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/TargetTransformInfoImpl.h" 11 #include "llvm/IR/CallSite.h" 12 #include "llvm/IR/CFG.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 "llvm/Analysis/CFG.h" 23 #include "llvm/Analysis/LoopIterator.h" 24 #include <utility> 25 26 using namespace llvm; 27 using namespace PatternMatch; 28 29 #define DEBUG_TYPE "tti" 30 31 static cl::opt<bool> EnableReduxCost("costmodel-reduxcost", cl::init(false), 32 cl::Hidden, 33 cl::desc("Recognize reduction patterns.")); 34 35 namespace { 36 /// No-op implementation of the TTI interface using the utility base 37 /// classes. 38 /// 39 /// This is used when no target specific information is available. 40 struct NoTTIImpl : TargetTransformInfoImplCRTPBase<NoTTIImpl> { 41 explicit NoTTIImpl(const DataLayout &DL) 42 : TargetTransformInfoImplCRTPBase<NoTTIImpl>(DL) {} 43 44 unsigned getCacheLineSize() const { return 0; } 45 46 llvm::Optional<unsigned> getCacheSize(TargetTransformInfo::CacheLevel Level) const { 47 switch (Level) { 48 case TargetTransformInfo::CacheLevel::L1D: 49 LLVM_FALLTHROUGH; 50 case TargetTransformInfo::CacheLevel::L2D: 51 return llvm::Optional<unsigned>(); 52 } 53 llvm_unreachable("Unknown TargetTransformInfo::CacheLevel"); 54 } 55 56 llvm::Optional<unsigned> getCacheAssociativity( 57 TargetTransformInfo::CacheLevel Level) const { 58 switch (Level) { 59 case TargetTransformInfo::CacheLevel::L1D: 60 LLVM_FALLTHROUGH; 61 case TargetTransformInfo::CacheLevel::L2D: 62 return llvm::Optional<unsigned>(); 63 } 64 65 llvm_unreachable("Unknown TargetTransformInfo::CacheLevel"); 66 } 67 68 unsigned getPrefetchDistance() const { return 0; } 69 unsigned getMinPrefetchStride() const { return 1; } 70 unsigned getMaxPrefetchIterationsAhead() const { return UINT_MAX; } 71 }; 72 } 73 74 bool HardwareLoopInfo::canAnalyze(LoopInfo &LI) { 75 // If the loop has irreducible control flow, it can not be converted to 76 // Hardware loop. 77 LoopBlocksRPO RPOT(L); 78 RPOT.perform(&LI); 79 if (containsIrreducibleCFG<const BasicBlock *>(RPOT, LI)) 80 return false; 81 return true; 82 } 83 84 bool HardwareLoopInfo::isHardwareLoopCandidate(ScalarEvolution &SE, 85 LoopInfo &LI, DominatorTree &DT, 86 bool ForceNestedLoop, 87 bool ForceHardwareLoopPHI) { 88 SmallVector<BasicBlock *, 4> ExitingBlocks; 89 L->getExitingBlocks(ExitingBlocks); 90 91 for (BasicBlock *BB : ExitingBlocks) { 92 // If we pass the updated counter back through a phi, we need to know 93 // which latch the updated value will be coming from. 94 if (!L->isLoopLatch(BB)) { 95 if (ForceHardwareLoopPHI || CounterInReg) 96 continue; 97 } 98 99 const SCEV *EC = SE.getExitCount(L, BB); 100 if (isa<SCEVCouldNotCompute>(EC)) 101 continue; 102 if (const SCEVConstant *ConstEC = dyn_cast<SCEVConstant>(EC)) { 103 if (ConstEC->getValue()->isZero()) 104 continue; 105 } else if (!SE.isLoopInvariant(EC, L)) 106 continue; 107 108 if (SE.getTypeSizeInBits(EC->getType()) > CountType->getBitWidth()) 109 continue; 110 111 // If this exiting block is contained in a nested loop, it is not eligible 112 // for insertion of the branch-and-decrement since the inner loop would 113 // end up messing up the value in the CTR. 114 if (!IsNestingLegal && LI.getLoopFor(BB) != L && !ForceNestedLoop) 115 continue; 116 117 // We now have a loop-invariant count of loop iterations (which is not the 118 // constant zero) for which we know that this loop will not exit via this 119 // existing block. 120 121 // We need to make sure that this block will run on every loop iteration. 122 // For this to be true, we must dominate all blocks with backedges. Such 123 // blocks are in-loop predecessors to the header block. 124 bool NotAlways = false; 125 for (BasicBlock *Pred : predecessors(L->getHeader())) { 126 if (!L->contains(Pred)) 127 continue; 128 129 if (!DT.dominates(BB, Pred)) { 130 NotAlways = true; 131 break; 132 } 133 } 134 135 if (NotAlways) 136 continue; 137 138 // Make sure this blocks ends with a conditional branch. 139 Instruction *TI = BB->getTerminator(); 140 if (!TI) 141 continue; 142 143 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) { 144 if (!BI->isConditional()) 145 continue; 146 147 ExitBranch = BI; 148 } else 149 continue; 150 151 // Note that this block may not be the loop latch block, even if the loop 152 // has a latch block. 153 ExitBlock = BB; 154 ExitCount = EC; 155 break; 156 } 157 158 if (!ExitBlock) 159 return false; 160 return true; 161 } 162 163 TargetTransformInfo::TargetTransformInfo(const DataLayout &DL) 164 : TTIImpl(new Model<NoTTIImpl>(NoTTIImpl(DL))) {} 165 166 TargetTransformInfo::~TargetTransformInfo() {} 167 168 TargetTransformInfo::TargetTransformInfo(TargetTransformInfo &&Arg) 169 : TTIImpl(std::move(Arg.TTIImpl)) {} 170 171 TargetTransformInfo &TargetTransformInfo::operator=(TargetTransformInfo &&RHS) { 172 TTIImpl = std::move(RHS.TTIImpl); 173 return *this; 174 } 175 176 int TargetTransformInfo::getOperationCost(unsigned Opcode, Type *Ty, 177 Type *OpTy) const { 178 int Cost = TTIImpl->getOperationCost(Opcode, Ty, OpTy); 179 assert(Cost >= 0 && "TTI should not produce negative costs!"); 180 return Cost; 181 } 182 183 int TargetTransformInfo::getCallCost(FunctionType *FTy, int NumArgs, 184 const User *U) const { 185 int Cost = TTIImpl->getCallCost(FTy, NumArgs, U); 186 assert(Cost >= 0 && "TTI should not produce negative costs!"); 187 return Cost; 188 } 189 190 int TargetTransformInfo::getCallCost(const Function *F, 191 ArrayRef<const Value *> Arguments, 192 const User *U) const { 193 int Cost = TTIImpl->getCallCost(F, Arguments, U); 194 assert(Cost >= 0 && "TTI should not produce negative costs!"); 195 return Cost; 196 } 197 198 unsigned TargetTransformInfo::getInliningThresholdMultiplier() const { 199 return TTIImpl->getInliningThresholdMultiplier(); 200 } 201 202 int TargetTransformInfo::getInlinerVectorBonusPercent() const { 203 return TTIImpl->getInlinerVectorBonusPercent(); 204 } 205 206 int TargetTransformInfo::getGEPCost(Type *PointeeType, const Value *Ptr, 207 ArrayRef<const Value *> Operands) const { 208 return TTIImpl->getGEPCost(PointeeType, Ptr, Operands); 209 } 210 211 int TargetTransformInfo::getExtCost(const Instruction *I, 212 const Value *Src) const { 213 return TTIImpl->getExtCost(I, Src); 214 } 215 216 int TargetTransformInfo::getIntrinsicCost( 217 Intrinsic::ID IID, Type *RetTy, ArrayRef<const Value *> Arguments, 218 const User *U) const { 219 int Cost = TTIImpl->getIntrinsicCost(IID, RetTy, Arguments, U); 220 assert(Cost >= 0 && "TTI should not produce negative costs!"); 221 return Cost; 222 } 223 224 unsigned 225 TargetTransformInfo::getEstimatedNumberOfCaseClusters(const SwitchInst &SI, 226 unsigned &JTSize) const { 227 return TTIImpl->getEstimatedNumberOfCaseClusters(SI, JTSize); 228 } 229 230 int TargetTransformInfo::getUserCost(const User *U, 231 ArrayRef<const Value *> Operands) const { 232 int Cost = TTIImpl->getUserCost(U, Operands); 233 assert(Cost >= 0 && "TTI should not produce negative costs!"); 234 return Cost; 235 } 236 237 bool TargetTransformInfo::hasBranchDivergence() const { 238 return TTIImpl->hasBranchDivergence(); 239 } 240 241 bool TargetTransformInfo::isSourceOfDivergence(const Value *V) const { 242 return TTIImpl->isSourceOfDivergence(V); 243 } 244 245 bool llvm::TargetTransformInfo::isAlwaysUniform(const Value *V) const { 246 return TTIImpl->isAlwaysUniform(V); 247 } 248 249 unsigned TargetTransformInfo::getFlatAddressSpace() const { 250 return TTIImpl->getFlatAddressSpace(); 251 } 252 253 bool TargetTransformInfo::collectFlatAddressOperands( 254 SmallVectorImpl<int> &OpIndexes, Intrinsic::ID IID) const { 255 return TTIImpl->collectFlatAddressOperands(OpIndexes, IID); 256 } 257 258 bool TargetTransformInfo::rewriteIntrinsicWithAddressSpace( 259 IntrinsicInst *II, Value *OldV, Value *NewV) const { 260 return TTIImpl->rewriteIntrinsicWithAddressSpace(II, OldV, NewV); 261 } 262 263 bool TargetTransformInfo::isLoweredToCall(const Function *F) const { 264 return TTIImpl->isLoweredToCall(F); 265 } 266 267 bool TargetTransformInfo::isHardwareLoopProfitable( 268 Loop *L, ScalarEvolution &SE, AssumptionCache &AC, 269 TargetLibraryInfo *LibInfo, HardwareLoopInfo &HWLoopInfo) const { 270 return TTIImpl->isHardwareLoopProfitable(L, SE, AC, LibInfo, HWLoopInfo); 271 } 272 273 void TargetTransformInfo::getUnrollingPreferences( 274 Loop *L, ScalarEvolution &SE, UnrollingPreferences &UP) const { 275 return TTIImpl->getUnrollingPreferences(L, SE, UP); 276 } 277 278 bool TargetTransformInfo::isLegalAddImmediate(int64_t Imm) const { 279 return TTIImpl->isLegalAddImmediate(Imm); 280 } 281 282 bool TargetTransformInfo::isLegalICmpImmediate(int64_t Imm) const { 283 return TTIImpl->isLegalICmpImmediate(Imm); 284 } 285 286 bool TargetTransformInfo::isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV, 287 int64_t BaseOffset, 288 bool HasBaseReg, 289 int64_t Scale, 290 unsigned AddrSpace, 291 Instruction *I) const { 292 return TTIImpl->isLegalAddressingMode(Ty, BaseGV, BaseOffset, HasBaseReg, 293 Scale, AddrSpace, I); 294 } 295 296 bool TargetTransformInfo::isLSRCostLess(LSRCost &C1, LSRCost &C2) const { 297 return TTIImpl->isLSRCostLess(C1, C2); 298 } 299 300 bool TargetTransformInfo::canMacroFuseCmp() const { 301 return TTIImpl->canMacroFuseCmp(); 302 } 303 304 bool TargetTransformInfo::canSaveCmp(Loop *L, BranchInst **BI, 305 ScalarEvolution *SE, LoopInfo *LI, 306 DominatorTree *DT, AssumptionCache *AC, 307 TargetLibraryInfo *LibInfo) const { 308 return TTIImpl->canSaveCmp(L, BI, SE, LI, DT, AC, LibInfo); 309 } 310 311 bool TargetTransformInfo::shouldFavorPostInc() const { 312 return TTIImpl->shouldFavorPostInc(); 313 } 314 315 bool TargetTransformInfo::shouldFavorBackedgeIndex(const Loop *L) const { 316 return TTIImpl->shouldFavorBackedgeIndex(L); 317 } 318 319 bool TargetTransformInfo::isLegalMaskedStore(Type *DataType) const { 320 return TTIImpl->isLegalMaskedStore(DataType); 321 } 322 323 bool TargetTransformInfo::isLegalMaskedLoad(Type *DataType) const { 324 return TTIImpl->isLegalMaskedLoad(DataType); 325 } 326 327 bool TargetTransformInfo::isLegalNTStore(Type *DataType, 328 Align Alignment) const { 329 return TTIImpl->isLegalNTStore(DataType, Alignment); 330 } 331 332 bool TargetTransformInfo::isLegalNTLoad(Type *DataType, Align Alignment) const { 333 return TTIImpl->isLegalNTLoad(DataType, Alignment); 334 } 335 336 bool TargetTransformInfo::isLegalMaskedGather(Type *DataType) const { 337 return TTIImpl->isLegalMaskedGather(DataType); 338 } 339 340 bool TargetTransformInfo::isLegalMaskedScatter(Type *DataType) const { 341 return TTIImpl->isLegalMaskedScatter(DataType); 342 } 343 344 bool TargetTransformInfo::isLegalMaskedCompressStore(Type *DataType) const { 345 return TTIImpl->isLegalMaskedCompressStore(DataType); 346 } 347 348 bool TargetTransformInfo::isLegalMaskedExpandLoad(Type *DataType) const { 349 return TTIImpl->isLegalMaskedExpandLoad(DataType); 350 } 351 352 bool TargetTransformInfo::hasDivRemOp(Type *DataType, bool IsSigned) const { 353 return TTIImpl->hasDivRemOp(DataType, IsSigned); 354 } 355 356 bool TargetTransformInfo::hasVolatileVariant(Instruction *I, 357 unsigned AddrSpace) const { 358 return TTIImpl->hasVolatileVariant(I, AddrSpace); 359 } 360 361 bool TargetTransformInfo::prefersVectorizedAddressing() const { 362 return TTIImpl->prefersVectorizedAddressing(); 363 } 364 365 int TargetTransformInfo::getScalingFactorCost(Type *Ty, GlobalValue *BaseGV, 366 int64_t BaseOffset, 367 bool HasBaseReg, 368 int64_t Scale, 369 unsigned AddrSpace) const { 370 int Cost = TTIImpl->getScalingFactorCost(Ty, BaseGV, BaseOffset, HasBaseReg, 371 Scale, AddrSpace); 372 assert(Cost >= 0 && "TTI should not produce negative costs!"); 373 return Cost; 374 } 375 376 bool TargetTransformInfo::LSRWithInstrQueries() const { 377 return TTIImpl->LSRWithInstrQueries(); 378 } 379 380 bool TargetTransformInfo::isTruncateFree(Type *Ty1, Type *Ty2) const { 381 return TTIImpl->isTruncateFree(Ty1, Ty2); 382 } 383 384 bool TargetTransformInfo::isProfitableToHoist(Instruction *I) const { 385 return TTIImpl->isProfitableToHoist(I); 386 } 387 388 bool TargetTransformInfo::useAA() const { return TTIImpl->useAA(); } 389 390 bool TargetTransformInfo::isTypeLegal(Type *Ty) const { 391 return TTIImpl->isTypeLegal(Ty); 392 } 393 394 bool TargetTransformInfo::shouldBuildLookupTables() const { 395 return TTIImpl->shouldBuildLookupTables(); 396 } 397 bool TargetTransformInfo::shouldBuildLookupTablesForConstant(Constant *C) const { 398 return TTIImpl->shouldBuildLookupTablesForConstant(C); 399 } 400 401 bool TargetTransformInfo::useColdCCForColdCall(Function &F) const { 402 return TTIImpl->useColdCCForColdCall(F); 403 } 404 405 unsigned TargetTransformInfo:: 406 getScalarizationOverhead(Type *Ty, bool Insert, bool Extract) const { 407 return TTIImpl->getScalarizationOverhead(Ty, Insert, Extract); 408 } 409 410 unsigned TargetTransformInfo:: 411 getOperandsScalarizationOverhead(ArrayRef<const Value *> Args, 412 unsigned VF) const { 413 return TTIImpl->getOperandsScalarizationOverhead(Args, VF); 414 } 415 416 bool TargetTransformInfo::supportsEfficientVectorElementLoadStore() const { 417 return TTIImpl->supportsEfficientVectorElementLoadStore(); 418 } 419 420 bool TargetTransformInfo::enableAggressiveInterleaving(bool LoopHasReductions) const { 421 return TTIImpl->enableAggressiveInterleaving(LoopHasReductions); 422 } 423 424 TargetTransformInfo::MemCmpExpansionOptions 425 TargetTransformInfo::enableMemCmpExpansion(bool OptSize, bool IsZeroCmp) const { 426 return TTIImpl->enableMemCmpExpansion(OptSize, IsZeroCmp); 427 } 428 429 bool TargetTransformInfo::enableInterleavedAccessVectorization() const { 430 return TTIImpl->enableInterleavedAccessVectorization(); 431 } 432 433 bool TargetTransformInfo::enableMaskedInterleavedAccessVectorization() const { 434 return TTIImpl->enableMaskedInterleavedAccessVectorization(); 435 } 436 437 bool TargetTransformInfo::isFPVectorizationPotentiallyUnsafe() const { 438 return TTIImpl->isFPVectorizationPotentiallyUnsafe(); 439 } 440 441 bool TargetTransformInfo::allowsMisalignedMemoryAccesses(LLVMContext &Context, 442 unsigned BitWidth, 443 unsigned AddressSpace, 444 unsigned Alignment, 445 bool *Fast) const { 446 return TTIImpl->allowsMisalignedMemoryAccesses(Context, BitWidth, AddressSpace, 447 Alignment, Fast); 448 } 449 450 TargetTransformInfo::PopcntSupportKind 451 TargetTransformInfo::getPopcntSupport(unsigned IntTyWidthInBit) const { 452 return TTIImpl->getPopcntSupport(IntTyWidthInBit); 453 } 454 455 bool TargetTransformInfo::haveFastSqrt(Type *Ty) const { 456 return TTIImpl->haveFastSqrt(Ty); 457 } 458 459 bool TargetTransformInfo::isFCmpOrdCheaperThanFCmpZero(Type *Ty) const { 460 return TTIImpl->isFCmpOrdCheaperThanFCmpZero(Ty); 461 } 462 463 int TargetTransformInfo::getFPOpCost(Type *Ty) const { 464 int Cost = TTIImpl->getFPOpCost(Ty); 465 assert(Cost >= 0 && "TTI should not produce negative costs!"); 466 return Cost; 467 } 468 469 int TargetTransformInfo::getIntImmCodeSizeCost(unsigned Opcode, unsigned Idx, 470 const APInt &Imm, 471 Type *Ty) const { 472 int Cost = TTIImpl->getIntImmCodeSizeCost(Opcode, Idx, Imm, Ty); 473 assert(Cost >= 0 && "TTI should not produce negative costs!"); 474 return Cost; 475 } 476 477 int TargetTransformInfo::getIntImmCost(const APInt &Imm, Type *Ty) const { 478 int Cost = TTIImpl->getIntImmCost(Imm, Ty); 479 assert(Cost >= 0 && "TTI should not produce negative costs!"); 480 return Cost; 481 } 482 483 int TargetTransformInfo::getIntImmCost(unsigned Opcode, unsigned Idx, 484 const APInt &Imm, Type *Ty) const { 485 int Cost = TTIImpl->getIntImmCost(Opcode, Idx, Imm, Ty); 486 assert(Cost >= 0 && "TTI should not produce negative costs!"); 487 return Cost; 488 } 489 490 int TargetTransformInfo::getIntImmCost(Intrinsic::ID IID, unsigned Idx, 491 const APInt &Imm, Type *Ty) const { 492 int Cost = TTIImpl->getIntImmCost(IID, Idx, Imm, Ty); 493 assert(Cost >= 0 && "TTI should not produce negative costs!"); 494 return Cost; 495 } 496 497 unsigned TargetTransformInfo::getNumberOfRegisters(bool Vector) const { 498 return TTIImpl->getNumberOfRegisters(Vector); 499 } 500 501 unsigned TargetTransformInfo::getRegisterBitWidth(bool Vector) const { 502 return TTIImpl->getRegisterBitWidth(Vector); 503 } 504 505 unsigned TargetTransformInfo::getMinVectorRegisterBitWidth() const { 506 return TTIImpl->getMinVectorRegisterBitWidth(); 507 } 508 509 bool TargetTransformInfo::shouldMaximizeVectorBandwidth(bool OptSize) const { 510 return TTIImpl->shouldMaximizeVectorBandwidth(OptSize); 511 } 512 513 unsigned TargetTransformInfo::getMinimumVF(unsigned ElemWidth) const { 514 return TTIImpl->getMinimumVF(ElemWidth); 515 } 516 517 bool TargetTransformInfo::shouldConsiderAddressTypePromotion( 518 const Instruction &I, bool &AllowPromotionWithoutCommonHeader) const { 519 return TTIImpl->shouldConsiderAddressTypePromotion( 520 I, AllowPromotionWithoutCommonHeader); 521 } 522 523 unsigned TargetTransformInfo::getCacheLineSize() const { 524 return TTIImpl->getCacheLineSize(); 525 } 526 527 llvm::Optional<unsigned> TargetTransformInfo::getCacheSize(CacheLevel Level) 528 const { 529 return TTIImpl->getCacheSize(Level); 530 } 531 532 llvm::Optional<unsigned> TargetTransformInfo::getCacheAssociativity( 533 CacheLevel Level) const { 534 return TTIImpl->getCacheAssociativity(Level); 535 } 536 537 unsigned TargetTransformInfo::getPrefetchDistance() const { 538 return TTIImpl->getPrefetchDistance(); 539 } 540 541 unsigned TargetTransformInfo::getMinPrefetchStride() const { 542 return TTIImpl->getMinPrefetchStride(); 543 } 544 545 unsigned TargetTransformInfo::getMaxPrefetchIterationsAhead() const { 546 return TTIImpl->getMaxPrefetchIterationsAhead(); 547 } 548 549 unsigned TargetTransformInfo::getMaxInterleaveFactor(unsigned VF) const { 550 return TTIImpl->getMaxInterleaveFactor(VF); 551 } 552 553 TargetTransformInfo::OperandValueKind 554 TargetTransformInfo::getOperandInfo(Value *V, OperandValueProperties &OpProps) { 555 OperandValueKind OpInfo = OK_AnyValue; 556 OpProps = OP_None; 557 558 if (auto *CI = dyn_cast<ConstantInt>(V)) { 559 if (CI->getValue().isPowerOf2()) 560 OpProps = OP_PowerOf2; 561 return OK_UniformConstantValue; 562 } 563 564 // A broadcast shuffle creates a uniform value. 565 // TODO: Add support for non-zero index broadcasts. 566 // TODO: Add support for different source vector width. 567 if (auto *ShuffleInst = dyn_cast<ShuffleVectorInst>(V)) 568 if (ShuffleInst->isZeroEltSplat()) 569 OpInfo = OK_UniformValue; 570 571 const Value *Splat = getSplatValue(V); 572 573 // Check for a splat of a constant or for a non uniform vector of constants 574 // and check if the constant(s) are all powers of two. 575 if (isa<ConstantVector>(V) || isa<ConstantDataVector>(V)) { 576 OpInfo = OK_NonUniformConstantValue; 577 if (Splat) { 578 OpInfo = OK_UniformConstantValue; 579 if (auto *CI = dyn_cast<ConstantInt>(Splat)) 580 if (CI->getValue().isPowerOf2()) 581 OpProps = OP_PowerOf2; 582 } else if (auto *CDS = dyn_cast<ConstantDataSequential>(V)) { 583 OpProps = OP_PowerOf2; 584 for (unsigned I = 0, E = CDS->getNumElements(); I != E; ++I) { 585 if (auto *CI = dyn_cast<ConstantInt>(CDS->getElementAsConstant(I))) 586 if (CI->getValue().isPowerOf2()) 587 continue; 588 OpProps = OP_None; 589 break; 590 } 591 } 592 } 593 594 // Check for a splat of a uniform value. This is not loop aware, so return 595 // true only for the obviously uniform cases (argument, globalvalue) 596 if (Splat && (isa<Argument>(Splat) || isa<GlobalValue>(Splat))) 597 OpInfo = OK_UniformValue; 598 599 return OpInfo; 600 } 601 602 int TargetTransformInfo::getArithmeticInstrCost( 603 unsigned Opcode, Type *Ty, OperandValueKind Opd1Info, 604 OperandValueKind Opd2Info, OperandValueProperties Opd1PropInfo, 605 OperandValueProperties Opd2PropInfo, 606 ArrayRef<const Value *> Args) const { 607 int Cost = TTIImpl->getArithmeticInstrCost(Opcode, Ty, Opd1Info, Opd2Info, 608 Opd1PropInfo, Opd2PropInfo, Args); 609 assert(Cost >= 0 && "TTI should not produce negative costs!"); 610 return Cost; 611 } 612 613 int TargetTransformInfo::getShuffleCost(ShuffleKind Kind, Type *Ty, int Index, 614 Type *SubTp) const { 615 int Cost = TTIImpl->getShuffleCost(Kind, Ty, Index, SubTp); 616 assert(Cost >= 0 && "TTI should not produce negative costs!"); 617 return Cost; 618 } 619 620 int TargetTransformInfo::getCastInstrCost(unsigned Opcode, Type *Dst, 621 Type *Src, const Instruction *I) const { 622 assert ((I == nullptr || I->getOpcode() == Opcode) && 623 "Opcode should reflect passed instruction."); 624 int Cost = TTIImpl->getCastInstrCost(Opcode, Dst, Src, I); 625 assert(Cost >= 0 && "TTI should not produce negative costs!"); 626 return Cost; 627 } 628 629 int TargetTransformInfo::getExtractWithExtendCost(unsigned Opcode, Type *Dst, 630 VectorType *VecTy, 631 unsigned Index) const { 632 int Cost = TTIImpl->getExtractWithExtendCost(Opcode, Dst, VecTy, Index); 633 assert(Cost >= 0 && "TTI should not produce negative costs!"); 634 return Cost; 635 } 636 637 int TargetTransformInfo::getCFInstrCost(unsigned Opcode) const { 638 int Cost = TTIImpl->getCFInstrCost(Opcode); 639 assert(Cost >= 0 && "TTI should not produce negative costs!"); 640 return Cost; 641 } 642 643 int TargetTransformInfo::getCmpSelInstrCost(unsigned Opcode, Type *ValTy, 644 Type *CondTy, const Instruction *I) const { 645 assert ((I == nullptr || I->getOpcode() == Opcode) && 646 "Opcode should reflect passed instruction."); 647 int Cost = TTIImpl->getCmpSelInstrCost(Opcode, ValTy, CondTy, I); 648 assert(Cost >= 0 && "TTI should not produce negative costs!"); 649 return Cost; 650 } 651 652 int TargetTransformInfo::getVectorInstrCost(unsigned Opcode, Type *Val, 653 unsigned Index) const { 654 int Cost = TTIImpl->getVectorInstrCost(Opcode, Val, Index); 655 assert(Cost >= 0 && "TTI should not produce negative costs!"); 656 return Cost; 657 } 658 659 int TargetTransformInfo::getMemoryOpCost(unsigned Opcode, Type *Src, 660 unsigned Alignment, 661 unsigned AddressSpace, 662 const Instruction *I) const { 663 assert ((I == nullptr || I->getOpcode() == Opcode) && 664 "Opcode should reflect passed instruction."); 665 int Cost = TTIImpl->getMemoryOpCost(Opcode, Src, Alignment, AddressSpace, I); 666 assert(Cost >= 0 && "TTI should not produce negative costs!"); 667 return Cost; 668 } 669 670 int TargetTransformInfo::getMaskedMemoryOpCost(unsigned Opcode, Type *Src, 671 unsigned Alignment, 672 unsigned AddressSpace) const { 673 int Cost = 674 TTIImpl->getMaskedMemoryOpCost(Opcode, Src, Alignment, AddressSpace); 675 assert(Cost >= 0 && "TTI should not produce negative costs!"); 676 return Cost; 677 } 678 679 int TargetTransformInfo::getGatherScatterOpCost(unsigned Opcode, Type *DataTy, 680 Value *Ptr, bool VariableMask, 681 unsigned Alignment) const { 682 int Cost = TTIImpl->getGatherScatterOpCost(Opcode, DataTy, Ptr, VariableMask, 683 Alignment); 684 assert(Cost >= 0 && "TTI should not produce negative costs!"); 685 return Cost; 686 } 687 688 int TargetTransformInfo::getInterleavedMemoryOpCost( 689 unsigned Opcode, Type *VecTy, unsigned Factor, ArrayRef<unsigned> Indices, 690 unsigned Alignment, unsigned AddressSpace, bool UseMaskForCond, 691 bool UseMaskForGaps) const { 692 int Cost = TTIImpl->getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices, 693 Alignment, AddressSpace, 694 UseMaskForCond, 695 UseMaskForGaps); 696 assert(Cost >= 0 && "TTI should not produce negative costs!"); 697 return Cost; 698 } 699 700 int TargetTransformInfo::getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy, 701 ArrayRef<Type *> Tys, FastMathFlags FMF, 702 unsigned ScalarizationCostPassed) const { 703 int Cost = TTIImpl->getIntrinsicInstrCost(ID, RetTy, Tys, FMF, 704 ScalarizationCostPassed); 705 assert(Cost >= 0 && "TTI should not produce negative costs!"); 706 return Cost; 707 } 708 709 int TargetTransformInfo::getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy, 710 ArrayRef<Value *> Args, FastMathFlags FMF, unsigned VF) const { 711 int Cost = TTIImpl->getIntrinsicInstrCost(ID, RetTy, Args, FMF, VF); 712 assert(Cost >= 0 && "TTI should not produce negative costs!"); 713 return Cost; 714 } 715 716 int TargetTransformInfo::getCallInstrCost(Function *F, Type *RetTy, 717 ArrayRef<Type *> Tys) const { 718 int Cost = TTIImpl->getCallInstrCost(F, RetTy, Tys); 719 assert(Cost >= 0 && "TTI should not produce negative costs!"); 720 return Cost; 721 } 722 723 unsigned TargetTransformInfo::getNumberOfParts(Type *Tp) const { 724 return TTIImpl->getNumberOfParts(Tp); 725 } 726 727 int TargetTransformInfo::getAddressComputationCost(Type *Tp, 728 ScalarEvolution *SE, 729 const SCEV *Ptr) const { 730 int Cost = TTIImpl->getAddressComputationCost(Tp, SE, Ptr); 731 assert(Cost >= 0 && "TTI should not produce negative costs!"); 732 return Cost; 733 } 734 735 int TargetTransformInfo::getMemcpyCost(const Instruction *I) const { 736 int Cost = TTIImpl->getMemcpyCost(I); 737 assert(Cost >= 0 && "TTI should not produce negative costs!"); 738 return Cost; 739 } 740 741 int TargetTransformInfo::getArithmeticReductionCost(unsigned Opcode, Type *Ty, 742 bool IsPairwiseForm) const { 743 int Cost = TTIImpl->getArithmeticReductionCost(Opcode, Ty, IsPairwiseForm); 744 assert(Cost >= 0 && "TTI should not produce negative costs!"); 745 return Cost; 746 } 747 748 int TargetTransformInfo::getMinMaxReductionCost(Type *Ty, Type *CondTy, 749 bool IsPairwiseForm, 750 bool IsUnsigned) const { 751 int Cost = 752 TTIImpl->getMinMaxReductionCost(Ty, CondTy, IsPairwiseForm, IsUnsigned); 753 assert(Cost >= 0 && "TTI should not produce negative costs!"); 754 return Cost; 755 } 756 757 unsigned 758 TargetTransformInfo::getCostOfKeepingLiveOverCall(ArrayRef<Type *> Tys) const { 759 return TTIImpl->getCostOfKeepingLiveOverCall(Tys); 760 } 761 762 bool TargetTransformInfo::getTgtMemIntrinsic(IntrinsicInst *Inst, 763 MemIntrinsicInfo &Info) const { 764 return TTIImpl->getTgtMemIntrinsic(Inst, Info); 765 } 766 767 unsigned TargetTransformInfo::getAtomicMemIntrinsicMaxElementSize() const { 768 return TTIImpl->getAtomicMemIntrinsicMaxElementSize(); 769 } 770 771 Value *TargetTransformInfo::getOrCreateResultFromMemIntrinsic( 772 IntrinsicInst *Inst, Type *ExpectedType) const { 773 return TTIImpl->getOrCreateResultFromMemIntrinsic(Inst, ExpectedType); 774 } 775 776 Type *TargetTransformInfo::getMemcpyLoopLoweringType(LLVMContext &Context, 777 Value *Length, 778 unsigned SrcAlign, 779 unsigned DestAlign) const { 780 return TTIImpl->getMemcpyLoopLoweringType(Context, Length, SrcAlign, 781 DestAlign); 782 } 783 784 void TargetTransformInfo::getMemcpyLoopResidualLoweringType( 785 SmallVectorImpl<Type *> &OpsOut, LLVMContext &Context, 786 unsigned RemainingBytes, unsigned SrcAlign, unsigned DestAlign) const { 787 TTIImpl->getMemcpyLoopResidualLoweringType(OpsOut, Context, RemainingBytes, 788 SrcAlign, DestAlign); 789 } 790 791 bool TargetTransformInfo::areInlineCompatible(const Function *Caller, 792 const Function *Callee) const { 793 return TTIImpl->areInlineCompatible(Caller, Callee); 794 } 795 796 bool TargetTransformInfo::areFunctionArgsABICompatible( 797 const Function *Caller, const Function *Callee, 798 SmallPtrSetImpl<Argument *> &Args) const { 799 return TTIImpl->areFunctionArgsABICompatible(Caller, Callee, Args); 800 } 801 802 bool TargetTransformInfo::isIndexedLoadLegal(MemIndexedMode Mode, 803 Type *Ty) const { 804 return TTIImpl->isIndexedLoadLegal(Mode, Ty); 805 } 806 807 bool TargetTransformInfo::isIndexedStoreLegal(MemIndexedMode Mode, 808 Type *Ty) const { 809 return TTIImpl->isIndexedStoreLegal(Mode, Ty); 810 } 811 812 unsigned TargetTransformInfo::getLoadStoreVecRegBitWidth(unsigned AS) const { 813 return TTIImpl->getLoadStoreVecRegBitWidth(AS); 814 } 815 816 bool TargetTransformInfo::isLegalToVectorizeLoad(LoadInst *LI) const { 817 return TTIImpl->isLegalToVectorizeLoad(LI); 818 } 819 820 bool TargetTransformInfo::isLegalToVectorizeStore(StoreInst *SI) const { 821 return TTIImpl->isLegalToVectorizeStore(SI); 822 } 823 824 bool TargetTransformInfo::isLegalToVectorizeLoadChain( 825 unsigned ChainSizeInBytes, unsigned Alignment, unsigned AddrSpace) const { 826 return TTIImpl->isLegalToVectorizeLoadChain(ChainSizeInBytes, Alignment, 827 AddrSpace); 828 } 829 830 bool TargetTransformInfo::isLegalToVectorizeStoreChain( 831 unsigned ChainSizeInBytes, unsigned Alignment, unsigned AddrSpace) const { 832 return TTIImpl->isLegalToVectorizeStoreChain(ChainSizeInBytes, Alignment, 833 AddrSpace); 834 } 835 836 unsigned TargetTransformInfo::getLoadVectorFactor(unsigned VF, 837 unsigned LoadSize, 838 unsigned ChainSizeInBytes, 839 VectorType *VecTy) const { 840 return TTIImpl->getLoadVectorFactor(VF, LoadSize, ChainSizeInBytes, VecTy); 841 } 842 843 unsigned TargetTransformInfo::getStoreVectorFactor(unsigned VF, 844 unsigned StoreSize, 845 unsigned ChainSizeInBytes, 846 VectorType *VecTy) const { 847 return TTIImpl->getStoreVectorFactor(VF, StoreSize, ChainSizeInBytes, VecTy); 848 } 849 850 bool TargetTransformInfo::useReductionIntrinsic(unsigned Opcode, 851 Type *Ty, ReductionFlags Flags) const { 852 return TTIImpl->useReductionIntrinsic(Opcode, Ty, Flags); 853 } 854 855 bool TargetTransformInfo::shouldExpandReduction(const IntrinsicInst *II) const { 856 return TTIImpl->shouldExpandReduction(II); 857 } 858 859 unsigned TargetTransformInfo::getGISelRematGlobalCost() const { 860 return TTIImpl->getGISelRematGlobalCost(); 861 } 862 863 int TargetTransformInfo::getInstructionLatency(const Instruction *I) const { 864 return TTIImpl->getInstructionLatency(I); 865 } 866 867 static bool matchPairwiseShuffleMask(ShuffleVectorInst *SI, bool IsLeft, 868 unsigned Level) { 869 // We don't need a shuffle if we just want to have element 0 in position 0 of 870 // the vector. 871 if (!SI && Level == 0 && IsLeft) 872 return true; 873 else if (!SI) 874 return false; 875 876 SmallVector<int, 32> Mask(SI->getType()->getVectorNumElements(), -1); 877 878 // Build a mask of 0, 2, ... (left) or 1, 3, ... (right) depending on whether 879 // we look at the left or right side. 880 for (unsigned i = 0, e = (1 << Level), val = !IsLeft; i != e; ++i, val += 2) 881 Mask[i] = val; 882 883 SmallVector<int, 16> ActualMask = SI->getShuffleMask(); 884 return Mask == ActualMask; 885 } 886 887 namespace { 888 /// Kind of the reduction data. 889 enum ReductionKind { 890 RK_None, /// Not a reduction. 891 RK_Arithmetic, /// Binary reduction data. 892 RK_MinMax, /// Min/max reduction data. 893 RK_UnsignedMinMax, /// Unsigned min/max reduction data. 894 }; 895 /// Contains opcode + LHS/RHS parts of the reduction operations. 896 struct ReductionData { 897 ReductionData() = delete; 898 ReductionData(ReductionKind Kind, unsigned Opcode, Value *LHS, Value *RHS) 899 : Opcode(Opcode), LHS(LHS), RHS(RHS), Kind(Kind) { 900 assert(Kind != RK_None && "expected binary or min/max reduction only."); 901 } 902 unsigned Opcode = 0; 903 Value *LHS = nullptr; 904 Value *RHS = nullptr; 905 ReductionKind Kind = RK_None; 906 bool hasSameData(ReductionData &RD) const { 907 return Kind == RD.Kind && Opcode == RD.Opcode; 908 } 909 }; 910 } // namespace 911 912 static Optional<ReductionData> getReductionData(Instruction *I) { 913 Value *L, *R; 914 if (m_BinOp(m_Value(L), m_Value(R)).match(I)) 915 return ReductionData(RK_Arithmetic, I->getOpcode(), L, R); 916 if (auto *SI = dyn_cast<SelectInst>(I)) { 917 if (m_SMin(m_Value(L), m_Value(R)).match(SI) || 918 m_SMax(m_Value(L), m_Value(R)).match(SI) || 919 m_OrdFMin(m_Value(L), m_Value(R)).match(SI) || 920 m_OrdFMax(m_Value(L), m_Value(R)).match(SI) || 921 m_UnordFMin(m_Value(L), m_Value(R)).match(SI) || 922 m_UnordFMax(m_Value(L), m_Value(R)).match(SI)) { 923 auto *CI = cast<CmpInst>(SI->getCondition()); 924 return ReductionData(RK_MinMax, CI->getOpcode(), L, R); 925 } 926 if (m_UMin(m_Value(L), m_Value(R)).match(SI) || 927 m_UMax(m_Value(L), m_Value(R)).match(SI)) { 928 auto *CI = cast<CmpInst>(SI->getCondition()); 929 return ReductionData(RK_UnsignedMinMax, CI->getOpcode(), L, R); 930 } 931 } 932 return llvm::None; 933 } 934 935 static ReductionKind matchPairwiseReductionAtLevel(Instruction *I, 936 unsigned Level, 937 unsigned NumLevels) { 938 // Match one level of pairwise operations. 939 // %rdx.shuf.0.0 = shufflevector <4 x float> %rdx, <4 x float> undef, 940 // <4 x i32> <i32 0, i32 2 , i32 undef, i32 undef> 941 // %rdx.shuf.0.1 = shufflevector <4 x float> %rdx, <4 x float> undef, 942 // <4 x i32> <i32 1, i32 3, i32 undef, i32 undef> 943 // %bin.rdx.0 = fadd <4 x float> %rdx.shuf.0.0, %rdx.shuf.0.1 944 if (!I) 945 return RK_None; 946 947 assert(I->getType()->isVectorTy() && "Expecting a vector type"); 948 949 Optional<ReductionData> RD = getReductionData(I); 950 if (!RD) 951 return RK_None; 952 953 ShuffleVectorInst *LS = dyn_cast<ShuffleVectorInst>(RD->LHS); 954 if (!LS && Level) 955 return RK_None; 956 ShuffleVectorInst *RS = dyn_cast<ShuffleVectorInst>(RD->RHS); 957 if (!RS && Level) 958 return RK_None; 959 960 // On level 0 we can omit one shufflevector instruction. 961 if (!Level && !RS && !LS) 962 return RK_None; 963 964 // Shuffle inputs must match. 965 Value *NextLevelOpL = LS ? LS->getOperand(0) : nullptr; 966 Value *NextLevelOpR = RS ? RS->getOperand(0) : nullptr; 967 Value *NextLevelOp = nullptr; 968 if (NextLevelOpR && NextLevelOpL) { 969 // If we have two shuffles their operands must match. 970 if (NextLevelOpL != NextLevelOpR) 971 return RK_None; 972 973 NextLevelOp = NextLevelOpL; 974 } else if (Level == 0 && (NextLevelOpR || NextLevelOpL)) { 975 // On the first level we can omit the shufflevector <0, undef,...>. So the 976 // input to the other shufflevector <1, undef> must match with one of the 977 // inputs to the current binary operation. 978 // Example: 979 // %NextLevelOpL = shufflevector %R, <1, undef ...> 980 // %BinOp = fadd %NextLevelOpL, %R 981 if (NextLevelOpL && NextLevelOpL != RD->RHS) 982 return RK_None; 983 else if (NextLevelOpR && NextLevelOpR != RD->LHS) 984 return RK_None; 985 986 NextLevelOp = NextLevelOpL ? RD->RHS : RD->LHS; 987 } else 988 return RK_None; 989 990 // Check that the next levels binary operation exists and matches with the 991 // current one. 992 if (Level + 1 != NumLevels) { 993 Optional<ReductionData> NextLevelRD = 994 getReductionData(cast<Instruction>(NextLevelOp)); 995 if (!NextLevelRD || !RD->hasSameData(*NextLevelRD)) 996 return RK_None; 997 } 998 999 // Shuffle mask for pairwise operation must match. 1000 if (matchPairwiseShuffleMask(LS, /*IsLeft=*/true, Level)) { 1001 if (!matchPairwiseShuffleMask(RS, /*IsLeft=*/false, Level)) 1002 return RK_None; 1003 } else if (matchPairwiseShuffleMask(RS, /*IsLeft=*/true, Level)) { 1004 if (!matchPairwiseShuffleMask(LS, /*IsLeft=*/false, Level)) 1005 return RK_None; 1006 } else { 1007 return RK_None; 1008 } 1009 1010 if (++Level == NumLevels) 1011 return RD->Kind; 1012 1013 // Match next level. 1014 return matchPairwiseReductionAtLevel(cast<Instruction>(NextLevelOp), Level, 1015 NumLevels); 1016 } 1017 1018 static ReductionKind matchPairwiseReduction(const ExtractElementInst *ReduxRoot, 1019 unsigned &Opcode, Type *&Ty) { 1020 if (!EnableReduxCost) 1021 return RK_None; 1022 1023 // Need to extract the first element. 1024 ConstantInt *CI = dyn_cast<ConstantInt>(ReduxRoot->getOperand(1)); 1025 unsigned Idx = ~0u; 1026 if (CI) 1027 Idx = CI->getZExtValue(); 1028 if (Idx != 0) 1029 return RK_None; 1030 1031 auto *RdxStart = dyn_cast<Instruction>(ReduxRoot->getOperand(0)); 1032 if (!RdxStart) 1033 return RK_None; 1034 Optional<ReductionData> RD = getReductionData(RdxStart); 1035 if (!RD) 1036 return RK_None; 1037 1038 Type *VecTy = RdxStart->getType(); 1039 unsigned NumVecElems = VecTy->getVectorNumElements(); 1040 if (!isPowerOf2_32(NumVecElems)) 1041 return RK_None; 1042 1043 // We look for a sequence of shuffle,shuffle,add triples like the following 1044 // that builds a pairwise reduction tree. 1045 // 1046 // (X0, X1, X2, X3) 1047 // (X0 + X1, X2 + X3, undef, undef) 1048 // ((X0 + X1) + (X2 + X3), undef, undef, undef) 1049 // 1050 // %rdx.shuf.0.0 = shufflevector <4 x float> %rdx, <4 x float> undef, 1051 // <4 x i32> <i32 0, i32 2 , i32 undef, i32 undef> 1052 // %rdx.shuf.0.1 = shufflevector <4 x float> %rdx, <4 x float> undef, 1053 // <4 x i32> <i32 1, i32 3, i32 undef, i32 undef> 1054 // %bin.rdx.0 = fadd <4 x float> %rdx.shuf.0.0, %rdx.shuf.0.1 1055 // %rdx.shuf.1.0 = shufflevector <4 x float> %bin.rdx.0, <4 x float> undef, 1056 // <4 x i32> <i32 0, i32 undef, i32 undef, i32 undef> 1057 // %rdx.shuf.1.1 = shufflevector <4 x float> %bin.rdx.0, <4 x float> undef, 1058 // <4 x i32> <i32 1, i32 undef, i32 undef, i32 undef> 1059 // %bin.rdx8 = fadd <4 x float> %rdx.shuf.1.0, %rdx.shuf.1.1 1060 // %r = extractelement <4 x float> %bin.rdx8, i32 0 1061 if (matchPairwiseReductionAtLevel(RdxStart, 0, Log2_32(NumVecElems)) == 1062 RK_None) 1063 return RK_None; 1064 1065 Opcode = RD->Opcode; 1066 Ty = VecTy; 1067 1068 return RD->Kind; 1069 } 1070 1071 static std::pair<Value *, ShuffleVectorInst *> 1072 getShuffleAndOtherOprd(Value *L, Value *R) { 1073 ShuffleVectorInst *S = nullptr; 1074 1075 if ((S = dyn_cast<ShuffleVectorInst>(L))) 1076 return std::make_pair(R, S); 1077 1078 S = dyn_cast<ShuffleVectorInst>(R); 1079 return std::make_pair(L, S); 1080 } 1081 1082 static ReductionKind 1083 matchVectorSplittingReduction(const ExtractElementInst *ReduxRoot, 1084 unsigned &Opcode, Type *&Ty) { 1085 if (!EnableReduxCost) 1086 return RK_None; 1087 1088 // Need to extract the first element. 1089 ConstantInt *CI = dyn_cast<ConstantInt>(ReduxRoot->getOperand(1)); 1090 unsigned Idx = ~0u; 1091 if (CI) 1092 Idx = CI->getZExtValue(); 1093 if (Idx != 0) 1094 return RK_None; 1095 1096 auto *RdxStart = dyn_cast<Instruction>(ReduxRoot->getOperand(0)); 1097 if (!RdxStart) 1098 return RK_None; 1099 Optional<ReductionData> RD = getReductionData(RdxStart); 1100 if (!RD) 1101 return RK_None; 1102 1103 Type *VecTy = ReduxRoot->getOperand(0)->getType(); 1104 unsigned NumVecElems = VecTy->getVectorNumElements(); 1105 if (!isPowerOf2_32(NumVecElems)) 1106 return RK_None; 1107 1108 // We look for a sequence of shuffles and adds like the following matching one 1109 // fadd, shuffle vector pair at a time. 1110 // 1111 // %rdx.shuf = shufflevector <4 x float> %rdx, <4 x float> undef, 1112 // <4 x i32> <i32 2, i32 3, i32 undef, i32 undef> 1113 // %bin.rdx = fadd <4 x float> %rdx, %rdx.shuf 1114 // %rdx.shuf7 = shufflevector <4 x float> %bin.rdx, <4 x float> undef, 1115 // <4 x i32> <i32 1, i32 undef, i32 undef, i32 undef> 1116 // %bin.rdx8 = fadd <4 x float> %bin.rdx, %rdx.shuf7 1117 // %r = extractelement <4 x float> %bin.rdx8, i32 0 1118 1119 unsigned MaskStart = 1; 1120 Instruction *RdxOp = RdxStart; 1121 SmallVector<int, 32> ShuffleMask(NumVecElems, 0); 1122 unsigned NumVecElemsRemain = NumVecElems; 1123 while (NumVecElemsRemain - 1) { 1124 // Check for the right reduction operation. 1125 if (!RdxOp) 1126 return RK_None; 1127 Optional<ReductionData> RDLevel = getReductionData(RdxOp); 1128 if (!RDLevel || !RDLevel->hasSameData(*RD)) 1129 return RK_None; 1130 1131 Value *NextRdxOp; 1132 ShuffleVectorInst *Shuffle; 1133 std::tie(NextRdxOp, Shuffle) = 1134 getShuffleAndOtherOprd(RDLevel->LHS, RDLevel->RHS); 1135 1136 // Check the current reduction operation and the shuffle use the same value. 1137 if (Shuffle == nullptr) 1138 return RK_None; 1139 if (Shuffle->getOperand(0) != NextRdxOp) 1140 return RK_None; 1141 1142 // Check that shuffle masks matches. 1143 for (unsigned j = 0; j != MaskStart; ++j) 1144 ShuffleMask[j] = MaskStart + j; 1145 // Fill the rest of the mask with -1 for undef. 1146 std::fill(&ShuffleMask[MaskStart], ShuffleMask.end(), -1); 1147 1148 SmallVector<int, 16> Mask = Shuffle->getShuffleMask(); 1149 if (ShuffleMask != Mask) 1150 return RK_None; 1151 1152 RdxOp = dyn_cast<Instruction>(NextRdxOp); 1153 NumVecElemsRemain /= 2; 1154 MaskStart *= 2; 1155 } 1156 1157 Opcode = RD->Opcode; 1158 Ty = VecTy; 1159 return RD->Kind; 1160 } 1161 1162 int TargetTransformInfo::getInstructionThroughput(const Instruction *I) const { 1163 switch (I->getOpcode()) { 1164 case Instruction::GetElementPtr: 1165 return getUserCost(I); 1166 1167 case Instruction::Ret: 1168 case Instruction::PHI: 1169 case Instruction::Br: { 1170 return getCFInstrCost(I->getOpcode()); 1171 } 1172 case Instruction::Add: 1173 case Instruction::FAdd: 1174 case Instruction::Sub: 1175 case Instruction::FSub: 1176 case Instruction::Mul: 1177 case Instruction::FMul: 1178 case Instruction::UDiv: 1179 case Instruction::SDiv: 1180 case Instruction::FDiv: 1181 case Instruction::URem: 1182 case Instruction::SRem: 1183 case Instruction::FRem: 1184 case Instruction::Shl: 1185 case Instruction::LShr: 1186 case Instruction::AShr: 1187 case Instruction::And: 1188 case Instruction::Or: 1189 case Instruction::Xor: { 1190 TargetTransformInfo::OperandValueKind Op1VK, Op2VK; 1191 TargetTransformInfo::OperandValueProperties Op1VP, Op2VP; 1192 Op1VK = getOperandInfo(I->getOperand(0), Op1VP); 1193 Op2VK = getOperandInfo(I->getOperand(1), Op2VP); 1194 SmallVector<const Value *, 2> Operands(I->operand_values()); 1195 return getArithmeticInstrCost(I->getOpcode(), I->getType(), Op1VK, Op2VK, 1196 Op1VP, Op2VP, Operands); 1197 } 1198 case Instruction::FNeg: { 1199 TargetTransformInfo::OperandValueKind Op1VK, Op2VK; 1200 TargetTransformInfo::OperandValueProperties Op1VP, Op2VP; 1201 Op1VK = getOperandInfo(I->getOperand(0), Op1VP); 1202 Op2VK = OK_AnyValue; 1203 Op2VP = OP_None; 1204 SmallVector<const Value *, 2> Operands(I->operand_values()); 1205 return getArithmeticInstrCost(I->getOpcode(), I->getType(), Op1VK, Op2VK, 1206 Op1VP, Op2VP, Operands); 1207 } 1208 case Instruction::Select: { 1209 const SelectInst *SI = cast<SelectInst>(I); 1210 Type *CondTy = SI->getCondition()->getType(); 1211 return getCmpSelInstrCost(I->getOpcode(), I->getType(), CondTy, I); 1212 } 1213 case Instruction::ICmp: 1214 case Instruction::FCmp: { 1215 Type *ValTy = I->getOperand(0)->getType(); 1216 return getCmpSelInstrCost(I->getOpcode(), ValTy, I->getType(), I); 1217 } 1218 case Instruction::Store: { 1219 const StoreInst *SI = cast<StoreInst>(I); 1220 Type *ValTy = SI->getValueOperand()->getType(); 1221 return getMemoryOpCost(I->getOpcode(), ValTy, 1222 SI->getAlignment(), 1223 SI->getPointerAddressSpace(), I); 1224 } 1225 case Instruction::Load: { 1226 const LoadInst *LI = cast<LoadInst>(I); 1227 return getMemoryOpCost(I->getOpcode(), I->getType(), 1228 LI->getAlignment(), 1229 LI->getPointerAddressSpace(), I); 1230 } 1231 case Instruction::ZExt: 1232 case Instruction::SExt: 1233 case Instruction::FPToUI: 1234 case Instruction::FPToSI: 1235 case Instruction::FPExt: 1236 case Instruction::PtrToInt: 1237 case Instruction::IntToPtr: 1238 case Instruction::SIToFP: 1239 case Instruction::UIToFP: 1240 case Instruction::Trunc: 1241 case Instruction::FPTrunc: 1242 case Instruction::BitCast: 1243 case Instruction::AddrSpaceCast: { 1244 Type *SrcTy = I->getOperand(0)->getType(); 1245 return getCastInstrCost(I->getOpcode(), I->getType(), SrcTy, I); 1246 } 1247 case Instruction::ExtractElement: { 1248 const ExtractElementInst * EEI = cast<ExtractElementInst>(I); 1249 ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1)); 1250 unsigned Idx = -1; 1251 if (CI) 1252 Idx = CI->getZExtValue(); 1253 1254 // Try to match a reduction sequence (series of shufflevector and vector 1255 // adds followed by a extractelement). 1256 unsigned ReduxOpCode; 1257 Type *ReduxType; 1258 1259 switch (matchVectorSplittingReduction(EEI, ReduxOpCode, ReduxType)) { 1260 case RK_Arithmetic: 1261 return getArithmeticReductionCost(ReduxOpCode, ReduxType, 1262 /*IsPairwiseForm=*/false); 1263 case RK_MinMax: 1264 return getMinMaxReductionCost( 1265 ReduxType, CmpInst::makeCmpResultType(ReduxType), 1266 /*IsPairwiseForm=*/false, /*IsUnsigned=*/false); 1267 case RK_UnsignedMinMax: 1268 return getMinMaxReductionCost( 1269 ReduxType, CmpInst::makeCmpResultType(ReduxType), 1270 /*IsPairwiseForm=*/false, /*IsUnsigned=*/true); 1271 case RK_None: 1272 break; 1273 } 1274 1275 switch (matchPairwiseReduction(EEI, ReduxOpCode, ReduxType)) { 1276 case RK_Arithmetic: 1277 return getArithmeticReductionCost(ReduxOpCode, ReduxType, 1278 /*IsPairwiseForm=*/true); 1279 case RK_MinMax: 1280 return getMinMaxReductionCost( 1281 ReduxType, CmpInst::makeCmpResultType(ReduxType), 1282 /*IsPairwiseForm=*/true, /*IsUnsigned=*/false); 1283 case RK_UnsignedMinMax: 1284 return getMinMaxReductionCost( 1285 ReduxType, CmpInst::makeCmpResultType(ReduxType), 1286 /*IsPairwiseForm=*/true, /*IsUnsigned=*/true); 1287 case RK_None: 1288 break; 1289 } 1290 1291 return getVectorInstrCost(I->getOpcode(), 1292 EEI->getOperand(0)->getType(), Idx); 1293 } 1294 case Instruction::InsertElement: { 1295 const InsertElementInst * IE = cast<InsertElementInst>(I); 1296 ConstantInt *CI = dyn_cast<ConstantInt>(IE->getOperand(2)); 1297 unsigned Idx = -1; 1298 if (CI) 1299 Idx = CI->getZExtValue(); 1300 return getVectorInstrCost(I->getOpcode(), 1301 IE->getType(), Idx); 1302 } 1303 case Instruction::ExtractValue: 1304 return 0; // Model all ExtractValue nodes as free. 1305 case Instruction::ShuffleVector: { 1306 const ShuffleVectorInst *Shuffle = cast<ShuffleVectorInst>(I); 1307 Type *Ty = Shuffle->getType(); 1308 Type *SrcTy = Shuffle->getOperand(0)->getType(); 1309 1310 // TODO: Identify and add costs for insert subvector, etc. 1311 int SubIndex; 1312 if (Shuffle->isExtractSubvectorMask(SubIndex)) 1313 return TTIImpl->getShuffleCost(SK_ExtractSubvector, SrcTy, SubIndex, Ty); 1314 1315 if (Shuffle->changesLength()) 1316 return -1; 1317 1318 if (Shuffle->isIdentity()) 1319 return 0; 1320 1321 if (Shuffle->isReverse()) 1322 return TTIImpl->getShuffleCost(SK_Reverse, Ty, 0, nullptr); 1323 1324 if (Shuffle->isSelect()) 1325 return TTIImpl->getShuffleCost(SK_Select, Ty, 0, nullptr); 1326 1327 if (Shuffle->isTranspose()) 1328 return TTIImpl->getShuffleCost(SK_Transpose, Ty, 0, nullptr); 1329 1330 if (Shuffle->isZeroEltSplat()) 1331 return TTIImpl->getShuffleCost(SK_Broadcast, Ty, 0, nullptr); 1332 1333 if (Shuffle->isSingleSource()) 1334 return TTIImpl->getShuffleCost(SK_PermuteSingleSrc, Ty, 0, nullptr); 1335 1336 return TTIImpl->getShuffleCost(SK_PermuteTwoSrc, Ty, 0, nullptr); 1337 } 1338 case Instruction::Call: 1339 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) { 1340 SmallVector<Value *, 4> Args(II->arg_operands()); 1341 1342 FastMathFlags FMF; 1343 if (auto *FPMO = dyn_cast<FPMathOperator>(II)) 1344 FMF = FPMO->getFastMathFlags(); 1345 1346 return getIntrinsicInstrCost(II->getIntrinsicID(), II->getType(), 1347 Args, FMF); 1348 } 1349 return -1; 1350 default: 1351 // We don't have any information on this instruction. 1352 return -1; 1353 } 1354 } 1355 1356 TargetTransformInfo::Concept::~Concept() {} 1357 1358 TargetIRAnalysis::TargetIRAnalysis() : TTICallback(&getDefaultTTI) {} 1359 1360 TargetIRAnalysis::TargetIRAnalysis( 1361 std::function<Result(const Function &)> TTICallback) 1362 : TTICallback(std::move(TTICallback)) {} 1363 1364 TargetIRAnalysis::Result TargetIRAnalysis::run(const Function &F, 1365 FunctionAnalysisManager &) { 1366 return TTICallback(F); 1367 } 1368 1369 AnalysisKey TargetIRAnalysis::Key; 1370 1371 TargetIRAnalysis::Result TargetIRAnalysis::getDefaultTTI(const Function &F) { 1372 return Result(F.getParent()->getDataLayout()); 1373 } 1374 1375 // Register the basic pass. 1376 INITIALIZE_PASS(TargetTransformInfoWrapperPass, "tti", 1377 "Target Transform Information", false, true) 1378 char TargetTransformInfoWrapperPass::ID = 0; 1379 1380 void TargetTransformInfoWrapperPass::anchor() {} 1381 1382 TargetTransformInfoWrapperPass::TargetTransformInfoWrapperPass() 1383 : ImmutablePass(ID) { 1384 initializeTargetTransformInfoWrapperPassPass( 1385 *PassRegistry::getPassRegistry()); 1386 } 1387 1388 TargetTransformInfoWrapperPass::TargetTransformInfoWrapperPass( 1389 TargetIRAnalysis TIRA) 1390 : ImmutablePass(ID), TIRA(std::move(TIRA)) { 1391 initializeTargetTransformInfoWrapperPassPass( 1392 *PassRegistry::getPassRegistry()); 1393 } 1394 1395 TargetTransformInfo &TargetTransformInfoWrapperPass::getTTI(const Function &F) { 1396 FunctionAnalysisManager DummyFAM; 1397 TTI = TIRA.run(F, DummyFAM); 1398 return *TTI; 1399 } 1400 1401 ImmutablePass * 1402 llvm::createTargetTransformInfoWrapperPass(TargetIRAnalysis TIRA) { 1403 return new TargetTransformInfoWrapperPass(std::move(TIRA)); 1404 } 1405