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