1 //===- llvm/Analysis/TargetTransformInfo.cpp ------------------------------===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 10 #include "llvm/Analysis/TargetTransformInfo.h" 11 #include "llvm/Analysis/TargetTransformInfoImpl.h" 12 #include "llvm/IR/CallSite.h" 13 #include "llvm/IR/DataLayout.h" 14 #include "llvm/IR/Instruction.h" 15 #include "llvm/IR/Instructions.h" 16 #include "llvm/IR/IntrinsicInst.h" 17 #include "llvm/IR/Module.h" 18 #include "llvm/IR/Operator.h" 19 #include "llvm/IR/PatternMatch.h" 20 #include "llvm/Support/CommandLine.h" 21 #include "llvm/Support/ErrorHandling.h" 22 #include <utility> 23 24 using namespace llvm; 25 using namespace PatternMatch; 26 27 #define DEBUG_TYPE "tti" 28 29 static cl::opt<bool> UseWideMemcpyLoopLowering( 30 "use-wide-memcpy-loop-lowering", cl::init(false), 31 cl::desc("Enables the new wide memcpy loop lowering in Transforms/Utils."), 32 cl::Hidden); 33 34 static cl::opt<bool> EnableReduxCost("costmodel-reduxcost", cl::init(false), 35 cl::Hidden, 36 cl::desc("Recognize reduction patterns.")); 37 38 namespace { 39 /// \brief No-op implementation of the TTI interface using the utility base 40 /// classes. 41 /// 42 /// This is used when no target specific information is available. 43 struct NoTTIImpl : TargetTransformInfoImplCRTPBase<NoTTIImpl> { 44 explicit NoTTIImpl(const DataLayout &DL) 45 : TargetTransformInfoImplCRTPBase<NoTTIImpl>(DL) {} 46 }; 47 } 48 49 TargetTransformInfo::TargetTransformInfo(const DataLayout &DL) 50 : TTIImpl(new Model<NoTTIImpl>(NoTTIImpl(DL))) {} 51 52 TargetTransformInfo::~TargetTransformInfo() {} 53 54 TargetTransformInfo::TargetTransformInfo(TargetTransformInfo &&Arg) 55 : TTIImpl(std::move(Arg.TTIImpl)) {} 56 57 TargetTransformInfo &TargetTransformInfo::operator=(TargetTransformInfo &&RHS) { 58 TTIImpl = std::move(RHS.TTIImpl); 59 return *this; 60 } 61 62 int TargetTransformInfo::getOperationCost(unsigned Opcode, Type *Ty, 63 Type *OpTy) const { 64 int Cost = TTIImpl->getOperationCost(Opcode, Ty, OpTy); 65 assert(Cost >= 0 && "TTI should not produce negative costs!"); 66 return Cost; 67 } 68 69 int TargetTransformInfo::getCallCost(FunctionType *FTy, int NumArgs) const { 70 int Cost = TTIImpl->getCallCost(FTy, NumArgs); 71 assert(Cost >= 0 && "TTI should not produce negative costs!"); 72 return Cost; 73 } 74 75 int TargetTransformInfo::getCallCost(const Function *F, 76 ArrayRef<const Value *> Arguments) const { 77 int Cost = TTIImpl->getCallCost(F, Arguments); 78 assert(Cost >= 0 && "TTI should not produce negative costs!"); 79 return Cost; 80 } 81 82 unsigned TargetTransformInfo::getInliningThresholdMultiplier() const { 83 return TTIImpl->getInliningThresholdMultiplier(); 84 } 85 86 int TargetTransformInfo::getGEPCost(Type *PointeeType, const Value *Ptr, 87 ArrayRef<const Value *> Operands) const { 88 return TTIImpl->getGEPCost(PointeeType, Ptr, Operands); 89 } 90 91 int TargetTransformInfo::getExtCost(const Instruction *I, 92 const Value *Src) const { 93 return TTIImpl->getExtCost(I, Src); 94 } 95 96 int TargetTransformInfo::getIntrinsicCost( 97 Intrinsic::ID IID, Type *RetTy, ArrayRef<const Value *> Arguments) const { 98 int Cost = TTIImpl->getIntrinsicCost(IID, RetTy, Arguments); 99 assert(Cost >= 0 && "TTI should not produce negative costs!"); 100 return Cost; 101 } 102 103 unsigned 104 TargetTransformInfo::getEstimatedNumberOfCaseClusters(const SwitchInst &SI, 105 unsigned &JTSize) const { 106 return TTIImpl->getEstimatedNumberOfCaseClusters(SI, JTSize); 107 } 108 109 int TargetTransformInfo::getUserCost(const User *U, 110 ArrayRef<const Value *> Operands) const { 111 int Cost = TTIImpl->getUserCost(U, Operands); 112 assert(Cost >= 0 && "TTI should not produce negative costs!"); 113 return Cost; 114 } 115 116 bool TargetTransformInfo::hasBranchDivergence() const { 117 return TTIImpl->hasBranchDivergence(); 118 } 119 120 bool TargetTransformInfo::isSourceOfDivergence(const Value *V) const { 121 return TTIImpl->isSourceOfDivergence(V); 122 } 123 124 bool llvm::TargetTransformInfo::isAlwaysUniform(const Value *V) const { 125 return TTIImpl->isAlwaysUniform(V); 126 } 127 128 unsigned TargetTransformInfo::getFlatAddressSpace() const { 129 return TTIImpl->getFlatAddressSpace(); 130 } 131 132 bool TargetTransformInfo::isLoweredToCall(const Function *F) const { 133 return TTIImpl->isLoweredToCall(F); 134 } 135 136 void TargetTransformInfo::getUnrollingPreferences( 137 Loop *L, ScalarEvolution &SE, UnrollingPreferences &UP) const { 138 return TTIImpl->getUnrollingPreferences(L, SE, UP); 139 } 140 141 bool TargetTransformInfo::isLegalAddImmediate(int64_t Imm) const { 142 return TTIImpl->isLegalAddImmediate(Imm); 143 } 144 145 bool TargetTransformInfo::isLegalICmpImmediate(int64_t Imm) const { 146 return TTIImpl->isLegalICmpImmediate(Imm); 147 } 148 149 bool TargetTransformInfo::isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV, 150 int64_t BaseOffset, 151 bool HasBaseReg, 152 int64_t Scale, 153 unsigned AddrSpace, 154 Instruction *I) const { 155 return TTIImpl->isLegalAddressingMode(Ty, BaseGV, BaseOffset, HasBaseReg, 156 Scale, AddrSpace, I); 157 } 158 159 bool TargetTransformInfo::isLSRCostLess(LSRCost &C1, LSRCost &C2) const { 160 return TTIImpl->isLSRCostLess(C1, C2); 161 } 162 163 bool TargetTransformInfo::isLegalMaskedStore(Type *DataType) const { 164 return TTIImpl->isLegalMaskedStore(DataType); 165 } 166 167 bool TargetTransformInfo::isLegalMaskedLoad(Type *DataType) const { 168 return TTIImpl->isLegalMaskedLoad(DataType); 169 } 170 171 bool TargetTransformInfo::isLegalMaskedGather(Type *DataType) const { 172 return TTIImpl->isLegalMaskedGather(DataType); 173 } 174 175 bool TargetTransformInfo::isLegalMaskedScatter(Type *DataType) const { 176 return TTIImpl->isLegalMaskedScatter(DataType); 177 } 178 179 bool TargetTransformInfo::hasDivRemOp(Type *DataType, bool IsSigned) const { 180 return TTIImpl->hasDivRemOp(DataType, IsSigned); 181 } 182 183 bool TargetTransformInfo::prefersVectorizedAddressing() const { 184 return TTIImpl->prefersVectorizedAddressing(); 185 } 186 187 int TargetTransformInfo::getScalingFactorCost(Type *Ty, GlobalValue *BaseGV, 188 int64_t BaseOffset, 189 bool HasBaseReg, 190 int64_t Scale, 191 unsigned AddrSpace) const { 192 int Cost = TTIImpl->getScalingFactorCost(Ty, BaseGV, BaseOffset, HasBaseReg, 193 Scale, AddrSpace); 194 assert(Cost >= 0 && "TTI should not produce negative costs!"); 195 return Cost; 196 } 197 198 bool TargetTransformInfo::LSRWithInstrQueries() const { 199 return TTIImpl->LSRWithInstrQueries(); 200 } 201 202 bool TargetTransformInfo::isTruncateFree(Type *Ty1, Type *Ty2) const { 203 return TTIImpl->isTruncateFree(Ty1, Ty2); 204 } 205 206 bool TargetTransformInfo::isProfitableToHoist(Instruction *I) const { 207 return TTIImpl->isProfitableToHoist(I); 208 } 209 210 bool TargetTransformInfo::isTypeLegal(Type *Ty) const { 211 return TTIImpl->isTypeLegal(Ty); 212 } 213 214 unsigned TargetTransformInfo::getJumpBufAlignment() const { 215 return TTIImpl->getJumpBufAlignment(); 216 } 217 218 unsigned TargetTransformInfo::getJumpBufSize() const { 219 return TTIImpl->getJumpBufSize(); 220 } 221 222 bool TargetTransformInfo::shouldBuildLookupTables() const { 223 return TTIImpl->shouldBuildLookupTables(); 224 } 225 bool TargetTransformInfo::shouldBuildLookupTablesForConstant(Constant *C) const { 226 return TTIImpl->shouldBuildLookupTablesForConstant(C); 227 } 228 229 unsigned TargetTransformInfo:: 230 getScalarizationOverhead(Type *Ty, bool Insert, bool Extract) const { 231 return TTIImpl->getScalarizationOverhead(Ty, Insert, Extract); 232 } 233 234 unsigned TargetTransformInfo:: 235 getOperandsScalarizationOverhead(ArrayRef<const Value *> Args, 236 unsigned VF) const { 237 return TTIImpl->getOperandsScalarizationOverhead(Args, VF); 238 } 239 240 bool TargetTransformInfo::supportsEfficientVectorElementLoadStore() const { 241 return TTIImpl->supportsEfficientVectorElementLoadStore(); 242 } 243 244 bool TargetTransformInfo::enableAggressiveInterleaving(bool LoopHasReductions) const { 245 return TTIImpl->enableAggressiveInterleaving(LoopHasReductions); 246 } 247 248 bool TargetTransformInfo::enableMemCmpExpansion(unsigned &MaxLoadSize) const { 249 return TTIImpl->enableMemCmpExpansion(MaxLoadSize); 250 } 251 252 bool TargetTransformInfo::enableInterleavedAccessVectorization() const { 253 return TTIImpl->enableInterleavedAccessVectorization(); 254 } 255 256 bool TargetTransformInfo::isFPVectorizationPotentiallyUnsafe() const { 257 return TTIImpl->isFPVectorizationPotentiallyUnsafe(); 258 } 259 260 bool TargetTransformInfo::allowsMisalignedMemoryAccesses(LLVMContext &Context, 261 unsigned BitWidth, 262 unsigned AddressSpace, 263 unsigned Alignment, 264 bool *Fast) const { 265 return TTIImpl->allowsMisalignedMemoryAccesses(Context, BitWidth, AddressSpace, 266 Alignment, Fast); 267 } 268 269 TargetTransformInfo::PopcntSupportKind 270 TargetTransformInfo::getPopcntSupport(unsigned IntTyWidthInBit) const { 271 return TTIImpl->getPopcntSupport(IntTyWidthInBit); 272 } 273 274 bool TargetTransformInfo::haveFastSqrt(Type *Ty) const { 275 return TTIImpl->haveFastSqrt(Ty); 276 } 277 278 int TargetTransformInfo::getFPOpCost(Type *Ty) const { 279 int Cost = TTIImpl->getFPOpCost(Ty); 280 assert(Cost >= 0 && "TTI should not produce negative costs!"); 281 return Cost; 282 } 283 284 int TargetTransformInfo::getIntImmCodeSizeCost(unsigned Opcode, unsigned Idx, 285 const APInt &Imm, 286 Type *Ty) const { 287 int Cost = TTIImpl->getIntImmCodeSizeCost(Opcode, Idx, Imm, Ty); 288 assert(Cost >= 0 && "TTI should not produce negative costs!"); 289 return Cost; 290 } 291 292 int TargetTransformInfo::getIntImmCost(const APInt &Imm, Type *Ty) const { 293 int Cost = TTIImpl->getIntImmCost(Imm, Ty); 294 assert(Cost >= 0 && "TTI should not produce negative costs!"); 295 return Cost; 296 } 297 298 int TargetTransformInfo::getIntImmCost(unsigned Opcode, unsigned Idx, 299 const APInt &Imm, Type *Ty) const { 300 int Cost = TTIImpl->getIntImmCost(Opcode, Idx, Imm, Ty); 301 assert(Cost >= 0 && "TTI should not produce negative costs!"); 302 return Cost; 303 } 304 305 int TargetTransformInfo::getIntImmCost(Intrinsic::ID IID, unsigned Idx, 306 const APInt &Imm, Type *Ty) const { 307 int Cost = TTIImpl->getIntImmCost(IID, Idx, Imm, Ty); 308 assert(Cost >= 0 && "TTI should not produce negative costs!"); 309 return Cost; 310 } 311 312 unsigned TargetTransformInfo::getNumberOfRegisters(bool Vector) const { 313 return TTIImpl->getNumberOfRegisters(Vector); 314 } 315 316 unsigned TargetTransformInfo::getRegisterBitWidth(bool Vector) const { 317 return TTIImpl->getRegisterBitWidth(Vector); 318 } 319 320 unsigned TargetTransformInfo::getMinVectorRegisterBitWidth() const { 321 return TTIImpl->getMinVectorRegisterBitWidth(); 322 } 323 324 bool TargetTransformInfo::shouldConsiderAddressTypePromotion( 325 const Instruction &I, bool &AllowPromotionWithoutCommonHeader) const { 326 return TTIImpl->shouldConsiderAddressTypePromotion( 327 I, AllowPromotionWithoutCommonHeader); 328 } 329 330 unsigned TargetTransformInfo::getCacheLineSize() const { 331 return TTIImpl->getCacheLineSize(); 332 } 333 334 llvm::Optional<unsigned> TargetTransformInfo::getCacheSize(CacheLevel Level) 335 const { 336 return TTIImpl->getCacheSize(Level); 337 } 338 339 llvm::Optional<unsigned> TargetTransformInfo::getCacheAssociativity( 340 CacheLevel Level) const { 341 return TTIImpl->getCacheAssociativity(Level); 342 } 343 344 unsigned TargetTransformInfo::getPrefetchDistance() const { 345 return TTIImpl->getPrefetchDistance(); 346 } 347 348 unsigned TargetTransformInfo::getMinPrefetchStride() const { 349 return TTIImpl->getMinPrefetchStride(); 350 } 351 352 unsigned TargetTransformInfo::getMaxPrefetchIterationsAhead() const { 353 return TTIImpl->getMaxPrefetchIterationsAhead(); 354 } 355 356 unsigned TargetTransformInfo::getMaxInterleaveFactor(unsigned VF) const { 357 return TTIImpl->getMaxInterleaveFactor(VF); 358 } 359 360 int TargetTransformInfo::getArithmeticInstrCost( 361 unsigned Opcode, Type *Ty, OperandValueKind Opd1Info, 362 OperandValueKind Opd2Info, OperandValueProperties Opd1PropInfo, 363 OperandValueProperties Opd2PropInfo, 364 ArrayRef<const Value *> Args) const { 365 int Cost = TTIImpl->getArithmeticInstrCost(Opcode, Ty, Opd1Info, Opd2Info, 366 Opd1PropInfo, Opd2PropInfo, Args); 367 assert(Cost >= 0 && "TTI should not produce negative costs!"); 368 return Cost; 369 } 370 371 int TargetTransformInfo::getShuffleCost(ShuffleKind Kind, Type *Ty, int Index, 372 Type *SubTp) const { 373 int Cost = TTIImpl->getShuffleCost(Kind, Ty, Index, SubTp); 374 assert(Cost >= 0 && "TTI should not produce negative costs!"); 375 return Cost; 376 } 377 378 int TargetTransformInfo::getCastInstrCost(unsigned Opcode, Type *Dst, 379 Type *Src, const Instruction *I) const { 380 assert ((I == nullptr || I->getOpcode() == Opcode) && 381 "Opcode should reflect passed instruction."); 382 int Cost = TTIImpl->getCastInstrCost(Opcode, Dst, Src, I); 383 assert(Cost >= 0 && "TTI should not produce negative costs!"); 384 return Cost; 385 } 386 387 int TargetTransformInfo::getExtractWithExtendCost(unsigned Opcode, Type *Dst, 388 VectorType *VecTy, 389 unsigned Index) const { 390 int Cost = TTIImpl->getExtractWithExtendCost(Opcode, Dst, VecTy, Index); 391 assert(Cost >= 0 && "TTI should not produce negative costs!"); 392 return Cost; 393 } 394 395 int TargetTransformInfo::getCFInstrCost(unsigned Opcode) const { 396 int Cost = TTIImpl->getCFInstrCost(Opcode); 397 assert(Cost >= 0 && "TTI should not produce negative costs!"); 398 return Cost; 399 } 400 401 int TargetTransformInfo::getCmpSelInstrCost(unsigned Opcode, Type *ValTy, 402 Type *CondTy, const Instruction *I) const { 403 assert ((I == nullptr || I->getOpcode() == Opcode) && 404 "Opcode should reflect passed instruction."); 405 int Cost = TTIImpl->getCmpSelInstrCost(Opcode, ValTy, CondTy, I); 406 assert(Cost >= 0 && "TTI should not produce negative costs!"); 407 return Cost; 408 } 409 410 int TargetTransformInfo::getVectorInstrCost(unsigned Opcode, Type *Val, 411 unsigned Index) const { 412 int Cost = TTIImpl->getVectorInstrCost(Opcode, Val, Index); 413 assert(Cost >= 0 && "TTI should not produce negative costs!"); 414 return Cost; 415 } 416 417 int TargetTransformInfo::getMemoryOpCost(unsigned Opcode, Type *Src, 418 unsigned Alignment, 419 unsigned AddressSpace, 420 const Instruction *I) const { 421 assert ((I == nullptr || I->getOpcode() == Opcode) && 422 "Opcode should reflect passed instruction."); 423 int Cost = TTIImpl->getMemoryOpCost(Opcode, Src, Alignment, AddressSpace, I); 424 assert(Cost >= 0 && "TTI should not produce negative costs!"); 425 return Cost; 426 } 427 428 int TargetTransformInfo::getMaskedMemoryOpCost(unsigned Opcode, Type *Src, 429 unsigned Alignment, 430 unsigned AddressSpace) const { 431 int Cost = 432 TTIImpl->getMaskedMemoryOpCost(Opcode, Src, Alignment, AddressSpace); 433 assert(Cost >= 0 && "TTI should not produce negative costs!"); 434 return Cost; 435 } 436 437 int TargetTransformInfo::getGatherScatterOpCost(unsigned Opcode, Type *DataTy, 438 Value *Ptr, bool VariableMask, 439 unsigned Alignment) const { 440 int Cost = TTIImpl->getGatherScatterOpCost(Opcode, DataTy, Ptr, VariableMask, 441 Alignment); 442 assert(Cost >= 0 && "TTI should not produce negative costs!"); 443 return Cost; 444 } 445 446 int TargetTransformInfo::getInterleavedMemoryOpCost( 447 unsigned Opcode, Type *VecTy, unsigned Factor, ArrayRef<unsigned> Indices, 448 unsigned Alignment, unsigned AddressSpace) const { 449 int Cost = TTIImpl->getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices, 450 Alignment, AddressSpace); 451 assert(Cost >= 0 && "TTI should not produce negative costs!"); 452 return Cost; 453 } 454 455 int TargetTransformInfo::getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy, 456 ArrayRef<Type *> Tys, FastMathFlags FMF, 457 unsigned ScalarizationCostPassed) const { 458 int Cost = TTIImpl->getIntrinsicInstrCost(ID, RetTy, Tys, FMF, 459 ScalarizationCostPassed); 460 assert(Cost >= 0 && "TTI should not produce negative costs!"); 461 return Cost; 462 } 463 464 int TargetTransformInfo::getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy, 465 ArrayRef<Value *> Args, FastMathFlags FMF, unsigned VF) const { 466 int Cost = TTIImpl->getIntrinsicInstrCost(ID, RetTy, Args, FMF, VF); 467 assert(Cost >= 0 && "TTI should not produce negative costs!"); 468 return Cost; 469 } 470 471 int TargetTransformInfo::getCallInstrCost(Function *F, Type *RetTy, 472 ArrayRef<Type *> Tys) const { 473 int Cost = TTIImpl->getCallInstrCost(F, RetTy, Tys); 474 assert(Cost >= 0 && "TTI should not produce negative costs!"); 475 return Cost; 476 } 477 478 unsigned TargetTransformInfo::getNumberOfParts(Type *Tp) const { 479 return TTIImpl->getNumberOfParts(Tp); 480 } 481 482 int TargetTransformInfo::getAddressComputationCost(Type *Tp, 483 ScalarEvolution *SE, 484 const SCEV *Ptr) const { 485 int Cost = TTIImpl->getAddressComputationCost(Tp, SE, Ptr); 486 assert(Cost >= 0 && "TTI should not produce negative costs!"); 487 return Cost; 488 } 489 490 int TargetTransformInfo::getArithmeticReductionCost(unsigned Opcode, Type *Ty, 491 bool IsPairwiseForm) const { 492 int Cost = TTIImpl->getArithmeticReductionCost(Opcode, Ty, IsPairwiseForm); 493 assert(Cost >= 0 && "TTI should not produce negative costs!"); 494 return Cost; 495 } 496 497 int TargetTransformInfo::getMinMaxReductionCost(Type *Ty, Type *CondTy, 498 bool IsPairwiseForm, 499 bool IsUnsigned) const { 500 int Cost = 501 TTIImpl->getMinMaxReductionCost(Ty, CondTy, IsPairwiseForm, IsUnsigned); 502 assert(Cost >= 0 && "TTI should not produce negative costs!"); 503 return Cost; 504 } 505 506 unsigned 507 TargetTransformInfo::getCostOfKeepingLiveOverCall(ArrayRef<Type *> Tys) const { 508 return TTIImpl->getCostOfKeepingLiveOverCall(Tys); 509 } 510 511 bool TargetTransformInfo::getTgtMemIntrinsic(IntrinsicInst *Inst, 512 MemIntrinsicInfo &Info) const { 513 return TTIImpl->getTgtMemIntrinsic(Inst, Info); 514 } 515 516 unsigned TargetTransformInfo::getAtomicMemIntrinsicMaxElementSize() const { 517 return TTIImpl->getAtomicMemIntrinsicMaxElementSize(); 518 } 519 520 Value *TargetTransformInfo::getOrCreateResultFromMemIntrinsic( 521 IntrinsicInst *Inst, Type *ExpectedType) const { 522 return TTIImpl->getOrCreateResultFromMemIntrinsic(Inst, ExpectedType); 523 } 524 525 Type *TargetTransformInfo::getMemcpyLoopLoweringType(LLVMContext &Context, 526 Value *Length, 527 unsigned SrcAlign, 528 unsigned DestAlign) const { 529 return TTIImpl->getMemcpyLoopLoweringType(Context, Length, SrcAlign, 530 DestAlign); 531 } 532 533 void TargetTransformInfo::getMemcpyLoopResidualLoweringType( 534 SmallVectorImpl<Type *> &OpsOut, LLVMContext &Context, 535 unsigned RemainingBytes, unsigned SrcAlign, unsigned DestAlign) const { 536 TTIImpl->getMemcpyLoopResidualLoweringType(OpsOut, Context, RemainingBytes, 537 SrcAlign, DestAlign); 538 } 539 540 bool TargetTransformInfo::useWideIRMemcpyLoopLowering() const { 541 return UseWideMemcpyLoopLowering; 542 } 543 544 bool TargetTransformInfo::areInlineCompatible(const Function *Caller, 545 const Function *Callee) const { 546 return TTIImpl->areInlineCompatible(Caller, Callee); 547 } 548 549 unsigned TargetTransformInfo::getLoadStoreVecRegBitWidth(unsigned AS) const { 550 return TTIImpl->getLoadStoreVecRegBitWidth(AS); 551 } 552 553 bool TargetTransformInfo::isLegalToVectorizeLoad(LoadInst *LI) const { 554 return TTIImpl->isLegalToVectorizeLoad(LI); 555 } 556 557 bool TargetTransformInfo::isLegalToVectorizeStore(StoreInst *SI) const { 558 return TTIImpl->isLegalToVectorizeStore(SI); 559 } 560 561 bool TargetTransformInfo::isLegalToVectorizeLoadChain( 562 unsigned ChainSizeInBytes, unsigned Alignment, unsigned AddrSpace) const { 563 return TTIImpl->isLegalToVectorizeLoadChain(ChainSizeInBytes, Alignment, 564 AddrSpace); 565 } 566 567 bool TargetTransformInfo::isLegalToVectorizeStoreChain( 568 unsigned ChainSizeInBytes, unsigned Alignment, unsigned AddrSpace) const { 569 return TTIImpl->isLegalToVectorizeStoreChain(ChainSizeInBytes, Alignment, 570 AddrSpace); 571 } 572 573 unsigned TargetTransformInfo::getLoadVectorFactor(unsigned VF, 574 unsigned LoadSize, 575 unsigned ChainSizeInBytes, 576 VectorType *VecTy) const { 577 return TTIImpl->getLoadVectorFactor(VF, LoadSize, ChainSizeInBytes, VecTy); 578 } 579 580 unsigned TargetTransformInfo::getStoreVectorFactor(unsigned VF, 581 unsigned StoreSize, 582 unsigned ChainSizeInBytes, 583 VectorType *VecTy) const { 584 return TTIImpl->getStoreVectorFactor(VF, StoreSize, ChainSizeInBytes, VecTy); 585 } 586 587 bool TargetTransformInfo::useReductionIntrinsic(unsigned Opcode, 588 Type *Ty, ReductionFlags Flags) const { 589 return TTIImpl->useReductionIntrinsic(Opcode, Ty, Flags); 590 } 591 592 bool TargetTransformInfo::shouldExpandReduction(const IntrinsicInst *II) const { 593 return TTIImpl->shouldExpandReduction(II); 594 } 595 596 int TargetTransformInfo::getInstructionLatency(const Instruction *I) const { 597 return TTIImpl->getInstructionLatency(I); 598 } 599 600 static bool isReverseVectorMask(ArrayRef<int> Mask) { 601 for (unsigned i = 0, MaskSize = Mask.size(); i < MaskSize; ++i) 602 if (Mask[i] >= 0 && Mask[i] != (int)(MaskSize - 1 - i)) 603 return false; 604 return true; 605 } 606 607 static bool isSingleSourceVectorMask(ArrayRef<int> Mask) { 608 bool Vec0 = false; 609 bool Vec1 = false; 610 for (unsigned i = 0, NumVecElts = Mask.size(); i < NumVecElts; ++i) { 611 if (Mask[i] >= 0) { 612 if ((unsigned)Mask[i] >= NumVecElts) 613 Vec1 = true; 614 else 615 Vec0 = true; 616 } 617 } 618 return !(Vec0 && Vec1); 619 } 620 621 static bool isZeroEltBroadcastVectorMask(ArrayRef<int> Mask) { 622 for (unsigned i = 0; i < Mask.size(); ++i) 623 if (Mask[i] > 0) 624 return false; 625 return true; 626 } 627 628 static bool isAlternateVectorMask(ArrayRef<int> Mask) { 629 bool isAlternate = true; 630 unsigned MaskSize = Mask.size(); 631 632 // Example: shufflevector A, B, <0,5,2,7> 633 for (unsigned i = 0; i < MaskSize && isAlternate; ++i) { 634 if (Mask[i] < 0) 635 continue; 636 isAlternate = Mask[i] == (int)((i & 1) ? MaskSize + i : i); 637 } 638 639 if (isAlternate) 640 return true; 641 642 isAlternate = true; 643 // Example: shufflevector A, B, <4,1,6,3> 644 for (unsigned i = 0; i < MaskSize && isAlternate; ++i) { 645 if (Mask[i] < 0) 646 continue; 647 isAlternate = Mask[i] == (int)((i & 1) ? i : MaskSize + i); 648 } 649 650 return isAlternate; 651 } 652 653 static TargetTransformInfo::OperandValueKind getOperandInfo(Value *V) { 654 TargetTransformInfo::OperandValueKind OpInfo = 655 TargetTransformInfo::OK_AnyValue; 656 657 // Check for a splat of a constant or for a non uniform vector of constants. 658 if (isa<ConstantVector>(V) || isa<ConstantDataVector>(V)) { 659 OpInfo = TargetTransformInfo::OK_NonUniformConstantValue; 660 if (cast<Constant>(V)->getSplatValue() != nullptr) 661 OpInfo = TargetTransformInfo::OK_UniformConstantValue; 662 } 663 664 // Check for a splat of a uniform value. This is not loop aware, so return 665 // true only for the obviously uniform cases (argument, globalvalue) 666 const Value *Splat = getSplatValue(V); 667 if (Splat && (isa<Argument>(Splat) || isa<GlobalValue>(Splat))) 668 OpInfo = TargetTransformInfo::OK_UniformValue; 669 670 return OpInfo; 671 } 672 673 static bool matchPairwiseShuffleMask(ShuffleVectorInst *SI, bool IsLeft, 674 unsigned Level) { 675 // We don't need a shuffle if we just want to have element 0 in position 0 of 676 // the vector. 677 if (!SI && Level == 0 && IsLeft) 678 return true; 679 else if (!SI) 680 return false; 681 682 SmallVector<int, 32> Mask(SI->getType()->getVectorNumElements(), -1); 683 684 // Build a mask of 0, 2, ... (left) or 1, 3, ... (right) depending on whether 685 // we look at the left or right side. 686 for (unsigned i = 0, e = (1 << Level), val = !IsLeft; i != e; ++i, val += 2) 687 Mask[i] = val; 688 689 SmallVector<int, 16> ActualMask = SI->getShuffleMask(); 690 return Mask == ActualMask; 691 } 692 693 namespace { 694 /// Kind of the reduction data. 695 enum ReductionKind { 696 RK_None, /// Not a reduction. 697 RK_Arithmetic, /// Binary reduction data. 698 RK_MinMax, /// Min/max reduction data. 699 RK_UnsignedMinMax, /// Unsigned min/max reduction data. 700 }; 701 /// Contains opcode + LHS/RHS parts of the reduction operations. 702 struct ReductionData { 703 ReductionData() = delete; 704 ReductionData(ReductionKind Kind, unsigned Opcode, Value *LHS, Value *RHS) 705 : Opcode(Opcode), LHS(LHS), RHS(RHS), Kind(Kind) { 706 assert(Kind != RK_None && "expected binary or min/max reduction only."); 707 } 708 unsigned Opcode = 0; 709 Value *LHS = nullptr; 710 Value *RHS = nullptr; 711 ReductionKind Kind = RK_None; 712 bool hasSameData(ReductionData &RD) const { 713 return Kind == RD.Kind && Opcode == RD.Opcode; 714 } 715 }; 716 } // namespace 717 718 static Optional<ReductionData> getReductionData(Instruction *I) { 719 Value *L, *R; 720 if (m_BinOp(m_Value(L), m_Value(R)).match(I)) 721 return ReductionData(RK_Arithmetic, I->getOpcode(), L, R); 722 if (auto *SI = dyn_cast<SelectInst>(I)) { 723 if (m_SMin(m_Value(L), m_Value(R)).match(SI) || 724 m_SMax(m_Value(L), m_Value(R)).match(SI) || 725 m_OrdFMin(m_Value(L), m_Value(R)).match(SI) || 726 m_OrdFMax(m_Value(L), m_Value(R)).match(SI) || 727 m_UnordFMin(m_Value(L), m_Value(R)).match(SI) || 728 m_UnordFMax(m_Value(L), m_Value(R)).match(SI)) { 729 auto *CI = cast<CmpInst>(SI->getCondition()); 730 return ReductionData(RK_MinMax, CI->getOpcode(), L, R); 731 } 732 if (m_UMin(m_Value(L), m_Value(R)).match(SI) || 733 m_UMax(m_Value(L), m_Value(R)).match(SI)) { 734 auto *CI = cast<CmpInst>(SI->getCondition()); 735 return ReductionData(RK_UnsignedMinMax, CI->getOpcode(), L, R); 736 } 737 } 738 return llvm::None; 739 } 740 741 static ReductionKind matchPairwiseReductionAtLevel(Instruction *I, 742 unsigned Level, 743 unsigned NumLevels) { 744 // Match one level of pairwise operations. 745 // %rdx.shuf.0.0 = shufflevector <4 x float> %rdx, <4 x float> undef, 746 // <4 x i32> <i32 0, i32 2 , i32 undef, i32 undef> 747 // %rdx.shuf.0.1 = shufflevector <4 x float> %rdx, <4 x float> undef, 748 // <4 x i32> <i32 1, i32 3, i32 undef, i32 undef> 749 // %bin.rdx.0 = fadd <4 x float> %rdx.shuf.0.0, %rdx.shuf.0.1 750 if (!I) 751 return RK_None; 752 753 assert(I->getType()->isVectorTy() && "Expecting a vector type"); 754 755 Optional<ReductionData> RD = getReductionData(I); 756 if (!RD) 757 return RK_None; 758 759 ShuffleVectorInst *LS = dyn_cast<ShuffleVectorInst>(RD->LHS); 760 if (!LS && Level) 761 return RK_None; 762 ShuffleVectorInst *RS = dyn_cast<ShuffleVectorInst>(RD->RHS); 763 if (!RS && Level) 764 return RK_None; 765 766 // On level 0 we can omit one shufflevector instruction. 767 if (!Level && !RS && !LS) 768 return RK_None; 769 770 // Shuffle inputs must match. 771 Value *NextLevelOpL = LS ? LS->getOperand(0) : nullptr; 772 Value *NextLevelOpR = RS ? RS->getOperand(0) : nullptr; 773 Value *NextLevelOp = nullptr; 774 if (NextLevelOpR && NextLevelOpL) { 775 // If we have two shuffles their operands must match. 776 if (NextLevelOpL != NextLevelOpR) 777 return RK_None; 778 779 NextLevelOp = NextLevelOpL; 780 } else if (Level == 0 && (NextLevelOpR || NextLevelOpL)) { 781 // On the first level we can omit the shufflevector <0, undef,...>. So the 782 // input to the other shufflevector <1, undef> must match with one of the 783 // inputs to the current binary operation. 784 // Example: 785 // %NextLevelOpL = shufflevector %R, <1, undef ...> 786 // %BinOp = fadd %NextLevelOpL, %R 787 if (NextLevelOpL && NextLevelOpL != RD->RHS) 788 return RK_None; 789 else if (NextLevelOpR && NextLevelOpR != RD->LHS) 790 return RK_None; 791 792 NextLevelOp = NextLevelOpL ? RD->RHS : RD->LHS; 793 } else 794 return RK_None; 795 796 // Check that the next levels binary operation exists and matches with the 797 // current one. 798 if (Level + 1 != NumLevels) { 799 Optional<ReductionData> NextLevelRD = 800 getReductionData(cast<Instruction>(NextLevelOp)); 801 if (!NextLevelRD || !RD->hasSameData(*NextLevelRD)) 802 return RK_None; 803 } 804 805 // Shuffle mask for pairwise operation must match. 806 if (matchPairwiseShuffleMask(LS, /*IsLeft=*/true, Level)) { 807 if (!matchPairwiseShuffleMask(RS, /*IsLeft=*/false, Level)) 808 return RK_None; 809 } else if (matchPairwiseShuffleMask(RS, /*IsLeft=*/true, Level)) { 810 if (!matchPairwiseShuffleMask(LS, /*IsLeft=*/false, Level)) 811 return RK_None; 812 } else { 813 return RK_None; 814 } 815 816 if (++Level == NumLevels) 817 return RD->Kind; 818 819 // Match next level. 820 return matchPairwiseReductionAtLevel(cast<Instruction>(NextLevelOp), Level, 821 NumLevels); 822 } 823 824 static ReductionKind matchPairwiseReduction(const ExtractElementInst *ReduxRoot, 825 unsigned &Opcode, Type *&Ty) { 826 if (!EnableReduxCost) 827 return RK_None; 828 829 // Need to extract the first element. 830 ConstantInt *CI = dyn_cast<ConstantInt>(ReduxRoot->getOperand(1)); 831 unsigned Idx = ~0u; 832 if (CI) 833 Idx = CI->getZExtValue(); 834 if (Idx != 0) 835 return RK_None; 836 837 auto *RdxStart = dyn_cast<Instruction>(ReduxRoot->getOperand(0)); 838 if (!RdxStart) 839 return RK_None; 840 Optional<ReductionData> RD = getReductionData(RdxStart); 841 if (!RD) 842 return RK_None; 843 844 Type *VecTy = RdxStart->getType(); 845 unsigned NumVecElems = VecTy->getVectorNumElements(); 846 if (!isPowerOf2_32(NumVecElems)) 847 return RK_None; 848 849 // We look for a sequence of shuffle,shuffle,add triples like the following 850 // that builds a pairwise reduction tree. 851 // 852 // (X0, X1, X2, X3) 853 // (X0 + X1, X2 + X3, undef, undef) 854 // ((X0 + X1) + (X2 + X3), undef, undef, undef) 855 // 856 // %rdx.shuf.0.0 = shufflevector <4 x float> %rdx, <4 x float> undef, 857 // <4 x i32> <i32 0, i32 2 , i32 undef, i32 undef> 858 // %rdx.shuf.0.1 = shufflevector <4 x float> %rdx, <4 x float> undef, 859 // <4 x i32> <i32 1, i32 3, i32 undef, i32 undef> 860 // %bin.rdx.0 = fadd <4 x float> %rdx.shuf.0.0, %rdx.shuf.0.1 861 // %rdx.shuf.1.0 = shufflevector <4 x float> %bin.rdx.0, <4 x float> undef, 862 // <4 x i32> <i32 0, i32 undef, i32 undef, i32 undef> 863 // %rdx.shuf.1.1 = shufflevector <4 x float> %bin.rdx.0, <4 x float> undef, 864 // <4 x i32> <i32 1, i32 undef, i32 undef, i32 undef> 865 // %bin.rdx8 = fadd <4 x float> %rdx.shuf.1.0, %rdx.shuf.1.1 866 // %r = extractelement <4 x float> %bin.rdx8, i32 0 867 if (matchPairwiseReductionAtLevel(RdxStart, 0, Log2_32(NumVecElems)) == 868 RK_None) 869 return RK_None; 870 871 Opcode = RD->Opcode; 872 Ty = VecTy; 873 874 return RD->Kind; 875 } 876 877 static std::pair<Value *, ShuffleVectorInst *> 878 getShuffleAndOtherOprd(Value *L, Value *R) { 879 ShuffleVectorInst *S = nullptr; 880 881 if ((S = dyn_cast<ShuffleVectorInst>(L))) 882 return std::make_pair(R, S); 883 884 S = dyn_cast<ShuffleVectorInst>(R); 885 return std::make_pair(L, S); 886 } 887 888 static ReductionKind 889 matchVectorSplittingReduction(const ExtractElementInst *ReduxRoot, 890 unsigned &Opcode, Type *&Ty) { 891 if (!EnableReduxCost) 892 return RK_None; 893 894 // Need to extract the first element. 895 ConstantInt *CI = dyn_cast<ConstantInt>(ReduxRoot->getOperand(1)); 896 unsigned Idx = ~0u; 897 if (CI) 898 Idx = CI->getZExtValue(); 899 if (Idx != 0) 900 return RK_None; 901 902 auto *RdxStart = dyn_cast<Instruction>(ReduxRoot->getOperand(0)); 903 if (!RdxStart) 904 return RK_None; 905 Optional<ReductionData> RD = getReductionData(RdxStart); 906 if (!RD) 907 return RK_None; 908 909 Type *VecTy = ReduxRoot->getOperand(0)->getType(); 910 unsigned NumVecElems = VecTy->getVectorNumElements(); 911 if (!isPowerOf2_32(NumVecElems)) 912 return RK_None; 913 914 // We look for a sequence of shuffles and adds like the following matching one 915 // fadd, shuffle vector pair at a time. 916 // 917 // %rdx.shuf = shufflevector <4 x float> %rdx, <4 x float> undef, 918 // <4 x i32> <i32 2, i32 3, i32 undef, i32 undef> 919 // %bin.rdx = fadd <4 x float> %rdx, %rdx.shuf 920 // %rdx.shuf7 = shufflevector <4 x float> %bin.rdx, <4 x float> undef, 921 // <4 x i32> <i32 1, i32 undef, i32 undef, i32 undef> 922 // %bin.rdx8 = fadd <4 x float> %bin.rdx, %rdx.shuf7 923 // %r = extractelement <4 x float> %bin.rdx8, i32 0 924 925 unsigned MaskStart = 1; 926 Instruction *RdxOp = RdxStart; 927 SmallVector<int, 32> ShuffleMask(NumVecElems, 0); 928 unsigned NumVecElemsRemain = NumVecElems; 929 while (NumVecElemsRemain - 1) { 930 // Check for the right reduction operation. 931 if (!RdxOp) 932 return RK_None; 933 Optional<ReductionData> RDLevel = getReductionData(RdxOp); 934 if (!RDLevel || !RDLevel->hasSameData(*RD)) 935 return RK_None; 936 937 Value *NextRdxOp; 938 ShuffleVectorInst *Shuffle; 939 std::tie(NextRdxOp, Shuffle) = 940 getShuffleAndOtherOprd(RDLevel->LHS, RDLevel->RHS); 941 942 // Check the current reduction operation and the shuffle use the same value. 943 if (Shuffle == nullptr) 944 return RK_None; 945 if (Shuffle->getOperand(0) != NextRdxOp) 946 return RK_None; 947 948 // Check that shuffle masks matches. 949 for (unsigned j = 0; j != MaskStart; ++j) 950 ShuffleMask[j] = MaskStart + j; 951 // Fill the rest of the mask with -1 for undef. 952 std::fill(&ShuffleMask[MaskStart], ShuffleMask.end(), -1); 953 954 SmallVector<int, 16> Mask = Shuffle->getShuffleMask(); 955 if (ShuffleMask != Mask) 956 return RK_None; 957 958 RdxOp = dyn_cast<Instruction>(NextRdxOp); 959 NumVecElemsRemain /= 2; 960 MaskStart *= 2; 961 } 962 963 Opcode = RD->Opcode; 964 Ty = VecTy; 965 return RD->Kind; 966 } 967 968 int TargetTransformInfo::getInstructionThroughput(const Instruction *I) const { 969 switch (I->getOpcode()) { 970 case Instruction::GetElementPtr: 971 return getUserCost(I); 972 973 case Instruction::Ret: 974 case Instruction::PHI: 975 case Instruction::Br: { 976 return getCFInstrCost(I->getOpcode()); 977 } 978 case Instruction::Add: 979 case Instruction::FAdd: 980 case Instruction::Sub: 981 case Instruction::FSub: 982 case Instruction::Mul: 983 case Instruction::FMul: 984 case Instruction::UDiv: 985 case Instruction::SDiv: 986 case Instruction::FDiv: 987 case Instruction::URem: 988 case Instruction::SRem: 989 case Instruction::FRem: 990 case Instruction::Shl: 991 case Instruction::LShr: 992 case Instruction::AShr: 993 case Instruction::And: 994 case Instruction::Or: 995 case Instruction::Xor: { 996 TargetTransformInfo::OperandValueKind Op1VK = 997 getOperandInfo(I->getOperand(0)); 998 TargetTransformInfo::OperandValueKind Op2VK = 999 getOperandInfo(I->getOperand(1)); 1000 SmallVector<const Value*, 2> Operands(I->operand_values()); 1001 return getArithmeticInstrCost(I->getOpcode(), I->getType(), Op1VK, 1002 Op2VK, TargetTransformInfo::OP_None, 1003 TargetTransformInfo::OP_None, 1004 Operands); 1005 } 1006 case Instruction::Select: { 1007 const SelectInst *SI = cast<SelectInst>(I); 1008 Type *CondTy = SI->getCondition()->getType(); 1009 return getCmpSelInstrCost(I->getOpcode(), I->getType(), CondTy, I); 1010 } 1011 case Instruction::ICmp: 1012 case Instruction::FCmp: { 1013 Type *ValTy = I->getOperand(0)->getType(); 1014 return getCmpSelInstrCost(I->getOpcode(), ValTy, I->getType(), I); 1015 } 1016 case Instruction::Store: { 1017 const StoreInst *SI = cast<StoreInst>(I); 1018 Type *ValTy = SI->getValueOperand()->getType(); 1019 return getMemoryOpCost(I->getOpcode(), ValTy, 1020 SI->getAlignment(), 1021 SI->getPointerAddressSpace(), I); 1022 } 1023 case Instruction::Load: { 1024 const LoadInst *LI = cast<LoadInst>(I); 1025 return getMemoryOpCost(I->getOpcode(), I->getType(), 1026 LI->getAlignment(), 1027 LI->getPointerAddressSpace(), I); 1028 } 1029 case Instruction::ZExt: 1030 case Instruction::SExt: 1031 case Instruction::FPToUI: 1032 case Instruction::FPToSI: 1033 case Instruction::FPExt: 1034 case Instruction::PtrToInt: 1035 case Instruction::IntToPtr: 1036 case Instruction::SIToFP: 1037 case Instruction::UIToFP: 1038 case Instruction::Trunc: 1039 case Instruction::FPTrunc: 1040 case Instruction::BitCast: 1041 case Instruction::AddrSpaceCast: { 1042 Type *SrcTy = I->getOperand(0)->getType(); 1043 return getCastInstrCost(I->getOpcode(), I->getType(), SrcTy, I); 1044 } 1045 case Instruction::ExtractElement: { 1046 const ExtractElementInst * EEI = cast<ExtractElementInst>(I); 1047 ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1)); 1048 unsigned Idx = -1; 1049 if (CI) 1050 Idx = CI->getZExtValue(); 1051 1052 // Try to match a reduction sequence (series of shufflevector and vector 1053 // adds followed by a extractelement). 1054 unsigned ReduxOpCode; 1055 Type *ReduxType; 1056 1057 switch (matchVectorSplittingReduction(EEI, ReduxOpCode, ReduxType)) { 1058 case RK_Arithmetic: 1059 return getArithmeticReductionCost(ReduxOpCode, ReduxType, 1060 /*IsPairwiseForm=*/false); 1061 case RK_MinMax: 1062 return getMinMaxReductionCost( 1063 ReduxType, CmpInst::makeCmpResultType(ReduxType), 1064 /*IsPairwiseForm=*/false, /*IsUnsigned=*/false); 1065 case RK_UnsignedMinMax: 1066 return getMinMaxReductionCost( 1067 ReduxType, CmpInst::makeCmpResultType(ReduxType), 1068 /*IsPairwiseForm=*/false, /*IsUnsigned=*/true); 1069 case RK_None: 1070 break; 1071 } 1072 1073 switch (matchPairwiseReduction(EEI, ReduxOpCode, ReduxType)) { 1074 case RK_Arithmetic: 1075 return getArithmeticReductionCost(ReduxOpCode, ReduxType, 1076 /*IsPairwiseForm=*/true); 1077 case RK_MinMax: 1078 return getMinMaxReductionCost( 1079 ReduxType, CmpInst::makeCmpResultType(ReduxType), 1080 /*IsPairwiseForm=*/true, /*IsUnsigned=*/false); 1081 case RK_UnsignedMinMax: 1082 return getMinMaxReductionCost( 1083 ReduxType, CmpInst::makeCmpResultType(ReduxType), 1084 /*IsPairwiseForm=*/true, /*IsUnsigned=*/true); 1085 case RK_None: 1086 break; 1087 } 1088 1089 return getVectorInstrCost(I->getOpcode(), 1090 EEI->getOperand(0)->getType(), Idx); 1091 } 1092 case Instruction::InsertElement: { 1093 const InsertElementInst * IE = cast<InsertElementInst>(I); 1094 ConstantInt *CI = dyn_cast<ConstantInt>(IE->getOperand(2)); 1095 unsigned Idx = -1; 1096 if (CI) 1097 Idx = CI->getZExtValue(); 1098 return getVectorInstrCost(I->getOpcode(), 1099 IE->getType(), Idx); 1100 } 1101 case Instruction::ShuffleVector: { 1102 const ShuffleVectorInst *Shuffle = cast<ShuffleVectorInst>(I); 1103 Type *VecTypOp0 = Shuffle->getOperand(0)->getType(); 1104 unsigned NumVecElems = VecTypOp0->getVectorNumElements(); 1105 SmallVector<int, 16> Mask = Shuffle->getShuffleMask(); 1106 1107 if (NumVecElems == Mask.size()) { 1108 if (isReverseVectorMask(Mask)) 1109 return getShuffleCost(TargetTransformInfo::SK_Reverse, VecTypOp0, 1110 0, nullptr); 1111 if (isAlternateVectorMask(Mask)) 1112 return getShuffleCost(TargetTransformInfo::SK_Alternate, 1113 VecTypOp0, 0, nullptr); 1114 1115 if (isZeroEltBroadcastVectorMask(Mask)) 1116 return getShuffleCost(TargetTransformInfo::SK_Broadcast, 1117 VecTypOp0, 0, nullptr); 1118 1119 if (isSingleSourceVectorMask(Mask)) 1120 return getShuffleCost(TargetTransformInfo::SK_PermuteSingleSrc, 1121 VecTypOp0, 0, nullptr); 1122 1123 return getShuffleCost(TargetTransformInfo::SK_PermuteTwoSrc, 1124 VecTypOp0, 0, nullptr); 1125 } 1126 1127 return -1; 1128 } 1129 case Instruction::Call: 1130 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) { 1131 SmallVector<Value *, 4> Args(II->arg_operands()); 1132 1133 FastMathFlags FMF; 1134 if (auto *FPMO = dyn_cast<FPMathOperator>(II)) 1135 FMF = FPMO->getFastMathFlags(); 1136 1137 return getIntrinsicInstrCost(II->getIntrinsicID(), II->getType(), 1138 Args, FMF); 1139 } 1140 return -1; 1141 default: 1142 // We don't have any information on this instruction. 1143 return -1; 1144 } 1145 } 1146 1147 TargetTransformInfo::Concept::~Concept() {} 1148 1149 TargetIRAnalysis::TargetIRAnalysis() : TTICallback(&getDefaultTTI) {} 1150 1151 TargetIRAnalysis::TargetIRAnalysis( 1152 std::function<Result(const Function &)> TTICallback) 1153 : TTICallback(std::move(TTICallback)) {} 1154 1155 TargetIRAnalysis::Result TargetIRAnalysis::run(const Function &F, 1156 FunctionAnalysisManager &) { 1157 return TTICallback(F); 1158 } 1159 1160 AnalysisKey TargetIRAnalysis::Key; 1161 1162 TargetIRAnalysis::Result TargetIRAnalysis::getDefaultTTI(const Function &F) { 1163 return Result(F.getParent()->getDataLayout()); 1164 } 1165 1166 // Register the basic pass. 1167 INITIALIZE_PASS(TargetTransformInfoWrapperPass, "tti", 1168 "Target Transform Information", false, true) 1169 char TargetTransformInfoWrapperPass::ID = 0; 1170 1171 void TargetTransformInfoWrapperPass::anchor() {} 1172 1173 TargetTransformInfoWrapperPass::TargetTransformInfoWrapperPass() 1174 : ImmutablePass(ID) { 1175 initializeTargetTransformInfoWrapperPassPass( 1176 *PassRegistry::getPassRegistry()); 1177 } 1178 1179 TargetTransformInfoWrapperPass::TargetTransformInfoWrapperPass( 1180 TargetIRAnalysis TIRA) 1181 : ImmutablePass(ID), TIRA(std::move(TIRA)) { 1182 initializeTargetTransformInfoWrapperPassPass( 1183 *PassRegistry::getPassRegistry()); 1184 } 1185 1186 TargetTransformInfo &TargetTransformInfoWrapperPass::getTTI(const Function &F) { 1187 FunctionAnalysisManager DummyFAM; 1188 TTI = TIRA.run(F, DummyFAM); 1189 return *TTI; 1190 } 1191 1192 ImmutablePass * 1193 llvm::createTargetTransformInfoWrapperPass(TargetIRAnalysis TIRA) { 1194 return new TargetTransformInfoWrapperPass(std::move(TIRA)); 1195 } 1196