xref: /llvm-project/llvm/lib/Analysis/TargetTransformInfo.cpp (revision 18f8106f310ee702046a11f360af47947c030d2e) 
1 //===- llvm/Analysis/TargetTransformInfo.cpp ------------------------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 
9 #include "llvm/Analysis/TargetTransformInfo.h"
10 #include "llvm/Analysis/CFG.h"
11 #include "llvm/Analysis/LoopIterator.h"
12 #include "llvm/Analysis/TargetLibraryInfo.h"
13 #include "llvm/Analysis/TargetTransformInfoImpl.h"
14 #include "llvm/IR/CFG.h"
15 #include "llvm/IR/Dominators.h"
16 #include "llvm/IR/Instruction.h"
17 #include "llvm/IR/Instructions.h"
18 #include "llvm/IR/IntrinsicInst.h"
19 #include "llvm/IR/Module.h"
20 #include "llvm/IR/Operator.h"
21 #include "llvm/InitializePasses.h"
22 #include "llvm/Support/CommandLine.h"
23 #include <optional>
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 static cl::opt<unsigned> CacheLineSize(
36     "cache-line-size", cl::init(0), cl::Hidden,
37     cl::desc("Use this to override the target cache line size when "
38              "specified by the user."));
39 
40 static cl::opt<unsigned> MinPageSize(
41     "min-page-size", cl::init(0), cl::Hidden,
42     cl::desc("Use this to override the target's minimum page size."));
43 
44 static cl::opt<unsigned> PredictableBranchThreshold(
45     "predictable-branch-threshold", cl::init(99), cl::Hidden,
46     cl::desc(
47         "Use this to override the target's predictable branch threshold (%)."));
48 
49 namespace {
50 /// No-op implementation of the TTI interface using the utility base
51 /// classes.
52 ///
53 /// This is used when no target specific information is available.
54 struct NoTTIImpl : TargetTransformInfoImplCRTPBase<NoTTIImpl> {
55   explicit NoTTIImpl(const DataLayout &DL)
56       : TargetTransformInfoImplCRTPBase<NoTTIImpl>(DL) {}
57 };
58 } // namespace
59 
60 bool HardwareLoopInfo::canAnalyze(LoopInfo &LI) {
61   // If the loop has irreducible control flow, it can not be converted to
62   // Hardware loop.
63   LoopBlocksRPO RPOT(L);
64   RPOT.perform(&LI);
65   if (containsIrreducibleCFG<const BasicBlock *>(RPOT, LI))
66     return false;
67   return true;
68 }
69 
70 IntrinsicCostAttributes::IntrinsicCostAttributes(
71     Intrinsic::ID Id, const CallBase &CI, InstructionCost ScalarizationCost,
72     bool TypeBasedOnly)
73     : II(dyn_cast<IntrinsicInst>(&CI)), RetTy(CI.getType()), IID(Id),
74       ScalarizationCost(ScalarizationCost) {
75 
76   if (const auto *FPMO = dyn_cast<FPMathOperator>(&CI))
77     FMF = FPMO->getFastMathFlags();
78 
79   if (!TypeBasedOnly)
80     Arguments.insert(Arguments.begin(), CI.arg_begin(), CI.arg_end());
81   FunctionType *FTy = CI.getCalledFunction()->getFunctionType();
82   ParamTys.insert(ParamTys.begin(), FTy->param_begin(), FTy->param_end());
83 }
84 
85 IntrinsicCostAttributes::IntrinsicCostAttributes(Intrinsic::ID Id, Type *RTy,
86                                                  ArrayRef<Type *> Tys,
87                                                  FastMathFlags Flags,
88                                                  const IntrinsicInst *I,
89                                                  InstructionCost ScalarCost)
90     : II(I), RetTy(RTy), IID(Id), FMF(Flags), ScalarizationCost(ScalarCost) {
91   ParamTys.insert(ParamTys.begin(), Tys.begin(), Tys.end());
92 }
93 
94 IntrinsicCostAttributes::IntrinsicCostAttributes(Intrinsic::ID Id, Type *Ty,
95                                                  ArrayRef<const Value *> Args)
96     : RetTy(Ty), IID(Id) {
97 
98   Arguments.insert(Arguments.begin(), Args.begin(), Args.end());
99   ParamTys.reserve(Arguments.size());
100   for (const Value *Argument : Arguments)
101     ParamTys.push_back(Argument->getType());
102 }
103 
104 IntrinsicCostAttributes::IntrinsicCostAttributes(Intrinsic::ID Id, Type *RTy,
105                                                  ArrayRef<const Value *> Args,
106                                                  ArrayRef<Type *> Tys,
107                                                  FastMathFlags Flags,
108                                                  const IntrinsicInst *I,
109                                                  InstructionCost ScalarCost)
110     : II(I), RetTy(RTy), IID(Id), FMF(Flags), ScalarizationCost(ScalarCost) {
111   ParamTys.insert(ParamTys.begin(), Tys.begin(), Tys.end());
112   Arguments.insert(Arguments.begin(), Args.begin(), Args.end());
113 }
114 
115 HardwareLoopInfo::HardwareLoopInfo(Loop *L) : L(L) {
116   // Match default options:
117   // - hardware-loop-counter-bitwidth = 32
118   // - hardware-loop-decrement = 1
119   CountType = Type::getInt32Ty(L->getHeader()->getContext());
120   LoopDecrement = ConstantInt::get(CountType, 1);
121 }
122 
123 bool HardwareLoopInfo::isHardwareLoopCandidate(ScalarEvolution &SE,
124                                                LoopInfo &LI, DominatorTree &DT,
125                                                bool ForceNestedLoop,
126                                                bool ForceHardwareLoopPHI) {
127   SmallVector<BasicBlock *, 4> ExitingBlocks;
128   L->getExitingBlocks(ExitingBlocks);
129 
130   for (BasicBlock *BB : ExitingBlocks) {
131     // If we pass the updated counter back through a phi, we need to know
132     // which latch the updated value will be coming from.
133     if (!L->isLoopLatch(BB)) {
134       if (ForceHardwareLoopPHI || CounterInReg)
135         continue;
136     }
137 
138     const SCEV *EC = SE.getExitCount(L, BB);
139     if (isa<SCEVCouldNotCompute>(EC))
140       continue;
141     if (const SCEVConstant *ConstEC = dyn_cast<SCEVConstant>(EC)) {
142       if (ConstEC->getValue()->isZero())
143         continue;
144     } else if (!SE.isLoopInvariant(EC, L))
145       continue;
146 
147     if (SE.getTypeSizeInBits(EC->getType()) > CountType->getBitWidth())
148       continue;
149 
150     // If this exiting block is contained in a nested loop, it is not eligible
151     // for insertion of the branch-and-decrement since the inner loop would
152     // end up messing up the value in the CTR.
153     if (!IsNestingLegal && LI.getLoopFor(BB) != L && !ForceNestedLoop)
154       continue;
155 
156     // We now have a loop-invariant count of loop iterations (which is not the
157     // constant zero) for which we know that this loop will not exit via this
158     // existing block.
159 
160     // We need to make sure that this block will run on every loop iteration.
161     // For this to be true, we must dominate all blocks with backedges. Such
162     // blocks are in-loop predecessors to the header block.
163     bool NotAlways = false;
164     for (BasicBlock *Pred : predecessors(L->getHeader())) {
165       if (!L->contains(Pred))
166         continue;
167 
168       if (!DT.dominates(BB, Pred)) {
169         NotAlways = true;
170         break;
171       }
172     }
173 
174     if (NotAlways)
175       continue;
176 
177     // Make sure this blocks ends with a conditional branch.
178     Instruction *TI = BB->getTerminator();
179     if (!TI)
180       continue;
181 
182     if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
183       if (!BI->isConditional())
184         continue;
185 
186       ExitBranch = BI;
187     } else
188       continue;
189 
190     // Note that this block may not be the loop latch block, even if the loop
191     // has a latch block.
192     ExitBlock = BB;
193     ExitCount = EC;
194     break;
195   }
196 
197   if (!ExitBlock)
198     return false;
199   return true;
200 }
201 
202 TargetTransformInfo::TargetTransformInfo(const DataLayout &DL)
203     : TTIImpl(new Model<NoTTIImpl>(NoTTIImpl(DL))) {}
204 
205 TargetTransformInfo::~TargetTransformInfo() = default;
206 
207 TargetTransformInfo::TargetTransformInfo(TargetTransformInfo &&Arg)
208     : TTIImpl(std::move(Arg.TTIImpl)) {}
209 
210 TargetTransformInfo &TargetTransformInfo::operator=(TargetTransformInfo &&RHS) {
211   TTIImpl = std::move(RHS.TTIImpl);
212   return *this;
213 }
214 
215 unsigned TargetTransformInfo::getInliningThresholdMultiplier() const {
216   return TTIImpl->getInliningThresholdMultiplier();
217 }
218 
219 unsigned
220 TargetTransformInfo::getInliningCostBenefitAnalysisSavingsMultiplier() const {
221   return TTIImpl->getInliningCostBenefitAnalysisSavingsMultiplier();
222 }
223 
224 unsigned
225 TargetTransformInfo::getInliningCostBenefitAnalysisProfitableMultiplier()
226     const {
227   return TTIImpl->getInliningCostBenefitAnalysisProfitableMultiplier();
228 }
229 
230 int TargetTransformInfo::getInliningLastCallToStaticBonus() const {
231   return TTIImpl->getInliningLastCallToStaticBonus();
232 }
233 
234 unsigned
235 TargetTransformInfo::adjustInliningThreshold(const CallBase *CB) const {
236   return TTIImpl->adjustInliningThreshold(CB);
237 }
238 
239 unsigned TargetTransformInfo::getCallerAllocaCost(const CallBase *CB,
240                                                   const AllocaInst *AI) const {
241   return TTIImpl->getCallerAllocaCost(CB, AI);
242 }
243 
244 int TargetTransformInfo::getInlinerVectorBonusPercent() const {
245   return TTIImpl->getInlinerVectorBonusPercent();
246 }
247 
248 InstructionCost TargetTransformInfo::getGEPCost(
249     Type *PointeeType, const Value *Ptr, ArrayRef<const Value *> Operands,
250     Type *AccessType, TTI::TargetCostKind CostKind) const {
251   return TTIImpl->getGEPCost(PointeeType, Ptr, Operands, AccessType, CostKind);
252 }
253 
254 InstructionCost TargetTransformInfo::getPointersChainCost(
255     ArrayRef<const Value *> Ptrs, const Value *Base,
256     const TTI::PointersChainInfo &Info, Type *AccessTy,
257     TTI::TargetCostKind CostKind) const {
258   assert((Base || !Info.isSameBase()) &&
259          "If pointers have same base address it has to be provided.");
260   return TTIImpl->getPointersChainCost(Ptrs, Base, Info, AccessTy, CostKind);
261 }
262 
263 unsigned TargetTransformInfo::getEstimatedNumberOfCaseClusters(
264     const SwitchInst &SI, unsigned &JTSize, ProfileSummaryInfo *PSI,
265     BlockFrequencyInfo *BFI) const {
266   return TTIImpl->getEstimatedNumberOfCaseClusters(SI, JTSize, PSI, BFI);
267 }
268 
269 InstructionCost
270 TargetTransformInfo::getInstructionCost(const User *U,
271                                         ArrayRef<const Value *> Operands,
272                                         enum TargetCostKind CostKind) const {
273   InstructionCost Cost = TTIImpl->getInstructionCost(U, Operands, CostKind);
274   assert((CostKind == TTI::TCK_RecipThroughput || Cost >= 0) &&
275          "TTI should not produce negative costs!");
276   return Cost;
277 }
278 
279 BranchProbability TargetTransformInfo::getPredictableBranchThreshold() const {
280   return PredictableBranchThreshold.getNumOccurrences() > 0
281              ? BranchProbability(PredictableBranchThreshold, 100)
282              : TTIImpl->getPredictableBranchThreshold();
283 }
284 
285 InstructionCost TargetTransformInfo::getBranchMispredictPenalty() const {
286   return TTIImpl->getBranchMispredictPenalty();
287 }
288 
289 bool TargetTransformInfo::hasBranchDivergence(const Function *F) const {
290   return TTIImpl->hasBranchDivergence(F);
291 }
292 
293 bool TargetTransformInfo::isSourceOfDivergence(const Value *V) const {
294   if (const auto *Call = dyn_cast<CallBase>(V)) {
295     if (Call->hasFnAttr(Attribute::NoDivergenceSource))
296       return false;
297   }
298   return TTIImpl->isSourceOfDivergence(V);
299 }
300 
301 bool llvm::TargetTransformInfo::isAlwaysUniform(const Value *V) const {
302   return TTIImpl->isAlwaysUniform(V);
303 }
304 
305 bool llvm::TargetTransformInfo::isValidAddrSpaceCast(unsigned FromAS,
306                                                      unsigned ToAS) const {
307   return TTIImpl->isValidAddrSpaceCast(FromAS, ToAS);
308 }
309 
310 bool llvm::TargetTransformInfo::addrspacesMayAlias(unsigned FromAS,
311                                                    unsigned ToAS) const {
312   return TTIImpl->addrspacesMayAlias(FromAS, ToAS);
313 }
314 
315 unsigned TargetTransformInfo::getFlatAddressSpace() const {
316   return TTIImpl->getFlatAddressSpace();
317 }
318 
319 bool TargetTransformInfo::collectFlatAddressOperands(
320     SmallVectorImpl<int> &OpIndexes, Intrinsic::ID IID) const {
321   return TTIImpl->collectFlatAddressOperands(OpIndexes, IID);
322 }
323 
324 bool TargetTransformInfo::isNoopAddrSpaceCast(unsigned FromAS,
325                                               unsigned ToAS) const {
326   return TTIImpl->isNoopAddrSpaceCast(FromAS, ToAS);
327 }
328 
329 bool TargetTransformInfo::canHaveNonUndefGlobalInitializerInAddressSpace(
330     unsigned AS) const {
331   return TTIImpl->canHaveNonUndefGlobalInitializerInAddressSpace(AS);
332 }
333 
334 unsigned TargetTransformInfo::getAssumedAddrSpace(const Value *V) const {
335   return TTIImpl->getAssumedAddrSpace(V);
336 }
337 
338 bool TargetTransformInfo::isSingleThreaded() const {
339   return TTIImpl->isSingleThreaded();
340 }
341 
342 std::pair<const Value *, unsigned>
343 TargetTransformInfo::getPredicatedAddrSpace(const Value *V) const {
344   return TTIImpl->getPredicatedAddrSpace(V);
345 }
346 
347 Value *TargetTransformInfo::rewriteIntrinsicWithAddressSpace(
348     IntrinsicInst *II, Value *OldV, Value *NewV) const {
349   return TTIImpl->rewriteIntrinsicWithAddressSpace(II, OldV, NewV);
350 }
351 
352 bool TargetTransformInfo::isLoweredToCall(const Function *F) const {
353   return TTIImpl->isLoweredToCall(F);
354 }
355 
356 bool TargetTransformInfo::isHardwareLoopProfitable(
357     Loop *L, ScalarEvolution &SE, AssumptionCache &AC,
358     TargetLibraryInfo *LibInfo, HardwareLoopInfo &HWLoopInfo) const {
359   return TTIImpl->isHardwareLoopProfitable(L, SE, AC, LibInfo, HWLoopInfo);
360 }
361 
362 unsigned TargetTransformInfo::getEpilogueVectorizationMinVF() const {
363   return TTIImpl->getEpilogueVectorizationMinVF();
364 }
365 
366 bool TargetTransformInfo::preferPredicateOverEpilogue(
367     TailFoldingInfo *TFI) const {
368   return TTIImpl->preferPredicateOverEpilogue(TFI);
369 }
370 
371 TailFoldingStyle TargetTransformInfo::getPreferredTailFoldingStyle(
372     bool IVUpdateMayOverflow) const {
373   return TTIImpl->getPreferredTailFoldingStyle(IVUpdateMayOverflow);
374 }
375 
376 std::optional<Instruction *>
377 TargetTransformInfo::instCombineIntrinsic(InstCombiner &IC,
378                                           IntrinsicInst &II) const {
379   return TTIImpl->instCombineIntrinsic(IC, II);
380 }
381 
382 std::optional<Value *> TargetTransformInfo::simplifyDemandedUseBitsIntrinsic(
383     InstCombiner &IC, IntrinsicInst &II, APInt DemandedMask, KnownBits &Known,
384     bool &KnownBitsComputed) const {
385   return TTIImpl->simplifyDemandedUseBitsIntrinsic(IC, II, DemandedMask, Known,
386                                                    KnownBitsComputed);
387 }
388 
389 std::optional<Value *> TargetTransformInfo::simplifyDemandedVectorEltsIntrinsic(
390     InstCombiner &IC, IntrinsicInst &II, APInt DemandedElts, APInt &UndefElts,
391     APInt &UndefElts2, APInt &UndefElts3,
392     std::function<void(Instruction *, unsigned, APInt, APInt &)>
393         SimplifyAndSetOp) const {
394   return TTIImpl->simplifyDemandedVectorEltsIntrinsic(
395       IC, II, DemandedElts, UndefElts, UndefElts2, UndefElts3,
396       SimplifyAndSetOp);
397 }
398 
399 void TargetTransformInfo::getUnrollingPreferences(
400     Loop *L, ScalarEvolution &SE, UnrollingPreferences &UP,
401     OptimizationRemarkEmitter *ORE) const {
402   return TTIImpl->getUnrollingPreferences(L, SE, UP, ORE);
403 }
404 
405 void TargetTransformInfo::getPeelingPreferences(Loop *L, ScalarEvolution &SE,
406                                                 PeelingPreferences &PP) const {
407   return TTIImpl->getPeelingPreferences(L, SE, PP);
408 }
409 
410 bool TargetTransformInfo::isLegalAddImmediate(int64_t Imm) const {
411   return TTIImpl->isLegalAddImmediate(Imm);
412 }
413 
414 bool TargetTransformInfo::isLegalAddScalableImmediate(int64_t Imm) const {
415   return TTIImpl->isLegalAddScalableImmediate(Imm);
416 }
417 
418 bool TargetTransformInfo::isLegalICmpImmediate(int64_t Imm) const {
419   return TTIImpl->isLegalICmpImmediate(Imm);
420 }
421 
422 bool TargetTransformInfo::isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV,
423                                                 int64_t BaseOffset,
424                                                 bool HasBaseReg, int64_t Scale,
425                                                 unsigned AddrSpace,
426                                                 Instruction *I,
427                                                 int64_t ScalableOffset) const {
428   return TTIImpl->isLegalAddressingMode(Ty, BaseGV, BaseOffset, HasBaseReg,
429                                         Scale, AddrSpace, I, ScalableOffset);
430 }
431 
432 bool TargetTransformInfo::isLSRCostLess(const LSRCost &C1,
433                                         const LSRCost &C2) const {
434   return TTIImpl->isLSRCostLess(C1, C2);
435 }
436 
437 bool TargetTransformInfo::isNumRegsMajorCostOfLSR() const {
438   return TTIImpl->isNumRegsMajorCostOfLSR();
439 }
440 
441 bool TargetTransformInfo::shouldDropLSRSolutionIfLessProfitable() const {
442   return TTIImpl->shouldDropLSRSolutionIfLessProfitable();
443 }
444 
445 bool TargetTransformInfo::isProfitableLSRChainElement(Instruction *I) const {
446   return TTIImpl->isProfitableLSRChainElement(I);
447 }
448 
449 bool TargetTransformInfo::canMacroFuseCmp() const {
450   return TTIImpl->canMacroFuseCmp();
451 }
452 
453 bool TargetTransformInfo::canSaveCmp(Loop *L, BranchInst **BI,
454                                      ScalarEvolution *SE, LoopInfo *LI,
455                                      DominatorTree *DT, AssumptionCache *AC,
456                                      TargetLibraryInfo *LibInfo) const {
457   return TTIImpl->canSaveCmp(L, BI, SE, LI, DT, AC, LibInfo);
458 }
459 
460 TTI::AddressingModeKind
461 TargetTransformInfo::getPreferredAddressingMode(const Loop *L,
462                                                 ScalarEvolution *SE) const {
463   return TTIImpl->getPreferredAddressingMode(L, SE);
464 }
465 
466 bool TargetTransformInfo::isLegalMaskedStore(Type *DataType,
467                                              Align Alignment) const {
468   return TTIImpl->isLegalMaskedStore(DataType, Alignment);
469 }
470 
471 bool TargetTransformInfo::isLegalMaskedLoad(Type *DataType,
472                                             Align Alignment) const {
473   return TTIImpl->isLegalMaskedLoad(DataType, Alignment);
474 }
475 
476 bool TargetTransformInfo::isLegalNTStore(Type *DataType,
477                                          Align Alignment) const {
478   return TTIImpl->isLegalNTStore(DataType, Alignment);
479 }
480 
481 bool TargetTransformInfo::isLegalNTLoad(Type *DataType, Align Alignment) const {
482   return TTIImpl->isLegalNTLoad(DataType, Alignment);
483 }
484 
485 bool TargetTransformInfo::isLegalBroadcastLoad(Type *ElementTy,
486                                                ElementCount NumElements) const {
487   return TTIImpl->isLegalBroadcastLoad(ElementTy, NumElements);
488 }
489 
490 bool TargetTransformInfo::isLegalMaskedGather(Type *DataType,
491                                               Align Alignment) const {
492   return TTIImpl->isLegalMaskedGather(DataType, Alignment);
493 }
494 
495 bool TargetTransformInfo::isLegalAltInstr(
496     VectorType *VecTy, unsigned Opcode0, unsigned Opcode1,
497     const SmallBitVector &OpcodeMask) const {
498   return TTIImpl->isLegalAltInstr(VecTy, Opcode0, Opcode1, OpcodeMask);
499 }
500 
501 bool TargetTransformInfo::isLegalMaskedScatter(Type *DataType,
502                                                Align Alignment) const {
503   return TTIImpl->isLegalMaskedScatter(DataType, Alignment);
504 }
505 
506 bool TargetTransformInfo::forceScalarizeMaskedGather(VectorType *DataType,
507                                                      Align Alignment) const {
508   return TTIImpl->forceScalarizeMaskedGather(DataType, Alignment);
509 }
510 
511 bool TargetTransformInfo::forceScalarizeMaskedScatter(VectorType *DataType,
512                                                       Align Alignment) const {
513   return TTIImpl->forceScalarizeMaskedScatter(DataType, Alignment);
514 }
515 
516 bool TargetTransformInfo::isLegalMaskedCompressStore(Type *DataType,
517                                                      Align Alignment) const {
518   return TTIImpl->isLegalMaskedCompressStore(DataType, Alignment);
519 }
520 
521 bool TargetTransformInfo::isLegalMaskedExpandLoad(Type *DataType,
522                                                   Align Alignment) const {
523   return TTIImpl->isLegalMaskedExpandLoad(DataType, Alignment);
524 }
525 
526 bool TargetTransformInfo::isLegalStridedLoadStore(Type *DataType,
527                                                   Align Alignment) const {
528   return TTIImpl->isLegalStridedLoadStore(DataType, Alignment);
529 }
530 
531 bool TargetTransformInfo::isLegalInterleavedAccessType(
532     VectorType *VTy, unsigned Factor, Align Alignment,
533     unsigned AddrSpace) const {
534   return TTIImpl->isLegalInterleavedAccessType(VTy, Factor, Alignment,
535                                                AddrSpace);
536 }
537 
538 bool TargetTransformInfo::isLegalMaskedVectorHistogram(Type *AddrType,
539                                                        Type *DataType) const {
540   return TTIImpl->isLegalMaskedVectorHistogram(AddrType, DataType);
541 }
542 
543 bool TargetTransformInfo::enableOrderedReductions() const {
544   return TTIImpl->enableOrderedReductions();
545 }
546 
547 bool TargetTransformInfo::hasDivRemOp(Type *DataType, bool IsSigned) const {
548   return TTIImpl->hasDivRemOp(DataType, IsSigned);
549 }
550 
551 bool TargetTransformInfo::hasVolatileVariant(Instruction *I,
552                                              unsigned AddrSpace) const {
553   return TTIImpl->hasVolatileVariant(I, AddrSpace);
554 }
555 
556 bool TargetTransformInfo::prefersVectorizedAddressing() const {
557   return TTIImpl->prefersVectorizedAddressing();
558 }
559 
560 InstructionCost TargetTransformInfo::getScalingFactorCost(
561     Type *Ty, GlobalValue *BaseGV, StackOffset BaseOffset, bool HasBaseReg,
562     int64_t Scale, unsigned AddrSpace) const {
563   InstructionCost Cost = TTIImpl->getScalingFactorCost(
564       Ty, BaseGV, BaseOffset, HasBaseReg, Scale, AddrSpace);
565   assert(Cost >= 0 && "TTI should not produce negative costs!");
566   return Cost;
567 }
568 
569 bool TargetTransformInfo::LSRWithInstrQueries() const {
570   return TTIImpl->LSRWithInstrQueries();
571 }
572 
573 bool TargetTransformInfo::isTruncateFree(Type *Ty1, Type *Ty2) const {
574   return TTIImpl->isTruncateFree(Ty1, Ty2);
575 }
576 
577 bool TargetTransformInfo::isProfitableToHoist(Instruction *I) const {
578   return TTIImpl->isProfitableToHoist(I);
579 }
580 
581 bool TargetTransformInfo::useAA() const { return TTIImpl->useAA(); }
582 
583 bool TargetTransformInfo::isTypeLegal(Type *Ty) const {
584   return TTIImpl->isTypeLegal(Ty);
585 }
586 
587 unsigned TargetTransformInfo::getRegUsageForType(Type *Ty) const {
588   return TTIImpl->getRegUsageForType(Ty);
589 }
590 
591 bool TargetTransformInfo::shouldBuildLookupTables() const {
592   return TTIImpl->shouldBuildLookupTables();
593 }
594 
595 bool TargetTransformInfo::shouldBuildLookupTablesForConstant(
596     Constant *C) const {
597   return TTIImpl->shouldBuildLookupTablesForConstant(C);
598 }
599 
600 bool TargetTransformInfo::shouldBuildRelLookupTables() const {
601   return TTIImpl->shouldBuildRelLookupTables();
602 }
603 
604 bool TargetTransformInfo::useColdCCForColdCall(Function &F) const {
605   return TTIImpl->useColdCCForColdCall(F);
606 }
607 
608 bool TargetTransformInfo::isTargetIntrinsicTriviallyScalarizable(
609     Intrinsic::ID ID) const {
610   return TTIImpl->isTargetIntrinsicTriviallyScalarizable(ID);
611 }
612 
613 bool TargetTransformInfo::isTargetIntrinsicWithScalarOpAtArg(
614     Intrinsic::ID ID, unsigned ScalarOpdIdx) const {
615   return TTIImpl->isTargetIntrinsicWithScalarOpAtArg(ID, ScalarOpdIdx);
616 }
617 
618 bool TargetTransformInfo::isTargetIntrinsicWithOverloadTypeAtArg(
619     Intrinsic::ID ID, int OpdIdx) const {
620   return TTIImpl->isTargetIntrinsicWithOverloadTypeAtArg(ID, OpdIdx);
621 }
622 
623 bool TargetTransformInfo::isTargetIntrinsicWithStructReturnOverloadAtField(
624     Intrinsic::ID ID, int RetIdx) const {
625   return TTIImpl->isTargetIntrinsicWithStructReturnOverloadAtField(ID, RetIdx);
626 }
627 
628 InstructionCost TargetTransformInfo::getScalarizationOverhead(
629     VectorType *Ty, const APInt &DemandedElts, bool Insert, bool Extract,
630     TTI::TargetCostKind CostKind, ArrayRef<Value *> VL) const {
631   return TTIImpl->getScalarizationOverhead(Ty, DemandedElts, Insert, Extract,
632                                            CostKind, VL);
633 }
634 
635 InstructionCost TargetTransformInfo::getOperandsScalarizationOverhead(
636     ArrayRef<const Value *> Args, ArrayRef<Type *> Tys,
637     TTI::TargetCostKind CostKind) const {
638   return TTIImpl->getOperandsScalarizationOverhead(Args, Tys, CostKind);
639 }
640 
641 bool TargetTransformInfo::supportsEfficientVectorElementLoadStore() const {
642   return TTIImpl->supportsEfficientVectorElementLoadStore();
643 }
644 
645 bool TargetTransformInfo::supportsTailCalls() const {
646   return TTIImpl->supportsTailCalls();
647 }
648 
649 bool TargetTransformInfo::supportsTailCallFor(const CallBase *CB) const {
650   return TTIImpl->supportsTailCallFor(CB);
651 }
652 
653 bool TargetTransformInfo::enableAggressiveInterleaving(
654     bool LoopHasReductions) const {
655   return TTIImpl->enableAggressiveInterleaving(LoopHasReductions);
656 }
657 
658 TargetTransformInfo::MemCmpExpansionOptions
659 TargetTransformInfo::enableMemCmpExpansion(bool OptSize, bool IsZeroCmp) const {
660   return TTIImpl->enableMemCmpExpansion(OptSize, IsZeroCmp);
661 }
662 
663 bool TargetTransformInfo::enableSelectOptimize() const {
664   return TTIImpl->enableSelectOptimize();
665 }
666 
667 bool TargetTransformInfo::shouldTreatInstructionLikeSelect(
668     const Instruction *I) const {
669   return TTIImpl->shouldTreatInstructionLikeSelect(I);
670 }
671 
672 bool TargetTransformInfo::enableInterleavedAccessVectorization() const {
673   return TTIImpl->enableInterleavedAccessVectorization();
674 }
675 
676 bool TargetTransformInfo::enableMaskedInterleavedAccessVectorization() const {
677   return TTIImpl->enableMaskedInterleavedAccessVectorization();
678 }
679 
680 bool TargetTransformInfo::isFPVectorizationPotentiallyUnsafe() const {
681   return TTIImpl->isFPVectorizationPotentiallyUnsafe();
682 }
683 
684 bool
685 TargetTransformInfo::allowsMisalignedMemoryAccesses(LLVMContext &Context,
686                                                     unsigned BitWidth,
687                                                     unsigned AddressSpace,
688                                                     Align Alignment,
689                                                     unsigned *Fast) const {
690   return TTIImpl->allowsMisalignedMemoryAccesses(Context, BitWidth,
691                                                  AddressSpace, Alignment, Fast);
692 }
693 
694 TargetTransformInfo::PopcntSupportKind
695 TargetTransformInfo::getPopcntSupport(unsigned IntTyWidthInBit) const {
696   return TTIImpl->getPopcntSupport(IntTyWidthInBit);
697 }
698 
699 bool TargetTransformInfo::haveFastSqrt(Type *Ty) const {
700   return TTIImpl->haveFastSqrt(Ty);
701 }
702 
703 bool TargetTransformInfo::isExpensiveToSpeculativelyExecute(
704     const Instruction *I) const {
705   return TTIImpl->isExpensiveToSpeculativelyExecute(I);
706 }
707 
708 bool TargetTransformInfo::isFCmpOrdCheaperThanFCmpZero(Type *Ty) const {
709   return TTIImpl->isFCmpOrdCheaperThanFCmpZero(Ty);
710 }
711 
712 InstructionCost TargetTransformInfo::getFPOpCost(Type *Ty) const {
713   InstructionCost Cost = TTIImpl->getFPOpCost(Ty);
714   assert(Cost >= 0 && "TTI should not produce negative costs!");
715   return Cost;
716 }
717 
718 InstructionCost TargetTransformInfo::getIntImmCodeSizeCost(unsigned Opcode,
719                                                            unsigned Idx,
720                                                            const APInt &Imm,
721                                                            Type *Ty) const {
722   InstructionCost Cost = TTIImpl->getIntImmCodeSizeCost(Opcode, Idx, Imm, Ty);
723   assert(Cost >= 0 && "TTI should not produce negative costs!");
724   return Cost;
725 }
726 
727 InstructionCost
728 TargetTransformInfo::getIntImmCost(const APInt &Imm, Type *Ty,
729                                    TTI::TargetCostKind CostKind) const {
730   InstructionCost Cost = TTIImpl->getIntImmCost(Imm, Ty, CostKind);
731   assert(Cost >= 0 && "TTI should not produce negative costs!");
732   return Cost;
733 }
734 
735 InstructionCost TargetTransformInfo::getIntImmCostInst(
736     unsigned Opcode, unsigned Idx, const APInt &Imm, Type *Ty,
737     TTI::TargetCostKind CostKind, Instruction *Inst) const {
738   InstructionCost Cost =
739       TTIImpl->getIntImmCostInst(Opcode, Idx, Imm, Ty, CostKind, Inst);
740   assert(Cost >= 0 && "TTI should not produce negative costs!");
741   return Cost;
742 }
743 
744 InstructionCost
745 TargetTransformInfo::getIntImmCostIntrin(Intrinsic::ID IID, unsigned Idx,
746                                          const APInt &Imm, Type *Ty,
747                                          TTI::TargetCostKind CostKind) const {
748   InstructionCost Cost =
749       TTIImpl->getIntImmCostIntrin(IID, Idx, Imm, Ty, CostKind);
750   assert(Cost >= 0 && "TTI should not produce negative costs!");
751   return Cost;
752 }
753 
754 bool TargetTransformInfo::preferToKeepConstantsAttached(
755     const Instruction &Inst, const Function &Fn) const {
756   return TTIImpl->preferToKeepConstantsAttached(Inst, Fn);
757 }
758 
759 unsigned TargetTransformInfo::getNumberOfRegisters(unsigned ClassID) const {
760   return TTIImpl->getNumberOfRegisters(ClassID);
761 }
762 
763 bool TargetTransformInfo::hasConditionalLoadStoreForType(Type *Ty) const {
764   return TTIImpl->hasConditionalLoadStoreForType(Ty);
765 }
766 
767 unsigned TargetTransformInfo::getRegisterClassForType(bool Vector,
768                                                       Type *Ty) const {
769   return TTIImpl->getRegisterClassForType(Vector, Ty);
770 }
771 
772 const char *TargetTransformInfo::getRegisterClassName(unsigned ClassID) const {
773   return TTIImpl->getRegisterClassName(ClassID);
774 }
775 
776 TypeSize TargetTransformInfo::getRegisterBitWidth(
777     TargetTransformInfo::RegisterKind K) const {
778   return TTIImpl->getRegisterBitWidth(K);
779 }
780 
781 unsigned TargetTransformInfo::getMinVectorRegisterBitWidth() const {
782   return TTIImpl->getMinVectorRegisterBitWidth();
783 }
784 
785 std::optional<unsigned> TargetTransformInfo::getMaxVScale() const {
786   return TTIImpl->getMaxVScale();
787 }
788 
789 std::optional<unsigned> TargetTransformInfo::getVScaleForTuning() const {
790   return TTIImpl->getVScaleForTuning();
791 }
792 
793 bool TargetTransformInfo::isVScaleKnownToBeAPowerOfTwo() const {
794   return TTIImpl->isVScaleKnownToBeAPowerOfTwo();
795 }
796 
797 bool TargetTransformInfo::shouldMaximizeVectorBandwidth(
798     TargetTransformInfo::RegisterKind K) const {
799   return TTIImpl->shouldMaximizeVectorBandwidth(K);
800 }
801 
802 ElementCount TargetTransformInfo::getMinimumVF(unsigned ElemWidth,
803                                                bool IsScalable) const {
804   return TTIImpl->getMinimumVF(ElemWidth, IsScalable);
805 }
806 
807 unsigned TargetTransformInfo::getMaximumVF(unsigned ElemWidth,
808                                            unsigned Opcode) const {
809   return TTIImpl->getMaximumVF(ElemWidth, Opcode);
810 }
811 
812 unsigned TargetTransformInfo::getStoreMinimumVF(unsigned VF, Type *ScalarMemTy,
813                                                 Type *ScalarValTy) const {
814   return TTIImpl->getStoreMinimumVF(VF, ScalarMemTy, ScalarValTy);
815 }
816 
817 bool TargetTransformInfo::shouldConsiderAddressTypePromotion(
818     const Instruction &I, bool &AllowPromotionWithoutCommonHeader) const {
819   return TTIImpl->shouldConsiderAddressTypePromotion(
820       I, AllowPromotionWithoutCommonHeader);
821 }
822 
823 unsigned TargetTransformInfo::getCacheLineSize() const {
824   return CacheLineSize.getNumOccurrences() > 0 ? CacheLineSize
825                                                : TTIImpl->getCacheLineSize();
826 }
827 
828 std::optional<unsigned>
829 TargetTransformInfo::getCacheSize(CacheLevel Level) const {
830   return TTIImpl->getCacheSize(Level);
831 }
832 
833 std::optional<unsigned>
834 TargetTransformInfo::getCacheAssociativity(CacheLevel Level) const {
835   return TTIImpl->getCacheAssociativity(Level);
836 }
837 
838 std::optional<unsigned> TargetTransformInfo::getMinPageSize() const {
839   return MinPageSize.getNumOccurrences() > 0 ? MinPageSize
840                                              : TTIImpl->getMinPageSize();
841 }
842 
843 unsigned TargetTransformInfo::getPrefetchDistance() const {
844   return TTIImpl->getPrefetchDistance();
845 }
846 
847 unsigned TargetTransformInfo::getMinPrefetchStride(
848     unsigned NumMemAccesses, unsigned NumStridedMemAccesses,
849     unsigned NumPrefetches, bool HasCall) const {
850   return TTIImpl->getMinPrefetchStride(NumMemAccesses, NumStridedMemAccesses,
851                                        NumPrefetches, HasCall);
852 }
853 
854 unsigned TargetTransformInfo::getMaxPrefetchIterationsAhead() const {
855   return TTIImpl->getMaxPrefetchIterationsAhead();
856 }
857 
858 bool TargetTransformInfo::enableWritePrefetching() const {
859   return TTIImpl->enableWritePrefetching();
860 }
861 
862 bool TargetTransformInfo::shouldPrefetchAddressSpace(unsigned AS) const {
863   return TTIImpl->shouldPrefetchAddressSpace(AS);
864 }
865 
866 InstructionCost TargetTransformInfo::getPartialReductionCost(
867     unsigned Opcode, Type *InputTypeA, Type *InputTypeB, Type *AccumType,
868     ElementCount VF, PartialReductionExtendKind OpAExtend,
869     PartialReductionExtendKind OpBExtend, std::optional<unsigned> BinOp) const {
870   return TTIImpl->getPartialReductionCost(Opcode, InputTypeA, InputTypeB,
871                                           AccumType, VF, OpAExtend, OpBExtend,
872                                           BinOp);
873 }
874 
875 unsigned TargetTransformInfo::getMaxInterleaveFactor(ElementCount VF) const {
876   return TTIImpl->getMaxInterleaveFactor(VF);
877 }
878 
879 TargetTransformInfo::OperandValueInfo
880 TargetTransformInfo::getOperandInfo(const Value *V) {
881   OperandValueKind OpInfo = OK_AnyValue;
882   OperandValueProperties OpProps = OP_None;
883 
884   if (isa<ConstantInt>(V) || isa<ConstantFP>(V)) {
885     if (const auto *CI = dyn_cast<ConstantInt>(V)) {
886       if (CI->getValue().isPowerOf2())
887         OpProps = OP_PowerOf2;
888       else if (CI->getValue().isNegatedPowerOf2())
889         OpProps = OP_NegatedPowerOf2;
890     }
891     return {OK_UniformConstantValue, OpProps};
892   }
893 
894   // A broadcast shuffle creates a uniform value.
895   // TODO: Add support for non-zero index broadcasts.
896   // TODO: Add support for different source vector width.
897   if (const auto *ShuffleInst = dyn_cast<ShuffleVectorInst>(V))
898     if (ShuffleInst->isZeroEltSplat())
899       OpInfo = OK_UniformValue;
900 
901   const Value *Splat = getSplatValue(V);
902 
903   // Check for a splat of a constant or for a non uniform vector of constants
904   // and check if the constant(s) are all powers of two.
905   if (isa<ConstantVector>(V) || isa<ConstantDataVector>(V)) {
906     OpInfo = OK_NonUniformConstantValue;
907     if (Splat) {
908       OpInfo = OK_UniformConstantValue;
909       if (auto *CI = dyn_cast<ConstantInt>(Splat)) {
910         if (CI->getValue().isPowerOf2())
911           OpProps = OP_PowerOf2;
912         else if (CI->getValue().isNegatedPowerOf2())
913           OpProps = OP_NegatedPowerOf2;
914       }
915     } else if (const auto *CDS = dyn_cast<ConstantDataSequential>(V)) {
916       bool AllPow2 = true, AllNegPow2 = true;
917       for (unsigned I = 0, E = CDS->getNumElements(); I != E; ++I) {
918         if (auto *CI = dyn_cast<ConstantInt>(CDS->getElementAsConstant(I))) {
919           AllPow2 &= CI->getValue().isPowerOf2();
920           AllNegPow2 &= CI->getValue().isNegatedPowerOf2();
921           if (AllPow2 || AllNegPow2)
922             continue;
923         }
924         AllPow2 = AllNegPow2 = false;
925         break;
926       }
927       OpProps = AllPow2 ? OP_PowerOf2 : OpProps;
928       OpProps = AllNegPow2 ? OP_NegatedPowerOf2 : OpProps;
929     }
930   }
931 
932   // Check for a splat of a uniform value. This is not loop aware, so return
933   // true only for the obviously uniform cases (argument, globalvalue)
934   if (Splat && (isa<Argument>(Splat) || isa<GlobalValue>(Splat)))
935     OpInfo = OK_UniformValue;
936 
937   return {OpInfo, OpProps};
938 }
939 
940 InstructionCost TargetTransformInfo::getArithmeticInstrCost(
941     unsigned Opcode, Type *Ty, TTI::TargetCostKind CostKind,
942     OperandValueInfo Op1Info, OperandValueInfo Op2Info,
943     ArrayRef<const Value *> Args, const Instruction *CxtI,
944     const TargetLibraryInfo *TLibInfo) const {
945 
946   // Use call cost for frem intructions that have platform specific vector math
947   // functions, as those will be replaced with calls later by SelectionDAG or
948   // ReplaceWithVecLib pass.
949   if (TLibInfo && Opcode == Instruction::FRem) {
950     VectorType *VecTy = dyn_cast<VectorType>(Ty);
951     LibFunc Func;
952     if (VecTy &&
953         TLibInfo->getLibFunc(Instruction::FRem, Ty->getScalarType(), Func) &&
954         TLibInfo->isFunctionVectorizable(TLibInfo->getName(Func),
955                                          VecTy->getElementCount()))
956       return getCallInstrCost(nullptr, VecTy, {VecTy, VecTy}, CostKind);
957   }
958 
959   InstructionCost Cost =
960       TTIImpl->getArithmeticInstrCost(Opcode, Ty, CostKind,
961                                       Op1Info, Op2Info,
962                                       Args, CxtI);
963   assert(Cost >= 0 && "TTI should not produce negative costs!");
964   return Cost;
965 }
966 
967 InstructionCost TargetTransformInfo::getAltInstrCost(
968     VectorType *VecTy, unsigned Opcode0, unsigned Opcode1,
969     const SmallBitVector &OpcodeMask, TTI::TargetCostKind CostKind) const {
970   InstructionCost Cost =
971       TTIImpl->getAltInstrCost(VecTy, Opcode0, Opcode1, OpcodeMask, CostKind);
972   assert(Cost >= 0 && "TTI should not produce negative costs!");
973   return Cost;
974 }
975 
976 InstructionCost TargetTransformInfo::getShuffleCost(
977     ShuffleKind Kind, VectorType *Ty, ArrayRef<int> Mask,
978     TTI::TargetCostKind CostKind, int Index, VectorType *SubTp,
979     ArrayRef<const Value *> Args, const Instruction *CxtI) const {
980   InstructionCost Cost = TTIImpl->getShuffleCost(Kind, Ty, Mask, CostKind,
981                                                  Index, SubTp, Args, CxtI);
982   assert(Cost >= 0 && "TTI should not produce negative costs!");
983   return Cost;
984 }
985 
986 TargetTransformInfo::PartialReductionExtendKind
987 TargetTransformInfo::getPartialReductionExtendKind(Instruction *I) {
988   if (isa<SExtInst>(I))
989     return PR_SignExtend;
990   if (isa<ZExtInst>(I))
991     return PR_ZeroExtend;
992   return PR_None;
993 }
994 
995 TTI::CastContextHint
996 TargetTransformInfo::getCastContextHint(const Instruction *I) {
997   if (!I)
998     return CastContextHint::None;
999 
1000   auto getLoadStoreKind = [](const Value *V, unsigned LdStOp, unsigned MaskedOp,
1001                              unsigned GatScatOp) {
1002     const Instruction *I = dyn_cast<Instruction>(V);
1003     if (!I)
1004       return CastContextHint::None;
1005 
1006     if (I->getOpcode() == LdStOp)
1007       return CastContextHint::Normal;
1008 
1009     if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1010       if (II->getIntrinsicID() == MaskedOp)
1011         return TTI::CastContextHint::Masked;
1012       if (II->getIntrinsicID() == GatScatOp)
1013         return TTI::CastContextHint::GatherScatter;
1014     }
1015 
1016     return TTI::CastContextHint::None;
1017   };
1018 
1019   switch (I->getOpcode()) {
1020   case Instruction::ZExt:
1021   case Instruction::SExt:
1022   case Instruction::FPExt:
1023     return getLoadStoreKind(I->getOperand(0), Instruction::Load,
1024                             Intrinsic::masked_load, Intrinsic::masked_gather);
1025   case Instruction::Trunc:
1026   case Instruction::FPTrunc:
1027     if (I->hasOneUse())
1028       return getLoadStoreKind(*I->user_begin(), Instruction::Store,
1029                               Intrinsic::masked_store,
1030                               Intrinsic::masked_scatter);
1031     break;
1032   default:
1033     return CastContextHint::None;
1034   }
1035 
1036   return TTI::CastContextHint::None;
1037 }
1038 
1039 InstructionCost TargetTransformInfo::getCastInstrCost(
1040     unsigned Opcode, Type *Dst, Type *Src, CastContextHint CCH,
1041     TTI::TargetCostKind CostKind, const Instruction *I) const {
1042   assert((I == nullptr || I->getOpcode() == Opcode) &&
1043          "Opcode should reflect passed instruction.");
1044   InstructionCost Cost =
1045       TTIImpl->getCastInstrCost(Opcode, Dst, Src, CCH, CostKind, I);
1046   assert(Cost >= 0 && "TTI should not produce negative costs!");
1047   return Cost;
1048 }
1049 
1050 InstructionCost TargetTransformInfo::getExtractWithExtendCost(
1051     unsigned Opcode, Type *Dst, VectorType *VecTy, unsigned Index) const {
1052   InstructionCost Cost =
1053       TTIImpl->getExtractWithExtendCost(Opcode, Dst, VecTy, Index);
1054   assert(Cost >= 0 && "TTI should not produce negative costs!");
1055   return Cost;
1056 }
1057 
1058 InstructionCost TargetTransformInfo::getCFInstrCost(
1059     unsigned Opcode, TTI::TargetCostKind CostKind, const Instruction *I) const {
1060   assert((I == nullptr || I->getOpcode() == Opcode) &&
1061          "Opcode should reflect passed instruction.");
1062   InstructionCost Cost = TTIImpl->getCFInstrCost(Opcode, CostKind, I);
1063   assert(Cost >= 0 && "TTI should not produce negative costs!");
1064   return Cost;
1065 }
1066 
1067 InstructionCost TargetTransformInfo::getCmpSelInstrCost(
1068     unsigned Opcode, Type *ValTy, Type *CondTy, CmpInst::Predicate VecPred,
1069     TTI::TargetCostKind CostKind, OperandValueInfo Op1Info,
1070     OperandValueInfo Op2Info, const Instruction *I) const {
1071   assert((I == nullptr || I->getOpcode() == Opcode) &&
1072          "Opcode should reflect passed instruction.");
1073   InstructionCost Cost = TTIImpl->getCmpSelInstrCost(
1074       Opcode, ValTy, CondTy, VecPred, CostKind, Op1Info, Op2Info, I);
1075   assert(Cost >= 0 && "TTI should not produce negative costs!");
1076   return Cost;
1077 }
1078 
1079 InstructionCost TargetTransformInfo::getVectorInstrCost(
1080     unsigned Opcode, Type *Val, TTI::TargetCostKind CostKind, unsigned Index,
1081     Value *Op0, Value *Op1) const {
1082   assert((Opcode == Instruction::InsertElement ||
1083           Opcode == Instruction::ExtractElement) &&
1084          "Expecting Opcode to be insertelement/extractelement.");
1085   InstructionCost Cost =
1086       TTIImpl->getVectorInstrCost(Opcode, Val, CostKind, Index, Op0, Op1);
1087   assert(Cost >= 0 && "TTI should not produce negative costs!");
1088   return Cost;
1089 }
1090 
1091 InstructionCost TargetTransformInfo::getVectorInstrCost(
1092     unsigned Opcode, Type *Val, TTI::TargetCostKind CostKind, unsigned Index,
1093     Value *Scalar,
1094     ArrayRef<std::tuple<Value *, User *, int>> ScalarUserAndIdx) const {
1095   assert((Opcode == Instruction::InsertElement ||
1096           Opcode == Instruction::ExtractElement) &&
1097          "Expecting Opcode to be insertelement/extractelement.");
1098   InstructionCost Cost = TTIImpl->getVectorInstrCost(
1099       Opcode, Val, CostKind, Index, Scalar, ScalarUserAndIdx);
1100   assert(Cost >= 0 && "TTI should not produce negative costs!");
1101   return Cost;
1102 }
1103 
1104 InstructionCost
1105 TargetTransformInfo::getVectorInstrCost(const Instruction &I, Type *Val,
1106                                         TTI::TargetCostKind CostKind,
1107                                         unsigned Index) const {
1108   // FIXME: Assert that Opcode is either InsertElement or ExtractElement.
1109   // This is mentioned in the interface description and respected by all
1110   // callers, but never asserted upon.
1111   InstructionCost Cost = TTIImpl->getVectorInstrCost(I, Val, CostKind, Index);
1112   assert(Cost >= 0 && "TTI should not produce negative costs!");
1113   return Cost;
1114 }
1115 
1116 InstructionCost TargetTransformInfo::getReplicationShuffleCost(
1117     Type *EltTy, int ReplicationFactor, int VF, const APInt &DemandedDstElts,
1118     TTI::TargetCostKind CostKind) const {
1119   InstructionCost Cost = TTIImpl->getReplicationShuffleCost(
1120       EltTy, ReplicationFactor, VF, DemandedDstElts, CostKind);
1121   assert(Cost >= 0 && "TTI should not produce negative costs!");
1122   return Cost;
1123 }
1124 
1125 InstructionCost TargetTransformInfo::getMemoryOpCost(
1126     unsigned Opcode, Type *Src, Align Alignment, unsigned AddressSpace,
1127     TTI::TargetCostKind CostKind, TTI::OperandValueInfo OpInfo,
1128     const Instruction *I) const {
1129   assert((I == nullptr || I->getOpcode() == Opcode) &&
1130          "Opcode should reflect passed instruction.");
1131   InstructionCost Cost = TTIImpl->getMemoryOpCost(
1132       Opcode, Src, Alignment, AddressSpace, CostKind, OpInfo, I);
1133   assert(Cost >= 0 && "TTI should not produce negative costs!");
1134   return Cost;
1135 }
1136 
1137 InstructionCost TargetTransformInfo::getMaskedMemoryOpCost(
1138     unsigned Opcode, Type *Src, Align Alignment, unsigned AddressSpace,
1139     TTI::TargetCostKind CostKind) const {
1140   InstructionCost Cost = TTIImpl->getMaskedMemoryOpCost(Opcode, Src, Alignment,
1141                                                         AddressSpace, CostKind);
1142   assert(Cost >= 0 && "TTI should not produce negative costs!");
1143   return Cost;
1144 }
1145 
1146 InstructionCost TargetTransformInfo::getGatherScatterOpCost(
1147     unsigned Opcode, Type *DataTy, const Value *Ptr, bool VariableMask,
1148     Align Alignment, TTI::TargetCostKind CostKind, const Instruction *I) const {
1149   InstructionCost Cost = TTIImpl->getGatherScatterOpCost(
1150       Opcode, DataTy, Ptr, VariableMask, Alignment, CostKind, I);
1151   assert((!Cost.isValid() || Cost >= 0) &&
1152          "TTI should not produce negative costs!");
1153   return Cost;
1154 }
1155 
1156 InstructionCost TargetTransformInfo::getStridedMemoryOpCost(
1157     unsigned Opcode, Type *DataTy, const Value *Ptr, bool VariableMask,
1158     Align Alignment, TTI::TargetCostKind CostKind, const Instruction *I) const {
1159   InstructionCost Cost = TTIImpl->getStridedMemoryOpCost(
1160       Opcode, DataTy, Ptr, VariableMask, Alignment, CostKind, I);
1161   assert(Cost >= 0 && "TTI should not produce negative costs!");
1162   return Cost;
1163 }
1164 
1165 InstructionCost TargetTransformInfo::getInterleavedMemoryOpCost(
1166     unsigned Opcode, Type *VecTy, unsigned Factor, ArrayRef<unsigned> Indices,
1167     Align Alignment, unsigned AddressSpace, TTI::TargetCostKind CostKind,
1168     bool UseMaskForCond, bool UseMaskForGaps) const {
1169   InstructionCost Cost = TTIImpl->getInterleavedMemoryOpCost(
1170       Opcode, VecTy, Factor, Indices, Alignment, AddressSpace, CostKind,
1171       UseMaskForCond, UseMaskForGaps);
1172   assert(Cost >= 0 && "TTI should not produce negative costs!");
1173   return Cost;
1174 }
1175 
1176 InstructionCost
1177 TargetTransformInfo::getIntrinsicInstrCost(const IntrinsicCostAttributes &ICA,
1178                                            TTI::TargetCostKind CostKind) const {
1179   InstructionCost Cost = TTIImpl->getIntrinsicInstrCost(ICA, CostKind);
1180   assert(Cost >= 0 && "TTI should not produce negative costs!");
1181   return Cost;
1182 }
1183 
1184 InstructionCost
1185 TargetTransformInfo::getCallInstrCost(Function *F, Type *RetTy,
1186                                       ArrayRef<Type *> Tys,
1187                                       TTI::TargetCostKind CostKind) const {
1188   InstructionCost Cost = TTIImpl->getCallInstrCost(F, RetTy, Tys, CostKind);
1189   assert(Cost >= 0 && "TTI should not produce negative costs!");
1190   return Cost;
1191 }
1192 
1193 unsigned TargetTransformInfo::getNumberOfParts(Type *Tp) const {
1194   return TTIImpl->getNumberOfParts(Tp);
1195 }
1196 
1197 InstructionCost
1198 TargetTransformInfo::getAddressComputationCost(Type *Tp, ScalarEvolution *SE,
1199                                                const SCEV *Ptr) const {
1200   InstructionCost Cost = TTIImpl->getAddressComputationCost(Tp, SE, Ptr);
1201   assert(Cost >= 0 && "TTI should not produce negative costs!");
1202   return Cost;
1203 }
1204 
1205 InstructionCost TargetTransformInfo::getMemcpyCost(const Instruction *I) const {
1206   InstructionCost Cost = TTIImpl->getMemcpyCost(I);
1207   assert(Cost >= 0 && "TTI should not produce negative costs!");
1208   return Cost;
1209 }
1210 
1211 uint64_t TargetTransformInfo::getMaxMemIntrinsicInlineSizeThreshold() const {
1212   return TTIImpl->getMaxMemIntrinsicInlineSizeThreshold();
1213 }
1214 
1215 InstructionCost TargetTransformInfo::getArithmeticReductionCost(
1216     unsigned Opcode, VectorType *Ty, std::optional<FastMathFlags> FMF,
1217     TTI::TargetCostKind CostKind) const {
1218   InstructionCost Cost =
1219       TTIImpl->getArithmeticReductionCost(Opcode, Ty, FMF, CostKind);
1220   assert(Cost >= 0 && "TTI should not produce negative costs!");
1221   return Cost;
1222 }
1223 
1224 InstructionCost TargetTransformInfo::getMinMaxReductionCost(
1225     Intrinsic::ID IID, VectorType *Ty, FastMathFlags FMF,
1226     TTI::TargetCostKind CostKind) const {
1227   InstructionCost Cost =
1228       TTIImpl->getMinMaxReductionCost(IID, Ty, FMF, CostKind);
1229   assert(Cost >= 0 && "TTI should not produce negative costs!");
1230   return Cost;
1231 }
1232 
1233 InstructionCost TargetTransformInfo::getExtendedReductionCost(
1234     unsigned Opcode, bool IsUnsigned, Type *ResTy, VectorType *Ty,
1235     FastMathFlags FMF, TTI::TargetCostKind CostKind) const {
1236   return TTIImpl->getExtendedReductionCost(Opcode, IsUnsigned, ResTy, Ty, FMF,
1237                                            CostKind);
1238 }
1239 
1240 InstructionCost TargetTransformInfo::getMulAccReductionCost(
1241     bool IsUnsigned, Type *ResTy, VectorType *Ty,
1242     TTI::TargetCostKind CostKind) const {
1243   return TTIImpl->getMulAccReductionCost(IsUnsigned, ResTy, Ty, CostKind);
1244 }
1245 
1246 InstructionCost
1247 TargetTransformInfo::getCostOfKeepingLiveOverCall(ArrayRef<Type *> Tys) const {
1248   return TTIImpl->getCostOfKeepingLiveOverCall(Tys);
1249 }
1250 
1251 bool TargetTransformInfo::getTgtMemIntrinsic(IntrinsicInst *Inst,
1252                                              MemIntrinsicInfo &Info) const {
1253   return TTIImpl->getTgtMemIntrinsic(Inst, Info);
1254 }
1255 
1256 unsigned TargetTransformInfo::getAtomicMemIntrinsicMaxElementSize() const {
1257   return TTIImpl->getAtomicMemIntrinsicMaxElementSize();
1258 }
1259 
1260 Value *TargetTransformInfo::getOrCreateResultFromMemIntrinsic(
1261     IntrinsicInst *Inst, Type *ExpectedType) const {
1262   return TTIImpl->getOrCreateResultFromMemIntrinsic(Inst, ExpectedType);
1263 }
1264 
1265 Type *TargetTransformInfo::getMemcpyLoopLoweringType(
1266     LLVMContext &Context, Value *Length, unsigned SrcAddrSpace,
1267     unsigned DestAddrSpace, Align SrcAlign, Align DestAlign,
1268     std::optional<uint32_t> AtomicElementSize) const {
1269   return TTIImpl->getMemcpyLoopLoweringType(Context, Length, SrcAddrSpace,
1270                                             DestAddrSpace, SrcAlign, DestAlign,
1271                                             AtomicElementSize);
1272 }
1273 
1274 void TargetTransformInfo::getMemcpyLoopResidualLoweringType(
1275     SmallVectorImpl<Type *> &OpsOut, LLVMContext &Context,
1276     unsigned RemainingBytes, unsigned SrcAddrSpace, unsigned DestAddrSpace,
1277     Align SrcAlign, Align DestAlign,
1278     std::optional<uint32_t> AtomicCpySize) const {
1279   TTIImpl->getMemcpyLoopResidualLoweringType(
1280       OpsOut, Context, RemainingBytes, SrcAddrSpace, DestAddrSpace, SrcAlign,
1281       DestAlign, AtomicCpySize);
1282 }
1283 
1284 bool TargetTransformInfo::areInlineCompatible(const Function *Caller,
1285                                               const Function *Callee) const {
1286   return TTIImpl->areInlineCompatible(Caller, Callee);
1287 }
1288 
1289 unsigned
1290 TargetTransformInfo::getInlineCallPenalty(const Function *F,
1291                                           const CallBase &Call,
1292                                           unsigned DefaultCallPenalty) const {
1293   return TTIImpl->getInlineCallPenalty(F, Call, DefaultCallPenalty);
1294 }
1295 
1296 bool TargetTransformInfo::areTypesABICompatible(
1297     const Function *Caller, const Function *Callee,
1298     const ArrayRef<Type *> &Types) const {
1299   return TTIImpl->areTypesABICompatible(Caller, Callee, Types);
1300 }
1301 
1302 bool TargetTransformInfo::isIndexedLoadLegal(MemIndexedMode Mode,
1303                                              Type *Ty) const {
1304   return TTIImpl->isIndexedLoadLegal(Mode, Ty);
1305 }
1306 
1307 bool TargetTransformInfo::isIndexedStoreLegal(MemIndexedMode Mode,
1308                                               Type *Ty) const {
1309   return TTIImpl->isIndexedStoreLegal(Mode, Ty);
1310 }
1311 
1312 unsigned TargetTransformInfo::getLoadStoreVecRegBitWidth(unsigned AS) const {
1313   return TTIImpl->getLoadStoreVecRegBitWidth(AS);
1314 }
1315 
1316 bool TargetTransformInfo::isLegalToVectorizeLoad(LoadInst *LI) const {
1317   return TTIImpl->isLegalToVectorizeLoad(LI);
1318 }
1319 
1320 bool TargetTransformInfo::isLegalToVectorizeStore(StoreInst *SI) const {
1321   return TTIImpl->isLegalToVectorizeStore(SI);
1322 }
1323 
1324 bool TargetTransformInfo::isLegalToVectorizeLoadChain(
1325     unsigned ChainSizeInBytes, Align Alignment, unsigned AddrSpace) const {
1326   return TTIImpl->isLegalToVectorizeLoadChain(ChainSizeInBytes, Alignment,
1327                                               AddrSpace);
1328 }
1329 
1330 bool TargetTransformInfo::isLegalToVectorizeStoreChain(
1331     unsigned ChainSizeInBytes, Align Alignment, unsigned AddrSpace) const {
1332   return TTIImpl->isLegalToVectorizeStoreChain(ChainSizeInBytes, Alignment,
1333                                                AddrSpace);
1334 }
1335 
1336 bool TargetTransformInfo::isLegalToVectorizeReduction(
1337     const RecurrenceDescriptor &RdxDesc, ElementCount VF) const {
1338   return TTIImpl->isLegalToVectorizeReduction(RdxDesc, VF);
1339 }
1340 
1341 bool TargetTransformInfo::isElementTypeLegalForScalableVector(Type *Ty) const {
1342   return TTIImpl->isElementTypeLegalForScalableVector(Ty);
1343 }
1344 
1345 unsigned TargetTransformInfo::getLoadVectorFactor(unsigned VF,
1346                                                   unsigned LoadSize,
1347                                                   unsigned ChainSizeInBytes,
1348                                                   VectorType *VecTy) const {
1349   return TTIImpl->getLoadVectorFactor(VF, LoadSize, ChainSizeInBytes, VecTy);
1350 }
1351 
1352 unsigned TargetTransformInfo::getStoreVectorFactor(unsigned VF,
1353                                                    unsigned StoreSize,
1354                                                    unsigned ChainSizeInBytes,
1355                                                    VectorType *VecTy) const {
1356   return TTIImpl->getStoreVectorFactor(VF, StoreSize, ChainSizeInBytes, VecTy);
1357 }
1358 
1359 bool TargetTransformInfo::preferFixedOverScalableIfEqualCost() const {
1360   return TTIImpl->preferFixedOverScalableIfEqualCost();
1361 }
1362 
1363 bool TargetTransformInfo::preferInLoopReduction(unsigned Opcode, Type *Ty,
1364                                                 ReductionFlags Flags) const {
1365   return TTIImpl->preferInLoopReduction(Opcode, Ty, Flags);
1366 }
1367 
1368 bool TargetTransformInfo::preferPredicatedReductionSelect(
1369     unsigned Opcode, Type *Ty, ReductionFlags Flags) const {
1370   return TTIImpl->preferPredicatedReductionSelect(Opcode, Ty, Flags);
1371 }
1372 
1373 bool TargetTransformInfo::preferEpilogueVectorization() const {
1374   return TTIImpl->preferEpilogueVectorization();
1375 }
1376 
1377 TargetTransformInfo::VPLegalization
1378 TargetTransformInfo::getVPLegalizationStrategy(const VPIntrinsic &VPI) const {
1379   return TTIImpl->getVPLegalizationStrategy(VPI);
1380 }
1381 
1382 bool TargetTransformInfo::hasArmWideBranch(bool Thumb) const {
1383   return TTIImpl->hasArmWideBranch(Thumb);
1384 }
1385 
1386 uint64_t TargetTransformInfo::getFeatureMask(const Function &F) const {
1387   return TTIImpl->getFeatureMask(F);
1388 }
1389 
1390 bool TargetTransformInfo::isMultiversionedFunction(const Function &F) const {
1391   return TTIImpl->isMultiversionedFunction(F);
1392 }
1393 
1394 unsigned TargetTransformInfo::getMaxNumArgs() const {
1395   return TTIImpl->getMaxNumArgs();
1396 }
1397 
1398 bool TargetTransformInfo::shouldExpandReduction(const IntrinsicInst *II) const {
1399   return TTIImpl->shouldExpandReduction(II);
1400 }
1401 
1402 TargetTransformInfo::ReductionShuffle
1403 TargetTransformInfo::getPreferredExpandedReductionShuffle(
1404     const IntrinsicInst *II) const {
1405   return TTIImpl->getPreferredExpandedReductionShuffle(II);
1406 }
1407 
1408 unsigned TargetTransformInfo::getGISelRematGlobalCost() const {
1409   return TTIImpl->getGISelRematGlobalCost();
1410 }
1411 
1412 unsigned TargetTransformInfo::getMinTripCountTailFoldingThreshold() const {
1413   return TTIImpl->getMinTripCountTailFoldingThreshold();
1414 }
1415 
1416 bool TargetTransformInfo::supportsScalableVectors() const {
1417   return TTIImpl->supportsScalableVectors();
1418 }
1419 
1420 bool TargetTransformInfo::enableScalableVectorization() const {
1421   return TTIImpl->enableScalableVectorization();
1422 }
1423 
1424 bool TargetTransformInfo::hasActiveVectorLength(unsigned Opcode, Type *DataType,
1425                                                 Align Alignment) const {
1426   return TTIImpl->hasActiveVectorLength(Opcode, DataType, Alignment);
1427 }
1428 
1429 bool TargetTransformInfo::isProfitableToSinkOperands(
1430     Instruction *I, SmallVectorImpl<Use *> &OpsToSink) const {
1431   return TTIImpl->isProfitableToSinkOperands(I, OpsToSink);
1432 }
1433 
1434 bool TargetTransformInfo::isVectorShiftByScalarCheap(Type *Ty) const {
1435   return TTIImpl->isVectorShiftByScalarCheap(Ty);
1436 }
1437 
1438 unsigned
1439 TargetTransformInfo::getNumBytesToPadGlobalArray(unsigned Size,
1440                                                  Type *ArrayType) const {
1441   return TTIImpl->getNumBytesToPadGlobalArray(Size, ArrayType);
1442 }
1443 
1444 void TargetTransformInfo::collectKernelLaunchBounds(
1445     const Function &F,
1446     SmallVectorImpl<std::pair<StringRef, int64_t>> &LB) const {
1447   return TTIImpl->collectKernelLaunchBounds(F, LB);
1448 }
1449 
1450 TargetTransformInfo::Concept::~Concept() = default;
1451 
1452 TargetIRAnalysis::TargetIRAnalysis() : TTICallback(&getDefaultTTI) {}
1453 
1454 TargetIRAnalysis::TargetIRAnalysis(
1455     std::function<Result(const Function &)> TTICallback)
1456     : TTICallback(std::move(TTICallback)) {}
1457 
1458 TargetIRAnalysis::Result TargetIRAnalysis::run(const Function &F,
1459                                                FunctionAnalysisManager &) {
1460   return TTICallback(F);
1461 }
1462 
1463 AnalysisKey TargetIRAnalysis::Key;
1464 
1465 TargetIRAnalysis::Result TargetIRAnalysis::getDefaultTTI(const Function &F) {
1466   return Result(F.getDataLayout());
1467 }
1468 
1469 // Register the basic pass.
1470 INITIALIZE_PASS(TargetTransformInfoWrapperPass, "tti",
1471                 "Target Transform Information", false, true)
1472 char TargetTransformInfoWrapperPass::ID = 0;
1473 
1474 void TargetTransformInfoWrapperPass::anchor() {}
1475 
1476 TargetTransformInfoWrapperPass::TargetTransformInfoWrapperPass()
1477     : ImmutablePass(ID) {
1478   initializeTargetTransformInfoWrapperPassPass(
1479       *PassRegistry::getPassRegistry());
1480 }
1481 
1482 TargetTransformInfoWrapperPass::TargetTransformInfoWrapperPass(
1483     TargetIRAnalysis TIRA)
1484     : ImmutablePass(ID), TIRA(std::move(TIRA)) {
1485   initializeTargetTransformInfoWrapperPassPass(
1486       *PassRegistry::getPassRegistry());
1487 }
1488 
1489 TargetTransformInfo &TargetTransformInfoWrapperPass::getTTI(const Function &F) {
1490   FunctionAnalysisManager DummyFAM;
1491   TTI = TIRA.run(F, DummyFAM);
1492   return *TTI;
1493 }
1494 
1495 ImmutablePass *
1496 llvm::createTargetTransformInfoWrapperPass(TargetIRAnalysis TIRA) {
1497   return new TargetTransformInfoWrapperPass(std::move(TIRA));
1498 }
1499