| File: | src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/IR/PatternMatch.h |
| Warning: | line 232, column 9 Called C++ object pointer is null |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
| 1 | //===- ValueTracking.cpp - Walk computations to compute properties --------===// | ||||||
| 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 contains routines that help analyze properties that chains of | ||||||
| 10 | // computations have. | ||||||
| 11 | // | ||||||
| 12 | //===----------------------------------------------------------------------===// | ||||||
| 13 | |||||||
| 14 | #include "llvm/Analysis/ValueTracking.h" | ||||||
| 15 | #include "llvm/ADT/APFloat.h" | ||||||
| 16 | #include "llvm/ADT/APInt.h" | ||||||
| 17 | #include "llvm/ADT/ArrayRef.h" | ||||||
| 18 | #include "llvm/ADT/None.h" | ||||||
| 19 | #include "llvm/ADT/Optional.h" | ||||||
| 20 | #include "llvm/ADT/STLExtras.h" | ||||||
| 21 | #include "llvm/ADT/SmallPtrSet.h" | ||||||
| 22 | #include "llvm/ADT/SmallSet.h" | ||||||
| 23 | #include "llvm/ADT/SmallVector.h" | ||||||
| 24 | #include "llvm/ADT/StringRef.h" | ||||||
| 25 | #include "llvm/ADT/iterator_range.h" | ||||||
| 26 | #include "llvm/Analysis/AliasAnalysis.h" | ||||||
| 27 | #include "llvm/Analysis/AssumeBundleQueries.h" | ||||||
| 28 | #include "llvm/Analysis/AssumptionCache.h" | ||||||
| 29 | #include "llvm/Analysis/EHPersonalities.h" | ||||||
| 30 | #include "llvm/Analysis/GuardUtils.h" | ||||||
| 31 | #include "llvm/Analysis/InstructionSimplify.h" | ||||||
| 32 | #include "llvm/Analysis/Loads.h" | ||||||
| 33 | #include "llvm/Analysis/LoopInfo.h" | ||||||
| 34 | #include "llvm/Analysis/OptimizationRemarkEmitter.h" | ||||||
| 35 | #include "llvm/Analysis/TargetLibraryInfo.h" | ||||||
| 36 | #include "llvm/IR/Argument.h" | ||||||
| 37 | #include "llvm/IR/Attributes.h" | ||||||
| 38 | #include "llvm/IR/BasicBlock.h" | ||||||
| 39 | #include "llvm/IR/Constant.h" | ||||||
| 40 | #include "llvm/IR/ConstantRange.h" | ||||||
| 41 | #include "llvm/IR/Constants.h" | ||||||
| 42 | #include "llvm/IR/DerivedTypes.h" | ||||||
| 43 | #include "llvm/IR/DiagnosticInfo.h" | ||||||
| 44 | #include "llvm/IR/Dominators.h" | ||||||
| 45 | #include "llvm/IR/Function.h" | ||||||
| 46 | #include "llvm/IR/GetElementPtrTypeIterator.h" | ||||||
| 47 | #include "llvm/IR/GlobalAlias.h" | ||||||
| 48 | #include "llvm/IR/GlobalValue.h" | ||||||
| 49 | #include "llvm/IR/GlobalVariable.h" | ||||||
| 50 | #include "llvm/IR/InstrTypes.h" | ||||||
| 51 | #include "llvm/IR/Instruction.h" | ||||||
| 52 | #include "llvm/IR/Instructions.h" | ||||||
| 53 | #include "llvm/IR/IntrinsicInst.h" | ||||||
| 54 | #include "llvm/IR/Intrinsics.h" | ||||||
| 55 | #include "llvm/IR/IntrinsicsAArch64.h" | ||||||
| 56 | #include "llvm/IR/IntrinsicsRISCV.h" | ||||||
| 57 | #include "llvm/IR/IntrinsicsX86.h" | ||||||
| 58 | #include "llvm/IR/LLVMContext.h" | ||||||
| 59 | #include "llvm/IR/Metadata.h" | ||||||
| 60 | #include "llvm/IR/Module.h" | ||||||
| 61 | #include "llvm/IR/Operator.h" | ||||||
| 62 | #include "llvm/IR/PatternMatch.h" | ||||||
| 63 | #include "llvm/IR/Type.h" | ||||||
| 64 | #include "llvm/IR/User.h" | ||||||
| 65 | #include "llvm/IR/Value.h" | ||||||
| 66 | #include "llvm/Support/Casting.h" | ||||||
| 67 | #include "llvm/Support/CommandLine.h" | ||||||
| 68 | #include "llvm/Support/Compiler.h" | ||||||
| 69 | #include "llvm/Support/ErrorHandling.h" | ||||||
| 70 | #include "llvm/Support/KnownBits.h" | ||||||
| 71 | #include "llvm/Support/MathExtras.h" | ||||||
| 72 | #include <algorithm> | ||||||
| 73 | #include <array> | ||||||
| 74 | #include <cassert> | ||||||
| 75 | #include <cstdint> | ||||||
| 76 | #include <iterator> | ||||||
| 77 | #include <utility> | ||||||
| 78 | |||||||
| 79 | using namespace llvm; | ||||||
| 80 | using namespace llvm::PatternMatch; | ||||||
| 81 | |||||||
| 82 | // Controls the number of uses of the value searched for possible | ||||||
| 83 | // dominating comparisons. | ||||||
| 84 | static cl::opt<unsigned> DomConditionsMaxUses("dom-conditions-max-uses", | ||||||
| 85 | cl::Hidden, cl::init(20)); | ||||||
| 86 | |||||||
| 87 | /// Returns the bitwidth of the given scalar or pointer type. For vector types, | ||||||
| 88 | /// returns the element type's bitwidth. | ||||||
| 89 | static unsigned getBitWidth(Type *Ty, const DataLayout &DL) { | ||||||
| 90 | if (unsigned BitWidth = Ty->getScalarSizeInBits()) | ||||||
| 91 | return BitWidth; | ||||||
| 92 | |||||||
| 93 | return DL.getPointerTypeSizeInBits(Ty); | ||||||
| 94 | } | ||||||
| 95 | |||||||
| 96 | namespace { | ||||||
| 97 | |||||||
| 98 | // Simplifying using an assume can only be done in a particular control-flow | ||||||
| 99 | // context (the context instruction provides that context). If an assume and | ||||||
| 100 | // the context instruction are not in the same block then the DT helps in | ||||||
| 101 | // figuring out if we can use it. | ||||||
| 102 | struct Query { | ||||||
| 103 | const DataLayout &DL; | ||||||
| 104 | AssumptionCache *AC; | ||||||
| 105 | const Instruction *CxtI; | ||||||
| 106 | const DominatorTree *DT; | ||||||
| 107 | |||||||
| 108 | // Unlike the other analyses, this may be a nullptr because not all clients | ||||||
| 109 | // provide it currently. | ||||||
| 110 | OptimizationRemarkEmitter *ORE; | ||||||
| 111 | |||||||
| 112 | /// If true, it is safe to use metadata during simplification. | ||||||
| 113 | InstrInfoQuery IIQ; | ||||||
| 114 | |||||||
| 115 | Query(const DataLayout &DL, AssumptionCache *AC, const Instruction *CxtI, | ||||||
| 116 | const DominatorTree *DT, bool UseInstrInfo, | ||||||
| 117 | OptimizationRemarkEmitter *ORE = nullptr) | ||||||
| 118 | : DL(DL), AC(AC), CxtI(CxtI), DT(DT), ORE(ORE), IIQ(UseInstrInfo) {} | ||||||
| 119 | }; | ||||||
| 120 | |||||||
| 121 | } // end anonymous namespace | ||||||
| 122 | |||||||
| 123 | // Given the provided Value and, potentially, a context instruction, return | ||||||
| 124 | // the preferred context instruction (if any). | ||||||
| 125 | static const Instruction *safeCxtI(const Value *V, const Instruction *CxtI) { | ||||||
| 126 | // If we've been provided with a context instruction, then use that (provided | ||||||
| 127 | // it has been inserted). | ||||||
| 128 | if (CxtI && CxtI->getParent()) | ||||||
| 129 | return CxtI; | ||||||
| 130 | |||||||
| 131 | // If the value is really an already-inserted instruction, then use that. | ||||||
| 132 | CxtI = dyn_cast<Instruction>(V); | ||||||
| 133 | if (CxtI && CxtI->getParent()) | ||||||
| 134 | return CxtI; | ||||||
| 135 | |||||||
| 136 | return nullptr; | ||||||
| 137 | } | ||||||
| 138 | |||||||
| 139 | static const Instruction *safeCxtI(const Value *V1, const Value *V2, const Instruction *CxtI) { | ||||||
| 140 | // If we've been provided with a context instruction, then use that (provided | ||||||
| 141 | // it has been inserted). | ||||||
| 142 | if (CxtI && CxtI->getParent()) | ||||||
| 143 | return CxtI; | ||||||
| 144 | |||||||
| 145 | // If the value is really an already-inserted instruction, then use that. | ||||||
| 146 | CxtI = dyn_cast<Instruction>(V1); | ||||||
| 147 | if (CxtI && CxtI->getParent()) | ||||||
| 148 | return CxtI; | ||||||
| 149 | |||||||
| 150 | CxtI = dyn_cast<Instruction>(V2); | ||||||
| 151 | if (CxtI && CxtI->getParent()) | ||||||
| 152 | return CxtI; | ||||||
| 153 | |||||||
| 154 | return nullptr; | ||||||
| 155 | } | ||||||
| 156 | |||||||
| 157 | static bool getShuffleDemandedElts(const ShuffleVectorInst *Shuf, | ||||||
| 158 | const APInt &DemandedElts, | ||||||
| 159 | APInt &DemandedLHS, APInt &DemandedRHS) { | ||||||
| 160 | // The length of scalable vectors is unknown at compile time, thus we | ||||||
| 161 | // cannot check their values | ||||||
| 162 | if (isa<ScalableVectorType>(Shuf->getType())) | ||||||
| 163 | return false; | ||||||
| 164 | |||||||
| 165 | int NumElts = | ||||||
| 166 | cast<FixedVectorType>(Shuf->getOperand(0)->getType())->getNumElements(); | ||||||
| 167 | int NumMaskElts = cast<FixedVectorType>(Shuf->getType())->getNumElements(); | ||||||
| 168 | DemandedLHS = DemandedRHS = APInt::getNullValue(NumElts); | ||||||
| 169 | if (DemandedElts.isNullValue()) | ||||||
| 170 | return true; | ||||||
| 171 | // Simple case of a shuffle with zeroinitializer. | ||||||
| 172 | if (all_of(Shuf->getShuffleMask(), [](int Elt) { return Elt == 0; })) { | ||||||
| 173 | DemandedLHS.setBit(0); | ||||||
| 174 | return true; | ||||||
| 175 | } | ||||||
| 176 | for (int i = 0; i != NumMaskElts; ++i) { | ||||||
| 177 | if (!DemandedElts[i]) | ||||||
| 178 | continue; | ||||||
| 179 | int M = Shuf->getMaskValue(i); | ||||||
| 180 | assert(M < (NumElts * 2) && "Invalid shuffle mask constant")((void)0); | ||||||
| 181 | |||||||
| 182 | // For undef elements, we don't know anything about the common state of | ||||||
| 183 | // the shuffle result. | ||||||
| 184 | if (M == -1) | ||||||
| 185 | return false; | ||||||
| 186 | if (M < NumElts) | ||||||
| 187 | DemandedLHS.setBit(M % NumElts); | ||||||
| 188 | else | ||||||
| 189 | DemandedRHS.setBit(M % NumElts); | ||||||
| 190 | } | ||||||
| 191 | |||||||
| 192 | return true; | ||||||
| 193 | } | ||||||
| 194 | |||||||
| 195 | static void computeKnownBits(const Value *V, const APInt &DemandedElts, | ||||||
| 196 | KnownBits &Known, unsigned Depth, const Query &Q); | ||||||
| 197 | |||||||
| 198 | static void computeKnownBits(const Value *V, KnownBits &Known, unsigned Depth, | ||||||
| 199 | const Query &Q) { | ||||||
| 200 | // FIXME: We currently have no way to represent the DemandedElts of a scalable | ||||||
| 201 | // vector | ||||||
| 202 | if (isa<ScalableVectorType>(V->getType())) { | ||||||
| 203 | Known.resetAll(); | ||||||
| 204 | return; | ||||||
| 205 | } | ||||||
| 206 | |||||||
| 207 | auto *FVTy = dyn_cast<FixedVectorType>(V->getType()); | ||||||
| 208 | APInt DemandedElts = | ||||||
| 209 | FVTy ? APInt::getAllOnesValue(FVTy->getNumElements()) : APInt(1, 1); | ||||||
| 210 | computeKnownBits(V, DemandedElts, Known, Depth, Q); | ||||||
| 211 | } | ||||||
| 212 | |||||||
| 213 | void llvm::computeKnownBits(const Value *V, KnownBits &Known, | ||||||
| 214 | const DataLayout &DL, unsigned Depth, | ||||||
| 215 | AssumptionCache *AC, const Instruction *CxtI, | ||||||
| 216 | const DominatorTree *DT, | ||||||
| 217 | OptimizationRemarkEmitter *ORE, bool UseInstrInfo) { | ||||||
| 218 | ::computeKnownBits(V, Known, Depth, | ||||||
| 219 | Query(DL, AC, safeCxtI(V, CxtI), DT, UseInstrInfo, ORE)); | ||||||
| 220 | } | ||||||
| 221 | |||||||
| 222 | void llvm::computeKnownBits(const Value *V, const APInt &DemandedElts, | ||||||
| 223 | KnownBits &Known, const DataLayout &DL, | ||||||
| 224 | unsigned Depth, AssumptionCache *AC, | ||||||
| 225 | const Instruction *CxtI, const DominatorTree *DT, | ||||||
| 226 | OptimizationRemarkEmitter *ORE, bool UseInstrInfo) { | ||||||
| 227 | ::computeKnownBits(V, DemandedElts, Known, Depth, | ||||||
| 228 | Query(DL, AC, safeCxtI(V, CxtI), DT, UseInstrInfo, ORE)); | ||||||
| 229 | } | ||||||
| 230 | |||||||
| 231 | static KnownBits computeKnownBits(const Value *V, const APInt &DemandedElts, | ||||||
| 232 | unsigned Depth, const Query &Q); | ||||||
| 233 | |||||||
| 234 | static KnownBits computeKnownBits(const Value *V, unsigned Depth, | ||||||
| 235 | const Query &Q); | ||||||
| 236 | |||||||
| 237 | KnownBits llvm::computeKnownBits(const Value *V, const DataLayout &DL, | ||||||
| 238 | unsigned Depth, AssumptionCache *AC, | ||||||
| 239 | const Instruction *CxtI, | ||||||
| 240 | const DominatorTree *DT, | ||||||
| 241 | OptimizationRemarkEmitter *ORE, | ||||||
| 242 | bool UseInstrInfo) { | ||||||
| 243 | return ::computeKnownBits( | ||||||
| 244 | V, Depth, Query(DL, AC, safeCxtI(V, CxtI), DT, UseInstrInfo, ORE)); | ||||||
| 245 | } | ||||||
| 246 | |||||||
| 247 | KnownBits llvm::computeKnownBits(const Value *V, const APInt &DemandedElts, | ||||||
| 248 | const DataLayout &DL, unsigned Depth, | ||||||
| 249 | AssumptionCache *AC, const Instruction *CxtI, | ||||||
| 250 | const DominatorTree *DT, | ||||||
| 251 | OptimizationRemarkEmitter *ORE, | ||||||
| 252 | bool UseInstrInfo) { | ||||||
| 253 | return ::computeKnownBits( | ||||||
| 254 | V, DemandedElts, Depth, | ||||||
| 255 | Query(DL, AC, safeCxtI(V, CxtI), DT, UseInstrInfo, ORE)); | ||||||
| 256 | } | ||||||
| 257 | |||||||
| 258 | bool llvm::haveNoCommonBitsSet(const Value *LHS, const Value *RHS, | ||||||
| 259 | const DataLayout &DL, AssumptionCache *AC, | ||||||
| 260 | const Instruction *CxtI, const DominatorTree *DT, | ||||||
| 261 | bool UseInstrInfo) { | ||||||
| 262 | assert(LHS->getType() == RHS->getType() &&((void)0) | ||||||
| 263 | "LHS and RHS should have the same type")((void)0); | ||||||
| 264 | assert(LHS->getType()->isIntOrIntVectorTy() &&((void)0) | ||||||
| 265 | "LHS and RHS should be integers")((void)0); | ||||||
| 266 | // Look for an inverted mask: (X & ~M) op (Y & M). | ||||||
| 267 | Value *M; | ||||||
| 268 | if (match(LHS, m_c_And(m_Not(m_Value(M)), m_Value())) && | ||||||
| 269 | match(RHS, m_c_And(m_Specific(M), m_Value()))) | ||||||
| 270 | return true; | ||||||
| 271 | if (match(RHS, m_c_And(m_Not(m_Value(M)), m_Value())) && | ||||||
| 272 | match(LHS, m_c_And(m_Specific(M), m_Value()))) | ||||||
| 273 | return true; | ||||||
| 274 | IntegerType *IT = cast<IntegerType>(LHS->getType()->getScalarType()); | ||||||
| 275 | KnownBits LHSKnown(IT->getBitWidth()); | ||||||
| 276 | KnownBits RHSKnown(IT->getBitWidth()); | ||||||
| 277 | computeKnownBits(LHS, LHSKnown, DL, 0, AC, CxtI, DT, nullptr, UseInstrInfo); | ||||||
| 278 | computeKnownBits(RHS, RHSKnown, DL, 0, AC, CxtI, DT, nullptr, UseInstrInfo); | ||||||
| 279 | return KnownBits::haveNoCommonBitsSet(LHSKnown, RHSKnown); | ||||||
| 280 | } | ||||||
| 281 | |||||||
| 282 | bool llvm::isOnlyUsedInZeroEqualityComparison(const Instruction *CxtI) { | ||||||
| 283 | for (const User *U : CxtI->users()) { | ||||||
| 284 | if (const ICmpInst *IC = dyn_cast<ICmpInst>(U)) | ||||||
| 285 | if (IC->isEquality()) | ||||||
| 286 | if (Constant *C = dyn_cast<Constant>(IC->getOperand(1))) | ||||||
| 287 | if (C->isNullValue()) | ||||||
| 288 | continue; | ||||||
| 289 | return false; | ||||||
| 290 | } | ||||||
| 291 | return true; | ||||||
| 292 | } | ||||||
| 293 | |||||||
| 294 | static bool isKnownToBeAPowerOfTwo(const Value *V, bool OrZero, unsigned Depth, | ||||||
| 295 | const Query &Q); | ||||||
| 296 | |||||||
| 297 | bool llvm::isKnownToBeAPowerOfTwo(const Value *V, const DataLayout &DL, | ||||||
| 298 | bool OrZero, unsigned Depth, | ||||||
| 299 | AssumptionCache *AC, const Instruction *CxtI, | ||||||
| 300 | const DominatorTree *DT, bool UseInstrInfo) { | ||||||
| 301 | return ::isKnownToBeAPowerOfTwo( | ||||||
| 302 | V, OrZero, Depth, Query(DL, AC, safeCxtI(V, CxtI), DT, UseInstrInfo)); | ||||||
| 303 | } | ||||||
| 304 | |||||||
| 305 | static bool isKnownNonZero(const Value *V, const APInt &DemandedElts, | ||||||
| 306 | unsigned Depth, const Query &Q); | ||||||
| 307 | |||||||
| 308 | static bool isKnownNonZero(const Value *V, unsigned Depth, const Query &Q); | ||||||
| 309 | |||||||
| 310 | bool llvm::isKnownNonZero(const Value *V, const DataLayout &DL, unsigned Depth, | ||||||
| 311 | AssumptionCache *AC, const Instruction *CxtI, | ||||||
| 312 | const DominatorTree *DT, bool UseInstrInfo) { | ||||||
| 313 | return ::isKnownNonZero(V, Depth, | ||||||
| 314 | Query(DL, AC, safeCxtI(V, CxtI), DT, UseInstrInfo)); | ||||||
| 315 | } | ||||||
| 316 | |||||||
| 317 | bool llvm::isKnownNonNegative(const Value *V, const DataLayout &DL, | ||||||
| 318 | unsigned Depth, AssumptionCache *AC, | ||||||
| 319 | const Instruction *CxtI, const DominatorTree *DT, | ||||||
| 320 | bool UseInstrInfo) { | ||||||
| 321 | KnownBits Known = | ||||||
| 322 | computeKnownBits(V, DL, Depth, AC, CxtI, DT, nullptr, UseInstrInfo); | ||||||
| 323 | return Known.isNonNegative(); | ||||||
| 324 | } | ||||||
| 325 | |||||||
| 326 | bool llvm::isKnownPositive(const Value *V, const DataLayout &DL, unsigned Depth, | ||||||
| 327 | AssumptionCache *AC, const Instruction *CxtI, | ||||||
| 328 | const DominatorTree *DT, bool UseInstrInfo) { | ||||||
| 329 | if (auto *CI = dyn_cast<ConstantInt>(V)) | ||||||
| 330 | return CI->getValue().isStrictlyPositive(); | ||||||
| 331 | |||||||
| 332 | // TODO: We'd doing two recursive queries here. We should factor this such | ||||||
| 333 | // that only a single query is needed. | ||||||
| 334 | return isKnownNonNegative(V, DL, Depth, AC, CxtI, DT, UseInstrInfo) && | ||||||
| 335 | isKnownNonZero(V, DL, Depth, AC, CxtI, DT, UseInstrInfo); | ||||||
| 336 | } | ||||||
| 337 | |||||||
| 338 | bool llvm::isKnownNegative(const Value *V, const DataLayout &DL, unsigned Depth, | ||||||
| 339 | AssumptionCache *AC, const Instruction *CxtI, | ||||||
| 340 | const DominatorTree *DT, bool UseInstrInfo) { | ||||||
| 341 | KnownBits Known = | ||||||
| 342 | computeKnownBits(V, DL, Depth, AC, CxtI, DT, nullptr, UseInstrInfo); | ||||||
| 343 | return Known.isNegative(); | ||||||
| 344 | } | ||||||
| 345 | |||||||
| 346 | static bool isKnownNonEqual(const Value *V1, const Value *V2, unsigned Depth, | ||||||
| 347 | const Query &Q); | ||||||
| 348 | |||||||
| 349 | bool llvm::isKnownNonEqual(const Value *V1, const Value *V2, | ||||||
| 350 | const DataLayout &DL, AssumptionCache *AC, | ||||||
| 351 | const Instruction *CxtI, const DominatorTree *DT, | ||||||
| 352 | bool UseInstrInfo) { | ||||||
| 353 | return ::isKnownNonEqual(V1, V2, 0, | ||||||
| 354 | Query(DL, AC, safeCxtI(V2, V1, CxtI), DT, | ||||||
| 355 | UseInstrInfo, /*ORE=*/nullptr)); | ||||||
| 356 | } | ||||||
| 357 | |||||||
| 358 | static bool MaskedValueIsZero(const Value *V, const APInt &Mask, unsigned Depth, | ||||||
| 359 | const Query &Q); | ||||||
| 360 | |||||||
| 361 | bool llvm::MaskedValueIsZero(const Value *V, const APInt &Mask, | ||||||
| 362 | const DataLayout &DL, unsigned Depth, | ||||||
| 363 | AssumptionCache *AC, const Instruction *CxtI, | ||||||
| 364 | const DominatorTree *DT, bool UseInstrInfo) { | ||||||
| 365 | return ::MaskedValueIsZero( | ||||||
| 366 | V, Mask, Depth, Query(DL, AC, safeCxtI(V, CxtI), DT, UseInstrInfo)); | ||||||
| 367 | } | ||||||
| 368 | |||||||
| 369 | static unsigned ComputeNumSignBits(const Value *V, const APInt &DemandedElts, | ||||||
| 370 | unsigned Depth, const Query &Q); | ||||||
| 371 | |||||||
| 372 | static unsigned ComputeNumSignBits(const Value *V, unsigned Depth, | ||||||
| 373 | const Query &Q) { | ||||||
| 374 | // FIXME: We currently have no way to represent the DemandedElts of a scalable | ||||||
| 375 | // vector | ||||||
| 376 | if (isa<ScalableVectorType>(V->getType())) | ||||||
| 377 | return 1; | ||||||
| 378 | |||||||
| 379 | auto *FVTy = dyn_cast<FixedVectorType>(V->getType()); | ||||||
| 380 | APInt DemandedElts = | ||||||
| 381 | FVTy ? APInt::getAllOnesValue(FVTy->getNumElements()) : APInt(1, 1); | ||||||
| 382 | return ComputeNumSignBits(V, DemandedElts, Depth, Q); | ||||||
| 383 | } | ||||||
| 384 | |||||||
| 385 | unsigned llvm::ComputeNumSignBits(const Value *V, const DataLayout &DL, | ||||||
| 386 | unsigned Depth, AssumptionCache *AC, | ||||||
| 387 | const Instruction *CxtI, | ||||||
| 388 | const DominatorTree *DT, bool UseInstrInfo) { | ||||||
| 389 | return ::ComputeNumSignBits( | ||||||
| 390 | V, Depth, Query(DL, AC, safeCxtI(V, CxtI), DT, UseInstrInfo)); | ||||||
| 391 | } | ||||||
| 392 | |||||||
| 393 | static void computeKnownBitsAddSub(bool Add, const Value *Op0, const Value *Op1, | ||||||
| 394 | bool NSW, const APInt &DemandedElts, | ||||||
| 395 | KnownBits &KnownOut, KnownBits &Known2, | ||||||
| 396 | unsigned Depth, const Query &Q) { | ||||||
| 397 | computeKnownBits(Op1, DemandedElts, KnownOut, Depth + 1, Q); | ||||||
| 398 | |||||||
| 399 | // If one operand is unknown and we have no nowrap information, | ||||||
| 400 | // the result will be unknown independently of the second operand. | ||||||
| 401 | if (KnownOut.isUnknown() && !NSW) | ||||||
| 402 | return; | ||||||
| 403 | |||||||
| 404 | computeKnownBits(Op0, DemandedElts, Known2, Depth + 1, Q); | ||||||
| 405 | KnownOut = KnownBits::computeForAddSub(Add, NSW, Known2, KnownOut); | ||||||
| 406 | } | ||||||
| 407 | |||||||
| 408 | static void computeKnownBitsMul(const Value *Op0, const Value *Op1, bool NSW, | ||||||
| 409 | const APInt &DemandedElts, KnownBits &Known, | ||||||
| 410 | KnownBits &Known2, unsigned Depth, | ||||||
| 411 | const Query &Q) { | ||||||
| 412 | computeKnownBits(Op1, DemandedElts, Known, Depth + 1, Q); | ||||||
| 413 | computeKnownBits(Op0, DemandedElts, Known2, Depth + 1, Q); | ||||||
| 414 | |||||||
| 415 | bool isKnownNegative = false; | ||||||
| 416 | bool isKnownNonNegative = false; | ||||||
| 417 | // If the multiplication is known not to overflow, compute the sign bit. | ||||||
| 418 | if (NSW) { | ||||||
| 419 | if (Op0 == Op1) { | ||||||
| 420 | // The product of a number with itself is non-negative. | ||||||
| 421 | isKnownNonNegative = true; | ||||||
| 422 | } else { | ||||||
| 423 | bool isKnownNonNegativeOp1 = Known.isNonNegative(); | ||||||
| 424 | bool isKnownNonNegativeOp0 = Known2.isNonNegative(); | ||||||
| 425 | bool isKnownNegativeOp1 = Known.isNegative(); | ||||||
| 426 | bool isKnownNegativeOp0 = Known2.isNegative(); | ||||||
| 427 | // The product of two numbers with the same sign is non-negative. | ||||||
| 428 | isKnownNonNegative = (isKnownNegativeOp1 && isKnownNegativeOp0) || | ||||||
| 429 | (isKnownNonNegativeOp1 && isKnownNonNegativeOp0); | ||||||
| 430 | // The product of a negative number and a non-negative number is either | ||||||
| 431 | // negative or zero. | ||||||
| 432 | if (!isKnownNonNegative) | ||||||
| 433 | isKnownNegative = | ||||||
| 434 | (isKnownNegativeOp1 && isKnownNonNegativeOp0 && | ||||||
| 435 | Known2.isNonZero()) || | ||||||
| 436 | (isKnownNegativeOp0 && isKnownNonNegativeOp1 && Known.isNonZero()); | ||||||
| 437 | } | ||||||
| 438 | } | ||||||
| 439 | |||||||
| 440 | Known = KnownBits::mul(Known, Known2); | ||||||
| 441 | |||||||
| 442 | // Only make use of no-wrap flags if we failed to compute the sign bit | ||||||
| 443 | // directly. This matters if the multiplication always overflows, in | ||||||
| 444 | // which case we prefer to follow the result of the direct computation, | ||||||
| 445 | // though as the program is invoking undefined behaviour we can choose | ||||||
| 446 | // whatever we like here. | ||||||
| 447 | if (isKnownNonNegative && !Known.isNegative()) | ||||||
| 448 | Known.makeNonNegative(); | ||||||
| 449 | else if (isKnownNegative && !Known.isNonNegative()) | ||||||
| 450 | Known.makeNegative(); | ||||||
| 451 | } | ||||||
| 452 | |||||||
| 453 | void llvm::computeKnownBitsFromRangeMetadata(const MDNode &Ranges, | ||||||
| 454 | KnownBits &Known) { | ||||||
| 455 | unsigned BitWidth = Known.getBitWidth(); | ||||||
| 456 | unsigned NumRanges = Ranges.getNumOperands() / 2; | ||||||
| 457 | assert(NumRanges >= 1)((void)0); | ||||||
| 458 | |||||||
| 459 | Known.Zero.setAllBits(); | ||||||
| 460 | Known.One.setAllBits(); | ||||||
| 461 | |||||||
| 462 | for (unsigned i = 0; i < NumRanges; ++i) { | ||||||
| 463 | ConstantInt *Lower = | ||||||
| 464 | mdconst::extract<ConstantInt>(Ranges.getOperand(2 * i + 0)); | ||||||
| 465 | ConstantInt *Upper = | ||||||
| 466 | mdconst::extract<ConstantInt>(Ranges.getOperand(2 * i + 1)); | ||||||
| 467 | ConstantRange Range(Lower->getValue(), Upper->getValue()); | ||||||
| 468 | |||||||
| 469 | // The first CommonPrefixBits of all values in Range are equal. | ||||||
| 470 | unsigned CommonPrefixBits = | ||||||
| 471 | (Range.getUnsignedMax() ^ Range.getUnsignedMin()).countLeadingZeros(); | ||||||
| 472 | APInt Mask = APInt::getHighBitsSet(BitWidth, CommonPrefixBits); | ||||||
| 473 | APInt UnsignedMax = Range.getUnsignedMax().zextOrTrunc(BitWidth); | ||||||
| 474 | Known.One &= UnsignedMax & Mask; | ||||||
| 475 | Known.Zero &= ~UnsignedMax & Mask; | ||||||
| 476 | } | ||||||
| 477 | } | ||||||
| 478 | |||||||
| 479 | static bool isEphemeralValueOf(const Instruction *I, const Value *E) { | ||||||
| 480 | SmallVector<const Value *, 16> WorkSet(1, I); | ||||||
| 481 | SmallPtrSet<const Value *, 32> Visited; | ||||||
| 482 | SmallPtrSet<const Value *, 16> EphValues; | ||||||
| 483 | |||||||
| 484 | // The instruction defining an assumption's condition itself is always | ||||||
| 485 | // considered ephemeral to that assumption (even if it has other | ||||||
| 486 | // non-ephemeral users). See r246696's test case for an example. | ||||||
| 487 | if (is_contained(I->operands(), E)) | ||||||
| 488 | return true; | ||||||
| 489 | |||||||
| 490 | while (!WorkSet.empty()) { | ||||||
| 491 | const Value *V = WorkSet.pop_back_val(); | ||||||
| 492 | if (!Visited.insert(V).second) | ||||||
| 493 | continue; | ||||||
| 494 | |||||||
| 495 | // If all uses of this value are ephemeral, then so is this value. | ||||||
| 496 | if (llvm::all_of(V->users(), [&](const User *U) { | ||||||
| 497 | return EphValues.count(U); | ||||||
| 498 | })) { | ||||||
| 499 | if (V == E) | ||||||
| 500 | return true; | ||||||
| 501 | |||||||
| 502 | if (V == I || isSafeToSpeculativelyExecute(V)) { | ||||||
| 503 | EphValues.insert(V); | ||||||
| 504 | if (const User *U = dyn_cast<User>(V)) | ||||||
| 505 | append_range(WorkSet, U->operands()); | ||||||
| 506 | } | ||||||
| 507 | } | ||||||
| 508 | } | ||||||
| 509 | |||||||
| 510 | return false; | ||||||
| 511 | } | ||||||
| 512 | |||||||
| 513 | // Is this an intrinsic that cannot be speculated but also cannot trap? | ||||||
| 514 | bool llvm::isAssumeLikeIntrinsic(const Instruction *I) { | ||||||
| 515 | if (const IntrinsicInst *CI = dyn_cast<IntrinsicInst>(I)) | ||||||
| 516 | return CI->isAssumeLikeIntrinsic(); | ||||||
| 517 | |||||||
| 518 | return false; | ||||||
| 519 | } | ||||||
| 520 | |||||||
| 521 | bool llvm::isValidAssumeForContext(const Instruction *Inv, | ||||||
| 522 | const Instruction *CxtI, | ||||||
| 523 | const DominatorTree *DT) { | ||||||
| 524 | // There are two restrictions on the use of an assume: | ||||||
| 525 | // 1. The assume must dominate the context (or the control flow must | ||||||
| 526 | // reach the assume whenever it reaches the context). | ||||||
| 527 | // 2. The context must not be in the assume's set of ephemeral values | ||||||
| 528 | // (otherwise we will use the assume to prove that the condition | ||||||
| 529 | // feeding the assume is trivially true, thus causing the removal of | ||||||
| 530 | // the assume). | ||||||
| 531 | |||||||
| 532 | if (Inv->getParent() == CxtI->getParent()) { | ||||||
| 533 | // If Inv and CtxI are in the same block, check if the assume (Inv) is first | ||||||
| 534 | // in the BB. | ||||||
| 535 | if (Inv->comesBefore(CxtI)) | ||||||
| 536 | return true; | ||||||
| 537 | |||||||
| 538 | // Don't let an assume affect itself - this would cause the problems | ||||||
| 539 | // `isEphemeralValueOf` is trying to prevent, and it would also make | ||||||
| 540 | // the loop below go out of bounds. | ||||||
| 541 | if (Inv == CxtI) | ||||||
| 542 | return false; | ||||||
| 543 | |||||||
| 544 | // The context comes first, but they're both in the same block. | ||||||
| 545 | // Make sure there is nothing in between that might interrupt | ||||||
| 546 | // the control flow, not even CxtI itself. | ||||||
| 547 | // We limit the scan distance between the assume and its context instruction | ||||||
| 548 | // to avoid a compile-time explosion. This limit is chosen arbitrarily, so | ||||||
| 549 | // it can be adjusted if needed (could be turned into a cl::opt). | ||||||
| 550 | unsigned ScanLimit = 15; | ||||||
| 551 | for (BasicBlock::const_iterator I(CxtI), IE(Inv); I != IE; ++I) | ||||||
| 552 | if (!isGuaranteedToTransferExecutionToSuccessor(&*I) || --ScanLimit == 0) | ||||||
| 553 | return false; | ||||||
| 554 | |||||||
| 555 | return !isEphemeralValueOf(Inv, CxtI); | ||||||
| 556 | } | ||||||
| 557 | |||||||
| 558 | // Inv and CxtI are in different blocks. | ||||||
| 559 | if (DT) { | ||||||
| 560 | if (DT->dominates(Inv, CxtI)) | ||||||
| 561 | return true; | ||||||
| 562 | } else if (Inv->getParent() == CxtI->getParent()->getSinglePredecessor()) { | ||||||
| 563 | // We don't have a DT, but this trivially dominates. | ||||||
| 564 | return true; | ||||||
| 565 | } | ||||||
| 566 | |||||||
| 567 | return false; | ||||||
| 568 | } | ||||||
| 569 | |||||||
| 570 | static bool cmpExcludesZero(CmpInst::Predicate Pred, const Value *RHS) { | ||||||
| 571 | // v u> y implies v != 0. | ||||||
| 572 | if (Pred == ICmpInst::ICMP_UGT) | ||||||
| 573 | return true; | ||||||
| 574 | |||||||
| 575 | // Special-case v != 0 to also handle v != null. | ||||||
| 576 | if (Pred == ICmpInst::ICMP_NE) | ||||||
| 577 | return match(RHS, m_Zero()); | ||||||
| 578 | |||||||
| 579 | // All other predicates - rely on generic ConstantRange handling. | ||||||
| 580 | const APInt *C; | ||||||
| 581 | if (!match(RHS, m_APInt(C))) | ||||||
| 582 | return false; | ||||||
| 583 | |||||||
| 584 | ConstantRange TrueValues = ConstantRange::makeExactICmpRegion(Pred, *C); | ||||||
| 585 | return !TrueValues.contains(APInt::getNullValue(C->getBitWidth())); | ||||||
| 586 | } | ||||||
| 587 | |||||||
| 588 | static bool isKnownNonZeroFromAssume(const Value *V, const Query &Q) { | ||||||
| 589 | // Use of assumptions is context-sensitive. If we don't have a context, we | ||||||
| 590 | // cannot use them! | ||||||
| 591 | if (!Q.AC || !Q.CxtI) | ||||||
| 592 | return false; | ||||||
| 593 | |||||||
| 594 | if (Q.CxtI && V->getType()->isPointerTy()) { | ||||||
| 595 | SmallVector<Attribute::AttrKind, 2> AttrKinds{Attribute::NonNull}; | ||||||
| 596 | if (!NullPointerIsDefined(Q.CxtI->getFunction(), | ||||||
| 597 | V->getType()->getPointerAddressSpace())) | ||||||
| 598 | AttrKinds.push_back(Attribute::Dereferenceable); | ||||||
| 599 | |||||||
| 600 | if (getKnowledgeValidInContext(V, AttrKinds, Q.CxtI, Q.DT, Q.AC)) | ||||||
| 601 | return true; | ||||||
| 602 | } | ||||||
| 603 | |||||||
| 604 | for (auto &AssumeVH : Q.AC->assumptionsFor(V)) { | ||||||
| 605 | if (!AssumeVH) | ||||||
| 606 | continue; | ||||||
| 607 | CallInst *I = cast<CallInst>(AssumeVH); | ||||||
| 608 | assert(I->getFunction() == Q.CxtI->getFunction() &&((void)0) | ||||||
| 609 | "Got assumption for the wrong function!")((void)0); | ||||||
| 610 | |||||||
| 611 | // Warning: This loop can end up being somewhat performance sensitive. | ||||||
| 612 | // We're running this loop for once for each value queried resulting in a | ||||||
| 613 | // runtime of ~O(#assumes * #values). | ||||||
| 614 | |||||||
| 615 | assert(I->getCalledFunction()->getIntrinsicID() == Intrinsic::assume &&((void)0) | ||||||
| 616 | "must be an assume intrinsic")((void)0); | ||||||
| 617 | |||||||
| 618 | Value *RHS; | ||||||
| 619 | CmpInst::Predicate Pred; | ||||||
| 620 | auto m_V = m_CombineOr(m_Specific(V), m_PtrToInt(m_Specific(V))); | ||||||
| 621 | if (!match(I->getArgOperand(0), m_c_ICmp(Pred, m_V, m_Value(RHS)))) | ||||||
| 622 | return false; | ||||||
| 623 | |||||||
| 624 | if (cmpExcludesZero(Pred, RHS) && isValidAssumeForContext(I, Q.CxtI, Q.DT)) | ||||||
| 625 | return true; | ||||||
| 626 | } | ||||||
| 627 | |||||||
| 628 | return false; | ||||||
| 629 | } | ||||||
| 630 | |||||||
| 631 | static void computeKnownBitsFromAssume(const Value *V, KnownBits &Known, | ||||||
| 632 | unsigned Depth, const Query &Q) { | ||||||
| 633 | // Use of assumptions is context-sensitive. If we don't have a context, we | ||||||
| 634 | // cannot use them! | ||||||
| 635 | if (!Q.AC || !Q.CxtI) | ||||||
| 636 | return; | ||||||
| 637 | |||||||
| 638 | unsigned BitWidth = Known.getBitWidth(); | ||||||
| 639 | |||||||
| 640 | // Refine Known set if the pointer alignment is set by assume bundles. | ||||||
| 641 | if (V->getType()->isPointerTy()) { | ||||||
| 642 | if (RetainedKnowledge RK = getKnowledgeValidInContext( | ||||||
| 643 | V, {Attribute::Alignment}, Q.CxtI, Q.DT, Q.AC)) { | ||||||
| 644 | Known.Zero.setLowBits(Log2_32(RK.ArgValue)); | ||||||
| 645 | } | ||||||
| 646 | } | ||||||
| 647 | |||||||
| 648 | // Note that the patterns below need to be kept in sync with the code | ||||||
| 649 | // in AssumptionCache::updateAffectedValues. | ||||||
| 650 | |||||||
| 651 | for (auto &AssumeVH : Q.AC->assumptionsFor(V)) { | ||||||
| 652 | if (!AssumeVH) | ||||||
| 653 | continue; | ||||||
| 654 | CallInst *I = cast<CallInst>(AssumeVH); | ||||||
| 655 | assert(I->getParent()->getParent() == Q.CxtI->getParent()->getParent() &&((void)0) | ||||||
| 656 | "Got assumption for the wrong function!")((void)0); | ||||||
| 657 | |||||||
| 658 | // Warning: This loop can end up being somewhat performance sensitive. | ||||||
| 659 | // We're running this loop for once for each value queried resulting in a | ||||||
| 660 | // runtime of ~O(#assumes * #values). | ||||||
| 661 | |||||||
| 662 | assert(I->getCalledFunction()->getIntrinsicID() == Intrinsic::assume &&((void)0) | ||||||
| 663 | "must be an assume intrinsic")((void)0); | ||||||
| 664 | |||||||
| 665 | Value *Arg = I->getArgOperand(0); | ||||||
| 666 | |||||||
| 667 | if (Arg == V && isValidAssumeForContext(I, Q.CxtI, Q.DT)) { | ||||||
| 668 | assert(BitWidth == 1 && "assume operand is not i1?")((void)0); | ||||||
| 669 | Known.setAllOnes(); | ||||||
| 670 | return; | ||||||
| 671 | } | ||||||
| 672 | if (match(Arg, m_Not(m_Specific(V))) && | ||||||
| 673 | isValidAssumeForContext(I, Q.CxtI, Q.DT)) { | ||||||
| 674 | assert(BitWidth == 1 && "assume operand is not i1?")((void)0); | ||||||
| 675 | Known.setAllZero(); | ||||||
| 676 | return; | ||||||
| 677 | } | ||||||
| 678 | |||||||
| 679 | // The remaining tests are all recursive, so bail out if we hit the limit. | ||||||
| 680 | if (Depth == MaxAnalysisRecursionDepth) | ||||||
| 681 | continue; | ||||||
| 682 | |||||||
| 683 | ICmpInst *Cmp = dyn_cast<ICmpInst>(Arg); | ||||||
| 684 | if (!Cmp) | ||||||
| 685 | continue; | ||||||
| 686 | |||||||
| 687 | // We are attempting to compute known bits for the operands of an assume. | ||||||
| 688 | // Do not try to use other assumptions for those recursive calls because | ||||||
| 689 | // that can lead to mutual recursion and a compile-time explosion. | ||||||
| 690 | // An example of the mutual recursion: computeKnownBits can call | ||||||
| 691 | // isKnownNonZero which calls computeKnownBitsFromAssume (this function) | ||||||
| 692 | // and so on. | ||||||
| 693 | Query QueryNoAC = Q; | ||||||
| 694 | QueryNoAC.AC = nullptr; | ||||||
| 695 | |||||||
| 696 | // Note that ptrtoint may change the bitwidth. | ||||||
| 697 | Value *A, *B; | ||||||
| 698 | auto m_V = m_CombineOr(m_Specific(V), m_PtrToInt(m_Specific(V))); | ||||||
| 699 | |||||||
| 700 | CmpInst::Predicate Pred; | ||||||
| 701 | uint64_t C; | ||||||
| 702 | switch (Cmp->getPredicate()) { | ||||||
| 703 | default: | ||||||
| 704 | break; | ||||||
| 705 | case ICmpInst::ICMP_EQ: | ||||||
| 706 | // assume(v = a) | ||||||
| 707 | if (match(Cmp, m_c_ICmp(Pred, m_V, m_Value(A))) && | ||||||
| 708 | isValidAssumeForContext(I, Q.CxtI, Q.DT)) { | ||||||
| 709 | KnownBits RHSKnown = | ||||||
| 710 | computeKnownBits(A, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth); | ||||||
| 711 | Known.Zero |= RHSKnown.Zero; | ||||||
| 712 | Known.One |= RHSKnown.One; | ||||||
| 713 | // assume(v & b = a) | ||||||
| 714 | } else if (match(Cmp, | ||||||
| 715 | m_c_ICmp(Pred, m_c_And(m_V, m_Value(B)), m_Value(A))) && | ||||||
| 716 | isValidAssumeForContext(I, Q.CxtI, Q.DT)) { | ||||||
| 717 | KnownBits RHSKnown = | ||||||
| 718 | computeKnownBits(A, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth); | ||||||
| 719 | KnownBits MaskKnown = | ||||||
| 720 | computeKnownBits(B, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth); | ||||||
| 721 | |||||||
| 722 | // For those bits in the mask that are known to be one, we can propagate | ||||||
| 723 | // known bits from the RHS to V. | ||||||
| 724 | Known.Zero |= RHSKnown.Zero & MaskKnown.One; | ||||||
| 725 | Known.One |= RHSKnown.One & MaskKnown.One; | ||||||
| 726 | // assume(~(v & b) = a) | ||||||
| 727 | } else if (match(Cmp, m_c_ICmp(Pred, m_Not(m_c_And(m_V, m_Value(B))), | ||||||
| 728 | m_Value(A))) && | ||||||
| 729 | isValidAssumeForContext(I, Q.CxtI, Q.DT)) { | ||||||
| 730 | KnownBits RHSKnown = | ||||||
| 731 | computeKnownBits(A, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth); | ||||||
| 732 | KnownBits MaskKnown = | ||||||
| 733 | computeKnownBits(B, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth); | ||||||
| 734 | |||||||
| 735 | // For those bits in the mask that are known to be one, we can propagate | ||||||
| 736 | // inverted known bits from the RHS to V. | ||||||
| 737 | Known.Zero |= RHSKnown.One & MaskKnown.One; | ||||||
| 738 | Known.One |= RHSKnown.Zero & MaskKnown.One; | ||||||
| 739 | // assume(v | b = a) | ||||||
| 740 | } else if (match(Cmp, | ||||||
| 741 | m_c_ICmp(Pred, m_c_Or(m_V, m_Value(B)), m_Value(A))) && | ||||||
| 742 | isValidAssumeForContext(I, Q.CxtI, Q.DT)) { | ||||||
| 743 | KnownBits RHSKnown = | ||||||
| 744 | computeKnownBits(A, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth); | ||||||
| 745 | KnownBits BKnown = | ||||||
| 746 | computeKnownBits(B, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth); | ||||||
| 747 | |||||||
| 748 | // For those bits in B that are known to be zero, we can propagate known | ||||||
| 749 | // bits from the RHS to V. | ||||||
| 750 | Known.Zero |= RHSKnown.Zero & BKnown.Zero; | ||||||
| 751 | Known.One |= RHSKnown.One & BKnown.Zero; | ||||||
| 752 | // assume(~(v | b) = a) | ||||||
| 753 | } else if (match(Cmp, m_c_ICmp(Pred, m_Not(m_c_Or(m_V, m_Value(B))), | ||||||
| 754 | m_Value(A))) && | ||||||
| 755 | isValidAssumeForContext(I, Q.CxtI, Q.DT)) { | ||||||
| 756 | KnownBits RHSKnown = | ||||||
| 757 | computeKnownBits(A, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth); | ||||||
| 758 | KnownBits BKnown = | ||||||
| 759 | computeKnownBits(B, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth); | ||||||
| 760 | |||||||
| 761 | // For those bits in B that are known to be zero, we can propagate | ||||||
| 762 | // inverted known bits from the RHS to V. | ||||||
| 763 | Known.Zero |= RHSKnown.One & BKnown.Zero; | ||||||
| 764 | Known.One |= RHSKnown.Zero & BKnown.Zero; | ||||||
| 765 | // assume(v ^ b = a) | ||||||
| 766 | } else if (match(Cmp, | ||||||
| 767 | m_c_ICmp(Pred, m_c_Xor(m_V, m_Value(B)), m_Value(A))) && | ||||||
| 768 | isValidAssumeForContext(I, Q.CxtI, Q.DT)) { | ||||||
| 769 | KnownBits RHSKnown = | ||||||
| 770 | computeKnownBits(A, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth); | ||||||
| 771 | KnownBits BKnown = | ||||||
| 772 | computeKnownBits(B, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth); | ||||||
| 773 | |||||||
| 774 | // For those bits in B that are known to be zero, we can propagate known | ||||||
| 775 | // bits from the RHS to V. For those bits in B that are known to be one, | ||||||
| 776 | // we can propagate inverted known bits from the RHS to V. | ||||||
| 777 | Known.Zero |= RHSKnown.Zero & BKnown.Zero; | ||||||
| 778 | Known.One |= RHSKnown.One & BKnown.Zero; | ||||||
| 779 | Known.Zero |= RHSKnown.One & BKnown.One; | ||||||
| 780 | Known.One |= RHSKnown.Zero & BKnown.One; | ||||||
| 781 | // assume(~(v ^ b) = a) | ||||||
| 782 | } else if (match(Cmp, m_c_ICmp(Pred, m_Not(m_c_Xor(m_V, m_Value(B))), | ||||||
| 783 | m_Value(A))) && | ||||||
| 784 | isValidAssumeForContext(I, Q.CxtI, Q.DT)) { | ||||||
| 785 | KnownBits RHSKnown = | ||||||
| 786 | computeKnownBits(A, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth); | ||||||
| 787 | KnownBits BKnown = | ||||||
| 788 | computeKnownBits(B, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth); | ||||||
| 789 | |||||||
| 790 | // For those bits in B that are known to be zero, we can propagate | ||||||
| 791 | // inverted known bits from the RHS to V. For those bits in B that are | ||||||
| 792 | // known to be one, we can propagate known bits from the RHS to V. | ||||||
| 793 | Known.Zero |= RHSKnown.One & BKnown.Zero; | ||||||
| 794 | Known.One |= RHSKnown.Zero & BKnown.Zero; | ||||||
| 795 | Known.Zero |= RHSKnown.Zero & BKnown.One; | ||||||
| 796 | Known.One |= RHSKnown.One & BKnown.One; | ||||||
| 797 | // assume(v << c = a) | ||||||
| 798 | } else if (match(Cmp, m_c_ICmp(Pred, m_Shl(m_V, m_ConstantInt(C)), | ||||||
| 799 | m_Value(A))) && | ||||||
| 800 | isValidAssumeForContext(I, Q.CxtI, Q.DT) && C < BitWidth) { | ||||||
| 801 | KnownBits RHSKnown = | ||||||
| 802 | computeKnownBits(A, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth); | ||||||
| 803 | |||||||
| 804 | // For those bits in RHS that are known, we can propagate them to known | ||||||
| 805 | // bits in V shifted to the right by C. | ||||||
| 806 | RHSKnown.Zero.lshrInPlace(C); | ||||||
| 807 | Known.Zero |= RHSKnown.Zero; | ||||||
| 808 | RHSKnown.One.lshrInPlace(C); | ||||||
| 809 | Known.One |= RHSKnown.One; | ||||||
| 810 | // assume(~(v << c) = a) | ||||||
| 811 | } else if (match(Cmp, m_c_ICmp(Pred, m_Not(m_Shl(m_V, m_ConstantInt(C))), | ||||||
| 812 | m_Value(A))) && | ||||||
| 813 | isValidAssumeForContext(I, Q.CxtI, Q.DT) && C < BitWidth) { | ||||||
| 814 | KnownBits RHSKnown = | ||||||
| 815 | computeKnownBits(A, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth); | ||||||
| 816 | // For those bits in RHS that are known, we can propagate them inverted | ||||||
| 817 | // to known bits in V shifted to the right by C. | ||||||
| 818 | RHSKnown.One.lshrInPlace(C); | ||||||
| 819 | Known.Zero |= RHSKnown.One; | ||||||
| 820 | RHSKnown.Zero.lshrInPlace(C); | ||||||
| 821 | Known.One |= RHSKnown.Zero; | ||||||
| 822 | // assume(v >> c = a) | ||||||
| 823 | } else if (match(Cmp, m_c_ICmp(Pred, m_Shr(m_V, m_ConstantInt(C)), | ||||||
| 824 | m_Value(A))) && | ||||||
| 825 | isValidAssumeForContext(I, Q.CxtI, Q.DT) && C < BitWidth) { | ||||||
| 826 | KnownBits RHSKnown = | ||||||
| 827 | computeKnownBits(A, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth); | ||||||
| 828 | // For those bits in RHS that are known, we can propagate them to known | ||||||
| 829 | // bits in V shifted to the right by C. | ||||||
| 830 | Known.Zero |= RHSKnown.Zero << C; | ||||||
| 831 | Known.One |= RHSKnown.One << C; | ||||||
| 832 | // assume(~(v >> c) = a) | ||||||
| 833 | } else if (match(Cmp, m_c_ICmp(Pred, m_Not(m_Shr(m_V, m_ConstantInt(C))), | ||||||
| 834 | m_Value(A))) && | ||||||
| 835 | isValidAssumeForContext(I, Q.CxtI, Q.DT) && C < BitWidth) { | ||||||
| 836 | KnownBits RHSKnown = | ||||||
| 837 | computeKnownBits(A, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth); | ||||||
| 838 | // For those bits in RHS that are known, we can propagate them inverted | ||||||
| 839 | // to known bits in V shifted to the right by C. | ||||||
| 840 | Known.Zero |= RHSKnown.One << C; | ||||||
| 841 | Known.One |= RHSKnown.Zero << C; | ||||||
| 842 | } | ||||||
| 843 | break; | ||||||
| 844 | case ICmpInst::ICMP_SGE: | ||||||
| 845 | // assume(v >=_s c) where c is non-negative | ||||||
| 846 | if (match(Cmp, m_ICmp(Pred, m_V, m_Value(A))) && | ||||||
| 847 | isValidAssumeForContext(I, Q.CxtI, Q.DT)) { | ||||||
| 848 | KnownBits RHSKnown = | ||||||
| 849 | computeKnownBits(A, Depth + 1, QueryNoAC).anyextOrTrunc(BitWidth); | ||||||
| 850 | |||||||
| 851 | if (RHSKnown.isNonNegative()) { | ||||||
| 852 | // We know that the sign bit is zero. | ||||||
| 853 | Known.makeNonNegative(); | ||||||
| 854 | } | ||||||
| 855 | } | ||||||
| 856 | break; | ||||||
| 857 | case ICmpInst::ICMP_SGT: | ||||||
| 858 | // assume(v >_s c) where c is at least -1. | ||||||
| 859 | if (match(Cmp, m_ICmp(Pred, m_V, m_Value(A))) && | ||||||
| 860 | isValidAssumeForContext(I, Q.CxtI, Q.DT)) { | ||||||
| 861 | KnownBits RHSKnown = | ||||||
| 862 | computeKnownBits(A, Depth + 1, QueryNoAC).anyextOrTrunc(BitWidth); | ||||||
| 863 | |||||||
| 864 | if (RHSKnown.isAllOnes() || RHSKnown.isNonNegative()) { | ||||||
| 865 | // We know that the sign bit is zero. | ||||||
| 866 | Known.makeNonNegative(); | ||||||
| 867 | } | ||||||
| 868 | } | ||||||
| 869 | break; | ||||||
| 870 | case ICmpInst::ICMP_SLE: | ||||||
| 871 | // assume(v <=_s c) where c is negative | ||||||
| 872 | if (match(Cmp, m_ICmp(Pred, m_V, m_Value(A))) && | ||||||
| 873 | isValidAssumeForContext(I, Q.CxtI, Q.DT)) { | ||||||
| 874 | KnownBits RHSKnown = | ||||||
| 875 | computeKnownBits(A, Depth + 1, QueryNoAC).anyextOrTrunc(BitWidth); | ||||||
| 876 | |||||||
| 877 | if (RHSKnown.isNegative()) { | ||||||
| 878 | // We know that the sign bit is one. | ||||||
| 879 | Known.makeNegative(); | ||||||
| 880 | } | ||||||
| 881 | } | ||||||
| 882 | break; | ||||||
| 883 | case ICmpInst::ICMP_SLT: | ||||||
| 884 | // assume(v <_s c) where c is non-positive | ||||||
| 885 | if (match(Cmp, m_ICmp(Pred, m_V, m_Value(A))) && | ||||||
| 886 | isValidAssumeForContext(I, Q.CxtI, Q.DT)) { | ||||||
| 887 | KnownBits RHSKnown = | ||||||
| 888 | computeKnownBits(A, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth); | ||||||
| 889 | |||||||
| 890 | if (RHSKnown.isZero() || RHSKnown.isNegative()) { | ||||||
| 891 | // We know that the sign bit is one. | ||||||
| 892 | Known.makeNegative(); | ||||||
| 893 | } | ||||||
| 894 | } | ||||||
| 895 | break; | ||||||
| 896 | case ICmpInst::ICMP_ULE: | ||||||
| 897 | // assume(v <=_u c) | ||||||
| 898 | if (match(Cmp, m_ICmp(Pred, m_V, m_Value(A))) && | ||||||
| 899 | isValidAssumeForContext(I, Q.CxtI, Q.DT)) { | ||||||
| 900 | KnownBits RHSKnown = | ||||||
| 901 | computeKnownBits(A, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth); | ||||||
| 902 | |||||||
| 903 | // Whatever high bits in c are zero are known to be zero. | ||||||
| 904 | Known.Zero.setHighBits(RHSKnown.countMinLeadingZeros()); | ||||||
| 905 | } | ||||||
| 906 | break; | ||||||
| 907 | case ICmpInst::ICMP_ULT: | ||||||
| 908 | // assume(v <_u c) | ||||||
| 909 | if (match(Cmp, m_ICmp(Pred, m_V, m_Value(A))) && | ||||||
| 910 | isValidAssumeForContext(I, Q.CxtI, Q.DT)) { | ||||||
| 911 | KnownBits RHSKnown = | ||||||
| 912 | computeKnownBits(A, Depth+1, QueryNoAC).anyextOrTrunc(BitWidth); | ||||||
| 913 | |||||||
| 914 | // If the RHS is known zero, then this assumption must be wrong (nothing | ||||||
| 915 | // is unsigned less than zero). Signal a conflict and get out of here. | ||||||
| 916 | if (RHSKnown.isZero()) { | ||||||
| 917 | Known.Zero.setAllBits(); | ||||||
| 918 | Known.One.setAllBits(); | ||||||
| 919 | break; | ||||||
| 920 | } | ||||||
| 921 | |||||||
| 922 | // Whatever high bits in c are zero are known to be zero (if c is a power | ||||||
| 923 | // of 2, then one more). | ||||||
| 924 | if (isKnownToBeAPowerOfTwo(A, false, Depth + 1, QueryNoAC)) | ||||||
| 925 | Known.Zero.setHighBits(RHSKnown.countMinLeadingZeros() + 1); | ||||||
| 926 | else | ||||||
| 927 | Known.Zero.setHighBits(RHSKnown.countMinLeadingZeros()); | ||||||
| 928 | } | ||||||
| 929 | break; | ||||||
| 930 | } | ||||||
| 931 | } | ||||||
| 932 | |||||||
| 933 | // If assumptions conflict with each other or previous known bits, then we | ||||||
| 934 | // have a logical fallacy. It's possible that the assumption is not reachable, | ||||||
| 935 | // so this isn't a real bug. On the other hand, the program may have undefined | ||||||
| 936 | // behavior, or we might have a bug in the compiler. We can't assert/crash, so | ||||||
| 937 | // clear out the known bits, try to warn the user, and hope for the best. | ||||||
| 938 | if (Known.Zero.intersects(Known.One)) { | ||||||
| 939 | Known.resetAll(); | ||||||
| 940 | |||||||
| 941 | if (Q.ORE) | ||||||
| 942 | Q.ORE->emit([&]() { | ||||||
| 943 | auto *CxtI = const_cast<Instruction *>(Q.CxtI); | ||||||
| 944 | return OptimizationRemarkAnalysis("value-tracking", "BadAssumption", | ||||||
| 945 | CxtI) | ||||||
| 946 | << "Detected conflicting code assumptions. Program may " | ||||||
| 947 | "have undefined behavior, or compiler may have " | ||||||
| 948 | "internal error."; | ||||||
| 949 | }); | ||||||
| 950 | } | ||||||
| 951 | } | ||||||
| 952 | |||||||
| 953 | /// Compute known bits from a shift operator, including those with a | ||||||
| 954 | /// non-constant shift amount. Known is the output of this function. Known2 is a | ||||||
| 955 | /// pre-allocated temporary with the same bit width as Known and on return | ||||||
| 956 | /// contains the known bit of the shift value source. KF is an | ||||||
| 957 | /// operator-specific function that, given the known-bits and a shift amount, | ||||||
| 958 | /// compute the implied known-bits of the shift operator's result respectively | ||||||
| 959 | /// for that shift amount. The results from calling KF are conservatively | ||||||
| 960 | /// combined for all permitted shift amounts. | ||||||
| 961 | static void computeKnownBitsFromShiftOperator( | ||||||
| 962 | const Operator *I, const APInt &DemandedElts, KnownBits &Known, | ||||||
| 963 | KnownBits &Known2, unsigned Depth, const Query &Q, | ||||||
| 964 | function_ref<KnownBits(const KnownBits &, const KnownBits &)> KF) { | ||||||
| 965 | unsigned BitWidth = Known.getBitWidth(); | ||||||
| 966 | computeKnownBits(I->getOperand(0), DemandedElts, Known2, Depth + 1, Q); | ||||||
| 967 | computeKnownBits(I->getOperand(1), DemandedElts, Known, Depth + 1, Q); | ||||||
| 968 | |||||||
| 969 | // Note: We cannot use Known.Zero.getLimitedValue() here, because if | ||||||
| 970 | // BitWidth > 64 and any upper bits are known, we'll end up returning the | ||||||
| 971 | // limit value (which implies all bits are known). | ||||||
| 972 | uint64_t ShiftAmtKZ = Known.Zero.zextOrTrunc(64).getZExtValue(); | ||||||
| 973 | uint64_t ShiftAmtKO = Known.One.zextOrTrunc(64).getZExtValue(); | ||||||
| 974 | bool ShiftAmtIsConstant = Known.isConstant(); | ||||||
| 975 | bool MaxShiftAmtIsOutOfRange = Known.getMaxValue().uge(BitWidth); | ||||||
| 976 | |||||||
| 977 | if (ShiftAmtIsConstant) { | ||||||
| 978 | Known = KF(Known2, Known); | ||||||
| 979 | |||||||
| 980 | // If the known bits conflict, this must be an overflowing left shift, so | ||||||
| 981 | // the shift result is poison. We can return anything we want. Choose 0 for | ||||||
| 982 | // the best folding opportunity. | ||||||
| 983 | if (Known.hasConflict()) | ||||||
| 984 | Known.setAllZero(); | ||||||
| 985 | |||||||
| 986 | return; | ||||||
| 987 | } | ||||||
| 988 | |||||||
| 989 | // If the shift amount could be greater than or equal to the bit-width of the | ||||||
| 990 | // LHS, the value could be poison, but bail out because the check below is | ||||||
| 991 | // expensive. | ||||||
| 992 | // TODO: Should we just carry on? | ||||||
| 993 | if (MaxShiftAmtIsOutOfRange) { | ||||||
| 994 | Known.resetAll(); | ||||||
| 995 | return; | ||||||
| 996 | } | ||||||
| 997 | |||||||
| 998 | // It would be more-clearly correct to use the two temporaries for this | ||||||
| 999 | // calculation. Reusing the APInts here to prevent unnecessary allocations. | ||||||
| 1000 | Known.resetAll(); | ||||||
| 1001 | |||||||
| 1002 | // If we know the shifter operand is nonzero, we can sometimes infer more | ||||||
| 1003 | // known bits. However this is expensive to compute, so be lazy about it and | ||||||
| 1004 | // only compute it when absolutely necessary. | ||||||
| 1005 | Optional<bool> ShifterOperandIsNonZero; | ||||||
| 1006 | |||||||
| 1007 | // Early exit if we can't constrain any well-defined shift amount. | ||||||
| 1008 | if (!(ShiftAmtKZ & (PowerOf2Ceil(BitWidth) - 1)) && | ||||||
| 1009 | !(ShiftAmtKO & (PowerOf2Ceil(BitWidth) - 1))) { | ||||||
| 1010 | ShifterOperandIsNonZero = | ||||||
| 1011 | isKnownNonZero(I->getOperand(1), DemandedElts, Depth + 1, Q); | ||||||
| 1012 | if (!*ShifterOperandIsNonZero) | ||||||
| 1013 | return; | ||||||
| 1014 | } | ||||||
| 1015 | |||||||
| 1016 | Known.Zero.setAllBits(); | ||||||
| 1017 | Known.One.setAllBits(); | ||||||
| 1018 | for (unsigned ShiftAmt = 0; ShiftAmt < BitWidth; ++ShiftAmt) { | ||||||
| 1019 | // Combine the shifted known input bits only for those shift amounts | ||||||
| 1020 | // compatible with its known constraints. | ||||||
| 1021 | if ((ShiftAmt & ~ShiftAmtKZ) != ShiftAmt) | ||||||
| 1022 | continue; | ||||||
| 1023 | if ((ShiftAmt | ShiftAmtKO) != ShiftAmt) | ||||||
| 1024 | continue; | ||||||
| 1025 | // If we know the shifter is nonzero, we may be able to infer more known | ||||||
| 1026 | // bits. This check is sunk down as far as possible to avoid the expensive | ||||||
| 1027 | // call to isKnownNonZero if the cheaper checks above fail. | ||||||
| 1028 | if (ShiftAmt == 0) { | ||||||
| 1029 | if (!ShifterOperandIsNonZero.hasValue()) | ||||||
| 1030 | ShifterOperandIsNonZero = | ||||||
| 1031 | isKnownNonZero(I->getOperand(1), DemandedElts, Depth + 1, Q); | ||||||
| 1032 | if (*ShifterOperandIsNonZero) | ||||||
| 1033 | continue; | ||||||
| 1034 | } | ||||||
| 1035 | |||||||
| 1036 | Known = KnownBits::commonBits( | ||||||
| 1037 | Known, KF(Known2, KnownBits::makeConstant(APInt(32, ShiftAmt)))); | ||||||
| 1038 | } | ||||||
| 1039 | |||||||
| 1040 | // If the known bits conflict, the result is poison. Return a 0 and hope the | ||||||
| 1041 | // caller can further optimize that. | ||||||
| 1042 | if (Known.hasConflict()) | ||||||
| 1043 | Known.setAllZero(); | ||||||
| 1044 | } | ||||||
| 1045 | |||||||
| 1046 | static void computeKnownBitsFromOperator(const Operator *I, | ||||||
| 1047 | const APInt &DemandedElts, | ||||||
| 1048 | KnownBits &Known, unsigned Depth, | ||||||
| 1049 | const Query &Q) { | ||||||
| 1050 | unsigned BitWidth = Known.getBitWidth(); | ||||||
| 1051 | |||||||
| 1052 | KnownBits Known2(BitWidth); | ||||||
| 1053 | switch (I->getOpcode()) { | ||||||
| 1054 | default: break; | ||||||
| 1055 | case Instruction::Load: | ||||||
| 1056 | if (MDNode *MD = | ||||||
| 1057 | Q.IIQ.getMetadata(cast<LoadInst>(I), LLVMContext::MD_range)) | ||||||
| 1058 | computeKnownBitsFromRangeMetadata(*MD, Known); | ||||||
| 1059 | break; | ||||||
| 1060 | case Instruction::And: { | ||||||
| 1061 | // If either the LHS or the RHS are Zero, the result is zero. | ||||||
| 1062 | computeKnownBits(I->getOperand(1), DemandedElts, Known, Depth + 1, Q); | ||||||
| 1063 | computeKnownBits(I->getOperand(0), DemandedElts, Known2, Depth + 1, Q); | ||||||
| 1064 | |||||||
| 1065 | Known &= Known2; | ||||||
| 1066 | |||||||
| 1067 | // and(x, add (x, -1)) is a common idiom that always clears the low bit; | ||||||
| 1068 | // here we handle the more general case of adding any odd number by | ||||||
| 1069 | // matching the form add(x, add(x, y)) where y is odd. | ||||||
| 1070 | // TODO: This could be generalized to clearing any bit set in y where the | ||||||
| 1071 | // following bit is known to be unset in y. | ||||||
| 1072 | Value *X = nullptr, *Y = nullptr; | ||||||
| 1073 | if (!Known.Zero[0] && !Known.One[0] && | ||||||
| 1074 | match(I, m_c_BinOp(m_Value(X), m_Add(m_Deferred(X), m_Value(Y))))) { | ||||||
| 1075 | Known2.resetAll(); | ||||||
| 1076 | computeKnownBits(Y, DemandedElts, Known2, Depth + 1, Q); | ||||||
| 1077 | if (Known2.countMinTrailingOnes() > 0) | ||||||
| 1078 | Known.Zero.setBit(0); | ||||||
| 1079 | } | ||||||
| 1080 | break; | ||||||
| 1081 | } | ||||||
| 1082 | case Instruction::Or: | ||||||
| 1083 | computeKnownBits(I->getOperand(1), DemandedElts, Known, Depth + 1, Q); | ||||||
| 1084 | computeKnownBits(I->getOperand(0), DemandedElts, Known2, Depth + 1, Q); | ||||||
| 1085 | |||||||
| 1086 | Known |= Known2; | ||||||
| 1087 | break; | ||||||
| 1088 | case Instruction::Xor: | ||||||
| 1089 | computeKnownBits(I->getOperand(1), DemandedElts, Known, Depth + 1, Q); | ||||||
| 1090 | computeKnownBits(I->getOperand(0), DemandedElts, Known2, Depth + 1, Q); | ||||||
| 1091 | |||||||
| 1092 | Known ^= Known2; | ||||||
| 1093 | break; | ||||||
| 1094 | case Instruction::Mul: { | ||||||
| 1095 | bool NSW = Q.IIQ.hasNoSignedWrap(cast<OverflowingBinaryOperator>(I)); | ||||||
| 1096 | computeKnownBitsMul(I->getOperand(0), I->getOperand(1), NSW, DemandedElts, | ||||||
| 1097 | Known, Known2, Depth, Q); | ||||||
| 1098 | break; | ||||||
| 1099 | } | ||||||
| 1100 | case Instruction::UDiv: { | ||||||
| 1101 | computeKnownBits(I->getOperand(0), Known, Depth + 1, Q); | ||||||
| 1102 | computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q); | ||||||
| 1103 | Known = KnownBits::udiv(Known, Known2); | ||||||
| 1104 | break; | ||||||
| 1105 | } | ||||||
| 1106 | case Instruction::Select: { | ||||||
| 1107 | const Value *LHS = nullptr, *RHS = nullptr; | ||||||
| 1108 | SelectPatternFlavor SPF = matchSelectPattern(I, LHS, RHS).Flavor; | ||||||
| 1109 | if (SelectPatternResult::isMinOrMax(SPF)) { | ||||||
| 1110 | computeKnownBits(RHS, Known, Depth + 1, Q); | ||||||
| 1111 | computeKnownBits(LHS, Known2, Depth + 1, Q); | ||||||
| 1112 | switch (SPF) { | ||||||
| 1113 | default: | ||||||
| 1114 | llvm_unreachable("Unhandled select pattern flavor!")__builtin_unreachable(); | ||||||
| 1115 | case SPF_SMAX: | ||||||
| 1116 | Known = KnownBits::smax(Known, Known2); | ||||||
| 1117 | break; | ||||||
| 1118 | case SPF_SMIN: | ||||||
| 1119 | Known = KnownBits::smin(Known, Known2); | ||||||
| 1120 | break; | ||||||
| 1121 | case SPF_UMAX: | ||||||
| 1122 | Known = KnownBits::umax(Known, Known2); | ||||||
| 1123 | break; | ||||||
| 1124 | case SPF_UMIN: | ||||||
| 1125 | Known = KnownBits::umin(Known, Known2); | ||||||
| 1126 | break; | ||||||
| 1127 | } | ||||||
| 1128 | break; | ||||||
| 1129 | } | ||||||
| 1130 | |||||||
| 1131 | computeKnownBits(I->getOperand(2), Known, Depth + 1, Q); | ||||||
| 1132 | computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q); | ||||||
| 1133 | |||||||
| 1134 | // Only known if known in both the LHS and RHS. | ||||||
| 1135 | Known = KnownBits::commonBits(Known, Known2); | ||||||
| 1136 | |||||||
| 1137 | if (SPF == SPF_ABS) { | ||||||
| 1138 | // RHS from matchSelectPattern returns the negation part of abs pattern. | ||||||
| 1139 | // If the negate has an NSW flag we can assume the sign bit of the result | ||||||
| 1140 | // will be 0 because that makes abs(INT_MIN) undefined. | ||||||
| 1141 | if (match(RHS, m_Neg(m_Specific(LHS))) && | ||||||
| 1142 | Q.IIQ.hasNoSignedWrap(cast<Instruction>(RHS))) | ||||||
| 1143 | Known.Zero.setSignBit(); | ||||||
| 1144 | } | ||||||
| 1145 | |||||||
| 1146 | break; | ||||||
| 1147 | } | ||||||
| 1148 | case Instruction::FPTrunc: | ||||||
| 1149 | case Instruction::FPExt: | ||||||
| 1150 | case Instruction::FPToUI: | ||||||
| 1151 | case Instruction::FPToSI: | ||||||
| 1152 | case Instruction::SIToFP: | ||||||
| 1153 | case Instruction::UIToFP: | ||||||
| 1154 | break; // Can't work with floating point. | ||||||
| 1155 | case Instruction::PtrToInt: | ||||||
| 1156 | case Instruction::IntToPtr: | ||||||
| 1157 | // Fall through and handle them the same as zext/trunc. | ||||||
| 1158 | LLVM_FALLTHROUGH[[gnu::fallthrough]]; | ||||||
| 1159 | case Instruction::ZExt: | ||||||
| 1160 | case Instruction::Trunc: { | ||||||
| 1161 | Type *SrcTy = I->getOperand(0)->getType(); | ||||||
| 1162 | |||||||
| 1163 | unsigned SrcBitWidth; | ||||||
| 1164 | // Note that we handle pointer operands here because of inttoptr/ptrtoint | ||||||
| 1165 | // which fall through here. | ||||||
| 1166 | Type *ScalarTy = SrcTy->getScalarType(); | ||||||
| 1167 | SrcBitWidth = ScalarTy->isPointerTy() ? | ||||||
| 1168 | Q.DL.getPointerTypeSizeInBits(ScalarTy) : | ||||||
| 1169 | Q.DL.getTypeSizeInBits(ScalarTy); | ||||||
| 1170 | |||||||
| 1171 | assert(SrcBitWidth && "SrcBitWidth can't be zero")((void)0); | ||||||
| 1172 | Known = Known.anyextOrTrunc(SrcBitWidth); | ||||||
| 1173 | computeKnownBits(I->getOperand(0), Known, Depth + 1, Q); | ||||||
| 1174 | Known = Known.zextOrTrunc(BitWidth); | ||||||
| 1175 | break; | ||||||
| 1176 | } | ||||||
| 1177 | case Instruction::BitCast: { | ||||||
| 1178 | Type *SrcTy = I->getOperand(0)->getType(); | ||||||
| 1179 | if (SrcTy->isIntOrPtrTy() && | ||||||
| 1180 | // TODO: For now, not handling conversions like: | ||||||
| 1181 | // (bitcast i64 %x to <2 x i32>) | ||||||
| 1182 | !I->getType()->isVectorTy()) { | ||||||
| 1183 | computeKnownBits(I->getOperand(0), Known, Depth + 1, Q); | ||||||
| 1184 | break; | ||||||
| 1185 | } | ||||||
| 1186 | |||||||
| 1187 | // Handle cast from vector integer type to scalar or vector integer. | ||||||
| 1188 | auto *SrcVecTy = dyn_cast<FixedVectorType>(SrcTy); | ||||||
| 1189 | if (!SrcVecTy || !SrcVecTy->getElementType()->isIntegerTy() || | ||||||
| 1190 | !I->getType()->isIntOrIntVectorTy()) | ||||||
| 1191 | break; | ||||||
| 1192 | |||||||
| 1193 | // Look through a cast from narrow vector elements to wider type. | ||||||
| 1194 | // Examples: v4i32 -> v2i64, v3i8 -> v24 | ||||||
| 1195 | unsigned SubBitWidth = SrcVecTy->getScalarSizeInBits(); | ||||||
| 1196 | if (BitWidth % SubBitWidth == 0) { | ||||||
| 1197 | // Known bits are automatically intersected across demanded elements of a | ||||||
| 1198 | // vector. So for example, if a bit is computed as known zero, it must be | ||||||
| 1199 | // zero across all demanded elements of the vector. | ||||||
| 1200 | // | ||||||
| 1201 | // For this bitcast, each demanded element of the output is sub-divided | ||||||
| 1202 | // across a set of smaller vector elements in the source vector. To get | ||||||
| 1203 | // the known bits for an entire element of the output, compute the known | ||||||
| 1204 | // bits for each sub-element sequentially. This is done by shifting the | ||||||
| 1205 | // one-set-bit demanded elements parameter across the sub-elements for | ||||||
| 1206 | // consecutive calls to computeKnownBits. We are using the demanded | ||||||
| 1207 | // elements parameter as a mask operator. | ||||||
| 1208 | // | ||||||
| 1209 | // The known bits of each sub-element are then inserted into place | ||||||
| 1210 | // (dependent on endian) to form the full result of known bits. | ||||||
| 1211 | unsigned NumElts = DemandedElts.getBitWidth(); | ||||||
| 1212 | unsigned SubScale = BitWidth / SubBitWidth; | ||||||
| 1213 | APInt SubDemandedElts = APInt::getNullValue(NumElts * SubScale); | ||||||
| 1214 | for (unsigned i = 0; i != NumElts; ++i) { | ||||||
| 1215 | if (DemandedElts[i]) | ||||||
| 1216 | SubDemandedElts.setBit(i * SubScale); | ||||||
| 1217 | } | ||||||
| 1218 | |||||||
| 1219 | KnownBits KnownSrc(SubBitWidth); | ||||||
| 1220 | for (unsigned i = 0; i != SubScale; ++i) { | ||||||
| 1221 | computeKnownBits(I->getOperand(0), SubDemandedElts.shl(i), KnownSrc, | ||||||
| 1222 | Depth + 1, Q); | ||||||
| 1223 | unsigned ShiftElt = Q.DL.isLittleEndian() ? i : SubScale - 1 - i; | ||||||
| 1224 | Known.insertBits(KnownSrc, ShiftElt * SubBitWidth); | ||||||
| 1225 | } | ||||||
| 1226 | } | ||||||
| 1227 | break; | ||||||
| 1228 | } | ||||||
| 1229 | case Instruction::SExt: { | ||||||
| 1230 | // Compute the bits in the result that are not present in the input. | ||||||
| 1231 | unsigned SrcBitWidth = I->getOperand(0)->getType()->getScalarSizeInBits(); | ||||||
| 1232 | |||||||
| 1233 | Known = Known.trunc(SrcBitWidth); | ||||||
| 1234 | computeKnownBits(I->getOperand(0), Known, Depth + 1, Q); | ||||||
| 1235 | // If the sign bit of the input is known set or clear, then we know the | ||||||
| 1236 | // top bits of the result. | ||||||
| 1237 | Known = Known.sext(BitWidth); | ||||||
| 1238 | break; | ||||||
| 1239 | } | ||||||
| 1240 | case Instruction::Shl: { | ||||||
| 1241 | bool NSW = Q.IIQ.hasNoSignedWrap(cast<OverflowingBinaryOperator>(I)); | ||||||
| 1242 | auto KF = [NSW](const KnownBits &KnownVal, const KnownBits &KnownAmt) { | ||||||
| 1243 | KnownBits Result = KnownBits::shl(KnownVal, KnownAmt); | ||||||
| 1244 | // If this shift has "nsw" keyword, then the result is either a poison | ||||||
| 1245 | // value or has the same sign bit as the first operand. | ||||||
| 1246 | if (NSW) { | ||||||
| 1247 | if (KnownVal.Zero.isSignBitSet()) | ||||||
| 1248 | Result.Zero.setSignBit(); | ||||||
| 1249 | if (KnownVal.One.isSignBitSet()) | ||||||
| 1250 | Result.One.setSignBit(); | ||||||
| 1251 | } | ||||||
| 1252 | return Result; | ||||||
| 1253 | }; | ||||||
| 1254 | computeKnownBitsFromShiftOperator(I, DemandedElts, Known, Known2, Depth, Q, | ||||||
| 1255 | KF); | ||||||
| 1256 | // Trailing zeros of a right-shifted constant never decrease. | ||||||
| 1257 | const APInt *C; | ||||||
| 1258 | if (match(I->getOperand(0), m_APInt(C))) | ||||||
| 1259 | Known.Zero.setLowBits(C->countTrailingZeros()); | ||||||
| 1260 | break; | ||||||
| 1261 | } | ||||||
| 1262 | case Instruction::LShr: { | ||||||
| 1263 | auto KF = [](const KnownBits &KnownVal, const KnownBits &KnownAmt) { | ||||||
| 1264 | return KnownBits::lshr(KnownVal, KnownAmt); | ||||||
| 1265 | }; | ||||||
| 1266 | computeKnownBitsFromShiftOperator(I, DemandedElts, Known, Known2, Depth, Q, | ||||||
| 1267 | KF); | ||||||
| 1268 | // Leading zeros of a left-shifted constant never decrease. | ||||||
| 1269 | const APInt *C; | ||||||
| 1270 | if (match(I->getOperand(0), m_APInt(C))) | ||||||
| 1271 | Known.Zero.setHighBits(C->countLeadingZeros()); | ||||||
| 1272 | break; | ||||||
| 1273 | } | ||||||
| 1274 | case Instruction::AShr: { | ||||||
| 1275 | auto KF = [](const KnownBits &KnownVal, const KnownBits &KnownAmt) { | ||||||
| 1276 | return KnownBits::ashr(KnownVal, KnownAmt); | ||||||
| 1277 | }; | ||||||
| 1278 | computeKnownBitsFromShiftOperator(I, DemandedElts, Known, Known2, Depth, Q, | ||||||
| 1279 | KF); | ||||||
| 1280 | break; | ||||||
| 1281 | } | ||||||
| 1282 | case Instruction::Sub: { | ||||||
| 1283 | bool NSW = Q.IIQ.hasNoSignedWrap(cast<OverflowingBinaryOperator>(I)); | ||||||
| 1284 | computeKnownBitsAddSub(false, I->getOperand(0), I->getOperand(1), NSW, | ||||||
| 1285 | DemandedElts, Known, Known2, Depth, Q); | ||||||
| 1286 | break; | ||||||
| 1287 | } | ||||||
| 1288 | case Instruction::Add: { | ||||||
| 1289 | bool NSW = Q.IIQ.hasNoSignedWrap(cast<OverflowingBinaryOperator>(I)); | ||||||
| 1290 | computeKnownBitsAddSub(true, I->getOperand(0), I->getOperand(1), NSW, | ||||||
| 1291 | DemandedElts, Known, Known2, Depth, Q); | ||||||
| 1292 | break; | ||||||
| 1293 | } | ||||||
| 1294 | case Instruction::SRem: | ||||||
| 1295 | computeKnownBits(I->getOperand(0), Known, Depth + 1, Q); | ||||||
| 1296 | computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q); | ||||||
| 1297 | Known = KnownBits::srem(Known, Known2); | ||||||
| 1298 | break; | ||||||
| 1299 | |||||||
| 1300 | case Instruction::URem: | ||||||
| 1301 | computeKnownBits(I->getOperand(0), Known, Depth + 1, Q); | ||||||
| 1302 | computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q); | ||||||
| 1303 | Known = KnownBits::urem(Known, Known2); | ||||||
| 1304 | break; | ||||||
| 1305 | case Instruction::Alloca: | ||||||
| 1306 | Known.Zero.setLowBits(Log2(cast<AllocaInst>(I)->getAlign())); | ||||||
| 1307 | break; | ||||||
| 1308 | case Instruction::GetElementPtr: { | ||||||
| 1309 | // Analyze all of the subscripts of this getelementptr instruction | ||||||
| 1310 | // to determine if we can prove known low zero bits. | ||||||
| 1311 | computeKnownBits(I->getOperand(0), Known, Depth + 1, Q); | ||||||
| 1312 | // Accumulate the constant indices in a separate variable | ||||||
| 1313 | // to minimize the number of calls to computeForAddSub. | ||||||
| 1314 | APInt AccConstIndices(BitWidth, 0, /*IsSigned*/ true); | ||||||
| 1315 | |||||||
| 1316 | gep_type_iterator GTI = gep_type_begin(I); | ||||||
| 1317 | for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i, ++GTI) { | ||||||
| 1318 | // TrailZ can only become smaller, short-circuit if we hit zero. | ||||||
| 1319 | if (Known.isUnknown()) | ||||||
| 1320 | break; | ||||||
| 1321 | |||||||
| 1322 | Value *Index = I->getOperand(i); | ||||||
| 1323 | |||||||
| 1324 | // Handle case when index is zero. | ||||||
| 1325 | Constant *CIndex = dyn_cast<Constant>(Index); | ||||||
| 1326 | if (CIndex && CIndex->isZeroValue()) | ||||||
| 1327 | continue; | ||||||
| 1328 | |||||||
| 1329 | if (StructType *STy = GTI.getStructTypeOrNull()) { | ||||||
| 1330 | // Handle struct member offset arithmetic. | ||||||
| 1331 | |||||||
| 1332 | assert(CIndex &&((void)0) | ||||||
| 1333 | "Access to structure field must be known at compile time")((void)0); | ||||||
| 1334 | |||||||
| 1335 | if (CIndex->getType()->isVectorTy()) | ||||||
| 1336 | Index = CIndex->getSplatValue(); | ||||||
| 1337 | |||||||
| 1338 | unsigned Idx = cast<ConstantInt>(Index)->getZExtValue(); | ||||||
| 1339 | const StructLayout *SL = Q.DL.getStructLayout(STy); | ||||||
| 1340 | uint64_t Offset = SL->getElementOffset(Idx); | ||||||
| 1341 | AccConstIndices += Offset; | ||||||
| 1342 | continue; | ||||||
| 1343 | } | ||||||
| 1344 | |||||||
| 1345 | // Handle array index arithmetic. | ||||||
| 1346 | Type *IndexedTy = GTI.getIndexedType(); | ||||||
| 1347 | if (!IndexedTy->isSized()) { | ||||||
| 1348 | Known.resetAll(); | ||||||
| 1349 | break; | ||||||
| 1350 | } | ||||||
| 1351 | |||||||
| 1352 | unsigned IndexBitWidth = Index->getType()->getScalarSizeInBits(); | ||||||
| 1353 | KnownBits IndexBits(IndexBitWidth); | ||||||
| 1354 | computeKnownBits(Index, IndexBits, Depth + 1, Q); | ||||||
| 1355 | TypeSize IndexTypeSize = Q.DL.getTypeAllocSize(IndexedTy); | ||||||
| 1356 | uint64_t TypeSizeInBytes = IndexTypeSize.getKnownMinSize(); | ||||||
| 1357 | KnownBits ScalingFactor(IndexBitWidth); | ||||||
| 1358 | // Multiply by current sizeof type. | ||||||
| 1359 | // &A[i] == A + i * sizeof(*A[i]). | ||||||
| 1360 | if (IndexTypeSize.isScalable()) { | ||||||
| 1361 | // For scalable types the only thing we know about sizeof is | ||||||
| 1362 | // that this is a multiple of the minimum size. | ||||||
| 1363 | ScalingFactor.Zero.setLowBits(countTrailingZeros(TypeSizeInBytes)); | ||||||
| 1364 | } else if (IndexBits.isConstant()) { | ||||||
| 1365 | APInt IndexConst = IndexBits.getConstant(); | ||||||
| 1366 | APInt ScalingFactor(IndexBitWidth, TypeSizeInBytes); | ||||||
| 1367 | IndexConst *= ScalingFactor; | ||||||
| 1368 | AccConstIndices += IndexConst.sextOrTrunc(BitWidth); | ||||||
| 1369 | continue; | ||||||
| 1370 | } else { | ||||||
| 1371 | ScalingFactor = | ||||||
| 1372 | KnownBits::makeConstant(APInt(IndexBitWidth, TypeSizeInBytes)); | ||||||
| 1373 | } | ||||||
| 1374 | IndexBits = KnownBits::mul(IndexBits, ScalingFactor); | ||||||
| 1375 | |||||||
| 1376 | // If the offsets have a different width from the pointer, according | ||||||
| 1377 | // to the language reference we need to sign-extend or truncate them | ||||||
| 1378 | // to the width of the pointer. | ||||||
| 1379 | IndexBits = IndexBits.sextOrTrunc(BitWidth); | ||||||
| 1380 | |||||||
| 1381 | // Note that inbounds does *not* guarantee nsw for the addition, as only | ||||||
| 1382 | // the offset is signed, while the base address is unsigned. | ||||||
| 1383 | Known = KnownBits::computeForAddSub( | ||||||
| 1384 | /*Add=*/true, /*NSW=*/false, Known, IndexBits); | ||||||
| 1385 | } | ||||||
| 1386 | if (!Known.isUnknown() && !AccConstIndices.isNullValue()) { | ||||||
| 1387 | KnownBits Index = KnownBits::makeConstant(AccConstIndices); | ||||||
| 1388 | Known = KnownBits::computeForAddSub( | ||||||
| 1389 | /*Add=*/true, /*NSW=*/false, Known, Index); | ||||||
| 1390 | } | ||||||
| 1391 | break; | ||||||
| 1392 | } | ||||||
| 1393 | case Instruction::PHI: { | ||||||
| 1394 | const PHINode *P = cast<PHINode>(I); | ||||||
| 1395 | BinaryOperator *BO = nullptr; | ||||||
| 1396 | Value *R = nullptr, *L = nullptr; | ||||||
| 1397 | if (matchSimpleRecurrence(P, BO, R, L)) { | ||||||
| 1398 | // Handle the case of a simple two-predecessor recurrence PHI. | ||||||
| 1399 | // There's a lot more that could theoretically be done here, but | ||||||
| 1400 | // this is sufficient to catch some interesting cases. | ||||||
| 1401 | unsigned Opcode = BO->getOpcode(); | ||||||
| 1402 | |||||||
| 1403 | // If this is a shift recurrence, we know the bits being shifted in. | ||||||
| 1404 | // We can combine that with information about the start value of the | ||||||
| 1405 | // recurrence to conclude facts about the result. | ||||||
| 1406 | if ((Opcode == Instruction::LShr || Opcode == Instruction::AShr || | ||||||
| 1407 | Opcode == Instruction::Shl) && | ||||||
| 1408 | BO->getOperand(0) == I) { | ||||||
| 1409 | |||||||
| 1410 | // We have matched a recurrence of the form: | ||||||
| 1411 | // %iv = [R, %entry], [%iv.next, %backedge] | ||||||
| 1412 | // %iv.next = shift_op %iv, L | ||||||
| 1413 | |||||||
| 1414 | // Recurse with the phi context to avoid concern about whether facts | ||||||
| 1415 | // inferred hold at original context instruction. TODO: It may be | ||||||
| 1416 | // correct to use the original context. IF warranted, explore and | ||||||
| 1417 | // add sufficient tests to cover. | ||||||
| 1418 | Query RecQ = Q; | ||||||
| 1419 | RecQ.CxtI = P; | ||||||
| 1420 | computeKnownBits(R, DemandedElts, Known2, Depth + 1, RecQ); | ||||||
| 1421 | switch (Opcode) { | ||||||
| 1422 | case Instruction::Shl: | ||||||
| 1423 | // A shl recurrence will only increase the tailing zeros | ||||||
| 1424 | Known.Zero.setLowBits(Known2.countMinTrailingZeros()); | ||||||
| 1425 | break; | ||||||
| 1426 | case Instruction::LShr: | ||||||
| 1427 | // A lshr recurrence will preserve the leading zeros of the | ||||||
| 1428 | // start value | ||||||
| 1429 | Known.Zero.setHighBits(Known2.countMinLeadingZeros()); | ||||||
| 1430 | break; | ||||||
| 1431 | case Instruction::AShr: | ||||||
| 1432 | // An ashr recurrence will extend the initial sign bit | ||||||
| 1433 | Known.Zero.setHighBits(Known2.countMinLeadingZeros()); | ||||||
| 1434 | Known.One.setHighBits(Known2.countMinLeadingOnes()); | ||||||
| 1435 | break; | ||||||
| 1436 | }; | ||||||
| 1437 | } | ||||||
| 1438 | |||||||
| 1439 | // Check for operations that have the property that if | ||||||
| 1440 | // both their operands have low zero bits, the result | ||||||
| 1441 | // will have low zero bits. | ||||||
| 1442 | if (Opcode == Instruction::Add || | ||||||
| 1443 | Opcode == Instruction::Sub || | ||||||
| 1444 | Opcode == Instruction::And || | ||||||
| 1445 | Opcode == Instruction::Or || | ||||||
| 1446 | Opcode == Instruction::Mul) { | ||||||
| 1447 | // Change the context instruction to the "edge" that flows into the | ||||||
| 1448 | // phi. This is important because that is where the value is actually | ||||||
| 1449 | // "evaluated" even though it is used later somewhere else. (see also | ||||||
| 1450 | // D69571). | ||||||
| 1451 | Query RecQ = Q; | ||||||
| 1452 | |||||||
| 1453 | unsigned OpNum = P->getOperand(0) == R ? 0 : 1; | ||||||
| 1454 | Instruction *RInst = P->getIncomingBlock(OpNum)->getTerminator(); | ||||||
| 1455 | Instruction *LInst = P->getIncomingBlock(1-OpNum)->getTerminator(); | ||||||
| 1456 | |||||||
| 1457 | // Ok, we have a PHI of the form L op= R. Check for low | ||||||
| 1458 | // zero bits. | ||||||
| 1459 | RecQ.CxtI = RInst; | ||||||
| 1460 | computeKnownBits(R, Known2, Depth + 1, RecQ); | ||||||
| 1461 | |||||||
| 1462 | // We need to take the minimum number of known bits | ||||||
| 1463 | KnownBits Known3(BitWidth); | ||||||
| 1464 | RecQ.CxtI = LInst; | ||||||
| 1465 | computeKnownBits(L, Known3, Depth + 1, RecQ); | ||||||
| 1466 | |||||||
| 1467 | Known.Zero.setLowBits(std::min(Known2.countMinTrailingZeros(), | ||||||
| 1468 | Known3.countMinTrailingZeros())); | ||||||
| 1469 | |||||||
| 1470 | auto *OverflowOp = dyn_cast<OverflowingBinaryOperator>(BO); | ||||||
| 1471 | if (OverflowOp && Q.IIQ.hasNoSignedWrap(OverflowOp)) { | ||||||
| 1472 | // If initial value of recurrence is nonnegative, and we are adding | ||||||
| 1473 | // a nonnegative number with nsw, the result can only be nonnegative | ||||||
| 1474 | // or poison value regardless of the number of times we execute the | ||||||
| 1475 | // add in phi recurrence. If initial value is negative and we are | ||||||
| 1476 | // adding a negative number with nsw, the result can only be | ||||||
| 1477 | // negative or poison value. Similar arguments apply to sub and mul. | ||||||
| 1478 | // | ||||||
| 1479 | // (add non-negative, non-negative) --> non-negative | ||||||
| 1480 | // (add negative, negative) --> negative | ||||||
| 1481 | if (Opcode == Instruction::Add) { | ||||||
| 1482 | if (Known2.isNonNegative() && Known3.isNonNegative()) | ||||||
| 1483 | Known.makeNonNegative(); | ||||||
| 1484 | else if (Known2.isNegative() && Known3.isNegative()) | ||||||
| 1485 | Known.makeNegative(); | ||||||
| 1486 | } | ||||||
| 1487 | |||||||
| 1488 | // (sub nsw non-negative, negative) --> non-negative | ||||||
| 1489 | // (sub nsw negative, non-negative) --> negative | ||||||
| 1490 | else if (Opcode == Instruction::Sub && BO->getOperand(0) == I) { | ||||||
| 1491 | if (Known2.isNonNegative() && Known3.isNegative()) | ||||||
| 1492 | Known.makeNonNegative(); | ||||||
| 1493 | else if (Known2.isNegative() && Known3.isNonNegative()) | ||||||
| 1494 | Known.makeNegative(); | ||||||
| 1495 | } | ||||||
| 1496 | |||||||
| 1497 | // (mul nsw non-negative, non-negative) --> non-negative | ||||||
| 1498 | else if (Opcode == Instruction::Mul && Known2.isNonNegative() && | ||||||
| 1499 | Known3.isNonNegative()) | ||||||
| 1500 | Known.makeNonNegative(); | ||||||
| 1501 | } | ||||||
| 1502 | |||||||
| 1503 | break; | ||||||
| 1504 | } | ||||||
| 1505 | } | ||||||
| 1506 | |||||||
| 1507 | // Unreachable blocks may have zero-operand PHI nodes. | ||||||
| 1508 | if (P->getNumIncomingValues() == 0) | ||||||
| 1509 | break; | ||||||
| 1510 | |||||||
| 1511 | // Otherwise take the unions of the known bit sets of the operands, | ||||||
| 1512 | // taking conservative care to avoid excessive recursion. | ||||||
| 1513 | if (Depth < MaxAnalysisRecursionDepth - 1 && !Known.Zero && !Known.One) { | ||||||
| 1514 | // Skip if every incoming value references to ourself. | ||||||
| 1515 | if (dyn_cast_or_null<UndefValue>(P->hasConstantValue())) | ||||||
| 1516 | break; | ||||||
| 1517 | |||||||
| 1518 | Known.Zero.setAllBits(); | ||||||
| 1519 | Known.One.setAllBits(); | ||||||
| 1520 | for (unsigned u = 0, e = P->getNumIncomingValues(); u < e; ++u) { | ||||||
| 1521 | Value *IncValue = P->getIncomingValue(u); | ||||||
| 1522 | // Skip direct self references. | ||||||
| 1523 | if (IncValue == P) continue; | ||||||
| 1524 | |||||||
| 1525 | // Change the context instruction to the "edge" that flows into the | ||||||
| 1526 | // phi. This is important because that is where the value is actually | ||||||
| 1527 | // "evaluated" even though it is used later somewhere else. (see also | ||||||
| 1528 | // D69571). | ||||||
| 1529 | Query RecQ = Q; | ||||||
| 1530 | RecQ.CxtI = P->getIncomingBlock(u)->getTerminator(); | ||||||
| 1531 | |||||||
| 1532 | Known2 = KnownBits(BitWidth); | ||||||
| 1533 | // Recurse, but cap the recursion to one level, because we don't | ||||||
| 1534 | // want to waste time spinning around in loops. | ||||||
| 1535 | computeKnownBits(IncValue, Known2, MaxAnalysisRecursionDepth - 1, RecQ); | ||||||
| 1536 | Known = KnownBits::commonBits(Known, Known2); | ||||||
| 1537 | // If all bits have been ruled out, there's no need to check | ||||||
| 1538 | // more operands. | ||||||
| 1539 | if (Known.isUnknown()) | ||||||
| 1540 | break; | ||||||
| 1541 | } | ||||||
| 1542 | } | ||||||
| 1543 | break; | ||||||
| 1544 | } | ||||||
| 1545 | case Instruction::Call: | ||||||
| 1546 | case Instruction::Invoke: | ||||||
| 1547 | // If range metadata is attached to this call, set known bits from that, | ||||||
| 1548 | // and then intersect with known bits based on other properties of the | ||||||
| 1549 | // function. | ||||||
| 1550 | if (MDNode *MD = | ||||||
| 1551 | Q.IIQ.getMetadata(cast<Instruction>(I), LLVMContext::MD_range)) | ||||||
| 1552 | computeKnownBitsFromRangeMetadata(*MD, Known); | ||||||
| 1553 | if (const Value *RV = cast<CallBase>(I)->getReturnedArgOperand()) { | ||||||
| 1554 | computeKnownBits(RV, Known2, Depth + 1, Q); | ||||||
| 1555 | Known.Zero |= Known2.Zero; | ||||||
| 1556 | Known.One |= Known2.One; | ||||||
| 1557 | } | ||||||
| 1558 | if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) { | ||||||
| 1559 | switch (II->getIntrinsicID()) { | ||||||
| 1560 | default: break; | ||||||
| 1561 | case Intrinsic::abs: { | ||||||
| 1562 | computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q); | ||||||
| 1563 | bool IntMinIsPoison = match(II->getArgOperand(1), m_One()); | ||||||
| 1564 | Known = Known2.abs(IntMinIsPoison); | ||||||
| 1565 | break; | ||||||
| 1566 | } | ||||||
| 1567 | case Intrinsic::bitreverse: | ||||||
| 1568 | computeKnownBits(I->getOperand(0), DemandedElts, Known2, Depth + 1, Q); | ||||||
| 1569 | Known.Zero |= Known2.Zero.reverseBits(); | ||||||
| 1570 | Known.One |= Known2.One.reverseBits(); | ||||||
| 1571 | break; | ||||||
| 1572 | case Intrinsic::bswap: | ||||||
| 1573 | computeKnownBits(I->getOperand(0), DemandedElts, Known2, Depth + 1, Q); | ||||||
| 1574 | Known.Zero |= Known2.Zero.byteSwap(); | ||||||
| 1575 | Known.One |= Known2.One.byteSwap(); | ||||||
| 1576 | break; | ||||||
| 1577 | case Intrinsic::ctlz: { | ||||||
| 1578 | computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q); | ||||||
| 1579 | // If we have a known 1, its position is our upper bound. | ||||||
| 1580 | unsigned PossibleLZ = Known2.countMaxLeadingZeros(); | ||||||
| 1581 | // If this call is undefined for 0, the result will be less than 2^n. | ||||||
| 1582 | if (II->getArgOperand(1) == ConstantInt::getTrue(II->getContext())) | ||||||
| 1583 | PossibleLZ = std::min(PossibleLZ, BitWidth - 1); | ||||||
| 1584 | unsigned LowBits = Log2_32(PossibleLZ)+1; | ||||||
| 1585 | Known.Zero.setBitsFrom(LowBits); | ||||||
| 1586 | break; | ||||||
| 1587 | } | ||||||
| 1588 | case Intrinsic::cttz: { | ||||||
| 1589 | computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q); | ||||||
| 1590 | // If we have a known 1, its position is our upper bound. | ||||||
| 1591 | unsigned PossibleTZ = Known2.countMaxTrailingZeros(); | ||||||
| 1592 | // If this call is undefined for 0, the result will be less than 2^n. | ||||||
| 1593 | if (II->getArgOperand(1) == ConstantInt::getTrue(II->getContext())) | ||||||
| 1594 | PossibleTZ = std::min(PossibleTZ, BitWidth - 1); | ||||||
| 1595 | unsigned LowBits = Log2_32(PossibleTZ)+1; | ||||||
| 1596 | Known.Zero.setBitsFrom(LowBits); | ||||||
| 1597 | break; | ||||||
| 1598 | } | ||||||
| 1599 | case Intrinsic::ctpop: { | ||||||
| 1600 | computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q); | ||||||
| 1601 | // We can bound the space the count needs. Also, bits known to be zero | ||||||
| 1602 | // can't contribute to the population. | ||||||
| 1603 | unsigned BitsPossiblySet = Known2.countMaxPopulation(); | ||||||
| 1604 | unsigned LowBits = Log2_32(BitsPossiblySet)+1; | ||||||
| 1605 | Known.Zero.setBitsFrom(LowBits); | ||||||
| 1606 | // TODO: we could bound KnownOne using the lower bound on the number | ||||||
| 1607 | // of bits which might be set provided by popcnt KnownOne2. | ||||||
| 1608 | break; | ||||||
| 1609 | } | ||||||
| 1610 | case Intrinsic::fshr: | ||||||
| 1611 | case Intrinsic::fshl: { | ||||||
| 1612 | const APInt *SA; | ||||||
| 1613 | if (!match(I->getOperand(2), m_APInt(SA))) | ||||||
| 1614 | break; | ||||||
| 1615 | |||||||
| 1616 | // Normalize to funnel shift left. | ||||||
| 1617 | uint64_t ShiftAmt = SA->urem(BitWidth); | ||||||
| 1618 | if (II->getIntrinsicID() == Intrinsic::fshr) | ||||||
| 1619 | ShiftAmt = BitWidth - ShiftAmt; | ||||||
| 1620 | |||||||
| 1621 | KnownBits Known3(BitWidth); | ||||||
| 1622 | computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q); | ||||||
| 1623 | computeKnownBits(I->getOperand(1), Known3, Depth + 1, Q); | ||||||
| 1624 | |||||||
| 1625 | Known.Zero = | ||||||
| 1626 | Known2.Zero.shl(ShiftAmt) | Known3.Zero.lshr(BitWidth - ShiftAmt); | ||||||
| 1627 | Known.One = | ||||||
| 1628 | Known2.One.shl(ShiftAmt) | Known3.One.lshr(BitWidth - ShiftAmt); | ||||||
| 1629 | break; | ||||||
| 1630 | } | ||||||
| 1631 | case Intrinsic::uadd_sat: | ||||||
| 1632 | case Intrinsic::usub_sat: { | ||||||
| 1633 | bool IsAdd = II->getIntrinsicID() == Intrinsic::uadd_sat; | ||||||
| 1634 | computeKnownBits(I->getOperand(0), Known, Depth + 1, Q); | ||||||
| 1635 | computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q); | ||||||
| 1636 | |||||||
| 1637 | // Add: Leading ones of either operand are preserved. | ||||||
| 1638 | // Sub: Leading zeros of LHS and leading ones of RHS are preserved | ||||||
| 1639 | // as leading zeros in the result. | ||||||
| 1640 | unsigned LeadingKnown; | ||||||
| 1641 | if (IsAdd) | ||||||
| 1642 | LeadingKnown = std::max(Known.countMinLeadingOnes(), | ||||||
| 1643 | Known2.countMinLeadingOnes()); | ||||||
| 1644 | else | ||||||
| 1645 | LeadingKnown = std::max(Known.countMinLeadingZeros(), | ||||||
| 1646 | Known2.countMinLeadingOnes()); | ||||||
| 1647 | |||||||
| 1648 | Known = KnownBits::computeForAddSub( | ||||||
| 1649 | IsAdd, /* NSW */ false, Known, Known2); | ||||||
| 1650 | |||||||
| 1651 | // We select between the operation result and all-ones/zero | ||||||
| 1652 | // respectively, so we can preserve known ones/zeros. | ||||||
| 1653 | if (IsAdd) { | ||||||
| 1654 | Known.One.setHighBits(LeadingKnown); | ||||||
| 1655 | Known.Zero.clearAllBits(); | ||||||
| 1656 | } else { | ||||||
| 1657 | Known.Zero.setHighBits(LeadingKnown); | ||||||
| 1658 | Known.One.clearAllBits(); | ||||||
| 1659 | } | ||||||
| 1660 | break; | ||||||
| 1661 | } | ||||||
| 1662 | case Intrinsic::umin: | ||||||
| 1663 | computeKnownBits(I->getOperand(0), Known, Depth + 1, Q); | ||||||
| 1664 | computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q); | ||||||
| 1665 | Known = KnownBits::umin(Known, Known2); | ||||||
| 1666 | break; | ||||||
| 1667 | case Intrinsic::umax: | ||||||
| 1668 | computeKnownBits(I->getOperand(0), Known, Depth + 1, Q); | ||||||
| 1669 | computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q); | ||||||
| 1670 | Known = KnownBits::umax(Known, Known2); | ||||||
| 1671 | break; | ||||||
| 1672 | case Intrinsic::smin: | ||||||
| 1673 | computeKnownBits(I->getOperand(0), Known, Depth + 1, Q); | ||||||
| 1674 | computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q); | ||||||
| 1675 | Known = KnownBits::smin(Known, Known2); | ||||||
| 1676 | break; | ||||||
| 1677 | case Intrinsic::smax: | ||||||
| 1678 | computeKnownBits(I->getOperand(0), Known, Depth + 1, Q); | ||||||
| 1679 | computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q); | ||||||
| 1680 | Known = KnownBits::smax(Known, Known2); | ||||||
| 1681 | break; | ||||||
| 1682 | case Intrinsic::x86_sse42_crc32_64_64: | ||||||
| 1683 | Known.Zero.setBitsFrom(32); | ||||||
| 1684 | break; | ||||||
| 1685 | case Intrinsic::riscv_vsetvli: | ||||||
| 1686 | case Intrinsic::riscv_vsetvlimax: | ||||||
| 1687 | // Assume that VL output is positive and would fit in an int32_t. | ||||||
| 1688 | // TODO: VLEN might be capped at 16 bits in a future V spec update. | ||||||
| 1689 | if (BitWidth >= 32) | ||||||
| 1690 | Known.Zero.setBitsFrom(31); | ||||||
| 1691 | break; | ||||||
| 1692 | } | ||||||
| 1693 | } | ||||||
| 1694 | break; | ||||||
| 1695 | case Instruction::ShuffleVector: { | ||||||
| 1696 | auto *Shuf = dyn_cast<ShuffleVectorInst>(I); | ||||||
| 1697 | // FIXME: Do we need to handle ConstantExpr involving shufflevectors? | ||||||
| 1698 | if (!Shuf) { | ||||||
| 1699 | Known.resetAll(); | ||||||
| 1700 | return; | ||||||
| 1701 | } | ||||||
| 1702 | // For undef elements, we don't know anything about the common state of | ||||||
| 1703 | // the shuffle result. | ||||||
| 1704 | APInt DemandedLHS, DemandedRHS; | ||||||
| 1705 | if (!getShuffleDemandedElts(Shuf, DemandedElts, DemandedLHS, DemandedRHS)) { | ||||||
| 1706 | Known.resetAll(); | ||||||
| 1707 | return; | ||||||
| 1708 | } | ||||||
| 1709 | Known.One.setAllBits(); | ||||||
| 1710 | Known.Zero.setAllBits(); | ||||||
| 1711 | if (!!DemandedLHS) { | ||||||
| 1712 | const Value *LHS = Shuf->getOperand(0); | ||||||
| 1713 | computeKnownBits(LHS, DemandedLHS, Known, Depth + 1, Q); | ||||||
| 1714 | // If we don't know any bits, early out. | ||||||
| 1715 | if (Known.isUnknown()) | ||||||
| 1716 | break; | ||||||
| 1717 | } | ||||||
| 1718 | if (!!DemandedRHS) { | ||||||
| 1719 | const Value *RHS = Shuf->getOperand(1); | ||||||
| 1720 | computeKnownBits(RHS, DemandedRHS, Known2, Depth + 1, Q); | ||||||
| 1721 | Known = KnownBits::commonBits(Known, Known2); | ||||||
| 1722 | } | ||||||
| 1723 | break; | ||||||
| 1724 | } | ||||||
| 1725 | case Instruction::InsertElement: { | ||||||
| 1726 | const Value *Vec = I->getOperand(0); | ||||||
| 1727 | const Value *Elt = I->getOperand(1); | ||||||
| 1728 | auto *CIdx = dyn_cast<ConstantInt>(I->getOperand(2)); | ||||||
| 1729 | // Early out if the index is non-constant or out-of-range. | ||||||
| 1730 | unsigned NumElts = DemandedElts.getBitWidth(); | ||||||
| 1731 | if (!CIdx || CIdx->getValue().uge(NumElts)) { | ||||||
| 1732 | Known.resetAll(); | ||||||
| 1733 | return; | ||||||
| 1734 | } | ||||||
| 1735 | Known.One.setAllBits(); | ||||||
| 1736 | Known.Zero.setAllBits(); | ||||||
| 1737 | unsigned EltIdx = CIdx->getZExtValue(); | ||||||
| 1738 | // Do we demand the inserted element? | ||||||
| 1739 | if (DemandedElts[EltIdx]) { | ||||||
| 1740 | computeKnownBits(Elt, Known, Depth + 1, Q); | ||||||
| 1741 | // If we don't know any bits, early out. | ||||||
| 1742 | if (Known.isUnknown()) | ||||||
| 1743 | break; | ||||||
| 1744 | } | ||||||
| 1745 | // We don't need the base vector element that has been inserted. | ||||||
| 1746 | APInt DemandedVecElts = DemandedElts; | ||||||
| 1747 | DemandedVecElts.clearBit(EltIdx); | ||||||
| 1748 | if (!!DemandedVecElts) { | ||||||
| 1749 | computeKnownBits(Vec, DemandedVecElts, Known2, Depth + 1, Q); | ||||||
| 1750 | Known = KnownBits::commonBits(Known, Known2); | ||||||
| 1751 | } | ||||||
| 1752 | break; | ||||||
| 1753 | } | ||||||
| 1754 | case Instruction::ExtractElement: { | ||||||
| 1755 | // Look through extract element. If the index is non-constant or | ||||||
| 1756 | // out-of-range demand all elements, otherwise just the extracted element. | ||||||
| 1757 | const Value *Vec = I->getOperand(0); | ||||||
| 1758 | const Value *Idx = I->getOperand(1); | ||||||
| 1759 | auto *CIdx = dyn_cast<ConstantInt>(Idx); | ||||||
| 1760 | if (isa<ScalableVectorType>(Vec->getType())) { | ||||||
| 1761 | // FIXME: there's probably *something* we can do with scalable vectors | ||||||
| 1762 | Known.resetAll(); | ||||||
| 1763 | break; | ||||||
| 1764 | } | ||||||
| 1765 | unsigned NumElts = cast<FixedVectorType>(Vec->getType())->getNumElements(); | ||||||
| 1766 | APInt DemandedVecElts = APInt::getAllOnesValue(NumElts); | ||||||
| 1767 | if (CIdx && CIdx->getValue().ult(NumElts)) | ||||||
| 1768 | DemandedVecElts = APInt::getOneBitSet(NumElts, CIdx->getZExtValue()); | ||||||
| 1769 | computeKnownBits(Vec, DemandedVecElts, Known, Depth + 1, Q); | ||||||
| 1770 | break; | ||||||
| 1771 | } | ||||||
| 1772 | case Instruction::ExtractValue: | ||||||
| 1773 | if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I->getOperand(0))) { | ||||||
| 1774 | const ExtractValueInst *EVI = cast<ExtractValueInst>(I); | ||||||
| 1775 | if (EVI->getNumIndices() != 1) break; | ||||||
| 1776 | if (EVI->getIndices()[0] == 0) { | ||||||
| 1777 | switch (II->getIntrinsicID()) { | ||||||
| 1778 | default: break; | ||||||
| 1779 | case Intrinsic::uadd_with_overflow: | ||||||
| 1780 | case Intrinsic::sadd_with_overflow: | ||||||
| 1781 | computeKnownBitsAddSub(true, II->getArgOperand(0), | ||||||
| 1782 | II->getArgOperand(1), false, DemandedElts, | ||||||
| 1783 | Known, Known2, Depth, Q); | ||||||
| 1784 | break; | ||||||
| 1785 | case Intrinsic::usub_with_overflow: | ||||||
| 1786 | case Intrinsic::ssub_with_overflow: | ||||||
| 1787 | computeKnownBitsAddSub(false, II->getArgOperand(0), | ||||||
| 1788 | II->getArgOperand(1), false, DemandedElts, | ||||||
| 1789 | Known, Known2, Depth, Q); | ||||||
| 1790 | break; | ||||||
| 1791 | case Intrinsic::umul_with_overflow: | ||||||
| 1792 | case Intrinsic::smul_with_overflow: | ||||||
| 1793 | computeKnownBitsMul(II->getArgOperand(0), II->getArgOperand(1), false, | ||||||
| 1794 | DemandedElts, Known, Known2, Depth, Q); | ||||||
| 1795 | break; | ||||||
| 1796 | } | ||||||
| 1797 | } | ||||||
| 1798 | } | ||||||
| 1799 | break; | ||||||
| 1800 | case Instruction::Freeze: | ||||||
| 1801 | if (isGuaranteedNotToBePoison(I->getOperand(0), Q.AC, Q.CxtI, Q.DT, | ||||||
| 1802 | Depth + 1)) | ||||||
| 1803 | computeKnownBits(I->getOperand(0), Known, Depth + 1, Q); | ||||||
| 1804 | break; | ||||||
| 1805 | } | ||||||
| 1806 | } | ||||||
| 1807 | |||||||
| 1808 | /// Determine which bits of V are known to be either zero or one and return | ||||||
| 1809 | /// them. | ||||||
| 1810 | KnownBits computeKnownBits(const Value *V, const APInt &DemandedElts, | ||||||
| 1811 | unsigned Depth, const Query &Q) { | ||||||
| 1812 | KnownBits Known(getBitWidth(V->getType(), Q.DL)); | ||||||
| 1813 | computeKnownBits(V, DemandedElts, Known, Depth, Q); | ||||||
| 1814 | return Known; | ||||||
| 1815 | } | ||||||
| 1816 | |||||||
| 1817 | /// Determine which bits of V are known to be either zero or one and return | ||||||
| 1818 | /// them. | ||||||
| 1819 | KnownBits computeKnownBits(const Value *V, unsigned Depth, const Query &Q) { | ||||||
| 1820 | KnownBits Known(getBitWidth(V->getType(), Q.DL)); | ||||||
| 1821 | computeKnownBits(V, Known, Depth, Q); | ||||||
| 1822 | return Known; | ||||||
| 1823 | } | ||||||
| 1824 | |||||||
| 1825 | /// Determine which bits of V are known to be either zero or one and return | ||||||
| 1826 | /// them in the Known bit set. | ||||||
| 1827 | /// | ||||||
| 1828 | /// NOTE: we cannot consider 'undef' to be "IsZero" here. The problem is that | ||||||
| 1829 | /// we cannot optimize based on the assumption that it is zero without changing | ||||||
| 1830 | /// it to be an explicit zero. If we don't change it to zero, other code could | ||||||
| 1831 | /// optimized based on the contradictory assumption that it is non-zero. | ||||||
| 1832 | /// Because instcombine aggressively folds operations with undef args anyway, | ||||||
| 1833 | /// this won't lose us code quality. | ||||||
| 1834 | /// | ||||||
| 1835 | /// This function is defined on values with integer type, values with pointer | ||||||
| 1836 | /// type, and vectors of integers. In the case | ||||||
| 1837 | /// where V is a vector, known zero, and known one values are the | ||||||
| 1838 | /// same width as the vector element, and the bit is set only if it is true | ||||||
| 1839 | /// for all of the demanded elements in the vector specified by DemandedElts. | ||||||
| 1840 | void computeKnownBits(const Value *V, const APInt &DemandedElts, | ||||||
| 1841 | KnownBits &Known, unsigned Depth, const Query &Q) { | ||||||
| 1842 | if (!DemandedElts || isa<ScalableVectorType>(V->getType())) { | ||||||
| 1843 | // No demanded elts or V is a scalable vector, better to assume we don't | ||||||
| 1844 | // know anything. | ||||||
| 1845 | Known.resetAll(); | ||||||
| 1846 | return; | ||||||
| 1847 | } | ||||||
| 1848 | |||||||
| 1849 | assert(V && "No Value?")((void)0); | ||||||
| 1850 | assert(Depth <= MaxAnalysisRecursionDepth && "Limit Search Depth")((void)0); | ||||||
| 1851 | |||||||
| 1852 | #ifndef NDEBUG1 | ||||||
| 1853 | Type *Ty = V->getType(); | ||||||
| 1854 | unsigned BitWidth = Known.getBitWidth(); | ||||||
| 1855 | |||||||
| 1856 | assert((Ty->isIntOrIntVectorTy(BitWidth) || Ty->isPtrOrPtrVectorTy()) &&((void)0) | ||||||
| 1857 | "Not integer or pointer type!")((void)0); | ||||||
| 1858 | |||||||
| 1859 | if (auto *FVTy = dyn_cast<FixedVectorType>(Ty)) { | ||||||
| 1860 | assert(((void)0) | ||||||
| 1861 | FVTy->getNumElements() == DemandedElts.getBitWidth() &&((void)0) | ||||||
| 1862 | "DemandedElt width should equal the fixed vector number of elements")((void)0); | ||||||
| 1863 | } else { | ||||||
| 1864 | assert(DemandedElts == APInt(1, 1) &&((void)0) | ||||||
| 1865 | "DemandedElt width should be 1 for scalars")((void)0); | ||||||
| 1866 | } | ||||||
| 1867 | |||||||
| 1868 | Type *ScalarTy = Ty->getScalarType(); | ||||||
| 1869 | if (ScalarTy->isPointerTy()) { | ||||||
| 1870 | assert(BitWidth == Q.DL.getPointerTypeSizeInBits(ScalarTy) &&((void)0) | ||||||
| 1871 | "V and Known should have same BitWidth")((void)0); | ||||||
| 1872 | } else { | ||||||
| 1873 | assert(BitWidth == Q.DL.getTypeSizeInBits(ScalarTy) &&((void)0) | ||||||
| 1874 | "V and Known should have same BitWidth")((void)0); | ||||||
| 1875 | } | ||||||
| 1876 | #endif | ||||||
| 1877 | |||||||
| 1878 | const APInt *C; | ||||||
| 1879 | if (match(V, m_APInt(C))) { | ||||||
| 1880 | // We know all of the bits for a scalar constant or a splat vector constant! | ||||||
| 1881 | Known = KnownBits::makeConstant(*C); | ||||||
| 1882 | return; | ||||||
| 1883 | } | ||||||
| 1884 | // Null and aggregate-zero are all-zeros. | ||||||
| 1885 | if (isa<ConstantPointerNull>(V) || isa<ConstantAggregateZero>(V)) { | ||||||
| 1886 | Known.setAllZero(); | ||||||
| 1887 | return; | ||||||
| 1888 | } | ||||||
| 1889 | // Handle a constant vector by taking the intersection of the known bits of | ||||||
| 1890 | // each element. | ||||||
| 1891 | if (const ConstantDataVector *CDV = dyn_cast<ConstantDataVector>(V)) { | ||||||
| 1892 | // We know that CDV must be a vector of integers. Take the intersection of | ||||||
| 1893 | // each element. | ||||||
| 1894 | Known.Zero.setAllBits(); Known.One.setAllBits(); | ||||||
| 1895 | for (unsigned i = 0, e = CDV->getNumElements(); i != e; ++i) { | ||||||
| 1896 | if (!DemandedElts[i]) | ||||||
| 1897 | continue; | ||||||
| 1898 | APInt Elt = CDV->getElementAsAPInt(i); | ||||||
| 1899 | Known.Zero &= ~Elt; | ||||||
| 1900 | Known.One &= Elt; | ||||||
| 1901 | } | ||||||
| 1902 | return; | ||||||
| 1903 | } | ||||||
| 1904 | |||||||
| 1905 | if (const auto *CV = dyn_cast<ConstantVector>(V)) { | ||||||
| 1906 | // We know that CV must be a vector of integers. Take the intersection of | ||||||
| 1907 | // each element. | ||||||
| 1908 | Known.Zero.setAllBits(); Known.One.setAllBits(); | ||||||
| 1909 | for (unsigned i = 0, e = CV->getNumOperands(); i != e; ++i) { | ||||||
| 1910 | if (!DemandedElts[i]) | ||||||
| 1911 | continue; | ||||||
| 1912 | Constant *Element = CV->getAggregateElement(i); | ||||||
| 1913 | auto *ElementCI = dyn_cast_or_null<ConstantInt>(Element); | ||||||
| 1914 | if (!ElementCI) { | ||||||
| 1915 | Known.resetAll(); | ||||||
| 1916 | return; | ||||||
| 1917 | } | ||||||
| 1918 | const APInt &Elt = ElementCI->getValue(); | ||||||
| 1919 | Known.Zero &= ~Elt; | ||||||
| 1920 | Known.One &= Elt; | ||||||
| 1921 | } | ||||||
| 1922 | return; | ||||||
| 1923 | } | ||||||
| 1924 | |||||||
| 1925 | // Start out not knowing anything. | ||||||
| 1926 | Known.resetAll(); | ||||||
| 1927 | |||||||
| 1928 | // We can't imply anything about undefs. | ||||||
| 1929 | if (isa<UndefValue>(V)) | ||||||
| 1930 | return; | ||||||
| 1931 | |||||||
| 1932 | // There's no point in looking through other users of ConstantData for | ||||||
| 1933 | // assumptions. Confirm that we've handled them all. | ||||||
| 1934 | assert(!isa<ConstantData>(V) && "Unhandled constant data!")((void)0); | ||||||
| 1935 | |||||||
| 1936 | // All recursive calls that increase depth must come after this. | ||||||
| 1937 | if (Depth == MaxAnalysisRecursionDepth) | ||||||
| 1938 | return; | ||||||
| 1939 | |||||||
| 1940 | // A weak GlobalAlias is totally unknown. A non-weak GlobalAlias has | ||||||
| 1941 | // the bits of its aliasee. | ||||||
| 1942 | if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) { | ||||||
| 1943 | if (!GA->isInterposable()) | ||||||
| 1944 | computeKnownBits(GA->getAliasee(), Known, Depth + 1, Q); | ||||||
| 1945 | return; | ||||||
| 1946 | } | ||||||
| 1947 | |||||||
| 1948 | if (const Operator *I = dyn_cast<Operator>(V)) | ||||||
| 1949 | computeKnownBitsFromOperator(I, DemandedElts, Known, Depth, Q); | ||||||
| 1950 | |||||||
| 1951 | // Aligned pointers have trailing zeros - refine Known.Zero set | ||||||
| 1952 | if (isa<PointerType>(V->getType())) { | ||||||
| 1953 | Align Alignment = V->getPointerAlignment(Q.DL); | ||||||
| 1954 | Known.Zero.setLowBits(Log2(Alignment)); | ||||||
| 1955 | } | ||||||
| 1956 | |||||||
| 1957 | // computeKnownBitsFromAssume strictly refines Known. | ||||||
| 1958 | // Therefore, we run them after computeKnownBitsFromOperator. | ||||||
| 1959 | |||||||
| 1960 | // Check whether a nearby assume intrinsic can determine some known bits. | ||||||
| 1961 | computeKnownBitsFromAssume(V, Known, Depth, Q); | ||||||
| 1962 | |||||||
| 1963 | assert((Known.Zero & Known.One) == 0 && "Bits known to be one AND zero?")((void)0); | ||||||
| 1964 | } | ||||||
| 1965 | |||||||
| 1966 | /// Return true if the given value is known to have exactly one | ||||||
| 1967 | /// bit set when defined. For vectors return true if every element is known to | ||||||
| 1968 | /// be a power of two when defined. Supports values with integer or pointer | ||||||
| 1969 | /// types and vectors of integers. | ||||||
| 1970 | bool isKnownToBeAPowerOfTwo(const Value *V, bool OrZero, unsigned Depth, | ||||||
| 1971 | const Query &Q) { | ||||||
| 1972 | assert(Depth <= MaxAnalysisRecursionDepth && "Limit Search Depth")((void)0); | ||||||
| 1973 | |||||||
| 1974 | // Attempt to match against constants. | ||||||
| 1975 | if (OrZero && match(V, m_Power2OrZero())) | ||||||
| 1976 | return true; | ||||||
| 1977 | if (match(V, m_Power2())) | ||||||
| 1978 | return true; | ||||||
| 1979 | |||||||
| 1980 | // 1 << X is clearly a power of two if the one is not shifted off the end. If | ||||||
| 1981 | // it is shifted off the end then the result is undefined. | ||||||
| 1982 | if (match(V, m_Shl(m_One(), m_Value()))) | ||||||
| 1983 | return true; | ||||||
| 1984 | |||||||
| 1985 | // (signmask) >>l X is clearly a power of two if the one is not shifted off | ||||||
| 1986 | // the bottom. If it is shifted off the bottom then the result is undefined. | ||||||
| 1987 | if (match(V, m_LShr(m_SignMask(), m_Value()))) | ||||||
| 1988 | return true; | ||||||
| 1989 | |||||||
| 1990 | // The remaining tests are all recursive, so bail out if we hit the limit. | ||||||
| 1991 | if (Depth++ == MaxAnalysisRecursionDepth) | ||||||
| 1992 | return false; | ||||||
| 1993 | |||||||
| 1994 | Value *X = nullptr, *Y = nullptr; | ||||||
| 1995 | // A shift left or a logical shift right of a power of two is a power of two | ||||||
| 1996 | // or zero. | ||||||
| 1997 | if (OrZero && (match(V, m_Shl(m_Value(X), m_Value())) || | ||||||
| 1998 | match(V, m_LShr(m_Value(X), m_Value())))) | ||||||
| 1999 | return isKnownToBeAPowerOfTwo(X, /*OrZero*/ true, Depth, Q); | ||||||
| 2000 | |||||||
| 2001 | if (const ZExtInst *ZI = dyn_cast<ZExtInst>(V)) | ||||||
| 2002 | return isKnownToBeAPowerOfTwo(ZI->getOperand(0), OrZero, Depth, Q); | ||||||
| 2003 | |||||||
| 2004 | if (const SelectInst *SI = dyn_cast<SelectInst>(V)) | ||||||
| 2005 | return isKnownToBeAPowerOfTwo(SI->getTrueValue(), OrZero, Depth, Q) && | ||||||
| 2006 | isKnownToBeAPowerOfTwo(SI->getFalseValue(), OrZero, Depth, Q); | ||||||
| 2007 | |||||||
| 2008 | // Peek through min/max. | ||||||
| 2009 | if (match(V, m_MaxOrMin(m_Value(X), m_Value(Y)))) { | ||||||
| 2010 | return isKnownToBeAPowerOfTwo(X, OrZero, Depth, Q) && | ||||||
| 2011 | isKnownToBeAPowerOfTwo(Y, OrZero, Depth, Q); | ||||||
| 2012 | } | ||||||
| 2013 | |||||||
| 2014 | if (OrZero && match(V, m_And(m_Value(X), m_Value(Y)))) { | ||||||
| 2015 | // A power of two and'd with anything is a power of two or zero. | ||||||
| 2016 | if (isKnownToBeAPowerOfTwo(X, /*OrZero*/ true, Depth, Q) || | ||||||
| 2017 | isKnownToBeAPowerOfTwo(Y, /*OrZero*/ true, Depth, Q)) | ||||||
| 2018 | return true; | ||||||
| 2019 | // X & (-X) is always a power of two or zero. | ||||||
| 2020 | if (match(X, m_Neg(m_Specific(Y))) || match(Y, m_Neg(m_Specific(X)))) | ||||||
| 2021 | return true; | ||||||
| 2022 | return false; | ||||||
| 2023 | } | ||||||
| 2024 | |||||||
| 2025 | // Adding a power-of-two or zero to the same power-of-two or zero yields | ||||||
| 2026 | // either the original power-of-two, a larger power-of-two or zero. | ||||||
| 2027 | if (match(V, m_Add(m_Value(X), m_Value(Y)))) { | ||||||
| 2028 | const OverflowingBinaryOperator *VOBO = cast<OverflowingBinaryOperator>(V); | ||||||
| 2029 | if (OrZero || Q.IIQ.hasNoUnsignedWrap(VOBO) || | ||||||
| 2030 | Q.IIQ.hasNoSignedWrap(VOBO)) { | ||||||
| 2031 | if (match(X, m_And(m_Specific(Y), m_Value())) || | ||||||
| 2032 | match(X, m_And(m_Value(), m_Specific(Y)))) | ||||||
| 2033 | if (isKnownToBeAPowerOfTwo(Y, OrZero, Depth, Q)) | ||||||
| 2034 | return true; | ||||||
| 2035 | if (match(Y, m_And(m_Specific(X), m_Value())) || | ||||||
| 2036 | match(Y, m_And(m_Value(), m_Specific(X)))) | ||||||
| 2037 | if (isKnownToBeAPowerOfTwo(X, OrZero, Depth, Q)) | ||||||
| 2038 | return true; | ||||||
| 2039 | |||||||
| 2040 | unsigned BitWidth = V->getType()->getScalarSizeInBits(); | ||||||
| 2041 | KnownBits LHSBits(BitWidth); | ||||||
| 2042 | computeKnownBits(X, LHSBits, Depth, Q); | ||||||
| 2043 | |||||||
| 2044 | KnownBits RHSBits(BitWidth); | ||||||
| 2045 | computeKnownBits(Y, RHSBits, Depth, Q); | ||||||
| 2046 | // If i8 V is a power of two or zero: | ||||||
| 2047 | // ZeroBits: 1 1 1 0 1 1 1 1 | ||||||
| 2048 | // ~ZeroBits: 0 0 0 1 0 0 0 0 | ||||||
| 2049 | if ((~(LHSBits.Zero & RHSBits.Zero)).isPowerOf2()) | ||||||
| 2050 | // If OrZero isn't set, we cannot give back a zero result. | ||||||
| 2051 | // Make sure either the LHS or RHS has a bit set. | ||||||
| 2052 | if (OrZero || RHSBits.One.getBoolValue() || LHSBits.One.getBoolValue()) | ||||||
| 2053 | return true; | ||||||
| 2054 | } | ||||||
| 2055 | } | ||||||
| 2056 | |||||||
| 2057 | // An exact divide or right shift can only shift off zero bits, so the result | ||||||
| 2058 | // is a power of two only if the first operand is a power of two and not | ||||||
| 2059 | // copying a sign bit (sdiv int_min, 2). | ||||||
| 2060 | if (match(V, m_Exact(m_LShr(m_Value(), m_Value()))) || | ||||||
| 2061 | match(V, m_Exact(m_UDiv(m_Value(), m_Value())))) { | ||||||
| 2062 | return isKnownToBeAPowerOfTwo(cast<Operator>(V)->getOperand(0), OrZero, | ||||||
| 2063 | Depth, Q); | ||||||
| 2064 | } | ||||||
| 2065 | |||||||
| 2066 | return false; | ||||||
| 2067 | } | ||||||
| 2068 | |||||||
| 2069 | /// Test whether a GEP's result is known to be non-null. | ||||||
| 2070 | /// | ||||||
| 2071 | /// Uses properties inherent in a GEP to try to determine whether it is known | ||||||
| 2072 | /// to be non-null. | ||||||
| 2073 | /// | ||||||
| 2074 | /// Currently this routine does not support vector GEPs. | ||||||
| 2075 | static bool isGEPKnownNonNull(const GEPOperator *GEP, unsigned Depth, | ||||||
| 2076 | const Query &Q) { | ||||||
| 2077 | const Function *F = nullptr; | ||||||
| 2078 | if (const Instruction *I = dyn_cast<Instruction>(GEP)) | ||||||
| 2079 | F = I->getFunction(); | ||||||
| 2080 | |||||||
| 2081 | if (!GEP->isInBounds() || | ||||||
| 2082 | NullPointerIsDefined(F, GEP->getPointerAddressSpace())) | ||||||
| 2083 | return false; | ||||||
| 2084 | |||||||
| 2085 | // FIXME: Support vector-GEPs. | ||||||
| 2086 | assert(GEP->getType()->isPointerTy() && "We only support plain pointer GEP")((void)0); | ||||||
| 2087 | |||||||
| 2088 | // If the base pointer is non-null, we cannot walk to a null address with an | ||||||
| 2089 | // inbounds GEP in address space zero. | ||||||
| 2090 | if (isKnownNonZero(GEP->getPointerOperand(), Depth, Q)) | ||||||
| 2091 | return true; | ||||||
| 2092 | |||||||
| 2093 | // Walk the GEP operands and see if any operand introduces a non-zero offset. | ||||||
| 2094 | // If so, then the GEP cannot produce a null pointer, as doing so would | ||||||
| 2095 | // inherently violate the inbounds contract within address space zero. | ||||||
| 2096 | for (gep_type_iterator GTI = gep_type_begin(GEP), GTE = gep_type_end(GEP); | ||||||
| 2097 | GTI != GTE; ++GTI) { | ||||||
| 2098 | // Struct types are easy -- they must always be indexed by a constant. | ||||||
| 2099 | if (StructType *STy = GTI.getStructTypeOrNull()) { | ||||||
| 2100 | ConstantInt *OpC = cast<ConstantInt>(GTI.getOperand()); | ||||||
| 2101 | unsigned ElementIdx = OpC->getZExtValue(); | ||||||
| 2102 | const StructLayout *SL = Q.DL.getStructLayout(STy); | ||||||
| 2103 | uint64_t ElementOffset = SL->getElementOffset(ElementIdx); | ||||||
| 2104 | if (ElementOffset > 0) | ||||||
| 2105 | return true; | ||||||
| 2106 | continue; | ||||||
| 2107 | } | ||||||
| 2108 | |||||||
| 2109 | // If we have a zero-sized type, the index doesn't matter. Keep looping. | ||||||
| 2110 | if (Q.DL.getTypeAllocSize(GTI.getIndexedType()).getKnownMinSize() == 0) | ||||||
| 2111 | continue; | ||||||
| 2112 | |||||||
| 2113 | // Fast path the constant operand case both for efficiency and so we don't | ||||||
| 2114 | // increment Depth when just zipping down an all-constant GEP. | ||||||
| 2115 | if (ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand())) { | ||||||
| 2116 | if (!OpC->isZero()) | ||||||
| 2117 | return true; | ||||||
| 2118 | continue; | ||||||
| 2119 | } | ||||||
| 2120 | |||||||
| 2121 | // We post-increment Depth here because while isKnownNonZero increments it | ||||||
| 2122 | // as well, when we pop back up that increment won't persist. We don't want | ||||||
| 2123 | // to recurse 10k times just because we have 10k GEP operands. We don't | ||||||
| 2124 | // bail completely out because we want to handle constant GEPs regardless | ||||||
| 2125 | // of depth. | ||||||
| 2126 | if (Depth++ >= MaxAnalysisRecursionDepth) | ||||||
| 2127 | continue; | ||||||
| 2128 | |||||||
| 2129 | if (isKnownNonZero(GTI.getOperand(), Depth, Q)) | ||||||
| 2130 | return true; | ||||||
| 2131 | } | ||||||
| 2132 | |||||||
| 2133 | return false; | ||||||
| 2134 | } | ||||||
| 2135 | |||||||
| 2136 | static bool isKnownNonNullFromDominatingCondition(const Value *V, | ||||||
| 2137 | const Instruction *CtxI, | ||||||
| 2138 | const DominatorTree *DT) { | ||||||
| 2139 | if (isa<Constant>(V)) | ||||||
| 2140 | return false; | ||||||
| 2141 | |||||||
| 2142 | if (!CtxI || !DT) | ||||||
| 2143 | return false; | ||||||
| 2144 | |||||||
| 2145 | unsigned NumUsesExplored = 0; | ||||||
| 2146 | for (auto *U : V->users()) { | ||||||
| 2147 | // Avoid massive lists | ||||||
| 2148 | if (NumUsesExplored >= DomConditionsMaxUses) | ||||||
| 2149 | break; | ||||||
| 2150 | NumUsesExplored++; | ||||||
| 2151 | |||||||
| 2152 | // If the value is used as an argument to a call or invoke, then argument | ||||||
| 2153 | // attributes may provide an answer about null-ness. | ||||||
| 2154 | if (const auto *CB = dyn_cast<CallBase>(U)) | ||||||
| 2155 | if (auto *CalledFunc = CB->getCalledFunction()) | ||||||
| 2156 | for (const Argument &Arg : CalledFunc->args()) | ||||||
| 2157 | if (CB->getArgOperand(Arg.getArgNo()) == V && | ||||||
| 2158 | Arg.hasNonNullAttr(/* AllowUndefOrPoison */ false) && | ||||||
| 2159 | DT->dominates(CB, CtxI)) | ||||||
| 2160 | return true; | ||||||
| 2161 | |||||||
| 2162 | // If the value is used as a load/store, then the pointer must be non null. | ||||||
| 2163 | if (V == getLoadStorePointerOperand(U)) { | ||||||
| 2164 | const Instruction *I = cast<Instruction>(U); | ||||||
| 2165 | if (!NullPointerIsDefined(I->getFunction(), | ||||||
| 2166 | V->getType()->getPointerAddressSpace()) && | ||||||
| 2167 | DT->dominates(I, CtxI)) | ||||||
| 2168 | return true; | ||||||
| 2169 | } | ||||||
| 2170 | |||||||
| 2171 | // Consider only compare instructions uniquely controlling a branch | ||||||
| 2172 | Value *RHS; | ||||||
| 2173 | CmpInst::Predicate Pred; | ||||||
| 2174 | if (!match(U, m_c_ICmp(Pred, m_Specific(V), m_Value(RHS)))) | ||||||
| 2175 | continue; | ||||||
| 2176 | |||||||
| 2177 | bool NonNullIfTrue; | ||||||
| 2178 | if (cmpExcludesZero(Pred, RHS)) | ||||||
| 2179 | NonNullIfTrue = true; | ||||||
| 2180 | else if (cmpExcludesZero(CmpInst::getInversePredicate(Pred), RHS)) | ||||||
| 2181 | NonNullIfTrue = false; | ||||||
| 2182 | else | ||||||
| 2183 | continue; | ||||||
| 2184 | |||||||
| 2185 | SmallVector<const User *, 4> WorkList; | ||||||
| 2186 | SmallPtrSet<const User *, 4> Visited; | ||||||
| 2187 | for (auto *CmpU : U->users()) { | ||||||
| 2188 | assert(WorkList.empty() && "Should be!")((void)0); | ||||||
| 2189 | if (Visited.insert(CmpU).second) | ||||||
| 2190 | WorkList.push_back(CmpU); | ||||||
| 2191 | |||||||
| 2192 | while (!WorkList.empty()) { | ||||||
| 2193 | auto *Curr = WorkList.pop_back_val(); | ||||||
| 2194 | |||||||
| 2195 | // If a user is an AND, add all its users to the work list. We only | ||||||
| 2196 | // propagate "pred != null" condition through AND because it is only | ||||||
| 2197 | // correct to assume that all conditions of AND are met in true branch. | ||||||
| 2198 | // TODO: Support similar logic of OR and EQ predicate? | ||||||
| 2199 | if (NonNullIfTrue) | ||||||
| 2200 | if (match(Curr, m_LogicalAnd(m_Value(), m_Value()))) { | ||||||
| 2201 | for (auto *CurrU : Curr->users()) | ||||||
| 2202 | if (Visited.insert(CurrU).second) | ||||||
| 2203 | WorkList.push_back(CurrU); | ||||||
| 2204 | continue; | ||||||
| 2205 | } | ||||||
| 2206 | |||||||
| 2207 | if (const BranchInst *BI = dyn_cast<BranchInst>(Curr)) { | ||||||
| 2208 | assert(BI->isConditional() && "uses a comparison!")((void)0); | ||||||
| 2209 | |||||||
| 2210 | BasicBlock *NonNullSuccessor = | ||||||
| 2211 | BI->getSuccessor(NonNullIfTrue ? 0 : 1); | ||||||
| 2212 | BasicBlockEdge Edge(BI->getParent(), NonNullSuccessor); | ||||||
| 2213 | if (Edge.isSingleEdge() && DT->dominates(Edge, CtxI->getParent())) | ||||||
| 2214 | return true; | ||||||
| 2215 | } else if (NonNullIfTrue && isGuard(Curr) && | ||||||
| 2216 | DT->dominates(cast<Instruction>(Curr), CtxI)) { | ||||||
| 2217 | return true; | ||||||
| 2218 | } | ||||||
| 2219 | } | ||||||
| 2220 | } | ||||||
| 2221 | } | ||||||
| 2222 | |||||||
| 2223 | return false; | ||||||
| 2224 | } | ||||||
| 2225 | |||||||
| 2226 | /// Does the 'Range' metadata (which must be a valid MD_range operand list) | ||||||
| 2227 | /// ensure that the value it's attached to is never Value? 'RangeType' is | ||||||
| 2228 | /// is the type of the value described by the range. | ||||||
| 2229 | static bool rangeMetadataExcludesValue(const MDNode* Ranges, const APInt& Value) { | ||||||
| 2230 | const unsigned NumRanges = Ranges->getNumOperands() / 2; | ||||||
| 2231 | assert(NumRanges >= 1)((void)0); | ||||||
| 2232 | for (unsigned i = 0; i < NumRanges; ++i) { | ||||||
| 2233 | ConstantInt *Lower = | ||||||
| 2234 | mdconst::extract<ConstantInt>(Ranges->getOperand(2 * i + 0)); | ||||||
| 2235 | ConstantInt *Upper = | ||||||
| 2236 | mdconst::extract<ConstantInt>(Ranges->getOperand(2 * i + 1)); | ||||||
| 2237 | ConstantRange Range(Lower->getValue(), Upper->getValue()); | ||||||
| 2238 | if (Range.contains(Value)) | ||||||
| 2239 | return false; | ||||||
| 2240 | } | ||||||
| 2241 | return true; | ||||||
| 2242 | } | ||||||
| 2243 | |||||||
| 2244 | /// Try to detect a recurrence that monotonically increases/decreases from a | ||||||
| 2245 | /// non-zero starting value. These are common as induction variables. | ||||||
| 2246 | static bool isNonZeroRecurrence(const PHINode *PN) { | ||||||
| 2247 | BinaryOperator *BO = nullptr; | ||||||
| 2248 | Value *Start = nullptr, *Step = nullptr; | ||||||
| 2249 | const APInt *StartC, *StepC; | ||||||
| 2250 | if (!matchSimpleRecurrence(PN, BO, Start, Step) || | ||||||
| 2251 | !match(Start, m_APInt(StartC)) || StartC->isNullValue()) | ||||||
| 2252 | return false; | ||||||
| 2253 | |||||||
| 2254 | switch (BO->getOpcode()) { | ||||||
| 2255 | case Instruction::Add: | ||||||
| 2256 | // Starting from non-zero and stepping away from zero can never wrap back | ||||||
| 2257 | // to zero. | ||||||
| 2258 | return BO->hasNoUnsignedWrap() || | ||||||
| 2259 | (BO->hasNoSignedWrap() && match(Step, m_APInt(StepC)) && | ||||||
| 2260 | StartC->isNegative() == StepC->isNegative()); | ||||||
| 2261 | case Instruction::Mul: | ||||||
| 2262 | return (BO->hasNoUnsignedWrap() || BO->hasNoSignedWrap()) && | ||||||
| 2263 | match(Step, m_APInt(StepC)) && !StepC->isNullValue(); | ||||||
| 2264 | case Instruction::Shl: | ||||||
| 2265 | return BO->hasNoUnsignedWrap() || BO->hasNoSignedWrap(); | ||||||
| 2266 | case Instruction::AShr: | ||||||
| 2267 | case Instruction::LShr: | ||||||
| 2268 | return BO->isExact(); | ||||||
| 2269 | default: | ||||||
| 2270 | return false; | ||||||
| 2271 | } | ||||||
| 2272 | } | ||||||
| 2273 | |||||||
| 2274 | /// Return true if the given value is known to be non-zero when defined. For | ||||||
| 2275 | /// vectors, return true if every demanded element is known to be non-zero when | ||||||
| 2276 | /// defined. For pointers, if the context instruction and dominator tree are | ||||||
| 2277 | /// specified, perform context-sensitive analysis and return true if the | ||||||
| 2278 | /// pointer couldn't possibly be null at the specified instruction. | ||||||
| 2279 | /// Supports values with integer or pointer type and vectors of integers. | ||||||
| 2280 | bool isKnownNonZero(const Value *V, const APInt &DemandedElts, unsigned Depth, | ||||||
| 2281 | const Query &Q) { | ||||||
| 2282 | // FIXME: We currently have no way to represent the DemandedElts of a scalable | ||||||
| 2283 | // vector | ||||||
| 2284 | if (isa<ScalableVectorType>(V->getType())) | ||||||
| 2285 | return false; | ||||||
| 2286 | |||||||
| 2287 | if (auto *C = dyn_cast<Constant>(V)) { | ||||||
| 2288 | if (C->isNullValue()) | ||||||
| 2289 | return false; | ||||||
| 2290 | if (isa<ConstantInt>(C)) | ||||||
| 2291 | // Must be non-zero due to null test above. | ||||||
| 2292 | return true; | ||||||
| 2293 | |||||||
| 2294 | if (auto *CE = dyn_cast<ConstantExpr>(C)) { | ||||||
| 2295 | // See the comment for IntToPtr/PtrToInt instructions below. | ||||||
| 2296 | if (CE->getOpcode() == Instruction::IntToPtr || | ||||||
| 2297 | CE->getOpcode() == Instruction::PtrToInt) | ||||||
| 2298 | if (Q.DL.getTypeSizeInBits(CE->getOperand(0)->getType()) | ||||||
| 2299 | .getFixedSize() <= | ||||||
| 2300 | Q.DL.getTypeSizeInBits(CE->getType()).getFixedSize()) | ||||||
| 2301 | return isKnownNonZero(CE->getOperand(0), Depth, Q); | ||||||
| 2302 | } | ||||||
| 2303 | |||||||
| 2304 | // For constant vectors, check that all elements are undefined or known | ||||||
| 2305 | // non-zero to determine that the whole vector is known non-zero. | ||||||
| 2306 | if (auto *VecTy = dyn_cast<FixedVectorType>(C->getType())) { | ||||||
| 2307 | for (unsigned i = 0, e = VecTy->getNumElements(); i != e; ++i) { | ||||||
| 2308 | if (!DemandedElts[i]) | ||||||
| 2309 | continue; | ||||||
| 2310 | Constant *Elt = C->getAggregateElement(i); | ||||||
| 2311 | if (!Elt || Elt->isNullValue()) | ||||||
| 2312 | return false; | ||||||
| 2313 | if (!isa<UndefValue>(Elt) && !isa<ConstantInt>(Elt)) | ||||||
| 2314 | return false; | ||||||
| 2315 | } | ||||||
| 2316 | return true; | ||||||
| 2317 | } | ||||||
| 2318 | |||||||
| 2319 | // A global variable in address space 0 is non null unless extern weak | ||||||
| 2320 | // or an absolute symbol reference. Other address spaces may have null as a | ||||||
| 2321 | // valid address for a global, so we can't assume anything. | ||||||
| 2322 | if (const GlobalValue *GV = dyn_cast<GlobalValue>(V)) { | ||||||
| 2323 | if (!GV->isAbsoluteSymbolRef() && !GV->hasExternalWeakLinkage() && | ||||||
| 2324 | GV->getType()->getAddressSpace() == 0) | ||||||
| 2325 | return true; | ||||||
| 2326 | } else | ||||||
| 2327 | return false; | ||||||
| 2328 | } | ||||||
| 2329 | |||||||
| 2330 | if (auto *I = dyn_cast<Instruction>(V)) { | ||||||
| 2331 | if (MDNode *Ranges = Q.IIQ.getMetadata(I, LLVMContext::MD_range)) { | ||||||
| 2332 | // If the possible ranges don't contain zero, then the value is | ||||||
| 2333 | // definitely non-zero. | ||||||
| 2334 | if (auto *Ty = dyn_cast<IntegerType>(V->getType())) { | ||||||
| 2335 | const APInt ZeroValue(Ty->getBitWidth(), 0); | ||||||
| 2336 | if (rangeMetadataExcludesValue(Ranges, ZeroValue)) | ||||||
| 2337 | return true; | ||||||
| 2338 | } | ||||||
| 2339 | } | ||||||
| 2340 | } | ||||||
| 2341 | |||||||
| 2342 | if (isKnownNonZeroFromAssume(V, Q)) | ||||||
| 2343 | return true; | ||||||
| 2344 | |||||||
| 2345 | // Some of the tests below are recursive, so bail out if we hit the limit. | ||||||
| 2346 | if (Depth++ >= MaxAnalysisRecursionDepth) | ||||||
| 2347 | return false; | ||||||
| 2348 | |||||||
| 2349 | // Check for pointer simplifications. | ||||||
| 2350 | |||||||
| 2351 | if (PointerType *PtrTy = dyn_cast<PointerType>(V->getType())) { | ||||||
| 2352 | // Alloca never returns null, malloc might. | ||||||
| 2353 | if (isa<AllocaInst>(V) && Q.DL.getAllocaAddrSpace() == 0) | ||||||
| 2354 | return true; | ||||||
| 2355 | |||||||
| 2356 | // A byval, inalloca may not be null in a non-default addres space. A | ||||||
| 2357 | // nonnull argument is assumed never 0. | ||||||
| 2358 | if (const Argument *A = dyn_cast<Argument>(V)) { | ||||||
| 2359 | if (((A->hasPassPointeeByValueCopyAttr() && | ||||||
| 2360 | !NullPointerIsDefined(A->getParent(), PtrTy->getAddressSpace())) || | ||||||
| 2361 | A->hasNonNullAttr())) | ||||||
| 2362 | return true; | ||||||
| 2363 | } | ||||||
| 2364 | |||||||
| 2365 | // A Load tagged with nonnull metadata is never null. | ||||||
| 2366 | if (const LoadInst *LI = dyn_cast<LoadInst>(V)) | ||||||
| 2367 | if (Q.IIQ.getMetadata(LI, LLVMContext::MD_nonnull)) | ||||||
| 2368 | return true; | ||||||
| 2369 | |||||||
| 2370 | if (const auto *Call = dyn_cast<CallBase>(V)) { | ||||||
| 2371 | if (Call->isReturnNonNull()) | ||||||
| 2372 | return true; | ||||||
| 2373 | if (const auto *RP = getArgumentAliasingToReturnedPointer(Call, true)) | ||||||
| 2374 | return isKnownNonZero(RP, Depth, Q); | ||||||
| 2375 | } | ||||||
| 2376 | } | ||||||
| 2377 | |||||||
| 2378 | if (isKnownNonNullFromDominatingCondition(V, Q.CxtI, Q.DT)) | ||||||
| 2379 | return true; | ||||||
| 2380 | |||||||
| 2381 | // Check for recursive pointer simplifications. | ||||||
| 2382 | if (V->getType()->isPointerTy()) { | ||||||
| 2383 | // Look through bitcast operations, GEPs, and int2ptr instructions as they | ||||||
| 2384 | // do not alter the value, or at least not the nullness property of the | ||||||
| 2385 | // value, e.g., int2ptr is allowed to zero/sign extend the value. | ||||||
| 2386 | // | ||||||
| 2387 | // Note that we have to take special care to avoid looking through | ||||||
| 2388 | // truncating casts, e.g., int2ptr/ptr2int with appropriate sizes, as well | ||||||
| 2389 | // as casts that can alter the value, e.g., AddrSpaceCasts. | ||||||
| 2390 | if (const GEPOperator *GEP = dyn_cast<GEPOperator>(V)) | ||||||
| 2391 | return isGEPKnownNonNull(GEP, Depth, Q); | ||||||
| 2392 | |||||||
| 2393 | if (auto *BCO = dyn_cast<BitCastOperator>(V)) | ||||||
| 2394 | return isKnownNonZero(BCO->getOperand(0), Depth, Q); | ||||||
| 2395 | |||||||
| 2396 | if (auto *I2P = dyn_cast<IntToPtrInst>(V)) | ||||||
| 2397 | if (Q.DL.getTypeSizeInBits(I2P->getSrcTy()).getFixedSize() <= | ||||||
| 2398 | Q.DL.getTypeSizeInBits(I2P->getDestTy()).getFixedSize()) | ||||||
| 2399 | return isKnownNonZero(I2P->getOperand(0), Depth, Q); | ||||||
| 2400 | } | ||||||
| 2401 | |||||||
| 2402 | // Similar to int2ptr above, we can look through ptr2int here if the cast | ||||||
| 2403 | // is a no-op or an extend and not a truncate. | ||||||
| 2404 | if (auto *P2I = dyn_cast<PtrToIntInst>(V)) | ||||||
| 2405 | if (Q.DL.getTypeSizeInBits(P2I->getSrcTy()).getFixedSize() <= | ||||||
| 2406 | Q.DL.getTypeSizeInBits(P2I->getDestTy()).getFixedSize()) | ||||||
| 2407 | return isKnownNonZero(P2I->getOperand(0), Depth, Q); | ||||||
| 2408 | |||||||
| 2409 | unsigned BitWidth = getBitWidth(V->getType()->getScalarType(), Q.DL); | ||||||
| 2410 | |||||||
| 2411 | // X | Y != 0 if X != 0 or Y != 0. | ||||||
| 2412 | Value *X = nullptr, *Y = nullptr; | ||||||
| 2413 | if (match(V, m_Or(m_Value(X), m_Value(Y)))) | ||||||
| 2414 | return isKnownNonZero(X, DemandedElts, Depth, Q) || | ||||||
| 2415 | isKnownNonZero(Y, DemandedElts, Depth, Q); | ||||||
| 2416 | |||||||
| 2417 | // ext X != 0 if X != 0. | ||||||
| 2418 | if (isa<SExtInst>(V) || isa<ZExtInst>(V)) | ||||||
| 2419 | return isKnownNonZero(cast<Instruction>(V)->getOperand(0), Depth, Q); | ||||||
| 2420 | |||||||
| 2421 | // shl X, Y != 0 if X is odd. Note that the value of the shift is undefined | ||||||
| 2422 | // if the lowest bit is shifted off the end. | ||||||
| 2423 | if (match(V, m_Shl(m_Value(X), m_Value(Y)))) { | ||||||
| 2424 | // shl nuw can't remove any non-zero bits. | ||||||
| 2425 | const OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V); | ||||||
| 2426 | if (Q.IIQ.hasNoUnsignedWrap(BO)) | ||||||
| 2427 | return isKnownNonZero(X, Depth, Q); | ||||||
| 2428 | |||||||
| 2429 | KnownBits Known(BitWidth); | ||||||
| 2430 | computeKnownBits(X, DemandedElts, Known, Depth, Q); | ||||||
| 2431 | if (Known.One[0]) | ||||||
| 2432 | return true; | ||||||
| 2433 | } | ||||||
| 2434 | // shr X, Y != 0 if X is negative. Note that the value of the shift is not | ||||||
| 2435 | // defined if the sign bit is shifted off the end. | ||||||
| 2436 | else if (match(V, m_Shr(m_Value(X), m_Value(Y)))) { | ||||||
| 2437 | // shr exact can only shift out zero bits. | ||||||
| 2438 | const PossiblyExactOperator *BO = cast<PossiblyExactOperator>(V); | ||||||
| 2439 | if (BO->isExact()) | ||||||
| 2440 | return isKnownNonZero(X, Depth, Q); | ||||||
| 2441 | |||||||
| 2442 | KnownBits Known = computeKnownBits(X, DemandedElts, Depth, Q); | ||||||
| 2443 | if (Known.isNegative()) | ||||||
| 2444 | return true; | ||||||
| 2445 | |||||||
| 2446 | // If the shifter operand is a constant, and all of the bits shifted | ||||||
| 2447 | // out are known to be zero, and X is known non-zero then at least one | ||||||
| 2448 | // non-zero bit must remain. | ||||||
| 2449 | if (ConstantInt *Shift = dyn_cast<ConstantInt>(Y)) { | ||||||
| 2450 | auto ShiftVal = Shift->getLimitedValue(BitWidth - 1); | ||||||
| 2451 | // Is there a known one in the portion not shifted out? | ||||||
| 2452 | if (Known.countMaxLeadingZeros() < BitWidth - ShiftVal) | ||||||
| 2453 | return true; | ||||||
| 2454 | // Are all the bits to be shifted out known zero? | ||||||
| 2455 | if (Known.countMinTrailingZeros() >= ShiftVal) | ||||||
| 2456 | return isKnownNonZero(X, DemandedElts, Depth, Q); | ||||||
| 2457 | } | ||||||
| 2458 | } | ||||||
| 2459 | // div exact can only produce a zero if the dividend is zero. | ||||||
| 2460 | else if (match(V, m_Exact(m_IDiv(m_Value(X), m_Value())))) { | ||||||
| 2461 | return isKnownNonZero(X, DemandedElts, Depth, Q); | ||||||
| 2462 | } | ||||||
| 2463 | // X + Y. | ||||||
| 2464 | else if (match(V, m_Add(m_Value(X), m_Value(Y)))) { | ||||||
| 2465 | KnownBits XKnown = computeKnownBits(X, DemandedElts, Depth, Q); | ||||||
| 2466 | KnownBits YKnown = computeKnownBits(Y, DemandedElts, Depth, Q); | ||||||
| 2467 | |||||||
| 2468 | // If X and Y are both non-negative (as signed values) then their sum is not | ||||||
| 2469 | // zero unless both X and Y are zero. | ||||||
| 2470 | if (XKnown.isNonNegative() && YKnown.isNonNegative()) | ||||||
| 2471 | if (isKnownNonZero(X, DemandedElts, Depth, Q) || | ||||||
| 2472 | isKnownNonZero(Y, DemandedElts, Depth, Q)) | ||||||
| 2473 | return true; | ||||||
| 2474 | |||||||
| 2475 | // If X and Y are both negative (as signed values) then their sum is not | ||||||
| 2476 | // zero unless both X and Y equal INT_MIN. | ||||||
| 2477 | if (XKnown.isNegative() && YKnown.isNegative()) { | ||||||
| 2478 | APInt Mask = APInt::getSignedMaxValue(BitWidth); | ||||||
| 2479 | // The sign bit of X is set. If some other bit is set then X is not equal | ||||||
| 2480 | // to INT_MIN. | ||||||
| 2481 | if (XKnown.One.intersects(Mask)) | ||||||
| 2482 | return true; | ||||||
| 2483 | // The sign bit of Y is set. If some other bit is set then Y is not equal | ||||||
| 2484 | // to INT_MIN. | ||||||
| 2485 | if (YKnown.One.intersects(Mask)) | ||||||
| 2486 | return true; | ||||||
| 2487 | } | ||||||
| 2488 | |||||||
| 2489 | // The sum of a non-negative number and a power of two is not zero. | ||||||
| 2490 | if (XKnown.isNonNegative() && | ||||||
| 2491 | isKnownToBeAPowerOfTwo(Y, /*OrZero*/ false, Depth, Q)) | ||||||
| 2492 | return true; | ||||||
| 2493 | if (YKnown.isNonNegative() && | ||||||
| 2494 | isKnownToBeAPowerOfTwo(X, /*OrZero*/ false, Depth, Q)) | ||||||
| 2495 | return true; | ||||||
| 2496 | } | ||||||
| 2497 | // X * Y. | ||||||
| 2498 | else if (match(V, m_Mul(m_Value(X), m_Value(Y)))) { | ||||||
| 2499 | const OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V); | ||||||
| 2500 | // If X and Y are non-zero then so is X * Y as long as the multiplication | ||||||
| 2501 | // does not overflow. | ||||||
| 2502 | if ((Q.IIQ.hasNoSignedWrap(BO) || Q.IIQ.hasNoUnsignedWrap(BO)) && | ||||||
| 2503 | isKnownNonZero(X, DemandedElts, Depth, Q) && | ||||||
| 2504 | isKnownNonZero(Y, DemandedElts, Depth, Q)) | ||||||
| 2505 | return true; | ||||||
| 2506 | } | ||||||
| 2507 | // (C ? X : Y) != 0 if X != 0 and Y != 0. | ||||||
| 2508 | else if (const SelectInst *SI = dyn_cast<SelectInst>(V)) { | ||||||
| 2509 | if (isKnownNonZero(SI->getTrueValue(), DemandedElts, Depth, Q) && | ||||||
| 2510 | isKnownNonZero(SI->getFalseValue(), DemandedElts, Depth, Q)) | ||||||
| 2511 | return true; | ||||||
| 2512 | } | ||||||
| 2513 | // PHI | ||||||
| 2514 | else if (const PHINode *PN = dyn_cast<PHINode>(V)) { | ||||||
| 2515 | if (Q.IIQ.UseInstrInfo && isNonZeroRecurrence(PN)) | ||||||
| 2516 | return true; | ||||||
| 2517 | |||||||
| 2518 | // Check if all incoming values are non-zero using recursion. | ||||||
| 2519 | Query RecQ = Q; | ||||||
| 2520 | unsigned NewDepth = std::max(Depth, MaxAnalysisRecursionDepth - 1); | ||||||
| 2521 | return llvm::all_of(PN->operands(), [&](const Use &U) { | ||||||
| 2522 | if (U.get() == PN) | ||||||
| 2523 | return true; | ||||||
| 2524 | RecQ.CxtI = PN->getIncomingBlock(U)->getTerminator(); | ||||||
| 2525 | return isKnownNonZero(U.get(), DemandedElts, NewDepth, RecQ); | ||||||
| 2526 | }); | ||||||
| 2527 | } | ||||||
| 2528 | // ExtractElement | ||||||
| 2529 | else if (const auto *EEI = dyn_cast<ExtractElementInst>(V)) { | ||||||
| 2530 | const Value *Vec = EEI->getVectorOperand(); | ||||||
| 2531 | const Value *Idx = EEI->getIndexOperand(); | ||||||
| 2532 | auto *CIdx = dyn_cast<ConstantInt>(Idx); | ||||||
| 2533 | if (auto *VecTy = dyn_cast<FixedVectorType>(Vec->getType())) { | ||||||
| 2534 | unsigned NumElts = VecTy->getNumElements(); | ||||||
| 2535 | APInt DemandedVecElts = APInt::getAllOnesValue(NumElts); | ||||||
| 2536 | if (CIdx && CIdx->getValue().ult(NumElts)) | ||||||
| 2537 | DemandedVecElts = APInt::getOneBitSet(NumElts, CIdx->getZExtValue()); | ||||||
| 2538 | return isKnownNonZero(Vec, DemandedVecElts, Depth, Q); | ||||||
| 2539 | } | ||||||
| 2540 | } | ||||||
| 2541 | // Freeze | ||||||
| 2542 | else if (const FreezeInst *FI = dyn_cast<FreezeInst>(V)) { | ||||||
| 2543 | auto *Op = FI->getOperand(0); | ||||||
| 2544 | if (isKnownNonZero(Op, Depth, Q) && | ||||||
| 2545 | isGuaranteedNotToBePoison(Op, Q.AC, Q.CxtI, Q.DT, Depth)) | ||||||
| 2546 | return true; | ||||||
| 2547 | } | ||||||
| 2548 | |||||||
| 2549 | KnownBits Known(BitWidth); | ||||||
| 2550 | computeKnownBits(V, DemandedElts, Known, Depth, Q); | ||||||
| 2551 | return Known.One != 0; | ||||||
| 2552 | } | ||||||
| 2553 | |||||||
| 2554 | bool isKnownNonZero(const Value* V, unsigned Depth, const Query& Q) { | ||||||
| 2555 | // FIXME: We currently have no way to represent the DemandedElts of a scalable | ||||||
| 2556 | // vector | ||||||
| 2557 | if (isa<ScalableVectorType>(V->getType())) | ||||||
| 2558 | return false; | ||||||
| 2559 | |||||||
| 2560 | auto *FVTy = dyn_cast<FixedVectorType>(V->getType()); | ||||||
| 2561 | APInt DemandedElts = | ||||||
| 2562 | FVTy ? APInt::getAllOnesValue(FVTy->getNumElements()) : APInt(1, 1); | ||||||
| 2563 | return isKnownNonZero(V, DemandedElts, Depth, Q); | ||||||
| 2564 | } | ||||||
| 2565 | |||||||
| 2566 | /// If the pair of operators are the same invertible function, return the | ||||||
| 2567 | /// the operands of the function corresponding to each input. Otherwise, | ||||||
| 2568 | /// return None. An invertible function is one that is 1-to-1 and maps | ||||||
| 2569 | /// every input value to exactly one output value. This is equivalent to | ||||||
| 2570 | /// saying that Op1 and Op2 are equal exactly when the specified pair of | ||||||
| 2571 | /// operands are equal, (except that Op1 and Op2 may be poison more often.) | ||||||
| 2572 | static Optional<std::pair<Value*, Value*>> | ||||||
| 2573 | getInvertibleOperands(const Operator *Op1, | ||||||
| 2574 | const Operator *Op2) { | ||||||
| 2575 | if (Op1->getOpcode() != Op2->getOpcode()) | ||||||
| 2576 | return None; | ||||||
| 2577 | |||||||
| 2578 | auto getOperands = [&](unsigned OpNum) -> auto { | ||||||
| 2579 | return std::make_pair(Op1->getOperand(OpNum), Op2->getOperand(OpNum)); | ||||||
| 2580 | }; | ||||||
| 2581 | |||||||
| 2582 | switch (Op1->getOpcode()) { | ||||||
| 2583 | default: | ||||||
| 2584 | break; | ||||||
| 2585 | case Instruction::Add: | ||||||
| 2586 | case Instruction::Sub: | ||||||
| 2587 | if (Op1->getOperand(0) == Op2->getOperand(0)) | ||||||
| 2588 | return getOperands(1); | ||||||
| 2589 | if (Op1->getOperand(1) == Op2->getOperand(1)) | ||||||
| 2590 | return getOperands(0); | ||||||
| 2591 | break; | ||||||
| 2592 | case Instruction::Mul: { | ||||||
| 2593 | // invertible if A * B == (A * B) mod 2^N where A, and B are integers | ||||||
| 2594 | // and N is the bitwdith. The nsw case is non-obvious, but proven by | ||||||
| 2595 | // alive2: https://alive2.llvm.org/ce/z/Z6D5qK | ||||||
| 2596 | auto *OBO1 = cast<OverflowingBinaryOperator>(Op1); | ||||||
| 2597 | auto *OBO2 = cast<OverflowingBinaryOperator>(Op2); | ||||||
| 2598 | if ((!OBO1->hasNoUnsignedWrap() || !OBO2->hasNoUnsignedWrap()) && | ||||||
| 2599 | (!OBO1->hasNoSignedWrap() || !OBO2->hasNoSignedWrap())) | ||||||
| 2600 | break; | ||||||
| 2601 | |||||||
| 2602 | // Assume operand order has been canonicalized | ||||||
| 2603 | if (Op1->getOperand(1) == Op2->getOperand(1) && | ||||||
| 2604 | isa<ConstantInt>(Op1->getOperand(1)) && | ||||||
| 2605 | !cast<ConstantInt>(Op1->getOperand(1))->isZero()) | ||||||
| 2606 | return getOperands(0); | ||||||
| 2607 | break; | ||||||
| 2608 | } | ||||||
| 2609 | case Instruction::Shl: { | ||||||
| 2610 | // Same as multiplies, with the difference that we don't need to check | ||||||
| 2611 | // for a non-zero multiply. Shifts always multiply by non-zero. | ||||||
| 2612 | auto *OBO1 = cast<OverflowingBinaryOperator>(Op1); | ||||||
| 2613 | auto *OBO2 = cast<OverflowingBinaryOperator>(Op2); | ||||||
| 2614 | if ((!OBO1->hasNoUnsignedWrap() || !OBO2->hasNoUnsignedWrap()) && | ||||||
| 2615 | (!OBO1->hasNoSignedWrap() || !OBO2->hasNoSignedWrap())) | ||||||
| 2616 | break; | ||||||
| 2617 | |||||||
| 2618 | if (Op1->getOperand(1) == Op2->getOperand(1)) | ||||||
| 2619 | return getOperands(0); | ||||||
| 2620 | break; | ||||||
| 2621 | } | ||||||
| 2622 | case Instruction::AShr: | ||||||
| 2623 | case Instruction::LShr: { | ||||||
| 2624 | auto *PEO1 = cast<PossiblyExactOperator>(Op1); | ||||||
| 2625 | auto *PEO2 = cast<PossiblyExactOperator>(Op2); | ||||||
| 2626 | if (!PEO1->isExact() || !PEO2->isExact()) | ||||||
| 2627 | break; | ||||||
| 2628 | |||||||
| 2629 | if (Op1->getOperand(1) == Op2->getOperand(1)) | ||||||
| 2630 | return getOperands(0); | ||||||
| 2631 | break; | ||||||
| 2632 | } | ||||||
| 2633 | case Instruction::SExt: | ||||||
| 2634 | case Instruction::ZExt: | ||||||
| 2635 | if (Op1->getOperand(0)->getType() == Op2->getOperand(0)->getType()) | ||||||
| 2636 | return getOperands(0); | ||||||
| 2637 | break; | ||||||
| 2638 | case Instruction::PHI: { | ||||||
| 2639 | const PHINode *PN1 = cast<PHINode>(Op1); | ||||||
| 2640 | const PHINode *PN2 = cast<PHINode>(Op2); | ||||||
| 2641 | |||||||
| 2642 | // If PN1 and PN2 are both recurrences, can we prove the entire recurrences | ||||||
| 2643 | // are a single invertible function of the start values? Note that repeated | ||||||
| 2644 | // application of an invertible function is also invertible | ||||||
| 2645 | BinaryOperator *BO1 = nullptr; | ||||||
| 2646 | Value *Start1 = nullptr, *Step1 = nullptr; | ||||||
| 2647 | BinaryOperator *BO2 = nullptr; | ||||||
| 2648 | Value *Start2 = nullptr, *Step2 = nullptr; | ||||||
| 2649 | if (PN1->getParent() != PN2->getParent() || | ||||||
| 2650 | !matchSimpleRecurrence(PN1, BO1, Start1, Step1) || | ||||||
| 2651 | !matchSimpleRecurrence(PN2, BO2, Start2, Step2)) | ||||||
| 2652 | break; | ||||||
| 2653 | |||||||
| 2654 | auto Values = getInvertibleOperands(cast<Operator>(BO1), | ||||||
| 2655 | cast<Operator>(BO2)); | ||||||
| 2656 | if (!Values) | ||||||
| 2657 | break; | ||||||
| 2658 | |||||||
| 2659 | // We have to be careful of mutually defined recurrences here. Ex: | ||||||
| 2660 | // * X_i = X_(i-1) OP Y_(i-1), and Y_i = X_(i-1) OP V | ||||||
| 2661 | // * X_i = Y_i = X_(i-1) OP Y_(i-1) | ||||||
| 2662 | // The invertibility of these is complicated, and not worth reasoning | ||||||
| 2663 | // about (yet?). | ||||||
| 2664 | if (Values->first != PN1 || Values->second != PN2) | ||||||
| 2665 | break; | ||||||
| 2666 | |||||||
| 2667 | return std::make_pair(Start1, Start2); | ||||||
| 2668 | } | ||||||
| 2669 | } | ||||||
| 2670 | return None; | ||||||
| 2671 | } | ||||||
| 2672 | |||||||
| 2673 | /// Return true if V2 == V1 + X, where X is known non-zero. | ||||||
| 2674 | static bool isAddOfNonZero(const Value *V1, const Value *V2, unsigned Depth, | ||||||
| 2675 | const Query &Q) { | ||||||
| 2676 | const BinaryOperator *BO = dyn_cast<BinaryOperator>(V1); | ||||||
| 2677 | if (!BO || BO->getOpcode() != Instruction::Add) | ||||||
| 2678 | return false; | ||||||
| 2679 | Value *Op = nullptr; | ||||||
| 2680 | if (V2 == BO->getOperand(0)) | ||||||
| 2681 | Op = BO->getOperand(1); | ||||||
| 2682 | else if (V2 == BO->getOperand(1)) | ||||||
| 2683 | Op = BO->getOperand(0); | ||||||
| 2684 | else | ||||||
| 2685 | return false; | ||||||
| 2686 | return isKnownNonZero(Op, Depth + 1, Q); | ||||||
| 2687 | } | ||||||
| 2688 | |||||||
| 2689 | /// Return true if V2 == V1 * C, where V1 is known non-zero, C is not 0/1 and | ||||||
| 2690 | /// the multiplication is nuw or nsw. | ||||||
| 2691 | static bool isNonEqualMul(const Value *V1, const Value *V2, unsigned Depth, | ||||||
| 2692 | const Query &Q) { | ||||||
| 2693 | if (auto *OBO = dyn_cast<OverflowingBinaryOperator>(V2)) { | ||||||
| 2694 | const APInt *C; | ||||||
| 2695 | return match(OBO, m_Mul(m_Specific(V1), m_APInt(C))) && | ||||||
| 2696 | (OBO->hasNoUnsignedWrap() || OBO->hasNoSignedWrap()) && | ||||||
| 2697 | !C->isNullValue() && !C->isOneValue() && | ||||||
| 2698 | isKnownNonZero(V1, Depth + 1, Q); | ||||||
| 2699 | } | ||||||
| 2700 | return false; | ||||||
| 2701 | } | ||||||
| 2702 | |||||||
| 2703 | /// Return true if V2 == V1 << C, where V1 is known non-zero, C is not 0 and | ||||||
| 2704 | /// the shift is nuw or nsw. | ||||||
| 2705 | static bool isNonEqualShl(const Value *V1, const Value *V2, unsigned Depth, | ||||||
| 2706 | const Query &Q) { | ||||||
| 2707 | if (auto *OBO = dyn_cast<OverflowingBinaryOperator>(V2)) { | ||||||
| 2708 | const APInt *C; | ||||||
| 2709 | return match(OBO, m_Shl(m_Specific(V1), m_APInt(C))) && | ||||||
| 2710 | (OBO->hasNoUnsignedWrap() || OBO->hasNoSignedWrap()) && | ||||||
| 2711 | !C->isNullValue() && isKnownNonZero(V1, Depth + 1, Q); | ||||||
| 2712 | } | ||||||
| 2713 | return false; | ||||||
| 2714 | } | ||||||
| 2715 | |||||||
| 2716 | static bool isNonEqualPHIs(const PHINode *PN1, const PHINode *PN2, | ||||||
| 2717 | unsigned Depth, const Query &Q) { | ||||||
| 2718 | // Check two PHIs are in same block. | ||||||
| 2719 | if (PN1->getParent() != PN2->getParent()) | ||||||
| 2720 | return false; | ||||||
| 2721 | |||||||
| 2722 | SmallPtrSet<const BasicBlock *, 8> VisitedBBs; | ||||||
| 2723 | bool UsedFullRecursion = false; | ||||||
| 2724 | for (const BasicBlock *IncomBB : PN1->blocks()) { | ||||||
| 2725 | if (!VisitedBBs.insert(IncomBB).second) | ||||||
| 2726 | continue; // Don't reprocess blocks that we have dealt with already. | ||||||
| 2727 | const Value *IV1 = PN1->getIncomingValueForBlock(IncomBB); | ||||||
| 2728 | const Value *IV2 = PN2->getIncomingValueForBlock(IncomBB); | ||||||
| 2729 | const APInt *C1, *C2; | ||||||
| 2730 | if (match(IV1, m_APInt(C1)) && match(IV2, m_APInt(C2)) && *C1 != *C2) | ||||||
| 2731 | continue; | ||||||
| 2732 | |||||||
| 2733 | // Only one pair of phi operands is allowed for full recursion. | ||||||
| 2734 | if (UsedFullRecursion) | ||||||
| 2735 | return false; | ||||||
| 2736 | |||||||
| 2737 | Query RecQ = Q; | ||||||
| 2738 | RecQ.CxtI = IncomBB->getTerminator(); | ||||||
| 2739 | if (!isKnownNonEqual(IV1, IV2, Depth + 1, RecQ)) | ||||||
| 2740 | return false; | ||||||
| 2741 | UsedFullRecursion = true; | ||||||
| 2742 | } | ||||||
| 2743 | return true; | ||||||
| 2744 | } | ||||||
| 2745 | |||||||
| 2746 | /// Return true if it is known that V1 != V2. | ||||||
| 2747 | static bool isKnownNonEqual(const Value *V1, const Value *V2, unsigned Depth, | ||||||
| 2748 | const Query &Q) { | ||||||
| 2749 | if (V1 == V2) | ||||||
| 2750 | return false; | ||||||
| 2751 | if (V1->getType() != V2->getType()) | ||||||
| 2752 | // We can't look through casts yet. | ||||||
| 2753 | return false; | ||||||
| 2754 | |||||||
| 2755 | if (Depth >= MaxAnalysisRecursionDepth) | ||||||
| 2756 | return false; | ||||||
| 2757 | |||||||
| 2758 | // See if we can recurse through (exactly one of) our operands. This | ||||||
| 2759 | // requires our operation be 1-to-1 and map every input value to exactly | ||||||
| 2760 | // one output value. Such an operation is invertible. | ||||||
| 2761 | auto *O1 = dyn_cast<Operator>(V1); | ||||||
| 2762 | auto *O2 = dyn_cast<Operator>(V2); | ||||||
| 2763 | if (O1 && O2 && O1->getOpcode() == O2->getOpcode()) { | ||||||
| 2764 | if (auto Values = getInvertibleOperands(O1, O2)) | ||||||
| 2765 | return isKnownNonEqual(Values->first, Values->second, Depth + 1, Q); | ||||||
| 2766 | |||||||
| 2767 | if (const PHINode *PN1 = dyn_cast<PHINode>(V1)) { | ||||||
| 2768 | const PHINode *PN2 = cast<PHINode>(V2); | ||||||
| 2769 | // FIXME: This is missing a generalization to handle the case where one is | ||||||
| 2770 | // a PHI and another one isn't. | ||||||
| 2771 | if (isNonEqualPHIs(PN1, PN2, Depth, Q)) | ||||||
| 2772 | return true; | ||||||
| 2773 | }; | ||||||
| 2774 | } | ||||||
| 2775 | |||||||
| 2776 | if (isAddOfNonZero(V1, V2, Depth, Q) || isAddOfNonZero(V2, V1, Depth, Q)) | ||||||
| 2777 | return true; | ||||||
| 2778 | |||||||
| 2779 | if (isNonEqualMul(V1, V2, Depth, Q) || isNonEqualMul(V2, V1, Depth, Q)) | ||||||
| 2780 | return true; | ||||||
| 2781 | |||||||
| 2782 | if (isNonEqualShl(V1, V2, Depth, Q) || isNonEqualShl(V2, V1, Depth, Q)) | ||||||
| 2783 | return true; | ||||||
| 2784 | |||||||
| 2785 | if (V1->getType()->isIntOrIntVectorTy()) { | ||||||
| 2786 | // Are any known bits in V1 contradictory to known bits in V2? If V1 | ||||||
| 2787 | // has a known zero where V2 has a known one, they must not be equal. | ||||||
| 2788 | KnownBits Known1 = computeKnownBits(V1, Depth, Q); | ||||||
| 2789 | KnownBits Known2 = computeKnownBits(V2, Depth, Q); | ||||||
| 2790 | |||||||
| 2791 | if (Known1.Zero.intersects(Known2.One) || | ||||||
| 2792 | Known2.Zero.intersects(Known1.One)) | ||||||
| 2793 | return true; | ||||||
| 2794 | } | ||||||
| 2795 | return false; | ||||||
| 2796 | } | ||||||
| 2797 | |||||||
| 2798 | /// Return true if 'V & Mask' is known to be zero. We use this predicate to | ||||||
| 2799 | /// simplify operations downstream. Mask is known to be zero for bits that V | ||||||
| 2800 | /// cannot have. | ||||||
| 2801 | /// | ||||||
| 2802 | /// This function is defined on values with integer type, values with pointer | ||||||
| 2803 | /// type, and vectors of integers. In the case | ||||||
| 2804 | /// where V is a vector, the mask, known zero, and known one values are the | ||||||
| 2805 | /// same width as the vector element, and the bit is set only if it is true | ||||||
| 2806 | /// for all of the elements in the vector. | ||||||
| 2807 | bool MaskedValueIsZero(const Value *V, const APInt &Mask, unsigned Depth, | ||||||
| 2808 | const Query &Q) { | ||||||
| 2809 | KnownBits Known(Mask.getBitWidth()); | ||||||
| 2810 | computeKnownBits(V, Known, Depth, Q); | ||||||
| 2811 | return Mask.isSubsetOf(Known.Zero); | ||||||
| 2812 | } | ||||||
| 2813 | |||||||
| 2814 | // Match a signed min+max clamp pattern like smax(smin(In, CHigh), CLow). | ||||||
| 2815 | // Returns the input and lower/upper bounds. | ||||||
| 2816 | static bool isSignedMinMaxClamp(const Value *Select, const Value *&In, | ||||||
| 2817 | const APInt *&CLow, const APInt *&CHigh) { | ||||||
| 2818 | assert(isa<Operator>(Select) &&((void)0) | ||||||
| 2819 | cast<Operator>(Select)->getOpcode() == Instruction::Select &&((void)0) | ||||||
| 2820 | "Input should be a Select!")((void)0); | ||||||
| 2821 | |||||||
| 2822 | const Value *LHS = nullptr, *RHS = nullptr; | ||||||
| 2823 | SelectPatternFlavor SPF = matchSelectPattern(Select, LHS, RHS).Flavor; | ||||||
| 2824 | if (SPF != SPF_SMAX && SPF != SPF_SMIN) | ||||||
| 2825 | return false; | ||||||
| 2826 | |||||||
| 2827 | if (!match(RHS, m_APInt(CLow))) | ||||||
| 2828 | return false; | ||||||
| 2829 | |||||||
| 2830 | const Value *LHS2 = nullptr, *RHS2 = nullptr; | ||||||
| 2831 | SelectPatternFlavor SPF2 = matchSelectPattern(LHS, LHS2, RHS2).Flavor; | ||||||
| 2832 | if (getInverseMinMaxFlavor(SPF) != SPF2) | ||||||
| 2833 | return false; | ||||||
| 2834 | |||||||
| 2835 | if (!match(RHS2, m_APInt(CHigh))) | ||||||
| 2836 | return false; | ||||||
| 2837 | |||||||
| 2838 | if (SPF == SPF_SMIN) | ||||||
| 2839 | std::swap(CLow, CHigh); | ||||||
| 2840 | |||||||
| 2841 | In = LHS2; | ||||||
| 2842 | return CLow->sle(*CHigh); | ||||||
| 2843 | } | ||||||
| 2844 | |||||||
| 2845 | /// For vector constants, loop over the elements and find the constant with the | ||||||
| 2846 | /// minimum number of sign bits. Return 0 if the value is not a vector constant | ||||||
| 2847 | /// or if any element was not analyzed; otherwise, return the count for the | ||||||
| 2848 | /// element with the minimum number of sign bits. | ||||||
| 2849 | static unsigned computeNumSignBitsVectorConstant(const Value *V, | ||||||
| 2850 | const APInt &DemandedElts, | ||||||
| 2851 | unsigned TyBits) { | ||||||
| 2852 | const auto *CV = dyn_cast<Constant>(V); | ||||||
| 2853 | if (!CV || !isa<FixedVectorType>(CV->getType())) | ||||||
| 2854 | return 0; | ||||||
| 2855 | |||||||
| 2856 | unsigned MinSignBits = TyBits; | ||||||
| 2857 | unsigned NumElts = cast<FixedVectorType>(CV->getType())->getNumElements(); | ||||||
| 2858 | for (unsigned i = 0; i != NumElts; ++i) { | ||||||
| 2859 | if (!DemandedElts[i]) | ||||||
| 2860 | continue; | ||||||
| 2861 | // If we find a non-ConstantInt, bail out. | ||||||
| 2862 | auto *Elt = dyn_cast_or_null<ConstantInt>(CV->getAggregateElement(i)); | ||||||
| 2863 | if (!Elt) | ||||||
| 2864 | return 0; | ||||||
| 2865 | |||||||
| 2866 | MinSignBits = std::min(MinSignBits, Elt->getValue().getNumSignBits()); | ||||||
| 2867 | } | ||||||
| 2868 | |||||||
| 2869 | return MinSignBits; | ||||||
| 2870 | } | ||||||
| 2871 | |||||||
| 2872 | static unsigned ComputeNumSignBitsImpl(const Value *V, | ||||||
| 2873 | const APInt &DemandedElts, | ||||||
| 2874 | unsigned Depth, const Query &Q); | ||||||
| 2875 | |||||||
| 2876 | static unsigned ComputeNumSignBits(const Value *V, const APInt &DemandedElts, | ||||||
| 2877 | unsigned Depth, const Query &Q) { | ||||||
| 2878 | unsigned Result = ComputeNumSignBitsImpl(V, DemandedElts, Depth, Q); | ||||||
| 2879 | assert(Result > 0 && "At least one sign bit needs to be present!")((void)0); | ||||||
| 2880 | return Result; | ||||||
| 2881 | } | ||||||
| 2882 | |||||||
| 2883 | /// Return the number of times the sign bit of the register is replicated into | ||||||
| 2884 | /// the other bits. We know that at least 1 bit is always equal to the sign bit | ||||||
| 2885 | /// (itself), but other cases can give us information. For example, immediately | ||||||
| 2886 | /// after an "ashr X, 2", we know that the top 3 bits are all equal to each | ||||||
| 2887 | /// other, so we return 3. For vectors, return the number of sign bits for the | ||||||
| 2888 | /// vector element with the minimum number of known sign bits of the demanded | ||||||
| 2889 | /// elements in the vector specified by DemandedElts. | ||||||
| 2890 | static unsigned ComputeNumSignBitsImpl(const Value *V, | ||||||
| 2891 | const APInt &DemandedElts, | ||||||
| 2892 | unsigned Depth, const Query &Q) { | ||||||
| 2893 | Type *Ty = V->getType(); | ||||||
| 2894 | |||||||
| 2895 | // FIXME: We currently have no way to represent the DemandedElts of a scalable | ||||||
| 2896 | // vector | ||||||
| 2897 | if (isa<ScalableVectorType>(Ty)) | ||||||
| 2898 | return 1; | ||||||
| 2899 | |||||||
| 2900 | #ifndef NDEBUG1 | ||||||
| 2901 | assert(Depth <= MaxAnalysisRecursionDepth && "Limit Search Depth")((void)0); | ||||||
| 2902 | |||||||
| 2903 | if (auto *FVTy = dyn_cast<FixedVectorType>(Ty)) { | ||||||
| 2904 | assert(((void)0) | ||||||
| 2905 | FVTy->getNumElements() == DemandedElts.getBitWidth() &&((void)0) | ||||||
| 2906 | "DemandedElt width should equal the fixed vector number of elements")((void)0); | ||||||
| 2907 | } else { | ||||||
| 2908 | assert(DemandedElts == APInt(1, 1) &&((void)0) | ||||||
| 2909 | "DemandedElt width should be 1 for scalars")((void)0); | ||||||
| 2910 | } | ||||||
| 2911 | #endif | ||||||
| 2912 | |||||||
| 2913 | // We return the minimum number of sign bits that are guaranteed to be present | ||||||
| 2914 | // in V, so for undef we have to conservatively return 1. We don't have the | ||||||
| 2915 | // same behavior for poison though -- that's a FIXME today. | ||||||
| 2916 | |||||||
| 2917 | Type *ScalarTy = Ty->getScalarType(); | ||||||
| 2918 | unsigned TyBits = ScalarTy->isPointerTy() ? | ||||||
| 2919 | Q.DL.getPointerTypeSizeInBits(ScalarTy) : | ||||||
| 2920 | Q.DL.getTypeSizeInBits(ScalarTy); | ||||||
| 2921 | |||||||
| 2922 | unsigned Tmp, Tmp2; | ||||||
| 2923 | unsigned FirstAnswer = 1; | ||||||
| 2924 | |||||||
| 2925 | // Note that ConstantInt is handled by the general computeKnownBits case | ||||||
| 2926 | // below. | ||||||
| 2927 | |||||||
| 2928 | if (Depth == MaxAnalysisRecursionDepth) | ||||||
| 2929 | return 1; | ||||||
| 2930 | |||||||
| 2931 | if (auto *U = dyn_cast<Operator>(V)) { | ||||||
| 2932 | switch (Operator::getOpcode(V)) { | ||||||
| 2933 | default: break; | ||||||
| 2934 | case Instruction::SExt: | ||||||
| 2935 | Tmp = TyBits - U->getOperand(0)->getType()->getScalarSizeInBits(); | ||||||
| 2936 | return ComputeNumSignBits(U->getOperand(0), Depth + 1, Q) + Tmp; | ||||||
| 2937 | |||||||
| 2938 | case Instruction::SDiv: { | ||||||
| 2939 | const APInt *Denominator; | ||||||
| 2940 | // sdiv X, C -> adds log(C) sign bits. | ||||||
| 2941 | if (match(U->getOperand(1), m_APInt(Denominator))) { | ||||||
| 2942 | |||||||
| 2943 | // Ignore non-positive denominator. | ||||||
| 2944 | if (!Denominator->isStrictlyPositive()) | ||||||
| 2945 | break; | ||||||
| 2946 | |||||||
| 2947 | // Calculate the incoming numerator bits. | ||||||
| 2948 | unsigned NumBits = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q); | ||||||
| 2949 | |||||||
| 2950 | // Add floor(log(C)) bits to the numerator bits. | ||||||
| 2951 | return std::min(TyBits, NumBits + Denominator->logBase2()); | ||||||
| 2952 | } | ||||||
| 2953 | break; | ||||||
| 2954 | } | ||||||
| 2955 | |||||||
| 2956 | case Instruction::SRem: { | ||||||
| 2957 | Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q); | ||||||
| 2958 | |||||||
| 2959 | const APInt *Denominator; | ||||||
| 2960 | // srem X, C -> we know that the result is within [-C+1,C) when C is a | ||||||
| 2961 | // positive constant. This let us put a lower bound on the number of sign | ||||||
| 2962 | // bits. | ||||||
| 2963 | if (match(U->getOperand(1), m_APInt(Denominator))) { | ||||||
| 2964 | |||||||
| 2965 | // Ignore non-positive denominator. | ||||||
| 2966 | if (Denominator->isStrictlyPositive()) { | ||||||
| 2967 | // Calculate the leading sign bit constraints by examining the | ||||||
| 2968 | // denominator. Given that the denominator is positive, there are two | ||||||
| 2969 | // cases: | ||||||
| 2970 | // | ||||||
| 2971 | // 1. The numerator is positive. The result range is [0,C) and | ||||||
| 2972 | // [0,C) u< (1 << ceilLogBase2(C)). | ||||||
| 2973 | // | ||||||
| 2974 | // 2. The numerator is negative. Then the result range is (-C,0] and | ||||||
| 2975 | // integers in (-C,0] are either 0 or >u (-1 << ceilLogBase2(C)). | ||||||
| 2976 | // | ||||||
| 2977 | // Thus a lower bound on the number of sign bits is `TyBits - | ||||||
| 2978 | // ceilLogBase2(C)`. | ||||||
| 2979 | |||||||
| 2980 | unsigned ResBits = TyBits - Denominator->ceilLogBase2(); | ||||||
| 2981 | Tmp = std::max(Tmp, ResBits); | ||||||
| 2982 | } | ||||||
| 2983 | } | ||||||
| 2984 | return Tmp; | ||||||
| 2985 | } | ||||||
| 2986 | |||||||
| 2987 | case Instruction::AShr: { | ||||||
| 2988 | Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q); | ||||||
| 2989 | // ashr X, C -> adds C sign bits. Vectors too. | ||||||
| 2990 | const APInt *ShAmt; | ||||||
| 2991 | if (match(U->getOperand(1), m_APInt(ShAmt))) { | ||||||
| 2992 | if (ShAmt->uge(TyBits)) | ||||||
| 2993 | break; // Bad shift. | ||||||
| 2994 | unsigned ShAmtLimited = ShAmt->getZExtValue(); | ||||||
| 2995 | Tmp += ShAmtLimited; | ||||||
| 2996 | if (Tmp > TyBits) Tmp = TyBits; | ||||||
| 2997 | } | ||||||
| 2998 | return Tmp; | ||||||
| 2999 | } | ||||||
| 3000 | case Instruction::Shl: { | ||||||
| 3001 | const APInt *ShAmt; | ||||||
| 3002 | if (match(U->getOperand(1), m_APInt(ShAmt))) { | ||||||
| 3003 | // shl destroys sign bits. | ||||||
| 3004 | Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q); | ||||||
| 3005 | if (ShAmt->uge(TyBits) || // Bad shift. | ||||||
| 3006 | ShAmt->uge(Tmp)) break; // Shifted all sign bits out. | ||||||
| 3007 | Tmp2 = ShAmt->getZExtValue(); | ||||||
| 3008 | return Tmp - Tmp2; | ||||||
| 3009 | } | ||||||
| 3010 | break; | ||||||
| 3011 | } | ||||||
| 3012 | case Instruction::And: | ||||||
| 3013 | case Instruction::Or: | ||||||
| 3014 | case Instruction::Xor: // NOT is handled here. | ||||||
| 3015 | // Logical binary ops preserve the number of sign bits at the worst. | ||||||
| 3016 | Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q); | ||||||
| 3017 | if (Tmp != 1) { | ||||||
| 3018 | Tmp2 = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q); | ||||||
| 3019 | FirstAnswer = std::min(Tmp, Tmp2); | ||||||
| 3020 | // We computed what we know about the sign bits as our first | ||||||
| 3021 | // answer. Now proceed to the generic code that uses | ||||||
| 3022 | // computeKnownBits, and pick whichever answer is better. | ||||||
| 3023 | } | ||||||
| 3024 | break; | ||||||
| 3025 | |||||||
| 3026 | case Instruction::Select: { | ||||||
| 3027 | // If we have a clamp pattern, we know that the number of sign bits will | ||||||
| 3028 | // be the minimum of the clamp min/max range. | ||||||
| 3029 | const Value *X; | ||||||
| 3030 | const APInt *CLow, *CHigh; | ||||||
| 3031 | if (isSignedMinMaxClamp(U, X, CLow, CHigh)) | ||||||
| 3032 | return std::min(CLow->getNumSignBits(), CHigh->getNumSignBits()); | ||||||
| 3033 | |||||||
| 3034 | Tmp = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q); | ||||||
| 3035 | if (Tmp == 1) break; | ||||||
| 3036 | Tmp2 = ComputeNumSignBits(U->getOperand(2), Depth + 1, Q); | ||||||
| 3037 | return std::min(Tmp, Tmp2); | ||||||
| 3038 | } | ||||||
| 3039 | |||||||
| 3040 | case Instruction::Add: | ||||||
| 3041 | // Add can have at most one carry bit. Thus we know that the output | ||||||
| 3042 | // is, at worst, one more bit than the inputs. | ||||||
| 3043 | Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q); | ||||||
| 3044 | if (Tmp == 1) break; | ||||||
| 3045 | |||||||
| 3046 | // Special case decrementing a value (ADD X, -1): | ||||||
| 3047 | if (const auto *CRHS = dyn_cast<Constant>(U->getOperand(1))) | ||||||
| 3048 | if (CRHS->isAllOnesValue()) { | ||||||
| 3049 | KnownBits Known(TyBits); | ||||||
| 3050 | computeKnownBits(U->getOperand(0), Known, Depth + 1, Q); | ||||||
| 3051 | |||||||
| 3052 | // If the input is known to be 0 or 1, the output is 0/-1, which is | ||||||
| 3053 | // all sign bits set. | ||||||
| 3054 | if ((Known.Zero | 1).isAllOnesValue()) | ||||||
| 3055 | return TyBits; | ||||||
| 3056 | |||||||
| 3057 | // If we are subtracting one from a positive number, there is no carry | ||||||
| 3058 | // out of the result. | ||||||
| 3059 | if (Known.isNonNegative()) | ||||||
| 3060 | return Tmp; | ||||||
| 3061 | } | ||||||
| 3062 | |||||||
| 3063 | Tmp2 = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q); | ||||||
| 3064 | if (Tmp2 == 1) break; | ||||||
| 3065 | return std::min(Tmp, Tmp2) - 1; | ||||||
| 3066 | |||||||
| 3067 | case Instruction::Sub: | ||||||
| 3068 | Tmp2 = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q); | ||||||
| 3069 | if (Tmp2 == 1) break; | ||||||
| 3070 | |||||||
| 3071 | // Handle NEG. | ||||||
| 3072 | if (const auto *CLHS = dyn_cast<Constant>(U->getOperand(0))) | ||||||
| 3073 | if (CLHS->isNullValue()) { | ||||||
| 3074 | KnownBits Known(TyBits); | ||||||
| 3075 | computeKnownBits(U->getOperand(1), Known, Depth + 1, Q); | ||||||
| 3076 | // If the input is known to be 0 or 1, the output is 0/-1, which is | ||||||
| 3077 | // all sign bits set. | ||||||
| 3078 | if ((Known.Zero | 1).isAllOnesValue()) | ||||||
| 3079 | return TyBits; | ||||||
| 3080 | |||||||
| 3081 | // If the input is known to be positive (the sign bit is known clear), | ||||||
| 3082 | // the output of the NEG has the same number of sign bits as the | ||||||
| 3083 | // input. | ||||||
| 3084 | if (Known.isNonNegative()) | ||||||
| 3085 | return Tmp2; | ||||||
| 3086 | |||||||
| 3087 | // Otherwise, we treat this like a SUB. | ||||||
| 3088 | } | ||||||
| 3089 | |||||||
| 3090 | // Sub can have at most one carry bit. Thus we know that the output | ||||||
| 3091 | // is, at worst, one more bit than the inputs. | ||||||
| 3092 | Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q); | ||||||
| 3093 | if (Tmp == 1) break; | ||||||
| 3094 | return std::min(Tmp, Tmp2) - 1; | ||||||
| 3095 | |||||||
| 3096 | case Instruction::Mul: { | ||||||
| 3097 | // The output of the Mul can be at most twice the valid bits in the | ||||||
| 3098 | // inputs. | ||||||
| 3099 | unsigned SignBitsOp0 = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q); | ||||||
| 3100 | if (SignBitsOp0 == 1) break; | ||||||
| 3101 | unsigned SignBitsOp1 = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q); | ||||||
| 3102 | if (SignBitsOp1 == 1) break; | ||||||
| 3103 | unsigned OutValidBits = | ||||||
| 3104 | (TyBits - SignBitsOp0 + 1) + (TyBits - SignBitsOp1 + 1); | ||||||
| 3105 | return OutValidBits > TyBits ? 1 : TyBits - OutValidBits + 1; | ||||||
| 3106 | } | ||||||
| 3107 | |||||||
| 3108 | case Instruction::PHI: { | ||||||
| 3109 | const PHINode *PN = cast<PHINode>(U); | ||||||
| 3110 | unsigned NumIncomingValues = PN->getNumIncomingValues(); | ||||||
| 3111 | // Don't analyze large in-degree PHIs. | ||||||
| 3112 | if (NumIncomingValues > 4) break; | ||||||
| 3113 | // Unreachable blocks may have zero-operand PHI nodes. | ||||||
| 3114 | if (NumIncomingValues == 0) break; | ||||||
| 3115 | |||||||
| 3116 | // Take the minimum of all incoming values. This can't infinitely loop | ||||||
| 3117 | // because of our depth threshold. | ||||||
| 3118 | Query RecQ = Q; | ||||||
| 3119 | Tmp = TyBits; | ||||||
| 3120 | for (unsigned i = 0, e = NumIncomingValues; i != e; ++i) { | ||||||
| 3121 | if (Tmp == 1) return Tmp; | ||||||
| 3122 | RecQ.CxtI = PN->getIncomingBlock(i)->getTerminator(); | ||||||
| 3123 | Tmp = std::min( | ||||||
| 3124 | Tmp, ComputeNumSignBits(PN->getIncomingValue(i), Depth + 1, RecQ)); | ||||||
| 3125 | } | ||||||
| 3126 | return Tmp; | ||||||
| 3127 | } | ||||||
| 3128 | |||||||
| 3129 | case Instruction::Trunc: | ||||||
| 3130 | // FIXME: it's tricky to do anything useful for this, but it is an | ||||||
| 3131 | // important case for targets like X86. | ||||||
| 3132 | break; | ||||||
| 3133 | |||||||
| 3134 | case Instruction::ExtractElement: | ||||||
| 3135 | // Look through extract element. At the moment we keep this simple and | ||||||
| 3136 | // skip tracking the specific element. But at least we might find | ||||||
| 3137 | // information valid for all elements of the vector (for example if vector | ||||||
| 3138 | // is sign extended, shifted, etc). | ||||||
| 3139 | return ComputeNumSignBits(U->getOperand(0), Depth + 1, Q); | ||||||
| 3140 | |||||||
| 3141 | case Instruction::ShuffleVector: { | ||||||
| 3142 | // Collect the minimum number of sign bits that are shared by every vector | ||||||
| 3143 | // element referenced by the shuffle. | ||||||
| 3144 | auto *Shuf = dyn_cast<ShuffleVectorInst>(U); | ||||||
| 3145 | if (!Shuf) { | ||||||
| 3146 | // FIXME: Add support for shufflevector constant expressions. | ||||||
| 3147 | return 1; | ||||||
| 3148 | } | ||||||
| 3149 | APInt DemandedLHS, DemandedRHS; | ||||||
| 3150 | // For undef elements, we don't know anything about the common state of | ||||||
| 3151 | // the shuffle result. | ||||||
| 3152 | if (!getShuffleDemandedElts(Shuf, DemandedElts, DemandedLHS, DemandedRHS)) | ||||||
| 3153 | return 1; | ||||||
| 3154 | Tmp = std::numeric_limits<unsigned>::max(); | ||||||
| 3155 | if (!!DemandedLHS) { | ||||||
| 3156 | const Value *LHS = Shuf->getOperand(0); | ||||||
| 3157 | Tmp = ComputeNumSignBits(LHS, DemandedLHS, Depth + 1, Q); | ||||||
| 3158 | } | ||||||
| 3159 | // If we don't know anything, early out and try computeKnownBits | ||||||
| 3160 | // fall-back. | ||||||
| 3161 | if (Tmp == 1) | ||||||
| 3162 | break; | ||||||
| 3163 | if (!!DemandedRHS) { | ||||||
| 3164 | const Value *RHS = Shuf->getOperand(1); | ||||||
| 3165 | Tmp2 = ComputeNumSignBits(RHS, DemandedRHS, Depth + 1, Q); | ||||||
| 3166 | Tmp = std::min(Tmp, Tmp2); | ||||||
| 3167 | } | ||||||
| 3168 | // If we don't know anything, early out and try computeKnownBits | ||||||
| 3169 | // fall-back. | ||||||
| 3170 | if (Tmp == 1) | ||||||
| 3171 | break; | ||||||
| 3172 | assert(Tmp <= TyBits && "Failed to determine minimum sign bits")((void)0); | ||||||
| 3173 | return Tmp; | ||||||
| 3174 | } | ||||||
| 3175 | case Instruction::Call: { | ||||||
| 3176 | if (const auto *II = dyn_cast<IntrinsicInst>(U)) { | ||||||
| 3177 | switch (II->getIntrinsicID()) { | ||||||
| 3178 | default: break; | ||||||
| 3179 | case Intrinsic::abs: | ||||||
| 3180 | Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q); | ||||||
| 3181 | if (Tmp == 1) break; | ||||||
| 3182 | |||||||
| 3183 | // Absolute value reduces number of sign bits by at most 1. | ||||||
| 3184 | return Tmp - 1; | ||||||
| 3185 | } | ||||||
| 3186 | } | ||||||
| 3187 | } | ||||||
| 3188 | } | ||||||
| 3189 | } | ||||||
| 3190 | |||||||
| 3191 | // Finally, if we can prove that the top bits of the result are 0's or 1's, | ||||||
| 3192 | // use this information. | ||||||
| 3193 | |||||||
| 3194 | // If we can examine all elements of a vector constant successfully, we're | ||||||
| 3195 | // done (we can't do any better than that). If not, keep trying. | ||||||
| 3196 | if (unsigned VecSignBits = | ||||||
| 3197 | computeNumSignBitsVectorConstant(V, DemandedElts, TyBits)) | ||||||
| 3198 | return VecSignBits; | ||||||
| 3199 | |||||||
| 3200 | KnownBits Known(TyBits); | ||||||
| 3201 | computeKnownBits(V, DemandedElts, Known, Depth, Q); | ||||||
| 3202 | |||||||
| 3203 | // If we know that the sign bit is either zero or one, determine the number of | ||||||
| 3204 | // identical bits in the top of the input value. | ||||||
| 3205 | return std::max(FirstAnswer, Known.countMinSignBits()); | ||||||
| 3206 | } | ||||||
| 3207 | |||||||
| 3208 | /// This function computes the integer multiple of Base that equals V. | ||||||
| 3209 | /// If successful, it returns true and returns the multiple in | ||||||
| 3210 | /// Multiple. If unsuccessful, it returns false. It looks | ||||||
| 3211 | /// through SExt instructions only if LookThroughSExt is true. | ||||||
| 3212 | bool llvm::ComputeMultiple(Value *V, unsigned Base, Value *&Multiple, | ||||||
| 3213 | bool LookThroughSExt, unsigned Depth) { | ||||||
| 3214 | assert(V && "No Value?")((void)0); | ||||||
| 3215 | assert(Depth <= MaxAnalysisRecursionDepth && "Limit Search Depth")((void)0); | ||||||
| 3216 | assert(V->getType()->isIntegerTy() && "Not integer or pointer type!")((void)0); | ||||||
| 3217 | |||||||
| 3218 | Type *T = V->getType(); | ||||||
| 3219 | |||||||
| 3220 | ConstantInt *CI = dyn_cast<ConstantInt>(V); | ||||||
| 3221 | |||||||
| 3222 | if (Base == 0) | ||||||
| 3223 | return false; | ||||||
| 3224 | |||||||
| 3225 | if (Base == 1) { | ||||||
| 3226 | Multiple = V; | ||||||
| 3227 | return true; | ||||||
| 3228 | } | ||||||
| 3229 | |||||||
| 3230 | ConstantExpr *CO = dyn_cast<ConstantExpr>(V); | ||||||
| 3231 | Constant *BaseVal = ConstantInt::get(T, Base); | ||||||
| 3232 | if (CO && CO == BaseVal) { | ||||||
| 3233 | // Multiple is 1. | ||||||
| 3234 | Multiple = ConstantInt::get(T, 1); | ||||||
| 3235 | return true; | ||||||
| 3236 | } | ||||||
| 3237 | |||||||
| 3238 | if (CI && CI->getZExtValue() % Base == 0) { | ||||||
| 3239 | Multiple = ConstantInt::get(T, CI->getZExtValue() / Base); | ||||||
| 3240 | return true; | ||||||
| 3241 | } | ||||||
| 3242 | |||||||
| 3243 | if (Depth == MaxAnalysisRecursionDepth) return false; | ||||||
| 3244 | |||||||
| 3245 | Operator *I = dyn_cast<Operator>(V); | ||||||
| 3246 | if (!I) return false; | ||||||
| 3247 | |||||||
| 3248 | switch (I->getOpcode()) { | ||||||
| 3249 | default: break; | ||||||
| 3250 | case Instruction::SExt: | ||||||
| 3251 | if (!LookThroughSExt) return false; | ||||||
| 3252 | // otherwise fall through to ZExt | ||||||
| 3253 | LLVM_FALLTHROUGH[[gnu::fallthrough]]; | ||||||
| 3254 | case Instruction::ZExt: | ||||||
| 3255 | return ComputeMultiple(I->getOperand(0), Base, Multiple, | ||||||
| 3256 | LookThroughSExt, Depth+1); | ||||||
| 3257 | case Instruction::Shl: | ||||||
| 3258 | case Instruction::Mul: { | ||||||
| 3259 | Value *Op0 = I->getOperand(0); | ||||||
| 3260 | Value *Op1 = I->getOperand(1); | ||||||
| 3261 | |||||||
| 3262 | if (I->getOpcode() == Instruction::Shl) { | ||||||
| 3263 | ConstantInt *Op1CI = dyn_cast<ConstantInt>(Op1); | ||||||
| 3264 | if (!Op1CI) return false; | ||||||
| 3265 | // Turn Op0 << Op1 into Op0 * 2^Op1 | ||||||
| 3266 | APInt Op1Int = Op1CI->getValue(); | ||||||
| 3267 | uint64_t BitToSet = Op1Int.getLimitedValue(Op1Int.getBitWidth() - 1); | ||||||
| 3268 | APInt API(Op1Int.getBitWidth(), 0); | ||||||
| 3269 | API.setBit(BitToSet); | ||||||
| 3270 | Op1 = ConstantInt::get(V->getContext(), API); | ||||||
| 3271 | } | ||||||
| 3272 | |||||||
| 3273 | Value *Mul0 = nullptr; | ||||||
| 3274 | if (ComputeMultiple(Op0, Base, Mul0, LookThroughSExt, Depth+1)) { | ||||||
| 3275 | if (Constant *Op1C = dyn_cast<Constant>(Op1)) | ||||||
| 3276 | if (Constant *MulC = dyn_cast<Constant>(Mul0)) { | ||||||
| 3277 | if (Op1C->getType()->getPrimitiveSizeInBits().getFixedSize() < | ||||||
| 3278 | MulC->getType()->getPrimitiveSizeInBits().getFixedSize()) | ||||||
| 3279 | Op1C = ConstantExpr::getZExt(Op1C, MulC->getType()); | ||||||
| 3280 | if (Op1C->getType()->getPrimitiveSizeInBits().getFixedSize() > | ||||||
| 3281 | MulC->getType()->getPrimitiveSizeInBits().getFixedSize()) | ||||||
| 3282 | MulC = ConstantExpr::getZExt(MulC, Op1C->getType()); | ||||||
| 3283 | |||||||
| 3284 | // V == Base * (Mul0 * Op1), so return (Mul0 * Op1) | ||||||
| 3285 | Multiple = ConstantExpr::getMul(MulC, Op1C); | ||||||
| 3286 | return true; | ||||||
| 3287 | } | ||||||
| 3288 | |||||||
| 3289 | if (ConstantInt *Mul0CI = dyn_cast<ConstantInt>(Mul0)) | ||||||
| 3290 | if (Mul0CI->getValue() == 1) { | ||||||
| 3291 | // V == Base * Op1, so return Op1 | ||||||
| 3292 | Multiple = Op1; | ||||||
| 3293 | return true; | ||||||
| 3294 | } | ||||||
| 3295 | } | ||||||
| 3296 | |||||||
| 3297 | Value *Mul1 = nullptr; | ||||||
| 3298 | if (ComputeMultiple(Op1, Base, Mul1, LookThroughSExt, Depth+1)) { | ||||||
| 3299 | if (Constant *Op0C = dyn_cast<Constant>(Op0)) | ||||||
| 3300 | if (Constant *MulC = dyn_cast<Constant>(Mul1)) { | ||||||
| 3301 | if (Op0C->getType()->getPrimitiveSizeInBits().getFixedSize() < | ||||||
| 3302 | MulC->getType()->getPrimitiveSizeInBits().getFixedSize()) | ||||||
| 3303 | Op0C = ConstantExpr::getZExt(Op0C, MulC->getType()); | ||||||
| 3304 | if (Op0C->getType()->getPrimitiveSizeInBits().getFixedSize() > | ||||||
| 3305 | MulC->getType()->getPrimitiveSizeInBits().getFixedSize()) | ||||||
| 3306 | MulC = ConstantExpr::getZExt(MulC, Op0C->getType()); | ||||||
| 3307 | |||||||
| 3308 | // V == Base * (Mul1 * Op0), so return (Mul1 * Op0) | ||||||
| 3309 | Multiple = ConstantExpr::getMul(MulC, Op0C); | ||||||
| 3310 | return true; | ||||||
| 3311 | } | ||||||
| 3312 | |||||||
| 3313 | if (ConstantInt *Mul1CI = dyn_cast<ConstantInt>(Mul1)) | ||||||
| 3314 | if (Mul1CI->getValue() == 1) { | ||||||
| 3315 | // V == Base * Op0, so return Op0 | ||||||
| 3316 | Multiple = Op0; | ||||||
| 3317 | return true; | ||||||
| 3318 | } | ||||||
| 3319 | } | ||||||
| 3320 | } | ||||||
| 3321 | } | ||||||
| 3322 | |||||||
| 3323 | // We could not determine if V is a multiple of Base. | ||||||
| 3324 | return false; | ||||||
| 3325 | } | ||||||
| 3326 | |||||||
| 3327 | Intrinsic::ID llvm::getIntrinsicForCallSite(const CallBase &CB, | ||||||
| 3328 | const TargetLibraryInfo *TLI) { | ||||||
| 3329 | const Function *F = CB.getCalledFunction(); | ||||||
| 3330 | if (!F) | ||||||
| 3331 | return Intrinsic::not_intrinsic; | ||||||
| 3332 | |||||||
| 3333 | if (F->isIntrinsic()) | ||||||
| 3334 | return F->getIntrinsicID(); | ||||||
| 3335 | |||||||
| 3336 | // We are going to infer semantics of a library function based on mapping it | ||||||
| 3337 | // to an LLVM intrinsic. Check that the library function is available from | ||||||
| 3338 | // this callbase and in this environment. | ||||||
| 3339 | LibFunc Func; | ||||||
| 3340 | if (F->hasLocalLinkage() || !TLI || !TLI->getLibFunc(CB, Func) || | ||||||
| 3341 | !CB.onlyReadsMemory()) | ||||||
| 3342 | return Intrinsic::not_intrinsic; | ||||||
| 3343 | |||||||
| 3344 | switch (Func) { | ||||||
| 3345 | default: | ||||||
| 3346 | break; | ||||||
| 3347 | case LibFunc_sin: | ||||||
| 3348 | case LibFunc_sinf: | ||||||
| 3349 | case LibFunc_sinl: | ||||||
| 3350 | return Intrinsic::sin; | ||||||
| 3351 | case LibFunc_cos: | ||||||
| 3352 | case LibFunc_cosf: | ||||||
| 3353 | case LibFunc_cosl: | ||||||
| 3354 | return Intrinsic::cos; | ||||||
| 3355 | case LibFunc_exp: | ||||||
| 3356 | case LibFunc_expf: | ||||||
| 3357 | case LibFunc_expl: | ||||||
| 3358 | return Intrinsic::exp; | ||||||
| 3359 | case LibFunc_exp2: | ||||||
| 3360 | case LibFunc_exp2f: | ||||||
| 3361 | case LibFunc_exp2l: | ||||||
| 3362 | return Intrinsic::exp2; | ||||||
| 3363 | case LibFunc_log: | ||||||
| 3364 | case LibFunc_logf: | ||||||
| 3365 | case LibFunc_logl: | ||||||
| 3366 | return Intrinsic::log; | ||||||
| 3367 | case LibFunc_log10: | ||||||
| 3368 | case LibFunc_log10f: | ||||||
| 3369 | case LibFunc_log10l: | ||||||
| 3370 | return Intrinsic::log10; | ||||||
| 3371 | case LibFunc_log2: | ||||||
| 3372 | case LibFunc_log2f: | ||||||
| 3373 | case LibFunc_log2l: | ||||||
| 3374 | return Intrinsic::log2; | ||||||
| 3375 | case LibFunc_fabs: | ||||||
| 3376 | case LibFunc_fabsf: | ||||||
| 3377 | case LibFunc_fabsl: | ||||||
| 3378 | return Intrinsic::fabs; | ||||||
| 3379 | case LibFunc_fmin: | ||||||
| 3380 | case LibFunc_fminf: | ||||||
| 3381 | case LibFunc_fminl: | ||||||
| 3382 | return Intrinsic::minnum; | ||||||
| 3383 | case LibFunc_fmax: | ||||||
| 3384 | case LibFunc_fmaxf: | ||||||
| 3385 | case LibFunc_fmaxl: | ||||||
| 3386 | return Intrinsic::maxnum; | ||||||
| 3387 | case LibFunc_copysign: | ||||||
| 3388 | case LibFunc_copysignf: | ||||||
| 3389 | case LibFunc_copysignl: | ||||||
| 3390 | return Intrinsic::copysign; | ||||||
| 3391 | case LibFunc_floor: | ||||||
| 3392 | case LibFunc_floorf: | ||||||
| 3393 | case LibFunc_floorl: | ||||||
| 3394 | return Intrinsic::floor; | ||||||
| 3395 | case LibFunc_ceil: | ||||||
| 3396 | case LibFunc_ceilf: | ||||||
| 3397 | case LibFunc_ceill: | ||||||
| 3398 | return Intrinsic::ceil; | ||||||
| 3399 | case LibFunc_trunc: | ||||||
| 3400 | case LibFunc_truncf: | ||||||
| 3401 | case LibFunc_truncl: | ||||||
| 3402 | return Intrinsic::trunc; | ||||||
| 3403 | case LibFunc_rint: | ||||||
| 3404 | case LibFunc_rintf: | ||||||
| 3405 | case LibFunc_rintl: | ||||||
| 3406 | return Intrinsic::rint; | ||||||
| 3407 | case LibFunc_nearbyint: | ||||||
| 3408 | case LibFunc_nearbyintf: | ||||||
| 3409 | case LibFunc_nearbyintl: | ||||||
| 3410 | return Intrinsic::nearbyint; | ||||||
| 3411 | case LibFunc_round: | ||||||
| 3412 | case LibFunc_roundf: | ||||||
| 3413 | case LibFunc_roundl: | ||||||
| 3414 | return Intrinsic::round; | ||||||
| 3415 | case LibFunc_roundeven: | ||||||
| 3416 | case LibFunc_roundevenf: | ||||||
| 3417 | case LibFunc_roundevenl: | ||||||
| 3418 | return Intrinsic::roundeven; | ||||||
| 3419 | case LibFunc_pow: | ||||||
| 3420 | case LibFunc_powf: | ||||||
| 3421 | case LibFunc_powl: | ||||||
| 3422 | return Intrinsic::pow; | ||||||
| 3423 | case LibFunc_sqrt: | ||||||
| 3424 | case LibFunc_sqrtf: | ||||||
| 3425 | case LibFunc_sqrtl: | ||||||
| 3426 | return Intrinsic::sqrt; | ||||||
| 3427 | } | ||||||
| 3428 | |||||||
| 3429 | return Intrinsic::not_intrinsic; | ||||||
| 3430 | } | ||||||
| 3431 | |||||||
| 3432 | /// Return true if we can prove that the specified FP value is never equal to | ||||||
| 3433 | /// -0.0. | ||||||
| 3434 | /// NOTE: Do not check 'nsz' here because that fast-math-flag does not guarantee | ||||||
| 3435 | /// that a value is not -0.0. It only guarantees that -0.0 may be treated | ||||||
| 3436 | /// the same as +0.0 in floating-point ops. | ||||||
| 3437 | /// | ||||||
| 3438 | /// NOTE: this function will need to be revisited when we support non-default | ||||||
| 3439 | /// rounding modes! | ||||||
| 3440 | bool llvm::CannotBeNegativeZero(const Value *V, const TargetLibraryInfo *TLI, | ||||||
| 3441 | unsigned Depth) { | ||||||
| 3442 | if (auto *CFP = dyn_cast<ConstantFP>(V)) | ||||||
| 3443 | return !CFP->getValueAPF().isNegZero(); | ||||||
| 3444 | |||||||
| 3445 | if (Depth == MaxAnalysisRecursionDepth) | ||||||
| 3446 | return false; | ||||||
| 3447 | |||||||
| 3448 | auto *Op = dyn_cast<Operator>(V); | ||||||
| 3449 | if (!Op) | ||||||
| 3450 | return false; | ||||||
| 3451 | |||||||
| 3452 | // (fadd x, 0.0) is guaranteed to return +0.0, not -0.0. | ||||||
| 3453 | if (match(Op, m_FAdd(m_Value(), m_PosZeroFP()))) | ||||||
| 3454 | return true; | ||||||
| 3455 | |||||||
| 3456 | // sitofp and uitofp turn into +0.0 for zero. | ||||||
| 3457 | if (isa<SIToFPInst>(Op) || isa<UIToFPInst>(Op)) | ||||||
| 3458 | return true; | ||||||
| 3459 | |||||||
| 3460 | if (auto *Call = dyn_cast<CallInst>(Op)) { | ||||||
| 3461 | Intrinsic::ID IID = getIntrinsicForCallSite(*Call, TLI); | ||||||
| 3462 | switch (IID) { | ||||||
| 3463 | default: | ||||||
| 3464 | break; | ||||||
| 3465 | // sqrt(-0.0) = -0.0, no other negative results are possible. | ||||||
| 3466 | case Intrinsic::sqrt: | ||||||
| 3467 | case Intrinsic::canonicalize: | ||||||
| 3468 | return CannotBeNegativeZero(Call->getArgOperand(0), TLI, Depth + 1); | ||||||
| 3469 | // fabs(x) != -0.0 | ||||||
| 3470 | case Intrinsic::fabs: | ||||||
| 3471 | return true; | ||||||
| 3472 | } | ||||||
| 3473 | } | ||||||
| 3474 | |||||||
| 3475 | return false; | ||||||
| 3476 | } | ||||||
| 3477 | |||||||
| 3478 | /// If \p SignBitOnly is true, test for a known 0 sign bit rather than a | ||||||
| 3479 | /// standard ordered compare. e.g. make -0.0 olt 0.0 be true because of the sign | ||||||
| 3480 | /// bit despite comparing equal. | ||||||
| 3481 | static bool cannotBeOrderedLessThanZeroImpl(const Value *V, | ||||||
| 3482 | const TargetLibraryInfo *TLI, | ||||||
| 3483 | bool SignBitOnly, | ||||||
| 3484 | unsigned Depth) { | ||||||
| 3485 | // TODO: This function does not do the right thing when SignBitOnly is true | ||||||
| 3486 | // and we're lowering to a hypothetical IEEE 754-compliant-but-evil platform | ||||||
| 3487 | // which flips the sign bits of NaNs. See | ||||||
| 3488 | // https://llvm.org/bugs/show_bug.cgi?id=31702. | ||||||
| 3489 | |||||||
| 3490 | if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V)) { | ||||||
| 3491 | return !CFP->getValueAPF().isNegative() || | ||||||
| 3492 | (!SignBitOnly && CFP->getValueAPF().isZero()); | ||||||
| 3493 | } | ||||||
| 3494 | |||||||
| 3495 | // Handle vector of constants. | ||||||
| 3496 | if (auto *CV = dyn_cast<Constant>(V)) { | ||||||
| 3497 | if (auto *CVFVTy = dyn_cast<FixedVectorType>(CV->getType())) { | ||||||
| 3498 | unsigned NumElts = CVFVTy->getNumElements(); | ||||||
| 3499 | for (unsigned i = 0; i != NumElts; ++i) { | ||||||
| 3500 | auto *CFP = dyn_cast_or_null<ConstantFP>(CV->getAggregateElement(i)); | ||||||
| 3501 | if (!CFP) | ||||||
| 3502 | return false; | ||||||
| 3503 | if (CFP->getValueAPF().isNegative() && | ||||||
| 3504 | (SignBitOnly || !CFP->getValueAPF().isZero())) | ||||||
| 3505 | return false; | ||||||
| 3506 | } | ||||||
| 3507 | |||||||
| 3508 | // All non-negative ConstantFPs. | ||||||
| 3509 | return true; | ||||||
| 3510 | } | ||||||
| 3511 | } | ||||||
| 3512 | |||||||
| 3513 | if (Depth == MaxAnalysisRecursionDepth) | ||||||
| 3514 | return false; | ||||||
| 3515 | |||||||
| 3516 | const Operator *I = dyn_cast<Operator>(V); | ||||||
| 3517 | if (!I) | ||||||
| 3518 | return false; | ||||||
| 3519 | |||||||
| 3520 | switch (I->getOpcode()) { | ||||||
| 3521 | default: | ||||||
| 3522 | break; | ||||||
| 3523 | // Unsigned integers are always nonnegative. | ||||||
| 3524 | case Instruction::UIToFP: | ||||||
| 3525 | return true; | ||||||
| 3526 | case Instruction::FMul: | ||||||
| 3527 | case Instruction::FDiv: | ||||||
| 3528 | // X * X is always non-negative or a NaN. | ||||||
| 3529 | // X / X is always exactly 1.0 or a NaN. | ||||||
| 3530 | if (I->getOperand(0) == I->getOperand(1) && | ||||||
| 3531 | (!SignBitOnly || cast<FPMathOperator>(I)->hasNoNaNs())) | ||||||
| 3532 | return true; | ||||||
| 3533 | |||||||
| 3534 | LLVM_FALLTHROUGH[[gnu::fallthrough]]; | ||||||
| 3535 | case Instruction::FAdd: | ||||||
| 3536 | case Instruction::FRem: | ||||||
| 3537 | return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly, | ||||||
| 3538 | Depth + 1) && | ||||||
| 3539 | cannotBeOrderedLessThanZeroImpl(I->getOperand(1), TLI, SignBitOnly, | ||||||
| 3540 | Depth + 1); | ||||||
| 3541 | case Instruction::Select: | ||||||
| 3542 | return cannotBeOrderedLessThanZeroImpl(I->getOperand(1), TLI, SignBitOnly, | ||||||
| 3543 | Depth + 1) && | ||||||
| 3544 | cannotBeOrderedLessThanZeroImpl(I->getOperand(2), TLI, SignBitOnly, | ||||||
| 3545 | Depth + 1); | ||||||
| 3546 | case Instruction::FPExt: | ||||||
| 3547 | case Instruction::FPTrunc: | ||||||
| 3548 | // Widening/narrowing never change sign. | ||||||
| 3549 | return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly, | ||||||
| 3550 | Depth + 1); | ||||||
| 3551 | case Instruction::ExtractElement: | ||||||
| 3552 | // Look through extract element. At the moment we keep this simple and skip | ||||||
| 3553 | // tracking the specific element. But at least we might find information | ||||||
| 3554 | // valid for all elements of the vector. | ||||||
| 3555 | return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly, | ||||||
| 3556 | Depth + 1); | ||||||
| 3557 | case Instruction::Call: | ||||||
| 3558 | const auto *CI = cast<CallInst>(I); | ||||||
| 3559 | Intrinsic::ID IID = getIntrinsicForCallSite(*CI, TLI); | ||||||
| 3560 | switch (IID) { | ||||||
| 3561 | default: | ||||||
| 3562 | break; | ||||||
| 3563 | case Intrinsic::maxnum: { | ||||||
| 3564 | Value *V0 = I->getOperand(0), *V1 = I->getOperand(1); | ||||||
| 3565 | auto isPositiveNum = [&](Value *V) { | ||||||
| 3566 | if (SignBitOnly) { | ||||||
| 3567 | // With SignBitOnly, this is tricky because the result of | ||||||
| 3568 | // maxnum(+0.0, -0.0) is unspecified. Just check if the operand is | ||||||
| 3569 | // a constant strictly greater than 0.0. | ||||||
| 3570 | const APFloat *C; | ||||||
| 3571 | return match(V, m_APFloat(C)) && | ||||||
| 3572 | *C > APFloat::getZero(C->getSemantics()); | ||||||
| 3573 | } | ||||||
| 3574 | |||||||
| 3575 | // -0.0 compares equal to 0.0, so if this operand is at least -0.0, | ||||||
| 3576 | // maxnum can't be ordered-less-than-zero. | ||||||
| 3577 | return isKnownNeverNaN(V, TLI) && | ||||||
| 3578 | cannotBeOrderedLessThanZeroImpl(V, TLI, false, Depth + 1); | ||||||
| 3579 | }; | ||||||
| 3580 | |||||||
| 3581 | // TODO: This could be improved. We could also check that neither operand | ||||||
| 3582 | // has its sign bit set (and at least 1 is not-NAN?). | ||||||
| 3583 | return isPositiveNum(V0) || isPositiveNum(V1); | ||||||
| 3584 | } | ||||||
| 3585 | |||||||
| 3586 | case Intrinsic::maximum: | ||||||
| 3587 | return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly, | ||||||
| 3588 | Depth + 1) || | ||||||
| 3589 | cannotBeOrderedLessThanZeroImpl(I->getOperand(1), TLI, SignBitOnly, | ||||||
| 3590 | Depth + 1); | ||||||
| 3591 | case Intrinsic::minnum: | ||||||
| 3592 | case Intrinsic::minimum: | ||||||
| 3593 | return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly, | ||||||
| 3594 | Depth + 1) && | ||||||
| 3595 | cannotBeOrderedLessThanZeroImpl(I->getOperand(1), TLI, SignBitOnly, | ||||||
| 3596 | Depth + 1); | ||||||
| 3597 | case Intrinsic::exp: | ||||||
| 3598 | case Intrinsic::exp2: | ||||||
| 3599 | case Intrinsic::fabs: | ||||||
| 3600 | return true; | ||||||
| 3601 | |||||||
| 3602 | case Intrinsic::sqrt: | ||||||
| 3603 | // sqrt(x) is always >= -0 or NaN. Moreover, sqrt(x) == -0 iff x == -0. | ||||||
| 3604 | if (!SignBitOnly) | ||||||
| 3605 | return true; | ||||||
| 3606 | return CI->hasNoNaNs() && (CI->hasNoSignedZeros() || | ||||||
| 3607 | CannotBeNegativeZero(CI->getOperand(0), TLI)); | ||||||
| 3608 | |||||||
| 3609 | case Intrinsic::powi: | ||||||
| 3610 | if (ConstantInt *Exponent = dyn_cast<ConstantInt>(I->getOperand(1))) { | ||||||
| 3611 | // powi(x,n) is non-negative if n is even. | ||||||
| 3612 | if (Exponent->getBitWidth() <= 64 && Exponent->getSExtValue() % 2u == 0) | ||||||
| 3613 | return true; | ||||||
| 3614 | } | ||||||
| 3615 | // TODO: This is not correct. Given that exp is an integer, here are the | ||||||
| 3616 | // ways that pow can return a negative value: | ||||||
| 3617 | // | ||||||
| 3618 | // pow(x, exp) --> negative if exp is odd and x is negative. | ||||||
| 3619 | // pow(-0, exp) --> -inf if exp is negative odd. | ||||||
| 3620 | // pow(-0, exp) --> -0 if exp is positive odd. | ||||||
| 3621 | // pow(-inf, exp) --> -0 if exp is negative odd. | ||||||
| 3622 | // pow(-inf, exp) --> -inf if exp is positive odd. | ||||||
| 3623 | // | ||||||
| 3624 | // Therefore, if !SignBitOnly, we can return true if x >= +0 or x is NaN, | ||||||
| 3625 | // but we must return false if x == -0. Unfortunately we do not currently | ||||||
| 3626 | // have a way of expressing this constraint. See details in | ||||||
| 3627 | // https://llvm.org/bugs/show_bug.cgi?id=31702. | ||||||
| 3628 | return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly, | ||||||
| 3629 | Depth + 1); | ||||||
| 3630 | |||||||
| 3631 | case Intrinsic::fma: | ||||||
| 3632 | case Intrinsic::fmuladd: | ||||||
| 3633 | // x*x+y is non-negative if y is non-negative. | ||||||
| 3634 | return I->getOperand(0) == I->getOperand(1) && | ||||||
| 3635 | (!SignBitOnly || cast<FPMathOperator>(I)->hasNoNaNs()) && | ||||||
| 3636 | cannotBeOrderedLessThanZeroImpl(I->getOperand(2), TLI, SignBitOnly, | ||||||
| 3637 | Depth + 1); | ||||||
| 3638 | } | ||||||
| 3639 | break; | ||||||
| 3640 | } | ||||||
| 3641 | return false; | ||||||
| 3642 | } | ||||||
| 3643 | |||||||
| 3644 | bool llvm::CannotBeOrderedLessThanZero(const Value *V, | ||||||
| 3645 | const TargetLibraryInfo *TLI) { | ||||||
| 3646 | return cannotBeOrderedLessThanZeroImpl(V, TLI, false, 0); | ||||||
| 3647 | } | ||||||
| 3648 | |||||||
| 3649 | bool llvm::SignBitMustBeZero(const Value *V, const TargetLibraryInfo *TLI) { | ||||||
| 3650 | return cannotBeOrderedLessThanZeroImpl(V, TLI, true, 0); | ||||||
| 3651 | } | ||||||
| 3652 | |||||||
| 3653 | bool llvm::isKnownNeverInfinity(const Value *V, const TargetLibraryInfo *TLI, | ||||||
| 3654 | unsigned Depth) { | ||||||
| 3655 | assert(V->getType()->isFPOrFPVectorTy() && "Querying for Inf on non-FP type")((void)0); | ||||||
| 3656 | |||||||
| 3657 | // If we're told that infinities won't happen, assume they won't. | ||||||
| 3658 | if (auto *FPMathOp = dyn_cast<FPMathOperator>(V)) | ||||||
| 3659 | if (FPMathOp->hasNoInfs()) | ||||||
| 3660 | return true; | ||||||
| 3661 | |||||||
| 3662 | // Handle scalar constants. | ||||||
| 3663 | if (auto *CFP = dyn_cast<ConstantFP>(V)) | ||||||
| 3664 | return !CFP->isInfinity(); | ||||||
| 3665 | |||||||
| 3666 | if (Depth == MaxAnalysisRecursionDepth) | ||||||
| 3667 | return false; | ||||||
| 3668 | |||||||
| 3669 | if (auto *Inst = dyn_cast<Instruction>(V)) { | ||||||
| 3670 | switch (Inst->getOpcode()) { | ||||||
| 3671 | case Instruction::Select: { | ||||||
| 3672 | return isKnownNeverInfinity(Inst->getOperand(1), TLI, Depth + 1) && | ||||||
| 3673 | isKnownNeverInfinity(Inst->getOperand(2), TLI, Depth + 1); | ||||||
| 3674 | } | ||||||
| 3675 | case Instruction::SIToFP: | ||||||
| 3676 | case Instruction::UIToFP: { | ||||||
| 3677 | // Get width of largest magnitude integer (remove a bit if signed). | ||||||
| 3678 | // This still works for a signed minimum value because the largest FP | ||||||
| 3679 | // value is scaled by some fraction close to 2.0 (1.0 + 0.xxxx). | ||||||
| 3680 | int IntSize = Inst->getOperand(0)->getType()->getScalarSizeInBits(); | ||||||
| 3681 | if (Inst->getOpcode() == Instruction::SIToFP) | ||||||
| 3682 | --IntSize; | ||||||
| 3683 | |||||||
| 3684 | // If the exponent of the largest finite FP value can hold the largest | ||||||
| 3685 | // integer, the result of the cast must be finite. | ||||||
| 3686 | Type *FPTy = Inst->getType()->getScalarType(); | ||||||
| 3687 | return ilogb(APFloat::getLargest(FPTy->getFltSemantics())) >= IntSize; | ||||||
| 3688 | } | ||||||
| 3689 | default: | ||||||
| 3690 | break; | ||||||
| 3691 | } | ||||||
| 3692 | } | ||||||
| 3693 | |||||||
| 3694 | // try to handle fixed width vector constants | ||||||
| 3695 | auto *VFVTy = dyn_cast<FixedVectorType>(V->getType()); | ||||||
| 3696 | if (VFVTy && isa<Constant>(V)) { | ||||||
| 3697 | // For vectors, verify that each element is not infinity. | ||||||
| 3698 | unsigned NumElts = VFVTy->getNumElements(); | ||||||
| 3699 | for (unsigned i = 0; i != NumElts; ++i) { | ||||||
| 3700 | Constant *Elt = cast<Constant>(V)->getAggregateElement(i); | ||||||
| 3701 | if (!Elt) | ||||||
| 3702 | return false; | ||||||
| 3703 | if (isa<UndefValue>(Elt)) | ||||||
| 3704 | continue; | ||||||
| 3705 | auto *CElt = dyn_cast<ConstantFP>(Elt); | ||||||
| 3706 | if (!CElt || CElt->isInfinity()) | ||||||
| 3707 | return false; | ||||||
| 3708 | } | ||||||
| 3709 | // All elements were confirmed non-infinity or undefined. | ||||||
| 3710 | return true; | ||||||
| 3711 | } | ||||||
| 3712 | |||||||
| 3713 | // was not able to prove that V never contains infinity | ||||||
| 3714 | return false; | ||||||
| 3715 | } | ||||||
| 3716 | |||||||
| 3717 | bool llvm::isKnownNeverNaN(const Value *V, const TargetLibraryInfo *TLI, | ||||||
| 3718 | unsigned Depth) { | ||||||
| 3719 | assert(V->getType()->isFPOrFPVectorTy() && "Querying for NaN on non-FP type")((void)0); | ||||||
| 3720 | |||||||
| 3721 | // If we're told that NaNs won't happen, assume they won't. | ||||||
| 3722 | if (auto *FPMathOp = dyn_cast<FPMathOperator>(V)) | ||||||
| 3723 | if (FPMathOp->hasNoNaNs()) | ||||||
| 3724 | return true; | ||||||
| 3725 | |||||||
| 3726 | // Handle scalar constants. | ||||||
| 3727 | if (auto *CFP = dyn_cast<ConstantFP>(V)) | ||||||
| 3728 | return !CFP->isNaN(); | ||||||
| 3729 | |||||||
| 3730 | if (Depth == MaxAnalysisRecursionDepth) | ||||||
| 3731 | return false; | ||||||
| 3732 | |||||||
| 3733 | if (auto *Inst = dyn_cast<Instruction>(V)) { | ||||||
| 3734 | switch (Inst->getOpcode()) { | ||||||
| 3735 | case Instruction::FAdd: | ||||||
| 3736 | case Instruction::FSub: | ||||||
| 3737 | // Adding positive and negative infinity produces NaN. | ||||||
| 3738 | return isKnownNeverNaN(Inst->getOperand(0), TLI, Depth + 1) && | ||||||
| 3739 | isKnownNeverNaN(Inst->getOperand(1), TLI, Depth + 1) && | ||||||
| 3740 | (isKnownNeverInfinity(Inst->getOperand(0), TLI, Depth + 1) || | ||||||
| 3741 | isKnownNeverInfinity(Inst->getOperand(1), TLI, Depth + 1)); | ||||||
| 3742 | |||||||
| 3743 | case Instruction::FMul: | ||||||
| 3744 | // Zero multiplied with infinity produces NaN. | ||||||
| 3745 | // FIXME: If neither side can be zero fmul never produces NaN. | ||||||
| 3746 | return isKnownNeverNaN(Inst->getOperand(0), TLI, Depth + 1) && | ||||||
| 3747 | isKnownNeverInfinity(Inst->getOperand(0), TLI, Depth + 1) && | ||||||
| 3748 | isKnownNeverNaN(Inst->getOperand(1), TLI, Depth + 1) && | ||||||
| 3749 | isKnownNeverInfinity(Inst->getOperand(1), TLI, Depth + 1); | ||||||
| 3750 | |||||||
| 3751 | case Instruction::FDiv: | ||||||
| 3752 | case Instruction::FRem: | ||||||
| 3753 | // FIXME: Only 0/0, Inf/Inf, Inf REM x and x REM 0 produce NaN. | ||||||
| 3754 | return false; | ||||||
| 3755 | |||||||
| 3756 | case Instruction::Select: { | ||||||
| 3757 | return isKnownNeverNaN(Inst->getOperand(1), TLI, Depth + 1) && | ||||||
| 3758 | isKnownNeverNaN(Inst->getOperand(2), TLI, Depth + 1); | ||||||
| 3759 | } | ||||||
| 3760 | case Instruction::SIToFP: | ||||||
| 3761 | case Instruction::UIToFP: | ||||||
| 3762 | return true; | ||||||
| 3763 | case Instruction::FPTrunc: | ||||||
| 3764 | case Instruction::FPExt: | ||||||
| 3765 | return isKnownNeverNaN(Inst->getOperand(0), TLI, Depth + 1); | ||||||
| 3766 | default: | ||||||
| 3767 | break; | ||||||
| 3768 | } | ||||||
| 3769 | } | ||||||
| 3770 | |||||||
| 3771 | if (const auto *II = dyn_cast<IntrinsicInst>(V)) { | ||||||
| 3772 | switch (II->getIntrinsicID()) { | ||||||
| 3773 | case Intrinsic::canonicalize: | ||||||
| 3774 | case Intrinsic::fabs: | ||||||
| 3775 | case Intrinsic::copysign: | ||||||
| 3776 | case Intrinsic::exp: | ||||||
| 3777 | case Intrinsic::exp2: | ||||||
| 3778 | case Intrinsic::floor: | ||||||
| 3779 | case Intrinsic::ceil: | ||||||
| 3780 | case Intrinsic::trunc: | ||||||
| 3781 | case Intrinsic::rint: | ||||||
| 3782 | case Intrinsic::nearbyint: | ||||||
| 3783 | case Intrinsic::round: | ||||||
| 3784 | case Intrinsic::roundeven: | ||||||
| 3785 | return isKnownNeverNaN(II->getArgOperand(0), TLI, Depth + 1); | ||||||
| 3786 | case Intrinsic::sqrt: | ||||||
| 3787 | return isKnownNeverNaN(II->getArgOperand(0), TLI, Depth + 1) && | ||||||
| 3788 | CannotBeOrderedLessThanZero(II->getArgOperand(0), TLI); | ||||||
| 3789 | case Intrinsic::minnum: | ||||||
| 3790 | case Intrinsic::maxnum: | ||||||
| 3791 | // If either operand is not NaN, the result is not NaN. | ||||||
| 3792 | return isKnownNeverNaN(II->getArgOperand(0), TLI, Depth + 1) || | ||||||
| 3793 | isKnownNeverNaN(II->getArgOperand(1), TLI, Depth + 1); | ||||||
| 3794 | default: | ||||||
| 3795 | return false; | ||||||
| 3796 | } | ||||||
| 3797 | } | ||||||
| 3798 | |||||||
| 3799 | // Try to handle fixed width vector constants | ||||||
| 3800 | auto *VFVTy = dyn_cast<FixedVectorType>(V->getType()); | ||||||
| 3801 | if (VFVTy && isa<Constant>(V)) { | ||||||
| 3802 | // For vectors, verify that each element is not NaN. | ||||||
| 3803 | unsigned NumElts = VFVTy->getNumElements(); | ||||||
| 3804 | for (unsigned i = 0; i != NumElts; ++i) { | ||||||
| 3805 | Constant *Elt = cast<Constant>(V)->getAggregateElement(i); | ||||||
| 3806 | if (!Elt) | ||||||
| 3807 | return false; | ||||||
| 3808 | if (isa<UndefValue>(Elt)) | ||||||
| 3809 | continue; | ||||||
| 3810 | auto *CElt = dyn_cast<ConstantFP>(Elt); | ||||||
| 3811 | if (!CElt || CElt->isNaN()) | ||||||
| 3812 | return false; | ||||||
| 3813 | } | ||||||
| 3814 | // All elements were confirmed not-NaN or undefined. | ||||||
| 3815 | return true; | ||||||
| 3816 | } | ||||||
| 3817 | |||||||
| 3818 | // Was not able to prove that V never contains NaN | ||||||
| 3819 | return false; | ||||||
| 3820 | } | ||||||
| 3821 | |||||||
| 3822 | Value *llvm::isBytewiseValue(Value *V, const DataLayout &DL) { | ||||||
| 3823 | |||||||
| 3824 | // All byte-wide stores are splatable, even of arbitrary variables. | ||||||
| 3825 | if (V->getType()->isIntegerTy(8)) | ||||||
| 3826 | return V; | ||||||
| 3827 | |||||||
| 3828 | LLVMContext &Ctx = V->getContext(); | ||||||
| 3829 | |||||||
| 3830 | // Undef don't care. | ||||||
| 3831 | auto *UndefInt8 = UndefValue::get(Type::getInt8Ty(Ctx)); | ||||||
| 3832 | if (isa<UndefValue>(V)) | ||||||
| 3833 | return UndefInt8; | ||||||
| 3834 | |||||||
| 3835 | // Return Undef for zero-sized type. | ||||||
| 3836 | if (!DL.getTypeStoreSize(V->getType()).isNonZero()) | ||||||
| 3837 | return UndefInt8; | ||||||
| 3838 | |||||||
| 3839 | Constant *C = dyn_cast<Constant>(V); | ||||||
| 3840 | if (!C) { | ||||||
| 3841 | // Conceptually, we could handle things like: | ||||||
| 3842 | // %a = zext i8 %X to i16 | ||||||
| 3843 | // %b = shl i16 %a, 8 | ||||||
| 3844 | // %c = or i16 %a, %b | ||||||
| 3845 | // but until there is an example that actually needs this, it doesn't seem | ||||||
| 3846 | // worth worrying about. | ||||||
| 3847 | return nullptr; | ||||||
| 3848 | } | ||||||
| 3849 | |||||||
| 3850 | // Handle 'null' ConstantArrayZero etc. | ||||||
| 3851 | if (C->isNullValue()) | ||||||
| 3852 | return Constant::getNullValue(Type::getInt8Ty(Ctx)); | ||||||
| 3853 | |||||||
| 3854 | // Constant floating-point values can be handled as integer values if the | ||||||
| 3855 | // corresponding integer value is "byteable". An important case is 0.0. | ||||||
| 3856 | if (ConstantFP *CFP = dyn_cast<ConstantFP>(C)) { | ||||||
| 3857 | Type *Ty = nullptr; | ||||||
| 3858 | if (CFP->getType()->isHalfTy()) | ||||||
| 3859 | Ty = Type::getInt16Ty(Ctx); | ||||||
| 3860 | else if (CFP->getType()->isFloatTy()) | ||||||
| 3861 | Ty = Type::getInt32Ty(Ctx); | ||||||
| 3862 | else if (CFP->getType()->isDoubleTy()) | ||||||
| 3863 | Ty = Type::getInt64Ty(Ctx); | ||||||
| 3864 | // Don't handle long double formats, which have strange constraints. | ||||||
| 3865 | return Ty ? isBytewiseValue(ConstantExpr::getBitCast(CFP, Ty), DL) | ||||||
| 3866 | : nullptr; | ||||||
| 3867 | } | ||||||
| 3868 | |||||||
| 3869 | // We can handle constant integers that are multiple of 8 bits. | ||||||
| 3870 | if (ConstantInt *CI = dyn_cast<ConstantInt>(C)) { | ||||||
| 3871 | if (CI->getBitWidth() % 8 == 0) { | ||||||
| 3872 | assert(CI->getBitWidth() > 8 && "8 bits should be handled above!")((void)0); | ||||||
| 3873 | if (!CI->getValue().isSplat(8)) | ||||||
| 3874 | return nullptr; | ||||||
| 3875 | return ConstantInt::get(Ctx, CI->getValue().trunc(8)); | ||||||
| 3876 | } | ||||||
| 3877 | } | ||||||
| 3878 | |||||||
| 3879 | if (auto *CE = dyn_cast<ConstantExpr>(C)) { | ||||||
| 3880 | if (CE->getOpcode() == Instruction::IntToPtr) { | ||||||
| 3881 | if (auto *PtrTy = dyn_cast<PointerType>(CE->getType())) { | ||||||
| 3882 | unsigned BitWidth = DL.getPointerSizeInBits(PtrTy->getAddressSpace()); | ||||||
| 3883 | return isBytewiseValue( | ||||||
| 3884 | ConstantExpr::getIntegerCast(CE->getOperand(0), | ||||||
| 3885 | Type::getIntNTy(Ctx, BitWidth), false), | ||||||
| 3886 | DL); | ||||||
| 3887 | } | ||||||
| 3888 | } | ||||||
| 3889 | } | ||||||
| 3890 | |||||||
| 3891 | auto Merge = [&](Value *LHS, Value *RHS) -> Value * { | ||||||
| 3892 | if (LHS == RHS) | ||||||
| 3893 | return LHS; | ||||||
| 3894 | if (!LHS || !RHS) | ||||||
| 3895 | return nullptr; | ||||||
| 3896 | if (LHS == UndefInt8) | ||||||
| 3897 | return RHS; | ||||||
| 3898 | if (RHS == UndefInt8) | ||||||
| 3899 | return LHS; | ||||||
| 3900 | return nullptr; | ||||||
| 3901 | }; | ||||||
| 3902 | |||||||
| 3903 | if (ConstantDataSequential *CA = dyn_cast<ConstantDataSequential>(C)) { | ||||||
| 3904 | Value *Val = UndefInt8; | ||||||
| 3905 | for (unsigned I = 0, E = CA->getNumElements(); I != E; ++I) | ||||||
| 3906 | if (!(Val = Merge(Val, isBytewiseValue(CA->getElementAsConstant(I), DL)))) | ||||||
| 3907 | return nullptr; | ||||||
| 3908 | return Val; | ||||||
| 3909 | } | ||||||
| 3910 | |||||||
| 3911 | if (isa<ConstantAggregate>(C)) { | ||||||
| 3912 | Value *Val = UndefInt8; | ||||||
| 3913 | for (unsigned I = 0, E = C->getNumOperands(); I != E; ++I) | ||||||
| 3914 | if (!(Val = Merge(Val, isBytewiseValue(C->getOperand(I), DL)))) | ||||||
| 3915 | return nullptr; | ||||||
| 3916 | return Val; | ||||||
| 3917 | } | ||||||
| 3918 | |||||||
| 3919 | // Don't try to handle the handful of other constants. | ||||||
| 3920 | return nullptr; | ||||||
| 3921 | } | ||||||
| 3922 | |||||||
| 3923 | // This is the recursive version of BuildSubAggregate. It takes a few different | ||||||
| 3924 | // arguments. Idxs is the index within the nested struct From that we are | ||||||
| 3925 | // looking at now (which is of type IndexedType). IdxSkip is the number of | ||||||
| 3926 | // indices from Idxs that should be left out when inserting into the resulting | ||||||
| 3927 | // struct. To is the result struct built so far, new insertvalue instructions | ||||||
| 3928 | // build on that. | ||||||
| 3929 | static Value *BuildSubAggregate(Value *From, Value* To, Type *IndexedType, | ||||||
| 3930 | SmallVectorImpl<unsigned> &Idxs, | ||||||
| 3931 | unsigned IdxSkip, | ||||||
| 3932 | Instruction *InsertBefore) { | ||||||
| 3933 | StructType *STy = dyn_cast<StructType>(IndexedType); | ||||||
| 3934 | if (STy) { | ||||||
| 3935 | // Save the original To argument so we can modify it | ||||||
| 3936 | Value *OrigTo = To; | ||||||
| 3937 | // General case, the type indexed by Idxs is a struct | ||||||
| 3938 | for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { | ||||||
| 3939 | // Process each struct element recursively | ||||||
| 3940 | Idxs.push_back(i); | ||||||
| 3941 | Value *PrevTo = To; | ||||||
| 3942 | To = BuildSubAggregate(From, To, STy->getElementType(i), Idxs, IdxSkip, | ||||||
| 3943 | InsertBefore); | ||||||
| 3944 | Idxs.pop_back(); | ||||||
| 3945 | if (!To) { | ||||||
| 3946 | // Couldn't find any inserted value for this index? Cleanup | ||||||
| 3947 | while (PrevTo != OrigTo) { | ||||||
| 3948 | InsertValueInst* Del = cast<InsertValueInst>(PrevTo); | ||||||
| 3949 | PrevTo = Del->getAggregateOperand(); | ||||||
| 3950 | Del->eraseFromParent(); | ||||||
| 3951 | } | ||||||
| 3952 | // Stop processing elements | ||||||
| 3953 | break; | ||||||
| 3954 | } | ||||||
| 3955 | } | ||||||
| 3956 | // If we successfully found a value for each of our subaggregates | ||||||
| 3957 | if (To) | ||||||
| 3958 | return To; | ||||||
| 3959 | } | ||||||
| 3960 | // Base case, the type indexed by SourceIdxs is not a struct, or not all of | ||||||
| 3961 | // the struct's elements had a value that was inserted directly. In the latter | ||||||
| 3962 | // case, perhaps we can't determine each of the subelements individually, but | ||||||
| 3963 | // we might be able to find the complete struct somewhere. | ||||||
| 3964 | |||||||
| 3965 | // Find the value that is at that particular spot | ||||||
| 3966 | Value *V = FindInsertedValue(From, Idxs); | ||||||
| 3967 | |||||||
| 3968 | if (!V) | ||||||
| 3969 | return nullptr; | ||||||
| 3970 | |||||||
| 3971 | // Insert the value in the new (sub) aggregate | ||||||
| 3972 | return InsertValueInst::Create(To, V, makeArrayRef(Idxs).slice(IdxSkip), | ||||||
| 3973 | "tmp", InsertBefore); | ||||||
| 3974 | } | ||||||
| 3975 | |||||||
| 3976 | // This helper takes a nested struct and extracts a part of it (which is again a | ||||||
| 3977 | // struct) into a new value. For example, given the struct: | ||||||
| 3978 | // { a, { b, { c, d }, e } } | ||||||
| 3979 | // and the indices "1, 1" this returns | ||||||
| 3980 | // { c, d }. | ||||||
| 3981 | // | ||||||
| 3982 | // It does this by inserting an insertvalue for each element in the resulting | ||||||
| 3983 | // struct, as opposed to just inserting a single struct. This will only work if | ||||||
| 3984 | // each of the elements of the substruct are known (ie, inserted into From by an | ||||||
| 3985 | // insertvalue instruction somewhere). | ||||||
| 3986 | // | ||||||
| 3987 | // All inserted insertvalue instructions are inserted before InsertBefore | ||||||
| 3988 | static Value *BuildSubAggregate(Value *From, ArrayRef<unsigned> idx_range, | ||||||
| 3989 | Instruction *InsertBefore) { | ||||||
| 3990 | assert(InsertBefore && "Must have someplace to insert!")((void)0); | ||||||
| 3991 | Type *IndexedType = ExtractValueInst::getIndexedType(From->getType(), | ||||||
| 3992 | idx_range); | ||||||
| 3993 | Value *To = UndefValue::get(IndexedType); | ||||||
| 3994 | SmallVector<unsigned, 10> Idxs(idx_range.begin(), idx_range.end()); | ||||||
| 3995 | unsigned IdxSkip = Idxs.size(); | ||||||
| 3996 | |||||||
| 3997 | return BuildSubAggregate(From, To, IndexedType, Idxs, IdxSkip, InsertBefore); | ||||||
| 3998 | } | ||||||
| 3999 | |||||||
| 4000 | /// Given an aggregate and a sequence of indices, see if the scalar value | ||||||
| 4001 | /// indexed is already around as a register, for example if it was inserted | ||||||
| 4002 | /// directly into the aggregate. | ||||||
| 4003 | /// | ||||||
| 4004 | /// If InsertBefore is not null, this function will duplicate (modified) | ||||||
| 4005 | /// insertvalues when a part of a nested struct is extracted. | ||||||
| 4006 | Value *llvm::FindInsertedValue(Value *V, ArrayRef<unsigned> idx_range, | ||||||
| 4007 | Instruction *InsertBefore) { | ||||||
| 4008 | // Nothing to index? Just return V then (this is useful at the end of our | ||||||
| 4009 | // recursion). | ||||||
| 4010 | if (idx_range.empty()) | ||||||
| 4011 | return V; | ||||||
| 4012 | // We have indices, so V should have an indexable type. | ||||||
| 4013 | assert((V->getType()->isStructTy() || V->getType()->isArrayTy()) &&((void)0) | ||||||
| 4014 | "Not looking at a struct or array?")((void)0); | ||||||
| 4015 | assert(ExtractValueInst::getIndexedType(V->getType(), idx_range) &&((void)0) | ||||||
| 4016 | "Invalid indices for type?")((void)0); | ||||||
| 4017 | |||||||
| 4018 | if (Constant *C = dyn_cast<Constant>(V)) { | ||||||
| 4019 | C = C->getAggregateElement(idx_range[0]); | ||||||
| 4020 | if (!C) return nullptr; | ||||||
| 4021 | return FindInsertedValue(C, idx_range.slice(1), InsertBefore); | ||||||
| 4022 | } | ||||||
| 4023 | |||||||
| 4024 | if (InsertValueInst *I = dyn_cast<InsertValueInst>(V)) { | ||||||
| 4025 | // Loop the indices for the insertvalue instruction in parallel with the | ||||||
| 4026 | // requested indices | ||||||
| 4027 | const unsigned *req_idx = idx_range.begin(); | ||||||
| 4028 | for (const unsigned *i = I->idx_begin(), *e = I->idx_end(); | ||||||
| 4029 | i != e; ++i, ++req_idx) { | ||||||
| 4030 | if (req_idx == idx_range.end()) { | ||||||
| 4031 | // We can't handle this without inserting insertvalues | ||||||
| 4032 | if (!InsertBefore) | ||||||
| 4033 | return nullptr; | ||||||
| 4034 | |||||||
| 4035 | // The requested index identifies a part of a nested aggregate. Handle | ||||||
| 4036 | // this specially. For example, | ||||||
| 4037 | // %A = insertvalue { i32, {i32, i32 } } undef, i32 10, 1, 0 | ||||||
| 4038 | // %B = insertvalue { i32, {i32, i32 } } %A, i32 11, 1, 1 | ||||||
| 4039 | // %C = extractvalue {i32, { i32, i32 } } %B, 1 | ||||||
| 4040 | // This can be changed into | ||||||
| 4041 | // %A = insertvalue {i32, i32 } undef, i32 10, 0 | ||||||
| 4042 | // %C = insertvalue {i32, i32 } %A, i32 11, 1 | ||||||
| 4043 | // which allows the unused 0,0 element from the nested struct to be | ||||||
| 4044 | // removed. | ||||||
| 4045 | return BuildSubAggregate(V, makeArrayRef(idx_range.begin(), req_idx), | ||||||
| 4046 | InsertBefore); | ||||||
| 4047 | } | ||||||
| 4048 | |||||||
| 4049 | // This insert value inserts something else than what we are looking for. | ||||||
| 4050 | // See if the (aggregate) value inserted into has the value we are | ||||||
| 4051 | // looking for, then. | ||||||
| 4052 | if (*req_idx != *i) | ||||||
| 4053 | return FindInsertedValue(I->getAggregateOperand(), idx_range, | ||||||
| 4054 | InsertBefore); | ||||||
| 4055 | } | ||||||
| 4056 | // If we end up here, the indices of the insertvalue match with those | ||||||
| 4057 | // requested (though possibly only partially). Now we recursively look at | ||||||
| 4058 | // the inserted value, passing any remaining indices. | ||||||
| 4059 | return FindInsertedValue(I->getInsertedValueOperand(), | ||||||
| 4060 | makeArrayRef(req_idx, idx_range.end()), | ||||||
| 4061 | InsertBefore); | ||||||
| 4062 | } | ||||||
| 4063 | |||||||
| 4064 | if (ExtractValueInst *I = dyn_cast<ExtractValueInst>(V)) { | ||||||
| 4065 | // If we're extracting a value from an aggregate that was extracted from | ||||||
| 4066 | // something else, we can extract from that something else directly instead. | ||||||
| 4067 | // However, we will need to chain I's indices with the requested indices. | ||||||
| 4068 | |||||||
| 4069 | // Calculate the number of indices required | ||||||
| 4070 | unsigned size = I->getNumIndices() + idx_range.size(); | ||||||
| 4071 | // Allocate some space to put the new indices in | ||||||
| 4072 | SmallVector<unsigned, 5> Idxs; | ||||||
| 4073 | Idxs.reserve(size); | ||||||
| 4074 | // Add indices from the extract value instruction | ||||||
| 4075 | Idxs.append(I->idx_begin(), I->idx_end()); | ||||||
| 4076 | |||||||
| 4077 | // Add requested indices | ||||||
| 4078 | Idxs.append(idx_range.begin(), idx_range.end()); | ||||||
| 4079 | |||||||
| 4080 | assert(Idxs.size() == size((void)0) | ||||||
| 4081 | && "Number of indices added not correct?")((void)0); | ||||||
| 4082 | |||||||
| 4083 | return FindInsertedValue(I->getAggregateOperand(), Idxs, InsertBefore); | ||||||
| 4084 | } | ||||||
| 4085 | // Otherwise, we don't know (such as, extracting from a function return value | ||||||
| 4086 | // or load instruction) | ||||||
| 4087 | return nullptr; | ||||||
| 4088 | } | ||||||
| 4089 | |||||||
| 4090 | bool llvm::isGEPBasedOnPointerToString(const GEPOperator *GEP, | ||||||
| 4091 | unsigned CharSize) { | ||||||
| 4092 | // Make sure the GEP has exactly three arguments. | ||||||
| 4093 | if (GEP->getNumOperands() != 3) | ||||||
| 4094 | return false; | ||||||
| 4095 | |||||||
| 4096 | // Make sure the index-ee is a pointer to array of \p CharSize integers. | ||||||
| 4097 | // CharSize. | ||||||
| 4098 | ArrayType *AT = dyn_cast<ArrayType>(GEP->getSourceElementType()); | ||||||
| 4099 | if (!AT || !AT->getElementType()->isIntegerTy(CharSize)) | ||||||
| 4100 | return false; | ||||||
| 4101 | |||||||
| 4102 | // Check to make sure that the first operand of the GEP is an integer and | ||||||
| 4103 | // has value 0 so that we are sure we're indexing into the initializer. | ||||||
| 4104 | const ConstantInt *FirstIdx = dyn_cast<ConstantInt>(GEP->getOperand(1)); | ||||||
| 4105 | if (!FirstIdx || !FirstIdx->isZero()) | ||||||
| 4106 | return false; | ||||||
| 4107 | |||||||
| 4108 | return true; | ||||||
| 4109 | } | ||||||
| 4110 | |||||||
| 4111 | bool llvm::getConstantDataArrayInfo(const Value *V, | ||||||
| 4112 | ConstantDataArraySlice &Slice, | ||||||
| 4113 | unsigned ElementSize, uint64_t Offset) { | ||||||
| 4114 | assert(V)((void)0); | ||||||
| 4115 | |||||||
| 4116 | // Look through bitcast instructions and geps. | ||||||
| 4117 | V = V->stripPointerCasts(); | ||||||
| 4118 | |||||||
| 4119 | // If the value is a GEP instruction or constant expression, treat it as an | ||||||
| 4120 | // offset. | ||||||
| 4121 | if (const GEPOperator *GEP = dyn_cast<GEPOperator>(V)) { | ||||||
| 4122 | // The GEP operator should be based on a pointer to string constant, and is | ||||||
| 4123 | // indexing into the string constant. | ||||||
| 4124 | if (!isGEPBasedOnPointerToString(GEP, ElementSize)) | ||||||
| 4125 | return false; | ||||||
| 4126 | |||||||
| 4127 | // If the second index isn't a ConstantInt, then this is a variable index | ||||||
| 4128 | // into the array. If this occurs, we can't say anything meaningful about | ||||||
| 4129 | // the string. | ||||||
| 4130 | uint64_t StartIdx = 0; | ||||||
| 4131 | if (const ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(2))) | ||||||
| 4132 | StartIdx = CI->getZExtValue(); | ||||||
| 4133 | else | ||||||
| 4134 | return false; | ||||||
| 4135 | return getConstantDataArrayInfo(GEP->getOperand(0), Slice, ElementSize, | ||||||
| 4136 | StartIdx + Offset); | ||||||
| 4137 | } | ||||||
| 4138 | |||||||
| 4139 | // The GEP instruction, constant or instruction, must reference a global | ||||||
| 4140 | // variable that is a constant and is initialized. The referenced constant | ||||||
| 4141 | // initializer is the array that we'll use for optimization. | ||||||
| 4142 | const GlobalVariable *GV = dyn_cast<GlobalVariable>(V); | ||||||
| 4143 | if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer()) | ||||||
| 4144 | return false; | ||||||
| 4145 | |||||||
| 4146 | const ConstantDataArray *Array; | ||||||
| 4147 | ArrayType *ArrayTy; | ||||||
| 4148 | if (GV->getInitializer()->isNullValue()) { | ||||||
| 4149 | Type *GVTy = GV->getValueType(); | ||||||
| 4150 | if ( (ArrayTy = dyn_cast<ArrayType>(GVTy)) ) { | ||||||
| 4151 | // A zeroinitializer for the array; there is no ConstantDataArray. | ||||||
| 4152 | Array = nullptr; | ||||||
| 4153 | } else { | ||||||
| 4154 | const DataLayout &DL = GV->getParent()->getDataLayout(); | ||||||
| 4155 | uint64_t SizeInBytes = DL.getTypeStoreSize(GVTy).getFixedSize(); | ||||||
| 4156 | uint64_t Length = SizeInBytes / (ElementSize / 8); | ||||||
| 4157 | if (Length <= Offset) | ||||||
| 4158 | return false; | ||||||
| 4159 | |||||||
| 4160 | Slice.Array = nullptr; | ||||||
| 4161 | Slice.Offset = 0; | ||||||
| 4162 | Slice.Length = Length - Offset; | ||||||
| 4163 | return true; | ||||||
| 4164 | } | ||||||
| 4165 | } else { | ||||||
| 4166 | // This must be a ConstantDataArray. | ||||||
| 4167 | Array = dyn_cast<ConstantDataArray>(GV->getInitializer()); | ||||||
| 4168 | if (!Array) | ||||||
| 4169 | return false; | ||||||
| 4170 | ArrayTy = Array->getType(); | ||||||
| 4171 | } | ||||||
| 4172 | if (!ArrayTy->getElementType()->isIntegerTy(ElementSize)) | ||||||
| 4173 | return false; | ||||||
| 4174 | |||||||
| 4175 | uint64_t NumElts = ArrayTy->getArrayNumElements(); | ||||||
| 4176 | if (Offset > NumElts) | ||||||
| 4177 | return false; | ||||||
| 4178 | |||||||
| 4179 | Slice.Array = Array; | ||||||
| 4180 | Slice.Offset = Offset; | ||||||
| 4181 | Slice.Length = NumElts - Offset; | ||||||
| 4182 | return true; | ||||||
| 4183 | } | ||||||
| 4184 | |||||||
| 4185 | /// This function computes the length of a null-terminated C string pointed to | ||||||
| 4186 | /// by V. If successful, it returns true and returns the string in Str. | ||||||
| 4187 | /// If unsuccessful, it returns false. | ||||||
| 4188 | bool llvm::getConstantStringInfo(const Value *V, StringRef &Str, | ||||||
| 4189 | uint64_t Offset, bool TrimAtNul) { | ||||||
| 4190 | ConstantDataArraySlice Slice; | ||||||
| 4191 | if (!getConstantDataArrayInfo(V, Slice, 8, Offset)) | ||||||
| 4192 | return false; | ||||||
| 4193 | |||||||
| 4194 | if (Slice.Array == nullptr) { | ||||||
| 4195 | if (TrimAtNul) { | ||||||
| 4196 | Str = StringRef(); | ||||||
| 4197 | return true; | ||||||
| 4198 | } | ||||||
| 4199 | if (Slice.Length == 1) { | ||||||
| 4200 | Str = StringRef("", 1); | ||||||
| 4201 | return true; | ||||||
| 4202 | } | ||||||
| 4203 | // We cannot instantiate a StringRef as we do not have an appropriate string | ||||||
| 4204 | // of 0s at hand. | ||||||
| 4205 | return false; | ||||||
| 4206 | } | ||||||
| 4207 | |||||||
| 4208 | // Start out with the entire array in the StringRef. | ||||||
| 4209 | Str = Slice.Array->getAsString(); | ||||||
| 4210 | // Skip over 'offset' bytes. | ||||||
| 4211 | Str = Str.substr(Slice.Offset); | ||||||
| 4212 | |||||||
| 4213 | if (TrimAtNul) { | ||||||
| 4214 | // Trim off the \0 and anything after it. If the array is not nul | ||||||
| 4215 | // terminated, we just return the whole end of string. The client may know | ||||||
| 4216 | // some other way that the string is length-bound. | ||||||
| 4217 | Str = Str.substr(0, Str.find('\0')); | ||||||
| 4218 | } | ||||||
| 4219 | return true; | ||||||
| 4220 | } | ||||||
| 4221 | |||||||
| 4222 | // These next two are very similar to the above, but also look through PHI | ||||||
| 4223 | // nodes. | ||||||
| 4224 | // TODO: See if we can integrate these two together. | ||||||
| 4225 | |||||||
| 4226 | /// If we can compute the length of the string pointed to by | ||||||
| 4227 | /// the specified pointer, return 'len+1'. If we can't, return 0. | ||||||
| 4228 | static uint64_t GetStringLengthH(const Value *V, | ||||||
| 4229 | SmallPtrSetImpl<const PHINode*> &PHIs, | ||||||
| 4230 | unsigned CharSize) { | ||||||
| 4231 | // Look through noop bitcast instructions. | ||||||
| 4232 | V = V->stripPointerCasts(); | ||||||
| 4233 | |||||||
| 4234 | // If this is a PHI node, there are two cases: either we have already seen it | ||||||
| 4235 | // or we haven't. | ||||||
| 4236 | if (const PHINode *PN = dyn_cast<PHINode>(V)) { | ||||||
| 4237 | if (!PHIs.insert(PN).second) | ||||||
| 4238 | return ~0ULL; // already in the set. | ||||||
| 4239 | |||||||
| 4240 | // If it was new, see if all the input strings are the same length. | ||||||
| 4241 | uint64_t LenSoFar = ~0ULL; | ||||||
| 4242 | for (Value *IncValue : PN->incoming_values()) { | ||||||
| 4243 | uint64_t Len = GetStringLengthH(IncValue, PHIs, CharSize); | ||||||
| 4244 | if (Len == 0) return 0; // Unknown length -> unknown. | ||||||
| 4245 | |||||||
| 4246 | if (Len == ~0ULL) continue; | ||||||
| 4247 | |||||||
| 4248 | if (Len != LenSoFar && LenSoFar != ~0ULL) | ||||||
| 4249 | return 0; // Disagree -> unknown. | ||||||
| 4250 | LenSoFar = Len; | ||||||
| 4251 | } | ||||||
| 4252 | |||||||
| 4253 | // Success, all agree. | ||||||
| 4254 | return LenSoFar; | ||||||
| 4255 | } | ||||||
| 4256 | |||||||
| 4257 | // strlen(select(c,x,y)) -> strlen(x) ^ strlen(y) | ||||||
| 4258 | if (const SelectInst *SI = dyn_cast<SelectInst>(V)) { | ||||||
| 4259 | uint64_t Len1 = GetStringLengthH(SI->getTrueValue(), PHIs, CharSize); | ||||||
| 4260 | if (Len1 == 0) return 0; | ||||||
| 4261 | uint64_t Len2 = GetStringLengthH(SI->getFalseValue(), PHIs, CharSize); | ||||||
| 4262 | if (Len2 == 0) return 0; | ||||||
| 4263 | if (Len1 == ~0ULL) return Len2; | ||||||
| 4264 | if (Len2 == ~0ULL) return Len1; | ||||||
| 4265 | if (Len1 != Len2) return 0; | ||||||
| 4266 | return Len1; | ||||||
| 4267 | } | ||||||
| 4268 | |||||||
| 4269 | // Otherwise, see if we can read the string. | ||||||
| 4270 | ConstantDataArraySlice Slice; | ||||||
| 4271 | if (!getConstantDataArrayInfo(V, Slice, CharSize)) | ||||||
| 4272 | return 0; | ||||||
| 4273 | |||||||
| 4274 | if (Slice.Array == nullptr) | ||||||
| 4275 | return 1; | ||||||
| 4276 | |||||||
| 4277 | // Search for nul characters | ||||||
| 4278 | unsigned NullIndex = 0; | ||||||
| 4279 | for (unsigned E = Slice.Length; NullIndex < E; ++NullIndex) { | ||||||
| 4280 | if (Slice.Array->getElementAsInteger(Slice.Offset + NullIndex) == 0) | ||||||
| 4281 | break; | ||||||
| 4282 | } | ||||||
| 4283 | |||||||
| 4284 | return NullIndex + 1; | ||||||
| 4285 | } | ||||||
| 4286 | |||||||
| 4287 | /// If we can compute the length of the string pointed to by | ||||||
| 4288 | /// the specified pointer, return 'len+1'. If we can't, return 0. | ||||||
| 4289 | uint64_t llvm::GetStringLength(const Value *V, unsigned CharSize) { | ||||||
| 4290 | if (!V->getType()->isPointerTy()) | ||||||
| 4291 | return 0; | ||||||
| 4292 | |||||||
| 4293 | SmallPtrSet<const PHINode*, 32> PHIs; | ||||||
| 4294 | uint64_t Len = GetStringLengthH(V, PHIs, CharSize); | ||||||
| 4295 | // If Len is ~0ULL, we had an infinite phi cycle: this is dead code, so return | ||||||
| 4296 | // an empty string as a length. | ||||||
| 4297 | return Len == ~0ULL ? 1 : Len; | ||||||
| 4298 | } | ||||||
| 4299 | |||||||
| 4300 | const Value * | ||||||
| 4301 | llvm::getArgumentAliasingToReturnedPointer(const CallBase *Call, | ||||||
| 4302 | bool MustPreserveNullness) { | ||||||
| 4303 | assert(Call &&((void)0) | ||||||
| 4304 | "getArgumentAliasingToReturnedPointer only works on nonnull calls")((void)0); | ||||||
| 4305 | if (const Value *RV = Call->getReturnedArgOperand()) | ||||||
| 4306 | return RV; | ||||||
| 4307 | // This can be used only as a aliasing property. | ||||||
| 4308 | if (isIntrinsicReturningPointerAliasingArgumentWithoutCapturing( | ||||||
| 4309 | Call, MustPreserveNullness)) | ||||||
| 4310 | return Call->getArgOperand(0); | ||||||
| 4311 | return nullptr; | ||||||
| 4312 | } | ||||||
| 4313 | |||||||
| 4314 | bool llvm::isIntrinsicReturningPointerAliasingArgumentWithoutCapturing( | ||||||
| 4315 | const CallBase *Call, bool MustPreserveNullness) { | ||||||
| 4316 | switch (Call->getIntrinsicID()) { | ||||||
| 4317 | case Intrinsic::launder_invariant_group: | ||||||
| 4318 | case Intrinsic::strip_invariant_group: | ||||||
| 4319 | case Intrinsic::aarch64_irg: | ||||||
| 4320 | case Intrinsic::aarch64_tagp: | ||||||
| 4321 | return true; | ||||||
| 4322 | case Intrinsic::ptrmask: | ||||||
| 4323 | return !MustPreserveNullness; | ||||||
| 4324 | default: | ||||||
| 4325 | return false; | ||||||
| 4326 | } | ||||||
| 4327 | } | ||||||
| 4328 | |||||||
| 4329 | /// \p PN defines a loop-variant pointer to an object. Check if the | ||||||
| 4330 | /// previous iteration of the loop was referring to the same object as \p PN. | ||||||
| 4331 | static bool isSameUnderlyingObjectInLoop(const PHINode *PN, | ||||||
| 4332 | const LoopInfo *LI) { | ||||||
| 4333 | // Find the loop-defined value. | ||||||
| 4334 | Loop *L = LI->getLoopFor(PN->getParent()); | ||||||
| 4335 | if (PN->getNumIncomingValues() != 2) | ||||||
| 4336 | return true; | ||||||
| 4337 | |||||||
| 4338 | // Find the value from previous iteration. | ||||||
| 4339 | auto *PrevValue = dyn_cast<Instruction>(PN->getIncomingValue(0)); | ||||||
| 4340 | if (!PrevValue || LI->getLoopFor(PrevValue->getParent()) != L) | ||||||
| 4341 | PrevValue = dyn_cast<Instruction>(PN->getIncomingValue(1)); | ||||||
| 4342 | if (!PrevValue || LI->getLoopFor(PrevValue->getParent()) != L) | ||||||
| 4343 | return true; | ||||||
| 4344 | |||||||
| 4345 | // If a new pointer is loaded in the loop, the pointer references a different | ||||||
| 4346 | // object in every iteration. E.g.: | ||||||
| 4347 | // for (i) | ||||||
| 4348 | // int *p = a[i]; | ||||||
| 4349 | // ... | ||||||
| 4350 | if (auto *Load = dyn_cast<LoadInst>(PrevValue)) | ||||||
| 4351 | if (!L->isLoopInvariant(Load->getPointerOperand())) | ||||||
| 4352 | return false; | ||||||
| 4353 | return true; | ||||||
| 4354 | } | ||||||
| 4355 | |||||||
| 4356 | const Value *llvm::getUnderlyingObject(const Value *V, unsigned MaxLookup) { | ||||||
| 4357 | if (!V->getType()->isPointerTy()) | ||||||
| 4358 | return V; | ||||||
| 4359 | for (unsigned Count = 0; MaxLookup == 0 || Count < MaxLookup; ++Count) { | ||||||
| 4360 | if (auto *GEP = dyn_cast<GEPOperator>(V)) { | ||||||
| 4361 | V = GEP->getPointerOperand(); | ||||||
| 4362 | } else if (Operator::getOpcode(V) == Instruction::BitCast || | ||||||
| 4363 | Operator::getOpcode(V) == Instruction::AddrSpaceCast) { | ||||||
| 4364 | V = cast<Operator>(V)->getOperand(0); | ||||||
| 4365 | if (!V->getType()->isPointerTy()) | ||||||
| 4366 | return V; | ||||||
| 4367 | } else if (auto *GA = dyn_cast<GlobalAlias>(V)) { | ||||||
| 4368 | if (GA->isInterposable()) | ||||||
| 4369 | return V; | ||||||
| 4370 | V = GA->getAliasee(); | ||||||
| 4371 | } else { | ||||||
| 4372 | if (auto *PHI = dyn_cast<PHINode>(V)) { | ||||||
| 4373 | // Look through single-arg phi nodes created by LCSSA. | ||||||
| 4374 | if (PHI->getNumIncomingValues() == 1) { | ||||||
| 4375 | V = PHI->getIncomingValue(0); | ||||||
| 4376 | continue; | ||||||
| 4377 | } | ||||||
| 4378 | } else if (auto *Call = dyn_cast<CallBase>(V)) { | ||||||
| 4379 | // CaptureTracking can know about special capturing properties of some | ||||||
| 4380 | // intrinsics like launder.invariant.group, that can't be expressed with | ||||||
| 4381 | // the attributes, but have properties like returning aliasing pointer. | ||||||
| 4382 | // Because some analysis may assume that nocaptured pointer is not | ||||||
| 4383 | // returned from some special intrinsic (because function would have to | ||||||
| 4384 | // be marked with returns attribute), it is crucial to use this function | ||||||
| 4385 | // because it should be in sync with CaptureTracking. Not using it may | ||||||
| 4386 | // cause weird miscompilations where 2 aliasing pointers are assumed to | ||||||
| 4387 | // noalias. | ||||||
| 4388 | if (auto *RP = getArgumentAliasingToReturnedPointer(Call, false)) { | ||||||
| 4389 | V = RP; | ||||||
| 4390 | continue; | ||||||
| 4391 | } | ||||||
| 4392 | } | ||||||
| 4393 | |||||||
| 4394 | return V; | ||||||
| 4395 | } | ||||||
| 4396 | assert(V->getType()->isPointerTy() && "Unexpected operand type!")((void)0); | ||||||
| 4397 | } | ||||||
| 4398 | return V; | ||||||
| 4399 | } | ||||||
| 4400 | |||||||
| 4401 | void llvm::getUnderlyingObjects(const Value *V, | ||||||
| 4402 | SmallVectorImpl<const Value *> &Objects, | ||||||
| 4403 | LoopInfo *LI, unsigned MaxLookup) { | ||||||
| 4404 | SmallPtrSet<const Value *, 4> Visited; | ||||||
| 4405 | SmallVector<const Value *, 4> Worklist; | ||||||
| 4406 | Worklist.push_back(V); | ||||||
| 4407 | do { | ||||||
| 4408 | const Value *P = Worklist.pop_back_val(); | ||||||
| 4409 | P = getUnderlyingObject(P, MaxLookup); | ||||||
| 4410 | |||||||
| 4411 | if (!Visited.insert(P).second) | ||||||
| 4412 | continue; | ||||||
| 4413 | |||||||
| 4414 | if (auto *SI = dyn_cast<SelectInst>(P)) { | ||||||
| 4415 | Worklist.push_back(SI->getTrueValue()); | ||||||
| 4416 | Worklist.push_back(SI->getFalseValue()); | ||||||
| 4417 | continue; | ||||||
| 4418 | } | ||||||
| 4419 | |||||||
| 4420 | if (auto *PN = dyn_cast<PHINode>(P)) { | ||||||
| 4421 | // If this PHI changes the underlying object in every iteration of the | ||||||
| 4422 | // loop, don't look through it. Consider: | ||||||
| 4423 | // int **A; | ||||||
| 4424 | // for (i) { | ||||||
| 4425 | // Prev = Curr; // Prev = PHI (Prev_0, Curr) | ||||||
| 4426 | // Curr = A[i]; | ||||||
| 4427 | // *Prev, *Curr; | ||||||
| 4428 | // | ||||||
| 4429 | // Prev is tracking Curr one iteration behind so they refer to different | ||||||
| 4430 | // underlying objects. | ||||||
| 4431 | if (!LI || !LI->isLoopHeader(PN->getParent()) || | ||||||
| 4432 | isSameUnderlyingObjectInLoop(PN, LI)) | ||||||
| 4433 | append_range(Worklist, PN->incoming_values()); | ||||||
| 4434 | continue; | ||||||
| 4435 | } | ||||||
| 4436 | |||||||
| 4437 | Objects.push_back(P); | ||||||
| 4438 | } while (!Worklist.empty()); | ||||||
| 4439 | } | ||||||
| 4440 | |||||||
| 4441 | /// This is the function that does the work of looking through basic | ||||||
| 4442 | /// ptrtoint+arithmetic+inttoptr sequences. | ||||||
| 4443 | static const Value *getUnderlyingObjectFromInt(const Value *V) { | ||||||
| 4444 | do { | ||||||
| 4445 | if (const Operator *U = dyn_cast<Operator>(V)) { | ||||||
| 4446 | // If we find a ptrtoint, we can transfer control back to the | ||||||
| 4447 | // regular getUnderlyingObjectFromInt. | ||||||
| 4448 | if (U->getOpcode() == Instruction::PtrToInt) | ||||||
| 4449 | return U->getOperand(0); | ||||||
| 4450 | // If we find an add of a constant, a multiplied value, or a phi, it's | ||||||
| 4451 | // likely that the other operand will lead us to the base | ||||||
| 4452 | // object. We don't have to worry about the case where the | ||||||
| 4453 | // object address is somehow being computed by the multiply, | ||||||
| 4454 | // because our callers only care when the result is an | ||||||
| 4455 | // identifiable object. | ||||||
| 4456 | if (U->getOpcode() != Instruction::Add || | ||||||
| 4457 | (!isa<ConstantInt>(U->getOperand(1)) && | ||||||
| 4458 | Operator::getOpcode(U->getOperand(1)) != Instruction::Mul && | ||||||
| 4459 | !isa<PHINode>(U->getOperand(1)))) | ||||||
| 4460 | return V; | ||||||
| 4461 | V = U->getOperand(0); | ||||||
| 4462 | } else { | ||||||
| 4463 | return V; | ||||||
| 4464 | } | ||||||
| 4465 | assert(V->getType()->isIntegerTy() && "Unexpected operand type!")((void)0); | ||||||
| 4466 | } while (true); | ||||||
| 4467 | } | ||||||
| 4468 | |||||||
| 4469 | /// This is a wrapper around getUnderlyingObjects and adds support for basic | ||||||
| 4470 | /// ptrtoint+arithmetic+inttoptr sequences. | ||||||
| 4471 | /// It returns false if unidentified object is found in getUnderlyingObjects. | ||||||
| 4472 | bool llvm::getUnderlyingObjectsForCodeGen(const Value *V, | ||||||
| 4473 | SmallVectorImpl<Value *> &Objects) { | ||||||
| 4474 | SmallPtrSet<const Value *, 16> Visited; | ||||||
| 4475 | SmallVector<const Value *, 4> Working(1, V); | ||||||
| 4476 | do { | ||||||
| 4477 | V = Working.pop_back_val(); | ||||||
| 4478 | |||||||
| 4479 | SmallVector<const Value *, 4> Objs; | ||||||
| 4480 | getUnderlyingObjects(V, Objs); | ||||||
| 4481 | |||||||
| 4482 | for (const Value *V : Objs) { | ||||||
| 4483 | if (!Visited.insert(V).second) | ||||||
| 4484 | continue; | ||||||
| 4485 | if (Operator::getOpcode(V) == Instruction::IntToPtr) { | ||||||
| 4486 | const Value *O = | ||||||
| 4487 | getUnderlyingObjectFromInt(cast<User>(V)->getOperand(0)); | ||||||
| 4488 | if (O->getType()->isPointerTy()) { | ||||||
| 4489 | Working.push_back(O); | ||||||
| 4490 | continue; | ||||||
| 4491 | } | ||||||
| 4492 | } | ||||||
| 4493 | // If getUnderlyingObjects fails to find an identifiable object, | ||||||
| 4494 | // getUnderlyingObjectsForCodeGen also fails for safety. | ||||||
| 4495 | if (!isIdentifiedObject(V)) { | ||||||
| 4496 | Objects.clear(); | ||||||
| 4497 | return false; | ||||||
| 4498 | } | ||||||
| 4499 | Objects.push_back(const_cast<Value *>(V)); | ||||||
| 4500 | } | ||||||
| 4501 | } while (!Working.empty()); | ||||||
| 4502 | return true; | ||||||
| 4503 | } | ||||||
| 4504 | |||||||
| 4505 | AllocaInst *llvm::findAllocaForValue(Value *V, bool OffsetZero) { | ||||||
| 4506 | AllocaInst *Result = nullptr; | ||||||
| 4507 | SmallPtrSet<Value *, 4> Visited; | ||||||
| 4508 | SmallVector<Value *, 4> Worklist; | ||||||
| 4509 | |||||||
| 4510 | auto AddWork = [&](Value *V) { | ||||||
| 4511 | if (Visited.insert(V).second) | ||||||
| 4512 | Worklist.push_back(V); | ||||||
| 4513 | }; | ||||||
| 4514 | |||||||
| 4515 | AddWork(V); | ||||||
| 4516 | do { | ||||||
| 4517 | V = Worklist.pop_back_val(); | ||||||
| 4518 | assert(Visited.count(V))((void)0); | ||||||
| 4519 | |||||||
| 4520 | if (AllocaInst *AI = dyn_cast<AllocaInst>(V)) { | ||||||
| 4521 | if (Result && Result != AI) | ||||||
| 4522 | return nullptr; | ||||||
| 4523 | Result = AI; | ||||||
| 4524 | } else if (CastInst *CI = dyn_cast<CastInst>(V)) { | ||||||
| 4525 | AddWork(CI->getOperand(0)); | ||||||
| 4526 | } else if (PHINode *PN = dyn_cast<PHINode>(V)) { | ||||||
| 4527 | for (Value *IncValue : PN->incoming_values()) | ||||||
| 4528 | AddWork(IncValue); | ||||||
| 4529 | } else if (auto *SI = dyn_cast<SelectInst>(V)) { | ||||||
| 4530 | AddWork(SI->getTrueValue()); | ||||||
| 4531 | AddWork(SI->getFalseValue()); | ||||||
| 4532 | } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(V)) { | ||||||
| 4533 | if (OffsetZero && !GEP->hasAllZeroIndices()) | ||||||
| 4534 | return nullptr; | ||||||
| 4535 | AddWork(GEP->getPointerOperand()); | ||||||
| 4536 | } else { | ||||||
| 4537 | return nullptr; | ||||||
| 4538 | } | ||||||
| 4539 | } while (!Worklist.empty()); | ||||||
| 4540 | |||||||
| 4541 | return Result; | ||||||
| 4542 | } | ||||||
| 4543 | |||||||
| 4544 | static bool onlyUsedByLifetimeMarkersOrDroppableInstsHelper( | ||||||
| 4545 | const Value *V, bool AllowLifetime, bool AllowDroppable) { | ||||||
| 4546 | for (const User *U : V->users()) { | ||||||
| 4547 | const IntrinsicInst *II = dyn_cast<IntrinsicInst>(U); | ||||||
| 4548 | if (!II) | ||||||
| 4549 | return false; | ||||||
| 4550 | |||||||
| 4551 | if (AllowLifetime && II->isLifetimeStartOrEnd()) | ||||||
| 4552 | continue; | ||||||
| 4553 | |||||||
| 4554 | if (AllowDroppable && II->isDroppable()) | ||||||
| 4555 | continue; | ||||||
| 4556 | |||||||
| 4557 | return false; | ||||||
| 4558 | } | ||||||
| 4559 | return true; | ||||||
| 4560 | } | ||||||
| 4561 | |||||||
| 4562 | bool llvm::onlyUsedByLifetimeMarkers(const Value *V) { | ||||||
| 4563 | return onlyUsedByLifetimeMarkersOrDroppableInstsHelper( | ||||||
| 4564 | V, /* AllowLifetime */ true, /* AllowDroppable */ false); | ||||||
| 4565 | } | ||||||
| 4566 | bool llvm::onlyUsedByLifetimeMarkersOrDroppableInsts(const Value *V) { | ||||||
| 4567 | return onlyUsedByLifetimeMarkersOrDroppableInstsHelper( | ||||||
| 4568 | V, /* AllowLifetime */ true, /* AllowDroppable */ true); | ||||||
| 4569 | } | ||||||
| 4570 | |||||||
| 4571 | bool llvm::mustSuppressSpeculation(const LoadInst &LI) { | ||||||
| 4572 | if (!LI.isUnordered()) | ||||||
| 4573 | return true; | ||||||
| 4574 | const Function &F = *LI.getFunction(); | ||||||
| 4575 | // Speculative load may create a race that did not exist in the source. | ||||||
| 4576 | return F.hasFnAttribute(Attribute::SanitizeThread) || | ||||||
| 4577 | // Speculative load may load data from dirty regions. | ||||||
| 4578 | F.hasFnAttribute(Attribute::SanitizeAddress) || | ||||||
| 4579 | F.hasFnAttribute(Attribute::SanitizeHWAddress); | ||||||
| 4580 | } | ||||||
| 4581 | |||||||
| 4582 | |||||||
| 4583 | bool llvm::isSafeToSpeculativelyExecute(const Value *V, | ||||||
| 4584 | const Instruction *CtxI, | ||||||
| 4585 | const DominatorTree *DT, | ||||||
| 4586 | const TargetLibraryInfo *TLI) { | ||||||
| 4587 | const Operator *Inst = dyn_cast<Operator>(V); | ||||||
| 4588 | if (!Inst) | ||||||
| 4589 | return false; | ||||||
| 4590 | |||||||
| 4591 | for (unsigned i = 0, e = Inst->getNumOperands(); i != e; ++i) | ||||||
| 4592 | if (Constant *C = dyn_cast<Constant>(Inst->getOperand(i))) | ||||||
| 4593 | if (C->canTrap()) | ||||||
| 4594 | return false; | ||||||
| 4595 | |||||||
| 4596 | switch (Inst->getOpcode()) { | ||||||
| 4597 | default: | ||||||
| 4598 | return true; | ||||||
| 4599 | case Instruction::UDiv: | ||||||
| 4600 | case Instruction::URem: { | ||||||
| 4601 | // x / y is undefined if y == 0. | ||||||
| 4602 | const APInt *V; | ||||||
| 4603 | if (match(Inst->getOperand(1), m_APInt(V))) | ||||||
| 4604 | return *V != 0; | ||||||
| 4605 | return false; | ||||||
| 4606 | } | ||||||
| 4607 | case Instruction::SDiv: | ||||||
| 4608 | case Instruction::SRem: { | ||||||
| 4609 | // x / y is undefined if y == 0 or x == INT_MIN and y == -1 | ||||||
| 4610 | const APInt *Numerator, *Denominator; | ||||||
| 4611 | if (!match(Inst->getOperand(1), m_APInt(Denominator))) | ||||||
| 4612 | return false; | ||||||
| 4613 | // We cannot hoist this division if the denominator is 0. | ||||||
| 4614 | if (*Denominator == 0) | ||||||
| 4615 | return false; | ||||||
| 4616 | // It's safe to hoist if the denominator is not 0 or -1. | ||||||
| 4617 | if (!Denominator->isAllOnesValue()) | ||||||
| 4618 | return true; | ||||||
| 4619 | // At this point we know that the denominator is -1. It is safe to hoist as | ||||||
| 4620 | // long we know that the numerator is not INT_MIN. | ||||||
| 4621 | if (match(Inst->getOperand(0), m_APInt(Numerator))) | ||||||
| 4622 | return !Numerator->isMinSignedValue(); | ||||||
| 4623 | // The numerator *might* be MinSignedValue. | ||||||
| 4624 | return false; | ||||||
| 4625 | } | ||||||
| 4626 | case Instruction::Load: { | ||||||
| 4627 | const LoadInst *LI = cast<LoadInst>(Inst); | ||||||
| 4628 | if (mustSuppressSpeculation(*LI)) | ||||||
| 4629 | return false; | ||||||
| 4630 | const DataLayout &DL = LI->getModule()->getDataLayout(); | ||||||
| 4631 | return isDereferenceableAndAlignedPointer( | ||||||
| 4632 | LI->getPointerOperand(), LI->getType(), MaybeAlign(LI->getAlignment()), | ||||||
| 4633 | DL, CtxI, DT, TLI); | ||||||
| 4634 | } | ||||||
| 4635 | case Instruction::Call: { | ||||||
| 4636 | auto *CI = cast<const CallInst>(Inst); | ||||||
| 4637 | const Function *Callee = CI->getCalledFunction(); | ||||||
| 4638 | |||||||
| 4639 | // The called function could have undefined behavior or side-effects, even | ||||||
| 4640 | // if marked readnone nounwind. | ||||||
| 4641 | return Callee && Callee->isSpeculatable(); | ||||||
| 4642 | } | ||||||
| 4643 | case Instruction::VAArg: | ||||||
| 4644 | case Instruction::Alloca: | ||||||
| 4645 | case Instruction::Invoke: | ||||||
| 4646 | case Instruction::CallBr: | ||||||
| 4647 | case Instruction::PHI: | ||||||
| 4648 | case Instruction::Store: | ||||||
| 4649 | case Instruction::Ret: | ||||||
| 4650 | case Instruction::Br: | ||||||
| 4651 | case Instruction::IndirectBr: | ||||||
| 4652 | case Instruction::Switch: | ||||||
| 4653 | case Instruction::Unreachable: | ||||||
| 4654 | case Instruction::Fence: | ||||||
| 4655 | case Instruction::AtomicRMW: | ||||||
| 4656 | case Instruction::AtomicCmpXchg: | ||||||
| 4657 | case Instruction::LandingPad: | ||||||
| 4658 | case Instruction::Resume: | ||||||
| 4659 | case Instruction::CatchSwitch: | ||||||
| 4660 | case Instruction::CatchPad: | ||||||
| 4661 | case Instruction::CatchRet: | ||||||
| 4662 | case Instruction::CleanupPad: | ||||||
| 4663 | case Instruction::CleanupRet: | ||||||
| 4664 | return false; // Misc instructions which have effects | ||||||
| 4665 | } | ||||||
| 4666 | } | ||||||
| 4667 | |||||||
| 4668 | bool llvm::mayBeMemoryDependent(const Instruction &I) { | ||||||
| 4669 | return I.mayReadOrWriteMemory() || !isSafeToSpeculativelyExecute(&I); | ||||||
| 4670 | } | ||||||
| 4671 | |||||||
| 4672 | /// Convert ConstantRange OverflowResult into ValueTracking OverflowResult. | ||||||
| 4673 | static OverflowResult mapOverflowResult(ConstantRange::OverflowResult OR) { | ||||||
| 4674 | switch (OR) { | ||||||
| 4675 | case ConstantRange::OverflowResult::MayOverflow: | ||||||
| 4676 | return OverflowResult::MayOverflow; | ||||||
| 4677 | case ConstantRange::OverflowResult::AlwaysOverflowsLow: | ||||||
| 4678 | return OverflowResult::AlwaysOverflowsLow; | ||||||
| 4679 | case ConstantRange::OverflowResult::AlwaysOverflowsHigh: | ||||||
| 4680 | return OverflowResult::AlwaysOverflowsHigh; | ||||||
| 4681 | case ConstantRange::OverflowResult::NeverOverflows: | ||||||
| 4682 | return OverflowResult::NeverOverflows; | ||||||
| 4683 | } | ||||||
| 4684 | llvm_unreachable("Unknown OverflowResult")__builtin_unreachable(); | ||||||
| 4685 | } | ||||||
| 4686 | |||||||
| 4687 | /// Combine constant ranges from computeConstantRange() and computeKnownBits(). | ||||||
| 4688 | static ConstantRange computeConstantRangeIncludingKnownBits( | ||||||
| 4689 | const Value *V, bool ForSigned, const DataLayout &DL, unsigned Depth, | ||||||
| 4690 | AssumptionCache *AC, const Instruction *CxtI, const DominatorTree *DT, | ||||||
| 4691 | OptimizationRemarkEmitter *ORE = nullptr, bool UseInstrInfo = true) { | ||||||
| 4692 | KnownBits Known = computeKnownBits( | ||||||
| 4693 | V, DL, Depth, AC, CxtI, DT, ORE, UseInstrInfo); | ||||||
| 4694 | ConstantRange CR1 = ConstantRange::fromKnownBits(Known, ForSigned); | ||||||
| 4695 | ConstantRange CR2 = computeConstantRange(V, UseInstrInfo); | ||||||
| 4696 | ConstantRange::PreferredRangeType RangeType = | ||||||
| 4697 | ForSigned ? ConstantRange::Signed : ConstantRange::Unsigned; | ||||||
| 4698 | return CR1.intersectWith(CR2, RangeType); | ||||||
| 4699 | } | ||||||
| 4700 | |||||||
| 4701 | OverflowResult llvm::computeOverflowForUnsignedMul( | ||||||
| 4702 | const Value *LHS, const Value *RHS, const DataLayout &DL, | ||||||
| 4703 | AssumptionCache *AC, const Instruction *CxtI, const DominatorTree *DT, | ||||||
| 4704 | bool UseInstrInfo) { | ||||||
| 4705 | KnownBits LHSKnown = computeKnownBits(LHS, DL, /*Depth=*/0, AC, CxtI, DT, | ||||||
| 4706 | nullptr, UseInstrInfo); | ||||||
| 4707 | KnownBits RHSKnown = computeKnownBits(RHS, DL, /*Depth=*/0, AC, CxtI, DT, | ||||||
| 4708 | nullptr, UseInstrInfo); | ||||||
| 4709 | ConstantRange LHSRange = ConstantRange::fromKnownBits(LHSKnown, false); | ||||||
| 4710 | ConstantRange RHSRange = ConstantRange::fromKnownBits(RHSKnown, false); | ||||||
| 4711 | return mapOverflowResult(LHSRange.unsignedMulMayOverflow(RHSRange)); | ||||||
| 4712 | } | ||||||
| 4713 | |||||||
| 4714 | OverflowResult | ||||||
| 4715 | llvm::computeOverflowForSignedMul(const Value *LHS, const Value *RHS, | ||||||
| 4716 | const DataLayout &DL, AssumptionCache *AC, | ||||||
| 4717 | const Instruction *CxtI, | ||||||
| 4718 | const DominatorTree *DT, bool UseInstrInfo) { | ||||||
| 4719 | // Multiplying n * m significant bits yields a result of n + m significant | ||||||
| 4720 | // bits. If the total number of significant bits does not exceed the | ||||||
| 4721 | // result bit width (minus 1), there is no overflow. | ||||||
| 4722 | // This means if we have enough leading sign bits in the operands | ||||||
| 4723 | // we can guarantee that the result does not overflow. | ||||||
| 4724 | // Ref: "Hacker's Delight" by Henry Warren | ||||||
| 4725 | unsigned BitWidth = LHS->getType()->getScalarSizeInBits(); | ||||||
| 4726 | |||||||
| 4727 | // Note that underestimating the number of sign bits gives a more | ||||||
| 4728 | // conservative answer. | ||||||
| 4729 | unsigned SignBits = ComputeNumSignBits(LHS, DL, 0, AC, CxtI, DT) + | ||||||
| 4730 | ComputeNumSignBits(RHS, DL, 0, AC, CxtI, DT); | ||||||
| 4731 | |||||||
| 4732 | // First handle the easy case: if we have enough sign bits there's | ||||||
| 4733 | // definitely no overflow. | ||||||
| 4734 | if (SignBits > BitWidth + 1) | ||||||
| 4735 | return OverflowResult::NeverOverflows; | ||||||
| 4736 | |||||||
| 4737 | // There are two ambiguous cases where there can be no overflow: | ||||||
| 4738 | // SignBits == BitWidth + 1 and | ||||||
| 4739 | // SignBits == BitWidth | ||||||
| 4740 | // The second case is difficult to check, therefore we only handle the | ||||||
| 4741 | // first case. | ||||||
| 4742 | if (SignBits == BitWidth + 1) { | ||||||
| 4743 | // It overflows only when both arguments are negative and the true | ||||||
| 4744 | // product is exactly the minimum negative number. | ||||||
| 4745 | // E.g. mul i16 with 17 sign bits: 0xff00 * 0xff80 = 0x8000 | ||||||
| 4746 | // For simplicity we just check if at least one side is not negative. | ||||||
| 4747 | KnownBits LHSKnown = computeKnownBits(LHS, DL, /*Depth=*/0, AC, CxtI, DT, | ||||||
| 4748 | nullptr, UseInstrInfo); | ||||||
| 4749 | KnownBits RHSKnown = computeKnownBits(RHS, DL, /*Depth=*/0, AC, CxtI, DT, | ||||||
| 4750 | nullptr, UseInstrInfo); | ||||||
| 4751 | if (LHSKnown.isNonNegative() || RHSKnown.isNonNegative()) | ||||||
| 4752 | return OverflowResult::NeverOverflows; | ||||||
| 4753 | } | ||||||
| 4754 | return OverflowResult::MayOverflow; | ||||||
| 4755 | } | ||||||
| 4756 | |||||||
| 4757 | OverflowResult llvm::computeOverflowForUnsignedAdd( | ||||||
| 4758 | const Value *LHS, const Value *RHS, const DataLayout &DL, | ||||||
| 4759 | AssumptionCache *AC, const Instruction *CxtI, const DominatorTree *DT, | ||||||
| 4760 | bool UseInstrInfo) { | ||||||
| 4761 | ConstantRange LHSRange = computeConstantRangeIncludingKnownBits( | ||||||
| 4762 | LHS, /*ForSigned=*/false, DL, /*Depth=*/0, AC, CxtI, DT, | ||||||
| 4763 | nullptr, UseInstrInfo); | ||||||
| 4764 | ConstantRange RHSRange = computeConstantRangeIncludingKnownBits( | ||||||
| 4765 | RHS, /*ForSigned=*/false, DL, /*Depth=*/0, AC, CxtI, DT, | ||||||
| 4766 | nullptr, UseInstrInfo); | ||||||
| 4767 | return mapOverflowResult(LHSRange.unsignedAddMayOverflow(RHSRange)); | ||||||
| 4768 | } | ||||||
| 4769 | |||||||
| 4770 | static OverflowResult computeOverflowForSignedAdd(const Value *LHS, | ||||||
| 4771 | const Value *RHS, | ||||||
| 4772 | const AddOperator *Add, | ||||||
| 4773 | const DataLayout &DL, | ||||||
| 4774 | AssumptionCache *AC, | ||||||
| 4775 | const Instruction *CxtI, | ||||||
| 4776 | const DominatorTree *DT) { | ||||||
| 4777 | if (Add && Add->hasNoSignedWrap()) { | ||||||
| 4778 | return OverflowResult::NeverOverflows; | ||||||
| 4779 | } | ||||||
| 4780 | |||||||
| 4781 | // If LHS and RHS each have at least two sign bits, the addition will look | ||||||
| 4782 | // like | ||||||
| 4783 | // | ||||||
| 4784 | // XX..... + | ||||||
| 4785 | // YY..... | ||||||
| 4786 | // | ||||||
| 4787 | // If the carry into the most significant position is 0, X and Y can't both | ||||||
| 4788 | // be 1 and therefore the carry out of the addition is also 0. | ||||||
| 4789 | // | ||||||
| 4790 | // If the carry into the most significant position is 1, X and Y can't both | ||||||
| 4791 | // be 0 and therefore the carry out of the addition is also 1. | ||||||
| 4792 | // | ||||||
| 4793 | // Since the carry into the most significant position is always equal to | ||||||
| 4794 | // the carry out of the addition, there is no signed overflow. | ||||||
| 4795 | if (ComputeNumSignBits(LHS, DL, 0, AC, CxtI, DT) > 1 && | ||||||
| 4796 | ComputeNumSignBits(RHS, DL, 0, AC, CxtI, DT) > 1) | ||||||
| 4797 | return OverflowResult::NeverOverflows; | ||||||
| 4798 | |||||||
| 4799 | ConstantRange LHSRange = computeConstantRangeIncludingKnownBits( | ||||||
| 4800 | LHS, /*ForSigned=*/true, DL, /*Depth=*/0, AC, CxtI, DT); | ||||||
| 4801 | ConstantRange RHSRange = computeConstantRangeIncludingKnownBits( | ||||||
| 4802 | RHS, /*ForSigned=*/true, DL, /*Depth=*/0, AC, CxtI, DT); | ||||||
| 4803 | OverflowResult OR = | ||||||
| 4804 | mapOverflowResult(LHSRange.signedAddMayOverflow(RHSRange)); | ||||||
| 4805 | if (OR != OverflowResult::MayOverflow) | ||||||
| 4806 | return OR; | ||||||
| 4807 | |||||||
| 4808 | // The remaining code needs Add to be available. Early returns if not so. | ||||||
| 4809 | if (!Add) | ||||||
| 4810 | return OverflowResult::MayOverflow; | ||||||
| 4811 | |||||||
| 4812 | // If the sign of Add is the same as at least one of the operands, this add | ||||||
| 4813 | // CANNOT overflow. If this can be determined from the known bits of the | ||||||
| 4814 | // operands the above signedAddMayOverflow() check will have already done so. | ||||||
| 4815 | // The only other way to improve on the known bits is from an assumption, so | ||||||
| 4816 | // call computeKnownBitsFromAssume() directly. | ||||||
| 4817 | bool LHSOrRHSKnownNonNegative = | ||||||
| 4818 | (LHSRange.isAllNonNegative() || RHSRange.isAllNonNegative()); | ||||||
| 4819 | bool LHSOrRHSKnownNegative = | ||||||
| 4820 | (LHSRange.isAllNegative() || RHSRange.isAllNegative()); | ||||||
| 4821 | if (LHSOrRHSKnownNonNegative || LHSOrRHSKnownNegative) { | ||||||
| 4822 | KnownBits AddKnown(LHSRange.getBitWidth()); | ||||||
| 4823 | computeKnownBitsFromAssume( | ||||||
| 4824 | Add, AddKnown, /*Depth=*/0, Query(DL, AC, CxtI, DT, true)); | ||||||
| 4825 | if ((AddKnown.isNonNegative() && LHSOrRHSKnownNonNegative) || | ||||||
| 4826 | (AddKnown.isNegative() && LHSOrRHSKnownNegative)) | ||||||
| 4827 | return OverflowResult::NeverOverflows; | ||||||
| 4828 | } | ||||||
| 4829 | |||||||
| 4830 | return OverflowResult::MayOverflow; | ||||||
| 4831 | } | ||||||
| 4832 | |||||||
| 4833 | OverflowResult llvm::computeOverflowForUnsignedSub(const Value *LHS, | ||||||
| 4834 | const Value *RHS, | ||||||
| 4835 | const DataLayout &DL, | ||||||
| 4836 | AssumptionCache *AC, | ||||||
| 4837 | const Instruction *CxtI, | ||||||
| 4838 | const DominatorTree *DT) { | ||||||
| 4839 | // Checking for conditions implied by dominating conditions may be expensive. | ||||||
| 4840 | // Limit it to usub_with_overflow calls for now. | ||||||
| 4841 | if (match(CxtI, | ||||||
| 4842 | m_Intrinsic<Intrinsic::usub_with_overflow>(m_Value(), m_Value()))) | ||||||
| 4843 | if (auto C = | ||||||
| 4844 | isImpliedByDomCondition(CmpInst::ICMP_UGE, LHS, RHS, CxtI, DL)) { | ||||||
| 4845 | if (*C) | ||||||
| 4846 | return OverflowResult::NeverOverflows; | ||||||
| 4847 | return OverflowResult::AlwaysOverflowsLow; | ||||||
| 4848 | } | ||||||
| 4849 | ConstantRange LHSRange = computeConstantRangeIncludingKnownBits( | ||||||
| 4850 | LHS, /*ForSigned=*/false, DL, /*Depth=*/0, AC, CxtI, DT); | ||||||
| 4851 | ConstantRange RHSRange = computeConstantRangeIncludingKnownBits( | ||||||
| 4852 | RHS, /*ForSigned=*/false, DL, /*Depth=*/0, AC, CxtI, DT); | ||||||
| 4853 | return mapOverflowResult(LHSRange.unsignedSubMayOverflow(RHSRange)); | ||||||
| 4854 | } | ||||||
| 4855 | |||||||
| 4856 | OverflowResult llvm::computeOverflowForSignedSub(const Value *LHS, | ||||||
| 4857 | const Value *RHS, | ||||||
| 4858 | const DataLayout &DL, | ||||||
| 4859 | AssumptionCache *AC, | ||||||
| 4860 | const Instruction *CxtI, | ||||||
| 4861 | const DominatorTree *DT) { | ||||||
| 4862 | // If LHS and RHS each have at least two sign bits, the subtraction | ||||||
| 4863 | // cannot overflow. | ||||||
| 4864 | if (ComputeNumSignBits(LHS, DL, 0, AC, CxtI, DT) > 1 && | ||||||
| 4865 | ComputeNumSignBits(RHS, DL, 0, AC, CxtI, DT) > 1) | ||||||
| 4866 | return OverflowResult::NeverOverflows; | ||||||
| 4867 | |||||||
| 4868 | ConstantRange LHSRange = computeConstantRangeIncludingKnownBits( | ||||||
| 4869 | LHS, /*ForSigned=*/true, DL, /*Depth=*/0, AC, CxtI, DT); | ||||||
| 4870 | ConstantRange RHSRange = computeConstantRangeIncludingKnownBits( | ||||||
| 4871 | RHS, /*ForSigned=*/true, DL, /*Depth=*/0, AC, CxtI, DT); | ||||||
| 4872 | return mapOverflowResult(LHSRange.signedSubMayOverflow(RHSRange)); | ||||||
| 4873 | } | ||||||
| 4874 | |||||||
| 4875 | bool llvm::isOverflowIntrinsicNoWrap(const WithOverflowInst *WO, | ||||||
| 4876 | const DominatorTree &DT) { | ||||||
| 4877 | SmallVector<const BranchInst *, 2> GuardingBranches; | ||||||
| 4878 | SmallVector<const ExtractValueInst *, 2> Results; | ||||||
| 4879 | |||||||
| 4880 | for (const User *U : WO->users()) { | ||||||
| 4881 | if (const auto *EVI = dyn_cast<ExtractValueInst>(U)) { | ||||||
| 4882 | assert(EVI->getNumIndices() == 1 && "Obvious from CI's type")((void)0); | ||||||
| 4883 | |||||||
| 4884 | if (EVI->getIndices()[0] == 0) | ||||||
| 4885 | Results.push_back(EVI); | ||||||
| 4886 | else { | ||||||
| 4887 | assert(EVI->getIndices()[0] == 1 && "Obvious from CI's type")((void)0); | ||||||
| 4888 | |||||||
| 4889 | for (const auto *U : EVI->users()) | ||||||
| 4890 | if (const auto *B = dyn_cast<BranchInst>(U)) { | ||||||
| 4891 | assert(B->isConditional() && "How else is it using an i1?")((void)0); | ||||||
| 4892 | GuardingBranches.push_back(B); | ||||||
| 4893 | } | ||||||
| 4894 | } | ||||||
| 4895 | } else { | ||||||
| 4896 | // We are using the aggregate directly in a way we don't want to analyze | ||||||
| 4897 | // here (storing it to a global, say). | ||||||
| 4898 | return false; | ||||||
| 4899 | } | ||||||
| 4900 | } | ||||||
| 4901 | |||||||
| 4902 | auto AllUsesGuardedByBranch = [&](const BranchInst *BI) { | ||||||
| 4903 | BasicBlockEdge NoWrapEdge(BI->getParent(), BI->getSuccessor(1)); | ||||||
| 4904 | if (!NoWrapEdge.isSingleEdge()) | ||||||
| 4905 | return false; | ||||||
| 4906 | |||||||
| 4907 | // Check if all users of the add are provably no-wrap. | ||||||
| 4908 | for (const auto *Result : Results) { | ||||||
| 4909 | // If the extractvalue itself is not executed on overflow, the we don't | ||||||
| 4910 | // need to check each use separately, since domination is transitive. | ||||||
| 4911 | if (DT.dominates(NoWrapEdge, Result->getParent())) | ||||||
| 4912 | continue; | ||||||
| 4913 | |||||||
| 4914 | for (auto &RU : Result->uses()) | ||||||
| 4915 | if (!DT.dominates(NoWrapEdge, RU)) | ||||||
| 4916 | return false; | ||||||
| 4917 | } | ||||||
| 4918 | |||||||
| 4919 | return true; | ||||||
| 4920 | }; | ||||||
| 4921 | |||||||
| 4922 | return llvm::any_of(GuardingBranches, AllUsesGuardedByBranch); | ||||||
| 4923 | } | ||||||
| 4924 | |||||||
| 4925 | static bool canCreateUndefOrPoison(const Operator *Op, bool PoisonOnly) { | ||||||
| 4926 | // See whether I has flags that may create poison | ||||||
| 4927 | if (const auto *OvOp = dyn_cast<OverflowingBinaryOperator>(Op)) { | ||||||
| 4928 | if (OvOp->hasNoSignedWrap() || OvOp->hasNoUnsignedWrap()) | ||||||
| 4929 | return true; | ||||||
| 4930 | } | ||||||
| 4931 | if (const auto *ExactOp = dyn_cast<PossiblyExactOperator>(Op)) | ||||||
| 4932 | if (ExactOp->isExact()) | ||||||
| 4933 | return true; | ||||||
| 4934 | if (const auto *FP = dyn_cast<FPMathOperator>(Op)) { | ||||||
| 4935 | auto FMF = FP->getFastMathFlags(); | ||||||
| 4936 | if (FMF.noNaNs() || FMF.noInfs()) | ||||||
| 4937 | return true; | ||||||
| 4938 | } | ||||||
| 4939 | |||||||
| 4940 | unsigned Opcode = Op->getOpcode(); | ||||||
| 4941 | |||||||
| 4942 | // Check whether opcode is a poison/undef-generating operation | ||||||
| 4943 | switch (Opcode) { | ||||||
| 4944 | case Instruction::Shl: | ||||||
| 4945 | case Instruction::AShr: | ||||||
| 4946 | case Instruction::LShr: { | ||||||
| 4947 | // Shifts return poison if shiftwidth is larger than the bitwidth. | ||||||
| 4948 | if (auto *C = dyn_cast<Constant>(Op->getOperand(1))) { | ||||||
| 4949 | SmallVector<Constant *, 4> ShiftAmounts; | ||||||
| 4950 | if (auto *FVTy = dyn_cast<FixedVectorType>(C->getType())) { | ||||||
| 4951 | unsigned NumElts = FVTy->getNumElements(); | ||||||
| 4952 | for (unsigned i = 0; i < NumElts; ++i) | ||||||
| 4953 | ShiftAmounts.push_back(C->getAggregateElement(i)); | ||||||
| 4954 | } else if (isa<ScalableVectorType>(C->getType())) | ||||||
| 4955 | return true; // Can't tell, just return true to be safe | ||||||
| 4956 | else | ||||||
| 4957 | ShiftAmounts.push_back(C); | ||||||
| 4958 | |||||||
| 4959 | bool Safe = llvm::all_of(ShiftAmounts, [](Constant *C) { | ||||||
| 4960 | auto *CI = dyn_cast_or_null<ConstantInt>(C); | ||||||
| 4961 | return CI && CI->getValue().ult(C->getType()->getIntegerBitWidth()); | ||||||
| 4962 | }); | ||||||
| 4963 | return !Safe; | ||||||
| 4964 | } | ||||||
| 4965 | return true; | ||||||
| 4966 | } | ||||||
| 4967 | case Instruction::FPToSI: | ||||||
| 4968 | case Instruction::FPToUI: | ||||||
| 4969 | // fptosi/ui yields poison if the resulting value does not fit in the | ||||||
| 4970 | // destination type. | ||||||
| 4971 | return true; | ||||||
| 4972 | case Instruction::Call: | ||||||
| 4973 | if (auto *II = dyn_cast<IntrinsicInst>(Op)) { | ||||||
| 4974 | switch (II->getIntrinsicID()) { | ||||||
| 4975 | // TODO: Add more intrinsics. | ||||||
| 4976 | case Intrinsic::ctpop: | ||||||
| 4977 | case Intrinsic::sadd_with_overflow: | ||||||
| 4978 | case Intrinsic::ssub_with_overflow: | ||||||
| 4979 | case Intrinsic::smul_with_overflow: | ||||||
| 4980 | case Intrinsic::uadd_with_overflow: | ||||||
| 4981 | case Intrinsic::usub_with_overflow: | ||||||
| 4982 | case Intrinsic::umul_with_overflow: | ||||||
| 4983 | return false; | ||||||
| 4984 | } | ||||||
| 4985 | } | ||||||
| 4986 | LLVM_FALLTHROUGH[[gnu::fallthrough]]; | ||||||
| 4987 | case Instruction::CallBr: | ||||||
| 4988 | case Instruction::Invoke: { | ||||||
| 4989 | const auto *CB = cast<CallBase>(Op); | ||||||
| 4990 | return !CB->hasRetAttr(Attribute::NoUndef); | ||||||
| 4991 | } | ||||||
| 4992 | case Instruction::InsertElement: | ||||||
| 4993 | case Instruction::ExtractElement: { | ||||||
| 4994 | // If index exceeds the length of the vector, it returns poison | ||||||
| 4995 | auto *VTy = cast<VectorType>(Op->getOperand(0)->getType()); | ||||||
| 4996 | unsigned IdxOp = Op->getOpcode() == Instruction::InsertElement ? 2 : 1; | ||||||
| 4997 | auto *Idx = dyn_cast<ConstantInt>(Op->getOperand(IdxOp)); | ||||||
| 4998 | if (!Idx || Idx->getValue().uge(VTy->getElementCount().getKnownMinValue())) | ||||||
| 4999 | return true; | ||||||
| 5000 | return false; | ||||||
| 5001 | } | ||||||
| 5002 | case Instruction::ShuffleVector: { | ||||||
| 5003 | // shufflevector may return undef. | ||||||
| 5004 | if (PoisonOnly) | ||||||
| 5005 | return false; | ||||||
| 5006 | ArrayRef<int> Mask = isa<ConstantExpr>(Op) | ||||||
| 5007 | ? cast<ConstantExpr>(Op)->getShuffleMask() | ||||||
| 5008 | : cast<ShuffleVectorInst>(Op)->getShuffleMask(); | ||||||
| 5009 | return is_contained(Mask, UndefMaskElem); | ||||||
| 5010 | } | ||||||
| 5011 | case Instruction::FNeg: | ||||||
| 5012 | case Instruction::PHI: | ||||||
| 5013 | case Instruction::Select: | ||||||
| 5014 | case Instruction::URem: | ||||||
| 5015 | case Instruction::SRem: | ||||||
| 5016 | case Instruction::ExtractValue: | ||||||
| 5017 | case Instruction::InsertValue: | ||||||
| 5018 | case Instruction::Freeze: | ||||||
| 5019 | case Instruction::ICmp: | ||||||
| 5020 | case Instruction::FCmp: | ||||||
| 5021 | return false; | ||||||
| 5022 | case Instruction::GetElementPtr: { | ||||||
| 5023 | const auto *GEP = cast<GEPOperator>(Op); | ||||||
| 5024 | return GEP->isInBounds(); | ||||||
| 5025 | } | ||||||
| 5026 | default: { | ||||||
| 5027 | const auto *CE = dyn_cast<ConstantExpr>(Op); | ||||||
| 5028 | if (isa<CastInst>(Op) || (CE && CE->isCast())) | ||||||
| 5029 | return false; | ||||||
| 5030 | else if (Instruction::isBinaryOp(Opcode)) | ||||||
| 5031 | return false; | ||||||
| 5032 | // Be conservative and return true. | ||||||
| 5033 | return true; | ||||||
| 5034 | } | ||||||
| 5035 | } | ||||||
| 5036 | } | ||||||
| 5037 | |||||||
| 5038 | bool llvm::canCreateUndefOrPoison(const Operator *Op) { | ||||||
| 5039 | return ::canCreateUndefOrPoison(Op, /*PoisonOnly=*/false); | ||||||
| 5040 | } | ||||||
| 5041 | |||||||
| 5042 | bool llvm::canCreatePoison(const Operator *Op) { | ||||||
| 5043 | return ::canCreateUndefOrPoison(Op, /*PoisonOnly=*/true); | ||||||
| 5044 | } | ||||||
| 5045 | |||||||
| 5046 | static bool directlyImpliesPoison(const Value *ValAssumedPoison, | ||||||
| 5047 | const Value *V, unsigned Depth) { | ||||||
| 5048 | if (ValAssumedPoison == V) | ||||||
| 5049 | return true; | ||||||
| 5050 | |||||||
| 5051 | const unsigned MaxDepth = 2; | ||||||
| 5052 | if (Depth >= MaxDepth) | ||||||
| 5053 | return false; | ||||||
| 5054 | |||||||
| 5055 | if (const auto *I = dyn_cast<Instruction>(V)) { | ||||||
| 5056 | if (propagatesPoison(cast<Operator>(I))) | ||||||
| 5057 | return any_of(I->operands(), [=](const Value *Op) { | ||||||
| 5058 | return directlyImpliesPoison(ValAssumedPoison, Op, Depth + 1); | ||||||
| 5059 | }); | ||||||
| 5060 | |||||||
| 5061 | // 'select ValAssumedPoison, _, _' is poison. | ||||||
| 5062 | if (const auto *SI = dyn_cast<SelectInst>(I)) | ||||||
| 5063 | return directlyImpliesPoison(ValAssumedPoison, SI->getCondition(), | ||||||
| 5064 | Depth + 1); | ||||||
| 5065 | // V = extractvalue V0, idx | ||||||
| 5066 | // V2 = extractvalue V0, idx2 | ||||||
| 5067 | // V0's elements are all poison or not. (e.g., add_with_overflow) | ||||||
| 5068 | const WithOverflowInst *II; | ||||||
| 5069 | if (match(I, m_ExtractValue(m_WithOverflowInst(II))) && | ||||||
| 5070 | (match(ValAssumedPoison, m_ExtractValue(m_Specific(II))) || | ||||||
| 5071 | llvm::is_contained(II->arg_operands(), ValAssumedPoison))) | ||||||
| 5072 | return true; | ||||||
| 5073 | } | ||||||
| 5074 | return false; | ||||||
| 5075 | } | ||||||
| 5076 | |||||||
| 5077 | static bool impliesPoison(const Value *ValAssumedPoison, const Value *V, | ||||||
| 5078 | unsigned Depth) { | ||||||
| 5079 | if (isGuaranteedNotToBeUndefOrPoison(ValAssumedPoison)) | ||||||
| 5080 | return true; | ||||||
| 5081 | |||||||
| 5082 | if (directlyImpliesPoison(ValAssumedPoison, V, /* Depth */ 0)) | ||||||
| 5083 | return true; | ||||||
| 5084 | |||||||
| 5085 | const unsigned MaxDepth = 2; | ||||||
| 5086 | if (Depth >= MaxDepth) | ||||||
| 5087 | return false; | ||||||
| 5088 | |||||||
| 5089 | const auto *I = dyn_cast<Instruction>(ValAssumedPoison); | ||||||
| 5090 | if (I && !canCreatePoison(cast<Operator>(I))) { | ||||||
| 5091 | return all_of(I->operands(), [=](const Value *Op) { | ||||||
| 5092 | return impliesPoison(Op, V, Depth + 1); | ||||||
| 5093 | }); | ||||||
| 5094 | } | ||||||
| 5095 | return false; | ||||||
| 5096 | } | ||||||
| 5097 | |||||||
| 5098 | bool llvm::impliesPoison(const Value *ValAssumedPoison, const Value *V) { | ||||||
| 5099 | return ::impliesPoison(ValAssumedPoison, V, /* Depth */ 0); | ||||||
| 5100 | } | ||||||
| 5101 | |||||||
| 5102 | static bool programUndefinedIfUndefOrPoison(const Value *V, | ||||||
| 5103 | bool PoisonOnly); | ||||||
| 5104 | |||||||
| 5105 | static bool isGuaranteedNotToBeUndefOrPoison(const Value *V, | ||||||
| 5106 | AssumptionCache *AC, | ||||||
| 5107 | const Instruction *CtxI, | ||||||
| 5108 | const DominatorTree *DT, | ||||||
| 5109 | unsigned Depth, bool PoisonOnly) { | ||||||
| 5110 | if (Depth >= MaxAnalysisRecursionDepth) | ||||||
| 5111 | return false; | ||||||
| 5112 | |||||||
| 5113 | if (isa<MetadataAsValue>(V)) | ||||||
| 5114 | return false; | ||||||
| 5115 | |||||||
| 5116 | if (const auto *A = dyn_cast<Argument>(V)) { | ||||||
| 5117 | if (A->hasAttribute(Attribute::NoUndef)) | ||||||
| 5118 | return true; | ||||||
| 5119 | } | ||||||
| 5120 | |||||||
| 5121 | if (auto *C = dyn_cast<Constant>(V)) { | ||||||
| 5122 | if (isa<UndefValue>(C)) | ||||||
| 5123 | return PoisonOnly && !isa<PoisonValue>(C); | ||||||
| 5124 | |||||||
| 5125 | if (isa<ConstantInt>(C) || isa<GlobalVariable>(C) || isa<ConstantFP>(V) || | ||||||
| 5126 | isa<ConstantPointerNull>(C) || isa<Function>(C)) | ||||||
| 5127 | return true; | ||||||
| 5128 | |||||||
| 5129 | if (C->getType()->isVectorTy() && !isa<ConstantExpr>(C)) | ||||||
| 5130 | return (PoisonOnly ? !C->containsPoisonElement() | ||||||
| 5131 | : !C->containsUndefOrPoisonElement()) && | ||||||
| 5132 | !C->containsConstantExpression(); | ||||||
| 5133 | } | ||||||
| 5134 | |||||||
| 5135 | // Strip cast operations from a pointer value. | ||||||
| 5136 | // Note that stripPointerCastsSameRepresentation can strip off getelementptr | ||||||
| 5137 | // inbounds with zero offset. To guarantee that the result isn't poison, the | ||||||
| 5138 | // stripped pointer is checked as it has to be pointing into an allocated | ||||||
| 5139 | // object or be null `null` to ensure `inbounds` getelement pointers with a | ||||||
| 5140 | // zero offset could not produce poison. | ||||||
| 5141 | // It can strip off addrspacecast that do not change bit representation as | ||||||
| 5142 | // well. We believe that such addrspacecast is equivalent to no-op. | ||||||
| 5143 | auto *StrippedV = V->stripPointerCastsSameRepresentation(); | ||||||
| 5144 | if (isa<AllocaInst>(StrippedV) || isa<GlobalVariable>(StrippedV) || | ||||||
| 5145 | isa<Function>(StrippedV) || isa<ConstantPointerNull>(StrippedV)) | ||||||
| 5146 | return true; | ||||||
| 5147 | |||||||
| 5148 | auto OpCheck = [&](const Value *V) { | ||||||
| 5149 | return isGuaranteedNotToBeUndefOrPoison(V, AC, CtxI, DT, Depth + 1, | ||||||
| 5150 | PoisonOnly); | ||||||
| 5151 | }; | ||||||
| 5152 | |||||||
| 5153 | if (auto *Opr = dyn_cast<Operator>(V)) { | ||||||
| 5154 | // If the value is a freeze instruction, then it can never | ||||||
| 5155 | // be undef or poison. | ||||||
| 5156 | if (isa<FreezeInst>(V)) | ||||||
| 5157 | return true; | ||||||
| 5158 | |||||||
| 5159 | if (const auto *CB = dyn_cast<CallBase>(V)) { | ||||||
| 5160 | if (CB->hasRetAttr(Attribute::NoUndef)) | ||||||
| 5161 | return true; | ||||||
| 5162 | } | ||||||
| 5163 | |||||||
| 5164 | if (const auto *PN = dyn_cast<PHINode>(V)) { | ||||||
| 5165 | unsigned Num = PN->getNumIncomingValues(); | ||||||
| 5166 | bool IsWellDefined = true; | ||||||
| 5167 | for (unsigned i = 0; i < Num; ++i) { | ||||||
| 5168 | auto *TI = PN->getIncomingBlock(i)->getTerminator(); | ||||||
| 5169 | if (!isGuaranteedNotToBeUndefOrPoison(PN->getIncomingValue(i), AC, TI, | ||||||
| 5170 | DT, Depth + 1, PoisonOnly)) { | ||||||
| 5171 | IsWellDefined = false; | ||||||
| 5172 | break; | ||||||
| 5173 | } | ||||||
| 5174 | } | ||||||
| 5175 | if (IsWellDefined) | ||||||
| 5176 | return true; | ||||||
| 5177 | } else if (!canCreateUndefOrPoison(Opr) && all_of(Opr->operands(), OpCheck)) | ||||||
| 5178 | return true; | ||||||
| 5179 | } | ||||||
| 5180 | |||||||
| 5181 | if (auto *I = dyn_cast<LoadInst>(V)) | ||||||
| 5182 | if (I->getMetadata(LLVMContext::MD_noundef)) | ||||||
| 5183 | return true; | ||||||
| 5184 | |||||||
| 5185 | if (programUndefinedIfUndefOrPoison(V, PoisonOnly)) | ||||||
| 5186 | return true; | ||||||
| 5187 | |||||||
| 5188 | // CxtI may be null or a cloned instruction. | ||||||
| 5189 | if (!CtxI || !CtxI->getParent() || !DT) | ||||||
| 5190 | return false; | ||||||
| 5191 | |||||||
| 5192 | auto *DNode = DT->getNode(CtxI->getParent()); | ||||||
| 5193 | if (!DNode) | ||||||
| 5194 | // Unreachable block | ||||||
| 5195 | return false; | ||||||
| 5196 | |||||||
| 5197 | // If V is used as a branch condition before reaching CtxI, V cannot be | ||||||
| 5198 | // undef or poison. | ||||||
| 5199 | // br V, BB1, BB2 | ||||||
| 5200 | // BB1: | ||||||
| 5201 | // CtxI ; V cannot be undef or poison here | ||||||
| 5202 | auto *Dominator = DNode->getIDom(); | ||||||
| 5203 | while (Dominator) { | ||||||
| 5204 | auto *TI = Dominator->getBlock()->getTerminator(); | ||||||
| 5205 | |||||||
| 5206 | Value *Cond = nullptr; | ||||||
| 5207 | if (auto BI = dyn_cast<BranchInst>(TI)) { | ||||||
| 5208 | if (BI->isConditional()) | ||||||
| 5209 | Cond = BI->getCondition(); | ||||||
| 5210 | } else if (auto SI = dyn_cast<SwitchInst>(TI)) { | ||||||
| 5211 | Cond = SI->getCondition(); | ||||||
| 5212 | } | ||||||
| 5213 | |||||||
| 5214 | if (Cond) { | ||||||
| 5215 | if (Cond == V) | ||||||
| 5216 | return true; | ||||||
| 5217 | else if (PoisonOnly && isa<Operator>(Cond)) { | ||||||
| 5218 | // For poison, we can analyze further | ||||||
| 5219 | auto *Opr = cast<Operator>(Cond); | ||||||
| 5220 | if (propagatesPoison(Opr) && is_contained(Opr->operand_values(), V)) | ||||||
| 5221 | return true; | ||||||
| 5222 | } | ||||||
| 5223 | } | ||||||
| 5224 | |||||||
| 5225 | Dominator = Dominator->getIDom(); | ||||||
| 5226 | } | ||||||
| 5227 | |||||||
| 5228 | SmallVector<Attribute::AttrKind, 2> AttrKinds{Attribute::NoUndef}; | ||||||
| 5229 | if (getKnowledgeValidInContext(V, AttrKinds, CtxI, DT, AC)) | ||||||
| 5230 | return true; | ||||||
| 5231 | |||||||
| 5232 | return false; | ||||||
| 5233 | } | ||||||
| 5234 | |||||||
| 5235 | bool llvm::isGuaranteedNotToBeUndefOrPoison(const Value *V, AssumptionCache *AC, | ||||||
| 5236 | const Instruction *CtxI, | ||||||
| 5237 | const DominatorTree *DT, | ||||||
| 5238 | unsigned Depth) { | ||||||
| 5239 | return ::isGuaranteedNotToBeUndefOrPoison(V, AC, CtxI, DT, Depth, false); | ||||||
| 5240 | } | ||||||
| 5241 | |||||||
| 5242 | bool llvm::isGuaranteedNotToBePoison(const Value *V, AssumptionCache *AC, | ||||||
| 5243 | const Instruction *CtxI, | ||||||
| 5244 | const DominatorTree *DT, unsigned Depth) { | ||||||
| 5245 | return ::isGuaranteedNotToBeUndefOrPoison(V, AC, CtxI, DT, Depth, true); | ||||||
| 5246 | } | ||||||
| 5247 | |||||||
| 5248 | OverflowResult llvm::computeOverflowForSignedAdd(const AddOperator *Add, | ||||||
| 5249 | const DataLayout &DL, | ||||||
| 5250 | AssumptionCache *AC, | ||||||
| 5251 | const Instruction *CxtI, | ||||||
| 5252 | const DominatorTree *DT) { | ||||||
| 5253 | return ::computeOverflowForSignedAdd(Add->getOperand(0), Add->getOperand(1), | ||||||
| 5254 | Add, DL, AC, CxtI, DT); | ||||||
| 5255 | } | ||||||
| 5256 | |||||||
| 5257 | OverflowResult llvm::computeOverflowForSignedAdd(const Value *LHS, | ||||||
| 5258 | const Value *RHS, | ||||||
| 5259 | const DataLayout &DL, | ||||||
| 5260 | AssumptionCache *AC, | ||||||
| 5261 | const Instruction *CxtI, | ||||||
| 5262 | const DominatorTree *DT) { | ||||||
| 5263 | return ::computeOverflowForSignedAdd(LHS, RHS, nullptr, DL, AC, CxtI, DT); | ||||||
| 5264 | } | ||||||
| 5265 | |||||||
| 5266 | bool llvm::isGuaranteedToTransferExecutionToSuccessor(const Instruction *I) { | ||||||
| 5267 | // Note: An atomic operation isn't guaranteed to return in a reasonable amount | ||||||
| 5268 | // of time because it's possible for another thread to interfere with it for an | ||||||
| 5269 | // arbitrary length of time, but programs aren't allowed to rely on that. | ||||||
| 5270 | |||||||
| 5271 | // If there is no successor, then execution can't transfer to it. | ||||||
| 5272 | if (isa<ReturnInst>(I)) | ||||||
| 5273 | return false; | ||||||
| 5274 | if (isa<UnreachableInst>(I)) | ||||||
| 5275 | return false; | ||||||
| 5276 | |||||||
| 5277 | // Note: Do not add new checks here; instead, change Instruction::mayThrow or | ||||||
| 5278 | // Instruction::willReturn. | ||||||
| 5279 | // | ||||||
| 5280 | // FIXME: Move this check into Instruction::willReturn. | ||||||
| 5281 | if (isa<CatchPadInst>(I)) { | ||||||
| 5282 | switch (classifyEHPersonality(I->getFunction()->getPersonalityFn())) { | ||||||
| 5283 | default: | ||||||
| 5284 | // A catchpad may invoke exception object constructors and such, which | ||||||
| 5285 | // in some languages can be arbitrary code, so be conservative by default. | ||||||
| 5286 | return false; | ||||||
| 5287 | case EHPersonality::CoreCLR: | ||||||
| 5288 | // For CoreCLR, it just involves a type test. | ||||||
| 5289 | return true; | ||||||
| 5290 | } | ||||||
| 5291 | } | ||||||
| 5292 | |||||||
| 5293 | // An instruction that returns without throwing must transfer control flow | ||||||
| 5294 | // to a successor. | ||||||
| 5295 | return !I->mayThrow() && I->willReturn(); | ||||||
| 5296 | } | ||||||
| 5297 | |||||||
| 5298 | bool llvm::isGuaranteedToTransferExecutionToSuccessor(const BasicBlock *BB) { | ||||||
| 5299 | // TODO: This is slightly conservative for invoke instruction since exiting | ||||||
| 5300 | // via an exception *is* normal control for them. | ||||||
| 5301 | for (const Instruction &I : *BB) | ||||||
| 5302 | if (!isGuaranteedToTransferExecutionToSuccessor(&I)) | ||||||
| 5303 | return false; | ||||||
| 5304 | return true; | ||||||
| 5305 | } | ||||||
| 5306 | |||||||
| 5307 | bool llvm::isGuaranteedToExecuteForEveryIteration(const Instruction *I, | ||||||
| 5308 | const Loop *L) { | ||||||
| 5309 | // The loop header is guaranteed to be executed for every iteration. | ||||||
| 5310 | // | ||||||
| 5311 | // FIXME: Relax this constraint to cover all basic blocks that are | ||||||
| 5312 | // guaranteed to be executed at every iteration. | ||||||
| 5313 | if (I->getParent() != L->getHeader()) return false; | ||||||
| 5314 | |||||||
| 5315 | for (const Instruction &LI : *L->getHeader()) { | ||||||
| 5316 | if (&LI == I) return true; | ||||||
| 5317 | if (!isGuaranteedToTransferExecutionToSuccessor(&LI)) return false; | ||||||
| 5318 | } | ||||||
| 5319 | llvm_unreachable("Instruction not contained in its own parent basic block.")__builtin_unreachable(); | ||||||
| 5320 | } | ||||||
| 5321 | |||||||
| 5322 | bool llvm::propagatesPoison(const Operator *I) { | ||||||
| 5323 | switch (I->getOpcode()) { | ||||||
| 5324 | case Instruction::Freeze: | ||||||
| 5325 | case Instruction::Select: | ||||||
| 5326 | case Instruction::PHI: | ||||||
| 5327 | case Instruction::Invoke: | ||||||
| 5328 | return false; | ||||||
| 5329 | case Instruction::Call: | ||||||
| 5330 | if (auto *II = dyn_cast<IntrinsicInst>(I)) { | ||||||
| 5331 | switch (II->getIntrinsicID()) { | ||||||
| 5332 | // TODO: Add more intrinsics. | ||||||
| 5333 | case Intrinsic::sadd_with_overflow: | ||||||
| 5334 | case Intrinsic::ssub_with_overflow: | ||||||
| 5335 | case Intrinsic::smul_with_overflow: | ||||||
| 5336 | case Intrinsic::uadd_with_overflow: | ||||||
| 5337 | case Intrinsic::usub_with_overflow: | ||||||
| 5338 | case Intrinsic::umul_with_overflow: | ||||||
| 5339 | // If an input is a vector containing a poison element, the | ||||||
| 5340 | // two output vectors (calculated results, overflow bits)' | ||||||
| 5341 | // corresponding lanes are poison. | ||||||
| 5342 | return true; | ||||||
| 5343 | case Intrinsic::ctpop: | ||||||
| 5344 | return true; | ||||||
| 5345 | } | ||||||
| 5346 | } | ||||||
| 5347 | return false; | ||||||
| 5348 | case Instruction::ICmp: | ||||||
| 5349 | case Instruction::FCmp: | ||||||
| 5350 | case Instruction::GetElementPtr: | ||||||
| 5351 | return true; | ||||||
| 5352 | default: | ||||||
| 5353 | if (isa<BinaryOperator>(I) || isa<UnaryOperator>(I) || isa<CastInst>(I)) | ||||||
| 5354 | return true; | ||||||
| 5355 | |||||||
| 5356 | // Be conservative and return false. | ||||||
| 5357 | return false; | ||||||
| 5358 | } | ||||||
| 5359 | } | ||||||
| 5360 | |||||||
| 5361 | void llvm::getGuaranteedWellDefinedOps( | ||||||
| 5362 | const Instruction *I, SmallPtrSetImpl<const Value *> &Operands) { | ||||||
| 5363 | switch (I->getOpcode()) { | ||||||
| 5364 | case Instruction::Store: | ||||||
| 5365 | Operands.insert(cast<StoreInst>(I)->getPointerOperand()); | ||||||
| 5366 | break; | ||||||
| 5367 | |||||||
| 5368 | case Instruction::Load: | ||||||
| 5369 | Operands.insert(cast<LoadInst>(I)->getPointerOperand()); | ||||||
| 5370 | break; | ||||||
| 5371 | |||||||
| 5372 | // Since dereferenceable attribute imply noundef, atomic operations | ||||||
| 5373 | // also implicitly have noundef pointers too | ||||||
| 5374 | case Instruction::AtomicCmpXchg: | ||||||
| 5375 | Operands.insert(cast<AtomicCmpXchgInst>(I)->getPointerOperand()); | ||||||
| 5376 | break; | ||||||
| 5377 | |||||||
| 5378 | case Instruction::AtomicRMW: | ||||||
| 5379 | Operands.insert(cast<AtomicRMWInst>(I)->getPointerOperand()); | ||||||
| 5380 | break; | ||||||
| 5381 | |||||||
| 5382 | case Instruction::Call: | ||||||
| 5383 | case Instruction::Invoke: { | ||||||
| 5384 | const CallBase *CB = cast<CallBase>(I); | ||||||
| 5385 | if (CB->isIndirectCall()) | ||||||
| 5386 | Operands.insert(CB->getCalledOperand()); | ||||||
| 5387 | for (unsigned i = 0; i < CB->arg_size(); ++i) { | ||||||
| 5388 | if (CB->paramHasAttr(i, Attribute::NoUndef) || | ||||||
| 5389 | CB->paramHasAttr(i, Attribute::Dereferenceable)) | ||||||
| 5390 | Operands.insert(CB->getArgOperand(i)); | ||||||
| 5391 | } | ||||||
| 5392 | break; | ||||||
| 5393 | } | ||||||
| 5394 | |||||||
| 5395 | default: | ||||||
| 5396 | break; | ||||||
| 5397 | } | ||||||
| 5398 | } | ||||||
| 5399 | |||||||
| 5400 | void llvm::getGuaranteedNonPoisonOps(const Instruction *I, | ||||||
| 5401 | SmallPtrSetImpl<const Value *> &Operands) { | ||||||
| 5402 | getGuaranteedWellDefinedOps(I, Operands); | ||||||
| 5403 | switch (I->getOpcode()) { | ||||||
| 5404 | // Divisors of these operations are allowed to be partially undef. | ||||||
| 5405 | case Instruction::UDiv: | ||||||
| 5406 | case Instruction::SDiv: | ||||||
| 5407 | case Instruction::URem: | ||||||
| 5408 | case Instruction::SRem: | ||||||
| 5409 | Operands.insert(I->getOperand(1)); | ||||||
| 5410 | break; | ||||||
| 5411 | |||||||
| 5412 | default: | ||||||
| 5413 | break; | ||||||
| 5414 | } | ||||||
| 5415 | } | ||||||
| 5416 | |||||||
| 5417 | bool llvm::mustTriggerUB(const Instruction *I, | ||||||
| 5418 | const SmallSet<const Value *, 16>& KnownPoison) { | ||||||
| 5419 | SmallPtrSet<const Value *, 4> NonPoisonOps; | ||||||
| 5420 | getGuaranteedNonPoisonOps(I, NonPoisonOps); | ||||||
| 5421 | |||||||
| 5422 | for (const auto *V : NonPoisonOps) | ||||||
| 5423 | if (KnownPoison.count(V)) | ||||||
| 5424 | return true; | ||||||
| 5425 | |||||||
| 5426 | return false; | ||||||
| 5427 | } | ||||||
| 5428 | |||||||
| 5429 | static bool programUndefinedIfUndefOrPoison(const Value *V, | ||||||
| 5430 | bool PoisonOnly) { | ||||||
| 5431 | // We currently only look for uses of values within the same basic | ||||||
| 5432 | // block, as that makes it easier to guarantee that the uses will be | ||||||
| 5433 | // executed given that Inst is executed. | ||||||
| 5434 | // | ||||||
| 5435 | // FIXME: Expand this to consider uses beyond the same basic block. To do | ||||||
| 5436 | // this, look out for the distinction between post-dominance and strong | ||||||
| 5437 | // post-dominance. | ||||||
| 5438 | const BasicBlock *BB = nullptr; | ||||||
| 5439 | BasicBlock::const_iterator Begin; | ||||||
| 5440 | if (const auto *Inst = dyn_cast<Instruction>(V)) { | ||||||
| 5441 | BB = Inst->getParent(); | ||||||
| 5442 | Begin = Inst->getIterator(); | ||||||
| 5443 | Begin++; | ||||||
| 5444 | } else if (const auto *Arg = dyn_cast<Argument>(V)) { | ||||||
| 5445 | BB = &Arg->getParent()->getEntryBlock(); | ||||||
| 5446 | Begin = BB->begin(); | ||||||
| 5447 | } else { | ||||||
| 5448 | return false; | ||||||
| 5449 | } | ||||||
| 5450 | |||||||
| 5451 | // Limit number of instructions we look at, to avoid scanning through large | ||||||
| 5452 | // blocks. The current limit is chosen arbitrarily. | ||||||
| 5453 | unsigned ScanLimit = 32; | ||||||
| 5454 | BasicBlock::const_iterator End = BB->end(); | ||||||
| 5455 | |||||||
| 5456 | if (!PoisonOnly) { | ||||||
| 5457 | // Since undef does not propagate eagerly, be conservative & just check | ||||||
| 5458 | // whether a value is directly passed to an instruction that must take | ||||||
| 5459 | // well-defined operands. | ||||||
| 5460 | |||||||
| 5461 | for (auto &I : make_range(Begin, End)) { | ||||||
| 5462 | if (isa<DbgInfoIntrinsic>(I)) | ||||||
| 5463 | continue; | ||||||
| 5464 | if (--ScanLimit == 0) | ||||||
| 5465 | break; | ||||||
| 5466 | |||||||
| 5467 | SmallPtrSet<const Value *, 4> WellDefinedOps; | ||||||
| 5468 | getGuaranteedWellDefinedOps(&I, WellDefinedOps); | ||||||
| 5469 | if (WellDefinedOps.contains(V)) | ||||||
| 5470 | return true; | ||||||
| 5471 | |||||||
| 5472 | if (!isGuaranteedToTransferExecutionToSuccessor(&I)) | ||||||
| 5473 | break; | ||||||
| 5474 | } | ||||||
| 5475 | return false; | ||||||
| 5476 | } | ||||||
| 5477 | |||||||
| 5478 | // Set of instructions that we have proved will yield poison if Inst | ||||||
| 5479 | // does. | ||||||
| 5480 | SmallSet<const Value *, 16> YieldsPoison; | ||||||
| 5481 | SmallSet<const BasicBlock *, 4> Visited; | ||||||
| 5482 | |||||||
| 5483 | YieldsPoison.insert(V); | ||||||
| 5484 | auto Propagate = [&](const User *User) { | ||||||
| 5485 | if (propagatesPoison(cast<Operator>(User))) | ||||||
| 5486 | YieldsPoison.insert(User); | ||||||
| 5487 | }; | ||||||
| 5488 | for_each(V->users(), Propagate); | ||||||
| 5489 | Visited.insert(BB); | ||||||
| 5490 | |||||||
| 5491 | while (true) { | ||||||
| 5492 | for (auto &I : make_range(Begin, End)) { | ||||||
| 5493 | if (isa<DbgInfoIntrinsic>(I)) | ||||||
| 5494 | continue; | ||||||
| 5495 | if (--ScanLimit == 0) | ||||||
| 5496 | return false; | ||||||
| 5497 | if (mustTriggerUB(&I, YieldsPoison)) | ||||||
| 5498 | return true; | ||||||
| 5499 | if (!isGuaranteedToTransferExecutionToSuccessor(&I)) | ||||||
| 5500 | return false; | ||||||
| 5501 | |||||||
| 5502 | // Mark poison that propagates from I through uses of I. | ||||||
| 5503 | if (YieldsPoison.count(&I)) | ||||||
| 5504 | for_each(I.users(), Propagate); | ||||||
| 5505 | } | ||||||
| 5506 | |||||||
| 5507 | BB = BB->getSingleSuccessor(); | ||||||
| 5508 | if (!BB || !Visited.insert(BB).second) | ||||||
| 5509 | break; | ||||||
| 5510 | |||||||
| 5511 | Begin = BB->getFirstNonPHI()->getIterator(); | ||||||
| 5512 | End = BB->end(); | ||||||
| 5513 | } | ||||||
| 5514 | return false; | ||||||
| 5515 | } | ||||||
| 5516 | |||||||
| 5517 | bool llvm::programUndefinedIfUndefOrPoison(const Instruction *Inst) { | ||||||
| 5518 | return ::programUndefinedIfUndefOrPoison(Inst, false); | ||||||
| 5519 | } | ||||||
| 5520 | |||||||
| 5521 | bool llvm::programUndefinedIfPoison(const Instruction *Inst) { | ||||||
| 5522 | return ::programUndefinedIfUndefOrPoison(Inst, true); | ||||||
| 5523 | } | ||||||
| 5524 | |||||||
| 5525 | static bool isKnownNonNaN(const Value *V, FastMathFlags FMF) { | ||||||
| 5526 | if (FMF.noNaNs()) | ||||||
| 5527 | return true; | ||||||
| 5528 | |||||||
| 5529 | if (auto *C = dyn_cast<ConstantFP>(V)) | ||||||
| 5530 | return !C->isNaN(); | ||||||
| 5531 | |||||||
| 5532 | if (auto *C = dyn_cast<ConstantDataVector>(V)) { | ||||||
| 5533 | if (!C->getElementType()->isFloatingPointTy()) | ||||||
| 5534 | return false; | ||||||
| 5535 | for (unsigned I = 0, E = C->getNumElements(); I < E; ++I) { | ||||||
| 5536 | if (C->getElementAsAPFloat(I).isNaN()) | ||||||
| 5537 | return false; | ||||||
| 5538 | } | ||||||
| 5539 | return true; | ||||||
| 5540 | } | ||||||
| 5541 | |||||||
| 5542 | if (isa<ConstantAggregateZero>(V)) | ||||||
| 5543 | return true; | ||||||
| 5544 | |||||||
| 5545 | return false; | ||||||
| 5546 | } | ||||||
| 5547 | |||||||
| 5548 | static bool isKnownNonZero(const Value *V) { | ||||||
| 5549 | if (auto *C = dyn_cast<ConstantFP>(V)) | ||||||
| 5550 | return !C->isZero(); | ||||||
| 5551 | |||||||
| 5552 | if (auto *C = dyn_cast<ConstantDataVector>(V)) { | ||||||
| 5553 | if (!C->getElementType()->isFloatingPointTy()) | ||||||
| 5554 | return false; | ||||||
| 5555 | for (unsigned I = 0, E = C->getNumElements(); I < E; ++I) { | ||||||
| 5556 | if (C->getElementAsAPFloat(I).isZero()) | ||||||
| 5557 | return false; | ||||||
| 5558 | } | ||||||
| 5559 | return true; | ||||||
| 5560 | } | ||||||
| 5561 | |||||||
| 5562 | return false; | ||||||
| 5563 | } | ||||||
| 5564 | |||||||
| 5565 | /// Match clamp pattern for float types without care about NaNs or signed zeros. | ||||||
| 5566 | /// Given non-min/max outer cmp/select from the clamp pattern this | ||||||
| 5567 | /// function recognizes if it can be substitued by a "canonical" min/max | ||||||
| 5568 | /// pattern. | ||||||
| 5569 | static SelectPatternResult matchFastFloatClamp(CmpInst::Predicate Pred, | ||||||
| 5570 | Value *CmpLHS, Value *CmpRHS, | ||||||
| 5571 | Value *TrueVal, Value *FalseVal, | ||||||
| 5572 | Value *&LHS, Value *&RHS) { | ||||||
| 5573 | // Try to match | ||||||
| 5574 | // X < C1 ? C1 : Min(X, C2) --> Max(C1, Min(X, C2)) | ||||||
| 5575 | // X > C1 ? C1 : Max(X, C2) --> Min(C1, Max(X, C2)) | ||||||
| 5576 | // and return description of the outer Max/Min. | ||||||
| 5577 | |||||||
| 5578 | // First, check if select has inverse order: | ||||||
| 5579 | if (CmpRHS == FalseVal) { | ||||||
| 5580 | std::swap(TrueVal, FalseVal); | ||||||
| 5581 | Pred = CmpInst::getInversePredicate(Pred); | ||||||
| 5582 | } | ||||||
| 5583 | |||||||
| 5584 | // Assume success now. If there's no match, callers should not use these anyway. | ||||||
| 5585 | LHS = TrueVal; | ||||||
| 5586 | RHS = FalseVal; | ||||||
| 5587 | |||||||
| 5588 | const APFloat *FC1; | ||||||
| 5589 | if (CmpRHS != TrueVal || !match(CmpRHS, m_APFloat(FC1)) || !FC1->isFinite()) | ||||||
| 5590 | return {SPF_UNKNOWN, SPNB_NA, false}; | ||||||
| 5591 | |||||||
| 5592 | const APFloat *FC2; | ||||||
| 5593 | switch (Pred) { | ||||||
| 5594 | case CmpInst::FCMP_OLT: | ||||||
| 5595 | case CmpInst::FCMP_OLE: | ||||||
| 5596 | case CmpInst::FCMP_ULT: | ||||||
| 5597 | case CmpInst::FCMP_ULE: | ||||||
| 5598 | if (match(FalseVal, | ||||||
| 5599 | m_CombineOr(m_OrdFMin(m_Specific(CmpLHS), m_APFloat(FC2)), | ||||||
| 5600 | m_UnordFMin(m_Specific(CmpLHS), m_APFloat(FC2)))) && | ||||||
| 5601 | *FC1 < *FC2) | ||||||
| 5602 | return {SPF_FMAXNUM, SPNB_RETURNS_ANY, false}; | ||||||
| 5603 | break; | ||||||
| 5604 | case CmpInst::FCMP_OGT: | ||||||
| 5605 | case CmpInst::FCMP_OGE: | ||||||
| 5606 | case CmpInst::FCMP_UGT: | ||||||
| 5607 | case CmpInst::FCMP_UGE: | ||||||
| 5608 | if (match(FalseVal, | ||||||
| 5609 | m_CombineOr(m_OrdFMax(m_Specific(CmpLHS), m_APFloat(FC2)), | ||||||
| 5610 | m_UnordFMax(m_Specific(CmpLHS), m_APFloat(FC2)))) && | ||||||
| 5611 | *FC1 > *FC2) | ||||||
| 5612 | return {SPF_FMINNUM, SPNB_RETURNS_ANY, false}; | ||||||
| 5613 | break; | ||||||
| 5614 | default: | ||||||
| 5615 | break; | ||||||
| 5616 | } | ||||||
| 5617 | |||||||
| 5618 | return {SPF_UNKNOWN, SPNB_NA, false}; | ||||||
| 5619 | } | ||||||
| 5620 | |||||||
| 5621 | /// Recognize variations of: | ||||||
| 5622 | /// CLAMP(v,l,h) ==> ((v) < (l) ? (l) : ((v) > (h) ? (h) : (v))) | ||||||
| 5623 | static SelectPatternResult matchClamp(CmpInst::Predicate Pred, | ||||||
| 5624 | Value *CmpLHS, Value *CmpRHS, | ||||||
| 5625 | Value *TrueVal, Value *FalseVal) { | ||||||
| 5626 | // Swap the select operands and predicate to match the patterns below. | ||||||
| 5627 | if (CmpRHS
| ||||||
| 5628 | Pred = ICmpInst::getSwappedPredicate(Pred); | ||||||
| 5629 | std::swap(TrueVal, FalseVal); | ||||||
| 5630 | } | ||||||
| 5631 | const APInt *C1; | ||||||
| 5632 | if (CmpRHS == TrueVal && match(CmpRHS, m_APInt(C1))) { | ||||||
| 5633 | const APInt *C2; | ||||||
| 5634 | // (X <s C1) ? C1 : SMIN(X, C2) ==> SMAX(SMIN(X, C2), C1) | ||||||
| 5635 | if (match(FalseVal, m_SMin(m_Specific(CmpLHS), m_APInt(C2))) && | ||||||
| 5636 | C1->slt(*C2) && Pred == CmpInst::ICMP_SLT) | ||||||
| 5637 | return {SPF_SMAX, SPNB_NA, false}; | ||||||
| 5638 | |||||||
| 5639 | // (X >s C1) ? C1 : SMAX(X, C2) ==> SMIN(SMAX(X, C2), C1) | ||||||
| 5640 | if (match(FalseVal, m_SMax(m_Specific(CmpLHS), m_APInt(C2))) && | ||||||
| 5641 | C1->sgt(*C2) && Pred == CmpInst::ICMP_SGT) | ||||||
| 5642 | return {SPF_SMIN, SPNB_NA, false}; | ||||||
| 5643 | |||||||
| 5644 | // (X <u C1) ? C1 : UMIN(X, C2) ==> UMAX(UMIN(X, C2), C1) | ||||||
| 5645 | if (match(FalseVal, m_UMin(m_Specific(CmpLHS), m_APInt(C2))) && | ||||||
| 5646 | C1->ult(*C2) && Pred == CmpInst::ICMP_ULT) | ||||||
| 5647 | return {SPF_UMAX, SPNB_NA, false}; | ||||||
| 5648 | |||||||
| 5649 | // (X >u C1) ? C1 : UMAX(X, C2) ==> UMIN(UMAX(X, C2), C1) | ||||||
| 5650 | if (match(FalseVal, m_UMax(m_Specific(CmpLHS), m_APInt(C2))) && | ||||||
| 5651 | C1->ugt(*C2) && Pred == CmpInst::ICMP_UGT) | ||||||
| 5652 | return {SPF_UMIN, SPNB_NA, false}; | ||||||
| 5653 | } | ||||||
| 5654 | return {SPF_UNKNOWN, SPNB_NA, false}; | ||||||
| 5655 | } | ||||||
| 5656 | |||||||
| 5657 | /// Recognize variations of: | ||||||
| 5658 | /// a < c ? min(a,b) : min(b,c) ==> min(min(a,b),min(b,c)) | ||||||
| 5659 | static SelectPatternResult matchMinMaxOfMinMax(CmpInst::Predicate Pred, | ||||||
| 5660 | Value *CmpLHS, Value *CmpRHS, | ||||||
| 5661 | Value *TVal, Value *FVal, | ||||||
| 5662 | unsigned Depth) { | ||||||
| 5663 | // TODO: Allow FP min/max with nnan/nsz. | ||||||
| 5664 | assert(CmpInst::isIntPredicate(Pred) && "Expected integer comparison")((void)0); | ||||||
| 5665 | |||||||
| 5666 | Value *A = nullptr, *B = nullptr; | ||||||
| 5667 | SelectPatternResult L = matchSelectPattern(TVal, A, B, nullptr, Depth + 1); | ||||||
| 5668 | if (!SelectPatternResult::isMinOrMax(L.Flavor)) | ||||||
| 5669 | return {SPF_UNKNOWN, SPNB_NA, false}; | ||||||
| 5670 | |||||||
| 5671 | Value *C = nullptr, *D = nullptr; | ||||||
| 5672 | SelectPatternResult R = matchSelectPattern(FVal, C, D, nullptr, Depth + 1); | ||||||
| 5673 | if (L.Flavor != R.Flavor) | ||||||
| 5674 | return {SPF_UNKNOWN, SPNB_NA, false}; | ||||||
| 5675 | |||||||
| 5676 | // We have something like: x Pred y ? min(a, b) : min(c, d). | ||||||
| 5677 | // Try to match the compare to the min/max operations of the select operands. | ||||||
| 5678 | // First, make sure we have the right compare predicate. | ||||||
| 5679 | switch (L.Flavor) { | ||||||
| 5680 | case SPF_SMIN: | ||||||
| 5681 | if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE) { | ||||||
| 5682 | Pred = ICmpInst::getSwappedPredicate(Pred); | ||||||
| 5683 | std::swap(CmpLHS, CmpRHS); | ||||||
| 5684 | } | ||||||
| 5685 | if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE) | ||||||
| 5686 | break; | ||||||
| 5687 | return {SPF_UNKNOWN, SPNB_NA, false}; | ||||||
| 5688 | case SPF_SMAX: | ||||||
| 5689 | if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE) { | ||||||
| 5690 | Pred = ICmpInst::getSwappedPredicate(Pred); | ||||||
| 5691 | std::swap(CmpLHS, CmpRHS); | ||||||
| 5692 | } | ||||||
| 5693 | if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE) | ||||||
| 5694 | break; | ||||||
| 5695 | return {SPF_UNKNOWN, SPNB_NA, false}; | ||||||
| 5696 | case SPF_UMIN: | ||||||
| 5697 | if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE) { | ||||||
| 5698 | Pred = ICmpInst::getSwappedPredicate(Pred); | ||||||
| 5699 | std::swap(CmpLHS, CmpRHS); | ||||||
| 5700 | } | ||||||
| 5701 | if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE) | ||||||
| 5702 | break; | ||||||
| 5703 | return {SPF_UNKNOWN, SPNB_NA, false}; | ||||||
| 5704 | case SPF_UMAX: | ||||||
| 5705 | if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE) { | ||||||
| 5706 | Pred = ICmpInst::getSwappedPredicate(Pred); | ||||||
| 5707 | std::swap(CmpLHS, CmpRHS); | ||||||
| 5708 | } | ||||||
| 5709 | if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE) | ||||||
| 5710 | break; | ||||||
| 5711 | return {SPF_UNKNOWN, SPNB_NA, false}; | ||||||
| 5712 | default: | ||||||
| 5713 | return {SPF_UNKNOWN, SPNB_NA, false}; | ||||||
| 5714 | } | ||||||
| 5715 | |||||||
| 5716 | // If there is a common operand in the already matched min/max and the other | ||||||
| 5717 | // min/max operands match the compare operands (either directly or inverted), | ||||||
| 5718 | // then this is min/max of the same flavor. | ||||||
| 5719 | |||||||
| 5720 | // a pred c ? m(a, b) : m(c, b) --> m(m(a, b), m(c, b)) | ||||||
| 5721 | // ~c pred ~a ? m(a, b) : m(c, b) --> m(m(a, b), m(c, b)) | ||||||
| 5722 | if (D == B) { | ||||||
| 5723 | if ((CmpLHS == A && CmpRHS == C) || (match(C, m_Not(m_Specific(CmpLHS))) && | ||||||
| 5724 | match(A, m_Not(m_Specific(CmpRHS))))) | ||||||
| 5725 | return {L.Flavor, SPNB_NA, false}; | ||||||
| 5726 | } | ||||||
| 5727 | // a pred d ? m(a, b) : m(b, d) --> m(m(a, b), m(b, d)) | ||||||
| 5728 | // ~d pred ~a ? m(a, b) : m(b, d) --> m(m(a, b), m(b, d)) | ||||||
| 5729 | if (C == B) { | ||||||
| 5730 | if ((CmpLHS == A && CmpRHS == D) || (match(D, m_Not(m_Specific(CmpLHS))) && | ||||||
| 5731 | match(A, m_Not(m_Specific(CmpRHS))))) | ||||||
| 5732 | return {L.Flavor, SPNB_NA, false}; | ||||||
| 5733 | } | ||||||
| 5734 | // b pred c ? m(a, b) : m(c, a) --> m(m(a, b), m(c, a)) | ||||||
| 5735 | // ~c pred ~b ? m(a, b) : m(c, a) --> m(m(a, b), m(c, a)) | ||||||
| 5736 | if (D == A) { | ||||||
| 5737 | if ((CmpLHS == B && CmpRHS == C) || (match(C, m_Not(m_Specific(CmpLHS))) && | ||||||
| 5738 | match(B, m_Not(m_Specific(CmpRHS))))) | ||||||
| 5739 | return {L.Flavor, SPNB_NA, false}; | ||||||
| 5740 | } | ||||||
| 5741 | // b pred d ? m(a, b) : m(a, d) --> m(m(a, b), m(a, d)) | ||||||
| 5742 | // ~d pred ~b ? m(a, b) : m(a, d) --> m(m(a, b), m(a, d)) | ||||||
| 5743 | if (C == A) { | ||||||
| 5744 | if ((CmpLHS == B && CmpRHS == D) || (match(D, m_Not(m_Specific(CmpLHS))) && | ||||||
| 5745 | match(B, m_Not(m_Specific(CmpRHS))))) | ||||||
| 5746 | return {L.Flavor, SPNB_NA, false}; | ||||||
| 5747 | } | ||||||
| 5748 | |||||||
| 5749 | return {SPF_UNKNOWN, SPNB_NA, false}; | ||||||
| 5750 | } | ||||||
| 5751 | |||||||
| 5752 | /// If the input value is the result of a 'not' op, constant integer, or vector | ||||||
| 5753 | /// splat of a constant integer, return the bitwise-not source value. | ||||||
| 5754 | /// TODO: This could be extended to handle non-splat vector integer constants. | ||||||
| 5755 | static Value *getNotValue(Value *V) { | ||||||
| 5756 | Value *NotV; | ||||||
| 5757 | if (match(V, m_Not(m_Value(NotV)))) | ||||||
| 5758 | return NotV; | ||||||
| 5759 | |||||||
| 5760 | const APInt *C; | ||||||
| 5761 | if (match(V, m_APInt(C))) | ||||||
| 5762 | return ConstantInt::get(V->getType(), ~(*C)); | ||||||
| 5763 | |||||||
| 5764 | return nullptr; | ||||||
| 5765 | } | ||||||
| 5766 | |||||||
| 5767 | /// Match non-obvious integer minimum and maximum sequences. | ||||||
| 5768 | static SelectPatternResult matchMinMax(CmpInst::Predicate Pred, | ||||||
| 5769 | Value *CmpLHS, Value *CmpRHS, | ||||||
| 5770 | Value *TrueVal, Value *FalseVal, | ||||||
| 5771 | Value *&LHS, Value *&RHS, | ||||||
| 5772 | unsigned Depth) { | ||||||
| 5773 | // Assume success. If there's no match, callers should not use these anyway. | ||||||
| 5774 | LHS = TrueVal; | ||||||
| 5775 | RHS = FalseVal; | ||||||
| 5776 | |||||||
| 5777 | SelectPatternResult SPR = matchClamp(Pred, CmpLHS, CmpRHS, TrueVal, FalseVal); | ||||||
| 5778 | if (SPR.Flavor != SelectPatternFlavor::SPF_UNKNOWN) | ||||||
| 5779 | return SPR; | ||||||
| 5780 | |||||||
| 5781 | SPR = matchMinMaxOfMinMax(Pred, CmpLHS, CmpRHS, TrueVal, FalseVal, Depth); | ||||||
| 5782 | if (SPR.Flavor != SelectPatternFlavor::SPF_UNKNOWN) | ||||||
| 5783 | return SPR; | ||||||
| 5784 | |||||||
| 5785 | // Look through 'not' ops to find disguised min/max. | ||||||
| 5786 | // (X > Y) ? ~X : ~Y ==> (~X < ~Y) ? ~X : ~Y ==> MIN(~X, ~Y) | ||||||
| 5787 | // (X < Y) ? ~X : ~Y ==> (~X > ~Y) ? ~X : ~Y ==> MAX(~X, ~Y) | ||||||
| 5788 | if (CmpLHS == getNotValue(TrueVal) && CmpRHS == getNotValue(FalseVal)) { | ||||||
| 5789 | switch (Pred) { | ||||||
| 5790 | case CmpInst::ICMP_SGT: return {SPF_SMIN, SPNB_NA, false}; | ||||||
| 5791 | case CmpInst::ICMP_SLT: return {SPF_SMAX, SPNB_NA, false}; | ||||||
| 5792 | case CmpInst::ICMP_UGT: return {SPF_UMIN, SPNB_NA, false}; | ||||||
| 5793 | case CmpInst::ICMP_ULT: return {SPF_UMAX, SPNB_NA, false}; | ||||||
| 5794 | default: break; | ||||||
| 5795 | } | ||||||
| 5796 | } | ||||||
| 5797 | |||||||
| 5798 | // (X > Y) ? ~Y : ~X ==> (~X < ~Y) ? ~Y : ~X ==> MAX(~Y, ~X) | ||||||
| 5799 | // (X < Y) ? ~Y : ~X ==> (~X > ~Y) ? ~Y : ~X ==> MIN(~Y, ~X) | ||||||
| 5800 | if (CmpLHS == getNotValue(FalseVal) && CmpRHS == getNotValue(TrueVal)) { | ||||||
| 5801 | switch (Pred) { | ||||||
| 5802 | case CmpInst::ICMP_SGT: return {SPF_SMAX, SPNB_NA, false}; | ||||||
| 5803 | case CmpInst::ICMP_SLT: return {SPF_SMIN, SPNB_NA, false}; | ||||||
| 5804 | case CmpInst::ICMP_UGT: return {SPF_UMAX, SPNB_NA, false}; | ||||||
| 5805 | case CmpInst::ICMP_ULT: return {SPF_UMIN, SPNB_NA, false}; | ||||||
| 5806 | default: break; | ||||||
| 5807 | } | ||||||
| 5808 | } | ||||||
| 5809 | |||||||
| 5810 | if (Pred != CmpInst::ICMP_SGT && Pred != CmpInst::ICMP_SLT) | ||||||
| 5811 | return {SPF_UNKNOWN, SPNB_NA, false}; | ||||||
| 5812 | |||||||
| 5813 | // Z = X -nsw Y | ||||||
| 5814 | // (X >s Y) ? 0 : Z ==> (Z >s 0) ? 0 : Z ==> SMIN(Z, 0) | ||||||
| 5815 | // (X <s Y) ? 0 : Z ==> (Z <s 0) ? 0 : Z ==> SMAX(Z, 0) | ||||||
| 5816 | if (match(TrueVal, m_Zero()) && | ||||||
| 5817 | match(FalseVal, m_NSWSub(m_Specific(CmpLHS), m_Specific(CmpRHS)))) | ||||||
| 5818 | return {Pred == CmpInst::ICMP_SGT ? SPF_SMIN : SPF_SMAX, SPNB_NA, false}; | ||||||
| 5819 | |||||||
| 5820 | // Z = X -nsw Y | ||||||
| 5821 | // (X >s Y) ? Z : 0 ==> (Z >s 0) ? Z : 0 ==> SMAX(Z, 0) | ||||||
| 5822 | // (X <s Y) ? Z : 0 ==> (Z <s 0) ? Z : 0 ==> SMIN(Z, 0) | ||||||
| 5823 | if (match(FalseVal, m_Zero()) && | ||||||
| 5824 | match(TrueVal, m_NSWSub(m_Specific(CmpLHS), m_Specific(CmpRHS)))) | ||||||
| 5825 | return {Pred == CmpInst::ICMP_SGT ? SPF_SMAX : SPF_SMIN, SPNB_NA, false}; | ||||||
| 5826 | |||||||
| 5827 | const APInt *C1; | ||||||
| 5828 | if (!match(CmpRHS, m_APInt(C1))) | ||||||
| 5829 | return {SPF_UNKNOWN, SPNB_NA, false}; | ||||||
| 5830 | |||||||
| 5831 | // An unsigned min/max can be written with a signed compare. | ||||||
| 5832 | const APInt *C2; | ||||||
| 5833 | if ((CmpLHS == TrueVal && match(FalseVal, m_APInt(C2))) || | ||||||
| 5834 | (CmpLHS == FalseVal && match(TrueVal, m_APInt(C2)))) { | ||||||
| 5835 | // Is the sign bit set? | ||||||
| 5836 | // (X <s 0) ? X : MAXVAL ==> (X >u MAXVAL) ? X : MAXVAL ==> UMAX | ||||||
| 5837 | // (X <s 0) ? MAXVAL : X ==> (X >u MAXVAL) ? MAXVAL : X ==> UMIN | ||||||
| 5838 | if (Pred == CmpInst::ICMP_SLT && C1->isNullValue() && | ||||||
| 5839 | C2->isMaxSignedValue()) | ||||||
| 5840 | return {CmpLHS == TrueVal ? SPF_UMAX : SPF_UMIN, SPNB_NA, false}; | ||||||
| 5841 | |||||||
| 5842 | // Is the sign bit clear? | ||||||
| 5843 | // (X >s -1) ? MINVAL : X ==> (X <u MINVAL) ? MINVAL : X ==> UMAX | ||||||
| 5844 | // (X >s -1) ? X : MINVAL ==> (X <u MINVAL) ? X : MINVAL ==> UMIN | ||||||
| 5845 | if (Pred == CmpInst::ICMP_SGT && C1->isAllOnesValue() && | ||||||
| 5846 | C2->isMinSignedValue()) | ||||||
| 5847 | return {CmpLHS == FalseVal ? SPF_UMAX : SPF_UMIN, SPNB_NA, false}; | ||||||
| 5848 | } | ||||||
| 5849 | |||||||
| 5850 | return {SPF_UNKNOWN, SPNB_NA, false}; | ||||||
| 5851 | } | ||||||
| 5852 | |||||||
| 5853 | bool llvm::isKnownNegation(const Value *X, const Value *Y, bool NeedNSW) { | ||||||
| 5854 | assert(X && Y && "Invalid operand")((void)0); | ||||||
| 5855 | |||||||
| 5856 | // X = sub (0, Y) || X = sub nsw (0, Y) | ||||||
| 5857 | if ((!NeedNSW
| ||||||
| 5858 | (NeedNSW && match(X, m_NSWSub(m_ZeroInt(), m_Specific(Y))))) | ||||||
| 5859 | return true; | ||||||
| 5860 | |||||||
| 5861 | // Y = sub (0, X) || Y = sub nsw (0, X) | ||||||
| 5862 | if ((!NeedNSW && match(Y, m_Sub(m_ZeroInt(), m_Specific(X)))) || | ||||||
| 5863 | (NeedNSW && match(Y, m_NSWSub(m_ZeroInt(), m_Specific(X))))) | ||||||
| 5864 | return true; | ||||||
| 5865 | |||||||
| 5866 | // X = sub (A, B), Y = sub (B, A) || X = sub nsw (A, B), Y = sub nsw (B, A) | ||||||
| 5867 | Value *A, *B; | ||||||
| 5868 | return (!NeedNSW && (match(X, m_Sub(m_Value(A), m_Value(B))) && | ||||||
| 5869 | match(Y, m_Sub(m_Specific(B), m_Specific(A))))) || | ||||||
| 5870 | (NeedNSW && (match(X, m_NSWSub(m_Value(A), m_Value(B))) && | ||||||
| 5871 | match(Y, m_NSWSub(m_Specific(B), m_Specific(A))))); | ||||||
| 5872 | } | ||||||
| 5873 | |||||||
| 5874 | static SelectPatternResult matchSelectPattern(CmpInst::Predicate Pred, | ||||||
| 5875 | FastMathFlags FMF, | ||||||
| 5876 | Value *CmpLHS, Value *CmpRHS, | ||||||
| 5877 | Value *TrueVal, Value *FalseVal, | ||||||
| 5878 | Value *&LHS, Value *&RHS, | ||||||
| 5879 | unsigned Depth) { | ||||||
| 5880 | if (CmpInst::isFPPredicate(Pred)) { | ||||||
| |||||||
| 5881 | // IEEE-754 ignores the sign of 0.0 in comparisons. So if the select has one | ||||||
| 5882 | // 0.0 operand, set the compare's 0.0 operands to that same value for the | ||||||
| 5883 | // purpose of identifying min/max. Disregard vector constants with undefined | ||||||
| 5884 | // elements because those can not be back-propagated for analysis. | ||||||
| 5885 | Value *OutputZeroVal = nullptr; | ||||||
| 5886 | if (match(TrueVal, m_AnyZeroFP()) && !match(FalseVal, m_AnyZeroFP()) && | ||||||
| 5887 | !cast<Constant>(TrueVal)->containsUndefOrPoisonElement()) | ||||||
| 5888 | OutputZeroVal = TrueVal; | ||||||
| 5889 | else if (match(FalseVal, m_AnyZeroFP()) && !match(TrueVal, m_AnyZeroFP()) && | ||||||
| 5890 | !cast<Constant>(FalseVal)->containsUndefOrPoisonElement()) | ||||||
| 5891 | OutputZeroVal = FalseVal; | ||||||
| 5892 | |||||||
| 5893 | if (OutputZeroVal) { | ||||||
| 5894 | if (match(CmpLHS, m_AnyZeroFP())) | ||||||
| 5895 | CmpLHS = OutputZeroVal; | ||||||
| 5896 | if (match(CmpRHS, m_AnyZeroFP())) | ||||||
| 5897 | CmpRHS = OutputZeroVal; | ||||||
| 5898 | } | ||||||
| 5899 | } | ||||||
| 5900 | |||||||
| 5901 | LHS = CmpLHS; | ||||||
| 5902 | RHS = CmpRHS; | ||||||
| 5903 | |||||||
| 5904 | // Signed zero may return inconsistent results between implementations. | ||||||
| 5905 | // (0.0 <= -0.0) ? 0.0 : -0.0 // Returns 0.0 | ||||||
| 5906 | // minNum(0.0, -0.0) // May return -0.0 or 0.0 (IEEE 754-2008 5.3.1) | ||||||
| 5907 | // Therefore, we behave conservatively and only proceed if at least one of the | ||||||
| 5908 | // operands is known to not be zero or if we don't care about signed zero. | ||||||
| 5909 | switch (Pred) { | ||||||
| 5910 | default: break; | ||||||
| 5911 | // FIXME: Include OGT/OLT/UGT/ULT. | ||||||
| 5912 | case CmpInst::FCMP_OGE: case CmpInst::FCMP_OLE: | ||||||
| 5913 | case CmpInst::FCMP_UGE: case CmpInst::FCMP_ULE: | ||||||
| 5914 | if (!FMF.noSignedZeros() && !isKnownNonZero(CmpLHS) && | ||||||
| 5915 | !isKnownNonZero(CmpRHS)) | ||||||
| 5916 | return {SPF_UNKNOWN, SPNB_NA, false}; | ||||||
| 5917 | } | ||||||
| 5918 | |||||||
| 5919 | SelectPatternNaNBehavior NaNBehavior = SPNB_NA; | ||||||
| 5920 | bool Ordered = false; | ||||||
| 5921 | |||||||
| 5922 | // When given one NaN and one non-NaN input: | ||||||
| 5923 | // - maxnum/minnum (C99 fmaxf()/fminf()) return the non-NaN input. | ||||||
| 5924 | // - A simple C99 (a < b ? a : b) construction will return 'b' (as the | ||||||
| 5925 | // ordered comparison fails), which could be NaN or non-NaN. | ||||||
| 5926 | // so here we discover exactly what NaN behavior is required/accepted. | ||||||
| 5927 | if (CmpInst::isFPPredicate(Pred)) { | ||||||
| 5928 | bool LHSSafe = isKnownNonNaN(CmpLHS, FMF); | ||||||
| 5929 | bool RHSSafe = isKnownNonNaN(CmpRHS, FMF); | ||||||
| 5930 | |||||||
| 5931 | if (LHSSafe && RHSSafe) { | ||||||
| 5932 | // Both operands are known non-NaN. | ||||||
| 5933 | NaNBehavior = SPNB_RETURNS_ANY; | ||||||
| 5934 | } else if (CmpInst::isOrdered(Pred)) { | ||||||
| 5935 | // An ordered comparison will return false when given a NaN, so it | ||||||
| 5936 | // returns the RHS. | ||||||
| 5937 | Ordered = true; | ||||||
| 5938 | if (LHSSafe) | ||||||
| 5939 | // LHS is non-NaN, so if RHS is NaN then NaN will be returned. | ||||||
| 5940 | NaNBehavior = SPNB_RETURNS_NAN; | ||||||
| 5941 | else if (RHSSafe) | ||||||
| 5942 | NaNBehavior = SPNB_RETURNS_OTHER; | ||||||
| 5943 | else | ||||||
| 5944 | // Completely unsafe. | ||||||
| 5945 | return {SPF_UNKNOWN, SPNB_NA, false}; | ||||||
| 5946 | } else { | ||||||
| 5947 | Ordered = false; | ||||||
| 5948 | // An unordered comparison will return true when given a NaN, so it | ||||||
| 5949 | // returns the LHS. | ||||||
| 5950 | if (LHSSafe) | ||||||
| 5951 | // LHS is non-NaN, so if RHS is NaN then non-NaN will be returned. | ||||||
| 5952 | NaNBehavior = SPNB_RETURNS_OTHER; | ||||||
| 5953 | else if (RHSSafe) | ||||||
| 5954 | NaNBehavior = SPNB_RETURNS_NAN; | ||||||
| 5955 | else | ||||||
| 5956 | // Completely unsafe. | ||||||
| 5957 | return {SPF_UNKNOWN, SPNB_NA, false}; | ||||||
| 5958 | } | ||||||
| 5959 | } | ||||||
| 5960 | |||||||
| 5961 | if (TrueVal == CmpRHS && FalseVal == CmpLHS) { | ||||||
| 5962 | std::swap(CmpLHS, CmpRHS); | ||||||
| 5963 | Pred = CmpInst::getSwappedPredicate(Pred); | ||||||
| 5964 | if (NaNBehavior == SPNB_RETURNS_NAN) | ||||||
| 5965 | NaNBehavior = SPNB_RETURNS_OTHER; | ||||||
| 5966 | else if (NaNBehavior == SPNB_RETURNS_OTHER) | ||||||
| 5967 | NaNBehavior = SPNB_RETURNS_NAN; | ||||||
| 5968 | Ordered = !Ordered; | ||||||
| 5969 | } | ||||||
| 5970 | |||||||
| 5971 | // ([if]cmp X, Y) ? X : Y | ||||||
| 5972 | if (TrueVal == CmpLHS && FalseVal == CmpRHS) { | ||||||
| 5973 | switch (Pred) { | ||||||
| 5974 | default: return {SPF_UNKNOWN, SPNB_NA, false}; // Equality. | ||||||
| 5975 | case ICmpInst::ICMP_UGT: | ||||||
| 5976 | case ICmpInst::ICMP_UGE: return {SPF_UMAX, SPNB_NA, false}; | ||||||
| 5977 | case ICmpInst::ICMP_SGT: | ||||||
| 5978 | case ICmpInst::ICMP_SGE: return {SPF_SMAX, SPNB_NA, false}; | ||||||
| 5979 | case ICmpInst::ICMP_ULT: | ||||||
| 5980 | case ICmpInst::ICMP_ULE: return {SPF_UMIN, SPNB_NA, false}; | ||||||
| 5981 | case ICmpInst::ICMP_SLT: | ||||||
| 5982 | case ICmpInst::ICMP_SLE: return {SPF_SMIN, SPNB_NA, false}; | ||||||
| 5983 | case FCmpInst::FCMP_UGT: | ||||||
| 5984 | case FCmpInst::FCMP_UGE: | ||||||
| 5985 | case FCmpInst::FCMP_OGT: | ||||||
| 5986 | case FCmpInst::FCMP_OGE: return {SPF_FMAXNUM, NaNBehavior, Ordered}; | ||||||
| 5987 | case FCmpInst::FCMP_ULT: | ||||||
| 5988 | case FCmpInst::FCMP_ULE: | ||||||
| 5989 | case FCmpInst::FCMP_OLT: | ||||||
| 5990 | case FCmpInst::FCMP_OLE: return {SPF_FMINNUM, NaNBehavior, Ordered}; | ||||||
| 5991 | } | ||||||
| 5992 | } | ||||||
| 5993 | |||||||
| 5994 | if (isKnownNegation(TrueVal, FalseVal)) { | ||||||
| 5995 | // Sign-extending LHS does not change its sign, so TrueVal/FalseVal can | ||||||
| 5996 | // match against either LHS or sext(LHS). | ||||||
| 5997 | auto MaybeSExtCmpLHS = | ||||||
| 5998 | m_CombineOr(m_Specific(CmpLHS), m_SExt(m_Specific(CmpLHS))); | ||||||
| 5999 | auto ZeroOrAllOnes = m_CombineOr(m_ZeroInt(), m_AllOnes()); | ||||||
| 6000 | auto ZeroOrOne = m_CombineOr(m_ZeroInt(), m_One()); | ||||||
| 6001 | if (match(TrueVal, MaybeSExtCmpLHS)) { | ||||||
| 6002 | // Set the return values. If the compare uses the negated value (-X >s 0), | ||||||
| 6003 | // swap the return values because the negated value is always 'RHS'. | ||||||
| 6004 | LHS = TrueVal; | ||||||
| 6005 | RHS = FalseVal; | ||||||
| 6006 | if (match(CmpLHS, m_Neg(m_Specific(FalseVal)))) | ||||||
| 6007 | std::swap(LHS, RHS); | ||||||
| 6008 | |||||||
| 6009 | // (X >s 0) ? X : -X or (X >s -1) ? X : -X --> ABS(X) | ||||||
| 6010 | // (-X >s 0) ? -X : X or (-X >s -1) ? -X : X --> ABS(X) | ||||||
| 6011 | if (Pred == ICmpInst::ICMP_SGT && match(CmpRHS, ZeroOrAllOnes)) | ||||||
| 6012 | return {SPF_ABS, SPNB_NA, false}; | ||||||
| 6013 | |||||||
| 6014 | // (X >=s 0) ? X : -X or (X >=s 1) ? X : -X --> ABS(X) | ||||||
| 6015 | if (Pred == ICmpInst::ICMP_SGE && match(CmpRHS, ZeroOrOne)) | ||||||
| 6016 | return {SPF_ABS, SPNB_NA, false}; | ||||||
| 6017 | |||||||
| 6018 | // (X <s 0) ? X : -X or (X <s 1) ? X : -X --> NABS(X) | ||||||
| 6019 | // (-X <s 0) ? -X : X or (-X <s 1) ? -X : X --> NABS(X) | ||||||
| 6020 | if (Pred == ICmpInst::ICMP_SLT && match(CmpRHS, ZeroOrOne)) | ||||||
| 6021 | return {SPF_NABS, SPNB_NA, false}; | ||||||
| 6022 | } | ||||||
| 6023 | else if (match(FalseVal, MaybeSExtCmpLHS)) { | ||||||
| 6024 | // Set the return values. If the compare uses the negated value (-X >s 0), | ||||||
| 6025 | // swap the return values because the negated value is always 'RHS'. | ||||||
| 6026 | LHS = FalseVal; | ||||||
| 6027 | RHS = TrueVal; | ||||||
| 6028 | if (match(CmpLHS, m_Neg(m_Specific(TrueVal)))) | ||||||
| 6029 | std::swap(LHS, RHS); | ||||||
| 6030 | |||||||
| 6031 | // (X >s 0) ? -X : X or (X >s -1) ? -X : X --> NABS(X) | ||||||
| 6032 | // (-X >s 0) ? X : -X or (-X >s -1) ? X : -X --> NABS(X) | ||||||
| 6033 | if (Pred == ICmpInst::ICMP_SGT && match(CmpRHS, ZeroOrAllOnes)) | ||||||
| 6034 | return {SPF_NABS, SPNB_NA, false}; | ||||||
| 6035 | |||||||
| 6036 | // (X <s 0) ? -X : X or (X <s 1) ? -X : X --> ABS(X) | ||||||
| 6037 | // (-X <s 0) ? X : -X or (-X <s 1) ? X : -X --> ABS(X) | ||||||
| 6038 | if (Pred == ICmpInst::ICMP_SLT && match(CmpRHS, ZeroOrOne)) | ||||||
| 6039 | return {SPF_ABS, SPNB_NA, false}; | ||||||
| 6040 | } | ||||||
| 6041 | } | ||||||
| 6042 | |||||||
| 6043 | if (CmpInst::isIntPredicate(Pred)) | ||||||
| 6044 | return matchMinMax(Pred, CmpLHS, CmpRHS, TrueVal, FalseVal, LHS, RHS, Depth); | ||||||
| 6045 | |||||||
| 6046 | // According to (IEEE 754-2008 5.3.1), minNum(0.0, -0.0) and similar | ||||||
| 6047 | // may return either -0.0 or 0.0, so fcmp/select pair has stricter | ||||||
| 6048 | // semantics than minNum. Be conservative in such case. | ||||||
| 6049 | if (NaNBehavior != SPNB_RETURNS_ANY || | ||||||
| 6050 | (!FMF.noSignedZeros() && !isKnownNonZero(CmpLHS) && | ||||||
| 6051 | !isKnownNonZero(CmpRHS))) | ||||||
| 6052 | return {SPF_UNKNOWN, SPNB_NA, false}; | ||||||
| 6053 | |||||||
| 6054 | return matchFastFloatClamp(Pred, CmpLHS, CmpRHS, TrueVal, FalseVal, LHS, RHS); | ||||||
| 6055 | } | ||||||
| 6056 | |||||||
| 6057 | /// Helps to match a select pattern in case of a type mismatch. | ||||||
| 6058 | /// | ||||||
| 6059 | /// The function processes the case when type of true and false values of a | ||||||
| 6060 | /// select instruction differs from type of the cmp instruction operands because | ||||||
| 6061 | /// of a cast instruction. The function checks if it is legal to move the cast | ||||||
| 6062 | /// operation after "select". If yes, it returns the new second value of | ||||||
| 6063 | /// "select" (with the assumption that cast is moved): | ||||||
| 6064 | /// 1. As operand of cast instruction when both values of "select" are same cast | ||||||
| 6065 | /// instructions. | ||||||
| 6066 | /// 2. As restored constant (by applying reverse cast operation) when the first | ||||||
| 6067 | /// value of the "select" is a cast operation and the second value is a | ||||||
| 6068 | /// constant. | ||||||
| 6069 | /// NOTE: We return only the new second value because the first value could be | ||||||
| 6070 | /// accessed as operand of cast instruction. | ||||||
| 6071 | static Value *lookThroughCast(CmpInst *CmpI, Value *V1, Value *V2, | ||||||
| 6072 | Instruction::CastOps *CastOp) { | ||||||
| 6073 | auto *Cast1 = dyn_cast<CastInst>(V1); | ||||||
| 6074 | if (!Cast1) | ||||||
| 6075 | return nullptr; | ||||||
| 6076 | |||||||
| 6077 | *CastOp = Cast1->getOpcode(); | ||||||
| 6078 | Type *SrcTy = Cast1->getSrcTy(); | ||||||
| 6079 | if (auto *Cast2 = dyn_cast<CastInst>(V2)) { | ||||||
| 6080 | // If V1 and V2 are both the same cast from the same type, look through V1. | ||||||
| 6081 | if (*CastOp == Cast2->getOpcode() && SrcTy == Cast2->getSrcTy()) | ||||||
| 6082 | return Cast2->getOperand(0); | ||||||
| 6083 | return nullptr; | ||||||
| 6084 | } | ||||||
| 6085 | |||||||
| 6086 | auto *C = dyn_cast<Constant>(V2); | ||||||
| 6087 | if (!C) | ||||||
| 6088 | return nullptr; | ||||||
| 6089 | |||||||
| 6090 | Constant *CastedTo = nullptr; | ||||||
| 6091 | switch (*CastOp) { | ||||||
| 6092 | case Instruction::ZExt: | ||||||
| 6093 | if (CmpI->isUnsigned()) | ||||||
| 6094 | CastedTo = ConstantExpr::getTrunc(C, SrcTy); | ||||||
| 6095 | break; | ||||||
| 6096 | case Instruction::SExt: | ||||||
| 6097 | if (CmpI->isSigned()) | ||||||
| 6098 | CastedTo = ConstantExpr::getTrunc(C, SrcTy, true); | ||||||
| 6099 | break; | ||||||
| 6100 | case Instruction::Trunc: | ||||||
| 6101 | Constant *CmpConst; | ||||||
| 6102 | if (match(CmpI->getOperand(1), m_Constant(CmpConst)) && | ||||||
| 6103 | CmpConst->getType() == SrcTy) { | ||||||
| 6104 | // Here we have the following case: | ||||||
| 6105 | // | ||||||
| 6106 | // %cond = cmp iN %x, CmpConst | ||||||
| 6107 | // %tr = trunc iN %x to iK | ||||||
| 6108 | // %narrowsel = select i1 %cond, iK %t, iK C | ||||||
| 6109 | // | ||||||
| 6110 | // We can always move trunc after select operation: | ||||||
| 6111 | // | ||||||
| 6112 | // %cond = cmp iN %x, CmpConst | ||||||
| 6113 | // %widesel = select i1 %cond, iN %x, iN CmpConst | ||||||
| 6114 | // %tr = trunc iN %widesel to iK | ||||||
| 6115 | // | ||||||
| 6116 | // Note that C could be extended in any way because we don't care about | ||||||
| 6117 | // upper bits after truncation. It can't be abs pattern, because it would | ||||||
| 6118 | // look like: | ||||||
| 6119 | // | ||||||
| 6120 | // select i1 %cond, x, -x. | ||||||
| 6121 | // | ||||||
| 6122 | // So only min/max pattern could be matched. Such match requires widened C | ||||||
| 6123 | // == CmpConst. That is why set widened C = CmpConst, condition trunc | ||||||
| 6124 | // CmpConst == C is checked below. | ||||||
| 6125 | CastedTo = CmpConst; | ||||||
| 6126 | } else { | ||||||
| 6127 | CastedTo = ConstantExpr::getIntegerCast(C, SrcTy, CmpI->isSigned()); | ||||||
| 6128 | } | ||||||
| 6129 | break; | ||||||
| 6130 | case Instruction::FPTrunc: | ||||||
| 6131 | CastedTo = ConstantExpr::getFPExtend(C, SrcTy, true); | ||||||
| 6132 | break; | ||||||
| 6133 | case Instruction::FPExt: | ||||||
| 6134 | CastedTo = ConstantExpr::getFPTrunc(C, SrcTy, true); | ||||||
| 6135 | break; | ||||||
| 6136 | case Instruction::FPToUI: | ||||||
| 6137 | CastedTo = ConstantExpr::getUIToFP(C, SrcTy, true); | ||||||
| 6138 | break; | ||||||
| 6139 | case Instruction::FPToSI: | ||||||
| 6140 | CastedTo = ConstantExpr::getSIToFP(C, SrcTy, true); | ||||||
| 6141 | break; | ||||||
| 6142 | case Instruction::UIToFP: | ||||||
| 6143 | CastedTo = ConstantExpr::getFPToUI(C, SrcTy, true); | ||||||
| 6144 | break; | ||||||
| 6145 | case Instruction::SIToFP: | ||||||
| 6146 | CastedTo = ConstantExpr::getFPToSI(C, SrcTy, true); | ||||||
| 6147 | break; | ||||||
| 6148 | default: | ||||||
| 6149 | break; | ||||||
| 6150 | } | ||||||
| 6151 | |||||||
| 6152 | if (!CastedTo) | ||||||
| 6153 | return nullptr; | ||||||
| 6154 | |||||||
| 6155 | // Make sure the cast doesn't lose any information. | ||||||
| 6156 | Constant *CastedBack = | ||||||
| 6157 | ConstantExpr::getCast(*CastOp, CastedTo, C->getType(), true); | ||||||
| 6158 | if (CastedBack != C) | ||||||
| 6159 | return nullptr; | ||||||
| 6160 | |||||||
| 6161 | return CastedTo; | ||||||
| 6162 | } | ||||||
| 6163 | |||||||
| 6164 | SelectPatternResult llvm::matchSelectPattern(Value *V, Value *&LHS, Value *&RHS, | ||||||
| 6165 | Instruction::CastOps *CastOp, | ||||||
| 6166 | unsigned Depth) { | ||||||
| 6167 | if (Depth >= MaxAnalysisRecursionDepth) | ||||||
| 6168 | return {SPF_UNKNOWN, SPNB_NA, false}; | ||||||
| 6169 | |||||||
| 6170 | SelectInst *SI = dyn_cast<SelectInst>(V); | ||||||
| 6171 | if (!SI) return {SPF_UNKNOWN, SPNB_NA, false}; | ||||||
| 6172 | |||||||
| 6173 | CmpInst *CmpI = dyn_cast<CmpInst>(SI->getCondition()); | ||||||
| 6174 | if (!CmpI) return {SPF_UNKNOWN, SPNB_NA, false}; | ||||||
| 6175 | |||||||
| 6176 | Value *TrueVal = SI->getTrueValue(); | ||||||
| 6177 | Value *FalseVal = SI->getFalseValue(); | ||||||
| 6178 | |||||||
| 6179 | return llvm::matchDecomposedSelectPattern(CmpI, TrueVal, FalseVal, LHS, RHS, | ||||||
| 6180 | CastOp, Depth); | ||||||
| 6181 | } | ||||||
| 6182 | |||||||
| 6183 | SelectPatternResult llvm::matchDecomposedSelectPattern( | ||||||
| 6184 | CmpInst *CmpI, Value *TrueVal, Value *FalseVal, Value *&LHS, Value *&RHS, | ||||||
| 6185 | Instruction::CastOps *CastOp, unsigned Depth) { | ||||||
| 6186 | CmpInst::Predicate Pred = CmpI->getPredicate(); | ||||||
| 6187 | Value *CmpLHS = CmpI->getOperand(0); | ||||||
| 6188 | Value *CmpRHS = CmpI->getOperand(1); | ||||||
| 6189 | FastMathFlags FMF; | ||||||
| 6190 | if (isa<FPMathOperator>(CmpI)) | ||||||
| 6191 | FMF = CmpI->getFastMathFlags(); | ||||||
| 6192 | |||||||
| 6193 | // Bail out early. | ||||||
| 6194 | if (CmpI->isEquality()) | ||||||
| 6195 | return {SPF_UNKNOWN, SPNB_NA, false}; | ||||||
| 6196 | |||||||
| 6197 | // Deal with type mismatches. | ||||||
| 6198 | if (CastOp && CmpLHS->getType() != TrueVal->getType()) { | ||||||
| 6199 | if (Value *C = lookThroughCast(CmpI, TrueVal, FalseVal, CastOp)) { | ||||||
| 6200 | // If this is a potential fmin/fmax with a cast to integer, then ignore | ||||||
| 6201 | // -0.0 because there is no corresponding integer value. | ||||||
| 6202 | if (*CastOp == Instruction::FPToSI || *CastOp == Instruction::FPToUI) | ||||||
| 6203 | FMF.setNoSignedZeros(); | ||||||
| 6204 | return ::matchSelectPattern(Pred, FMF, CmpLHS, CmpRHS, | ||||||
| 6205 | cast<CastInst>(TrueVal)->getOperand(0), C, | ||||||
| 6206 | LHS, RHS, Depth); | ||||||
| 6207 | } | ||||||
| 6208 | if (Value *C = lookThroughCast(CmpI, FalseVal, TrueVal, CastOp)) { | ||||||
| 6209 | // If this is a potential fmin/fmax with a cast to integer, then ignore | ||||||
| 6210 | // -0.0 because there is no corresponding integer value. | ||||||
| 6211 | if (*CastOp == Instruction::FPToSI || *CastOp == Instruction::FPToUI) | ||||||
| 6212 | FMF.setNoSignedZeros(); | ||||||
| 6213 | return ::matchSelectPattern(Pred, FMF, CmpLHS, CmpRHS, | ||||||
| 6214 | C, cast<CastInst>(FalseVal)->getOperand(0), | ||||||
| 6215 | LHS, RHS, Depth); | ||||||
| 6216 | } | ||||||
| 6217 | } | ||||||
| 6218 | return ::matchSelectPattern(Pred, FMF, CmpLHS, CmpRHS, TrueVal, FalseVal, | ||||||
| 6219 | LHS, RHS, Depth); | ||||||
| 6220 | } | ||||||
| 6221 | |||||||
| 6222 | CmpInst::Predicate llvm::getMinMaxPred(SelectPatternFlavor SPF, bool Ordered) { | ||||||
| 6223 | if (SPF == SPF_SMIN) return ICmpInst::ICMP_SLT; | ||||||
| 6224 | if (SPF == SPF_UMIN) return ICmpInst::ICMP_ULT; | ||||||
| 6225 | if (SPF == SPF_SMAX) return ICmpInst::ICMP_SGT; | ||||||
| 6226 | if (SPF == SPF_UMAX) return ICmpInst::ICMP_UGT; | ||||||
| 6227 | if (SPF == SPF_FMINNUM) | ||||||
| 6228 | return Ordered ? FCmpInst::FCMP_OLT : FCmpInst::FCMP_ULT; | ||||||
| 6229 | if (SPF == SPF_FMAXNUM) | ||||||
| 6230 | return Ordered ? FCmpInst::FCMP_OGT : FCmpInst::FCMP_UGT; | ||||||
| 6231 | llvm_unreachable("unhandled!")__builtin_unreachable(); | ||||||
| 6232 | } | ||||||
| 6233 | |||||||
| 6234 | SelectPatternFlavor llvm::getInverseMinMaxFlavor(SelectPatternFlavor SPF) { | ||||||
| 6235 | if (SPF == SPF_SMIN) return SPF_SMAX; | ||||||
| 6236 | if (SPF == SPF_UMIN) return SPF_UMAX; | ||||||
| 6237 | if (SPF == SPF_SMAX) return SPF_SMIN; | ||||||
| 6238 | if (SPF == SPF_UMAX) return SPF_UMIN; | ||||||
| 6239 | llvm_unreachable("unhandled!")__builtin_unreachable(); | ||||||
| 6240 | } | ||||||
| 6241 | |||||||
| 6242 | Intrinsic::ID llvm::getInverseMinMaxIntrinsic(Intrinsic::ID MinMaxID) { | ||||||
| 6243 | switch (MinMaxID) { | ||||||
| 6244 | case Intrinsic::smax: return Intrinsic::smin; | ||||||
| 6245 | case Intrinsic::smin: return Intrinsic::smax; | ||||||
| 6246 | case Intrinsic::umax: return Intrinsic::umin; | ||||||
| 6247 | case Intrinsic::umin: return Intrinsic::umax; | ||||||
| 6248 | default: llvm_unreachable("Unexpected intrinsic")__builtin_unreachable(); | ||||||
| 6249 | } | ||||||
| 6250 | } | ||||||
| 6251 | |||||||
| 6252 | CmpInst::Predicate llvm::getInverseMinMaxPred(SelectPatternFlavor SPF) { | ||||||
| 6253 | return getMinMaxPred(getInverseMinMaxFlavor(SPF)); | ||||||
| 6254 | } | ||||||
| 6255 | |||||||
| 6256 | APInt llvm::getMinMaxLimit(SelectPatternFlavor SPF, unsigned BitWidth) { | ||||||
| 6257 | switch (SPF) { | ||||||
| 6258 | case SPF_SMAX: return APInt::getSignedMaxValue(BitWidth); | ||||||
| 6259 | case SPF_SMIN: return APInt::getSignedMinValue(BitWidth); | ||||||
| 6260 | case SPF_UMAX: return APInt::getMaxValue(BitWidth); | ||||||
| 6261 | case SPF_UMIN: return APInt::getMinValue(BitWidth); | ||||||
| 6262 | default: llvm_unreachable("Unexpected flavor")__builtin_unreachable(); | ||||||
| 6263 | } | ||||||
| 6264 | } | ||||||
| 6265 | |||||||
| 6266 | std::pair<Intrinsic::ID, bool> | ||||||
| 6267 | llvm::canConvertToMinOrMaxIntrinsic(ArrayRef<Value *> VL) { | ||||||
| 6268 | // Check if VL contains select instructions that can be folded into a min/max | ||||||
| 6269 | // vector intrinsic and return the intrinsic if it is possible. | ||||||
| 6270 | // TODO: Support floating point min/max. | ||||||
| 6271 | bool AllCmpSingleUse = true; | ||||||
| 6272 | SelectPatternResult SelectPattern; | ||||||
| 6273 | SelectPattern.Flavor = SPF_UNKNOWN; | ||||||
| 6274 | if (all_of(VL, [&SelectPattern, &AllCmpSingleUse](Value *I) { | ||||||
| 6275 | Value *LHS, *RHS; | ||||||
| 6276 | auto CurrentPattern = matchSelectPattern(I, LHS, RHS); | ||||||
| 6277 | if (!SelectPatternResult::isMinOrMax(CurrentPattern.Flavor) || | ||||||
| 6278 | CurrentPattern.Flavor == SPF_FMINNUM || | ||||||
| 6279 | CurrentPattern.Flavor == SPF_FMAXNUM || | ||||||
| 6280 | !I->getType()->isIntOrIntVectorTy()) | ||||||
| 6281 | return false; | ||||||
| 6282 | if (SelectPattern.Flavor != SPF_UNKNOWN && | ||||||
| 6283 | SelectPattern.Flavor != CurrentPattern.Flavor) | ||||||
| 6284 | return false; | ||||||
| 6285 | SelectPattern = CurrentPattern; | ||||||
| 6286 | AllCmpSingleUse &= | ||||||
| 6287 | match(I, m_Select(m_OneUse(m_Value()), m_Value(), m_Value())); | ||||||
| 6288 | return true; | ||||||
| 6289 | })) { | ||||||
| 6290 | switch (SelectPattern.Flavor) { | ||||||
| 6291 | case SPF_SMIN: | ||||||
| 6292 | return {Intrinsic::smin, AllCmpSingleUse}; | ||||||
| 6293 | case SPF_UMIN: | ||||||
| 6294 | return {Intrinsic::umin, AllCmpSingleUse}; | ||||||
| 6295 | case SPF_SMAX: | ||||||
| 6296 | return {Intrinsic::smax, AllCmpSingleUse}; | ||||||
| 6297 | case SPF_UMAX: | ||||||
| 6298 | return {Intrinsic::umax, AllCmpSingleUse}; | ||||||
| 6299 | default: | ||||||
| 6300 | llvm_unreachable("unexpected select pattern flavor")__builtin_unreachable(); | ||||||
| 6301 | } | ||||||
| 6302 | } | ||||||
| 6303 | return {Intrinsic::not_intrinsic, false}; | ||||||
| 6304 | } | ||||||
| 6305 | |||||||
| 6306 | bool llvm::matchSimpleRecurrence(const PHINode *P, BinaryOperator *&BO, | ||||||
| 6307 | Value *&Start, Value *&Step) { | ||||||
| 6308 | // Handle the case of a simple two-predecessor recurrence PHI. | ||||||
| 6309 | // There's a lot more that could theoretically be done here, but | ||||||
| 6310 | // this is sufficient to catch some interesting cases. | ||||||
| 6311 | if (P->getNumIncomingValues() != 2) | ||||||
| 6312 | return false; | ||||||
| 6313 | |||||||
| 6314 | for (unsigned i = 0; i != 2; ++i) { | ||||||
| 6315 | Value *L = P->getIncomingValue(i); | ||||||
| 6316 | Value *R = P->getIncomingValue(!i); | ||||||
| 6317 | Operator *LU = dyn_cast<Operator>(L); | ||||||
| 6318 | if (!LU) | ||||||
| 6319 | continue; | ||||||
| 6320 | unsigned Opcode = LU->getOpcode(); | ||||||
| 6321 | |||||||
| 6322 | switch (Opcode) { | ||||||
| 6323 | default: | ||||||
| 6324 | continue; | ||||||
| 6325 | // TODO: Expand list -- xor, div, gep, uaddo, etc.. | ||||||
| 6326 | case Instruction::LShr: | ||||||
| 6327 | case Instruction::AShr: | ||||||
| 6328 | case Instruction::Shl: | ||||||
| 6329 | case Instruction::Add: | ||||||
| 6330 | case Instruction::Sub: | ||||||
| 6331 | case Instruction::And: | ||||||
| 6332 | case Instruction::Or: | ||||||
| 6333 | case Instruction::Mul: { | ||||||
| 6334 | Value *LL = LU->getOperand(0); | ||||||
| 6335 | Value *LR = LU->getOperand(1); | ||||||
| 6336 | // Find a recurrence. | ||||||
| 6337 | if (LL == P) | ||||||
| 6338 | L = LR; | ||||||
| 6339 | else if (LR == P) | ||||||
| 6340 | L = LL; | ||||||
| 6341 | else | ||||||
| 6342 | continue; // Check for recurrence with L and R flipped. | ||||||
| 6343 | |||||||
| 6344 | break; // Match! | ||||||
| 6345 | } | ||||||
| 6346 | }; | ||||||
| 6347 | |||||||
| 6348 | // We have matched a recurrence of the form: | ||||||
| 6349 | // %iv = [R, %entry], [%iv.next, %backedge] | ||||||
| 6350 | // %iv.next = binop %iv, L | ||||||
| 6351 | // OR | ||||||
| 6352 | // %iv = [R, %entry], [%iv.next, %backedge] | ||||||
| 6353 | // %iv.next = binop L, %iv | ||||||
| 6354 | BO = cast<BinaryOperator>(LU); | ||||||
| 6355 | Start = R; | ||||||
| 6356 | Step = L; | ||||||
| 6357 | return true; | ||||||
| 6358 | } | ||||||
| 6359 | return false; | ||||||
| 6360 | } | ||||||
| 6361 | |||||||
| 6362 | bool llvm::matchSimpleRecurrence(const BinaryOperator *I, PHINode *&P, | ||||||
| 6363 | Value *&Start, Value *&Step) { | ||||||
| 6364 | BinaryOperator *BO = nullptr; | ||||||
| 6365 | P = dyn_cast<PHINode>(I->getOperand(0)); | ||||||
| 6366 | if (!P) | ||||||
| 6367 | P = dyn_cast<PHINode>(I->getOperand(1)); | ||||||
| 6368 | return P && matchSimpleRecurrence(P, BO, Start, Step) && BO == I; | ||||||
| 6369 | } | ||||||
| 6370 | |||||||
| 6371 | /// Return true if "icmp Pred LHS RHS" is always true. | ||||||
| 6372 | static bool isTruePredicate(CmpInst::Predicate Pred, const Value *LHS, | ||||||
| 6373 | const Value *RHS, const DataLayout &DL, | ||||||
| 6374 | unsigned Depth) { | ||||||
| 6375 | assert(!LHS->getType()->isVectorTy() && "TODO: extend to handle vectors!")((void)0); | ||||||
| 6376 | if (ICmpInst::isTrueWhenEqual(Pred) && LHS == RHS) | ||||||
| 6377 | return true; | ||||||
| 6378 | |||||||
| 6379 | switch (Pred) { | ||||||
| 6380 | default: | ||||||
| 6381 | return false; | ||||||
| 6382 | |||||||
| 6383 | case CmpInst::ICMP_SLE: { | ||||||
| 6384 | const APInt *C; | ||||||
| 6385 | |||||||
| 6386 | // LHS s<= LHS +_{nsw} C if C >= 0 | ||||||
| 6387 | if (match(RHS, m_NSWAdd(m_Specific(LHS), m_APInt(C)))) | ||||||
| 6388 | return !C->isNegative(); | ||||||
| 6389 | return false; | ||||||
| 6390 | } | ||||||
| 6391 | |||||||
| 6392 | case CmpInst::ICMP_ULE: { | ||||||
| 6393 | const APInt *C; | ||||||
| 6394 | |||||||
| 6395 | // LHS u<= LHS +_{nuw} C for any C | ||||||
| 6396 | if (match(RHS, m_NUWAdd(m_Specific(LHS), m_APInt(C)))) | ||||||
| 6397 | return true; | ||||||
| 6398 | |||||||
| 6399 | // Match A to (X +_{nuw} CA) and B to (X +_{nuw} CB) | ||||||
| 6400 | auto MatchNUWAddsToSameValue = [&](const Value *A, const Value *B, | ||||||
| 6401 | const Value *&X, | ||||||
| 6402 | const APInt *&CA, const APInt *&CB) { | ||||||
| 6403 | if (match(A, m_NUWAdd(m_Value(X), m_APInt(CA))) && | ||||||
| 6404 | match(B, m_NUWAdd(m_Specific(X), m_APInt(CB)))) | ||||||
| 6405 | return true; | ||||||
| 6406 | |||||||
| 6407 | // If X & C == 0 then (X | C) == X +_{nuw} C | ||||||
| 6408 | if (match(A, m_Or(m_Value(X), m_APInt(CA))) && | ||||||
| 6409 | match(B, m_Or(m_Specific(X), m_APInt(CB)))) { | ||||||
| 6410 | KnownBits Known(CA->getBitWidth()); | ||||||
| 6411 | computeKnownBits(X, Known, DL, Depth + 1, /*AC*/ nullptr, | ||||||
| 6412 | /*CxtI*/ nullptr, /*DT*/ nullptr); | ||||||
| 6413 | if (CA->isSubsetOf(Known.Zero) && CB->isSubsetOf(Known.Zero)) | ||||||
| 6414 | return true; | ||||||
| 6415 | } | ||||||
| 6416 | |||||||
| 6417 | return false; | ||||||
| 6418 | }; | ||||||
| 6419 | |||||||
| 6420 | const Value *X; | ||||||
| 6421 | const APInt *CLHS, *CRHS; | ||||||
| 6422 | if (MatchNUWAddsToSameValue(LHS, RHS, X, CLHS, CRHS)) | ||||||
| 6423 | return CLHS->ule(*CRHS); | ||||||
| 6424 | |||||||
| 6425 | return false; | ||||||
| 6426 | } | ||||||
| 6427 | } | ||||||
| 6428 | } | ||||||
| 6429 | |||||||
| 6430 | /// Return true if "icmp Pred BLHS BRHS" is true whenever "icmp Pred | ||||||
| 6431 | /// ALHS ARHS" is true. Otherwise, return None. | ||||||
| 6432 | static Optional<bool> | ||||||
| 6433 | isImpliedCondOperands(CmpInst::Predicate Pred, const Value *ALHS, | ||||||
| 6434 | const Value *ARHS, const Value *BLHS, const Value *BRHS, | ||||||
| 6435 | const DataLayout &DL, unsigned Depth) { | ||||||
| 6436 | switch (Pred) { | ||||||
| 6437 | default: | ||||||
| 6438 | return None; | ||||||
| 6439 | |||||||
| 6440 | case CmpInst::ICMP_SLT: | ||||||
| 6441 | case CmpInst::ICMP_SLE: | ||||||
| 6442 | if (isTruePredicate(CmpInst::ICMP_SLE, BLHS, ALHS, DL, Depth) && | ||||||
| 6443 | isTruePredicate(CmpInst::ICMP_SLE, ARHS, BRHS, DL, Depth)) | ||||||
| 6444 | return true; | ||||||
| 6445 | return None; | ||||||
| 6446 | |||||||
| 6447 | case CmpInst::ICMP_ULT: | ||||||
| 6448 | case CmpInst::ICMP_ULE: | ||||||
| 6449 | if (isTruePredicate(CmpInst::ICMP_ULE, BLHS, ALHS, DL, Depth) && | ||||||
| 6450 | isTruePredicate(CmpInst::ICMP_ULE, ARHS, BRHS, DL, Depth)) | ||||||
| 6451 | return true; | ||||||
| 6452 | return None; | ||||||
| 6453 | } | ||||||
| 6454 | } | ||||||
| 6455 | |||||||
| 6456 | /// Return true if the operands of the two compares match. IsSwappedOps is true | ||||||
| 6457 | /// when the operands match, but are swapped. | ||||||
| 6458 | static bool isMatchingOps(const Value *ALHS, const Value *ARHS, | ||||||
| 6459 | const Value *BLHS, const Value *BRHS, | ||||||
| 6460 | bool &IsSwappedOps) { | ||||||
| 6461 | |||||||
| 6462 | bool IsMatchingOps = (ALHS == BLHS && ARHS == BRHS); | ||||||
| 6463 | IsSwappedOps = (ALHS == BRHS && ARHS == BLHS); | ||||||
| 6464 | return IsMatchingOps || IsSwappedOps; | ||||||
| 6465 | } | ||||||
| 6466 | |||||||
| 6467 | /// Return true if "icmp1 APred X, Y" implies "icmp2 BPred X, Y" is true. | ||||||
| 6468 | /// Return false if "icmp1 APred X, Y" implies "icmp2 BPred X, Y" is false. | ||||||
| 6469 | /// Otherwise, return None if we can't infer anything. | ||||||
| 6470 | static Optional<bool> isImpliedCondMatchingOperands(CmpInst::Predicate APred, | ||||||
| 6471 | CmpInst::Predicate BPred, | ||||||
| 6472 | bool AreSwappedOps) { | ||||||
| 6473 | // Canonicalize the predicate as if the operands were not commuted. | ||||||
| 6474 | if (AreSwappedOps) | ||||||
| 6475 | BPred = ICmpInst::getSwappedPredicate(BPred); | ||||||
| 6476 | |||||||
| 6477 | if (CmpInst::isImpliedTrueByMatchingCmp(APred, BPred)) | ||||||
| 6478 | return true; | ||||||
| 6479 | if (CmpInst::isImpliedFalseByMatchingCmp(APred, BPred)) | ||||||
| 6480 | return false; | ||||||
| 6481 | |||||||
| 6482 | return None; | ||||||
| 6483 | } | ||||||
| 6484 | |||||||
| 6485 | /// Return true if "icmp APred X, C1" implies "icmp BPred X, C2" is true. | ||||||
| 6486 | /// Return false if "icmp APred X, C1" implies "icmp BPred X, C2" is false. | ||||||
| 6487 | /// Otherwise, return None if we can't infer anything. | ||||||
| 6488 | static Optional<bool> | ||||||
| 6489 | isImpliedCondMatchingImmOperands(CmpInst::Predicate APred, | ||||||
| 6490 | const ConstantInt *C1, | ||||||
| 6491 | CmpInst::Predicate BPred, | ||||||
| 6492 | const ConstantInt *C2) { | ||||||
| 6493 | ConstantRange DomCR = | ||||||
| 6494 | ConstantRange::makeExactICmpRegion(APred, C1->getValue()); | ||||||
| 6495 | ConstantRange CR = ConstantRange::makeExactICmpRegion(BPred, C2->getValue()); | ||||||
| 6496 | ConstantRange Intersection = DomCR.intersectWith(CR); | ||||||
| 6497 | ConstantRange Difference = DomCR.difference(CR); | ||||||
| 6498 | if (Intersection.isEmptySet()) | ||||||
| 6499 | return false; | ||||||
| 6500 | if (Difference.isEmptySet()) | ||||||
| 6501 | return true; | ||||||
| 6502 | return None; | ||||||
| 6503 | } | ||||||
| 6504 | |||||||
| 6505 | /// Return true if LHS implies RHS is true. Return false if LHS implies RHS is | ||||||
| 6506 | /// false. Otherwise, return None if we can't infer anything. | ||||||
| 6507 | static Optional<bool> isImpliedCondICmps(const ICmpInst *LHS, | ||||||
| 6508 | CmpInst::Predicate BPred, | ||||||
| 6509 | const Value *BLHS, const Value *BRHS, | ||||||
| 6510 | const DataLayout &DL, bool LHSIsTrue, | ||||||
| 6511 | unsigned Depth) { | ||||||
| 6512 | Value *ALHS = LHS->getOperand(0); | ||||||
| 6513 | Value *ARHS = LHS->getOperand(1); | ||||||
| 6514 | |||||||
| 6515 | // The rest of the logic assumes the LHS condition is true. If that's not the | ||||||
| 6516 | // case, invert the predicate to make it so. | ||||||
| 6517 | CmpInst::Predicate APred = | ||||||
| 6518 | LHSIsTrue ? LHS->getPredicate() : LHS->getInversePredicate(); | ||||||
| 6519 | |||||||
| 6520 | // Can we infer anything when the two compares have matching operands? | ||||||
| 6521 | bool AreSwappedOps; | ||||||
| 6522 | if (isMatchingOps(ALHS, ARHS, BLHS, BRHS, AreSwappedOps)) { | ||||||
| 6523 | if (Optional<bool> Implication = isImpliedCondMatchingOperands( | ||||||
| 6524 | APred, BPred, AreSwappedOps)) | ||||||
| 6525 | return Implication; | ||||||
| 6526 | // No amount of additional analysis will infer the second condition, so | ||||||
| 6527 | // early exit. | ||||||
| 6528 | return None; | ||||||
| 6529 | } | ||||||
| 6530 | |||||||
| 6531 | // Can we infer anything when the LHS operands match and the RHS operands are | ||||||
| 6532 | // constants (not necessarily matching)? | ||||||
| 6533 | if (ALHS == BLHS && isa<ConstantInt>(ARHS) && isa<ConstantInt>(BRHS)) { | ||||||
| 6534 | if (Optional<bool> Implication = isImpliedCondMatchingImmOperands( | ||||||
| 6535 | APred, cast<ConstantInt>(ARHS), BPred, cast<ConstantInt>(BRHS))) | ||||||
| 6536 | return Implication; | ||||||
| 6537 | // No amount of additional analysis will infer the second condition, so | ||||||
| 6538 | // early exit. | ||||||
| 6539 | return None; | ||||||
| 6540 | } | ||||||
| 6541 | |||||||
| 6542 | if (APred == BPred) | ||||||
| 6543 | return isImpliedCondOperands(APred, ALHS, ARHS, BLHS, BRHS, DL, Depth); | ||||||
| 6544 | return None; | ||||||
| 6545 | } | ||||||
| 6546 | |||||||
| 6547 | /// Return true if LHS implies RHS is true. Return false if LHS implies RHS is | ||||||
| 6548 | /// false. Otherwise, return None if we can't infer anything. We expect the | ||||||
| 6549 | /// RHS to be an icmp and the LHS to be an 'and', 'or', or a 'select' instruction. | ||||||
| 6550 | static Optional<bool> | ||||||
| 6551 | isImpliedCondAndOr(const Instruction *LHS, CmpInst::Predicate RHSPred, | ||||||
| 6552 | const Value *RHSOp0, const Value *RHSOp1, | ||||||
| 6553 | const DataLayout &DL, bool LHSIsTrue, unsigned Depth) { | ||||||
| 6554 | // The LHS must be an 'or', 'and', or a 'select' instruction. | ||||||
| 6555 | assert((LHS->getOpcode() == Instruction::And ||((void)0) | ||||||
| 6556 | LHS->getOpcode() == Instruction::Or ||((void)0) | ||||||
| 6557 | LHS->getOpcode() == Instruction::Select) &&((void)0) | ||||||
| 6558 | "Expected LHS to be 'and', 'or', or 'select'.")((void)0); | ||||||
| 6559 | |||||||
| 6560 | assert(Depth <= MaxAnalysisRecursionDepth && "Hit recursion limit")((void)0); | ||||||
| 6561 | |||||||
| 6562 | // If the result of an 'or' is false, then we know both legs of the 'or' are | ||||||
| 6563 | // false. Similarly, if the result of an 'and' is true, then we know both | ||||||
| 6564 | // legs of the 'and' are true. | ||||||
| 6565 | const Value *ALHS, *ARHS; | ||||||
| 6566 | if ((!LHSIsTrue && match(LHS, m_LogicalOr(m_Value(ALHS), m_Value(ARHS)))) || | ||||||
| 6567 | (LHSIsTrue && match(LHS, m_LogicalAnd(m_Value(ALHS), m_Value(ARHS))))) { | ||||||
| 6568 | // FIXME: Make this non-recursion. | ||||||
| 6569 | if (Optional<bool> Implication = isImpliedCondition( | ||||||
| 6570 | ALHS, RHSPred, RHSOp0, RHSOp1, DL, LHSIsTrue, Depth + 1)) | ||||||
| 6571 | return Implication; | ||||||
| 6572 | if (Optional<bool> Implication = isImpliedCondition( | ||||||
| 6573 | ARHS, RHSPred, RHSOp0, RHSOp1, DL, LHSIsTrue, Depth + 1)) | ||||||
| 6574 | return Implication; | ||||||
| 6575 | return None; | ||||||
| 6576 | } | ||||||
| 6577 | return None; | ||||||
| 6578 | } | ||||||
| 6579 | |||||||
| 6580 | Optional<bool> | ||||||
| 6581 | llvm::isImpliedCondition(const Value *LHS, CmpInst::Predicate RHSPred, | ||||||
| 6582 | const Value *RHSOp0, const Value *RHSOp1, | ||||||
| 6583 | const DataLayout &DL, bool LHSIsTrue, unsigned Depth) { | ||||||
| 6584 | // Bail out when we hit the limit. | ||||||
| 6585 | if (Depth == MaxAnalysisRecursionDepth) | ||||||
| 6586 | return None; | ||||||
| 6587 | |||||||
| 6588 | // A mismatch occurs when we compare a scalar cmp to a vector cmp, for | ||||||
| 6589 | // example. | ||||||
| 6590 | if (RHSOp0->getType()->isVectorTy() != LHS->getType()->isVectorTy()) | ||||||
| 6591 | return None; | ||||||
| 6592 | |||||||
| 6593 | Type *OpTy = LHS->getType(); | ||||||
| 6594 | assert(OpTy->isIntOrIntVectorTy(1) && "Expected integer type only!")((void)0); | ||||||
| 6595 | |||||||
| 6596 | // FIXME: Extending the code below to handle vectors. | ||||||
| 6597 | if (OpTy->isVectorTy()) | ||||||
| 6598 | return None; | ||||||
| 6599 | |||||||
| 6600 | assert(OpTy->isIntegerTy(1) && "implied by above")((void)0); | ||||||
| 6601 | |||||||
| 6602 | // Both LHS and RHS are icmps. | ||||||
| 6603 | const ICmpInst *LHSCmp = dyn_cast<ICmpInst>(LHS); | ||||||
| 6604 | if (LHSCmp) | ||||||
| 6605 | return isImpliedCondICmps(LHSCmp, RHSPred, RHSOp0, RHSOp1, DL, LHSIsTrue, | ||||||
| 6606 | Depth); | ||||||
| 6607 | |||||||
| 6608 | /// The LHS should be an 'or', 'and', or a 'select' instruction. We expect | ||||||
| 6609 | /// the RHS to be an icmp. | ||||||
| 6610 | /// FIXME: Add support for and/or/select on the RHS. | ||||||
| 6611 | if (const Instruction *LHSI = dyn_cast<Instruction>(LHS)) { | ||||||
| 6612 | if ((LHSI->getOpcode() == Instruction::And || | ||||||
| 6613 | LHSI->getOpcode() == Instruction::Or || | ||||||
| 6614 | LHSI->getOpcode() == Instruction::Select)) | ||||||
| 6615 | return isImpliedCondAndOr(LHSI, RHSPred, RHSOp0, RHSOp1, DL, LHSIsTrue, | ||||||
| 6616 | Depth); | ||||||
| 6617 | } | ||||||
| 6618 | return None; | ||||||
| 6619 | } | ||||||
| 6620 | |||||||
| 6621 | Optional<bool> llvm::isImpliedCondition(const Value *LHS, const Value *RHS, | ||||||
| 6622 | const DataLayout &DL, bool LHSIsTrue, | ||||||
| 6623 | unsigned Depth) { | ||||||
| 6624 | // LHS ==> RHS by definition | ||||||
| 6625 | if (LHS == RHS) | ||||||
| 6626 | return LHSIsTrue; | ||||||
| 6627 | |||||||
| 6628 | const ICmpInst *RHSCmp = dyn_cast<ICmpInst>(RHS); | ||||||
| 6629 | if (RHSCmp) | ||||||
| 6630 | return isImpliedCondition(LHS, RHSCmp->getPredicate(), | ||||||
| 6631 | RHSCmp->getOperand(0), RHSCmp->getOperand(1), DL, | ||||||
| 6632 | LHSIsTrue, Depth); | ||||||
| 6633 | return None; | ||||||
| 6634 | } | ||||||
| 6635 | |||||||
| 6636 | // Returns a pair (Condition, ConditionIsTrue), where Condition is a branch | ||||||
| 6637 | // condition dominating ContextI or nullptr, if no condition is found. | ||||||
| 6638 | static std::pair<Value *, bool> | ||||||
| 6639 | getDomPredecessorCondition(const Instruction *ContextI) { | ||||||
| 6640 | if (!ContextI || !ContextI->getParent()) | ||||||
| 6641 | return {nullptr, false}; | ||||||
| 6642 | |||||||
| 6643 | // TODO: This is a poor/cheap way to determine dominance. Should we use a | ||||||
| 6644 | // dominator tree (eg, from a SimplifyQuery) instead? | ||||||
| 6645 | const BasicBlock *ContextBB = ContextI->getParent(); | ||||||
| 6646 | const BasicBlock *PredBB = ContextBB->getSinglePredecessor(); | ||||||
| 6647 | if (!PredBB) | ||||||
| 6648 | return {nullptr, false}; | ||||||
| 6649 | |||||||
| 6650 | // We need a conditional branch in the predecessor. | ||||||
| 6651 | Value *PredCond; | ||||||
| 6652 | BasicBlock *TrueBB, *FalseBB; | ||||||
| 6653 | if (!match(PredBB->getTerminator(), m_Br(m_Value(PredCond), TrueBB, FalseBB))) | ||||||
| 6654 | return {nullptr, false}; | ||||||
| 6655 | |||||||
| 6656 | // The branch should get simplified. Don't bother simplifying this condition. | ||||||
| 6657 | if (TrueBB == FalseBB) | ||||||
| 6658 | return {nullptr, false}; | ||||||
| 6659 | |||||||
| 6660 | assert((TrueBB == ContextBB || FalseBB == ContextBB) &&((void)0) | ||||||
| 6661 | "Predecessor block does not point to successor?")((void)0); | ||||||
| 6662 | |||||||
| 6663 | // Is this condition implied by the predecessor condition? | ||||||
| 6664 | return {PredCond, TrueBB == ContextBB}; | ||||||
| 6665 | } | ||||||
| 6666 | |||||||
| 6667 | Optional<bool> llvm::isImpliedByDomCondition(const Value *Cond, | ||||||
| 6668 | const Instruction *ContextI, | ||||||
| 6669 | const DataLayout &DL) { | ||||||
| 6670 | assert(Cond->getType()->isIntOrIntVectorTy(1) && "Condition must be bool")((void)0); | ||||||
| 6671 | auto PredCond = getDomPredecessorCondition(ContextI); | ||||||
| 6672 | if (PredCond.first) | ||||||
| 6673 | return isImpliedCondition(PredCond.first, Cond, DL, PredCond.second); | ||||||
| 6674 | return None; | ||||||
| 6675 | } | ||||||
| 6676 | |||||||
| 6677 | Optional<bool> llvm::isImpliedByDomCondition(CmpInst::Predicate Pred, | ||||||
| 6678 | const Value *LHS, const Value *RHS, | ||||||
| 6679 | const Instruction *ContextI, | ||||||
| 6680 | const DataLayout &DL) { | ||||||
| 6681 | auto PredCond = getDomPredecessorCondition(ContextI); | ||||||
| 6682 | if (PredCond.first) | ||||||
| 6683 | return isImpliedCondition(PredCond.first, Pred, LHS, RHS, DL, | ||||||
| 6684 | PredCond.second); | ||||||
| 6685 | return None; | ||||||
| 6686 | } | ||||||
| 6687 | |||||||
| 6688 | static void setLimitsForBinOp(const BinaryOperator &BO, APInt &Lower, | ||||||
| 6689 | APInt &Upper, const InstrInfoQuery &IIQ) { | ||||||
| 6690 | unsigned Width = Lower.getBitWidth(); | ||||||
| 6691 | const APInt *C; | ||||||
| 6692 | switch (BO.getOpcode()) { | ||||||
| 6693 | case Instruction::Add: | ||||||
| 6694 | if (match(BO.getOperand(1), m_APInt(C)) && !C->isNullValue()) { | ||||||
| 6695 | // FIXME: If we have both nuw and nsw, we should reduce the range further. | ||||||
| 6696 | if (IIQ.hasNoUnsignedWrap(cast<OverflowingBinaryOperator>(&BO))) { | ||||||
| 6697 | // 'add nuw x, C' produces [C, UINT_MAX]. | ||||||
| 6698 | Lower = *C; | ||||||
| 6699 | } else if (IIQ.hasNoSignedWrap(cast<OverflowingBinaryOperator>(&BO))) { | ||||||
| 6700 | if (C->isNegative()) { | ||||||
| 6701 | // 'add nsw x, -C' produces [SINT_MIN, SINT_MAX - C]. | ||||||
| 6702 | Lower = APInt::getSignedMinValue(Width); | ||||||
| 6703 | Upper = APInt::getSignedMaxValue(Width) + *C + 1; | ||||||
| 6704 | } else { | ||||||
| 6705 | // 'add nsw x, +C' produces [SINT_MIN + C, SINT_MAX]. | ||||||
| 6706 | Lower = APInt::getSignedMinValue(Width) + *C; | ||||||
| 6707 | Upper = APInt::getSignedMaxValue(Width) + 1; | ||||||
| 6708 | } | ||||||
| 6709 | } | ||||||
| 6710 | } | ||||||
| 6711 | break; | ||||||
| 6712 | |||||||
| 6713 | case Instruction::And: | ||||||
| 6714 | if (match(BO.getOperand(1), m_APInt(C))) | ||||||
| 6715 | // 'and x, C' produces [0, C]. | ||||||
| 6716 | Upper = *C + 1; | ||||||
| 6717 | break; | ||||||
| 6718 | |||||||
| 6719 | case Instruction::Or: | ||||||
| 6720 | if (match(BO.getOperand(1), m_APInt(C))) | ||||||
| 6721 | // 'or x, C' produces [C, UINT_MAX]. | ||||||
| 6722 | Lower = *C; | ||||||
| 6723 | break; | ||||||
| 6724 | |||||||
| 6725 | case Instruction::AShr: | ||||||
| 6726 | if (match(BO.getOperand(1), m_APInt(C)) && C->ult(Width)) { | ||||||
| 6727 | // 'ashr x, C' produces [INT_MIN >> C, INT_MAX >> C]. | ||||||
| 6728 | Lower = APInt::getSignedMinValue(Width).ashr(*C); | ||||||
| 6729 | Upper = APInt::getSignedMaxValue(Width).ashr(*C) + 1; | ||||||
| 6730 | } else if (match(BO.getOperand(0), m_APInt(C))) { | ||||||
| 6731 | unsigned ShiftAmount = Width - 1; | ||||||
| 6732 | if (!C->isNullValue() && IIQ.isExact(&BO)) | ||||||
| 6733 | ShiftAmount = C->countTrailingZeros(); | ||||||
| 6734 | if (C->isNegative()) { | ||||||
| 6735 | // 'ashr C, x' produces [C, C >> (Width-1)] | ||||||
| 6736 | Lower = *C; | ||||||
| 6737 | Upper = C->ashr(ShiftAmount) + 1; | ||||||
| 6738 | } else { | ||||||
| 6739 | // 'ashr C, x' produces [C >> (Width-1), C] | ||||||
| 6740 | Lower = C->ashr(ShiftAmount); | ||||||
| 6741 | Upper = *C + 1; | ||||||
| 6742 | } | ||||||
| 6743 | } | ||||||
| 6744 | break; | ||||||
| 6745 | |||||||
| 6746 | case Instruction::LShr: | ||||||
| 6747 | if (match(BO.getOperand(1), m_APInt(C)) && C->ult(Width)) { | ||||||
| 6748 | // 'lshr x, C' produces [0, UINT_MAX >> C]. | ||||||
| 6749 | Upper = APInt::getAllOnesValue(Width).lshr(*C) + 1; | ||||||
| 6750 | } else if (match(BO.getOperand(0), m_APInt(C))) { | ||||||
| 6751 | // 'lshr C, x' produces [C >> (Width-1), C]. | ||||||
| 6752 | unsigned ShiftAmount = Width - 1; | ||||||
| 6753 | if (!C->isNullValue() && IIQ.isExact(&BO)) | ||||||
| 6754 | ShiftAmount = C->countTrailingZeros(); | ||||||
| 6755 | Lower = C->lshr(ShiftAmount); | ||||||
| 6756 | Upper = *C + 1; | ||||||
| 6757 | } | ||||||
| 6758 | break; | ||||||
| 6759 | |||||||
| 6760 | case Instruction::Shl: | ||||||
| 6761 | if (match(BO.getOperand(0), m_APInt(C))) { | ||||||
| 6762 | if (IIQ.hasNoUnsignedWrap(&BO)) { | ||||||
| 6763 | // 'shl nuw C, x' produces [C, C << CLZ(C)] | ||||||
| 6764 | Lower = *C; | ||||||
| 6765 | Upper = Lower.shl(Lower.countLeadingZeros()) + 1; | ||||||
| 6766 | } else if (BO.hasNoSignedWrap()) { // TODO: What if both nuw+nsw? | ||||||
| 6767 | if (C->isNegative()) { | ||||||
| 6768 | // 'shl nsw C, x' produces [C << CLO(C)-1, C] | ||||||
| 6769 | unsigned ShiftAmount = C->countLeadingOnes() - 1; | ||||||
| 6770 | Lower = C->shl(ShiftAmount); | ||||||
| 6771 | Upper = *C + 1; | ||||||
| 6772 | } else { | ||||||
| 6773 | // 'shl nsw C, x' produces [C, C << CLZ(C)-1] | ||||||
| 6774 | unsigned ShiftAmount = C->countLeadingZeros() - 1; | ||||||
| 6775 | Lower = *C; | ||||||
| 6776 | Upper = C->shl(ShiftAmount) + 1; | ||||||
| 6777 | } | ||||||
| 6778 | } | ||||||
| 6779 | } | ||||||
| 6780 | break; | ||||||
| 6781 | |||||||
| 6782 | case Instruction::SDiv: | ||||||
| 6783 | if (match(BO.getOperand(1), m_APInt(C))) { | ||||||
| 6784 | APInt IntMin = APInt::getSignedMinValue(Width); | ||||||
| 6785 | APInt IntMax = APInt::getSignedMaxValue(Width); | ||||||
| 6786 | if (C->isAllOnesValue()) { | ||||||
| 6787 | // 'sdiv x, -1' produces [INT_MIN + 1, INT_MAX] | ||||||
| 6788 | // where C != -1 and C != 0 and C != 1 | ||||||
| 6789 | Lower = IntMin + 1; | ||||||
| 6790 | Upper = IntMax + 1; | ||||||
| 6791 | } else if (C->countLeadingZeros() < Width - 1) { | ||||||
| 6792 | // 'sdiv x, C' produces [INT_MIN / C, INT_MAX / C] | ||||||
| 6793 | // where C != -1 and C != 0 and C != 1 | ||||||
| 6794 | Lower = IntMin.sdiv(*C); | ||||||
| 6795 | Upper = IntMax.sdiv(*C); | ||||||
| 6796 | if (Lower.sgt(Upper)) | ||||||
| 6797 | std::swap(Lower, Upper); | ||||||
| 6798 | Upper = Upper + 1; | ||||||
| 6799 | assert(Upper != Lower && "Upper part of range has wrapped!")((void)0); | ||||||
| 6800 | } | ||||||
| 6801 | } else if (match(BO.getOperand(0), m_APInt(C))) { | ||||||
| 6802 | if (C->isMinSignedValue()) { | ||||||
| 6803 | // 'sdiv INT_MIN, x' produces [INT_MIN, INT_MIN / -2]. | ||||||
| 6804 | Lower = *C; | ||||||
| 6805 | Upper = Lower.lshr(1) + 1; | ||||||
| 6806 | } else { | ||||||
| 6807 | // 'sdiv C, x' produces [-|C|, |C|]. | ||||||
| 6808 | Upper = C->abs() + 1; | ||||||
| 6809 | Lower = (-Upper) + 1; | ||||||
| 6810 | } | ||||||
| 6811 | } | ||||||
| 6812 | break; | ||||||
| 6813 | |||||||
| 6814 | case Instruction::UDiv: | ||||||
| 6815 | if (match(BO.getOperand(1), m_APInt(C)) && !C->isNullValue()) { | ||||||
| 6816 | // 'udiv x, C' produces [0, UINT_MAX / C]. | ||||||
| 6817 | Upper = APInt::getMaxValue(Width).udiv(*C) + 1; | ||||||
| 6818 | } else if (match(BO.getOperand(0), m_APInt(C))) { | ||||||
| 6819 | // 'udiv C, x' produces [0, C]. | ||||||
| 6820 | Upper = *C + 1; | ||||||
| 6821 | } | ||||||
| 6822 | break; | ||||||
| 6823 | |||||||
| 6824 | case Instruction::SRem: | ||||||
| 6825 | if (match(BO.getOperand(1), m_APInt(C))) { | ||||||
| 6826 | // 'srem x, C' produces (-|C|, |C|). | ||||||
| 6827 | Upper = C->abs(); | ||||||
| 6828 | Lower = (-Upper) + 1; | ||||||
| 6829 | } | ||||||
| 6830 | break; | ||||||
| 6831 | |||||||
| 6832 | case Instruction::URem: | ||||||
| 6833 | if (match(BO.getOperand(1), m_APInt(C))) | ||||||
| 6834 | // 'urem x, C' produces [0, C). | ||||||
| 6835 | Upper = *C; | ||||||
| 6836 | break; | ||||||
| 6837 | |||||||
| 6838 | default: | ||||||
| 6839 | break; | ||||||
| 6840 | } | ||||||
| 6841 | } | ||||||
| 6842 | |||||||
| 6843 | static void setLimitsForIntrinsic(const IntrinsicInst &II, APInt &Lower, | ||||||
| 6844 | APInt &Upper) { | ||||||
| 6845 | unsigned Width = Lower.getBitWidth(); | ||||||
| 6846 | const APInt *C; | ||||||
| 6847 | switch (II.getIntrinsicID()) { | ||||||
| 6848 | case Intrinsic::ctpop: | ||||||
| 6849 | case Intrinsic::ctlz: | ||||||
| 6850 | case Intrinsic::cttz: | ||||||
| 6851 | // Maximum of set/clear bits is the bit width. | ||||||
| 6852 | assert(Lower == 0 && "Expected lower bound to be zero")((void)0); | ||||||
| 6853 | Upper = Width + 1; | ||||||
| 6854 | break; | ||||||
| 6855 | case Intrinsic::uadd_sat: | ||||||
| 6856 | // uadd.sat(x, C) produces [C, UINT_MAX]. | ||||||
| 6857 | if (match(II.getOperand(0), m_APInt(C)) || | ||||||
| 6858 | match(II.getOperand(1), m_APInt(C))) | ||||||
| 6859 | Lower = *C; | ||||||
| 6860 | break; | ||||||
| 6861 | case Intrinsic::sadd_sat: | ||||||
| 6862 | if (match(II.getOperand(0), m_APInt(C)) || | ||||||
| 6863 | match(II.getOperand(1), m_APInt(C))) { | ||||||
| 6864 | if (C->isNegative()) { | ||||||
| 6865 | // sadd.sat(x, -C) produces [SINT_MIN, SINT_MAX + (-C)]. | ||||||
| 6866 | Lower = APInt::getSignedMinValue(Width); | ||||||
| 6867 | Upper = APInt::getSignedMaxValue(Width) + *C + 1; | ||||||
| 6868 | } else { | ||||||
| 6869 | // sadd.sat(x, +C) produces [SINT_MIN + C, SINT_MAX]. | ||||||
| 6870 | Lower = APInt::getSignedMinValue(Width) + *C; | ||||||
| 6871 | Upper = APInt::getSignedMaxValue(Width) + 1; | ||||||
| 6872 | } | ||||||
| 6873 | } | ||||||
| 6874 | break; | ||||||
| 6875 | case Intrinsic::usub_sat: | ||||||
| 6876 | // usub.sat(C, x) produces [0, C]. | ||||||
| 6877 | if (match(II.getOperand(0), m_APInt(C))) | ||||||
| 6878 | Upper = *C + 1; | ||||||
| 6879 | // usub.sat(x, C) produces [0, UINT_MAX - C]. | ||||||
| 6880 | else if (match(II.getOperand(1), m_APInt(C))) | ||||||
| 6881 | Upper = APInt::getMaxValue(Width) - *C + 1; | ||||||
| 6882 | break; | ||||||
| 6883 | case Intrinsic::ssub_sat: | ||||||
| 6884 | if (match(II.getOperand(0), m_APInt(C))) { | ||||||
| 6885 | if (C->isNegative()) { | ||||||
| 6886 | // ssub.sat(-C, x) produces [SINT_MIN, -SINT_MIN + (-C)]. | ||||||
| 6887 | Lower = APInt::getSignedMinValue(Width); | ||||||
| 6888 | Upper = *C - APInt::getSignedMinValue(Width) + 1; | ||||||
| 6889 | } else { | ||||||
| 6890 | // ssub.sat(+C, x) produces [-SINT_MAX + C, SINT_MAX]. | ||||||
| 6891 | Lower = *C - APInt::getSignedMaxValue(Width); | ||||||
| 6892 | Upper = APInt::getSignedMaxValue(Width) + 1; | ||||||
| 6893 | } | ||||||
| 6894 | } else if (match(II.getOperand(1), m_APInt(C))) { | ||||||
| 6895 | if (C->isNegative()) { | ||||||
| 6896 | // ssub.sat(x, -C) produces [SINT_MIN - (-C), SINT_MAX]: | ||||||
| 6897 | Lower = APInt::getSignedMinValue(Width) - *C; | ||||||
| 6898 | Upper = APInt::getSignedMaxValue(Width) + 1; | ||||||
| 6899 | } else { | ||||||
| 6900 | // ssub.sat(x, +C) produces [SINT_MIN, SINT_MAX - C]. | ||||||
| 6901 | Lower = APInt::getSignedMinValue(Width); | ||||||
| 6902 | Upper = APInt::getSignedMaxValue(Width) - *C + 1; | ||||||
| 6903 | } | ||||||
| 6904 | } | ||||||
| 6905 | break; | ||||||
| 6906 | case Intrinsic::umin: | ||||||
| 6907 | case Intrinsic::umax: | ||||||
| 6908 | case Intrinsic::smin: | ||||||
| 6909 | case Intrinsic::smax: | ||||||
| 6910 | if (!match(II.getOperand(0), m_APInt(C)) && | ||||||
| 6911 | !match(II.getOperand(1), m_APInt(C))) | ||||||
| 6912 | break; | ||||||
| 6913 | |||||||
| 6914 | switch (II.getIntrinsicID()) { | ||||||
| 6915 | case Intrinsic::umin: | ||||||
| 6916 | Upper = *C + 1; | ||||||
| 6917 | break; | ||||||
| 6918 | case Intrinsic::umax: | ||||||
| 6919 | Lower = *C; | ||||||
| 6920 | break; | ||||||
| 6921 | case Intrinsic::smin: | ||||||
| 6922 | Lower = APInt::getSignedMinValue(Width); | ||||||
| 6923 | Upper = *C + 1; | ||||||
| 6924 | break; | ||||||
| 6925 | case Intrinsic::smax: | ||||||
| 6926 | Lower = *C; | ||||||
| 6927 | Upper = APInt::getSignedMaxValue(Width) + 1; | ||||||
| 6928 | break; | ||||||
| 6929 | default: | ||||||
| 6930 | llvm_unreachable("Must be min/max intrinsic")__builtin_unreachable(); | ||||||
| 6931 | } | ||||||
| 6932 | break; | ||||||
| 6933 | case Intrinsic::abs: | ||||||
| 6934 | // If abs of SIGNED_MIN is poison, then the result is [0..SIGNED_MAX], | ||||||
| 6935 | // otherwise it is [0..SIGNED_MIN], as -SIGNED_MIN == SIGNED_MIN. | ||||||
| 6936 | if (match(II.getOperand(1), m_One())) | ||||||
| 6937 | Upper = APInt::getSignedMaxValue(Width) + 1; | ||||||
| 6938 | else | ||||||
| 6939 | Upper = APInt::getSignedMinValue(Width) + 1; | ||||||
| 6940 | break; | ||||||
| 6941 | default: | ||||||
| 6942 | break; | ||||||
| 6943 | } | ||||||
| 6944 | } | ||||||
| 6945 | |||||||
| 6946 | static void setLimitsForSelectPattern(const SelectInst &SI, APInt &Lower, | ||||||
| 6947 | APInt &Upper, const InstrInfoQuery &IIQ) { | ||||||
| 6948 | const Value *LHS = nullptr, *RHS = nullptr; | ||||||
| 6949 | SelectPatternResult R = matchSelectPattern(&SI, LHS, RHS); | ||||||
| 6950 | if (R.Flavor == SPF_UNKNOWN) | ||||||
| 6951 | return; | ||||||
| 6952 | |||||||
| 6953 | unsigned BitWidth = SI.getType()->getScalarSizeInBits(); | ||||||
| 6954 | |||||||
| 6955 | if (R.Flavor == SelectPatternFlavor::SPF_ABS) { | ||||||
| 6956 | // If the negation part of the abs (in RHS) has the NSW flag, | ||||||
| 6957 | // then the result of abs(X) is [0..SIGNED_MAX], | ||||||
| 6958 | // otherwise it is [0..SIGNED_MIN], as -SIGNED_MIN == SIGNED_MIN. | ||||||
| 6959 | Lower = APInt::getNullValue(BitWidth); | ||||||
| 6960 | if (match(RHS, m_Neg(m_Specific(LHS))) && | ||||||
| 6961 | IIQ.hasNoSignedWrap(cast<Instruction>(RHS))) | ||||||
| 6962 | Upper = APInt::getSignedMaxValue(BitWidth) + 1; | ||||||
| 6963 | else | ||||||
| 6964 | Upper = APInt::getSignedMinValue(BitWidth) + 1; | ||||||
| 6965 | return; | ||||||
| 6966 | } | ||||||
| 6967 | |||||||
| 6968 | if (R.Flavor == SelectPatternFlavor::SPF_NABS) { | ||||||
| 6969 | // The result of -abs(X) is <= 0. | ||||||
| 6970 | Lower = APInt::getSignedMinValue(BitWidth); | ||||||
| 6971 | Upper = APInt(BitWidth, 1); | ||||||
| 6972 | return; | ||||||
| 6973 | } | ||||||
| 6974 | |||||||
| 6975 | const APInt *C; | ||||||
| 6976 | if (!match(LHS, m_APInt(C)) && !match(RHS, m_APInt(C))) | ||||||
| 6977 | return; | ||||||
| 6978 | |||||||
| 6979 | switch (R.Flavor) { | ||||||
| 6980 | case SPF_UMIN: | ||||||
| 6981 | Upper = *C + 1; | ||||||
| 6982 | break; | ||||||
| 6983 | case SPF_UMAX: | ||||||
| 6984 | Lower = *C; | ||||||
| 6985 | break; | ||||||
| 6986 | case SPF_SMIN: | ||||||
| 6987 | Lower = APInt::getSignedMinValue(BitWidth); | ||||||
| 6988 | Upper = *C + 1; | ||||||
| 6989 | break; | ||||||
| 6990 | case SPF_SMAX: | ||||||
| 6991 | Lower = *C; | ||||||
| 6992 | Upper = APInt::getSignedMaxValue(BitWidth) + 1; | ||||||
| 6993 | break; | ||||||
| 6994 | default: | ||||||
| 6995 | break; | ||||||
| 6996 | } | ||||||
| 6997 | } | ||||||
| 6998 | |||||||
| 6999 | ConstantRange llvm::computeConstantRange(const Value *V, bool UseInstrInfo, | ||||||
| 7000 | AssumptionCache *AC, | ||||||
| 7001 | const Instruction *CtxI, | ||||||
| 7002 | unsigned Depth) { | ||||||
| 7003 | assert(V->getType()->isIntOrIntVectorTy() && "Expected integer instruction")((void)0); | ||||||
| 7004 | |||||||
| 7005 | if (Depth == MaxAnalysisRecursionDepth) | ||||||
| 7006 | return ConstantRange::getFull(V->getType()->getScalarSizeInBits()); | ||||||
| 7007 | |||||||
| 7008 | const APInt *C; | ||||||
| 7009 | if (match(V, m_APInt(C))) | ||||||
| 7010 | return ConstantRange(*C); | ||||||
| 7011 | |||||||
| 7012 | InstrInfoQuery IIQ(UseInstrInfo); | ||||||
| 7013 | unsigned BitWidth = V->getType()->getScalarSizeInBits(); | ||||||
| 7014 | APInt Lower = APInt(BitWidth, 0); | ||||||
| 7015 | APInt Upper = APInt(BitWidth, 0); | ||||||
| 7016 | if (auto *BO = dyn_cast<BinaryOperator>(V)) | ||||||
| 7017 | setLimitsForBinOp(*BO, Lower, Upper, IIQ); | ||||||
| 7018 | else if (auto *II = dyn_cast<IntrinsicInst>(V)) | ||||||
| 7019 | setLimitsForIntrinsic(*II, Lower, Upper); | ||||||
| 7020 | else if (auto *SI = dyn_cast<SelectInst>(V)) | ||||||
| 7021 | setLimitsForSelectPattern(*SI, Lower, Upper, IIQ); | ||||||
| 7022 | |||||||
| 7023 | ConstantRange CR = ConstantRange::getNonEmpty(Lower, Upper); | ||||||
| 7024 | |||||||
| 7025 | if (auto *I = dyn_cast<Instruction>(V)) | ||||||
| 7026 | if (auto *Range = IIQ.getMetadata(I, LLVMContext::MD_range)) | ||||||
| 7027 | CR = CR.intersectWith(getConstantRangeFromMetadata(*Range)); | ||||||
| 7028 | |||||||
| 7029 | if (CtxI && AC) { | ||||||
| 7030 | // Try to restrict the range based on information from assumptions. | ||||||
| 7031 | for (auto &AssumeVH : AC->assumptionsFor(V)) { | ||||||
| 7032 | if (!AssumeVH) | ||||||
| 7033 | continue; | ||||||
| 7034 | CallInst *I = cast<CallInst>(AssumeVH); | ||||||
| 7035 | assert(I->getParent()->getParent() == CtxI->getParent()->getParent() &&((void)0) | ||||||
| 7036 | "Got assumption for the wrong function!")((void)0); | ||||||
| 7037 | assert(I->getCalledFunction()->getIntrinsicID() == Intrinsic::assume &&((void)0) | ||||||
| 7038 | "must be an assume intrinsic")((void)0); | ||||||
| 7039 | |||||||
| 7040 | if (!isValidAssumeForContext(I, CtxI, nullptr)) | ||||||
| 7041 | continue; | ||||||
| 7042 | Value *Arg = I->getArgOperand(0); | ||||||
| 7043 | ICmpInst *Cmp = dyn_cast<ICmpInst>(Arg); | ||||||
| 7044 | // Currently we just use information from comparisons. | ||||||
| 7045 | if (!Cmp || Cmp->getOperand(0) != V) | ||||||
| 7046 | continue; | ||||||
| 7047 | ConstantRange RHS = computeConstantRange(Cmp->getOperand(1), UseInstrInfo, | ||||||
| 7048 | AC, I, Depth + 1); | ||||||
| 7049 | CR = CR.intersectWith( | ||||||
| 7050 | ConstantRange::makeSatisfyingICmpRegion(Cmp->getPredicate(), RHS)); | ||||||
| 7051 | } | ||||||
| 7052 | } | ||||||
| 7053 | |||||||
| 7054 | return CR; | ||||||
| 7055 | } | ||||||
| 7056 | |||||||
| 7057 | static Optional<int64_t> | ||||||
| 7058 | getOffsetFromIndex(const GEPOperator *GEP, unsigned Idx, const DataLayout &DL) { | ||||||
| 7059 | // Skip over the first indices. | ||||||
| 7060 | gep_type_iterator GTI = gep_type_begin(GEP); | ||||||
| 7061 | for (unsigned i = 1; i != Idx; ++i, ++GTI) | ||||||
| 7062 | /*skip along*/; | ||||||
| 7063 | |||||||
| 7064 | // Compute the offset implied by the rest of the indices. | ||||||
| 7065 | int64_t Offset = 0; | ||||||
| 7066 | for (unsigned i = Idx, e = GEP->getNumOperands(); i != e; ++i, ++GTI) { | ||||||
| 7067 | ConstantInt *OpC = dyn_cast<ConstantInt>(GEP->getOperand(i)); | ||||||
| 7068 | if (!OpC) | ||||||
| 7069 | return None; | ||||||
| 7070 | if (OpC->isZero()) | ||||||
| 7071 | continue; // No offset. | ||||||
| 7072 | |||||||
| 7073 | // Handle struct indices, which add their field offset to the pointer. | ||||||
| 7074 | if (StructType *STy = GTI.getStructTypeOrNull()) { | ||||||
| 7075 | Offset += DL.getStructLayout(STy)->getElementOffset(OpC->getZExtValue()); | ||||||
| 7076 | continue; | ||||||
| 7077 | } | ||||||
| 7078 | |||||||
| 7079 | // Otherwise, we have a sequential type like an array or fixed-length | ||||||
| 7080 | // vector. Multiply the index by the ElementSize. | ||||||
| 7081 | TypeSize Size = DL.getTypeAllocSize(GTI.getIndexedType()); | ||||||
| 7082 | if (Size.isScalable()) | ||||||
| 7083 | return None; | ||||||
| 7084 | Offset += Size.getFixedSize() * OpC->getSExtValue(); | ||||||
| 7085 | } | ||||||
| 7086 | |||||||
| 7087 | return Offset; | ||||||
| 7088 | } | ||||||
| 7089 | |||||||
| 7090 | Optional<int64_t> llvm::isPointerOffset(const Value *Ptr1, const Value *Ptr2, | ||||||
| 7091 | const DataLayout &DL) { | ||||||
| 7092 | Ptr1 = Ptr1->stripPointerCasts(); | ||||||
| 7093 | Ptr2 = Ptr2->stripPointerCasts(); | ||||||
| 7094 | |||||||
| 7095 | // Handle the trivial case first. | ||||||
| 7096 | if (Ptr1 == Ptr2) { | ||||||
| 7097 | return 0; | ||||||
| 7098 | } | ||||||
| 7099 | |||||||
| 7100 | const GEPOperator *GEP1 = dyn_cast<GEPOperator>(Ptr1); | ||||||
| 7101 | const GEPOperator *GEP2 = dyn_cast<GEPOperator>(Ptr2); | ||||||
| 7102 | |||||||
| 7103 | // If one pointer is a GEP see if the GEP is a constant offset from the base, | ||||||
| 7104 | // as in "P" and "gep P, 1". | ||||||
| 7105 | // Also do this iteratively to handle the the following case: | ||||||
| 7106 | // Ptr_t1 = GEP Ptr1, c1 | ||||||
| 7107 | // Ptr_t2 = GEP Ptr_t1, c2 | ||||||
| 7108 | // Ptr2 = GEP Ptr_t2, c3 | ||||||
| 7109 | // where we will return c1+c2+c3. | ||||||
| 7110 | // TODO: Handle the case when both Ptr1 and Ptr2 are GEPs of some common base | ||||||
| 7111 | // -- replace getOffsetFromBase with getOffsetAndBase, check that the bases | ||||||
| 7112 | // are the same, and return the difference between offsets. | ||||||
| 7113 | auto getOffsetFromBase = [&DL](const GEPOperator *GEP, | ||||||
| 7114 | const Value *Ptr) -> Optional<int64_t> { | ||||||
| 7115 | const GEPOperator *GEP_T = GEP; | ||||||
| 7116 | int64_t OffsetVal = 0; | ||||||
| 7117 | bool HasSameBase = false; | ||||||
| 7118 | while (GEP_T) { | ||||||
| 7119 | auto Offset = getOffsetFromIndex(GEP_T, 1, DL); | ||||||
| 7120 | if (!Offset) | ||||||
| 7121 | return None; | ||||||
| 7122 | OffsetVal += *Offset; | ||||||
| 7123 | auto Op0 = GEP_T->getOperand(0)->stripPointerCasts(); | ||||||
| 7124 | if (Op0 == Ptr) { | ||||||
| 7125 | HasSameBase = true; | ||||||
| 7126 | break; | ||||||
| 7127 | } | ||||||
| 7128 | GEP_T = dyn_cast<GEPOperator>(Op0); | ||||||
| 7129 | } | ||||||
| 7130 | if (!HasSameBase) | ||||||
| 7131 | return None; | ||||||
| 7132 | return OffsetVal; | ||||||
| 7133 | }; | ||||||
| 7134 | |||||||
| 7135 | if (GEP1) { | ||||||
| 7136 | auto Offset = getOffsetFromBase(GEP1, Ptr2); | ||||||
| 7137 | if (Offset) | ||||||
| 7138 | return -*Offset; | ||||||
| 7139 | } | ||||||
| 7140 | if (GEP2) { | ||||||
| 7141 | auto Offset = getOffsetFromBase(GEP2, Ptr1); | ||||||
| 7142 | if (Offset) | ||||||
| 7143 | return Offset; | ||||||
| 7144 | } | ||||||
| 7145 | |||||||
| 7146 | // Right now we handle the case when Ptr1/Ptr2 are both GEPs with an identical | ||||||
| 7147 | // base. After that base, they may have some number of common (and | ||||||
| 7148 | // potentially variable) indices. After that they handle some constant | ||||||
| 7149 | // offset, which determines their offset from each other. At this point, we | ||||||
| 7150 | // handle no other case. | ||||||
| 7151 | if (!GEP1 || !GEP2 || GEP1->getOperand(0) != GEP2->getOperand(0)) | ||||||
| 7152 | return None; | ||||||
| 7153 | |||||||
| 7154 | // Skip any common indices and track the GEP types. | ||||||
| 7155 | unsigned Idx = 1; | ||||||
| 7156 | for (; Idx != GEP1->getNumOperands() && Idx != GEP2->getNumOperands(); ++Idx) | ||||||
| 7157 | if (GEP1->getOperand(Idx) != GEP2->getOperand(Idx)) | ||||||
| 7158 | break; | ||||||
| 7159 | |||||||
| 7160 | auto Offset1 = getOffsetFromIndex(GEP1, Idx, DL); | ||||||
| 7161 | auto Offset2 = getOffsetFromIndex(GEP2, Idx, DL); | ||||||
| 7162 | if (!Offset1 || !Offset2) | ||||||
| 7163 | return None; | ||||||
| 7164 | return *Offset2 - *Offset1; | ||||||
| 7165 | } |
| 1 | //===- llvm/InstrTypes.h - Important Instruction subclasses -----*- C++ -*-===// | ||||||
| 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 various meta classes of instructions that exist in the VM | ||||||
| 10 | // representation. Specific concrete subclasses of these may be found in the | ||||||
| 11 | // i*.h files... | ||||||
| 12 | // | ||||||
| 13 | //===----------------------------------------------------------------------===// | ||||||
| 14 | |||||||
| 15 | #ifndef LLVM_IR_INSTRTYPES_H | ||||||
| 16 | #define LLVM_IR_INSTRTYPES_H | ||||||
| 17 | |||||||
| 18 | #include "llvm/ADT/ArrayRef.h" | ||||||
| 19 | #include "llvm/ADT/None.h" | ||||||
| 20 | #include "llvm/ADT/Optional.h" | ||||||
| 21 | #include "llvm/ADT/STLExtras.h" | ||||||
| 22 | #include "llvm/ADT/StringMap.h" | ||||||
| 23 | #include "llvm/ADT/StringRef.h" | ||||||
| 24 | #include "llvm/ADT/Twine.h" | ||||||
| 25 | #include "llvm/ADT/iterator_range.h" | ||||||
| 26 | #include "llvm/IR/Attributes.h" | ||||||
| 27 | #include "llvm/IR/CallingConv.h" | ||||||
| 28 | #include "llvm/IR/Constants.h" | ||||||
| 29 | #include "llvm/IR/DerivedTypes.h" | ||||||
| 30 | #include "llvm/IR/Function.h" | ||||||
| 31 | #include "llvm/IR/Instruction.h" | ||||||
| 32 | #include "llvm/IR/LLVMContext.h" | ||||||
| 33 | #include "llvm/IR/OperandTraits.h" | ||||||
| 34 | #include "llvm/IR/Type.h" | ||||||
| 35 | #include "llvm/IR/User.h" | ||||||
| 36 | #include "llvm/IR/Value.h" | ||||||
| 37 | #include "llvm/Support/Casting.h" | ||||||
| 38 | #include "llvm/Support/ErrorHandling.h" | ||||||
| 39 | #include <algorithm> | ||||||
| 40 | #include <cassert> | ||||||
| 41 | #include <cstddef> | ||||||
| 42 | #include <cstdint> | ||||||
| 43 | #include <iterator> | ||||||
| 44 | #include <string> | ||||||
| 45 | #include <vector> | ||||||
| 46 | |||||||
| 47 | namespace llvm { | ||||||
| 48 | |||||||
| 49 | namespace Intrinsic { | ||||||
| 50 | typedef unsigned ID; | ||||||
| 51 | } | ||||||
| 52 | |||||||
| 53 | //===----------------------------------------------------------------------===// | ||||||
| 54 | // UnaryInstruction Class | ||||||
| 55 | //===----------------------------------------------------------------------===// | ||||||
| 56 | |||||||
| 57 | class UnaryInstruction : public Instruction { | ||||||
| 58 | protected: | ||||||
| 59 | UnaryInstruction(Type *Ty, unsigned iType, Value *V, | ||||||
| 60 | Instruction *IB = nullptr) | ||||||
| 61 | : Instruction(Ty, iType, &Op<0>(), 1, IB) { | ||||||
| 62 | Op<0>() = V; | ||||||
| 63 | } | ||||||
| 64 | UnaryInstruction(Type *Ty, unsigned iType, Value *V, BasicBlock *IAE) | ||||||
| 65 | : Instruction(Ty, iType, &Op<0>(), 1, IAE) { | ||||||
| 66 | Op<0>() = V; | ||||||
| 67 | } | ||||||
| 68 | |||||||
| 69 | public: | ||||||
| 70 | // allocate space for exactly one operand | ||||||
| 71 | void *operator new(size_t S) { return User::operator new(S, 1); } | ||||||
| 72 | void operator delete(void *Ptr) { User::operator delete(Ptr); } | ||||||
| 73 | |||||||
| 74 | /// Transparently provide more efficient getOperand methods. | ||||||
| 75 | DECLARE_TRANSPARENT_OPERAND_ACCESSORS(Value)public: inline Value *getOperand(unsigned) const; inline void setOperand(unsigned, Value*); inline op_iterator op_begin(); inline const_op_iterator op_begin() const; inline op_iterator op_end(); inline const_op_iterator op_end() const; protected : template <int> inline Use &Op(); template <int > inline const Use &Op() const; public: inline unsigned getNumOperands() const; | ||||||
| 76 | |||||||
| 77 | // Methods for support type inquiry through isa, cast, and dyn_cast: | ||||||
| 78 | static bool classof(const Instruction *I) { | ||||||
| 79 | return I->isUnaryOp() || | ||||||
| 80 | I->getOpcode() == Instruction::Alloca || | ||||||
| 81 | I->getOpcode() == Instruction::Load || | ||||||
| 82 | I->getOpcode() == Instruction::VAArg || | ||||||
| 83 | I->getOpcode() == Instruction::ExtractValue || | ||||||
| 84 | (I->getOpcode() >= CastOpsBegin && I->getOpcode() < CastOpsEnd); | ||||||
| 85 | } | ||||||
| 86 | static bool classof(const Value *V) { | ||||||
| 87 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); | ||||||
| 88 | } | ||||||
| 89 | }; | ||||||
| 90 | |||||||
| 91 | template <> | ||||||
| 92 | struct OperandTraits<UnaryInstruction> : | ||||||
| 93 | public FixedNumOperandTraits<UnaryInstruction, 1> { | ||||||
| 94 | }; | ||||||
| 95 | |||||||
| 96 | DEFINE_TRANSPARENT_OPERAND_ACCESSORS(UnaryInstruction, Value)UnaryInstruction::op_iterator UnaryInstruction::op_begin() { return OperandTraits<UnaryInstruction>::op_begin(this); } UnaryInstruction ::const_op_iterator UnaryInstruction::op_begin() const { return OperandTraits<UnaryInstruction>::op_begin(const_cast< UnaryInstruction*>(this)); } UnaryInstruction::op_iterator UnaryInstruction::op_end() { return OperandTraits<UnaryInstruction >::op_end(this); } UnaryInstruction::const_op_iterator UnaryInstruction ::op_end() const { return OperandTraits<UnaryInstruction> ::op_end(const_cast<UnaryInstruction*>(this)); } Value * UnaryInstruction::getOperand(unsigned i_nocapture) const { (( void)0); return cast_or_null<Value>( OperandTraits<UnaryInstruction >::op_begin(const_cast<UnaryInstruction*>(this))[i_nocapture ].get()); } void UnaryInstruction::setOperand(unsigned i_nocapture , Value *Val_nocapture) { ((void)0); OperandTraits<UnaryInstruction >::op_begin(this)[i_nocapture] = Val_nocapture; } unsigned UnaryInstruction::getNumOperands() const { return OperandTraits <UnaryInstruction>::operands(this); } template <int Idx_nocapture > Use &UnaryInstruction::Op() { return this->OpFrom <Idx_nocapture>(this); } template <int Idx_nocapture > const Use &UnaryInstruction::Op() const { return this ->OpFrom<Idx_nocapture>(this); } | ||||||
| 97 | |||||||
| 98 | //===----------------------------------------------------------------------===// | ||||||
| 99 | // UnaryOperator Class | ||||||
| 100 | //===----------------------------------------------------------------------===// | ||||||
| 101 | |||||||
| 102 | class UnaryOperator : public UnaryInstruction { | ||||||
| 103 | void AssertOK(); | ||||||
| 104 | |||||||
| 105 | protected: | ||||||
| 106 | UnaryOperator(UnaryOps iType, Value *S, Type *Ty, | ||||||
| 107 | const Twine &Name, Instruction *InsertBefore); | ||||||
| 108 | UnaryOperator(UnaryOps iType, Value *S, Type *Ty, | ||||||
| 109 | const Twine &Name, BasicBlock *InsertAtEnd); | ||||||
| 110 | |||||||
| 111 | // Note: Instruction needs to be a friend here to call cloneImpl. | ||||||
| 112 | friend class Instruction; | ||||||
| 113 | |||||||
| 114 | UnaryOperator *cloneImpl() const; | ||||||
| 115 | |||||||
| 116 | public: | ||||||
| 117 | |||||||
| 118 | /// Construct a unary instruction, given the opcode and an operand. | ||||||
| 119 | /// Optionally (if InstBefore is specified) insert the instruction | ||||||
| 120 | /// into a BasicBlock right before the specified instruction. The specified | ||||||
| 121 | /// Instruction is allowed to be a dereferenced end iterator. | ||||||
| 122 | /// | ||||||
| 123 | static UnaryOperator *Create(UnaryOps Op, Value *S, | ||||||
| 124 | const Twine &Name = Twine(), | ||||||
| 125 | Instruction *InsertBefore = nullptr); | ||||||
| 126 | |||||||
| 127 | /// Construct a unary instruction, given the opcode and an operand. | ||||||
| 128 | /// Also automatically insert this instruction to the end of the | ||||||
| 129 | /// BasicBlock specified. | ||||||
| 130 | /// | ||||||
| 131 | static UnaryOperator *Create(UnaryOps Op, Value *S, | ||||||
| 132 | const Twine &Name, | ||||||
| 133 | BasicBlock *InsertAtEnd); | ||||||
| 134 | |||||||
| 135 | /// These methods just forward to Create, and are useful when you | ||||||
| 136 | /// statically know what type of instruction you're going to create. These | ||||||
| 137 | /// helpers just save some typing. | ||||||
| 138 | #define HANDLE_UNARY_INST(N, OPC, CLASS) \ | ||||||
| 139 | static UnaryOperator *Create##OPC(Value *V, const Twine &Name = "") {\ | ||||||
| 140 | return Create(Instruction::OPC, V, Name);\ | ||||||
| 141 | } | ||||||
| 142 | #include "llvm/IR/Instruction.def" | ||||||
| 143 | #define HANDLE_UNARY_INST(N, OPC, CLASS) \ | ||||||
| 144 | static UnaryOperator *Create##OPC(Value *V, const Twine &Name, \ | ||||||
| 145 | BasicBlock *BB) {\ | ||||||
| 146 | return Create(Instruction::OPC, V, Name, BB);\ | ||||||
| 147 | } | ||||||
| 148 | #include "llvm/IR/Instruction.def" | ||||||
| 149 | #define HANDLE_UNARY_INST(N, OPC, CLASS) \ | ||||||
| 150 | static UnaryOperator *Create##OPC(Value *V, const Twine &Name, \ | ||||||
| 151 | Instruction *I) {\ | ||||||
| 152 | return Create(Instruction::OPC, V, Name, I);\ | ||||||
| 153 | } | ||||||
| 154 | #include "llvm/IR/Instruction.def" | ||||||
| 155 | |||||||
| 156 | static UnaryOperator * | ||||||
| 157 | CreateWithCopiedFlags(UnaryOps Opc, Value *V, Instruction *CopyO, | ||||||
| 158 | const Twine &Name = "", | ||||||
| 159 | Instruction *InsertBefore = nullptr) { | ||||||
| 160 | UnaryOperator *UO = Create(Opc, V, Name, InsertBefore); | ||||||
| 161 | UO->copyIRFlags(CopyO); | ||||||
| 162 | return UO; | ||||||
| 163 | } | ||||||
| 164 | |||||||
| 165 | static UnaryOperator *CreateFNegFMF(Value *Op, Instruction *FMFSource, | ||||||
| 166 | const Twine &Name = "", | ||||||
| 167 | Instruction *InsertBefore = nullptr) { | ||||||
| 168 | return CreateWithCopiedFlags(Instruction::FNeg, Op, FMFSource, Name, | ||||||
| 169 | InsertBefore); | ||||||
| 170 | } | ||||||
| 171 | |||||||
| 172 | UnaryOps getOpcode() const { | ||||||
| 173 | return static_cast<UnaryOps>(Instruction::getOpcode()); | ||||||
| 174 | } | ||||||
| 175 | |||||||
| 176 | // Methods for support type inquiry through isa, cast, and dyn_cast: | ||||||
| 177 | static bool classof(const Instruction *I) { | ||||||
| 178 | return I->isUnaryOp(); | ||||||
| 179 | } | ||||||
| 180 | static bool classof(const Value *V) { | ||||||
| 181 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); | ||||||
| 182 | } | ||||||
| 183 | }; | ||||||
| 184 | |||||||
| 185 | //===----------------------------------------------------------------------===// | ||||||
| 186 | // BinaryOperator Class | ||||||
| 187 | //===----------------------------------------------------------------------===// | ||||||
| 188 | |||||||
| 189 | class BinaryOperator : public Instruction { | ||||||
| 190 | void AssertOK(); | ||||||
| 191 | |||||||
| 192 | protected: | ||||||
| 193 | BinaryOperator(BinaryOps iType, Value *S1, Value *S2, Type *Ty, | ||||||
| 194 | const Twine &Name, Instruction *InsertBefore); | ||||||
| 195 | BinaryOperator(BinaryOps iType, Value *S1, Value *S2, Type *Ty, | ||||||
| 196 | const Twine &Name, BasicBlock *InsertAtEnd); | ||||||
| 197 | |||||||
| 198 | // Note: Instruction needs to be a friend here to call cloneImpl. | ||||||
| 199 | friend class Instruction; | ||||||
| 200 | |||||||
| 201 | BinaryOperator *cloneImpl() const; | ||||||
| 202 | |||||||
| 203 | public: | ||||||
| 204 | // allocate space for exactly two operands | ||||||
| 205 | void *operator new(size_t S) { return User::operator new(S, 2); } | ||||||
| 206 | void operator delete(void *Ptr) { User::operator delete(Ptr); } | ||||||
| 207 | |||||||
| 208 | /// Transparently provide more efficient getOperand methods. | ||||||
| 209 | DECLARE_TRANSPARENT_OPERAND_ACCESSORS(Value)public: inline Value *getOperand(unsigned) const; inline void setOperand(unsigned, Value*); inline op_iterator op_begin(); inline const_op_iterator op_begin() const; inline op_iterator op_end(); inline const_op_iterator op_end() const; protected : template <int> inline Use &Op(); template <int > inline const Use &Op() const; public: inline unsigned getNumOperands() const; | ||||||
| 210 | |||||||
| 211 | /// Construct a binary instruction, given the opcode and the two | ||||||
| 212 | /// operands. Optionally (if InstBefore is specified) insert the instruction | ||||||
| 213 | /// into a BasicBlock right before the specified instruction. The specified | ||||||
| 214 | /// Instruction is allowed to be a dereferenced end iterator. | ||||||
| 215 | /// | ||||||
| 216 | static BinaryOperator *Create(BinaryOps Op, Value *S1, Value *S2, | ||||||
| 217 | const Twine &Name = Twine(), | ||||||
| 218 | Instruction *InsertBefore = nullptr); | ||||||
| 219 | |||||||
| 220 | /// Construct a binary instruction, given the opcode and the two | ||||||
| 221 | /// operands. Also automatically insert this instruction to the end of the | ||||||
| 222 | /// BasicBlock specified. | ||||||
| 223 | /// | ||||||
| 224 | static BinaryOperator *Create(BinaryOps Op, Value *S1, Value *S2, | ||||||
| 225 | const Twine &Name, BasicBlock *InsertAtEnd); | ||||||
| 226 | |||||||
| 227 | /// These methods just forward to Create, and are useful when you | ||||||
| 228 | /// statically know what type of instruction you're going to create. These | ||||||
| 229 | /// helpers just save some typing. | ||||||
| 230 | #define HANDLE_BINARY_INST(N, OPC, CLASS) \ | ||||||
| 231 | static BinaryOperator *Create##OPC(Value *V1, Value *V2, \ | ||||||
| 232 | const Twine &Name = "") {\ | ||||||
| 233 | return Create(Instruction::OPC, V1, V2, Name);\ | ||||||
| 234 | } | ||||||
| 235 | #include "llvm/IR/Instruction.def" | ||||||
| 236 | #define HANDLE_BINARY_INST(N, OPC, CLASS) \ | ||||||
| 237 | static BinaryOperator *Create##OPC(Value *V1, Value *V2, \ | ||||||
| 238 | const Twine &Name, BasicBlock *BB) {\ | ||||||
| 239 | return Create(Instruction::OPC, V1, V2, Name, BB);\ | ||||||
| 240 | } | ||||||
| 241 | #include "llvm/IR/Instruction.def" | ||||||
| 242 | #define HANDLE_BINARY_INST(N, OPC, CLASS) \ | ||||||
| 243 | static BinaryOperator *Create##OPC(Value *V1, Value *V2, \ | ||||||
| 244 | const Twine &Name, Instruction *I) {\ | ||||||
| 245 | return Create(Instruction::OPC, V1, V2, Name, I);\ | ||||||
| 246 | } | ||||||
| 247 | #include "llvm/IR/Instruction.def" | ||||||
| 248 | |||||||
| 249 | static BinaryOperator * | ||||||
| 250 | CreateWithCopiedFlags(BinaryOps Opc, Value *V1, Value *V2, Instruction *CopyO, | ||||||
| 251 | const Twine &Name = "", | ||||||
| 252 | Instruction *InsertBefore = nullptr) { | ||||||
| 253 | BinaryOperator *BO = Create(Opc, V1, V2, Name, InsertBefore); | ||||||
| 254 | BO->copyIRFlags(CopyO); | ||||||
| 255 | return BO; | ||||||
| 256 | } | ||||||
| 257 | |||||||
| 258 | static BinaryOperator *CreateFAddFMF(Value *V1, Value *V2, | ||||||
| 259 | Instruction *FMFSource, | ||||||
| 260 | const Twine &Name = "") { | ||||||
| 261 | return CreateWithCopiedFlags(Instruction::FAdd, V1, V2, FMFSource, Name); | ||||||
| 262 | } | ||||||
| 263 | static BinaryOperator *CreateFSubFMF(Value *V1, Value *V2, | ||||||
| 264 | Instruction *FMFSource, | ||||||
| 265 | const Twine &Name = "") { | ||||||
| 266 | return CreateWithCopiedFlags(Instruction::FSub, V1, V2, FMFSource, Name); | ||||||
| 267 | } | ||||||
| 268 | static BinaryOperator *CreateFMulFMF(Value *V1, Value *V2, | ||||||
| 269 | Instruction *FMFSource, | ||||||
| 270 | const Twine &Name = "") { | ||||||
| 271 | return CreateWithCopiedFlags(Instruction::FMul, V1, V2, FMFSource, Name); | ||||||
| 272 | } | ||||||
| 273 | static BinaryOperator *CreateFDivFMF(Value *V1, Value *V2, | ||||||
| 274 | Instruction *FMFSource, | ||||||
| 275 | const Twine &Name = "") { | ||||||
| 276 | return CreateWithCopiedFlags(Instruction::FDiv, V1, V2, FMFSource, Name); | ||||||
| 277 | } | ||||||
| 278 | static BinaryOperator *CreateFRemFMF(Value *V1, Value *V2, | ||||||
| 279 | Instruction *FMFSource, | ||||||
| 280 | const Twine &Name = "") { | ||||||
| 281 | return CreateWithCopiedFlags(Instruction::FRem, V1, V2, FMFSource, Name); | ||||||
| 282 | } | ||||||
| 283 | |||||||
| 284 | static BinaryOperator *CreateNSW(BinaryOps Opc, Value *V1, Value *V2, | ||||||
| 285 | const Twine &Name = "") { | ||||||
| 286 | BinaryOperator *BO = Create(Opc, V1, V2, Name); | ||||||
| 287 | BO->setHasNoSignedWrap(true); | ||||||
| 288 | return BO; | ||||||
| 289 | } | ||||||
| 290 | static BinaryOperator *CreateNSW(BinaryOps Opc, Value *V1, Value *V2, | ||||||
| 291 | const Twine &Name, BasicBlock *BB) { | ||||||
| 292 | BinaryOperator *BO = Create(Opc, V1, V2, Name, BB); | ||||||
| 293 | BO->setHasNoSignedWrap(true); | ||||||
| 294 | return BO; | ||||||
| 295 | } | ||||||
| 296 | static BinaryOperator *CreateNSW(BinaryOps Opc, Value *V1, Value *V2, | ||||||
| 297 | const Twine &Name, Instruction *I) { | ||||||
| 298 | BinaryOperator *BO = Create(Opc, V1, V2, Name, I); | ||||||
| 299 | BO->setHasNoSignedWrap(true); | ||||||
| 300 | return BO; | ||||||
| 301 | } | ||||||
| 302 | |||||||
| 303 | static BinaryOperator *CreateNUW(BinaryOps Opc, Value *V1, Value *V2, | ||||||
| 304 | const Twine &Name = "") { | ||||||
| 305 | BinaryOperator *BO = Create(Opc, V1, V2, Name); | ||||||
| 306 | BO->setHasNoUnsignedWrap(true); | ||||||
| 307 | return BO; | ||||||
| 308 | } | ||||||
| 309 | static BinaryOperator *CreateNUW(BinaryOps Opc, Value *V1, Value *V2, | ||||||
| 310 | const Twine &Name, BasicBlock *BB) { | ||||||
| 311 | BinaryOperator *BO = Create(Opc, V1, V2, Name, BB); | ||||||
| 312 | BO->setHasNoUnsignedWrap(true); | ||||||
| 313 | return BO; | ||||||
| 314 | } | ||||||
| 315 | static BinaryOperator *CreateNUW(BinaryOps Opc, Value *V1, Value *V2, | ||||||
| 316 | const Twine &Name, Instruction *I) { | ||||||
| 317 | BinaryOperator *BO = Create(Opc, V1, V2, Name, I); | ||||||
| 318 | BO->setHasNoUnsignedWrap(true); | ||||||
| 319 | return BO; | ||||||
| 320 | } | ||||||
| 321 | |||||||
| 322 | static BinaryOperator *CreateExact(BinaryOps Opc, Value *V1, Value *V2, | ||||||
| 323 | const Twine &Name = "") { | ||||||
| 324 | BinaryOperator *BO = Create(Opc, V1, V2, Name); | ||||||
| 325 | BO->setIsExact(true); | ||||||
| 326 | return BO; | ||||||
| 327 | } | ||||||
| 328 | static BinaryOperator *CreateExact(BinaryOps Opc, Value *V1, Value *V2, | ||||||
| 329 | const Twine &Name, BasicBlock *BB) { | ||||||
| 330 | BinaryOperator *BO = Create(Opc, V1, V2, Name, BB); | ||||||
| 331 | BO->setIsExact(true); | ||||||
| 332 | return BO; | ||||||
| 333 | } | ||||||
| 334 | static BinaryOperator *CreateExact(BinaryOps Opc, Value *V1, Value *V2, | ||||||
| 335 | const Twine &Name, Instruction *I) { | ||||||
| 336 | BinaryOperator *BO = Create(Opc, V1, V2, Name, I); | ||||||
| 337 | BO->setIsExact(true); | ||||||
| 338 | return BO; | ||||||
| 339 | } | ||||||
| 340 | |||||||
| 341 | #define DEFINE_HELPERS(OPC, NUWNSWEXACT) \ | ||||||
| 342 | static BinaryOperator *Create##NUWNSWEXACT##OPC(Value *V1, Value *V2, \ | ||||||
| 343 | const Twine &Name = "") { \ | ||||||
| 344 | return Create##NUWNSWEXACT(Instruction::OPC, V1, V2, Name); \ | ||||||
| 345 | } \ | ||||||
| 346 | static BinaryOperator *Create##NUWNSWEXACT##OPC( \ | ||||||
| 347 | Value *V1, Value *V2, const Twine &Name, BasicBlock *BB) { \ | ||||||
| 348 | return Create##NUWNSWEXACT(Instruction::OPC, V1, V2, Name, BB); \ | ||||||
| 349 | } \ | ||||||
| 350 | static BinaryOperator *Create##NUWNSWEXACT##OPC( \ | ||||||
| 351 | Value *V1, Value *V2, const Twine &Name, Instruction *I) { \ | ||||||
| 352 | return Create##NUWNSWEXACT(Instruction::OPC, V1, V2, Name, I); \ | ||||||
| 353 | } | ||||||
| 354 | |||||||
| 355 | DEFINE_HELPERS(Add, NSW) // CreateNSWAdd | ||||||
| 356 | DEFINE_HELPERS(Add, NUW) // CreateNUWAdd | ||||||
| 357 | DEFINE_HELPERS(Sub, NSW) // CreateNSWSub | ||||||
| 358 | DEFINE_HELPERS(Sub, NUW) // CreateNUWSub | ||||||
| 359 | DEFINE_HELPERS(Mul, NSW) // CreateNSWMul | ||||||
| 360 | DEFINE_HELPERS(Mul, NUW) // CreateNUWMul | ||||||
| 361 | DEFINE_HELPERS(Shl, NSW) // CreateNSWShl | ||||||
| 362 | DEFINE_HELPERS(Shl, NUW) // CreateNUWShl | ||||||
| 363 | |||||||
| 364 | DEFINE_HELPERS(SDiv, Exact) // CreateExactSDiv | ||||||
| 365 | DEFINE_HELPERS(UDiv, Exact) // CreateExactUDiv | ||||||
| 366 | DEFINE_HELPERS(AShr, Exact) // CreateExactAShr | ||||||
| 367 | DEFINE_HELPERS(LShr, Exact) // CreateExactLShr | ||||||
| 368 | |||||||
| 369 | #undef DEFINE_HELPERS | ||||||
| 370 | |||||||
| 371 | /// Helper functions to construct and inspect unary operations (NEG and NOT) | ||||||
| 372 | /// via binary operators SUB and XOR: | ||||||
| 373 | /// | ||||||
| 374 | /// Create the NEG and NOT instructions out of SUB and XOR instructions. | ||||||
| 375 | /// | ||||||
| 376 | static BinaryOperator *CreateNeg(Value *Op, const Twine &Name = "", | ||||||
| 377 | Instruction *InsertBefore = nullptr); | ||||||
| 378 | static BinaryOperator *CreateNeg(Value *Op, const Twine &Name, | ||||||
| 379 | BasicBlock *InsertAtEnd); | ||||||
| 380 | static BinaryOperator *CreateNSWNeg(Value *Op, const Twine &Name = "", | ||||||
| 381 | Instruction *InsertBefore = nullptr); | ||||||
| 382 | static BinaryOperator *CreateNSWNeg(Value *Op, const Twine &Name, | ||||||
| 383 | BasicBlock *InsertAtEnd); | ||||||
| 384 | static BinaryOperator *CreateNUWNeg(Value *Op, const Twine &Name = "", | ||||||
| 385 | Instruction *InsertBefore = nullptr); | ||||||
| 386 | static BinaryOperator *CreateNUWNeg(Value *Op, const Twine &Name, | ||||||
| 387 | BasicBlock *InsertAtEnd); | ||||||
| 388 | static BinaryOperator *CreateNot(Value *Op, const Twine &Name = "", | ||||||
| 389 | Instruction *InsertBefore = nullptr); | ||||||
| 390 | static BinaryOperator *CreateNot(Value *Op, const Twine &Name, | ||||||
| 391 | BasicBlock *InsertAtEnd); | ||||||
| 392 | |||||||
| 393 | BinaryOps getOpcode() const { | ||||||
| 394 | return static_cast<BinaryOps>(Instruction::getOpcode()); | ||||||
| 395 | } | ||||||
| 396 | |||||||
| 397 | /// Exchange the two operands to this instruction. | ||||||
| 398 | /// This instruction is safe to use on any binary instruction and | ||||||
| 399 | /// does not modify the semantics of the instruction. If the instruction | ||||||
| 400 | /// cannot be reversed (ie, it's a Div), then return true. | ||||||
| 401 | /// | ||||||
| 402 | bool swapOperands(); | ||||||
| 403 | |||||||
| 404 | // Methods for support type inquiry through isa, cast, and dyn_cast: | ||||||
| 405 | static bool classof(const Instruction *I) { | ||||||
| 406 | return I->isBinaryOp(); | ||||||
| 407 | } | ||||||
| 408 | static bool classof(const Value *V) { | ||||||
| 409 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); | ||||||
| 410 | } | ||||||
| 411 | }; | ||||||
| 412 | |||||||
| 413 | template <> | ||||||
| 414 | struct OperandTraits<BinaryOperator> : | ||||||
| 415 | public FixedNumOperandTraits<BinaryOperator, 2> { | ||||||
| 416 | }; | ||||||
| 417 | |||||||
| 418 | DEFINE_TRANSPARENT_OPERAND_ACCESSORS(BinaryOperator, Value)BinaryOperator::op_iterator BinaryOperator::op_begin() { return OperandTraits<BinaryOperator>::op_begin(this); } BinaryOperator ::const_op_iterator BinaryOperator::op_begin() const { return OperandTraits<BinaryOperator>::op_begin(const_cast< BinaryOperator*>(this)); } BinaryOperator::op_iterator BinaryOperator ::op_end() { return OperandTraits<BinaryOperator>::op_end (this); } BinaryOperator::const_op_iterator BinaryOperator::op_end () const { return OperandTraits<BinaryOperator>::op_end (const_cast<BinaryOperator*>(this)); } Value *BinaryOperator ::getOperand(unsigned i_nocapture) const { ((void)0); return cast_or_null <Value>( OperandTraits<BinaryOperator>::op_begin( const_cast<BinaryOperator*>(this))[i_nocapture].get()); } void BinaryOperator::setOperand(unsigned i_nocapture, Value *Val_nocapture) { ((void)0); OperandTraits<BinaryOperator >::op_begin(this)[i_nocapture] = Val_nocapture; } unsigned BinaryOperator::getNumOperands() const { return OperandTraits <BinaryOperator>::operands(this); } template <int Idx_nocapture > Use &BinaryOperator::Op() { return this->OpFrom< Idx_nocapture>(this); } template <int Idx_nocapture> const Use &BinaryOperator::Op() const { return this-> OpFrom<Idx_nocapture>(this); } | ||||||
| 419 | |||||||
| 420 | //===----------------------------------------------------------------------===// | ||||||
| 421 | // CastInst Class | ||||||
| 422 | //===----------------------------------------------------------------------===// | ||||||
| 423 | |||||||
| 424 | /// This is the base class for all instructions that perform data | ||||||
| 425 | /// casts. It is simply provided so that instruction category testing | ||||||
| 426 | /// can be performed with code like: | ||||||
| 427 | /// | ||||||
| 428 | /// if (isa<CastInst>(Instr)) { ... } | ||||||
| 429 | /// Base class of casting instructions. | ||||||
| 430 | class CastInst : public UnaryInstruction { | ||||||
| 431 | protected: | ||||||
| 432 | /// Constructor with insert-before-instruction semantics for subclasses | ||||||
| 433 | CastInst(Type *Ty, unsigned iType, Value *S, | ||||||
| 434 | const Twine &NameStr = "", Instruction *InsertBefore = nullptr) | ||||||
| 435 | : UnaryInstruction(Ty, iType, S, InsertBefore) { | ||||||
| 436 | setName(NameStr); | ||||||
| 437 | } | ||||||
| 438 | /// Constructor with insert-at-end-of-block semantics for subclasses | ||||||
| 439 | CastInst(Type *Ty, unsigned iType, Value *S, | ||||||
| 440 | const Twine &NameStr, BasicBlock *InsertAtEnd) | ||||||
| 441 | : UnaryInstruction(Ty, iType, S, InsertAtEnd) { | ||||||
| 442 | setName(NameStr); | ||||||
| 443 | } | ||||||
| 444 | |||||||
| 445 | public: | ||||||
| 446 | /// Provides a way to construct any of the CastInst subclasses using an | ||||||
| 447 | /// opcode instead of the subclass's constructor. The opcode must be in the | ||||||
| 448 | /// CastOps category (Instruction::isCast(opcode) returns true). This | ||||||
| 449 | /// constructor has insert-before-instruction semantics to automatically | ||||||
| 450 | /// insert the new CastInst before InsertBefore (if it is non-null). | ||||||
| 451 | /// Construct any of the CastInst subclasses | ||||||
| 452 | static CastInst *Create( | ||||||
| 453 | Instruction::CastOps, ///< The opcode of the cast instruction | ||||||
| 454 | Value *S, ///< The value to be casted (operand 0) | ||||||
| 455 | Type *Ty, ///< The type to which cast should be made | ||||||
| 456 | const Twine &Name = "", ///< Name for the instruction | ||||||
| 457 | Instruction *InsertBefore = nullptr ///< Place to insert the instruction | ||||||
| 458 | ); | ||||||
| 459 | /// Provides a way to construct any of the CastInst subclasses using an | ||||||
| 460 | /// opcode instead of the subclass's constructor. The opcode must be in the | ||||||
| 461 | /// CastOps category. This constructor has insert-at-end-of-block semantics | ||||||
| 462 | /// to automatically insert the new CastInst at the end of InsertAtEnd (if | ||||||
| 463 | /// its non-null). | ||||||
| 464 | /// Construct any of the CastInst subclasses | ||||||
| 465 | static CastInst *Create( | ||||||
| 466 | Instruction::CastOps, ///< The opcode for the cast instruction | ||||||
| 467 | Value *S, ///< The value to be casted (operand 0) | ||||||
| 468 | Type *Ty, ///< The type to which operand is casted | ||||||
| 469 | const Twine &Name, ///< The name for the instruction | ||||||
| 470 | BasicBlock *InsertAtEnd ///< The block to insert the instruction into | ||||||
| 471 | ); | ||||||
| 472 | |||||||
| 473 | /// Create a ZExt or BitCast cast instruction | ||||||
| 474 | static CastInst *CreateZExtOrBitCast( | ||||||
| 475 | Value *S, ///< The value to be casted (operand 0) | ||||||
| 476 | Type *Ty, ///< The type to which cast should be made | ||||||
| 477 | const Twine &Name = "", ///< Name for the instruction | ||||||
| 478 | Instruction *InsertBefore = nullptr ///< Place to insert the instruction | ||||||
| 479 | ); | ||||||
| 480 | |||||||
| 481 | /// Create a ZExt or BitCast cast instruction | ||||||
| 482 | static CastInst *CreateZExtOrBitCast( | ||||||
| 483 | Value *S, ///< The value to be casted (operand 0) | ||||||
| 484 | Type *Ty, ///< The type to which operand is casted | ||||||
| 485 | const Twine &Name, ///< The name for the instruction | ||||||
| 486 | BasicBlock *InsertAtEnd ///< The block to insert the instruction into | ||||||
| 487 | ); | ||||||
| 488 | |||||||
| 489 | /// Create a SExt or BitCast cast instruction | ||||||
| 490 | static CastInst *CreateSExtOrBitCast( | ||||||
| 491 | Value *S, ///< The value to be casted (operand 0) | ||||||
| 492 | Type *Ty, ///< The type to which cast should be made | ||||||
| 493 | const Twine &Name = "", ///< Name for the instruction | ||||||
| 494 | Instruction *InsertBefore = nullptr ///< Place to insert the instruction | ||||||
| 495 | ); | ||||||
| 496 | |||||||
| 497 | /// Create a SExt or BitCast cast instruction | ||||||
| 498 | static CastInst *CreateSExtOrBitCast( | ||||||
| 499 | Value *S, ///< The value to be casted (operand 0) | ||||||
| 500 | Type *Ty, ///< The type to which operand is casted | ||||||
| 501 | const Twine &Name, ///< The name for the instruction | ||||||
| 502 | BasicBlock *InsertAtEnd ///< The block to insert the instruction into | ||||||
| 503 | ); | ||||||
| 504 | |||||||
| 505 | /// Create a BitCast AddrSpaceCast, or a PtrToInt cast instruction. | ||||||
| 506 | static CastInst *CreatePointerCast( | ||||||
| 507 | Value *S, ///< The pointer value to be casted (operand 0) | ||||||
| 508 | Type *Ty, ///< The type to which operand is casted | ||||||
| 509 | const Twine &Name, ///< The name for the instruction | ||||||
| 510 | BasicBlock *InsertAtEnd ///< The block to insert the instruction into | ||||||
| 511 | ); | ||||||
| 512 | |||||||
| 513 | /// Create a BitCast, AddrSpaceCast or a PtrToInt cast instruction. | ||||||
| 514 | static CastInst *CreatePointerCast( | ||||||
| 515 | Value *S, ///< The pointer value to be casted (operand 0) | ||||||
| 516 | Type *Ty, ///< The type to which cast should be made | ||||||
| 517 | const Twine &Name = "", ///< Name for the instruction | ||||||
| 518 | Instruction *InsertBefore = nullptr ///< Place to insert the instruction | ||||||
| 519 | ); | ||||||
| 520 | |||||||
| 521 | /// Create a BitCast or an AddrSpaceCast cast instruction. | ||||||
| 522 | static CastInst *CreatePointerBitCastOrAddrSpaceCast( | ||||||
| 523 | Value *S, ///< The pointer value to be casted (operand 0) | ||||||
| 524 | Type *Ty, ///< The type to which operand is casted | ||||||
| 525 | const Twine &Name, ///< The name for the instruction | ||||||
| 526 | BasicBlock *InsertAtEnd ///< The block to insert the instruction into | ||||||
| 527 | ); | ||||||
| 528 | |||||||
| 529 | /// Create a BitCast or an AddrSpaceCast cast instruction. | ||||||
| 530 | static CastInst *CreatePointerBitCastOrAddrSpaceCast( | ||||||
| 531 | Value *S, ///< The pointer value to be casted (operand 0) | ||||||
| 532 | Type *Ty, ///< The type to which cast should be made | ||||||
| 533 | const Twine &Name = "", ///< Name for the instruction | ||||||
| 534 | Instruction *InsertBefore = nullptr ///< Place to insert the instruction | ||||||
| 535 | ); | ||||||
| 536 | |||||||
| 537 | /// Create a BitCast, a PtrToInt, or an IntToPTr cast instruction. | ||||||
| 538 | /// | ||||||
| 539 | /// If the value is a pointer type and the destination an integer type, | ||||||
| 540 | /// creates a PtrToInt cast. If the value is an integer type and the | ||||||
| 541 | /// destination a pointer type, creates an IntToPtr cast. Otherwise, creates | ||||||
| 542 | /// a bitcast. | ||||||
| 543 | static CastInst *CreateBitOrPointerCast( | ||||||
| 544 | Value *S, ///< The pointer value to be casted (operand 0) | ||||||
| 545 | Type *Ty, ///< The type to which cast should be made | ||||||
| 546 | const Twine &Name = "", ///< Name for the instruction | ||||||
| 547 | Instruction *InsertBefore = nullptr ///< Place to insert the instruction | ||||||
| 548 | ); | ||||||
| 549 | |||||||
| 550 | /// Create a ZExt, BitCast, or Trunc for int -> int casts. | ||||||
| 551 | static CastInst *CreateIntegerCast( | ||||||
| 552 | Value *S, ///< The pointer value to be casted (operand 0) | ||||||
| 553 | Type *Ty, ///< The type to which cast should be made | ||||||
| 554 | bool isSigned, ///< Whether to regard S as signed or not | ||||||
| 555 | const Twine &Name = "", ///< Name for the instruction | ||||||
| 556 | Instruction *InsertBefore = nullptr ///< Place to insert the instruction | ||||||
| 557 | ); | ||||||
| 558 | |||||||
| 559 | /// Create a ZExt, BitCast, or Trunc for int -> int casts. | ||||||
| 560 | static CastInst *CreateIntegerCast( | ||||||
| 561 | Value *S, ///< The integer value to be casted (operand 0) | ||||||
| 562 | Type *Ty, ///< The integer type to which operand is casted | ||||||
| 563 | bool isSigned, ///< Whether to regard S as signed or not | ||||||
| 564 | const Twine &Name, ///< The name for the instruction | ||||||
| 565 | BasicBlock *InsertAtEnd ///< The block to insert the instruction into | ||||||
| 566 | ); | ||||||
| 567 | |||||||
| 568 | /// Create an FPExt, BitCast, or FPTrunc for fp -> fp casts | ||||||
| 569 | static CastInst *CreateFPCast( | ||||||
| 570 | Value *S, ///< The floating point value to be casted | ||||||
| 571 | Type *Ty, ///< The floating point type to cast to | ||||||
| 572 | const Twine &Name = "", ///< Name for the instruction | ||||||
| 573 | Instruction *InsertBefore = nullptr ///< Place to insert the instruction | ||||||
| 574 | ); | ||||||
| 575 | |||||||
| 576 | /// Create an FPExt, BitCast, or FPTrunc for fp -> fp casts | ||||||
| 577 | static CastInst *CreateFPCast( | ||||||
| 578 | Value *S, ///< The floating point value to be casted | ||||||
| 579 | Type *Ty, ///< The floating point type to cast to | ||||||
| 580 | const Twine &Name, ///< The name for the instruction | ||||||
| 581 | BasicBlock *InsertAtEnd ///< The block to insert the instruction into | ||||||
| 582 | ); | ||||||
| 583 | |||||||
| 584 | /// Create a Trunc or BitCast cast instruction | ||||||
| 585 | static CastInst *CreateTruncOrBitCast( | ||||||
| 586 | Value *S, ///< The value to be casted (operand 0) | ||||||
| 587 | Type *Ty, ///< The type to which cast should be made | ||||||
| 588 | const Twine &Name = "", ///< Name for the instruction | ||||||
| 589 | Instruction *InsertBefore = nullptr ///< Place to insert the instruction | ||||||
| 590 | ); | ||||||
| 591 | |||||||
| 592 | /// Create a Trunc or BitCast cast instruction | ||||||
| 593 | static CastInst *CreateTruncOrBitCast( | ||||||
| 594 | Value *S, ///< The value to be casted (operand 0) | ||||||
| 595 | Type *Ty, ///< The type to which operand is casted | ||||||
| 596 | const Twine &Name, ///< The name for the instruction | ||||||
| 597 | BasicBlock *InsertAtEnd ///< The block to insert the instruction into | ||||||
| 598 | ); | ||||||
| 599 | |||||||
| 600 | /// Check whether a bitcast between these types is valid | ||||||
| 601 | static bool isBitCastable( | ||||||
| 602 | Type *SrcTy, ///< The Type from which the value should be cast. | ||||||
| 603 | Type *DestTy ///< The Type to which the value should be cast. | ||||||
| 604 | ); | ||||||
| 605 | |||||||
| 606 | /// Check whether a bitcast, inttoptr, or ptrtoint cast between these | ||||||
| 607 | /// types is valid and a no-op. | ||||||
| 608 | /// | ||||||
| 609 | /// This ensures that any pointer<->integer cast has enough bits in the | ||||||
| 610 | /// integer and any other cast is a bitcast. | ||||||
| 611 | static bool isBitOrNoopPointerCastable( | ||||||
| 612 | Type *SrcTy, ///< The Type from which the value should be cast. | ||||||
| 613 | Type *DestTy, ///< The Type to which the value should be cast. | ||||||
| 614 | const DataLayout &DL); | ||||||
| 615 | |||||||
| 616 | /// Returns the opcode necessary to cast Val into Ty using usual casting | ||||||
| 617 | /// rules. | ||||||
| 618 | /// Infer the opcode for cast operand and type | ||||||
| 619 | static Instruction::CastOps getCastOpcode( | ||||||
| 620 | const Value *Val, ///< The value to cast | ||||||
| 621 | bool SrcIsSigned, ///< Whether to treat the source as signed | ||||||
| 622 | Type *Ty, ///< The Type to which the value should be casted | ||||||
| 623 | bool DstIsSigned ///< Whether to treate the dest. as signed | ||||||
| 624 | ); | ||||||
| 625 | |||||||
| 626 | /// There are several places where we need to know if a cast instruction | ||||||
| 627 | /// only deals with integer source and destination types. To simplify that | ||||||
| 628 | /// logic, this method is provided. | ||||||
| 629 | /// @returns true iff the cast has only integral typed operand and dest type. | ||||||
| 630 | /// Determine if this is an integer-only cast. | ||||||
| 631 | bool isIntegerCast() const; | ||||||
| 632 | |||||||
| 633 | /// A lossless cast is one that does not alter the basic value. It implies | ||||||
| 634 | /// a no-op cast but is more stringent, preventing things like int->float, | ||||||
| 635 | /// long->double, or int->ptr. | ||||||
| 636 | /// @returns true iff the cast is lossless. | ||||||
| 637 | /// Determine if this is a lossless cast. | ||||||
| 638 | bool isLosslessCast() const; | ||||||
| 639 | |||||||
| 640 | /// A no-op cast is one that can be effected without changing any bits. | ||||||
| 641 | /// It implies that the source and destination types are the same size. The | ||||||
| 642 | /// DataLayout argument is to determine the pointer size when examining casts | ||||||
| 643 | /// involving Integer and Pointer types. They are no-op casts if the integer | ||||||
| 644 | /// is the same size as the pointer. However, pointer size varies with | ||||||
| 645 | /// platform. Note that a precondition of this method is that the cast is | ||||||
| 646 | /// legal - i.e. the instruction formed with these operands would verify. | ||||||
| 647 | static bool isNoopCast( | ||||||
| 648 | Instruction::CastOps Opcode, ///< Opcode of cast | ||||||
| 649 | Type *SrcTy, ///< SrcTy of cast | ||||||
| 650 | Type *DstTy, ///< DstTy of cast | ||||||
| 651 | const DataLayout &DL ///< DataLayout to get the Int Ptr type from. | ||||||
| 652 | ); | ||||||
| 653 | |||||||
| 654 | /// Determine if this cast is a no-op cast. | ||||||
| 655 | /// | ||||||
| 656 | /// \param DL is the DataLayout to determine pointer size. | ||||||
| 657 | bool isNoopCast(const DataLayout &DL) const; | ||||||
| 658 | |||||||
| 659 | /// Determine how a pair of casts can be eliminated, if they can be at all. | ||||||
| 660 | /// This is a helper function for both CastInst and ConstantExpr. | ||||||
| 661 | /// @returns 0 if the CastInst pair can't be eliminated, otherwise | ||||||
| 662 | /// returns Instruction::CastOps value for a cast that can replace | ||||||
| 663 | /// the pair, casting SrcTy to DstTy. | ||||||
| 664 | /// Determine if a cast pair is eliminable | ||||||
| 665 | static unsigned isEliminableCastPair( | ||||||
| 666 | Instruction::CastOps firstOpcode, ///< Opcode of first cast | ||||||
| 667 | Instruction::CastOps secondOpcode, ///< Opcode of second cast | ||||||
| 668 | Type *SrcTy, ///< SrcTy of 1st cast | ||||||
| 669 | Type *MidTy, ///< DstTy of 1st cast & SrcTy of 2nd cast | ||||||
| 670 | Type *DstTy, ///< DstTy of 2nd cast | ||||||
| 671 | Type *SrcIntPtrTy, ///< Integer type corresponding to Ptr SrcTy, or null | ||||||
| 672 | Type *MidIntPtrTy, ///< Integer type corresponding to Ptr MidTy, or null | ||||||
| 673 | Type *DstIntPtrTy ///< Integer type corresponding to Ptr DstTy, or null | ||||||
| 674 | ); | ||||||
| 675 | |||||||
| 676 | /// Return the opcode of this CastInst | ||||||
| 677 | Instruction::CastOps getOpcode() const { | ||||||
| 678 | return Instruction::CastOps(Instruction::getOpcode()); | ||||||
| 679 | } | ||||||
| 680 | |||||||
| 681 | /// Return the source type, as a convenience | ||||||
| 682 | Type* getSrcTy() const { return getOperand(0)->getType(); } | ||||||
| 683 | /// Return the destination type, as a convenience | ||||||
| 684 | Type* getDestTy() const { return getType(); } | ||||||
| 685 | |||||||
| 686 | /// This method can be used to determine if a cast from SrcTy to DstTy using | ||||||
| 687 | /// Opcode op is valid or not. | ||||||
| 688 | /// @returns true iff the proposed cast is valid. | ||||||
| 689 | /// Determine if a cast is valid without creating one. | ||||||
| 690 | static bool castIsValid(Instruction::CastOps op, Type *SrcTy, Type *DstTy); | ||||||
| 691 | static bool castIsValid(Instruction::CastOps op, Value *S, Type *DstTy) { | ||||||
| 692 | return castIsValid(op, S->getType(), DstTy); | ||||||
| 693 | } | ||||||
| 694 | |||||||
| 695 | /// Methods for support type inquiry through isa, cast, and dyn_cast: | ||||||
| 696 | static bool classof(const Instruction *I) { | ||||||
| 697 | return I->isCast(); | ||||||
| 698 | } | ||||||
| 699 | static bool classof(const Value *V) { | ||||||
| 700 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); | ||||||
| 701 | } | ||||||
| 702 | }; | ||||||
| 703 | |||||||
| 704 | //===----------------------------------------------------------------------===// | ||||||
| 705 | // CmpInst Class | ||||||
| 706 | //===----------------------------------------------------------------------===// | ||||||
| 707 | |||||||
| 708 | /// This class is the base class for the comparison instructions. | ||||||
| 709 | /// Abstract base class of comparison instructions. | ||||||
| 710 | class CmpInst : public Instruction { | ||||||
| 711 | public: | ||||||
| 712 | /// This enumeration lists the possible predicates for CmpInst subclasses. | ||||||
| 713 | /// Values in the range 0-31 are reserved for FCmpInst, while values in the | ||||||
| 714 | /// range 32-64 are reserved for ICmpInst. This is necessary to ensure the | ||||||
| 715 | /// predicate values are not overlapping between the classes. | ||||||
| 716 | /// | ||||||
| 717 | /// Some passes (e.g. InstCombine) depend on the bit-wise characteristics of | ||||||
| 718 | /// FCMP_* values. Changing the bit patterns requires a potential change to | ||||||
| 719 | /// those passes. | ||||||
| 720 | enum Predicate : unsigned { | ||||||
| 721 | // Opcode U L G E Intuitive operation | ||||||
| 722 | FCMP_FALSE = 0, ///< 0 0 0 0 Always false (always folded) | ||||||
| 723 | FCMP_OEQ = 1, ///< 0 0 0 1 True if ordered and equal | ||||||
| 724 | FCMP_OGT = 2, ///< 0 0 1 0 True if ordered and greater than | ||||||
| 725 | FCMP_OGE = 3, ///< 0 0 1 1 True if ordered and greater than or equal | ||||||
| 726 | FCMP_OLT = 4, ///< 0 1 0 0 True if ordered and less than | ||||||
| 727 | FCMP_OLE = 5, ///< 0 1 0 1 True if ordered and less than or equal | ||||||
| 728 | FCMP_ONE = 6, ///< 0 1 1 0 True if ordered and operands are unequal | ||||||
| 729 | FCMP_ORD = 7, ///< 0 1 1 1 True if ordered (no nans) | ||||||
| 730 | FCMP_UNO = 8, ///< 1 0 0 0 True if unordered: isnan(X) | isnan(Y) | ||||||
| 731 | FCMP_UEQ = 9, ///< 1 0 0 1 True if unordered or equal | ||||||
| 732 | FCMP_UGT = 10, ///< 1 0 1 0 True if unordered or greater than | ||||||
| 733 | FCMP_UGE = 11, ///< 1 0 1 1 True if unordered, greater than, or equal | ||||||
| 734 | FCMP_ULT = 12, ///< 1 1 0 0 True if unordered or less than | ||||||
| 735 | FCMP_ULE = 13, ///< 1 1 0 1 True if unordered, less than, or equal | ||||||
| 736 | FCMP_UNE = 14, ///< 1 1 1 0 True if unordered or not equal | ||||||
| 737 | FCMP_TRUE = 15, ///< 1 1 1 1 Always true (always folded) | ||||||
| 738 | FIRST_FCMP_PREDICATE = FCMP_FALSE, | ||||||
| 739 | LAST_FCMP_PREDICATE = FCMP_TRUE, | ||||||
| 740 | BAD_FCMP_PREDICATE = FCMP_TRUE + 1, | ||||||
| 741 | ICMP_EQ = 32, ///< equal | ||||||
| 742 | ICMP_NE = 33, ///< not equal | ||||||
| 743 | ICMP_UGT = 34, ///< unsigned greater than | ||||||
| 744 | ICMP_UGE = 35, ///< unsigned greater or equal | ||||||
| 745 | ICMP_ULT = 36, ///< unsigned less than | ||||||
| 746 | ICMP_ULE = 37, ///< unsigned less or equal | ||||||
| 747 | ICMP_SGT = 38, ///< signed greater than | ||||||
| 748 | ICMP_SGE = 39, ///< signed greater or equal | ||||||
| 749 | ICMP_SLT = 40, ///< signed less than | ||||||
| 750 | ICMP_SLE = 41, ///< signed less or equal | ||||||
| 751 | FIRST_ICMP_PREDICATE = ICMP_EQ, | ||||||
| 752 | LAST_ICMP_PREDICATE = ICMP_SLE, | ||||||
| 753 | BAD_ICMP_PREDICATE = ICMP_SLE + 1 | ||||||
| 754 | }; | ||||||
| 755 | using PredicateField = | ||||||
| 756 | Bitfield::Element<Predicate, 0, 6, LAST_ICMP_PREDICATE>; | ||||||
| 757 | |||||||
| 758 | protected: | ||||||
| 759 | CmpInst(Type *ty, Instruction::OtherOps op, Predicate pred, | ||||||
| 760 | Value *LHS, Value *RHS, const Twine &Name = "", | ||||||
| 761 | Instruction *InsertBefore = nullptr, | ||||||
| 762 | Instruction *FlagsSource = nullptr); | ||||||
| 763 | |||||||
| 764 | CmpInst(Type *ty, Instruction::OtherOps op, Predicate pred, | ||||||
| 765 | Value *LHS, Value *RHS, const Twine &Name, | ||||||
| 766 | BasicBlock *InsertAtEnd); | ||||||
| 767 | |||||||
| 768 | public: | ||||||
| 769 | // allocate space for exactly two operands | ||||||
| 770 | void *operator new(size_t S) { return User::operator new(S, 2); } | ||||||
| 771 | void operator delete(void *Ptr) { User::operator delete(Ptr); } | ||||||
| 772 | |||||||
| 773 | /// Construct a compare instruction, given the opcode, the predicate and | ||||||
| 774 | /// the two operands. Optionally (if InstBefore is specified) insert the | ||||||
| 775 | /// instruction into a BasicBlock right before the specified instruction. | ||||||
| 776 | /// The specified Instruction is allowed to be a dereferenced end iterator. | ||||||
| 777 | /// Create a CmpInst | ||||||
| 778 | static CmpInst *Create(OtherOps Op, | ||||||
| 779 | Predicate predicate, Value *S1, | ||||||
| 780 | Value *S2, const Twine &Name = "", | ||||||
| 781 | Instruction *InsertBefore = nullptr); | ||||||
| 782 | |||||||
| 783 | /// Construct a compare instruction, given the opcode, the predicate and the | ||||||
| 784 | /// two operands. Also automatically insert this instruction to the end of | ||||||
| 785 | /// the BasicBlock specified. | ||||||
| 786 | /// Create a CmpInst | ||||||
| 787 | static CmpInst *Create(OtherOps Op, Predicate predicate, Value *S1, | ||||||
| 788 | Value *S2, const Twine &Name, BasicBlock *InsertAtEnd); | ||||||
| 789 | |||||||
| 790 | /// Get the opcode casted to the right type | ||||||
| 791 | OtherOps getOpcode() const { | ||||||
| 792 | return static_cast<OtherOps>(Instruction::getOpcode()); | ||||||
| 793 | } | ||||||
| 794 | |||||||
| 795 | /// Return the predicate for this instruction. | ||||||
| 796 | Predicate getPredicate() const { return getSubclassData<PredicateField>(); } | ||||||
| 797 | |||||||
| 798 | /// Set the predicate for this instruction to the specified value. | ||||||
| 799 | void setPredicate(Predicate P) { setSubclassData<PredicateField>(P); } | ||||||
| 800 | |||||||
| 801 | static bool isFPPredicate(Predicate P) { | ||||||
| 802 | static_assert(FIRST_FCMP_PREDICATE == 0, | ||||||
| 803 | "FIRST_FCMP_PREDICATE is required to be 0"); | ||||||
| 804 | return P
| ||||||
| 805 | } | ||||||
| 806 | |||||||
| 807 | static bool isIntPredicate(Predicate P) { | ||||||
| 808 | return P >= FIRST_ICMP_PREDICATE && P <= LAST_ICMP_PREDICATE; | ||||||
| 809 | } | ||||||
| 810 | |||||||
| 811 | static StringRef getPredicateName(Predicate P); | ||||||
| 812 | |||||||
| 813 | bool isFPPredicate() const { return isFPPredicate(getPredicate()); } | ||||||
| 814 | bool isIntPredicate() const { return isIntPredicate(getPredicate()); } | ||||||
| 815 | |||||||
| 816 | /// For example, EQ -> NE, UGT -> ULE, SLT -> SGE, | ||||||
| 817 | /// OEQ -> UNE, UGT -> OLE, OLT -> UGE, etc. | ||||||
| 818 | /// @returns the inverse predicate for the instruction's current predicate. | ||||||
| 819 | /// Return the inverse of the instruction's predicate. | ||||||
| 820 | Predicate getInversePredicate() const { | ||||||
| 821 | return getInversePredicate(getPredicate()); | ||||||
| 822 | } | ||||||
| 823 | |||||||
| 824 | /// For example, EQ -> NE, UGT -> ULE, SLT -> SGE, | ||||||
| 825 | /// OEQ -> UNE, UGT -> OLE, OLT -> UGE, etc. | ||||||
| 826 | /// @returns the inverse predicate for predicate provided in \p pred. | ||||||
| 827 | /// Return the inverse of a given predicate | ||||||
| 828 | static Predicate getInversePredicate(Predicate pred); | ||||||
| 829 | |||||||
| 830 | /// For example, EQ->EQ, SLE->SGE, ULT->UGT, | ||||||
| 831 | /// OEQ->OEQ, ULE->UGE, OLT->OGT, etc. | ||||||
| 832 | /// @returns the predicate that would be the result of exchanging the two | ||||||
| 833 | /// operands of the CmpInst instruction without changing the result | ||||||
| 834 | /// produced. | ||||||
| 835 | /// Return the predicate as if the operands were swapped | ||||||
| 836 | Predicate getSwappedPredicate() const { | ||||||
| 837 | return getSwappedPredicate(getPredicate()); | ||||||
| 838 | } | ||||||
| 839 | |||||||
| 840 | /// This is a static version that you can use without an instruction | ||||||
| 841 | /// available. | ||||||
| 842 | /// Return the predicate as if the operands were swapped. | ||||||
| 843 | static Predicate getSwappedPredicate(Predicate pred); | ||||||
| 844 | |||||||
| 845 | /// This is a static version that you can use without an instruction | ||||||
| 846 | /// available. | ||||||
| 847 | /// @returns true if the comparison predicate is strict, false otherwise. | ||||||
| 848 | static bool isStrictPredicate(Predicate predicate); | ||||||
| 849 | |||||||
| 850 | /// @returns true if the comparison predicate is strict, false otherwise. | ||||||
| 851 | /// Determine if this instruction is using an strict comparison predicate. | ||||||
| 852 | bool isStrictPredicate() const { return isStrictPredicate(getPredicate()); } | ||||||
| 853 | |||||||
| 854 | /// This is a static version that you can use without an instruction | ||||||
| 855 | /// available. | ||||||
| 856 | /// @returns true if the comparison predicate is non-strict, false otherwise. | ||||||
| 857 | static bool isNonStrictPredicate(Predicate predicate); | ||||||
| 858 | |||||||
| 859 | /// @returns true if the comparison predicate is non-strict, false otherwise. | ||||||
| 860 | /// Determine if this instruction is using an non-strict comparison predicate. | ||||||
| 861 | bool isNonStrictPredicate() const { | ||||||
| 862 | return isNonStrictPredicate(getPredicate()); | ||||||
| 863 | } | ||||||
| 864 | |||||||
| 865 | /// For example, SGE -> SGT, SLE -> SLT, ULE -> ULT, UGE -> UGT. | ||||||
| 866 | /// Returns the strict version of non-strict comparisons. | ||||||
| 867 | Predicate getStrictPredicate() const { | ||||||
| 868 | return getStrictPredicate(getPredicate()); | ||||||
| 869 | } | ||||||
| 870 | |||||||
| 871 | /// This is a static version that you can use without an instruction | ||||||
| 872 | /// available. | ||||||
| 873 | /// @returns the strict version of comparison provided in \p pred. | ||||||
| 874 | /// If \p pred is not a strict comparison predicate, returns \p pred. | ||||||
| 875 | /// Returns the strict version of non-strict comparisons. | ||||||
| 876 | static Predicate getStrictPredicate(Predicate pred); | ||||||
| 877 | |||||||
| 878 | /// For example, SGT -> SGE, SLT -> SLE, ULT -> ULE, UGT -> UGE. | ||||||
| 879 | /// Returns the non-strict version of strict comparisons. | ||||||
| 880 | Predicate getNonStrictPredicate() const { | ||||||
| 881 | return getNonStrictPredicate(getPredicate()); | ||||||
| 882 | } | ||||||
| 883 | |||||||
| 884 | /// This is a static version that you can use without an instruction | ||||||
| 885 | /// available. | ||||||
| 886 | /// @returns the non-strict version of comparison provided in \p pred. | ||||||
| 887 | /// If \p pred is not a strict comparison predicate, returns \p pred. | ||||||
| 888 | /// Returns the non-strict version of strict comparisons. | ||||||
| 889 | static Predicate getNonStrictPredicate(Predicate pred); | ||||||
| 890 | |||||||
| 891 | /// This is a static version that you can use without an instruction | ||||||
| 892 | /// available. | ||||||
| 893 | /// Return the flipped strictness of predicate | ||||||
| 894 | static Predicate getFlippedStrictnessPredicate(Predicate pred); | ||||||
| 895 | |||||||
| 896 | /// For predicate of kind "is X or equal to 0" returns the predicate "is X". | ||||||
| 897 | /// For predicate of kind "is X" returns the predicate "is X or equal to 0". | ||||||
| 898 | /// does not support other kind of predicates. | ||||||
| 899 | /// @returns the predicate that does not contains is equal to zero if | ||||||
| 900 | /// it had and vice versa. | ||||||
| 901 | /// Return the flipped strictness of predicate | ||||||
| 902 | Predicate getFlippedStrictnessPredicate() const { | ||||||
| 903 | return getFlippedStrictnessPredicate(getPredicate()); | ||||||
| 904 | } | ||||||
| 905 | |||||||
| 906 | /// Provide more efficient getOperand methods. | ||||||
| 907 | DECLARE_TRANSPARENT_OPERAND_ACCESSORS(Value)public: inline Value *getOperand(unsigned) const; inline void setOperand(unsigned, Value*); inline op_iterator op_begin(); inline const_op_iterator op_begin() const; inline op_iterator op_end(); inline const_op_iterator op_end() const; protected : template <int> inline Use &Op(); template <int > inline const Use &Op() const; public: inline unsigned getNumOperands() const; | ||||||
| 908 | |||||||
| 909 | /// This is just a convenience that dispatches to the subclasses. | ||||||
| 910 | /// Swap the operands and adjust predicate accordingly to retain | ||||||
| 911 | /// the same comparison. | ||||||
| 912 | void swapOperands(); | ||||||
| 913 | |||||||
| 914 | /// This is just a convenience that dispatches to the subclasses. | ||||||
| 915 | /// Determine if this CmpInst is commutative. | ||||||
| 916 | bool isCommutative() const; | ||||||
| 917 | |||||||
| 918 | /// Determine if this is an equals/not equals predicate. | ||||||
| 919 | /// This is a static version that you can use without an instruction | ||||||
| 920 | /// available. | ||||||
| 921 | static bool isEquality(Predicate pred); | ||||||
| 922 | |||||||
| 923 | /// Determine if this is an equals/not equals predicate. | ||||||
| 924 | bool isEquality() const { return isEquality(getPredicate()); } | ||||||
| 925 | |||||||
| 926 | /// Return true if the predicate is relational (not EQ or NE). | ||||||
| 927 | static bool isRelational(Predicate P) { return !isEquality(P); } | ||||||
| 928 | |||||||
| 929 | /// Return true if the predicate is relational (not EQ or NE). | ||||||
| 930 | bool isRelational() const { return !isEquality(); } | ||||||
| 931 | |||||||
| 932 | /// @returns true if the comparison is signed, false otherwise. | ||||||
| 933 | /// Determine if this instruction is using a signed comparison. | ||||||
| 934 | bool isSigned() const { | ||||||
| 935 | return isSigned(getPredicate()); | ||||||
| 936 | } | ||||||
| 937 | |||||||
| 938 | /// @returns true if the comparison is unsigned, false otherwise. | ||||||
| 939 | /// Determine if this instruction is using an unsigned comparison. | ||||||
| 940 | bool isUnsigned() const { | ||||||
| 941 | return isUnsigned(getPredicate()); | ||||||
| 942 | } | ||||||
| 943 | |||||||
| 944 | /// For example, ULT->SLT, ULE->SLE, UGT->SGT, UGE->SGE, SLT->Failed assert | ||||||
| 945 | /// @returns the signed version of the unsigned predicate pred. | ||||||
| 946 | /// return the signed version of a predicate | ||||||
| 947 | static Predicate getSignedPredicate(Predicate pred); | ||||||
| 948 | |||||||
| 949 | /// For example, ULT->SLT, ULE->SLE, UGT->SGT, UGE->SGE, SLT->Failed assert | ||||||
| 950 | /// @returns the signed version of the predicate for this instruction (which | ||||||
| 951 | /// has to be an unsigned predicate). | ||||||
| 952 | /// return the signed version of a predicate | ||||||
| 953 | Predicate getSignedPredicate() { | ||||||
| 954 | return getSignedPredicate(getPredicate()); | ||||||
| 955 | } | ||||||
| 956 | |||||||
| 957 | /// For example, SLT->ULT, SLE->ULE, SGT->UGT, SGE->UGE, ULT->Failed assert | ||||||
| 958 | /// @returns the unsigned version of the signed predicate pred. | ||||||
| 959 | static Predicate getUnsignedPredicate(Predicate pred); | ||||||
| 960 | |||||||
| 961 | /// For example, SLT->ULT, SLE->ULE, SGT->UGT, SGE->UGE, ULT->Failed assert | ||||||
| 962 | /// @returns the unsigned version of the predicate for this instruction (which | ||||||
| 963 | /// has to be an signed predicate). | ||||||
| 964 | /// return the unsigned version of a predicate | ||||||
| 965 | Predicate getUnsignedPredicate() { | ||||||
| 966 | return getUnsignedPredicate(getPredicate()); | ||||||
| 967 | } | ||||||
| 968 | |||||||
| 969 | /// For example, SLT->ULT, ULT->SLT, SLE->ULE, ULE->SLE, EQ->Failed assert | ||||||
| 970 | /// @returns the unsigned version of the signed predicate pred or | ||||||
| 971 | /// the signed version of the signed predicate pred. | ||||||
| 972 | static Predicate getFlippedSignednessPredicate(Predicate pred); | ||||||
| 973 | |||||||
| 974 | /// For example, SLT->ULT, ULT->SLT, SLE->ULE, ULE->SLE, EQ->Failed assert | ||||||
| 975 | /// @returns the unsigned version of the signed predicate pred or | ||||||
| 976 | /// the signed version of the signed predicate pred. | ||||||
| 977 | Predicate getFlippedSignednessPredicate() { | ||||||
| 978 | return getFlippedSignednessPredicate(getPredicate()); | ||||||
| 979 | } | ||||||
| 980 | |||||||
| 981 | /// This is just a convenience. | ||||||
| 982 | /// Determine if this is true when both operands are the same. | ||||||
| 983 | bool isTrueWhenEqual() const { | ||||||
| 984 | return isTrueWhenEqual(getPredicate()); | ||||||
| 985 | } | ||||||
| 986 | |||||||
| 987 | /// This is just a convenience. | ||||||
| 988 | /// Determine if this is false when both operands are the same. | ||||||
| 989 | bool isFalseWhenEqual() const { | ||||||
| 990 | return isFalseWhenEqual(getPredicate()); | ||||||
| 991 | } | ||||||
| 992 | |||||||
| 993 | /// @returns true if the predicate is unsigned, false otherwise. | ||||||
| 994 | /// Determine if the predicate is an unsigned operation. | ||||||
| 995 | static bool isUnsigned(Predicate predicate); | ||||||
| 996 | |||||||
| 997 | /// @returns true if the predicate is signed, false otherwise. | ||||||
| 998 | /// Determine if the predicate is an signed operation. | ||||||
| 999 | static bool isSigned(Predicate predicate); | ||||||
| 1000 | |||||||
| 1001 | /// Determine if the predicate is an ordered operation. | ||||||
| 1002 | static bool isOrdered(Predicate predicate); | ||||||
| 1003 | |||||||
| 1004 | /// Determine if the predicate is an unordered operation. | ||||||
| 1005 | static bool isUnordered(Predicate predicate); | ||||||
| 1006 | |||||||
| 1007 | /// Determine if the predicate is true when comparing a value with itself. | ||||||
| 1008 | static bool isTrueWhenEqual(Predicate predicate); | ||||||
| 1009 | |||||||
| 1010 | /// Determine if the predicate is false when comparing a value with itself. | ||||||
| 1011 | static bool isFalseWhenEqual(Predicate predicate); | ||||||
| 1012 | |||||||
| 1013 | /// Determine if Pred1 implies Pred2 is true when two compares have matching | ||||||
| 1014 | /// operands. | ||||||
| 1015 | static bool isImpliedTrueByMatchingCmp(Predicate Pred1, Predicate Pred2); | ||||||
| 1016 | |||||||
| 1017 | /// Determine if Pred1 implies Pred2 is false when two compares have matching | ||||||
| 1018 | /// operands. | ||||||
| 1019 | static bool isImpliedFalseByMatchingCmp(Predicate Pred1, Predicate Pred2); | ||||||
| 1020 | |||||||
| 1021 | /// Methods for support type inquiry through isa, cast, and dyn_cast: | ||||||
| 1022 | static bool classof(const Instruction *I) { | ||||||
| 1023 | return I->getOpcode() == Instruction::ICmp || | ||||||
| 1024 | I->getOpcode() == Instruction::FCmp; | ||||||
| 1025 | } | ||||||
| 1026 | static bool classof(const Value *V) { | ||||||
| 1027 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); | ||||||
| 1028 | } | ||||||
| 1029 | |||||||
| 1030 | /// Create a result type for fcmp/icmp | ||||||
| 1031 | static Type* makeCmpResultType(Type* opnd_type) { | ||||||
| 1032 | if (VectorType* vt = dyn_cast<VectorType>(opnd_type)) { | ||||||
| 1033 | return VectorType::get(Type::getInt1Ty(opnd_type->getContext()), | ||||||
| 1034 | vt->getElementCount()); | ||||||
| 1035 | } | ||||||
| 1036 | return Type::getInt1Ty(opnd_type->getContext()); | ||||||
| 1037 | } | ||||||
| 1038 | |||||||
| 1039 | private: | ||||||
| 1040 | // Shadow Value::setValueSubclassData with a private forwarding method so that | ||||||
| 1041 | // subclasses cannot accidentally use it. | ||||||
| 1042 | void setValueSubclassData(unsigned short D) { | ||||||
| 1043 | Value::setValueSubclassData(D); | ||||||
| 1044 | } | ||||||
| 1045 | }; | ||||||
| 1046 | |||||||
| 1047 | // FIXME: these are redundant if CmpInst < BinaryOperator | ||||||
| 1048 | template <> | ||||||
| 1049 | struct OperandTraits<CmpInst> : public FixedNumOperandTraits<CmpInst, 2> { | ||||||
| 1050 | }; | ||||||
| 1051 | |||||||
| 1052 | DEFINE_TRANSPARENT_OPERAND_ACCESSORS(CmpInst, Value)CmpInst::op_iterator CmpInst::op_begin() { return OperandTraits <CmpInst>::op_begin(this); } CmpInst::const_op_iterator CmpInst::op_begin() const { return OperandTraits<CmpInst> ::op_begin(const_cast<CmpInst*>(this)); } CmpInst::op_iterator CmpInst::op_end() { return OperandTraits<CmpInst>::op_end (this); } CmpInst::const_op_iterator CmpInst::op_end() const { return OperandTraits<CmpInst>::op_end(const_cast<CmpInst *>(this)); } Value *CmpInst::getOperand(unsigned i_nocapture ) const { ((void)0); return cast_or_null<Value>( OperandTraits <CmpInst>::op_begin(const_cast<CmpInst*>(this))[i_nocapture ].get()); } void CmpInst::setOperand(unsigned i_nocapture, Value *Val_nocapture) { ((void)0); OperandTraits<CmpInst>::op_begin (this)[i_nocapture] = Val_nocapture; } unsigned CmpInst::getNumOperands () const { return OperandTraits<CmpInst>::operands(this ); } template <int Idx_nocapture> Use &CmpInst::Op( ) { return this->OpFrom<Idx_nocapture>(this); } template <int Idx_nocapture> const Use &CmpInst::Op() const { return this->OpFrom<Idx_nocapture>(this); } | ||||||
| 1053 | |||||||
| 1054 | /// A lightweight accessor for an operand bundle meant to be passed | ||||||
| 1055 | /// around by value. | ||||||
| 1056 | struct OperandBundleUse { | ||||||
| 1057 | ArrayRef<Use> Inputs; | ||||||
| 1058 | |||||||
| 1059 | OperandBundleUse() = default; | ||||||
| 1060 | explicit OperandBundleUse(StringMapEntry<uint32_t> *Tag, ArrayRef<Use> Inputs) | ||||||
| 1061 | : Inputs(Inputs), Tag(Tag) {} | ||||||
| 1062 | |||||||
| 1063 | /// Return true if the operand at index \p Idx in this operand bundle | ||||||
| 1064 | /// has the attribute A. | ||||||
| 1065 | bool operandHasAttr(unsigned Idx, Attribute::AttrKind A) const { | ||||||
| 1066 | if (isDeoptOperandBundle()) | ||||||
| 1067 | if (A == Attribute::ReadOnly || A == Attribute::NoCapture) | ||||||
| 1068 | return Inputs[Idx]->getType()->isPointerTy(); | ||||||
| 1069 | |||||||
| 1070 | // Conservative answer: no operands have any attributes. | ||||||
| 1071 | return false; | ||||||
| 1072 | } | ||||||
| 1073 | |||||||
| 1074 | /// Return the tag of this operand bundle as a string. | ||||||
| 1075 | StringRef getTagName() const { | ||||||
| 1076 | return Tag->getKey(); | ||||||
| 1077 | } | ||||||
| 1078 | |||||||
| 1079 | /// Return the tag of this operand bundle as an integer. | ||||||
| 1080 | /// | ||||||
| 1081 | /// Operand bundle tags are interned by LLVMContextImpl::getOrInsertBundleTag, | ||||||
| 1082 | /// and this function returns the unique integer getOrInsertBundleTag | ||||||
| 1083 | /// associated the tag of this operand bundle to. | ||||||
| 1084 | uint32_t getTagID() const { | ||||||
| 1085 | return Tag->getValue(); | ||||||
| 1086 | } | ||||||
| 1087 | |||||||
| 1088 | /// Return true if this is a "deopt" operand bundle. | ||||||
| 1089 | bool isDeoptOperandBundle() const { | ||||||
| 1090 | return getTagID() == LLVMContext::OB_deopt; | ||||||
| 1091 | } | ||||||
| 1092 | |||||||
| 1093 | /// Return true if this is a "funclet" operand bundle. | ||||||
| 1094 | bool isFuncletOperandBundle() const { | ||||||
| 1095 | return getTagID() == LLVMContext::OB_funclet; | ||||||
| 1096 | } | ||||||
| 1097 | |||||||
| 1098 | /// Return true if this is a "cfguardtarget" operand bundle. | ||||||
| 1099 | bool isCFGuardTargetOperandBundle() const { | ||||||
| 1100 | return getTagID() == LLVMContext::OB_cfguardtarget; | ||||||
| 1101 | } | ||||||
| 1102 | |||||||
| 1103 | private: | ||||||
| 1104 | /// Pointer to an entry in LLVMContextImpl::getOrInsertBundleTag. | ||||||
| 1105 | StringMapEntry<uint32_t> *Tag; | ||||||
| 1106 | }; | ||||||
| 1107 | |||||||
| 1108 | /// A container for an operand bundle being viewed as a set of values | ||||||
| 1109 | /// rather than a set of uses. | ||||||
| 1110 | /// | ||||||
| 1111 | /// Unlike OperandBundleUse, OperandBundleDefT owns the memory it carries, and | ||||||
| 1112 | /// so it is possible to create and pass around "self-contained" instances of | ||||||
| 1113 | /// OperandBundleDef and ConstOperandBundleDef. | ||||||
| 1114 | template <typename InputTy> class OperandBundleDefT { | ||||||
| 1115 | std::string Tag; | ||||||
| 1116 | std::vector<InputTy> Inputs; | ||||||
| 1117 | |||||||
| 1118 | public: | ||||||
| 1119 | explicit OperandBundleDefT(std::string Tag, std::vector<InputTy> Inputs) | ||||||
| 1120 | : Tag(std::move(Tag)), Inputs(std::move(Inputs)) {} | ||||||
| 1121 | explicit OperandBundleDefT(std::string Tag, ArrayRef<InputTy> Inputs) | ||||||
| 1122 | : Tag(std::move(Tag)), Inputs(Inputs) {} | ||||||
| 1123 | |||||||
| 1124 | explicit OperandBundleDefT(const OperandBundleUse &OBU) { | ||||||
| 1125 | Tag = std::string(OBU.getTagName()); | ||||||
| 1126 | llvm::append_range(Inputs, OBU.Inputs); | ||||||
| 1127 | } | ||||||
| 1128 | |||||||
| 1129 | ArrayRef<InputTy> inputs() const { return Inputs; } | ||||||
| 1130 | |||||||
| 1131 | using input_iterator = typename std::vector<InputTy>::const_iterator; | ||||||
| 1132 | |||||||
| 1133 | size_t input_size() const { return Inputs.size(); } | ||||||
| 1134 | input_iterator input_begin() const { return Inputs.begin(); } | ||||||
| 1135 | input_iterator input_end() const { return Inputs.end(); } | ||||||
| 1136 | |||||||
| 1137 | StringRef getTag() const { return Tag; } | ||||||
| 1138 | }; | ||||||
| 1139 | |||||||
| 1140 | using OperandBundleDef = OperandBundleDefT<Value *>; | ||||||
| 1141 | using ConstOperandBundleDef = OperandBundleDefT<const Value *>; | ||||||
| 1142 | |||||||
| 1143 | //===----------------------------------------------------------------------===// | ||||||
| 1144 | // CallBase Class | ||||||
| 1145 | //===----------------------------------------------------------------------===// | ||||||
| 1146 | |||||||
| 1147 | /// Base class for all callable instructions (InvokeInst and CallInst) | ||||||
| 1148 | /// Holds everything related to calling a function. | ||||||
| 1149 | /// | ||||||
| 1150 | /// All call-like instructions are required to use a common operand layout: | ||||||
| 1151 | /// - Zero or more arguments to the call, | ||||||
| 1152 | /// - Zero or more operand bundles with zero or more operand inputs each | ||||||
| 1153 | /// bundle, | ||||||
| 1154 | /// - Zero or more subclass controlled operands | ||||||
| 1155 | /// - The called function. | ||||||
| 1156 | /// | ||||||
| 1157 | /// This allows this base class to easily access the called function and the | ||||||
| 1158 | /// start of the arguments without knowing how many other operands a particular | ||||||
| 1159 | /// subclass requires. Note that accessing the end of the argument list isn't | ||||||
| 1160 | /// as cheap as most other operations on the base class. | ||||||
| 1161 | class CallBase : public Instruction { | ||||||
| 1162 | protected: | ||||||
| 1163 | // The first two bits are reserved by CallInst for fast retrieval, | ||||||
| 1164 | using CallInstReservedField = Bitfield::Element<unsigned, 0, 2>; | ||||||
| 1165 | using CallingConvField = | ||||||
| 1166 | Bitfield::Element<CallingConv::ID, CallInstReservedField::NextBit, 10, | ||||||
| 1167 | CallingConv::MaxID>; | ||||||
| 1168 | static_assert( | ||||||
| 1169 | Bitfield::areContiguous<CallInstReservedField, CallingConvField>(), | ||||||
| 1170 | "Bitfields must be contiguous"); | ||||||
| 1171 | |||||||
| 1172 | /// The last operand is the called operand. | ||||||
| 1173 | static constexpr int CalledOperandOpEndIdx = -1; | ||||||
| 1174 | |||||||
| 1175 | AttributeList Attrs; ///< parameter attributes for callable | ||||||
| 1176 | FunctionType *FTy; | ||||||
| 1177 | |||||||
| 1178 | template <class... ArgsTy> | ||||||
| 1179 | CallBase(AttributeList const &A, FunctionType *FT, ArgsTy &&... Args) | ||||||
| 1180 | : Instruction(std::forward<ArgsTy>(Args)...), Attrs(A), FTy(FT) {} | ||||||
| 1181 | |||||||
| 1182 | using Instruction::Instruction; | ||||||
| 1183 | |||||||
| 1184 | bool hasDescriptor() const { return Value::HasDescriptor; } | ||||||
| 1185 | |||||||
| 1186 | unsigned getNumSubclassExtraOperands() const { | ||||||
| 1187 | switch (getOpcode()) { | ||||||
| 1188 | case Instruction::Call: | ||||||
| 1189 | return 0; | ||||||
| 1190 | case Instruction::Invoke: | ||||||
| 1191 | return 2; | ||||||
| 1192 | case Instruction::CallBr: | ||||||
| 1193 | return getNumSubclassExtraOperandsDynamic(); | ||||||
| 1194 | } | ||||||
| 1195 | llvm_unreachable("Invalid opcode!")__builtin_unreachable(); | ||||||
| 1196 | } | ||||||
| 1197 | |||||||
| 1198 | /// Get the number of extra operands for instructions that don't have a fixed | ||||||
| 1199 | /// number of extra operands. | ||||||
| 1200 | unsigned getNumSubclassExtraOperandsDynamic() const; | ||||||
| 1201 | |||||||
| 1202 | public: | ||||||
| 1203 | using Instruction::getContext; | ||||||
| 1204 | |||||||
| 1205 | /// Create a clone of \p CB with a different set of operand bundles and | ||||||
| 1206 | /// insert it before \p InsertPt. | ||||||
| 1207 | /// | ||||||
| 1208 | /// The returned call instruction is identical \p CB in every way except that | ||||||
| 1209 | /// the operand bundles for the new instruction are set to the operand bundles | ||||||
| 1210 | /// in \p Bundles. | ||||||
| 1211 | static CallBase *Create(CallBase *CB, ArrayRef<OperandBundleDef> Bundles, | ||||||
| 1212 | Instruction *InsertPt = nullptr); | ||||||
| 1213 | |||||||
| 1214 | /// Create a clone of \p CB with the operand bundle with the tag matching | ||||||
| 1215 | /// \p Bundle's tag replaced with Bundle, and insert it before \p InsertPt. | ||||||
| 1216 | /// | ||||||
| 1217 | /// The returned call instruction is identical \p CI in every way except that | ||||||
| 1218 | /// the specified operand bundle has been replaced. | ||||||
| 1219 | static CallBase *Create(CallBase *CB, | ||||||
| 1220 | OperandBundleDef Bundle, | ||||||
| 1221 | Instruction *InsertPt = nullptr); | ||||||
| 1222 | |||||||
| 1223 | /// Create a clone of \p CB with operand bundle \p OB added. | ||||||
| 1224 | static CallBase *addOperandBundle(CallBase *CB, uint32_t ID, | ||||||
| 1225 | OperandBundleDef OB, | ||||||
| 1226 | Instruction *InsertPt = nullptr); | ||||||
| 1227 | |||||||
| 1228 | /// Create a clone of \p CB with operand bundle \p ID removed. | ||||||
| 1229 | static CallBase *removeOperandBundle(CallBase *CB, uint32_t ID, | ||||||
| 1230 | Instruction *InsertPt = nullptr); | ||||||
| 1231 | |||||||
| 1232 | static bool classof(const Instruction *I) { | ||||||
| 1233 | return I->getOpcode() == Instruction::Call || | ||||||
| 1234 | I->getOpcode() == Instruction::Invoke || | ||||||
| 1235 | I->getOpcode() == Instruction::CallBr; | ||||||
| 1236 | } | ||||||
| 1237 | static bool classof(const Value *V) { | ||||||
| 1238 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); | ||||||
| 1239 | } | ||||||
| 1240 | |||||||
| 1241 | FunctionType *getFunctionType() const { return FTy; } | ||||||
| 1242 | |||||||
| 1243 | void mutateFunctionType(FunctionType *FTy) { | ||||||
| 1244 | Value::mutateType(FTy->getReturnType()); | ||||||
| 1245 | this->FTy = FTy; | ||||||
| 1246 | } | ||||||
| 1247 | |||||||
| 1248 | DECLARE_TRANSPARENT_OPERAND_ACCESSORS(Value)public: inline Value *getOperand(unsigned) const; inline void setOperand(unsigned, Value*); inline op_iterator op_begin(); inline const_op_iterator op_begin() const; inline op_iterator op_end(); inline const_op_iterator op_end() const; protected : template <int> inline Use &Op(); template <int > inline const Use &Op() const; public: inline unsigned getNumOperands() const; | ||||||
| 1249 | |||||||
| 1250 | /// data_operands_begin/data_operands_end - Return iterators iterating over | ||||||
| 1251 | /// the call / invoke argument list and bundle operands. For invokes, this is | ||||||
| 1252 | /// the set of instruction operands except the invoke target and the two | ||||||
| 1253 | /// successor blocks; and for calls this is the set of instruction operands | ||||||
| 1254 | /// except the call target. | ||||||
| 1255 | User::op_iterator data_operands_begin() { return op_begin(); } | ||||||
| 1256 | User::const_op_iterator data_operands_begin() const { | ||||||
| 1257 | return const_cast<CallBase *>(this)->data_operands_begin(); | ||||||
| 1258 | } | ||||||
| 1259 | User::op_iterator data_operands_end() { | ||||||
| 1260 | // Walk from the end of the operands over the called operand and any | ||||||
| 1261 | // subclass operands. | ||||||
| 1262 | return op_end() - getNumSubclassExtraOperands() - 1; | ||||||
| 1263 | } | ||||||
| 1264 | User::const_op_iterator data_operands_end() const { | ||||||
| 1265 | return const_cast<CallBase *>(this)->data_operands_end(); | ||||||
| 1266 | } | ||||||
| 1267 | iterator_range<User::op_iterator> data_ops() { | ||||||
| 1268 | return make_range(data_operands_begin(), data_operands_end()); | ||||||
| 1269 | } | ||||||
| 1270 | iterator_range<User::const_op_iterator> data_ops() const { | ||||||
| 1271 | return make_range(data_operands_begin(), data_operands_end()); | ||||||
| 1272 | } | ||||||
| 1273 | bool data_operands_empty() const { | ||||||
| 1274 | return data_operands_end() == data_operands_begin(); | ||||||
| 1275 | } | ||||||
| 1276 | unsigned data_operands_size() const { | ||||||
| 1277 | return std::distance(data_operands_begin(), data_operands_end()); | ||||||
| 1278 | } | ||||||
| 1279 | |||||||
| 1280 | bool isDataOperand(const Use *U) const { | ||||||
| 1281 | assert(this == U->getUser() &&((void)0) | ||||||
| 1282 | "Only valid to query with a use of this instruction!")((void)0); | ||||||
| 1283 | return data_operands_begin() <= U && U < data_operands_end(); | ||||||
| 1284 | } | ||||||
| 1285 | bool isDataOperand(Value::const_user_iterator UI) const { | ||||||
| 1286 | return isDataOperand(&UI.getUse()); | ||||||
| 1287 | } | ||||||
| 1288 | |||||||
| 1289 | /// Given a value use iterator, return the data operand corresponding to it. | ||||||
| 1290 | /// Iterator must actually correspond to a data operand. | ||||||
| 1291 | unsigned getDataOperandNo(Value::const_user_iterator UI) const { | ||||||
| 1292 | return getDataOperandNo(&UI.getUse()); | ||||||
| 1293 | } | ||||||
| 1294 | |||||||
| 1295 | /// Given a use for a data operand, get the data operand number that | ||||||
| 1296 | /// corresponds to it. | ||||||
| 1297 | unsigned getDataOperandNo(const Use *U) const { | ||||||
| 1298 | assert(isDataOperand(U) && "Data operand # out of range!")((void)0); | ||||||
| 1299 | return U - data_operands_begin(); | ||||||
| 1300 | } | ||||||
| 1301 | |||||||
| 1302 | /// Return the iterator pointing to the beginning of the argument list. | ||||||
| 1303 | User::op_iterator arg_begin() { return op_begin(); } | ||||||
| 1304 | User::const_op_iterator arg_begin() const { | ||||||
| 1305 | return const_cast<CallBase *>(this)->arg_begin(); | ||||||
| 1306 | } | ||||||
| 1307 | |||||||
| 1308 | /// Return the iterator pointing to the end of the argument list. | ||||||
| 1309 | User::op_iterator arg_end() { | ||||||
| 1310 | // From the end of the data operands, walk backwards past the bundle | ||||||
| 1311 | // operands. | ||||||
| 1312 | return data_operands_end() - getNumTotalBundleOperands(); | ||||||
| 1313 | } | ||||||
| 1314 | User::const_op_iterator arg_end() const { | ||||||
| 1315 | return const_cast<CallBase *>(this)->arg_end(); | ||||||
| 1316 | } | ||||||
| 1317 | |||||||
| 1318 | /// Iteration adapter for range-for loops. | ||||||
| 1319 | iterator_range<User::op_iterator> args() { | ||||||
| 1320 | return make_range(arg_begin(), arg_end()); | ||||||
| 1321 | } | ||||||
| 1322 | iterator_range<User::const_op_iterator> args() const { | ||||||
| 1323 | return make_range(arg_begin(), arg_end()); | ||||||
| 1324 | } | ||||||
| 1325 | bool arg_empty() const { return arg_end() == arg_begin(); } | ||||||
| 1326 | unsigned arg_size() const { return arg_end() - arg_begin(); } | ||||||
| 1327 | |||||||
| 1328 | // Legacy API names that duplicate the above and will be removed once users | ||||||
| 1329 | // are migrated. | ||||||
| 1330 | iterator_range<User::op_iterator> arg_operands() { | ||||||
| 1331 | return make_range(arg_begin(), arg_end()); | ||||||
| 1332 | } | ||||||
| 1333 | iterator_range<User::const_op_iterator> arg_operands() const { | ||||||
| 1334 | return make_range(arg_begin(), arg_end()); | ||||||
| 1335 | } | ||||||
| 1336 | unsigned getNumArgOperands() const { return arg_size(); } | ||||||
| 1337 | |||||||
| 1338 | Value *getArgOperand(unsigned i) const { | ||||||
| 1339 | assert(i < getNumArgOperands() && "Out of bounds!")((void)0); | ||||||
| 1340 | return getOperand(i); | ||||||
| 1341 | } | ||||||
| 1342 | |||||||
| 1343 | void setArgOperand(unsigned i, Value *v) { | ||||||
| 1344 | assert(i < getNumArgOperands() && "Out of bounds!")((void)0); | ||||||
| 1345 | setOperand(i, v); | ||||||
| 1346 | } | ||||||
| 1347 | |||||||
| 1348 | /// Wrappers for getting the \c Use of a call argument. | ||||||
| 1349 | const Use &getArgOperandUse(unsigned i) const { | ||||||
| 1350 | assert(i < getNumArgOperands() && "Out of bounds!")((void)0); | ||||||
| 1351 | return User::getOperandUse(i); | ||||||
| 1352 | } | ||||||
| 1353 | Use &getArgOperandUse(unsigned i) { | ||||||
| 1354 | assert(i < getNumArgOperands() && "Out of bounds!")((void)0); | ||||||
| 1355 | return User::getOperandUse(i); | ||||||
| 1356 | } | ||||||
| 1357 | |||||||
| 1358 | bool isArgOperand(const Use *U) const { | ||||||
| 1359 | assert(this == U->getUser() &&((void)0) | ||||||
| 1360 | "Only valid to query with a use of this instruction!")((void)0); | ||||||
| 1361 | return arg_begin() <= U && U < arg_end(); | ||||||
| 1362 | } | ||||||
| 1363 | bool isArgOperand(Value::const_user_iterator UI) const { | ||||||
| 1364 | return isArgOperand(&UI.getUse()); | ||||||
| 1365 | } | ||||||
| 1366 | |||||||
| 1367 | /// Given a use for a arg operand, get the arg operand number that | ||||||
| 1368 | /// corresponds to it. | ||||||
| 1369 | unsigned getArgOperandNo(const Use *U) const { | ||||||
| 1370 | assert(isArgOperand(U) && "Arg operand # out of range!")((void)0); | ||||||
| 1371 | return U - arg_begin(); | ||||||
| 1372 | } | ||||||
| 1373 | |||||||
| 1374 | /// Given a value use iterator, return the arg operand number corresponding to | ||||||
| 1375 | /// it. Iterator must actually correspond to a data operand. | ||||||
| 1376 | unsigned getArgOperandNo(Value::const_user_iterator UI) const { | ||||||
| 1377 | return getArgOperandNo(&UI.getUse()); | ||||||
| 1378 | } | ||||||
| 1379 | |||||||
| 1380 | /// Returns true if this CallSite passes the given Value* as an argument to | ||||||
| 1381 | /// the called function. | ||||||
| 1382 | bool hasArgument(const Value *V) const { | ||||||
| 1383 | return llvm::is_contained(args(), V); | ||||||
| 1384 | } | ||||||
| 1385 | |||||||
| 1386 | Value *getCalledOperand() const { return Op<CalledOperandOpEndIdx>(); } | ||||||
| 1387 | |||||||
| 1388 | const Use &getCalledOperandUse() const { return Op<CalledOperandOpEndIdx>(); } | ||||||
| 1389 | Use &getCalledOperandUse() { return Op<CalledOperandOpEndIdx>(); } | ||||||
| 1390 | |||||||
| 1391 | /// Returns the function called, or null if this is an | ||||||
| 1392 | /// indirect function invocation. | ||||||
| 1393 | Function *getCalledFunction() const { | ||||||
| 1394 | return dyn_cast_or_null<Function>(getCalledOperand()); | ||||||
| 1395 | } | ||||||
| 1396 | |||||||
| 1397 | /// Return true if the callsite is an indirect call. | ||||||
| 1398 | bool isIndirectCall() const; | ||||||
| 1399 | |||||||
| 1400 | /// Determine whether the passed iterator points to the callee operand's Use. | ||||||
| 1401 | bool isCallee(Value::const_user_iterator UI) const { | ||||||
| 1402 | return isCallee(&UI.getUse()); | ||||||
| 1403 | } | ||||||
| 1404 | |||||||
| 1405 | /// Determine whether this Use is the callee operand's Use. | ||||||
| 1406 | bool isCallee(const Use *U) const { return &getCalledOperandUse() == U; } | ||||||
| 1407 | |||||||
| 1408 | /// Helper to get the caller (the parent function). | ||||||
| 1409 | Function *getCaller(); | ||||||
| 1410 | const Function *getCaller() const { | ||||||
| 1411 | return const_cast<CallBase *>(this)->getCaller(); | ||||||
| 1412 | } | ||||||
| 1413 | |||||||
| 1414 | /// Tests if this call site must be tail call optimized. Only a CallInst can | ||||||
| 1415 | /// be tail call optimized. | ||||||
| 1416 | bool isMustTailCall() const; | ||||||
| 1417 | |||||||
| 1418 | /// Tests if this call site is marked as a tail call. | ||||||
| 1419 | bool isTailCall() const; | ||||||
| 1420 | |||||||
| 1421 | /// Returns the intrinsic ID of the intrinsic called or | ||||||
| 1422 | /// Intrinsic::not_intrinsic if the called function is not an intrinsic, or if | ||||||
| 1423 | /// this is an indirect call. | ||||||
| 1424 | Intrinsic::ID getIntrinsicID() const; | ||||||
| 1425 | |||||||
| 1426 | void setCalledOperand(Value *V) { Op<CalledOperandOpEndIdx>() = V; } | ||||||
| 1427 | |||||||
| 1428 | /// Sets the function called, including updating the function type. | ||||||
| 1429 | void setCalledFunction(Function *Fn) { | ||||||
| 1430 | setCalledFunction(Fn->getFunctionType(), Fn); | ||||||
| 1431 | } | ||||||
| 1432 | |||||||
| 1433 | /// Sets the function called, including updating the function type. | ||||||
| 1434 | void setCalledFunction(FunctionCallee Fn) { | ||||||
| 1435 | setCalledFunction(Fn.getFunctionType(), Fn.getCallee()); | ||||||
| 1436 | } | ||||||
| 1437 | |||||||
| 1438 | /// Sets the function called, including updating to the specified function | ||||||
| 1439 | /// type. | ||||||
| 1440 | void setCalledFunction(FunctionType *FTy, Value *Fn) { | ||||||
| 1441 | this->FTy = FTy; | ||||||
| 1442 | assert(cast<PointerType>(Fn->getType())->isOpaqueOrPointeeTypeMatches(FTy))((void)0); | ||||||
| 1443 | // This function doesn't mutate the return type, only the function | ||||||
| 1444 | // type. Seems broken, but I'm just gonna stick an assert in for now. | ||||||
| 1445 | assert(getType() == FTy->getReturnType())((void)0); | ||||||
| 1446 | setCalledOperand(Fn); | ||||||
| 1447 | } | ||||||
| 1448 | |||||||
| 1449 | CallingConv::ID getCallingConv() const { | ||||||
| 1450 | return getSubclassData<CallingConvField>(); | ||||||
| 1451 | } | ||||||
| 1452 | |||||||
| 1453 | void setCallingConv(CallingConv::ID CC) { | ||||||
| 1454 | setSubclassData<CallingConvField>(CC); | ||||||
| 1455 | } | ||||||
| 1456 | |||||||
| 1457 | /// Check if this call is an inline asm statement. | ||||||
| 1458 | bool isInlineAsm() const { return isa<InlineAsm>(getCalledOperand()); } | ||||||
| 1459 | |||||||
| 1460 | /// \name Attribute API | ||||||
| 1461 | /// | ||||||
| 1462 | /// These methods access and modify attributes on this call (including | ||||||
| 1463 | /// looking through to the attributes on the called function when necessary). | ||||||
| 1464 | ///@{ | ||||||
| 1465 | |||||||
| 1466 | /// Return the parameter attributes for this call. | ||||||
| 1467 | /// | ||||||
| 1468 | AttributeList getAttributes() const { return Attrs; } | ||||||
| 1469 | |||||||
| 1470 | /// Set the parameter attributes for this call. | ||||||
| 1471 | /// | ||||||
| 1472 | void setAttributes(AttributeList A) { Attrs = A; } | ||||||
| 1473 | |||||||
| 1474 | /// Determine whether this call has the given attribute. If it does not | ||||||
| 1475 | /// then determine if the called function has the attribute, but only if | ||||||
| 1476 | /// the attribute is allowed for the call. | ||||||
| 1477 | bool hasFnAttr(Attribute::AttrKind Kind) const { | ||||||
| 1478 | assert(Kind != Attribute::NoBuiltin &&((void)0) | ||||||
| 1479 | "Use CallBase::isNoBuiltin() to check for Attribute::NoBuiltin")((void)0); | ||||||
| 1480 | return hasFnAttrImpl(Kind); | ||||||
| 1481 | } | ||||||
| 1482 | |||||||
| 1483 | /// Determine whether this call has the given attribute. If it does not | ||||||
| 1484 | /// then determine if the called function has the attribute, but only if | ||||||
| 1485 | /// the attribute is allowed for the call. | ||||||
| 1486 | bool hasFnAttr(StringRef Kind) const { return hasFnAttrImpl(Kind); } | ||||||
| 1487 | |||||||
| 1488 | /// adds the attribute to the list of attributes. | ||||||
| 1489 | void addAttribute(unsigned i, Attribute::AttrKind Kind) { | ||||||
| 1490 | AttributeList PAL = getAttributes(); | ||||||
| 1491 | PAL = PAL.addAttribute(getContext(), i, Kind); | ||||||
| 1492 | setAttributes(PAL); | ||||||
| 1493 | } | ||||||
| 1494 | |||||||
| 1495 | /// adds the attribute to the list of attributes. | ||||||
| 1496 | void addAttribute(unsigned i, Attribute Attr) { | ||||||
| 1497 | AttributeList PAL = getAttributes(); | ||||||
| 1498 | PAL = PAL.addAttribute(getContext(), i, Attr); | ||||||
| 1499 | setAttributes(PAL); | ||||||
| 1500 | } | ||||||
| 1501 | |||||||
| 1502 | /// Adds the attribute to the indicated argument | ||||||
| 1503 | void addParamAttr(unsigned ArgNo, Attribute::AttrKind Kind) { | ||||||
| 1504 | assert(ArgNo < getNumArgOperands() && "Out of bounds")((void)0); | ||||||
| 1505 | AttributeList PAL = getAttributes(); | ||||||
| 1506 | PAL = PAL.addParamAttribute(getContext(), ArgNo, Kind); | ||||||
| 1507 | setAttributes(PAL); | ||||||
| 1508 | } | ||||||
| 1509 | |||||||
| 1510 | /// Adds the attribute to the indicated argument | ||||||
| 1511 | void addParamAttr(unsigned ArgNo, Attribute Attr) { | ||||||
| 1512 | assert(ArgNo < getNumArgOperands() && "Out of bounds")((void)0); | ||||||
| 1513 | AttributeList PAL = getAttributes(); | ||||||
| 1514 | PAL = PAL.addParamAttribute(getContext(), ArgNo, Attr); | ||||||
| 1515 | setAttributes(PAL); | ||||||
| 1516 | } | ||||||
| 1517 | |||||||
| 1518 | /// removes the attribute from the list of attributes. | ||||||
| 1519 | void removeAttribute(unsigned i, Attribute::AttrKind Kind) { | ||||||
| 1520 | AttributeList PAL = getAttributes(); | ||||||
| 1521 | PAL = PAL.removeAttribute(getContext(), i, Kind); | ||||||
| 1522 | setAttributes(PAL); | ||||||
| 1523 | } | ||||||
| 1524 | |||||||
| 1525 | /// removes the attribute from the list of attributes. | ||||||
| 1526 | void removeAttribute(unsigned i, StringRef Kind) { | ||||||
| 1527 | AttributeList PAL = getAttributes(); | ||||||
| 1528 | PAL = PAL.removeAttribute(getContext(), i, Kind); | ||||||
| 1529 | setAttributes(PAL); | ||||||
| 1530 | } | ||||||
| 1531 | |||||||
| 1532 | void removeAttributes(unsigned i, const AttrBuilder &Attrs) { | ||||||
| 1533 | AttributeList PAL = getAttributes(); | ||||||
| 1534 | PAL = PAL.removeAttributes(getContext(), i, Attrs); | ||||||
| 1535 | setAttributes(PAL); | ||||||
| 1536 | } | ||||||
| 1537 | |||||||
| 1538 | /// Removes the attribute from the given argument | ||||||
| 1539 | void removeParamAttr(unsigned ArgNo, Attribute::AttrKind Kind) { | ||||||
| 1540 | assert(ArgNo < getNumArgOperands() && "Out of bounds")((void)0); | ||||||
| 1541 | AttributeList PAL = getAttributes(); | ||||||
| 1542 | PAL = PAL.removeParamAttribute(getContext(), ArgNo, Kind); | ||||||
| 1543 | setAttributes(PAL); | ||||||
| 1544 | } | ||||||
| 1545 | |||||||
| 1546 | /// Removes the attribute from the given argument | ||||||
| 1547 | void removeParamAttr(unsigned ArgNo, StringRef Kind) { | ||||||
| 1548 | assert(ArgNo < getNumArgOperands() && "Out of bounds")((void)0); | ||||||
| 1549 | AttributeList PAL = getAttributes(); | ||||||
| 1550 | PAL = PAL.removeParamAttribute(getContext(), ArgNo, Kind); | ||||||
| 1551 | setAttributes(PAL); | ||||||
| 1552 | } | ||||||
| 1553 | |||||||
| 1554 | /// Removes the attributes from the given argument | ||||||
| 1555 | void removeParamAttrs(unsigned ArgNo, const AttrBuilder &Attrs) { | ||||||
| 1556 | AttributeList PAL = getAttributes(); | ||||||
| 1557 | PAL = PAL.removeParamAttributes(getContext(), ArgNo, Attrs); | ||||||
| 1558 | setAttributes(PAL); | ||||||
| 1559 | } | ||||||
| 1560 | |||||||
| 1561 | /// adds the dereferenceable attribute to the list of attributes. | ||||||
| 1562 | void addDereferenceableAttr(unsigned i, uint64_t Bytes) { | ||||||
| 1563 | AttributeList PAL = getAttributes(); | ||||||
| 1564 | PAL = PAL.addDereferenceableAttr(getContext(), i, Bytes); | ||||||
| 1565 | setAttributes(PAL); | ||||||
| 1566 | } | ||||||
| 1567 | |||||||
| 1568 | /// adds the dereferenceable_or_null attribute to the list of | ||||||
| 1569 | /// attributes. | ||||||
| 1570 | void addDereferenceableOrNullAttr(unsigned i, uint64_t Bytes) { | ||||||
| 1571 | AttributeList PAL = getAttributes(); | ||||||
| 1572 | PAL = PAL.addDereferenceableOrNullAttr(getContext(), i, Bytes); | ||||||
| 1573 | setAttributes(PAL); | ||||||
| 1574 | } | ||||||
| 1575 | |||||||
| 1576 | /// Determine whether the return value has the given attribute. | ||||||
| 1577 | bool hasRetAttr(Attribute::AttrKind Kind) const { | ||||||
| 1578 | return hasRetAttrImpl(Kind); | ||||||
| 1579 | } | ||||||
| 1580 | /// Determine whether the return value has the given attribute. | ||||||
| 1581 | bool hasRetAttr(StringRef Kind) const { return hasRetAttrImpl(Kind); } | ||||||
| 1582 | |||||||
| 1583 | /// Determine whether the argument or parameter has the given attribute. | ||||||
| 1584 | bool paramHasAttr(unsigned ArgNo, Attribute::AttrKind Kind) const; | ||||||
| 1585 | |||||||
| 1586 | /// Get the attribute of a given kind at a position. | ||||||
| 1587 | Attribute getAttribute(unsigned i, Attribute::AttrKind Kind) const { | ||||||
| 1588 | return getAttributes().getAttribute(i, Kind); | ||||||
| 1589 | } | ||||||
| 1590 | |||||||
| 1591 | /// Get the attribute of a given kind at a position. | ||||||
| 1592 | Attribute getAttribute(unsigned i, StringRef Kind) const { | ||||||
| 1593 | return getAttributes().getAttribute(i, Kind); | ||||||
| 1594 | } | ||||||
| 1595 | |||||||
| 1596 | /// Get the attribute of a given kind from a given arg | ||||||
| 1597 | Attribute getParamAttr(unsigned ArgNo, Attribute::AttrKind Kind) const { | ||||||
| 1598 | assert(ArgNo < getNumArgOperands() && "Out of bounds")((void)0); | ||||||
| 1599 | return getAttributes().getParamAttr(ArgNo, Kind); | ||||||
| 1600 | } | ||||||
| 1601 | |||||||
| 1602 | /// Get the attribute of a given kind from a given arg | ||||||
| 1603 | Attribute getParamAttr(unsigned ArgNo, StringRef Kind) const { | ||||||
| 1604 | assert(ArgNo < getNumArgOperands() && "Out of bounds")((void)0); | ||||||
| 1605 | return getAttributes().getParamAttr(ArgNo, Kind); | ||||||
| 1606 | } | ||||||
| 1607 | |||||||
| 1608 | /// Return true if the data operand at index \p i has the attribute \p | ||||||
| 1609 | /// A. | ||||||
| 1610 | /// | ||||||
| 1611 | /// Data operands include call arguments and values used in operand bundles, | ||||||
| 1612 | /// but does not include the callee operand. This routine dispatches to the | ||||||
| 1613 | /// underlying AttributeList or the OperandBundleUser as appropriate. | ||||||
| 1614 | /// | ||||||
| 1615 | /// The index \p i is interpreted as | ||||||
| 1616 | /// | ||||||
| 1617 | /// \p i == Attribute::ReturnIndex -> the return value | ||||||
| 1618 | /// \p i in [1, arg_size + 1) -> argument number (\p i - 1) | ||||||
| 1619 | /// \p i in [arg_size + 1, data_operand_size + 1) -> bundle operand at index | ||||||
| 1620 | /// (\p i - 1) in the operand list. | ||||||
| 1621 | bool dataOperandHasImpliedAttr(unsigned i, Attribute::AttrKind Kind) const { | ||||||
| 1622 | // Note that we have to add one because `i` isn't zero-indexed. | ||||||
| 1623 | assert(i < (getNumArgOperands() + getNumTotalBundleOperands() + 1) &&((void)0) | ||||||
| 1624 | "Data operand index out of bounds!")((void)0); | ||||||
| 1625 | |||||||
| 1626 | // The attribute A can either be directly specified, if the operand in | ||||||
| 1627 | // question is a call argument; or be indirectly implied by the kind of its | ||||||
| 1628 | // containing operand bundle, if the operand is a bundle operand. | ||||||
| 1629 | |||||||
| 1630 | if (i == AttributeList::ReturnIndex) | ||||||
| 1631 | return hasRetAttr(Kind); | ||||||
| 1632 | |||||||
| 1633 | // FIXME: Avoid these i - 1 calculations and update the API to use | ||||||
| 1634 | // zero-based indices. | ||||||
| 1635 | if (i < (getNumArgOperands() + 1)) | ||||||
| 1636 | return paramHasAttr(i - 1, Kind); | ||||||
| 1637 | |||||||
| 1638 | assert(hasOperandBundles() && i >= (getBundleOperandsStartIndex() + 1) &&((void)0) | ||||||
| 1639 | "Must be either a call argument or an operand bundle!")((void)0); | ||||||
| 1640 | return bundleOperandHasAttr(i - 1, Kind); | ||||||
| 1641 | } | ||||||
| 1642 | |||||||
| 1643 | /// Determine whether this data operand is not captured. | ||||||
| 1644 | // FIXME: Once this API is no longer duplicated in `CallSite`, rename this to | ||||||
| 1645 | // better indicate that this may return a conservative answer. | ||||||
| 1646 | bool doesNotCapture(unsigned OpNo) const { | ||||||
| 1647 | return dataOperandHasImpliedAttr(OpNo + 1, Attribute::NoCapture); | ||||||
| 1648 | } | ||||||
| 1649 | |||||||
| 1650 | /// Determine whether this argument is passed by value. | ||||||
| 1651 | bool isByValArgument(unsigned ArgNo) const { | ||||||
| 1652 | return paramHasAttr(ArgNo, Attribute::ByVal); | ||||||
| 1653 | } | ||||||
| 1654 | |||||||
| 1655 | /// Determine whether this argument is passed in an alloca. | ||||||
| 1656 | bool isInAllocaArgument(unsigned ArgNo) const { | ||||||
| 1657 | return paramHasAttr(ArgNo, Attribute::InAlloca); | ||||||
| 1658 | } | ||||||
| 1659 | |||||||
| 1660 | /// Determine whether this argument is passed by value, in an alloca, or is | ||||||
| 1661 | /// preallocated. | ||||||
| 1662 | bool isPassPointeeByValueArgument(unsigned ArgNo) const { | ||||||
| 1663 | return paramHasAttr(ArgNo, Attribute::ByVal) || | ||||||
| 1664 | paramHasAttr(ArgNo, Attribute::InAlloca) || | ||||||
| 1665 | paramHasAttr(ArgNo, Attribute::Preallocated); | ||||||
| 1666 | } | ||||||
| 1667 | |||||||
| 1668 | /// Determine whether passing undef to this argument is undefined behavior. | ||||||
| 1669 | /// If passing undef to this argument is UB, passing poison is UB as well | ||||||
| 1670 | /// because poison is more undefined than undef. | ||||||
| 1671 | bool isPassingUndefUB(unsigned ArgNo) const { | ||||||
| 1672 | return paramHasAttr(ArgNo, Attribute::NoUndef) || | ||||||
| 1673 | // dereferenceable implies noundef. | ||||||
| 1674 | paramHasAttr(ArgNo, Attribute::Dereferenceable) || | ||||||
| 1675 | // dereferenceable implies noundef, and null is a well-defined value. | ||||||
| 1676 | paramHasAttr(ArgNo, Attribute::DereferenceableOrNull); | ||||||
| 1677 | } | ||||||
| 1678 | |||||||
| 1679 | /// Determine if there are is an inalloca argument. Only the last argument can | ||||||
| 1680 | /// have the inalloca attribute. | ||||||
| 1681 | bool hasInAllocaArgument() const { | ||||||
| 1682 | return !arg_empty() && paramHasAttr(arg_size() - 1, Attribute::InAlloca); | ||||||
| 1683 | } | ||||||
| 1684 | |||||||
| 1685 | // FIXME: Once this API is no longer duplicated in `CallSite`, rename this to | ||||||
| 1686 | // better indicate that this may return a conservative answer. | ||||||
| 1687 | bool doesNotAccessMemory(unsigned OpNo) const { | ||||||
| 1688 | return dataOperandHasImpliedAttr(OpNo + 1, Attribute::ReadNone); | ||||||
| 1689 | } | ||||||
| 1690 | |||||||
| 1691 | // FIXME: Once this API is no longer duplicated in `CallSite`, rename this to | ||||||
| 1692 | // better indicate that this may return a conservative answer. | ||||||
| 1693 | bool onlyReadsMemory(unsigned OpNo) const { | ||||||
| 1694 | return dataOperandHasImpliedAttr(OpNo + 1, Attribute::ReadOnly) || | ||||||
| 1695 | dataOperandHasImpliedAttr(OpNo + 1, Attribute::ReadNone); | ||||||
| 1696 | } | ||||||
| 1697 | |||||||
| 1698 | // FIXME: Once this API is no longer duplicated in `CallSite`, rename this to | ||||||
| 1699 | // better indicate that this may return a conservative answer. | ||||||
| 1700 | bool doesNotReadMemory(unsigned OpNo) const { | ||||||
| 1701 | return dataOperandHasImpliedAttr(OpNo + 1, Attribute::WriteOnly) || | ||||||
| 1702 | dataOperandHasImpliedAttr(OpNo + 1, Attribute::ReadNone); | ||||||
| 1703 | } | ||||||
| 1704 | |||||||
| 1705 | /// Extract the alignment of the return value. | ||||||
| 1706 | MaybeAlign getRetAlign() const { return Attrs.getRetAlignment(); } | ||||||
| 1707 | |||||||
| 1708 | /// Extract the alignment for a call or parameter (0=unknown). | ||||||
| 1709 | MaybeAlign getParamAlign(unsigned ArgNo) const { | ||||||
| 1710 | return Attrs.getParamAlignment(ArgNo); | ||||||
| 1711 | } | ||||||
| 1712 | |||||||
| 1713 | MaybeAlign getParamStackAlign(unsigned ArgNo) const { | ||||||
| 1714 | return Attrs.getParamStackAlignment(ArgNo); | ||||||
| 1715 | } | ||||||
| 1716 | |||||||
| 1717 | /// Extract the byval type for a call or parameter. | ||||||
| 1718 | Type *getParamByValType(unsigned ArgNo) const { | ||||||
| 1719 | if (auto *Ty = Attrs.getParamByValType(ArgNo)) | ||||||
| 1720 | return Ty; | ||||||
| 1721 | if (const Function *F = getCalledFunction()) | ||||||
| 1722 | return F->getAttributes().getParamByValType(ArgNo); | ||||||
| 1723 | return nullptr; | ||||||
| 1724 | } | ||||||
| 1725 | |||||||
| 1726 | /// Extract the preallocated type for a call or parameter. | ||||||
| 1727 | Type *getParamPreallocatedType(unsigned ArgNo) const { | ||||||
| 1728 | if (auto *Ty = Attrs.getParamPreallocatedType(ArgNo)) | ||||||
| 1729 | return Ty; | ||||||
| 1730 | if (const Function *F = getCalledFunction()) | ||||||
| 1731 | return F->getAttributes().getParamPreallocatedType(ArgNo); | ||||||
| 1732 | return nullptr; | ||||||
| 1733 | } | ||||||
| 1734 | |||||||
| 1735 | /// Extract the preallocated type for a call or parameter. | ||||||
| 1736 | Type *getParamInAllocaType(unsigned ArgNo) const { | ||||||
| 1737 | if (auto *Ty = Attrs.getParamInAllocaType(ArgNo)) | ||||||
| 1738 | return Ty; | ||||||
| 1739 | if (const Function *F = getCalledFunction()) | ||||||
| 1740 | return F->getAttributes().getParamInAllocaType(ArgNo); | ||||||
| 1741 | return nullptr; | ||||||
| 1742 | } | ||||||
| 1743 | |||||||
| 1744 | /// Extract the number of dereferenceable bytes for a call or | ||||||
| 1745 | /// parameter (0=unknown). | ||||||
| 1746 | uint64_t getDereferenceableBytes(unsigned i) const { | ||||||
| 1747 | return Attrs.getDereferenceableBytes(i); | ||||||
| 1748 | } | ||||||
| 1749 | |||||||
| 1750 | /// Extract the number of dereferenceable_or_null bytes for a call or | ||||||
| 1751 | /// parameter (0=unknown). | ||||||
| 1752 | uint64_t getDereferenceableOrNullBytes(unsigned i) const { | ||||||
| 1753 | return Attrs.getDereferenceableOrNullBytes(i); | ||||||
| 1754 | } | ||||||
| 1755 | |||||||
| 1756 | /// Return true if the return value is known to be not null. | ||||||
| 1757 | /// This may be because it has the nonnull attribute, or because at least | ||||||
| 1758 | /// one byte is dereferenceable and the pointer is in addrspace(0). | ||||||
| 1759 | bool isReturnNonNull() const; | ||||||
| 1760 | |||||||
| 1761 | /// Determine if the return value is marked with NoAlias attribute. | ||||||
| 1762 | bool returnDoesNotAlias() const { | ||||||
| 1763 | return Attrs.hasAttribute(AttributeList::ReturnIndex, Attribute::NoAlias); | ||||||
| 1764 | } | ||||||
| 1765 | |||||||
| 1766 | /// If one of the arguments has the 'returned' attribute, returns its | ||||||
| 1767 | /// operand value. Otherwise, return nullptr. | ||||||
| 1768 | Value *getReturnedArgOperand() const; | ||||||
| 1769 | |||||||
| 1770 | /// Return true if the call should not be treated as a call to a | ||||||
| 1771 | /// builtin. | ||||||
| 1772 | bool isNoBuiltin() const { | ||||||
| 1773 | return hasFnAttrImpl(Attribute::NoBuiltin) && | ||||||
| 1774 | !hasFnAttrImpl(Attribute::Builtin); | ||||||
| 1775 | } | ||||||
| 1776 | |||||||
| 1777 | /// Determine if the call requires strict floating point semantics. | ||||||
| 1778 | bool isStrictFP() const { return hasFnAttr(Attribute::StrictFP); } | ||||||
| 1779 | |||||||
| 1780 | /// Return true if the call should not be inlined. | ||||||
| 1781 | bool isNoInline() const { return hasFnAttr(Attribute::NoInline); } | ||||||
| 1782 | void setIsNoInline() { | ||||||
| 1783 | addAttribute(AttributeList::FunctionIndex, Attribute::NoInline); | ||||||
| 1784 | } | ||||||
| 1785 | /// Determine if the call does not access memory. | ||||||
| 1786 | bool doesNotAccessMemory() const { return hasFnAttr(Attribute::ReadNone); } | ||||||
| 1787 | void setDoesNotAccessMemory() { | ||||||
| 1788 | addAttribute(AttributeList::FunctionIndex, Attribute::ReadNone); | ||||||
| 1789 | } | ||||||
| 1790 | |||||||
| 1791 | /// Determine if the call does not access or only reads memory. | ||||||
| 1792 | bool onlyReadsMemory() const { | ||||||
| 1793 | return doesNotAccessMemory() || hasFnAttr(Attribute::ReadOnly); | ||||||
| 1794 | } | ||||||
| 1795 | |||||||
| 1796 | void setOnlyReadsMemory() { | ||||||
| 1797 | addAttribute(AttributeList::FunctionIndex, Attribute::ReadOnly); | ||||||
| 1798 | } | ||||||
| 1799 | |||||||
| 1800 | /// Determine if the call does not access or only writes memory. | ||||||
| 1801 | bool doesNotReadMemory() const { | ||||||
| 1802 | return doesNotAccessMemory() || hasFnAttr(Attribute::WriteOnly); | ||||||
| 1803 | } | ||||||
| 1804 | void setDoesNotReadMemory() { | ||||||
| 1805 | addAttribute(AttributeList::FunctionIndex, Attribute::WriteOnly); | ||||||
| 1806 | } | ||||||
| 1807 | |||||||
| 1808 | /// Determine if the call can access memmory only using pointers based | ||||||
| 1809 | /// on its arguments. | ||||||
| 1810 | bool onlyAccessesArgMemory() const { | ||||||
| 1811 | return hasFnAttr(Attribute::ArgMemOnly); | ||||||
| 1812 | } | ||||||
| 1813 | void setOnlyAccessesArgMemory() { | ||||||
| 1814 | addAttribute(AttributeList::FunctionIndex, Attribute::ArgMemOnly); | ||||||
| 1815 | } | ||||||
| 1816 | |||||||
| 1817 | /// Determine if the function may only access memory that is | ||||||
| 1818 | /// inaccessible from the IR. | ||||||
| 1819 | bool onlyAccessesInaccessibleMemory() const { | ||||||
| 1820 | return hasFnAttr(Attribute::InaccessibleMemOnly); | ||||||
| 1821 | } | ||||||
| 1822 | void setOnlyAccessesInaccessibleMemory() { | ||||||
| 1823 | addAttribute(AttributeList::FunctionIndex, Attribute::InaccessibleMemOnly); | ||||||
| 1824 | } | ||||||
| 1825 | |||||||
| 1826 | /// Determine if the function may only access memory that is | ||||||
| 1827 | /// either inaccessible from the IR or pointed to by its arguments. | ||||||
| 1828 | bool onlyAccessesInaccessibleMemOrArgMem() const { | ||||||
| 1829 | return hasFnAttr(Attribute::InaccessibleMemOrArgMemOnly); | ||||||
| 1830 | } | ||||||
| 1831 | void setOnlyAccessesInaccessibleMemOrArgMem() { | ||||||
| 1832 | addAttribute(AttributeList::FunctionIndex, | ||||||
| 1833 | Attribute::InaccessibleMemOrArgMemOnly); | ||||||
| 1834 | } | ||||||
| 1835 | /// Determine if the call cannot return. | ||||||
| 1836 | bool doesNotReturn() const { return hasFnAttr(Attribute::NoReturn); } | ||||||
| 1837 | void setDoesNotReturn() { | ||||||
| 1838 | addAttribute(AttributeList::FunctionIndex, Attribute::NoReturn); | ||||||
| 1839 | } | ||||||
| 1840 | |||||||
| 1841 | /// Determine if the call should not perform indirect branch tracking. | ||||||
| 1842 | bool doesNoCfCheck() const { return hasFnAttr(Attribute::NoCfCheck); } | ||||||
| 1843 | |||||||
| 1844 | /// Determine if the call cannot unwind. | ||||||
| 1845 | bool doesNotThrow() const { return hasFnAttr(Attribute::NoUnwind); } | ||||||
| 1846 | void setDoesNotThrow() { | ||||||
| 1847 | addAttribute(AttributeList::FunctionIndex, Attribute::NoUnwind); | ||||||
| 1848 | } | ||||||
| 1849 | |||||||
| 1850 | /// Determine if the invoke cannot be duplicated. | ||||||
| 1851 | bool cannotDuplicate() const { return hasFnAttr(Attribute::NoDuplicate); } | ||||||
| 1852 | void setCannotDuplicate() { | ||||||
| 1853 | addAttribute(AttributeList::FunctionIndex, Attribute::NoDuplicate); | ||||||
| 1854 | } | ||||||
| 1855 | |||||||
| 1856 | /// Determine if the call cannot be tail merged. | ||||||
| 1857 | bool cannotMerge() const { return hasFnAttr(Attribute::NoMerge); } | ||||||
| 1858 | void setCannotMerge() { | ||||||
| 1859 | addAttribute(AttributeList::FunctionIndex, Attribute::NoMerge); | ||||||
| 1860 | } | ||||||
| 1861 | |||||||
| 1862 | /// Determine if the invoke is convergent | ||||||
| 1863 | bool isConvergent() const { return hasFnAttr(Attribute::Convergent); } | ||||||
| 1864 | void setConvergent() { | ||||||
| 1865 | addAttribute(AttributeList::FunctionIndex, Attribute::Convergent); | ||||||
| 1866 | } | ||||||
| 1867 | void setNotConvergent() { | ||||||
| 1868 | removeAttribute(AttributeList::FunctionIndex, Attribute::Convergent); | ||||||
| 1869 | } | ||||||
| 1870 | |||||||
| 1871 | /// Determine if the call returns a structure through first | ||||||
| 1872 | /// pointer argument. | ||||||
| 1873 | bool hasStructRetAttr() const { | ||||||
| 1874 | if (getNumArgOperands() == 0) | ||||||
| 1875 | return false; | ||||||
| 1876 | |||||||
| 1877 | // Be friendly and also check the callee. | ||||||
| 1878 | return paramHasAttr(0, Attribute::StructRet); | ||||||
| 1879 | } | ||||||
| 1880 | |||||||
| 1881 | /// Determine if any call argument is an aggregate passed by value. | ||||||
| 1882 | bool hasByValArgument() const { | ||||||
| 1883 | return Attrs.hasAttrSomewhere(Attribute::ByVal); | ||||||
| 1884 | } | ||||||
| 1885 | |||||||
| 1886 | ///@{ | ||||||
| 1887 | // End of attribute API. | ||||||
| 1888 | |||||||
| 1889 | /// \name Operand Bundle API | ||||||
| 1890 | /// | ||||||
| 1891 | /// This group of methods provides the API to access and manipulate operand | ||||||
| 1892 | /// bundles on this call. | ||||||
| 1893 | /// @{ | ||||||
| 1894 | |||||||
| 1895 | /// Return the number of operand bundles associated with this User. | ||||||
| 1896 | unsigned getNumOperandBundles() const { | ||||||
| 1897 | return std::distance(bundle_op_info_begin(), bundle_op_info_end()); | ||||||
| 1898 | } | ||||||
| 1899 | |||||||
| 1900 | /// Return true if this User has any operand bundles. | ||||||
| 1901 | bool hasOperandBundles() const { return getNumOperandBundles() != 0; } | ||||||
| 1902 | |||||||
| 1903 | /// Return the index of the first bundle operand in the Use array. | ||||||
| 1904 | unsigned getBundleOperandsStartIndex() const { | ||||||
| 1905 | assert(hasOperandBundles() && "Don't call otherwise!")((void)0); | ||||||
| 1906 | return bundle_op_info_begin()->Begin; | ||||||
| 1907 | } | ||||||
| 1908 | |||||||
| 1909 | /// Return the index of the last bundle operand in the Use array. | ||||||
| 1910 | unsigned getBundleOperandsEndIndex() const { | ||||||
| 1911 | assert(hasOperandBundles() && "Don't call otherwise!")((void)0); | ||||||
| 1912 | return bundle_op_info_end()[-1].End; | ||||||
| 1913 | } | ||||||
| 1914 | |||||||
| 1915 | /// Return true if the operand at index \p Idx is a bundle operand. | ||||||
| 1916 | bool isBundleOperand(unsigned Idx) const { | ||||||
| 1917 | return hasOperandBundles() && Idx >= getBundleOperandsStartIndex() && | ||||||
| 1918 | Idx < getBundleOperandsEndIndex(); | ||||||
| 1919 | } | ||||||
| 1920 | |||||||
| 1921 | /// Returns true if the use is a bundle operand. | ||||||
| 1922 | bool isBundleOperand(const Use *U) const { | ||||||
| 1923 | assert(this == U->getUser() &&((void)0) | ||||||
| 1924 | "Only valid to query with a use of this instruction!")((void)0); | ||||||
| 1925 | return hasOperandBundles() && isBundleOperand(U - op_begin()); | ||||||
| 1926 | } | ||||||
| 1927 | bool isBundleOperand(Value::const_user_iterator UI) const { | ||||||
| 1928 | return isBundleOperand(&UI.getUse()); | ||||||
| 1929 | } | ||||||
| 1930 | |||||||
| 1931 | /// Return the total number operands (not operand bundles) used by | ||||||
| 1932 | /// every operand bundle in this OperandBundleUser. | ||||||
| 1933 | unsigned getNumTotalBundleOperands() const { | ||||||
| 1934 | if (!hasOperandBundles()) | ||||||
| 1935 | return 0; | ||||||
| 1936 | |||||||
| 1937 | unsigned Begin = getBundleOperandsStartIndex(); | ||||||
| 1938 | unsigned End = getBundleOperandsEndIndex(); | ||||||
| 1939 | |||||||
| 1940 | assert(Begin <= End && "Should be!")((void)0); | ||||||
| 1941 | return End - Begin; | ||||||
| 1942 | } | ||||||
| 1943 | |||||||
| 1944 | /// Return the operand bundle at a specific index. | ||||||
| 1945 | OperandBundleUse getOperandBundleAt(unsigned Index) const { | ||||||
| 1946 | assert(Index < getNumOperandBundles() && "Index out of bounds!")((void)0); | ||||||
| 1947 | return operandBundleFromBundleOpInfo(*(bundle_op_info_begin() + Index)); | ||||||
| 1948 | } | ||||||
| 1949 | |||||||
| 1950 | /// Return the number of operand bundles with the tag Name attached to | ||||||
| 1951 | /// this instruction. | ||||||
| 1952 | unsigned countOperandBundlesOfType(StringRef Name) const { | ||||||
| 1953 | unsigned Count = 0; | ||||||
| 1954 | for (unsigned i = 0, e = getNumOperandBundles(); i != e; ++i) | ||||||
| 1955 | if (getOperandBundleAt(i).getTagName() == Name) | ||||||
| 1956 | Count++; | ||||||
| 1957 | |||||||
| 1958 | return Count; | ||||||
| 1959 | } | ||||||
| 1960 | |||||||
| 1961 | /// Return the number of operand bundles with the tag ID attached to | ||||||
| 1962 | /// this instruction. | ||||||
| 1963 | unsigned countOperandBundlesOfType(uint32_t ID) const { | ||||||
| 1964 | unsigned Count = 0; | ||||||
| 1965 | for (unsigned i = 0, e = getNumOperandBundles(); i != e; ++i) | ||||||
| 1966 | if (getOperandBundleAt(i).getTagID() == ID) | ||||||
| 1967 | Count++; | ||||||
| 1968 | |||||||
| 1969 | return Count; | ||||||
| 1970 | } | ||||||
| 1971 | |||||||
| 1972 | /// Return an operand bundle by name, if present. | ||||||
| 1973 | /// | ||||||
| 1974 | /// It is an error to call this for operand bundle types that may have | ||||||
| 1975 | /// multiple instances of them on the same instruction. | ||||||
| 1976 | Optional<OperandBundleUse> getOperandBundle(StringRef Name) const { | ||||||
| 1977 | assert(countOperandBundlesOfType(Name) < 2 && "Precondition violated!")((void)0); | ||||||
| 1978 | |||||||
| 1979 | for (unsigned i = 0, e = getNumOperandBundles(); i != e; ++i) { | ||||||
| 1980 | OperandBundleUse U = getOperandBundleAt(i); | ||||||
| 1981 | if (U.getTagName() == Name) | ||||||
| 1982 | return U; | ||||||
| 1983 | } | ||||||
| 1984 | |||||||
| 1985 | return None; | ||||||
| 1986 | } | ||||||
| 1987 | |||||||
| 1988 | /// Return an operand bundle by tag ID, if present. | ||||||
| 1989 | /// | ||||||
| 1990 | /// It is an error to call this for operand bundle types that may have | ||||||
| 1991 | /// multiple instances of them on the same instruction. | ||||||
| 1992 | Optional<OperandBundleUse> getOperandBundle(uint32_t ID) const { | ||||||
| 1993 | assert(countOperandBundlesOfType(ID) < 2 && "Precondition violated!")((void)0); | ||||||
| 1994 | |||||||
| 1995 | for (unsigned i = 0, e = getNumOperandBundles(); i != e; ++i) { | ||||||
| 1996 | OperandBundleUse U = getOperandBundleAt(i); | ||||||
| 1997 | if (U.getTagID() == ID) | ||||||
| 1998 | return U; | ||||||
| 1999 | } | ||||||
| 2000 | |||||||
| 2001 | return None; | ||||||
| 2002 | } | ||||||
| 2003 | |||||||
| 2004 | /// Return the list of operand bundles attached to this instruction as | ||||||
| 2005 | /// a vector of OperandBundleDefs. | ||||||
| 2006 | /// | ||||||
| 2007 | /// This function copies the OperandBundeUse instances associated with this | ||||||
| 2008 | /// OperandBundleUser to a vector of OperandBundleDefs. Note: | ||||||
| 2009 | /// OperandBundeUses and OperandBundleDefs are non-trivially *different* | ||||||
| 2010 | /// representations of operand bundles (see documentation above). | ||||||
| 2011 | void getOperandBundlesAsDefs(SmallVectorImpl<OperandBundleDef> &Defs) const; | ||||||
| 2012 | |||||||
| 2013 | /// Return the operand bundle for the operand at index OpIdx. | ||||||
| 2014 | /// | ||||||
| 2015 | /// It is an error to call this with an OpIdx that does not correspond to an | ||||||
| 2016 | /// bundle operand. | ||||||
| 2017 | OperandBundleUse getOperandBundleForOperand(unsigned OpIdx) const { | ||||||
| 2018 | return operandBundleFromBundleOpInfo(getBundleOpInfoForOperand(OpIdx)); | ||||||
| 2019 | } | ||||||
| 2020 | |||||||
| 2021 | /// Return true if this operand bundle user has operand bundles that | ||||||
| 2022 | /// may read from the heap. | ||||||
| 2023 | bool hasReadingOperandBundles() const; | ||||||
| 2024 | |||||||
| 2025 | /// Return true if this operand bundle user has operand bundles that | ||||||
| 2026 | /// may write to the heap. | ||||||
| 2027 | bool hasClobberingOperandBundles() const { | ||||||
| 2028 | for (auto &BOI : bundle_op_infos()) { | ||||||
| 2029 | if (BOI.Tag->second == LLVMContext::OB_deopt || | ||||||
| 2030 | BOI.Tag->second == LLVMContext::OB_funclet) | ||||||
| 2031 | continue; | ||||||
| 2032 | |||||||
| 2033 | // This instruction has an operand bundle that is not known to us. | ||||||
| 2034 | // Assume the worst. | ||||||
| 2035 | return true; | ||||||
| 2036 | } | ||||||
| 2037 | |||||||
| 2038 | return false; | ||||||
| 2039 | } | ||||||
| 2040 | |||||||
| 2041 | /// Return true if the bundle operand at index \p OpIdx has the | ||||||
| 2042 | /// attribute \p A. | ||||||
| 2043 | bool bundleOperandHasAttr(unsigned OpIdx, Attribute::AttrKind A) const { | ||||||
| 2044 | auto &BOI = getBundleOpInfoForOperand(OpIdx); | ||||||
| 2045 | auto OBU = operandBundleFromBundleOpInfo(BOI); | ||||||
| 2046 | return OBU.operandHasAttr(OpIdx - BOI.Begin, A); | ||||||
| 2047 | } | ||||||
| 2048 | |||||||
| 2049 | /// Return true if \p Other has the same sequence of operand bundle | ||||||
| 2050 | /// tags with the same number of operands on each one of them as this | ||||||
| 2051 | /// OperandBundleUser. | ||||||
| 2052 | bool hasIdenticalOperandBundleSchema(const CallBase &Other) const { | ||||||
| 2053 | if (getNumOperandBundles() != Other.getNumOperandBundles()) | ||||||
| 2054 | return false; | ||||||
| 2055 | |||||||
| 2056 | return std::equal(bundle_op_info_begin(), bundle_op_info_end(), | ||||||
| 2057 | Other.bundle_op_info_begin()); | ||||||
| 2058 | } | ||||||
| 2059 | |||||||
| 2060 | /// Return true if this operand bundle user contains operand bundles | ||||||
| 2061 | /// with tags other than those specified in \p IDs. | ||||||
| 2062 | bool hasOperandBundlesOtherThan(ArrayRef<uint32_t> IDs) const { | ||||||
| 2063 | for (unsigned i = 0, e = getNumOperandBundles(); i != e; ++i) { | ||||||
| 2064 | uint32_t ID = getOperandBundleAt(i).getTagID(); | ||||||
| 2065 | if (!is_contained(IDs, ID)) | ||||||
| 2066 | return true; | ||||||
| 2067 | } | ||||||
| 2068 | return false; | ||||||
| 2069 | } | ||||||
| 2070 | |||||||
| 2071 | /// Is the function attribute S disallowed by some operand bundle on | ||||||
| 2072 | /// this operand bundle user? | ||||||
| 2073 | bool isFnAttrDisallowedByOpBundle(StringRef S) const { | ||||||
| 2074 | // Operand bundles only possibly disallow readnone, readonly and argmemonly | ||||||
| 2075 | // attributes. All String attributes are fine. | ||||||
| 2076 | return false; | ||||||
| 2077 | } | ||||||
| 2078 | |||||||
| 2079 | /// Is the function attribute A disallowed by some operand bundle on | ||||||
| 2080 | /// this operand bundle user? | ||||||
| 2081 | bool isFnAttrDisallowedByOpBundle(Attribute::AttrKind A) const { | ||||||
| 2082 | switch (A) { | ||||||
| 2083 | default: | ||||||
| 2084 | return false; | ||||||
| 2085 | |||||||
| 2086 | case Attribute::InaccessibleMemOrArgMemOnly: | ||||||
| 2087 | return hasReadingOperandBundles(); | ||||||
| 2088 | |||||||
| 2089 | case Attribute::InaccessibleMemOnly: | ||||||
| 2090 | return hasReadingOperandBundles(); | ||||||
| 2091 | |||||||
| 2092 | case Attribute::ArgMemOnly: | ||||||
| 2093 | return hasReadingOperandBundles(); | ||||||
| 2094 | |||||||
| 2095 | case Attribute::ReadNone: | ||||||
| 2096 | return hasReadingOperandBundles(); | ||||||
| 2097 | |||||||
| 2098 | case Attribute::ReadOnly: | ||||||
| 2099 | return hasClobberingOperandBundles(); | ||||||
| 2100 | } | ||||||
| 2101 | |||||||
| 2102 | llvm_unreachable("switch has a default case!")__builtin_unreachable(); | ||||||
| 2103 | } | ||||||
| 2104 | |||||||
| 2105 | /// Used to keep track of an operand bundle. See the main comment on | ||||||
| 2106 | /// OperandBundleUser above. | ||||||
| 2107 | struct BundleOpInfo { | ||||||
| 2108 | /// The operand bundle tag, interned by | ||||||
| 2109 | /// LLVMContextImpl::getOrInsertBundleTag. | ||||||
| 2110 | StringMapEntry<uint32_t> *Tag; | ||||||
| 2111 | |||||||
| 2112 | /// The index in the Use& vector where operands for this operand | ||||||
| 2113 | /// bundle starts. | ||||||
| 2114 | uint32_t Begin; | ||||||
| 2115 | |||||||
| 2116 | /// The index in the Use& vector where operands for this operand | ||||||
| 2117 | /// bundle ends. | ||||||
| 2118 | uint32_t End; | ||||||
| 2119 | |||||||
| 2120 | bool operator==(const BundleOpInfo &Other) const { | ||||||
| 2121 | return Tag == Other.Tag && Begin == Other.Begin && End == Other.End; | ||||||
| 2122 | } | ||||||
| 2123 | }; | ||||||
| 2124 | |||||||
| 2125 | /// Simple helper function to map a BundleOpInfo to an | ||||||
| 2126 | /// OperandBundleUse. | ||||||
| 2127 | OperandBundleUse | ||||||
| 2128 | operandBundleFromBundleOpInfo(const BundleOpInfo &BOI) const { | ||||||
| 2129 | auto begin = op_begin(); | ||||||
| 2130 | ArrayRef<Use> Inputs(begin + BOI.Begin, begin + BOI.End); | ||||||
| 2131 | return OperandBundleUse(BOI.Tag, Inputs); | ||||||
| 2132 | } | ||||||
| 2133 | |||||||
| 2134 | using bundle_op_iterator = BundleOpInfo *; | ||||||
| 2135 | using const_bundle_op_iterator = const BundleOpInfo *; | ||||||
| 2136 | |||||||
| 2137 | /// Return the start of the list of BundleOpInfo instances associated | ||||||
| 2138 | /// with this OperandBundleUser. | ||||||
| 2139 | /// | ||||||
| 2140 | /// OperandBundleUser uses the descriptor area co-allocated with the host User | ||||||
| 2141 | /// to store some meta information about which operands are "normal" operands, | ||||||
| 2142 | /// and which ones belong to some operand bundle. | ||||||
| 2143 | /// | ||||||
| 2144 | /// The layout of an operand bundle user is | ||||||
| 2145 | /// | ||||||
| 2146 | /// +-----------uint32_t End-------------------------------------+ | ||||||
| 2147 | /// | | | ||||||
| 2148 | /// | +--------uint32_t Begin--------------------+ | | ||||||
| 2149 | /// | | | | | ||||||
| 2150 | /// ^ ^ v v | ||||||
| 2151 | /// |------|------|----|----|----|----|----|---------|----|---------|----|----- | ||||||
| 2152 | /// | BOI0 | BOI1 | .. | DU | U0 | U1 | .. | BOI0_U0 | .. | BOI1_U0 | .. | Un | ||||||
| 2153 | /// |------|------|----|----|----|----|----|---------|----|---------|----|----- | ||||||
| 2154 | /// v v ^ ^ | ||||||
| 2155 | /// | | | | | ||||||
| 2156 | /// | +--------uint32_t Begin------------+ | | ||||||
| 2157 | /// | | | ||||||
| 2158 | /// +-----------uint32_t End-----------------------------+ | ||||||
| 2159 | /// | ||||||
| 2160 | /// | ||||||
| 2161 | /// BOI0, BOI1 ... are descriptions of operand bundles in this User's use | ||||||
| 2162 | /// list. These descriptions are installed and managed by this class, and | ||||||
| 2163 | /// they're all instances of OperandBundleUser<T>::BundleOpInfo. | ||||||
| 2164 | /// | ||||||
| 2165 | /// DU is an additional descriptor installed by User's 'operator new' to keep | ||||||
| 2166 | /// track of the 'BOI0 ... BOIN' co-allocation. OperandBundleUser does not | ||||||
| 2167 | /// access or modify DU in any way, it's an implementation detail private to | ||||||
| 2168 | /// User. | ||||||
| 2169 | /// | ||||||
| 2170 | /// The regular Use& vector for the User starts at U0. The operand bundle | ||||||
| 2171 | /// uses are part of the Use& vector, just like normal uses. In the diagram | ||||||
| 2172 | /// above, the operand bundle uses start at BOI0_U0. Each instance of | ||||||
| 2173 | /// BundleOpInfo has information about a contiguous set of uses constituting | ||||||
| 2174 | /// an operand bundle, and the total set of operand bundle uses themselves | ||||||
| 2175 | /// form a contiguous set of uses (i.e. there are no gaps between uses | ||||||
| 2176 | /// corresponding to individual operand bundles). | ||||||
| 2177 | /// | ||||||
| 2178 | /// This class does not know the location of the set of operand bundle uses | ||||||
| 2179 | /// within the use list -- that is decided by the User using this class via | ||||||
| 2180 | /// the BeginIdx argument in populateBundleOperandInfos. | ||||||
| 2181 | /// | ||||||
| 2182 | /// Currently operand bundle users with hung-off operands are not supported. | ||||||
| 2183 | bundle_op_iterator bundle_op_info_begin() { | ||||||
| 2184 | if (!hasDescriptor()) | ||||||
| 2185 | return nullptr; | ||||||
| 2186 | |||||||
| 2187 | uint8_t *BytesBegin = getDescriptor().begin(); | ||||||
| 2188 | return reinterpret_cast<bundle_op_iterator>(BytesBegin); | ||||||
| 2189 | } | ||||||
| 2190 | |||||||
| 2191 | /// Return the start of the list of BundleOpInfo instances associated | ||||||
| 2192 | /// with this OperandBundleUser. | ||||||
| 2193 | const_bundle_op_iterator bundle_op_info_begin() const { | ||||||
| 2194 | auto *NonConstThis = const_cast<CallBase *>(this); | ||||||
| 2195 | return NonConstThis->bundle_op_info_begin(); | ||||||
| 2196 | } | ||||||
| 2197 | |||||||
| 2198 | /// Return the end of the list of BundleOpInfo instances associated | ||||||
| 2199 | /// with this OperandBundleUser. | ||||||
| 2200 | bundle_op_iterator bundle_op_info_end() { | ||||||
| 2201 | if (!hasDescriptor()) | ||||||
| 2202 | return nullptr; | ||||||
| 2203 | |||||||
| 2204 | uint8_t *BytesEnd = getDescriptor().end(); | ||||||
| 2205 | return reinterpret_cast<bundle_op_iterator>(BytesEnd); | ||||||
| 2206 | } | ||||||
| 2207 | |||||||
| 2208 | /// Return the end of the list of BundleOpInfo instances associated | ||||||
| 2209 | /// with this OperandBundleUser. | ||||||
| 2210 | const_bundle_op_iterator bundle_op_info_end() const { | ||||||
| 2211 | auto *NonConstThis = const_cast<CallBase *>(this); | ||||||
| 2212 | return NonConstThis->bundle_op_info_end(); | ||||||
| 2213 | } | ||||||
| 2214 | |||||||
| 2215 | /// Return the range [\p bundle_op_info_begin, \p bundle_op_info_end). | ||||||
| 2216 | iterator_range<bundle_op_iterator> bundle_op_infos() { | ||||||
| 2217 | return make_range(bundle_op_info_begin(), bundle_op_info_end()); | ||||||
| 2218 | } | ||||||
| 2219 | |||||||
| 2220 | /// Return the range [\p bundle_op_info_begin, \p bundle_op_info_end). | ||||||
| 2221 | iterator_range<const_bundle_op_iterator> bundle_op_infos() const { | ||||||
| 2222 | return make_range(bundle_op_info_begin(), bundle_op_info_end()); | ||||||
| 2223 | } | ||||||
| 2224 | |||||||
| 2225 | /// Populate the BundleOpInfo instances and the Use& vector from \p | ||||||
| 2226 | /// Bundles. Return the op_iterator pointing to the Use& one past the last | ||||||
| 2227 | /// last bundle operand use. | ||||||
| 2228 | /// | ||||||
| 2229 | /// Each \p OperandBundleDef instance is tracked by a OperandBundleInfo | ||||||
| 2230 | /// instance allocated in this User's descriptor. | ||||||
| 2231 | op_iterator populateBundleOperandInfos(ArrayRef<OperandBundleDef> Bundles, | ||||||
| 2232 | const unsigned BeginIndex); | ||||||
| 2233 | |||||||
| 2234 | public: | ||||||
| 2235 | /// Return the BundleOpInfo for the operand at index OpIdx. | ||||||
| 2236 | /// | ||||||
| 2237 | /// It is an error to call this with an OpIdx that does not correspond to an | ||||||
| 2238 | /// bundle operand. | ||||||
| 2239 | BundleOpInfo &getBundleOpInfoForOperand(unsigned OpIdx); | ||||||
| 2240 | const BundleOpInfo &getBundleOpInfoForOperand(unsigned OpIdx) const { | ||||||
| 2241 | return const_cast<CallBase *>(this)->getBundleOpInfoForOperand(OpIdx); | ||||||
| 2242 | } | ||||||
| 2243 | |||||||
| 2244 | protected: | ||||||
| 2245 | /// Return the total number of values used in \p Bundles. | ||||||
| 2246 | static unsigned CountBundleInputs(ArrayRef<OperandBundleDef> Bundles) { | ||||||
| 2247 | unsigned Total = 0; | ||||||
| 2248 | for (auto &B : Bundles) | ||||||
| 2249 | Total += B.input_size(); | ||||||
| 2250 | return Total; | ||||||
| 2251 | } | ||||||
| 2252 | |||||||
| 2253 | /// @} | ||||||
| 2254 | // End of operand bundle API. | ||||||
| 2255 | |||||||
| 2256 | private: | ||||||
| 2257 | bool hasFnAttrOnCalledFunction(Attribute::AttrKind Kind) const; | ||||||
| 2258 | bool hasFnAttrOnCalledFunction(StringRef Kind) const; | ||||||
| 2259 | |||||||
| 2260 | template <typename AttrKind> bool hasFnAttrImpl(AttrKind Kind) const { | ||||||
| 2261 | if (Attrs.hasFnAttribute(Kind)) | ||||||
| 2262 | return true; | ||||||
| 2263 | |||||||
| 2264 | // Operand bundles override attributes on the called function, but don't | ||||||
| 2265 | // override attributes directly present on the call instruction. | ||||||
| 2266 | if (isFnAttrDisallowedByOpBundle(Kind)) | ||||||
| 2267 | return false; | ||||||
| 2268 | |||||||
| 2269 | return hasFnAttrOnCalledFunction(Kind); | ||||||
| 2270 | } | ||||||
| 2271 | |||||||
| 2272 | /// Determine whether the return value has the given attribute. Supports | ||||||
| 2273 | /// Attribute::AttrKind and StringRef as \p AttrKind types. | ||||||
| 2274 | template <typename AttrKind> bool hasRetAttrImpl(AttrKind Kind) const { | ||||||
| 2275 | if (Attrs.hasAttribute(AttributeList::ReturnIndex, Kind)) | ||||||
| 2276 | return true; | ||||||
| 2277 | |||||||
| 2278 | // Look at the callee, if available. | ||||||
| 2279 | if (const Function *F = getCalledFunction()) | ||||||
| 2280 | return F->getAttributes().hasAttribute(AttributeList::ReturnIndex, Kind); | ||||||
| 2281 | return false; | ||||||
| 2282 | } | ||||||
| 2283 | }; | ||||||
| 2284 | |||||||
| 2285 | template <> | ||||||
| 2286 | struct OperandTraits<CallBase> : public VariadicOperandTraits<CallBase, 1> {}; | ||||||
| 2287 | |||||||
| 2288 | DEFINE_TRANSPARENT_OPERAND_ACCESSORS(CallBase, Value)CallBase::op_iterator CallBase::op_begin() { return OperandTraits <CallBase>::op_begin(this); } CallBase::const_op_iterator CallBase::op_begin() const { return OperandTraits<CallBase >::op_begin(const_cast<CallBase*>(this)); } CallBase ::op_iterator CallBase::op_end() { return OperandTraits<CallBase >::op_end(this); } CallBase::const_op_iterator CallBase::op_end () const { return OperandTraits<CallBase>::op_end(const_cast <CallBase*>(this)); } Value *CallBase::getOperand(unsigned i_nocapture) const { ((void)0); return cast_or_null<Value >( OperandTraits<CallBase>::op_begin(const_cast<CallBase *>(this))[i_nocapture].get()); } void CallBase::setOperand (unsigned i_nocapture, Value *Val_nocapture) { ((void)0); OperandTraits <CallBase>::op_begin(this)[i_nocapture] = Val_nocapture ; } unsigned CallBase::getNumOperands() const { return OperandTraits <CallBase>::operands(this); } template <int Idx_nocapture > Use &CallBase::Op() { return this->OpFrom<Idx_nocapture >(this); } template <int Idx_nocapture> const Use & CallBase::Op() const { return this->OpFrom<Idx_nocapture >(this); } | ||||||
| 2289 | |||||||
| 2290 | //===----------------------------------------------------------------------===// | ||||||
| 2291 | // FuncletPadInst Class | ||||||
| 2292 | //===----------------------------------------------------------------------===// | ||||||
| 2293 | class FuncletPadInst : public Instruction { | ||||||
| 2294 | private: | ||||||
| 2295 | FuncletPadInst(const FuncletPadInst &CPI); | ||||||
| 2296 | |||||||
| 2297 | explicit FuncletPadInst(Instruction::FuncletPadOps Op, Value *ParentPad, | ||||||
| 2298 | ArrayRef<Value *> Args, unsigned Values, | ||||||
| 2299 | const Twine &NameStr, Instruction *InsertBefore); | ||||||
| 2300 | explicit FuncletPadInst(Instruction::FuncletPadOps Op, Value *ParentPad, | ||||||
| 2301 | ArrayRef<Value *> Args, unsigned Values, | ||||||
| 2302 | const Twine &NameStr, BasicBlock *InsertAtEnd); | ||||||
| 2303 | |||||||
| 2304 | void init(Value *ParentPad, ArrayRef<Value *> Args, const Twine &NameStr); | ||||||
| 2305 | |||||||
| 2306 | protected: | ||||||
| 2307 | // Note: Instruction needs to be a friend here to call cloneImpl. | ||||||
| 2308 | friend class Instruction; | ||||||
| 2309 | friend class CatchPadInst; | ||||||
| 2310 | friend class CleanupPadInst; | ||||||
| 2311 | |||||||
| 2312 | FuncletPadInst *cloneImpl() const; | ||||||
| 2313 | |||||||
| 2314 | public: | ||||||
| 2315 | /// Provide fast operand accessors | ||||||
| 2316 | DECLARE_TRANSPARENT_OPERAND_ACCESSORS(Value)public: inline Value *getOperand(unsigned) const; inline void setOperand(unsigned, Value*); inline op_iterator op_begin(); inline const_op_iterator op_begin() const; inline op_iterator op_end(); inline const_op_iterator op_end() const; protected : template <int> inline Use &Op(); template <int > inline const Use &Op() const; public: inline unsigned getNumOperands() const; | ||||||
| 2317 | |||||||
| 2318 | /// getNumArgOperands - Return the number of funcletpad arguments. | ||||||
| 2319 | /// | ||||||
| 2320 | unsigned getNumArgOperands() const { return getNumOperands() - 1; } | ||||||
| 2321 | |||||||
| 2322 | /// Convenience accessors | ||||||
| 2323 | |||||||
| 2324 | /// Return the outer EH-pad this funclet is nested within. | ||||||
| 2325 | /// | ||||||
| 2326 | /// Note: This returns the associated CatchSwitchInst if this FuncletPadInst | ||||||
| 2327 | /// is a CatchPadInst. | ||||||
| 2328 | Value *getParentPad() const { return Op<-1>(); } | ||||||
| 2329 | void setParentPad(Value *ParentPad) { | ||||||
| 2330 | assert(ParentPad)((void)0); | ||||||
| 2331 | Op<-1>() = ParentPad; | ||||||
| 2332 | } | ||||||
| 2333 | |||||||
| 2334 | /// getArgOperand/setArgOperand - Return/set the i-th funcletpad argument. | ||||||
| 2335 | /// | ||||||
| 2336 | Value *getArgOperand(unsigned i) const { return getOperand(i); } | ||||||
| 2337 | void setArgOperand(unsigned i, Value *v) { setOperand(i, v); } | ||||||
| 2338 | |||||||
| 2339 | /// arg_operands - iteration adapter for range-for loops. | ||||||
| 2340 | op_range arg_operands() { return op_range(op_begin(), op_end() - 1); } | ||||||
| 2341 | |||||||
| 2342 | /// arg_operands - iteration adapter for range-for loops. | ||||||
| 2343 | const_op_range arg_operands() const { | ||||||
| 2344 | return const_op_range(op_begin(), op_end() - 1); | ||||||
| 2345 | } | ||||||
| 2346 | |||||||
| 2347 | // Methods for support type inquiry through isa, cast, and dyn_cast: | ||||||
| 2348 | static bool classof(const Instruction *I) { return I->isFuncletPad(); } | ||||||
| 2349 | static bool classof(const Value *V) { | ||||||
| 2350 | return isa<Instruction>(V) && classof(cast<Instruction>(V)); | ||||||
| 2351 | } | ||||||
| 2352 | }; | ||||||
| 2353 | |||||||
| 2354 | template <> | ||||||
| 2355 | struct OperandTraits<FuncletPadInst> | ||||||
| 2356 | : public VariadicOperandTraits<FuncletPadInst, /*MINARITY=*/1> {}; | ||||||
| 2357 | |||||||
| 2358 | DEFINE_TRANSPARENT_OPERAND_ACCESSORS(FuncletPadInst, Value)FuncletPadInst::op_iterator FuncletPadInst::op_begin() { return OperandTraits<FuncletPadInst>::op_begin(this); } FuncletPadInst ::const_op_iterator FuncletPadInst::op_begin() const { return OperandTraits<FuncletPadInst>::op_begin(const_cast< FuncletPadInst*>(this)); } FuncletPadInst::op_iterator FuncletPadInst ::op_end() { return OperandTraits<FuncletPadInst>::op_end (this); } FuncletPadInst::const_op_iterator FuncletPadInst::op_end () const { return OperandTraits<FuncletPadInst>::op_end (const_cast<FuncletPadInst*>(this)); } Value *FuncletPadInst ::getOperand(unsigned i_nocapture) const { ((void)0); return cast_or_null <Value>( OperandTraits<FuncletPadInst>::op_begin( const_cast<FuncletPadInst*>(this))[i_nocapture].get()); } void FuncletPadInst::setOperand(unsigned i_nocapture, Value *Val_nocapture) { ((void)0); OperandTraits<FuncletPadInst >::op_begin(this)[i_nocapture] = Val_nocapture; } unsigned FuncletPadInst::getNumOperands() const { return OperandTraits <FuncletPadInst>::operands(this); } template <int Idx_nocapture > Use &FuncletPadInst::Op() { return this->OpFrom< Idx_nocapture>(this); } template <int Idx_nocapture> const Use &FuncletPadInst::Op() const { return this-> OpFrom<Idx_nocapture>(this); } | ||||||
| 2359 | |||||||
| 2360 | } // end namespace llvm | ||||||
| 2361 | |||||||
| 2362 | #endif // LLVM_IR_INSTRTYPES_H |
| 1 | //===- PatternMatch.h - Match on the LLVM IR --------------------*- C++ -*-===// | ||||||
| 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 provides a simple and efficient mechanism for performing general | ||||||
| 10 | // tree-based pattern matches on the LLVM IR. The power of these routines is | ||||||
| 11 | // that it allows you to write concise patterns that are expressive and easy to | ||||||
| 12 | // understand. The other major advantage of this is that it allows you to | ||||||
| 13 | // trivially capture/bind elements in the pattern to variables. For example, | ||||||
| 14 | // you can do something like this: | ||||||
| 15 | // | ||||||
| 16 | // Value *Exp = ... | ||||||
| 17 | // Value *X, *Y; ConstantInt *C1, *C2; // (X & C1) | (Y & C2) | ||||||
| 18 | // if (match(Exp, m_Or(m_And(m_Value(X), m_ConstantInt(C1)), | ||||||
| 19 | // m_And(m_Value(Y), m_ConstantInt(C2))))) { | ||||||
| 20 | // ... Pattern is matched and variables are bound ... | ||||||
| 21 | // } | ||||||
| 22 | // | ||||||
| 23 | // This is primarily useful to things like the instruction combiner, but can | ||||||
| 24 | // also be useful for static analysis tools or code generators. | ||||||
| 25 | // | ||||||
| 26 | //===----------------------------------------------------------------------===// | ||||||
| 27 | |||||||
| 28 | #ifndef LLVM_IR_PATTERNMATCH_H | ||||||
| 29 | #define LLVM_IR_PATTERNMATCH_H | ||||||
| 30 | |||||||
| 31 | #include "llvm/ADT/APFloat.h" | ||||||
| 32 | #include "llvm/ADT/APInt.h" | ||||||
| 33 | #include "llvm/IR/Constant.h" | ||||||
| 34 | #include "llvm/IR/Constants.h" | ||||||
| 35 | #include "llvm/IR/DataLayout.h" | ||||||
| 36 | #include "llvm/IR/InstrTypes.h" | ||||||
| 37 | #include "llvm/IR/Instruction.h" | ||||||
| 38 | #include "llvm/IR/Instructions.h" | ||||||
| 39 | #include "llvm/IR/IntrinsicInst.h" | ||||||
| 40 | #include "llvm/IR/Intrinsics.h" | ||||||
| 41 | #include "llvm/IR/Operator.h" | ||||||
| 42 | #include "llvm/IR/Value.h" | ||||||
| 43 | #include "llvm/Support/Casting.h" | ||||||
| 44 | #include <cstdint> | ||||||
| 45 | |||||||
| 46 | namespace llvm { | ||||||
| 47 | namespace PatternMatch { | ||||||
| 48 | |||||||
| 49 | template <typename Val, typename Pattern> bool match(Val *V, const Pattern &P) { | ||||||
| 50 | return const_cast<Pattern &>(P).match(V); | ||||||
| 51 | } | ||||||
| 52 | |||||||
| 53 | template <typename Pattern> bool match(ArrayRef<int> Mask, const Pattern &P) { | ||||||
| 54 | return const_cast<Pattern &>(P).match(Mask); | ||||||
| 55 | } | ||||||
| 56 | |||||||
| 57 | template <typename SubPattern_t> struct OneUse_match { | ||||||
| 58 | SubPattern_t SubPattern; | ||||||
| 59 | |||||||
| 60 | OneUse_match(const SubPattern_t &SP) : SubPattern(SP) {} | ||||||
| 61 | |||||||
| 62 | template <typename OpTy> bool match(OpTy *V) { | ||||||
| 63 | return V->hasOneUse() && SubPattern.match(V); | ||||||
| 64 | } | ||||||
| 65 | }; | ||||||
| 66 | |||||||
| 67 | template <typename T> inline OneUse_match<T> m_OneUse(const T &SubPattern) { | ||||||
| 68 | return SubPattern; | ||||||
| 69 | } | ||||||
| 70 | |||||||
| 71 | template <typename Class> struct class_match { | ||||||
| 72 | template <typename ITy> bool match(ITy *V) { return isa<Class>(V); } | ||||||
| 73 | }; | ||||||
| 74 | |||||||
| 75 | /// Match an arbitrary value and ignore it. | ||||||
| 76 | inline class_match<Value> m_Value() { return class_match<Value>(); } | ||||||
| 77 | |||||||
| 78 | /// Match an arbitrary unary operation and ignore it. | ||||||
| 79 | inline class_match<UnaryOperator> m_UnOp() { | ||||||
| 80 | return class_match<UnaryOperator>(); | ||||||
| 81 | } | ||||||
| 82 | |||||||
| 83 | /// Match an arbitrary binary operation and ignore it. | ||||||
| 84 | inline class_match<BinaryOperator> m_BinOp() { | ||||||
| 85 | return class_match<BinaryOperator>(); | ||||||
| 86 | } | ||||||
| 87 | |||||||
| 88 | /// Matches any compare instruction and ignore it. | ||||||
| 89 | inline class_match<CmpInst> m_Cmp() { return class_match<CmpInst>(); } | ||||||
| 90 | |||||||
| 91 | struct undef_match { | ||||||
| 92 | static bool check(const Value *V) { | ||||||
| 93 | if (isa<UndefValue>(V)) | ||||||
| 94 | return true; | ||||||
| 95 | |||||||
| 96 | const auto *CA = dyn_cast<ConstantAggregate>(V); | ||||||
| 97 | if (!CA) | ||||||
| 98 | return false; | ||||||
| 99 | |||||||
| 100 | SmallPtrSet<const ConstantAggregate *, 8> Seen; | ||||||
| 101 | SmallVector<const ConstantAggregate *, 8> Worklist; | ||||||
| 102 | |||||||
| 103 | // Either UndefValue, PoisonValue, or an aggregate that only contains | ||||||
| 104 | // these is accepted by matcher. | ||||||
| 105 | // CheckValue returns false if CA cannot satisfy this constraint. | ||||||
| 106 | auto CheckValue = [&](const ConstantAggregate *CA) { | ||||||
| 107 | for (const Value *Op : CA->operand_values()) { | ||||||
| 108 | if (isa<UndefValue>(Op)) | ||||||
| 109 | continue; | ||||||
| 110 | |||||||
| 111 | const auto *CA = dyn_cast<ConstantAggregate>(Op); | ||||||
| 112 | if (!CA) | ||||||
| 113 | return false; | ||||||
| 114 | if (Seen.insert(CA).second) | ||||||
| 115 | Worklist.emplace_back(CA); | ||||||
| 116 | } | ||||||
| 117 | |||||||
| 118 | return true; | ||||||
| 119 | }; | ||||||
| 120 | |||||||
| 121 | if (!CheckValue(CA)) | ||||||
| 122 | return false; | ||||||
| 123 | |||||||
| 124 | while (!Worklist.empty()) { | ||||||
| 125 | if (!CheckValue(Worklist.pop_back_val())) | ||||||
| 126 | return false; | ||||||
| 127 | } | ||||||
| 128 | return true; | ||||||
| 129 | } | ||||||
| 130 | template <typename ITy> bool match(ITy *V) { return check(V); } | ||||||
| 131 | }; | ||||||
| 132 | |||||||
| 133 | /// Match an arbitrary undef constant. This matches poison as well. | ||||||
| 134 | /// If this is an aggregate and contains a non-aggregate element that is | ||||||
| 135 | /// neither undef nor poison, the aggregate is not matched. | ||||||
| 136 | inline auto m_Undef() { return undef_match(); } | ||||||
| 137 | |||||||
| 138 | /// Match an arbitrary poison constant. | ||||||
| 139 | inline class_match<PoisonValue> m_Poison() { return class_match<PoisonValue>(); } | ||||||
| 140 | |||||||
| 141 | /// Match an arbitrary Constant and ignore it. | ||||||
| 142 | inline class_match<Constant> m_Constant() { return class_match<Constant>(); } | ||||||
| 143 | |||||||
| 144 | /// Match an arbitrary ConstantInt and ignore it. | ||||||
| 145 | inline class_match<ConstantInt> m_ConstantInt() { | ||||||
| 146 | return class_match<ConstantInt>(); | ||||||
| 147 | } | ||||||
| 148 | |||||||
| 149 | /// Match an arbitrary ConstantFP and ignore it. | ||||||
| 150 | inline class_match<ConstantFP> m_ConstantFP() { | ||||||
| 151 | return class_match<ConstantFP>(); | ||||||
| 152 | } | ||||||
| 153 | |||||||
| 154 | /// Match an arbitrary ConstantExpr and ignore it. | ||||||
| 155 | inline class_match<ConstantExpr> m_ConstantExpr() { | ||||||
| 156 | return class_match<ConstantExpr>(); | ||||||
| 157 | } | ||||||
| 158 | |||||||
| 159 | /// Match an arbitrary basic block value and ignore it. | ||||||
| 160 | inline class_match<BasicBlock> m_BasicBlock() { | ||||||
| 161 | return class_match<BasicBlock>(); | ||||||
| 162 | } | ||||||
| 163 | |||||||
| 164 | /// Inverting matcher | ||||||
| 165 | template <typename Ty> struct match_unless { | ||||||
| 166 | Ty M; | ||||||
| 167 | |||||||
| 168 | match_unless(const Ty &Matcher) : M(Matcher) {} | ||||||
| 169 | |||||||
| 170 | template <typename ITy> bool match(ITy *V) { return !M.match(V); } | ||||||
| 171 | }; | ||||||
| 172 | |||||||
| 173 | /// Match if the inner matcher does *NOT* match. | ||||||
| 174 | template <typename Ty> inline match_unless<Ty> m_Unless(const Ty &M) { | ||||||
| 175 | return match_unless<Ty>(M); | ||||||
| 176 | } | ||||||
| 177 | |||||||
| 178 | /// Matching combinators | ||||||
| 179 | template <typename LTy, typename RTy> struct match_combine_or { | ||||||
| 180 | LTy L; | ||||||
| 181 | RTy R; | ||||||
| 182 | |||||||
| 183 | match_combine_or(const LTy &Left, const RTy &Right) : L(Left), R(Right) {} | ||||||
| 184 | |||||||
| 185 | template <typename ITy> bool match(ITy *V) { | ||||||
| 186 | if (L.match(V)) | ||||||
| 187 | return true; | ||||||
| 188 | if (R.match(V)) | ||||||
| 189 | return true; | ||||||
| 190 | return false; | ||||||
| 191 | } | ||||||
| 192 | }; | ||||||
| 193 | |||||||
| 194 | template <typename LTy, typename RTy> struct match_combine_and { | ||||||
| 195 | LTy L; | ||||||
| 196 | RTy R; | ||||||
| 197 | |||||||
| 198 | match_combine_and(const LTy &Left, const RTy &Right) : L(Left), R(Right) {} | ||||||
| 199 | |||||||
| 200 | template <typename ITy> bool match(ITy *V) { | ||||||
| 201 | if (L.match(V)) | ||||||
| 202 | if (R.match(V)) | ||||||
| 203 | return true; | ||||||
| 204 | return false; | ||||||
| 205 | } | ||||||
| 206 | }; | ||||||
| 207 | |||||||
| 208 | /// Combine two pattern matchers matching L || R | ||||||
| 209 | template <typename LTy, typename RTy> | ||||||
| 210 | inline match_combine_or<LTy, RTy> m_CombineOr(const LTy &L, const RTy &R) { | ||||||
| 211 | return match_combine_or<LTy, RTy>(L, R); | ||||||
| 212 | } | ||||||
| 213 | |||||||
| 214 | /// Combine two pattern matchers matching L && R | ||||||
| 215 | template <typename LTy, typename RTy> | ||||||
| 216 | inline match_combine_and<LTy, RTy> m_CombineAnd(const LTy &L, const RTy &R) { | ||||||
| 217 | return match_combine_and<LTy, RTy>(L, R); | ||||||
| 218 | } | ||||||
| 219 | |||||||
| 220 | struct apint_match { | ||||||
| 221 | const APInt *&Res; | ||||||
| 222 | bool AllowUndef; | ||||||
| 223 | |||||||
| 224 | apint_match(const APInt *&Res, bool AllowUndef) | ||||||
| 225 | : Res(Res), AllowUndef(AllowUndef) {} | ||||||
| 226 | |||||||
| 227 | template <typename ITy> bool match(ITy *V) { | ||||||
| 228 | if (auto *CI = dyn_cast<ConstantInt>(V)) { | ||||||
| 229 | Res = &CI->getValue(); | ||||||
| 230 | return true; | ||||||
| 231 | } | ||||||
| 232 | if (V->getType()->isVectorTy()) | ||||||
| |||||||
| 233 | if (const auto *C = dyn_cast<Constant>(V)) | ||||||
| 234 | if (auto *CI = dyn_cast_or_null<ConstantInt>( | ||||||
| 235 | C->getSplatValue(AllowUndef))) { | ||||||
| 236 | Res = &CI->getValue(); | ||||||
| 237 | return true; | ||||||
| 238 | } | ||||||
| 239 | return false; | ||||||
| 240 | } | ||||||
| 241 | }; | ||||||
| 242 | // Either constexpr if or renaming ConstantFP::getValueAPF to | ||||||
| 243 | // ConstantFP::getValue is needed to do it via single template | ||||||
| 244 | // function for both apint/apfloat. | ||||||
| 245 | struct apfloat_match { | ||||||
| 246 | const APFloat *&Res; | ||||||
| 247 | bool AllowUndef; | ||||||
| 248 | |||||||
| 249 | apfloat_match(const APFloat *&Res, bool AllowUndef) | ||||||
| 250 | : Res(Res), AllowUndef(AllowUndef) {} | ||||||
| 251 | |||||||
| 252 | template <typename ITy> bool match(ITy *V) { | ||||||
| 253 | if (auto *CI = dyn_cast<ConstantFP>(V)) { | ||||||
| 254 | Res = &CI->getValueAPF(); | ||||||
| 255 | return true; | ||||||
| 256 | } | ||||||
| 257 | if (V->getType()->isVectorTy()) | ||||||
| 258 | if (const auto *C = dyn_cast<Constant>(V)) | ||||||
| 259 | if (auto *CI = dyn_cast_or_null<ConstantFP>( | ||||||
| 260 | C->getSplatValue(AllowUndef))) { | ||||||
| 261 | Res = &CI->getValueAPF(); | ||||||
| 262 | return true; | ||||||
| 263 | } | ||||||
| 264 | return false; | ||||||
| 265 | } | ||||||
| 266 | }; | ||||||
| 267 | |||||||
| 268 | /// Match a ConstantInt or splatted ConstantVector, binding the | ||||||
| 269 | /// specified pointer to the contained APInt. | ||||||
| 270 | inline apint_match m_APInt(const APInt *&Res) { | ||||||
| 271 | // Forbid undefs by default to maintain previous behavior. | ||||||
| 272 | return apint_match(Res, /* AllowUndef */ false); | ||||||
| 273 | } | ||||||
| 274 | |||||||
| 275 | /// Match APInt while allowing undefs in splat vector constants. | ||||||
| 276 | inline apint_match m_APIntAllowUndef(const APInt *&Res) { | ||||||
| 277 | return apint_match(Res, /* AllowUndef */ true); | ||||||
| 278 | } | ||||||
| 279 | |||||||
| 280 | /// Match APInt while forbidding undefs in splat vector constants. | ||||||
| 281 | inline apint_match m_APIntForbidUndef(const APInt *&Res) { | ||||||
| 282 | return apint_match(Res, /* AllowUndef */ false); | ||||||
| 283 | } | ||||||
| 284 | |||||||
| 285 | /// Match a ConstantFP or splatted ConstantVector, binding the | ||||||
| 286 | /// specified pointer to the contained APFloat. | ||||||
| 287 | inline apfloat_match m_APFloat(const APFloat *&Res) { | ||||||
| 288 | // Forbid undefs by default to maintain previous behavior. | ||||||
| 289 | return apfloat_match(Res, /* AllowUndef */ false); | ||||||
| 290 | } | ||||||
| 291 | |||||||
| 292 | /// Match APFloat while allowing undefs in splat vector constants. | ||||||
| 293 | inline apfloat_match m_APFloatAllowUndef(const APFloat *&Res) { | ||||||
| 294 | return apfloat_match(Res, /* AllowUndef */ true); | ||||||
| 295 | } | ||||||
| 296 | |||||||
| 297 | /// Match APFloat while forbidding undefs in splat vector constants. | ||||||
| 298 | inline apfloat_match m_APFloatForbidUndef(const APFloat *&Res) { | ||||||
| 299 | return apfloat_match(Res, /* AllowUndef */ false); | ||||||
| 300 | } | ||||||
| 301 | |||||||
| 302 | template <int64_t Val> struct constantint_match { | ||||||
| 303 | template <typename ITy> bool match(ITy *V) { | ||||||
| 304 | if (const auto *CI = dyn_cast<ConstantInt>(V)) { | ||||||
| 305 | const APInt &CIV = CI->getValue(); | ||||||
| 306 | if (Val >= 0) | ||||||
| 307 | return CIV == static_cast<uint64_t>(Val); | ||||||
| 308 | // If Val is negative, and CI is shorter than it, truncate to the right | ||||||
| 309 | // number of bits. If it is larger, then we have to sign extend. Just | ||||||
| 310 | // compare their negated values. | ||||||
| 311 | return -CIV == -Val; | ||||||
| 312 | } | ||||||
| 313 | return false; | ||||||
| 314 | } | ||||||
| 315 | }; | ||||||
| 316 | |||||||
| 317 | /// Match a ConstantInt with a specific value. | ||||||
| 318 | template <int64_t Val> inline constantint_match<Val> m_ConstantInt() { | ||||||
| 319 | return constantint_match<Val>(); | ||||||
| 320 | } | ||||||
| 321 | |||||||
| 322 | /// This helper class is used to match constant scalars, vector splats, | ||||||
| 323 | /// and fixed width vectors that satisfy a specified predicate. | ||||||
| 324 | /// For fixed width vector constants, undefined elements are ignored. | ||||||
| 325 | template <typename Predicate, typename ConstantVal> | ||||||
| 326 | struct cstval_pred_ty : public Predicate { | ||||||
| 327 | template <typename ITy> bool match(ITy *V) { | ||||||
| 328 | if (const auto *CV = dyn_cast<ConstantVal>(V)) | ||||||
| 329 | return this->isValue(CV->getValue()); | ||||||
| 330 | if (const auto *VTy = dyn_cast<VectorType>(V->getType())) { | ||||||
| 331 | if (const auto *C = dyn_cast<Constant>(V)) { | ||||||
| 332 | if (const auto *CV = dyn_cast_or_null<ConstantVal>(C->getSplatValue())) | ||||||
| 333 | return this->isValue(CV->getValue()); | ||||||
| 334 | |||||||
| 335 | // Number of elements of a scalable vector unknown at compile time | ||||||
| 336 | auto *FVTy = dyn_cast<FixedVectorType>(VTy); | ||||||
| 337 | if (!FVTy) | ||||||
| 338 | return false; | ||||||
| 339 | |||||||
| 340 | // Non-splat vector constant: check each element for a match. | ||||||
| 341 | unsigned NumElts = FVTy->getNumElements(); | ||||||
| 342 | assert(NumElts != 0 && "Constant vector with no elements?")((void)0); | ||||||
| 343 | bool HasNonUndefElements = false; | ||||||
| 344 | for (unsigned i = 0; i != NumElts; ++i) { | ||||||
| 345 | Constant *Elt = C->getAggregateElement(i); | ||||||
| 346 | if (!Elt) | ||||||
| 347 | return false; | ||||||
| 348 | if (isa<UndefValue>(Elt)) | ||||||
| 349 | continue; | ||||||
| 350 | auto *CV = dyn_cast<ConstantVal>(Elt); | ||||||
| 351 | if (!CV || !this->isValue(CV->getValue())) | ||||||
| 352 | return false; | ||||||
| 353 | HasNonUndefElements = true; | ||||||
| 354 | } | ||||||
| 355 | return HasNonUndefElements; | ||||||
| 356 | } | ||||||
| 357 | } | ||||||
| 358 | return false; | ||||||
| 359 | } | ||||||
| 360 | }; | ||||||
| 361 | |||||||
| 362 | /// specialization of cstval_pred_ty for ConstantInt | ||||||
| 363 | template <typename Predicate> | ||||||
| 364 | using cst_pred_ty = cstval_pred_ty<Predicate, ConstantInt>; | ||||||
| 365 | |||||||
| 366 | /// specialization of cstval_pred_ty for ConstantFP | ||||||
| 367 | template <typename Predicate> | ||||||
| 368 | using cstfp_pred_ty = cstval_pred_ty<Predicate, ConstantFP>; | ||||||
| 369 | |||||||
| 370 | /// This helper class is used to match scalar and vector constants that | ||||||
| 371 | /// satisfy a specified predicate, and bind them to an APInt. | ||||||
| 372 | template <typename Predicate> struct api_pred_ty : public Predicate { | ||||||
| 373 | const APInt *&Res; | ||||||
| 374 | |||||||
| 375 | api_pred_ty(const APInt *&R) : Res(R) {} | ||||||
| 376 | |||||||
| 377 | template <typename ITy> bool match(ITy *V) { | ||||||
| 378 | if (const auto *CI = dyn_cast<ConstantInt>(V)) | ||||||
| 379 | if (this->isValue(CI->getValue())) { | ||||||
| 380 | Res = &CI->getValue(); | ||||||
| 381 | return true; | ||||||
| 382 | } | ||||||
| 383 | if (V->getType()->isVectorTy()) | ||||||
| 384 | if (const auto *C = dyn_cast<Constant>(V)) | ||||||
| 385 | if (auto *CI = dyn_cast_or_null<ConstantInt>(C->getSplatValue())) | ||||||
| 386 | if (this->isValue(CI->getValue())) { | ||||||
| 387 | Res = &CI->getValue(); | ||||||
| 388 | return true; | ||||||
| 389 | } | ||||||
| 390 | |||||||
| 391 | return false; | ||||||
| 392 | } | ||||||
| 393 | }; | ||||||
| 394 | |||||||
| 395 | /// This helper class is used to match scalar and vector constants that | ||||||
| 396 | /// satisfy a specified predicate, and bind them to an APFloat. | ||||||
| 397 | /// Undefs are allowed in splat vector constants. | ||||||
| 398 | template <typename Predicate> struct apf_pred_ty : public Predicate { | ||||||
| 399 | const APFloat *&Res; | ||||||
| 400 | |||||||
| 401 | apf_pred_ty(const APFloat *&R) : Res(R) {} | ||||||
| 402 | |||||||
| 403 | template <typename ITy> bool match(ITy *V) { | ||||||
| 404 | if (const auto *CI = dyn_cast<ConstantFP>(V)) | ||||||
| 405 | if (this->isValue(CI->getValue())) { | ||||||
| 406 | Res = &CI->getValue(); | ||||||
| 407 | return true; | ||||||
| 408 | } | ||||||
| 409 | if (V->getType()->isVectorTy()) | ||||||
| 410 | if (const auto *C = dyn_cast<Constant>(V)) | ||||||
| 411 | if (auto *CI = dyn_cast_or_null<ConstantFP>( | ||||||
| 412 | C->getSplatValue(/* AllowUndef */ true))) | ||||||
| 413 | if (this->isValue(CI->getValue())) { | ||||||
| 414 | Res = &CI->getValue(); | ||||||
| 415 | return true; | ||||||
| 416 | } | ||||||
| 417 | |||||||
| 418 | return false; | ||||||
| 419 | } | ||||||
| 420 | }; | ||||||
| 421 | |||||||
| 422 | /////////////////////////////////////////////////////////////////////////////// | ||||||
| 423 | // | ||||||
| 424 | // Encapsulate constant value queries for use in templated predicate matchers. | ||||||
| 425 | // This allows checking if constants match using compound predicates and works | ||||||
| 426 | // with vector constants, possibly with relaxed constraints. For example, ignore | ||||||
| 427 | // undef values. | ||||||
| 428 | // | ||||||
| 429 | /////////////////////////////////////////////////////////////////////////////// | ||||||
| 430 | |||||||
| 431 | struct is_any_apint { | ||||||
| 432 | bool isValue(const APInt &C) { return true; } | ||||||
| 433 | }; | ||||||
| 434 | /// Match an integer or vector with any integral constant. | ||||||
| 435 | /// For vectors, this includes constants with undefined elements. | ||||||
| 436 | inline cst_pred_ty<is_any_apint> m_AnyIntegralConstant() { | ||||||
| 437 | return cst_pred_ty<is_any_apint>(); | ||||||
| 438 | } | ||||||
| 439 | |||||||
| 440 | struct is_all_ones { | ||||||
| 441 | bool isValue(const APInt &C) { return C.isAllOnesValue(); } | ||||||
| 442 | }; | ||||||
| 443 | /// Match an integer or vector with all bits set. | ||||||
| 444 | /// For vectors, this includes constants with undefined elements. | ||||||
| 445 | inline cst_pred_ty<is_all_ones> m_AllOnes() { | ||||||
| 446 | return cst_pred_ty<is_all_ones>(); | ||||||
| 447 | } | ||||||
| 448 | |||||||
| 449 | struct is_maxsignedvalue { | ||||||
| 450 | bool isValue(const APInt &C) { return C.isMaxSignedValue(); } | ||||||
| 451 | }; | ||||||
| 452 | /// Match an integer or vector with values having all bits except for the high | ||||||
| 453 | /// bit set (0x7f...). | ||||||
| 454 | /// For vectors, this includes constants with undefined elements. | ||||||
| 455 | inline cst_pred_ty<is_maxsignedvalue> m_MaxSignedValue() { | ||||||
| 456 | return cst_pred_ty<is_maxsignedvalue>(); | ||||||
| 457 | } | ||||||
| 458 | inline api_pred_ty<is_maxsignedvalue> m_MaxSignedValue(const APInt *&V) { | ||||||
| 459 | return V; | ||||||
| 460 | } | ||||||
| 461 | |||||||
| 462 | struct is_negative { | ||||||
| 463 | bool isValue(const APInt &C) { return C.isNegative(); } | ||||||
| 464 | }; | ||||||
| 465 | /// Match an integer or vector of negative values. | ||||||
| 466 | /// For vectors, this includes constants with undefined elements. | ||||||
| 467 | inline cst_pred_ty<is_negative> m_Negative() { | ||||||
| 468 | return cst_pred_ty<is_negative>(); | ||||||
| 469 | } | ||||||
| 470 | inline api_pred_ty<is_negative> m_Negative(const APInt *&V) { | ||||||
| 471 | return V; | ||||||
| 472 | } | ||||||
| 473 | |||||||
| 474 | struct is_nonnegative { | ||||||
| 475 | bool isValue(const APInt &C) { return C.isNonNegative(); } | ||||||
| 476 | }; | ||||||
| 477 | /// Match an integer or vector of non-negative values. | ||||||
| 478 | /// For vectors, this includes constants with undefined elements. | ||||||
| 479 | inline cst_pred_ty<is_nonnegative> m_NonNegative() { | ||||||
| 480 | return cst_pred_ty<is_nonnegative>(); | ||||||
| 481 | } | ||||||
| 482 | inline api_pred_ty<is_nonnegative> m_NonNegative(const APInt *&V) { | ||||||
| 483 | return V; | ||||||
| 484 | } | ||||||
| 485 | |||||||
| 486 | struct is_strictlypositive { | ||||||
| 487 | bool isValue(const APInt &C) { return C.isStrictlyPositive(); } | ||||||
| 488 | }; | ||||||
| 489 | /// Match an integer or vector of strictly positive values. | ||||||
| 490 | /// For vectors, this includes constants with undefined elements. | ||||||
| 491 | inline cst_pred_ty<is_strictlypositive> m_StrictlyPositive() { | ||||||
| 492 | return cst_pred_ty<is_strictlypositive>(); | ||||||
| 493 | } | ||||||
| 494 | inline api_pred_ty<is_strictlypositive> m_StrictlyPositive(const APInt *&V) { | ||||||
| 495 | return V; | ||||||
| 496 | } | ||||||
| 497 | |||||||
| 498 | struct is_nonpositive { | ||||||
| 499 | bool isValue(const APInt &C) { return C.isNonPositive(); } | ||||||
| 500 | }; | ||||||
| 501 | /// Match an integer or vector of non-positive values. | ||||||
| 502 | /// For vectors, this includes constants with undefined elements. | ||||||
| 503 | inline cst_pred_ty<is_nonpositive> m_NonPositive() { | ||||||
| 504 | return cst_pred_ty<is_nonpositive>(); | ||||||
| 505 | } | ||||||
| 506 | inline api_pred_ty<is_nonpositive> m_NonPositive(const APInt *&V) { return V; } | ||||||
| 507 | |||||||
| 508 | struct is_one { | ||||||
| 509 | bool isValue(const APInt &C) { return C.isOneValue(); } | ||||||
| 510 | }; | ||||||
| 511 | /// Match an integer 1 or a vector with all elements equal to 1. | ||||||
| 512 | /// For vectors, this includes constants with undefined elements. | ||||||
| 513 | inline cst_pred_ty<is_one> m_One() { | ||||||
| 514 | return cst_pred_ty<is_one>(); | ||||||
| 515 | } | ||||||
| 516 | |||||||
| 517 | struct is_zero_int { | ||||||
| 518 | bool isValue(const APInt &C) { return C.isNullValue(); } | ||||||
| 519 | }; | ||||||
| 520 | /// Match an integer 0 or a vector with all elements equal to 0. | ||||||
| 521 | /// For vectors, this includes constants with undefined elements. | ||||||
| 522 | inline cst_pred_ty<is_zero_int> m_ZeroInt() { | ||||||
| 523 | return cst_pred_ty<is_zero_int>(); | ||||||
| 524 | } | ||||||
| 525 | |||||||
| 526 | struct is_zero { | ||||||
| 527 | template <typename ITy> bool match(ITy *V) { | ||||||
| 528 | auto *C = dyn_cast<Constant>(V); | ||||||
| 529 | // FIXME: this should be able to do something for scalable vectors | ||||||
| 530 | return C && (C->isNullValue() || cst_pred_ty<is_zero_int>().match(C)); | ||||||
| 531 | } | ||||||
| 532 | }; | ||||||
| 533 | /// Match any null constant or a vector with all elements equal to 0. | ||||||
| 534 | /// For vectors, this includes constants with undefined elements. | ||||||
| 535 | inline is_zero m_Zero() { | ||||||
| 536 | return is_zero(); | ||||||
| 537 | } | ||||||
| 538 | |||||||
| 539 | struct is_power2 { | ||||||
| 540 | bool isValue(const APInt &C) { return C.isPowerOf2(); } | ||||||
| 541 | }; | ||||||
| 542 | /// Match an integer or vector power-of-2. | ||||||
| 543 | /// For vectors, this includes constants with undefined elements. | ||||||
| 544 | inline cst_pred_ty<is_power2> m_Power2() { | ||||||
| 545 | return cst_pred_ty<is_power2>(); | ||||||
| 546 | } | ||||||
| 547 | inline api_pred_ty<is_power2> m_Power2(const APInt *&V) { | ||||||
| 548 | return V; | ||||||
| 549 | } | ||||||
| 550 | |||||||
| 551 | struct is_negated_power2 { | ||||||
| 552 | bool isValue(const APInt &C) { return (-C).isPowerOf2(); } | ||||||
| 553 | }; | ||||||
| 554 | /// Match a integer or vector negated power-of-2. | ||||||
| 555 | /// For vectors, this includes constants with undefined elements. | ||||||
| 556 | inline cst_pred_ty<is_negated_power2> m_NegatedPower2() { | ||||||
| 557 | return cst_pred_ty<is_negated_power2>(); | ||||||
| 558 | } | ||||||
| 559 | inline api_pred_ty<is_negated_power2> m_NegatedPower2(const APInt *&V) { | ||||||
| 560 | return V; | ||||||
| 561 | } | ||||||
| 562 | |||||||
| 563 | struct is_power2_or_zero { | ||||||
| 564 | bool isValue(const APInt &C) { return !C || C.isPowerOf2(); } | ||||||
| 565 | }; | ||||||
| 566 | /// Match an integer or vector of 0 or power-of-2 values. | ||||||
| 567 | /// For vectors, this includes constants with undefined elements. | ||||||
| 568 | inline cst_pred_ty<is_power2_or_zero> m_Power2OrZero() { | ||||||
| 569 | return cst_pred_ty<is_power2_or_zero>(); | ||||||
| 570 | } | ||||||
| 571 | inline api_pred_ty<is_power2_or_zero> m_Power2OrZero(const APInt *&V) { | ||||||
| 572 | return V; | ||||||
| 573 | } | ||||||
| 574 | |||||||
| 575 | struct is_sign_mask { | ||||||
| 576 | bool isValue(const APInt &C) { return C.isSignMask(); } | ||||||
| 577 | }; | ||||||
| 578 | /// Match an integer or vector with only the sign bit(s) set. | ||||||
| 579 | /// For vectors, this includes constants with undefined elements. | ||||||
| 580 | inline cst_pred_ty<is_sign_mask> m_SignMask() { | ||||||
| 581 | return cst_pred_ty<is_sign_mask>(); | ||||||
| 582 | } | ||||||
| 583 | |||||||
| 584 | struct is_lowbit_mask { | ||||||
| 585 | bool isValue(const APInt &C) { return C.isMask(); } | ||||||
| 586 | }; | ||||||
| 587 | /// Match an integer or vector with only the low bit(s) set. | ||||||
| 588 | /// For vectors, this includes constants with undefined elements. | ||||||
| 589 | inline cst_pred_ty<is_lowbit_mask> m_LowBitMask() { | ||||||
| 590 | return cst_pred_ty<is_lowbit_mask>(); | ||||||
| 591 | } | ||||||
| 592 | |||||||
| 593 | struct icmp_pred_with_threshold { | ||||||
| 594 | ICmpInst::Predicate Pred; | ||||||
| 595 | const APInt *Thr; | ||||||
| 596 | bool isValue(const APInt &C) { | ||||||
| 597 | switch (Pred) { | ||||||
| 598 | case ICmpInst::Predicate::ICMP_EQ: | ||||||
| 599 | return C.eq(*Thr); | ||||||
| 600 | case ICmpInst::Predicate::ICMP_NE: | ||||||
| 601 | return C.ne(*Thr); | ||||||
| 602 | case ICmpInst::Predicate::ICMP_UGT: | ||||||
| 603 | return C.ugt(*Thr); | ||||||
| 604 | case ICmpInst::Predicate::ICMP_UGE: | ||||||
| 605 | return C.uge(*Thr); | ||||||
| 606 | case ICmpInst::Predicate::ICMP_ULT: | ||||||
| 607 | return C.ult(*Thr); | ||||||
| 608 | case ICmpInst::Predicate::ICMP_ULE: | ||||||
| 609 | return C.ule(*Thr); | ||||||
| 610 | case ICmpInst::Predicate::ICMP_SGT: | ||||||
| 611 | return C.sgt(*Thr); | ||||||
| 612 | case ICmpInst::Predicate::ICMP_SGE: | ||||||
| 613 | return C.sge(*Thr); | ||||||
| 614 | case ICmpInst::Predicate::ICMP_SLT: | ||||||
| 615 | return C.slt(*Thr); | ||||||
| 616 | case ICmpInst::Predicate::ICMP_SLE: | ||||||
| 617 | return C.sle(*Thr); | ||||||
| 618 | default: | ||||||
| 619 | llvm_unreachable("Unhandled ICmp predicate")__builtin_unreachable(); | ||||||
| 620 | } | ||||||
| 621 | } | ||||||
| 622 | }; | ||||||
| 623 | /// Match an integer or vector with every element comparing 'pred' (eg/ne/...) | ||||||
| 624 | /// to Threshold. For vectors, this includes constants with undefined elements. | ||||||
| 625 | inline cst_pred_ty<icmp_pred_with_threshold> | ||||||
| 626 | m_SpecificInt_ICMP(ICmpInst::Predicate Predicate, const APInt &Threshold) { | ||||||
| 627 | cst_pred_ty<icmp_pred_with_threshold> P; | ||||||
| 628 | P.Pred = Predicate; | ||||||
| 629 | P.Thr = &Threshold; | ||||||
| 630 | return P; | ||||||
| 631 | } | ||||||
| 632 | |||||||
| 633 | struct is_nan { | ||||||
| 634 | bool isValue(const APFloat &C) { return C.isNaN(); } | ||||||
| 635 | }; | ||||||
| 636 | /// Match an arbitrary NaN constant. This includes quiet and signalling nans. | ||||||
| 637 | /// For vectors, this includes constants with undefined elements. | ||||||
| 638 | inline cstfp_pred_ty<is_nan> m_NaN() { | ||||||
| 639 | return cstfp_pred_ty<is_nan>(); | ||||||
| 640 | } | ||||||
| 641 | |||||||
| 642 | struct is_nonnan { | ||||||
| 643 | bool isValue(const APFloat &C) { return !C.isNaN(); } | ||||||
| 644 | }; | ||||||
| 645 | /// Match a non-NaN FP constant. | ||||||
| 646 | /// For vectors, this includes constants with undefined elements. | ||||||
| 647 | inline cstfp_pred_ty<is_nonnan> m_NonNaN() { | ||||||
| 648 | return cstfp_pred_ty<is_nonnan>(); | ||||||
| 649 | } | ||||||
| 650 | |||||||
| 651 | struct is_inf { | ||||||
| 652 | bool isValue(const APFloat &C) { return C.isInfinity(); } | ||||||
| 653 | }; | ||||||
| 654 | /// Match a positive or negative infinity FP constant. | ||||||
| 655 | /// For vectors, this includes constants with undefined elements. | ||||||
| 656 | inline cstfp_pred_ty<is_inf> m_Inf() { | ||||||
| 657 | return cstfp_pred_ty<is_inf>(); | ||||||
| 658 | } | ||||||
| 659 | |||||||
| 660 | struct is_noninf { | ||||||
| 661 | bool isValue(const APFloat &C) { return !C.isInfinity(); } | ||||||
| 662 | }; | ||||||
| 663 | /// Match a non-infinity FP constant, i.e. finite or NaN. | ||||||
| 664 | /// For vectors, this includes constants with undefined elements. | ||||||
| 665 | inline cstfp_pred_ty<is_noninf> m_NonInf() { | ||||||
| 666 | return cstfp_pred_ty<is_noninf>(); | ||||||
| 667 | } | ||||||
| 668 | |||||||
| 669 | struct is_finite { | ||||||
| 670 | bool isValue(const APFloat &C) { return C.isFinite(); } | ||||||
| 671 | }; | ||||||
| 672 | /// Match a finite FP constant, i.e. not infinity or NaN. | ||||||
| 673 | /// For vectors, this includes constants with undefined elements. | ||||||
| 674 | inline cstfp_pred_ty<is_finite> m_Finite() { | ||||||
| 675 | return cstfp_pred_ty<is_finite>(); | ||||||
| 676 | } | ||||||
| 677 | inline apf_pred_ty<is_finite> m_Finite(const APFloat *&V) { return V; } | ||||||
| 678 | |||||||
| 679 | struct is_finitenonzero { | ||||||
| 680 | bool isValue(const APFloat &C) { return C.isFiniteNonZero(); } | ||||||
| 681 | }; | ||||||
| 682 | /// Match a finite non-zero FP constant. | ||||||
| 683 | /// For vectors, this includes constants with undefined elements. | ||||||
| 684 | inline cstfp_pred_ty<is_finitenonzero> m_FiniteNonZero() { | ||||||
| 685 | return cstfp_pred_ty<is_finitenonzero>(); | ||||||
| 686 | } | ||||||
| 687 | inline apf_pred_ty<is_finitenonzero> m_FiniteNonZero(const APFloat *&V) { | ||||||
| 688 | return V; | ||||||
| 689 | } | ||||||
| 690 | |||||||
| 691 | struct is_any_zero_fp { | ||||||
| 692 | bool isValue(const APFloat &C) { return C.isZero(); } | ||||||
| 693 | }; | ||||||
| 694 | /// Match a floating-point negative zero or positive zero. | ||||||
| 695 | /// For vectors, this includes constants with undefined elements. | ||||||
| 696 | inline cstfp_pred_ty<is_any_zero_fp> m_AnyZeroFP() { | ||||||
| 697 | return cstfp_pred_ty<is_any_zero_fp>(); | ||||||
| 698 | } | ||||||
| 699 | |||||||
| 700 | struct is_pos_zero_fp { | ||||||
| 701 | bool isValue(const APFloat &C) { return C.isPosZero(); } | ||||||
| 702 | }; | ||||||
| 703 | /// Match a floating-point positive zero. | ||||||
| 704 | /// For vectors, this includes constants with undefined elements. | ||||||
| 705 | inline cstfp_pred_ty<is_pos_zero_fp> m_PosZeroFP() { | ||||||
| 706 | return cstfp_pred_ty<is_pos_zero_fp>(); | ||||||
| 707 | } | ||||||
| 708 | |||||||
| 709 | struct is_neg_zero_fp { | ||||||
| 710 | bool isValue(const APFloat &C) { return C.isNegZero(); } | ||||||
| 711 | }; | ||||||
| 712 | /// Match a floating-point negative zero. | ||||||
| 713 | /// For vectors, this includes constants with undefined elements. | ||||||
| 714 | inline cstfp_pred_ty<is_neg_zero_fp> m_NegZeroFP() { | ||||||
| 715 | return cstfp_pred_ty<is_neg_zero_fp>(); | ||||||
| 716 | } | ||||||
| 717 | |||||||
| 718 | struct is_non_zero_fp { | ||||||
| 719 | bool isValue(const APFloat &C) { return C.isNonZero(); } | ||||||
| 720 | }; | ||||||
| 721 | /// Match a floating-point non-zero. | ||||||
| 722 | /// For vectors, this includes constants with undefined elements. | ||||||
| 723 | inline cstfp_pred_ty<is_non_zero_fp> m_NonZeroFP() { | ||||||
| 724 | return cstfp_pred_ty<is_non_zero_fp>(); | ||||||
| 725 | } | ||||||
| 726 | |||||||
| 727 | /////////////////////////////////////////////////////////////////////////////// | ||||||
| 728 | |||||||
| 729 | template <typename Class> struct bind_ty { | ||||||
| 730 | Class *&VR; | ||||||
| 731 | |||||||
| 732 | bind_ty(Class *&V) : VR(V) {} | ||||||
| 733 | |||||||
| 734 | template <typename ITy> bool match(ITy *V) { | ||||||
| 735 | if (auto *CV = dyn_cast<Class>(V)) { | ||||||
| 736 | VR = CV; | ||||||
| 737 | return true; | ||||||
| 738 | } | ||||||
| 739 | return false; | ||||||
| 740 | } | ||||||
| 741 | }; | ||||||
| 742 | |||||||
| 743 | /// Match a value, capturing it if we match. | ||||||
| 744 | inline bind_ty<Value> m_Value(Value *&V) { return V; } | ||||||
| 745 | inline bind_ty<const Value> m_Value(const Value *&V) { return V; } | ||||||
| 746 | |||||||
| 747 | /// Match an instruction, capturing it if we match. | ||||||
| 748 | inline bind_ty<Instruction> m_Instruction(Instruction *&I) { return I; } | ||||||
| 749 | /// Match a unary operator, capturing it if we match. | ||||||
| 750 | inline bind_ty<UnaryOperator> m_UnOp(UnaryOperator *&I) { return I; } | ||||||
| 751 | /// Match a binary operator, capturing it if we match. | ||||||
| 752 | inline bind_ty<BinaryOperator> m_BinOp(BinaryOperator *&I) { return I; } | ||||||
| 753 | /// Match a with overflow intrinsic, capturing it if we match. | ||||||
| 754 | inline bind_ty<WithOverflowInst> m_WithOverflowInst(WithOverflowInst *&I) { return I; } | ||||||
| 755 | inline bind_ty<const WithOverflowInst> | ||||||
| 756 | m_WithOverflowInst(const WithOverflowInst *&I) { | ||||||
| 757 | return I; | ||||||
| 758 | } | ||||||
| 759 | |||||||
| 760 | /// Match a Constant, capturing the value if we match. | ||||||
| 761 | inline bind_ty<Constant> m_Constant(Constant *&C) { return C; } | ||||||
| 762 | |||||||
| 763 | /// Match a ConstantInt, capturing the value if we match. | ||||||
| 764 | inline bind_ty<ConstantInt> m_ConstantInt(ConstantInt *&CI) { return CI; } | ||||||
| 765 | |||||||
| 766 | /// Match a ConstantFP, capturing the value if we match. | ||||||
| 767 | inline bind_ty<ConstantFP> m_ConstantFP(ConstantFP *&C) { return C; } | ||||||
| 768 | |||||||
| 769 | /// Match a ConstantExpr, capturing the value if we match. | ||||||
| 770 | inline bind_ty<ConstantExpr> m_ConstantExpr(ConstantExpr *&C) { return C; } | ||||||
| 771 | |||||||
| 772 | /// Match a basic block value, capturing it if we match. | ||||||
| 773 | inline bind_ty<BasicBlock> m_BasicBlock(BasicBlock *&V) { return V; } | ||||||
| 774 | inline bind_ty<const BasicBlock> m_BasicBlock(const BasicBlock *&V) { | ||||||
| 775 | return V; | ||||||
| 776 | } | ||||||
| 777 | |||||||
| 778 | /// Match an arbitrary immediate Constant and ignore it. | ||||||
| 779 | inline match_combine_and<class_match<Constant>, | ||||||
| 780 | match_unless<class_match<ConstantExpr>>> | ||||||
| 781 | m_ImmConstant() { | ||||||
| 782 | return m_CombineAnd(m_Constant(), m_Unless(m_ConstantExpr())); | ||||||
| 783 | } | ||||||
| 784 | |||||||
| 785 | /// Match an immediate Constant, capturing the value if we match. | ||||||
| 786 | inline match_combine_and<bind_ty<Constant>, | ||||||
| 787 | match_unless<class_match<ConstantExpr>>> | ||||||
| 788 | m_ImmConstant(Constant *&C) { | ||||||
| 789 | return m_CombineAnd(m_Constant(C), m_Unless(m_ConstantExpr())); | ||||||
| 790 | } | ||||||
| 791 | |||||||
| 792 | /// Match a specified Value*. | ||||||
| 793 | struct specificval_ty { | ||||||
| 794 | const Value *Val; | ||||||
| 795 | |||||||
| 796 | specificval_ty(const Value *V) : Val(V) {} | ||||||
| 797 | |||||||
| 798 | template <typename ITy> bool match(ITy *V) { return V
| ||||||
| 799 | }; | ||||||
| 800 | |||||||
| 801 | /// Match if we have a specific specified value. | ||||||
| 802 | inline specificval_ty m_Specific(const Value *V) { return V; } | ||||||
| 803 | |||||||
| 804 | /// Stores a reference to the Value *, not the Value * itself, | ||||||
| 805 | /// thus can be used in commutative matchers. | ||||||
| 806 | template <typename Class> struct deferredval_ty { | ||||||
| 807 | Class *const &Val; | ||||||
| 808 | |||||||
| 809 | deferredval_ty(Class *const &V) : Val(V) {} | ||||||
| 810 | |||||||
| 811 | template <typename ITy> bool match(ITy *const V) { return V == Val; } | ||||||
| 812 | }; | ||||||
| 813 | |||||||
| 814 | /// Like m_Specific(), but works if the specific value to match is determined | ||||||
| 815 | /// as part of the same match() expression. For example: | ||||||
| 816 | /// m_Add(m_Value(X), m_Specific(X)) is incorrect, because m_Specific() will | ||||||
| 817 | /// bind X before the pattern match starts. | ||||||
| 818 | /// m_Add(m_Value(X), m_Deferred(X)) is correct, and will check against | ||||||
| 819 | /// whichever value m_Value(X) populated. | ||||||
| 820 | inline deferredval_ty<Value> m_Deferred(Value *const &V) { return V; } | ||||||
| 821 | inline deferredval_ty<const Value> m_Deferred(const Value *const &V) { | ||||||
| 822 | return V; | ||||||
| 823 | } | ||||||
| 824 | |||||||
| 825 | /// Match a specified floating point value or vector of all elements of | ||||||
| 826 | /// that value. | ||||||
| 827 | struct specific_fpval { | ||||||
| 828 | double Val; | ||||||
| 829 | |||||||
| 830 | specific_fpval(double V) : Val(V) {} | ||||||
| 831 | |||||||
| 832 | template <typename ITy> bool match(ITy *V) { | ||||||
| 833 | if (const auto *CFP = dyn_cast<ConstantFP>(V)) | ||||||
| 834 | return CFP->isExactlyValue(Val); | ||||||
| 835 | if (V->getType()->isVectorTy()) | ||||||
| 836 | if (const auto *C = dyn_cast<Constant>(V)) | ||||||
| 837 | if (auto *CFP = dyn_cast_or_null<ConstantFP>(C->getSplatValue())) | ||||||
| 838 | return CFP->isExactlyValue(Val); | ||||||
| 839 | return false; | ||||||
| 840 | } | ||||||
| 841 | }; | ||||||
| 842 | |||||||
| 843 | /// Match a specific floating point value or vector with all elements | ||||||
| 844 | /// equal to the value. | ||||||
| 845 | inline specific_fpval m_SpecificFP(double V) { return specific_fpval(V); } | ||||||
| 846 | |||||||
| 847 | /// Match a float 1.0 or vector with all elements equal to 1.0. | ||||||
| 848 | inline specific_fpval m_FPOne() { return m_SpecificFP(1.0); } | ||||||
| 849 | |||||||
| 850 | struct bind_const_intval_ty { | ||||||
| 851 | uint64_t &VR; | ||||||
| 852 | |||||||
| 853 | bind_const_intval_ty(uint64_t &V) : VR(V) {} | ||||||
| 854 | |||||||
| 855 | template <typename ITy> bool match(ITy *V) { | ||||||
| 856 | if (const auto *CV = dyn_cast<ConstantInt>(V)) | ||||||
| 857 | if (CV->getValue().ule(UINT64_MAX0xffffffffffffffffULL)) { | ||||||
| 858 | VR = CV->getZExtValue(); | ||||||
| 859 | return true; | ||||||
| 860 | } | ||||||
| 861 | return false; | ||||||
| 862 | } | ||||||
| 863 | }; | ||||||
| 864 | |||||||
| 865 | /// Match a specified integer value or vector of all elements of that | ||||||
| 866 | /// value. | ||||||
| 867 | template <bool AllowUndefs> | ||||||
| 868 | struct specific_intval { | ||||||
| 869 | APInt Val; | ||||||
| 870 | |||||||
| 871 | specific_intval(APInt V) : Val(std::move(V)) {} | ||||||
| 872 | |||||||
| 873 | template <typename ITy> bool match(ITy *V) { | ||||||
| 874 | const auto *CI = dyn_cast<ConstantInt>(V); | ||||||
| 875 | if (!CI && V->getType()->isVectorTy()) | ||||||
| 876 | if (const auto *C = dyn_cast<Constant>(V)) | ||||||
| 877 | CI = dyn_cast_or_null<ConstantInt>(C->getSplatValue(AllowUndefs)); | ||||||
| 878 | |||||||
| 879 | return CI && APInt::isSameValue(CI->getValue(), Val); | ||||||
| 880 | } | ||||||
| 881 | }; | ||||||
| 882 | |||||||
| 883 | /// Match a specific integer value or vector with all elements equal to | ||||||
| 884 | /// the value. | ||||||
| 885 | inline specific_intval<false> m_SpecificInt(APInt V) { | ||||||
| 886 | return specific_intval<false>(std::move(V)); | ||||||
| 887 | } | ||||||
| 888 | |||||||
| 889 | inline specific_intval<false> m_SpecificInt(uint64_t V) { | ||||||
| 890 | return m_SpecificInt(APInt(64, V)); | ||||||
| 891 | } | ||||||
| 892 | |||||||
| 893 | inline specific_intval<true> m_SpecificIntAllowUndef(APInt V) { | ||||||
| 894 | return specific_intval<true>(std::move(V)); | ||||||
| 895 | } | ||||||
| 896 | |||||||
| 897 | inline specific_intval<true> m_SpecificIntAllowUndef(uint64_t V) { | ||||||
| 898 | return m_SpecificIntAllowUndef(APInt(64, V)); | ||||||
| 899 | } | ||||||
| 900 | |||||||
| 901 | /// Match a ConstantInt and bind to its value. This does not match | ||||||
| 902 | /// ConstantInts wider than 64-bits. | ||||||
| 903 | inline bind_const_intval_ty m_ConstantInt(uint64_t &V) { return V; } | ||||||
| 904 | |||||||
| 905 | /// Match a specified basic block value. | ||||||
| 906 | struct specific_bbval { | ||||||
| 907 | BasicBlock *Val; | ||||||
| 908 | |||||||
| 909 | specific_bbval(BasicBlock *Val) : Val(Val) {} | ||||||
| 910 | |||||||
| 911 | template <typename ITy> bool match(ITy *V) { | ||||||
| 912 | const auto *BB = dyn_cast<BasicBlock>(V); | ||||||
| 913 | return BB && BB == Val; | ||||||
| 914 | } | ||||||
| 915 | }; | ||||||
| 916 | |||||||
| 917 | /// Match a specific basic block value. | ||||||
| 918 | inline specific_bbval m_SpecificBB(BasicBlock *BB) { | ||||||
| 919 | return specific_bbval(BB); | ||||||
| 920 | } | ||||||
| 921 | |||||||
| 922 | /// A commutative-friendly version of m_Specific(). | ||||||
| 923 | inline deferredval_ty<BasicBlock> m_Deferred(BasicBlock *const &BB) { | ||||||
| 924 | return BB; | ||||||
| 925 | } | ||||||
| 926 | inline deferredval_ty<const BasicBlock> | ||||||
| 927 | m_Deferred(const BasicBlock *const &BB) { | ||||||
| 928 | return BB; | ||||||
| 929 | } | ||||||
| 930 | |||||||
| 931 | //===----------------------------------------------------------------------===// | ||||||
| 932 | // Matcher for any binary operator. | ||||||
| 933 | // | ||||||
| 934 | template <typename LHS_t, typename RHS_t, bool Commutable = false> | ||||||
| 935 | struct AnyBinaryOp_match { | ||||||
| 936 | LHS_t L; | ||||||
| 937 | RHS_t R; | ||||||
| 938 | |||||||
| 939 | // The evaluation order is always stable, regardless of Commutability. | ||||||
| 940 | // The LHS is always matched first. | ||||||
| 941 | AnyBinaryOp_match(const LHS_t &LHS, const RHS_t &RHS) : L(LHS), R(RHS) {} | ||||||
| 942 | |||||||
| 943 | template <typename OpTy> bool match(OpTy *V) { | ||||||
| 944 | if (auto *I = dyn_cast<BinaryOperator>(V)) | ||||||
| 945 | return (L.match(I->getOperand(0)) && R.match(I->getOperand(1))) || | ||||||
| 946 | (Commutable && L.match(I->getOperand(1)) && | ||||||
| 947 | R.match(I->getOperand(0))); | ||||||
| 948 | return false; | ||||||
| 949 | } | ||||||
| 950 | }; | ||||||
| 951 | |||||||
| 952 | template <typename LHS, typename RHS> | ||||||
| 953 | inline AnyBinaryOp_match<LHS, RHS> m_BinOp(const LHS &L, const RHS &R) { | ||||||
| 954 | return AnyBinaryOp_match<LHS, RHS>(L, R); | ||||||
| 955 | } | ||||||
| 956 | |||||||
| 957 | //===----------------------------------------------------------------------===// | ||||||
| 958 | // Matcher for any unary operator. | ||||||
| 959 | // TODO fuse unary, binary matcher into n-ary matcher | ||||||
| 960 | // | ||||||
| 961 | template <typename OP_t> struct AnyUnaryOp_match { | ||||||
| 962 | OP_t X; | ||||||
| 963 | |||||||
| 964 | AnyUnaryOp_match(const OP_t &X) : X(X) {} | ||||||
| 965 | |||||||
| 966 | template <typename OpTy> bool match(OpTy *V) { | ||||||
| 967 | if (auto *I = dyn_cast<UnaryOperator>(V)) | ||||||
| 968 | return X.match(I->getOperand(0)); | ||||||
| 969 | return false; | ||||||
| 970 | } | ||||||
| 971 | }; | ||||||
| 972 | |||||||
| 973 | template <typename OP_t> inline AnyUnaryOp_match<OP_t> m_UnOp(const OP_t &X) { | ||||||
| 974 | return AnyUnaryOp_match<OP_t>(X); | ||||||
| 975 | } | ||||||
| 976 | |||||||
| 977 | //===----------------------------------------------------------------------===// | ||||||
| 978 | // Matchers for specific binary operators. | ||||||
| 979 | // | ||||||
| 980 | |||||||
| 981 | template <typename LHS_t, typename RHS_t, unsigned Opcode, | ||||||
| 982 | bool Commutable = false> | ||||||
| 983 | struct BinaryOp_match { | ||||||
| 984 | LHS_t L; | ||||||
| 985 | RHS_t R; | ||||||
| 986 | |||||||
| 987 | // The evaluation order is always stable, regardless of Commutability. | ||||||
| 988 | // The LHS is always matched first. | ||||||
| 989 | BinaryOp_match(const LHS_t &LHS, const RHS_t &RHS) : L(LHS), R(RHS) {} | ||||||
| 990 | |||||||
| 991 | template <typename OpTy> bool match(OpTy *V) { | ||||||
| 992 | if (V->getValueID() == Value::InstructionVal + Opcode) { | ||||||
| 993 | auto *I = cast<BinaryOperator>(V); | ||||||
| 994 | return (L.match(I->getOperand(0)) && R.match(I->getOperand(1))) || | ||||||
| 995 | (Commutable && L.match(I->getOperand(1)) && | ||||||
| 996 | R.match(I->getOperand(0))); | ||||||
| 997 | } | ||||||
| 998 | if (auto *CE = dyn_cast<ConstantExpr>(V)) | ||||||
| 999 | return CE->getOpcode() == Opcode && | ||||||
| 1000 | ((L.match(CE->getOperand(0)) && R.match(CE->getOperand(1))) || | ||||||
| 1001 | (Commutable && L.match(CE->getOperand(1)) && | ||||||
| 1002 | R.match(CE->getOperand(0)))); | ||||||
| 1003 | return false; | ||||||
| 1004 | } | ||||||
| 1005 | }; | ||||||
| 1006 | |||||||
| 1007 | template <typename LHS, typename RHS> | ||||||
| 1008 | inline BinaryOp_match<LHS, RHS, Instruction::Add> m_Add(const LHS &L, | ||||||
| 1009 | const RHS &R) { | ||||||
| 1010 | return BinaryOp_match<LHS, RHS, Instruction::Add>(L, R); | ||||||
| 1011 | } | ||||||
| 1012 | |||||||
| 1013 | template <typename LHS, typename RHS> | ||||||
| 1014 | inline BinaryOp_match<LHS, RHS, Instruction::FAdd> m_FAdd(const LHS &L, | ||||||
| 1015 | const RHS &R) { | ||||||
| 1016 | return BinaryOp_match<LHS, RHS, Instruction::FAdd>(L, R); | ||||||
| 1017 | } | ||||||
| 1018 | |||||||
| 1019 | template <typename LHS, typename RHS> | ||||||
| 1020 | inline BinaryOp_match<LHS, RHS, Instruction::Sub> m_Sub(const LHS &L, | ||||||
| 1021 | const RHS &R) { | ||||||
| 1022 | return BinaryOp_match<LHS, RHS, Instruction::Sub>(L, R); | ||||||
| 1023 | } | ||||||
| 1024 | |||||||
| 1025 | template <typename LHS, typename RHS> | ||||||
| 1026 | inline BinaryOp_match<LHS, RHS, Instruction::FSub> m_FSub(const LHS &L, | ||||||
| 1027 | const RHS &R) { | ||||||
| 1028 | return BinaryOp_match<LHS, RHS, Instruction::FSub>(L, R); | ||||||
| 1029 | } | ||||||
| 1030 | |||||||
| 1031 | template <typename Op_t> struct FNeg_match { | ||||||
| 1032 | Op_t X; | ||||||
| 1033 | |||||||
| 1034 | FNeg_match(const Op_t &Op) : X(Op) {} | ||||||
| 1035 | template <typename OpTy> bool match(OpTy *V) { | ||||||
| 1036 | auto *FPMO = dyn_cast<FPMathOperator>(V); | ||||||
| 1037 | if (!FPMO) return false; | ||||||
| 1038 | |||||||
| 1039 | if (FPMO->getOpcode() == Instruction::FNeg) | ||||||
| 1040 | return X.match(FPMO->getOperand(0)); | ||||||
| 1041 | |||||||
| 1042 | if (FPMO->getOpcode() == Instruction::FSub) { | ||||||
| 1043 | if (FPMO->hasNoSignedZeros()) { | ||||||
| 1044 | // With 'nsz', any zero goes. | ||||||
| 1045 | if (!cstfp_pred_ty<is_any_zero_fp>().match(FPMO->getOperand(0))) | ||||||
| 1046 | return false; | ||||||
| 1047 | } else { | ||||||
| 1048 | // Without 'nsz', we need fsub -0.0, X exactly. | ||||||
| 1049 | if (!cstfp_pred_ty<is_neg_zero_fp>().match(FPMO->getOperand(0))) | ||||||
| 1050 | return false; | ||||||
| 1051 | } | ||||||
| 1052 | |||||||
| 1053 | return X.match(FPMO->getOperand(1)); | ||||||
| 1054 | } | ||||||
| 1055 | |||||||
| 1056 | return false; | ||||||
| 1057 | } | ||||||
| 1058 | }; | ||||||
| 1059 | |||||||
| 1060 | /// Match 'fneg X' as 'fsub -0.0, X'. | ||||||
| 1061 | template <typename OpTy> | ||||||
| 1062 | inline FNeg_match<OpTy> | ||||||
| 1063 | m_FNeg(const OpTy &X) { | ||||||
| 1064 | return FNeg_match<OpTy>(X); | ||||||
| 1065 | } | ||||||
| 1066 | |||||||
| 1067 | /// Match 'fneg X' as 'fsub +-0.0, X'. | ||||||
| 1068 | template <typename RHS> | ||||||
| 1069 | inline BinaryOp_match<cstfp_pred_ty<is_any_zero_fp>, RHS, Instruction::FSub> | ||||||
| 1070 | m_FNegNSZ(const RHS &X) { | ||||||
| 1071 | return m_FSub(m_AnyZeroFP(), X); | ||||||
| 1072 | } | ||||||
| 1073 | |||||||
| 1074 | template <typename LHS, typename RHS> | ||||||
| 1075 | inline BinaryOp_match<LHS, RHS, Instruction::Mul> m_Mul(const LHS &L, | ||||||
| 1076 | const RHS &R) { | ||||||
| 1077 | return BinaryOp_match<LHS, RHS, Instruction::Mul>(L, R); | ||||||
| 1078 | } | ||||||
| 1079 | |||||||
| 1080 | template <typename LHS, typename RHS> | ||||||
| 1081 | inline BinaryOp_match<LHS, RHS, Instruction::FMul> m_FMul(const LHS &L, | ||||||
| 1082 | const RHS &R) { | ||||||
| 1083 | return BinaryOp_match<LHS, RHS, Instruction::FMul>(L, R); | ||||||
| 1084 | } | ||||||
| 1085 | |||||||
| 1086 | template <typename LHS, typename RHS> | ||||||
| 1087 | inline BinaryOp_match<LHS, RHS, Instruction::UDiv> m_UDiv(const LHS &L, | ||||||
| 1088 | const RHS &R) { | ||||||
| 1089 | return BinaryOp_match<LHS, RHS, Instruction::UDiv>(L, R); | ||||||
| 1090 | } | ||||||
| 1091 | |||||||
| 1092 | template <typename LHS, typename RHS> | ||||||
| 1093 | inline BinaryOp_match<LHS, RHS, Instruction::SDiv> m_SDiv(const LHS &L, | ||||||
| 1094 | const RHS &R) { | ||||||
| 1095 | return BinaryOp_match<LHS, RHS, Instruction::SDiv>(L, R); | ||||||
| 1096 | } | ||||||
| 1097 | |||||||
| 1098 | template <typename LHS, typename RHS> | ||||||
| 1099 | inline BinaryOp_match<LHS, RHS, Instruction::FDiv> m_FDiv(const LHS &L, | ||||||
| 1100 | const RHS &R) { | ||||||
| 1101 | return BinaryOp_match<LHS, RHS, Instruction::FDiv>(L, R); | ||||||
| 1102 | } | ||||||
| 1103 | |||||||
| 1104 | template <typename LHS, typename RHS> | ||||||
| 1105 | inline BinaryOp_match<LHS, RHS, Instruction::URem> m_URem(const LHS &L, | ||||||
| 1106 | const RHS &R) { | ||||||
| 1107 | return BinaryOp_match<LHS, RHS, Instruction::URem>(L, R); | ||||||
| 1108 | } | ||||||
| 1109 | |||||||
| 1110 | template <typename LHS, typename RHS> | ||||||
| 1111 | inline BinaryOp_match<LHS, RHS, Instruction::SRem> m_SRem(const LHS &L, | ||||||
| 1112 | const RHS &R) { | ||||||
| 1113 | return BinaryOp_match<LHS, RHS, Instruction::SRem>(L, R); | ||||||
| 1114 | } | ||||||
| 1115 | |||||||
| 1116 | template <typename LHS, typename RHS> | ||||||
| 1117 | inline BinaryOp_match<LHS, RHS, Instruction::FRem> m_FRem(const LHS &L, | ||||||
| 1118 | const RHS &R) { | ||||||
| 1119 | return BinaryOp_match<LHS, RHS, Instruction::FRem>(L, R); | ||||||
| 1120 | } | ||||||
| 1121 | |||||||
| 1122 | template <typename LHS, typename RHS> | ||||||
| 1123 | inline BinaryOp_match<LHS, RHS, Instruction::And> m_And(const LHS &L, | ||||||
| 1124 | const RHS &R) { | ||||||
| 1125 | return BinaryOp_match<LHS, RHS, Instruction::And>(L, R); | ||||||
| 1126 | } | ||||||
| 1127 | |||||||
| 1128 | template <typename LHS, typename RHS> | ||||||
| 1129 | inline BinaryOp_match<LHS, RHS, Instruction::Or> m_Or(const LHS &L, | ||||||
| 1130 | const RHS &R) { | ||||||
| 1131 | return BinaryOp_match<LHS, RHS, Instruction::Or>(L, R); | ||||||
| 1132 | } | ||||||
| 1133 | |||||||
| 1134 | template <typename LHS, typename RHS> | ||||||
| 1135 | inline BinaryOp_match<LHS, RHS, Instruction::Xor> m_Xor(const LHS &L, | ||||||
| 1136 | const RHS &R) { | ||||||
| 1137 | return BinaryOp_match<LHS, RHS, Instruction::Xor>(L, R); | ||||||
| 1138 | } | ||||||
| 1139 | |||||||
| 1140 | template <typename LHS, typename RHS> | ||||||
| 1141 | inline BinaryOp_match<LHS, RHS, Instruction::Shl> m_Shl(const LHS &L, | ||||||
| 1142 | const RHS &R) { | ||||||
| 1143 | return BinaryOp_match<LHS, RHS, Instruction::Shl>(L, R); | ||||||
| 1144 | } | ||||||
| 1145 | |||||||
| 1146 | template <typename LHS, typename RHS> | ||||||
| 1147 | inline BinaryOp_match<LHS, RHS, Instruction::LShr> m_LShr(const LHS &L, | ||||||
| 1148 | const RHS &R) { | ||||||
| 1149 | return BinaryOp_match<LHS, RHS, Instruction::LShr>(L, R); | ||||||
| 1150 | } | ||||||
| 1151 | |||||||
| 1152 | template <typename LHS, typename RHS> | ||||||
| 1153 | inline BinaryOp_match<LHS, RHS, Instruction::AShr> m_AShr(const LHS &L, | ||||||
| 1154 | const RHS &R) { | ||||||
| 1155 | return BinaryOp_match<LHS, RHS, Instruction::AShr>(L, R); | ||||||
| 1156 | } | ||||||
| 1157 | |||||||
| 1158 | template <typename LHS_t, typename RHS_t, unsigned Opcode, | ||||||
| 1159 | unsigned WrapFlags = 0> | ||||||
| 1160 | struct OverflowingBinaryOp_match { | ||||||
| 1161 | LHS_t L; | ||||||
| 1162 | RHS_t R; | ||||||
| 1163 | |||||||
| 1164 | OverflowingBinaryOp_match(const LHS_t &LHS, const RHS_t &RHS) | ||||||
| 1165 | : L(LHS), R(RHS) {} | ||||||
| 1166 | |||||||
| 1167 | template <typename OpTy> bool match(OpTy *V) { | ||||||
| 1168 | if (auto *Op = dyn_cast<OverflowingBinaryOperator>(V)) { | ||||||
| 1169 | if (Op->getOpcode() != Opcode) | ||||||
| 1170 | return false; | ||||||
| 1171 | if ((WrapFlags & OverflowingBinaryOperator::NoUnsignedWrap) && | ||||||
| 1172 | !Op->hasNoUnsignedWrap()) | ||||||
| 1173 | return false; | ||||||
| 1174 | if ((WrapFlags & OverflowingBinaryOperator::NoSignedWrap) && | ||||||
| 1175 | !Op->hasNoSignedWrap()) | ||||||
| 1176 | return false; | ||||||
| 1177 | return L.match(Op->getOperand(0)) && R.match(Op->getOperand(1)); | ||||||
| 1178 | } | ||||||
| 1179 | return false; | ||||||
| 1180 | } | ||||||
| 1181 | }; | ||||||
| 1182 | |||||||
| 1183 | template <typename LHS, typename RHS> | ||||||
| 1184 | inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Add, | ||||||
| 1185 | OverflowingBinaryOperator::NoSignedWrap> | ||||||
| 1186 | m_NSWAdd(const LHS &L, const RHS &R) { | ||||||
| 1187 | return OverflowingBinaryOp_match<LHS, RHS, Instruction::Add, | ||||||
| 1188 | OverflowingBinaryOperator::NoSignedWrap>( | ||||||
| 1189 | L, R); | ||||||
| 1190 | } | ||||||
| 1191 | template <typename LHS, typename RHS> | ||||||
| 1192 | inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Sub, | ||||||
| 1193 | OverflowingBinaryOperator::NoSignedWrap> | ||||||
| 1194 | m_NSWSub(const LHS &L, const RHS &R) { | ||||||
| 1195 | return OverflowingBinaryOp_match<LHS, RHS, Instruction::Sub, | ||||||
| 1196 | OverflowingBinaryOperator::NoSignedWrap>( | ||||||
| 1197 | L, R); | ||||||
| 1198 | } | ||||||
| 1199 | template <typename LHS, typename RHS> | ||||||
| 1200 | inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Mul, | ||||||
| 1201 | OverflowingBinaryOperator::NoSignedWrap> | ||||||
| 1202 | m_NSWMul(const LHS &L, const RHS &R) { | ||||||
| 1203 | return OverflowingBinaryOp_match<LHS, RHS, Instruction::Mul, | ||||||
| 1204 | OverflowingBinaryOperator::NoSignedWrap>( | ||||||
| 1205 | L, R); | ||||||
| 1206 | } | ||||||
| 1207 | template <typename LHS, typename RHS> | ||||||
| 1208 | inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Shl, | ||||||
| 1209 | OverflowingBinaryOperator::NoSignedWrap> | ||||||
| 1210 | m_NSWShl(const LHS &L, const RHS &R) { | ||||||
| 1211 | return OverflowingBinaryOp_match<LHS, RHS, Instruction::Shl, | ||||||
| 1212 | OverflowingBinaryOperator::NoSignedWrap>( | ||||||
| 1213 | L, R); | ||||||
| 1214 | } | ||||||
| 1215 | |||||||
| 1216 | template <typename LHS, typename RHS> | ||||||
| 1217 | inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Add, | ||||||
| 1218 | OverflowingBinaryOperator::NoUnsignedWrap> | ||||||
| 1219 | m_NUWAdd(const LHS &L, const RHS &R) { | ||||||
| 1220 | return OverflowingBinaryOp_match<LHS, RHS, Instruction::Add, | ||||||
| 1221 | OverflowingBinaryOperator::NoUnsignedWrap>( | ||||||
| 1222 | L, R); | ||||||
| 1223 | } | ||||||
| 1224 | template <typename LHS, typename RHS> | ||||||
| 1225 | inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Sub, | ||||||
| 1226 | OverflowingBinaryOperator::NoUnsignedWrap> | ||||||
| 1227 | m_NUWSub(const LHS &L, const RHS &R) { | ||||||
| 1228 | return OverflowingBinaryOp_match<LHS, RHS, Instruction::Sub, | ||||||
| 1229 | OverflowingBinaryOperator::NoUnsignedWrap>( | ||||||
| 1230 | L, R); | ||||||
| 1231 | } | ||||||
| 1232 | template <typename LHS, typename RHS> | ||||||
| 1233 | inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Mul, | ||||||
| 1234 | OverflowingBinaryOperator::NoUnsignedWrap> | ||||||
| 1235 | m_NUWMul(const LHS &L, const RHS &R) { | ||||||
| 1236 | return OverflowingBinaryOp_match<LHS, RHS, Instruction::Mul, | ||||||
| 1237 | OverflowingBinaryOperator::NoUnsignedWrap>( | ||||||
| 1238 | L, R); | ||||||
| 1239 | } | ||||||
| 1240 | template <typename LHS, typename RHS> | ||||||
| 1241 | inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Shl, | ||||||
| 1242 | OverflowingBinaryOperator::NoUnsignedWrap> | ||||||
| 1243 | m_NUWShl(const LHS &L, const RHS &R) { | ||||||
| 1244 | return OverflowingBinaryOp_match<LHS, RHS, Instruction::Shl, | ||||||
| 1245 | OverflowingBinaryOperator::NoUnsignedWrap>( | ||||||
| 1246 | L, R); | ||||||
| 1247 | } | ||||||
| 1248 | |||||||
| 1249 | //===----------------------------------------------------------------------===// | ||||||
| 1250 | // Class that matches a group of binary opcodes. | ||||||
| 1251 | // | ||||||
| 1252 | template <typename LHS_t, typename RHS_t, typename Predicate> | ||||||
| 1253 | struct BinOpPred_match : Predicate { | ||||||
| 1254 | LHS_t L; | ||||||
| 1255 | RHS_t R; | ||||||
| 1256 | |||||||
| 1257 | BinOpPred_match(const LHS_t &LHS, const RHS_t &RHS) : L(LHS), R(RHS) {} | ||||||
| 1258 | |||||||
| 1259 | template <typename OpTy> bool match(OpTy *V) { | ||||||
| 1260 | if (auto *I = dyn_cast<Instruction>(V)) | ||||||
| 1261 | return this->isOpType(I->getOpcode()) && L.match(I->getOperand(0)) && | ||||||
| 1262 | R.match(I->getOperand(1)); | ||||||
| 1263 | if (auto *CE = dyn_cast<ConstantExpr>(V)) | ||||||
| 1264 | return this->isOpType(CE->getOpcode()) && L.match(CE->getOperand(0)) && | ||||||
| 1265 | R.match(CE->getOperand(1)); | ||||||
| 1266 | return false; | ||||||
| 1267 | } | ||||||
| 1268 | }; | ||||||
| 1269 | |||||||
| 1270 | struct is_shift_op { | ||||||
| 1271 | bool isOpType(unsigned Opcode) { return Instruction::isShift(Opcode); } | ||||||
| 1272 | }; | ||||||
| 1273 | |||||||
| 1274 | struct is_right_shift_op { | ||||||
| 1275 | bool isOpType(unsigned Opcode) { | ||||||
| 1276 | return Opcode == Instruction::LShr || Opcode == Instruction::AShr; | ||||||
| 1277 | } | ||||||
| 1278 | }; | ||||||
| 1279 | |||||||
| 1280 | struct is_logical_shift_op { | ||||||
| 1281 | bool isOpType(unsigned Opcode) { | ||||||
| 1282 | return Opcode == Instruction::LShr || Opcode == Instruction::Shl; | ||||||
| 1283 | } | ||||||
| 1284 | }; | ||||||
| 1285 | |||||||
| 1286 | struct is_bitwiselogic_op { | ||||||
| 1287 | bool isOpType(unsigned Opcode) { | ||||||
| 1288 | return Instruction::isBitwiseLogicOp(Opcode); | ||||||
| 1289 | } | ||||||
| 1290 | }; | ||||||
| 1291 | |||||||
| 1292 | struct is_idiv_op { | ||||||
| 1293 | bool isOpType(unsigned Opcode) { | ||||||
| 1294 | return Opcode == Instruction::SDiv || Opcode == Instruction::UDiv; | ||||||
| 1295 | } | ||||||
| 1296 | }; | ||||||
| 1297 | |||||||
| 1298 | struct is_irem_op { | ||||||
| 1299 | bool isOpType(unsigned Opcode) { | ||||||
| 1300 | return Opcode == Instruction::SRem || Opcode == Instruction::URem; | ||||||
| 1301 | } | ||||||
| 1302 | }; | ||||||
| 1303 | |||||||
| 1304 | /// Matches shift operations. | ||||||
| 1305 | template <typename LHS, typename RHS> | ||||||
| 1306 | inline BinOpPred_match<LHS, RHS, is_shift_op> m_Shift(const LHS &L, | ||||||
| 1307 | const RHS &R) { | ||||||
| 1308 | return BinOpPred_match<LHS, RHS, is_shift_op>(L, R); | ||||||
| 1309 | } | ||||||
| 1310 | |||||||
| 1311 | /// Matches logical shift operations. | ||||||
| 1312 | template <typename LHS, typename RHS> | ||||||
| 1313 | inline BinOpPred_match<LHS, RHS, is_right_shift_op> m_Shr(const LHS &L, | ||||||
| 1314 | const RHS &R) { | ||||||
| 1315 | return BinOpPred_match<LHS, RHS, is_right_shift_op>(L, R); | ||||||
| 1316 | } | ||||||
| 1317 | |||||||
| 1318 | /// Matches logical shift operations. | ||||||
| 1319 | template <typename LHS, typename RHS> | ||||||
| 1320 | inline BinOpPred_match<LHS, RHS, is_logical_shift_op> | ||||||
| 1321 | m_LogicalShift(const LHS &L, const RHS &R) { | ||||||
| 1322 | return BinOpPred_match<LHS, RHS, is_logical_shift_op>(L, R); | ||||||
| 1323 | } | ||||||
| 1324 | |||||||
| 1325 | /// Matches bitwise logic operations. | ||||||
| 1326 | template <typename LHS, typename RHS> | ||||||
| 1327 | inline BinOpPred_match<LHS, RHS, is_bitwiselogic_op> | ||||||
| 1328 | m_BitwiseLogic(const LHS &L, const RHS &R) { | ||||||
| 1329 | return BinOpPred_match<LHS, RHS, is_bitwiselogic_op>(L, R); | ||||||
| 1330 | } | ||||||
| 1331 | |||||||
| 1332 | /// Matches integer division operations. | ||||||
| 1333 | template <typename LHS, typename RHS> | ||||||
| 1334 | inline BinOpPred_match<LHS, RHS, is_idiv_op> m_IDiv(const LHS &L, | ||||||
| 1335 | const RHS &R) { | ||||||
| 1336 | return BinOpPred_match<LHS, RHS, is_idiv_op>(L, R); | ||||||
| 1337 | } | ||||||
| 1338 | |||||||
| 1339 | /// Matches integer remainder operations. | ||||||
| 1340 | template <typename LHS, typename RHS> | ||||||
| 1341 | inline BinOpPred_match<LHS, RHS, is_irem_op> m_IRem(const LHS &L, | ||||||
| 1342 | const RHS &R) { | ||||||
| 1343 | return BinOpPred_match<LHS, RHS, is_irem_op>(L, R); | ||||||
| 1344 | } | ||||||
| 1345 | |||||||
| 1346 | //===----------------------------------------------------------------------===// | ||||||
| 1347 | // Class that matches exact binary ops. | ||||||
| 1348 | // | ||||||
| 1349 | template <typename SubPattern_t> struct Exact_match { | ||||||
| 1350 | SubPattern_t SubPattern; | ||||||
| 1351 | |||||||
| 1352 | Exact_match(const SubPattern_t &SP) : SubPattern(SP) {} | ||||||
| 1353 | |||||||
| 1354 | template <typename OpTy> bool match(OpTy *V) { | ||||||
| 1355 | if (auto *PEO = dyn_cast<PossiblyExactOperator>(V)) | ||||||
| 1356 | return PEO->isExact() && SubPattern.match(V); | ||||||
| 1357 | return false; | ||||||
| 1358 | } | ||||||
| 1359 | }; | ||||||
| 1360 | |||||||
| 1361 | template <typename T> inline Exact_match<T> m_Exact(const T &SubPattern) { | ||||||
| 1362 | return SubPattern; | ||||||
| 1363 | } | ||||||
| 1364 | |||||||
| 1365 | //===----------------------------------------------------------------------===// | ||||||
| 1366 | // Matchers for CmpInst classes | ||||||
| 1367 | // | ||||||
| 1368 | |||||||
| 1369 | template <typename LHS_t, typename RHS_t, typename Class, typename PredicateTy, | ||||||
| 1370 | bool Commutable = false> | ||||||
| 1371 | struct CmpClass_match { | ||||||
| 1372 | PredicateTy &Predicate; | ||||||
| 1373 | LHS_t L; | ||||||
| 1374 | RHS_t R; | ||||||
| 1375 | |||||||
| 1376 | // The evaluation order is always stable, regardless of Commutability. | ||||||
| 1377 | // The LHS is always matched first. | ||||||
| 1378 | CmpClass_match(PredicateTy &Pred, const LHS_t &LHS, const RHS_t &RHS) | ||||||
| 1379 | : Predicate(Pred), L(LHS), R(RHS) {} | ||||||
| 1380 | |||||||
| 1381 | template <typename OpTy> bool match(OpTy *V) { | ||||||
| 1382 | if (auto *I = dyn_cast<Class>(V)) { | ||||||
| 1383 | if (L.match(I->getOperand(0)) && R.match(I->getOperand(1))) { | ||||||
| 1384 | Predicate = I->getPredicate(); | ||||||
| 1385 | return true; | ||||||
| 1386 | } else if (Commutable && L.match(I->getOperand(1)) && | ||||||
| 1387 | R.match(I->getOperand(0))) { | ||||||
| 1388 | Predicate = I->getSwappedPredicate(); | ||||||
| 1389 | return true; | ||||||
| 1390 | } | ||||||
| 1391 | } | ||||||
| 1392 | return false; | ||||||
| 1393 | } | ||||||
| 1394 | }; | ||||||
| 1395 | |||||||
| 1396 | template <typename LHS, typename RHS> | ||||||
| 1397 | inline CmpClass_match<LHS, RHS, CmpInst, CmpInst::Predicate> | ||||||
| 1398 | m_Cmp(CmpInst::Predicate &Pred, const LHS &L, const RHS &R) { | ||||||
| 1399 | return CmpClass_match<LHS, RHS, CmpInst, CmpInst::Predicate>(Pred, L, R); | ||||||
| 1400 | } | ||||||
| 1401 | |||||||
| 1402 | template <typename LHS, typename RHS> | ||||||
| 1403 | inline CmpClass_match<LHS, RHS, ICmpInst, ICmpInst::Predicate> | ||||||
| 1404 | m_ICmp(ICmpInst::Predicate &Pred, const LHS &L, const RHS &R) { | ||||||
| 1405 | return CmpClass_match<LHS, RHS, ICmpInst, ICmpInst::Predicate>(Pred, L, R); | ||||||
| 1406 | } | ||||||
| 1407 | |||||||
| 1408 | template <typename LHS, typename RHS> | ||||||
| 1409 | inline CmpClass_match<LHS, RHS, FCmpInst, FCmpInst::Predicate> | ||||||
| 1410 | m_FCmp(FCmpInst::Predicate &Pred, const LHS &L, const RHS &R) { | ||||||
| 1411 | return CmpClass_match<LHS, RHS, FCmpInst, FCmpInst::Predicate>(Pred, L, R); | ||||||
| 1412 | } | ||||||
| 1413 | |||||||
| 1414 | //===----------------------------------------------------------------------===// | ||||||
| 1415 | // Matchers for instructions with a given opcode and number of operands. | ||||||
| 1416 | // | ||||||
| 1417 | |||||||
| 1418 | /// Matches instructions with Opcode and three operands. | ||||||
| 1419 | template <typename T0, unsigned Opcode> struct OneOps_match { | ||||||
| 1420 | T0 Op1; | ||||||
| 1421 | |||||||
| 1422 | OneOps_match(const T0 &Op1) : Op1(Op1) {} | ||||||
| 1423 | |||||||
| 1424 | template <typename OpTy> bool match(OpTy *V) { | ||||||
| 1425 | if (V->getValueID() == Value::InstructionVal + Opcode) { | ||||||
| 1426 | auto *I = cast<Instruction>(V); | ||||||
| 1427 | return Op1.match(I->getOperand(0)); | ||||||
| 1428 | } | ||||||
| 1429 | return false; | ||||||
| 1430 | } | ||||||
| 1431 | }; | ||||||
| 1432 | |||||||
| 1433 | /// Matches instructions with Opcode and three operands. | ||||||
| 1434 | template <typename T0, typename T1, unsigned Opcode> struct TwoOps_match { | ||||||
| 1435 | T0 Op1; | ||||||
| 1436 | T1 Op2; | ||||||
| 1437 | |||||||
| 1438 | TwoOps_match(const T0 &Op1, const T1 &Op2) : Op1(Op1), Op2(Op2) {} | ||||||
| 1439 | |||||||
| 1440 | template <typename OpTy> bool match(OpTy *V) { | ||||||
| 1441 | if (V->getValueID() == Value::InstructionVal + Opcode) { | ||||||
| 1442 | auto *I = cast<Instruction>(V); | ||||||
| 1443 | return Op1.match(I->getOperand(0)) && Op2.match(I->getOperand(1)); | ||||||
| 1444 | } | ||||||
| 1445 | return false; | ||||||
| 1446 | } | ||||||
| 1447 | }; | ||||||
| 1448 | |||||||
| 1449 | /// Matches instructions with Opcode and three operands. | ||||||
| 1450 | template <typename T0, typename T1, typename T2, unsigned Opcode> | ||||||
| 1451 | struct ThreeOps_match { | ||||||
| 1452 | T0 Op1; | ||||||
| 1453 | T1 Op2; | ||||||
| 1454 | T2 Op3; | ||||||
| 1455 | |||||||
| 1456 | ThreeOps_match(const T0 &Op1, const T1 &Op2, const T2 &Op3) | ||||||
| 1457 | : Op1(Op1), Op2(Op2), Op3(Op3) {} | ||||||
| 1458 | |||||||
| 1459 | template <typename OpTy> bool match(OpTy *V) { | ||||||
| 1460 | if (V->getValueID() == Value::InstructionVal + Opcode) { | ||||||
| 1461 | auto *I = cast<Instruction>(V); | ||||||
| 1462 | return Op1.match(I->getOperand(0)) && Op2.match(I->getOperand(1)) && | ||||||
| 1463 | Op3.match(I->getOperand(2)); | ||||||
| 1464 | } | ||||||
| 1465 | return false; | ||||||
| 1466 | } | ||||||
| 1467 | }; | ||||||
| 1468 | |||||||
| 1469 | /// Matches SelectInst. | ||||||
| 1470 | template <typename Cond, typename LHS, typename RHS> | ||||||
| 1471 | inline ThreeOps_match<Cond, LHS, RHS, Instruction::Select> | ||||||
| 1472 | m_Select(const Cond &C, const LHS &L, const RHS &R) { | ||||||
| 1473 | return ThreeOps_match<Cond, LHS, RHS, Instruction::Select>(C, L, R); | ||||||
| 1474 | } | ||||||
| 1475 | |||||||
| 1476 | /// This matches a select of two constants, e.g.: | ||||||
| 1477 | /// m_SelectCst<-1, 0>(m_Value(V)) | ||||||
| 1478 | template <int64_t L, int64_t R, typename Cond> | ||||||
| 1479 | inline ThreeOps_match<Cond, constantint_match<L>, constantint_match<R>, | ||||||
| 1480 | Instruction::Select> | ||||||
| 1481 | m_SelectCst(const Cond &C) { | ||||||
| 1482 | return m_Select(C, m_ConstantInt<L>(), m_ConstantInt<R>()); | ||||||
| 1483 | } | ||||||
| 1484 | |||||||
| 1485 | /// Matches FreezeInst. | ||||||
| 1486 | template <typename OpTy> | ||||||
| 1487 | inline OneOps_match<OpTy, Instruction::Freeze> m_Freeze(const OpTy &Op) { | ||||||
| 1488 | return OneOps_match<OpTy, Instruction::Freeze>(Op); | ||||||
| 1489 | } | ||||||
| 1490 | |||||||
| 1491 | /// Matches InsertElementInst. | ||||||
| 1492 | template <typename Val_t, typename Elt_t, typename Idx_t> | ||||||
| 1493 | inline ThreeOps_match<Val_t, Elt_t, Idx_t, Instruction::InsertElement> | ||||||
| 1494 | m_InsertElt(const Val_t &Val, const Elt_t &Elt, const Idx_t &Idx) { | ||||||
| 1495 | return ThreeOps_match<Val_t, Elt_t, Idx_t, Instruction::InsertElement>( | ||||||
| 1496 | Val, Elt, Idx); | ||||||
| 1497 | } | ||||||
| 1498 | |||||||
| 1499 | /// Matches ExtractElementInst. | ||||||
| 1500 | template <typename Val_t, typename Idx_t> | ||||||
| 1501 | inline TwoOps_match<Val_t, Idx_t, Instruction::ExtractElement> | ||||||
| 1502 | m_ExtractElt(const Val_t &Val, const Idx_t &Idx) { | ||||||
| 1503 | return TwoOps_match<Val_t, Idx_t, Instruction::ExtractElement>(Val, Idx); | ||||||
| 1504 | } | ||||||
| 1505 | |||||||
| 1506 | /// Matches shuffle. | ||||||
| 1507 | template <typename T0, typename T1, typename T2> struct Shuffle_match { | ||||||
| 1508 | T0 Op1; | ||||||
| 1509 | T1 Op2; | ||||||
| 1510 | T2 Mask; | ||||||
| 1511 | |||||||
| 1512 | Shuffle_match(const T0 &Op1, const T1 &Op2, const T2 &Mask) | ||||||
| 1513 | : Op1(Op1), Op2(Op2), Mask(Mask) {} | ||||||
| 1514 | |||||||
| 1515 | template <typename OpTy> bool match(OpTy *V) { | ||||||
| 1516 | if (auto *I = dyn_cast<ShuffleVectorInst>(V)) { | ||||||
| 1517 | return Op1.match(I->getOperand(0)) && Op2.match(I->getOperand(1)) && | ||||||
| 1518 | Mask.match(I->getShuffleMask()); | ||||||
| 1519 | } | ||||||
| 1520 | return false; | ||||||
| 1521 | } | ||||||
| 1522 | }; | ||||||
| 1523 | |||||||
| 1524 | struct m_Mask { | ||||||
| 1525 | ArrayRef<int> &MaskRef; | ||||||
| 1526 | m_Mask(ArrayRef<int> &MaskRef) : MaskRef(MaskRef) {} | ||||||
| 1527 | bool match(ArrayRef<int> Mask) { | ||||||
| 1528 | MaskRef = Mask; | ||||||
| 1529 | return true; | ||||||
| 1530 | } | ||||||
| 1531 | }; | ||||||
| 1532 | |||||||
| 1533 | struct m_ZeroMask { | ||||||
| 1534 | bool match(ArrayRef<int> Mask) { | ||||||
| 1535 | return all_of(Mask, [](int Elem) { return Elem == 0 || Elem == -1; }); | ||||||
| 1536 | } | ||||||
| 1537 | }; | ||||||
| 1538 | |||||||
| 1539 | struct m_SpecificMask { | ||||||
| 1540 | ArrayRef<int> &MaskRef; | ||||||
| 1541 | m_SpecificMask(ArrayRef<int> &MaskRef) : MaskRef(MaskRef) {} | ||||||
| 1542 | bool match(ArrayRef<int> Mask) { return MaskRef == Mask; } | ||||||
| 1543 | }; | ||||||
| 1544 | |||||||
| 1545 | struct m_SplatOrUndefMask { | ||||||
| 1546 | int &SplatIndex; | ||||||
| 1547 | m_SplatOrUndefMask(int &SplatIndex) : SplatIndex(SplatIndex) {} | ||||||
| 1548 | bool match(ArrayRef<int> Mask) { | ||||||
| 1549 | auto First = find_if(Mask, [](int Elem) { return Elem != -1; }); | ||||||
| 1550 | if (First == Mask.end()) | ||||||
| 1551 | return false; | ||||||
| 1552 | SplatIndex = *First; | ||||||
| 1553 | return all_of(Mask, | ||||||
| 1554 | [First](int Elem) { return Elem == *First || Elem == -1; }); | ||||||
| 1555 | } | ||||||
| 1556 | }; | ||||||
| 1557 | |||||||
| 1558 | /// Matches ShuffleVectorInst independently of mask value. | ||||||
| 1559 | template <typename V1_t, typename V2_t> | ||||||
| 1560 | inline TwoOps_match<V1_t, V2_t, Instruction::ShuffleVector> | ||||||
| 1561 | m_Shuffle(const V1_t &v1, const V2_t &v2) { | ||||||
| 1562 | return TwoOps_match<V1_t, V2_t, Instruction::ShuffleVector>(v1, v2); | ||||||
| 1563 | } | ||||||
| 1564 | |||||||
| 1565 | template <typename V1_t, typename V2_t, typename Mask_t> | ||||||
| 1566 | inline Shuffle_match<V1_t, V2_t, Mask_t> | ||||||
| 1567 | m_Shuffle(const V1_t &v1, const V2_t &v2, const Mask_t &mask) { | ||||||
| 1568 | return Shuffle_match<V1_t, V2_t, Mask_t>(v1, v2, mask); | ||||||
| 1569 | } | ||||||
| 1570 | |||||||
| 1571 | /// Matches LoadInst. | ||||||
| 1572 | template <typename OpTy> | ||||||
| 1573 | inline OneOps_match<OpTy, Instruction::Load> m_Load(const OpTy &Op) { | ||||||
| 1574 | return OneOps_match<OpTy, Instruction::Load>(Op); | ||||||
| 1575 | } | ||||||
| 1576 | |||||||
| 1577 | /// Matches StoreInst. | ||||||
| 1578 | template <typename ValueOpTy, typename PointerOpTy> | ||||||
| 1579 | inline TwoOps_match<ValueOpTy, PointerOpTy, Instruction::Store> | ||||||
| 1580 | m_Store(const ValueOpTy &ValueOp, const PointerOpTy &PointerOp) { | ||||||
| 1581 | return TwoOps_match<ValueOpTy, PointerOpTy, Instruction::Store>(ValueOp, | ||||||
| 1582 | PointerOp); | ||||||
| 1583 | } | ||||||
| 1584 | |||||||
| 1585 | //===----------------------------------------------------------------------===// | ||||||
| 1586 | // Matchers for CastInst classes | ||||||
| 1587 | // | ||||||
| 1588 | |||||||
| 1589 | template <typename Op_t, unsigned Opcode> struct CastClass_match { | ||||||
| 1590 | Op_t Op; | ||||||
| 1591 | |||||||
| 1592 | CastClass_match(const Op_t &OpMatch) : Op(OpMatch) {} | ||||||
| 1593 | |||||||
| 1594 | template <typename OpTy> bool match(OpTy *V) { | ||||||
| 1595 | if (auto *O
| ||||||
| 1596 | return O->getOpcode() == Opcode && Op.match(O->getOperand(0)); | ||||||
| 1597 | return false; | ||||||
| 1598 | } | ||||||
| 1599 | }; | ||||||
| 1600 | |||||||
| 1601 | /// Matches BitCast. | ||||||
| 1602 | template <typename OpTy> | ||||||
| 1603 | inline CastClass_match<OpTy, Instruction::BitCast> m_BitCast(const OpTy &Op) { | ||||||
| 1604 | return CastClass_match<OpTy, Instruction::BitCast>(Op); | ||||||
| 1605 | } | ||||||
| 1606 | |||||||
| 1607 | /// Matches PtrToInt. | ||||||
| 1608 | template <typename OpTy> | ||||||
| 1609 | inline CastClass_match<OpTy, Instruction::PtrToInt> m_PtrToInt(const OpTy &Op) { | ||||||
| 1610 | return CastClass_match<OpTy, Instruction::PtrToInt>(Op); | ||||||
| 1611 | } | ||||||
| 1612 | |||||||
| 1613 | /// Matches IntToPtr. | ||||||
| 1614 | template <typename OpTy> | ||||||
| 1615 | inline CastClass_match<OpTy, Instruction::IntToPtr> m_IntToPtr(const OpTy &Op) { | ||||||
| 1616 | return CastClass_match<OpTy, Instruction::IntToPtr>(Op); | ||||||
| 1617 | } | ||||||
| 1618 | |||||||
| 1619 | /// Matches Trunc. | ||||||
| 1620 | template <typename OpTy> | ||||||
| 1621 | inline CastClass_match<OpTy, Instruction::Trunc> m_Trunc(const OpTy &Op) { | ||||||
| 1622 | return CastClass_match<OpTy, Instruction::Trunc>(Op); | ||||||
| 1623 | } | ||||||
| 1624 | |||||||
| 1625 | template <typename OpTy> | ||||||
| 1626 | inline match_combine_or<CastClass_match<OpTy, Instruction::Trunc>, OpTy> | ||||||
| 1627 | m_TruncOrSelf(const OpTy &Op) { | ||||||
| 1628 | return m_CombineOr(m_Trunc(Op), Op); | ||||||
| 1629 | } | ||||||
| 1630 | |||||||
| 1631 | /// Matches SExt. | ||||||
| 1632 | template <typename OpTy> | ||||||
| 1633 | inline CastClass_match<OpTy, Instruction::SExt> m_SExt(const OpTy &Op) { | ||||||
| 1634 | return CastClass_match<OpTy, Instruction::SExt>(Op); | ||||||
| 1635 | } | ||||||
| 1636 | |||||||
| 1637 | /// Matches ZExt. | ||||||
| 1638 | template <typename OpTy> | ||||||
| 1639 | inline CastClass_match<OpTy, Instruction::ZExt> m_ZExt(const OpTy &Op) { | ||||||
| 1640 | return CastClass_match<OpTy, Instruction::ZExt>(Op); | ||||||
| 1641 | } | ||||||
| 1642 | |||||||
| 1643 | template <typename OpTy> | ||||||
| 1644 | inline match_combine_or<CastClass_match<OpTy, Instruction::ZExt>, OpTy> | ||||||
| 1645 | m_ZExtOrSelf(const OpTy &Op) { | ||||||
| 1646 | return m_CombineOr(m_ZExt(Op), Op); | ||||||
| 1647 | } | ||||||
| 1648 | |||||||
| 1649 | template <typename OpTy> | ||||||
| 1650 | inline match_combine_or<CastClass_match<OpTy, Instruction::SExt>, OpTy> | ||||||
| 1651 | m_SExtOrSelf(const OpTy &Op) { | ||||||
| 1652 | return m_CombineOr(m_SExt(Op), Op); | ||||||
| 1653 | } | ||||||
| 1654 | |||||||
| 1655 | template <typename OpTy> | ||||||
| 1656 | inline match_combine_or<CastClass_match<OpTy, Instruction::ZExt>, | ||||||
| 1657 | CastClass_match<OpTy, Instruction::SExt>> | ||||||
| 1658 | m_ZExtOrSExt(const OpTy &Op) { | ||||||
| 1659 | return m_CombineOr(m_ZExt(Op), m_SExt(Op)); | ||||||
| 1660 | } | ||||||
| 1661 | |||||||
| 1662 | template <typename OpTy> | ||||||
| 1663 | inline match_combine_or< | ||||||
| 1664 | match_combine_or<CastClass_match<OpTy, Instruction::ZExt>, | ||||||
| 1665 | CastClass_match<OpTy, Instruction::SExt>>, | ||||||
| 1666 | OpTy> | ||||||
| 1667 | m_ZExtOrSExtOrSelf(const OpTy &Op) { | ||||||
| 1668 | return m_CombineOr(m_ZExtOrSExt(Op), Op); | ||||||
| 1669 | } | ||||||
| 1670 | |||||||
| 1671 | template <typename OpTy> | ||||||
| 1672 | inline CastClass_match<OpTy, Instruction::UIToFP> m_UIToFP(const OpTy &Op) { | ||||||
| 1673 | return CastClass_match<OpTy, Instruction::UIToFP>(Op); | ||||||
| 1674 | } | ||||||
| 1675 | |||||||
| 1676 | template <typename OpTy> | ||||||
| 1677 | inline CastClass_match<OpTy, Instruction::SIToFP> m_SIToFP(const OpTy &Op) { | ||||||
| 1678 | return CastClass_match<OpTy, Instruction::SIToFP>(Op); | ||||||
| 1679 | } | ||||||
| 1680 | |||||||
| 1681 | template <typename OpTy> | ||||||
| 1682 | inline CastClass_match<OpTy, Instruction::FPToUI> m_FPToUI(const OpTy &Op) { | ||||||
| 1683 | return CastClass_match<OpTy, Instruction::FPToUI>(Op); | ||||||
| 1684 | } | ||||||
| 1685 | |||||||
| 1686 | template <typename OpTy> | ||||||
| 1687 | inline CastClass_match<OpTy, Instruction::FPToSI> m_FPToSI(const OpTy &Op) { | ||||||
| 1688 | return CastClass_match<OpTy, Instruction::FPToSI>(Op); | ||||||
| 1689 | } | ||||||
| 1690 | |||||||
| 1691 | template <typename OpTy> | ||||||
| 1692 | inline CastClass_match<OpTy, Instruction::FPTrunc> m_FPTrunc(const OpTy &Op) { | ||||||
| 1693 | return CastClass_match<OpTy, Instruction::FPTrunc>(Op); | ||||||
| 1694 | } | ||||||
| 1695 | |||||||
| 1696 | template <typename OpTy> | ||||||
| 1697 | inline CastClass_match<OpTy, Instruction::FPExt> m_FPExt(const OpTy &Op) { | ||||||
| 1698 | return CastClass_match<OpTy, Instruction::FPExt>(Op); | ||||||
| 1699 | } | ||||||
| 1700 | |||||||
| 1701 | //===----------------------------------------------------------------------===// | ||||||
| 1702 | // Matchers for control flow. | ||||||
| 1703 | // | ||||||
| 1704 | |||||||
| 1705 | struct br_match { | ||||||
| 1706 | BasicBlock *&Succ; | ||||||
| 1707 | |||||||
| 1708 | br_match(BasicBlock *&Succ) : Succ(Succ) {} | ||||||
| 1709 | |||||||
| 1710 | template <typename OpTy> bool match(OpTy *V) { | ||||||
| 1711 | if (auto *BI = dyn_cast<BranchInst>(V)) | ||||||
| 1712 | if (BI->isUnconditional()) { | ||||||
| 1713 | Succ = BI->getSuccessor(0); | ||||||
| 1714 | return true; | ||||||
| 1715 | } | ||||||
| 1716 | return false; | ||||||
| 1717 | } | ||||||
| 1718 | }; | ||||||
| 1719 | |||||||
| 1720 | inline br_match m_UnconditionalBr(BasicBlock *&Succ) { return br_match(Succ); } | ||||||
| 1721 | |||||||
| 1722 | template <typename Cond_t, typename TrueBlock_t, typename FalseBlock_t> | ||||||
| 1723 | struct brc_match { | ||||||
| 1724 | Cond_t Cond; | ||||||
| 1725 | TrueBlock_t T; | ||||||
| 1726 | FalseBlock_t F; | ||||||
| 1727 | |||||||
| 1728 | brc_match(const Cond_t &C, const TrueBlock_t &t, const FalseBlock_t &f) | ||||||
| 1729 | : Cond(C), T(t), F(f) {} | ||||||
| 1730 | |||||||
| 1731 | template <typename OpTy> bool match(OpTy *V) { | ||||||
| 1732 | if (auto *BI = dyn_cast<BranchInst>(V)) | ||||||
| 1733 | if (BI->isConditional() && Cond.match(BI->getCondition())) | ||||||
| 1734 | return T.match(BI->getSuccessor(0)) && F.match(BI->getSuccessor(1)); | ||||||
| 1735 | return false; | ||||||
| 1736 | } | ||||||
| 1737 | }; | ||||||
| 1738 | |||||||
| 1739 | template <typename Cond_t> | ||||||
| 1740 | inline brc_match<Cond_t, bind_ty<BasicBlock>, bind_ty<BasicBlock>> | ||||||
| 1741 | m_Br(const Cond_t &C, BasicBlock *&T, BasicBlock *&F) { | ||||||
| 1742 | return brc_match<Cond_t, bind_ty<BasicBlock>, bind_ty<BasicBlock>>( | ||||||
| 1743 | C, m_BasicBlock(T), m_BasicBlock(F)); | ||||||
| 1744 | } | ||||||
| 1745 | |||||||
| 1746 | template <typename Cond_t, typename TrueBlock_t, typename FalseBlock_t> | ||||||
| 1747 | inline brc_match<Cond_t, TrueBlock_t, FalseBlock_t> | ||||||
| 1748 | m_Br(const Cond_t &C, const TrueBlock_t &T, const FalseBlock_t &F) { | ||||||
| 1749 | return brc_match<Cond_t, TrueBlock_t, FalseBlock_t>(C, T, F); | ||||||
| 1750 | } | ||||||
| 1751 | |||||||
| 1752 | //===----------------------------------------------------------------------===// | ||||||
| 1753 | // Matchers for max/min idioms, eg: "select (sgt x, y), x, y" -> smax(x,y). | ||||||
| 1754 | // | ||||||
| 1755 | |||||||
| 1756 | template <typename CmpInst_t, typename LHS_t, typename RHS_t, typename Pred_t, | ||||||
| 1757 | bool Commutable = false> | ||||||
| 1758 | struct MaxMin_match { | ||||||
| 1759 | using PredType = Pred_t; | ||||||
| 1760 | LHS_t L; | ||||||
| 1761 | RHS_t R; | ||||||
| 1762 | |||||||
| 1763 | // The evaluation order is always stable, regardless of Commutability. | ||||||
| 1764 | // The LHS is always matched first. | ||||||
| 1765 | MaxMin_match(const LHS_t &LHS, const RHS_t &RHS) : L(LHS), R(RHS) {} | ||||||
| 1766 | |||||||
| 1767 | template <typename OpTy> bool match(OpTy *V) { | ||||||
| 1768 | if (auto *II = dyn_cast<IntrinsicInst>(V)) { | ||||||
| 1769 | Intrinsic::ID IID = II->getIntrinsicID(); | ||||||
| 1770 | if ((IID == Intrinsic::smax && Pred_t::match(ICmpInst::ICMP_SGT)) || | ||||||
| 1771 | (IID == Intrinsic::smin && Pred_t::match(ICmpInst::ICMP_SLT)) || | ||||||
| 1772 | (IID == Intrinsic::umax && Pred_t::match(ICmpInst::ICMP_UGT)) || | ||||||
| 1773 | (IID == Intrinsic::umin && Pred_t::match(ICmpInst::ICMP_ULT))) { | ||||||
| 1774 | Value *LHS = II->getOperand(0), *RHS = II->getOperand(1); | ||||||
| 1775 | return (L.match(LHS) && R.match(RHS)) || | ||||||
| 1776 | (Commutable && L.match(RHS) && R.match(LHS)); | ||||||
| 1777 | } | ||||||
| 1778 | } | ||||||
| 1779 | // Look for "(x pred y) ? x : y" or "(x pred y) ? y : x". | ||||||
| 1780 | auto *SI = dyn_cast<SelectInst>(V); | ||||||
| 1781 | if (!SI) | ||||||
| 1782 | return false; | ||||||
| 1783 | auto *Cmp = dyn_cast<CmpInst_t>(SI->getCondition()); | ||||||
| 1784 | if (!Cmp) | ||||||
| 1785 | return false; | ||||||
| 1786 | // At this point we have a select conditioned on a comparison. Check that | ||||||
| 1787 | // it is the values returned by the select that are being compared. | ||||||
| 1788 | auto *TrueVal = SI->getTrueValue(); | ||||||
| 1789 | auto *FalseVal = SI->getFalseValue(); | ||||||
| 1790 | auto *LHS = Cmp->getOperand(0); | ||||||
| 1791 | auto *RHS = Cmp->getOperand(1); | ||||||
| 1792 | if ((TrueVal != LHS || FalseVal != RHS) && | ||||||
| 1793 | (TrueVal != RHS || FalseVal != LHS)) | ||||||
| 1794 | return false; | ||||||
| 1795 | typename CmpInst_t::Predicate Pred = | ||||||
| 1796 | LHS == TrueVal ? Cmp->getPredicate() : Cmp->getInversePredicate(); | ||||||
| 1797 | // Does "(x pred y) ? x : y" represent the desired max/min operation? | ||||||
| 1798 | if (!Pred_t::match(Pred)) | ||||||
| 1799 | return false; | ||||||
| 1800 | // It does! Bind the operands. | ||||||
| 1801 | return (L.match(LHS) && R.match(RHS)) || | ||||||
| 1802 | (Commutable && L.match(RHS) && R.match(LHS)); | ||||||
| 1803 | } | ||||||
| 1804 | }; | ||||||
| 1805 | |||||||
| 1806 | /// Helper class for identifying signed max predicates. | ||||||
| 1807 | struct smax_pred_ty { | ||||||
| 1808 | static bool match(ICmpInst::Predicate Pred) { | ||||||
| 1809 | return Pred == CmpInst::ICMP_SGT || Pred == CmpInst::ICMP_SGE; | ||||||
| 1810 | } | ||||||
| 1811 | }; | ||||||
| 1812 | |||||||
| 1813 | /// Helper class for identifying signed min predicates. | ||||||
| 1814 | struct smin_pred_ty { | ||||||
| 1815 | static bool match(ICmpInst::Predicate Pred) { | ||||||
| 1816 | return Pred == CmpInst::ICMP_SLT || Pred == CmpInst::ICMP_SLE; | ||||||
| 1817 | } | ||||||
| 1818 | }; | ||||||
| 1819 | |||||||
| 1820 | /// Helper class for identifying unsigned max predicates. | ||||||
| 1821 | struct umax_pred_ty { | ||||||
| 1822 | static bool match(ICmpInst::Predicate Pred) { | ||||||
| 1823 | return Pred == CmpInst::ICMP_UGT || Pred == CmpInst::ICMP_UGE; | ||||||
| 1824 | } | ||||||
| 1825 | }; | ||||||
| 1826 | |||||||
| 1827 | /// Helper class for identifying unsigned min predicates. | ||||||
| 1828 | struct umin_pred_ty { | ||||||
| 1829 | static bool match(ICmpInst::Predicate Pred) { | ||||||
| 1830 | return Pred == CmpInst::ICMP_ULT || Pred == CmpInst::ICMP_ULE; | ||||||
| 1831 | } | ||||||
| 1832 | }; | ||||||
| 1833 | |||||||
| 1834 | /// Helper class for identifying ordered max predicates. | ||||||
| 1835 | struct ofmax_pred_ty { | ||||||
| 1836 | static bool match(FCmpInst::Predicate Pred) { | ||||||
| 1837 | return Pred == CmpInst::FCMP_OGT || Pred == CmpInst::FCMP_OGE; | ||||||
| 1838 | } | ||||||
| 1839 | }; | ||||||
| 1840 | |||||||
| 1841 | /// Helper class for identifying ordered min predicates. | ||||||
| 1842 | struct ofmin_pred_ty { | ||||||
| 1843 | static bool match(FCmpInst::Predicate Pred) { | ||||||
| 1844 | return Pred == CmpInst::FCMP_OLT || Pred == CmpInst::FCMP_OLE; | ||||||
| 1845 | } | ||||||
| 1846 | }; | ||||||
| 1847 | |||||||
| 1848 | /// Helper class for identifying unordered max predicates. | ||||||
| 1849 | struct ufmax_pred_ty { | ||||||
| 1850 | static bool match(FCmpInst::Predicate Pred) { | ||||||
| 1851 | return Pred == CmpInst::FCMP_UGT || Pred == CmpInst::FCMP_UGE; | ||||||
| 1852 | } | ||||||
| 1853 | }; | ||||||
| 1854 | |||||||
| 1855 | /// Helper class for identifying unordered min predicates. | ||||||
| 1856 | struct ufmin_pred_ty { | ||||||
| 1857 | static bool match(FCmpInst::Predicate Pred) { | ||||||
| 1858 | return Pred == CmpInst::FCMP_ULT || Pred == CmpInst::FCMP_ULE; | ||||||
| 1859 | } | ||||||
| 1860 | }; | ||||||
| 1861 | |||||||
| 1862 | template <typename LHS, typename RHS> | ||||||
| 1863 | inline MaxMin_match<ICmpInst, LHS, RHS, smax_pred_ty> m_SMax(const LHS &L, | ||||||
| 1864 | const RHS &R) { | ||||||
| 1865 | return MaxMin_match<ICmpInst, LHS, RHS, smax_pred_ty>(L, R); | ||||||
| 1866 | } | ||||||
| 1867 | |||||||
| 1868 | template <typename LHS, typename RHS> | ||||||
| 1869 | inline MaxMin_match<ICmpInst, LHS, RHS, smin_pred_ty> m_SMin(const LHS &L, | ||||||
| 1870 | const RHS &R) { | ||||||
| 1871 | return MaxMin_match<ICmpInst, LHS, RHS, smin_pred_ty>(L, R); | ||||||
| 1872 | } | ||||||
| 1873 | |||||||
| 1874 | template <typename LHS, typename RHS> | ||||||
| 1875 | inline MaxMin_match<ICmpInst, LHS, RHS, umax_pred_ty> m_UMax(const LHS &L, | ||||||
| 1876 | const RHS &R) { | ||||||
| 1877 | return MaxMin_match<ICmpInst, LHS, RHS, umax_pred_ty>(L, R); | ||||||
| 1878 | } | ||||||
| 1879 | |||||||
| 1880 | template <typename LHS, typename RHS> | ||||||
| 1881 | inline MaxMin_match<ICmpInst, LHS, RHS, umin_pred_ty> m_UMin(const LHS &L, | ||||||
| 1882 | const RHS &R) { | ||||||
| 1883 | return MaxMin_match<ICmpInst, LHS, RHS, umin_pred_ty>(L, R); | ||||||
| 1884 | } | ||||||
| 1885 | |||||||
| 1886 | template <typename LHS, typename RHS> | ||||||
| 1887 | inline match_combine_or< | ||||||
| 1888 | match_combine_or<MaxMin_match<ICmpInst, LHS, RHS, smax_pred_ty>, | ||||||
| 1889 | MaxMin_match<ICmpInst, LHS, RHS, smin_pred_ty>>, | ||||||
| 1890 | match_combine_or<MaxMin_match<ICmpInst, LHS, RHS, umax_pred_ty>, | ||||||
| 1891 | MaxMin_match<ICmpInst, LHS, RHS, umin_pred_ty>>> | ||||||
| 1892 | m_MaxOrMin(const LHS &L, const RHS &R) { | ||||||
| 1893 | return m_CombineOr(m_CombineOr(m_SMax(L, R), m_SMin(L, R)), | ||||||
| 1894 | m_CombineOr(m_UMax(L, R), m_UMin(L, R))); | ||||||
| 1895 | } | ||||||
| 1896 | |||||||
| 1897 | /// Match an 'ordered' floating point maximum function. | ||||||
| 1898 | /// Floating point has one special value 'NaN'. Therefore, there is no total | ||||||
| 1899 | /// order. However, if we can ignore the 'NaN' value (for example, because of a | ||||||
| 1900 | /// 'no-nans-float-math' flag) a combination of a fcmp and select has 'maximum' | ||||||
| 1901 | /// semantics. In the presence of 'NaN' we have to preserve the original | ||||||
| 1902 | /// select(fcmp(ogt/ge, L, R), L, R) semantics matched by this predicate. | ||||||
| 1903 | /// | ||||||
| 1904 | /// max(L, R) iff L and R are not NaN | ||||||
| 1905 | /// m_OrdFMax(L, R) = R iff L or R are NaN | ||||||
| 1906 | template <typename LHS, typename RHS> | ||||||
| 1907 | inline MaxMin_match<FCmpInst, LHS, RHS, ofmax_pred_ty> m_OrdFMax(const LHS &L, | ||||||
| 1908 | const RHS &R) { | ||||||
| 1909 | return MaxMin_match<FCmpInst, LHS, RHS, ofmax_pred_ty>(L, R); | ||||||
| 1910 | } | ||||||
| 1911 | |||||||
| 1912 | /// Match an 'ordered' floating point minimum function. | ||||||
| 1913 | /// Floating point has one special value 'NaN'. Therefore, there is no total | ||||||
| 1914 | /// order. However, if we can ignore the 'NaN' value (for example, because of a | ||||||
| 1915 | /// 'no-nans-float-math' flag) a combination of a fcmp and select has 'minimum' | ||||||
| 1916 | /// semantics. In the presence of 'NaN' we have to preserve the original | ||||||
| 1917 | /// select(fcmp(olt/le, L, R), L, R) semantics matched by this predicate. | ||||||
| 1918 | /// | ||||||
| 1919 | /// min(L, R) iff L and R are not NaN | ||||||
| 1920 | /// m_OrdFMin(L, R) = R iff L or R are NaN | ||||||
| 1921 | template <typename LHS, typename RHS> | ||||||
| 1922 | inline MaxMin_match<FCmpInst, LHS, RHS, ofmin_pred_ty> m_OrdFMin(const LHS &L, | ||||||
| 1923 | const RHS &R) { | ||||||
| 1924 | return MaxMin_match<FCmpInst, LHS, RHS, ofmin_pred_ty>(L, R); | ||||||
| 1925 | } | ||||||
| 1926 | |||||||
| 1927 | /// Match an 'unordered' floating point maximum function. | ||||||
| 1928 | /// Floating point has one special value 'NaN'. Therefore, there is no total | ||||||
| 1929 | /// order. However, if we can ignore the 'NaN' value (for example, because of a | ||||||
| 1930 | /// 'no-nans-float-math' flag) a combination of a fcmp and select has 'maximum' | ||||||
| 1931 | /// semantics. In the presence of 'NaN' we have to preserve the original | ||||||
| 1932 | /// select(fcmp(ugt/ge, L, R), L, R) semantics matched by this predicate. | ||||||
| 1933 | /// | ||||||
| 1934 | /// max(L, R) iff L and R are not NaN | ||||||
| 1935 | /// m_UnordFMax(L, R) = L iff L or R are NaN | ||||||
| 1936 | template <typename LHS, typename RHS> | ||||||
| 1937 | inline MaxMin_match<FCmpInst, LHS, RHS, ufmax_pred_ty> | ||||||
| 1938 | m_UnordFMax(const LHS &L, const RHS &R) { | ||||||
| 1939 | return MaxMin_match<FCmpInst, LHS, RHS, ufmax_pred_ty>(L, R); | ||||||
| 1940 | } | ||||||
| 1941 | |||||||
| 1942 | /// Match an 'unordered' floating point minimum function. | ||||||
| 1943 | /// Floating point has one special value 'NaN'. Therefore, there is no total | ||||||
| 1944 | /// order. However, if we can ignore the 'NaN' value (for example, because of a | ||||||
| 1945 | /// 'no-nans-float-math' flag) a combination of a fcmp and select has 'minimum' | ||||||
| 1946 | /// semantics. In the presence of 'NaN' we have to preserve the original | ||||||
| 1947 | /// select(fcmp(ult/le, L, R), L, R) semantics matched by this predicate. | ||||||
| 1948 | /// | ||||||
| 1949 | /// min(L, R) iff L and R are not NaN | ||||||
| 1950 | /// m_UnordFMin(L, R) = L iff L or R are NaN | ||||||
| 1951 | template <typename LHS, typename RHS> | ||||||
| 1952 | inline MaxMin_match<FCmpInst, LHS, RHS, ufmin_pred_ty> | ||||||
| 1953 | m_UnordFMin(const LHS &L, const RHS &R) { | ||||||
| 1954 | return MaxMin_match<FCmpInst, LHS, RHS, ufmin_pred_ty>(L, R); | ||||||
| 1955 | } | ||||||
| 1956 | |||||||
| 1957 | //===----------------------------------------------------------------------===// | ||||||
| 1958 | // Matchers for overflow check patterns: e.g. (a + b) u< a, (a ^ -1) <u b | ||||||
| 1959 | // Note that S might be matched to other instructions than AddInst. | ||||||
| 1960 | // | ||||||
| 1961 | |||||||
| 1962 | template <typename LHS_t, typename RHS_t, typename Sum_t> | ||||||
| 1963 | struct UAddWithOverflow_match { | ||||||
| 1964 | LHS_t L; | ||||||
| 1965 | RHS_t R; | ||||||
| 1966 | Sum_t S; | ||||||
| 1967 | |||||||
| 1968 | UAddWithOverflow_match(const LHS_t &L, const RHS_t &R, const Sum_t &S) | ||||||
| 1969 | : L(L), R(R), S(S) {} | ||||||
| 1970 | |||||||
| 1971 | template <typename OpTy> bool match(OpTy *V) { | ||||||
| 1972 | Value *ICmpLHS, *ICmpRHS; | ||||||
| 1973 | ICmpInst::Predicate Pred; | ||||||
| 1974 | if (!m_ICmp(Pred, m_Value(ICmpLHS), m_Value(ICmpRHS)).match(V)) | ||||||
| 1975 | return false; | ||||||
| 1976 | |||||||
| 1977 | Value *AddLHS, *AddRHS; | ||||||
| 1978 | auto AddExpr = m_Add(m_Value(AddLHS), m_Value(AddRHS)); | ||||||
| 1979 | |||||||
| 1980 | // (a + b) u< a, (a + b) u< b | ||||||
| 1981 | if (Pred == ICmpInst::ICMP_ULT) | ||||||
| 1982 | if (AddExpr.match(ICmpLHS) && (ICmpRHS == AddLHS || ICmpRHS == AddRHS)) | ||||||
| 1983 | return L.match(AddLHS) && R.match(AddRHS) && S.match(ICmpLHS); | ||||||
| 1984 | |||||||
| 1985 | // a >u (a + b), b >u (a + b) | ||||||
| 1986 | if (Pred == ICmpInst::ICMP_UGT) | ||||||
| 1987 | if (AddExpr.match(ICmpRHS) && (ICmpLHS == AddLHS || ICmpLHS == AddRHS)) | ||||||
| 1988 | return L.match(AddLHS) && R.match(AddRHS) && S.match(ICmpRHS); | ||||||
| 1989 | |||||||
| 1990 | Value *Op1; | ||||||
| 1991 | auto XorExpr = m_OneUse(m_Xor(m_Value(Op1), m_AllOnes())); | ||||||
| 1992 | // (a ^ -1) <u b | ||||||
| 1993 | if (Pred == ICmpInst::ICMP_ULT) { | ||||||
| 1994 | if (XorExpr.match(ICmpLHS)) | ||||||
| 1995 | return L.match(Op1) && R.match(ICmpRHS) && S.match(ICmpLHS); | ||||||
| 1996 | } | ||||||
| 1997 | // b > u (a ^ -1) | ||||||
| 1998 | if (Pred == ICmpInst::ICMP_UGT) { | ||||||
| 1999 | if (XorExpr.match(ICmpRHS)) | ||||||
| 2000 | return L.match(Op1) && R.match(ICmpLHS) && S.match(ICmpRHS); | ||||||
| 2001 | } | ||||||
| 2002 | |||||||
| 2003 | // Match special-case for increment-by-1. | ||||||
| 2004 | if (Pred == ICmpInst::ICMP_EQ) { | ||||||
| 2005 | // (a + 1) == 0 | ||||||
| 2006 | // (1 + a) == 0 | ||||||
| 2007 | if (AddExpr.match(ICmpLHS) && m_ZeroInt().match(ICmpRHS) && | ||||||
| 2008 | (m_One().match(AddLHS) || m_One().match(AddRHS))) | ||||||
| 2009 | return L.match(AddLHS) && R.match(AddRHS) && S.match(ICmpLHS); | ||||||
| 2010 | // 0 == (a + 1) | ||||||
| 2011 | // 0 == (1 + a) | ||||||
| 2012 | if (m_ZeroInt().match(ICmpLHS) && AddExpr.match(ICmpRHS) && | ||||||
| 2013 | (m_One().match(AddLHS) || m_One().match(AddRHS))) | ||||||
| 2014 | return L.match(AddLHS) && R.match(AddRHS) && S.match(ICmpRHS); | ||||||
| 2015 | } | ||||||
| 2016 | |||||||
| 2017 | return false; | ||||||
| 2018 | } | ||||||
| 2019 | }; | ||||||
| 2020 | |||||||
| 2021 | /// Match an icmp instruction checking for unsigned overflow on addition. | ||||||
| 2022 | /// | ||||||
| 2023 | /// S is matched to the addition whose result is being checked for overflow, and | ||||||
| 2024 | /// L and R are matched to the LHS and RHS of S. | ||||||
| 2025 | template <typename LHS_t, typename RHS_t, typename Sum_t> | ||||||
| 2026 | UAddWithOverflow_match<LHS_t, RHS_t, Sum_t> | ||||||
| 2027 | m_UAddWithOverflow(const LHS_t &L, const RHS_t &R, const Sum_t &S) { | ||||||
| 2028 | return UAddWithOverflow_match<LHS_t, RHS_t, Sum_t>(L, R, S); | ||||||
| 2029 | } | ||||||
| 2030 | |||||||
| 2031 | template <typename Opnd_t> struct Argument_match { | ||||||
| 2032 | unsigned OpI; | ||||||
| 2033 | Opnd_t Val; | ||||||
| 2034 | |||||||
| 2035 | Argument_match(unsigned OpIdx, const Opnd_t &V) : OpI(OpIdx), Val(V) {} | ||||||
| 2036 | |||||||
| 2037 | template <typename OpTy> bool match(OpTy *V) { | ||||||
| 2038 | // FIXME: Should likely be switched to use `CallBase`. | ||||||
| 2039 | if (const auto *CI = dyn_cast<CallInst>(V)) | ||||||
| 2040 | return Val.match(CI->getArgOperand(OpI)); | ||||||
| 2041 | return false; | ||||||
| 2042 | } | ||||||
| 2043 | }; | ||||||
| 2044 | |||||||
| 2045 | /// Match an argument. | ||||||
| 2046 | template <unsigned OpI, typename Opnd_t> | ||||||
| 2047 | inline Argument_match<Opnd_t> m_Argument(const Opnd_t &Op) { | ||||||
| 2048 | return Argument_match<Opnd_t>(OpI, Op); | ||||||
| 2049 | } | ||||||
| 2050 | |||||||
| 2051 | /// Intrinsic matchers. | ||||||
| 2052 | struct IntrinsicID_match { | ||||||
| 2053 | unsigned ID; | ||||||
| 2054 | |||||||
| 2055 | IntrinsicID_match(Intrinsic::ID IntrID) : ID(IntrID) {} | ||||||
| 2056 | |||||||
| 2057 | template <typename OpTy> bool match(OpTy *V) { | ||||||
| 2058 | if (const auto *CI = dyn_cast<CallInst>(V)) | ||||||
| 2059 | if (const auto *F = CI->getCalledFunction()) | ||||||
| 2060 | return F->getIntrinsicID() == ID; | ||||||
| 2061 | return false; | ||||||
| 2062 | } | ||||||
| 2063 | }; | ||||||
| 2064 | |||||||
| 2065 | /// Intrinsic matches are combinations of ID matchers, and argument | ||||||
| 2066 | /// matchers. Higher arity matcher are defined recursively in terms of and-ing | ||||||
| 2067 | /// them with lower arity matchers. Here's some convenient typedefs for up to | ||||||
| 2068 | /// several arguments, and more can be added as needed | ||||||
| 2069 | template <typename T0 = void, typename T1 = void, typename T2 = void, | ||||||
| 2070 | typename T3 = void, typename T4 = void, typename T5 = void, | ||||||
| 2071 | typename T6 = void, typename T7 = void, typename T8 = void, | ||||||
| 2072 | typename T9 = void, typename T10 = void> | ||||||
| 2073 | struct m_Intrinsic_Ty; | ||||||
| 2074 | template <typename T0> struct m_Intrinsic_Ty<T0> { | ||||||
| 2075 | using Ty = match_combine_and<IntrinsicID_match, Argument_match<T0>>; | ||||||
| 2076 | }; | ||||||
| 2077 | template <typename T0, typename T1> struct m_Intrinsic_Ty<T0, T1> { | ||||||
| 2078 | using Ty = | ||||||
| 2079 | match_combine_and<typename m_Intrinsic_Ty<T0>::Ty, Argument_match<T1>>; | ||||||
| 2080 | }; | ||||||
| 2081 | template <typename T0, typename T1, typename T2> | ||||||
| 2082 | struct m_Intrinsic_Ty<T0, T1, T2> { | ||||||
| 2083 | using Ty = | ||||||
| 2084 | match_combine_and<typename m_Intrinsic_Ty<T0, T1>::Ty, | ||||||
| 2085 | Argument_match<T2>>; | ||||||
| 2086 | }; | ||||||
| 2087 | template <typename T0, typename T1, typename T2, typename T3> | ||||||
| 2088 | struct m_Intrinsic_Ty<T0, T1, T2, T3> { | ||||||
| 2089 | using Ty = | ||||||
| 2090 | match_combine_and<typename m_Intrinsic_Ty<T0, T1, T2>::Ty, | ||||||
| 2091 | Argument_match<T3>>; | ||||||
| 2092 | }; | ||||||
| 2093 | |||||||
| 2094 | template <typename T0, typename T1, typename T2, typename T3, typename T4> | ||||||
| 2095 | struct m_Intrinsic_Ty<T0, T1, T2, T3, T4> { | ||||||
| 2096 | using Ty = match_combine_and<typename m_Intrinsic_Ty<T0, T1, T2, T3>::Ty, | ||||||
| 2097 | Argument_match<T4>>; | ||||||
| 2098 | }; | ||||||
| 2099 | |||||||
| 2100 | template <typename T0, typename T1, typename T2, typename T3, typename T4, typename T5> | ||||||
| 2101 | struct m_Intrinsic_Ty<T0, T1, T2, T3, T4, T5> { | ||||||
| 2102 | using Ty = match_combine_and<typename m_Intrinsic_Ty<T0, T1, T2, T3, T4>::Ty, | ||||||
| 2103 | Argument_match<T5>>; | ||||||
| 2104 | }; | ||||||
| 2105 | |||||||
| 2106 | /// Match intrinsic calls like this: | ||||||
| 2107 | /// m_Intrinsic<Intrinsic::fabs>(m_Value(X)) | ||||||
| 2108 | template <Intrinsic::ID IntrID> inline IntrinsicID_match m_Intrinsic() { | ||||||
| 2109 | return IntrinsicID_match(IntrID); | ||||||
| 2110 | } | ||||||
| 2111 | |||||||
| 2112 | /// Matches MaskedLoad Intrinsic. | ||||||
| 2113 | template <typename Opnd0, typename Opnd1, typename Opnd2, typename Opnd3> | ||||||
| 2114 | inline typename m_Intrinsic_Ty<Opnd0, Opnd1, Opnd2, Opnd3>::Ty | ||||||
| 2115 | m_MaskedLoad(const Opnd0 &Op0, const Opnd1 &Op1, const Opnd2 &Op2, | ||||||
| 2116 | const Opnd3 &Op3) { | ||||||
| 2117 | return m_Intrinsic<Intrinsic::masked_load>(Op0, Op1, Op2, Op3); | ||||||
| 2118 | } | ||||||
| 2119 | |||||||
| 2120 | template <Intrinsic::ID IntrID, typename T0> | ||||||
| 2121 | inline typename m_Intrinsic_Ty<T0>::Ty m_Intrinsic(const T0 &Op0) { | ||||||
| 2122 | return m_CombineAnd(m_Intrinsic<IntrID>(), m_Argument<0>(Op0)); | ||||||
| 2123 | } | ||||||
| 2124 | |||||||
| 2125 | template <Intrinsic::ID IntrID, typename T0, typename T1> | ||||||
| 2126 | inline typename m_Intrinsic_Ty<T0, T1>::Ty m_Intrinsic(const T0 &Op0, | ||||||
| 2127 | const T1 &Op1) { | ||||||
| 2128 | return m_CombineAnd(m_Intrinsic<IntrID>(Op0), m_Argument<1>(Op1)); | ||||||
| 2129 | } | ||||||
| 2130 | |||||||
| 2131 | template <Intrinsic::ID IntrID, typename T0, typename T1, typename T2> | ||||||
| 2132 | inline typename m_Intrinsic_Ty<T0, T1, T2>::Ty | ||||||
| 2133 | m_Intrinsic(const T0 &Op0, const T1 &Op1, const T2 &Op2) { | ||||||
| 2134 | return m_CombineAnd(m_Intrinsic<IntrID>(Op0, Op1), m_Argument<2>(Op2)); | ||||||
| 2135 | } | ||||||
| 2136 | |||||||
| 2137 | template <Intrinsic::ID IntrID, typename T0, typename T1, typename T2, | ||||||
| 2138 | typename T3> | ||||||
| 2139 | inline typename m_Intrinsic_Ty<T0, T1, T2, T3>::Ty | ||||||
| 2140 | m_Intrinsic(const T0 &Op0, const T1 &Op1, const T2 &Op2, const T3 &Op3) { | ||||||
| 2141 | return m_CombineAnd(m_Intrinsic<IntrID>(Op0, Op1, Op2), m_Argument<3>(Op3)); | ||||||
| 2142 | } | ||||||
| 2143 | |||||||
| 2144 | template <Intrinsic::ID IntrID, typename T0, typename T1, typename T2, | ||||||
| 2145 | typename T3, typename T4> | ||||||
| 2146 | inline typename m_Intrinsic_Ty<T0, T1, T2, T3, T4>::Ty | ||||||
| 2147 | m_Intrinsic(const T0 &Op0, const T1 &Op1, const T2 &Op2, const T3 &Op3, | ||||||
| 2148 | const T4 &Op4) { | ||||||
| 2149 | return m_CombineAnd(m_Intrinsic<IntrID>(Op0, Op1, Op2, Op3), | ||||||
| 2150 | m_Argument<4>(Op4)); | ||||||
| 2151 | } | ||||||
| 2152 | |||||||
| 2153 | template <Intrinsic::ID IntrID, typename T0, typename T1, typename T2, | ||||||
| 2154 | typename T3, typename T4, typename T5> | ||||||
| 2155 | inline typename m_Intrinsic_Ty<T0, T1, T2, T3, T4, T5>::Ty | ||||||
| 2156 | m_Intrinsic(const T0 &Op0, const T1 &Op1, const T2 &Op2, const T3 &Op3, | ||||||
| 2157 | const T4 &Op4, const T5 &Op5) { | ||||||
| 2158 | return m_CombineAnd(m_Intrinsic<IntrID>(Op0, Op1, Op2, Op3, Op4), | ||||||
| 2159 | m_Argument<5>(Op5)); | ||||||
| 2160 | } | ||||||
| 2161 | |||||||
| 2162 | // Helper intrinsic matching specializations. | ||||||
| 2163 | template <typename Opnd0> | ||||||
| 2164 | inline typename m_Intrinsic_Ty<Opnd0>::Ty m_BitReverse(const Opnd0 &Op0) { | ||||||
| 2165 | return m_Intrinsic<Intrinsic::bitreverse>(Op0); | ||||||
| 2166 | } | ||||||
| 2167 | |||||||
| 2168 | template <typename Opnd0> | ||||||
| 2169 | inline typename m_Intrinsic_Ty<Opnd0>::Ty m_BSwap(const Opnd0 &Op0) { | ||||||
| 2170 | return m_Intrinsic<Intrinsic::bswap>(Op0); | ||||||
| 2171 | } | ||||||
| 2172 | |||||||
| 2173 | template <typename Opnd0> | ||||||
| 2174 | inline typename m_Intrinsic_Ty<Opnd0>::Ty m_FAbs(const Opnd0 &Op0) { | ||||||
| 2175 | return m_Intrinsic<Intrinsic::fabs>(Op0); | ||||||
| 2176 | } | ||||||
| 2177 | |||||||
| 2178 | template <typename Opnd0> | ||||||
| 2179 | inline typename m_Intrinsic_Ty<Opnd0>::Ty m_FCanonicalize(const Opnd0 &Op0) { | ||||||
| 2180 | return m_Intrinsic<Intrinsic::canonicalize>(Op0); | ||||||
| 2181 | } | ||||||
| 2182 | |||||||
| 2183 | template <typename Opnd0, typename Opnd1> | ||||||
| 2184 | inline typename m_Intrinsic_Ty<Opnd0, Opnd1>::Ty m_FMin(const Opnd0 &Op0, | ||||||
| 2185 | const Opnd1 &Op1) { | ||||||
| 2186 | return m_Intrinsic<Intrinsic::minnum>(Op0, Op1); | ||||||
| 2187 | } | ||||||
| 2188 | |||||||
| 2189 | template <typename Opnd0, typename Opnd1> | ||||||
| 2190 | inline typename m_Intrinsic_Ty<Opnd0, Opnd1>::Ty m_FMax(const Opnd0 &Op0, | ||||||
| 2191 | const Opnd1 &Op1) { | ||||||
| 2192 | return m_Intrinsic<Intrinsic::maxnum>(Op0, Op1); | ||||||
| 2193 | } | ||||||
| 2194 | |||||||
| 2195 | template <typename Opnd0, typename Opnd1, typename Opnd2> | ||||||
| 2196 | inline typename m_Intrinsic_Ty<Opnd0, Opnd1, Opnd2>::Ty | ||||||
| 2197 | m_FShl(const Opnd0 &Op0, const Opnd1 &Op1, const Opnd2 &Op2) { | ||||||
| 2198 | return m_Intrinsic<Intrinsic::fshl>(Op0, Op1, Op2); | ||||||
| 2199 | } | ||||||
| 2200 | |||||||
| 2201 | template <typename Opnd0, typename Opnd1, typename Opnd2> | ||||||
| 2202 | inline typename m_Intrinsic_Ty<Opnd0, Opnd1, Opnd2>::Ty | ||||||
| 2203 | m_FShr(const Opnd0 &Op0, const Opnd1 &Op1, const Opnd2 &Op2) { | ||||||
| 2204 | return m_Intrinsic<Intrinsic::fshr>(Op0, Op1, Op2); | ||||||
| 2205 | } | ||||||
| 2206 | |||||||
| 2207 | //===----------------------------------------------------------------------===// | ||||||
| 2208 | // Matchers for two-operands operators with the operators in either order | ||||||
| 2209 | // | ||||||
| 2210 | |||||||
| 2211 | /// Matches a BinaryOperator with LHS and RHS in either order. | ||||||
| 2212 | template <typename LHS, typename RHS> | ||||||
| 2213 | inline AnyBinaryOp_match<LHS, RHS, true> m_c_BinOp(const LHS &L, const RHS &R) { | ||||||
| 2214 | return AnyBinaryOp_match<LHS, RHS, true>(L, R); | ||||||
| 2215 | } | ||||||
| 2216 | |||||||
| 2217 | /// Matches an ICmp with a predicate over LHS and RHS in either order. | ||||||
| 2218 | /// Swaps the predicate if operands are commuted. | ||||||
| 2219 | template <typename LHS, typename RHS> | ||||||
| 2220 | inline CmpClass_match<LHS, RHS, ICmpInst, ICmpInst::Predicate, true> | ||||||
| 2221 | m_c_ICmp(ICmpInst::Predicate &Pred, const LHS &L, const RHS &R) { | ||||||
| 2222 | return CmpClass_match<LHS, RHS, ICmpInst, ICmpInst::Predicate, true>(Pred, L, | ||||||
| 2223 | R); | ||||||
| 2224 | } | ||||||
| 2225 | |||||||
| 2226 | /// Matches a Add with LHS and RHS in either order. | ||||||
| 2227 | template <typename LHS, typename RHS> | ||||||
| 2228 | inline BinaryOp_match<LHS, RHS, Instruction::Add, true> m_c_Add(const LHS &L, | ||||||
| 2229 | const RHS &R) { | ||||||
| 2230 | return BinaryOp_match<LHS, RHS, Instruction::Add, true>(L, R); | ||||||
| 2231 | } | ||||||
| 2232 | |||||||
| 2233 | /// Matches a Mul with LHS and RHS in either order. | ||||||
| 2234 | template <typename LHS, typename RHS> | ||||||
| 2235 | inline BinaryOp_match<LHS, RHS, Instruction::Mul, true> m_c_Mul(const LHS &L, | ||||||
| 2236 | const RHS &R) { | ||||||
| 2237 | return BinaryOp_match<LHS, RHS, Instruction::Mul, true>(L, R); | ||||||
| 2238 | } | ||||||
| 2239 | |||||||
| 2240 | /// Matches an And with LHS and RHS in either order. | ||||||
| 2241 | template <typename LHS, typename RHS> | ||||||
| 2242 | inline BinaryOp_match<LHS, RHS, Instruction::And, true> m_c_And(const LHS &L, | ||||||
| 2243 | const RHS &R) { | ||||||
| 2244 | return BinaryOp_match<LHS, RHS, Instruction::And, true>(L, R); | ||||||
| 2245 | } | ||||||
| 2246 | |||||||
| 2247 | /// Matches an Or with LHS and RHS in either order. | ||||||
| 2248 | template <typename LHS, typename RHS> | ||||||
| 2249 | inline BinaryOp_match<LHS, RHS, Instruction::Or, true> m_c_Or(const LHS &L, | ||||||
| 2250 | const RHS &R) { | ||||||
| 2251 | return BinaryOp_match<LHS, RHS, Instruction::Or, true>(L, R); | ||||||
| 2252 | } | ||||||
| 2253 | |||||||
| 2254 | /// Matches an Xor with LHS and RHS in either order. | ||||||
| 2255 | template <typename LHS, typename RHS> | ||||||
| 2256 | inline BinaryOp_match<LHS, RHS, Instruction::Xor, true> m_c_Xor(const LHS &L, | ||||||
| 2257 | const RHS &R) { | ||||||
| 2258 | return BinaryOp_match<LHS, RHS, Instruction::Xor, true>(L, R); | ||||||
| 2259 | } | ||||||
| 2260 | |||||||
| 2261 | /// Matches a 'Neg' as 'sub 0, V'. | ||||||
| 2262 | template <typename ValTy> | ||||||
| 2263 | inline BinaryOp_match<cst_pred_ty<is_zero_int>, ValTy, Instruction::Sub> | ||||||
| 2264 | m_Neg(const ValTy &V) { | ||||||
| 2265 | return m_Sub(m_ZeroInt(), V); | ||||||
| 2266 | } | ||||||
| 2267 | |||||||
| 2268 | /// Matches a 'Neg' as 'sub nsw 0, V'. | ||||||
| 2269 | template <typename ValTy> | ||||||
| 2270 | inline OverflowingBinaryOp_match<cst_pred_ty<is_zero_int>, ValTy, | ||||||
| 2271 | Instruction::Sub, | ||||||
| 2272 | OverflowingBinaryOperator::NoSignedWrap> | ||||||
| 2273 | m_NSWNeg(const ValTy &V) { | ||||||
| 2274 | return m_NSWSub(m_ZeroInt(), V); | ||||||
| 2275 | } | ||||||
| 2276 | |||||||
| 2277 | /// Matches a 'Not' as 'xor V, -1' or 'xor -1, V'. | ||||||
| 2278 | template <typename ValTy> | ||||||
| 2279 | inline BinaryOp_match<ValTy, cst_pred_ty<is_all_ones>, Instruction::Xor, true> | ||||||
| 2280 | m_Not(const ValTy &V) { | ||||||
| 2281 | return m_c_Xor(V, m_AllOnes()); | ||||||
| 2282 | } | ||||||
| 2283 | |||||||
| 2284 | /// Matches an SMin with LHS and RHS in either order. | ||||||
| 2285 | template <typename LHS, typename RHS> | ||||||
| 2286 | inline MaxMin_match<ICmpInst, LHS, RHS, smin_pred_ty, true> | ||||||
| 2287 | m_c_SMin(const LHS &L, const RHS &R) { | ||||||
| 2288 | return MaxMin_match<ICmpInst, LHS, RHS, smin_pred_ty, true>(L, R); | ||||||
| 2289 | } | ||||||
| 2290 | /// Matches an SMax with LHS and RHS in either order. | ||||||
| 2291 | template <typename LHS, typename RHS> | ||||||
| 2292 | inline MaxMin_match<ICmpInst, LHS, RHS, smax_pred_ty, true> | ||||||
| 2293 | m_c_SMax(const LHS &L, const RHS &R) { | ||||||
| 2294 | return MaxMin_match<ICmpInst, LHS, RHS, smax_pred_ty, true>(L, R); | ||||||
| 2295 | } | ||||||
| 2296 | /// Matches a UMin with LHS and RHS in either order. | ||||||
| 2297 | template <typename LHS, typename RHS> | ||||||
| 2298 | inline MaxMin_match<ICmpInst, LHS, RHS, umin_pred_ty, true> | ||||||
| 2299 | m_c_UMin(const LHS &L, const RHS &R) { | ||||||
| 2300 | return MaxMin_match<ICmpInst, LHS, RHS, umin_pred_ty, true>(L, R); | ||||||
| 2301 | } | ||||||
| 2302 | /// Matches a UMax with LHS and RHS in either order. | ||||||
| 2303 | template <typename LHS, typename RHS> | ||||||
| 2304 | inline MaxMin_match<ICmpInst, LHS, RHS, umax_pred_ty, true> | ||||||
| 2305 | m_c_UMax(const LHS &L, const RHS &R) { | ||||||
| 2306 | return MaxMin_match<ICmpInst, LHS, RHS, umax_pred_ty, true>(L, R); | ||||||
| 2307 | } | ||||||
| 2308 | |||||||
| 2309 | template <typename LHS, typename RHS> | ||||||
| 2310 | inline match_combine_or< | ||||||
| 2311 | match_combine_or<MaxMin_match<ICmpInst, LHS, RHS, smax_pred_ty, true>, | ||||||
| 2312 | MaxMin_match<ICmpInst, LHS, RHS, smin_pred_ty, true>>, | ||||||
| 2313 | match_combine_or<MaxMin_match<ICmpInst, LHS, RHS, umax_pred_ty, true>, | ||||||
| 2314 | MaxMin_match<ICmpInst, LHS, RHS, umin_pred_ty, true>>> | ||||||
| 2315 | m_c_MaxOrMin(const LHS &L, const RHS &R) { | ||||||
| 2316 | return m_CombineOr(m_CombineOr(m_c_SMax(L, R), m_c_SMin(L, R)), | ||||||
| 2317 | m_CombineOr(m_c_UMax(L, R), m_c_UMin(L, R))); | ||||||
| 2318 | } | ||||||
| 2319 | |||||||
| 2320 | /// Matches FAdd with LHS and RHS in either order. | ||||||
| 2321 | template <typename LHS, typename RHS> | ||||||
| 2322 | inline BinaryOp_match<LHS, RHS, Instruction::FAdd, true> | ||||||
| 2323 | m_c_FAdd(const LHS &L, const RHS &R) { | ||||||
| 2324 | return BinaryOp_match<LHS, RHS, Instruction::FAdd, true>(L, R); | ||||||
| 2325 | } | ||||||
| 2326 | |||||||
| 2327 | /// Matches FMul with LHS and RHS in either order. | ||||||
| 2328 | template <typename LHS, typename RHS> | ||||||
| 2329 | inline BinaryOp_match<LHS, RHS, Instruction::FMul, true> | ||||||
| 2330 | m_c_FMul(const LHS &L, const RHS &R) { | ||||||
| 2331 | return BinaryOp_match<LHS, RHS, Instruction::FMul, true>(L, R); | ||||||
| 2332 | } | ||||||
| 2333 | |||||||
| 2334 | template <typename Opnd_t> struct Signum_match { | ||||||
| 2335 | Opnd_t Val; | ||||||
| 2336 | Signum_match(const Opnd_t &V) : Val(V) {} | ||||||
| 2337 | |||||||
| 2338 | template <typename OpTy> bool match(OpTy *V) { | ||||||
| 2339 | unsigned TypeSize = V->getType()->getScalarSizeInBits(); | ||||||
| 2340 | if (TypeSize == 0) | ||||||
| 2341 | return false; | ||||||
| 2342 | |||||||
| 2343 | unsigned ShiftWidth = TypeSize - 1; | ||||||
| 2344 | Value *OpL = nullptr, *OpR = nullptr; | ||||||
| 2345 | |||||||
| 2346 | // This is the representation of signum we match: | ||||||
| 2347 | // | ||||||
| 2348 | // signum(x) == (x >> 63) | (-x >>u 63) | ||||||
| 2349 | // | ||||||
| 2350 | // An i1 value is its own signum, so it's correct to match | ||||||
| 2351 | // | ||||||
| 2352 | // signum(x) == (x >> 0) | (-x >>u 0) | ||||||
| 2353 | // | ||||||
| 2354 | // for i1 values. | ||||||
| 2355 | |||||||
| 2356 | auto LHS = m_AShr(m_Value(OpL), m_SpecificInt(ShiftWidth)); | ||||||
| 2357 | auto RHS = m_LShr(m_Neg(m_Value(OpR)), m_SpecificInt(ShiftWidth)); | ||||||
| 2358 | auto Signum = m_Or(LHS, RHS); | ||||||
| 2359 | |||||||
| 2360 | return Signum.match(V) && OpL == OpR && Val.match(OpL); | ||||||
| 2361 | } | ||||||
| 2362 | }; | ||||||
| 2363 | |||||||
| 2364 | /// Matches a signum pattern. | ||||||
| 2365 | /// | ||||||
| 2366 | /// signum(x) = | ||||||
| 2367 | /// x > 0 -> 1 | ||||||
| 2368 | /// x == 0 -> 0 | ||||||
| 2369 | /// x < 0 -> -1 | ||||||
| 2370 | template <typename Val_t> inline Signum_match<Val_t> m_Signum(const Val_t &V) { | ||||||
| 2371 | return Signum_match<Val_t>(V); | ||||||
| 2372 | } | ||||||
| 2373 | |||||||
| 2374 | template <int Ind, typename Opnd_t> struct ExtractValue_match { | ||||||
| 2375 | Opnd_t Val; | ||||||
| 2376 | ExtractValue_match(const Opnd_t &V) : Val(V) {} | ||||||
| 2377 | |||||||
| 2378 | template <typename OpTy> bool match(OpTy *V) { | ||||||
| 2379 | if (auto *I = dyn_cast<ExtractValueInst>(V)) { | ||||||
| 2380 | // If Ind is -1, don't inspect indices | ||||||
| 2381 | if (Ind != -1 && | ||||||
| 2382 | !(I->getNumIndices() == 1 && I->getIndices()[0] == (unsigned)Ind)) | ||||||
| 2383 | return false; | ||||||
| 2384 | return Val.match(I->getAggregateOperand()); | ||||||
| 2385 | } | ||||||
| 2386 | return false; | ||||||
| 2387 | } | ||||||
| 2388 | }; | ||||||
| 2389 | |||||||
| 2390 | /// Match a single index ExtractValue instruction. | ||||||
| 2391 | /// For example m_ExtractValue<1>(...) | ||||||
| 2392 | template <int Ind, typename Val_t> | ||||||
| 2393 | inline ExtractValue_match<Ind, Val_t> m_ExtractValue(const Val_t &V) { | ||||||
| 2394 | return ExtractValue_match<Ind, Val_t>(V); | ||||||
| 2395 | } | ||||||
| 2396 | |||||||
| 2397 | /// Match an ExtractValue instruction with any index. | ||||||
| 2398 | /// For example m_ExtractValue(...) | ||||||
| 2399 | template <typename Val_t> | ||||||
| 2400 | inline ExtractValue_match<-1, Val_t> m_ExtractValue(const Val_t &V) { | ||||||
| 2401 | return ExtractValue_match<-1, Val_t>(V); | ||||||
| 2402 | } | ||||||
| 2403 | |||||||
| 2404 | /// Matcher for a single index InsertValue instruction. | ||||||
| 2405 | template <int Ind, typename T0, typename T1> struct InsertValue_match { | ||||||
| 2406 | T0 Op0; | ||||||
| 2407 | T1 Op1; | ||||||
| 2408 | |||||||
| 2409 | InsertValue_match(const T0 &Op0, const T1 &Op1) : Op0(Op0), Op1(Op1) {} | ||||||
| 2410 | |||||||
| 2411 | template <typename OpTy> bool match(OpTy *V) { | ||||||
| 2412 | if (auto *I = dyn_cast<InsertValueInst>(V)) { | ||||||
| 2413 | return Op0.match(I->getOperand(0)) && Op1.match(I->getOperand(1)) && | ||||||
| 2414 | I->getNumIndices() == 1 && Ind == I->getIndices()[0]; | ||||||
| 2415 | } | ||||||
| 2416 | return false; | ||||||
| 2417 | } | ||||||
| 2418 | }; | ||||||
| 2419 | |||||||
| 2420 | /// Matches a single index InsertValue instruction. | ||||||
| 2421 | template <int Ind, typename Val_t, typename Elt_t> | ||||||
| 2422 | inline InsertValue_match<Ind, Val_t, Elt_t> m_InsertValue(const Val_t &Val, | ||||||
| 2423 | const Elt_t &Elt) { | ||||||
| 2424 | return InsertValue_match<Ind, Val_t, Elt_t>(Val, Elt); | ||||||
| 2425 | } | ||||||
| 2426 | |||||||
| 2427 | /// Matches patterns for `vscale`. This can either be a call to `llvm.vscale` or | ||||||
| 2428 | /// the constant expression | ||||||
| 2429 | /// `ptrtoint(gep <vscale x 1 x i8>, <vscale x 1 x i8>* null, i32 1>` | ||||||
| 2430 | /// under the right conditions determined by DataLayout. | ||||||
| 2431 | struct VScaleVal_match { | ||||||
| 2432 | const DataLayout &DL; | ||||||
| 2433 | VScaleVal_match(const DataLayout &DL) : DL(DL) {} | ||||||
| 2434 | |||||||
| 2435 | template <typename ITy> bool match(ITy *V) { | ||||||
| 2436 | if (m_Intrinsic<Intrinsic::vscale>().match(V)) | ||||||
| 2437 | return true; | ||||||
| 2438 | |||||||
| 2439 | Value *Ptr; | ||||||
| 2440 | if (m_PtrToInt(m_Value(Ptr)).match(V)) { | ||||||
| 2441 | if (auto *GEP = dyn_cast<GEPOperator>(Ptr)) { | ||||||
| 2442 | auto *DerefTy = GEP->getSourceElementType(); | ||||||
| 2443 | if (GEP->getNumIndices() == 1 && isa<ScalableVectorType>(DerefTy) && | ||||||
| 2444 | m_Zero().match(GEP->getPointerOperand()) && | ||||||
| 2445 | m_SpecificInt(1).match(GEP->idx_begin()->get()) && | ||||||
| 2446 | DL.getTypeAllocSizeInBits(DerefTy).getKnownMinSize() == 8) | ||||||
| 2447 | return true; | ||||||
| 2448 | } | ||||||
| 2449 | } | ||||||
| 2450 | |||||||
| 2451 | return false; | ||||||
| 2452 | } | ||||||
| 2453 | }; | ||||||
| 2454 | |||||||
| 2455 | inline VScaleVal_match m_VScale(const DataLayout &DL) { | ||||||
| 2456 | return VScaleVal_match(DL); | ||||||
| 2457 | } | ||||||
| 2458 | |||||||
| 2459 | template <typename LHS, typename RHS, unsigned Opcode> | ||||||
| 2460 | struct LogicalOp_match { | ||||||
| 2461 | LHS L; | ||||||
| 2462 | RHS R; | ||||||
| 2463 | |||||||
| 2464 | LogicalOp_match(const LHS &L, const RHS &R) : L(L), R(R) {} | ||||||
| 2465 | |||||||
| 2466 | template <typename T> bool match(T *V) { | ||||||
| 2467 | if (auto *I = dyn_cast<Instruction>(V)) { | ||||||
| 2468 | if (!I->getType()->isIntOrIntVectorTy(1)) | ||||||
| 2469 | return false; | ||||||
| 2470 | |||||||
| 2471 | if (I->getOpcode() == Opcode && L.match(I->getOperand(0)) && | ||||||
| 2472 | R.match(I->getOperand(1))) | ||||||
| 2473 | return true; | ||||||
| 2474 | |||||||
| 2475 | if (auto *SI = dyn_cast<SelectInst>(I)) { | ||||||
| 2476 | if (Opcode == Instruction::And) { | ||||||
| 2477 | if (const auto *C = dyn_cast<Constant>(SI->getFalseValue())) | ||||||
| 2478 | if (C->isNullValue() && L.match(SI->getCondition()) && | ||||||
| 2479 | R.match(SI->getTrueValue())) | ||||||
| 2480 | return true; | ||||||
| 2481 | } else { | ||||||
| 2482 | assert(Opcode == Instruction::Or)((void)0); | ||||||
| 2483 | if (const auto *C = dyn_cast<Constant>(SI->getTrueValue())) | ||||||
| 2484 | if (C->isOneValue() && L.match(SI->getCondition()) && | ||||||
| 2485 | R.match(SI->getFalseValue())) | ||||||
| 2486 | return true; | ||||||
| 2487 | } | ||||||
| 2488 | } | ||||||
| 2489 | } | ||||||
| 2490 | |||||||
| 2491 | return false; | ||||||
| 2492 | } | ||||||
| 2493 | }; | ||||||
| 2494 | |||||||
| 2495 | /// Matches L && R either in the form of L & R or L ? R : false. | ||||||
| 2496 | /// Note that the latter form is poison-blocking. | ||||||
| 2497 | template <typename LHS, typename RHS> | ||||||
| 2498 | inline LogicalOp_match<LHS, RHS, Instruction::And> | ||||||
| 2499 | m_LogicalAnd(const LHS &L, const RHS &R) { | ||||||
| 2500 | return LogicalOp_match<LHS, RHS, Instruction::And>(L, R); | ||||||
| 2501 | } | ||||||
| 2502 | |||||||
| 2503 | /// Matches L && R where L and R are arbitrary values. | ||||||
| 2504 | inline auto m_LogicalAnd() { return m_LogicalAnd(m_Value(), m_Value()); } | ||||||
| 2505 | |||||||
| 2506 | /// Matches L || R either in the form of L | R or L ? true : R. | ||||||
| 2507 | /// Note that the latter form is poison-blocking. | ||||||
| 2508 | template <typename LHS, typename RHS> | ||||||
| 2509 | inline LogicalOp_match<LHS, RHS, Instruction::Or> | ||||||
| 2510 | m_LogicalOr(const LHS &L, const RHS &R) { | ||||||
| 2511 | return LogicalOp_match<LHS, RHS, Instruction::Or>(L, R); | ||||||
| 2512 | } | ||||||
| 2513 | |||||||
| 2514 | /// Matches L || R where L and R are arbitrary values. | ||||||
| 2515 | inline auto m_LogicalOr() { | ||||||
| 2516 | return m_LogicalOr(m_Value(), m_Value()); | ||||||
| 2517 | } | ||||||
| 2518 | |||||||
| 2519 | } // end namespace PatternMatch | ||||||
| 2520 | } // end namespace llvm | ||||||
| 2521 | |||||||
| 2522 | #endif // LLVM_IR_PATTERNMATCH_H |