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