Bug Summary

File:src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/lib/Target/X86/X86TargetTransformInfo.cpp
Warning:line 4650, column 27
Called C++ object pointer is null

Annotated Source Code

Press '?' to see keyboard shortcuts

clang -cc1 -cc1 -triple amd64-unknown-openbsd7.0 -analyze -disable-free -disable-llvm-verifier -discard-value-names -main-file-name X86TargetTransformInfo.cpp -analyzer-store=region -analyzer-opt-analyze-nested-blocks -analyzer-checker=core -analyzer-checker=apiModeling -analyzer-checker=unix -analyzer-checker=deadcode -analyzer-checker=cplusplus -analyzer-checker=security.insecureAPI.UncheckedReturn -analyzer-checker=security.insecureAPI.getpw -analyzer-checker=security.insecureAPI.gets -analyzer-checker=security.insecureAPI.mktemp -analyzer-checker=security.insecureAPI.mkstemp -analyzer-checker=security.insecureAPI.vfork -analyzer-checker=nullability.NullPassedToNonnull -analyzer-checker=nullability.NullReturnedFromNonnull -analyzer-output plist -w -setup-static-analyzer -mrelocation-model static -mframe-pointer=all -relaxed-aliasing -fno-rounding-math -mconstructor-aliases -munwind-tables -target-cpu x86-64 -tune-cpu generic -debugger-tuning=gdb -fcoverage-compilation-dir=/usr/src/gnu/usr.bin/clang/libLLVM/obj -resource-dir /usr/local/lib/clang/13.0.0 -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/Transforms -I /usr/src/gnu/usr.bin/clang/libLLVM/obj/../include/llvm/AMDGPU -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/lib/Target/AMDGPU -I /usr/src/gnu/usr.bin/clang/libLLVM/obj/../include/llvm/AMDGPU -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/lib/Target/AMDGPU -I /usr/src/gnu/usr.bin/clang/libLLVM/obj/../include/llvm/AMDGPU -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/lib/Target/AMDGPU -I /usr/src/gnu/usr.bin/clang/libLLVM/obj/../include/llvm/AMDGPU -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/lib/Target/AMDGPU -I /usr/src/gnu/usr.bin/clang/libLLVM/obj/../include/llvm/AMDGPU -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/lib/Target/AMDGPU -I /usr/src/gnu/usr.bin/clang/libLLVM/obj/../include/llvm/AMDGPU -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/lib/Target/AMDGPU -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/Analysis -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/ASMParser -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/BinaryFormat -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/Bitcode -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/Bitcode -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/Bitstream -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/Transforms -I /include/llvm/CodeGen -I /include/llvm/CodeGen/PBQP -I /usr/src/gnu/usr.bin/clang/libLLVM/obj/../include/llvm/IR -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/IR -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/Transforms -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/Transforms/Coroutines -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/ProfileData/Coverage -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/DebugInfo/CodeView -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/DebugInfo/DWARF -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/DebugInfo -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/DebugInfo/MSF -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/DebugInfo/PDB -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/Demangle -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/ExecutionEngine -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/ExecutionEngine/JITLink -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/ExecutionEngine/Orc -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/Frontend -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/Frontend/OpenACC -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/Frontend -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/Frontend/OpenMP -I /include/llvm/CodeGen/GlobalISel -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/IRReader -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/Transforms -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/Transforms/InstCombine -I /usr/src/gnu/usr.bin/clang/libLLVM/obj/../include/llvm/Transforms/InstCombine -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/Transforms -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/LTO -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/Linker -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/MC -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/MC/MCParser -I /include/llvm/CodeGen/MIRParser -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/Transforms -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/Object -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/Option -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/Passes -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/ -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/ProfileData -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/Transforms -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/Transforms/Scalar -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/ADT -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/Support -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/DebugInfo/Symbolize -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/Target -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/Transforms -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/Transforms/Utils -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/Transforms -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/Transforms/Vectorize -I /usr/src/gnu/usr.bin/clang/libLLVM/obj/../include/llvm/X86 -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/lib/Target/X86 -I /usr/src/gnu/usr.bin/clang/libLLVM/obj/../include/llvm/X86 -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/lib/Target/X86 -I /usr/src/gnu/usr.bin/clang/libLLVM/obj/../include/llvm/X86 -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/lib/Target/X86 -I /usr/src/gnu/usr.bin/clang/libLLVM/obj/../include/llvm/X86 -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/lib/Target/X86 -I /usr/src/gnu/usr.bin/clang/libLLVM/obj/../include/llvm/X86 -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/lib/Target/X86 -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/Transforms -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/Transforms/IPO -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include -I /usr/src/gnu/usr.bin/clang/libLLVM/../include -I /usr/src/gnu/usr.bin/clang/libLLVM/obj -I /usr/src/gnu/usr.bin/clang/libLLVM/obj/../include -D NDEBUG -D __STDC_LIMIT_MACROS -D __STDC_CONSTANT_MACROS -D __STDC_FORMAT_MACROS -D LLVM_PREFIX="/usr" -internal-isystem /usr/include/c++/v1 -internal-isystem /usr/local/lib/clang/13.0.0/include -internal-externc-isystem /usr/include -O2 -Wno-unused-parameter -Wwrite-strings -Wno-missing-field-initializers -Wno-long-long -Wno-comment -std=c++14 -fdeprecated-macro -fdebug-compilation-dir=/usr/src/gnu/usr.bin/clang/libLLVM/obj -ferror-limit 19 -fvisibility-inlines-hidden -fwrapv -stack-protector 2 -fno-rtti -fgnuc-version=4.2.1 -vectorize-loops -vectorize-slp -fno-builtin-malloc -fno-builtin-calloc -fno-builtin-realloc -fno-builtin-valloc -fno-builtin-free -fno-builtin-strdup -fno-builtin-strndup -analyzer-output=html -faddrsig -D__GCC_HAVE_DWARF2_CFI_ASM=1 -o /home/ben/Projects/vmm/scan-build/2022-01-12-194120-40624-1 -x c++ /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/lib/Target/X86/X86TargetTransformInfo.cpp
1//===-- X86TargetTransformInfo.cpp - X86 specific TTI pass ----------------===//
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/// \file
9/// This file implements a TargetTransformInfo analysis pass specific to the
10/// X86 target machine. It uses the target's detailed information to provide
11/// more precise answers to certain TTI queries, while letting the target
12/// independent and default TTI implementations handle the rest.
13///
14//===----------------------------------------------------------------------===//
15/// About Cost Model numbers used below it's necessary to say the following:
16/// the numbers correspond to some "generic" X86 CPU instead of usage of
17/// concrete CPU model. Usually the numbers correspond to CPU where the feature
18/// apeared at the first time. For example, if we do Subtarget.hasSSE42() in
19/// the lookups below the cost is based on Nehalem as that was the first CPU
20/// to support that feature level and thus has most likely the worst case cost.
21/// Some examples of other technologies/CPUs:
22/// SSE 3 - Pentium4 / Athlon64
23/// SSE 4.1 - Penryn
24/// SSE 4.2 - Nehalem
25/// AVX - Sandy Bridge
26/// AVX2 - Haswell
27/// AVX-512 - Xeon Phi / Skylake
28/// And some examples of instruction target dependent costs (latency)
29/// divss sqrtss rsqrtss
30/// AMD K7 11-16 19 3
31/// Piledriver 9-24 13-15 5
32/// Jaguar 14 16 2
33/// Pentium II,III 18 30 2
34/// Nehalem 7-14 7-18 3
35/// Haswell 10-13 11 5
36/// TODO: Develop and implement the target dependent cost model and
37/// specialize cost numbers for different Cost Model Targets such as throughput,
38/// code size, latency and uop count.
39//===----------------------------------------------------------------------===//
40
41#include "X86TargetTransformInfo.h"
42#include "llvm/Analysis/TargetTransformInfo.h"
43#include "llvm/CodeGen/BasicTTIImpl.h"
44#include "llvm/CodeGen/CostTable.h"
45#include "llvm/CodeGen/TargetLowering.h"
46#include "llvm/IR/IntrinsicInst.h"
47#include "llvm/Support/Debug.h"
48
49using namespace llvm;
50
51#define DEBUG_TYPE"x86tti" "x86tti"
52
53//===----------------------------------------------------------------------===//
54//
55// X86 cost model.
56//
57//===----------------------------------------------------------------------===//
58
59TargetTransformInfo::PopcntSupportKind
60X86TTIImpl::getPopcntSupport(unsigned TyWidth) {
61 assert(isPowerOf2_32(TyWidth) && "Ty width must be power of 2")((void)0);
62 // TODO: Currently the __builtin_popcount() implementation using SSE3
63 // instructions is inefficient. Once the problem is fixed, we should
64 // call ST->hasSSE3() instead of ST->hasPOPCNT().
65 return ST->hasPOPCNT() ? TTI::PSK_FastHardware : TTI::PSK_Software;
66}
67
68llvm::Optional<unsigned> X86TTIImpl::getCacheSize(
69 TargetTransformInfo::CacheLevel Level) const {
70 switch (Level) {
71 case TargetTransformInfo::CacheLevel::L1D:
72 // - Penryn
73 // - Nehalem
74 // - Westmere
75 // - Sandy Bridge
76 // - Ivy Bridge
77 // - Haswell
78 // - Broadwell
79 // - Skylake
80 // - Kabylake
81 return 32 * 1024; // 32 KByte
82 case TargetTransformInfo::CacheLevel::L2D:
83 // - Penryn
84 // - Nehalem
85 // - Westmere
86 // - Sandy Bridge
87 // - Ivy Bridge
88 // - Haswell
89 // - Broadwell
90 // - Skylake
91 // - Kabylake
92 return 256 * 1024; // 256 KByte
93 }
94
95 llvm_unreachable("Unknown TargetTransformInfo::CacheLevel")__builtin_unreachable();
96}
97
98llvm::Optional<unsigned> X86TTIImpl::getCacheAssociativity(
99 TargetTransformInfo::CacheLevel Level) const {
100 // - Penryn
101 // - Nehalem
102 // - Westmere
103 // - Sandy Bridge
104 // - Ivy Bridge
105 // - Haswell
106 // - Broadwell
107 // - Skylake
108 // - Kabylake
109 switch (Level) {
110 case TargetTransformInfo::CacheLevel::L1D:
111 LLVM_FALLTHROUGH[[gnu::fallthrough]];
112 case TargetTransformInfo::CacheLevel::L2D:
113 return 8;
114 }
115
116 llvm_unreachable("Unknown TargetTransformInfo::CacheLevel")__builtin_unreachable();
117}
118
119unsigned X86TTIImpl::getNumberOfRegisters(unsigned ClassID) const {
120 bool Vector = (ClassID == 1);
121 if (Vector && !ST->hasSSE1())
122 return 0;
123
124 if (ST->is64Bit()) {
125 if (Vector && ST->hasAVX512())
126 return 32;
127 return 16;
128 }
129 return 8;
130}
131
132TypeSize
133X86TTIImpl::getRegisterBitWidth(TargetTransformInfo::RegisterKind K) const {
134 unsigned PreferVectorWidth = ST->getPreferVectorWidth();
135 switch (K) {
136 case TargetTransformInfo::RGK_Scalar:
137 return TypeSize::getFixed(ST->is64Bit() ? 64 : 32);
138 case TargetTransformInfo::RGK_FixedWidthVector:
139 if (ST->hasAVX512() && PreferVectorWidth >= 512)
140 return TypeSize::getFixed(512);
141 if (ST->hasAVX() && PreferVectorWidth >= 256)
142 return TypeSize::getFixed(256);
143 if (ST->hasSSE1() && PreferVectorWidth >= 128)
144 return TypeSize::getFixed(128);
145 return TypeSize::getFixed(0);
146 case TargetTransformInfo::RGK_ScalableVector:
147 return TypeSize::getScalable(0);
148 }
149
150 llvm_unreachable("Unsupported register kind")__builtin_unreachable();
151}
152
153unsigned X86TTIImpl::getLoadStoreVecRegBitWidth(unsigned) const {
154 return getRegisterBitWidth(TargetTransformInfo::RGK_FixedWidthVector)
155 .getFixedSize();
156}
157
158unsigned X86TTIImpl::getMaxInterleaveFactor(unsigned VF) {
159 // If the loop will not be vectorized, don't interleave the loop.
160 // Let regular unroll to unroll the loop, which saves the overflow
161 // check and memory check cost.
162 if (VF == 1)
163 return 1;
164
165 if (ST->isAtom())
166 return 1;
167
168 // Sandybridge and Haswell have multiple execution ports and pipelined
169 // vector units.
170 if (ST->hasAVX())
171 return 4;
172
173 return 2;
174}
175
176InstructionCost X86TTIImpl::getArithmeticInstrCost(
177 unsigned Opcode, Type *Ty, TTI::TargetCostKind CostKind,
178 TTI::OperandValueKind Op1Info, TTI::OperandValueKind Op2Info,
179 TTI::OperandValueProperties Opd1PropInfo,
180 TTI::OperandValueProperties Opd2PropInfo, ArrayRef<const Value *> Args,
181 const Instruction *CxtI) {
182 // TODO: Handle more cost kinds.
183 if (CostKind != TTI::TCK_RecipThroughput)
184 return BaseT::getArithmeticInstrCost(Opcode, Ty, CostKind, Op1Info,
185 Op2Info, Opd1PropInfo,
186 Opd2PropInfo, Args, CxtI);
187
188 // vXi8 multiplications are always promoted to vXi16.
189 if (Opcode == Instruction::Mul && Ty->isVectorTy() &&
190 Ty->getScalarSizeInBits() == 8) {
191 Type *WideVecTy =
192 VectorType::getExtendedElementVectorType(cast<VectorType>(Ty));
193 return getCastInstrCost(Instruction::ZExt, WideVecTy, Ty,
194 TargetTransformInfo::CastContextHint::None,
195 CostKind) +
196 getCastInstrCost(Instruction::Trunc, Ty, WideVecTy,
197 TargetTransformInfo::CastContextHint::None,
198 CostKind) +
199 getArithmeticInstrCost(Opcode, WideVecTy, CostKind, Op1Info, Op2Info,
200 Opd1PropInfo, Opd2PropInfo);
201 }
202
203 // Legalize the type.
204 std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, Ty);
205
206 int ISD = TLI->InstructionOpcodeToISD(Opcode);
207 assert(ISD && "Invalid opcode")((void)0);
208
209 static const CostTblEntry GLMCostTable[] = {
210 { ISD::FDIV, MVT::f32, 18 }, // divss
211 { ISD::FDIV, MVT::v4f32, 35 }, // divps
212 { ISD::FDIV, MVT::f64, 33 }, // divsd
213 { ISD::FDIV, MVT::v2f64, 65 }, // divpd
214 };
215
216 if (ST->useGLMDivSqrtCosts())
217 if (const auto *Entry = CostTableLookup(GLMCostTable, ISD,
218 LT.second))
219 return LT.first * Entry->Cost;
220
221 static const CostTblEntry SLMCostTable[] = {
222 { ISD::MUL, MVT::v4i32, 11 }, // pmulld
223 { ISD::MUL, MVT::v8i16, 2 }, // pmullw
224 { ISD::FMUL, MVT::f64, 2 }, // mulsd
225 { ISD::FMUL, MVT::v2f64, 4 }, // mulpd
226 { ISD::FMUL, MVT::v4f32, 2 }, // mulps
227 { ISD::FDIV, MVT::f32, 17 }, // divss
228 { ISD::FDIV, MVT::v4f32, 39 }, // divps
229 { ISD::FDIV, MVT::f64, 32 }, // divsd
230 { ISD::FDIV, MVT::v2f64, 69 }, // divpd
231 { ISD::FADD, MVT::v2f64, 2 }, // addpd
232 { ISD::FSUB, MVT::v2f64, 2 }, // subpd
233 // v2i64/v4i64 mul is custom lowered as a series of long:
234 // multiplies(3), shifts(3) and adds(2)
235 // slm muldq version throughput is 2 and addq throughput 4
236 // thus: 3X2 (muldq throughput) + 3X1 (shift throughput) +
237 // 3X4 (addq throughput) = 17
238 { ISD::MUL, MVT::v2i64, 17 },
239 // slm addq\subq throughput is 4
240 { ISD::ADD, MVT::v2i64, 4 },
241 { ISD::SUB, MVT::v2i64, 4 },
242 };
243
244 if (ST->isSLM()) {
245 if (Args.size() == 2 && ISD == ISD::MUL && LT.second == MVT::v4i32) {
246 // Check if the operands can be shrinked into a smaller datatype.
247 bool Op1Signed = false;
248 unsigned Op1MinSize = BaseT::minRequiredElementSize(Args[0], Op1Signed);
249 bool Op2Signed = false;
250 unsigned Op2MinSize = BaseT::minRequiredElementSize(Args[1], Op2Signed);
251
252 bool SignedMode = Op1Signed || Op2Signed;
253 unsigned OpMinSize = std::max(Op1MinSize, Op2MinSize);
254
255 if (OpMinSize <= 7)
256 return LT.first * 3; // pmullw/sext
257 if (!SignedMode && OpMinSize <= 8)
258 return LT.first * 3; // pmullw/zext
259 if (OpMinSize <= 15)
260 return LT.first * 5; // pmullw/pmulhw/pshuf
261 if (!SignedMode && OpMinSize <= 16)
262 return LT.first * 5; // pmullw/pmulhw/pshuf
263 }
264
265 if (const auto *Entry = CostTableLookup(SLMCostTable, ISD,
266 LT.second)) {
267 return LT.first * Entry->Cost;
268 }
269 }
270
271 if ((ISD == ISD::SDIV || ISD == ISD::SREM || ISD == ISD::UDIV ||
272 ISD == ISD::UREM) &&
273 (Op2Info == TargetTransformInfo::OK_UniformConstantValue ||
274 Op2Info == TargetTransformInfo::OK_NonUniformConstantValue) &&
275 Opd2PropInfo == TargetTransformInfo::OP_PowerOf2) {
276 if (ISD == ISD::SDIV || ISD == ISD::SREM) {
277 // On X86, vector signed division by constants power-of-two are
278 // normally expanded to the sequence SRA + SRL + ADD + SRA.
279 // The OperandValue properties may not be the same as that of the previous
280 // operation; conservatively assume OP_None.
281 InstructionCost Cost =
282 2 * getArithmeticInstrCost(Instruction::AShr, Ty, CostKind, Op1Info,
283 Op2Info, TargetTransformInfo::OP_None,
284 TargetTransformInfo::OP_None);
285 Cost += getArithmeticInstrCost(Instruction::LShr, Ty, CostKind, Op1Info,
286 Op2Info,
287 TargetTransformInfo::OP_None,
288 TargetTransformInfo::OP_None);
289 Cost += getArithmeticInstrCost(Instruction::Add, Ty, CostKind, Op1Info,
290 Op2Info,
291 TargetTransformInfo::OP_None,
292 TargetTransformInfo::OP_None);
293
294 if (ISD == ISD::SREM) {
295 // For SREM: (X % C) is the equivalent of (X - (X/C)*C)
296 Cost += getArithmeticInstrCost(Instruction::Mul, Ty, CostKind, Op1Info,
297 Op2Info);
298 Cost += getArithmeticInstrCost(Instruction::Sub, Ty, CostKind, Op1Info,
299 Op2Info);
300 }
301
302 return Cost;
303 }
304
305 // Vector unsigned division/remainder will be simplified to shifts/masks.
306 if (ISD == ISD::UDIV)
307 return getArithmeticInstrCost(Instruction::LShr, Ty, CostKind,
308 Op1Info, Op2Info,
309 TargetTransformInfo::OP_None,
310 TargetTransformInfo::OP_None);
311
312 else // UREM
313 return getArithmeticInstrCost(Instruction::And, Ty, CostKind,
314 Op1Info, Op2Info,
315 TargetTransformInfo::OP_None,
316 TargetTransformInfo::OP_None);
317 }
318
319 static const CostTblEntry AVX512BWUniformConstCostTable[] = {
320 { ISD::SHL, MVT::v64i8, 2 }, // psllw + pand.
321 { ISD::SRL, MVT::v64i8, 2 }, // psrlw + pand.
322 { ISD::SRA, MVT::v64i8, 4 }, // psrlw, pand, pxor, psubb.
323 };
324
325 if (Op2Info == TargetTransformInfo::OK_UniformConstantValue &&
326 ST->hasBWI()) {
327 if (const auto *Entry = CostTableLookup(AVX512BWUniformConstCostTable, ISD,
328 LT.second))
329 return LT.first * Entry->Cost;
330 }
331
332 static const CostTblEntry AVX512UniformConstCostTable[] = {
333 { ISD::SRA, MVT::v2i64, 1 },
334 { ISD::SRA, MVT::v4i64, 1 },
335 { ISD::SRA, MVT::v8i64, 1 },
336
337 { ISD::SHL, MVT::v64i8, 4 }, // psllw + pand.
338 { ISD::SRL, MVT::v64i8, 4 }, // psrlw + pand.
339 { ISD::SRA, MVT::v64i8, 8 }, // psrlw, pand, pxor, psubb.
340
341 { ISD::SDIV, MVT::v16i32, 6 }, // pmuludq sequence
342 { ISD::SREM, MVT::v16i32, 8 }, // pmuludq+mul+sub sequence
343 { ISD::UDIV, MVT::v16i32, 5 }, // pmuludq sequence
344 { ISD::UREM, MVT::v16i32, 7 }, // pmuludq+mul+sub sequence
345 };
346
347 if (Op2Info == TargetTransformInfo::OK_UniformConstantValue &&
348 ST->hasAVX512()) {
349 if (const auto *Entry = CostTableLookup(AVX512UniformConstCostTable, ISD,
350 LT.second))
351 return LT.first * Entry->Cost;
352 }
353
354 static const CostTblEntry AVX2UniformConstCostTable[] = {
355 { ISD::SHL, MVT::v32i8, 2 }, // psllw + pand.
356 { ISD::SRL, MVT::v32i8, 2 }, // psrlw + pand.
357 { ISD::SRA, MVT::v32i8, 4 }, // psrlw, pand, pxor, psubb.
358
359 { ISD::SRA, MVT::v4i64, 4 }, // 2 x psrad + shuffle.
360
361 { ISD::SDIV, MVT::v8i32, 6 }, // pmuludq sequence
362 { ISD::SREM, MVT::v8i32, 8 }, // pmuludq+mul+sub sequence
363 { ISD::UDIV, MVT::v8i32, 5 }, // pmuludq sequence
364 { ISD::UREM, MVT::v8i32, 7 }, // pmuludq+mul+sub sequence
365 };
366
367 if (Op2Info == TargetTransformInfo::OK_UniformConstantValue &&
368 ST->hasAVX2()) {
369 if (const auto *Entry = CostTableLookup(AVX2UniformConstCostTable, ISD,
370 LT.second))
371 return LT.first * Entry->Cost;
372 }
373
374 static const CostTblEntry SSE2UniformConstCostTable[] = {
375 { ISD::SHL, MVT::v16i8, 2 }, // psllw + pand.
376 { ISD::SRL, MVT::v16i8, 2 }, // psrlw + pand.
377 { ISD::SRA, MVT::v16i8, 4 }, // psrlw, pand, pxor, psubb.
378
379 { ISD::SHL, MVT::v32i8, 4+2 }, // 2*(psllw + pand) + split.
380 { ISD::SRL, MVT::v32i8, 4+2 }, // 2*(psrlw + pand) + split.
381 { ISD::SRA, MVT::v32i8, 8+2 }, // 2*(psrlw, pand, pxor, psubb) + split.
382
383 { ISD::SDIV, MVT::v8i32, 12+2 }, // 2*pmuludq sequence + split.
384 { ISD::SREM, MVT::v8i32, 16+2 }, // 2*pmuludq+mul+sub sequence + split.
385 { ISD::SDIV, MVT::v4i32, 6 }, // pmuludq sequence
386 { ISD::SREM, MVT::v4i32, 8 }, // pmuludq+mul+sub sequence
387 { ISD::UDIV, MVT::v8i32, 10+2 }, // 2*pmuludq sequence + split.
388 { ISD::UREM, MVT::v8i32, 14+2 }, // 2*pmuludq+mul+sub sequence + split.
389 { ISD::UDIV, MVT::v4i32, 5 }, // pmuludq sequence
390 { ISD::UREM, MVT::v4i32, 7 }, // pmuludq+mul+sub sequence
391 };
392
393 // XOP has faster vXi8 shifts.
394 if (Op2Info == TargetTransformInfo::OK_UniformConstantValue &&
395 ST->hasSSE2() && !ST->hasXOP()) {
396 if (const auto *Entry =
397 CostTableLookup(SSE2UniformConstCostTable, ISD, LT.second))
398 return LT.first * Entry->Cost;
399 }
400
401 static const CostTblEntry AVX512BWConstCostTable[] = {
402 { ISD::SDIV, MVT::v64i8, 14 }, // 2*ext+2*pmulhw sequence
403 { ISD::SREM, MVT::v64i8, 16 }, // 2*ext+2*pmulhw+mul+sub sequence
404 { ISD::UDIV, MVT::v64i8, 14 }, // 2*ext+2*pmulhw sequence
405 { ISD::UREM, MVT::v64i8, 16 }, // 2*ext+2*pmulhw+mul+sub sequence
406 { ISD::SDIV, MVT::v32i16, 6 }, // vpmulhw sequence
407 { ISD::SREM, MVT::v32i16, 8 }, // vpmulhw+mul+sub sequence
408 { ISD::UDIV, MVT::v32i16, 6 }, // vpmulhuw sequence
409 { ISD::UREM, MVT::v32i16, 8 }, // vpmulhuw+mul+sub sequence
410 };
411
412 if ((Op2Info == TargetTransformInfo::OK_UniformConstantValue ||
413 Op2Info == TargetTransformInfo::OK_NonUniformConstantValue) &&
414 ST->hasBWI()) {
415 if (const auto *Entry =
416 CostTableLookup(AVX512BWConstCostTable, ISD, LT.second))
417 return LT.first * Entry->Cost;
418 }
419
420 static const CostTblEntry AVX512ConstCostTable[] = {
421 { ISD::SDIV, MVT::v16i32, 15 }, // vpmuldq sequence
422 { ISD::SREM, MVT::v16i32, 17 }, // vpmuldq+mul+sub sequence
423 { ISD::UDIV, MVT::v16i32, 15 }, // vpmuludq sequence
424 { ISD::UREM, MVT::v16i32, 17 }, // vpmuludq+mul+sub sequence
425 { ISD::SDIV, MVT::v64i8, 28 }, // 4*ext+4*pmulhw sequence
426 { ISD::SREM, MVT::v64i8, 32 }, // 4*ext+4*pmulhw+mul+sub sequence
427 { ISD::UDIV, MVT::v64i8, 28 }, // 4*ext+4*pmulhw sequence
428 { ISD::UREM, MVT::v64i8, 32 }, // 4*ext+4*pmulhw+mul+sub sequence
429 { ISD::SDIV, MVT::v32i16, 12 }, // 2*vpmulhw sequence
430 { ISD::SREM, MVT::v32i16, 16 }, // 2*vpmulhw+mul+sub sequence
431 { ISD::UDIV, MVT::v32i16, 12 }, // 2*vpmulhuw sequence
432 { ISD::UREM, MVT::v32i16, 16 }, // 2*vpmulhuw+mul+sub sequence
433 };
434
435 if ((Op2Info == TargetTransformInfo::OK_UniformConstantValue ||
436 Op2Info == TargetTransformInfo::OK_NonUniformConstantValue) &&
437 ST->hasAVX512()) {
438 if (const auto *Entry =
439 CostTableLookup(AVX512ConstCostTable, ISD, LT.second))
440 return LT.first * Entry->Cost;
441 }
442
443 static const CostTblEntry AVX2ConstCostTable[] = {
444 { ISD::SDIV, MVT::v32i8, 14 }, // 2*ext+2*pmulhw sequence
445 { ISD::SREM, MVT::v32i8, 16 }, // 2*ext+2*pmulhw+mul+sub sequence
446 { ISD::UDIV, MVT::v32i8, 14 }, // 2*ext+2*pmulhw sequence
447 { ISD::UREM, MVT::v32i8, 16 }, // 2*ext+2*pmulhw+mul+sub sequence
448 { ISD::SDIV, MVT::v16i16, 6 }, // vpmulhw sequence
449 { ISD::SREM, MVT::v16i16, 8 }, // vpmulhw+mul+sub sequence
450 { ISD::UDIV, MVT::v16i16, 6 }, // vpmulhuw sequence
451 { ISD::UREM, MVT::v16i16, 8 }, // vpmulhuw+mul+sub sequence
452 { ISD::SDIV, MVT::v8i32, 15 }, // vpmuldq sequence
453 { ISD::SREM, MVT::v8i32, 19 }, // vpmuldq+mul+sub sequence
454 { ISD::UDIV, MVT::v8i32, 15 }, // vpmuludq sequence
455 { ISD::UREM, MVT::v8i32, 19 }, // vpmuludq+mul+sub sequence
456 };
457
458 if ((Op2Info == TargetTransformInfo::OK_UniformConstantValue ||
459 Op2Info == TargetTransformInfo::OK_NonUniformConstantValue) &&
460 ST->hasAVX2()) {
461 if (const auto *Entry = CostTableLookup(AVX2ConstCostTable, ISD, LT.second))
462 return LT.first * Entry->Cost;
463 }
464
465 static const CostTblEntry SSE2ConstCostTable[] = {
466 { ISD::SDIV, MVT::v32i8, 28+2 }, // 4*ext+4*pmulhw sequence + split.
467 { ISD::SREM, MVT::v32i8, 32+2 }, // 4*ext+4*pmulhw+mul+sub sequence + split.
468 { ISD::SDIV, MVT::v16i8, 14 }, // 2*ext+2*pmulhw sequence
469 { ISD::SREM, MVT::v16i8, 16 }, // 2*ext+2*pmulhw+mul+sub sequence
470 { ISD::UDIV, MVT::v32i8, 28+2 }, // 4*ext+4*pmulhw sequence + split.
471 { ISD::UREM, MVT::v32i8, 32+2 }, // 4*ext+4*pmulhw+mul+sub sequence + split.
472 { ISD::UDIV, MVT::v16i8, 14 }, // 2*ext+2*pmulhw sequence
473 { ISD::UREM, MVT::v16i8, 16 }, // 2*ext+2*pmulhw+mul+sub sequence
474 { ISD::SDIV, MVT::v16i16, 12+2 }, // 2*pmulhw sequence + split.
475 { ISD::SREM, MVT::v16i16, 16+2 }, // 2*pmulhw+mul+sub sequence + split.
476 { ISD::SDIV, MVT::v8i16, 6 }, // pmulhw sequence
477 { ISD::SREM, MVT::v8i16, 8 }, // pmulhw+mul+sub sequence
478 { ISD::UDIV, MVT::v16i16, 12+2 }, // 2*pmulhuw sequence + split.
479 { ISD::UREM, MVT::v16i16, 16+2 }, // 2*pmulhuw+mul+sub sequence + split.
480 { ISD::UDIV, MVT::v8i16, 6 }, // pmulhuw sequence
481 { ISD::UREM, MVT::v8i16, 8 }, // pmulhuw+mul+sub sequence
482 { ISD::SDIV, MVT::v8i32, 38+2 }, // 2*pmuludq sequence + split.
483 { ISD::SREM, MVT::v8i32, 48+2 }, // 2*pmuludq+mul+sub sequence + split.
484 { ISD::SDIV, MVT::v4i32, 19 }, // pmuludq sequence
485 { ISD::SREM, MVT::v4i32, 24 }, // pmuludq+mul+sub sequence
486 { ISD::UDIV, MVT::v8i32, 30+2 }, // 2*pmuludq sequence + split.
487 { ISD::UREM, MVT::v8i32, 40+2 }, // 2*pmuludq+mul+sub sequence + split.
488 { ISD::UDIV, MVT::v4i32, 15 }, // pmuludq sequence
489 { ISD::UREM, MVT::v4i32, 20 }, // pmuludq+mul+sub sequence
490 };
491
492 if ((Op2Info == TargetTransformInfo::OK_UniformConstantValue ||
493 Op2Info == TargetTransformInfo::OK_NonUniformConstantValue) &&
494 ST->hasSSE2()) {
495 // pmuldq sequence.
496 if (ISD == ISD::SDIV && LT.second == MVT::v8i32 && ST->hasAVX())
497 return LT.first * 32;
498 if (ISD == ISD::SREM && LT.second == MVT::v8i32 && ST->hasAVX())
499 return LT.first * 38;
500 if (ISD == ISD::SDIV && LT.second == MVT::v4i32 && ST->hasSSE41())
501 return LT.first * 15;
502 if (ISD == ISD::SREM && LT.second == MVT::v4i32 && ST->hasSSE41())
503 return LT.first * 20;
504
505 if (const auto *Entry = CostTableLookup(SSE2ConstCostTable, ISD, LT.second))
506 return LT.first * Entry->Cost;
507 }
508
509 static const CostTblEntry AVX512BWShiftCostTable[] = {
510 { ISD::SHL, MVT::v16i8, 4 }, // extend/vpsllvw/pack sequence.
511 { ISD::SRL, MVT::v16i8, 4 }, // extend/vpsrlvw/pack sequence.
512 { ISD::SRA, MVT::v16i8, 4 }, // extend/vpsravw/pack sequence.
513 { ISD::SHL, MVT::v32i8, 4 }, // extend/vpsllvw/pack sequence.
514 { ISD::SRL, MVT::v32i8, 4 }, // extend/vpsrlvw/pack sequence.
515 { ISD::SRA, MVT::v32i8, 6 }, // extend/vpsravw/pack sequence.
516 { ISD::SHL, MVT::v64i8, 6 }, // extend/vpsllvw/pack sequence.
517 { ISD::SRL, MVT::v64i8, 7 }, // extend/vpsrlvw/pack sequence.
518 { ISD::SRA, MVT::v64i8, 15 }, // extend/vpsravw/pack sequence.
519
520 { ISD::SHL, MVT::v8i16, 1 }, // vpsllvw
521 { ISD::SRL, MVT::v8i16, 1 }, // vpsrlvw
522 { ISD::SRA, MVT::v8i16, 1 }, // vpsravw
523 { ISD::SHL, MVT::v16i16, 1 }, // vpsllvw
524 { ISD::SRL, MVT::v16i16, 1 }, // vpsrlvw
525 { ISD::SRA, MVT::v16i16, 1 }, // vpsravw
526 { ISD::SHL, MVT::v32i16, 1 }, // vpsllvw
527 { ISD::SRL, MVT::v32i16, 1 }, // vpsrlvw
528 { ISD::SRA, MVT::v32i16, 1 }, // vpsravw
529 };
530
531 if (ST->hasBWI())
532 if (const auto *Entry = CostTableLookup(AVX512BWShiftCostTable, ISD, LT.second))
533 return LT.first * Entry->Cost;
534
535 static const CostTblEntry AVX2UniformCostTable[] = {
536 // Uniform splats are cheaper for the following instructions.
537 { ISD::SHL, MVT::v16i16, 1 }, // psllw.
538 { ISD::SRL, MVT::v16i16, 1 }, // psrlw.
539 { ISD::SRA, MVT::v16i16, 1 }, // psraw.
540 { ISD::SHL, MVT::v32i16, 2 }, // 2*psllw.
541 { ISD::SRL, MVT::v32i16, 2 }, // 2*psrlw.
542 { ISD::SRA, MVT::v32i16, 2 }, // 2*psraw.
543
544 { ISD::SHL, MVT::v8i32, 1 }, // pslld
545 { ISD::SRL, MVT::v8i32, 1 }, // psrld
546 { ISD::SRA, MVT::v8i32, 1 }, // psrad
547 { ISD::SHL, MVT::v4i64, 1 }, // psllq
548 { ISD::SRL, MVT::v4i64, 1 }, // psrlq
549 };
550
551 if (ST->hasAVX2() &&
552 ((Op2Info == TargetTransformInfo::OK_UniformConstantValue) ||
553 (Op2Info == TargetTransformInfo::OK_UniformValue))) {
554 if (const auto *Entry =
555 CostTableLookup(AVX2UniformCostTable, ISD, LT.second))
556 return LT.first * Entry->Cost;
557 }
558
559 static const CostTblEntry SSE2UniformCostTable[] = {
560 // Uniform splats are cheaper for the following instructions.
561 { ISD::SHL, MVT::v8i16, 1 }, // psllw.
562 { ISD::SHL, MVT::v4i32, 1 }, // pslld
563 { ISD::SHL, MVT::v2i64, 1 }, // psllq.
564
565 { ISD::SRL, MVT::v8i16, 1 }, // psrlw.
566 { ISD::SRL, MVT::v4i32, 1 }, // psrld.
567 { ISD::SRL, MVT::v2i64, 1 }, // psrlq.
568
569 { ISD::SRA, MVT::v8i16, 1 }, // psraw.
570 { ISD::SRA, MVT::v4i32, 1 }, // psrad.
571 };
572
573 if (ST->hasSSE2() &&
574 ((Op2Info == TargetTransformInfo::OK_UniformConstantValue) ||
575 (Op2Info == TargetTransformInfo::OK_UniformValue))) {
576 if (const auto *Entry =
577 CostTableLookup(SSE2UniformCostTable, ISD, LT.second))
578 return LT.first * Entry->Cost;
579 }
580
581 static const CostTblEntry AVX512DQCostTable[] = {
582 { ISD::MUL, MVT::v2i64, 2 }, // pmullq
583 { ISD::MUL, MVT::v4i64, 2 }, // pmullq
584 { ISD::MUL, MVT::v8i64, 2 } // pmullq
585 };
586
587 // Look for AVX512DQ lowering tricks for custom cases.
588 if (ST->hasDQI())
589 if (const auto *Entry = CostTableLookup(AVX512DQCostTable, ISD, LT.second))
590 return LT.first * Entry->Cost;
591
592 static const CostTblEntry AVX512BWCostTable[] = {
593 { ISD::SHL, MVT::v64i8, 11 }, // vpblendvb sequence.
594 { ISD::SRL, MVT::v64i8, 11 }, // vpblendvb sequence.
595 { ISD::SRA, MVT::v64i8, 24 }, // vpblendvb sequence.
596 };
597
598 // Look for AVX512BW lowering tricks for custom cases.
599 if (ST->hasBWI())
600 if (const auto *Entry = CostTableLookup(AVX512BWCostTable, ISD, LT.second))
601 return LT.first * Entry->Cost;
602
603 static const CostTblEntry AVX512CostTable[] = {
604 { ISD::SHL, MVT::v4i32, 1 },
605 { ISD::SRL, MVT::v4i32, 1 },
606 { ISD::SRA, MVT::v4i32, 1 },
607 { ISD::SHL, MVT::v8i32, 1 },
608 { ISD::SRL, MVT::v8i32, 1 },
609 { ISD::SRA, MVT::v8i32, 1 },
610 { ISD::SHL, MVT::v16i32, 1 },
611 { ISD::SRL, MVT::v16i32, 1 },
612 { ISD::SRA, MVT::v16i32, 1 },
613
614 { ISD::SHL, MVT::v2i64, 1 },
615 { ISD::SRL, MVT::v2i64, 1 },
616 { ISD::SHL, MVT::v4i64, 1 },
617 { ISD::SRL, MVT::v4i64, 1 },
618 { ISD::SHL, MVT::v8i64, 1 },
619 { ISD::SRL, MVT::v8i64, 1 },
620
621 { ISD::SRA, MVT::v2i64, 1 },
622 { ISD::SRA, MVT::v4i64, 1 },
623 { ISD::SRA, MVT::v8i64, 1 },
624
625 { ISD::MUL, MVT::v16i32, 1 }, // pmulld (Skylake from agner.org)
626 { ISD::MUL, MVT::v8i32, 1 }, // pmulld (Skylake from agner.org)
627 { ISD::MUL, MVT::v4i32, 1 }, // pmulld (Skylake from agner.org)
628 { ISD::MUL, MVT::v8i64, 6 }, // 3*pmuludq/3*shift/2*add
629
630 { ISD::FNEG, MVT::v8f64, 1 }, // Skylake from http://www.agner.org/
631 { ISD::FADD, MVT::v8f64, 1 }, // Skylake from http://www.agner.org/
632 { ISD::FSUB, MVT::v8f64, 1 }, // Skylake from http://www.agner.org/
633 { ISD::FMUL, MVT::v8f64, 1 }, // Skylake from http://www.agner.org/
634 { ISD::FDIV, MVT::f64, 4 }, // Skylake from http://www.agner.org/
635 { ISD::FDIV, MVT::v2f64, 4 }, // Skylake from http://www.agner.org/
636 { ISD::FDIV, MVT::v4f64, 8 }, // Skylake from http://www.agner.org/
637 { ISD::FDIV, MVT::v8f64, 16 }, // Skylake from http://www.agner.org/
638
639 { ISD::FNEG, MVT::v16f32, 1 }, // Skylake from http://www.agner.org/
640 { ISD::FADD, MVT::v16f32, 1 }, // Skylake from http://www.agner.org/
641 { ISD::FSUB, MVT::v16f32, 1 }, // Skylake from http://www.agner.org/
642 { ISD::FMUL, MVT::v16f32, 1 }, // Skylake from http://www.agner.org/
643 { ISD::FDIV, MVT::f32, 3 }, // Skylake from http://www.agner.org/
644 { ISD::FDIV, MVT::v4f32, 3 }, // Skylake from http://www.agner.org/
645 { ISD::FDIV, MVT::v8f32, 5 }, // Skylake from http://www.agner.org/
646 { ISD::FDIV, MVT::v16f32, 10 }, // Skylake from http://www.agner.org/
647 };
648
649 if (ST->hasAVX512())
650 if (const auto *Entry = CostTableLookup(AVX512CostTable, ISD, LT.second))
651 return LT.first * Entry->Cost;
652
653 static const CostTblEntry AVX2ShiftCostTable[] = {
654 // Shifts on vXi64/vXi32 on AVX2 is legal even though we declare to
655 // customize them to detect the cases where shift amount is a scalar one.
656 { ISD::SHL, MVT::v4i32, 2 }, // vpsllvd (Haswell from agner.org)
657 { ISD::SRL, MVT::v4i32, 2 }, // vpsrlvd (Haswell from agner.org)
658 { ISD::SRA, MVT::v4i32, 2 }, // vpsravd (Haswell from agner.org)
659 { ISD::SHL, MVT::v8i32, 2 }, // vpsllvd (Haswell from agner.org)
660 { ISD::SRL, MVT::v8i32, 2 }, // vpsrlvd (Haswell from agner.org)
661 { ISD::SRA, MVT::v8i32, 2 }, // vpsravd (Haswell from agner.org)
662 { ISD::SHL, MVT::v2i64, 1 }, // vpsllvq (Haswell from agner.org)
663 { ISD::SRL, MVT::v2i64, 1 }, // vpsrlvq (Haswell from agner.org)
664 { ISD::SHL, MVT::v4i64, 1 }, // vpsllvq (Haswell from agner.org)
665 { ISD::SRL, MVT::v4i64, 1 }, // vpsrlvq (Haswell from agner.org)
666 };
667
668 if (ST->hasAVX512()) {
669 if (ISD == ISD::SHL && LT.second == MVT::v32i16 &&
670 (Op2Info == TargetTransformInfo::OK_UniformConstantValue ||
671 Op2Info == TargetTransformInfo::OK_NonUniformConstantValue))
672 // On AVX512, a packed v32i16 shift left by a constant build_vector
673 // is lowered into a vector multiply (vpmullw).
674 return getArithmeticInstrCost(Instruction::Mul, Ty, CostKind,
675 Op1Info, Op2Info,
676 TargetTransformInfo::OP_None,
677 TargetTransformInfo::OP_None);
678 }
679
680 // Look for AVX2 lowering tricks (XOP is always better at v4i32 shifts).
681 if (ST->hasAVX2() && !(ST->hasXOP() && LT.second == MVT::v4i32)) {
682 if (ISD == ISD::SHL && LT.second == MVT::v16i16 &&
683 (Op2Info == TargetTransformInfo::OK_UniformConstantValue ||
684 Op2Info == TargetTransformInfo::OK_NonUniformConstantValue))
685 // On AVX2, a packed v16i16 shift left by a constant build_vector
686 // is lowered into a vector multiply (vpmullw).
687 return getArithmeticInstrCost(Instruction::Mul, Ty, CostKind,
688 Op1Info, Op2Info,
689 TargetTransformInfo::OP_None,
690 TargetTransformInfo::OP_None);
691
692 if (const auto *Entry = CostTableLookup(AVX2ShiftCostTable, ISD, LT.second))
693 return LT.first * Entry->Cost;
694 }
695
696 static const CostTblEntry XOPShiftCostTable[] = {
697 // 128bit shifts take 1cy, but right shifts require negation beforehand.
698 { ISD::SHL, MVT::v16i8, 1 },
699 { ISD::SRL, MVT::v16i8, 2 },
700 { ISD::SRA, MVT::v16i8, 2 },
701 { ISD::SHL, MVT::v8i16, 1 },
702 { ISD::SRL, MVT::v8i16, 2 },
703 { ISD::SRA, MVT::v8i16, 2 },
704 { ISD::SHL, MVT::v4i32, 1 },
705 { ISD::SRL, MVT::v4i32, 2 },
706 { ISD::SRA, MVT::v4i32, 2 },
707 { ISD::SHL, MVT::v2i64, 1 },
708 { ISD::SRL, MVT::v2i64, 2 },
709 { ISD::SRA, MVT::v2i64, 2 },
710 // 256bit shifts require splitting if AVX2 didn't catch them above.
711 { ISD::SHL, MVT::v32i8, 2+2 },
712 { ISD::SRL, MVT::v32i8, 4+2 },
713 { ISD::SRA, MVT::v32i8, 4+2 },
714 { ISD::SHL, MVT::v16i16, 2+2 },
715 { ISD::SRL, MVT::v16i16, 4+2 },
716 { ISD::SRA, MVT::v16i16, 4+2 },
717 { ISD::SHL, MVT::v8i32, 2+2 },
718 { ISD::SRL, MVT::v8i32, 4+2 },
719 { ISD::SRA, MVT::v8i32, 4+2 },
720 { ISD::SHL, MVT::v4i64, 2+2 },
721 { ISD::SRL, MVT::v4i64, 4+2 },
722 { ISD::SRA, MVT::v4i64, 4+2 },
723 };
724
725 // Look for XOP lowering tricks.
726 if (ST->hasXOP()) {
727 // If the right shift is constant then we'll fold the negation so
728 // it's as cheap as a left shift.
729 int ShiftISD = ISD;
730 if ((ShiftISD == ISD::SRL || ShiftISD == ISD::SRA) &&
731 (Op2Info == TargetTransformInfo::OK_UniformConstantValue ||
732 Op2Info == TargetTransformInfo::OK_NonUniformConstantValue))
733 ShiftISD = ISD::SHL;
734 if (const auto *Entry =
735 CostTableLookup(XOPShiftCostTable, ShiftISD, LT.second))
736 return LT.first * Entry->Cost;
737 }
738
739 static const CostTblEntry SSE2UniformShiftCostTable[] = {
740 // Uniform splats are cheaper for the following instructions.
741 { ISD::SHL, MVT::v16i16, 2+2 }, // 2*psllw + split.
742 { ISD::SHL, MVT::v8i32, 2+2 }, // 2*pslld + split.
743 { ISD::SHL, MVT::v4i64, 2+2 }, // 2*psllq + split.
744
745 { ISD::SRL, MVT::v16i16, 2+2 }, // 2*psrlw + split.
746 { ISD::SRL, MVT::v8i32, 2+2 }, // 2*psrld + split.
747 { ISD::SRL, MVT::v4i64, 2+2 }, // 2*psrlq + split.
748
749 { ISD::SRA, MVT::v16i16, 2+2 }, // 2*psraw + split.
750 { ISD::SRA, MVT::v8i32, 2+2 }, // 2*psrad + split.
751 { ISD::SRA, MVT::v2i64, 4 }, // 2*psrad + shuffle.
752 { ISD::SRA, MVT::v4i64, 8+2 }, // 2*(2*psrad + shuffle) + split.
753 };
754
755 if (ST->hasSSE2() &&
756 ((Op2Info == TargetTransformInfo::OK_UniformConstantValue) ||
757 (Op2Info == TargetTransformInfo::OK_UniformValue))) {
758
759 // Handle AVX2 uniform v4i64 ISD::SRA, it's not worth a table.
760 if (ISD == ISD::SRA && LT.second == MVT::v4i64 && ST->hasAVX2())
761 return LT.first * 4; // 2*psrad + shuffle.
762
763 if (const auto *Entry =
764 CostTableLookup(SSE2UniformShiftCostTable, ISD, LT.second))
765 return LT.first * Entry->Cost;
766 }
767
768 if (ISD == ISD::SHL &&
769 Op2Info == TargetTransformInfo::OK_NonUniformConstantValue) {
770 MVT VT = LT.second;
771 // Vector shift left by non uniform constant can be lowered
772 // into vector multiply.
773 if (((VT == MVT::v8i16 || VT == MVT::v4i32) && ST->hasSSE2()) ||
774 ((VT == MVT::v16i16 || VT == MVT::v8i32) && ST->hasAVX()))
775 ISD = ISD::MUL;
776 }
777
778 static const CostTblEntry AVX2CostTable[] = {
779 { ISD::SHL, MVT::v16i8, 6 }, // vpblendvb sequence.
780 { ISD::SHL, MVT::v32i8, 6 }, // vpblendvb sequence.
781 { ISD::SHL, MVT::v64i8, 12 }, // 2*vpblendvb sequence.
782 { ISD::SHL, MVT::v8i16, 5 }, // extend/vpsrlvd/pack sequence.
783 { ISD::SHL, MVT::v16i16, 7 }, // extend/vpsrlvd/pack sequence.
784 { ISD::SHL, MVT::v32i16, 14 }, // 2*extend/vpsrlvd/pack sequence.
785
786 { ISD::SRL, MVT::v16i8, 6 }, // vpblendvb sequence.
787 { ISD::SRL, MVT::v32i8, 6 }, // vpblendvb sequence.
788 { ISD::SRL, MVT::v64i8, 12 }, // 2*vpblendvb sequence.
789 { ISD::SRL, MVT::v8i16, 5 }, // extend/vpsrlvd/pack sequence.
790 { ISD::SRL, MVT::v16i16, 7 }, // extend/vpsrlvd/pack sequence.
791 { ISD::SRL, MVT::v32i16, 14 }, // 2*extend/vpsrlvd/pack sequence.
792
793 { ISD::SRA, MVT::v16i8, 17 }, // vpblendvb sequence.
794 { ISD::SRA, MVT::v32i8, 17 }, // vpblendvb sequence.
795 { ISD::SRA, MVT::v64i8, 34 }, // 2*vpblendvb sequence.
796 { ISD::SRA, MVT::v8i16, 5 }, // extend/vpsravd/pack sequence.
797 { ISD::SRA, MVT::v16i16, 7 }, // extend/vpsravd/pack sequence.
798 { ISD::SRA, MVT::v32i16, 14 }, // 2*extend/vpsravd/pack sequence.
799 { ISD::SRA, MVT::v2i64, 2 }, // srl/xor/sub sequence.
800 { ISD::SRA, MVT::v4i64, 2 }, // srl/xor/sub sequence.
801
802 { ISD::SUB, MVT::v32i8, 1 }, // psubb
803 { ISD::ADD, MVT::v32i8, 1 }, // paddb
804 { ISD::SUB, MVT::v16i16, 1 }, // psubw
805 { ISD::ADD, MVT::v16i16, 1 }, // paddw
806 { ISD::SUB, MVT::v8i32, 1 }, // psubd
807 { ISD::ADD, MVT::v8i32, 1 }, // paddd
808 { ISD::SUB, MVT::v4i64, 1 }, // psubq
809 { ISD::ADD, MVT::v4i64, 1 }, // paddq
810
811 { ISD::MUL, MVT::v16i16, 1 }, // pmullw
812 { ISD::MUL, MVT::v8i32, 2 }, // pmulld (Haswell from agner.org)
813 { ISD::MUL, MVT::v4i64, 6 }, // 3*pmuludq/3*shift/2*add
814
815 { ISD::FNEG, MVT::v4f64, 1 }, // Haswell from http://www.agner.org/
816 { ISD::FNEG, MVT::v8f32, 1 }, // Haswell from http://www.agner.org/
817 { ISD::FADD, MVT::v4f64, 1 }, // Haswell from http://www.agner.org/
818 { ISD::FADD, MVT::v8f32, 1 }, // Haswell from http://www.agner.org/
819 { ISD::FSUB, MVT::v4f64, 1 }, // Haswell from http://www.agner.org/
820 { ISD::FSUB, MVT::v8f32, 1 }, // Haswell from http://www.agner.org/
821 { ISD::FMUL, MVT::f64, 1 }, // Haswell from http://www.agner.org/
822 { ISD::FMUL, MVT::v2f64, 1 }, // Haswell from http://www.agner.org/
823 { ISD::FMUL, MVT::v4f64, 1 }, // Haswell from http://www.agner.org/
824 { ISD::FMUL, MVT::v8f32, 1 }, // Haswell from http://www.agner.org/
825
826 { ISD::FDIV, MVT::f32, 7 }, // Haswell from http://www.agner.org/
827 { ISD::FDIV, MVT::v4f32, 7 }, // Haswell from http://www.agner.org/
828 { ISD::FDIV, MVT::v8f32, 14 }, // Haswell from http://www.agner.org/
829 { ISD::FDIV, MVT::f64, 14 }, // Haswell from http://www.agner.org/
830 { ISD::FDIV, MVT::v2f64, 14 }, // Haswell from http://www.agner.org/
831 { ISD::FDIV, MVT::v4f64, 28 }, // Haswell from http://www.agner.org/
832 };
833
834 // Look for AVX2 lowering tricks for custom cases.
835 if (ST->hasAVX2())
836 if (const auto *Entry = CostTableLookup(AVX2CostTable, ISD, LT.second))
837 return LT.first * Entry->Cost;
838
839 static const CostTblEntry AVX1CostTable[] = {
840 // We don't have to scalarize unsupported ops. We can issue two half-sized
841 // operations and we only need to extract the upper YMM half.
842 // Two ops + 1 extract + 1 insert = 4.
843 { ISD::MUL, MVT::v16i16, 4 },
844 { ISD::MUL, MVT::v8i32, 5 }, // BTVER2 from http://www.agner.org/
845 { ISD::MUL, MVT::v4i64, 12 },
846
847 { ISD::SUB, MVT::v32i8, 4 },
848 { ISD::ADD, MVT::v32i8, 4 },
849 { ISD::SUB, MVT::v16i16, 4 },
850 { ISD::ADD, MVT::v16i16, 4 },
851 { ISD::SUB, MVT::v8i32, 4 },
852 { ISD::ADD, MVT::v8i32, 4 },
853 { ISD::SUB, MVT::v4i64, 4 },
854 { ISD::ADD, MVT::v4i64, 4 },
855
856 { ISD::SHL, MVT::v32i8, 22 }, // pblendvb sequence + split.
857 { ISD::SHL, MVT::v8i16, 6 }, // pblendvb sequence.
858 { ISD::SHL, MVT::v16i16, 13 }, // pblendvb sequence + split.
859 { ISD::SHL, MVT::v4i32, 3 }, // pslld/paddd/cvttps2dq/pmulld
860 { ISD::SHL, MVT::v8i32, 9 }, // pslld/paddd/cvttps2dq/pmulld + split
861 { ISD::SHL, MVT::v2i64, 2 }, // Shift each lane + blend.
862 { ISD::SHL, MVT::v4i64, 6 }, // Shift each lane + blend + split.
863
864 { ISD::SRL, MVT::v32i8, 23 }, // pblendvb sequence + split.
865 { ISD::SRL, MVT::v16i16, 28 }, // pblendvb sequence + split.
866 { ISD::SRL, MVT::v4i32, 6 }, // Shift each lane + blend.
867 { ISD::SRL, MVT::v8i32, 14 }, // Shift each lane + blend + split.
868 { ISD::SRL, MVT::v2i64, 2 }, // Shift each lane + blend.
869 { ISD::SRL, MVT::v4i64, 6 }, // Shift each lane + blend + split.
870
871 { ISD::SRA, MVT::v32i8, 44 }, // pblendvb sequence + split.
872 { ISD::SRA, MVT::v16i16, 28 }, // pblendvb sequence + split.
873 { ISD::SRA, MVT::v4i32, 6 }, // Shift each lane + blend.
874 { ISD::SRA, MVT::v8i32, 14 }, // Shift each lane + blend + split.
875 { ISD::SRA, MVT::v2i64, 5 }, // Shift each lane + blend.
876 { ISD::SRA, MVT::v4i64, 12 }, // Shift each lane + blend + split.
877
878 { ISD::FNEG, MVT::v4f64, 2 }, // BTVER2 from http://www.agner.org/
879 { ISD::FNEG, MVT::v8f32, 2 }, // BTVER2 from http://www.agner.org/
880
881 { ISD::FMUL, MVT::f64, 2 }, // BTVER2 from http://www.agner.org/
882 { ISD::FMUL, MVT::v2f64, 2 }, // BTVER2 from http://www.agner.org/
883 { ISD::FMUL, MVT::v4f64, 4 }, // BTVER2 from http://www.agner.org/
884
885 { ISD::FDIV, MVT::f32, 14 }, // SNB from http://www.agner.org/
886 { ISD::FDIV, MVT::v4f32, 14 }, // SNB from http://www.agner.org/
887 { ISD::FDIV, MVT::v8f32, 28 }, // SNB from http://www.agner.org/
888 { ISD::FDIV, MVT::f64, 22 }, // SNB from http://www.agner.org/
889 { ISD::FDIV, MVT::v2f64, 22 }, // SNB from http://www.agner.org/
890 { ISD::FDIV, MVT::v4f64, 44 }, // SNB from http://www.agner.org/
891 };
892
893 if (ST->hasAVX())
894 if (const auto *Entry = CostTableLookup(AVX1CostTable, ISD, LT.second))
895 return LT.first * Entry->Cost;
896
897 static const CostTblEntry SSE42CostTable[] = {
898 { ISD::FADD, MVT::f64, 1 }, // Nehalem from http://www.agner.org/
899 { ISD::FADD, MVT::f32, 1 }, // Nehalem from http://www.agner.org/
900 { ISD::FADD, MVT::v2f64, 1 }, // Nehalem from http://www.agner.org/
901 { ISD::FADD, MVT::v4f32, 1 }, // Nehalem from http://www.agner.org/
902
903 { ISD::FSUB, MVT::f64, 1 }, // Nehalem from http://www.agner.org/
904 { ISD::FSUB, MVT::f32 , 1 }, // Nehalem from http://www.agner.org/
905 { ISD::FSUB, MVT::v2f64, 1 }, // Nehalem from http://www.agner.org/
906 { ISD::FSUB, MVT::v4f32, 1 }, // Nehalem from http://www.agner.org/
907
908 { ISD::FMUL, MVT::f64, 1 }, // Nehalem from http://www.agner.org/
909 { ISD::FMUL, MVT::f32, 1 }, // Nehalem from http://www.agner.org/
910 { ISD::FMUL, MVT::v2f64, 1 }, // Nehalem from http://www.agner.org/
911 { ISD::FMUL, MVT::v4f32, 1 }, // Nehalem from http://www.agner.org/
912
913 { ISD::FDIV, MVT::f32, 14 }, // Nehalem from http://www.agner.org/
914 { ISD::FDIV, MVT::v4f32, 14 }, // Nehalem from http://www.agner.org/
915 { ISD::FDIV, MVT::f64, 22 }, // Nehalem from http://www.agner.org/
916 { ISD::FDIV, MVT::v2f64, 22 }, // Nehalem from http://www.agner.org/
917
918 { ISD::MUL, MVT::v2i64, 6 } // 3*pmuludq/3*shift/2*add
919 };
920
921 if (ST->hasSSE42())
922 if (const auto *Entry = CostTableLookup(SSE42CostTable, ISD, LT.second))
923 return LT.first * Entry->Cost;
924
925 static const CostTblEntry SSE41CostTable[] = {
926 { ISD::SHL, MVT::v16i8, 10 }, // pblendvb sequence.
927 { ISD::SHL, MVT::v8i16, 11 }, // pblendvb sequence.
928 { ISD::SHL, MVT::v4i32, 4 }, // pslld/paddd/cvttps2dq/pmulld
929
930 { ISD::SRL, MVT::v16i8, 11 }, // pblendvb sequence.
931 { ISD::SRL, MVT::v8i16, 13 }, // pblendvb sequence.
932 { ISD::SRL, MVT::v4i32, 16 }, // Shift each lane + blend.
933
934 { ISD::SRA, MVT::v16i8, 21 }, // pblendvb sequence.
935 { ISD::SRA, MVT::v8i16, 13 }, // pblendvb sequence.
936
937 { ISD::MUL, MVT::v4i32, 2 } // pmulld (Nehalem from agner.org)
938 };
939
940 if (ST->hasSSE41())
941 if (const auto *Entry = CostTableLookup(SSE41CostTable, ISD, LT.second))
942 return LT.first * Entry->Cost;
943
944 static const CostTblEntry SSE2CostTable[] = {
945 // We don't correctly identify costs of casts because they are marked as
946 // custom.
947 { ISD::SHL, MVT::v16i8, 13 }, // cmpgtb sequence.
948 { ISD::SHL, MVT::v8i16, 25 }, // cmpgtw sequence.
949 { ISD::SHL, MVT::v4i32, 16 }, // pslld/paddd/cvttps2dq/pmuludq.
950 { ISD::SHL, MVT::v2i64, 4 }, // splat+shuffle sequence.
951
952 { ISD::SRL, MVT::v16i8, 14 }, // cmpgtb sequence.
953 { ISD::SRL, MVT::v8i16, 16 }, // cmpgtw sequence.
954 { ISD::SRL, MVT::v4i32, 12 }, // Shift each lane + blend.
955 { ISD::SRL, MVT::v2i64, 4 }, // splat+shuffle sequence.
956
957 { ISD::SRA, MVT::v16i8, 27 }, // unpacked cmpgtb sequence.
958 { ISD::SRA, MVT::v8i16, 16 }, // cmpgtw sequence.
959 { ISD::SRA, MVT::v4i32, 12 }, // Shift each lane + blend.
960 { ISD::SRA, MVT::v2i64, 8 }, // srl/xor/sub splat+shuffle sequence.
961
962 { ISD::MUL, MVT::v8i16, 1 }, // pmullw
963 { ISD::MUL, MVT::v4i32, 6 }, // 3*pmuludq/4*shuffle
964 { ISD::MUL, MVT::v2i64, 8 }, // 3*pmuludq/3*shift/2*add
965
966 { ISD::FDIV, MVT::f32, 23 }, // Pentium IV from http://www.agner.org/
967 { ISD::FDIV, MVT::v4f32, 39 }, // Pentium IV from http://www.agner.org/
968 { ISD::FDIV, MVT::f64, 38 }, // Pentium IV from http://www.agner.org/
969 { ISD::FDIV, MVT::v2f64, 69 }, // Pentium IV from http://www.agner.org/
970
971 { ISD::FNEG, MVT::f32, 1 }, // Pentium IV from http://www.agner.org/
972 { ISD::FNEG, MVT::f64, 1 }, // Pentium IV from http://www.agner.org/
973 { ISD::FNEG, MVT::v4f32, 1 }, // Pentium IV from http://www.agner.org/
974 { ISD::FNEG, MVT::v2f64, 1 }, // Pentium IV from http://www.agner.org/
975
976 { ISD::FADD, MVT::f32, 2 }, // Pentium IV from http://www.agner.org/
977 { ISD::FADD, MVT::f64, 2 }, // Pentium IV from http://www.agner.org/
978
979 { ISD::FSUB, MVT::f32, 2 }, // Pentium IV from http://www.agner.org/
980 { ISD::FSUB, MVT::f64, 2 }, // Pentium IV from http://www.agner.org/
981 };
982
983 if (ST->hasSSE2())
984 if (const auto *Entry = CostTableLookup(SSE2CostTable, ISD, LT.second))
985 return LT.first * Entry->Cost;
986
987 static const CostTblEntry SSE1CostTable[] = {
988 { ISD::FDIV, MVT::f32, 17 }, // Pentium III from http://www.agner.org/
989 { ISD::FDIV, MVT::v4f32, 34 }, // Pentium III from http://www.agner.org/
990
991 { ISD::FNEG, MVT::f32, 2 }, // Pentium III from http://www.agner.org/
992 { ISD::FNEG, MVT::v4f32, 2 }, // Pentium III from http://www.agner.org/
993
994 { ISD::FADD, MVT::f32, 1 }, // Pentium III from http://www.agner.org/
995 { ISD::FADD, MVT::v4f32, 2 }, // Pentium III from http://www.agner.org/
996
997 { ISD::FSUB, MVT::f32, 1 }, // Pentium III from http://www.agner.org/
998 { ISD::FSUB, MVT::v4f32, 2 }, // Pentium III from http://www.agner.org/
999 };
1000
1001 if (ST->hasSSE1())
1002 if (const auto *Entry = CostTableLookup(SSE1CostTable, ISD, LT.second))
1003 return LT.first * Entry->Cost;
1004
1005 static const CostTblEntry X64CostTbl[] = { // 64-bit targets
1006 { ISD::ADD, MVT::i64, 1 }, // Core (Merom) from http://www.agner.org/
1007 { ISD::SUB, MVT::i64, 1 }, // Core (Merom) from http://www.agner.org/
1008 };
1009
1010 if (ST->is64Bit())
1011 if (const auto *Entry = CostTableLookup(X64CostTbl, ISD, LT.second))
1012 return LT.first * Entry->Cost;
1013
1014 static const CostTblEntry X86CostTbl[] = { // 32 or 64-bit targets
1015 { ISD::ADD, MVT::i8, 1 }, // Pentium III from http://www.agner.org/
1016 { ISD::ADD, MVT::i16, 1 }, // Pentium III from http://www.agner.org/
1017 { ISD::ADD, MVT::i32, 1 }, // Pentium III from http://www.agner.org/
1018
1019 { ISD::SUB, MVT::i8, 1 }, // Pentium III from http://www.agner.org/
1020 { ISD::SUB, MVT::i16, 1 }, // Pentium III from http://www.agner.org/
1021 { ISD::SUB, MVT::i32, 1 }, // Pentium III from http://www.agner.org/
1022 };
1023
1024 if (const auto *Entry = CostTableLookup(X86CostTbl, ISD, LT.second))
1025 return LT.first * Entry->Cost;
1026
1027 // It is not a good idea to vectorize division. We have to scalarize it and
1028 // in the process we will often end up having to spilling regular
1029 // registers. The overhead of division is going to dominate most kernels
1030 // anyways so try hard to prevent vectorization of division - it is
1031 // generally a bad idea. Assume somewhat arbitrarily that we have to be able
1032 // to hide "20 cycles" for each lane.
1033 if (LT.second.isVector() && (ISD == ISD::SDIV || ISD == ISD::SREM ||
1034 ISD == ISD::UDIV || ISD == ISD::UREM)) {
1035 InstructionCost ScalarCost = getArithmeticInstrCost(
1036 Opcode, Ty->getScalarType(), CostKind, Op1Info, Op2Info,
1037 TargetTransformInfo::OP_None, TargetTransformInfo::OP_None);
1038 return 20 * LT.first * LT.second.getVectorNumElements() * ScalarCost;
1039 }
1040
1041 // Fallback to the default implementation.
1042 return BaseT::getArithmeticInstrCost(Opcode, Ty, CostKind, Op1Info, Op2Info);
1043}
1044
1045InstructionCost X86TTIImpl::getShuffleCost(TTI::ShuffleKind Kind,
1046 VectorType *BaseTp,
1047 ArrayRef<int> Mask, int Index,
1048 VectorType *SubTp) {
1049 // 64-bit packed float vectors (v2f32) are widened to type v4f32.
1050 // 64-bit packed integer vectors (v2i32) are widened to type v4i32.
1051 std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, BaseTp);
1052
1053 Kind = improveShuffleKindFromMask(Kind, Mask);
1054 // Treat Transpose as 2-op shuffles - there's no difference in lowering.
1055 if (Kind == TTI::SK_Transpose)
1056 Kind = TTI::SK_PermuteTwoSrc;
1057
1058 // For Broadcasts we are splatting the first element from the first input
1059 // register, so only need to reference that input and all the output
1060 // registers are the same.
1061 if (Kind == TTI::SK_Broadcast)
1062 LT.first = 1;
1063
1064 // Subvector extractions are free if they start at the beginning of a
1065 // vector and cheap if the subvectors are aligned.
1066 if (Kind == TTI::SK_ExtractSubvector && LT.second.isVector()) {
1067 int NumElts = LT.second.getVectorNumElements();
1068 if ((Index % NumElts) == 0)
1069 return 0;
1070 std::pair<InstructionCost, MVT> SubLT =
1071 TLI->getTypeLegalizationCost(DL, SubTp);
1072 if (SubLT.second.isVector()) {
1073 int NumSubElts = SubLT.second.getVectorNumElements();
1074 if ((Index % NumSubElts) == 0 && (NumElts % NumSubElts) == 0)
1075 return SubLT.first;
1076 // Handle some cases for widening legalization. For now we only handle
1077 // cases where the original subvector was naturally aligned and evenly
1078 // fit in its legalized subvector type.
1079 // FIXME: Remove some of the alignment restrictions.
1080 // FIXME: We can use permq for 64-bit or larger extracts from 256-bit
1081 // vectors.
1082 int OrigSubElts = cast<FixedVectorType>(SubTp)->getNumElements();
1083 if (NumSubElts > OrigSubElts && (Index % OrigSubElts) == 0 &&
1084 (NumSubElts % OrigSubElts) == 0 &&
1085 LT.second.getVectorElementType() ==
1086 SubLT.second.getVectorElementType() &&
1087 LT.second.getVectorElementType().getSizeInBits() ==
1088 BaseTp->getElementType()->getPrimitiveSizeInBits()) {
1089 assert(NumElts >= NumSubElts && NumElts > OrigSubElts &&((void)0)
1090 "Unexpected number of elements!")((void)0);
1091 auto *VecTy = FixedVectorType::get(BaseTp->getElementType(),
1092 LT.second.getVectorNumElements());
1093 auto *SubTy = FixedVectorType::get(BaseTp->getElementType(),
1094 SubLT.second.getVectorNumElements());
1095 int ExtractIndex = alignDown((Index % NumElts), NumSubElts);
1096 InstructionCost ExtractCost = getShuffleCost(
1097 TTI::SK_ExtractSubvector, VecTy, None, ExtractIndex, SubTy);
1098
1099 // If the original size is 32-bits or more, we can use pshufd. Otherwise
1100 // if we have SSSE3 we can use pshufb.
1101 if (SubTp->getPrimitiveSizeInBits() >= 32 || ST->hasSSSE3())
1102 return ExtractCost + 1; // pshufd or pshufb
1103
1104 assert(SubTp->getPrimitiveSizeInBits() == 16 &&((void)0)
1105 "Unexpected vector size")((void)0);
1106
1107 return ExtractCost + 2; // worst case pshufhw + pshufd
1108 }
1109 }
1110 }
1111
1112 // Subvector insertions are cheap if the subvectors are aligned.
1113 // Note that in general, the insertion starting at the beginning of a vector
1114 // isn't free, because we need to preserve the rest of the wide vector.
1115 if (Kind == TTI::SK_InsertSubvector && LT.second.isVector()) {
1116 int NumElts = LT.second.getVectorNumElements();
1117 std::pair<InstructionCost, MVT> SubLT =
1118 TLI->getTypeLegalizationCost(DL, SubTp);
1119 if (SubLT.second.isVector()) {
1120 int NumSubElts = SubLT.second.getVectorNumElements();
1121 if ((Index % NumSubElts) == 0 && (NumElts % NumSubElts) == 0)
1122 return SubLT.first;
1123 }
1124 }
1125
1126 // Handle some common (illegal) sub-vector types as they are often very cheap
1127 // to shuffle even on targets without PSHUFB.
1128 EVT VT = TLI->getValueType(DL, BaseTp);
1129 if (VT.isSimple() && VT.isVector() && VT.getSizeInBits() < 128 &&
1130 !ST->hasSSSE3()) {
1131 static const CostTblEntry SSE2SubVectorShuffleTbl[] = {
1132 {TTI::SK_Broadcast, MVT::v4i16, 1}, // pshuflw
1133 {TTI::SK_Broadcast, MVT::v2i16, 1}, // pshuflw
1134 {TTI::SK_Broadcast, MVT::v8i8, 2}, // punpck/pshuflw
1135 {TTI::SK_Broadcast, MVT::v4i8, 2}, // punpck/pshuflw
1136 {TTI::SK_Broadcast, MVT::v2i8, 1}, // punpck
1137
1138 {TTI::SK_Reverse, MVT::v4i16, 1}, // pshuflw
1139 {TTI::SK_Reverse, MVT::v2i16, 1}, // pshuflw
1140 {TTI::SK_Reverse, MVT::v4i8, 3}, // punpck/pshuflw/packus
1141 {TTI::SK_Reverse, MVT::v2i8, 1}, // punpck
1142
1143 {TTI::SK_PermuteTwoSrc, MVT::v4i16, 2}, // punpck/pshuflw
1144 {TTI::SK_PermuteTwoSrc, MVT::v2i16, 2}, // punpck/pshuflw
1145 {TTI::SK_PermuteTwoSrc, MVT::v8i8, 7}, // punpck/pshuflw
1146 {TTI::SK_PermuteTwoSrc, MVT::v4i8, 4}, // punpck/pshuflw
1147 {TTI::SK_PermuteTwoSrc, MVT::v2i8, 2}, // punpck
1148
1149 {TTI::SK_PermuteSingleSrc, MVT::v4i16, 1}, // pshuflw
1150 {TTI::SK_PermuteSingleSrc, MVT::v2i16, 1}, // pshuflw
1151 {TTI::SK_PermuteSingleSrc, MVT::v8i8, 5}, // punpck/pshuflw
1152 {TTI::SK_PermuteSingleSrc, MVT::v4i8, 3}, // punpck/pshuflw
1153 {TTI::SK_PermuteSingleSrc, MVT::v2i8, 1}, // punpck
1154 };
1155
1156 if (ST->hasSSE2())
1157 if (const auto *Entry =
1158 CostTableLookup(SSE2SubVectorShuffleTbl, Kind, VT.getSimpleVT()))
1159 return Entry->Cost;
1160 }
1161
1162 // We are going to permute multiple sources and the result will be in multiple
1163 // destinations. Providing an accurate cost only for splits where the element
1164 // type remains the same.
1165 if (Kind == TTI::SK_PermuteSingleSrc && LT.first != 1) {
1166 MVT LegalVT = LT.second;
1167 if (LegalVT.isVector() &&
1168 LegalVT.getVectorElementType().getSizeInBits() ==
1169 BaseTp->getElementType()->getPrimitiveSizeInBits() &&
1170 LegalVT.getVectorNumElements() <
1171 cast<FixedVectorType>(BaseTp)->getNumElements()) {
1172
1173 unsigned VecTySize = DL.getTypeStoreSize(BaseTp);
1174 unsigned LegalVTSize = LegalVT.getStoreSize();
1175 // Number of source vectors after legalization:
1176 unsigned NumOfSrcs = (VecTySize + LegalVTSize - 1) / LegalVTSize;
1177 // Number of destination vectors after legalization:
1178 InstructionCost NumOfDests = LT.first;
1179
1180 auto *SingleOpTy = FixedVectorType::get(BaseTp->getElementType(),
1181 LegalVT.getVectorNumElements());
1182
1183 InstructionCost NumOfShuffles = (NumOfSrcs - 1) * NumOfDests;
1184 return NumOfShuffles * getShuffleCost(TTI::SK_PermuteTwoSrc, SingleOpTy,
1185 None, 0, nullptr);
1186 }
1187
1188 return BaseT::getShuffleCost(Kind, BaseTp, Mask, Index, SubTp);
1189 }
1190
1191 // For 2-input shuffles, we must account for splitting the 2 inputs into many.
1192 if (Kind == TTI::SK_PermuteTwoSrc && LT.first != 1) {
1193 // We assume that source and destination have the same vector type.
1194 InstructionCost NumOfDests = LT.first;
1195 InstructionCost NumOfShufflesPerDest = LT.first * 2 - 1;
1196 LT.first = NumOfDests * NumOfShufflesPerDest;
1197 }
1198
1199 static const CostTblEntry AVX512VBMIShuffleTbl[] = {
1200 {TTI::SK_Reverse, MVT::v64i8, 1}, // vpermb
1201 {TTI::SK_Reverse, MVT::v32i8, 1}, // vpermb
1202
1203 {TTI::SK_PermuteSingleSrc, MVT::v64i8, 1}, // vpermb
1204 {TTI::SK_PermuteSingleSrc, MVT::v32i8, 1}, // vpermb
1205
1206 {TTI::SK_PermuteTwoSrc, MVT::v64i8, 2}, // vpermt2b
1207 {TTI::SK_PermuteTwoSrc, MVT::v32i8, 2}, // vpermt2b
1208 {TTI::SK_PermuteTwoSrc, MVT::v16i8, 2} // vpermt2b
1209 };
1210
1211 if (ST->hasVBMI())
1212 if (const auto *Entry =
1213 CostTableLookup(AVX512VBMIShuffleTbl, Kind, LT.second))
1214 return LT.first * Entry->Cost;
1215
1216 static const CostTblEntry AVX512BWShuffleTbl[] = {
1217 {TTI::SK_Broadcast, MVT::v32i16, 1}, // vpbroadcastw
1218 {TTI::SK_Broadcast, MVT::v64i8, 1}, // vpbroadcastb
1219
1220 {TTI::SK_Reverse, MVT::v32i16, 2}, // vpermw
1221 {TTI::SK_Reverse, MVT::v16i16, 2}, // vpermw
1222 {TTI::SK_Reverse, MVT::v64i8, 2}, // pshufb + vshufi64x2
1223
1224 {TTI::SK_PermuteSingleSrc, MVT::v32i16, 2}, // vpermw
1225 {TTI::SK_PermuteSingleSrc, MVT::v16i16, 2}, // vpermw
1226 {TTI::SK_PermuteSingleSrc, MVT::v64i8, 8}, // extend to v32i16
1227
1228 {TTI::SK_PermuteTwoSrc, MVT::v32i16, 2}, // vpermt2w
1229 {TTI::SK_PermuteTwoSrc, MVT::v16i16, 2}, // vpermt2w
1230 {TTI::SK_PermuteTwoSrc, MVT::v8i16, 2}, // vpermt2w
1231 {TTI::SK_PermuteTwoSrc, MVT::v64i8, 19}, // 6 * v32i8 + 1
1232
1233 {TTI::SK_Select, MVT::v32i16, 1}, // vblendmw
1234 {TTI::SK_Select, MVT::v64i8, 1}, // vblendmb
1235 };
1236
1237 if (ST->hasBWI())
1238 if (const auto *Entry =
1239 CostTableLookup(AVX512BWShuffleTbl, Kind, LT.second))
1240 return LT.first * Entry->Cost;
1241
1242 static const CostTblEntry AVX512ShuffleTbl[] = {
1243 {TTI::SK_Broadcast, MVT::v8f64, 1}, // vbroadcastpd
1244 {TTI::SK_Broadcast, MVT::v16f32, 1}, // vbroadcastps
1245 {TTI::SK_Broadcast, MVT::v8i64, 1}, // vpbroadcastq
1246 {TTI::SK_Broadcast, MVT::v16i32, 1}, // vpbroadcastd
1247 {TTI::SK_Broadcast, MVT::v32i16, 1}, // vpbroadcastw
1248 {TTI::SK_Broadcast, MVT::v64i8, 1}, // vpbroadcastb
1249
1250 {TTI::SK_Reverse, MVT::v8f64, 1}, // vpermpd
1251 {TTI::SK_Reverse, MVT::v16f32, 1}, // vpermps
1252 {TTI::SK_Reverse, MVT::v8i64, 1}, // vpermq
1253 {TTI::SK_Reverse, MVT::v16i32, 1}, // vpermd
1254 {TTI::SK_Reverse, MVT::v32i16, 7}, // per mca
1255 {TTI::SK_Reverse, MVT::v64i8, 7}, // per mca
1256
1257 {TTI::SK_PermuteSingleSrc, MVT::v8f64, 1}, // vpermpd
1258 {TTI::SK_PermuteSingleSrc, MVT::v4f64, 1}, // vpermpd
1259 {TTI::SK_PermuteSingleSrc, MVT::v2f64, 1}, // vpermpd
1260 {TTI::SK_PermuteSingleSrc, MVT::v16f32, 1}, // vpermps
1261 {TTI::SK_PermuteSingleSrc, MVT::v8f32, 1}, // vpermps
1262 {TTI::SK_PermuteSingleSrc, MVT::v4f32, 1}, // vpermps
1263 {TTI::SK_PermuteSingleSrc, MVT::v8i64, 1}, // vpermq
1264 {TTI::SK_PermuteSingleSrc, MVT::v4i64, 1}, // vpermq
1265 {TTI::SK_PermuteSingleSrc, MVT::v2i64, 1}, // vpermq
1266 {TTI::SK_PermuteSingleSrc, MVT::v16i32, 1}, // vpermd
1267 {TTI::SK_PermuteSingleSrc, MVT::v8i32, 1}, // vpermd
1268 {TTI::SK_PermuteSingleSrc, MVT::v4i32, 1}, // vpermd
1269 {TTI::SK_PermuteSingleSrc, MVT::v16i8, 1}, // pshufb
1270
1271 {TTI::SK_PermuteTwoSrc, MVT::v8f64, 1}, // vpermt2pd
1272 {TTI::SK_PermuteTwoSrc, MVT::v16f32, 1}, // vpermt2ps
1273 {TTI::SK_PermuteTwoSrc, MVT::v8i64, 1}, // vpermt2q
1274 {TTI::SK_PermuteTwoSrc, MVT::v16i32, 1}, // vpermt2d
1275 {TTI::SK_PermuteTwoSrc, MVT::v4f64, 1}, // vpermt2pd
1276 {TTI::SK_PermuteTwoSrc, MVT::v8f32, 1}, // vpermt2ps
1277 {TTI::SK_PermuteTwoSrc, MVT::v4i64, 1}, // vpermt2q
1278 {TTI::SK_PermuteTwoSrc, MVT::v8i32, 1}, // vpermt2d
1279 {TTI::SK_PermuteTwoSrc, MVT::v2f64, 1}, // vpermt2pd
1280 {TTI::SK_PermuteTwoSrc, MVT::v4f32, 1}, // vpermt2ps
1281 {TTI::SK_PermuteTwoSrc, MVT::v2i64, 1}, // vpermt2q
1282 {TTI::SK_PermuteTwoSrc, MVT::v4i32, 1}, // vpermt2d
1283
1284 // FIXME: This just applies the type legalization cost rules above
1285 // assuming these completely split.
1286 {TTI::SK_PermuteSingleSrc, MVT::v32i16, 14},
1287 {TTI::SK_PermuteSingleSrc, MVT::v64i8, 14},
1288 {TTI::SK_PermuteTwoSrc, MVT::v32i16, 42},
1289 {TTI::SK_PermuteTwoSrc, MVT::v64i8, 42},
1290
1291 {TTI::SK_Select, MVT::v32i16, 1}, // vpternlogq
1292 {TTI::SK_Select, MVT::v64i8, 1}, // vpternlogq
1293 {TTI::SK_Select, MVT::v8f64, 1}, // vblendmpd
1294 {TTI::SK_Select, MVT::v16f32, 1}, // vblendmps
1295 {TTI::SK_Select, MVT::v8i64, 1}, // vblendmq
1296 {TTI::SK_Select, MVT::v16i32, 1}, // vblendmd
1297 };
1298
1299 if (ST->hasAVX512())
1300 if (const auto *Entry = CostTableLookup(AVX512ShuffleTbl, Kind, LT.second))
1301 return LT.first * Entry->Cost;
1302
1303 static const CostTblEntry AVX2ShuffleTbl[] = {
1304 {TTI::SK_Broadcast, MVT::v4f64, 1}, // vbroadcastpd
1305 {TTI::SK_Broadcast, MVT::v8f32, 1}, // vbroadcastps
1306 {TTI::SK_Broadcast, MVT::v4i64, 1}, // vpbroadcastq
1307 {TTI::SK_Broadcast, MVT::v8i32, 1}, // vpbroadcastd
1308 {TTI::SK_Broadcast, MVT::v16i16, 1}, // vpbroadcastw
1309 {TTI::SK_Broadcast, MVT::v32i8, 1}, // vpbroadcastb
1310
1311 {TTI::SK_Reverse, MVT::v4f64, 1}, // vpermpd
1312 {TTI::SK_Reverse, MVT::v8f32, 1}, // vpermps
1313 {TTI::SK_Reverse, MVT::v4i64, 1}, // vpermq
1314 {TTI::SK_Reverse, MVT::v8i32, 1}, // vpermd
1315 {TTI::SK_Reverse, MVT::v16i16, 2}, // vperm2i128 + pshufb
1316 {TTI::SK_Reverse, MVT::v32i8, 2}, // vperm2i128 + pshufb
1317
1318 {TTI::SK_Select, MVT::v16i16, 1}, // vpblendvb
1319 {TTI::SK_Select, MVT::v32i8, 1}, // vpblendvb
1320
1321 {TTI::SK_PermuteSingleSrc, MVT::v4f64, 1}, // vpermpd
1322 {TTI::SK_PermuteSingleSrc, MVT::v8f32, 1}, // vpermps
1323 {TTI::SK_PermuteSingleSrc, MVT::v4i64, 1}, // vpermq
1324 {TTI::SK_PermuteSingleSrc, MVT::v8i32, 1}, // vpermd
1325 {TTI::SK_PermuteSingleSrc, MVT::v16i16, 4}, // vperm2i128 + 2*vpshufb
1326 // + vpblendvb
1327 {TTI::SK_PermuteSingleSrc, MVT::v32i8, 4}, // vperm2i128 + 2*vpshufb
1328 // + vpblendvb
1329
1330 {TTI::SK_PermuteTwoSrc, MVT::v4f64, 3}, // 2*vpermpd + vblendpd
1331 {TTI::SK_PermuteTwoSrc, MVT::v8f32, 3}, // 2*vpermps + vblendps
1332 {TTI::SK_PermuteTwoSrc, MVT::v4i64, 3}, // 2*vpermq + vpblendd
1333 {TTI::SK_PermuteTwoSrc, MVT::v8i32, 3}, // 2*vpermd + vpblendd
1334 {TTI::SK_PermuteTwoSrc, MVT::v16i16, 7}, // 2*vperm2i128 + 4*vpshufb
1335 // + vpblendvb
1336 {TTI::SK_PermuteTwoSrc, MVT::v32i8, 7}, // 2*vperm2i128 + 4*vpshufb
1337 // + vpblendvb
1338 };
1339
1340 if (ST->hasAVX2())
1341 if (const auto *Entry = CostTableLookup(AVX2ShuffleTbl, Kind, LT.second))
1342 return LT.first * Entry->Cost;
1343
1344 static const CostTblEntry XOPShuffleTbl[] = {
1345 {TTI::SK_PermuteSingleSrc, MVT::v4f64, 2}, // vperm2f128 + vpermil2pd
1346 {TTI::SK_PermuteSingleSrc, MVT::v8f32, 2}, // vperm2f128 + vpermil2ps
1347 {TTI::SK_PermuteSingleSrc, MVT::v4i64, 2}, // vperm2f128 + vpermil2pd
1348 {TTI::SK_PermuteSingleSrc, MVT::v8i32, 2}, // vperm2f128 + vpermil2ps
1349 {TTI::SK_PermuteSingleSrc, MVT::v16i16, 4}, // vextractf128 + 2*vpperm
1350 // + vinsertf128
1351 {TTI::SK_PermuteSingleSrc, MVT::v32i8, 4}, // vextractf128 + 2*vpperm
1352 // + vinsertf128
1353
1354 {TTI::SK_PermuteTwoSrc, MVT::v16i16, 9}, // 2*vextractf128 + 6*vpperm
1355 // + vinsertf128
1356 {TTI::SK_PermuteTwoSrc, MVT::v8i16, 1}, // vpperm
1357 {TTI::SK_PermuteTwoSrc, MVT::v32i8, 9}, // 2*vextractf128 + 6*vpperm
1358 // + vinsertf128
1359 {TTI::SK_PermuteTwoSrc, MVT::v16i8, 1}, // vpperm
1360 };
1361
1362 if (ST->hasXOP())
1363 if (const auto *Entry = CostTableLookup(XOPShuffleTbl, Kind, LT.second))
1364 return LT.first * Entry->Cost;
1365
1366 static const CostTblEntry AVX1ShuffleTbl[] = {
1367 {TTI::SK_Broadcast, MVT::v4f64, 2}, // vperm2f128 + vpermilpd
1368 {TTI::SK_Broadcast, MVT::v8f32, 2}, // vperm2f128 + vpermilps
1369 {TTI::SK_Broadcast, MVT::v4i64, 2}, // vperm2f128 + vpermilpd
1370 {TTI::SK_Broadcast, MVT::v8i32, 2}, // vperm2f128 + vpermilps
1371 {TTI::SK_Broadcast, MVT::v16i16, 3}, // vpshuflw + vpshufd + vinsertf128
1372 {TTI::SK_Broadcast, MVT::v32i8, 2}, // vpshufb + vinsertf128
1373
1374 {TTI::SK_Reverse, MVT::v4f64, 2}, // vperm2f128 + vpermilpd
1375 {TTI::SK_Reverse, MVT::v8f32, 2}, // vperm2f128 + vpermilps
1376 {TTI::SK_Reverse, MVT::v4i64, 2}, // vperm2f128 + vpermilpd
1377 {TTI::SK_Reverse, MVT::v8i32, 2}, // vperm2f128 + vpermilps
1378 {TTI::SK_Reverse, MVT::v16i16, 4}, // vextractf128 + 2*pshufb
1379 // + vinsertf128
1380 {TTI::SK_Reverse, MVT::v32i8, 4}, // vextractf128 + 2*pshufb
1381 // + vinsertf128
1382
1383 {TTI::SK_Select, MVT::v4i64, 1}, // vblendpd
1384 {TTI::SK_Select, MVT::v4f64, 1}, // vblendpd
1385 {TTI::SK_Select, MVT::v8i32, 1}, // vblendps
1386 {TTI::SK_Select, MVT::v8f32, 1}, // vblendps
1387 {TTI::SK_Select, MVT::v16i16, 3}, // vpand + vpandn + vpor
1388 {TTI::SK_Select, MVT::v32i8, 3}, // vpand + vpandn + vpor
1389
1390 {TTI::SK_PermuteSingleSrc, MVT::v4f64, 2}, // vperm2f128 + vshufpd
1391 {TTI::SK_PermuteSingleSrc, MVT::v4i64, 2}, // vperm2f128 + vshufpd
1392 {TTI::SK_PermuteSingleSrc, MVT::v8f32, 4}, // 2*vperm2f128 + 2*vshufps
1393 {TTI::SK_PermuteSingleSrc, MVT::v8i32, 4}, // 2*vperm2f128 + 2*vshufps
1394 {TTI::SK_PermuteSingleSrc, MVT::v16i16, 8}, // vextractf128 + 4*pshufb
1395 // + 2*por + vinsertf128
1396 {TTI::SK_PermuteSingleSrc, MVT::v32i8, 8}, // vextractf128 + 4*pshufb
1397 // + 2*por + vinsertf128
1398
1399 {TTI::SK_PermuteTwoSrc, MVT::v4f64, 3}, // 2*vperm2f128 + vshufpd
1400 {TTI::SK_PermuteTwoSrc, MVT::v4i64, 3}, // 2*vperm2f128 + vshufpd
1401 {TTI::SK_PermuteTwoSrc, MVT::v8f32, 4}, // 2*vperm2f128 + 2*vshufps
1402 {TTI::SK_PermuteTwoSrc, MVT::v8i32, 4}, // 2*vperm2f128 + 2*vshufps
1403 {TTI::SK_PermuteTwoSrc, MVT::v16i16, 15}, // 2*vextractf128 + 8*pshufb
1404 // + 4*por + vinsertf128
1405 {TTI::SK_PermuteTwoSrc, MVT::v32i8, 15}, // 2*vextractf128 + 8*pshufb
1406 // + 4*por + vinsertf128
1407 };
1408
1409 if (ST->hasAVX())
1410 if (const auto *Entry = CostTableLookup(AVX1ShuffleTbl, Kind, LT.second))
1411 return LT.first * Entry->Cost;
1412
1413 static const CostTblEntry SSE41ShuffleTbl[] = {
1414 {TTI::SK_Select, MVT::v2i64, 1}, // pblendw
1415 {TTI::SK_Select, MVT::v2f64, 1}, // movsd
1416 {TTI::SK_Select, MVT::v4i32, 1}, // pblendw
1417 {TTI::SK_Select, MVT::v4f32, 1}, // blendps
1418 {TTI::SK_Select, MVT::v8i16, 1}, // pblendw
1419 {TTI::SK_Select, MVT::v16i8, 1} // pblendvb
1420 };
1421
1422 if (ST->hasSSE41())
1423 if (const auto *Entry = CostTableLookup(SSE41ShuffleTbl, Kind, LT.second))
1424 return LT.first * Entry->Cost;
1425
1426 static const CostTblEntry SSSE3ShuffleTbl[] = {
1427 {TTI::SK_Broadcast, MVT::v8i16, 1}, // pshufb
1428 {TTI::SK_Broadcast, MVT::v16i8, 1}, // pshufb
1429
1430 {TTI::SK_Reverse, MVT::v8i16, 1}, // pshufb
1431 {TTI::SK_Reverse, MVT::v16i8, 1}, // pshufb
1432
1433 {TTI::SK_Select, MVT::v8i16, 3}, // 2*pshufb + por
1434 {TTI::SK_Select, MVT::v16i8, 3}, // 2*pshufb + por
1435
1436 {TTI::SK_PermuteSingleSrc, MVT::v8i16, 1}, // pshufb
1437 {TTI::SK_PermuteSingleSrc, MVT::v16i8, 1}, // pshufb
1438
1439 {TTI::SK_PermuteTwoSrc, MVT::v8i16, 3}, // 2*pshufb + por
1440 {TTI::SK_PermuteTwoSrc, MVT::v16i8, 3}, // 2*pshufb + por
1441 };
1442
1443 if (ST->hasSSSE3())
1444 if (const auto *Entry = CostTableLookup(SSSE3ShuffleTbl, Kind, LT.second))
1445 return LT.first * Entry->Cost;
1446
1447 static const CostTblEntry SSE2ShuffleTbl[] = {
1448 {TTI::SK_Broadcast, MVT::v2f64, 1}, // shufpd
1449 {TTI::SK_Broadcast, MVT::v2i64, 1}, // pshufd
1450 {TTI::SK_Broadcast, MVT::v4i32, 1}, // pshufd
1451 {TTI::SK_Broadcast, MVT::v8i16, 2}, // pshuflw + pshufd
1452 {TTI::SK_Broadcast, MVT::v16i8, 3}, // unpck + pshuflw + pshufd
1453
1454 {TTI::SK_Reverse, MVT::v2f64, 1}, // shufpd
1455 {TTI::SK_Reverse, MVT::v2i64, 1}, // pshufd
1456 {TTI::SK_Reverse, MVT::v4i32, 1}, // pshufd
1457 {TTI::SK_Reverse, MVT::v8i16, 3}, // pshuflw + pshufhw + pshufd
1458 {TTI::SK_Reverse, MVT::v16i8, 9}, // 2*pshuflw + 2*pshufhw
1459 // + 2*pshufd + 2*unpck + packus
1460
1461 {TTI::SK_Select, MVT::v2i64, 1}, // movsd
1462 {TTI::SK_Select, MVT::v2f64, 1}, // movsd
1463 {TTI::SK_Select, MVT::v4i32, 2}, // 2*shufps
1464 {TTI::SK_Select, MVT::v8i16, 3}, // pand + pandn + por
1465 {TTI::SK_Select, MVT::v16i8, 3}, // pand + pandn + por
1466
1467 {TTI::SK_PermuteSingleSrc, MVT::v2f64, 1}, // shufpd
1468 {TTI::SK_PermuteSingleSrc, MVT::v2i64, 1}, // pshufd
1469 {TTI::SK_PermuteSingleSrc, MVT::v4i32, 1}, // pshufd
1470 {TTI::SK_PermuteSingleSrc, MVT::v8i16, 5}, // 2*pshuflw + 2*pshufhw
1471 // + pshufd/unpck
1472 { TTI::SK_PermuteSingleSrc, MVT::v16i8, 10 }, // 2*pshuflw + 2*pshufhw
1473 // + 2*pshufd + 2*unpck + 2*packus
1474
1475 { TTI::SK_PermuteTwoSrc, MVT::v2f64, 1 }, // shufpd
1476 { TTI::SK_PermuteTwoSrc, MVT::v2i64, 1 }, // shufpd
1477 { TTI::SK_PermuteTwoSrc, MVT::v4i32, 2 }, // 2*{unpck,movsd,pshufd}
1478 { TTI::SK_PermuteTwoSrc, MVT::v8i16, 8 }, // blend+permute
1479 { TTI::SK_PermuteTwoSrc, MVT::v16i8, 13 }, // blend+permute
1480 };
1481
1482 if (ST->hasSSE2())
1483 if (const auto *Entry = CostTableLookup(SSE2ShuffleTbl, Kind, LT.second))
1484 return LT.first * Entry->Cost;
1485
1486 static const CostTblEntry SSE1ShuffleTbl[] = {
1487 { TTI::SK_Broadcast, MVT::v4f32, 1 }, // shufps
1488 { TTI::SK_Reverse, MVT::v4f32, 1 }, // shufps
1489 { TTI::SK_Select, MVT::v4f32, 2 }, // 2*shufps
1490 { TTI::SK_PermuteSingleSrc, MVT::v4f32, 1 }, // shufps
1491 { TTI::SK_PermuteTwoSrc, MVT::v4f32, 2 }, // 2*shufps
1492 };
1493
1494 if (ST->hasSSE1())
1495 if (const auto *Entry = CostTableLookup(SSE1ShuffleTbl, Kind, LT.second))
1496 return LT.first * Entry->Cost;
1497
1498 return BaseT::getShuffleCost(Kind, BaseTp, Mask, Index, SubTp);
1499}
1500
1501InstructionCost X86TTIImpl::getCastInstrCost(unsigned Opcode, Type *Dst,
1502 Type *Src,
1503 TTI::CastContextHint CCH,
1504 TTI::TargetCostKind CostKind,
1505 const Instruction *I) {
1506 int ISD = TLI->InstructionOpcodeToISD(Opcode);
1507 assert(ISD && "Invalid opcode")((void)0);
1508
1509 // TODO: Allow non-throughput costs that aren't binary.
1510 auto AdjustCost = [&CostKind](InstructionCost Cost) -> InstructionCost {
1511 if (CostKind != TTI::TCK_RecipThroughput)
1512 return Cost == 0 ? 0 : 1;
1513 return Cost;
1514 };
1515
1516 // The cost tables include both specific, custom (non-legal) src/dst type
1517 // conversions and generic, legalized types. We test for customs first, before
1518 // falling back to legalization.
1519 // FIXME: Need a better design of the cost table to handle non-simple types of
1520 // potential massive combinations (elem_num x src_type x dst_type).
1521 static const TypeConversionCostTblEntry AVX512BWConversionTbl[] {
1522 { ISD::SIGN_EXTEND, MVT::v32i16, MVT::v32i8, 1 },
1523 { ISD::ZERO_EXTEND, MVT::v32i16, MVT::v32i8, 1 },
1524
1525 // Mask sign extend has an instruction.
1526 { ISD::SIGN_EXTEND, MVT::v2i8, MVT::v2i1, 1 },
1527 { ISD::SIGN_EXTEND, MVT::v2i16, MVT::v2i1, 1 },
1528 { ISD::SIGN_EXTEND, MVT::v4i8, MVT::v4i1, 1 },
1529 { ISD::SIGN_EXTEND, MVT::v4i16, MVT::v4i1, 1 },
1530 { ISD::SIGN_EXTEND, MVT::v8i8, MVT::v8i1, 1 },
1531 { ISD::SIGN_EXTEND, MVT::v8i16, MVT::v8i1, 1 },
1532 { ISD::SIGN_EXTEND, MVT::v16i8, MVT::v16i1, 1 },
1533 { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i1, 1 },
1534 { ISD::SIGN_EXTEND, MVT::v32i8, MVT::v32i1, 1 },
1535 { ISD::SIGN_EXTEND, MVT::v32i16, MVT::v32i1, 1 },
1536 { ISD::SIGN_EXTEND, MVT::v64i8, MVT::v64i1, 1 },
1537
1538 // Mask zero extend is a sext + shift.
1539 { ISD::ZERO_EXTEND, MVT::v2i8, MVT::v2i1, 2 },
1540 { ISD::ZERO_EXTEND, MVT::v2i16, MVT::v2i1, 2 },
1541 { ISD::ZERO_EXTEND, MVT::v4i8, MVT::v4i1, 2 },
1542 { ISD::ZERO_EXTEND, MVT::v4i16, MVT::v4i1, 2 },
1543 { ISD::ZERO_EXTEND, MVT::v8i8, MVT::v8i1, 2 },
1544 { ISD::ZERO_EXTEND, MVT::v8i16, MVT::v8i1, 2 },
1545 { ISD::ZERO_EXTEND, MVT::v16i8, MVT::v16i1, 2 },
1546 { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i1, 2 },
1547 { ISD::ZERO_EXTEND, MVT::v32i8, MVT::v32i1, 2 },
1548 { ISD::ZERO_EXTEND, MVT::v32i16, MVT::v32i1, 2 },
1549 { ISD::ZERO_EXTEND, MVT::v64i8, MVT::v64i1, 2 },
1550
1551 { ISD::TRUNCATE, MVT::v32i8, MVT::v32i16, 2 },
1552 { ISD::TRUNCATE, MVT::v16i8, MVT::v16i16, 2 }, // widen to zmm
1553 { ISD::TRUNCATE, MVT::v2i1, MVT::v2i8, 2 }, // widen to zmm
1554 { ISD::TRUNCATE, MVT::v2i1, MVT::v2i16, 2 }, // widen to zmm
1555 { ISD::TRUNCATE, MVT::v2i8, MVT::v2i16, 2 }, // vpmovwb
1556 { ISD::TRUNCATE, MVT::v4i1, MVT::v4i8, 2 }, // widen to zmm
1557 { ISD::TRUNCATE, MVT::v4i1, MVT::v4i16, 2 }, // widen to zmm
1558 { ISD::TRUNCATE, MVT::v4i8, MVT::v4i16, 2 }, // vpmovwb
1559 { ISD::TRUNCATE, MVT::v8i1, MVT::v8i8, 2 }, // widen to zmm
1560 { ISD::TRUNCATE, MVT::v8i1, MVT::v8i16, 2 }, // widen to zmm
1561 { ISD::TRUNCATE, MVT::v8i8, MVT::v8i16, 2 }, // vpmovwb
1562 { ISD::TRUNCATE, MVT::v16i1, MVT::v16i8, 2 }, // widen to zmm
1563 { ISD::TRUNCATE, MVT::v16i1, MVT::v16i16, 2 }, // widen to zmm
1564 { ISD::TRUNCATE, MVT::v32i1, MVT::v32i8, 2 }, // widen to zmm
1565 { ISD::TRUNCATE, MVT::v32i1, MVT::v32i16, 2 },
1566 { ISD::TRUNCATE, MVT::v64i1, MVT::v64i8, 2 },
1567 };
1568
1569 static const TypeConversionCostTblEntry AVX512DQConversionTbl[] = {
1570 { ISD::SINT_TO_FP, MVT::v8f32, MVT::v8i64, 1 },
1571 { ISD::SINT_TO_FP, MVT::v8f64, MVT::v8i64, 1 },
1572
1573 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i64, 1 },
1574 { ISD::UINT_TO_FP, MVT::v8f64, MVT::v8i64, 1 },
1575
1576 { ISD::FP_TO_SINT, MVT::v8i64, MVT::v8f32, 1 },
1577 { ISD::FP_TO_SINT, MVT::v8i64, MVT::v8f64, 1 },
1578
1579 { ISD::FP_TO_UINT, MVT::v8i64, MVT::v8f32, 1 },
1580 { ISD::FP_TO_UINT, MVT::v8i64, MVT::v8f64, 1 },
1581 };
1582
1583 // TODO: For AVX512DQ + AVX512VL, we also have cheap casts for 128-bit and
1584 // 256-bit wide vectors.
1585
1586 static const TypeConversionCostTblEntry AVX512FConversionTbl[] = {
1587 { ISD::FP_EXTEND, MVT::v8f64, MVT::v8f32, 1 },
1588 { ISD::FP_EXTEND, MVT::v8f64, MVT::v16f32, 3 },
1589 { ISD::FP_ROUND, MVT::v8f32, MVT::v8f64, 1 },
1590
1591 { ISD::TRUNCATE, MVT::v2i1, MVT::v2i8, 3 }, // sext+vpslld+vptestmd
1592 { ISD::TRUNCATE, MVT::v4i1, MVT::v4i8, 3 }, // sext+vpslld+vptestmd
1593 { ISD::TRUNCATE, MVT::v8i1, MVT::v8i8, 3 }, // sext+vpslld+vptestmd
1594 { ISD::TRUNCATE, MVT::v16i1, MVT::v16i8, 3 }, // sext+vpslld+vptestmd
1595 { ISD::TRUNCATE, MVT::v2i1, MVT::v2i16, 3 }, // sext+vpsllq+vptestmq
1596 { ISD::TRUNCATE, MVT::v4i1, MVT::v4i16, 3 }, // sext+vpsllq+vptestmq
1597 { ISD::TRUNCATE, MVT::v8i1, MVT::v8i16, 3 }, // sext+vpsllq+vptestmq
1598 { ISD::TRUNCATE, MVT::v16i1, MVT::v16i16, 3 }, // sext+vpslld+vptestmd
1599 { ISD::TRUNCATE, MVT::v2i1, MVT::v2i32, 2 }, // zmm vpslld+vptestmd
1600 { ISD::TRUNCATE, MVT::v4i1, MVT::v4i32, 2 }, // zmm vpslld+vptestmd
1601 { ISD::TRUNCATE, MVT::v8i1, MVT::v8i32, 2 }, // zmm vpslld+vptestmd
1602 { ISD::TRUNCATE, MVT::v16i1, MVT::v16i32, 2 }, // vpslld+vptestmd
1603 { ISD::TRUNCATE, MVT::v2i1, MVT::v2i64, 2 }, // zmm vpsllq+vptestmq
1604 { ISD::TRUNCATE, MVT::v4i1, MVT::v4i64, 2 }, // zmm vpsllq+vptestmq
1605 { ISD::TRUNCATE, MVT::v8i1, MVT::v8i64, 2 }, // vpsllq+vptestmq
1606 { ISD::TRUNCATE, MVT::v2i8, MVT::v2i32, 2 }, // vpmovdb
1607 { ISD::TRUNCATE, MVT::v4i8, MVT::v4i32, 2 }, // vpmovdb
1608 { ISD::TRUNCATE, MVT::v16i8, MVT::v16i32, 2 }, // vpmovdb
1609 { ISD::TRUNCATE, MVT::v16i16, MVT::v16i32, 2 }, // vpmovdb
1610 { ISD::TRUNCATE, MVT::v2i8, MVT::v2i64, 2 }, // vpmovqb
1611 { ISD::TRUNCATE, MVT::v2i16, MVT::v2i64, 1 }, // vpshufb
1612 { ISD::TRUNCATE, MVT::v8i8, MVT::v8i64, 2 }, // vpmovqb
1613 { ISD::TRUNCATE, MVT::v8i16, MVT::v8i64, 2 }, // vpmovqw
1614 { ISD::TRUNCATE, MVT::v8i32, MVT::v8i64, 1 }, // vpmovqd
1615 { ISD::TRUNCATE, MVT::v4i32, MVT::v4i64, 1 }, // zmm vpmovqd
1616 { ISD::TRUNCATE, MVT::v16i8, MVT::v16i64, 5 },// 2*vpmovqd+concat+vpmovdb
1617
1618 { ISD::TRUNCATE, MVT::v16i8, MVT::v16i16, 3 }, // extend to v16i32
1619 { ISD::TRUNCATE, MVT::v32i8, MVT::v32i16, 8 },
1620
1621 // Sign extend is zmm vpternlogd+vptruncdb.
1622 // Zero extend is zmm broadcast load+vptruncdw.
1623 { ISD::SIGN_EXTEND, MVT::v2i8, MVT::v2i1, 3 },
1624 { ISD::ZERO_EXTEND, MVT::v2i8, MVT::v2i1, 4 },
1625 { ISD::SIGN_EXTEND, MVT::v4i8, MVT::v4i1, 3 },
1626 { ISD::ZERO_EXTEND, MVT::v4i8, MVT::v4i1, 4 },
1627 { ISD::SIGN_EXTEND, MVT::v8i8, MVT::v8i1, 3 },
1628 { ISD::ZERO_EXTEND, MVT::v8i8, MVT::v8i1, 4 },
1629 { ISD::SIGN_EXTEND, MVT::v16i8, MVT::v16i1, 3 },
1630 { ISD::ZERO_EXTEND, MVT::v16i8, MVT::v16i1, 4 },
1631
1632 // Sign extend is zmm vpternlogd+vptruncdw.
1633 // Zero extend is zmm vpternlogd+vptruncdw+vpsrlw.
1634 { ISD::SIGN_EXTEND, MVT::v2i16, MVT::v2i1, 3 },
1635 { ISD::ZERO_EXTEND, MVT::v2i16, MVT::v2i1, 4 },
1636 { ISD::SIGN_EXTEND, MVT::v4i16, MVT::v4i1, 3 },
1637 { ISD::ZERO_EXTEND, MVT::v4i16, MVT::v4i1, 4 },
1638 { ISD::SIGN_EXTEND, MVT::v8i16, MVT::v8i1, 3 },
1639 { ISD::ZERO_EXTEND, MVT::v8i16, MVT::v8i1, 4 },
1640 { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i1, 3 },
1641 { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i1, 4 },
1642
1643 { ISD::SIGN_EXTEND, MVT::v2i32, MVT::v2i1, 1 }, // zmm vpternlogd
1644 { ISD::ZERO_EXTEND, MVT::v2i32, MVT::v2i1, 2 }, // zmm vpternlogd+psrld
1645 { ISD::SIGN_EXTEND, MVT::v4i32, MVT::v4i1, 1 }, // zmm vpternlogd
1646 { ISD::ZERO_EXTEND, MVT::v4i32, MVT::v4i1, 2 }, // zmm vpternlogd+psrld
1647 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i1, 1 }, // zmm vpternlogd
1648 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i1, 2 }, // zmm vpternlogd+psrld
1649 { ISD::SIGN_EXTEND, MVT::v2i64, MVT::v2i1, 1 }, // zmm vpternlogq
1650 { ISD::ZERO_EXTEND, MVT::v2i64, MVT::v2i1, 2 }, // zmm vpternlogq+psrlq
1651 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i1, 1 }, // zmm vpternlogq
1652 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i1, 2 }, // zmm vpternlogq+psrlq
1653
1654 { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i1, 1 }, // vpternlogd
1655 { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i1, 2 }, // vpternlogd+psrld
1656 { ISD::SIGN_EXTEND, MVT::v8i64, MVT::v8i1, 1 }, // vpternlogq
1657 { ISD::ZERO_EXTEND, MVT::v8i64, MVT::v8i1, 2 }, // vpternlogq+psrlq
1658
1659 { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i8, 1 },
1660 { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i8, 1 },
1661 { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i16, 1 },
1662 { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i16, 1 },
1663 { ISD::SIGN_EXTEND, MVT::v8i64, MVT::v8i8, 1 },
1664 { ISD::ZERO_EXTEND, MVT::v8i64, MVT::v8i8, 1 },
1665 { ISD::SIGN_EXTEND, MVT::v8i64, MVT::v8i16, 1 },
1666 { ISD::ZERO_EXTEND, MVT::v8i64, MVT::v8i16, 1 },
1667 { ISD::SIGN_EXTEND, MVT::v8i64, MVT::v8i32, 1 },
1668 { ISD::ZERO_EXTEND, MVT::v8i64, MVT::v8i32, 1 },
1669
1670 { ISD::SIGN_EXTEND, MVT::v32i16, MVT::v32i8, 3 }, // FIXME: May not be right
1671 { ISD::ZERO_EXTEND, MVT::v32i16, MVT::v32i8, 3 }, // FIXME: May not be right
1672
1673 { ISD::SINT_TO_FP, MVT::v8f64, MVT::v8i1, 4 },
1674 { ISD::SINT_TO_FP, MVT::v16f32, MVT::v16i1, 3 },
1675 { ISD::SINT_TO_FP, MVT::v8f64, MVT::v16i8, 2 },
1676 { ISD::SINT_TO_FP, MVT::v16f32, MVT::v16i8, 1 },
1677 { ISD::SINT_TO_FP, MVT::v8f64, MVT::v8i16, 2 },
1678 { ISD::SINT_TO_FP, MVT::v16f32, MVT::v16i16, 1 },
1679 { ISD::SINT_TO_FP, MVT::v8f64, MVT::v8i32, 1 },
1680 { ISD::SINT_TO_FP, MVT::v16f32, MVT::v16i32, 1 },
1681
1682 { ISD::UINT_TO_FP, MVT::v8f64, MVT::v8i1, 4 },
1683 { ISD::UINT_TO_FP, MVT::v16f32, MVT::v16i1, 3 },
1684 { ISD::UINT_TO_FP, MVT::v8f64, MVT::v16i8, 2 },
1685 { ISD::UINT_TO_FP, MVT::v16f32, MVT::v16i8, 1 },
1686 { ISD::UINT_TO_FP, MVT::v8f64, MVT::v8i16, 2 },
1687 { ISD::UINT_TO_FP, MVT::v16f32, MVT::v16i16, 1 },
1688 { ISD::UINT_TO_FP, MVT::v8f64, MVT::v8i32, 1 },
1689 { ISD::UINT_TO_FP, MVT::v16f32, MVT::v16i32, 1 },
1690 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i64, 26 },
1691 { ISD::UINT_TO_FP, MVT::v8f64, MVT::v8i64, 5 },
1692
1693 { ISD::FP_TO_SINT, MVT::v16i8, MVT::v16f32, 2 },
1694 { ISD::FP_TO_SINT, MVT::v16i8, MVT::v16f64, 7 },
1695 { ISD::FP_TO_SINT, MVT::v32i8, MVT::v32f64,15 },
1696 { ISD::FP_TO_SINT, MVT::v64i8, MVT::v64f32,11 },
1697 { ISD::FP_TO_SINT, MVT::v64i8, MVT::v64f64,31 },
1698 { ISD::FP_TO_SINT, MVT::v8i16, MVT::v8f64, 3 },
1699 { ISD::FP_TO_SINT, MVT::v16i16, MVT::v16f64, 7 },
1700 { ISD::FP_TO_SINT, MVT::v32i16, MVT::v32f32, 5 },
1701 { ISD::FP_TO_SINT, MVT::v32i16, MVT::v32f64,15 },
1702 { ISD::FP_TO_SINT, MVT::v8i32, MVT::v8f64, 1 },
1703 { ISD::FP_TO_SINT, MVT::v16i32, MVT::v16f64, 3 },
1704
1705 { ISD::FP_TO_UINT, MVT::v8i32, MVT::v8f64, 1 },
1706 { ISD::FP_TO_UINT, MVT::v8i16, MVT::v8f64, 3 },
1707 { ISD::FP_TO_UINT, MVT::v8i8, MVT::v8f64, 3 },
1708 { ISD::FP_TO_UINT, MVT::v16i32, MVT::v16f32, 1 },
1709 { ISD::FP_TO_UINT, MVT::v16i16, MVT::v16f32, 3 },
1710 { ISD::FP_TO_UINT, MVT::v16i8, MVT::v16f32, 3 },
1711 };
1712
1713 static const TypeConversionCostTblEntry AVX512BWVLConversionTbl[] {
1714 // Mask sign extend has an instruction.
1715 { ISD::SIGN_EXTEND, MVT::v2i8, MVT::v2i1, 1 },
1716 { ISD::SIGN_EXTEND, MVT::v2i16, MVT::v2i1, 1 },
1717 { ISD::SIGN_EXTEND, MVT::v4i8, MVT::v4i1, 1 },
1718 { ISD::SIGN_EXTEND, MVT::v4i16, MVT::v4i1, 1 },
1719 { ISD::SIGN_EXTEND, MVT::v8i8, MVT::v8i1, 1 },
1720 { ISD::SIGN_EXTEND, MVT::v8i16, MVT::v8i1, 1 },
1721 { ISD::SIGN_EXTEND, MVT::v16i8, MVT::v16i1, 1 },
1722 { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i1, 1 },
1723 { ISD::SIGN_EXTEND, MVT::v32i8, MVT::v32i1, 1 },
1724
1725 // Mask zero extend is a sext + shift.
1726 { ISD::ZERO_EXTEND, MVT::v2i8, MVT::v2i1, 2 },
1727 { ISD::ZERO_EXTEND, MVT::v2i16, MVT::v2i1, 2 },
1728 { ISD::ZERO_EXTEND, MVT::v4i8, MVT::v4i1, 2 },
1729 { ISD::ZERO_EXTEND, MVT::v4i16, MVT::v4i1, 2 },
1730 { ISD::ZERO_EXTEND, MVT::v8i8, MVT::v8i1, 2 },
1731 { ISD::ZERO_EXTEND, MVT::v8i16, MVT::v8i1, 2 },
1732 { ISD::ZERO_EXTEND, MVT::v16i8, MVT::v16i1, 2 },
1733 { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i1, 2 },
1734 { ISD::ZERO_EXTEND, MVT::v32i8, MVT::v32i1, 2 },
1735
1736 { ISD::TRUNCATE, MVT::v16i8, MVT::v16i16, 2 },
1737 { ISD::TRUNCATE, MVT::v2i1, MVT::v2i8, 2 }, // vpsllw+vptestmb
1738 { ISD::TRUNCATE, MVT::v2i1, MVT::v2i16, 2 }, // vpsllw+vptestmw
1739 { ISD::TRUNCATE, MVT::v4i1, MVT::v4i8, 2 }, // vpsllw+vptestmb
1740 { ISD::TRUNCATE, MVT::v4i1, MVT::v4i16, 2 }, // vpsllw+vptestmw
1741 { ISD::TRUNCATE, MVT::v8i1, MVT::v8i8, 2 }, // vpsllw+vptestmb
1742 { ISD::TRUNCATE, MVT::v8i1, MVT::v8i16, 2 }, // vpsllw+vptestmw
1743 { ISD::TRUNCATE, MVT::v16i1, MVT::v16i8, 2 }, // vpsllw+vptestmb
1744 { ISD::TRUNCATE, MVT::v16i1, MVT::v16i16, 2 }, // vpsllw+vptestmw
1745 { ISD::TRUNCATE, MVT::v32i1, MVT::v32i8, 2 }, // vpsllw+vptestmb
1746 };
1747
1748 static const TypeConversionCostTblEntry AVX512DQVLConversionTbl[] = {
1749 { ISD::SINT_TO_FP, MVT::v2f32, MVT::v2i64, 1 },
1750 { ISD::SINT_TO_FP, MVT::v2f64, MVT::v2i64, 1 },
1751 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i64, 1 },
1752 { ISD::SINT_TO_FP, MVT::v4f64, MVT::v4i64, 1 },
1753
1754 { ISD::UINT_TO_FP, MVT::v2f32, MVT::v2i64, 1 },
1755 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i64, 1 },
1756 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i64, 1 },
1757 { ISD::UINT_TO_FP, MVT::v4f64, MVT::v4i64, 1 },
1758
1759 { ISD::FP_TO_SINT, MVT::v2i64, MVT::v4f32, 1 },
1760 { ISD::FP_TO_SINT, MVT::v4i64, MVT::v4f32, 1 },
1761 { ISD::FP_TO_SINT, MVT::v2i64, MVT::v2f64, 1 },
1762 { ISD::FP_TO_SINT, MVT::v4i64, MVT::v4f64, 1 },
1763
1764 { ISD::FP_TO_UINT, MVT::v2i64, MVT::v4f32, 1 },
1765 { ISD::FP_TO_UINT, MVT::v4i64, MVT::v4f32, 1 },
1766 { ISD::FP_TO_UINT, MVT::v2i64, MVT::v2f64, 1 },
1767 { ISD::FP_TO_UINT, MVT::v4i64, MVT::v4f64, 1 },
1768 };
1769
1770 static const TypeConversionCostTblEntry AVX512VLConversionTbl[] = {
1771 { ISD::TRUNCATE, MVT::v2i1, MVT::v2i8, 3 }, // sext+vpslld+vptestmd
1772 { ISD::TRUNCATE, MVT::v4i1, MVT::v4i8, 3 }, // sext+vpslld+vptestmd
1773 { ISD::TRUNCATE, MVT::v8i1, MVT::v8i8, 3 }, // sext+vpslld+vptestmd
1774 { ISD::TRUNCATE, MVT::v16i1, MVT::v16i8, 8 }, // split+2*v8i8
1775 { ISD::TRUNCATE, MVT::v2i1, MVT::v2i16, 3 }, // sext+vpsllq+vptestmq
1776 { ISD::TRUNCATE, MVT::v4i1, MVT::v4i16, 3 }, // sext+vpsllq+vptestmq
1777 { ISD::TRUNCATE, MVT::v8i1, MVT::v8i16, 3 }, // sext+vpsllq+vptestmq
1778 { ISD::TRUNCATE, MVT::v16i1, MVT::v16i16, 8 }, // split+2*v8i16
1779 { ISD::TRUNCATE, MVT::v2i1, MVT::v2i32, 2 }, // vpslld+vptestmd
1780 { ISD::TRUNCATE, MVT::v4i1, MVT::v4i32, 2 }, // vpslld+vptestmd
1781 { ISD::TRUNCATE, MVT::v8i1, MVT::v8i32, 2 }, // vpslld+vptestmd
1782 { ISD::TRUNCATE, MVT::v2i1, MVT::v2i64, 2 }, // vpsllq+vptestmq
1783 { ISD::TRUNCATE, MVT::v4i1, MVT::v4i64, 2 }, // vpsllq+vptestmq
1784 { ISD::TRUNCATE, MVT::v4i32, MVT::v4i64, 1 }, // vpmovqd
1785 { ISD::TRUNCATE, MVT::v4i8, MVT::v4i64, 2 }, // vpmovqb
1786 { ISD::TRUNCATE, MVT::v4i16, MVT::v4i64, 2 }, // vpmovqw
1787 { ISD::TRUNCATE, MVT::v8i8, MVT::v8i32, 2 }, // vpmovwb
1788
1789 // sign extend is vpcmpeq+maskedmove+vpmovdw+vpacksswb
1790 // zero extend is vpcmpeq+maskedmove+vpmovdw+vpsrlw+vpackuswb
1791 { ISD::SIGN_EXTEND, MVT::v2i8, MVT::v2i1, 5 },
1792 { ISD::ZERO_EXTEND, MVT::v2i8, MVT::v2i1, 6 },
1793 { ISD::SIGN_EXTEND, MVT::v4i8, MVT::v4i1, 5 },
1794 { ISD::ZERO_EXTEND, MVT::v4i8, MVT::v4i1, 6 },
1795 { ISD::SIGN_EXTEND, MVT::v8i8, MVT::v8i1, 5 },
1796 { ISD::ZERO_EXTEND, MVT::v8i8, MVT::v8i1, 6 },
1797 { ISD::SIGN_EXTEND, MVT::v16i8, MVT::v16i1, 10 },
1798 { ISD::ZERO_EXTEND, MVT::v16i8, MVT::v16i1, 12 },
1799
1800 // sign extend is vpcmpeq+maskedmove+vpmovdw
1801 // zero extend is vpcmpeq+maskedmove+vpmovdw+vpsrlw
1802 { ISD::SIGN_EXTEND, MVT::v2i16, MVT::v2i1, 4 },
1803 { ISD::ZERO_EXTEND, MVT::v2i16, MVT::v2i1, 5 },
1804 { ISD::SIGN_EXTEND, MVT::v4i16, MVT::v4i1, 4 },
1805 { ISD::ZERO_EXTEND, MVT::v4i16, MVT::v4i1, 5 },
1806 { ISD::SIGN_EXTEND, MVT::v8i16, MVT::v8i1, 4 },
1807 { ISD::ZERO_EXTEND, MVT::v8i16, MVT::v8i1, 5 },
1808 { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i1, 10 },
1809 { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i1, 12 },
1810
1811 { ISD::SIGN_EXTEND, MVT::v2i32, MVT::v2i1, 1 }, // vpternlogd
1812 { ISD::ZERO_EXTEND, MVT::v2i32, MVT::v2i1, 2 }, // vpternlogd+psrld
1813 { ISD::SIGN_EXTEND, MVT::v4i32, MVT::v4i1, 1 }, // vpternlogd
1814 { ISD::ZERO_EXTEND, MVT::v4i32, MVT::v4i1, 2 }, // vpternlogd+psrld
1815 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i1, 1 }, // vpternlogd
1816 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i1, 2 }, // vpternlogd+psrld
1817 { ISD::SIGN_EXTEND, MVT::v2i64, MVT::v2i1, 1 }, // vpternlogq
1818 { ISD::ZERO_EXTEND, MVT::v2i64, MVT::v2i1, 2 }, // vpternlogq+psrlq
1819 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i1, 1 }, // vpternlogq
1820 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i1, 2 }, // vpternlogq+psrlq
1821
1822 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v16i8, 1 },
1823 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v16i8, 1 },
1824 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v16i8, 1 },
1825 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v16i8, 1 },
1826 { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i8, 1 },
1827 { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i8, 1 },
1828 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v8i16, 1 },
1829 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v8i16, 1 },
1830 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i16, 1 },
1831 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i16, 1 },
1832 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i32, 1 },
1833 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i32, 1 },
1834
1835 { ISD::SINT_TO_FP, MVT::v2f64, MVT::v16i8, 1 },
1836 { ISD::SINT_TO_FP, MVT::v8f32, MVT::v16i8, 1 },
1837 { ISD::SINT_TO_FP, MVT::v2f64, MVT::v8i16, 1 },
1838 { ISD::SINT_TO_FP, MVT::v8f32, MVT::v8i16, 1 },
1839
1840 { ISD::UINT_TO_FP, MVT::f32, MVT::i64, 1 },
1841 { ISD::UINT_TO_FP, MVT::f64, MVT::i64, 1 },
1842 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v16i8, 1 },
1843 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v16i8, 1 },
1844 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v8i16, 1 },
1845 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i16, 1 },
1846 { ISD::UINT_TO_FP, MVT::v2f32, MVT::v2i32, 1 },
1847 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i32, 1 },
1848 { ISD::UINT_TO_FP, MVT::v4f64, MVT::v4i32, 1 },
1849 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i32, 1 },
1850 { ISD::UINT_TO_FP, MVT::v2f32, MVT::v2i64, 5 },
1851 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i64, 5 },
1852 { ISD::UINT_TO_FP, MVT::v4f64, MVT::v4i64, 5 },
1853
1854 { ISD::FP_TO_SINT, MVT::v16i8, MVT::v8f32, 2 },
1855 { ISD::FP_TO_SINT, MVT::v16i8, MVT::v16f32, 2 },
1856 { ISD::FP_TO_SINT, MVT::v32i8, MVT::v32f32, 5 },
1857
1858 { ISD::FP_TO_UINT, MVT::i64, MVT::f32, 1 },
1859 { ISD::FP_TO_UINT, MVT::i64, MVT::f64, 1 },
1860 { ISD::FP_TO_UINT, MVT::v4i32, MVT::v4f32, 1 },
1861 { ISD::FP_TO_UINT, MVT::v4i32, MVT::v2f64, 1 },
1862 { ISD::FP_TO_UINT, MVT::v4i32, MVT::v4f64, 1 },
1863 { ISD::FP_TO_UINT, MVT::v8i32, MVT::v8f32, 1 },
1864 { ISD::FP_TO_UINT, MVT::v8i32, MVT::v8f64, 1 },
1865 };
1866
1867 static const TypeConversionCostTblEntry AVX2ConversionTbl[] = {
1868 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i1, 3 },
1869 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i1, 3 },
1870 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i1, 3 },
1871 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i1, 3 },
1872 { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i1, 1 },
1873 { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i1, 1 },
1874
1875 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v16i8, 2 },
1876 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v16i8, 2 },
1877 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v16i8, 2 },
1878 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v16i8, 2 },
1879 { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i8, 2 },
1880 { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i8, 2 },
1881 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v8i16, 2 },
1882 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v8i16, 2 },
1883 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i16, 2 },
1884 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i16, 2 },
1885 { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i16, 3 },
1886 { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i16, 3 },
1887 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i32, 2 },
1888 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i32, 2 },
1889
1890 { ISD::TRUNCATE, MVT::v8i1, MVT::v8i32, 2 },
1891
1892 { ISD::TRUNCATE, MVT::v16i8, MVT::v8i16, 1 },
1893 { ISD::TRUNCATE, MVT::v16i8, MVT::v4i32, 1 },
1894 { ISD::TRUNCATE, MVT::v16i8, MVT::v2i64, 1 },
1895 { ISD::TRUNCATE, MVT::v16i8, MVT::v8i32, 4 },
1896 { ISD::TRUNCATE, MVT::v16i8, MVT::v4i64, 4 },
1897 { ISD::TRUNCATE, MVT::v8i16, MVT::v4i32, 1 },
1898 { ISD::TRUNCATE, MVT::v8i16, MVT::v2i64, 1 },
1899 { ISD::TRUNCATE, MVT::v8i16, MVT::v4i64, 5 },
1900 { ISD::TRUNCATE, MVT::v4i32, MVT::v4i64, 1 },
1901 { ISD::TRUNCATE, MVT::v8i16, MVT::v8i32, 2 },
1902
1903 { ISD::FP_EXTEND, MVT::v8f64, MVT::v8f32, 3 },
1904 { ISD::FP_ROUND, MVT::v8f32, MVT::v8f64, 3 },
1905
1906 { ISD::FP_TO_SINT, MVT::v16i16, MVT::v8f32, 1 },
1907 { ISD::FP_TO_SINT, MVT::v4i32, MVT::v4f64, 1 },
1908 { ISD::FP_TO_SINT, MVT::v8i32, MVT::v8f32, 1 },
1909 { ISD::FP_TO_SINT, MVT::v8i32, MVT::v8f64, 3 },
1910
1911 { ISD::FP_TO_UINT, MVT::i64, MVT::f32, 3 },
1912 { ISD::FP_TO_UINT, MVT::i64, MVT::f64, 3 },
1913 { ISD::FP_TO_UINT, MVT::v16i16, MVT::v8f32, 1 },
1914 { ISD::FP_TO_UINT, MVT::v4i32, MVT::v4f32, 3 },
1915 { ISD::FP_TO_UINT, MVT::v4i32, MVT::v2f64, 4 },
1916 { ISD::FP_TO_UINT, MVT::v4i32, MVT::v4f64, 4 },
1917 { ISD::FP_TO_UINT, MVT::v8i32, MVT::v8f32, 3 },
1918 { ISD::FP_TO_UINT, MVT::v8i32, MVT::v4f64, 4 },
1919
1920 { ISD::SINT_TO_FP, MVT::v2f64, MVT::v16i8, 2 },
1921 { ISD::SINT_TO_FP, MVT::v8f32, MVT::v16i8, 2 },
1922 { ISD::SINT_TO_FP, MVT::v2f64, MVT::v8i16, 2 },
1923 { ISD::SINT_TO_FP, MVT::v8f32, MVT::v8i16, 2 },
1924 { ISD::SINT_TO_FP, MVT::v4f64, MVT::v4i32, 1 },
1925 { ISD::SINT_TO_FP, MVT::v8f32, MVT::v8i32, 1 },
1926 { ISD::SINT_TO_FP, MVT::v8f64, MVT::v8i32, 3 },
1927
1928 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v16i8, 2 },
1929 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v16i8, 2 },
1930 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v8i16, 2 },
1931 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i16, 2 },
1932 { ISD::UINT_TO_FP, MVT::v2f32, MVT::v2i32, 2 },
1933 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i32, 1 },
1934 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i32, 2 },
1935 { ISD::UINT_TO_FP, MVT::v4f64, MVT::v4i32, 2 },
1936 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i32, 2 },
1937 { ISD::UINT_TO_FP, MVT::v8f64, MVT::v8i32, 4 },
1938 };
1939
1940 static const TypeConversionCostTblEntry AVXConversionTbl[] = {
1941 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i1, 6 },
1942 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i1, 4 },
1943 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i1, 7 },
1944 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i1, 4 },
1945 { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i1, 4 },
1946 { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i1, 4 },
1947
1948 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v16i8, 3 },
1949 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v16i8, 3 },
1950 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v16i8, 3 },
1951 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v16i8, 3 },
1952 { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i8, 3 },
1953 { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i8, 3 },
1954 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v8i16, 3 },
1955 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v8i16, 3 },
1956 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i16, 3 },
1957 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i16, 3 },
1958 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i32, 3 },
1959 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i32, 3 },
1960
1961 { ISD::TRUNCATE, MVT::v4i1, MVT::v4i64, 4 },
1962 { ISD::TRUNCATE, MVT::v8i1, MVT::v8i32, 5 },
1963 { ISD::TRUNCATE, MVT::v16i1, MVT::v16i16, 4 },
1964 { ISD::TRUNCATE, MVT::v8i1, MVT::v8i64, 9 },
1965 { ISD::TRUNCATE, MVT::v16i1, MVT::v16i64, 11 },
1966
1967 { ISD::TRUNCATE, MVT::v16i8, MVT::v16i16, 2 }, // and+extract+packuswb
1968 { ISD::TRUNCATE, MVT::v16i8, MVT::v8i32, 5 },
1969 { ISD::TRUNCATE, MVT::v8i16, MVT::v8i32, 5 },
1970 { ISD::TRUNCATE, MVT::v16i8, MVT::v4i64, 5 },
1971 { ISD::TRUNCATE, MVT::v8i16, MVT::v4i64, 3 }, // and+extract+2*packusdw
1972 { ISD::TRUNCATE, MVT::v4i32, MVT::v4i64, 2 },
1973
1974 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i1, 3 },
1975 { ISD::SINT_TO_FP, MVT::v4f64, MVT::v4i1, 3 },
1976 { ISD::SINT_TO_FP, MVT::v8f32, MVT::v8i1, 8 },
1977 { ISD::SINT_TO_FP, MVT::v8f32, MVT::v16i8, 4 },
1978 { ISD::SINT_TO_FP, MVT::v4f64, MVT::v16i8, 2 },
1979 { ISD::SINT_TO_FP, MVT::v8f32, MVT::v8i16, 4 },
1980 { ISD::SINT_TO_FP, MVT::v4f64, MVT::v8i16, 2 },
1981 { ISD::SINT_TO_FP, MVT::v4f64, MVT::v4i32, 2 },
1982 { ISD::SINT_TO_FP, MVT::v8f32, MVT::v8i32, 2 },
1983 { ISD::SINT_TO_FP, MVT::v8f64, MVT::v8i32, 4 },
1984 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v2i64, 5 },
1985 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i64, 8 },
1986
1987 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i1, 7 },
1988 { ISD::UINT_TO_FP, MVT::v4f64, MVT::v4i1, 7 },
1989 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i1, 6 },
1990 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v16i8, 4 },
1991 { ISD::UINT_TO_FP, MVT::v4f64, MVT::v16i8, 2 },
1992 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i16, 4 },
1993 { ISD::UINT_TO_FP, MVT::v4f64, MVT::v8i16, 2 },
1994 { ISD::UINT_TO_FP, MVT::v2f32, MVT::v2i32, 4 },
1995 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i32, 4 },
1996 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i32, 5 },
1997 { ISD::UINT_TO_FP, MVT::v4f64, MVT::v4i32, 6 },
1998 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i32, 8 },
1999 { ISD::UINT_TO_FP, MVT::v8f64, MVT::v8i32, 10 },
2000 { ISD::UINT_TO_FP, MVT::v2f32, MVT::v2i64, 10 },
2001 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i64, 18 },
2002 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i64, 5 },
2003 { ISD::UINT_TO_FP, MVT::v4f64, MVT::v4i64, 10 },
2004
2005 { ISD::FP_TO_SINT, MVT::v16i8, MVT::v8f32, 2 },
2006 { ISD::FP_TO_SINT, MVT::v16i8, MVT::v4f64, 2 },
2007 { ISD::FP_TO_SINT, MVT::v32i8, MVT::v8f32, 2 },
2008 { ISD::FP_TO_SINT, MVT::v32i8, MVT::v4f64, 2 },
2009 { ISD::FP_TO_SINT, MVT::v8i16, MVT::v8f32, 2 },
2010 { ISD::FP_TO_SINT, MVT::v8i16, MVT::v4f64, 2 },
2011 { ISD::FP_TO_SINT, MVT::v16i16, MVT::v8f32, 2 },
2012 { ISD::FP_TO_SINT, MVT::v16i16, MVT::v4f64, 2 },
2013 { ISD::FP_TO_SINT, MVT::v4i32, MVT::v4f64, 2 },
2014 { ISD::FP_TO_SINT, MVT::v8i32, MVT::v8f32, 2 },
2015 { ISD::FP_TO_SINT, MVT::v8i32, MVT::v8f64, 5 },
2016
2017 { ISD::FP_TO_UINT, MVT::v16i8, MVT::v8f32, 2 },
2018 { ISD::FP_TO_UINT, MVT::v16i8, MVT::v4f64, 2 },
2019 { ISD::FP_TO_UINT, MVT::v32i8, MVT::v8f32, 2 },
2020 { ISD::FP_TO_UINT, MVT::v32i8, MVT::v4f64, 2 },
2021 { ISD::FP_TO_UINT, MVT::v8i16, MVT::v8f32, 2 },
2022 { ISD::FP_TO_UINT, MVT::v8i16, MVT::v4f64, 2 },
2023 { ISD::FP_TO_UINT, MVT::v16i16, MVT::v8f32, 2 },
2024 { ISD::FP_TO_UINT, MVT::v16i16, MVT::v4f64, 2 },
2025 { ISD::FP_TO_UINT, MVT::v4i32, MVT::v4f32, 3 },
2026 { ISD::FP_TO_UINT, MVT::v4i32, MVT::v2f64, 4 },
2027 { ISD::FP_TO_UINT, MVT::v4i32, MVT::v4f64, 6 },
2028 { ISD::FP_TO_UINT, MVT::v8i32, MVT::v8f32, 7 },
2029 { ISD::FP_TO_UINT, MVT::v8i32, MVT::v4f64, 7 },
2030
2031 { ISD::FP_EXTEND, MVT::v4f64, MVT::v4f32, 1 },
2032 { ISD::FP_ROUND, MVT::v4f32, MVT::v4f64, 1 },
2033 };
2034
2035 static const TypeConversionCostTblEntry SSE41ConversionTbl[] = {
2036 { ISD::ZERO_EXTEND, MVT::v2i64, MVT::v16i8, 1 },
2037 { ISD::SIGN_EXTEND, MVT::v2i64, MVT::v16i8, 1 },
2038 { ISD::ZERO_EXTEND, MVT::v4i32, MVT::v16i8, 1 },
2039 { ISD::SIGN_EXTEND, MVT::v4i32, MVT::v16i8, 1 },
2040 { ISD::ZERO_EXTEND, MVT::v8i16, MVT::v16i8, 1 },
2041 { ISD::SIGN_EXTEND, MVT::v8i16, MVT::v16i8, 1 },
2042 { ISD::ZERO_EXTEND, MVT::v2i64, MVT::v8i16, 1 },
2043 { ISD::SIGN_EXTEND, MVT::v2i64, MVT::v8i16, 1 },
2044 { ISD::ZERO_EXTEND, MVT::v4i32, MVT::v8i16, 1 },
2045 { ISD::SIGN_EXTEND, MVT::v4i32, MVT::v8i16, 1 },
2046 { ISD::ZERO_EXTEND, MVT::v2i64, MVT::v4i32, 1 },
2047 { ISD::SIGN_EXTEND, MVT::v2i64, MVT::v4i32, 1 },
2048
2049 // These truncates end up widening elements.
2050 { ISD::TRUNCATE, MVT::v2i1, MVT::v2i8, 1 }, // PMOVXZBQ
2051 { ISD::TRUNCATE, MVT::v2i1, MVT::v2i16, 1 }, // PMOVXZWQ
2052 { ISD::TRUNCATE, MVT::v4i1, MVT::v4i8, 1 }, // PMOVXZBD
2053
2054 { ISD::TRUNCATE, MVT::v16i8, MVT::v4i32, 2 },
2055 { ISD::TRUNCATE, MVT::v8i16, MVT::v4i32, 2 },
2056 { ISD::TRUNCATE, MVT::v16i8, MVT::v2i64, 2 },
2057
2058 { ISD::SINT_TO_FP, MVT::f32, MVT::i32, 1 },
2059 { ISD::SINT_TO_FP, MVT::f64, MVT::i32, 1 },
2060 { ISD::SINT_TO_FP, MVT::f32, MVT::i64, 1 },
2061 { ISD::SINT_TO_FP, MVT::f64, MVT::i64, 1 },
2062 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v16i8, 1 },
2063 { ISD::SINT_TO_FP, MVT::v2f64, MVT::v16i8, 1 },
2064 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v8i16, 1 },
2065 { ISD::SINT_TO_FP, MVT::v2f64, MVT::v8i16, 1 },
2066 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i32, 1 },
2067 { ISD::SINT_TO_FP, MVT::v2f64, MVT::v4i32, 1 },
2068 { ISD::SINT_TO_FP, MVT::v4f64, MVT::v4i32, 2 },
2069
2070 { ISD::UINT_TO_FP, MVT::f32, MVT::i32, 1 },
2071 { ISD::UINT_TO_FP, MVT::f64, MVT::i32, 1 },
2072 { ISD::UINT_TO_FP, MVT::f32, MVT::i64, 4 },
2073 { ISD::UINT_TO_FP, MVT::f64, MVT::i64, 4 },
2074 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v16i8, 1 },
2075 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v16i8, 1 },
2076 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v8i16, 1 },
2077 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v8i16, 1 },
2078 { ISD::UINT_TO_FP, MVT::v2f32, MVT::v2i32, 3 },
2079 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i32, 3 },
2080 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v4i32, 2 },
2081 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v2i64, 12 },
2082 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i64, 22 },
2083 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i64, 4 },
2084
2085 { ISD::FP_TO_SINT, MVT::i32, MVT::f32, 1 },
2086 { ISD::FP_TO_SINT, MVT::i64, MVT::f32, 1 },
2087 { ISD::FP_TO_SINT, MVT::i32, MVT::f64, 1 },
2088 { ISD::FP_TO_SINT, MVT::i64, MVT::f64, 1 },
2089 { ISD::FP_TO_SINT, MVT::v16i8, MVT::v4f32, 2 },
2090 { ISD::FP_TO_SINT, MVT::v16i8, MVT::v2f64, 2 },
2091 { ISD::FP_TO_SINT, MVT::v8i16, MVT::v4f32, 1 },
2092 { ISD::FP_TO_SINT, MVT::v8i16, MVT::v2f64, 1 },
2093 { ISD::FP_TO_SINT, MVT::v4i32, MVT::v4f32, 1 },
2094 { ISD::FP_TO_SINT, MVT::v4i32, MVT::v2f64, 1 },
2095
2096 { ISD::FP_TO_UINT, MVT::i32, MVT::f32, 1 },
2097 { ISD::FP_TO_UINT, MVT::i64, MVT::f32, 4 },
2098 { ISD::FP_TO_UINT, MVT::i32, MVT::f64, 1 },
2099 { ISD::FP_TO_UINT, MVT::i64, MVT::f64, 4 },
2100 { ISD::FP_TO_UINT, MVT::v16i8, MVT::v4f32, 2 },
2101 { ISD::FP_TO_UINT, MVT::v16i8, MVT::v2f64, 2 },
2102 { ISD::FP_TO_UINT, MVT::v8i16, MVT::v4f32, 1 },
2103 { ISD::FP_TO_UINT, MVT::v8i16, MVT::v2f64, 1 },
2104 { ISD::FP_TO_UINT, MVT::v4i32, MVT::v4f32, 4 },
2105 { ISD::FP_TO_UINT, MVT::v4i32, MVT::v2f64, 4 },
2106 };
2107
2108 static const TypeConversionCostTblEntry SSE2ConversionTbl[] = {
2109 // These are somewhat magic numbers justified by comparing the
2110 // output of llvm-mca for our various supported scheduler models
2111 // and basing it off the worst case scenario.
2112 { ISD::SINT_TO_FP, MVT::f32, MVT::i32, 3 },
2113 { ISD::SINT_TO_FP, MVT::f64, MVT::i32, 3 },
2114 { ISD::SINT_TO_FP, MVT::f32, MVT::i64, 3 },
2115 { ISD::SINT_TO_FP, MVT::f64, MVT::i64, 3 },
2116 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v16i8, 3 },
2117 { ISD::SINT_TO_FP, MVT::v2f64, MVT::v16i8, 4 },
2118 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v8i16, 3 },
2119 { ISD::SINT_TO_FP, MVT::v2f64, MVT::v8i16, 4 },
2120 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i32, 3 },
2121 { ISD::SINT_TO_FP, MVT::v2f64, MVT::v4i32, 4 },
2122 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v2i64, 8 },
2123 { ISD::SINT_TO_FP, MVT::v2f64, MVT::v2i64, 8 },
2124
2125 { ISD::UINT_TO_FP, MVT::f32, MVT::i32, 3 },
2126 { ISD::UINT_TO_FP, MVT::f64, MVT::i32, 3 },
2127 { ISD::UINT_TO_FP, MVT::f32, MVT::i64, 8 },
2128 { ISD::UINT_TO_FP, MVT::f64, MVT::i64, 9 },
2129 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v16i8, 4 },
2130 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v16i8, 4 },
2131 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v8i16, 4 },
2132 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v8i16, 4 },
2133 { ISD::UINT_TO_FP, MVT::v2f32, MVT::v2i32, 7 },
2134 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v4i32, 7 },
2135 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i32, 5 },
2136 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i64, 15 },
2137 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v2i64, 18 },
2138
2139 { ISD::FP_TO_SINT, MVT::i32, MVT::f32, 4 },
2140 { ISD::FP_TO_SINT, MVT::i64, MVT::f32, 4 },
2141 { ISD::FP_TO_SINT, MVT::i32, MVT::f64, 4 },
2142 { ISD::FP_TO_SINT, MVT::i64, MVT::f64, 4 },
2143 { ISD::FP_TO_SINT, MVT::v16i8, MVT::v4f32, 6 },
2144 { ISD::FP_TO_SINT, MVT::v16i8, MVT::v2f64, 6 },
2145 { ISD::FP_TO_SINT, MVT::v8i16, MVT::v4f32, 5 },
2146 { ISD::FP_TO_SINT, MVT::v8i16, MVT::v2f64, 5 },
2147 { ISD::FP_TO_SINT, MVT::v4i32, MVT::v4f32, 4 },
2148 { ISD::FP_TO_SINT, MVT::v4i32, MVT::v2f64, 4 },
2149
2150 { ISD::FP_TO_UINT, MVT::i32, MVT::f32, 4 },
2151 { ISD::FP_TO_UINT, MVT::i64, MVT::f32, 4 },
2152 { ISD::FP_TO_UINT, MVT::i32, MVT::f64, 4 },
2153 { ISD::FP_TO_UINT, MVT::i64, MVT::f64, 15 },
2154 { ISD::FP_TO_UINT, MVT::v16i8, MVT::v4f32, 6 },
2155 { ISD::FP_TO_UINT, MVT::v16i8, MVT::v2f64, 6 },
2156 { ISD::FP_TO_UINT, MVT::v8i16, MVT::v4f32, 5 },
2157 { ISD::FP_TO_UINT, MVT::v8i16, MVT::v2f64, 5 },
2158 { ISD::FP_TO_UINT, MVT::v4i32, MVT::v4f32, 8 },
2159 { ISD::FP_TO_UINT, MVT::v4i32, MVT::v2f64, 8 },
2160
2161 { ISD::ZERO_EXTEND, MVT::v2i64, MVT::v16i8, 4 },
2162 { ISD::SIGN_EXTEND, MVT::v2i64, MVT::v16i8, 4 },
2163 { ISD::ZERO_EXTEND, MVT::v4i32, MVT::v16i8, 2 },
2164 { ISD::SIGN_EXTEND, MVT::v4i32, MVT::v16i8, 3 },
2165 { ISD::ZERO_EXTEND, MVT::v8i16, MVT::v16i8, 1 },
2166 { ISD::SIGN_EXTEND, MVT::v8i16, MVT::v16i8, 2 },
2167 { ISD::ZERO_EXTEND, MVT::v2i64, MVT::v8i16, 2 },
2168 { ISD::SIGN_EXTEND, MVT::v2i64, MVT::v8i16, 3 },
2169 { ISD::ZERO_EXTEND, MVT::v4i32, MVT::v8i16, 1 },
2170 { ISD::SIGN_EXTEND, MVT::v4i32, MVT::v8i16, 2 },
2171 { ISD::ZERO_EXTEND, MVT::v2i64, MVT::v4i32, 1 },
2172 { ISD::SIGN_EXTEND, MVT::v2i64, MVT::v4i32, 2 },
2173
2174 // These truncates are really widening elements.
2175 { ISD::TRUNCATE, MVT::v2i1, MVT::v2i32, 1 }, // PSHUFD
2176 { ISD::TRUNCATE, MVT::v2i1, MVT::v2i16, 2 }, // PUNPCKLWD+DQ
2177 { ISD::TRUNCATE, MVT::v2i1, MVT::v2i8, 3 }, // PUNPCKLBW+WD+PSHUFD
2178 { ISD::TRUNCATE, MVT::v4i1, MVT::v4i16, 1 }, // PUNPCKLWD
2179 { ISD::TRUNCATE, MVT::v4i1, MVT::v4i8, 2 }, // PUNPCKLBW+WD
2180 { ISD::TRUNCATE, MVT::v8i1, MVT::v8i8, 1 }, // PUNPCKLBW
2181
2182 { ISD::TRUNCATE, MVT::v16i8, MVT::v8i16, 2 }, // PAND+PACKUSWB
2183 { ISD::TRUNCATE, MVT::v16i8, MVT::v16i16, 3 },
2184 { ISD::TRUNCATE, MVT::v16i8, MVT::v4i32, 3 }, // PAND+2*PACKUSWB
2185 { ISD::TRUNCATE, MVT::v16i8, MVT::v16i32, 7 },
2186 { ISD::TRUNCATE, MVT::v2i16, MVT::v2i32, 1 },
2187 { ISD::TRUNCATE, MVT::v8i16, MVT::v4i32, 3 },
2188 { ISD::TRUNCATE, MVT::v8i16, MVT::v8i32, 5 },
2189 { ISD::TRUNCATE, MVT::v16i16, MVT::v16i32,10 },
2190 { ISD::TRUNCATE, MVT::v16i8, MVT::v2i64, 4 }, // PAND+3*PACKUSWB
2191 { ISD::TRUNCATE, MVT::v8i16, MVT::v2i64, 2 }, // PSHUFD+PSHUFLW
2192 { ISD::TRUNCATE, MVT::v4i32, MVT::v2i64, 1 }, // PSHUFD
2193 };
2194
2195 // Attempt to map directly to (simple) MVT types to let us match custom entries.
2196 EVT SrcTy = TLI->getValueType(DL, Src);
2197 EVT DstTy = TLI->getValueType(DL, Dst);
2198
2199 // The function getSimpleVT only handles simple value types.
2200 if (SrcTy.isSimple() && DstTy.isSimple()) {
2201 MVT SimpleSrcTy = SrcTy.getSimpleVT();
2202 MVT SimpleDstTy = DstTy.getSimpleVT();
2203
2204 if (ST->useAVX512Regs()) {
2205 if (ST->hasBWI())
2206 if (const auto *Entry = ConvertCostTableLookup(
2207 AVX512BWConversionTbl, ISD, SimpleDstTy, SimpleSrcTy))
2208 return AdjustCost(Entry->Cost);
2209
2210 if (ST->hasDQI())
2211 if (const auto *Entry = ConvertCostTableLookup(
2212 AVX512DQConversionTbl, ISD, SimpleDstTy, SimpleSrcTy))
2213 return AdjustCost(Entry->Cost);
2214
2215 if (ST->hasAVX512())
2216 if (const auto *Entry = ConvertCostTableLookup(
2217 AVX512FConversionTbl, ISD, SimpleDstTy, SimpleSrcTy))
2218 return AdjustCost(Entry->Cost);
2219 }
2220
2221 if (ST->hasBWI())
2222 if (const auto *Entry = ConvertCostTableLookup(
2223 AVX512BWVLConversionTbl, ISD, SimpleDstTy, SimpleSrcTy))
2224 return AdjustCost(Entry->Cost);
2225
2226 if (ST->hasDQI())
2227 if (const auto *Entry = ConvertCostTableLookup(
2228 AVX512DQVLConversionTbl, ISD, SimpleDstTy, SimpleSrcTy))
2229 return AdjustCost(Entry->Cost);
2230
2231 if (ST->hasAVX512())
2232 if (const auto *Entry = ConvertCostTableLookup(AVX512VLConversionTbl, ISD,
2233 SimpleDstTy, SimpleSrcTy))
2234 return AdjustCost(Entry->Cost);
2235
2236 if (ST->hasAVX2()) {
2237 if (const auto *Entry = ConvertCostTableLookup(AVX2ConversionTbl, ISD,
2238 SimpleDstTy, SimpleSrcTy))
2239 return AdjustCost(Entry->Cost);
2240 }
2241
2242 if (ST->hasAVX()) {
2243 if (const auto *Entry = ConvertCostTableLookup(AVXConversionTbl, ISD,
2244 SimpleDstTy, SimpleSrcTy))
2245 return AdjustCost(Entry->Cost);
2246 }
2247
2248 if (ST->hasSSE41()) {
2249 if (const auto *Entry = ConvertCostTableLookup(SSE41ConversionTbl, ISD,
2250 SimpleDstTy, SimpleSrcTy))
2251 return AdjustCost(Entry->Cost);
2252 }
2253
2254 if (ST->hasSSE2()) {
2255 if (const auto *Entry = ConvertCostTableLookup(SSE2ConversionTbl, ISD,
2256 SimpleDstTy, SimpleSrcTy))
2257 return AdjustCost(Entry->Cost);
2258 }
2259 }
2260
2261 // Fall back to legalized types.
2262 std::pair<InstructionCost, MVT> LTSrc = TLI->getTypeLegalizationCost(DL, Src);
2263 std::pair<InstructionCost, MVT> LTDest =
2264 TLI->getTypeLegalizationCost(DL, Dst);
2265
2266 if (ST->useAVX512Regs()) {
2267 if (ST->hasBWI())
2268 if (const auto *Entry = ConvertCostTableLookup(
2269 AVX512BWConversionTbl, ISD, LTDest.second, LTSrc.second))
2270 return AdjustCost(std::max(LTSrc.first, LTDest.first) * Entry->Cost);
2271
2272 if (ST->hasDQI())
2273 if (const auto *Entry = ConvertCostTableLookup(
2274 AVX512DQConversionTbl, ISD, LTDest.second, LTSrc.second))
2275 return AdjustCost(std::max(LTSrc.first, LTDest.first) * Entry->Cost);
2276
2277 if (ST->hasAVX512())
2278 if (const auto *Entry = ConvertCostTableLookup(
2279 AVX512FConversionTbl, ISD, LTDest.second, LTSrc.second))
2280 return AdjustCost(std::max(LTSrc.first, LTDest.first) * Entry->Cost);
2281 }
2282
2283 if (ST->hasBWI())
2284 if (const auto *Entry = ConvertCostTableLookup(AVX512BWVLConversionTbl, ISD,
2285 LTDest.second, LTSrc.second))
2286 return AdjustCost(std::max(LTSrc.first, LTDest.first) * Entry->Cost);
2287
2288 if (ST->hasDQI())
2289 if (const auto *Entry = ConvertCostTableLookup(AVX512DQVLConversionTbl, ISD,
2290 LTDest.second, LTSrc.second))
2291 return AdjustCost(std::max(LTSrc.first, LTDest.first) * Entry->Cost);
2292
2293 if (ST->hasAVX512())
2294 if (const auto *Entry = ConvertCostTableLookup(AVX512VLConversionTbl, ISD,
2295 LTDest.second, LTSrc.second))
2296 return AdjustCost(std::max(LTSrc.first, LTDest.first) * Entry->Cost);
2297
2298 if (ST->hasAVX2())
2299 if (const auto *Entry = ConvertCostTableLookup(AVX2ConversionTbl, ISD,
2300 LTDest.second, LTSrc.second))
2301 return AdjustCost(std::max(LTSrc.first, LTDest.first) * Entry->Cost);
2302
2303 if (ST->hasAVX())
2304 if (const auto *Entry = ConvertCostTableLookup(AVXConversionTbl, ISD,
2305 LTDest.second, LTSrc.second))
2306 return AdjustCost(std::max(LTSrc.first, LTDest.first) * Entry->Cost);
2307
2308 if (ST->hasSSE41())
2309 if (const auto *Entry = ConvertCostTableLookup(SSE41ConversionTbl, ISD,
2310 LTDest.second, LTSrc.second))
2311 return AdjustCost(std::max(LTSrc.first, LTDest.first) * Entry->Cost);
2312
2313 if (ST->hasSSE2())
2314 if (const auto *Entry = ConvertCostTableLookup(SSE2ConversionTbl, ISD,
2315 LTDest.second, LTSrc.second))
2316 return AdjustCost(std::max(LTSrc.first, LTDest.first) * Entry->Cost);
2317
2318 // Fallback, for i8/i16 sitofp/uitofp cases we need to extend to i32 for
2319 // sitofp.
2320 if ((ISD == ISD::SINT_TO_FP || ISD == ISD::UINT_TO_FP) &&
2321 1 < Src->getScalarSizeInBits() && Src->getScalarSizeInBits() < 32) {
2322 Type *ExtSrc = Src->getWithNewBitWidth(32);
2323 unsigned ExtOpc =
2324 (ISD == ISD::SINT_TO_FP) ? Instruction::SExt : Instruction::ZExt;
2325
2326 // For scalar loads the extend would be free.
2327 InstructionCost ExtCost = 0;
2328 if (!(Src->isIntegerTy() && I && isa<LoadInst>(I->getOperand(0))))
2329 ExtCost = getCastInstrCost(ExtOpc, ExtSrc, Src, CCH, CostKind);
2330
2331 return ExtCost + getCastInstrCost(Instruction::SIToFP, Dst, ExtSrc,
2332 TTI::CastContextHint::None, CostKind);
2333 }
2334
2335 // Fallback for fptosi/fptoui i8/i16 cases we need to truncate from fptosi
2336 // i32.
2337 if ((ISD == ISD::FP_TO_SINT || ISD == ISD::FP_TO_UINT) &&
2338 1 < Dst->getScalarSizeInBits() && Dst->getScalarSizeInBits() < 32) {
2339 Type *TruncDst = Dst->getWithNewBitWidth(32);
2340 return getCastInstrCost(Instruction::FPToSI, TruncDst, Src, CCH, CostKind) +
2341 getCastInstrCost(Instruction::Trunc, Dst, TruncDst,
2342 TTI::CastContextHint::None, CostKind);
2343 }
2344
2345 return AdjustCost(
2346 BaseT::getCastInstrCost(Opcode, Dst, Src, CCH, CostKind, I));
2347}
2348
2349InstructionCost X86TTIImpl::getCmpSelInstrCost(unsigned Opcode, Type *ValTy,
2350 Type *CondTy,
2351 CmpInst::Predicate VecPred,
2352 TTI::TargetCostKind CostKind,
2353 const Instruction *I) {
2354 // TODO: Handle other cost kinds.
2355 if (CostKind != TTI::TCK_RecipThroughput)
2356 return BaseT::getCmpSelInstrCost(Opcode, ValTy, CondTy, VecPred, CostKind,
2357 I);
2358
2359 // Legalize the type.
2360 std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, ValTy);
2361
2362 MVT MTy = LT.second;
2363
2364 int ISD = TLI->InstructionOpcodeToISD(Opcode);
2365 assert(ISD && "Invalid opcode")((void)0);
2366
2367 unsigned ExtraCost = 0;
2368 if (I && (Opcode == Instruction::ICmp || Opcode == Instruction::FCmp)) {
2369 // Some vector comparison predicates cost extra instructions.
2370 if (MTy.isVector() &&
2371 !((ST->hasXOP() && (!ST->hasAVX2() || MTy.is128BitVector())) ||
2372 (ST->hasAVX512() && 32 <= MTy.getScalarSizeInBits()) ||
2373 ST->hasBWI())) {
2374 switch (cast<CmpInst>(I)->getPredicate()) {
2375 case CmpInst::Predicate::ICMP_NE:
2376 // xor(cmpeq(x,y),-1)
2377 ExtraCost = 1;
2378 break;
2379 case CmpInst::Predicate::ICMP_SGE:
2380 case CmpInst::Predicate::ICMP_SLE:
2381 // xor(cmpgt(x,y),-1)
2382 ExtraCost = 1;
2383 break;
2384 case CmpInst::Predicate::ICMP_ULT:
2385 case CmpInst::Predicate::ICMP_UGT:
2386 // cmpgt(xor(x,signbit),xor(y,signbit))
2387 // xor(cmpeq(pmaxu(x,y),x),-1)
2388 ExtraCost = 2;
2389 break;
2390 case CmpInst::Predicate::ICMP_ULE:
2391 case CmpInst::Predicate::ICMP_UGE:
2392 if ((ST->hasSSE41() && MTy.getScalarSizeInBits() == 32) ||
2393 (ST->hasSSE2() && MTy.getScalarSizeInBits() < 32)) {
2394 // cmpeq(psubus(x,y),0)
2395 // cmpeq(pminu(x,y),x)
2396 ExtraCost = 1;
2397 } else {
2398 // xor(cmpgt(xor(x,signbit),xor(y,signbit)),-1)
2399 ExtraCost = 3;
2400 }
2401 break;
2402 default:
2403 break;
2404 }
2405 }
2406 }
2407
2408 static const CostTblEntry SLMCostTbl[] = {
2409 // slm pcmpeq/pcmpgt throughput is 2
2410 { ISD::SETCC, MVT::v2i64, 2 },
2411 };
2412
2413 static const CostTblEntry AVX512BWCostTbl[] = {
2414 { ISD::SETCC, MVT::v32i16, 1 },
2415 { ISD::SETCC, MVT::v64i8, 1 },
2416
2417 { ISD::SELECT, MVT::v32i16, 1 },
2418 { ISD::SELECT, MVT::v64i8, 1 },
2419 };
2420
2421 static const CostTblEntry AVX512CostTbl[] = {
2422 { ISD::SETCC, MVT::v8i64, 1 },
2423 { ISD::SETCC, MVT::v16i32, 1 },
2424 { ISD::SETCC, MVT::v8f64, 1 },
2425 { ISD::SETCC, MVT::v16f32, 1 },
2426
2427 { ISD::SELECT, MVT::v8i64, 1 },
2428 { ISD::SELECT, MVT::v16i32, 1 },
2429 { ISD::SELECT, MVT::v8f64, 1 },
2430 { ISD::SELECT, MVT::v16f32, 1 },
2431
2432 { ISD::SETCC, MVT::v32i16, 2 }, // FIXME: should probably be 4
2433 { ISD::SETCC, MVT::v64i8, 2 }, // FIXME: should probably be 4
2434
2435 { ISD::SELECT, MVT::v32i16, 2 }, // FIXME: should be 3
2436 { ISD::SELECT, MVT::v64i8, 2 }, // FIXME: should be 3
2437 };
2438
2439 static const CostTblEntry AVX2CostTbl[] = {
2440 { ISD::SETCC, MVT::v4i64, 1 },
2441 { ISD::SETCC, MVT::v8i32, 1 },
2442 { ISD::SETCC, MVT::v16i16, 1 },
2443 { ISD::SETCC, MVT::v32i8, 1 },
2444
2445 { ISD::SELECT, MVT::v4i64, 1 }, // pblendvb
2446 { ISD::SELECT, MVT::v8i32, 1 }, // pblendvb
2447 { ISD::SELECT, MVT::v16i16, 1 }, // pblendvb
2448 { ISD::SELECT, MVT::v32i8, 1 }, // pblendvb
2449 };
2450
2451 static const CostTblEntry AVX1CostTbl[] = {
2452 { ISD::SETCC, MVT::v4f64, 1 },
2453 { ISD::SETCC, MVT::v8f32, 1 },
2454 // AVX1 does not support 8-wide integer compare.
2455 { ISD::SETCC, MVT::v4i64, 4 },
2456 { ISD::SETCC, MVT::v8i32, 4 },
2457 { ISD::SETCC, MVT::v16i16, 4 },
2458 { ISD::SETCC, MVT::v32i8, 4 },
2459
2460 { ISD::SELECT, MVT::v4f64, 1 }, // vblendvpd
2461 { ISD::SELECT, MVT::v8f32, 1 }, // vblendvps
2462 { ISD::SELECT, MVT::v4i64, 1 }, // vblendvpd
2463 { ISD::SELECT, MVT::v8i32, 1 }, // vblendvps
2464 { ISD::SELECT, MVT::v16i16, 3 }, // vandps + vandnps + vorps
2465 { ISD::SELECT, MVT::v32i8, 3 }, // vandps + vandnps + vorps
2466 };
2467
2468 static const CostTblEntry SSE42CostTbl[] = {
2469 { ISD::SETCC, MVT::v2f64, 1 },
2470 { ISD::SETCC, MVT::v4f32, 1 },
2471 { ISD::SETCC, MVT::v2i64, 1 },
2472 };
2473
2474 static const CostTblEntry SSE41CostTbl[] = {
2475 { ISD::SELECT, MVT::v2f64, 1 }, // blendvpd
2476 { ISD::SELECT, MVT::v4f32, 1 }, // blendvps
2477 { ISD::SELECT, MVT::v2i64, 1 }, // pblendvb
2478 { ISD::SELECT, MVT::v4i32, 1 }, // pblendvb
2479 { ISD::SELECT, MVT::v8i16, 1 }, // pblendvb
2480 { ISD::SELECT, MVT::v16i8, 1 }, // pblendvb
2481 };
2482
2483 static const CostTblEntry SSE2CostTbl[] = {
2484 { ISD::SETCC, MVT::v2f64, 2 },
2485 { ISD::SETCC, MVT::f64, 1 },
2486 { ISD::SETCC, MVT::v2i64, 8 },
2487 { ISD::SETCC, MVT::v4i32, 1 },
2488 { ISD::SETCC, MVT::v8i16, 1 },
2489 { ISD::SETCC, MVT::v16i8, 1 },
2490
2491 { ISD::SELECT, MVT::v2f64, 3 }, // andpd + andnpd + orpd
2492 { ISD::SELECT, MVT::v2i64, 3 }, // pand + pandn + por
2493 { ISD::SELECT, MVT::v4i32, 3 }, // pand + pandn + por
2494 { ISD::SELECT, MVT::v8i16, 3 }, // pand + pandn + por
2495 { ISD::SELECT, MVT::v16i8, 3 }, // pand + pandn + por
2496 };
2497
2498 static const CostTblEntry SSE1CostTbl[] = {
2499 { ISD::SETCC, MVT::v4f32, 2 },
2500 { ISD::SETCC, MVT::f32, 1 },
2501
2502 { ISD::SELECT, MVT::v4f32, 3 }, // andps + andnps + orps
2503 };
2504
2505 if (ST->isSLM())
2506 if (const auto *Entry = CostTableLookup(SLMCostTbl, ISD, MTy))
2507 return LT.first * (ExtraCost + Entry->Cost);
2508
2509 if (ST->hasBWI())
2510 if (const auto *Entry = CostTableLookup(AVX512BWCostTbl, ISD, MTy))
2511 return LT.first * (ExtraCost + Entry->Cost);
2512
2513 if (ST->hasAVX512())
2514 if (const auto *Entry = CostTableLookup(AVX512CostTbl, ISD, MTy))
2515 return LT.first * (ExtraCost + Entry->Cost);
2516
2517 if (ST->hasAVX2())
2518 if (const auto *Entry = CostTableLookup(AVX2CostTbl, ISD, MTy))
2519 return LT.first * (ExtraCost + Entry->Cost);
2520
2521 if (ST->hasAVX())
2522 if (const auto *Entry = CostTableLookup(AVX1CostTbl, ISD, MTy))
2523 return LT.first * (ExtraCost + Entry->Cost);
2524
2525 if (ST->hasSSE42())
2526 if (const auto *Entry = CostTableLookup(SSE42CostTbl, ISD, MTy))
2527 return LT.first * (ExtraCost + Entry->Cost);
2528
2529 if (ST->hasSSE41())
2530 if (const auto *Entry = CostTableLookup(SSE41CostTbl, ISD, MTy))
2531 return LT.first * (ExtraCost + Entry->Cost);
2532
2533 if (ST->hasSSE2())
2534 if (const auto *Entry = CostTableLookup(SSE2CostTbl, ISD, MTy))
2535 return LT.first * (ExtraCost + Entry->Cost);
2536
2537 if (ST->hasSSE1())
2538 if (const auto *Entry = CostTableLookup(SSE1CostTbl, ISD, MTy))
2539 return LT.first * (ExtraCost + Entry->Cost);
2540
2541 return BaseT::getCmpSelInstrCost(Opcode, ValTy, CondTy, VecPred, CostKind, I);
2542}
2543
2544unsigned X86TTIImpl::getAtomicMemIntrinsicMaxElementSize() const { return 16; }
2545
2546InstructionCost
2547X86TTIImpl::getTypeBasedIntrinsicInstrCost(const IntrinsicCostAttributes &ICA,
2548 TTI::TargetCostKind CostKind) {
2549
2550 // Costs should match the codegen from:
2551 // BITREVERSE: llvm\test\CodeGen\X86\vector-bitreverse.ll
2552 // BSWAP: llvm\test\CodeGen\X86\bswap-vector.ll
2553 // CTLZ: llvm\test\CodeGen\X86\vector-lzcnt-*.ll
2554 // CTPOP: llvm\test\CodeGen\X86\vector-popcnt-*.ll
2555 // CTTZ: llvm\test\CodeGen\X86\vector-tzcnt-*.ll
2556
2557 // TODO: Overflow intrinsics (*ADDO, *SUBO, *MULO) with vector types are not
2558 // specialized in these tables yet.
2559 static const CostTblEntry AVX512CDCostTbl[] = {
2560 { ISD::CTLZ, MVT::v8i64, 1 },
2561 { ISD::CTLZ, MVT::v16i32, 1 },
2562 { ISD::CTLZ, MVT::v32i16, 8 },
2563 { ISD::CTLZ, MVT::v64i8, 20 },
2564 { ISD::CTLZ, MVT::v4i64, 1 },
2565 { ISD::CTLZ, MVT::v8i32, 1 },
2566 { ISD::CTLZ, MVT::v16i16, 4 },
2567 { ISD::CTLZ, MVT::v32i8, 10 },
2568 { ISD::CTLZ, MVT::v2i64, 1 },
2569 { ISD::CTLZ, MVT::v4i32, 1 },
2570 { ISD::CTLZ, MVT::v8i16, 4 },
2571 { ISD::CTLZ, MVT::v16i8, 4 },
2572 };
2573 static const CostTblEntry AVX512BWCostTbl[] = {
2574 { ISD::ABS, MVT::v32i16, 1 },
2575 { ISD::ABS, MVT::v64i8, 1 },
2576 { ISD::BITREVERSE, MVT::v8i64, 5 },
2577 { ISD::BITREVERSE, MVT::v16i32, 5 },
2578 { ISD::BITREVERSE, MVT::v32i16, 5 },
2579 { ISD::BITREVERSE, MVT::v64i8, 5 },
2580 { ISD::BSWAP, MVT::v8i64, 1 },
2581 { ISD::BSWAP, MVT::v16i32, 1 },
2582 { ISD::BSWAP, MVT::v32i16, 1 },
2583 { ISD::CTLZ, MVT::v8i64, 23 },
2584 { ISD::CTLZ, MVT::v16i32, 22 },
2585 { ISD::CTLZ, MVT::v32i16, 18 },
2586 { ISD::CTLZ, MVT::v64i8, 17 },
2587 { ISD::CTPOP, MVT::v8i64, 7 },
2588 { ISD::CTPOP, MVT::v16i32, 11 },
2589 { ISD::CTPOP, MVT::v32i16, 9 },
2590 { ISD::CTPOP, MVT::v64i8, 6 },
2591 { ISD::CTTZ, MVT::v8i64, 10 },
2592 { ISD::CTTZ, MVT::v16i32, 14 },
2593 { ISD::CTTZ, MVT::v32i16, 12 },
2594 { ISD::CTTZ, MVT::v64i8, 9 },
2595 { ISD::SADDSAT, MVT::v32i16, 1 },
2596 { ISD::SADDSAT, MVT::v64i8, 1 },
2597 { ISD::SMAX, MVT::v32i16, 1 },
2598 { ISD::SMAX, MVT::v64i8, 1 },
2599 { ISD::SMIN, MVT::v32i16, 1 },
2600 { ISD::SMIN, MVT::v64i8, 1 },
2601 { ISD::SSUBSAT, MVT::v32i16, 1 },
2602 { ISD::SSUBSAT, MVT::v64i8, 1 },
2603 { ISD::UADDSAT, MVT::v32i16, 1 },
2604 { ISD::UADDSAT, MVT::v64i8, 1 },
2605 { ISD::UMAX, MVT::v32i16, 1 },
2606 { ISD::UMAX, MVT::v64i8, 1 },
2607 { ISD::UMIN, MVT::v32i16, 1 },
2608 { ISD::UMIN, MVT::v64i8, 1 },
2609 { ISD::USUBSAT, MVT::v32i16, 1 },
2610 { ISD::USUBSAT, MVT::v64i8, 1 },
2611 };
2612 static const CostTblEntry AVX512CostTbl[] = {
2613 { ISD::ABS, MVT::v8i64, 1 },
2614 { ISD::ABS, MVT::v16i32, 1 },
2615 { ISD::ABS, MVT::v32i16, 2 }, // FIXME: include split
2616 { ISD::ABS, MVT::v64i8, 2 }, // FIXME: include split
2617 { ISD::ABS, MVT::v4i64, 1 },
2618 { ISD::ABS, MVT::v2i64, 1 },
2619 { ISD::BITREVERSE, MVT::v8i64, 36 },
2620 { ISD::BITREVERSE, MVT::v16i32, 24 },
2621 { ISD::BITREVERSE, MVT::v32i16, 10 },
2622 { ISD::BITREVERSE, MVT::v64i8, 10 },
2623 { ISD::BSWAP, MVT::v8i64, 4 },
2624 { ISD::BSWAP, MVT::v16i32, 4 },
2625 { ISD::BSWAP, MVT::v32i16, 4 },
2626 { ISD::CTLZ, MVT::v8i64, 29 },
2627 { ISD::CTLZ, MVT::v16i32, 35 },
2628 { ISD::CTLZ, MVT::v32i16, 28 },
2629 { ISD::CTLZ, MVT::v64i8, 18 },
2630 { ISD::CTPOP, MVT::v8i64, 16 },
2631 { ISD::CTPOP, MVT::v16i32, 24 },
2632 { ISD::CTPOP, MVT::v32i16, 18 },
2633 { ISD::CTPOP, MVT::v64i8, 12 },
2634 { ISD::CTTZ, MVT::v8i64, 20 },
2635 { ISD::CTTZ, MVT::v16i32, 28 },
2636 { ISD::CTTZ, MVT::v32i16, 24 },
2637 { ISD::CTTZ, MVT::v64i8, 18 },
2638 { ISD::SMAX, MVT::v8i64, 1 },
2639 { ISD::SMAX, MVT::v16i32, 1 },
2640 { ISD::SMAX, MVT::v32i16, 2 }, // FIXME: include split
2641 { ISD::SMAX, MVT::v64i8, 2 }, // FIXME: include split
2642 { ISD::SMAX, MVT::v4i64, 1 },
2643 { ISD::SMAX, MVT::v2i64, 1 },
2644 { ISD::SMIN, MVT::v8i64, 1 },
2645 { ISD::SMIN, MVT::v16i32, 1 },
2646 { ISD::SMIN, MVT::v32i16, 2 }, // FIXME: include split
2647 { ISD::SMIN, MVT::v64i8, 2 }, // FIXME: include split
2648 { ISD::SMIN, MVT::v4i64, 1 },
2649 { ISD::SMIN, MVT::v2i64, 1 },
2650 { ISD::UMAX, MVT::v8i64, 1 },
2651 { ISD::UMAX, MVT::v16i32, 1 },
2652 { ISD::UMAX, MVT::v32i16, 2 }, // FIXME: include split
2653 { ISD::UMAX, MVT::v64i8, 2 }, // FIXME: include split
2654 { ISD::UMAX, MVT::v4i64, 1 },
2655 { ISD::UMAX, MVT::v2i64, 1 },
2656 { ISD::UMIN, MVT::v8i64, 1 },
2657 { ISD::UMIN, MVT::v16i32, 1 },
2658 { ISD::UMIN, MVT::v32i16, 2 }, // FIXME: include split
2659 { ISD::UMIN, MVT::v64i8, 2 }, // FIXME: include split
2660 { ISD::UMIN, MVT::v4i64, 1 },
2661 { ISD::UMIN, MVT::v2i64, 1 },
2662 { ISD::USUBSAT, MVT::v16i32, 2 }, // pmaxud + psubd
2663 { ISD::USUBSAT, MVT::v2i64, 2 }, // pmaxuq + psubq
2664 { ISD::USUBSAT, MVT::v4i64, 2 }, // pmaxuq + psubq
2665 { ISD::USUBSAT, MVT::v8i64, 2 }, // pmaxuq + psubq
2666 { ISD::UADDSAT, MVT::v16i32, 3 }, // not + pminud + paddd
2667 { ISD::UADDSAT, MVT::v2i64, 3 }, // not + pminuq + paddq
2668 { ISD::UADDSAT, MVT::v4i64, 3 }, // not + pminuq + paddq
2669 { ISD::UADDSAT, MVT::v8i64, 3 }, // not + pminuq + paddq
2670 { ISD::SADDSAT, MVT::v32i16, 2 }, // FIXME: include split
2671 { ISD::SADDSAT, MVT::v64i8, 2 }, // FIXME: include split
2672 { ISD::SSUBSAT, MVT::v32i16, 2 }, // FIXME: include split
2673 { ISD::SSUBSAT, MVT::v64i8, 2 }, // FIXME: include split
2674 { ISD::UADDSAT, MVT::v32i16, 2 }, // FIXME: include split
2675 { ISD::UADDSAT, MVT::v64i8, 2 }, // FIXME: include split
2676 { ISD::USUBSAT, MVT::v32i16, 2 }, // FIXME: include split
2677 { ISD::USUBSAT, MVT::v64i8, 2 }, // FIXME: include split
2678 { ISD::FMAXNUM, MVT::f32, 2 },
2679 { ISD::FMAXNUM, MVT::v4f32, 2 },
2680 { ISD::FMAXNUM, MVT::v8f32, 2 },
2681 { ISD::FMAXNUM, MVT::v16f32, 2 },
2682 { ISD::FMAXNUM, MVT::f64, 2 },
2683 { ISD::FMAXNUM, MVT::v2f64, 2 },
2684 { ISD::FMAXNUM, MVT::v4f64, 2 },
2685 { ISD::FMAXNUM, MVT::v8f64, 2 },
2686 };
2687 static const CostTblEntry XOPCostTbl[] = {
2688 { ISD::BITREVERSE, MVT::v4i64, 4 },
2689 { ISD::BITREVERSE, MVT::v8i32, 4 },
2690 { ISD::BITREVERSE, MVT::v16i16, 4 },
2691 { ISD::BITREVERSE, MVT::v32i8, 4 },
2692 { ISD::BITREVERSE, MVT::v2i64, 1 },
2693 { ISD::BITREVERSE, MVT::v4i32, 1 },
2694 { ISD::BITREVERSE, MVT::v8i16, 1 },
2695 { ISD::BITREVERSE, MVT::v16i8, 1 },
2696 { ISD::BITREVERSE, MVT::i64, 3 },
2697 { ISD::BITREVERSE, MVT::i32, 3 },
2698 { ISD::BITREVERSE, MVT::i16, 3 },
2699 { ISD::BITREVERSE, MVT::i8, 3 }
2700 };
2701 static const CostTblEntry AVX2CostTbl[] = {
2702 { ISD::ABS, MVT::v4i64, 2 }, // VBLENDVPD(X,VPSUBQ(0,X),X)
2703 { ISD::ABS, MVT::v8i32, 1 },
2704 { ISD::ABS, MVT::v16i16, 1 },
2705 { ISD::ABS, MVT::v32i8, 1 },
2706 { ISD::BITREVERSE, MVT::v4i64, 5 },
2707 { ISD::BITREVERSE, MVT::v8i32, 5 },
2708 { ISD::BITREVERSE, MVT::v16i16, 5 },
2709 { ISD::BITREVERSE, MVT::v32i8, 5 },
2710 { ISD::BSWAP, MVT::v4i64, 1 },
2711 { ISD::BSWAP, MVT::v8i32, 1 },
2712 { ISD::BSWAP, MVT::v16i16, 1 },
2713 { ISD::CTLZ, MVT::v4i64, 23 },
2714 { ISD::CTLZ, MVT::v8i32, 18 },
2715 { ISD::CTLZ, MVT::v16i16, 14 },
2716 { ISD::CTLZ, MVT::v32i8, 9 },
2717 { ISD::CTPOP, MVT::v4i64, 7 },
2718 { ISD::CTPOP, MVT::v8i32, 11 },
2719 { ISD::CTPOP, MVT::v16i16, 9 },
2720 { ISD::CTPOP, MVT::v32i8, 6 },
2721 { ISD::CTTZ, MVT::v4i64, 10 },
2722 { ISD::CTTZ, MVT::v8i32, 14 },
2723 { ISD::CTTZ, MVT::v16i16, 12 },
2724 { ISD::CTTZ, MVT::v32i8, 9 },
2725 { ISD::SADDSAT, MVT::v16i16, 1 },
2726 { ISD::SADDSAT, MVT::v32i8, 1 },
2727 { ISD::SMAX, MVT::v8i32, 1 },
2728 { ISD::SMAX, MVT::v16i16, 1 },
2729 { ISD::SMAX, MVT::v32i8, 1 },
2730 { ISD::SMIN, MVT::v8i32, 1 },
2731 { ISD::SMIN, MVT::v16i16, 1 },
2732 { ISD::SMIN, MVT::v32i8, 1 },
2733 { ISD::SSUBSAT, MVT::v16i16, 1 },
2734 { ISD::SSUBSAT, MVT::v32i8, 1 },
2735 { ISD::UADDSAT, MVT::v16i16, 1 },
2736 { ISD::UADDSAT, MVT::v32i8, 1 },
2737 { ISD::UADDSAT, MVT::v8i32, 3 }, // not + pminud + paddd
2738 { ISD::UMAX, MVT::v8i32, 1 },
2739 { ISD::UMAX, MVT::v16i16, 1 },
2740 { ISD::UMAX, MVT::v32i8, 1 },
2741 { ISD::UMIN, MVT::v8i32, 1 },
2742 { ISD::UMIN, MVT::v16i16, 1 },
2743 { ISD::UMIN, MVT::v32i8, 1 },
2744 { ISD::USUBSAT, MVT::v16i16, 1 },
2745 { ISD::USUBSAT, MVT::v32i8, 1 },
2746 { ISD::USUBSAT, MVT::v8i32, 2 }, // pmaxud + psubd
2747 { ISD::FMAXNUM, MVT::v8f32, 3 }, // MAXPS + CMPUNORDPS + BLENDVPS
2748 { ISD::FMAXNUM, MVT::v4f64, 3 }, // MAXPD + CMPUNORDPD + BLENDVPD
2749 { ISD::FSQRT, MVT::f32, 7 }, // Haswell from http://www.agner.org/
2750 { ISD::FSQRT, MVT::v4f32, 7 }, // Haswell from http://www.agner.org/
2751 { ISD::FSQRT, MVT::v8f32, 14 }, // Haswell from http://www.agner.org/
2752 { ISD::FSQRT, MVT::f64, 14 }, // Haswell from http://www.agner.org/
2753 { ISD::FSQRT, MVT::v2f64, 14 }, // Haswell from http://www.agner.org/
2754 { ISD::FSQRT, MVT::v4f64, 28 }, // Haswell from http://www.agner.org/
2755 };
2756 static const CostTblEntry AVX1CostTbl[] = {
2757 { ISD::ABS, MVT::v4i64, 5 }, // VBLENDVPD(X,VPSUBQ(0,X),X)
2758 { ISD::ABS, MVT::v8i32, 3 },
2759 { ISD::ABS, MVT::v16i16, 3 },
2760 { ISD::ABS, MVT::v32i8, 3 },
2761 { ISD::BITREVERSE, MVT::v4i64, 12 }, // 2 x 128-bit Op + extract/insert
2762 { ISD::BITREVERSE, MVT::v8i32, 12 }, // 2 x 128-bit Op + extract/insert
2763 { ISD::BITREVERSE, MVT::v16i16, 12 }, // 2 x 128-bit Op + extract/insert
2764 { ISD::BITREVERSE, MVT::v32i8, 12 }, // 2 x 128-bit Op + extract/insert
2765 { ISD::BSWAP, MVT::v4i64, 4 },
2766 { ISD::BSWAP, MVT::v8i32, 4 },
2767 { ISD::BSWAP, MVT::v16i16, 4 },
2768 { ISD::CTLZ, MVT::v4i64, 48 }, // 2 x 128-bit Op + extract/insert
2769 { ISD::CTLZ, MVT::v8i32, 38 }, // 2 x 128-bit Op + extract/insert
2770 { ISD::CTLZ, MVT::v16i16, 30 }, // 2 x 128-bit Op + extract/insert
2771 { ISD::CTLZ, MVT::v32i8, 20 }, // 2 x 128-bit Op + extract/insert
2772 { ISD::CTPOP, MVT::v4i64, 16 }, // 2 x 128-bit Op + extract/insert
2773 { ISD::CTPOP, MVT::v8i32, 24 }, // 2 x 128-bit Op + extract/insert
2774 { ISD::CTPOP, MVT::v16i16, 20 }, // 2 x 128-bit Op + extract/insert
2775 { ISD::CTPOP, MVT::v32i8, 14 }, // 2 x 128-bit Op + extract/insert
2776 { ISD::CTTZ, MVT::v4i64, 22 }, // 2 x 128-bit Op + extract/insert
2777 { ISD::CTTZ, MVT::v8i32, 30 }, // 2 x 128-bit Op + extract/insert
2778 { ISD::CTTZ, MVT::v16i16, 26 }, // 2 x 128-bit Op + extract/insert
2779 { ISD::CTTZ, MVT::v32i8, 20 }, // 2 x 128-bit Op + extract/insert
2780 { ISD::SADDSAT, MVT::v16i16, 4 }, // 2 x 128-bit Op + extract/insert
2781 { ISD::SADDSAT, MVT::v32i8, 4 }, // 2 x 128-bit Op + extract/insert
2782 { ISD::SMAX, MVT::v8i32, 4 }, // 2 x 128-bit Op + extract/insert
2783 { ISD::SMAX, MVT::v16i16, 4 }, // 2 x 128-bit Op + extract/insert
2784 { ISD::SMAX, MVT::v32i8, 4 }, // 2 x 128-bit Op + extract/insert
2785 { ISD::SMIN, MVT::v8i32, 4 }, // 2 x 128-bit Op + extract/insert
2786 { ISD::SMIN, MVT::v16i16, 4 }, // 2 x 128-bit Op + extract/insert
2787 { ISD::SMIN, MVT::v32i8, 4 }, // 2 x 128-bit Op + extract/insert
2788 { ISD::SSUBSAT, MVT::v16i16, 4 }, // 2 x 128-bit Op + extract/insert
2789 { ISD::SSUBSAT, MVT::v32i8, 4 }, // 2 x 128-bit Op + extract/insert
2790 { ISD::UADDSAT, MVT::v16i16, 4 }, // 2 x 128-bit Op + extract/insert
2791 { ISD::UADDSAT, MVT::v32i8, 4 }, // 2 x 128-bit Op + extract/insert
2792 { ISD::UADDSAT, MVT::v8i32, 8 }, // 2 x 128-bit Op + extract/insert
2793 { ISD::UMAX, MVT::v8i32, 4 }, // 2 x 128-bit Op + extract/insert
2794 { ISD::UMAX, MVT::v16i16, 4 }, // 2 x 128-bit Op + extract/insert
2795 { ISD::UMAX, MVT::v32i8, 4 }, // 2 x 128-bit Op + extract/insert
2796 { ISD::UMIN, MVT::v8i32, 4 }, // 2 x 128-bit Op + extract/insert
2797 { ISD::UMIN, MVT::v16i16, 4 }, // 2 x 128-bit Op + extract/insert
2798 { ISD::UMIN, MVT::v32i8, 4 }, // 2 x 128-bit Op + extract/insert
2799 { ISD::USUBSAT, MVT::v16i16, 4 }, // 2 x 128-bit Op + extract/insert
2800 { ISD::USUBSAT, MVT::v32i8, 4 }, // 2 x 128-bit Op + extract/insert
2801 { ISD::USUBSAT, MVT::v8i32, 6 }, // 2 x 128-bit Op + extract/insert
2802 { ISD::FMAXNUM, MVT::f32, 3 }, // MAXSS + CMPUNORDSS + BLENDVPS
2803 { ISD::FMAXNUM, MVT::v4f32, 3 }, // MAXPS + CMPUNORDPS + BLENDVPS
2804 { ISD::FMAXNUM, MVT::v8f32, 5 }, // MAXPS + CMPUNORDPS + BLENDVPS + ?
2805 { ISD::FMAXNUM, MVT::f64, 3 }, // MAXSD + CMPUNORDSD + BLENDVPD
2806 { ISD::FMAXNUM, MVT::v2f64, 3 }, // MAXPD + CMPUNORDPD + BLENDVPD
2807 { ISD::FMAXNUM, MVT::v4f64, 5 }, // MAXPD + CMPUNORDPD + BLENDVPD + ?
2808 { ISD::FSQRT, MVT::f32, 14 }, // SNB from http://www.agner.org/
2809 { ISD::FSQRT, MVT::v4f32, 14 }, // SNB from http://www.agner.org/
2810 { ISD::FSQRT, MVT::v8f32, 28 }, // SNB from http://www.agner.org/
2811 { ISD::FSQRT, MVT::f64, 21 }, // SNB from http://www.agner.org/
2812 { ISD::FSQRT, MVT::v2f64, 21 }, // SNB from http://www.agner.org/
2813 { ISD::FSQRT, MVT::v4f64, 43 }, // SNB from http://www.agner.org/
2814 };
2815 static const CostTblEntry GLMCostTbl[] = {
2816 { ISD::FSQRT, MVT::f32, 19 }, // sqrtss
2817 { ISD::FSQRT, MVT::v4f32, 37 }, // sqrtps
2818 { ISD::FSQRT, MVT::f64, 34 }, // sqrtsd
2819 { ISD::FSQRT, MVT::v2f64, 67 }, // sqrtpd
2820 };
2821 static const CostTblEntry SLMCostTbl[] = {
2822 { ISD::FSQRT, MVT::f32, 20 }, // sqrtss
2823 { ISD::FSQRT, MVT::v4f32, 40 }, // sqrtps
2824 { ISD::FSQRT, MVT::f64, 35 }, // sqrtsd
2825 { ISD::FSQRT, MVT::v2f64, 70 }, // sqrtpd
2826 };
2827 static const CostTblEntry SSE42CostTbl[] = {
2828 { ISD::USUBSAT, MVT::v4i32, 2 }, // pmaxud + psubd
2829 { ISD::UADDSAT, MVT::v4i32, 3 }, // not + pminud + paddd
2830 { ISD::FSQRT, MVT::f32, 18 }, // Nehalem from http://www.agner.org/
2831 { ISD::FSQRT, MVT::v4f32, 18 }, // Nehalem from http://www.agner.org/
2832 };
2833 static const CostTblEntry SSE41CostTbl[] = {
2834 { ISD::ABS, MVT::v2i64, 2 }, // BLENDVPD(X,PSUBQ(0,X),X)
2835 { ISD::SMAX, MVT::v4i32, 1 },
2836 { ISD::SMAX, MVT::v16i8, 1 },
2837 { ISD::SMIN, MVT::v4i32, 1 },
2838 { ISD::SMIN, MVT::v16i8, 1 },
2839 { ISD::UMAX, MVT::v4i32, 1 },
2840 { ISD::UMAX, MVT::v8i16, 1 },
2841 { ISD::UMIN, MVT::v4i32, 1 },
2842 { ISD::UMIN, MVT::v8i16, 1 },
2843 };
2844 static const CostTblEntry SSSE3CostTbl[] = {
2845 { ISD::ABS, MVT::v4i32, 1 },
2846 { ISD::ABS, MVT::v8i16, 1 },
2847 { ISD::ABS, MVT::v16i8, 1 },
2848 { ISD::BITREVERSE, MVT::v2i64, 5 },
2849 { ISD::BITREVERSE, MVT::v4i32, 5 },
2850 { ISD::BITREVERSE, MVT::v8i16, 5 },
2851 { ISD::BITREVERSE, MVT::v16i8, 5 },
2852 { ISD::BSWAP, MVT::v2i64, 1 },
2853 { ISD::BSWAP, MVT::v4i32, 1 },
2854 { ISD::BSWAP, MVT::v8i16, 1 },
2855 { ISD::CTLZ, MVT::v2i64, 23 },
2856 { ISD::CTLZ, MVT::v4i32, 18 },
2857 { ISD::CTLZ, MVT::v8i16, 14 },
2858 { ISD::CTLZ, MVT::v16i8, 9 },
2859 { ISD::CTPOP, MVT::v2i64, 7 },
2860 { ISD::CTPOP, MVT::v4i32, 11 },
2861 { ISD::CTPOP, MVT::v8i16, 9 },
2862 { ISD::CTPOP, MVT::v16i8, 6 },
2863 { ISD::CTTZ, MVT::v2i64, 10 },
2864 { ISD::CTTZ, MVT::v4i32, 14 },
2865 { ISD::CTTZ, MVT::v8i16, 12 },
2866 { ISD::CTTZ, MVT::v16i8, 9 }
2867 };
2868 static const CostTblEntry SSE2CostTbl[] = {
2869 { ISD::ABS, MVT::v2i64, 4 },
2870 { ISD::ABS, MVT::v4i32, 3 },
2871 { ISD::ABS, MVT::v8i16, 2 },
2872 { ISD::ABS, MVT::v16i8, 2 },
2873 { ISD::BITREVERSE, MVT::v2i64, 29 },
2874 { ISD::BITREVERSE, MVT::v4i32, 27 },
2875 { ISD::BITREVERSE, MVT::v8i16, 27 },
2876 { ISD::BITREVERSE, MVT::v16i8, 20 },
2877 { ISD::BSWAP, MVT::v2i64, 7 },
2878 { ISD::BSWAP, MVT::v4i32, 7 },
2879 { ISD::BSWAP, MVT::v8i16, 7 },
2880 { ISD::CTLZ, MVT::v2i64, 25 },
2881 { ISD::CTLZ, MVT::v4i32, 26 },
2882 { ISD::CTLZ, MVT::v8i16, 20 },
2883 { ISD::CTLZ, MVT::v16i8, 17 },
2884 { ISD::CTPOP, MVT::v2i64, 12 },
2885 { ISD::CTPOP, MVT::v4i32, 15 },
2886 { ISD::CTPOP, MVT::v8i16, 13 },
2887 { ISD::CTPOP, MVT::v16i8, 10 },
2888 { ISD::CTTZ, MVT::v2i64, 14 },
2889 { ISD::CTTZ, MVT::v4i32, 18 },
2890 { ISD::CTTZ, MVT::v8i16, 16 },
2891 { ISD::CTTZ, MVT::v16i8, 13 },
2892 { ISD::SADDSAT, MVT::v8i16, 1 },
2893 { ISD::SADDSAT, MVT::v16i8, 1 },
2894 { ISD::SMAX, MVT::v8i16, 1 },
2895 { ISD::SMIN, MVT::v8i16, 1 },
2896 { ISD::SSUBSAT, MVT::v8i16, 1 },
2897 { ISD::SSUBSAT, MVT::v16i8, 1 },
2898 { ISD::UADDSAT, MVT::v8i16, 1 },
2899 { ISD::UADDSAT, MVT::v16i8, 1 },
2900 { ISD::UMAX, MVT::v8i16, 2 },
2901 { ISD::UMAX, MVT::v16i8, 1 },
2902 { ISD::UMIN, MVT::v8i16, 2 },
2903 { ISD::UMIN, MVT::v16i8, 1 },
2904 { ISD::USUBSAT, MVT::v8i16, 1 },
2905 { ISD::USUBSAT, MVT::v16i8, 1 },
2906 { ISD::FMAXNUM, MVT::f64, 4 },
2907 { ISD::FMAXNUM, MVT::v2f64, 4 },
2908 { ISD::FSQRT, MVT::f64, 32 }, // Nehalem from http://www.agner.org/
2909 { ISD::FSQRT, MVT::v2f64, 32 }, // Nehalem from http://www.agner.org/
2910 };
2911 static const CostTblEntry SSE1CostTbl[] = {
2912 { ISD::FMAXNUM, MVT::f32, 4 },
2913 { ISD::FMAXNUM, MVT::v4f32, 4 },
2914 { ISD::FSQRT, MVT::f32, 28 }, // Pentium III from http://www.agner.org/
2915 { ISD::FSQRT, MVT::v4f32, 56 }, // Pentium III from http://www.agner.org/
2916 };
2917 static const CostTblEntry BMI64CostTbl[] = { // 64-bit targets
2918 { ISD::CTTZ, MVT::i64, 1 },
2919 };
2920 static const CostTblEntry BMI32CostTbl[] = { // 32 or 64-bit targets
2921 { ISD::CTTZ, MVT::i32, 1 },
2922 { ISD::CTTZ, MVT::i16, 1 },
2923 { ISD::CTTZ, MVT::i8, 1 },
2924 };
2925 static const CostTblEntry LZCNT64CostTbl[] = { // 64-bit targets
2926 { ISD::CTLZ, MVT::i64, 1 },
2927 };
2928 static const CostTblEntry LZCNT32CostTbl[] = { // 32 or 64-bit targets
2929 { ISD::CTLZ, MVT::i32, 1 },
2930 { ISD::CTLZ, MVT::i16, 1 },
2931 { ISD::CTLZ, MVT::i8, 1 },
2932 };
2933 static const CostTblEntry POPCNT64CostTbl[] = { // 64-bit targets
2934 { ISD::CTPOP, MVT::i64, 1 },
2935 };
2936 static const CostTblEntry POPCNT32CostTbl[] = { // 32 or 64-bit targets
2937 { ISD::CTPOP, MVT::i32, 1 },
2938 { ISD::CTPOP, MVT::i16, 1 },
2939 { ISD::CTPOP, MVT::i8, 1 },
2940 };
2941 static const CostTblEntry X64CostTbl[] = { // 64-bit targets
2942 { ISD::ABS, MVT::i64, 2 }, // SUB+CMOV
2943 { ISD::BITREVERSE, MVT::i64, 14 },
2944 { ISD::BSWAP, MVT::i64, 1 },
2945 { ISD::CTLZ, MVT::i64, 4 }, // BSR+XOR or BSR+XOR+CMOV
2946 { ISD::CTTZ, MVT::i64, 3 }, // TEST+BSF+CMOV/BRANCH
2947 { ISD::CTPOP, MVT::i64, 10 },
2948 { ISD::SADDO, MVT::i64, 1 },
2949 { ISD::UADDO, MVT::i64, 1 },
2950 { ISD::UMULO, MVT::i64, 2 }, // mulq + seto
2951 };
2952 static const CostTblEntry X86CostTbl[] = { // 32 or 64-bit targets
2953 { ISD::ABS, MVT::i32, 2 }, // SUB+CMOV
2954 { ISD::ABS, MVT::i16, 2 }, // SUB+CMOV
2955 { ISD::BITREVERSE, MVT::i32, 14 },
2956 { ISD::BITREVERSE, MVT::i16, 14 },
2957 { ISD::BITREVERSE, MVT::i8, 11 },
2958 { ISD::BSWAP, MVT::i32, 1 },
2959 { ISD::BSWAP, MVT::i16, 1 }, // ROL
2960 { ISD::CTLZ, MVT::i32, 4 }, // BSR+XOR or BSR+XOR+CMOV
2961 { ISD::CTLZ, MVT::i16, 4 }, // BSR+XOR or BSR+XOR+CMOV
2962 { ISD::CTLZ, MVT::i8, 4 }, // BSR+XOR or BSR+XOR+CMOV
2963 { ISD::CTTZ, MVT::i32, 3 }, // TEST+BSF+CMOV/BRANCH
2964 { ISD::CTTZ, MVT::i16, 3 }, // TEST+BSF+CMOV/BRANCH
2965 { ISD::CTTZ, MVT::i8, 3 }, // TEST+BSF+CMOV/BRANCH
2966 { ISD::CTPOP, MVT::i32, 8 },
2967 { ISD::CTPOP, MVT::i16, 9 },
2968 { ISD::CTPOP, MVT::i8, 7 },
2969 { ISD::SADDO, MVT::i32, 1 },
2970 { ISD::SADDO, MVT::i16, 1 },
2971 { ISD::SADDO, MVT::i8, 1 },
2972 { ISD::UADDO, MVT::i32, 1 },
2973 { ISD::UADDO, MVT::i16, 1 },
2974 { ISD::UADDO, MVT::i8, 1 },
2975 { ISD::UMULO, MVT::i32, 2 }, // mul + seto
2976 { ISD::UMULO, MVT::i16, 2 },
2977 { ISD::UMULO, MVT::i8, 2 },
2978 };
2979
2980 Type *RetTy = ICA.getReturnType();
2981 Type *OpTy = RetTy;
2982 Intrinsic::ID IID = ICA.getID();
2983 unsigned ISD = ISD::DELETED_NODE;
2984 switch (IID) {
2985 default:
2986 break;
2987 case Intrinsic::abs:
2988 ISD = ISD::ABS;
2989 break;
2990 case Intrinsic::bitreverse:
2991 ISD = ISD::BITREVERSE;
2992 break;
2993 case Intrinsic::bswap:
2994 ISD = ISD::BSWAP;
2995 break;
2996 case Intrinsic::ctlz:
2997 ISD = ISD::CTLZ;
2998 break;
2999 case Intrinsic::ctpop:
3000 ISD = ISD::CTPOP;
3001 break;
3002 case Intrinsic::cttz:
3003 ISD = ISD::CTTZ;
3004 break;
3005 case Intrinsic::maxnum:
3006 case Intrinsic::minnum:
3007 // FMINNUM has same costs so don't duplicate.
3008 ISD = ISD::FMAXNUM;
3009 break;
3010 case Intrinsic::sadd_sat:
3011 ISD = ISD::SADDSAT;
3012 break;
3013 case Intrinsic::smax:
3014 ISD = ISD::SMAX;
3015 break;
3016 case Intrinsic::smin:
3017 ISD = ISD::SMIN;
3018 break;
3019 case Intrinsic::ssub_sat:
3020 ISD = ISD::SSUBSAT;
3021 break;
3022 case Intrinsic::uadd_sat:
3023 ISD = ISD::UADDSAT;
3024 break;
3025 case Intrinsic::umax:
3026 ISD = ISD::UMAX;
3027 break;
3028 case Intrinsic::umin:
3029 ISD = ISD::UMIN;
3030 break;
3031 case Intrinsic::usub_sat:
3032 ISD = ISD::USUBSAT;
3033 break;
3034 case Intrinsic::sqrt:
3035 ISD = ISD::FSQRT;
3036 break;
3037 case Intrinsic::sadd_with_overflow:
3038 case Intrinsic::ssub_with_overflow:
3039 // SSUBO has same costs so don't duplicate.
3040 ISD = ISD::SADDO;
3041 OpTy = RetTy->getContainedType(0);
3042 break;
3043 case Intrinsic::uadd_with_overflow:
3044 case Intrinsic::usub_with_overflow:
3045 // USUBO has same costs so don't duplicate.
3046 ISD = ISD::UADDO;
3047 OpTy = RetTy->getContainedType(0);
3048 break;
3049 case Intrinsic::umul_with_overflow:
3050 case Intrinsic::smul_with_overflow:
3051 // SMULO has same costs so don't duplicate.
3052 ISD = ISD::UMULO;
3053 OpTy = RetTy->getContainedType(0);
3054 break;
3055 }
3056
3057 if (ISD != ISD::DELETED_NODE) {
3058 // Legalize the type.
3059 std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, OpTy);
3060 MVT MTy = LT.second;
3061
3062 // Attempt to lookup cost.
3063 if (ISD == ISD::BITREVERSE && ST->hasGFNI() && ST->hasSSSE3() &&
3064 MTy.isVector()) {
3065 // With PSHUFB the code is very similar for all types. If we have integer
3066 // byte operations, we just need a GF2P8AFFINEQB for vXi8. For other types
3067 // we also need a PSHUFB.
3068 unsigned Cost = MTy.getVectorElementType() == MVT::i8 ? 1 : 2;
3069
3070 // Without byte operations, we need twice as many GF2P8AFFINEQB and PSHUFB
3071 // instructions. We also need an extract and an insert.
3072 if (!(MTy.is128BitVector() || (ST->hasAVX2() && MTy.is256BitVector()) ||
3073 (ST->hasBWI() && MTy.is512BitVector())))
3074 Cost = Cost * 2 + 2;
3075
3076 return LT.first * Cost;
3077 }
3078
3079 auto adjustTableCost = [](const CostTblEntry &Entry,
3080 InstructionCost LegalizationCost,
3081 FastMathFlags FMF) {
3082 // If there are no NANs to deal with, then these are reduced to a
3083 // single MIN** or MAX** instruction instead of the MIN/CMP/SELECT that we
3084 // assume is used in the non-fast case.
3085 if (Entry.ISD == ISD::FMAXNUM || Entry.ISD == ISD::FMINNUM) {
3086 if (FMF.noNaNs())
3087 return LegalizationCost * 1;
3088 }
3089 return LegalizationCost * (int)Entry.Cost;
3090 };
3091
3092 if (ST->useGLMDivSqrtCosts())
3093 if (const auto *Entry = CostTableLookup(GLMCostTbl, ISD, MTy))
3094 return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3095
3096 if (ST->isSLM())
3097 if (const auto *Entry = CostTableLookup(SLMCostTbl, ISD, MTy))
3098 return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3099
3100 if (ST->hasCDI())
3101 if (const auto *Entry = CostTableLookup(AVX512CDCostTbl, ISD, MTy))
3102 return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3103
3104 if (ST->hasBWI())
3105 if (const auto *Entry = CostTableLookup(AVX512BWCostTbl, ISD, MTy))
3106 return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3107
3108 if (ST->hasAVX512())
3109 if (const auto *Entry = CostTableLookup(AVX512CostTbl, ISD, MTy))
3110 return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3111
3112 if (ST->hasXOP())
3113 if (const auto *Entry = CostTableLookup(XOPCostTbl, ISD, MTy))
3114 return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3115
3116 if (ST->hasAVX2())
3117 if (const auto *Entry = CostTableLookup(AVX2CostTbl, ISD, MTy))
3118 return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3119
3120 if (ST->hasAVX())
3121 if (const auto *Entry = CostTableLookup(AVX1CostTbl, ISD, MTy))
3122 return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3123
3124 if (ST->hasSSE42())
3125 if (const auto *Entry = CostTableLookup(SSE42CostTbl, ISD, MTy))
3126 return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3127
3128 if (ST->hasSSE41())
3129 if (const auto *Entry = CostTableLookup(SSE41CostTbl, ISD, MTy))
3130 return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3131
3132 if (ST->hasSSSE3())
3133 if (const auto *Entry = CostTableLookup(SSSE3CostTbl, ISD, MTy))
3134 return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3135
3136 if (ST->hasSSE2())
3137 if (const auto *Entry = CostTableLookup(SSE2CostTbl, ISD, MTy))
3138 return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3139
3140 if (ST->hasSSE1())
3141 if (const auto *Entry = CostTableLookup(SSE1CostTbl, ISD, MTy))
3142 return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3143
3144 if (ST->hasBMI()) {
3145 if (ST->is64Bit())
3146 if (const auto *Entry = CostTableLookup(BMI64CostTbl, ISD, MTy))
3147 return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3148
3149 if (const auto *Entry = CostTableLookup(BMI32CostTbl, ISD, MTy))
3150 return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3151 }
3152
3153 if (ST->hasLZCNT()) {
3154 if (ST->is64Bit())
3155 if (const auto *Entry = CostTableLookup(LZCNT64CostTbl, ISD, MTy))
3156 return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3157
3158 if (const auto *Entry = CostTableLookup(LZCNT32CostTbl, ISD, MTy))
3159 return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3160 }
3161
3162 if (ST->hasPOPCNT()) {
3163 if (ST->is64Bit())
3164 if (const auto *Entry = CostTableLookup(POPCNT64CostTbl, ISD, MTy))
3165 return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3166
3167 if (const auto *Entry = CostTableLookup(POPCNT32CostTbl, ISD, MTy))
3168 return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3169 }
3170
3171 if (ISD == ISD::BSWAP && ST->hasMOVBE() && ST->hasFastMOVBE()) {
3172 if (const Instruction *II = ICA.getInst()) {
3173 if (II->hasOneUse() && isa<StoreInst>(II->user_back()))
3174 return TTI::TCC_Free;
3175 if (auto *LI = dyn_cast<LoadInst>(II->getOperand(0))) {
3176 if (LI->hasOneUse())
3177 return TTI::TCC_Free;
3178 }
3179 }
3180 }
3181
3182 // TODO - add BMI (TZCNT) scalar handling
3183
3184 if (ST->is64Bit())
3185 if (const auto *Entry = CostTableLookup(X64CostTbl, ISD, MTy))
3186 return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3187
3188 if (const auto *Entry = CostTableLookup(X86CostTbl, ISD, MTy))
3189 return adjustTableCost(*Entry, LT.first, ICA.getFlags());
3190 }
3191
3192 return BaseT::getIntrinsicInstrCost(ICA, CostKind);
3193}
3194
3195InstructionCost
3196X86TTIImpl::getIntrinsicInstrCost(const IntrinsicCostAttributes &ICA,
3197 TTI::TargetCostKind CostKind) {
3198 if (ICA.isTypeBasedOnly())
3199 return getTypeBasedIntrinsicInstrCost(ICA, CostKind);
3200
3201 static const CostTblEntry AVX512CostTbl[] = {
3202 { ISD::ROTL, MVT::v8i64, 1 },
3203 { ISD::ROTL, MVT::v4i64, 1 },
3204 { ISD::ROTL, MVT::v2i64, 1 },
3205 { ISD::ROTL, MVT::v16i32, 1 },
3206 { ISD::ROTL, MVT::v8i32, 1 },
3207 { ISD::ROTL, MVT::v4i32, 1 },
3208 { ISD::ROTR, MVT::v8i64, 1 },
3209 { ISD::ROTR, MVT::v4i64, 1 },
3210 { ISD::ROTR, MVT::v2i64, 1 },
3211 { ISD::ROTR, MVT::v16i32, 1 },
3212 { ISD::ROTR, MVT::v8i32, 1 },
3213 { ISD::ROTR, MVT::v4i32, 1 }
3214 };
3215 // XOP: ROTL = VPROT(X,Y), ROTR = VPROT(X,SUB(0,Y))
3216 static const CostTblEntry XOPCostTbl[] = {
3217 { ISD::ROTL, MVT::v4i64, 4 },
3218 { ISD::ROTL, MVT::v8i32, 4 },
3219 { ISD::ROTL, MVT::v16i16, 4 },
3220 { ISD::ROTL, MVT::v32i8, 4 },
3221 { ISD::ROTL, MVT::v2i64, 1 },
3222 { ISD::ROTL, MVT::v4i32, 1 },
3223 { ISD::ROTL, MVT::v8i16, 1 },
3224 { ISD::ROTL, MVT::v16i8, 1 },
3225 { ISD::ROTR, MVT::v4i64, 6 },
3226 { ISD::ROTR, MVT::v8i32, 6 },
3227 { ISD::ROTR, MVT::v16i16, 6 },
3228 { ISD::ROTR, MVT::v32i8, 6 },
3229 { ISD::ROTR, MVT::v2i64, 2 },
3230 { ISD::ROTR, MVT::v4i32, 2 },
3231 { ISD::ROTR, MVT::v8i16, 2 },
3232 { ISD::ROTR, MVT::v16i8, 2 }
3233 };
3234 static const CostTblEntry X64CostTbl[] = { // 64-bit targets
3235 { ISD::ROTL, MVT::i64, 1 },
3236 { ISD::ROTR, MVT::i64, 1 },
3237 { ISD::FSHL, MVT::i64, 4 }
3238 };
3239 static const CostTblEntry X86CostTbl[] = { // 32 or 64-bit targets
3240 { ISD::ROTL, MVT::i32, 1 },
3241 { ISD::ROTL, MVT::i16, 1 },
3242 { ISD::ROTL, MVT::i8, 1 },
3243 { ISD::ROTR, MVT::i32, 1 },
3244 { ISD::ROTR, MVT::i16, 1 },
3245 { ISD::ROTR, MVT::i8, 1 },
3246 { ISD::FSHL, MVT::i32, 4 },
3247 { ISD::FSHL, MVT::i16, 4 },
3248 { ISD::FSHL, MVT::i8, 4 }
3249 };
3250
3251 Intrinsic::ID IID = ICA.getID();
3252 Type *RetTy = ICA.getReturnType();
3253 const SmallVectorImpl<const Value *> &Args = ICA.getArgs();
3254 unsigned ISD = ISD::DELETED_NODE;
3255 switch (IID) {
3256 default:
3257 break;
3258 case Intrinsic::fshl:
3259 ISD = ISD::FSHL;
3260 if (Args[0] == Args[1])
3261 ISD = ISD::ROTL;
3262 break;
3263 case Intrinsic::fshr:
3264 // FSHR has same costs so don't duplicate.
3265 ISD = ISD::FSHL;
3266 if (Args[0] == Args[1])
3267 ISD = ISD::ROTR;
3268 break;
3269 }
3270
3271 if (ISD != ISD::DELETED_NODE) {
3272 // Legalize the type.
3273 std::pair<InstructionCost, MVT> LT =
3274 TLI->getTypeLegalizationCost(DL, RetTy);
3275 MVT MTy = LT.second;
3276
3277 // Attempt to lookup cost.
3278 if (ST->hasAVX512())
3279 if (const auto *Entry = CostTableLookup(AVX512CostTbl, ISD, MTy))
3280 return LT.first * Entry->Cost;
3281
3282 if (ST->hasXOP())
3283 if (const auto *Entry = CostTableLookup(XOPCostTbl, ISD, MTy))
3284 return LT.first * Entry->Cost;
3285
3286 if (ST->is64Bit())
3287 if (const auto *Entry = CostTableLookup(X64CostTbl, ISD, MTy))
3288 return LT.first * Entry->Cost;
3289
3290 if (const auto *Entry = CostTableLookup(X86CostTbl, ISD, MTy))
3291 return LT.first * Entry->Cost;
3292 }
3293
3294 return BaseT::getIntrinsicInstrCost(ICA, CostKind);
3295}
3296
3297InstructionCost X86TTIImpl::getVectorInstrCost(unsigned Opcode, Type *Val,
3298 unsigned Index) {
3299 static const CostTblEntry SLMCostTbl[] = {
3300 { ISD::EXTRACT_VECTOR_ELT, MVT::i8, 4 },
3301 { ISD::EXTRACT_VECTOR_ELT, MVT::i16, 4 },
3302 { ISD::EXTRACT_VECTOR_ELT, MVT::i32, 4 },
3303 { ISD::EXTRACT_VECTOR_ELT, MVT::i64, 7 }
3304 };
3305
3306 assert(Val->isVectorTy() && "This must be a vector type")((void)0);
3307 Type *ScalarType = Val->getScalarType();
3308 int RegisterFileMoveCost = 0;
3309
3310 // Non-immediate extraction/insertion can be handled as a sequence of
3311 // aliased loads+stores via the stack.
3312 if (Index == -1U && (Opcode == Instruction::ExtractElement ||
3313 Opcode == Instruction::InsertElement)) {
3314 // TODO: On some SSE41+ targets, we expand to cmp+splat+select patterns:
3315 // inselt N0, N1, N2 --> select (SplatN2 == {0,1,2...}) ? SplatN1 : N0.
3316
3317 // TODO: Move this to BasicTTIImpl.h? We'd need better gep + index handling.
3318 assert(isa<FixedVectorType>(Val) && "Fixed vector type expected")((void)0);
3319 Align VecAlign = DL.getPrefTypeAlign(Val);
3320 Align SclAlign = DL.getPrefTypeAlign(ScalarType);
3321
3322 // Extract - store vector to stack, load scalar.
3323 if (Opcode == Instruction::ExtractElement) {
3324 return getMemoryOpCost(Instruction::Store, Val, VecAlign, 0,
3325 TTI::TargetCostKind::TCK_RecipThroughput) +
3326 getMemoryOpCost(Instruction::Load, ScalarType, SclAlign, 0,
3327 TTI::TargetCostKind::TCK_RecipThroughput);
3328 }
3329 // Insert - store vector to stack, store scalar, load vector.
3330 if (Opcode == Instruction::InsertElement) {
3331 return getMemoryOpCost(Instruction::Store, Val, VecAlign, 0,
3332 TTI::TargetCostKind::TCK_RecipThroughput) +
3333 getMemoryOpCost(Instruction::Store, ScalarType, SclAlign, 0,
3334 TTI::TargetCostKind::TCK_RecipThroughput) +
3335 getMemoryOpCost(Instruction::Load, Val, VecAlign, 0,
3336 TTI::TargetCostKind::TCK_RecipThroughput);
3337 }
3338 }
3339
3340 if (Index != -1U && (Opcode == Instruction::ExtractElement ||
3341 Opcode == Instruction::InsertElement)) {
3342 // Legalize the type.
3343 std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, Val);
3344
3345 // This type is legalized to a scalar type.
3346 if (!LT.second.isVector())
3347 return 0;
3348
3349 // The type may be split. Normalize the index to the new type.
3350 unsigned NumElts = LT.second.getVectorNumElements();
3351 unsigned SubNumElts = NumElts;
3352 Index = Index % NumElts;
3353
3354 // For >128-bit vectors, we need to extract higher 128-bit subvectors.
3355 // For inserts, we also need to insert the subvector back.
3356 if (LT.second.getSizeInBits() > 128) {
3357 assert((LT.second.getSizeInBits() % 128) == 0 && "Illegal vector")((void)0);
3358 unsigned NumSubVecs = LT.second.getSizeInBits() / 128;
3359 SubNumElts = NumElts / NumSubVecs;
3360 if (SubNumElts <= Index) {
3361 RegisterFileMoveCost += (Opcode == Instruction::InsertElement ? 2 : 1);
3362 Index %= SubNumElts;
3363 }
3364 }
3365
3366 if (Index == 0) {
3367 // Floating point scalars are already located in index #0.
3368 // Many insertions to #0 can fold away for scalar fp-ops, so let's assume
3369 // true for all.
3370 if (ScalarType->isFloatingPointTy())
3371 return RegisterFileMoveCost;
3372
3373 // Assume movd/movq XMM -> GPR is relatively cheap on all targets.
3374 if (ScalarType->isIntegerTy() && Opcode == Instruction::ExtractElement)
3375 return 1 + RegisterFileMoveCost;
3376 }
3377
3378 int ISD = TLI->InstructionOpcodeToISD(Opcode);
3379 assert(ISD && "Unexpected vector opcode")((void)0);
3380 MVT MScalarTy = LT.second.getScalarType();
3381 if (ST->isSLM())
3382 if (auto *Entry = CostTableLookup(SLMCostTbl, ISD, MScalarTy))
3383 return Entry->Cost + RegisterFileMoveCost;
3384
3385 // Assume pinsr/pextr XMM <-> GPR is relatively cheap on all targets.
3386 if ((MScalarTy == MVT::i16 && ST->hasSSE2()) ||
3387 (MScalarTy.isInteger() && ST->hasSSE41()))
3388 return 1 + RegisterFileMoveCost;
3389
3390 // Assume insertps is relatively cheap on all targets.
3391 if (MScalarTy == MVT::f32 && ST->hasSSE41() &&
3392 Opcode == Instruction::InsertElement)
3393 return 1 + RegisterFileMoveCost;
3394
3395 // For extractions we just need to shuffle the element to index 0, which
3396 // should be very cheap (assume cost = 1). For insertions we need to shuffle
3397 // the elements to its destination. In both cases we must handle the
3398 // subvector move(s).
3399 // If the vector type is already less than 128-bits then don't reduce it.
3400 // TODO: Under what circumstances should we shuffle using the full width?
3401 InstructionCost ShuffleCost = 1;
3402 if (Opcode == Instruction::InsertElement) {
3403 auto *SubTy = cast<VectorType>(Val);
3404 EVT VT = TLI->getValueType(DL, Val);
3405 if (VT.getScalarType() != MScalarTy || VT.getSizeInBits() >= 128)
3406 SubTy = FixedVectorType::get(ScalarType, SubNumElts);
3407 ShuffleCost =
3408 getShuffleCost(TTI::SK_PermuteTwoSrc, SubTy, None, 0, SubTy);
3409 }
3410 int IntOrFpCost = ScalarType->isFloatingPointTy() ? 0 : 1;
3411 return ShuffleCost + IntOrFpCost + RegisterFileMoveCost;
3412 }
3413
3414 // Add to the base cost if we know that the extracted element of a vector is
3415 // destined to be moved to and used in the integer register file.
3416 if (Opcode == Instruction::ExtractElement && ScalarType->isPointerTy())
3417 RegisterFileMoveCost += 1;
3418
3419 return BaseT::getVectorInstrCost(Opcode, Val, Index) + RegisterFileMoveCost;
3420}
3421
3422InstructionCost X86TTIImpl::getScalarizationOverhead(VectorType *Ty,
3423 const APInt &DemandedElts,
3424 bool Insert,
3425 bool Extract) {
3426 InstructionCost Cost = 0;
3427
3428 // For insertions, a ISD::BUILD_VECTOR style vector initialization can be much
3429 // cheaper than an accumulation of ISD::INSERT_VECTOR_ELT.
3430 if (Insert) {
3431 std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, Ty);
3432 MVT MScalarTy = LT.second.getScalarType();
3433
3434 if ((MScalarTy == MVT::i16 && ST->hasSSE2()) ||
3435 (MScalarTy.isInteger() && ST->hasSSE41()) ||
3436 (MScalarTy == MVT::f32 && ST->hasSSE41())) {
3437 // For types we can insert directly, insertion into 128-bit sub vectors is
3438 // cheap, followed by a cheap chain of concatenations.
3439 if (LT.second.getSizeInBits() <= 128) {
3440 Cost +=
3441 BaseT::getScalarizationOverhead(Ty, DemandedElts, Insert, false);
3442 } else {
3443 // In each 128-lane, if at least one index is demanded but not all
3444 // indices are demanded and this 128-lane is not the first 128-lane of
3445 // the legalized-vector, then this 128-lane needs a extracti128; If in
3446 // each 128-lane, there is at least one demanded index, this 128-lane
3447 // needs a inserti128.
3448
3449 // The following cases will help you build a better understanding:
3450 // Assume we insert several elements into a v8i32 vector in avx2,
3451 // Case#1: inserting into 1th index needs vpinsrd + inserti128.
3452 // Case#2: inserting into 5th index needs extracti128 + vpinsrd +
3453 // inserti128.
3454 // Case#3: inserting into 4,5,6,7 index needs 4*vpinsrd + inserti128.
3455 const int CostValue = *LT.first.getValue();
3456 assert(CostValue >= 0 && "Negative cost!")((void)0);
3457 unsigned Num128Lanes = LT.second.getSizeInBits() / 128 * CostValue;
3458 unsigned NumElts = LT.second.getVectorNumElements() * CostValue;
3459 APInt WidenedDemandedElts = DemandedElts.zextOrSelf(NumElts);
3460 unsigned Scale = NumElts / Num128Lanes;
3461 // We iterate each 128-lane, and check if we need a
3462 // extracti128/inserti128 for this 128-lane.
3463 for (unsigned I = 0; I < NumElts; I += Scale) {
3464 APInt Mask = WidenedDemandedElts.getBitsSet(NumElts, I, I + Scale);
3465 APInt MaskedDE = Mask & WidenedDemandedElts;
3466 unsigned Population = MaskedDE.countPopulation();
3467 Cost += (Population > 0 && Population != Scale &&
3468 I % LT.second.getVectorNumElements() != 0);
3469 Cost += Population > 0;
3470 }
3471 Cost += DemandedElts.countPopulation();
3472
3473 // For vXf32 cases, insertion into the 0'th index in each v4f32
3474 // 128-bit vector is free.
3475 // NOTE: This assumes legalization widens vXf32 vectors.
3476 if (MScalarTy == MVT::f32)
3477 for (unsigned i = 0, e = cast<FixedVectorType>(Ty)->getNumElements();
3478 i < e; i += 4)
3479 if (DemandedElts[i])
3480 Cost--;
3481 }
3482 } else if (LT.second.isVector()) {
3483 // Without fast insertion, we need to use MOVD/MOVQ to pass each demanded
3484 // integer element as a SCALAR_TO_VECTOR, then we build the vector as a
3485 // series of UNPCK followed by CONCAT_VECTORS - all of these can be
3486 // considered cheap.
3487 if (Ty->isIntOrIntVectorTy())
3488 Cost += DemandedElts.countPopulation();
3489
3490 // Get the smaller of the legalized or original pow2-extended number of
3491 // vector elements, which represents the number of unpacks we'll end up
3492 // performing.
3493 unsigned NumElts = LT.second.getVectorNumElements();
3494 unsigned Pow2Elts =
3495 PowerOf2Ceil(cast<FixedVectorType>(Ty)->getNumElements());
3496 Cost += (std::min<unsigned>(NumElts, Pow2Elts) - 1) * LT.first;
3497 }
3498 }
3499
3500 // TODO: Use default extraction for now, but we should investigate extending this
3501 // to handle repeated subvector extraction.
3502 if (Extract)
3503 Cost += BaseT::getScalarizationOverhead(Ty, DemandedElts, false, Extract);
3504
3505 return Cost;
3506}
3507
3508InstructionCost X86TTIImpl::getMemoryOpCost(unsigned Opcode, Type *Src,
3509 MaybeAlign Alignment,
3510 unsigned AddressSpace,
3511 TTI::TargetCostKind CostKind,
3512 const Instruction *I) {
3513 // TODO: Handle other cost kinds.
3514 if (CostKind != TTI::TCK_RecipThroughput) {
3515 if (auto *SI = dyn_cast_or_null<StoreInst>(I)) {
3516 // Store instruction with index and scale costs 2 Uops.
3517 // Check the preceding GEP to identify non-const indices.
3518 if (auto *GEP = dyn_cast<GetElementPtrInst>(SI->getPointerOperand())) {
3519 if (!all_of(GEP->indices(), [](Value *V) { return isa<Constant>(V); }))
3520 return TTI::TCC_Basic * 2;
3521 }
3522 }
3523 return TTI::TCC_Basic;
3524 }
3525
3526 assert((Opcode == Instruction::Load || Opcode == Instruction::Store) &&((void)0)
3527 "Invalid Opcode")((void)0);
3528 // Type legalization can't handle structs
3529 if (TLI->getValueType(DL, Src, true) == MVT::Other)
3530 return BaseT::getMemoryOpCost(Opcode, Src, Alignment, AddressSpace,
3531 CostKind);
3532
3533 // Legalize the type.
3534 std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, Src);
3535
3536 auto *VTy = dyn_cast<FixedVectorType>(Src);
3537
3538 // Handle the simple case of non-vectors.
3539 // NOTE: this assumes that legalization never creates vector from scalars!
3540 if (!VTy || !LT.second.isVector())
3541 // Each load/store unit costs 1.
3542 return LT.first * 1;
3543
3544 bool IsLoad = Opcode == Instruction::Load;
3545
3546 Type *EltTy = VTy->getElementType();
3547
3548 const int EltTyBits = DL.getTypeSizeInBits(EltTy);
3549
3550 InstructionCost Cost = 0;
3551
3552 // Source of truth: how many elements were there in the original IR vector?
3553 const unsigned SrcNumElt = VTy->getNumElements();
3554
3555 // How far have we gotten?
3556 int NumEltRemaining = SrcNumElt;
3557 // Note that we intentionally capture by-reference, NumEltRemaining changes.
3558 auto NumEltDone = [&]() { return SrcNumElt - NumEltRemaining; };
3559
3560 const int MaxLegalOpSizeBytes = divideCeil(LT.second.getSizeInBits(), 8);
3561
3562 // Note that even if we can store 64 bits of an XMM, we still operate on XMM.
3563 const unsigned XMMBits = 128;
3564 if (XMMBits % EltTyBits != 0)
3565 // Vector size must be a multiple of the element size. I.e. no padding.
3566 return BaseT::getMemoryOpCost(Opcode, Src, Alignment, AddressSpace,
3567 CostKind);
3568 const int NumEltPerXMM = XMMBits / EltTyBits;
3569
3570 auto *XMMVecTy = FixedVectorType::get(EltTy, NumEltPerXMM);
3571
3572 for (int CurrOpSizeBytes = MaxLegalOpSizeBytes, SubVecEltsLeft = 0;
3573 NumEltRemaining > 0; CurrOpSizeBytes /= 2) {
3574 // How many elements would a single op deal with at once?
3575 if ((8 * CurrOpSizeBytes) % EltTyBits != 0)
3576 // Vector size must be a multiple of the element size. I.e. no padding.
3577 return BaseT::getMemoryOpCost(Opcode, Src, Alignment, AddressSpace,
3578 CostKind);
3579 int CurrNumEltPerOp = (8 * CurrOpSizeBytes) / EltTyBits;
3580
3581 assert(CurrOpSizeBytes > 0 && CurrNumEltPerOp > 0 && "How'd we get here?")((void)0);
3582 assert((((NumEltRemaining * EltTyBits) < (2 * 8 * CurrOpSizeBytes)) ||((void)0)
3583 (CurrOpSizeBytes == MaxLegalOpSizeBytes)) &&((void)0)
3584 "Unless we haven't halved the op size yet, "((void)0)
3585 "we have less than two op's sized units of work left.")((void)0);
3586
3587 auto *CurrVecTy = CurrNumEltPerOp > NumEltPerXMM
3588 ? FixedVectorType::get(EltTy, CurrNumEltPerOp)
3589 : XMMVecTy;
3590
3591 assert(CurrVecTy->getNumElements() % CurrNumEltPerOp == 0 &&((void)0)
3592 "After halving sizes, the vector elt count is no longer a multiple "((void)0)
3593 "of number of elements per operation?")((void)0);
3594 auto *CoalescedVecTy =
3595 CurrNumEltPerOp == 1
3596 ? CurrVecTy
3597 : FixedVectorType::get(
3598 IntegerType::get(Src->getContext(),
3599 EltTyBits * CurrNumEltPerOp),
3600 CurrVecTy->getNumElements() / CurrNumEltPerOp);
3601 assert(DL.getTypeSizeInBits(CoalescedVecTy) ==((void)0)
3602 DL.getTypeSizeInBits(CurrVecTy) &&((void)0)
3603 "coalesciing elements doesn't change vector width.")((void)0);
3604
3605 while (NumEltRemaining > 0) {
3606 assert(SubVecEltsLeft >= 0 && "Subreg element count overconsumtion?")((void)0);
3607
3608 // Can we use this vector size, as per the remaining element count?
3609 // Iff the vector is naturally aligned, we can do a wide load regardless.
3610 if (NumEltRemaining < CurrNumEltPerOp &&
3611 (!IsLoad || Alignment.valueOrOne() < CurrOpSizeBytes) &&
3612 CurrOpSizeBytes != 1)
3613 break; // Try smalled vector size.
3614
3615 bool Is0thSubVec = (NumEltDone() % LT.second.getVectorNumElements()) == 0;
3616
3617 // If we have fully processed the previous reg, we need to replenish it.
3618 if (SubVecEltsLeft == 0) {
3619 SubVecEltsLeft += CurrVecTy->getNumElements();
3620 // And that's free only for the 0'th subvector of a legalized vector.
3621 if (!Is0thSubVec)
3622 Cost += getShuffleCost(IsLoad ? TTI::ShuffleKind::SK_InsertSubvector
3623 : TTI::ShuffleKind::SK_ExtractSubvector,
3624 VTy, None, NumEltDone(), CurrVecTy);
3625 }
3626
3627 // While we can directly load/store ZMM, YMM, and 64-bit halves of XMM,
3628 // for smaller widths (32/16/8) we have to insert/extract them separately.
3629 // Again, it's free for the 0'th subreg (if op is 32/64 bit wide,
3630 // but let's pretend that it is also true for 16/8 bit wide ops...)
3631 if (CurrOpSizeBytes <= 32 / 8 && !Is0thSubVec) {
3632 int NumEltDoneInCurrXMM = NumEltDone() % NumEltPerXMM;
3633 assert(NumEltDoneInCurrXMM % CurrNumEltPerOp == 0 && "")((void)0);
3634 int CoalescedVecEltIdx = NumEltDoneInCurrXMM / CurrNumEltPerOp;
3635 APInt DemandedElts =
3636 APInt::getBitsSet(CoalescedVecTy->getNumElements(),
3637 CoalescedVecEltIdx, CoalescedVecEltIdx + 1);
3638 assert(DemandedElts.countPopulation() == 1 && "Inserting single value")((void)0);
3639 Cost += getScalarizationOverhead(CoalescedVecTy, DemandedElts, IsLoad,
3640 !IsLoad);
3641 }
3642
3643 // This isn't exactly right. We're using slow unaligned 32-byte accesses
3644 // as a proxy for a double-pumped AVX memory interface such as on
3645 // Sandybridge.
3646 if (CurrOpSizeBytes == 32 && ST->isUnalignedMem32Slow())
3647 Cost += 2;
3648 else
3649 Cost += 1;
3650
3651 SubVecEltsLeft -= CurrNumEltPerOp;
3652 NumEltRemaining -= CurrNumEltPerOp;
3653 Alignment = commonAlignment(Alignment.valueOrOne(), CurrOpSizeBytes);
3654 }
3655 }
3656
3657 assert(NumEltRemaining <= 0 && "Should have processed all the elements.")((void)0);
3658
3659 return Cost;
3660}
3661
3662InstructionCost
3663X86TTIImpl::getMaskedMemoryOpCost(unsigned Opcode, Type *SrcTy, Align Alignment,
3664 unsigned AddressSpace,
3665 TTI::TargetCostKind CostKind) {
3666 bool IsLoad = (Instruction::Load == Opcode);
3667 bool IsStore = (Instruction::Store == Opcode);
3668
3669 auto *SrcVTy = dyn_cast<FixedVectorType>(SrcTy);
3670 if (!SrcVTy)
3671 // To calculate scalar take the regular cost, without mask
3672 return getMemoryOpCost(Opcode, SrcTy, Alignment, AddressSpace, CostKind);
3673
3674 unsigned NumElem = SrcVTy->getNumElements();
3675 auto *MaskTy =
3676 FixedVectorType::get(Type::getInt8Ty(SrcVTy->getContext()), NumElem);
3677 if ((IsLoad && !isLegalMaskedLoad(SrcVTy, Alignment)) ||
3678 (IsStore && !isLegalMaskedStore(SrcVTy, Alignment))) {
3679 // Scalarization
3680 APInt DemandedElts = APInt::getAllOnesValue(NumElem);
3681 InstructionCost MaskSplitCost =
3682 getScalarizationOverhead(MaskTy, DemandedElts, false, true);
3683 InstructionCost ScalarCompareCost = getCmpSelInstrCost(
3684 Instruction::ICmp, Type::getInt8Ty(SrcVTy->getContext()), nullptr,
3685 CmpInst::BAD_ICMP_PREDICATE, CostKind);
3686 InstructionCost BranchCost = getCFInstrCost(Instruction::Br, CostKind);
3687 InstructionCost MaskCmpCost = NumElem * (BranchCost + ScalarCompareCost);
3688 InstructionCost ValueSplitCost =
3689 getScalarizationOverhead(SrcVTy, DemandedElts, IsLoad, IsStore);
3690 InstructionCost MemopCost =
3691 NumElem * BaseT::getMemoryOpCost(Opcode, SrcVTy->getScalarType(),
3692 Alignment, AddressSpace, CostKind);
3693 return MemopCost + ValueSplitCost + MaskSplitCost + MaskCmpCost;
3694 }
3695
3696 // Legalize the type.
3697 std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, SrcVTy);
3698 auto VT = TLI->getValueType(DL, SrcVTy);
3699 InstructionCost Cost = 0;
3700 if (VT.isSimple() && LT.second != VT.getSimpleVT() &&
3701 LT.second.getVectorNumElements() == NumElem)
3702 // Promotion requires extend/truncate for data and a shuffle for mask.
3703 Cost += getShuffleCost(TTI::SK_PermuteTwoSrc, SrcVTy, None, 0, nullptr) +
3704 getShuffleCost(TTI::SK_PermuteTwoSrc, MaskTy, None, 0, nullptr);
3705
3706 else if (LT.first * LT.second.getVectorNumElements() > NumElem) {
3707 auto *NewMaskTy = FixedVectorType::get(MaskTy->getElementType(),
3708 LT.second.getVectorNumElements());
3709 // Expanding requires fill mask with zeroes
3710 Cost += getShuffleCost(TTI::SK_InsertSubvector, NewMaskTy, None, 0, MaskTy);
3711 }
3712
3713 // Pre-AVX512 - each maskmov load costs 2 + store costs ~8.
3714 if (!ST->hasAVX512())
3715 return Cost + LT.first * (IsLoad ? 2 : 8);
3716
3717 // AVX-512 masked load/store is cheapper
3718 return Cost + LT.first;
3719}
3720
3721InstructionCost X86TTIImpl::getAddressComputationCost(Type *Ty,
3722 ScalarEvolution *SE,
3723 const SCEV *Ptr) {
3724 // Address computations in vectorized code with non-consecutive addresses will
3725 // likely result in more instructions compared to scalar code where the
3726 // computation can more often be merged into the index mode. The resulting
3727 // extra micro-ops can significantly decrease throughput.
3728 const unsigned NumVectorInstToHideOverhead = 10;
3729
3730 // Cost modeling of Strided Access Computation is hidden by the indexing
3731 // modes of X86 regardless of the stride value. We dont believe that there
3732 // is a difference between constant strided access in gerenal and constant
3733 // strided value which is less than or equal to 64.
3734 // Even in the case of (loop invariant) stride whose value is not known at
3735 // compile time, the address computation will not incur more than one extra
3736 // ADD instruction.
3737 if (Ty->isVectorTy() && SE) {
3738 if (!BaseT::isStridedAccess(Ptr))
3739 return NumVectorInstToHideOverhead;
3740 if (!BaseT::getConstantStrideStep(SE, Ptr))
3741 return 1;
3742 }
3743
3744 return BaseT::getAddressComputationCost(Ty, SE, Ptr);
3745}
3746
3747InstructionCost
3748X86TTIImpl::getArithmeticReductionCost(unsigned Opcode, VectorType *ValTy,
3749 Optional<FastMathFlags> FMF,
3750 TTI::TargetCostKind CostKind) {
3751 if (TTI::requiresOrderedReduction(FMF))
3752 return BaseT::getArithmeticReductionCost(Opcode, ValTy, FMF, CostKind);
3753
3754 // We use the Intel Architecture Code Analyzer(IACA) to measure the throughput
3755 // and make it as the cost.
3756
3757 static const CostTblEntry SLMCostTblNoPairWise[] = {
3758 { ISD::FADD, MVT::v2f64, 3 },
3759 { ISD::ADD, MVT::v2i64, 5 },
3760 };
3761
3762 static const CostTblEntry SSE2CostTblNoPairWise[] = {
3763 { ISD::FADD, MVT::v2f64, 2 },
3764 { ISD::FADD, MVT::v2f32, 2 },
3765 { ISD::FADD, MVT::v4f32, 4 },
3766 { ISD::ADD, MVT::v2i64, 2 }, // The data reported by the IACA tool is "1.6".
3767 { ISD::ADD, MVT::v2i32, 2 }, // FIXME: chosen to be less than v4i32
3768 { ISD::ADD, MVT::v4i32, 3 }, // The data reported by the IACA tool is "3.3".
3769 { ISD::ADD, MVT::v2i16, 2 }, // The data reported by the IACA tool is "4.3".
3770 { ISD::ADD, MVT::v4i16, 3 }, // The data reported by the IACA tool is "4.3".
3771 { ISD::ADD, MVT::v8i16, 4 }, // The data reported by the IACA tool is "4.3".
3772 { ISD::ADD, MVT::v2i8, 2 },
3773 { ISD::ADD, MVT::v4i8, 2 },
3774 { ISD::ADD, MVT::v8i8, 2 },
3775 { ISD::ADD, MVT::v16i8, 3 },
3776 };
3777
3778 static const CostTblEntry AVX1CostTblNoPairWise[] = {
3779 { ISD::FADD, MVT::v4f64, 3 },
3780 { ISD::FADD, MVT::v4f32, 3 },
3781 { ISD::FADD, MVT::v8f32, 4 },
3782 { ISD::ADD, MVT::v2i64, 1 }, // The data reported by the IACA tool is "1.5".
3783 { ISD::ADD, MVT::v4i64, 3 },
3784 { ISD::ADD, MVT::v8i32, 5 },
3785 { ISD::ADD, MVT::v16i16, 5 },
3786 { ISD::ADD, MVT::v32i8, 4 },
3787 };
3788
3789 int ISD = TLI->InstructionOpcodeToISD(Opcode);
3790 assert(ISD && "Invalid opcode")((void)0);
3791
3792 // Before legalizing the type, give a chance to look up illegal narrow types
3793 // in the table.
3794 // FIXME: Is there a better way to do this?
3795 EVT VT = TLI->getValueType(DL, ValTy);
3796 if (VT.isSimple()) {
3797 MVT MTy = VT.getSimpleVT();
3798 if (ST->isSLM())
3799 if (const auto *Entry = CostTableLookup(SLMCostTblNoPairWise, ISD, MTy))
3800 return Entry->Cost;
3801
3802 if (ST->hasAVX())
3803 if (const auto *Entry = CostTableLookup(AVX1CostTblNoPairWise, ISD, MTy))
3804 return Entry->Cost;
3805
3806 if (ST->hasSSE2())
3807 if (const auto *Entry = CostTableLookup(SSE2CostTblNoPairWise, ISD, MTy))
3808 return Entry->Cost;
3809 }
3810
3811 std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, ValTy);
3812
3813 MVT MTy = LT.second;
3814
3815 auto *ValVTy = cast<FixedVectorType>(ValTy);
3816
3817 // Special case: vXi8 mul reductions are performed as vXi16.
3818 if (ISD == ISD::MUL && MTy.getScalarType() == MVT::i8) {
3819 auto *WideSclTy = IntegerType::get(ValVTy->getContext(), 16);
3820 auto *WideVecTy = FixedVectorType::get(WideSclTy, ValVTy->getNumElements());
3821 return getCastInstrCost(Instruction::ZExt, WideVecTy, ValTy,
3822 TargetTransformInfo::CastContextHint::None,
3823 CostKind) +
3824 getArithmeticReductionCost(Opcode, WideVecTy, FMF, CostKind);
3825 }
3826
3827 InstructionCost ArithmeticCost = 0;
3828 if (LT.first != 1 && MTy.isVector() &&
3829 MTy.getVectorNumElements() < ValVTy->getNumElements()) {
3830 // Type needs to be split. We need LT.first - 1 arithmetic ops.
3831 auto *SingleOpTy = FixedVectorType::get(ValVTy->getElementType(),
3832 MTy.getVectorNumElements());
3833 ArithmeticCost = getArithmeticInstrCost(Opcode, SingleOpTy, CostKind);
3834 ArithmeticCost *= LT.first - 1;
3835 }
3836
3837 if (ST->isSLM())
3838 if (const auto *Entry = CostTableLookup(SLMCostTblNoPairWise, ISD, MTy))
3839 return ArithmeticCost + Entry->Cost;
3840
3841 if (ST->hasAVX())
3842 if (const auto *Entry = CostTableLookup(AVX1CostTblNoPairWise, ISD, MTy))
3843 return ArithmeticCost + Entry->Cost;
3844
3845 if (ST->hasSSE2())
3846 if (const auto *Entry = CostTableLookup(SSE2CostTblNoPairWise, ISD, MTy))
3847 return ArithmeticCost + Entry->Cost;
3848
3849 // FIXME: These assume a naive kshift+binop lowering, which is probably
3850 // conservative in most cases.
3851 static const CostTblEntry AVX512BoolReduction[] = {
3852 { ISD::AND, MVT::v2i1, 3 },
3853 { ISD::AND, MVT::v4i1, 5 },
3854 { ISD::AND, MVT::v8i1, 7 },
3855 { ISD::AND, MVT::v16i1, 9 },
3856 { ISD::AND, MVT::v32i1, 11 },
3857 { ISD::AND, MVT::v64i1, 13 },
3858 { ISD::OR, MVT::v2i1, 3 },
3859 { ISD::OR, MVT::v4i1, 5 },
3860 { ISD::OR, MVT::v8i1, 7 },
3861 { ISD::OR, MVT::v16i1, 9 },
3862 { ISD::OR, MVT::v32i1, 11 },
3863 { ISD::OR, MVT::v64i1, 13 },
3864 };
3865
3866 static const CostTblEntry AVX2BoolReduction[] = {
3867 { ISD::AND, MVT::v16i16, 2 }, // vpmovmskb + cmp
3868 { ISD::AND, MVT::v32i8, 2 }, // vpmovmskb + cmp
3869 { ISD::OR, MVT::v16i16, 2 }, // vpmovmskb + cmp
3870 { ISD::OR, MVT::v32i8, 2 }, // vpmovmskb + cmp
3871 };
3872
3873 static const CostTblEntry AVX1BoolReduction[] = {
3874 { ISD::AND, MVT::v4i64, 2 }, // vmovmskpd + cmp
3875 { ISD::AND, MVT::v8i32, 2 }, // vmovmskps + cmp
3876 { ISD::AND, MVT::v16i16, 4 }, // vextractf128 + vpand + vpmovmskb + cmp
3877 { ISD::AND, MVT::v32i8, 4 }, // vextractf128 + vpand + vpmovmskb + cmp
3878 { ISD::OR, MVT::v4i64, 2 }, // vmovmskpd + cmp
3879 { ISD::OR, MVT::v8i32, 2 }, // vmovmskps + cmp
3880 { ISD::OR, MVT::v16i16, 4 }, // vextractf128 + vpor + vpmovmskb + cmp
3881 { ISD::OR, MVT::v32i8, 4 }, // vextractf128 + vpor + vpmovmskb + cmp
3882 };
3883
3884 static const CostTblEntry SSE2BoolReduction[] = {
3885 { ISD::AND, MVT::v2i64, 2 }, // movmskpd + cmp
3886 { ISD::AND, MVT::v4i32, 2 }, // movmskps + cmp
3887 { ISD::AND, MVT::v8i16, 2 }, // pmovmskb + cmp
3888 { ISD::AND, MVT::v16i8, 2 }, // pmovmskb + cmp
3889 { ISD::OR, MVT::v2i64, 2 }, // movmskpd + cmp
3890 { ISD::OR, MVT::v4i32, 2 }, // movmskps + cmp
3891 { ISD::OR, MVT::v8i16, 2 }, // pmovmskb + cmp
3892 { ISD::OR, MVT::v16i8, 2 }, // pmovmskb + cmp
3893 };
3894
3895 // Handle bool allof/anyof patterns.
3896 if (ValVTy->getElementType()->isIntegerTy(1)) {
3897 InstructionCost ArithmeticCost = 0;
3898 if (LT.first != 1 && MTy.isVector() &&
3899 MTy.getVectorNumElements() < ValVTy->getNumElements()) {
3900 // Type needs to be split. We need LT.first - 1 arithmetic ops.
3901 auto *SingleOpTy = FixedVectorType::get(ValVTy->getElementType(),
3902 MTy.getVectorNumElements());
3903 ArithmeticCost = getArithmeticInstrCost(Opcode, SingleOpTy, CostKind);
3904 ArithmeticCost *= LT.first - 1;
3905 }
3906
3907 if (ST->hasAVX512())
3908 if (const auto *Entry = CostTableLookup(AVX512BoolReduction, ISD, MTy))
3909 return ArithmeticCost + Entry->Cost;
3910 if (ST->hasAVX2())
3911 if (const auto *Entry = CostTableLookup(AVX2BoolReduction, ISD, MTy))
3912 return ArithmeticCost + Entry->Cost;
3913 if (ST->hasAVX())
3914 if (const auto *Entry = CostTableLookup(AVX1BoolReduction, ISD, MTy))
3915 return ArithmeticCost + Entry->Cost;
3916 if (ST->hasSSE2())
3917 if (const auto *Entry = CostTableLookup(SSE2BoolReduction, ISD, MTy))
3918 return ArithmeticCost + Entry->Cost;
3919
3920 return BaseT::getArithmeticReductionCost(Opcode, ValVTy, FMF, CostKind);
3921 }
3922
3923 unsigned NumVecElts = ValVTy->getNumElements();
3924 unsigned ScalarSize = ValVTy->getScalarSizeInBits();
3925
3926 // Special case power of 2 reductions where the scalar type isn't changed
3927 // by type legalization.
3928 if (!isPowerOf2_32(NumVecElts) || ScalarSize != MTy.getScalarSizeInBits())
3929 return BaseT::getArithmeticReductionCost(Opcode, ValVTy, FMF, CostKind);
3930
3931 InstructionCost ReductionCost = 0;
3932
3933 auto *Ty = ValVTy;
3934 if (LT.first != 1 && MTy.isVector() &&
3935 MTy.getVectorNumElements() < ValVTy->getNumElements()) {
3936 // Type needs to be split. We need LT.first - 1 arithmetic ops.
3937 Ty = FixedVectorType::get(ValVTy->getElementType(),
3938 MTy.getVectorNumElements());
3939 ReductionCost = getArithmeticInstrCost(Opcode, Ty, CostKind);
3940 ReductionCost *= LT.first - 1;
3941 NumVecElts = MTy.getVectorNumElements();
3942 }
3943
3944 // Now handle reduction with the legal type, taking into account size changes
3945 // at each level.
3946 while (NumVecElts > 1) {
3947 // Determine the size of the remaining vector we need to reduce.
3948 unsigned Size = NumVecElts * ScalarSize;
3949 NumVecElts /= 2;
3950 // If we're reducing from 256/512 bits, use an extract_subvector.
3951 if (Size > 128) {
3952 auto *SubTy = FixedVectorType::get(ValVTy->getElementType(), NumVecElts);
3953 ReductionCost +=
3954 getShuffleCost(TTI::SK_ExtractSubvector, Ty, None, NumVecElts, SubTy);
3955 Ty = SubTy;
3956 } else if (Size == 128) {
3957 // Reducing from 128 bits is a permute of v2f64/v2i64.
3958 FixedVectorType *ShufTy;
3959 if (ValVTy->isFloatingPointTy())
3960 ShufTy =
3961 FixedVectorType::get(Type::getDoubleTy(ValVTy->getContext()), 2);
3962 else
3963 ShufTy =
3964 FixedVectorType::get(Type::getInt64Ty(ValVTy->getContext()), 2);
3965 ReductionCost +=
3966 getShuffleCost(TTI::SK_PermuteSingleSrc, ShufTy, None, 0, nullptr);
3967 } else if (Size == 64) {
3968 // Reducing from 64 bits is a shuffle of v4f32/v4i32.
3969 FixedVectorType *ShufTy;
3970 if (ValVTy->isFloatingPointTy())
3971 ShufTy =
3972 FixedVectorType::get(Type::getFloatTy(ValVTy->getContext()), 4);
3973 else
3974 ShufTy =
3975 FixedVectorType::get(Type::getInt32Ty(ValVTy->getContext()), 4);
3976 ReductionCost +=
3977 getShuffleCost(TTI::SK_PermuteSingleSrc, ShufTy, None, 0, nullptr);
3978 } else {
3979 // Reducing from smaller size is a shift by immediate.
3980 auto *ShiftTy = FixedVectorType::get(
3981 Type::getIntNTy(ValVTy->getContext(), Size), 128 / Size);
3982 ReductionCost += getArithmeticInstrCost(
3983 Instruction::LShr, ShiftTy, CostKind,
3984 TargetTransformInfo::OK_AnyValue,
3985 TargetTransformInfo::OK_UniformConstantValue,
3986 TargetTransformInfo::OP_None, TargetTransformInfo::OP_None);
3987 }
3988
3989 // Add the arithmetic op for this level.
3990 ReductionCost += getArithmeticInstrCost(Opcode, Ty, CostKind);
3991 }
3992
3993 // Add the final extract element to the cost.
3994 return ReductionCost + getVectorInstrCost(Instruction::ExtractElement, Ty, 0);
3995}
3996
3997InstructionCost X86TTIImpl::getMinMaxCost(Type *Ty, Type *CondTy,
3998 bool IsUnsigned) {
3999 std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, Ty);
4000
4001 MVT MTy = LT.second;
4002
4003 int ISD;
4004 if (Ty->isIntOrIntVectorTy()) {
4005 ISD = IsUnsigned ? ISD::UMIN : ISD::SMIN;
4006 } else {
4007 assert(Ty->isFPOrFPVectorTy() &&((void)0)
4008 "Expected float point or integer vector type.")((void)0);
4009 ISD = ISD::FMINNUM;
4010 }
4011
4012 static const CostTblEntry SSE1CostTbl[] = {
4013 {ISD::FMINNUM, MVT::v4f32, 1},
4014 };
4015
4016 static const CostTblEntry SSE2CostTbl[] = {
4017 {ISD::FMINNUM, MVT::v2f64, 1},
4018 {ISD::SMIN, MVT::v8i16, 1},
4019 {ISD::UMIN, MVT::v16i8, 1},
4020 };
4021
4022 static const CostTblEntry SSE41CostTbl[] = {
4023 {ISD::SMIN, MVT::v4i32, 1},
4024 {ISD::UMIN, MVT::v4i32, 1},
4025 {ISD::UMIN, MVT::v8i16, 1},
4026 {ISD::SMIN, MVT::v16i8, 1},
4027 };
4028
4029 static const CostTblEntry SSE42CostTbl[] = {
4030 {ISD::UMIN, MVT::v2i64, 3}, // xor+pcmpgtq+blendvpd
4031 };
4032
4033 static const CostTblEntry AVX1CostTbl[] = {
4034 {ISD::FMINNUM, MVT::v8f32, 1},
4035 {ISD::FMINNUM, MVT::v4f64, 1},
4036 {ISD::SMIN, MVT::v8i32, 3},
4037 {ISD::UMIN, MVT::v8i32, 3},
4038 {ISD::SMIN, MVT::v16i16, 3},
4039 {ISD::UMIN, MVT::v16i16, 3},
4040 {ISD::SMIN, MVT::v32i8, 3},
4041 {ISD::UMIN, MVT::v32i8, 3},
4042 };
4043
4044 static const CostTblEntry AVX2CostTbl[] = {
4045 {ISD::SMIN, MVT::v8i32, 1},
4046 {ISD::UMIN, MVT::v8i32, 1},
4047 {ISD::SMIN, MVT::v16i16, 1},
4048 {ISD::UMIN, MVT::v16i16, 1},
4049 {ISD::SMIN, MVT::v32i8, 1},
4050 {ISD::UMIN, MVT::v32i8, 1},
4051 };
4052
4053 static const CostTblEntry AVX512CostTbl[] = {
4054 {ISD::FMINNUM, MVT::v16f32, 1},
4055 {ISD::FMINNUM, MVT::v8f64, 1},
4056 {ISD::SMIN, MVT::v2i64, 1},
4057 {ISD::UMIN, MVT::v2i64, 1},
4058 {ISD::SMIN, MVT::v4i64, 1},
4059 {ISD::UMIN, MVT::v4i64, 1},
4060 {ISD::SMIN, MVT::v8i64, 1},
4061 {ISD::UMIN, MVT::v8i64, 1},
4062 {ISD::SMIN, MVT::v16i32, 1},
4063 {ISD::UMIN, MVT::v16i32, 1},
4064 };
4065
4066 static const CostTblEntry AVX512BWCostTbl[] = {
4067 {ISD::SMIN, MVT::v32i16, 1},
4068 {ISD::UMIN, MVT::v32i16, 1},
4069 {ISD::SMIN, MVT::v64i8, 1},
4070 {ISD::UMIN, MVT::v64i8, 1},
4071 };
4072
4073 // If we have a native MIN/MAX instruction for this type, use it.
4074 if (ST->hasBWI())
4075 if (const auto *Entry = CostTableLookup(AVX512BWCostTbl, ISD, MTy))
4076 return LT.first * Entry->Cost;
4077
4078 if (ST->hasAVX512())
4079 if (const auto *Entry = CostTableLookup(AVX512CostTbl, ISD, MTy))
4080 return LT.first * Entry->Cost;
4081
4082 if (ST->hasAVX2())
4083 if (const auto *Entry = CostTableLookup(AVX2CostTbl, ISD, MTy))
4084 return LT.first * Entry->Cost;
4085
4086 if (ST->hasAVX())
4087 if (const auto *Entry = CostTableLookup(AVX1CostTbl, ISD, MTy))
4088 return LT.first * Entry->Cost;
4089
4090 if (ST->hasSSE42())
4091 if (const auto *Entry = CostTableLookup(SSE42CostTbl, ISD, MTy))
4092 return LT.first * Entry->Cost;
4093
4094 if (ST->hasSSE41())
4095 if (const auto *Entry = CostTableLookup(SSE41CostTbl, ISD, MTy))
4096 return LT.first * Entry->Cost;
4097
4098 if (ST->hasSSE2())
4099 if (const auto *Entry = CostTableLookup(SSE2CostTbl, ISD, MTy))
4100 return LT.first * Entry->Cost;
4101
4102 if (ST->hasSSE1())
4103 if (const auto *Entry = CostTableLookup(SSE1CostTbl, ISD, MTy))
4104 return LT.first * Entry->Cost;
4105
4106 unsigned CmpOpcode;
4107 if (Ty->isFPOrFPVectorTy()) {
4108 CmpOpcode = Instruction::FCmp;
4109 } else {
4110 assert(Ty->isIntOrIntVectorTy() &&((void)0)
4111 "expecting floating point or integer type for min/max reduction")((void)0);
4112 CmpOpcode = Instruction::ICmp;
4113 }
4114
4115 TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput;
4116 // Otherwise fall back to cmp+select.
4117 InstructionCost Result =
4118 getCmpSelInstrCost(CmpOpcode, Ty, CondTy, CmpInst::BAD_ICMP_PREDICATE,
4119 CostKind) +
4120 getCmpSelInstrCost(Instruction::Select, Ty, CondTy,
4121 CmpInst::BAD_ICMP_PREDICATE, CostKind);
4122 return Result;
4123}
4124
4125InstructionCost
4126X86TTIImpl::getMinMaxReductionCost(VectorType *ValTy, VectorType *CondTy,
4127 bool IsUnsigned,
4128 TTI::TargetCostKind CostKind) {
4129 std::pair<InstructionCost, MVT> LT = TLI->getTypeLegalizationCost(DL, ValTy);
4130
4131 MVT MTy = LT.second;
4132
4133 int ISD;
4134 if (ValTy->isIntOrIntVectorTy()) {
4135 ISD = IsUnsigned ? ISD::UMIN : ISD::SMIN;
4136 } else {
4137 assert(ValTy->isFPOrFPVectorTy() &&((void)0)
4138 "Expected float point or integer vector type.")((void)0);
4139 ISD = ISD::FMINNUM;
4140 }
4141
4142 // We use the Intel Architecture Code Analyzer(IACA) to measure the throughput
4143 // and make it as the cost.
4144
4145 static const CostTblEntry SSE2CostTblNoPairWise[] = {
4146 {ISD::UMIN, MVT::v2i16, 5}, // need pxors to use pminsw/pmaxsw
4147 {ISD::UMIN, MVT::v4i16, 7}, // need pxors to use pminsw/pmaxsw
4148 {ISD::UMIN, MVT::v8i16, 9}, // need pxors to use pminsw/pmaxsw
4149 };
4150
4151 static const CostTblEntry SSE41CostTblNoPairWise[] = {
4152 {ISD::SMIN, MVT::v2i16, 3}, // same as sse2
4153 {ISD::SMIN, MVT::v4i16, 5}, // same as sse2
4154 {ISD::UMIN, MVT::v2i16, 5}, // same as sse2
4155 {ISD::UMIN, MVT::v4i16, 7}, // same as sse2
4156 {ISD::SMIN, MVT::v8i16, 4}, // phminposuw+xor
4157 {ISD::UMIN, MVT::v8i16, 4}, // FIXME: umin is cheaper than umax
4158 {ISD::SMIN, MVT::v2i8, 3}, // pminsb
4159 {ISD::SMIN, MVT::v4i8, 5}, // pminsb
4160 {ISD::SMIN, MVT::v8i8, 7}, // pminsb
4161 {ISD::SMIN, MVT::v16i8, 6},
4162 {ISD::UMIN, MVT::v2i8, 3}, // same as sse2
4163 {ISD::UMIN, MVT::v4i8, 5}, // same as sse2
4164 {ISD::UMIN, MVT::v8i8, 7}, // same as sse2
4165 {ISD::UMIN, MVT::v16i8, 6}, // FIXME: umin is cheaper than umax
4166 };
4167
4168 static const CostTblEntry AVX1CostTblNoPairWise[] = {
4169 {ISD::SMIN, MVT::v16i16, 6},
4170 {ISD::UMIN, MVT::v16i16, 6}, // FIXME: umin is cheaper than umax
4171 {ISD::SMIN, MVT::v32i8, 8},
4172 {ISD::UMIN, MVT::v32i8, 8},
4173 };
4174
4175 static const CostTblEntry AVX512BWCostTblNoPairWise[] = {
4176 {ISD::SMIN, MVT::v32i16, 8},
4177 {ISD::UMIN, MVT::v32i16, 8}, // FIXME: umin is cheaper than umax
4178 {ISD::SMIN, MVT::v64i8, 10},
4179 {ISD::UMIN, MVT::v64i8, 10},
4180 };
4181
4182 // Before legalizing the type, give a chance to look up illegal narrow types
4183 // in the table.
4184 // FIXME: Is there a better way to do this?
4185 EVT VT = TLI->getValueType(DL, ValTy);
4186 if (VT.isSimple()) {
4187 MVT MTy = VT.getSimpleVT();
4188 if (ST->hasBWI())
4189 if (const auto *Entry = CostTableLookup(AVX512BWCostTblNoPairWise, ISD, MTy))
4190 return Entry->Cost;
4191
4192 if (ST->hasAVX())
4193 if (const auto *Entry = CostTableLookup(AVX1CostTblNoPairWise, ISD, MTy))
4194 return Entry->Cost;
4195
4196 if (ST->hasSSE41())
4197 if (const auto *Entry = CostTableLookup(SSE41CostTblNoPairWise, ISD, MTy))
4198 return Entry->Cost;
4199
4200 if (ST->hasSSE2())
4201 if (const auto *Entry = CostTableLookup(SSE2CostTblNoPairWise, ISD, MTy))
4202 return Entry->Cost;
4203 }
4204
4205 auto *ValVTy = cast<FixedVectorType>(ValTy);
4206 unsigned NumVecElts = ValVTy->getNumElements();
4207
4208 auto *Ty = ValVTy;
4209 InstructionCost MinMaxCost = 0;
4210 if (LT.first != 1 && MTy.isVector() &&
4211 MTy.getVectorNumElements() < ValVTy->getNumElements()) {
4212 // Type needs to be split. We need LT.first - 1 operations ops.
4213 Ty = FixedVectorType::get(ValVTy->getElementType(),
4214 MTy.getVectorNumElements());
4215 auto *SubCondTy = FixedVectorType::get(CondTy->getElementType(),
4216 MTy.getVectorNumElements());
4217 MinMaxCost = getMinMaxCost(Ty, SubCondTy, IsUnsigned);
4218 MinMaxCost *= LT.first - 1;
4219 NumVecElts = MTy.getVectorNumElements();
4220 }
4221
4222 if (ST->hasBWI())
4223 if (const auto *Entry = CostTableLookup(AVX512BWCostTblNoPairWise, ISD, MTy))
4224 return MinMaxCost + Entry->Cost;
4225
4226 if (ST->hasAVX())
4227 if (const auto *Entry = CostTableLookup(AVX1CostTblNoPairWise, ISD, MTy))
4228 return MinMaxCost + Entry->Cost;
4229
4230 if (ST->hasSSE41())
4231 if (const auto *Entry = CostTableLookup(SSE41CostTblNoPairWise, ISD, MTy))
4232 return MinMaxCost + Entry->Cost;
4233
4234 if (ST->hasSSE2())
4235 if (const auto *Entry = CostTableLookup(SSE2CostTblNoPairWise, ISD, MTy))
4236 return MinMaxCost + Entry->Cost;
4237
4238 unsigned ScalarSize = ValTy->getScalarSizeInBits();
4239
4240 // Special case power of 2 reductions where the scalar type isn't changed
4241 // by type legalization.
4242 if (!isPowerOf2_32(ValVTy->getNumElements()) ||
4243 ScalarSize != MTy.getScalarSizeInBits())
4244 return BaseT::getMinMaxReductionCost(ValTy, CondTy, IsUnsigned, CostKind);
4245
4246 // Now handle reduction with the legal type, taking into account size changes
4247 // at each level.
4248 while (NumVecElts > 1) {
4249 // Determine the size of the remaining vector we need to reduce.
4250 unsigned Size = NumVecElts * ScalarSize;
4251 NumVecElts /= 2;
4252 // If we're reducing from 256/512 bits, use an extract_subvector.
4253 if (Size > 128) {
4254 auto *SubTy = FixedVectorType::get(ValVTy->getElementType(), NumVecElts);
4255 MinMaxCost +=
4256 getShuffleCost(TTI::SK_ExtractSubvector, Ty, None, NumVecElts, SubTy);
4257 Ty = SubTy;
4258 } else if (Size == 128) {
4259 // Reducing from 128 bits is a permute of v2f64/v2i64.
4260 VectorType *ShufTy;
4261 if (ValTy->isFloatingPointTy())
4262 ShufTy =
4263 FixedVectorType::get(Type::getDoubleTy(ValTy->getContext()), 2);
4264 else
4265 ShufTy = FixedVectorType::get(Type::getInt64Ty(ValTy->getContext()), 2);
4266 MinMaxCost +=
4267 getShuffleCost(TTI::SK_PermuteSingleSrc, ShufTy, None, 0, nullptr);
4268 } else if (Size == 64) {
4269 // Reducing from 64 bits is a shuffle of v4f32/v4i32.
4270 FixedVectorType *ShufTy;
4271 if (ValTy->isFloatingPointTy())
4272 ShufTy = FixedVectorType::get(Type::getFloatTy(ValTy->getContext()), 4);
4273 else
4274 ShufTy = FixedVectorType::get(Type::getInt32Ty(ValTy->getContext()), 4);
4275 MinMaxCost +=
4276 getShuffleCost(TTI::SK_PermuteSingleSrc, ShufTy, None, 0, nullptr);
4277 } else {
4278 // Reducing from smaller size is a shift by immediate.
4279 auto *ShiftTy = FixedVectorType::get(
4280 Type::getIntNTy(ValTy->getContext(), Size), 128 / Size);
4281 MinMaxCost += getArithmeticInstrCost(
4282 Instruction::LShr, ShiftTy, TTI::TCK_RecipThroughput,
4283 TargetTransformInfo::OK_AnyValue,
4284 TargetTransformInfo::OK_UniformConstantValue,
4285 TargetTransformInfo::OP_None, TargetTransformInfo::OP_None);
4286 }
4287
4288 // Add the arithmetic op for this level.
4289 auto *SubCondTy =
4290 FixedVectorType::get(CondTy->getElementType(), Ty->getNumElements());
4291 MinMaxCost += getMinMaxCost(Ty, SubCondTy, IsUnsigned);
4292 }
4293
4294 // Add the final extract element to the cost.
4295 return MinMaxCost + getVectorInstrCost(Instruction::ExtractElement, Ty, 0);
4296}
4297
4298/// Calculate the cost of materializing a 64-bit value. This helper
4299/// method might only calculate a fraction of a larger immediate. Therefore it
4300/// is valid to return a cost of ZERO.
4301InstructionCost X86TTIImpl::getIntImmCost(int64_t Val) {
4302 if (Val == 0)
4303 return TTI::TCC_Free;
4304
4305 if (isInt<32>(Val))
4306 return TTI::TCC_Basic;
4307
4308 return 2 * TTI::TCC_Basic;
4309}
4310
4311InstructionCost X86TTIImpl::getIntImmCost(const APInt &Imm, Type *Ty,
4312 TTI::TargetCostKind CostKind) {
4313 assert(Ty->isIntegerTy())((void)0);
4314
4315 unsigned BitSize = Ty->getPrimitiveSizeInBits();
4316 if (BitSize == 0)
4317 return ~0U;
4318
4319 // Never hoist constants larger than 128bit, because this might lead to
4320 // incorrect code generation or assertions in codegen.
4321 // Fixme: Create a cost model for types larger than i128 once the codegen
4322 // issues have been fixed.
4323 if (BitSize > 128)
4324 return TTI::TCC_Free;
4325
4326 if (Imm == 0)
4327 return TTI::TCC_Free;
4328
4329 // Sign-extend all constants to a multiple of 64-bit.
4330 APInt ImmVal = Imm;
4331 if (BitSize % 64 != 0)
4332 ImmVal = Imm.sext(alignTo(BitSize, 64));
4333
4334 // Split the constant into 64-bit chunks and calculate the cost for each
4335 // chunk.
4336 InstructionCost Cost = 0;
4337 for (unsigned ShiftVal = 0; ShiftVal < BitSize; ShiftVal += 64) {
4338 APInt Tmp = ImmVal.ashr(ShiftVal).sextOrTrunc(64);
4339 int64_t Val = Tmp.getSExtValue();
4340 Cost += getIntImmCost(Val);
4341 }
4342 // We need at least one instruction to materialize the constant.
4343 return std::max<InstructionCost>(1, Cost);
4344}
4345
4346InstructionCost X86TTIImpl::getIntImmCostInst(unsigned Opcode, unsigned Idx,
4347 const APInt &Imm, Type *Ty,
4348 TTI::TargetCostKind CostKind,
4349 Instruction *Inst) {
4350 assert(Ty->isIntegerTy())((void)0);
4351
4352 unsigned BitSize = Ty->getPrimitiveSizeInBits();
4353 // There is no cost model for constants with a bit size of 0. Return TCC_Free
4354 // here, so that constant hoisting will ignore this constant.
4355 if (BitSize == 0)
4356 return TTI::TCC_Free;
4357
4358 unsigned ImmIdx = ~0U;
4359 switch (Opcode) {
4360 default:
4361 return TTI::TCC_Free;
4362 case Instruction::GetElementPtr:
4363 // Always hoist the base address of a GetElementPtr. This prevents the
4364 // creation of new constants for every base constant that gets constant
4365 // folded with the offset.
4366 if (Idx == 0)
4367 return 2 * TTI::TCC_Basic;
4368 return TTI::TCC_Free;
4369 case Instruction::Store:
4370 ImmIdx = 0;
4371 break;
4372 case Instruction::ICmp:
4373 // This is an imperfect hack to prevent constant hoisting of
4374 // compares that might be trying to check if a 64-bit value fits in
4375 // 32-bits. The backend can optimize these cases using a right shift by 32.
4376 // Ideally we would check the compare predicate here. There also other
4377 // similar immediates the backend can use shifts for.
4378 if (Idx == 1 && Imm.getBitWidth() == 64) {
4379 uint64_t ImmVal = Imm.getZExtValue();
4380 if (ImmVal == 0x100000000ULL || ImmVal == 0xffffffff)
4381 return TTI::TCC_Free;
4382 }
4383 ImmIdx = 1;
4384 break;
4385 case Instruction::And:
4386 // We support 64-bit ANDs with immediates with 32-bits of leading zeroes
4387 // by using a 32-bit operation with implicit zero extension. Detect such
4388 // immediates here as the normal path expects bit 31 to be sign extended.
4389 if (Idx == 1 && Imm.getBitWidth() == 64 && isUInt<32>(Imm.getZExtValue()))
4390 return TTI::TCC_Free;
4391 ImmIdx = 1;
4392 break;
4393 case Instruction::Add:
4394 case Instruction::Sub:
4395 // For add/sub, we can use the opposite instruction for INT32_MIN.
4396 if (Idx == 1 && Imm.getBitWidth() == 64 && Imm.getZExtValue() == 0x80000000)
4397 return TTI::TCC_Free;
4398 ImmIdx = 1;
4399 break;
4400 case Instruction::UDiv:
4401 case Instruction::SDiv:
4402 case Instruction::URem:
4403 case Instruction::SRem:
4404 // Division by constant is typically expanded later into a different
4405 // instruction sequence. This completely changes the constants.
4406 // Report them as "free" to stop ConstantHoist from marking them as opaque.
4407 return TTI::TCC_Free;
4408 case Instruction::Mul:
4409 case Instruction::Or:
4410 case Instruction::Xor:
4411 ImmIdx = 1;
4412 break;
4413 // Always return TCC_Free for the shift value of a shift instruction.
4414 case Instruction::Shl:
4415 case Instruction::LShr:
4416 case Instruction::AShr:
4417 if (Idx == 1)
4418 return TTI::TCC_Free;
4419 break;
4420 case Instruction::Trunc:
4421 case Instruction::ZExt:
4422 case Instruction::SExt:
4423 case Instruction::IntToPtr:
4424 case Instruction::PtrToInt:
4425 case Instruction::BitCast:
4426 case Instruction::PHI:
4427 case Instruction::Call:
4428 case Instruction::Select:
4429 case Instruction::Ret:
4430 case Instruction::Load:
4431 break;
4432 }
4433
4434 if (Idx == ImmIdx) {
4435 int NumConstants = divideCeil(BitSize, 64);
4436 InstructionCost Cost = X86TTIImpl::getIntImmCost(Imm, Ty, CostKind);
4437 return (Cost <= NumConstants * TTI::TCC_Basic)
4438 ? static_cast<int>(TTI::TCC_Free)
4439 : Cost;
4440 }
4441
4442 return X86TTIImpl::getIntImmCost(Imm, Ty, CostKind);
4443}
4444
4445InstructionCost X86TTIImpl::getIntImmCostIntrin(Intrinsic::ID IID, unsigned Idx,
4446 const APInt &Imm, Type *Ty,
4447 TTI::TargetCostKind CostKind) {
4448 assert(Ty->isIntegerTy())((void)0);
4449
4450 unsigned BitSize = Ty->getPrimitiveSizeInBits();
4451 // There is no cost model for constants with a bit size of 0. Return TCC_Free
4452 // here, so that constant hoisting will ignore this constant.
4453 if (BitSize == 0)
4454 return TTI::TCC_Free;
4455
4456 switch (IID) {
4457 default:
4458 return TTI::TCC_Free;
4459 case Intrinsic::sadd_with_overflow:
4460 case Intrinsic::uadd_with_overflow:
4461 case Intrinsic::ssub_with_overflow:
4462 case Intrinsic::usub_with_overflow:
4463 case Intrinsic::smul_with_overflow:
4464 case Intrinsic::umul_with_overflow:
4465 if ((Idx == 1) && Imm.getBitWidth() <= 64 && isInt<32>(Imm.getSExtValue()))
4466 return TTI::TCC_Free;
4467 break;
4468 case Intrinsic::experimental_stackmap:
4469 if ((Idx < 2) || (Imm.getBitWidth() <= 64 && isInt<64>(Imm.getSExtValue())))
4470 return TTI::TCC_Free;
4471 break;
4472 case Intrinsic::experimental_patchpoint_void:
4473 case Intrinsic::experimental_patchpoint_i64:
4474 if ((Idx < 4) || (Imm.getBitWidth() <= 64 && isInt<64>(Imm.getSExtValue())))
4475 return TTI::TCC_Free;
4476 break;
4477 }
4478 return X86TTIImpl::getIntImmCost(Imm, Ty, CostKind);
4479}
4480
4481InstructionCost X86TTIImpl::getCFInstrCost(unsigned Opcode,
4482 TTI::TargetCostKind CostKind,
4483 const Instruction *I) {
4484 if (CostKind != TTI::TCK_RecipThroughput)
4485 return Opcode == Instruction::PHI ? 0 : 1;
4486 // Branches are assumed to be predicted.
4487 return 0;
4488}
4489
4490int X86TTIImpl::getGatherOverhead() const {
4491 // Some CPUs have more overhead for gather. The specified overhead is relative
4492 // to the Load operation. "2" is the number provided by Intel architects. This
4493 // parameter is used for cost estimation of Gather Op and comparison with
4494 // other alternatives.
4495 // TODO: Remove the explicit hasAVX512()?, That would mean we would only
4496 // enable gather with a -march.
4497 if (ST->hasAVX512() || (ST->hasAVX2() && ST->hasFastGather()))
4498 return 2;
4499
4500 return 1024;
4501}
4502
4503int X86TTIImpl::getScatterOverhead() const {
4504 if (ST->hasAVX512())
4505 return 2;
4506
4507 return 1024;
4508}
4509
4510// Return an average cost of Gather / Scatter instruction, maybe improved later.
4511// FIXME: Add TargetCostKind support.
4512InstructionCost X86TTIImpl::getGSVectorCost(unsigned Opcode, Type *SrcVTy,
4513 const Value *Ptr, Align Alignment,
4514 unsigned AddressSpace) {
4515
4516 assert(isa<VectorType>(SrcVTy) && "Unexpected type in getGSVectorCost")((void)0);
4517 unsigned VF = cast<FixedVectorType>(SrcVTy)->getNumElements();
4518
4519 // Try to reduce index size from 64 bit (default for GEP)
4520 // to 32. It is essential for VF 16. If the index can't be reduced to 32, the
4521 // operation will use 16 x 64 indices which do not fit in a zmm and needs
4522 // to split. Also check that the base pointer is the same for all lanes,
4523 // and that there's at most one variable index.
4524 auto getIndexSizeInBits = [](const Value *Ptr, const DataLayout &DL) {
4525 unsigned IndexSize = DL.getPointerSizeInBits();
4526 const GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr);
4527 if (IndexSize < 64 || !GEP)
4528 return IndexSize;
4529
4530 unsigned NumOfVarIndices = 0;
4531 const Value *Ptrs = GEP->getPointerOperand();
4532 if (Ptrs->getType()->isVectorTy() && !getSplatValue(Ptrs))
4533 return IndexSize;
4534 for (unsigned i = 1; i < GEP->getNumOperands(); ++i) {
4535 if (isa<Constant>(GEP->getOperand(i)))
4536 continue;
4537 Type *IndxTy = GEP->getOperand(i)->getType();
4538 if (auto *IndexVTy = dyn_cast<VectorType>(IndxTy))
4539 IndxTy = IndexVTy->getElementType();
4540 if ((IndxTy->getPrimitiveSizeInBits() == 64 &&
4541 !isa<SExtInst>(GEP->getOperand(i))) ||
4542 ++NumOfVarIndices > 1)
4543 return IndexSize; // 64
4544 }
4545 return (unsigned)32;
4546 };
4547
4548 // Trying to reduce IndexSize to 32 bits for vector 16.
4549 // By default the IndexSize is equal to pointer size.
4550 unsigned IndexSize = (ST->hasAVX512() && VF >= 16)
4551 ? getIndexSizeInBits(Ptr, DL)
4552 : DL.getPointerSizeInBits();
4553
4554 auto *IndexVTy = FixedVectorType::get(
4555 IntegerType::get(SrcVTy->getContext(), IndexSize), VF);
4556 std::pair<InstructionCost, MVT> IdxsLT =
4557 TLI->getTypeLegalizationCost(DL, IndexVTy);
4558 std::pair<InstructionCost, MVT> SrcLT =
4559 TLI->getTypeLegalizationCost(DL, SrcVTy);
4560 InstructionCost::CostType SplitFactor =
4561 *std::max(IdxsLT.first, SrcLT.first).getValue();
4562 if (SplitFactor > 1) {
4563 // Handle splitting of vector of pointers
4564 auto *SplitSrcTy =
4565 FixedVectorType::get(SrcVTy->getScalarType(), VF / SplitFactor);
4566 return SplitFactor * getGSVectorCost(Opcode, SplitSrcTy, Ptr, Alignment,
4567 AddressSpace);
4568 }
4569
4570 // The gather / scatter cost is given by Intel architects. It is a rough
4571 // number since we are looking at one instruction in a time.
4572 const int GSOverhead = (Opcode == Instruction::Load)
4573 ? getGatherOverhead()
4574 : getScatterOverhead();
4575 return GSOverhead + VF * getMemoryOpCost(Opcode, SrcVTy->getScalarType(),
4576 MaybeAlign(Alignment), AddressSpace,
4577 TTI::TCK_RecipThroughput);
4578}
4579
4580/// Return the cost of full scalarization of gather / scatter operation.
4581///
4582/// Opcode - Load or Store instruction.
4583/// SrcVTy - The type of the data vector that should be gathered or scattered.
4584/// VariableMask - The mask is non-constant at compile time.
4585/// Alignment - Alignment for one element.
4586/// AddressSpace - pointer[s] address space.
4587///
4588/// FIXME: Add TargetCostKind support.
4589InstructionCost X86TTIImpl::getGSScalarCost(unsigned Opcode, Type *SrcVTy,
4590 bool VariableMask, Align Alignment,
4591 unsigned AddressSpace) {
4592 unsigned VF = cast<FixedVectorType>(SrcVTy)->getNumElements();
4593 APInt DemandedElts = APInt::getAllOnesValue(VF);
4594 TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput;
4595
4596 InstructionCost MaskUnpackCost = 0;
4597 if (VariableMask) {
4598 auto *MaskTy =
4599 FixedVectorType::get(Type::getInt1Ty(SrcVTy->getContext()), VF);
4600 MaskUnpackCost =
4601 getScalarizationOverhead(MaskTy, DemandedElts, false, true);
4602 InstructionCost ScalarCompareCost = getCmpSelInstrCost(
4603 Instruction::ICmp, Type::getInt1Ty(SrcVTy->getContext()), nullptr,
4604 CmpInst::BAD_ICMP_PREDICATE, CostKind);
4605 InstructionCost BranchCost = getCFInstrCost(Instruction::Br, CostKind);
4606 MaskUnpackCost += VF * (BranchCost + ScalarCompareCost);
4607 }
4608
4609 // The cost of the scalar loads/stores.
4610 InstructionCost MemoryOpCost =
4611 VF * getMemoryOpCost(Opcode, SrcVTy->getScalarType(),
4612 MaybeAlign(Alignment), AddressSpace, CostKind);
4613
4614 InstructionCost InsertExtractCost = 0;
4615 if (Opcode == Instruction::Load)
4616 for (unsigned i = 0; i < VF; ++i)
4617 // Add the cost of inserting each scalar load into the vector
4618 InsertExtractCost +=
4619 getVectorInstrCost(Instruction::InsertElement, SrcVTy, i);
4620 else
4621 for (unsigned i = 0; i < VF; ++i)
4622 // Add the cost of extracting each element out of the data vector
4623 InsertExtractCost +=
4624 getVectorInstrCost(Instruction::ExtractElement, SrcVTy, i);
4625
4626 return MemoryOpCost + MaskUnpackCost + InsertExtractCost;
4627}
4628
4629/// Calculate the cost of Gather / Scatter operation
4630InstructionCost X86TTIImpl::getGatherScatterOpCost(
4631 unsigned Opcode, Type *SrcVTy, const Value *Ptr, bool VariableMask,
4632 Align Alignment, TTI::TargetCostKind CostKind,
4633 const Instruction *I = nullptr) {
4634 if (CostKind != TTI::TCK_RecipThroughput) {
1
Assuming 'CostKind' is equal to TCK_RecipThroughput
→
2
←
Taking false branch
→
4635 if ((Opcode == Instruction::Load &&
4636 isLegalMaskedGather(SrcVTy, Align(Alignment))) ||
4637 (Opcode == Instruction::Store &&
4638 isLegalMaskedScatter(SrcVTy, Align(Alignment))))
4639 return 1;
4640 return BaseT::getGatherScatterOpCost(Opcode, SrcVTy, Ptr, VariableMask,
4641 Alignment, CostKind, I);
4642 }
4643
4644 assert(SrcVTy->isVectorTy() && "Unexpected data type for Gather/Scatter")((void)0);
4645 PointerType *PtrTy = dyn_cast<PointerType>(Ptr->getType());
3
←
Assuming the object is not a 'PointerType'
→
4
←
'PtrTy' initialized to a null pointer value
→
4646 if (!PtrTy
4.1
'PtrTy' is null
 && Ptr->getType()->isVectorTy())
5
←
Taking false branch
→
4647 PtrTy = dyn_cast<PointerType>(
4648 cast<VectorType>(Ptr->getType())->getElementType());
4649 assert(PtrTy && "Unexpected type for Ptr argument")((void)0);
4650 unsigned AddressSpace = PtrTy->getAddressSpace();
6
←
Called C++ object pointer is null
4651
4652 if ((Opcode == Instruction::Load &&
4653 !isLegalMaskedGather(SrcVTy, Align(Alignment))) ||
4654 (Opcode == Instruction::Store &&
4655 !isLegalMaskedScatter(SrcVTy, Align(Alignment))))
4656 return getGSScalarCost(Opcode, SrcVTy, VariableMask, Alignment,
4657 AddressSpace);
4658
4659 return getGSVectorCost(Opcode, SrcVTy, Ptr, Alignment, AddressSpace);
4660}
4661
4662bool X86TTIImpl::isLSRCostLess(TargetTransformInfo::LSRCost &C1,
4663 TargetTransformInfo::LSRCost &C2) {
4664 // X86 specific here are "instruction number 1st priority".
4665 return std::tie(C1.Insns, C1.NumRegs, C1.AddRecCost,
4666 C1.NumIVMuls, C1.NumBaseAdds,
4667 C1.ScaleCost, C1.ImmCost, C1.SetupCost) <
4668 std::tie(C2.Insns, C2.NumRegs, C2.AddRecCost,
4669 C2.NumIVMuls, C2.NumBaseAdds,
4670 C2.ScaleCost, C2.ImmCost, C2.SetupCost);
4671}
4672
4673bool X86TTIImpl::canMacroFuseCmp() {
4674 return ST->hasMacroFusion() || ST->hasBranchFusion();
4675}
4676
4677bool X86TTIImpl::isLegalMaskedLoad(Type *DataTy, Align Alignment) {
4678 if (!ST->hasAVX())
4679 return false;
4680
4681 // The backend can't handle a single element vector.
4682 if (isa<VectorType>(DataTy) &&
4683 cast<FixedVectorType>(DataTy)->getNumElements() == 1)
4684 return false;
4685 Type *ScalarTy = DataTy->getScalarType();
4686
4687 if (ScalarTy->isPointerTy())
4688 return true;
4689
4690 if (ScalarTy->isFloatTy() || ScalarTy->isDoubleTy())
4691 return true;
4692
4693 if (!ScalarTy->isIntegerTy())
4694 return false;
4695
4696 unsigned IntWidth = ScalarTy->getIntegerBitWidth();
4697 return IntWidth == 32 || IntWidth == 64 ||
4698 ((IntWidth == 8 || IntWidth == 16) && ST->hasBWI());
4699}
4700
4701bool X86TTIImpl::isLegalMaskedStore(Type *DataType, Align Alignment) {
4702 return isLegalMaskedLoad(DataType, Alignment);
4703}
4704
4705bool X86TTIImpl::isLegalNTLoad(Type *DataType, Align Alignment) {
4706 unsigned DataSize = DL.getTypeStoreSize(DataType);
4707 // The only supported nontemporal loads are for aligned vectors of 16 or 32
4708 // bytes. Note that 32-byte nontemporal vector loads are supported by AVX2
4709 // (the equivalent stores only require AVX).
4710 if (Alignment >= DataSize && (DataSize == 16 || DataSize == 32))
4711 return DataSize == 16 ? ST->hasSSE1() : ST->hasAVX2();
4712
4713 return false;
4714}
4715
4716bool X86TTIImpl::isLegalNTStore(Type *DataType, Align Alignment) {
4717 unsigned DataSize = DL.getTypeStoreSize(DataType);
4718
4719 // SSE4A supports nontemporal stores of float and double at arbitrary
4720 // alignment.
4721 if (ST->hasSSE4A() && (DataType->isFloatTy() || DataType->isDoubleTy()))
4722 return true;
4723
4724 // Besides the SSE4A subtarget exception above, only aligned stores are
4725 // available nontemporaly on any other subtarget. And only stores with a size
4726 // of 4..32 bytes (powers of 2, only) are permitted.
4727 if (Alignment < DataSize || DataSize < 4 || DataSize > 32 ||
4728 !isPowerOf2_32(DataSize))
4729 return false;
4730
4731 // 32-byte vector nontemporal stores are supported by AVX (the equivalent
4732 // loads require AVX2).
4733 if (DataSize == 32)
4734 return ST->hasAVX();
4735 else if (DataSize == 16)
4736 return ST->hasSSE1();
4737 return true;
4738}
4739
4740bool X86TTIImpl::isLegalMaskedExpandLoad(Type *DataTy) {
4741 if (!isa<VectorType>(DataTy))
4742 return false;
4743
4744 if (!ST->hasAVX512())
4745 return false;
4746
4747 // The backend can't handle a single element vector.
4748 if (cast<FixedVectorType>(DataTy)->getNumElements() == 1)
4749 return false;
4750
4751 Type *ScalarTy = cast<VectorType>(DataTy)->getElementType();
4752
4753 if (ScalarTy->isFloatTy() || ScalarTy->isDoubleTy())
4754 return true;
4755
4756 if (!ScalarTy->isIntegerTy())
4757 return false;
4758
4759 unsigned IntWidth = ScalarTy->getIntegerBitWidth();
4760 return IntWidth == 32 || IntWidth == 64 ||
4761 ((IntWidth == 8 || IntWidth == 16) && ST->hasVBMI2());
4762}
4763
4764bool X86TTIImpl::isLegalMaskedCompressStore(Type *DataTy) {
4765 return isLegalMaskedExpandLoad(DataTy);
4766}
4767
4768bool X86TTIImpl::isLegalMaskedGather(Type *DataTy, Align Alignment) {
4769 // Some CPUs have better gather performance than others.
4770 // TODO: Remove the explicit ST->hasAVX512()?, That would mean we would only
4771 // enable gather with a -march.
4772 if (!(ST->hasAVX512() || (ST->hasFastGather() && ST->hasAVX2())))
4773 return false;
4774
4775 // This function is called now in two cases: from the Loop Vectorizer
4776 // and from the Scalarizer.
4777 // When the Loop Vectorizer asks about legality of the feature,
4778 // the vectorization factor is not calculated yet. The Loop Vectorizer
4779 // sends a scalar type and the decision is based on the width of the
4780 // scalar element.
4781 // Later on, the cost model will estimate usage this intrinsic based on
4782 // the vector type.
4783 // The Scalarizer asks again about legality. It sends a vector type.
4784 // In this case we can reject non-power-of-2 vectors.
4785 // We also reject single element vectors as the type legalizer can't
4786 // scalarize it.
4787 if (auto *DataVTy = dyn_cast<FixedVectorType>(DataTy)) {
4788 unsigned NumElts = DataVTy->getNumElements();
4789 if (NumElts == 1)
4790 return false;
4791 // Gather / Scatter for vector 2 is not profitable on KNL / SKX
4792 // Vector-4 of gather/scatter instruction does not exist on KNL.
4793 // We can extend it to 8 elements, but zeroing upper bits of
4794 // the mask vector will add more instructions. Right now we give the scalar
4795 // cost of vector-4 for KNL. TODO: Check, maybe the gather/scatter
4796 // instruction is better in the VariableMask case.
4797 if (ST->hasAVX512() && (NumElts == 2 || (NumElts == 4 && !ST->hasVLX())))
4798 return false;
4799 }
4800 Type *ScalarTy = DataTy->getScalarType();
4801 if (ScalarTy->isPointerTy())
4802 return true;
4803
4804 if (ScalarTy->isFloatTy() || ScalarTy->isDoubleTy())
4805 return true;
4806
4807 if (!ScalarTy->isIntegerTy())
4808 return false;
4809
4810 unsigned IntWidth = ScalarTy->getIntegerBitWidth();
4811 return IntWidth == 32 || IntWidth == 64;
4812}
4813
4814bool X86TTIImpl::isLegalMaskedScatter(Type *DataType, Align Alignment) {
4815 // AVX2 doesn't support scatter
4816 if (!ST->hasAVX512())
4817 return false;
4818 return isLegalMaskedGather(DataType, Alignment);
4819}
4820
4821bool X86TTIImpl::hasDivRemOp(Type *DataType, bool IsSigned) {
4822 EVT VT = TLI->getValueType(DL, DataType);
4823 return TLI->isOperationLegal(IsSigned ? ISD::SDIVREM : ISD::UDIVREM, VT);
4824}
4825
4826bool X86TTIImpl::isFCmpOrdCheaperThanFCmpZero(Type *Ty) {
4827 return false;
4828}
4829
4830bool X86TTIImpl::areInlineCompatible(const Function *Caller,
4831 const Function *Callee) const {
4832 const TargetMachine &TM = getTLI()->getTargetMachine();
4833
4834 // Work this as a subsetting of subtarget features.
4835 const FeatureBitset &CallerBits =
4836 TM.getSubtargetImpl(*Caller)->getFeatureBits();
4837 const FeatureBitset &CalleeBits =
4838 TM.getSubtargetImpl(*Callee)->getFeatureBits();
4839
4840 FeatureBitset RealCallerBits = CallerBits & ~InlineFeatureIgnoreList;
4841 FeatureBitset RealCalleeBits = CalleeBits & ~InlineFeatureIgnoreList;
4842 return (RealCallerBits & RealCalleeBits) == RealCalleeBits;
4843}
4844
4845bool X86TTIImpl::areFunctionArgsABICompatible(
4846 const Function *Caller, const Function *Callee,
4847 SmallPtrSetImpl<Argument *> &Args) const {
4848 if (!BaseT::areFunctionArgsABICompatible(Caller, Callee, Args))
4849 return false;
4850
4851 // If we get here, we know the target features match. If one function
4852 // considers 512-bit vectors legal and the other does not, consider them
4853 // incompatible.
4854 const TargetMachine &TM = getTLI()->getTargetMachine();
4855
4856 if (TM.getSubtarget<X86Subtarget>(*Caller).useAVX512Regs() ==
4857 TM.getSubtarget<X86Subtarget>(*Callee).useAVX512Regs())
4858 return true;
4859
4860 // Consider the arguments compatible if they aren't vectors or aggregates.
4861 // FIXME: Look at the size of vectors.
4862 // FIXME: Look at the element types of aggregates to see if there are vectors.
4863 // FIXME: The API of this function seems intended to allow arguments
4864 // to be removed from the set, but the caller doesn't check if the set
4865 // becomes empty so that may not work in practice.
4866 return llvm::none_of(Args, [](Argument *A) {
4867 auto *EltTy = cast<PointerType>(A->getType())->getElementType();
4868 return EltTy->isVectorTy() || EltTy->isAggregateType();
4869 });
4870}
4871
4872X86TTIImpl::TTI::MemCmpExpansionOptions
4873X86TTIImpl::enableMemCmpExpansion(bool OptSize, bool IsZeroCmp) const {
4874 TTI::MemCmpExpansionOptions Options;
4875 Options.MaxNumLoads = TLI->getMaxExpandSizeMemcmp(OptSize);
4876 Options.NumLoadsPerBlock = 2;
4877 // All GPR and vector loads can be unaligned.
4878 Options.AllowOverlappingLoads = true;
4879 if (IsZeroCmp) {
4880 // Only enable vector loads for equality comparison. Right now the vector
4881 // version is not as fast for three way compare (see #33329).
4882 const unsigned PreferredWidth = ST->getPreferVectorWidth();
4883 if (PreferredWidth >= 512 && ST->hasAVX512()) Options.LoadSizes.push_back(64);
4884 if (PreferredWidth >= 256 && ST->hasAVX()) Options.LoadSizes.push_back(32);
4885 if (PreferredWidth >= 128 && ST->hasSSE2()) Options.LoadSizes.push_back(16);
4886 }
4887 if (ST->is64Bit()) {
4888 Options.LoadSizes.push_back(8);
4889 }
4890 Options.LoadSizes.push_back(4);
4891 Options.LoadSizes.push_back(2);
4892 Options.LoadSizes.push_back(1);
4893 return Options;
4894}
4895
4896bool X86TTIImpl::enableInterleavedAccessVectorization() {
4897 // TODO: We expect this to be beneficial regardless of arch,
4898 // but there are currently some unexplained performance artifacts on Atom.
4899 // As a temporary solution, disable on Atom.
4900 return !(ST->isAtom());
4901}
4902
4903// Get estimation for interleaved load/store operations for AVX2.
4904// \p Factor is the interleaved-access factor (stride) - number of
4905// (interleaved) elements in the group.
4906// \p Indices contains the indices for a strided load: when the
4907// interleaved load has gaps they indicate which elements are used.
4908// If Indices is empty (or if the number of indices is equal to the size
4909// of the interleaved-access as given in \p Factor) the access has no gaps.
4910//
4911// As opposed to AVX-512, AVX2 does not have generic shuffles that allow
4912// computing the cost using a generic formula as a function of generic
4913// shuffles. We therefore use a lookup table instead, filled according to
4914// the instruction sequences that codegen currently generates.
4915InstructionCost X86TTIImpl::getInterleavedMemoryOpCostAVX2(
4916 unsigned Opcode, FixedVectorType *VecTy, unsigned Factor,
4917 ArrayRef<unsigned> Indices, Align Alignment, unsigned AddressSpace,
4918 TTI::TargetCostKind CostKind, bool UseMaskForCond, bool UseMaskForGaps) {
4919
4920 if (UseMaskForCond || UseMaskForGaps)
4921 return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices,
4922 Alignment, AddressSpace, CostKind,
4923 UseMaskForCond, UseMaskForGaps);
4924
4925 // We currently Support only fully-interleaved groups, with no gaps.
4926 // TODO: Support also strided loads (interleaved-groups with gaps).
4927 if (Indices.size() && Indices.size() != Factor)
4928 return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices,
4929 Alignment, AddressSpace, CostKind);
4930
4931 // VecTy for interleave memop is <VF*Factor x Elt>.
4932 // So, for VF=4, Interleave Factor = 3, Element type = i32 we have
4933 // VecTy = <12 x i32>.
4934 MVT LegalVT = getTLI()->getTypeLegalizationCost(DL, VecTy).second;
4935
4936 // This function can be called with VecTy=<6xi128>, Factor=3, in which case
4937 // the VF=2, while v2i128 is an unsupported MVT vector type
4938 // (see MachineValueType.h::getVectorVT()).
4939 if (!LegalVT.isVector())
4940 return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices,
4941 Alignment, AddressSpace, CostKind);
4942
4943 unsigned VF = VecTy->getNumElements() / Factor;
4944 Type *ScalarTy = VecTy->getElementType();
4945 // Deduplicate entries, model floats/pointers as appropriately-sized integers.
4946 if (!ScalarTy->isIntegerTy())
4947 ScalarTy =
4948 Type::getIntNTy(ScalarTy->getContext(), DL.getTypeSizeInBits(ScalarTy));
4949
4950 // Get the cost of all the memory operations.
4951 InstructionCost MemOpCosts = getMemoryOpCost(
4952 Opcode, VecTy, MaybeAlign(Alignment), AddressSpace, CostKind);
4953
4954 auto *VT = FixedVectorType::get(ScalarTy, VF);
4955 EVT ETy = TLI->getValueType(DL, VT);
4956 if (!ETy.isSimple())
4957 return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices,
4958 Alignment, AddressSpace, CostKind);
4959
4960 // TODO: Complete for other data-types and strides.
4961 // Each combination of Stride, element bit width and VF results in a different
4962 // sequence; The cost tables are therefore accessed with:
4963 // Factor (stride) and VectorType=VFxiN.
4964 // The Cost accounts only for the shuffle sequence;
4965 // The cost of the loads/stores is accounted for separately.
4966 //
4967 static const CostTblEntry AVX2InterleavedLoadTbl[] = {
4968 {2, MVT::v4i64, 6}, // (load 8i64 and) deinterleave into 2 x 4i64
4969
4970 {3, MVT::v2i8, 10}, // (load 6i8 and) deinterleave into 3 x 2i8
4971 {3, MVT::v4i8, 4}, // (load 12i8 and) deinterleave into 3 x 4i8
4972 {3, MVT::v8i8, 9}, // (load 24i8 and) deinterleave into 3 x 8i8
4973 {3, MVT::v16i8, 11}, // (load 48i8 and) deinterleave into 3 x 16i8
4974 {3, MVT::v32i8, 13}, // (load 96i8 and) deinterleave into 3 x 32i8
4975
4976 {3, MVT::v8i32, 17}, // (load 24i32 and) deinterleave into 3 x 8i32
4977
4978 {4, MVT::v2i8, 12}, // (load 8i8 and) deinterleave into 4 x 2i8
4979 {4, MVT::v4i8, 4}, // (load 16i8 and) deinterleave into 4 x 4i8
4980 {4, MVT::v8i8, 20}, // (load 32i8 and) deinterleave into 4 x 8i8
4981 {4, MVT::v16i8, 39}, // (load 64i8 and) deinterleave into 4 x 16i8
4982 {4, MVT::v32i8, 80}, // (load 128i8 and) deinterleave into 4 x 32i8
4983
4984 {8, MVT::v8i32, 40} // (load 64i32 and) deinterleave into 8 x 8i32
4985 };
4986
4987 static const CostTblEntry AVX2InterleavedStoreTbl[] = {
4988 {2, MVT::v4i64, 6}, // interleave 2 x 4i64 into 8i64 (and store)
4989
4990 {3, MVT::v2i8, 7}, // interleave 3 x 2i8 into 6i8 (and store)
4991 {3, MVT::v4i8, 8}, // interleave 3 x 4i8 into 12i8 (and store)
4992 {3, MVT::v8i8, 11}, // interleave 3 x 8i8 into 24i8 (and store)
4993 {3, MVT::v16i8, 11}, // interleave 3 x 16i8 into 48i8 (and store)
4994 {3, MVT::v32i8, 13}, // interleave 3 x 32i8 into 96i8 (and store)
4995
4996 {4, MVT::v2i8, 12}, // interleave 4 x 2i8 into 8i8 (and store)
4997 {4, MVT::v4i8, 9}, // interleave 4 x 4i8 into 16i8 (and store)
4998 {4, MVT::v8i8, 10}, // interleave 4 x 8i8 into 32i8 (and store)
4999 {4, MVT::v16i8, 10}, // interleave 4 x 16i8 into 64i8 (and store)
5000 {4, MVT::v32i8, 12} // interleave 4 x 32i8 into 128i8 (and store)
5001 };
5002
5003 if (Opcode == Instruction::Load) {
5004 if (const auto *Entry =
5005 CostTableLookup(AVX2InterleavedLoadTbl, Factor, ETy.getSimpleVT()))
5006 return MemOpCosts + Entry->Cost;
5007 } else {
5008 assert(Opcode == Instruction::Store &&((void)0)
5009 "Expected Store Instruction at this point")((void)0);
5010 if (const auto *Entry =
5011 CostTableLookup(AVX2InterleavedStoreTbl, Factor, ETy.getSimpleVT()))
5012 return MemOpCosts + Entry->Cost;
5013 }
5014
5015 return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices,
5016 Alignment, AddressSpace, CostKind);
5017}
5018
5019// Get estimation for interleaved load/store operations and strided load.
5020// \p Indices contains indices for strided load.
5021// \p Factor - the factor of interleaving.
5022// AVX-512 provides 3-src shuffles that significantly reduces the cost.
5023InstructionCost X86TTIImpl::getInterleavedMemoryOpCostAVX512(
5024 unsigned Opcode, FixedVectorType *VecTy, unsigned Factor,
5025 ArrayRef<unsigned> Indices, Align Alignment, unsigned AddressSpace,
5026 TTI::TargetCostKind CostKind, bool UseMaskForCond, bool UseMaskForGaps) {
5027
5028 if (UseMaskForCond || UseMaskForGaps)
5029 return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices,
5030 Alignment, AddressSpace, CostKind,
5031 UseMaskForCond, UseMaskForGaps);
5032
5033 // VecTy for interleave memop is <VF*Factor x Elt>.
5034 // So, for VF=4, Interleave Factor = 3, Element type = i32 we have
5035 // VecTy = <12 x i32>.
5036
5037 // Calculate the number of memory operations (NumOfMemOps), required
5038 // for load/store the VecTy.
5039 MVT LegalVT = getTLI()->getTypeLegalizationCost(DL, VecTy).second;
5040 unsigned VecTySize = DL.getTypeStoreSize(VecTy);
5041 unsigned LegalVTSize = LegalVT.getStoreSize();
5042 unsigned NumOfMemOps = (VecTySize + LegalVTSize - 1) / LegalVTSize;
5043
5044 // Get the cost of one memory operation.
5045 auto *SingleMemOpTy = FixedVectorType::get(VecTy->getElementType(),
5046 LegalVT.getVectorNumElements());
5047 InstructionCost MemOpCost = getMemoryOpCost(
5048 Opcode, SingleMemOpTy, MaybeAlign(Alignment), AddressSpace, CostKind);
5049
5050 unsigned VF = VecTy->getNumElements() / Factor;
5051 MVT VT = MVT::getVectorVT(MVT::getVT(VecTy->getScalarType()), VF);
5052
5053 if (Opcode == Instruction::Load) {
5054 // The tables (AVX512InterleavedLoadTbl and AVX512InterleavedStoreTbl)
5055 // contain the cost of the optimized shuffle sequence that the
5056 // X86InterleavedAccess pass will generate.
5057 // The cost of loads and stores are computed separately from the table.
5058
5059 // X86InterleavedAccess support only the following interleaved-access group.
5060 static const CostTblEntry AVX512InterleavedLoadTbl[] = {
5061 {3, MVT::v16i8, 12}, //(load 48i8 and) deinterleave into 3 x 16i8
5062 {3, MVT::v32i8, 14}, //(load 96i8 and) deinterleave into 3 x 32i8
5063 {3, MVT::v64i8, 22}, //(load 96i8 and) deinterleave into 3 x 32i8
5064 };
5065
5066 if (const auto *Entry =
5067 CostTableLookup(AVX512InterleavedLoadTbl, Factor, VT))
5068 return NumOfMemOps * MemOpCost + Entry->Cost;
5069 //If an entry does not exist, fallback to the default implementation.
5070
5071 // Kind of shuffle depends on number of loaded values.
5072 // If we load the entire data in one register, we can use a 1-src shuffle.
5073 // Otherwise, we'll merge 2 sources in each operation.
5074 TTI::ShuffleKind ShuffleKind =
5075 (NumOfMemOps > 1) ? TTI::SK_PermuteTwoSrc : TTI::SK_PermuteSingleSrc;
5076
5077 InstructionCost ShuffleCost =
5078 getShuffleCost(ShuffleKind, SingleMemOpTy, None, 0, nullptr);
5079
5080 unsigned NumOfLoadsInInterleaveGrp =
5081 Indices.size() ? Indices.size() : Factor;
5082 auto *ResultTy = FixedVectorType::get(VecTy->getElementType(),
5083 VecTy->getNumElements() / Factor);
5084 InstructionCost NumOfResults =
5085 getTLI()->getTypeLegalizationCost(DL, ResultTy).first *
5086 NumOfLoadsInInterleaveGrp;
5087
5088 // About a half of the loads may be folded in shuffles when we have only
5089 // one result. If we have more than one result, we do not fold loads at all.
5090 unsigned NumOfUnfoldedLoads =
5091 NumOfResults > 1 ? NumOfMemOps : NumOfMemOps / 2;
5092
5093 // Get a number of shuffle operations per result.
5094 unsigned NumOfShufflesPerResult =
5095 std::max((unsigned)1, (unsigned)(NumOfMemOps - 1));
5096
5097 // The SK_MergeTwoSrc shuffle clobbers one of src operands.
5098 // When we have more than one destination, we need additional instructions
5099 // to keep sources.
5100 InstructionCost NumOfMoves = 0;
5101 if (NumOfResults > 1 && ShuffleKind == TTI::SK_PermuteTwoSrc)
5102 NumOfMoves = NumOfResults * NumOfShufflesPerResult / 2;
5103
5104 InstructionCost Cost = NumOfResults * NumOfShufflesPerResult * ShuffleCost +
5105 NumOfUnfoldedLoads * MemOpCost + NumOfMoves;
5106
5107 return Cost;
5108 }
5109
5110 // Store.
5111 assert(Opcode == Instruction::Store &&((void)0)
5112 "Expected Store Instruction at this point")((void)0);
5113 // X86InterleavedAccess support only the following interleaved-access group.
5114 static const CostTblEntry AVX512InterleavedStoreTbl[] = {
5115 {3, MVT::v16i8, 12}, // interleave 3 x 16i8 into 48i8 (and store)
5116 {3, MVT::v32i8, 14}, // interleave 3 x 32i8 into 96i8 (and store)
5117 {3, MVT::v64i8, 26}, // interleave 3 x 64i8 into 96i8 (and store)
5118
5119 {4, MVT::v8i8, 10}, // interleave 4 x 8i8 into 32i8 (and store)
5120 {4, MVT::v16i8, 11}, // interleave 4 x 16i8 into 64i8 (and store)
5121 {4, MVT::v32i8, 14}, // interleave 4 x 32i8 into 128i8 (and store)
5122 {4, MVT::v64i8, 24} // interleave 4 x 32i8 into 256i8 (and store)
5123 };
5124
5125 if (const auto *Entry =
5126 CostTableLookup(AVX512InterleavedStoreTbl, Factor, VT))
5127 return NumOfMemOps * MemOpCost + Entry->Cost;
5128 //If an entry does not exist, fallback to the default implementation.
5129
5130 // There is no strided stores meanwhile. And store can't be folded in
5131 // shuffle.
5132 unsigned NumOfSources = Factor; // The number of values to be merged.
5133 InstructionCost ShuffleCost =
5134 getShuffleCost(TTI::SK_PermuteTwoSrc, SingleMemOpTy, None, 0, nullptr);
5135 unsigned NumOfShufflesPerStore = NumOfSources - 1;
5136
5137 // The SK_MergeTwoSrc shuffle clobbers one of src operands.
5138 // We need additional instructions to keep sources.
5139 unsigned NumOfMoves = NumOfMemOps * NumOfShufflesPerStore / 2;
5140 InstructionCost Cost =
5141 NumOfMemOps * (MemOpCost + NumOfShufflesPerStore * ShuffleCost) +
5142 NumOfMoves;
5143 return Cost;
5144}
5145
5146InstructionCost X86TTIImpl::getInterleavedMemoryOpCost(
5147 unsigned Opcode, Type *VecTy, unsigned Factor, ArrayRef<unsigned> Indices,
5148 Align Alignment, unsigned AddressSpace, TTI::TargetCostKind CostKind,
5149 bool UseMaskForCond, bool UseMaskForGaps) {
5150 auto isSupportedOnAVX512 = [](Type *VecTy, bool HasBW) {
5151 Type *EltTy = cast<VectorType>(VecTy)->getElementType();
5152 if (EltTy->isFloatTy() || EltTy->isDoubleTy() || EltTy->isIntegerTy(64) ||
5153 EltTy->isIntegerTy(32) || EltTy->isPointerTy())
5154 return true;
5155 if (EltTy->isIntegerTy(16) || EltTy->isIntegerTy(8))
5156 return HasBW;
5157 return false;
5158 };
5159 if (ST->hasAVX512() && isSupportedOnAVX512(VecTy, ST->hasBWI()))
5160 return getInterleavedMemoryOpCostAVX512(
5161 Opcode, cast<FixedVectorType>(VecTy), Factor, Indices, Alignment,
5162 AddressSpace, CostKind, UseMaskForCond, UseMaskForGaps);
5163 if (ST->hasAVX2())
5164 return getInterleavedMemoryOpCostAVX2(
5165 Opcode, cast<FixedVectorType>(VecTy), Factor, Indices, Alignment,
5166 AddressSpace, CostKind, UseMaskForCond, UseMaskForGaps);
5167
5168 return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices,
5169 Alignment, AddressSpace, CostKind,
5170 UseMaskForCond, UseMaskForGaps);
5171}