1 //===- llvm/Analysis/TargetTransformInfo.cpp ------------------------------===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 9 #include "llvm/Analysis/TargetTransformInfo.h" 10 #include "llvm/Analysis/CFG.h" 11 #include "llvm/Analysis/LoopIterator.h" 12 #include "llvm/Analysis/TargetTransformInfoImpl.h" 13 #include "llvm/IR/CFG.h" 14 #include "llvm/IR/CallSite.h" 15 #include "llvm/IR/DataLayout.h" 16 #include "llvm/IR/Instruction.h" 17 #include "llvm/IR/Instructions.h" 18 #include "llvm/IR/IntrinsicInst.h" 19 #include "llvm/IR/Module.h" 20 #include "llvm/IR/Operator.h" 21 #include "llvm/IR/PatternMatch.h" 22 #include "llvm/InitializePasses.h" 23 #include "llvm/Support/CommandLine.h" 24 #include "llvm/Support/ErrorHandling.h" 25 #include <utility> 26 27 using namespace llvm; 28 using namespace PatternMatch; 29 30 #define DEBUG_TYPE "tti" 31 32 static cl::opt<bool> EnableReduxCost("costmodel-reduxcost", cl::init(false), 33 cl::Hidden, 34 cl::desc("Recognize reduction patterns.")); 35 36 namespace { 37 /// No-op implementation of the TTI interface using the utility base 38 /// classes. 39 /// 40 /// This is used when no target specific information is available. 41 struct NoTTIImpl : TargetTransformInfoImplCRTPBase<NoTTIImpl> { 42 explicit NoTTIImpl(const DataLayout &DL) 43 : TargetTransformInfoImplCRTPBase<NoTTIImpl>(DL) {} 44 }; 45 } // namespace 46 47 bool HardwareLoopInfo::canAnalyze(LoopInfo &LI) { 48 // If the loop has irreducible control flow, it can not be converted to 49 // Hardware loop. 50 LoopBlocksRPO RPOT(L); 51 RPOT.perform(&LI); 52 if (containsIrreducibleCFG<const BasicBlock *>(RPOT, LI)) 53 return false; 54 return true; 55 } 56 57 bool HardwareLoopInfo::isHardwareLoopCandidate(ScalarEvolution &SE, 58 LoopInfo &LI, DominatorTree &DT, 59 bool ForceNestedLoop, 60 bool ForceHardwareLoopPHI) { 61 SmallVector<BasicBlock *, 4> ExitingBlocks; 62 L->getExitingBlocks(ExitingBlocks); 63 64 for (BasicBlock *BB : ExitingBlocks) { 65 // If we pass the updated counter back through a phi, we need to know 66 // which latch the updated value will be coming from. 67 if (!L->isLoopLatch(BB)) { 68 if (ForceHardwareLoopPHI || CounterInReg) 69 continue; 70 } 71 72 const SCEV *EC = SE.getExitCount(L, BB); 73 if (isa<SCEVCouldNotCompute>(EC)) 74 continue; 75 if (const SCEVConstant *ConstEC = dyn_cast<SCEVConstant>(EC)) { 76 if (ConstEC->getValue()->isZero()) 77 continue; 78 } else if (!SE.isLoopInvariant(EC, L)) 79 continue; 80 81 if (SE.getTypeSizeInBits(EC->getType()) > CountType->getBitWidth()) 82 continue; 83 84 // If this exiting block is contained in a nested loop, it is not eligible 85 // for insertion of the branch-and-decrement since the inner loop would 86 // end up messing up the value in the CTR. 87 if (!IsNestingLegal && LI.getLoopFor(BB) != L && !ForceNestedLoop) 88 continue; 89 90 // We now have a loop-invariant count of loop iterations (which is not the 91 // constant zero) for which we know that this loop will not exit via this 92 // existing block. 93 94 // We need to make sure that this block will run on every loop iteration. 95 // For this to be true, we must dominate all blocks with backedges. Such 96 // blocks are in-loop predecessors to the header block. 97 bool NotAlways = false; 98 for (BasicBlock *Pred : predecessors(L->getHeader())) { 99 if (!L->contains(Pred)) 100 continue; 101 102 if (!DT.dominates(BB, Pred)) { 103 NotAlways = true; 104 break; 105 } 106 } 107 108 if (NotAlways) 109 continue; 110 111 // Make sure this blocks ends with a conditional branch. 112 Instruction *TI = BB->getTerminator(); 113 if (!TI) 114 continue; 115 116 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) { 117 if (!BI->isConditional()) 118 continue; 119 120 ExitBranch = BI; 121 } else 122 continue; 123 124 // Note that this block may not be the loop latch block, even if the loop 125 // has a latch block. 126 ExitBlock = BB; 127 ExitCount = EC; 128 break; 129 } 130 131 if (!ExitBlock) 132 return false; 133 return true; 134 } 135 136 TargetTransformInfo::TargetTransformInfo(const DataLayout &DL) 137 : TTIImpl(new Model<NoTTIImpl>(NoTTIImpl(DL))) {} 138 139 TargetTransformInfo::~TargetTransformInfo() {} 140 141 TargetTransformInfo::TargetTransformInfo(TargetTransformInfo &&Arg) 142 : TTIImpl(std::move(Arg.TTIImpl)) {} 143 144 TargetTransformInfo &TargetTransformInfo::operator=(TargetTransformInfo &&RHS) { 145 TTIImpl = std::move(RHS.TTIImpl); 146 return *this; 147 } 148 149 int TargetTransformInfo::getOperationCost(unsigned Opcode, Type *Ty, 150 Type *OpTy) const { 151 int Cost = TTIImpl->getOperationCost(Opcode, Ty, OpTy); 152 assert(Cost >= 0 && "TTI should not produce negative costs!"); 153 return Cost; 154 } 155 156 unsigned TargetTransformInfo::getInliningThresholdMultiplier() const { 157 return TTIImpl->getInliningThresholdMultiplier(); 158 } 159 160 int TargetTransformInfo::getInlinerVectorBonusPercent() const { 161 return TTIImpl->getInlinerVectorBonusPercent(); 162 } 163 164 int TargetTransformInfo::getGEPCost(Type *PointeeType, const Value *Ptr, 165 ArrayRef<const Value *> Operands) const { 166 return TTIImpl->getGEPCost(PointeeType, Ptr, Operands); 167 } 168 169 int TargetTransformInfo::getExtCost(const Instruction *I, 170 const Value *Src) const { 171 return TTIImpl->getExtCost(I, Src); 172 } 173 174 int TargetTransformInfo::getIntrinsicCost(Intrinsic::ID IID, Type *RetTy, 175 ArrayRef<const Value *> Arguments, 176 const User *U) const { 177 int Cost = TTIImpl->getIntrinsicCost(IID, RetTy, Arguments, U); 178 assert(Cost >= 0 && "TTI should not produce negative costs!"); 179 return Cost; 180 } 181 182 unsigned TargetTransformInfo::getEstimatedNumberOfCaseClusters( 183 const SwitchInst &SI, unsigned &JTSize, ProfileSummaryInfo *PSI, 184 BlockFrequencyInfo *BFI) const { 185 return TTIImpl->getEstimatedNumberOfCaseClusters(SI, JTSize, PSI, BFI); 186 } 187 188 int TargetTransformInfo::getUserCost(const User *U, 189 ArrayRef<const Value *> Operands) const { 190 int Cost = TTIImpl->getUserCost(U, Operands); 191 assert(Cost >= 0 && "TTI should not produce negative costs!"); 192 return Cost; 193 } 194 195 bool TargetTransformInfo::hasBranchDivergence() const { 196 return TTIImpl->hasBranchDivergence(); 197 } 198 199 bool TargetTransformInfo::useGPUDivergenceAnalysis() const { 200 return TTIImpl->useGPUDivergenceAnalysis(); 201 } 202 203 bool TargetTransformInfo::isSourceOfDivergence(const Value *V) const { 204 return TTIImpl->isSourceOfDivergence(V); 205 } 206 207 bool llvm::TargetTransformInfo::isAlwaysUniform(const Value *V) const { 208 return TTIImpl->isAlwaysUniform(V); 209 } 210 211 unsigned TargetTransformInfo::getFlatAddressSpace() const { 212 return TTIImpl->getFlatAddressSpace(); 213 } 214 215 bool TargetTransformInfo::collectFlatAddressOperands( 216 SmallVectorImpl<int> &OpIndexes, Intrinsic::ID IID) const { 217 return TTIImpl->collectFlatAddressOperands(OpIndexes, IID); 218 } 219 220 bool TargetTransformInfo::rewriteIntrinsicWithAddressSpace(IntrinsicInst *II, 221 Value *OldV, 222 Value *NewV) const { 223 return TTIImpl->rewriteIntrinsicWithAddressSpace(II, OldV, NewV); 224 } 225 226 bool TargetTransformInfo::isLoweredToCall(const Function *F) const { 227 return TTIImpl->isLoweredToCall(F); 228 } 229 230 bool TargetTransformInfo::isHardwareLoopProfitable( 231 Loop *L, ScalarEvolution &SE, AssumptionCache &AC, 232 TargetLibraryInfo *LibInfo, HardwareLoopInfo &HWLoopInfo) const { 233 return TTIImpl->isHardwareLoopProfitable(L, SE, AC, LibInfo, HWLoopInfo); 234 } 235 236 bool TargetTransformInfo::preferPredicateOverEpilogue( 237 Loop *L, LoopInfo *LI, ScalarEvolution &SE, AssumptionCache &AC, 238 TargetLibraryInfo *TLI, DominatorTree *DT, 239 const LoopAccessInfo *LAI) const { 240 return TTIImpl->preferPredicateOverEpilogue(L, LI, SE, AC, TLI, DT, LAI); 241 } 242 243 void TargetTransformInfo::getUnrollingPreferences( 244 Loop *L, ScalarEvolution &SE, UnrollingPreferences &UP) const { 245 return TTIImpl->getUnrollingPreferences(L, SE, UP); 246 } 247 248 bool TargetTransformInfo::isLegalAddImmediate(int64_t Imm) const { 249 return TTIImpl->isLegalAddImmediate(Imm); 250 } 251 252 bool TargetTransformInfo::isLegalICmpImmediate(int64_t Imm) const { 253 return TTIImpl->isLegalICmpImmediate(Imm); 254 } 255 256 bool TargetTransformInfo::isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV, 257 int64_t BaseOffset, 258 bool HasBaseReg, int64_t Scale, 259 unsigned AddrSpace, 260 Instruction *I) const { 261 return TTIImpl->isLegalAddressingMode(Ty, BaseGV, BaseOffset, HasBaseReg, 262 Scale, AddrSpace, I); 263 } 264 265 bool TargetTransformInfo::isLSRCostLess(LSRCost &C1, LSRCost &C2) const { 266 return TTIImpl->isLSRCostLess(C1, C2); 267 } 268 269 bool TargetTransformInfo::canMacroFuseCmp() const { 270 return TTIImpl->canMacroFuseCmp(); 271 } 272 273 bool TargetTransformInfo::canSaveCmp(Loop *L, BranchInst **BI, 274 ScalarEvolution *SE, LoopInfo *LI, 275 DominatorTree *DT, AssumptionCache *AC, 276 TargetLibraryInfo *LibInfo) const { 277 return TTIImpl->canSaveCmp(L, BI, SE, LI, DT, AC, LibInfo); 278 } 279 280 bool TargetTransformInfo::shouldFavorPostInc() const { 281 return TTIImpl->shouldFavorPostInc(); 282 } 283 284 bool TargetTransformInfo::shouldFavorBackedgeIndex(const Loop *L) const { 285 return TTIImpl->shouldFavorBackedgeIndex(L); 286 } 287 288 bool TargetTransformInfo::isLegalMaskedStore(Type *DataType, 289 MaybeAlign Alignment) const { 290 return TTIImpl->isLegalMaskedStore(DataType, Alignment); 291 } 292 293 bool TargetTransformInfo::isLegalMaskedLoad(Type *DataType, 294 MaybeAlign Alignment) const { 295 return TTIImpl->isLegalMaskedLoad(DataType, Alignment); 296 } 297 298 bool TargetTransformInfo::isLegalNTStore(Type *DataType, 299 Align Alignment) const { 300 return TTIImpl->isLegalNTStore(DataType, Alignment); 301 } 302 303 bool TargetTransformInfo::isLegalNTLoad(Type *DataType, Align Alignment) const { 304 return TTIImpl->isLegalNTLoad(DataType, Alignment); 305 } 306 307 bool TargetTransformInfo::isLegalMaskedGather(Type *DataType, 308 MaybeAlign Alignment) const { 309 return TTIImpl->isLegalMaskedGather(DataType, Alignment); 310 } 311 312 bool TargetTransformInfo::isLegalMaskedScatter(Type *DataType, 313 MaybeAlign Alignment) const { 314 return TTIImpl->isLegalMaskedScatter(DataType, Alignment); 315 } 316 317 bool TargetTransformInfo::isLegalMaskedCompressStore(Type *DataType) const { 318 return TTIImpl->isLegalMaskedCompressStore(DataType); 319 } 320 321 bool TargetTransformInfo::isLegalMaskedExpandLoad(Type *DataType) const { 322 return TTIImpl->isLegalMaskedExpandLoad(DataType); 323 } 324 325 bool TargetTransformInfo::hasDivRemOp(Type *DataType, bool IsSigned) const { 326 return TTIImpl->hasDivRemOp(DataType, IsSigned); 327 } 328 329 bool TargetTransformInfo::hasVolatileVariant(Instruction *I, 330 unsigned AddrSpace) const { 331 return TTIImpl->hasVolatileVariant(I, AddrSpace); 332 } 333 334 bool TargetTransformInfo::prefersVectorizedAddressing() const { 335 return TTIImpl->prefersVectorizedAddressing(); 336 } 337 338 int TargetTransformInfo::getScalingFactorCost(Type *Ty, GlobalValue *BaseGV, 339 int64_t BaseOffset, 340 bool HasBaseReg, int64_t Scale, 341 unsigned AddrSpace) const { 342 int Cost = TTIImpl->getScalingFactorCost(Ty, BaseGV, BaseOffset, HasBaseReg, 343 Scale, AddrSpace); 344 assert(Cost >= 0 && "TTI should not produce negative costs!"); 345 return Cost; 346 } 347 348 bool TargetTransformInfo::LSRWithInstrQueries() const { 349 return TTIImpl->LSRWithInstrQueries(); 350 } 351 352 bool TargetTransformInfo::isTruncateFree(Type *Ty1, Type *Ty2) const { 353 return TTIImpl->isTruncateFree(Ty1, Ty2); 354 } 355 356 bool TargetTransformInfo::isProfitableToHoist(Instruction *I) const { 357 return TTIImpl->isProfitableToHoist(I); 358 } 359 360 bool TargetTransformInfo::useAA() const { return TTIImpl->useAA(); } 361 362 bool TargetTransformInfo::isTypeLegal(Type *Ty) const { 363 return TTIImpl->isTypeLegal(Ty); 364 } 365 366 bool TargetTransformInfo::shouldBuildLookupTables() const { 367 return TTIImpl->shouldBuildLookupTables(); 368 } 369 bool TargetTransformInfo::shouldBuildLookupTablesForConstant( 370 Constant *C) const { 371 return TTIImpl->shouldBuildLookupTablesForConstant(C); 372 } 373 374 bool TargetTransformInfo::useColdCCForColdCall(Function &F) const { 375 return TTIImpl->useColdCCForColdCall(F); 376 } 377 378 unsigned TargetTransformInfo::getScalarizationOverhead(Type *Ty, bool Insert, 379 bool Extract) const { 380 return TTIImpl->getScalarizationOverhead(Ty, Insert, Extract); 381 } 382 383 unsigned TargetTransformInfo::getOperandsScalarizationOverhead( 384 ArrayRef<const Value *> Args, unsigned VF) const { 385 return TTIImpl->getOperandsScalarizationOverhead(Args, VF); 386 } 387 388 bool TargetTransformInfo::supportsEfficientVectorElementLoadStore() const { 389 return TTIImpl->supportsEfficientVectorElementLoadStore(); 390 } 391 392 bool TargetTransformInfo::enableAggressiveInterleaving( 393 bool LoopHasReductions) const { 394 return TTIImpl->enableAggressiveInterleaving(LoopHasReductions); 395 } 396 397 TargetTransformInfo::MemCmpExpansionOptions 398 TargetTransformInfo::enableMemCmpExpansion(bool OptSize, bool IsZeroCmp) const { 399 return TTIImpl->enableMemCmpExpansion(OptSize, IsZeroCmp); 400 } 401 402 bool TargetTransformInfo::enableInterleavedAccessVectorization() const { 403 return TTIImpl->enableInterleavedAccessVectorization(); 404 } 405 406 bool TargetTransformInfo::enableMaskedInterleavedAccessVectorization() const { 407 return TTIImpl->enableMaskedInterleavedAccessVectorization(); 408 } 409 410 bool TargetTransformInfo::isFPVectorizationPotentiallyUnsafe() const { 411 return TTIImpl->isFPVectorizationPotentiallyUnsafe(); 412 } 413 414 bool TargetTransformInfo::allowsMisalignedMemoryAccesses(LLVMContext &Context, 415 unsigned BitWidth, 416 unsigned AddressSpace, 417 unsigned Alignment, 418 bool *Fast) const { 419 return TTIImpl->allowsMisalignedMemoryAccesses(Context, BitWidth, 420 AddressSpace, Alignment, Fast); 421 } 422 423 TargetTransformInfo::PopcntSupportKind 424 TargetTransformInfo::getPopcntSupport(unsigned IntTyWidthInBit) const { 425 return TTIImpl->getPopcntSupport(IntTyWidthInBit); 426 } 427 428 bool TargetTransformInfo::haveFastSqrt(Type *Ty) const { 429 return TTIImpl->haveFastSqrt(Ty); 430 } 431 432 bool TargetTransformInfo::isFCmpOrdCheaperThanFCmpZero(Type *Ty) const { 433 return TTIImpl->isFCmpOrdCheaperThanFCmpZero(Ty); 434 } 435 436 int TargetTransformInfo::getFPOpCost(Type *Ty) const { 437 int Cost = TTIImpl->getFPOpCost(Ty); 438 assert(Cost >= 0 && "TTI should not produce negative costs!"); 439 return Cost; 440 } 441 442 int TargetTransformInfo::getIntImmCodeSizeCost(unsigned Opcode, unsigned Idx, 443 const APInt &Imm, 444 Type *Ty) const { 445 int Cost = TTIImpl->getIntImmCodeSizeCost(Opcode, Idx, Imm, Ty); 446 assert(Cost >= 0 && "TTI should not produce negative costs!"); 447 return Cost; 448 } 449 450 int TargetTransformInfo::getIntImmCost(const APInt &Imm, Type *Ty) const { 451 int Cost = TTIImpl->getIntImmCost(Imm, Ty); 452 assert(Cost >= 0 && "TTI should not produce negative costs!"); 453 return Cost; 454 } 455 456 int TargetTransformInfo::getIntImmCostInst(unsigned Opcode, unsigned Idx, 457 const APInt &Imm, Type *Ty) const { 458 int Cost = TTIImpl->getIntImmCostInst(Opcode, Idx, Imm, Ty); 459 assert(Cost >= 0 && "TTI should not produce negative costs!"); 460 return Cost; 461 } 462 463 int TargetTransformInfo::getIntImmCostIntrin(Intrinsic::ID IID, unsigned Idx, 464 const APInt &Imm, Type *Ty) const { 465 int Cost = TTIImpl->getIntImmCostIntrin(IID, Idx, Imm, Ty); 466 assert(Cost >= 0 && "TTI should not produce negative costs!"); 467 return Cost; 468 } 469 470 unsigned TargetTransformInfo::getNumberOfRegisters(unsigned ClassID) const { 471 return TTIImpl->getNumberOfRegisters(ClassID); 472 } 473 474 unsigned TargetTransformInfo::getRegisterClassForType(bool Vector, 475 Type *Ty) const { 476 return TTIImpl->getRegisterClassForType(Vector, Ty); 477 } 478 479 const char *TargetTransformInfo::getRegisterClassName(unsigned ClassID) const { 480 return TTIImpl->getRegisterClassName(ClassID); 481 } 482 483 unsigned TargetTransformInfo::getRegisterBitWidth(bool Vector) const { 484 return TTIImpl->getRegisterBitWidth(Vector); 485 } 486 487 unsigned TargetTransformInfo::getMinVectorRegisterBitWidth() const { 488 return TTIImpl->getMinVectorRegisterBitWidth(); 489 } 490 491 bool TargetTransformInfo::shouldMaximizeVectorBandwidth(bool OptSize) const { 492 return TTIImpl->shouldMaximizeVectorBandwidth(OptSize); 493 } 494 495 unsigned TargetTransformInfo::getMinimumVF(unsigned ElemWidth) const { 496 return TTIImpl->getMinimumVF(ElemWidth); 497 } 498 499 bool TargetTransformInfo::shouldConsiderAddressTypePromotion( 500 const Instruction &I, bool &AllowPromotionWithoutCommonHeader) const { 501 return TTIImpl->shouldConsiderAddressTypePromotion( 502 I, AllowPromotionWithoutCommonHeader); 503 } 504 505 unsigned TargetTransformInfo::getCacheLineSize() const { 506 return TTIImpl->getCacheLineSize(); 507 } 508 509 llvm::Optional<unsigned> 510 TargetTransformInfo::getCacheSize(CacheLevel Level) const { 511 return TTIImpl->getCacheSize(Level); 512 } 513 514 llvm::Optional<unsigned> 515 TargetTransformInfo::getCacheAssociativity(CacheLevel Level) const { 516 return TTIImpl->getCacheAssociativity(Level); 517 } 518 519 unsigned TargetTransformInfo::getPrefetchDistance() const { 520 return TTIImpl->getPrefetchDistance(); 521 } 522 523 unsigned TargetTransformInfo::getMinPrefetchStride( 524 unsigned NumMemAccesses, unsigned NumStridedMemAccesses, 525 unsigned NumPrefetches, bool HasCall) const { 526 return TTIImpl->getMinPrefetchStride(NumMemAccesses, NumStridedMemAccesses, 527 NumPrefetches, HasCall); 528 } 529 530 unsigned TargetTransformInfo::getMaxPrefetchIterationsAhead() const { 531 return TTIImpl->getMaxPrefetchIterationsAhead(); 532 } 533 534 bool TargetTransformInfo::enableWritePrefetching() const { 535 return TTIImpl->enableWritePrefetching(); 536 } 537 538 unsigned TargetTransformInfo::getMaxInterleaveFactor(unsigned VF) const { 539 return TTIImpl->getMaxInterleaveFactor(VF); 540 } 541 542 TargetTransformInfo::OperandValueKind 543 TargetTransformInfo::getOperandInfo(Value *V, OperandValueProperties &OpProps) { 544 OperandValueKind OpInfo = OK_AnyValue; 545 OpProps = OP_None; 546 547 if (auto *CI = dyn_cast<ConstantInt>(V)) { 548 if (CI->getValue().isPowerOf2()) 549 OpProps = OP_PowerOf2; 550 return OK_UniformConstantValue; 551 } 552 553 // A broadcast shuffle creates a uniform value. 554 // TODO: Add support for non-zero index broadcasts. 555 // TODO: Add support for different source vector width. 556 if (auto *ShuffleInst = dyn_cast<ShuffleVectorInst>(V)) 557 if (ShuffleInst->isZeroEltSplat()) 558 OpInfo = OK_UniformValue; 559 560 const Value *Splat = getSplatValue(V); 561 562 // Check for a splat of a constant or for a non uniform vector of constants 563 // and check if the constant(s) are all powers of two. 564 if (isa<ConstantVector>(V) || isa<ConstantDataVector>(V)) { 565 OpInfo = OK_NonUniformConstantValue; 566 if (Splat) { 567 OpInfo = OK_UniformConstantValue; 568 if (auto *CI = dyn_cast<ConstantInt>(Splat)) 569 if (CI->getValue().isPowerOf2()) 570 OpProps = OP_PowerOf2; 571 } else if (auto *CDS = dyn_cast<ConstantDataSequential>(V)) { 572 OpProps = OP_PowerOf2; 573 for (unsigned I = 0, E = CDS->getNumElements(); I != E; ++I) { 574 if (auto *CI = dyn_cast<ConstantInt>(CDS->getElementAsConstant(I))) 575 if (CI->getValue().isPowerOf2()) 576 continue; 577 OpProps = OP_None; 578 break; 579 } 580 } 581 } 582 583 // Check for a splat of a uniform value. This is not loop aware, so return 584 // true only for the obviously uniform cases (argument, globalvalue) 585 if (Splat && (isa<Argument>(Splat) || isa<GlobalValue>(Splat))) 586 OpInfo = OK_UniformValue; 587 588 return OpInfo; 589 } 590 591 int TargetTransformInfo::getArithmeticInstrCost( 592 unsigned Opcode, Type *Ty, OperandValueKind Opd1Info, 593 OperandValueKind Opd2Info, OperandValueProperties Opd1PropInfo, 594 OperandValueProperties Opd2PropInfo, ArrayRef<const Value *> Args, 595 const Instruction *CxtI) const { 596 int Cost = TTIImpl->getArithmeticInstrCost( 597 Opcode, Ty, Opd1Info, Opd2Info, Opd1PropInfo, Opd2PropInfo, Args, CxtI); 598 assert(Cost >= 0 && "TTI should not produce negative costs!"); 599 return Cost; 600 } 601 602 int TargetTransformInfo::getShuffleCost(ShuffleKind Kind, Type *Ty, int Index, 603 Type *SubTp) const { 604 int Cost = TTIImpl->getShuffleCost(Kind, Ty, Index, SubTp); 605 assert(Cost >= 0 && "TTI should not produce negative costs!"); 606 return Cost; 607 } 608 609 int TargetTransformInfo::getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src, 610 const Instruction *I) const { 611 assert((I == nullptr || I->getOpcode() == Opcode) && 612 "Opcode should reflect passed instruction."); 613 int Cost = TTIImpl->getCastInstrCost(Opcode, Dst, Src, I); 614 assert(Cost >= 0 && "TTI should not produce negative costs!"); 615 return Cost; 616 } 617 618 int TargetTransformInfo::getExtractWithExtendCost(unsigned Opcode, Type *Dst, 619 VectorType *VecTy, 620 unsigned Index) const { 621 int Cost = TTIImpl->getExtractWithExtendCost(Opcode, Dst, VecTy, Index); 622 assert(Cost >= 0 && "TTI should not produce negative costs!"); 623 return Cost; 624 } 625 626 int TargetTransformInfo::getCFInstrCost(unsigned Opcode) const { 627 int Cost = TTIImpl->getCFInstrCost(Opcode); 628 assert(Cost >= 0 && "TTI should not produce negative costs!"); 629 return Cost; 630 } 631 632 int TargetTransformInfo::getCmpSelInstrCost(unsigned Opcode, Type *ValTy, 633 Type *CondTy, 634 const Instruction *I) const { 635 assert((I == nullptr || I->getOpcode() == Opcode) && 636 "Opcode should reflect passed instruction."); 637 int Cost = TTIImpl->getCmpSelInstrCost(Opcode, ValTy, CondTy, I); 638 assert(Cost >= 0 && "TTI should not produce negative costs!"); 639 return Cost; 640 } 641 642 int TargetTransformInfo::getVectorInstrCost(unsigned Opcode, Type *Val, 643 unsigned Index) const { 644 int Cost = TTIImpl->getVectorInstrCost(Opcode, Val, Index); 645 assert(Cost >= 0 && "TTI should not produce negative costs!"); 646 return Cost; 647 } 648 649 int TargetTransformInfo::getMemoryOpCost(unsigned Opcode, Type *Src, 650 MaybeAlign Alignment, 651 unsigned AddressSpace, 652 const Instruction *I) const { 653 assert((I == nullptr || I->getOpcode() == Opcode) && 654 "Opcode should reflect passed instruction."); 655 int Cost = TTIImpl->getMemoryOpCost(Opcode, Src, Alignment, AddressSpace, I); 656 assert(Cost >= 0 && "TTI should not produce negative costs!"); 657 return Cost; 658 } 659 660 int TargetTransformInfo::getMaskedMemoryOpCost(unsigned Opcode, Type *Src, 661 unsigned Alignment, 662 unsigned AddressSpace) const { 663 int Cost = 664 TTIImpl->getMaskedMemoryOpCost(Opcode, Src, Alignment, AddressSpace); 665 assert(Cost >= 0 && "TTI should not produce negative costs!"); 666 return Cost; 667 } 668 669 int TargetTransformInfo::getGatherScatterOpCost(unsigned Opcode, Type *DataTy, 670 Value *Ptr, bool VariableMask, 671 unsigned Alignment, 672 const Instruction *I) const { 673 int Cost = TTIImpl->getGatherScatterOpCost(Opcode, DataTy, Ptr, VariableMask, 674 Alignment, I); 675 assert(Cost >= 0 && "TTI should not produce negative costs!"); 676 return Cost; 677 } 678 679 int TargetTransformInfo::getInterleavedMemoryOpCost( 680 unsigned Opcode, Type *VecTy, unsigned Factor, ArrayRef<unsigned> Indices, 681 unsigned Alignment, unsigned AddressSpace, bool UseMaskForCond, 682 bool UseMaskForGaps) const { 683 int Cost = TTIImpl->getInterleavedMemoryOpCost( 684 Opcode, VecTy, Factor, Indices, Alignment, AddressSpace, UseMaskForCond, 685 UseMaskForGaps); 686 assert(Cost >= 0 && "TTI should not produce negative costs!"); 687 return Cost; 688 } 689 690 int TargetTransformInfo::getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy, 691 ArrayRef<Type *> Tys, 692 FastMathFlags FMF, 693 unsigned ScalarizationCostPassed, 694 const Instruction *I) const { 695 int Cost = TTIImpl->getIntrinsicInstrCost(ID, RetTy, Tys, FMF, 696 ScalarizationCostPassed, I); 697 assert(Cost >= 0 && "TTI should not produce negative costs!"); 698 return Cost; 699 } 700 701 int TargetTransformInfo::getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy, 702 ArrayRef<Value *> Args, 703 FastMathFlags FMF, unsigned VF, 704 const Instruction *I) const { 705 int Cost = TTIImpl->getIntrinsicInstrCost(ID, RetTy, Args, FMF, VF, I); 706 assert(Cost >= 0 && "TTI should not produce negative costs!"); 707 return Cost; 708 } 709 710 int TargetTransformInfo::getCallInstrCost(Function *F, Type *RetTy, 711 ArrayRef<Type *> Tys) const { 712 int Cost = TTIImpl->getCallInstrCost(F, RetTy, Tys); 713 assert(Cost >= 0 && "TTI should not produce negative costs!"); 714 return Cost; 715 } 716 717 unsigned TargetTransformInfo::getNumberOfParts(Type *Tp) const { 718 return TTIImpl->getNumberOfParts(Tp); 719 } 720 721 int TargetTransformInfo::getAddressComputationCost(Type *Tp, 722 ScalarEvolution *SE, 723 const SCEV *Ptr) const { 724 int Cost = TTIImpl->getAddressComputationCost(Tp, SE, Ptr); 725 assert(Cost >= 0 && "TTI should not produce negative costs!"); 726 return Cost; 727 } 728 729 int TargetTransformInfo::getMemcpyCost(const Instruction *I) const { 730 int Cost = TTIImpl->getMemcpyCost(I); 731 assert(Cost >= 0 && "TTI should not produce negative costs!"); 732 return Cost; 733 } 734 735 int TargetTransformInfo::getArithmeticReductionCost(unsigned Opcode, Type *Ty, 736 bool IsPairwiseForm) const { 737 int Cost = TTIImpl->getArithmeticReductionCost(Opcode, Ty, IsPairwiseForm); 738 assert(Cost >= 0 && "TTI should not produce negative costs!"); 739 return Cost; 740 } 741 742 int TargetTransformInfo::getMinMaxReductionCost(Type *Ty, Type *CondTy, 743 bool IsPairwiseForm, 744 bool IsUnsigned) const { 745 int Cost = 746 TTIImpl->getMinMaxReductionCost(Ty, CondTy, IsPairwiseForm, IsUnsigned); 747 assert(Cost >= 0 && "TTI should not produce negative costs!"); 748 return Cost; 749 } 750 751 unsigned 752 TargetTransformInfo::getCostOfKeepingLiveOverCall(ArrayRef<Type *> Tys) const { 753 return TTIImpl->getCostOfKeepingLiveOverCall(Tys); 754 } 755 756 bool TargetTransformInfo::getTgtMemIntrinsic(IntrinsicInst *Inst, 757 MemIntrinsicInfo &Info) const { 758 return TTIImpl->getTgtMemIntrinsic(Inst, Info); 759 } 760 761 unsigned TargetTransformInfo::getAtomicMemIntrinsicMaxElementSize() const { 762 return TTIImpl->getAtomicMemIntrinsicMaxElementSize(); 763 } 764 765 Value *TargetTransformInfo::getOrCreateResultFromMemIntrinsic( 766 IntrinsicInst *Inst, Type *ExpectedType) const { 767 return TTIImpl->getOrCreateResultFromMemIntrinsic(Inst, ExpectedType); 768 } 769 770 Type *TargetTransformInfo::getMemcpyLoopLoweringType( 771 LLVMContext &Context, Value *Length, unsigned SrcAddrSpace, 772 unsigned DestAddrSpace, unsigned SrcAlign, unsigned DestAlign) const { 773 return TTIImpl->getMemcpyLoopLoweringType(Context, Length, SrcAddrSpace, 774 DestAddrSpace, SrcAlign, DestAlign); 775 } 776 777 void TargetTransformInfo::getMemcpyLoopResidualLoweringType( 778 SmallVectorImpl<Type *> &OpsOut, LLVMContext &Context, 779 unsigned RemainingBytes, unsigned SrcAddrSpace, unsigned DestAddrSpace, 780 unsigned SrcAlign, unsigned DestAlign) const { 781 TTIImpl->getMemcpyLoopResidualLoweringType(OpsOut, Context, RemainingBytes, 782 SrcAddrSpace, DestAddrSpace, 783 SrcAlign, DestAlign); 784 } 785 786 bool TargetTransformInfo::areInlineCompatible(const Function *Caller, 787 const Function *Callee) const { 788 return TTIImpl->areInlineCompatible(Caller, Callee); 789 } 790 791 bool TargetTransformInfo::areFunctionArgsABICompatible( 792 const Function *Caller, const Function *Callee, 793 SmallPtrSetImpl<Argument *> &Args) const { 794 return TTIImpl->areFunctionArgsABICompatible(Caller, Callee, Args); 795 } 796 797 bool TargetTransformInfo::isIndexedLoadLegal(MemIndexedMode Mode, 798 Type *Ty) const { 799 return TTIImpl->isIndexedLoadLegal(Mode, Ty); 800 } 801 802 bool TargetTransformInfo::isIndexedStoreLegal(MemIndexedMode Mode, 803 Type *Ty) const { 804 return TTIImpl->isIndexedStoreLegal(Mode, Ty); 805 } 806 807 unsigned TargetTransformInfo::getLoadStoreVecRegBitWidth(unsigned AS) const { 808 return TTIImpl->getLoadStoreVecRegBitWidth(AS); 809 } 810 811 bool TargetTransformInfo::isLegalToVectorizeLoad(LoadInst *LI) const { 812 return TTIImpl->isLegalToVectorizeLoad(LI); 813 } 814 815 bool TargetTransformInfo::isLegalToVectorizeStore(StoreInst *SI) const { 816 return TTIImpl->isLegalToVectorizeStore(SI); 817 } 818 819 bool TargetTransformInfo::isLegalToVectorizeLoadChain( 820 unsigned ChainSizeInBytes, unsigned Alignment, unsigned AddrSpace) const { 821 return TTIImpl->isLegalToVectorizeLoadChain(ChainSizeInBytes, Alignment, 822 AddrSpace); 823 } 824 825 bool TargetTransformInfo::isLegalToVectorizeStoreChain( 826 unsigned ChainSizeInBytes, unsigned Alignment, unsigned AddrSpace) const { 827 return TTIImpl->isLegalToVectorizeStoreChain(ChainSizeInBytes, Alignment, 828 AddrSpace); 829 } 830 831 unsigned TargetTransformInfo::getLoadVectorFactor(unsigned VF, 832 unsigned LoadSize, 833 unsigned ChainSizeInBytes, 834 VectorType *VecTy) const { 835 return TTIImpl->getLoadVectorFactor(VF, LoadSize, ChainSizeInBytes, VecTy); 836 } 837 838 unsigned TargetTransformInfo::getStoreVectorFactor(unsigned VF, 839 unsigned StoreSize, 840 unsigned ChainSizeInBytes, 841 VectorType *VecTy) const { 842 return TTIImpl->getStoreVectorFactor(VF, StoreSize, ChainSizeInBytes, VecTy); 843 } 844 845 bool TargetTransformInfo::useReductionIntrinsic(unsigned Opcode, Type *Ty, 846 ReductionFlags Flags) const { 847 return TTIImpl->useReductionIntrinsic(Opcode, Ty, Flags); 848 } 849 850 bool TargetTransformInfo::shouldExpandReduction(const IntrinsicInst *II) const { 851 return TTIImpl->shouldExpandReduction(II); 852 } 853 854 unsigned TargetTransformInfo::getGISelRematGlobalCost() const { 855 return TTIImpl->getGISelRematGlobalCost(); 856 } 857 858 int TargetTransformInfo::getInstructionLatency(const Instruction *I) const { 859 return TTIImpl->getInstructionLatency(I); 860 } 861 862 static bool matchPairwiseShuffleMask(ShuffleVectorInst *SI, bool IsLeft, 863 unsigned Level) { 864 // We don't need a shuffle if we just want to have element 0 in position 0 of 865 // the vector. 866 if (!SI && Level == 0 && IsLeft) 867 return true; 868 else if (!SI) 869 return false; 870 871 SmallVector<int, 32> Mask(SI->getType()->getNumElements(), -1); 872 873 // Build a mask of 0, 2, ... (left) or 1, 3, ... (right) depending on whether 874 // we look at the left or right side. 875 for (unsigned i = 0, e = (1 << Level), val = !IsLeft; i != e; ++i, val += 2) 876 Mask[i] = val; 877 878 ArrayRef<int> ActualMask = SI->getShuffleMask(); 879 return Mask == ActualMask; 880 } 881 882 namespace { 883 /// Kind of the reduction data. 884 enum ReductionKind { 885 RK_None, /// Not a reduction. 886 RK_Arithmetic, /// Binary reduction data. 887 RK_MinMax, /// Min/max reduction data. 888 RK_UnsignedMinMax, /// Unsigned min/max reduction data. 889 }; 890 /// Contains opcode + LHS/RHS parts of the reduction operations. 891 struct ReductionData { 892 ReductionData() = delete; 893 ReductionData(ReductionKind Kind, unsigned Opcode, Value *LHS, Value *RHS) 894 : Opcode(Opcode), LHS(LHS), RHS(RHS), Kind(Kind) { 895 assert(Kind != RK_None && "expected binary or min/max reduction only."); 896 } 897 unsigned Opcode = 0; 898 Value *LHS = nullptr; 899 Value *RHS = nullptr; 900 ReductionKind Kind = RK_None; 901 bool hasSameData(ReductionData &RD) const { 902 return Kind == RD.Kind && Opcode == RD.Opcode; 903 } 904 }; 905 } // namespace 906 907 static Optional<ReductionData> getReductionData(Instruction *I) { 908 Value *L, *R; 909 if (m_BinOp(m_Value(L), m_Value(R)).match(I)) 910 return ReductionData(RK_Arithmetic, I->getOpcode(), L, R); 911 if (auto *SI = dyn_cast<SelectInst>(I)) { 912 if (m_SMin(m_Value(L), m_Value(R)).match(SI) || 913 m_SMax(m_Value(L), m_Value(R)).match(SI) || 914 m_OrdFMin(m_Value(L), m_Value(R)).match(SI) || 915 m_OrdFMax(m_Value(L), m_Value(R)).match(SI) || 916 m_UnordFMin(m_Value(L), m_Value(R)).match(SI) || 917 m_UnordFMax(m_Value(L), m_Value(R)).match(SI)) { 918 auto *CI = cast<CmpInst>(SI->getCondition()); 919 return ReductionData(RK_MinMax, CI->getOpcode(), L, R); 920 } 921 if (m_UMin(m_Value(L), m_Value(R)).match(SI) || 922 m_UMax(m_Value(L), m_Value(R)).match(SI)) { 923 auto *CI = cast<CmpInst>(SI->getCondition()); 924 return ReductionData(RK_UnsignedMinMax, CI->getOpcode(), L, R); 925 } 926 } 927 return llvm::None; 928 } 929 930 static ReductionKind matchPairwiseReductionAtLevel(Instruction *I, 931 unsigned Level, 932 unsigned NumLevels) { 933 // Match one level of pairwise operations. 934 // %rdx.shuf.0.0 = shufflevector <4 x float> %rdx, <4 x float> undef, 935 // <4 x i32> <i32 0, i32 2 , i32 undef, i32 undef> 936 // %rdx.shuf.0.1 = shufflevector <4 x float> %rdx, <4 x float> undef, 937 // <4 x i32> <i32 1, i32 3, i32 undef, i32 undef> 938 // %bin.rdx.0 = fadd <4 x float> %rdx.shuf.0.0, %rdx.shuf.0.1 939 if (!I) 940 return RK_None; 941 942 assert(I->getType()->isVectorTy() && "Expecting a vector type"); 943 944 Optional<ReductionData> RD = getReductionData(I); 945 if (!RD) 946 return RK_None; 947 948 ShuffleVectorInst *LS = dyn_cast<ShuffleVectorInst>(RD->LHS); 949 if (!LS && Level) 950 return RK_None; 951 ShuffleVectorInst *RS = dyn_cast<ShuffleVectorInst>(RD->RHS); 952 if (!RS && Level) 953 return RK_None; 954 955 // On level 0 we can omit one shufflevector instruction. 956 if (!Level && !RS && !LS) 957 return RK_None; 958 959 // Shuffle inputs must match. 960 Value *NextLevelOpL = LS ? LS->getOperand(0) : nullptr; 961 Value *NextLevelOpR = RS ? RS->getOperand(0) : nullptr; 962 Value *NextLevelOp = nullptr; 963 if (NextLevelOpR && NextLevelOpL) { 964 // If we have two shuffles their operands must match. 965 if (NextLevelOpL != NextLevelOpR) 966 return RK_None; 967 968 NextLevelOp = NextLevelOpL; 969 } else if (Level == 0 && (NextLevelOpR || NextLevelOpL)) { 970 // On the first level we can omit the shufflevector <0, undef,...>. So the 971 // input to the other shufflevector <1, undef> must match with one of the 972 // inputs to the current binary operation. 973 // Example: 974 // %NextLevelOpL = shufflevector %R, <1, undef ...> 975 // %BinOp = fadd %NextLevelOpL, %R 976 if (NextLevelOpL && NextLevelOpL != RD->RHS) 977 return RK_None; 978 else if (NextLevelOpR && NextLevelOpR != RD->LHS) 979 return RK_None; 980 981 NextLevelOp = NextLevelOpL ? RD->RHS : RD->LHS; 982 } else 983 return RK_None; 984 985 // Check that the next levels binary operation exists and matches with the 986 // current one. 987 if (Level + 1 != NumLevels) { 988 Optional<ReductionData> NextLevelRD = 989 getReductionData(cast<Instruction>(NextLevelOp)); 990 if (!NextLevelRD || !RD->hasSameData(*NextLevelRD)) 991 return RK_None; 992 } 993 994 // Shuffle mask for pairwise operation must match. 995 if (matchPairwiseShuffleMask(LS, /*IsLeft=*/true, Level)) { 996 if (!matchPairwiseShuffleMask(RS, /*IsLeft=*/false, Level)) 997 return RK_None; 998 } else if (matchPairwiseShuffleMask(RS, /*IsLeft=*/true, Level)) { 999 if (!matchPairwiseShuffleMask(LS, /*IsLeft=*/false, Level)) 1000 return RK_None; 1001 } else { 1002 return RK_None; 1003 } 1004 1005 if (++Level == NumLevels) 1006 return RD->Kind; 1007 1008 // Match next level. 1009 return matchPairwiseReductionAtLevel(cast<Instruction>(NextLevelOp), Level, 1010 NumLevels); 1011 } 1012 1013 static ReductionKind matchPairwiseReduction(const ExtractElementInst *ReduxRoot, 1014 unsigned &Opcode, Type *&Ty) { 1015 if (!EnableReduxCost) 1016 return RK_None; 1017 1018 // Need to extract the first element. 1019 ConstantInt *CI = dyn_cast<ConstantInt>(ReduxRoot->getOperand(1)); 1020 unsigned Idx = ~0u; 1021 if (CI) 1022 Idx = CI->getZExtValue(); 1023 if (Idx != 0) 1024 return RK_None; 1025 1026 auto *RdxStart = dyn_cast<Instruction>(ReduxRoot->getOperand(0)); 1027 if (!RdxStart) 1028 return RK_None; 1029 Optional<ReductionData> RD = getReductionData(RdxStart); 1030 if (!RD) 1031 return RK_None; 1032 1033 auto *VecTy = cast<VectorType>(RdxStart->getType()); 1034 unsigned NumVecElems = VecTy->getNumElements(); 1035 if (!isPowerOf2_32(NumVecElems)) 1036 return RK_None; 1037 1038 // We look for a sequence of shuffle,shuffle,add triples like the following 1039 // that builds a pairwise reduction tree. 1040 // 1041 // (X0, X1, X2, X3) 1042 // (X0 + X1, X2 + X3, undef, undef) 1043 // ((X0 + X1) + (X2 + X3), undef, undef, undef) 1044 // 1045 // %rdx.shuf.0.0 = shufflevector <4 x float> %rdx, <4 x float> undef, 1046 // <4 x i32> <i32 0, i32 2 , i32 undef, i32 undef> 1047 // %rdx.shuf.0.1 = shufflevector <4 x float> %rdx, <4 x float> undef, 1048 // <4 x i32> <i32 1, i32 3, i32 undef, i32 undef> 1049 // %bin.rdx.0 = fadd <4 x float> %rdx.shuf.0.0, %rdx.shuf.0.1 1050 // %rdx.shuf.1.0 = shufflevector <4 x float> %bin.rdx.0, <4 x float> undef, 1051 // <4 x i32> <i32 0, i32 undef, i32 undef, i32 undef> 1052 // %rdx.shuf.1.1 = shufflevector <4 x float> %bin.rdx.0, <4 x float> undef, 1053 // <4 x i32> <i32 1, i32 undef, i32 undef, i32 undef> 1054 // %bin.rdx8 = fadd <4 x float> %rdx.shuf.1.0, %rdx.shuf.1.1 1055 // %r = extractelement <4 x float> %bin.rdx8, i32 0 1056 if (matchPairwiseReductionAtLevel(RdxStart, 0, Log2_32(NumVecElems)) == 1057 RK_None) 1058 return RK_None; 1059 1060 Opcode = RD->Opcode; 1061 Ty = VecTy; 1062 1063 return RD->Kind; 1064 } 1065 1066 static std::pair<Value *, ShuffleVectorInst *> 1067 getShuffleAndOtherOprd(Value *L, Value *R) { 1068 ShuffleVectorInst *S = nullptr; 1069 1070 if ((S = dyn_cast<ShuffleVectorInst>(L))) 1071 return std::make_pair(R, S); 1072 1073 S = dyn_cast<ShuffleVectorInst>(R); 1074 return std::make_pair(L, S); 1075 } 1076 1077 static ReductionKind 1078 matchVectorSplittingReduction(const ExtractElementInst *ReduxRoot, 1079 unsigned &Opcode, Type *&Ty) { 1080 if (!EnableReduxCost) 1081 return RK_None; 1082 1083 // Need to extract the first element. 1084 ConstantInt *CI = dyn_cast<ConstantInt>(ReduxRoot->getOperand(1)); 1085 unsigned Idx = ~0u; 1086 if (CI) 1087 Idx = CI->getZExtValue(); 1088 if (Idx != 0) 1089 return RK_None; 1090 1091 auto *RdxStart = dyn_cast<Instruction>(ReduxRoot->getOperand(0)); 1092 if (!RdxStart) 1093 return RK_None; 1094 Optional<ReductionData> RD = getReductionData(RdxStart); 1095 if (!RD) 1096 return RK_None; 1097 1098 auto *VecTy = cast<VectorType>(ReduxRoot->getOperand(0)->getType()); 1099 unsigned NumVecElems = VecTy->getNumElements(); 1100 if (!isPowerOf2_32(NumVecElems)) 1101 return RK_None; 1102 1103 // We look for a sequence of shuffles and adds like the following matching one 1104 // fadd, shuffle vector pair at a time. 1105 // 1106 // %rdx.shuf = shufflevector <4 x float> %rdx, <4 x float> undef, 1107 // <4 x i32> <i32 2, i32 3, i32 undef, i32 undef> 1108 // %bin.rdx = fadd <4 x float> %rdx, %rdx.shuf 1109 // %rdx.shuf7 = shufflevector <4 x float> %bin.rdx, <4 x float> undef, 1110 // <4 x i32> <i32 1, i32 undef, i32 undef, i32 undef> 1111 // %bin.rdx8 = fadd <4 x float> %bin.rdx, %rdx.shuf7 1112 // %r = extractelement <4 x float> %bin.rdx8, i32 0 1113 1114 unsigned MaskStart = 1; 1115 Instruction *RdxOp = RdxStart; 1116 SmallVector<int, 32> ShuffleMask(NumVecElems, 0); 1117 unsigned NumVecElemsRemain = NumVecElems; 1118 while (NumVecElemsRemain - 1) { 1119 // Check for the right reduction operation. 1120 if (!RdxOp) 1121 return RK_None; 1122 Optional<ReductionData> RDLevel = getReductionData(RdxOp); 1123 if (!RDLevel || !RDLevel->hasSameData(*RD)) 1124 return RK_None; 1125 1126 Value *NextRdxOp; 1127 ShuffleVectorInst *Shuffle; 1128 std::tie(NextRdxOp, Shuffle) = 1129 getShuffleAndOtherOprd(RDLevel->LHS, RDLevel->RHS); 1130 1131 // Check the current reduction operation and the shuffle use the same value. 1132 if (Shuffle == nullptr) 1133 return RK_None; 1134 if (Shuffle->getOperand(0) != NextRdxOp) 1135 return RK_None; 1136 1137 // Check that shuffle masks matches. 1138 for (unsigned j = 0; j != MaskStart; ++j) 1139 ShuffleMask[j] = MaskStart + j; 1140 // Fill the rest of the mask with -1 for undef. 1141 std::fill(&ShuffleMask[MaskStart], ShuffleMask.end(), -1); 1142 1143 ArrayRef<int> Mask = Shuffle->getShuffleMask(); 1144 if (ShuffleMask != Mask) 1145 return RK_None; 1146 1147 RdxOp = dyn_cast<Instruction>(NextRdxOp); 1148 NumVecElemsRemain /= 2; 1149 MaskStart *= 2; 1150 } 1151 1152 Opcode = RD->Opcode; 1153 Ty = VecTy; 1154 return RD->Kind; 1155 } 1156 1157 int TargetTransformInfo::getInstructionThroughput(const Instruction *I) const { 1158 switch (I->getOpcode()) { 1159 case Instruction::GetElementPtr: 1160 return getUserCost(I); 1161 1162 case Instruction::Ret: 1163 case Instruction::PHI: 1164 case Instruction::Br: { 1165 return getCFInstrCost(I->getOpcode()); 1166 } 1167 case Instruction::Add: 1168 case Instruction::FAdd: 1169 case Instruction::Sub: 1170 case Instruction::FSub: 1171 case Instruction::Mul: 1172 case Instruction::FMul: 1173 case Instruction::UDiv: 1174 case Instruction::SDiv: 1175 case Instruction::FDiv: 1176 case Instruction::URem: 1177 case Instruction::SRem: 1178 case Instruction::FRem: 1179 case Instruction::Shl: 1180 case Instruction::LShr: 1181 case Instruction::AShr: 1182 case Instruction::And: 1183 case Instruction::Or: 1184 case Instruction::Xor: { 1185 TargetTransformInfo::OperandValueKind Op1VK, Op2VK; 1186 TargetTransformInfo::OperandValueProperties Op1VP, Op2VP; 1187 Op1VK = getOperandInfo(I->getOperand(0), Op1VP); 1188 Op2VK = getOperandInfo(I->getOperand(1), Op2VP); 1189 SmallVector<const Value *, 2> Operands(I->operand_values()); 1190 return getArithmeticInstrCost(I->getOpcode(), I->getType(), Op1VK, Op2VK, 1191 Op1VP, Op2VP, Operands, I); 1192 } 1193 case Instruction::FNeg: { 1194 TargetTransformInfo::OperandValueKind Op1VK, Op2VK; 1195 TargetTransformInfo::OperandValueProperties Op1VP, Op2VP; 1196 Op1VK = getOperandInfo(I->getOperand(0), Op1VP); 1197 Op2VK = OK_AnyValue; 1198 Op2VP = OP_None; 1199 SmallVector<const Value *, 2> Operands(I->operand_values()); 1200 return getArithmeticInstrCost(I->getOpcode(), I->getType(), Op1VK, Op2VK, 1201 Op1VP, Op2VP, Operands, I); 1202 } 1203 case Instruction::Select: { 1204 const SelectInst *SI = cast<SelectInst>(I); 1205 Type *CondTy = SI->getCondition()->getType(); 1206 return getCmpSelInstrCost(I->getOpcode(), I->getType(), CondTy, I); 1207 } 1208 case Instruction::ICmp: 1209 case Instruction::FCmp: { 1210 Type *ValTy = I->getOperand(0)->getType(); 1211 return getCmpSelInstrCost(I->getOpcode(), ValTy, I->getType(), I); 1212 } 1213 case Instruction::Store: { 1214 const StoreInst *SI = cast<StoreInst>(I); 1215 Type *ValTy = SI->getValueOperand()->getType(); 1216 return getMemoryOpCost(I->getOpcode(), ValTy, 1217 MaybeAlign(SI->getAlignment()), 1218 SI->getPointerAddressSpace(), I); 1219 } 1220 case Instruction::Load: { 1221 const LoadInst *LI = cast<LoadInst>(I); 1222 return getMemoryOpCost(I->getOpcode(), I->getType(), 1223 MaybeAlign(LI->getAlignment()), 1224 LI->getPointerAddressSpace(), I); 1225 } 1226 case Instruction::ZExt: 1227 case Instruction::SExt: 1228 case Instruction::FPToUI: 1229 case Instruction::FPToSI: 1230 case Instruction::FPExt: 1231 case Instruction::PtrToInt: 1232 case Instruction::IntToPtr: 1233 case Instruction::SIToFP: 1234 case Instruction::UIToFP: 1235 case Instruction::Trunc: 1236 case Instruction::FPTrunc: 1237 case Instruction::BitCast: 1238 case Instruction::AddrSpaceCast: { 1239 Type *SrcTy = I->getOperand(0)->getType(); 1240 return getCastInstrCost(I->getOpcode(), I->getType(), SrcTy, I); 1241 } 1242 case Instruction::ExtractElement: { 1243 const ExtractElementInst *EEI = cast<ExtractElementInst>(I); 1244 ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1)); 1245 unsigned Idx = -1; 1246 if (CI) 1247 Idx = CI->getZExtValue(); 1248 1249 // Try to match a reduction sequence (series of shufflevector and vector 1250 // adds followed by a extractelement). 1251 unsigned ReduxOpCode; 1252 Type *ReduxType; 1253 1254 switch (matchVectorSplittingReduction(EEI, ReduxOpCode, ReduxType)) { 1255 case RK_Arithmetic: 1256 return getArithmeticReductionCost(ReduxOpCode, ReduxType, 1257 /*IsPairwiseForm=*/false); 1258 case RK_MinMax: 1259 return getMinMaxReductionCost( 1260 ReduxType, CmpInst::makeCmpResultType(ReduxType), 1261 /*IsPairwiseForm=*/false, /*IsUnsigned=*/false); 1262 case RK_UnsignedMinMax: 1263 return getMinMaxReductionCost( 1264 ReduxType, CmpInst::makeCmpResultType(ReduxType), 1265 /*IsPairwiseForm=*/false, /*IsUnsigned=*/true); 1266 case RK_None: 1267 break; 1268 } 1269 1270 switch (matchPairwiseReduction(EEI, ReduxOpCode, ReduxType)) { 1271 case RK_Arithmetic: 1272 return getArithmeticReductionCost(ReduxOpCode, ReduxType, 1273 /*IsPairwiseForm=*/true); 1274 case RK_MinMax: 1275 return getMinMaxReductionCost( 1276 ReduxType, CmpInst::makeCmpResultType(ReduxType), 1277 /*IsPairwiseForm=*/true, /*IsUnsigned=*/false); 1278 case RK_UnsignedMinMax: 1279 return getMinMaxReductionCost( 1280 ReduxType, CmpInst::makeCmpResultType(ReduxType), 1281 /*IsPairwiseForm=*/true, /*IsUnsigned=*/true); 1282 case RK_None: 1283 break; 1284 } 1285 1286 return getVectorInstrCost(I->getOpcode(), EEI->getOperand(0)->getType(), 1287 Idx); 1288 } 1289 case Instruction::InsertElement: { 1290 const InsertElementInst *IE = cast<InsertElementInst>(I); 1291 ConstantInt *CI = dyn_cast<ConstantInt>(IE->getOperand(2)); 1292 unsigned Idx = -1; 1293 if (CI) 1294 Idx = CI->getZExtValue(); 1295 return getVectorInstrCost(I->getOpcode(), IE->getType(), Idx); 1296 } 1297 case Instruction::ExtractValue: 1298 return 0; // Model all ExtractValue nodes as free. 1299 case Instruction::ShuffleVector: { 1300 const ShuffleVectorInst *Shuffle = cast<ShuffleVectorInst>(I); 1301 Type *Ty = Shuffle->getType(); 1302 Type *SrcTy = Shuffle->getOperand(0)->getType(); 1303 1304 // TODO: Identify and add costs for insert subvector, etc. 1305 int SubIndex; 1306 if (Shuffle->isExtractSubvectorMask(SubIndex)) 1307 return TTIImpl->getShuffleCost(SK_ExtractSubvector, SrcTy, SubIndex, Ty); 1308 1309 if (Shuffle->changesLength()) 1310 return -1; 1311 1312 if (Shuffle->isIdentity()) 1313 return 0; 1314 1315 if (Shuffle->isReverse()) 1316 return TTIImpl->getShuffleCost(SK_Reverse, Ty, 0, nullptr); 1317 1318 if (Shuffle->isSelect()) 1319 return TTIImpl->getShuffleCost(SK_Select, Ty, 0, nullptr); 1320 1321 if (Shuffle->isTranspose()) 1322 return TTIImpl->getShuffleCost(SK_Transpose, Ty, 0, nullptr); 1323 1324 if (Shuffle->isZeroEltSplat()) 1325 return TTIImpl->getShuffleCost(SK_Broadcast, Ty, 0, nullptr); 1326 1327 if (Shuffle->isSingleSource()) 1328 return TTIImpl->getShuffleCost(SK_PermuteSingleSrc, Ty, 0, nullptr); 1329 1330 return TTIImpl->getShuffleCost(SK_PermuteTwoSrc, Ty, 0, nullptr); 1331 } 1332 case Instruction::Call: 1333 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) { 1334 SmallVector<Value *, 4> Args(II->arg_operands()); 1335 1336 FastMathFlags FMF; 1337 if (auto *FPMO = dyn_cast<FPMathOperator>(II)) 1338 FMF = FPMO->getFastMathFlags(); 1339 1340 return getIntrinsicInstrCost(II->getIntrinsicID(), II->getType(), Args, 1341 FMF, 1, II); 1342 } 1343 return -1; 1344 default: 1345 // We don't have any information on this instruction. 1346 return -1; 1347 } 1348 } 1349 1350 TargetTransformInfo::Concept::~Concept() {} 1351 1352 TargetIRAnalysis::TargetIRAnalysis() : TTICallback(&getDefaultTTI) {} 1353 1354 TargetIRAnalysis::TargetIRAnalysis( 1355 std::function<Result(const Function &)> TTICallback) 1356 : TTICallback(std::move(TTICallback)) {} 1357 1358 TargetIRAnalysis::Result TargetIRAnalysis::run(const Function &F, 1359 FunctionAnalysisManager &) { 1360 return TTICallback(F); 1361 } 1362 1363 AnalysisKey TargetIRAnalysis::Key; 1364 1365 TargetIRAnalysis::Result TargetIRAnalysis::getDefaultTTI(const Function &F) { 1366 return Result(F.getParent()->getDataLayout()); 1367 } 1368 1369 // Register the basic pass. 1370 INITIALIZE_PASS(TargetTransformInfoWrapperPass, "tti", 1371 "Target Transform Information", false, true) 1372 char TargetTransformInfoWrapperPass::ID = 0; 1373 1374 void TargetTransformInfoWrapperPass::anchor() {} 1375 1376 TargetTransformInfoWrapperPass::TargetTransformInfoWrapperPass() 1377 : ImmutablePass(ID) { 1378 initializeTargetTransformInfoWrapperPassPass( 1379 *PassRegistry::getPassRegistry()); 1380 } 1381 1382 TargetTransformInfoWrapperPass::TargetTransformInfoWrapperPass( 1383 TargetIRAnalysis TIRA) 1384 : ImmutablePass(ID), TIRA(std::move(TIRA)) { 1385 initializeTargetTransformInfoWrapperPassPass( 1386 *PassRegistry::getPassRegistry()); 1387 } 1388 1389 TargetTransformInfo &TargetTransformInfoWrapperPass::getTTI(const Function &F) { 1390 FunctionAnalysisManager DummyFAM; 1391 TTI = TIRA.run(F, DummyFAM); 1392 return *TTI; 1393 } 1394 1395 ImmutablePass * 1396 llvm::createTargetTransformInfoWrapperPass(TargetIRAnalysis TIRA) { 1397 return new TargetTransformInfoWrapperPass(std::move(TIRA)); 1398 } 1399