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 /// 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::useAA() const { return TTIImpl->useAA(); } 219 220 bool TargetTransformInfo::isTypeLegal(Type *Ty) const { 221 return TTIImpl->isTypeLegal(Ty); 222 } 223 224 unsigned TargetTransformInfo::getJumpBufAlignment() const { 225 return TTIImpl->getJumpBufAlignment(); 226 } 227 228 unsigned TargetTransformInfo::getJumpBufSize() const { 229 return TTIImpl->getJumpBufSize(); 230 } 231 232 bool TargetTransformInfo::shouldBuildLookupTables() const { 233 return TTIImpl->shouldBuildLookupTables(); 234 } 235 bool TargetTransformInfo::shouldBuildLookupTablesForConstant(Constant *C) const { 236 return TTIImpl->shouldBuildLookupTablesForConstant(C); 237 } 238 239 bool TargetTransformInfo::useColdCCForColdCall(Function &F) const { 240 return TTIImpl->useColdCCForColdCall(F); 241 } 242 243 unsigned TargetTransformInfo:: 244 getScalarizationOverhead(Type *Ty, bool Insert, bool Extract) const { 245 return TTIImpl->getScalarizationOverhead(Ty, Insert, Extract); 246 } 247 248 unsigned TargetTransformInfo:: 249 getOperandsScalarizationOverhead(ArrayRef<const Value *> Args, 250 unsigned VF) const { 251 return TTIImpl->getOperandsScalarizationOverhead(Args, VF); 252 } 253 254 bool TargetTransformInfo::supportsEfficientVectorElementLoadStore() const { 255 return TTIImpl->supportsEfficientVectorElementLoadStore(); 256 } 257 258 bool TargetTransformInfo::enableAggressiveInterleaving(bool LoopHasReductions) const { 259 return TTIImpl->enableAggressiveInterleaving(LoopHasReductions); 260 } 261 262 const TargetTransformInfo::MemCmpExpansionOptions * 263 TargetTransformInfo::enableMemCmpExpansion(bool IsZeroCmp) const { 264 return TTIImpl->enableMemCmpExpansion(IsZeroCmp); 265 } 266 267 bool TargetTransformInfo::enableInterleavedAccessVectorization() const { 268 return TTIImpl->enableInterleavedAccessVectorization(); 269 } 270 271 bool TargetTransformInfo::enableMaskedInterleavedAccessVectorization() const { 272 return TTIImpl->enableMaskedInterleavedAccessVectorization(); 273 } 274 275 bool TargetTransformInfo::isFPVectorizationPotentiallyUnsafe() const { 276 return TTIImpl->isFPVectorizationPotentiallyUnsafe(); 277 } 278 279 bool TargetTransformInfo::allowsMisalignedMemoryAccesses(LLVMContext &Context, 280 unsigned BitWidth, 281 unsigned AddressSpace, 282 unsigned Alignment, 283 bool *Fast) const { 284 return TTIImpl->allowsMisalignedMemoryAccesses(Context, BitWidth, AddressSpace, 285 Alignment, Fast); 286 } 287 288 TargetTransformInfo::PopcntSupportKind 289 TargetTransformInfo::getPopcntSupport(unsigned IntTyWidthInBit) const { 290 return TTIImpl->getPopcntSupport(IntTyWidthInBit); 291 } 292 293 bool TargetTransformInfo::haveFastSqrt(Type *Ty) const { 294 return TTIImpl->haveFastSqrt(Ty); 295 } 296 297 bool TargetTransformInfo::isFCmpOrdCheaperThanFCmpZero(Type *Ty) const { 298 return TTIImpl->isFCmpOrdCheaperThanFCmpZero(Ty); 299 } 300 301 int TargetTransformInfo::getFPOpCost(Type *Ty) const { 302 int Cost = TTIImpl->getFPOpCost(Ty); 303 assert(Cost >= 0 && "TTI should not produce negative costs!"); 304 return Cost; 305 } 306 307 int TargetTransformInfo::getIntImmCodeSizeCost(unsigned Opcode, unsigned Idx, 308 const APInt &Imm, 309 Type *Ty) const { 310 int Cost = TTIImpl->getIntImmCodeSizeCost(Opcode, Idx, Imm, Ty); 311 assert(Cost >= 0 && "TTI should not produce negative costs!"); 312 return Cost; 313 } 314 315 int TargetTransformInfo::getIntImmCost(const APInt &Imm, Type *Ty) const { 316 int Cost = TTIImpl->getIntImmCost(Imm, Ty); 317 assert(Cost >= 0 && "TTI should not produce negative costs!"); 318 return Cost; 319 } 320 321 int TargetTransformInfo::getIntImmCost(unsigned Opcode, unsigned Idx, 322 const APInt &Imm, Type *Ty) const { 323 int Cost = TTIImpl->getIntImmCost(Opcode, Idx, Imm, Ty); 324 assert(Cost >= 0 && "TTI should not produce negative costs!"); 325 return Cost; 326 } 327 328 int TargetTransformInfo::getIntImmCost(Intrinsic::ID IID, unsigned Idx, 329 const APInt &Imm, Type *Ty) const { 330 int Cost = TTIImpl->getIntImmCost(IID, Idx, Imm, Ty); 331 assert(Cost >= 0 && "TTI should not produce negative costs!"); 332 return Cost; 333 } 334 335 unsigned TargetTransformInfo::getNumberOfRegisters(bool Vector) const { 336 return TTIImpl->getNumberOfRegisters(Vector); 337 } 338 339 unsigned TargetTransformInfo::getRegisterBitWidth(bool Vector) const { 340 return TTIImpl->getRegisterBitWidth(Vector); 341 } 342 343 unsigned TargetTransformInfo::getMinVectorRegisterBitWidth() const { 344 return TTIImpl->getMinVectorRegisterBitWidth(); 345 } 346 347 bool TargetTransformInfo::shouldMaximizeVectorBandwidth(bool OptSize) const { 348 return TTIImpl->shouldMaximizeVectorBandwidth(OptSize); 349 } 350 351 unsigned TargetTransformInfo::getMinimumVF(unsigned ElemWidth) const { 352 return TTIImpl->getMinimumVF(ElemWidth); 353 } 354 355 bool TargetTransformInfo::shouldConsiderAddressTypePromotion( 356 const Instruction &I, bool &AllowPromotionWithoutCommonHeader) const { 357 return TTIImpl->shouldConsiderAddressTypePromotion( 358 I, AllowPromotionWithoutCommonHeader); 359 } 360 361 unsigned TargetTransformInfo::getCacheLineSize() const { 362 return TTIImpl->getCacheLineSize(); 363 } 364 365 llvm::Optional<unsigned> TargetTransformInfo::getCacheSize(CacheLevel Level) 366 const { 367 return TTIImpl->getCacheSize(Level); 368 } 369 370 llvm::Optional<unsigned> TargetTransformInfo::getCacheAssociativity( 371 CacheLevel Level) const { 372 return TTIImpl->getCacheAssociativity(Level); 373 } 374 375 unsigned TargetTransformInfo::getPrefetchDistance() const { 376 return TTIImpl->getPrefetchDistance(); 377 } 378 379 unsigned TargetTransformInfo::getMinPrefetchStride() const { 380 return TTIImpl->getMinPrefetchStride(); 381 } 382 383 unsigned TargetTransformInfo::getMaxPrefetchIterationsAhead() const { 384 return TTIImpl->getMaxPrefetchIterationsAhead(); 385 } 386 387 unsigned TargetTransformInfo::getMaxInterleaveFactor(unsigned VF) const { 388 return TTIImpl->getMaxInterleaveFactor(VF); 389 } 390 391 TargetTransformInfo::OperandValueKind 392 TargetTransformInfo::getOperandInfo(Value *V, OperandValueProperties &OpProps) { 393 OperandValueKind OpInfo = OK_AnyValue; 394 OpProps = OP_None; 395 396 if (auto *CI = dyn_cast<ConstantInt>(V)) { 397 if (CI->getValue().isPowerOf2()) 398 OpProps = OP_PowerOf2; 399 return OK_UniformConstantValue; 400 } 401 402 const Value *Splat = getSplatValue(V); 403 404 // Check for a splat of a constant or for a non uniform vector of constants 405 // and check if the constant(s) are all powers of two. 406 if (isa<ConstantVector>(V) || isa<ConstantDataVector>(V)) { 407 OpInfo = OK_NonUniformConstantValue; 408 if (Splat) { 409 OpInfo = OK_UniformConstantValue; 410 if (auto *CI = dyn_cast<ConstantInt>(Splat)) 411 if (CI->getValue().isPowerOf2()) 412 OpProps = OP_PowerOf2; 413 } else if (auto *CDS = dyn_cast<ConstantDataSequential>(V)) { 414 OpProps = OP_PowerOf2; 415 for (unsigned I = 0, E = CDS->getNumElements(); I != E; ++I) { 416 if (auto *CI = dyn_cast<ConstantInt>(CDS->getElementAsConstant(I))) 417 if (CI->getValue().isPowerOf2()) 418 continue; 419 OpProps = OP_None; 420 break; 421 } 422 } 423 } 424 425 // Check for a splat of a uniform value. This is not loop aware, so return 426 // true only for the obviously uniform cases (argument, globalvalue) 427 if (Splat && (isa<Argument>(Splat) || isa<GlobalValue>(Splat))) 428 OpInfo = OK_UniformValue; 429 430 return OpInfo; 431 } 432 433 int TargetTransformInfo::getArithmeticInstrCost( 434 unsigned Opcode, Type *Ty, OperandValueKind Opd1Info, 435 OperandValueKind Opd2Info, OperandValueProperties Opd1PropInfo, 436 OperandValueProperties Opd2PropInfo, 437 ArrayRef<const Value *> Args) const { 438 int Cost = TTIImpl->getArithmeticInstrCost(Opcode, Ty, Opd1Info, Opd2Info, 439 Opd1PropInfo, Opd2PropInfo, Args); 440 assert(Cost >= 0 && "TTI should not produce negative costs!"); 441 return Cost; 442 } 443 444 int TargetTransformInfo::getShuffleCost(ShuffleKind Kind, Type *Ty, int Index, 445 Type *SubTp) const { 446 int Cost = TTIImpl->getShuffleCost(Kind, Ty, Index, SubTp); 447 assert(Cost >= 0 && "TTI should not produce negative costs!"); 448 return Cost; 449 } 450 451 int TargetTransformInfo::getCastInstrCost(unsigned Opcode, Type *Dst, 452 Type *Src, const Instruction *I) const { 453 assert ((I == nullptr || I->getOpcode() == Opcode) && 454 "Opcode should reflect passed instruction."); 455 int Cost = TTIImpl->getCastInstrCost(Opcode, Dst, Src, I); 456 assert(Cost >= 0 && "TTI should not produce negative costs!"); 457 return Cost; 458 } 459 460 int TargetTransformInfo::getExtractWithExtendCost(unsigned Opcode, Type *Dst, 461 VectorType *VecTy, 462 unsigned Index) const { 463 int Cost = TTIImpl->getExtractWithExtendCost(Opcode, Dst, VecTy, Index); 464 assert(Cost >= 0 && "TTI should not produce negative costs!"); 465 return Cost; 466 } 467 468 int TargetTransformInfo::getCFInstrCost(unsigned Opcode) const { 469 int Cost = TTIImpl->getCFInstrCost(Opcode); 470 assert(Cost >= 0 && "TTI should not produce negative costs!"); 471 return Cost; 472 } 473 474 int TargetTransformInfo::getCmpSelInstrCost(unsigned Opcode, Type *ValTy, 475 Type *CondTy, const Instruction *I) const { 476 assert ((I == nullptr || I->getOpcode() == Opcode) && 477 "Opcode should reflect passed instruction."); 478 int Cost = TTIImpl->getCmpSelInstrCost(Opcode, ValTy, CondTy, I); 479 assert(Cost >= 0 && "TTI should not produce negative costs!"); 480 return Cost; 481 } 482 483 int TargetTransformInfo::getVectorInstrCost(unsigned Opcode, Type *Val, 484 unsigned Index) const { 485 int Cost = TTIImpl->getVectorInstrCost(Opcode, Val, Index); 486 assert(Cost >= 0 && "TTI should not produce negative costs!"); 487 return Cost; 488 } 489 490 int TargetTransformInfo::getMemoryOpCost(unsigned Opcode, Type *Src, 491 unsigned Alignment, 492 unsigned AddressSpace, 493 const Instruction *I) const { 494 assert ((I == nullptr || I->getOpcode() == Opcode) && 495 "Opcode should reflect passed instruction."); 496 int Cost = TTIImpl->getMemoryOpCost(Opcode, Src, Alignment, AddressSpace, I); 497 assert(Cost >= 0 && "TTI should not produce negative costs!"); 498 return Cost; 499 } 500 501 int TargetTransformInfo::getMaskedMemoryOpCost(unsigned Opcode, Type *Src, 502 unsigned Alignment, 503 unsigned AddressSpace) const { 504 int Cost = 505 TTIImpl->getMaskedMemoryOpCost(Opcode, Src, Alignment, AddressSpace); 506 assert(Cost >= 0 && "TTI should not produce negative costs!"); 507 return Cost; 508 } 509 510 int TargetTransformInfo::getGatherScatterOpCost(unsigned Opcode, Type *DataTy, 511 Value *Ptr, bool VariableMask, 512 unsigned Alignment) const { 513 int Cost = TTIImpl->getGatherScatterOpCost(Opcode, DataTy, Ptr, VariableMask, 514 Alignment); 515 assert(Cost >= 0 && "TTI should not produce negative costs!"); 516 return Cost; 517 } 518 519 int TargetTransformInfo::getInterleavedMemoryOpCost( 520 unsigned Opcode, Type *VecTy, unsigned Factor, ArrayRef<unsigned> Indices, 521 unsigned Alignment, unsigned AddressSpace, bool UseMaskForCond, 522 bool UseMaskForGaps) const { 523 int Cost = TTIImpl->getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices, 524 Alignment, AddressSpace, 525 UseMaskForCond, 526 UseMaskForGaps); 527 assert(Cost >= 0 && "TTI should not produce negative costs!"); 528 return Cost; 529 } 530 531 int TargetTransformInfo::getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy, 532 ArrayRef<Type *> Tys, FastMathFlags FMF, 533 unsigned ScalarizationCostPassed) const { 534 int Cost = TTIImpl->getIntrinsicInstrCost(ID, RetTy, Tys, FMF, 535 ScalarizationCostPassed); 536 assert(Cost >= 0 && "TTI should not produce negative costs!"); 537 return Cost; 538 } 539 540 int TargetTransformInfo::getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy, 541 ArrayRef<Value *> Args, FastMathFlags FMF, unsigned VF) const { 542 int Cost = TTIImpl->getIntrinsicInstrCost(ID, RetTy, Args, FMF, VF); 543 assert(Cost >= 0 && "TTI should not produce negative costs!"); 544 return Cost; 545 } 546 547 int TargetTransformInfo::getCallInstrCost(Function *F, Type *RetTy, 548 ArrayRef<Type *> Tys) const { 549 int Cost = TTIImpl->getCallInstrCost(F, RetTy, Tys); 550 assert(Cost >= 0 && "TTI should not produce negative costs!"); 551 return Cost; 552 } 553 554 unsigned TargetTransformInfo::getNumberOfParts(Type *Tp) const { 555 return TTIImpl->getNumberOfParts(Tp); 556 } 557 558 int TargetTransformInfo::getAddressComputationCost(Type *Tp, 559 ScalarEvolution *SE, 560 const SCEV *Ptr) const { 561 int Cost = TTIImpl->getAddressComputationCost(Tp, SE, Ptr); 562 assert(Cost >= 0 && "TTI should not produce negative costs!"); 563 return Cost; 564 } 565 566 int TargetTransformInfo::getArithmeticReductionCost(unsigned Opcode, Type *Ty, 567 bool IsPairwiseForm) const { 568 int Cost = TTIImpl->getArithmeticReductionCost(Opcode, Ty, IsPairwiseForm); 569 assert(Cost >= 0 && "TTI should not produce negative costs!"); 570 return Cost; 571 } 572 573 int TargetTransformInfo::getMinMaxReductionCost(Type *Ty, Type *CondTy, 574 bool IsPairwiseForm, 575 bool IsUnsigned) const { 576 int Cost = 577 TTIImpl->getMinMaxReductionCost(Ty, CondTy, IsPairwiseForm, IsUnsigned); 578 assert(Cost >= 0 && "TTI should not produce negative costs!"); 579 return Cost; 580 } 581 582 unsigned 583 TargetTransformInfo::getCostOfKeepingLiveOverCall(ArrayRef<Type *> Tys) const { 584 return TTIImpl->getCostOfKeepingLiveOverCall(Tys); 585 } 586 587 bool TargetTransformInfo::getTgtMemIntrinsic(IntrinsicInst *Inst, 588 MemIntrinsicInfo &Info) const { 589 return TTIImpl->getTgtMemIntrinsic(Inst, Info); 590 } 591 592 unsigned TargetTransformInfo::getAtomicMemIntrinsicMaxElementSize() const { 593 return TTIImpl->getAtomicMemIntrinsicMaxElementSize(); 594 } 595 596 Value *TargetTransformInfo::getOrCreateResultFromMemIntrinsic( 597 IntrinsicInst *Inst, Type *ExpectedType) const { 598 return TTIImpl->getOrCreateResultFromMemIntrinsic(Inst, ExpectedType); 599 } 600 601 Type *TargetTransformInfo::getMemcpyLoopLoweringType(LLVMContext &Context, 602 Value *Length, 603 unsigned SrcAlign, 604 unsigned DestAlign) const { 605 return TTIImpl->getMemcpyLoopLoweringType(Context, Length, SrcAlign, 606 DestAlign); 607 } 608 609 void TargetTransformInfo::getMemcpyLoopResidualLoweringType( 610 SmallVectorImpl<Type *> &OpsOut, LLVMContext &Context, 611 unsigned RemainingBytes, unsigned SrcAlign, unsigned DestAlign) const { 612 TTIImpl->getMemcpyLoopResidualLoweringType(OpsOut, Context, RemainingBytes, 613 SrcAlign, DestAlign); 614 } 615 616 bool TargetTransformInfo::areInlineCompatible(const Function *Caller, 617 const Function *Callee) const { 618 return TTIImpl->areInlineCompatible(Caller, Callee); 619 } 620 621 bool TargetTransformInfo::isIndexedLoadLegal(MemIndexedMode Mode, 622 Type *Ty) const { 623 return TTIImpl->isIndexedLoadLegal(Mode, Ty); 624 } 625 626 bool TargetTransformInfo::isIndexedStoreLegal(MemIndexedMode Mode, 627 Type *Ty) const { 628 return TTIImpl->isIndexedStoreLegal(Mode, Ty); 629 } 630 631 unsigned TargetTransformInfo::getLoadStoreVecRegBitWidth(unsigned AS) const { 632 return TTIImpl->getLoadStoreVecRegBitWidth(AS); 633 } 634 635 bool TargetTransformInfo::isLegalToVectorizeLoad(LoadInst *LI) const { 636 return TTIImpl->isLegalToVectorizeLoad(LI); 637 } 638 639 bool TargetTransformInfo::isLegalToVectorizeStore(StoreInst *SI) const { 640 return TTIImpl->isLegalToVectorizeStore(SI); 641 } 642 643 bool TargetTransformInfo::isLegalToVectorizeLoadChain( 644 unsigned ChainSizeInBytes, unsigned Alignment, unsigned AddrSpace) const { 645 return TTIImpl->isLegalToVectorizeLoadChain(ChainSizeInBytes, Alignment, 646 AddrSpace); 647 } 648 649 bool TargetTransformInfo::isLegalToVectorizeStoreChain( 650 unsigned ChainSizeInBytes, unsigned Alignment, unsigned AddrSpace) const { 651 return TTIImpl->isLegalToVectorizeStoreChain(ChainSizeInBytes, Alignment, 652 AddrSpace); 653 } 654 655 unsigned TargetTransformInfo::getLoadVectorFactor(unsigned VF, 656 unsigned LoadSize, 657 unsigned ChainSizeInBytes, 658 VectorType *VecTy) const { 659 return TTIImpl->getLoadVectorFactor(VF, LoadSize, ChainSizeInBytes, VecTy); 660 } 661 662 unsigned TargetTransformInfo::getStoreVectorFactor(unsigned VF, 663 unsigned StoreSize, 664 unsigned ChainSizeInBytes, 665 VectorType *VecTy) const { 666 return TTIImpl->getStoreVectorFactor(VF, StoreSize, ChainSizeInBytes, VecTy); 667 } 668 669 bool TargetTransformInfo::useReductionIntrinsic(unsigned Opcode, 670 Type *Ty, ReductionFlags Flags) const { 671 return TTIImpl->useReductionIntrinsic(Opcode, Ty, Flags); 672 } 673 674 bool TargetTransformInfo::shouldExpandReduction(const IntrinsicInst *II) const { 675 return TTIImpl->shouldExpandReduction(II); 676 } 677 678 int TargetTransformInfo::getInstructionLatency(const Instruction *I) const { 679 return TTIImpl->getInstructionLatency(I); 680 } 681 682 static bool matchPairwiseShuffleMask(ShuffleVectorInst *SI, bool IsLeft, 683 unsigned Level) { 684 // We don't need a shuffle if we just want to have element 0 in position 0 of 685 // the vector. 686 if (!SI && Level == 0 && IsLeft) 687 return true; 688 else if (!SI) 689 return false; 690 691 SmallVector<int, 32> Mask(SI->getType()->getVectorNumElements(), -1); 692 693 // Build a mask of 0, 2, ... (left) or 1, 3, ... (right) depending on whether 694 // we look at the left or right side. 695 for (unsigned i = 0, e = (1 << Level), val = !IsLeft; i != e; ++i, val += 2) 696 Mask[i] = val; 697 698 SmallVector<int, 16> ActualMask = SI->getShuffleMask(); 699 return Mask == ActualMask; 700 } 701 702 namespace { 703 /// Kind of the reduction data. 704 enum ReductionKind { 705 RK_None, /// Not a reduction. 706 RK_Arithmetic, /// Binary reduction data. 707 RK_MinMax, /// Min/max reduction data. 708 RK_UnsignedMinMax, /// Unsigned min/max reduction data. 709 }; 710 /// Contains opcode + LHS/RHS parts of the reduction operations. 711 struct ReductionData { 712 ReductionData() = delete; 713 ReductionData(ReductionKind Kind, unsigned Opcode, Value *LHS, Value *RHS) 714 : Opcode(Opcode), LHS(LHS), RHS(RHS), Kind(Kind) { 715 assert(Kind != RK_None && "expected binary or min/max reduction only."); 716 } 717 unsigned Opcode = 0; 718 Value *LHS = nullptr; 719 Value *RHS = nullptr; 720 ReductionKind Kind = RK_None; 721 bool hasSameData(ReductionData &RD) const { 722 return Kind == RD.Kind && Opcode == RD.Opcode; 723 } 724 }; 725 } // namespace 726 727 static Optional<ReductionData> getReductionData(Instruction *I) { 728 Value *L, *R; 729 if (m_BinOp(m_Value(L), m_Value(R)).match(I)) 730 return ReductionData(RK_Arithmetic, I->getOpcode(), L, R); 731 if (auto *SI = dyn_cast<SelectInst>(I)) { 732 if (m_SMin(m_Value(L), m_Value(R)).match(SI) || 733 m_SMax(m_Value(L), m_Value(R)).match(SI) || 734 m_OrdFMin(m_Value(L), m_Value(R)).match(SI) || 735 m_OrdFMax(m_Value(L), m_Value(R)).match(SI) || 736 m_UnordFMin(m_Value(L), m_Value(R)).match(SI) || 737 m_UnordFMax(m_Value(L), m_Value(R)).match(SI)) { 738 auto *CI = cast<CmpInst>(SI->getCondition()); 739 return ReductionData(RK_MinMax, CI->getOpcode(), L, R); 740 } 741 if (m_UMin(m_Value(L), m_Value(R)).match(SI) || 742 m_UMax(m_Value(L), m_Value(R)).match(SI)) { 743 auto *CI = cast<CmpInst>(SI->getCondition()); 744 return ReductionData(RK_UnsignedMinMax, CI->getOpcode(), L, R); 745 } 746 } 747 return llvm::None; 748 } 749 750 static ReductionKind matchPairwiseReductionAtLevel(Instruction *I, 751 unsigned Level, 752 unsigned NumLevels) { 753 // Match one level of pairwise operations. 754 // %rdx.shuf.0.0 = shufflevector <4 x float> %rdx, <4 x float> undef, 755 // <4 x i32> <i32 0, i32 2 , i32 undef, i32 undef> 756 // %rdx.shuf.0.1 = shufflevector <4 x float> %rdx, <4 x float> undef, 757 // <4 x i32> <i32 1, i32 3, i32 undef, i32 undef> 758 // %bin.rdx.0 = fadd <4 x float> %rdx.shuf.0.0, %rdx.shuf.0.1 759 if (!I) 760 return RK_None; 761 762 assert(I->getType()->isVectorTy() && "Expecting a vector type"); 763 764 Optional<ReductionData> RD = getReductionData(I); 765 if (!RD) 766 return RK_None; 767 768 ShuffleVectorInst *LS = dyn_cast<ShuffleVectorInst>(RD->LHS); 769 if (!LS && Level) 770 return RK_None; 771 ShuffleVectorInst *RS = dyn_cast<ShuffleVectorInst>(RD->RHS); 772 if (!RS && Level) 773 return RK_None; 774 775 // On level 0 we can omit one shufflevector instruction. 776 if (!Level && !RS && !LS) 777 return RK_None; 778 779 // Shuffle inputs must match. 780 Value *NextLevelOpL = LS ? LS->getOperand(0) : nullptr; 781 Value *NextLevelOpR = RS ? RS->getOperand(0) : nullptr; 782 Value *NextLevelOp = nullptr; 783 if (NextLevelOpR && NextLevelOpL) { 784 // If we have two shuffles their operands must match. 785 if (NextLevelOpL != NextLevelOpR) 786 return RK_None; 787 788 NextLevelOp = NextLevelOpL; 789 } else if (Level == 0 && (NextLevelOpR || NextLevelOpL)) { 790 // On the first level we can omit the shufflevector <0, undef,...>. So the 791 // input to the other shufflevector <1, undef> must match with one of the 792 // inputs to the current binary operation. 793 // Example: 794 // %NextLevelOpL = shufflevector %R, <1, undef ...> 795 // %BinOp = fadd %NextLevelOpL, %R 796 if (NextLevelOpL && NextLevelOpL != RD->RHS) 797 return RK_None; 798 else if (NextLevelOpR && NextLevelOpR != RD->LHS) 799 return RK_None; 800 801 NextLevelOp = NextLevelOpL ? RD->RHS : RD->LHS; 802 } else 803 return RK_None; 804 805 // Check that the next levels binary operation exists and matches with the 806 // current one. 807 if (Level + 1 != NumLevels) { 808 Optional<ReductionData> NextLevelRD = 809 getReductionData(cast<Instruction>(NextLevelOp)); 810 if (!NextLevelRD || !RD->hasSameData(*NextLevelRD)) 811 return RK_None; 812 } 813 814 // Shuffle mask for pairwise operation must match. 815 if (matchPairwiseShuffleMask(LS, /*IsLeft=*/true, Level)) { 816 if (!matchPairwiseShuffleMask(RS, /*IsLeft=*/false, Level)) 817 return RK_None; 818 } else if (matchPairwiseShuffleMask(RS, /*IsLeft=*/true, Level)) { 819 if (!matchPairwiseShuffleMask(LS, /*IsLeft=*/false, Level)) 820 return RK_None; 821 } else { 822 return RK_None; 823 } 824 825 if (++Level == NumLevels) 826 return RD->Kind; 827 828 // Match next level. 829 return matchPairwiseReductionAtLevel(cast<Instruction>(NextLevelOp), Level, 830 NumLevels); 831 } 832 833 static ReductionKind matchPairwiseReduction(const ExtractElementInst *ReduxRoot, 834 unsigned &Opcode, Type *&Ty) { 835 if (!EnableReduxCost) 836 return RK_None; 837 838 // Need to extract the first element. 839 ConstantInt *CI = dyn_cast<ConstantInt>(ReduxRoot->getOperand(1)); 840 unsigned Idx = ~0u; 841 if (CI) 842 Idx = CI->getZExtValue(); 843 if (Idx != 0) 844 return RK_None; 845 846 auto *RdxStart = dyn_cast<Instruction>(ReduxRoot->getOperand(0)); 847 if (!RdxStart) 848 return RK_None; 849 Optional<ReductionData> RD = getReductionData(RdxStart); 850 if (!RD) 851 return RK_None; 852 853 Type *VecTy = RdxStart->getType(); 854 unsigned NumVecElems = VecTy->getVectorNumElements(); 855 if (!isPowerOf2_32(NumVecElems)) 856 return RK_None; 857 858 // We look for a sequence of shuffle,shuffle,add triples like the following 859 // that builds a pairwise reduction tree. 860 // 861 // (X0, X1, X2, X3) 862 // (X0 + X1, X2 + X3, undef, undef) 863 // ((X0 + X1) + (X2 + X3), undef, undef, undef) 864 // 865 // %rdx.shuf.0.0 = shufflevector <4 x float> %rdx, <4 x float> undef, 866 // <4 x i32> <i32 0, i32 2 , i32 undef, i32 undef> 867 // %rdx.shuf.0.1 = shufflevector <4 x float> %rdx, <4 x float> undef, 868 // <4 x i32> <i32 1, i32 3, i32 undef, i32 undef> 869 // %bin.rdx.0 = fadd <4 x float> %rdx.shuf.0.0, %rdx.shuf.0.1 870 // %rdx.shuf.1.0 = shufflevector <4 x float> %bin.rdx.0, <4 x float> undef, 871 // <4 x i32> <i32 0, i32 undef, i32 undef, i32 undef> 872 // %rdx.shuf.1.1 = shufflevector <4 x float> %bin.rdx.0, <4 x float> undef, 873 // <4 x i32> <i32 1, i32 undef, i32 undef, i32 undef> 874 // %bin.rdx8 = fadd <4 x float> %rdx.shuf.1.0, %rdx.shuf.1.1 875 // %r = extractelement <4 x float> %bin.rdx8, i32 0 876 if (matchPairwiseReductionAtLevel(RdxStart, 0, Log2_32(NumVecElems)) == 877 RK_None) 878 return RK_None; 879 880 Opcode = RD->Opcode; 881 Ty = VecTy; 882 883 return RD->Kind; 884 } 885 886 static std::pair<Value *, ShuffleVectorInst *> 887 getShuffleAndOtherOprd(Value *L, Value *R) { 888 ShuffleVectorInst *S = nullptr; 889 890 if ((S = dyn_cast<ShuffleVectorInst>(L))) 891 return std::make_pair(R, S); 892 893 S = dyn_cast<ShuffleVectorInst>(R); 894 return std::make_pair(L, S); 895 } 896 897 static ReductionKind 898 matchVectorSplittingReduction(const ExtractElementInst *ReduxRoot, 899 unsigned &Opcode, Type *&Ty) { 900 if (!EnableReduxCost) 901 return RK_None; 902 903 // Need to extract the first element. 904 ConstantInt *CI = dyn_cast<ConstantInt>(ReduxRoot->getOperand(1)); 905 unsigned Idx = ~0u; 906 if (CI) 907 Idx = CI->getZExtValue(); 908 if (Idx != 0) 909 return RK_None; 910 911 auto *RdxStart = dyn_cast<Instruction>(ReduxRoot->getOperand(0)); 912 if (!RdxStart) 913 return RK_None; 914 Optional<ReductionData> RD = getReductionData(RdxStart); 915 if (!RD) 916 return RK_None; 917 918 Type *VecTy = ReduxRoot->getOperand(0)->getType(); 919 unsigned NumVecElems = VecTy->getVectorNumElements(); 920 if (!isPowerOf2_32(NumVecElems)) 921 return RK_None; 922 923 // We look for a sequence of shuffles and adds like the following matching one 924 // fadd, shuffle vector pair at a time. 925 // 926 // %rdx.shuf = shufflevector <4 x float> %rdx, <4 x float> undef, 927 // <4 x i32> <i32 2, i32 3, i32 undef, i32 undef> 928 // %bin.rdx = fadd <4 x float> %rdx, %rdx.shuf 929 // %rdx.shuf7 = shufflevector <4 x float> %bin.rdx, <4 x float> undef, 930 // <4 x i32> <i32 1, i32 undef, i32 undef, i32 undef> 931 // %bin.rdx8 = fadd <4 x float> %bin.rdx, %rdx.shuf7 932 // %r = extractelement <4 x float> %bin.rdx8, i32 0 933 934 unsigned MaskStart = 1; 935 Instruction *RdxOp = RdxStart; 936 SmallVector<int, 32> ShuffleMask(NumVecElems, 0); 937 unsigned NumVecElemsRemain = NumVecElems; 938 while (NumVecElemsRemain - 1) { 939 // Check for the right reduction operation. 940 if (!RdxOp) 941 return RK_None; 942 Optional<ReductionData> RDLevel = getReductionData(RdxOp); 943 if (!RDLevel || !RDLevel->hasSameData(*RD)) 944 return RK_None; 945 946 Value *NextRdxOp; 947 ShuffleVectorInst *Shuffle; 948 std::tie(NextRdxOp, Shuffle) = 949 getShuffleAndOtherOprd(RDLevel->LHS, RDLevel->RHS); 950 951 // Check the current reduction operation and the shuffle use the same value. 952 if (Shuffle == nullptr) 953 return RK_None; 954 if (Shuffle->getOperand(0) != NextRdxOp) 955 return RK_None; 956 957 // Check that shuffle masks matches. 958 for (unsigned j = 0; j != MaskStart; ++j) 959 ShuffleMask[j] = MaskStart + j; 960 // Fill the rest of the mask with -1 for undef. 961 std::fill(&ShuffleMask[MaskStart], ShuffleMask.end(), -1); 962 963 SmallVector<int, 16> Mask = Shuffle->getShuffleMask(); 964 if (ShuffleMask != Mask) 965 return RK_None; 966 967 RdxOp = dyn_cast<Instruction>(NextRdxOp); 968 NumVecElemsRemain /= 2; 969 MaskStart *= 2; 970 } 971 972 Opcode = RD->Opcode; 973 Ty = VecTy; 974 return RD->Kind; 975 } 976 977 int TargetTransformInfo::getInstructionThroughput(const Instruction *I) const { 978 switch (I->getOpcode()) { 979 case Instruction::GetElementPtr: 980 return getUserCost(I); 981 982 case Instruction::Ret: 983 case Instruction::PHI: 984 case Instruction::Br: { 985 return getCFInstrCost(I->getOpcode()); 986 } 987 case Instruction::Add: 988 case Instruction::FAdd: 989 case Instruction::Sub: 990 case Instruction::FSub: 991 case Instruction::Mul: 992 case Instruction::FMul: 993 case Instruction::UDiv: 994 case Instruction::SDiv: 995 case Instruction::FDiv: 996 case Instruction::URem: 997 case Instruction::SRem: 998 case Instruction::FRem: 999 case Instruction::Shl: 1000 case Instruction::LShr: 1001 case Instruction::AShr: 1002 case Instruction::And: 1003 case Instruction::Or: 1004 case Instruction::Xor: { 1005 TargetTransformInfo::OperandValueKind Op1VK, Op2VK; 1006 TargetTransformInfo::OperandValueProperties Op1VP, Op2VP; 1007 Op1VK = getOperandInfo(I->getOperand(0), Op1VP); 1008 Op2VK = getOperandInfo(I->getOperand(1), Op2VP); 1009 SmallVector<const Value *, 2> Operands(I->operand_values()); 1010 return getArithmeticInstrCost(I->getOpcode(), I->getType(), Op1VK, Op2VK, 1011 Op1VP, Op2VP, Operands); 1012 } 1013 case Instruction::Select: { 1014 const SelectInst *SI = cast<SelectInst>(I); 1015 Type *CondTy = SI->getCondition()->getType(); 1016 return getCmpSelInstrCost(I->getOpcode(), I->getType(), CondTy, I); 1017 } 1018 case Instruction::ICmp: 1019 case Instruction::FCmp: { 1020 Type *ValTy = I->getOperand(0)->getType(); 1021 return getCmpSelInstrCost(I->getOpcode(), ValTy, I->getType(), I); 1022 } 1023 case Instruction::Store: { 1024 const StoreInst *SI = cast<StoreInst>(I); 1025 Type *ValTy = SI->getValueOperand()->getType(); 1026 return getMemoryOpCost(I->getOpcode(), ValTy, 1027 SI->getAlignment(), 1028 SI->getPointerAddressSpace(), I); 1029 } 1030 case Instruction::Load: { 1031 const LoadInst *LI = cast<LoadInst>(I); 1032 return getMemoryOpCost(I->getOpcode(), I->getType(), 1033 LI->getAlignment(), 1034 LI->getPointerAddressSpace(), I); 1035 } 1036 case Instruction::ZExt: 1037 case Instruction::SExt: 1038 case Instruction::FPToUI: 1039 case Instruction::FPToSI: 1040 case Instruction::FPExt: 1041 case Instruction::PtrToInt: 1042 case Instruction::IntToPtr: 1043 case Instruction::SIToFP: 1044 case Instruction::UIToFP: 1045 case Instruction::Trunc: 1046 case Instruction::FPTrunc: 1047 case Instruction::BitCast: 1048 case Instruction::AddrSpaceCast: { 1049 Type *SrcTy = I->getOperand(0)->getType(); 1050 return getCastInstrCost(I->getOpcode(), I->getType(), SrcTy, I); 1051 } 1052 case Instruction::ExtractElement: { 1053 const ExtractElementInst * EEI = cast<ExtractElementInst>(I); 1054 ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1)); 1055 unsigned Idx = -1; 1056 if (CI) 1057 Idx = CI->getZExtValue(); 1058 1059 // Try to match a reduction sequence (series of shufflevector and vector 1060 // adds followed by a extractelement). 1061 unsigned ReduxOpCode; 1062 Type *ReduxType; 1063 1064 switch (matchVectorSplittingReduction(EEI, ReduxOpCode, ReduxType)) { 1065 case RK_Arithmetic: 1066 return getArithmeticReductionCost(ReduxOpCode, ReduxType, 1067 /*IsPairwiseForm=*/false); 1068 case RK_MinMax: 1069 return getMinMaxReductionCost( 1070 ReduxType, CmpInst::makeCmpResultType(ReduxType), 1071 /*IsPairwiseForm=*/false, /*IsUnsigned=*/false); 1072 case RK_UnsignedMinMax: 1073 return getMinMaxReductionCost( 1074 ReduxType, CmpInst::makeCmpResultType(ReduxType), 1075 /*IsPairwiseForm=*/false, /*IsUnsigned=*/true); 1076 case RK_None: 1077 break; 1078 } 1079 1080 switch (matchPairwiseReduction(EEI, ReduxOpCode, ReduxType)) { 1081 case RK_Arithmetic: 1082 return getArithmeticReductionCost(ReduxOpCode, ReduxType, 1083 /*IsPairwiseForm=*/true); 1084 case RK_MinMax: 1085 return getMinMaxReductionCost( 1086 ReduxType, CmpInst::makeCmpResultType(ReduxType), 1087 /*IsPairwiseForm=*/true, /*IsUnsigned=*/false); 1088 case RK_UnsignedMinMax: 1089 return getMinMaxReductionCost( 1090 ReduxType, CmpInst::makeCmpResultType(ReduxType), 1091 /*IsPairwiseForm=*/true, /*IsUnsigned=*/true); 1092 case RK_None: 1093 break; 1094 } 1095 1096 return getVectorInstrCost(I->getOpcode(), 1097 EEI->getOperand(0)->getType(), Idx); 1098 } 1099 case Instruction::InsertElement: { 1100 const InsertElementInst * IE = cast<InsertElementInst>(I); 1101 ConstantInt *CI = dyn_cast<ConstantInt>(IE->getOperand(2)); 1102 unsigned Idx = -1; 1103 if (CI) 1104 Idx = CI->getZExtValue(); 1105 return getVectorInstrCost(I->getOpcode(), 1106 IE->getType(), Idx); 1107 } 1108 case Instruction::ShuffleVector: { 1109 const ShuffleVectorInst *Shuffle = cast<ShuffleVectorInst>(I); 1110 Type *Ty = Shuffle->getType(); 1111 Type *SrcTy = Shuffle->getOperand(0)->getType(); 1112 1113 // TODO: Identify and add costs for insert subvector, etc. 1114 int SubIndex; 1115 if (Shuffle->isExtractSubvectorMask(SubIndex)) 1116 return TTIImpl->getShuffleCost(SK_ExtractSubvector, SrcTy, SubIndex, Ty); 1117 1118 if (Shuffle->changesLength()) 1119 return -1; 1120 1121 if (Shuffle->isIdentity()) 1122 return 0; 1123 1124 if (Shuffle->isReverse()) 1125 return TTIImpl->getShuffleCost(SK_Reverse, Ty, 0, nullptr); 1126 1127 if (Shuffle->isSelect()) 1128 return TTIImpl->getShuffleCost(SK_Select, Ty, 0, nullptr); 1129 1130 if (Shuffle->isTranspose()) 1131 return TTIImpl->getShuffleCost(SK_Transpose, Ty, 0, nullptr); 1132 1133 if (Shuffle->isZeroEltSplat()) 1134 return TTIImpl->getShuffleCost(SK_Broadcast, Ty, 0, nullptr); 1135 1136 if (Shuffle->isSingleSource()) 1137 return TTIImpl->getShuffleCost(SK_PermuteSingleSrc, Ty, 0, nullptr); 1138 1139 return TTIImpl->getShuffleCost(SK_PermuteTwoSrc, Ty, 0, nullptr); 1140 } 1141 case Instruction::Call: 1142 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) { 1143 SmallVector<Value *, 4> Args(II->arg_operands()); 1144 1145 FastMathFlags FMF; 1146 if (auto *FPMO = dyn_cast<FPMathOperator>(II)) 1147 FMF = FPMO->getFastMathFlags(); 1148 1149 return getIntrinsicInstrCost(II->getIntrinsicID(), II->getType(), 1150 Args, FMF); 1151 } 1152 return -1; 1153 default: 1154 // We don't have any information on this instruction. 1155 return -1; 1156 } 1157 } 1158 1159 TargetTransformInfo::Concept::~Concept() {} 1160 1161 TargetIRAnalysis::TargetIRAnalysis() : TTICallback(&getDefaultTTI) {} 1162 1163 TargetIRAnalysis::TargetIRAnalysis( 1164 std::function<Result(const Function &)> TTICallback) 1165 : TTICallback(std::move(TTICallback)) {} 1166 1167 TargetIRAnalysis::Result TargetIRAnalysis::run(const Function &F, 1168 FunctionAnalysisManager &) { 1169 return TTICallback(F); 1170 } 1171 1172 AnalysisKey TargetIRAnalysis::Key; 1173 1174 TargetIRAnalysis::Result TargetIRAnalysis::getDefaultTTI(const Function &F) { 1175 return Result(F.getParent()->getDataLayout()); 1176 } 1177 1178 // Register the basic pass. 1179 INITIALIZE_PASS(TargetTransformInfoWrapperPass, "tti", 1180 "Target Transform Information", false, true) 1181 char TargetTransformInfoWrapperPass::ID = 0; 1182 1183 void TargetTransformInfoWrapperPass::anchor() {} 1184 1185 TargetTransformInfoWrapperPass::TargetTransformInfoWrapperPass() 1186 : ImmutablePass(ID) { 1187 initializeTargetTransformInfoWrapperPassPass( 1188 *PassRegistry::getPassRegistry()); 1189 } 1190 1191 TargetTransformInfoWrapperPass::TargetTransformInfoWrapperPass( 1192 TargetIRAnalysis TIRA) 1193 : ImmutablePass(ID), TIRA(std::move(TIRA)) { 1194 initializeTargetTransformInfoWrapperPassPass( 1195 *PassRegistry::getPassRegistry()); 1196 } 1197 1198 TargetTransformInfo &TargetTransformInfoWrapperPass::getTTI(const Function &F) { 1199 FunctionAnalysisManager DummyFAM; 1200 TTI = TIRA.run(F, DummyFAM); 1201 return *TTI; 1202 } 1203 1204 ImmutablePass * 1205 llvm::createTargetTransformInfoWrapperPass(TargetIRAnalysis TIRA) { 1206 return new TargetTransformInfoWrapperPass(std::move(TIRA)); 1207 } 1208