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> EnableReduxCost("costmodel-reduxcost", cl::init(false), 30 cl::Hidden, 31 cl::desc("Recognize reduction patterns.")); 32 33 namespace { 34 /// \brief No-op implementation of the TTI interface using the utility base 35 /// classes. 36 /// 37 /// This is used when no target specific information is available. 38 struct NoTTIImpl : TargetTransformInfoImplCRTPBase<NoTTIImpl> { 39 explicit NoTTIImpl(const DataLayout &DL) 40 : TargetTransformInfoImplCRTPBase<NoTTIImpl>(DL) {} 41 }; 42 } 43 44 TargetTransformInfo::TargetTransformInfo(const DataLayout &DL) 45 : TTIImpl(new Model<NoTTIImpl>(NoTTIImpl(DL))) {} 46 47 TargetTransformInfo::~TargetTransformInfo() {} 48 49 TargetTransformInfo::TargetTransformInfo(TargetTransformInfo &&Arg) 50 : TTIImpl(std::move(Arg.TTIImpl)) {} 51 52 TargetTransformInfo &TargetTransformInfo::operator=(TargetTransformInfo &&RHS) { 53 TTIImpl = std::move(RHS.TTIImpl); 54 return *this; 55 } 56 57 int TargetTransformInfo::getOperationCost(unsigned Opcode, Type *Ty, 58 Type *OpTy) const { 59 int Cost = TTIImpl->getOperationCost(Opcode, Ty, OpTy); 60 assert(Cost >= 0 && "TTI should not produce negative costs!"); 61 return Cost; 62 } 63 64 int TargetTransformInfo::getCallCost(FunctionType *FTy, int NumArgs) const { 65 int Cost = TTIImpl->getCallCost(FTy, NumArgs); 66 assert(Cost >= 0 && "TTI should not produce negative costs!"); 67 return Cost; 68 } 69 70 int TargetTransformInfo::getCallCost(const Function *F, 71 ArrayRef<const Value *> Arguments) const { 72 int Cost = TTIImpl->getCallCost(F, Arguments); 73 assert(Cost >= 0 && "TTI should not produce negative costs!"); 74 return Cost; 75 } 76 77 unsigned TargetTransformInfo::getInliningThresholdMultiplier() const { 78 return TTIImpl->getInliningThresholdMultiplier(); 79 } 80 81 int TargetTransformInfo::getGEPCost(Type *PointeeType, const Value *Ptr, 82 ArrayRef<const Value *> Operands) const { 83 return TTIImpl->getGEPCost(PointeeType, Ptr, Operands); 84 } 85 86 int TargetTransformInfo::getExtCost(const Instruction *I, 87 const Value *Src) const { 88 return TTIImpl->getExtCost(I, Src); 89 } 90 91 int TargetTransformInfo::getIntrinsicCost( 92 Intrinsic::ID IID, Type *RetTy, ArrayRef<const Value *> Arguments) const { 93 int Cost = TTIImpl->getIntrinsicCost(IID, RetTy, Arguments); 94 assert(Cost >= 0 && "TTI should not produce negative costs!"); 95 return Cost; 96 } 97 98 unsigned 99 TargetTransformInfo::getEstimatedNumberOfCaseClusters(const SwitchInst &SI, 100 unsigned &JTSize) const { 101 return TTIImpl->getEstimatedNumberOfCaseClusters(SI, JTSize); 102 } 103 104 int TargetTransformInfo::getUserCost(const User *U, 105 ArrayRef<const Value *> Operands) const { 106 int Cost = TTIImpl->getUserCost(U, Operands); 107 assert(Cost >= 0 && "TTI should not produce negative costs!"); 108 return Cost; 109 } 110 111 bool TargetTransformInfo::hasBranchDivergence() const { 112 return TTIImpl->hasBranchDivergence(); 113 } 114 115 bool TargetTransformInfo::isSourceOfDivergence(const Value *V) const { 116 return TTIImpl->isSourceOfDivergence(V); 117 } 118 119 bool llvm::TargetTransformInfo::isAlwaysUniform(const Value *V) const { 120 return TTIImpl->isAlwaysUniform(V); 121 } 122 123 unsigned TargetTransformInfo::getFlatAddressSpace() const { 124 return TTIImpl->getFlatAddressSpace(); 125 } 126 127 bool TargetTransformInfo::isLoweredToCall(const Function *F) const { 128 return TTIImpl->isLoweredToCall(F); 129 } 130 131 void TargetTransformInfo::getUnrollingPreferences( 132 Loop *L, ScalarEvolution &SE, UnrollingPreferences &UP) const { 133 return TTIImpl->getUnrollingPreferences(L, SE, UP); 134 } 135 136 bool TargetTransformInfo::isLegalAddImmediate(int64_t Imm) const { 137 return TTIImpl->isLegalAddImmediate(Imm); 138 } 139 140 bool TargetTransformInfo::isLegalICmpImmediate(int64_t Imm) const { 141 return TTIImpl->isLegalICmpImmediate(Imm); 142 } 143 144 bool TargetTransformInfo::isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV, 145 int64_t BaseOffset, 146 bool HasBaseReg, 147 int64_t Scale, 148 unsigned AddrSpace, 149 Instruction *I) const { 150 return TTIImpl->isLegalAddressingMode(Ty, BaseGV, BaseOffset, HasBaseReg, 151 Scale, AddrSpace, I); 152 } 153 154 bool TargetTransformInfo::isLSRCostLess(LSRCost &C1, LSRCost &C2) const { 155 return TTIImpl->isLSRCostLess(C1, C2); 156 } 157 158 bool TargetTransformInfo::canMacroFuseCmp() const { 159 return TTIImpl->canMacroFuseCmp(); 160 } 161 162 bool TargetTransformInfo::shouldFavorPostInc() const { 163 return TTIImpl->shouldFavorPostInc(); 164 } 165 166 bool TargetTransformInfo::isLegalMaskedStore(Type *DataType) const { 167 return TTIImpl->isLegalMaskedStore(DataType); 168 } 169 170 bool TargetTransformInfo::isLegalMaskedLoad(Type *DataType) const { 171 return TTIImpl->isLegalMaskedLoad(DataType); 172 } 173 174 bool TargetTransformInfo::isLegalMaskedGather(Type *DataType) const { 175 return TTIImpl->isLegalMaskedGather(DataType); 176 } 177 178 bool TargetTransformInfo::isLegalMaskedScatter(Type *DataType) const { 179 return TTIImpl->isLegalMaskedScatter(DataType); 180 } 181 182 bool TargetTransformInfo::hasDivRemOp(Type *DataType, bool IsSigned) const { 183 return TTIImpl->hasDivRemOp(DataType, IsSigned); 184 } 185 186 bool TargetTransformInfo::hasVolatileVariant(Instruction *I, 187 unsigned AddrSpace) const { 188 return TTIImpl->hasVolatileVariant(I, AddrSpace); 189 } 190 191 bool TargetTransformInfo::prefersVectorizedAddressing() const { 192 return TTIImpl->prefersVectorizedAddressing(); 193 } 194 195 int TargetTransformInfo::getScalingFactorCost(Type *Ty, GlobalValue *BaseGV, 196 int64_t BaseOffset, 197 bool HasBaseReg, 198 int64_t Scale, 199 unsigned AddrSpace) const { 200 int Cost = TTIImpl->getScalingFactorCost(Ty, BaseGV, BaseOffset, HasBaseReg, 201 Scale, AddrSpace); 202 assert(Cost >= 0 && "TTI should not produce negative costs!"); 203 return Cost; 204 } 205 206 bool TargetTransformInfo::LSRWithInstrQueries() const { 207 return TTIImpl->LSRWithInstrQueries(); 208 } 209 210 bool TargetTransformInfo::isTruncateFree(Type *Ty1, Type *Ty2) const { 211 return TTIImpl->isTruncateFree(Ty1, Ty2); 212 } 213 214 bool TargetTransformInfo::isProfitableToHoist(Instruction *I) const { 215 return TTIImpl->isProfitableToHoist(I); 216 } 217 218 bool TargetTransformInfo::isTypeLegal(Type *Ty) const { 219 return TTIImpl->isTypeLegal(Ty); 220 } 221 222 unsigned TargetTransformInfo::getJumpBufAlignment() const { 223 return TTIImpl->getJumpBufAlignment(); 224 } 225 226 unsigned TargetTransformInfo::getJumpBufSize() const { 227 return TTIImpl->getJumpBufSize(); 228 } 229 230 bool TargetTransformInfo::shouldBuildLookupTables() const { 231 return TTIImpl->shouldBuildLookupTables(); 232 } 233 bool TargetTransformInfo::shouldBuildLookupTablesForConstant(Constant *C) const { 234 return TTIImpl->shouldBuildLookupTablesForConstant(C); 235 } 236 237 bool TargetTransformInfo::useColdCCForColdCall(Function &F) const { 238 return TTIImpl->useColdCCForColdCall(F); 239 } 240 241 unsigned TargetTransformInfo:: 242 getScalarizationOverhead(Type *Ty, bool Insert, bool Extract) const { 243 return TTIImpl->getScalarizationOverhead(Ty, Insert, Extract); 244 } 245 246 unsigned TargetTransformInfo:: 247 getOperandsScalarizationOverhead(ArrayRef<const Value *> Args, 248 unsigned VF) const { 249 return TTIImpl->getOperandsScalarizationOverhead(Args, VF); 250 } 251 252 bool TargetTransformInfo::supportsEfficientVectorElementLoadStore() const { 253 return TTIImpl->supportsEfficientVectorElementLoadStore(); 254 } 255 256 bool TargetTransformInfo::enableAggressiveInterleaving(bool LoopHasReductions) const { 257 return TTIImpl->enableAggressiveInterleaving(LoopHasReductions); 258 } 259 260 const TargetTransformInfo::MemCmpExpansionOptions * 261 TargetTransformInfo::enableMemCmpExpansion(bool IsZeroCmp) const { 262 return TTIImpl->enableMemCmpExpansion(IsZeroCmp); 263 } 264 265 bool TargetTransformInfo::enableInterleavedAccessVectorization() const { 266 return TTIImpl->enableInterleavedAccessVectorization(); 267 } 268 269 bool TargetTransformInfo::isFPVectorizationPotentiallyUnsafe() const { 270 return TTIImpl->isFPVectorizationPotentiallyUnsafe(); 271 } 272 273 bool TargetTransformInfo::allowsMisalignedMemoryAccesses(LLVMContext &Context, 274 unsigned BitWidth, 275 unsigned AddressSpace, 276 unsigned Alignment, 277 bool *Fast) const { 278 return TTIImpl->allowsMisalignedMemoryAccesses(Context, BitWidth, AddressSpace, 279 Alignment, Fast); 280 } 281 282 TargetTransformInfo::PopcntSupportKind 283 TargetTransformInfo::getPopcntSupport(unsigned IntTyWidthInBit) const { 284 return TTIImpl->getPopcntSupport(IntTyWidthInBit); 285 } 286 287 bool TargetTransformInfo::haveFastSqrt(Type *Ty) const { 288 return TTIImpl->haveFastSqrt(Ty); 289 } 290 291 bool TargetTransformInfo::isFCmpOrdCheaperThanFCmpZero(Type *Ty) const { 292 return TTIImpl->isFCmpOrdCheaperThanFCmpZero(Ty); 293 } 294 295 int TargetTransformInfo::getFPOpCost(Type *Ty) const { 296 int Cost = TTIImpl->getFPOpCost(Ty); 297 assert(Cost >= 0 && "TTI should not produce negative costs!"); 298 return Cost; 299 } 300 301 int TargetTransformInfo::getIntImmCodeSizeCost(unsigned Opcode, unsigned Idx, 302 const APInt &Imm, 303 Type *Ty) const { 304 int Cost = TTIImpl->getIntImmCodeSizeCost(Opcode, Idx, Imm, Ty); 305 assert(Cost >= 0 && "TTI should not produce negative costs!"); 306 return Cost; 307 } 308 309 int TargetTransformInfo::getIntImmCost(const APInt &Imm, Type *Ty) const { 310 int Cost = TTIImpl->getIntImmCost(Imm, Ty); 311 assert(Cost >= 0 && "TTI should not produce negative costs!"); 312 return Cost; 313 } 314 315 int TargetTransformInfo::getIntImmCost(unsigned Opcode, unsigned Idx, 316 const APInt &Imm, Type *Ty) const { 317 int Cost = TTIImpl->getIntImmCost(Opcode, Idx, Imm, Ty); 318 assert(Cost >= 0 && "TTI should not produce negative costs!"); 319 return Cost; 320 } 321 322 int TargetTransformInfo::getIntImmCost(Intrinsic::ID IID, unsigned Idx, 323 const APInt &Imm, Type *Ty) const { 324 int Cost = TTIImpl->getIntImmCost(IID, Idx, Imm, Ty); 325 assert(Cost >= 0 && "TTI should not produce negative costs!"); 326 return Cost; 327 } 328 329 unsigned TargetTransformInfo::getNumberOfRegisters(bool Vector) const { 330 return TTIImpl->getNumberOfRegisters(Vector); 331 } 332 333 unsigned TargetTransformInfo::getRegisterBitWidth(bool Vector) const { 334 return TTIImpl->getRegisterBitWidth(Vector); 335 } 336 337 unsigned TargetTransformInfo::getMinVectorRegisterBitWidth() const { 338 return TTIImpl->getMinVectorRegisterBitWidth(); 339 } 340 341 bool TargetTransformInfo::shouldConsiderAddressTypePromotion( 342 const Instruction &I, bool &AllowPromotionWithoutCommonHeader) const { 343 return TTIImpl->shouldConsiderAddressTypePromotion( 344 I, AllowPromotionWithoutCommonHeader); 345 } 346 347 unsigned TargetTransformInfo::getCacheLineSize() const { 348 return TTIImpl->getCacheLineSize(); 349 } 350 351 llvm::Optional<unsigned> TargetTransformInfo::getCacheSize(CacheLevel Level) 352 const { 353 return TTIImpl->getCacheSize(Level); 354 } 355 356 llvm::Optional<unsigned> TargetTransformInfo::getCacheAssociativity( 357 CacheLevel Level) const { 358 return TTIImpl->getCacheAssociativity(Level); 359 } 360 361 unsigned TargetTransformInfo::getPrefetchDistance() const { 362 return TTIImpl->getPrefetchDistance(); 363 } 364 365 unsigned TargetTransformInfo::getMinPrefetchStride() const { 366 return TTIImpl->getMinPrefetchStride(); 367 } 368 369 unsigned TargetTransformInfo::getMaxPrefetchIterationsAhead() const { 370 return TTIImpl->getMaxPrefetchIterationsAhead(); 371 } 372 373 unsigned TargetTransformInfo::getMaxInterleaveFactor(unsigned VF) const { 374 return TTIImpl->getMaxInterleaveFactor(VF); 375 } 376 377 int TargetTransformInfo::getArithmeticInstrCost( 378 unsigned Opcode, Type *Ty, OperandValueKind Opd1Info, 379 OperandValueKind Opd2Info, OperandValueProperties Opd1PropInfo, 380 OperandValueProperties Opd2PropInfo, 381 ArrayRef<const Value *> Args) const { 382 int Cost = TTIImpl->getArithmeticInstrCost(Opcode, Ty, Opd1Info, Opd2Info, 383 Opd1PropInfo, Opd2PropInfo, Args); 384 assert(Cost >= 0 && "TTI should not produce negative costs!"); 385 return Cost; 386 } 387 388 int TargetTransformInfo::getShuffleCost(ShuffleKind Kind, Type *Ty, int Index, 389 Type *SubTp) const { 390 int Cost = TTIImpl->getShuffleCost(Kind, Ty, Index, SubTp); 391 assert(Cost >= 0 && "TTI should not produce negative costs!"); 392 return Cost; 393 } 394 395 int TargetTransformInfo::getCastInstrCost(unsigned Opcode, Type *Dst, 396 Type *Src, const Instruction *I) const { 397 assert ((I == nullptr || I->getOpcode() == Opcode) && 398 "Opcode should reflect passed instruction."); 399 int Cost = TTIImpl->getCastInstrCost(Opcode, Dst, Src, I); 400 assert(Cost >= 0 && "TTI should not produce negative costs!"); 401 return Cost; 402 } 403 404 int TargetTransformInfo::getExtractWithExtendCost(unsigned Opcode, Type *Dst, 405 VectorType *VecTy, 406 unsigned Index) const { 407 int Cost = TTIImpl->getExtractWithExtendCost(Opcode, Dst, VecTy, Index); 408 assert(Cost >= 0 && "TTI should not produce negative costs!"); 409 return Cost; 410 } 411 412 int TargetTransformInfo::getCFInstrCost(unsigned Opcode) const { 413 int Cost = TTIImpl->getCFInstrCost(Opcode); 414 assert(Cost >= 0 && "TTI should not produce negative costs!"); 415 return Cost; 416 } 417 418 int TargetTransformInfo::getCmpSelInstrCost(unsigned Opcode, Type *ValTy, 419 Type *CondTy, const Instruction *I) const { 420 assert ((I == nullptr || I->getOpcode() == Opcode) && 421 "Opcode should reflect passed instruction."); 422 int Cost = TTIImpl->getCmpSelInstrCost(Opcode, ValTy, CondTy, I); 423 assert(Cost >= 0 && "TTI should not produce negative costs!"); 424 return Cost; 425 } 426 427 int TargetTransformInfo::getVectorInstrCost(unsigned Opcode, Type *Val, 428 unsigned Index) const { 429 int Cost = TTIImpl->getVectorInstrCost(Opcode, Val, Index); 430 assert(Cost >= 0 && "TTI should not produce negative costs!"); 431 return Cost; 432 } 433 434 int TargetTransformInfo::getMemoryOpCost(unsigned Opcode, Type *Src, 435 unsigned Alignment, 436 unsigned AddressSpace, 437 const Instruction *I) const { 438 assert ((I == nullptr || I->getOpcode() == Opcode) && 439 "Opcode should reflect passed instruction."); 440 int Cost = TTIImpl->getMemoryOpCost(Opcode, Src, Alignment, AddressSpace, I); 441 assert(Cost >= 0 && "TTI should not produce negative costs!"); 442 return Cost; 443 } 444 445 int TargetTransformInfo::getMaskedMemoryOpCost(unsigned Opcode, Type *Src, 446 unsigned Alignment, 447 unsigned AddressSpace) const { 448 int Cost = 449 TTIImpl->getMaskedMemoryOpCost(Opcode, Src, Alignment, AddressSpace); 450 assert(Cost >= 0 && "TTI should not produce negative costs!"); 451 return Cost; 452 } 453 454 int TargetTransformInfo::getGatherScatterOpCost(unsigned Opcode, Type *DataTy, 455 Value *Ptr, bool VariableMask, 456 unsigned Alignment) const { 457 int Cost = TTIImpl->getGatherScatterOpCost(Opcode, DataTy, Ptr, VariableMask, 458 Alignment); 459 assert(Cost >= 0 && "TTI should not produce negative costs!"); 460 return Cost; 461 } 462 463 int TargetTransformInfo::getInterleavedMemoryOpCost( 464 unsigned Opcode, Type *VecTy, unsigned Factor, ArrayRef<unsigned> Indices, 465 unsigned Alignment, unsigned AddressSpace) const { 466 int Cost = TTIImpl->getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices, 467 Alignment, AddressSpace); 468 assert(Cost >= 0 && "TTI should not produce negative costs!"); 469 return Cost; 470 } 471 472 int TargetTransformInfo::getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy, 473 ArrayRef<Type *> Tys, FastMathFlags FMF, 474 unsigned ScalarizationCostPassed) const { 475 int Cost = TTIImpl->getIntrinsicInstrCost(ID, RetTy, Tys, FMF, 476 ScalarizationCostPassed); 477 assert(Cost >= 0 && "TTI should not produce negative costs!"); 478 return Cost; 479 } 480 481 int TargetTransformInfo::getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy, 482 ArrayRef<Value *> Args, FastMathFlags FMF, unsigned VF) const { 483 int Cost = TTIImpl->getIntrinsicInstrCost(ID, RetTy, Args, FMF, VF); 484 assert(Cost >= 0 && "TTI should not produce negative costs!"); 485 return Cost; 486 } 487 488 int TargetTransformInfo::getCallInstrCost(Function *F, Type *RetTy, 489 ArrayRef<Type *> Tys) const { 490 int Cost = TTIImpl->getCallInstrCost(F, RetTy, Tys); 491 assert(Cost >= 0 && "TTI should not produce negative costs!"); 492 return Cost; 493 } 494 495 unsigned TargetTransformInfo::getNumberOfParts(Type *Tp) const { 496 return TTIImpl->getNumberOfParts(Tp); 497 } 498 499 int TargetTransformInfo::getAddressComputationCost(Type *Tp, 500 ScalarEvolution *SE, 501 const SCEV *Ptr) const { 502 int Cost = TTIImpl->getAddressComputationCost(Tp, SE, Ptr); 503 assert(Cost >= 0 && "TTI should not produce negative costs!"); 504 return Cost; 505 } 506 507 int TargetTransformInfo::getArithmeticReductionCost(unsigned Opcode, Type *Ty, 508 bool IsPairwiseForm) const { 509 int Cost = TTIImpl->getArithmeticReductionCost(Opcode, Ty, IsPairwiseForm); 510 assert(Cost >= 0 && "TTI should not produce negative costs!"); 511 return Cost; 512 } 513 514 int TargetTransformInfo::getMinMaxReductionCost(Type *Ty, Type *CondTy, 515 bool IsPairwiseForm, 516 bool IsUnsigned) const { 517 int Cost = 518 TTIImpl->getMinMaxReductionCost(Ty, CondTy, IsPairwiseForm, IsUnsigned); 519 assert(Cost >= 0 && "TTI should not produce negative costs!"); 520 return Cost; 521 } 522 523 unsigned 524 TargetTransformInfo::getCostOfKeepingLiveOverCall(ArrayRef<Type *> Tys) const { 525 return TTIImpl->getCostOfKeepingLiveOverCall(Tys); 526 } 527 528 bool TargetTransformInfo::getTgtMemIntrinsic(IntrinsicInst *Inst, 529 MemIntrinsicInfo &Info) const { 530 return TTIImpl->getTgtMemIntrinsic(Inst, Info); 531 } 532 533 unsigned TargetTransformInfo::getAtomicMemIntrinsicMaxElementSize() const { 534 return TTIImpl->getAtomicMemIntrinsicMaxElementSize(); 535 } 536 537 Value *TargetTransformInfo::getOrCreateResultFromMemIntrinsic( 538 IntrinsicInst *Inst, Type *ExpectedType) const { 539 return TTIImpl->getOrCreateResultFromMemIntrinsic(Inst, ExpectedType); 540 } 541 542 Type *TargetTransformInfo::getMemcpyLoopLoweringType(LLVMContext &Context, 543 Value *Length, 544 unsigned SrcAlign, 545 unsigned DestAlign) const { 546 return TTIImpl->getMemcpyLoopLoweringType(Context, Length, SrcAlign, 547 DestAlign); 548 } 549 550 void TargetTransformInfo::getMemcpyLoopResidualLoweringType( 551 SmallVectorImpl<Type *> &OpsOut, LLVMContext &Context, 552 unsigned RemainingBytes, unsigned SrcAlign, unsigned DestAlign) const { 553 TTIImpl->getMemcpyLoopResidualLoweringType(OpsOut, Context, RemainingBytes, 554 SrcAlign, DestAlign); 555 } 556 557 bool TargetTransformInfo::areInlineCompatible(const Function *Caller, 558 const Function *Callee) const { 559 return TTIImpl->areInlineCompatible(Caller, Callee); 560 } 561 562 bool TargetTransformInfo::isIndexedLoadLegal(MemIndexedMode Mode, 563 Type *Ty) const { 564 return TTIImpl->isIndexedLoadLegal(Mode, Ty); 565 } 566 567 bool TargetTransformInfo::isIndexedStoreLegal(MemIndexedMode Mode, 568 Type *Ty) const { 569 return TTIImpl->isIndexedStoreLegal(Mode, Ty); 570 } 571 572 unsigned TargetTransformInfo::getLoadStoreVecRegBitWidth(unsigned AS) const { 573 return TTIImpl->getLoadStoreVecRegBitWidth(AS); 574 } 575 576 bool TargetTransformInfo::isLegalToVectorizeLoad(LoadInst *LI) const { 577 return TTIImpl->isLegalToVectorizeLoad(LI); 578 } 579 580 bool TargetTransformInfo::isLegalToVectorizeStore(StoreInst *SI) const { 581 return TTIImpl->isLegalToVectorizeStore(SI); 582 } 583 584 bool TargetTransformInfo::isLegalToVectorizeLoadChain( 585 unsigned ChainSizeInBytes, unsigned Alignment, unsigned AddrSpace) const { 586 return TTIImpl->isLegalToVectorizeLoadChain(ChainSizeInBytes, Alignment, 587 AddrSpace); 588 } 589 590 bool TargetTransformInfo::isLegalToVectorizeStoreChain( 591 unsigned ChainSizeInBytes, unsigned Alignment, unsigned AddrSpace) const { 592 return TTIImpl->isLegalToVectorizeStoreChain(ChainSizeInBytes, Alignment, 593 AddrSpace); 594 } 595 596 unsigned TargetTransformInfo::getLoadVectorFactor(unsigned VF, 597 unsigned LoadSize, 598 unsigned ChainSizeInBytes, 599 VectorType *VecTy) const { 600 return TTIImpl->getLoadVectorFactor(VF, LoadSize, ChainSizeInBytes, VecTy); 601 } 602 603 unsigned TargetTransformInfo::getStoreVectorFactor(unsigned VF, 604 unsigned StoreSize, 605 unsigned ChainSizeInBytes, 606 VectorType *VecTy) const { 607 return TTIImpl->getStoreVectorFactor(VF, StoreSize, ChainSizeInBytes, VecTy); 608 } 609 610 bool TargetTransformInfo::useReductionIntrinsic(unsigned Opcode, 611 Type *Ty, ReductionFlags Flags) const { 612 return TTIImpl->useReductionIntrinsic(Opcode, Ty, Flags); 613 } 614 615 bool TargetTransformInfo::shouldExpandReduction(const IntrinsicInst *II) const { 616 return TTIImpl->shouldExpandReduction(II); 617 } 618 619 int TargetTransformInfo::getInstructionLatency(const Instruction *I) const { 620 return TTIImpl->getInstructionLatency(I); 621 } 622 623 static bool isReverseVectorMask(ArrayRef<int> Mask) { 624 for (unsigned i = 0, MaskSize = Mask.size(); i < MaskSize; ++i) 625 if (Mask[i] >= 0 && Mask[i] != (int)(MaskSize - 1 - i)) 626 return false; 627 return true; 628 } 629 630 static bool isSingleSourceVectorMask(ArrayRef<int> Mask) { 631 bool Vec0 = false; 632 bool Vec1 = false; 633 for (unsigned i = 0, NumVecElts = Mask.size(); i < NumVecElts; ++i) { 634 if (Mask[i] >= 0) { 635 if ((unsigned)Mask[i] >= NumVecElts) 636 Vec1 = true; 637 else 638 Vec0 = true; 639 } 640 } 641 return !(Vec0 && Vec1); 642 } 643 644 static bool isZeroEltBroadcastVectorMask(ArrayRef<int> Mask) { 645 for (unsigned i = 0; i < Mask.size(); ++i) 646 if (Mask[i] > 0) 647 return false; 648 return true; 649 } 650 651 static bool isAlternateVectorMask(ArrayRef<int> Mask) { 652 bool isAlternate = true; 653 unsigned MaskSize = Mask.size(); 654 655 // Example: shufflevector A, B, <0,5,2,7> 656 for (unsigned i = 0; i < MaskSize && isAlternate; ++i) { 657 if (Mask[i] < 0) 658 continue; 659 isAlternate = Mask[i] == (int)((i & 1) ? MaskSize + i : i); 660 } 661 662 if (isAlternate) 663 return true; 664 665 isAlternate = true; 666 // Example: shufflevector A, B, <4,1,6,3> 667 for (unsigned i = 0; i < MaskSize && isAlternate; ++i) { 668 if (Mask[i] < 0) 669 continue; 670 isAlternate = Mask[i] == (int)((i & 1) ? i : MaskSize + i); 671 } 672 673 return isAlternate; 674 } 675 676 static TargetTransformInfo::OperandValueKind getOperandInfo(Value *V) { 677 TargetTransformInfo::OperandValueKind OpInfo = 678 TargetTransformInfo::OK_AnyValue; 679 680 // Check for a splat of a constant or for a non uniform vector of constants. 681 if (isa<ConstantVector>(V) || isa<ConstantDataVector>(V)) { 682 OpInfo = TargetTransformInfo::OK_NonUniformConstantValue; 683 if (cast<Constant>(V)->getSplatValue() != nullptr) 684 OpInfo = TargetTransformInfo::OK_UniformConstantValue; 685 } 686 687 // Check for a splat of a uniform value. This is not loop aware, so return 688 // true only for the obviously uniform cases (argument, globalvalue) 689 const Value *Splat = getSplatValue(V); 690 if (Splat && (isa<Argument>(Splat) || isa<GlobalValue>(Splat))) 691 OpInfo = TargetTransformInfo::OK_UniformValue; 692 693 return OpInfo; 694 } 695 696 static bool matchPairwiseShuffleMask(ShuffleVectorInst *SI, bool IsLeft, 697 unsigned Level) { 698 // We don't need a shuffle if we just want to have element 0 in position 0 of 699 // the vector. 700 if (!SI && Level == 0 && IsLeft) 701 return true; 702 else if (!SI) 703 return false; 704 705 SmallVector<int, 32> Mask(SI->getType()->getVectorNumElements(), -1); 706 707 // Build a mask of 0, 2, ... (left) or 1, 3, ... (right) depending on whether 708 // we look at the left or right side. 709 for (unsigned i = 0, e = (1 << Level), val = !IsLeft; i != e; ++i, val += 2) 710 Mask[i] = val; 711 712 SmallVector<int, 16> ActualMask = SI->getShuffleMask(); 713 return Mask == ActualMask; 714 } 715 716 namespace { 717 /// Kind of the reduction data. 718 enum ReductionKind { 719 RK_None, /// Not a reduction. 720 RK_Arithmetic, /// Binary reduction data. 721 RK_MinMax, /// Min/max reduction data. 722 RK_UnsignedMinMax, /// Unsigned min/max reduction data. 723 }; 724 /// Contains opcode + LHS/RHS parts of the reduction operations. 725 struct ReductionData { 726 ReductionData() = delete; 727 ReductionData(ReductionKind Kind, unsigned Opcode, Value *LHS, Value *RHS) 728 : Opcode(Opcode), LHS(LHS), RHS(RHS), Kind(Kind) { 729 assert(Kind != RK_None && "expected binary or min/max reduction only."); 730 } 731 unsigned Opcode = 0; 732 Value *LHS = nullptr; 733 Value *RHS = nullptr; 734 ReductionKind Kind = RK_None; 735 bool hasSameData(ReductionData &RD) const { 736 return Kind == RD.Kind && Opcode == RD.Opcode; 737 } 738 }; 739 } // namespace 740 741 static Optional<ReductionData> getReductionData(Instruction *I) { 742 Value *L, *R; 743 if (m_BinOp(m_Value(L), m_Value(R)).match(I)) 744 return ReductionData(RK_Arithmetic, I->getOpcode(), L, R); 745 if (auto *SI = dyn_cast<SelectInst>(I)) { 746 if (m_SMin(m_Value(L), m_Value(R)).match(SI) || 747 m_SMax(m_Value(L), m_Value(R)).match(SI) || 748 m_OrdFMin(m_Value(L), m_Value(R)).match(SI) || 749 m_OrdFMax(m_Value(L), m_Value(R)).match(SI) || 750 m_UnordFMin(m_Value(L), m_Value(R)).match(SI) || 751 m_UnordFMax(m_Value(L), m_Value(R)).match(SI)) { 752 auto *CI = cast<CmpInst>(SI->getCondition()); 753 return ReductionData(RK_MinMax, CI->getOpcode(), L, R); 754 } 755 if (m_UMin(m_Value(L), m_Value(R)).match(SI) || 756 m_UMax(m_Value(L), m_Value(R)).match(SI)) { 757 auto *CI = cast<CmpInst>(SI->getCondition()); 758 return ReductionData(RK_UnsignedMinMax, CI->getOpcode(), L, R); 759 } 760 } 761 return llvm::None; 762 } 763 764 static ReductionKind matchPairwiseReductionAtLevel(Instruction *I, 765 unsigned Level, 766 unsigned NumLevels) { 767 // Match one level of pairwise operations. 768 // %rdx.shuf.0.0 = shufflevector <4 x float> %rdx, <4 x float> undef, 769 // <4 x i32> <i32 0, i32 2 , i32 undef, i32 undef> 770 // %rdx.shuf.0.1 = shufflevector <4 x float> %rdx, <4 x float> undef, 771 // <4 x i32> <i32 1, i32 3, i32 undef, i32 undef> 772 // %bin.rdx.0 = fadd <4 x float> %rdx.shuf.0.0, %rdx.shuf.0.1 773 if (!I) 774 return RK_None; 775 776 assert(I->getType()->isVectorTy() && "Expecting a vector type"); 777 778 Optional<ReductionData> RD = getReductionData(I); 779 if (!RD) 780 return RK_None; 781 782 ShuffleVectorInst *LS = dyn_cast<ShuffleVectorInst>(RD->LHS); 783 if (!LS && Level) 784 return RK_None; 785 ShuffleVectorInst *RS = dyn_cast<ShuffleVectorInst>(RD->RHS); 786 if (!RS && Level) 787 return RK_None; 788 789 // On level 0 we can omit one shufflevector instruction. 790 if (!Level && !RS && !LS) 791 return RK_None; 792 793 // Shuffle inputs must match. 794 Value *NextLevelOpL = LS ? LS->getOperand(0) : nullptr; 795 Value *NextLevelOpR = RS ? RS->getOperand(0) : nullptr; 796 Value *NextLevelOp = nullptr; 797 if (NextLevelOpR && NextLevelOpL) { 798 // If we have two shuffles their operands must match. 799 if (NextLevelOpL != NextLevelOpR) 800 return RK_None; 801 802 NextLevelOp = NextLevelOpL; 803 } else if (Level == 0 && (NextLevelOpR || NextLevelOpL)) { 804 // On the first level we can omit the shufflevector <0, undef,...>. So the 805 // input to the other shufflevector <1, undef> must match with one of the 806 // inputs to the current binary operation. 807 // Example: 808 // %NextLevelOpL = shufflevector %R, <1, undef ...> 809 // %BinOp = fadd %NextLevelOpL, %R 810 if (NextLevelOpL && NextLevelOpL != RD->RHS) 811 return RK_None; 812 else if (NextLevelOpR && NextLevelOpR != RD->LHS) 813 return RK_None; 814 815 NextLevelOp = NextLevelOpL ? RD->RHS : RD->LHS; 816 } else 817 return RK_None; 818 819 // Check that the next levels binary operation exists and matches with the 820 // current one. 821 if (Level + 1 != NumLevels) { 822 Optional<ReductionData> NextLevelRD = 823 getReductionData(cast<Instruction>(NextLevelOp)); 824 if (!NextLevelRD || !RD->hasSameData(*NextLevelRD)) 825 return RK_None; 826 } 827 828 // Shuffle mask for pairwise operation must match. 829 if (matchPairwiseShuffleMask(LS, /*IsLeft=*/true, Level)) { 830 if (!matchPairwiseShuffleMask(RS, /*IsLeft=*/false, Level)) 831 return RK_None; 832 } else if (matchPairwiseShuffleMask(RS, /*IsLeft=*/true, Level)) { 833 if (!matchPairwiseShuffleMask(LS, /*IsLeft=*/false, Level)) 834 return RK_None; 835 } else { 836 return RK_None; 837 } 838 839 if (++Level == NumLevels) 840 return RD->Kind; 841 842 // Match next level. 843 return matchPairwiseReductionAtLevel(cast<Instruction>(NextLevelOp), Level, 844 NumLevels); 845 } 846 847 static ReductionKind matchPairwiseReduction(const ExtractElementInst *ReduxRoot, 848 unsigned &Opcode, Type *&Ty) { 849 if (!EnableReduxCost) 850 return RK_None; 851 852 // Need to extract the first element. 853 ConstantInt *CI = dyn_cast<ConstantInt>(ReduxRoot->getOperand(1)); 854 unsigned Idx = ~0u; 855 if (CI) 856 Idx = CI->getZExtValue(); 857 if (Idx != 0) 858 return RK_None; 859 860 auto *RdxStart = dyn_cast<Instruction>(ReduxRoot->getOperand(0)); 861 if (!RdxStart) 862 return RK_None; 863 Optional<ReductionData> RD = getReductionData(RdxStart); 864 if (!RD) 865 return RK_None; 866 867 Type *VecTy = RdxStart->getType(); 868 unsigned NumVecElems = VecTy->getVectorNumElements(); 869 if (!isPowerOf2_32(NumVecElems)) 870 return RK_None; 871 872 // We look for a sequence of shuffle,shuffle,add triples like the following 873 // that builds a pairwise reduction tree. 874 // 875 // (X0, X1, X2, X3) 876 // (X0 + X1, X2 + X3, undef, undef) 877 // ((X0 + X1) + (X2 + X3), undef, undef, undef) 878 // 879 // %rdx.shuf.0.0 = shufflevector <4 x float> %rdx, <4 x float> undef, 880 // <4 x i32> <i32 0, i32 2 , i32 undef, i32 undef> 881 // %rdx.shuf.0.1 = shufflevector <4 x float> %rdx, <4 x float> undef, 882 // <4 x i32> <i32 1, i32 3, i32 undef, i32 undef> 883 // %bin.rdx.0 = fadd <4 x float> %rdx.shuf.0.0, %rdx.shuf.0.1 884 // %rdx.shuf.1.0 = shufflevector <4 x float> %bin.rdx.0, <4 x float> undef, 885 // <4 x i32> <i32 0, i32 undef, i32 undef, i32 undef> 886 // %rdx.shuf.1.1 = shufflevector <4 x float> %bin.rdx.0, <4 x float> undef, 887 // <4 x i32> <i32 1, i32 undef, i32 undef, i32 undef> 888 // %bin.rdx8 = fadd <4 x float> %rdx.shuf.1.0, %rdx.shuf.1.1 889 // %r = extractelement <4 x float> %bin.rdx8, i32 0 890 if (matchPairwiseReductionAtLevel(RdxStart, 0, Log2_32(NumVecElems)) == 891 RK_None) 892 return RK_None; 893 894 Opcode = RD->Opcode; 895 Ty = VecTy; 896 897 return RD->Kind; 898 } 899 900 static std::pair<Value *, ShuffleVectorInst *> 901 getShuffleAndOtherOprd(Value *L, Value *R) { 902 ShuffleVectorInst *S = nullptr; 903 904 if ((S = dyn_cast<ShuffleVectorInst>(L))) 905 return std::make_pair(R, S); 906 907 S = dyn_cast<ShuffleVectorInst>(R); 908 return std::make_pair(L, S); 909 } 910 911 static ReductionKind 912 matchVectorSplittingReduction(const ExtractElementInst *ReduxRoot, 913 unsigned &Opcode, Type *&Ty) { 914 if (!EnableReduxCost) 915 return RK_None; 916 917 // Need to extract the first element. 918 ConstantInt *CI = dyn_cast<ConstantInt>(ReduxRoot->getOperand(1)); 919 unsigned Idx = ~0u; 920 if (CI) 921 Idx = CI->getZExtValue(); 922 if (Idx != 0) 923 return RK_None; 924 925 auto *RdxStart = dyn_cast<Instruction>(ReduxRoot->getOperand(0)); 926 if (!RdxStart) 927 return RK_None; 928 Optional<ReductionData> RD = getReductionData(RdxStart); 929 if (!RD) 930 return RK_None; 931 932 Type *VecTy = ReduxRoot->getOperand(0)->getType(); 933 unsigned NumVecElems = VecTy->getVectorNumElements(); 934 if (!isPowerOf2_32(NumVecElems)) 935 return RK_None; 936 937 // We look for a sequence of shuffles and adds like the following matching one 938 // fadd, shuffle vector pair at a time. 939 // 940 // %rdx.shuf = shufflevector <4 x float> %rdx, <4 x float> undef, 941 // <4 x i32> <i32 2, i32 3, i32 undef, i32 undef> 942 // %bin.rdx = fadd <4 x float> %rdx, %rdx.shuf 943 // %rdx.shuf7 = shufflevector <4 x float> %bin.rdx, <4 x float> undef, 944 // <4 x i32> <i32 1, i32 undef, i32 undef, i32 undef> 945 // %bin.rdx8 = fadd <4 x float> %bin.rdx, %rdx.shuf7 946 // %r = extractelement <4 x float> %bin.rdx8, i32 0 947 948 unsigned MaskStart = 1; 949 Instruction *RdxOp = RdxStart; 950 SmallVector<int, 32> ShuffleMask(NumVecElems, 0); 951 unsigned NumVecElemsRemain = NumVecElems; 952 while (NumVecElemsRemain - 1) { 953 // Check for the right reduction operation. 954 if (!RdxOp) 955 return RK_None; 956 Optional<ReductionData> RDLevel = getReductionData(RdxOp); 957 if (!RDLevel || !RDLevel->hasSameData(*RD)) 958 return RK_None; 959 960 Value *NextRdxOp; 961 ShuffleVectorInst *Shuffle; 962 std::tie(NextRdxOp, Shuffle) = 963 getShuffleAndOtherOprd(RDLevel->LHS, RDLevel->RHS); 964 965 // Check the current reduction operation and the shuffle use the same value. 966 if (Shuffle == nullptr) 967 return RK_None; 968 if (Shuffle->getOperand(0) != NextRdxOp) 969 return RK_None; 970 971 // Check that shuffle masks matches. 972 for (unsigned j = 0; j != MaskStart; ++j) 973 ShuffleMask[j] = MaskStart + j; 974 // Fill the rest of the mask with -1 for undef. 975 std::fill(&ShuffleMask[MaskStart], ShuffleMask.end(), -1); 976 977 SmallVector<int, 16> Mask = Shuffle->getShuffleMask(); 978 if (ShuffleMask != Mask) 979 return RK_None; 980 981 RdxOp = dyn_cast<Instruction>(NextRdxOp); 982 NumVecElemsRemain /= 2; 983 MaskStart *= 2; 984 } 985 986 Opcode = RD->Opcode; 987 Ty = VecTy; 988 return RD->Kind; 989 } 990 991 int TargetTransformInfo::getInstructionThroughput(const Instruction *I) const { 992 switch (I->getOpcode()) { 993 case Instruction::GetElementPtr: 994 return getUserCost(I); 995 996 case Instruction::Ret: 997 case Instruction::PHI: 998 case Instruction::Br: { 999 return getCFInstrCost(I->getOpcode()); 1000 } 1001 case Instruction::Add: 1002 case Instruction::FAdd: 1003 case Instruction::Sub: 1004 case Instruction::FSub: 1005 case Instruction::Mul: 1006 case Instruction::FMul: 1007 case Instruction::UDiv: 1008 case Instruction::SDiv: 1009 case Instruction::FDiv: 1010 case Instruction::URem: 1011 case Instruction::SRem: 1012 case Instruction::FRem: 1013 case Instruction::Shl: 1014 case Instruction::LShr: 1015 case Instruction::AShr: 1016 case Instruction::And: 1017 case Instruction::Or: 1018 case Instruction::Xor: { 1019 TargetTransformInfo::OperandValueKind Op1VK = 1020 getOperandInfo(I->getOperand(0)); 1021 TargetTransformInfo::OperandValueKind Op2VK = 1022 getOperandInfo(I->getOperand(1)); 1023 SmallVector<const Value*, 2> Operands(I->operand_values()); 1024 return getArithmeticInstrCost(I->getOpcode(), I->getType(), Op1VK, 1025 Op2VK, TargetTransformInfo::OP_None, 1026 TargetTransformInfo::OP_None, 1027 Operands); 1028 } 1029 case Instruction::Select: { 1030 const SelectInst *SI = cast<SelectInst>(I); 1031 Type *CondTy = SI->getCondition()->getType(); 1032 return getCmpSelInstrCost(I->getOpcode(), I->getType(), CondTy, I); 1033 } 1034 case Instruction::ICmp: 1035 case Instruction::FCmp: { 1036 Type *ValTy = I->getOperand(0)->getType(); 1037 return getCmpSelInstrCost(I->getOpcode(), ValTy, I->getType(), I); 1038 } 1039 case Instruction::Store: { 1040 const StoreInst *SI = cast<StoreInst>(I); 1041 Type *ValTy = SI->getValueOperand()->getType(); 1042 return getMemoryOpCost(I->getOpcode(), ValTy, 1043 SI->getAlignment(), 1044 SI->getPointerAddressSpace(), I); 1045 } 1046 case Instruction::Load: { 1047 const LoadInst *LI = cast<LoadInst>(I); 1048 return getMemoryOpCost(I->getOpcode(), I->getType(), 1049 LI->getAlignment(), 1050 LI->getPointerAddressSpace(), I); 1051 } 1052 case Instruction::ZExt: 1053 case Instruction::SExt: 1054 case Instruction::FPToUI: 1055 case Instruction::FPToSI: 1056 case Instruction::FPExt: 1057 case Instruction::PtrToInt: 1058 case Instruction::IntToPtr: 1059 case Instruction::SIToFP: 1060 case Instruction::UIToFP: 1061 case Instruction::Trunc: 1062 case Instruction::FPTrunc: 1063 case Instruction::BitCast: 1064 case Instruction::AddrSpaceCast: { 1065 Type *SrcTy = I->getOperand(0)->getType(); 1066 return getCastInstrCost(I->getOpcode(), I->getType(), SrcTy, I); 1067 } 1068 case Instruction::ExtractElement: { 1069 const ExtractElementInst * EEI = cast<ExtractElementInst>(I); 1070 ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1)); 1071 unsigned Idx = -1; 1072 if (CI) 1073 Idx = CI->getZExtValue(); 1074 1075 // Try to match a reduction sequence (series of shufflevector and vector 1076 // adds followed by a extractelement). 1077 unsigned ReduxOpCode; 1078 Type *ReduxType; 1079 1080 switch (matchVectorSplittingReduction(EEI, ReduxOpCode, ReduxType)) { 1081 case RK_Arithmetic: 1082 return getArithmeticReductionCost(ReduxOpCode, ReduxType, 1083 /*IsPairwiseForm=*/false); 1084 case RK_MinMax: 1085 return getMinMaxReductionCost( 1086 ReduxType, CmpInst::makeCmpResultType(ReduxType), 1087 /*IsPairwiseForm=*/false, /*IsUnsigned=*/false); 1088 case RK_UnsignedMinMax: 1089 return getMinMaxReductionCost( 1090 ReduxType, CmpInst::makeCmpResultType(ReduxType), 1091 /*IsPairwiseForm=*/false, /*IsUnsigned=*/true); 1092 case RK_None: 1093 break; 1094 } 1095 1096 switch (matchPairwiseReduction(EEI, ReduxOpCode, ReduxType)) { 1097 case RK_Arithmetic: 1098 return getArithmeticReductionCost(ReduxOpCode, ReduxType, 1099 /*IsPairwiseForm=*/true); 1100 case RK_MinMax: 1101 return getMinMaxReductionCost( 1102 ReduxType, CmpInst::makeCmpResultType(ReduxType), 1103 /*IsPairwiseForm=*/true, /*IsUnsigned=*/false); 1104 case RK_UnsignedMinMax: 1105 return getMinMaxReductionCost( 1106 ReduxType, CmpInst::makeCmpResultType(ReduxType), 1107 /*IsPairwiseForm=*/true, /*IsUnsigned=*/true); 1108 case RK_None: 1109 break; 1110 } 1111 1112 return getVectorInstrCost(I->getOpcode(), 1113 EEI->getOperand(0)->getType(), Idx); 1114 } 1115 case Instruction::InsertElement: { 1116 const InsertElementInst * IE = cast<InsertElementInst>(I); 1117 ConstantInt *CI = dyn_cast<ConstantInt>(IE->getOperand(2)); 1118 unsigned Idx = -1; 1119 if (CI) 1120 Idx = CI->getZExtValue(); 1121 return getVectorInstrCost(I->getOpcode(), 1122 IE->getType(), Idx); 1123 } 1124 case Instruction::ShuffleVector: { 1125 const ShuffleVectorInst *Shuffle = cast<ShuffleVectorInst>(I); 1126 Type *VecTypOp0 = Shuffle->getOperand(0)->getType(); 1127 unsigned NumVecElems = VecTypOp0->getVectorNumElements(); 1128 SmallVector<int, 16> Mask = Shuffle->getShuffleMask(); 1129 1130 if (NumVecElems == Mask.size()) { 1131 if (isReverseVectorMask(Mask)) 1132 return getShuffleCost(TargetTransformInfo::SK_Reverse, VecTypOp0, 1133 0, nullptr); 1134 if (isAlternateVectorMask(Mask)) 1135 return getShuffleCost(TargetTransformInfo::SK_Alternate, 1136 VecTypOp0, 0, nullptr); 1137 1138 if (isZeroEltBroadcastVectorMask(Mask)) 1139 return getShuffleCost(TargetTransformInfo::SK_Broadcast, 1140 VecTypOp0, 0, nullptr); 1141 1142 if (isSingleSourceVectorMask(Mask)) 1143 return getShuffleCost(TargetTransformInfo::SK_PermuteSingleSrc, 1144 VecTypOp0, 0, nullptr); 1145 1146 return getShuffleCost(TargetTransformInfo::SK_PermuteTwoSrc, 1147 VecTypOp0, 0, nullptr); 1148 } 1149 1150 return -1; 1151 } 1152 case Instruction::Call: 1153 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) { 1154 SmallVector<Value *, 4> Args(II->arg_operands()); 1155 1156 FastMathFlags FMF; 1157 if (auto *FPMO = dyn_cast<FPMathOperator>(II)) 1158 FMF = FPMO->getFastMathFlags(); 1159 1160 return getIntrinsicInstrCost(II->getIntrinsicID(), II->getType(), 1161 Args, FMF); 1162 } 1163 return -1; 1164 default: 1165 // We don't have any information on this instruction. 1166 return -1; 1167 } 1168 } 1169 1170 TargetTransformInfo::Concept::~Concept() {} 1171 1172 TargetIRAnalysis::TargetIRAnalysis() : TTICallback(&getDefaultTTI) {} 1173 1174 TargetIRAnalysis::TargetIRAnalysis( 1175 std::function<Result(const Function &)> TTICallback) 1176 : TTICallback(std::move(TTICallback)) {} 1177 1178 TargetIRAnalysis::Result TargetIRAnalysis::run(const Function &F, 1179 FunctionAnalysisManager &) { 1180 return TTICallback(F); 1181 } 1182 1183 AnalysisKey TargetIRAnalysis::Key; 1184 1185 TargetIRAnalysis::Result TargetIRAnalysis::getDefaultTTI(const Function &F) { 1186 return Result(F.getParent()->getDataLayout()); 1187 } 1188 1189 // Register the basic pass. 1190 INITIALIZE_PASS(TargetTransformInfoWrapperPass, "tti", 1191 "Target Transform Information", false, true) 1192 char TargetTransformInfoWrapperPass::ID = 0; 1193 1194 void TargetTransformInfoWrapperPass::anchor() {} 1195 1196 TargetTransformInfoWrapperPass::TargetTransformInfoWrapperPass() 1197 : ImmutablePass(ID) { 1198 initializeTargetTransformInfoWrapperPassPass( 1199 *PassRegistry::getPassRegistry()); 1200 } 1201 1202 TargetTransformInfoWrapperPass::TargetTransformInfoWrapperPass( 1203 TargetIRAnalysis TIRA) 1204 : ImmutablePass(ID), TIRA(std::move(TIRA)) { 1205 initializeTargetTransformInfoWrapperPassPass( 1206 *PassRegistry::getPassRegistry()); 1207 } 1208 1209 TargetTransformInfo &TargetTransformInfoWrapperPass::getTTI(const Function &F) { 1210 FunctionAnalysisManager DummyFAM; 1211 TTI = TIRA.run(F, DummyFAM); 1212 return *TTI; 1213 } 1214 1215 ImmutablePass * 1216 llvm::createTargetTransformInfoWrapperPass(TargetIRAnalysis TIRA) { 1217 return new TargetTransformInfoWrapperPass(std::move(TIRA)); 1218 } 1219