current processor. As such, the location | |||
| 29305 | // referenced is not relevant for the ordering properties of the instruction. | |||
| 29306 | // See: Intel® 64 and IA-32 ArchitecturesSoftware Developer’s Manual, | |||
| 29307 | // 8.2.3.9 Loads and Stores Are Not Reordered with Locked Instructions | |||
| 29308 | // 2) Using an immediate operand appears to be the best encoding choice | |||
| 29309 | // here since it doesn't require an extra register. | |||
| 29310 | // 3) OR appears to be very slightly faster than ADD. (Though, the difference | |||
| 29311 | // is small enough it might just be measurement noise.) | |||
| 29312 | // 4) When choosing offsets, there are several contributing factors: | |||
| 29313 | // a) If there's no redzone, we default to TOS. (We could allocate a cache | |||
| 29314 | // line aligned stack object to improve this case.) | |||
| 29315 | // b) To minimize our chances of introducing a false dependence, we prefer | |||
| 29316 | // to offset the stack usage from TOS slightly. | |||
| 29317 | // c) To minimize concerns about cross thread stack usage - in particular, | |||
| 29318 | // the idiomatic MyThreadPool.run([&StackVars]() {...}) pattern which | |||
| 29319 | // captures state in the TOS frame and accesses it from many threads - | |||
| 29320 | // we want to use an offset such that the offset is in a distinct cache | |||
| 29321 | // line from the TOS frame. | |||
| 29322 | // | |||
| 29323 | // For a general discussion of the tradeoffs and benchmark results, see: | |||
| 29324 | // https://shipilev.net/blog/2014/on-the-fence-with-dependencies/ | |||
| 29325 | ||||
| 29326 | auto &MF = DAG.getMachineFunction(); | |||
| 29327 | auto &TFL = *Subtarget.getFrameLowering(); | |||
| 29328 | const unsigned SPOffset = TFL.has128ByteRedZone(MF) ? -64 : 0; | |||
| 29329 | ||||
| 29330 | if (Subtarget.is64Bit()) { | |||
| 29331 | SDValue Zero = DAG.getTargetConstant(0, DL, MVT::i32); | |||
| 29332 | SDValue Ops[] = { | |||
| 29333 | DAG.getRegister(X86::RSP, MVT::i64), // Base | |||
| 29334 | DAG.getTargetConstant(1, DL, MVT::i8), // Scale | |||
| 29335 | DAG.getRegister(0, MVT::i64), // Index | |||
| 29336 | DAG.getTargetConstant(SPOffset, DL, MVT::i32), // Disp | |||
| 29337 | DAG.getRegister(0, MVT::i16), // Segment. | |||
| 29338 | Zero, | |||
| 29339 | Chain}; | |||
| 29340 | SDNode *Res = DAG.getMachineNode(X86::OR32mi8Locked, DL, MVT::i32, | |||
| 29341 | MVT::Other, Ops); | |||
| 29342 | return SDValue(Res, 1); | |||
| 29343 | } | |||
| 29344 | ||||
| 29345 | SDValue Zero = DAG.getTargetConstant(0, DL, MVT::i32); | |||
| 29346 | SDValue Ops[] = { | |||
| 29347 | DAG.getRegister(X86::ESP, MVT::i32), // Base | |||
| 29348 | DAG.getTargetConstant(1, DL, MVT::i8), // Scale | |||
| 29349 | DAG.getRegister(0, MVT::i32), // Index | |||
| 29350 | DAG.getTargetConstant(SPOffset, DL, MVT::i32), // Disp | |||
| 29351 | DAG.getRegister(0, MVT::i16), // Segment. | |||
| 29352 | Zero, | |||
| 29353 | Chain | |||
| 29354 | }; | |||
| 29355 | SDNode *Res = DAG.getMachineNode(X86::OR32mi8Locked, DL, MVT::i32, | |||
| 29356 | MVT::Other, Ops); | |||
| 29357 | return SDValue(Res, 1); | |||
| 29358 | } | |||
| 29359 | ||||
| 29360 | static SDValue LowerATOMIC_FENCE(SDValue Op, const X86Subtarget &Subtarget, | |||
| 29361 | SelectionDAG &DAG) { | |||
| 29362 | SDLoc dl(Op); | |||
| 29363 | AtomicOrdering FenceOrdering = | |||
| 29364 | static_cast<AtomicOrdering>(Op.getConstantOperandVal(1)); | |||
| 29365 | SyncScope::ID FenceSSID = | |||
| 29366 | static_cast<SyncScope::ID>(Op.getConstantOperandVal(2)); | |||
| 29367 | ||||
| 29368 | // The only fence that needs an instruction is a sequentially-consistent | |||
| 29369 | // cross-thread fence. | |||
| 29370 | if (FenceOrdering == AtomicOrdering::SequentiallyConsistent && | |||
| 29371 | FenceSSID == SyncScope::System) { | |||
| 29372 | if (Subtarget.hasMFence()) | |||
| 29373 | return DAG.getNode(X86ISD::MFENCE, dl, MVT::Other, Op.getOperand(0)); | |||
| 29374 | ||||
| 29375 | SDValue Chain = Op.getOperand(0); | |||
| 29376 | return emitLockedStackOp(DAG, Subtarget, Chain, dl); | |||
| 29377 | } | |||
| 29378 | ||||
| 29379 | // MEMBARRIER is a compiler barrier; it codegens to a no-op. | |||
| 29380 | return DAG.getNode(X86ISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0)); | |||
| 29381 | } | |||
| 29382 | ||||
| 29383 | static SDValue LowerCMP_SWAP(SDValue Op, const X86Subtarget &Subtarget, | |||
| 29384 | SelectionDAG &DAG) { | |||
| 29385 | MVT T = Op.getSimpleValueType(); | |||
| 29386 | SDLoc DL(Op); | |||
| 29387 | unsigned Reg = 0; | |||
| 29388 | unsigned size = 0; | |||
| 29389 | switch(T.SimpleTy) { | |||
| 29390 | default: llvm_unreachable("Invalid value type!")__builtin_unreachable(); | |||
| 29391 | case MVT::i8: Reg = X86::AL; size = 1; break; | |||
| 29392 | case MVT::i16: Reg = X86::AX; size = 2; break; | |||
| 29393 | case MVT::i32: Reg = X86::EAX; size = 4; break; | |||
| 29394 | case MVT::i64: | |||
| 29395 | assert(Subtarget.is64Bit() && "Node not type legal!")((void)0); | |||
| 29396 | Reg = X86::RAX; size = 8; | |||
| 29397 | break; | |||
| 29398 | } | |||
| 29399 | SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), DL, Reg, | |||
| 29400 | Op.getOperand(2), SDValue()); | |||
| 29401 | SDValue Ops[] = { cpIn.getValue(0), | |||
| 29402 | Op.getOperand(1), | |||
| 29403 | Op.getOperand(3), | |||
| 29404 | DAG.getTargetConstant(size, DL, MVT::i8), | |||
| 29405 | cpIn.getValue(1) }; | |||
| 29406 | SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue); | |||
| 29407 | MachineMemOperand *MMO = cast<AtomicSDNode>(Op)->getMemOperand(); | |||
| 29408 | SDValue Result = DAG.getMemIntrinsicNode(X86ISD::LCMPXCHG_DAG, DL, Tys, | |||
| 29409 | Ops, T, MMO); | |||
| 29410 | ||||
| 29411 | SDValue cpOut = | |||
| 29412 | DAG.getCopyFromReg(Result.getValue(0), DL, Reg, T, Result.getValue(1)); | |||
| 29413 | SDValue EFLAGS = DAG.getCopyFromReg(cpOut.getValue(1), DL, X86::EFLAGS, | |||
| 29414 | MVT::i32, cpOut.getValue(2)); | |||
| 29415 | SDValue Success = getSETCC(X86::COND_E, EFLAGS, DL, DAG); | |||
| 29416 | ||||
| 29417 | return DAG.getNode(ISD::MERGE_VALUES, DL, Op->getVTList(), | |||
| 29418 | cpOut, Success, EFLAGS.getValue(1)); | |||
| 29419 | } | |||
| 29420 | ||||
| 29421 | // Create MOVMSKB, taking into account whether we need to split for AVX1. | |||
| 29422 | static SDValue getPMOVMSKB(const SDLoc &DL, SDValue V, SelectionDAG &DAG, | |||
| 29423 | const X86Subtarget &Subtarget) { | |||
| 29424 | MVT InVT = V.getSimpleValueType(); | |||
| 29425 | ||||
| 29426 | if (InVT == MVT::v64i8) { | |||
| 29427 | SDValue Lo, Hi; | |||
| 29428 | std::tie(Lo, Hi) = DAG.SplitVector(V, DL); | |||
| 29429 | Lo = getPMOVMSKB(DL, Lo, DAG, Subtarget); | |||
| 29430 | Hi = getPMOVMSKB(DL, Hi, DAG, Subtarget); | |||
| 29431 | Lo = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i64, Lo); | |||
| 29432 | Hi = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, Hi); | |||
| 29433 | Hi = DAG.getNode(ISD::SHL, DL, MVT::i64, Hi, | |||
| 29434 | DAG.getConstant(32, DL, MVT::i8)); | |||
| 29435 | return DAG.getNode(ISD::OR, DL, MVT::i64, Lo, Hi); | |||
| 29436 | } | |||
| 29437 | if (InVT == MVT::v32i8 && !Subtarget.hasInt256()) { | |||
| 29438 | SDValue Lo, Hi; | |||
| 29439 | std::tie(Lo, Hi) = DAG.SplitVector(V, DL); | |||
| 29440 | Lo = DAG.getNode(X86ISD::MOVMSK, DL, MVT::i32, Lo); | |||
| 29441 | Hi = DAG.getNode(X86ISD::MOVMSK, DL, MVT::i32, Hi); | |||
| 29442 | Hi = DAG.getNode(ISD::SHL, DL, MVT::i32, Hi, | |||
| 29443 | DAG.getConstant(16, DL, MVT::i8)); | |||
| 29444 | return DAG.getNode(ISD::OR, DL, MVT::i32, Lo, Hi); | |||
| 29445 | } | |||
| 29446 | ||||
| 29447 | return DAG.getNode(X86ISD::MOVMSK, DL, MVT::i32, V); | |||
| 29448 | } | |||
| 29449 | ||||
| 29450 | static SDValue LowerBITCAST(SDValue Op, const X86Subtarget &Subtarget, | |||
| 29451 | SelectionDAG &DAG) { | |||
| 29452 | SDValue Src = Op.getOperand(0); | |||
| 29453 | MVT SrcVT = Src.getSimpleValueType(); | |||
| 29454 | MVT DstVT = Op.getSimpleValueType(); | |||
| 29455 | ||||
| 29456 | // Legalize (v64i1 (bitcast i64 (X))) by splitting the i64, bitcasting each | |||
| 29457 | // half to v32i1 and concatenating the result. | |||
| 29458 | if (SrcVT == MVT::i64 && DstVT == MVT::v64i1) { | |||
| 29459 | assert(!Subtarget.is64Bit() && "Expected 32-bit mode")((void)0); | |||
| 29460 | assert(Subtarget.hasBWI() && "Expected BWI target")((void)0); | |||
| 29461 | SDLoc dl(Op); | |||
| 29462 | SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Src, | |||
| 29463 | DAG.getIntPtrConstant(0, dl)); | |||
| 29464 | Lo = DAG.getBitcast(MVT::v32i1, Lo); | |||
| 29465 | SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Src, | |||
| 29466 | DAG.getIntPtrConstant(1, dl)); | |||
| 29467 | Hi = DAG.getBitcast(MVT::v32i1, Hi); | |||
| 29468 | return DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v64i1, Lo, Hi); | |||
| 29469 | } | |||
| 29470 | ||||
| 29471 | // Use MOVMSK for vector to scalar conversion to prevent scalarization. | |||
| 29472 | if ((SrcVT == MVT::v16i1 || SrcVT == MVT::v32i1) && DstVT.isScalarInteger()) { | |||
| 29473 | assert(!Subtarget.hasAVX512() && "Should use K-registers with AVX512")((void)0); | |||
| 29474 | MVT SExtVT = SrcVT == MVT::v16i1 ? MVT::v16i8 : MVT::v32i8; | |||
| 29475 | SDLoc DL(Op); | |||
| 29476 | SDValue V = DAG.getSExtOrTrunc(Src, DL, SExtVT); | |||
| 29477 | V = getPMOVMSKB(DL, V, DAG, Subtarget); | |||
| 29478 | return DAG.getZExtOrTrunc(V, DL, DstVT); | |||
| 29479 | } | |||
| 29480 | ||||
| 29481 | assert((SrcVT == MVT::v2i32 || SrcVT == MVT::v4i16 || SrcVT == MVT::v8i8 ||((void)0) | |||
| 29482 | SrcVT == MVT::i64) && "Unexpected VT!")((void)0); | |||
| 29483 | ||||
| 29484 | assert(Subtarget.hasSSE2() && "Requires at least SSE2!")((void)0); | |||
| 29485 | if (!(DstVT == MVT::f64 && SrcVT == MVT::i64) && | |||
| 29486 | !(DstVT == MVT::x86mmx && SrcVT.isVector())) | |||
| 29487 | // This conversion needs to be expanded. | |||
| 29488 | return SDValue(); | |||
| 29489 | ||||
| 29490 | SDLoc dl(Op); | |||
| 29491 | if (SrcVT.isVector()) { | |||
| 29492 | // Widen the vector in input in the case of MVT::v2i32. | |||
| 29493 | // Example: from MVT::v2i32 to MVT::v4i32. | |||
| 29494 | MVT NewVT = MVT::getVectorVT(SrcVT.getVectorElementType(), | |||
| 29495 | SrcVT.getVectorNumElements() * 2); | |||
| 29496 | Src = DAG.getNode(ISD::CONCAT_VECTORS, dl, NewVT, Src, | |||
| 29497 | DAG.getUNDEF(SrcVT)); | |||
| 29498 | } else { | |||
| 29499 | assert(SrcVT == MVT::i64 && !Subtarget.is64Bit() &&((void)0) | |||
| 29500 | "Unexpected source type in LowerBITCAST")((void)0); | |||
| 29501 | Src = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Src); | |||
| 29502 | } | |||
| 29503 | ||||
| 29504 | MVT V2X64VT = DstVT == MVT::f64 ? MVT::v2f64 : MVT::v2i64; | |||
| 29505 | Src = DAG.getNode(ISD::BITCAST, dl, V2X64VT, Src); | |||
| 29506 | ||||
| 29507 | if (DstVT == MVT::x86mmx) | |||
| 29508 | return DAG.getNode(X86ISD::MOVDQ2Q, dl, DstVT, Src); | |||
| 29509 | ||||
| 29510 | return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, DstVT, Src, | |||
| 29511 | DAG.getIntPtrConstant(0, dl)); | |||
| 29512 | } | |||
| 29513 | ||||
| 29514 | /// Compute the horizontal sum of bytes in V for the elements of VT. | |||
| 29515 | /// | |||
| 29516 | /// Requires V to be a byte vector and VT to be an integer vector type with | |||
| 29517 | /// wider elements than V's type. The width of the elements of VT determines | |||
| 29518 | /// how many bytes of V are summed horizontally to produce each element of the | |||
| 29519 | /// result. | |||
| 29520 | static SDValue LowerHorizontalByteSum(SDValue V, MVT VT, | |||
| 29521 | const X86Subtarget &Subtarget, | |||
| 29522 | SelectionDAG &DAG) { | |||
| 29523 | SDLoc DL(V); | |||
| 29524 | MVT ByteVecVT = V.getSimpleValueType(); | |||
| 29525 | MVT EltVT = VT.getVectorElementType(); | |||
| 29526 | assert(ByteVecVT.getVectorElementType() == MVT::i8 &&((void)0) | |||
| 29527 | "Expected value to have byte element type.")((void)0); | |||
| 29528 | assert(EltVT != MVT::i8 &&((void)0) | |||
| 29529 | "Horizontal byte sum only makes sense for wider elements!")((void)0); | |||
| 29530 | unsigned VecSize = VT.getSizeInBits(); | |||
| 29531 | assert(ByteVecVT.getSizeInBits() == VecSize && "Cannot change vector size!")((void)0); | |||
| 29532 | ||||
| 29533 | // PSADBW instruction horizontally add all bytes and leave the result in i64 | |||
| 29534 | // chunks, thus directly computes the pop count for v2i64 and v4i64. | |||
| 29535 | if (EltVT == MVT::i64) { | |||
| 29536 | SDValue Zeros = DAG.getConstant(0, DL, ByteVecVT); | |||
| 29537 | MVT SadVecVT = MVT::getVectorVT(MVT::i64, VecSize / 64); | |||
| 29538 | V = DAG.getNode(X86ISD::PSADBW, DL, SadVecVT, V, Zeros); | |||
| 29539 | return DAG.getBitcast(VT, V); | |||
| 29540 | } | |||
| 29541 | ||||
| 29542 | if (EltVT == MVT::i32) { | |||
| 29543 | // We unpack the low half and high half into i32s interleaved with zeros so | |||
| 29544 | // that we can use PSADBW to horizontally sum them. The most useful part of | |||
| 29545 | // this is that it lines up the results of two PSADBW instructions to be | |||
| 29546 | // two v2i64 vectors which concatenated are the 4 population counts. We can | |||
| 29547 | // then use PACKUSWB to shrink and concatenate them into a v4i32 again. | |||
| 29548 | SDValue Zeros = DAG.getConstant(0, DL, VT); | |||
| 29549 | SDValue V32 = DAG.getBitcast(VT, V); | |||
| 29550 | SDValue Low = getUnpackl(DAG, DL, VT, V32, Zeros); | |||
| 29551 | SDValue High = getUnpackh(DAG, DL, VT, V32, Zeros); | |||
| 29552 | ||||
| 29553 | // Do the horizontal sums into two v2i64s. | |||
| 29554 | Zeros = DAG.getConstant(0, DL, ByteVecVT); | |||
| 29555 | MVT SadVecVT = MVT::getVectorVT(MVT::i64, VecSize / 64); | |||
| 29556 | Low = DAG.getNode(X86ISD::PSADBW, DL, SadVecVT, | |||
| 29557 | DAG.getBitcast(ByteVecVT, Low), Zeros); | |||
| 29558 | High = DAG.getNode(X86ISD::PSADBW, DL, SadVecVT, | |||
| 29559 | DAG.getBitcast(ByteVecVT, High), Zeros); | |||
| 29560 | ||||
| 29561 | // Merge them together. | |||
| 29562 | MVT ShortVecVT = MVT::getVectorVT(MVT::i16, VecSize / 16); | |||
| 29563 | V = DAG.getNode(X86ISD::PACKUS, DL, ByteVecVT, | |||
| 29564 | DAG.getBitcast(ShortVecVT, Low), | |||
| 29565 | DAG.getBitcast(ShortVecVT, High)); | |||
| 29566 | ||||
| 29567 | return DAG.getBitcast(VT, V); | |||
| 29568 | } | |||
| 29569 | ||||
| 29570 | // The only element type left is i16. | |||
| 29571 | assert(EltVT == MVT::i16 && "Unknown how to handle type")((void)0); | |||
| 29572 | ||||
| 29573 | // To obtain pop count for each i16 element starting from the pop count for | |||
| 29574 | // i8 elements, shift the i16s left by 8, sum as i8s, and then shift as i16s | |||
| 29575 | // right by 8. It is important to shift as i16s as i8 vector shift isn't | |||
| 29576 | // directly supported. | |||
| 29577 | SDValue ShifterV = DAG.getConstant(8, DL, VT); | |||
| 29578 | SDValue Shl = DAG.getNode(ISD::SHL, DL, VT, DAG.getBitcast(VT, V), ShifterV); | |||
| 29579 | V = DAG.getNode(ISD::ADD, DL, ByteVecVT, DAG.getBitcast(ByteVecVT, Shl), | |||
| 29580 | DAG.getBitcast(ByteVecVT, V)); | |||
| 29581 | return DAG.getNode(ISD::SRL, DL, VT, DAG.getBitcast(VT, V), ShifterV); | |||
| 29582 | } | |||
| 29583 | ||||
| 29584 | static SDValue LowerVectorCTPOPInRegLUT(SDValue Op, const SDLoc &DL, | |||
| 29585 | const X86Subtarget &Subtarget, | |||
| 29586 | SelectionDAG &DAG) { | |||
| 29587 | MVT VT = Op.getSimpleValueType(); | |||
| 29588 | MVT EltVT = VT.getVectorElementType(); | |||
| 29589 | int NumElts = VT.getVectorNumElements(); | |||
| 29590 | (void)EltVT; | |||
| 29591 | assert(EltVT == MVT::i8 && "Only vXi8 vector CTPOP lowering supported.")((void)0); | |||
| 29592 | ||||
| 29593 | // Implement a lookup table in register by using an algorithm based on: | |||
| 29594 | // http://wm.ite.pl/articles/sse-popcount.html | |||
| 29595 | // | |||
| 29596 | // The general idea is that every lower byte nibble in the input vector is an | |||
| 29597 | // index into a in-register pre-computed pop count table. We then split up the | |||
| 29598 | // input vector in two new ones: (1) a vector with only the shifted-right | |||
| 29599 | // higher nibbles for each byte and (2) a vector with the lower nibbles (and | |||
| 29600 | // masked out higher ones) for each byte. PSHUFB is used separately with both | |||
| 29601 | // to index the in-register table. Next, both are added and the result is a | |||
| 29602 | // i8 vector where each element contains the pop count for input byte. | |||
| 29603 | const int LUT[16] = {/* 0 */ 0, /* 1 */ 1, /* 2 */ 1, /* 3 */ 2, | |||
| 29604 | /* 4 */ 1, /* 5 */ 2, /* 6 */ 2, /* 7 */ 3, | |||
| 29605 | /* 8 */ 1, /* 9 */ 2, /* a */ 2, /* b */ 3, | |||
| 29606 | /* c */ 2, /* d */ 3, /* e */ 3, /* f */ 4}; | |||
| 29607 | ||||
| 29608 | SmallVector<SDValue, 64> LUTVec; | |||
| 29609 | for (int i = 0; i < NumElts; ++i) | |||
| 29610 | LUTVec.push_back(DAG.getConstant(LUT[i % 16], DL, MVT::i8)); | |||
| 29611 | SDValue InRegLUT = DAG.getBuildVector(VT, DL, LUTVec); | |||
| 29612 | SDValue M0F = DAG.getConstant(0x0F, DL, VT); | |||
| 29613 | ||||
| 29614 | // High nibbles | |||
| 29615 | SDValue FourV = DAG.getConstant(4, DL, VT); | |||
| 29616 | SDValue HiNibbles = DAG.getNode(ISD::SRL, DL, VT, Op, FourV); | |||
| 29617 | ||||
| 29618 | // Low nibbles | |||
| 29619 | SDValue LoNibbles = DAG.getNode(ISD::AND, DL, VT, Op, M0F); | |||
| 29620 | ||||
| 29621 | // The input vector is used as the shuffle mask that index elements into the | |||
| 29622 | // LUT. After counting low and high nibbles, add the vector to obtain the | |||
| 29623 | // final pop count per i8 element. | |||
| 29624 | SDValue HiPopCnt = DAG.getNode(X86ISD::PSHUFB, DL, VT, InRegLUT, HiNibbles); | |||
| 29625 | SDValue LoPopCnt = DAG.getNode(X86ISD::PSHUFB, DL, VT, InRegLUT, LoNibbles); | |||
| 29626 | return DAG.getNode(ISD::ADD, DL, VT, HiPopCnt, LoPopCnt); | |||
| 29627 | } | |||
| 29628 | ||||
| 29629 | // Please ensure that any codegen change from LowerVectorCTPOP is reflected in | |||
| 29630 | // updated cost models in X86TTIImpl::getIntrinsicInstrCost. | |||
| 29631 | static SDValue LowerVectorCTPOP(SDValue Op, const X86Subtarget &Subtarget, | |||
| 29632 | SelectionDAG &DAG) { | |||
| 29633 | MVT VT = Op.getSimpleValueType(); | |||
| 29634 | assert((VT.is512BitVector() || VT.is256BitVector() || VT.is128BitVector()) &&((void)0) | |||
| 29635 | "Unknown CTPOP type to handle")((void)0); | |||
| 29636 | SDLoc DL(Op.getNode()); | |||
| 29637 | SDValue Op0 = Op.getOperand(0); | |||
| 29638 | ||||
| 29639 | // TRUNC(CTPOP(ZEXT(X))) to make use of vXi32/vXi64 VPOPCNT instructions. | |||
| 29640 | if (Subtarget.hasVPOPCNTDQ()) { | |||
| 29641 | unsigned NumElems = VT.getVectorNumElements(); | |||
| 29642 | assert((VT.getVectorElementType() == MVT::i8 ||((void)0) | |||
| 29643 | VT.getVectorElementType() == MVT::i16) && "Unexpected type")((void)0); | |||
| 29644 | if (NumElems < 16 || (NumElems == 16 && Subtarget.canExtendTo512DQ())) { | |||
| 29645 | MVT NewVT = MVT::getVectorVT(MVT::i32, NumElems); | |||
| 29646 | Op = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, Op0); | |||
| 29647 | Op = DAG.getNode(ISD::CTPOP, DL, NewVT, Op); | |||
| 29648 | return DAG.getNode(ISD::TRUNCATE, DL, VT, Op); | |||
| 29649 | } | |||
| 29650 | } | |||
| 29651 | ||||
| 29652 | // Decompose 256-bit ops into smaller 128-bit ops. | |||
| 29653 | if (VT.is256BitVector() && !Subtarget.hasInt256()) | |||
| 29654 | return splitVectorIntUnary(Op, DAG); | |||
| 29655 | ||||
| 29656 | // Decompose 512-bit ops into smaller 256-bit ops. | |||
| 29657 | if (VT.is512BitVector() && !Subtarget.hasBWI()) | |||
| 29658 | return splitVectorIntUnary(Op, DAG); | |||
| 29659 | ||||
| 29660 | // For element types greater than i8, do vXi8 pop counts and a bytesum. | |||
| 29661 | if (VT.getScalarType() != MVT::i8) { | |||
| 29662 | MVT ByteVT = MVT::getVectorVT(MVT::i8, VT.getSizeInBits() / 8); | |||
| 29663 | SDValue ByteOp = DAG.getBitcast(ByteVT, Op0); | |||
| 29664 | SDValue PopCnt8 = DAG.getNode(ISD::CTPOP, DL, ByteVT, ByteOp); | |||
| 29665 | return LowerHorizontalByteSum(PopCnt8, VT, Subtarget, DAG); | |||
| 29666 | } | |||
| 29667 | ||||
| 29668 | // We can't use the fast LUT approach, so fall back on LegalizeDAG. | |||
| 29669 | if (!Subtarget.hasSSSE3()) | |||
| 29670 | return SDValue(); | |||
| 29671 | ||||
| 29672 | return LowerVectorCTPOPInRegLUT(Op0, DL, Subtarget, DAG); | |||
| 29673 | } | |||
| 29674 | ||||
| 29675 | static SDValue LowerCTPOP(SDValue Op, const X86Subtarget &Subtarget, | |||
| 29676 | SelectionDAG &DAG) { | |||
| 29677 | assert(Op.getSimpleValueType().isVector() &&((void)0) | |||
| 29678 | "We only do custom lowering for vector population count.")((void)0); | |||
| 29679 | return LowerVectorCTPOP(Op, Subtarget, DAG); | |||
| 29680 | } | |||
| 29681 | ||||
| 29682 | static SDValue LowerBITREVERSE_XOP(SDValue Op, SelectionDAG &DAG) { | |||
| 29683 | MVT VT = Op.getSimpleValueType(); | |||
| 29684 | SDValue In = Op.getOperand(0); | |||
| 29685 | SDLoc DL(Op); | |||
| 29686 | ||||
| 29687 | // For scalars, its still beneficial to transfer to/from the SIMD unit to | |||
| 29688 | // perform the BITREVERSE. | |||
| 29689 | if (!VT.isVector()) { | |||
| 29690 | MVT VecVT = MVT::getVectorVT(VT, 128 / VT.getSizeInBits()); | |||
| 29691 | SDValue Res = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, In); | |||
| 29692 | Res = DAG.getNode(ISD::BITREVERSE, DL, VecVT, Res); | |||
| 29693 | return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT, Res, | |||
| 29694 | DAG.getIntPtrConstant(0, DL)); | |||
| 29695 | } | |||
| 29696 | ||||
| 29697 | int NumElts = VT.getVectorNumElements(); | |||
| 29698 | int ScalarSizeInBytes = VT.getScalarSizeInBits() / 8; | |||
| 29699 | ||||
| 29700 | // Decompose 256-bit ops into smaller 128-bit ops. | |||
| 29701 | if (VT.is256BitVector()) | |||
| 29702 | return splitVectorIntUnary(Op, DAG); | |||
| 29703 | ||||
| 29704 | assert(VT.is128BitVector() &&((void)0) | |||
| 29705 | "Only 128-bit vector bitreverse lowering supported.")((void)0); | |||
| 29706 | ||||
| 29707 | // VPPERM reverses the bits of a byte with the permute Op (2 << 5), and we | |||
| 29708 | // perform the BSWAP in the shuffle. | |||
| 29709 | // Its best to shuffle using the second operand as this will implicitly allow | |||
| 29710 | // memory folding for multiple vectors. | |||
| 29711 | SmallVector<SDValue, 16> MaskElts; | |||
| 29712 | for (int i = 0; i != NumElts; ++i) { | |||
| 29713 | for (int j = ScalarSizeInBytes - 1; j >= 0; --j) { | |||
| 29714 | int SourceByte = 16 + (i * ScalarSizeInBytes) + j; | |||
| 29715 | int PermuteByte = SourceByte | (2 << 5); | |||
| 29716 | MaskElts.push_back(DAG.getConstant(PermuteByte, DL, MVT::i8)); | |||
| 29717 | } | |||
| 29718 | } | |||
| 29719 | ||||
| 29720 | SDValue Mask = DAG.getBuildVector(MVT::v16i8, DL, MaskElts); | |||
| 29721 | SDValue Res = DAG.getBitcast(MVT::v16i8, In); | |||
| 29722 | Res = DAG.getNode(X86ISD::VPPERM, DL, MVT::v16i8, DAG.getUNDEF(MVT::v16i8), | |||
| 29723 | Res, Mask); | |||
| 29724 | return DAG.getBitcast(VT, Res); | |||
| 29725 | } | |||
| 29726 | ||||
| 29727 | static SDValue LowerBITREVERSE(SDValue Op, const X86Subtarget &Subtarget, | |||
| 29728 | SelectionDAG &DAG) { | |||
| 29729 | MVT VT = Op.getSimpleValueType(); | |||
| 29730 | ||||
| 29731 | if (Subtarget.hasXOP() && !VT.is512BitVector()) | |||
| 29732 | return LowerBITREVERSE_XOP(Op, DAG); | |||
| 29733 | ||||
| 29734 | assert(Subtarget.hasSSSE3() && "SSSE3 required for BITREVERSE")((void)0); | |||
| 29735 | ||||
| 29736 | SDValue In = Op.getOperand(0); | |||
| 29737 | SDLoc DL(Op); | |||
| 29738 | ||||
| 29739 | assert(VT.getScalarType() == MVT::i8 &&((void)0) | |||
| 29740 | "Only byte vector BITREVERSE supported")((void)0); | |||
| 29741 | ||||
| 29742 | // Split v64i8 without BWI so that we can still use the PSHUFB lowering. | |||
| 29743 | if (VT == MVT::v64i8 && !Subtarget.hasBWI()) | |||
| 29744 | return splitVectorIntUnary(Op, DAG); | |||
| 29745 | ||||
| 29746 | // Decompose 256-bit ops into smaller 128-bit ops on pre-AVX2. | |||
| 29747 | if (VT == MVT::v32i8 && !Subtarget.hasInt256()) | |||
| 29748 | return splitVectorIntUnary(Op, DAG); | |||
| 29749 | ||||
| 29750 | unsigned NumElts = VT.getVectorNumElements(); | |||
| 29751 | ||||
| 29752 | // If we have GFNI, we can use GF2P8AFFINEQB to reverse the bits. | |||
| 29753 | if (Subtarget.hasGFNI()) { | |||
| 29754 | MVT MatrixVT = MVT::getVectorVT(MVT::i64, NumElts / 8); | |||
| 29755 | SDValue Matrix = DAG.getConstant(0x8040201008040201ULL, DL, MatrixVT); | |||
| 29756 | Matrix = DAG.getBitcast(VT, Matrix); | |||
| 29757 | return DAG.getNode(X86ISD::GF2P8AFFINEQB, DL, VT, In, Matrix, | |||
| 29758 | DAG.getTargetConstant(0, DL, MVT::i8)); | |||
| 29759 | } | |||
| 29760 | ||||
| 29761 | // Perform BITREVERSE using PSHUFB lookups. Each byte is split into | |||
| 29762 | // two nibbles and a PSHUFB lookup to find the bitreverse of each | |||
| 29763 | // 0-15 value (moved to the other nibble). | |||
| 29764 | SDValue NibbleMask = DAG.getConstant(0xF, DL, VT); | |||
| 29765 | SDValue Lo = DAG.getNode(ISD::AND, DL, VT, In, NibbleMask); | |||
| 29766 | SDValue Hi = DAG.getNode(ISD::SRL, DL, VT, In, DAG.getConstant(4, DL, VT)); | |||
| 29767 | ||||
| 29768 | const int LoLUT[16] = { | |||
| 29769 | /* 0 */ 0x00, /* 1 */ 0x80, /* 2 */ 0x40, /* 3 */ 0xC0, | |||
| 29770 | /* 4 */ 0x20, /* 5 */ 0xA0, /* 6 */ 0x60, /* 7 */ 0xE0, | |||
| 29771 | /* 8 */ 0x10, /* 9 */ 0x90, /* a */ 0x50, /* b */ 0xD0, | |||
| 29772 | /* c */ 0x30, /* d */ 0xB0, /* e */ 0x70, /* f */ 0xF0}; | |||
| 29773 | const int HiLUT[16] = { | |||
| 29774 | /* 0 */ 0x00, /* 1 */ 0x08, /* 2 */ 0x04, /* 3 */ 0x0C, | |||
| 29775 | /* 4 */ 0x02, /* 5 */ 0x0A, /* 6 */ 0x06, /* 7 */ 0x0E, | |||
| 29776 | /* 8 */ 0x01, /* 9 */ 0x09, /* a */ 0x05, /* b */ 0x0D, | |||
| 29777 | /* c */ 0x03, /* d */ 0x0B, /* e */ 0x07, /* f */ 0x0F}; | |||
| 29778 | ||||
| 29779 | SmallVector<SDValue, 16> LoMaskElts, HiMaskElts; | |||
| 29780 | for (unsigned i = 0; i < NumElts; ++i) { | |||
| 29781 | LoMaskElts.push_back(DAG.getConstant(LoLUT[i % 16], DL, MVT::i8)); | |||
| 29782 | HiMaskElts.push_back(DAG.getConstant(HiLUT[i % 16], DL, MVT::i8)); | |||
| 29783 | } | |||
| 29784 | ||||
| 29785 | SDValue LoMask = DAG.getBuildVector(VT, DL, LoMaskElts); | |||
| 29786 | SDValue HiMask = DAG.getBuildVector(VT, DL, HiMaskElts); | |||
| 29787 | Lo = DAG.getNode(X86ISD::PSHUFB, DL, VT, LoMask, Lo); | |||
| 29788 | Hi = DAG.getNode(X86ISD::PSHUFB, DL, VT, HiMask, Hi); | |||
| 29789 | return DAG.getNode(ISD::OR, DL, VT, Lo, Hi); | |||
| 29790 | } | |||
| 29791 | ||||
| 29792 | static SDValue LowerPARITY(SDValue Op, const X86Subtarget &Subtarget, | |||
| 29793 | SelectionDAG &DAG) { | |||
| 29794 | SDLoc DL(Op); | |||
| 29795 | SDValue X = Op.getOperand(0); | |||
| 29796 | MVT VT = Op.getSimpleValueType(); | |||
| 29797 | ||||
| 29798 | // Special case. If the input fits in 8-bits we can use a single 8-bit TEST. | |||
| 29799 | if (VT == MVT::i8 || | |||
| 29800 | DAG.MaskedValueIsZero(X, APInt::getBitsSetFrom(VT.getSizeInBits(), 8))) { | |||
| 29801 | X = DAG.getNode(ISD::TRUNCATE, DL, MVT::i8, X); | |||
| 29802 | SDValue Flags = DAG.getNode(X86ISD::CMP, DL, MVT::i32, X, | |||
| 29803 | DAG.getConstant(0, DL, MVT::i8)); | |||
| 29804 | // Copy the inverse of the parity flag into a register with setcc. | |||
| 29805 | SDValue Setnp = getSETCC(X86::COND_NP, Flags, DL, DAG); | |||
| 29806 | // Extend to the original type. | |||
| 29807 | return DAG.getNode(ISD::ZERO_EXTEND, DL, VT, Setnp); | |||
| 29808 | } | |||
| 29809 | ||||
| 29810 | if (VT == MVT::i64) { | |||
| 29811 | // Xor the high and low 16-bits together using a 32-bit operation. | |||
| 29812 | SDValue Hi = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, | |||
| 29813 | DAG.getNode(ISD::SRL, DL, MVT::i64, X, | |||
| 29814 | DAG.getConstant(32, DL, MVT::i8))); | |||
| 29815 | SDValue Lo = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, X); | |||
| 29816 | X = DAG.getNode(ISD::XOR, DL, MVT::i32, Lo, Hi); | |||
| 29817 | } | |||
| 29818 | ||||
| 29819 | if (VT != MVT::i16) { | |||
| 29820 | // Xor the high and low 16-bits together using a 32-bit operation. | |||
| 29821 | SDValue Hi16 = DAG.getNode(ISD::SRL, DL, MVT::i32, X, | |||
| 29822 | DAG.getConstant(16, DL, MVT::i8)); | |||
| 29823 | X = DAG.getNode(ISD::XOR, DL, MVT::i32, X, Hi16); | |||
| 29824 | } else { | |||
| 29825 | // If the input is 16-bits, we need to extend to use an i32 shift below. | |||
| 29826 | X = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, X); | |||
| 29827 | } | |||
| 29828 | ||||
| 29829 | // Finally xor the low 2 bytes together and use a 8-bit flag setting xor. | |||
| 29830 | // This should allow an h-reg to be used to save a shift. | |||
| 29831 | SDValue Hi = DAG.getNode( | |||
| 29832 | ISD::TRUNCATE, DL, MVT::i8, | |||
| 29833 | DAG.getNode(ISD::SRL, DL, MVT::i32, X, DAG.getConstant(8, DL, MVT::i8))); | |||
| 29834 | SDValue Lo = DAG.getNode(ISD::TRUNCATE, DL, MVT::i8, X); | |||
| 29835 | SDVTList VTs = DAG.getVTList(MVT::i8, MVT::i32); | |||
| 29836 | SDValue Flags = DAG.getNode(X86ISD::XOR, DL, VTs, Lo, Hi).getValue(1); | |||
| 29837 | ||||
| 29838 | // Copy the inverse of the parity flag into a register with setcc. | |||
| 29839 | SDValue Setnp = getSETCC(X86::COND_NP, Flags, DL, DAG); | |||
| 29840 | // Extend to the original type. | |||
| 29841 | return DAG.getNode(ISD::ZERO_EXTEND, DL, VT, Setnp); | |||
| 29842 | } | |||
| 29843 | ||||
| 29844 | static SDValue lowerAtomicArithWithLOCK(SDValue N, SelectionDAG &DAG, | |||
| 29845 | const X86Subtarget &Subtarget) { | |||
| 29846 | unsigned NewOpc = 0; | |||
| 29847 | switch (N->getOpcode()) { | |||
| 29848 | case ISD::ATOMIC_LOAD_ADD: | |||
| 29849 | NewOpc = X86ISD::LADD; | |||
| 29850 | break; | |||
| 29851 | case ISD::ATOMIC_LOAD_SUB: | |||
| 29852 | NewOpc = X86ISD::LSUB; | |||
| 29853 | break; | |||
| 29854 | case ISD::ATOMIC_LOAD_OR: | |||
| 29855 | NewOpc = X86ISD::LOR; | |||
| 29856 | break; | |||
| 29857 | case ISD::ATOMIC_LOAD_XOR: | |||
| 29858 | NewOpc = X86ISD::LXOR; | |||
| 29859 | break; | |||
| 29860 | case ISD::ATOMIC_LOAD_AND: | |||
| 29861 | NewOpc = X86ISD::LAND; | |||
| 29862 | break; | |||
| 29863 | default: | |||
| 29864 | llvm_unreachable("Unknown ATOMIC_LOAD_ opcode")__builtin_unreachable(); | |||
| 29865 | } | |||
| 29866 | ||||
| 29867 | MachineMemOperand *MMO = cast<MemSDNode>(N)->getMemOperand(); | |||
| 29868 | ||||
| 29869 | return DAG.getMemIntrinsicNode( | |||
| 29870 | NewOpc, SDLoc(N), DAG.getVTList(MVT::i32, MVT::Other), | |||
| 29871 | {N->getOperand(0), N->getOperand(1), N->getOperand(2)}, | |||
| 29872 | /*MemVT=*/N->getSimpleValueType(0), MMO); | |||
| 29873 | } | |||
| 29874 | ||||
| 29875 | /// Lower atomic_load_ops into LOCK-prefixed operations. | |||
| 29876 | static SDValue lowerAtomicArith(SDValue N, SelectionDAG &DAG, | |||
| 29877 | const X86Subtarget &Subtarget) { | |||
| 29878 | AtomicSDNode *AN = cast<AtomicSDNode>(N.getNode()); | |||
| 29879 | SDValue Chain = N->getOperand(0); | |||
| 29880 | SDValue LHS = N->getOperand(1); | |||
| 29881 | SDValue RHS = N->getOperand(2); | |||
| 29882 | unsigned Opc = N->getOpcode(); | |||
| 29883 | MVT VT = N->getSimpleValueType(0); | |||
| 29884 | SDLoc DL(N); | |||
| 29885 | ||||
| 29886 | // We can lower atomic_load_add into LXADD. However, any other atomicrmw op | |||
| 29887 | // can only be lowered when the result is unused. They should have already | |||
| 29888 | // been transformed into a cmpxchg loop in AtomicExpand. | |||
| 29889 | if (N->hasAnyUseOfValue(0)) { | |||
| 29890 | // Handle (atomic_load_sub p, v) as (atomic_load_add p, -v), to be able to | |||
| 29891 | // select LXADD if LOCK_SUB can't be selected. | |||
| 29892 | if (Opc == ISD::ATOMIC_LOAD_SUB) { | |||
| 29893 | RHS = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, DL, VT), RHS); | |||
| 29894 | return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, DL, VT, Chain, LHS, | |||
| 29895 | RHS, AN->getMemOperand()); | |||
| 29896 | } | |||
| 29897 | assert(Opc == ISD::ATOMIC_LOAD_ADD &&((void)0) | |||
| 29898 | "Used AtomicRMW ops other than Add should have been expanded!")((void)0); | |||
| 29899 | return N; | |||
| 29900 | } | |||
| 29901 | ||||
| 29902 | // Specialized lowering for the canonical form of an idemptotent atomicrmw. | |||
| 29903 | // The core idea here is that since the memory location isn't actually | |||
| 29904 | // changing, all we need is a lowering for the *ordering* impacts of the | |||
| 29905 | // atomicrmw. As such, we can chose a different operation and memory | |||
| 29906 | // location to minimize impact on other code. | |||
| 29907 | if (Opc == ISD::ATOMIC_LOAD_OR && isNullConstant(RHS)) { | |||
| 29908 | // On X86, the only ordering which actually requires an instruction is | |||
| 29909 | // seq_cst which isn't SingleThread, everything just needs to be preserved | |||
| 29910 | // during codegen and then dropped. Note that we expect (but don't assume), | |||
| 29911 | // that orderings other than seq_cst and acq_rel have been canonicalized to | |||
| 29912 | // a store or load. | |||
| 29913 | if (AN->getSuccessOrdering() == AtomicOrdering::SequentiallyConsistent && | |||
| 29914 | AN->getSyncScopeID() == SyncScope::System) { | |||
| 29915 | // Prefer a locked operation against a stack location to minimize cache | |||
| 29916 | // traffic. This assumes that stack locations are very likely to be | |||
| 29917 | // accessed only by the owning thread. | |||
| 29918 | SDValue NewChain = emitLockedStackOp(DAG, Subtarget, Chain, DL); | |||
| 29919 | assert(!N->hasAnyUseOfValue(0))((void)0); | |||
| 29920 | // NOTE: The getUNDEF is needed to give something for the unused result 0. | |||
| 29921 | return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), | |||
| 29922 | DAG.getUNDEF(VT), NewChain); | |||
| 29923 | } | |||
| 29924 | // MEMBARRIER is a compiler barrier; it codegens to a no-op. | |||
| 29925 | SDValue NewChain = DAG.getNode(X86ISD::MEMBARRIER, DL, MVT::Other, Chain); | |||
| 29926 | assert(!N->hasAnyUseOfValue(0))((void)0); | |||
| 29927 | // NOTE: The getUNDEF is needed to give something for the unused result 0. | |||
| 29928 | return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), | |||
| 29929 | DAG.getUNDEF(VT), NewChain); | |||
| 29930 | } | |||
| 29931 | ||||
| 29932 | SDValue LockOp = lowerAtomicArithWithLOCK(N, DAG, Subtarget); | |||
| 29933 | // RAUW the chain, but don't worry about the result, as it's unused. | |||
| 29934 | assert(!N->hasAnyUseOfValue(0))((void)0); | |||
| 29935 | // NOTE: The getUNDEF is needed to give something for the unused result 0. | |||
| 29936 | return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), | |||
| 29937 | DAG.getUNDEF(VT), LockOp.getValue(1)); | |||
| 29938 | } | |||
| 29939 | ||||
| 29940 | static SDValue LowerATOMIC_STORE(SDValue Op, SelectionDAG &DAG, | |||
| 29941 | const X86Subtarget &Subtarget) { | |||
| 29942 | auto *Node = cast<AtomicSDNode>(Op.getNode()); | |||
| 29943 | SDLoc dl(Node); | |||
| 29944 | EVT VT = Node->getMemoryVT(); | |||
| 29945 | ||||
| 29946 | bool IsSeqCst = | |||
| 29947 | Node->getSuccessOrdering() == AtomicOrdering::SequentiallyConsistent; | |||
| 29948 | bool IsTypeLegal = DAG.getTargetLoweringInfo().isTypeLegal(VT); | |||
| 29949 | ||||
| 29950 | // If this store is not sequentially consistent and the type is legal | |||
| 29951 | // we can just keep it. | |||
| 29952 | if (!IsSeqCst && IsTypeLegal) | |||
| 29953 | return Op; | |||
| 29954 | ||||
| 29955 | if (VT == MVT::i64 && !IsTypeLegal) { | |||
| 29956 | // For illegal i64 atomic_stores, we can try to use MOVQ or MOVLPS if SSE | |||
| 29957 | // is enabled. | |||
| 29958 | bool NoImplicitFloatOps = | |||
| 29959 | DAG.getMachineFunction().getFunction().hasFnAttribute( | |||
| 29960 | Attribute::NoImplicitFloat); | |||
| 29961 | if (!Subtarget.useSoftFloat() && !NoImplicitFloatOps) { | |||
| 29962 | SDValue Chain; | |||
| 29963 | if (Subtarget.hasSSE1()) { | |||
| 29964 | SDValue SclToVec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, | |||
| 29965 | Node->getOperand(2)); | |||
| 29966 | MVT StVT = Subtarget.hasSSE2() ? MVT::v2i64 : MVT::v4f32; | |||
| 29967 | SclToVec = DAG.getBitcast(StVT, SclToVec); | |||
| 29968 | SDVTList Tys = DAG.getVTList(MVT::Other); | |||
| 29969 | SDValue Ops[] = {Node->getChain(), SclToVec, Node->getBasePtr()}; | |||
| 29970 | Chain = DAG.getMemIntrinsicNode(X86ISD::VEXTRACT_STORE, dl, Tys, Ops, | |||
| 29971 | MVT::i64, Node->getMemOperand()); | |||
| 29972 | } else if (Subtarget.hasX87()) { | |||
| 29973 | // First load this into an 80-bit X87 register using a stack temporary. | |||
| 29974 | // This will put the whole integer into the significand. | |||
| 29975 | SDValue StackPtr = DAG.CreateStackTemporary(MVT::i64); | |||
| 29976 | int SPFI = cast<FrameIndexSDNode>(StackPtr.getNode())->getIndex(); | |||
| 29977 | MachinePointerInfo MPI = | |||
| 29978 | MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SPFI); | |||
| 29979 | Chain = | |||
| 29980 | DAG.getStore(Node->getChain(), dl, Node->getOperand(2), StackPtr, | |||
| 29981 | MPI, MaybeAlign(), MachineMemOperand::MOStore); | |||
| 29982 | SDVTList Tys = DAG.getVTList(MVT::f80, MVT::Other); | |||
| 29983 | SDValue LdOps[] = {Chain, StackPtr}; | |||
| 29984 | SDValue Value = | |||
| 29985 | DAG.getMemIntrinsicNode(X86ISD::FILD, dl, Tys, LdOps, MVT::i64, MPI, | |||
| 29986 | /*Align*/ None, MachineMemOperand::MOLoad); | |||
| 29987 | Chain = Value.getValue(1); | |||
| 29988 | ||||
| 29989 | // Now use an FIST to do the atomic store. | |||
| 29990 | SDValue StoreOps[] = {Chain, Value, Node->getBasePtr()}; | |||
| 29991 | Chain = | |||
| 29992 | DAG.getMemIntrinsicNode(X86ISD::FIST, dl, DAG.getVTList(MVT::Other), | |||
| 29993 | StoreOps, MVT::i64, Node->getMemOperand()); | |||
| 29994 | } | |||
| 29995 | ||||
| 29996 | if (Chain) { | |||
| 29997 | // If this is a sequentially consistent store, also emit an appropriate | |||
| 29998 | // barrier. | |||
| 29999 | if (IsSeqCst) | |||
| 30000 | Chain = emitLockedStackOp(DAG, Subtarget, Chain, dl); | |||
| 30001 | ||||
| 30002 | return Chain; | |||
| 30003 | } | |||
| 30004 | } | |||
| 30005 | } | |||
| 30006 | ||||
| 30007 | // Convert seq_cst store -> xchg | |||
| 30008 | // Convert wide store -> swap (-> cmpxchg8b/cmpxchg16b) | |||
| 30009 | // FIXME: 16-byte ATOMIC_SWAP isn't actually hooked up at the moment. | |||
| 30010 | SDValue Swap = DAG.getAtomic(ISD::ATOMIC_SWAP, dl, | |||
| 30011 | Node->getMemoryVT(), | |||
| 30012 | Node->getOperand(0), | |||
| 30013 | Node->getOperand(1), Node->getOperand(2), | |||
| 30014 | Node->getMemOperand()); | |||
| 30015 | return Swap.getValue(1); | |||
| 30016 | } | |||
| 30017 | ||||
| 30018 | static SDValue LowerADDSUBCARRY(SDValue Op, SelectionDAG &DAG) { | |||
| 30019 | SDNode *N = Op.getNode(); | |||
| 30020 | MVT VT = N->getSimpleValueType(0); | |||
| 30021 | unsigned Opc = Op.getOpcode(); | |||
| 30022 | ||||
| 30023 | // Let legalize expand this if it isn't a legal type yet. | |||
| 30024 | if (!DAG.getTargetLoweringInfo().isTypeLegal(VT)) | |||
| 30025 | return SDValue(); | |||
| 30026 | ||||
| 30027 | SDVTList VTs = DAG.getVTList(VT, MVT::i32); | |||
| 30028 | SDLoc DL(N); | |||
| 30029 | ||||
| 30030 | // Set the carry flag. | |||
| 30031 | SDValue Carry = Op.getOperand(2); | |||
| 30032 | EVT CarryVT = Carry.getValueType(); | |||
| 30033 | Carry = DAG.getNode(X86ISD::ADD, DL, DAG.getVTList(CarryVT, MVT::i32), | |||
| 30034 | Carry, DAG.getAllOnesConstant(DL, CarryVT)); | |||
| 30035 | ||||
| 30036 | bool IsAdd = Opc == ISD::ADDCARRY || Opc == ISD::SADDO_CARRY; | |||
| 30037 | SDValue Sum = DAG.getNode(IsAdd ? X86ISD::ADC : X86ISD::SBB, DL, VTs, | |||
| 30038 | Op.getOperand(0), Op.getOperand(1), | |||
| 30039 | Carry.getValue(1)); | |||
| 30040 | ||||
| 30041 | bool IsSigned = Opc == ISD::SADDO_CARRY || Opc == ISD::SSUBO_CARRY; | |||
| 30042 | SDValue SetCC = getSETCC(IsSigned ? X86::COND_O : X86::COND_B, | |||
| 30043 | Sum.getValue(1), DL, DAG); | |||
| 30044 | if (N->getValueType(1) == MVT::i1) | |||
| 30045 | SetCC = DAG.getNode(ISD::TRUNCATE, DL, MVT::i1, SetCC); | |||
| 30046 | ||||
| 30047 | return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC); | |||
| 30048 | } | |||
| 30049 | ||||
| 30050 | static SDValue LowerFSINCOS(SDValue Op, const X86Subtarget &Subtarget, | |||
| 30051 | SelectionDAG &DAG) { | |||
| 30052 | assert(Subtarget.isTargetDarwin() && Subtarget.is64Bit())((void)0); | |||
| 30053 | ||||
| 30054 | // For MacOSX, we want to call an alternative entry point: __sincos_stret, | |||
| 30055 | // which returns the values as { float, float } (in XMM0) or | |||
| 30056 | // { double, double } (which is returned in XMM0, XMM1). | |||
| 30057 | SDLoc dl(Op); | |||
| 30058 | SDValue Arg = Op.getOperand(0); | |||
| 30059 | EVT ArgVT = Arg.getValueType(); | |||
| 30060 | Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext()); | |||
| 30061 | ||||
| 30062 | TargetLowering::ArgListTy Args; | |||
| 30063 | TargetLowering::ArgListEntry Entry; | |||
| 30064 | ||||
| 30065 | Entry.Node = Arg; | |||
| 30066 | Entry.Ty = ArgTy; | |||
| 30067 | Entry.IsSExt = false; | |||
| 30068 | Entry.IsZExt = false; | |||
| 30069 | Args.push_back(Entry); | |||
| 30070 | ||||
| 30071 | bool isF64 = ArgVT == MVT::f64; | |||
| 30072 | // Only optimize x86_64 for now. i386 is a bit messy. For f32, | |||
| 30073 | // the small struct {f32, f32} is returned in (eax, edx). For f64, | |||
| 30074 | // the results are returned via SRet in memory. | |||
| 30075 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
| 30076 | RTLIB::Libcall LC = isF64 ? RTLIB::SINCOS_STRET_F64 : RTLIB::SINCOS_STRET_F32; | |||
| 30077 | const char *LibcallName = TLI.getLibcallName(LC); | |||
| 30078 | SDValue Callee = | |||
| 30079 | DAG.getExternalSymbol(LibcallName, TLI.getPointerTy(DAG.getDataLayout())); | |||
| 30080 | ||||
| 30081 | Type *RetTy = isF64 ? (Type *)StructType::get(ArgTy, ArgTy) | |||
| 30082 | : (Type *)FixedVectorType::get(ArgTy, 4); | |||
| 30083 | ||||
| 30084 | TargetLowering::CallLoweringInfo CLI(DAG); | |||
| 30085 | CLI.setDebugLoc(dl) | |||
| 30086 | .setChain(DAG.getEntryNode()) | |||
| 30087 | .setLibCallee(CallingConv::C, RetTy, Callee, std::move(Args)); | |||
| 30088 | ||||
| 30089 | std::pair<SDValue, SDValue> CallResult = TLI.LowerCallTo(CLI); | |||
| 30090 | ||||
| 30091 | if (isF64) | |||
| 30092 | // Returned in xmm0 and xmm1. | |||
| 30093 | return CallResult.first; | |||
| 30094 | ||||
| 30095 | // Returned in bits 0:31 and 32:64 xmm0. | |||
| 30096 | SDValue SinVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ArgVT, | |||
| 30097 | CallResult.first, DAG.getIntPtrConstant(0, dl)); | |||
| 30098 | SDValue CosVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ArgVT, | |||
| 30099 | CallResult.first, DAG.getIntPtrConstant(1, dl)); | |||
| 30100 | SDVTList Tys = DAG.getVTList(ArgVT, ArgVT); | |||
| 30101 | return DAG.getNode(ISD::MERGE_VALUES, dl, Tys, SinVal, CosVal); | |||
| 30102 | } | |||
| 30103 | ||||
| 30104 | /// Widen a vector input to a vector of NVT. The | |||
| 30105 | /// input vector must have the same element type as NVT. | |||
| 30106 | static SDValue ExtendToType(SDValue InOp, MVT NVT, SelectionDAG &DAG, | |||
| 30107 | bool FillWithZeroes = false) { | |||
| 30108 | // Check if InOp already has the right width. | |||
| 30109 | MVT InVT = InOp.getSimpleValueType(); | |||
| 30110 | if (InVT == NVT) | |||
| 30111 | return InOp; | |||
| 30112 | ||||
| 30113 | if (InOp.isUndef()) | |||
| 30114 | return DAG.getUNDEF(NVT); | |||
| 30115 | ||||
| 30116 | assert(InVT.getVectorElementType() == NVT.getVectorElementType() &&((void)0) | |||
| 30117 | "input and widen element type must match")((void)0); | |||
| 30118 | ||||
| 30119 | unsigned InNumElts = InVT.getVectorNumElements(); | |||
| 30120 | unsigned WidenNumElts = NVT.getVectorNumElements(); | |||
| 30121 | assert(WidenNumElts > InNumElts && WidenNumElts % InNumElts == 0 &&((void)0) | |||
| 30122 | "Unexpected request for vector widening")((void)0); | |||
| 30123 | ||||
| 30124 | SDLoc dl(InOp); | |||
| 30125 | if (InOp.getOpcode() == ISD::CONCAT_VECTORS && | |||
| 30126 | InOp.getNumOperands() == 2) { | |||
| 30127 | SDValue N1 = InOp.getOperand(1); | |||
| 30128 | if ((ISD::isBuildVectorAllZeros(N1.getNode()) && FillWithZeroes) || | |||
| 30129 | N1.isUndef()) { | |||
| 30130 | InOp = InOp.getOperand(0); | |||
| 30131 | InVT = InOp.getSimpleValueType(); | |||
| 30132 | InNumElts = InVT.getVectorNumElements(); | |||
| 30133 | } | |||
| 30134 | } | |||
| 30135 | if (ISD::isBuildVectorOfConstantSDNodes(InOp.getNode()) || | |||
| 30136 | ISD::isBuildVectorOfConstantFPSDNodes(InOp.getNode())) { | |||
| 30137 | SmallVector<SDValue, 16> Ops; | |||
| 30138 | for (unsigned i = 0; i < InNumElts; ++i) | |||
| 30139 | Ops.push_back(InOp.getOperand(i)); | |||
| 30140 | ||||
| 30141 | EVT EltVT = InOp.getOperand(0).getValueType(); | |||
| 30142 | ||||
| 30143 | SDValue FillVal = FillWithZeroes ? DAG.getConstant(0, dl, EltVT) : | |||
| 30144 | DAG.getUNDEF(EltVT); | |||
| 30145 | for (unsigned i = 0; i < WidenNumElts - InNumElts; ++i) | |||
| 30146 | Ops.push_back(FillVal); | |||
| 30147 | return DAG.getBuildVector(NVT, dl, Ops); | |||
| 30148 | } | |||
| 30149 | SDValue FillVal = FillWithZeroes ? DAG.getConstant(0, dl, NVT) : | |||
| 30150 | DAG.getUNDEF(NVT); | |||
| 30151 | return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, NVT, FillVal, | |||
| 30152 | InOp, DAG.getIntPtrConstant(0, dl)); | |||
| 30153 | } | |||
| 30154 | ||||
| 30155 | static SDValue LowerMSCATTER(SDValue Op, const X86Subtarget &Subtarget, | |||
| 30156 | SelectionDAG &DAG) { | |||
| 30157 | assert(Subtarget.hasAVX512() &&((void)0) | |||
| 30158 | "MGATHER/MSCATTER are supported on AVX-512 arch only")((void)0); | |||
| 30159 | ||||
| 30160 | MaskedScatterSDNode *N = cast<MaskedScatterSDNode>(Op.getNode()); | |||
| 30161 | SDValue Src = N->getValue(); | |||
| 30162 | MVT VT = Src.getSimpleValueType(); | |||
| 30163 | assert(VT.getScalarSizeInBits() >= 32 && "Unsupported scatter op")((void)0); | |||
| 30164 | SDLoc dl(Op); | |||
| 30165 | ||||
| 30166 | SDValue Scale = N->getScale(); | |||
| 30167 | SDValue Index = N->getIndex(); | |||
| 30168 | SDValue Mask = N->getMask(); | |||
| 30169 | SDValue Chain = N->getChain(); | |||
| 30170 | SDValue BasePtr = N->getBasePtr(); | |||
| 30171 | ||||
| 30172 | if (VT == MVT::v2f32 || VT == MVT::v2i32) { | |||
| 30173 | assert(Mask.getValueType() == MVT::v2i1 && "Unexpected mask type")((void)0); | |||
| 30174 | // If the index is v2i64 and we have VLX we can use xmm for data and index. | |||
| 30175 | if (Index.getValueType() == MVT::v2i64 && Subtarget.hasVLX()) { | |||
| 30176 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | |||
| 30177 | EVT WideVT = TLI.getTypeToTransformTo(*DAG.getContext(), VT); | |||
| 30178 | Src = DAG.getNode(ISD::CONCAT_VECTORS, dl, WideVT, Src, DAG.getUNDEF(VT)); | |||
| 30179 | SDVTList VTs = DAG.getVTList(MVT::Other); | |||
| 30180 | SDValue Ops[] = {Chain, Src, Mask, BasePtr, Index, Scale}; | |||
| 30181 | return DAG.getMemIntrinsicNode(X86ISD::MSCATTER, dl, VTs, Ops, | |||
| 30182 | N->getMemoryVT(), N->getMemOperand()); | |||
| 30183 | } | |||
| 30184 | return SDValue(); | |||
| 30185 | } | |||
| 30186 | ||||
| 30187 | MVT IndexVT = Index.getSimpleValueType(); | |||
| 30188 | ||||
| 30189 | // If the index is v2i32, we're being called by type legalization and we | |||
| 30190 | // should just let the default handling take care of it. | |||
| 30191 | if (IndexVT == MVT::v2i32) | |||
| 30192 | return SDValue(); | |||
| 30193 | ||||
| 30194 | // If we don't have VLX and neither the passthru or index is 512-bits, we | |||
| 30195 | // need to widen until one is. | |||
| 30196 | if (!Subtarget.hasVLX() && !VT.is512BitVector() && | |||
| 30197 | !Index.getSimpleValueType().is512BitVector()) { | |||
| 30198 | // Determine how much we need to widen by to get a 512-bit type. | |||
| 30199 | unsigned Factor = std::min(512/VT.getSizeInBits(), | |||
| 30200 | 512/IndexVT.getSizeInBits()); | |||
| 30201 | unsigned NumElts = VT.getVectorNumElements() * Factor; | |||
| 30202 | ||||
| 30203 | VT = MVT::getVectorVT(VT.getVectorElementType(), NumElts); | |||
| 30204 | IndexVT = MVT::getVectorVT(IndexVT.getVectorElementType(), NumElts); | |||
| 30205 | MVT MaskVT = MVT::getVectorVT(MVT::i1, NumElts); | |||
| 30206 | ||||
| 30207 | Src = ExtendToType(Src, VT, DAG); | |||
| 30208 | Index = ExtendToType(Index, IndexVT, DAG); | |||
| 30209 | Mask = ExtendToType(Mask, MaskVT, DAG, true); | |||
| 30210 | } | |||
| 30211 | ||||
| 30212 | SDVTList VTs = DAG.getVTList(MVT::Other); | |||
| 30213 | SDValue Ops[] = {Chain, Src, Mask, BasePtr, Index, Scale}; | |||
| 30214 | return DAG.getMemIntrinsicNode(X86ISD::MSCATTER, dl, VTs, Ops, | |||
| 30215 | N->getMemoryVT(), N->getMemOperand()); | |||
| 30216 | } | |||
| 30217 | ||||
| 30218 | static SDValue LowerMLOAD(SDValue Op, const X86Subtarget &Subtarget, | |||
| 30219 | SelectionDAG &DAG) { | |||
| 30220 | ||||
| 30221 | MaskedLoadSDNode *N = cast<MaskedLoadSDNode>(Op.getNode()); | |||
| 30222 | MVT VT = Op.getSimpleValueType(); | |||
| 30223 | MVT ScalarVT = VT.getScalarType(); | |||
| 30224 | SDValue Mask = N->getMask(); | |||
| 30225 | MVT MaskVT = Mask.getSimpleValueType(); | |||
| 30226 | SDValue PassThru = N->getPassThru(); | |||
| 30227 | SDLoc dl(Op); | |||
| 30228 | ||||
| 30229 | // Handle AVX masked loads which don't support passthru other than 0. | |||
| 30230 | if (MaskVT.getVectorElementType() != MVT::i1) { | |||
| 30231 | // We also allow undef in the isel pattern. | |||
| 30232 | if (PassThru.isUndef() || ISD::isBuildVectorAllZeros(PassThru.getNode())) | |||
| 30233 | return Op; | |||
| 30234 | ||||
| 30235 | SDValue NewLoad = DAG.getMaskedLoad( | |||
| 30236 | VT, dl, N->getChain(), N->getBasePtr(), N->getOffset(), Mask, | |||
| 30237 | getZeroVector(VT, Subtarget, DAG, dl), N->getMemoryVT(), | |||
| 30238 | N->getMemOperand(), N->getAddressingMode(), N->getExtensionType(), | |||
| 30239 | N->isExpandingLoad()); | |||
| 30240 | // Emit a blend. | |||
| 30241 | SDValue Select = DAG.getNode(ISD::VSELECT, dl, VT, Mask, NewLoad, PassThru); | |||
| 30242 | return DAG.getMergeValues({ Select, NewLoad.getValue(1) }, dl); | |||
| 30243 | } | |||
| 30244 | ||||
| 30245 | assert((!N->isExpandingLoad() || Subtarget.hasAVX512()) &&((void)0) | |||
| 30246 | "Expanding masked load is supported on AVX-512 target only!")((void)0); | |||
| 30247 | ||||
| 30248 | assert((!N->isExpandingLoad() || ScalarVT.getSizeInBits() >= 32) &&((void)0) | |||
| 30249 | "Expanding masked load is supported for 32 and 64-bit types only!")((void)0); | |||
| 30250 | ||||
| 30251 | assert(Subtarget.hasAVX512() && !Subtarget.hasVLX() && !VT.is512BitVector() &&((void)0) | |||
| 30252 | "Cannot lower masked load op.")((void)0); | |||
| 30253 | ||||
| 30254 | assert((ScalarVT.getSizeInBits() >= 32 ||((void)0) | |||
| 30255 | (Subtarget.hasBWI() &&((void)0) | |||
| 30256 | (ScalarVT == MVT::i8 || ScalarVT == MVT::i16))) &&((void)0) | |||
| 30257 | "Unsupported masked load op.")((void)0); | |||
| 30258 | ||||
| 30259 | // This operation is legal for targets with VLX, but without | |||
| 30260 | // VLX the vector should be widened to 512 bit | |||
| 30261 | unsigned NumEltsInWideVec = 512 / VT.getScalarSizeInBits(); | |||
| 30262 | MVT WideDataVT = MVT::getVectorVT(ScalarVT, NumEltsInWideVec); | |||
| 30263 | PassThru = ExtendToType(PassThru, WideDataVT, DAG); | |||
| 30264 | ||||
| 30265 | // Mask element has to be i1. | |||
| 30266 | assert(Mask.getSimpleValueType().getScalarType() == MVT::i1 &&((void)0) | |||
| 30267 | "Unexpected mask type")((void)0); | |||
| 30268 | ||||
| 30269 | MVT WideMaskVT = MVT::getVectorVT(MVT::i1, NumEltsInWideVec); | |||
| 30270 | ||||
| 30271 | Mask = ExtendToType(Mask, WideMaskVT, DAG, true); | |||
| 30272 | SDValue NewLoad = DAG.getMaskedLoad( | |||
| 30273 | WideDataVT, dl, N->getChain(), N->getBasePtr(), N->getOffset(), Mask, | |||
| 30274 | PassThru, N->getMemoryVT(), N->getMemOperand(), N->getAddressingMode(), | |||
| 30275 | N->getExtensionType(), N->isExpandingLoad()); | |||
| 30276 | ||||
| 30277 | SDValue Extract = | |||
| 30278 | DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, NewLoad.getValue(0), | |||
| 30279 | DAG.getIntPtrConstant(0, dl)); | |||
| 30280 | SDValue RetOps[] = {Extract, NewLoad.getValue(1)}; | |||
| 30281 | return DAG.getMergeValues(RetOps, dl); | |||
| 30282 | } | |||
| 30283 | ||||
| 30284 | static SDValue LowerMSTORE(SDValue Op, const X86Subtarget &Subtarget, | |||
| 30285 | SelectionDAG &DAG) { | |||
| 30286 | MaskedStoreSDNode *N = cast<MaskedStoreSDNode>(Op.getNode()); | |||
| 30287 | SDValue DataToStore = N->getValue(); | |||
| 30288 | MVT VT = DataToStore.getSimpleValueType(); | |||
| 30289 | MVT ScalarVT = VT.getScalarType(); | |||
| 30290 | SDValue Mask = N->getMask(); | |||
| 30291 | SDLoc dl(Op); | |||
| 30292 | ||||
| 30293 | assert((!N->isCompressingStore() || Subtarget.hasAVX512()) &&((void)0) | |||
| 30294 | "Expanding masked load is supported on AVX-512 target only!")((void)0); | |||
| 30295 |