1 //===- llvm/Analysis/TargetTransformInfo.cpp ------------------------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 
9 #include "llvm/Analysis/TargetTransformInfo.h"
10 #include "llvm/Analysis/CFG.h"
11 #include "llvm/Analysis/LoopIterator.h"
12 #include "llvm/Analysis/TargetTransformInfoImpl.h"
13 #include "llvm/IR/CFG.h"
14 #include "llvm/IR/DataLayout.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/IR/PatternMatch.h"
22 #include "llvm/InitializePasses.h"
23 #include "llvm/Support/CommandLine.h"
24 #include "llvm/Support/ErrorHandling.h"
25 #include <utility>
26 
27 using namespace llvm;
28 using namespace PatternMatch;
29 
30 #define DEBUG_TYPE "tti"
31 
32 static cl::opt<bool> EnableReduxCost("costmodel-reduxcost", cl::init(false),
33                                      cl::Hidden,
34                                      cl::desc("Recognize reduction patterns."));
35 
36 namespace {
37 /// No-op implementation of the TTI interface using the utility base
38 /// classes.
39 ///
40 /// This is used when no target specific information is available.
41 struct NoTTIImpl : TargetTransformInfoImplCRTPBase<NoTTIImpl> {
42   explicit NoTTIImpl(const DataLayout &DL)
43       : TargetTransformInfoImplCRTPBase<NoTTIImpl>(DL) {}
44 };
45 } // namespace
46 
47 bool HardwareLoopInfo::canAnalyze(LoopInfo &LI) {
48   // If the loop has irreducible control flow, it can not be converted to
49   // Hardware loop.
50   LoopBlocksRPO RPOT(L);
51   RPOT.perform(&LI);
52   if (containsIrreducibleCFG<const BasicBlock *>(RPOT, LI))
53     return false;
54   return true;
55 }
56 
57 IntrinsicCostAttributes::IntrinsicCostAttributes(const IntrinsicInst &I) :
58     II(&I), RetTy(I.getType()), IID(I.getIntrinsicID()) {
59 
60  FunctionType *FTy = I.getCalledFunction()->getFunctionType();
61  ParamTys.insert(ParamTys.begin(), FTy->param_begin(), FTy->param_end());
62  Arguments.insert(Arguments.begin(), I.arg_begin(), I.arg_end());
63  if (auto *FPMO = dyn_cast<FPMathOperator>(&I))
64    FMF = FPMO->getFastMathFlags();
65 }
66 
67 IntrinsicCostAttributes::IntrinsicCostAttributes(Intrinsic::ID Id,
68                                                  const CallBase &CI) :
69   II(dyn_cast<IntrinsicInst>(&CI)),  RetTy(CI.getType()), IID(Id) {
70 
71   if (const auto *FPMO = dyn_cast<FPMathOperator>(&CI))
72     FMF = FPMO->getFastMathFlags();
73 
74   Arguments.insert(Arguments.begin(), CI.arg_begin(), CI.arg_end());
75   FunctionType *FTy =
76     CI.getCalledFunction()->getFunctionType();
77   ParamTys.insert(ParamTys.begin(), FTy->param_begin(), FTy->param_end());
78 }
79 
80 IntrinsicCostAttributes::IntrinsicCostAttributes(Intrinsic::ID Id,
81                                                  const CallBase &CI,
82                                                  ElementCount Factor)
83     : RetTy(CI.getType()), IID(Id), VF(Factor) {
84 
85   assert(!Factor.isScalable() && "Scalable vectors are not yet supported");
86   if (auto *FPMO = dyn_cast<FPMathOperator>(&CI))
87     FMF = FPMO->getFastMathFlags();
88 
89   Arguments.insert(Arguments.begin(), CI.arg_begin(), CI.arg_end());
90   FunctionType *FTy =
91     CI.getCalledFunction()->getFunctionType();
92   ParamTys.insert(ParamTys.begin(), FTy->param_begin(), FTy->param_end());
93 }
94 
95 IntrinsicCostAttributes::IntrinsicCostAttributes(Intrinsic::ID Id,
96                                                  const CallBase &CI,
97                                                  ElementCount Factor,
98                                                  unsigned ScalarCost)
99     : RetTy(CI.getType()), IID(Id), VF(Factor), ScalarizationCost(ScalarCost) {
100 
101   if (const auto *FPMO = dyn_cast<FPMathOperator>(&CI))
102     FMF = FPMO->getFastMathFlags();
103 
104   Arguments.insert(Arguments.begin(), CI.arg_begin(), CI.arg_end());
105   FunctionType *FTy =
106     CI.getCalledFunction()->getFunctionType();
107   ParamTys.insert(ParamTys.begin(), FTy->param_begin(), FTy->param_end());
108 }
109 
110 IntrinsicCostAttributes::IntrinsicCostAttributes(Intrinsic::ID Id, Type *RTy,
111                                                  ArrayRef<Type *> Tys,
112                                                  FastMathFlags Flags) :
113     RetTy(RTy), IID(Id), FMF(Flags) {
114   ParamTys.insert(ParamTys.begin(), Tys.begin(), Tys.end());
115 }
116 
117 IntrinsicCostAttributes::IntrinsicCostAttributes(Intrinsic::ID Id, Type *RTy,
118                                                  ArrayRef<Type *> Tys,
119                                                  FastMathFlags Flags,
120                                                  unsigned ScalarCost) :
121     RetTy(RTy), IID(Id), FMF(Flags), ScalarizationCost(ScalarCost) {
122   ParamTys.insert(ParamTys.begin(), Tys.begin(), Tys.end());
123 }
124 
125 IntrinsicCostAttributes::IntrinsicCostAttributes(Intrinsic::ID Id, Type *RTy,
126                                                  ArrayRef<Type *> Tys,
127                                                  FastMathFlags Flags,
128                                                  unsigned ScalarCost,
129                                                  const IntrinsicInst *I) :
130     II(I), RetTy(RTy), IID(Id), FMF(Flags), ScalarizationCost(ScalarCost) {
131   ParamTys.insert(ParamTys.begin(), Tys.begin(), Tys.end());
132 }
133 
134 IntrinsicCostAttributes::IntrinsicCostAttributes(Intrinsic::ID Id, Type *RTy,
135                                                  ArrayRef<Type *> Tys) :
136     RetTy(RTy), IID(Id) {
137   ParamTys.insert(ParamTys.begin(), Tys.begin(), Tys.end());
138 }
139 
140 IntrinsicCostAttributes::IntrinsicCostAttributes(Intrinsic::ID Id, Type *Ty,
141                                                  ArrayRef<const Value *> Args)
142     : RetTy(Ty), IID(Id) {
143 
144   Arguments.insert(Arguments.begin(), Args.begin(), Args.end());
145   ParamTys.reserve(Arguments.size());
146   for (unsigned Idx = 0, Size = Arguments.size(); Idx != Size; ++Idx)
147     ParamTys.push_back(Arguments[Idx]->getType());
148 }
149 
150 bool HardwareLoopInfo::isHardwareLoopCandidate(ScalarEvolution &SE,
151                                                LoopInfo &LI, DominatorTree &DT,
152                                                bool ForceNestedLoop,
153                                                bool ForceHardwareLoopPHI) {
154   SmallVector<BasicBlock *, 4> ExitingBlocks;
155   L->getExitingBlocks(ExitingBlocks);
156 
157   for (BasicBlock *BB : ExitingBlocks) {
158     // If we pass the updated counter back through a phi, we need to know
159     // which latch the updated value will be coming from.
160     if (!L->isLoopLatch(BB)) {
161       if (ForceHardwareLoopPHI || CounterInReg)
162         continue;
163     }
164 
165     const SCEV *EC = SE.getExitCount(L, BB);
166     if (isa<SCEVCouldNotCompute>(EC))
167       continue;
168     if (const SCEVConstant *ConstEC = dyn_cast<SCEVConstant>(EC)) {
169       if (ConstEC->getValue()->isZero())
170         continue;
171     } else if (!SE.isLoopInvariant(EC, L))
172       continue;
173 
174     if (SE.getTypeSizeInBits(EC->getType()) > CountType->getBitWidth())
175       continue;
176 
177     // If this exiting block is contained in a nested loop, it is not eligible
178     // for insertion of the branch-and-decrement since the inner loop would
179     // end up messing up the value in the CTR.
180     if (!IsNestingLegal && LI.getLoopFor(BB) != L && !ForceNestedLoop)
181       continue;
182 
183     // We now have a loop-invariant count of loop iterations (which is not the
184     // constant zero) for which we know that this loop will not exit via this
185     // existing block.
186 
187     // We need to make sure that this block will run on every loop iteration.
188     // For this to be true, we must dominate all blocks with backedges. Such
189     // blocks are in-loop predecessors to the header block.
190     bool NotAlways = false;
191     for (BasicBlock *Pred : predecessors(L->getHeader())) {
192       if (!L->contains(Pred))
193         continue;
194 
195       if (!DT.dominates(BB, Pred)) {
196         NotAlways = true;
197         break;
198       }
199     }
200 
201     if (NotAlways)
202       continue;
203 
204     // Make sure this blocks ends with a conditional branch.
205     Instruction *TI = BB->getTerminator();
206     if (!TI)
207       continue;
208 
209     if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
210       if (!BI->isConditional())
211         continue;
212 
213       ExitBranch = BI;
214     } else
215       continue;
216 
217     // Note that this block may not be the loop latch block, even if the loop
218     // has a latch block.
219     ExitBlock = BB;
220     TripCount = SE.getAddExpr(EC, SE.getOne(EC->getType()));
221 
222     if (!EC->getType()->isPointerTy() && EC->getType() != CountType)
223       TripCount = SE.getZeroExtendExpr(TripCount, CountType);
224 
225     break;
226   }
227 
228   if (!ExitBlock)
229     return false;
230   return true;
231 }
232 
233 TargetTransformInfo::TargetTransformInfo(const DataLayout &DL)
234     : TTIImpl(new Model<NoTTIImpl>(NoTTIImpl(DL))) {}
235 
236 TargetTransformInfo::~TargetTransformInfo() {}
237 
238 TargetTransformInfo::TargetTransformInfo(TargetTransformInfo &&Arg)
239     : TTIImpl(std::move(Arg.TTIImpl)) {}
240 
241 TargetTransformInfo &TargetTransformInfo::operator=(TargetTransformInfo &&RHS) {
242   TTIImpl = std::move(RHS.TTIImpl);
243   return *this;
244 }
245 
246 unsigned TargetTransformInfo::getInliningThresholdMultiplier() const {
247   return TTIImpl->getInliningThresholdMultiplier();
248 }
249 
250 int TargetTransformInfo::getInlinerVectorBonusPercent() const {
251   return TTIImpl->getInlinerVectorBonusPercent();
252 }
253 
254 int TargetTransformInfo::getGEPCost(Type *PointeeType, const Value *Ptr,
255                                     ArrayRef<const Value *> Operands,
256                                     TTI::TargetCostKind CostKind) const {
257   return TTIImpl->getGEPCost(PointeeType, Ptr, Operands, CostKind);
258 }
259 
260 unsigned TargetTransformInfo::getEstimatedNumberOfCaseClusters(
261     const SwitchInst &SI, unsigned &JTSize, ProfileSummaryInfo *PSI,
262     BlockFrequencyInfo *BFI) const {
263   return TTIImpl->getEstimatedNumberOfCaseClusters(SI, JTSize, PSI, BFI);
264 }
265 
266 int TargetTransformInfo::getUserCost(const User *U,
267                                      ArrayRef<const Value *> Operands,
268                                      enum TargetCostKind CostKind) const {
269   int Cost = TTIImpl->getUserCost(U, Operands, CostKind);
270   assert((CostKind == TTI::TCK_RecipThroughput || Cost >= 0) &&
271          "TTI should not produce negative costs!");
272   return Cost;
273 }
274 
275 bool TargetTransformInfo::hasBranchDivergence() const {
276   return TTIImpl->hasBranchDivergence();
277 }
278 
279 bool TargetTransformInfo::useGPUDivergenceAnalysis() const {
280   return TTIImpl->useGPUDivergenceAnalysis();
281 }
282 
283 bool TargetTransformInfo::isSourceOfDivergence(const Value *V) const {
284   return TTIImpl->isSourceOfDivergence(V);
285 }
286 
287 bool llvm::TargetTransformInfo::isAlwaysUniform(const Value *V) const {
288   return TTIImpl->isAlwaysUniform(V);
289 }
290 
291 unsigned TargetTransformInfo::getFlatAddressSpace() const {
292   return TTIImpl->getFlatAddressSpace();
293 }
294 
295 bool TargetTransformInfo::collectFlatAddressOperands(
296     SmallVectorImpl<int> &OpIndexes, Intrinsic::ID IID) const {
297   return TTIImpl->collectFlatAddressOperands(OpIndexes, IID);
298 }
299 
300 bool TargetTransformInfo::isNoopAddrSpaceCast(unsigned FromAS,
301                                               unsigned ToAS) const {
302   return TTIImpl->isNoopAddrSpaceCast(FromAS, ToAS);
303 }
304 
305 unsigned TargetTransformInfo::getAssumedAddrSpace(const Value *V) const {
306   return TTIImpl->getAssumedAddrSpace(V);
307 }
308 
309 Value *TargetTransformInfo::rewriteIntrinsicWithAddressSpace(
310     IntrinsicInst *II, Value *OldV, Value *NewV) const {
311   return TTIImpl->rewriteIntrinsicWithAddressSpace(II, OldV, NewV);
312 }
313 
314 bool TargetTransformInfo::isLoweredToCall(const Function *F) const {
315   return TTIImpl->isLoweredToCall(F);
316 }
317 
318 bool TargetTransformInfo::isHardwareLoopProfitable(
319     Loop *L, ScalarEvolution &SE, AssumptionCache &AC,
320     TargetLibraryInfo *LibInfo, HardwareLoopInfo &HWLoopInfo) const {
321   return TTIImpl->isHardwareLoopProfitable(L, SE, AC, LibInfo, HWLoopInfo);
322 }
323 
324 bool TargetTransformInfo::preferPredicateOverEpilogue(
325     Loop *L, LoopInfo *LI, ScalarEvolution &SE, AssumptionCache &AC,
326     TargetLibraryInfo *TLI, DominatorTree *DT,
327     const LoopAccessInfo *LAI) const {
328   return TTIImpl->preferPredicateOverEpilogue(L, LI, SE, AC, TLI, DT, LAI);
329 }
330 
331 bool TargetTransformInfo::emitGetActiveLaneMask() const {
332   return TTIImpl->emitGetActiveLaneMask();
333 }
334 
335 Optional<Instruction *>
336 TargetTransformInfo::instCombineIntrinsic(InstCombiner &IC,
337                                           IntrinsicInst &II) const {
338   return TTIImpl->instCombineIntrinsic(IC, II);
339 }
340 
341 Optional<Value *> TargetTransformInfo::simplifyDemandedUseBitsIntrinsic(
342     InstCombiner &IC, IntrinsicInst &II, APInt DemandedMask, KnownBits &Known,
343     bool &KnownBitsComputed) const {
344   return TTIImpl->simplifyDemandedUseBitsIntrinsic(IC, II, DemandedMask, Known,
345                                                    KnownBitsComputed);
346 }
347 
348 Optional<Value *> TargetTransformInfo::simplifyDemandedVectorEltsIntrinsic(
349     InstCombiner &IC, IntrinsicInst &II, APInt DemandedElts, APInt &UndefElts,
350     APInt &UndefElts2, APInt &UndefElts3,
351     std::function<void(Instruction *, unsigned, APInt, APInt &)>
352         SimplifyAndSetOp) const {
353   return TTIImpl->simplifyDemandedVectorEltsIntrinsic(
354       IC, II, DemandedElts, UndefElts, UndefElts2, UndefElts3,
355       SimplifyAndSetOp);
356 }
357 
358 void TargetTransformInfo::getUnrollingPreferences(
359     Loop *L, ScalarEvolution &SE, UnrollingPreferences &UP) const {
360   return TTIImpl->getUnrollingPreferences(L, SE, UP);
361 }
362 
363 void TargetTransformInfo::getPeelingPreferences(Loop *L, ScalarEvolution &SE,
364                                                 PeelingPreferences &PP) const {
365   return TTIImpl->getPeelingPreferences(L, SE, PP);
366 }
367 
368 bool TargetTransformInfo::isLegalAddImmediate(int64_t Imm) const {
369   return TTIImpl->isLegalAddImmediate(Imm);
370 }
371 
372 bool TargetTransformInfo::isLegalICmpImmediate(int64_t Imm) const {
373   return TTIImpl->isLegalICmpImmediate(Imm);
374 }
375 
376 bool TargetTransformInfo::isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV,
377                                                 int64_t BaseOffset,
378                                                 bool HasBaseReg, int64_t Scale,
379                                                 unsigned AddrSpace,
380                                                 Instruction *I) const {
381   return TTIImpl->isLegalAddressingMode(Ty, BaseGV, BaseOffset, HasBaseReg,
382                                         Scale, AddrSpace, I);
383 }
384 
385 bool TargetTransformInfo::isLSRCostLess(LSRCost &C1, LSRCost &C2) const {
386   return TTIImpl->isLSRCostLess(C1, C2);
387 }
388 
389 bool TargetTransformInfo::isNumRegsMajorCostOfLSR() const {
390   return TTIImpl->isNumRegsMajorCostOfLSR();
391 }
392 
393 bool TargetTransformInfo::isProfitableLSRChainElement(Instruction *I) const {
394   return TTIImpl->isProfitableLSRChainElement(I);
395 }
396 
397 bool TargetTransformInfo::canMacroFuseCmp() const {
398   return TTIImpl->canMacroFuseCmp();
399 }
400 
401 bool TargetTransformInfo::canSaveCmp(Loop *L, BranchInst **BI,
402                                      ScalarEvolution *SE, LoopInfo *LI,
403                                      DominatorTree *DT, AssumptionCache *AC,
404                                      TargetLibraryInfo *LibInfo) const {
405   return TTIImpl->canSaveCmp(L, BI, SE, LI, DT, AC, LibInfo);
406 }
407 
408 bool TargetTransformInfo::shouldFavorPostInc() const {
409   return TTIImpl->shouldFavorPostInc();
410 }
411 
412 bool TargetTransformInfo::shouldFavorBackedgeIndex(const Loop *L) const {
413   return TTIImpl->shouldFavorBackedgeIndex(L);
414 }
415 
416 bool TargetTransformInfo::isLegalMaskedStore(Type *DataType,
417                                              Align Alignment) const {
418   return TTIImpl->isLegalMaskedStore(DataType, Alignment);
419 }
420 
421 bool TargetTransformInfo::isLegalMaskedLoad(Type *DataType,
422                                             Align Alignment) const {
423   return TTIImpl->isLegalMaskedLoad(DataType, Alignment);
424 }
425 
426 bool TargetTransformInfo::isLegalNTStore(Type *DataType,
427                                          Align Alignment) const {
428   return TTIImpl->isLegalNTStore(DataType, Alignment);
429 }
430 
431 bool TargetTransformInfo::isLegalNTLoad(Type *DataType, Align Alignment) const {
432   return TTIImpl->isLegalNTLoad(DataType, Alignment);
433 }
434 
435 bool TargetTransformInfo::isLegalMaskedGather(Type *DataType,
436                                               Align Alignment) const {
437   return TTIImpl->isLegalMaskedGather(DataType, Alignment);
438 }
439 
440 bool TargetTransformInfo::isLegalMaskedScatter(Type *DataType,
441                                                Align Alignment) const {
442   return TTIImpl->isLegalMaskedScatter(DataType, Alignment);
443 }
444 
445 bool TargetTransformInfo::isLegalMaskedCompressStore(Type *DataType) const {
446   return TTIImpl->isLegalMaskedCompressStore(DataType);
447 }
448 
449 bool TargetTransformInfo::isLegalMaskedExpandLoad(Type *DataType) const {
450   return TTIImpl->isLegalMaskedExpandLoad(DataType);
451 }
452 
453 bool TargetTransformInfo::hasDivRemOp(Type *DataType, bool IsSigned) const {
454   return TTIImpl->hasDivRemOp(DataType, IsSigned);
455 }
456 
457 bool TargetTransformInfo::hasVolatileVariant(Instruction *I,
458                                              unsigned AddrSpace) const {
459   return TTIImpl->hasVolatileVariant(I, AddrSpace);
460 }
461 
462 bool TargetTransformInfo::prefersVectorizedAddressing() const {
463   return TTIImpl->prefersVectorizedAddressing();
464 }
465 
466 int TargetTransformInfo::getScalingFactorCost(Type *Ty, GlobalValue *BaseGV,
467                                               int64_t BaseOffset,
468                                               bool HasBaseReg, int64_t Scale,
469                                               unsigned AddrSpace) const {
470   int Cost = TTIImpl->getScalingFactorCost(Ty, BaseGV, BaseOffset, HasBaseReg,
471                                            Scale, AddrSpace);
472   assert(Cost >= 0 && "TTI should not produce negative costs!");
473   return Cost;
474 }
475 
476 bool TargetTransformInfo::LSRWithInstrQueries() const {
477   return TTIImpl->LSRWithInstrQueries();
478 }
479 
480 bool TargetTransformInfo::isTruncateFree(Type *Ty1, Type *Ty2) const {
481   return TTIImpl->isTruncateFree(Ty1, Ty2);
482 }
483 
484 bool TargetTransformInfo::isProfitableToHoist(Instruction *I) const {
485   return TTIImpl->isProfitableToHoist(I);
486 }
487 
488 bool TargetTransformInfo::useAA() const { return TTIImpl->useAA(); }
489 
490 bool TargetTransformInfo::isTypeLegal(Type *Ty) const {
491   return TTIImpl->isTypeLegal(Ty);
492 }
493 
494 unsigned TargetTransformInfo::getRegUsageForType(Type *Ty) const {
495   return TTIImpl->getRegUsageForType(Ty);
496 }
497 
498 bool TargetTransformInfo::shouldBuildLookupTables() const {
499   return TTIImpl->shouldBuildLookupTables();
500 }
501 bool TargetTransformInfo::shouldBuildLookupTablesForConstant(
502     Constant *C) const {
503   return TTIImpl->shouldBuildLookupTablesForConstant(C);
504 }
505 
506 bool TargetTransformInfo::useColdCCForColdCall(Function &F) const {
507   return TTIImpl->useColdCCForColdCall(F);
508 }
509 
510 unsigned
511 TargetTransformInfo::getScalarizationOverhead(VectorType *Ty,
512                                               const APInt &DemandedElts,
513                                               bool Insert, bool Extract) const {
514   return TTIImpl->getScalarizationOverhead(Ty, DemandedElts, Insert, Extract);
515 }
516 
517 unsigned TargetTransformInfo::getOperandsScalarizationOverhead(
518     ArrayRef<const Value *> Args, unsigned VF) const {
519   return TTIImpl->getOperandsScalarizationOverhead(Args, VF);
520 }
521 
522 bool TargetTransformInfo::supportsEfficientVectorElementLoadStore() const {
523   return TTIImpl->supportsEfficientVectorElementLoadStore();
524 }
525 
526 bool TargetTransformInfo::enableAggressiveInterleaving(
527     bool LoopHasReductions) const {
528   return TTIImpl->enableAggressiveInterleaving(LoopHasReductions);
529 }
530 
531 TargetTransformInfo::MemCmpExpansionOptions
532 TargetTransformInfo::enableMemCmpExpansion(bool OptSize, bool IsZeroCmp) const {
533   return TTIImpl->enableMemCmpExpansion(OptSize, IsZeroCmp);
534 }
535 
536 bool TargetTransformInfo::enableInterleavedAccessVectorization() const {
537   return TTIImpl->enableInterleavedAccessVectorization();
538 }
539 
540 bool TargetTransformInfo::enableMaskedInterleavedAccessVectorization() const {
541   return TTIImpl->enableMaskedInterleavedAccessVectorization();
542 }
543 
544 bool TargetTransformInfo::isFPVectorizationPotentiallyUnsafe() const {
545   return TTIImpl->isFPVectorizationPotentiallyUnsafe();
546 }
547 
548 bool TargetTransformInfo::allowsMisalignedMemoryAccesses(LLVMContext &Context,
549                                                          unsigned BitWidth,
550                                                          unsigned AddressSpace,
551                                                          unsigned Alignment,
552                                                          bool *Fast) const {
553   return TTIImpl->allowsMisalignedMemoryAccesses(Context, BitWidth,
554                                                  AddressSpace, Alignment, Fast);
555 }
556 
557 TargetTransformInfo::PopcntSupportKind
558 TargetTransformInfo::getPopcntSupport(unsigned IntTyWidthInBit) const {
559   return TTIImpl->getPopcntSupport(IntTyWidthInBit);
560 }
561 
562 bool TargetTransformInfo::haveFastSqrt(Type *Ty) const {
563   return TTIImpl->haveFastSqrt(Ty);
564 }
565 
566 bool TargetTransformInfo::isFCmpOrdCheaperThanFCmpZero(Type *Ty) const {
567   return TTIImpl->isFCmpOrdCheaperThanFCmpZero(Ty);
568 }
569 
570 int TargetTransformInfo::getFPOpCost(Type *Ty) const {
571   int Cost = TTIImpl->getFPOpCost(Ty);
572   assert(Cost >= 0 && "TTI should not produce negative costs!");
573   return Cost;
574 }
575 
576 int TargetTransformInfo::getIntImmCodeSizeCost(unsigned Opcode, unsigned Idx,
577                                                const APInt &Imm,
578                                                Type *Ty) const {
579   int Cost = TTIImpl->getIntImmCodeSizeCost(Opcode, Idx, Imm, Ty);
580   assert(Cost >= 0 && "TTI should not produce negative costs!");
581   return Cost;
582 }
583 
584 int TargetTransformInfo::getIntImmCost(const APInt &Imm, Type *Ty,
585                                        TTI::TargetCostKind CostKind) const {
586   int Cost = TTIImpl->getIntImmCost(Imm, Ty, CostKind);
587   assert(Cost >= 0 && "TTI should not produce negative costs!");
588   return Cost;
589 }
590 
591 int TargetTransformInfo::getIntImmCostInst(unsigned Opcode, unsigned Idx,
592                                            const APInt &Imm, Type *Ty,
593                                            TTI::TargetCostKind CostKind,
594                                            Instruction *Inst) const {
595   int Cost = TTIImpl->getIntImmCostInst(Opcode, Idx, Imm, Ty, CostKind, Inst);
596   assert(Cost >= 0 && "TTI should not produce negative costs!");
597   return Cost;
598 }
599 
600 int
601 TargetTransformInfo::getIntImmCostIntrin(Intrinsic::ID IID, unsigned Idx,
602                                          const APInt &Imm, Type *Ty,
603                                          TTI::TargetCostKind CostKind) const {
604   int Cost = TTIImpl->getIntImmCostIntrin(IID, Idx, Imm, Ty, CostKind);
605   assert(Cost >= 0 && "TTI should not produce negative costs!");
606   return Cost;
607 }
608 
609 unsigned TargetTransformInfo::getNumberOfRegisters(unsigned ClassID) const {
610   return TTIImpl->getNumberOfRegisters(ClassID);
611 }
612 
613 unsigned TargetTransformInfo::getRegisterClassForType(bool Vector,
614                                                       Type *Ty) const {
615   return TTIImpl->getRegisterClassForType(Vector, Ty);
616 }
617 
618 const char *TargetTransformInfo::getRegisterClassName(unsigned ClassID) const {
619   return TTIImpl->getRegisterClassName(ClassID);
620 }
621 
622 unsigned TargetTransformInfo::getRegisterBitWidth(bool Vector) const {
623   return TTIImpl->getRegisterBitWidth(Vector);
624 }
625 
626 unsigned TargetTransformInfo::getMinVectorRegisterBitWidth() const {
627   return TTIImpl->getMinVectorRegisterBitWidth();
628 }
629 
630 Optional<unsigned> TargetTransformInfo::getMaxVScale() const {
631   return TTIImpl->getMaxVScale();
632 }
633 
634 bool TargetTransformInfo::shouldMaximizeVectorBandwidth(bool OptSize) const {
635   return TTIImpl->shouldMaximizeVectorBandwidth(OptSize);
636 }
637 
638 unsigned TargetTransformInfo::getMinimumVF(unsigned ElemWidth) const {
639   return TTIImpl->getMinimumVF(ElemWidth);
640 }
641 
642 unsigned TargetTransformInfo::getMaximumVF(unsigned ElemWidth,
643                                            unsigned Opcode) const {
644   return TTIImpl->getMaximumVF(ElemWidth, Opcode);
645 }
646 
647 bool TargetTransformInfo::shouldConsiderAddressTypePromotion(
648     const Instruction &I, bool &AllowPromotionWithoutCommonHeader) const {
649   return TTIImpl->shouldConsiderAddressTypePromotion(
650       I, AllowPromotionWithoutCommonHeader);
651 }
652 
653 unsigned TargetTransformInfo::getCacheLineSize() const {
654   return TTIImpl->getCacheLineSize();
655 }
656 
657 llvm::Optional<unsigned>
658 TargetTransformInfo::getCacheSize(CacheLevel Level) const {
659   return TTIImpl->getCacheSize(Level);
660 }
661 
662 llvm::Optional<unsigned>
663 TargetTransformInfo::getCacheAssociativity(CacheLevel Level) const {
664   return TTIImpl->getCacheAssociativity(Level);
665 }
666 
667 unsigned TargetTransformInfo::getPrefetchDistance() const {
668   return TTIImpl->getPrefetchDistance();
669 }
670 
671 unsigned TargetTransformInfo::getMinPrefetchStride(
672     unsigned NumMemAccesses, unsigned NumStridedMemAccesses,
673     unsigned NumPrefetches, bool HasCall) const {
674   return TTIImpl->getMinPrefetchStride(NumMemAccesses, NumStridedMemAccesses,
675                                        NumPrefetches, HasCall);
676 }
677 
678 unsigned TargetTransformInfo::getMaxPrefetchIterationsAhead() const {
679   return TTIImpl->getMaxPrefetchIterationsAhead();
680 }
681 
682 bool TargetTransformInfo::enableWritePrefetching() const {
683   return TTIImpl->enableWritePrefetching();
684 }
685 
686 unsigned TargetTransformInfo::getMaxInterleaveFactor(unsigned VF) const {
687   return TTIImpl->getMaxInterleaveFactor(VF);
688 }
689 
690 TargetTransformInfo::OperandValueKind
691 TargetTransformInfo::getOperandInfo(const Value *V,
692                                     OperandValueProperties &OpProps) {
693   OperandValueKind OpInfo = OK_AnyValue;
694   OpProps = OP_None;
695 
696   if (const auto *CI = dyn_cast<ConstantInt>(V)) {
697     if (CI->getValue().isPowerOf2())
698       OpProps = OP_PowerOf2;
699     return OK_UniformConstantValue;
700   }
701 
702   // A broadcast shuffle creates a uniform value.
703   // TODO: Add support for non-zero index broadcasts.
704   // TODO: Add support for different source vector width.
705   if (const auto *ShuffleInst = dyn_cast<ShuffleVectorInst>(V))
706     if (ShuffleInst->isZeroEltSplat())
707       OpInfo = OK_UniformValue;
708 
709   const Value *Splat = getSplatValue(V);
710 
711   // Check for a splat of a constant or for a non uniform vector of constants
712   // and check if the constant(s) are all powers of two.
713   if (isa<ConstantVector>(V) || isa<ConstantDataVector>(V)) {
714     OpInfo = OK_NonUniformConstantValue;
715     if (Splat) {
716       OpInfo = OK_UniformConstantValue;
717       if (auto *CI = dyn_cast<ConstantInt>(Splat))
718         if (CI->getValue().isPowerOf2())
719           OpProps = OP_PowerOf2;
720     } else if (const auto *CDS = dyn_cast<ConstantDataSequential>(V)) {
721       OpProps = OP_PowerOf2;
722       for (unsigned I = 0, E = CDS->getNumElements(); I != E; ++I) {
723         if (auto *CI = dyn_cast<ConstantInt>(CDS->getElementAsConstant(I)))
724           if (CI->getValue().isPowerOf2())
725             continue;
726         OpProps = OP_None;
727         break;
728       }
729     }
730   }
731 
732   // Check for a splat of a uniform value. This is not loop aware, so return
733   // true only for the obviously uniform cases (argument, globalvalue)
734   if (Splat && (isa<Argument>(Splat) || isa<GlobalValue>(Splat)))
735     OpInfo = OK_UniformValue;
736 
737   return OpInfo;
738 }
739 
740 int TargetTransformInfo::getArithmeticInstrCost(
741     unsigned Opcode, Type *Ty, TTI::TargetCostKind CostKind,
742     OperandValueKind Opd1Info,
743     OperandValueKind Opd2Info, OperandValueProperties Opd1PropInfo,
744     OperandValueProperties Opd2PropInfo, ArrayRef<const Value *> Args,
745     const Instruction *CxtI) const {
746   int Cost = TTIImpl->getArithmeticInstrCost(
747       Opcode, Ty, CostKind, Opd1Info, Opd2Info, Opd1PropInfo, Opd2PropInfo,
748       Args, CxtI);
749   assert(Cost >= 0 && "TTI should not produce negative costs!");
750   return Cost;
751 }
752 
753 int TargetTransformInfo::getShuffleCost(ShuffleKind Kind, VectorType *Ty,
754                                         int Index, VectorType *SubTp) const {
755   int Cost = TTIImpl->getShuffleCost(Kind, Ty, Index, SubTp);
756   assert(Cost >= 0 && "TTI should not produce negative costs!");
757   return Cost;
758 }
759 
760 TTI::CastContextHint
761 TargetTransformInfo::getCastContextHint(const Instruction *I) {
762   if (!I)
763     return CastContextHint::None;
764 
765   auto getLoadStoreKind = [](const Value *V, unsigned LdStOp, unsigned MaskedOp,
766                              unsigned GatScatOp) {
767     const Instruction *I = dyn_cast<Instruction>(V);
768     if (!I)
769       return CastContextHint::None;
770 
771     if (I->getOpcode() == LdStOp)
772       return CastContextHint::Normal;
773 
774     if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
775       if (II->getIntrinsicID() == MaskedOp)
776         return TTI::CastContextHint::Masked;
777       if (II->getIntrinsicID() == GatScatOp)
778         return TTI::CastContextHint::GatherScatter;
779     }
780 
781     return TTI::CastContextHint::None;
782   };
783 
784   switch (I->getOpcode()) {
785   case Instruction::ZExt:
786   case Instruction::SExt:
787   case Instruction::FPExt:
788     return getLoadStoreKind(I->getOperand(0), Instruction::Load,
789                             Intrinsic::masked_load, Intrinsic::masked_gather);
790   case Instruction::Trunc:
791   case Instruction::FPTrunc:
792     if (I->hasOneUse())
793       return getLoadStoreKind(*I->user_begin(), Instruction::Store,
794                               Intrinsic::masked_store,
795                               Intrinsic::masked_scatter);
796     break;
797   default:
798     return CastContextHint::None;
799   }
800 
801   return TTI::CastContextHint::None;
802 }
803 
804 int TargetTransformInfo::getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src,
805                                           CastContextHint CCH,
806                                           TTI::TargetCostKind CostKind,
807                                           const Instruction *I) const {
808   assert((I == nullptr || I->getOpcode() == Opcode) &&
809          "Opcode should reflect passed instruction.");
810   int Cost = TTIImpl->getCastInstrCost(Opcode, Dst, Src, CCH, CostKind, I);
811   assert(Cost >= 0 && "TTI should not produce negative costs!");
812   return Cost;
813 }
814 
815 int TargetTransformInfo::getExtractWithExtendCost(unsigned Opcode, Type *Dst,
816                                                   VectorType *VecTy,
817                                                   unsigned Index) const {
818   int Cost = TTIImpl->getExtractWithExtendCost(Opcode, Dst, VecTy, Index);
819   assert(Cost >= 0 && "TTI should not produce negative costs!");
820   return Cost;
821 }
822 
823 int TargetTransformInfo::getCFInstrCost(unsigned Opcode,
824                                         TTI::TargetCostKind CostKind) const {
825   int Cost = TTIImpl->getCFInstrCost(Opcode, CostKind);
826   assert(Cost >= 0 && "TTI should not produce negative costs!");
827   return Cost;
828 }
829 
830 int TargetTransformInfo::getCmpSelInstrCost(unsigned Opcode, Type *ValTy,
831                                             Type *CondTy,
832                                             CmpInst::Predicate VecPred,
833                                             TTI::TargetCostKind CostKind,
834                                             const Instruction *I) const {
835   assert((I == nullptr || I->getOpcode() == Opcode) &&
836          "Opcode should reflect passed instruction.");
837   int Cost =
838       TTIImpl->getCmpSelInstrCost(Opcode, ValTy, CondTy, VecPred, CostKind, I);
839   assert(Cost >= 0 && "TTI should not produce negative costs!");
840   return Cost;
841 }
842 
843 int TargetTransformInfo::getVectorInstrCost(unsigned Opcode, Type *Val,
844                                             unsigned Index) const {
845   int Cost = TTIImpl->getVectorInstrCost(Opcode, Val, Index);
846   assert(Cost >= 0 && "TTI should not produce negative costs!");
847   return Cost;
848 }
849 
850 int TargetTransformInfo::getMemoryOpCost(unsigned Opcode, Type *Src,
851                                          Align Alignment, unsigned AddressSpace,
852                                          TTI::TargetCostKind CostKind,
853                                          const Instruction *I) const {
854   assert((I == nullptr || I->getOpcode() == Opcode) &&
855          "Opcode should reflect passed instruction.");
856   int Cost = TTIImpl->getMemoryOpCost(Opcode, Src, Alignment, AddressSpace,
857                                       CostKind, I);
858   assert(Cost >= 0 && "TTI should not produce negative costs!");
859   return Cost;
860 }
861 
862 int TargetTransformInfo::getMaskedMemoryOpCost(
863     unsigned Opcode, Type *Src, Align Alignment, unsigned AddressSpace,
864     TTI::TargetCostKind CostKind) const {
865   int Cost =
866       TTIImpl->getMaskedMemoryOpCost(Opcode, Src, Alignment, AddressSpace,
867                                      CostKind);
868   assert(Cost >= 0 && "TTI should not produce negative costs!");
869   return Cost;
870 }
871 
872 int TargetTransformInfo::getGatherScatterOpCost(
873     unsigned Opcode, Type *DataTy, const Value *Ptr, bool VariableMask,
874     Align Alignment, TTI::TargetCostKind CostKind, const Instruction *I) const {
875   int Cost = TTIImpl->getGatherScatterOpCost(Opcode, DataTy, Ptr, VariableMask,
876                                              Alignment, CostKind, I);
877   assert(Cost >= 0 && "TTI should not produce negative costs!");
878   return Cost;
879 }
880 
881 int TargetTransformInfo::getInterleavedMemoryOpCost(
882     unsigned Opcode, Type *VecTy, unsigned Factor, ArrayRef<unsigned> Indices,
883     Align Alignment, unsigned AddressSpace, TTI::TargetCostKind CostKind,
884     bool UseMaskForCond, bool UseMaskForGaps) const {
885   int Cost = TTIImpl->getInterleavedMemoryOpCost(
886       Opcode, VecTy, Factor, Indices, Alignment, AddressSpace, CostKind,
887       UseMaskForCond, UseMaskForGaps);
888   assert(Cost >= 0 && "TTI should not produce negative costs!");
889   return Cost;
890 }
891 
892 int
893 TargetTransformInfo::getIntrinsicInstrCost(const IntrinsicCostAttributes &ICA,
894                                            TTI::TargetCostKind CostKind) const {
895   int Cost = TTIImpl->getIntrinsicInstrCost(ICA, CostKind);
896   assert(Cost >= 0 && "TTI should not produce negative costs!");
897   return Cost;
898 }
899 
900 int TargetTransformInfo::getCallInstrCost(Function *F, Type *RetTy,
901                                           ArrayRef<Type *> Tys,
902                                           TTI::TargetCostKind CostKind) const {
903   int Cost = TTIImpl->getCallInstrCost(F, RetTy, Tys, CostKind);
904   assert(Cost >= 0 && "TTI should not produce negative costs!");
905   return Cost;
906 }
907 
908 unsigned TargetTransformInfo::getNumberOfParts(Type *Tp) const {
909   return TTIImpl->getNumberOfParts(Tp);
910 }
911 
912 int TargetTransformInfo::getAddressComputationCost(Type *Tp,
913                                                    ScalarEvolution *SE,
914                                                    const SCEV *Ptr) const {
915   int Cost = TTIImpl->getAddressComputationCost(Tp, SE, Ptr);
916   assert(Cost >= 0 && "TTI should not produce negative costs!");
917   return Cost;
918 }
919 
920 int TargetTransformInfo::getMemcpyCost(const Instruction *I) const {
921   int Cost = TTIImpl->getMemcpyCost(I);
922   assert(Cost >= 0 && "TTI should not produce negative costs!");
923   return Cost;
924 }
925 
926 int TargetTransformInfo::getArithmeticReductionCost(unsigned Opcode,
927                                                     VectorType *Ty,
928                                                     bool IsPairwiseForm,
929                                                     TTI::TargetCostKind CostKind) const {
930   int Cost = TTIImpl->getArithmeticReductionCost(Opcode, Ty, IsPairwiseForm,
931                                                  CostKind);
932   assert(Cost >= 0 && "TTI should not produce negative costs!");
933   return Cost;
934 }
935 
936 int TargetTransformInfo::getMinMaxReductionCost(
937     VectorType *Ty, VectorType *CondTy, bool IsPairwiseForm, bool IsUnsigned,
938     TTI::TargetCostKind CostKind) const {
939   int Cost =
940       TTIImpl->getMinMaxReductionCost(Ty, CondTy, IsPairwiseForm, IsUnsigned,
941                                       CostKind);
942   assert(Cost >= 0 && "TTI should not produce negative costs!");
943   return Cost;
944 }
945 
946 unsigned
947 TargetTransformInfo::getCostOfKeepingLiveOverCall(ArrayRef<Type *> Tys) const {
948   return TTIImpl->getCostOfKeepingLiveOverCall(Tys);
949 }
950 
951 bool TargetTransformInfo::getTgtMemIntrinsic(IntrinsicInst *Inst,
952                                              MemIntrinsicInfo &Info) const {
953   return TTIImpl->getTgtMemIntrinsic(Inst, Info);
954 }
955 
956 unsigned TargetTransformInfo::getAtomicMemIntrinsicMaxElementSize() const {
957   return TTIImpl->getAtomicMemIntrinsicMaxElementSize();
958 }
959 
960 Value *TargetTransformInfo::getOrCreateResultFromMemIntrinsic(
961     IntrinsicInst *Inst, Type *ExpectedType) const {
962   return TTIImpl->getOrCreateResultFromMemIntrinsic(Inst, ExpectedType);
963 }
964 
965 Type *TargetTransformInfo::getMemcpyLoopLoweringType(
966     LLVMContext &Context, Value *Length, unsigned SrcAddrSpace,
967     unsigned DestAddrSpace, unsigned SrcAlign, unsigned DestAlign) const {
968   return TTIImpl->getMemcpyLoopLoweringType(Context, Length, SrcAddrSpace,
969                                             DestAddrSpace, SrcAlign, DestAlign);
970 }
971 
972 void TargetTransformInfo::getMemcpyLoopResidualLoweringType(
973     SmallVectorImpl<Type *> &OpsOut, LLVMContext &Context,
974     unsigned RemainingBytes, unsigned SrcAddrSpace, unsigned DestAddrSpace,
975     unsigned SrcAlign, unsigned DestAlign) const {
976   TTIImpl->getMemcpyLoopResidualLoweringType(OpsOut, Context, RemainingBytes,
977                                              SrcAddrSpace, DestAddrSpace,
978                                              SrcAlign, DestAlign);
979 }
980 
981 bool TargetTransformInfo::areInlineCompatible(const Function *Caller,
982                                               const Function *Callee) const {
983   return TTIImpl->areInlineCompatible(Caller, Callee);
984 }
985 
986 bool TargetTransformInfo::areFunctionArgsABICompatible(
987     const Function *Caller, const Function *Callee,
988     SmallPtrSetImpl<Argument *> &Args) const {
989   return TTIImpl->areFunctionArgsABICompatible(Caller, Callee, Args);
990 }
991 
992 bool TargetTransformInfo::isIndexedLoadLegal(MemIndexedMode Mode,
993                                              Type *Ty) const {
994   return TTIImpl->isIndexedLoadLegal(Mode, Ty);
995 }
996 
997 bool TargetTransformInfo::isIndexedStoreLegal(MemIndexedMode Mode,
998                                               Type *Ty) const {
999   return TTIImpl->isIndexedStoreLegal(Mode, Ty);
1000 }
1001 
1002 unsigned TargetTransformInfo::getLoadStoreVecRegBitWidth(unsigned AS) const {
1003   return TTIImpl->getLoadStoreVecRegBitWidth(AS);
1004 }
1005 
1006 bool TargetTransformInfo::isLegalToVectorizeLoad(LoadInst *LI) const {
1007   return TTIImpl->isLegalToVectorizeLoad(LI);
1008 }
1009 
1010 bool TargetTransformInfo::isLegalToVectorizeStore(StoreInst *SI) const {
1011   return TTIImpl->isLegalToVectorizeStore(SI);
1012 }
1013 
1014 bool TargetTransformInfo::isLegalToVectorizeLoadChain(
1015     unsigned ChainSizeInBytes, Align Alignment, unsigned AddrSpace) const {
1016   return TTIImpl->isLegalToVectorizeLoadChain(ChainSizeInBytes, Alignment,
1017                                               AddrSpace);
1018 }
1019 
1020 bool TargetTransformInfo::isLegalToVectorizeStoreChain(
1021     unsigned ChainSizeInBytes, Align Alignment, unsigned AddrSpace) const {
1022   return TTIImpl->isLegalToVectorizeStoreChain(ChainSizeInBytes, Alignment,
1023                                                AddrSpace);
1024 }
1025 
1026 unsigned TargetTransformInfo::getLoadVectorFactor(unsigned VF,
1027                                                   unsigned LoadSize,
1028                                                   unsigned ChainSizeInBytes,
1029                                                   VectorType *VecTy) const {
1030   return TTIImpl->getLoadVectorFactor(VF, LoadSize, ChainSizeInBytes, VecTy);
1031 }
1032 
1033 unsigned TargetTransformInfo::getStoreVectorFactor(unsigned VF,
1034                                                    unsigned StoreSize,
1035                                                    unsigned ChainSizeInBytes,
1036                                                    VectorType *VecTy) const {
1037   return TTIImpl->getStoreVectorFactor(VF, StoreSize, ChainSizeInBytes, VecTy);
1038 }
1039 
1040 bool TargetTransformInfo::useReductionIntrinsic(unsigned Opcode, Type *Ty,
1041                                                 ReductionFlags Flags) const {
1042   return TTIImpl->useReductionIntrinsic(Opcode, Ty, Flags);
1043 }
1044 
1045 bool TargetTransformInfo::preferInLoopReduction(unsigned Opcode, Type *Ty,
1046                                                 ReductionFlags Flags) const {
1047   return TTIImpl->preferInLoopReduction(Opcode, Ty, Flags);
1048 }
1049 
1050 bool TargetTransformInfo::preferPredicatedReductionSelect(
1051     unsigned Opcode, Type *Ty, ReductionFlags Flags) const {
1052   return TTIImpl->preferPredicatedReductionSelect(Opcode, Ty, Flags);
1053 }
1054 
1055 bool TargetTransformInfo::shouldExpandReduction(const IntrinsicInst *II) const {
1056   return TTIImpl->shouldExpandReduction(II);
1057 }
1058 
1059 unsigned TargetTransformInfo::getGISelRematGlobalCost() const {
1060   return TTIImpl->getGISelRematGlobalCost();
1061 }
1062 
1063 bool TargetTransformInfo::supportsScalableVectors() const {
1064   return TTIImpl->supportsScalableVectors();
1065 }
1066 
1067 int TargetTransformInfo::getInstructionLatency(const Instruction *I) const {
1068   return TTIImpl->getInstructionLatency(I);
1069 }
1070 
1071 static bool matchPairwiseShuffleMask(ShuffleVectorInst *SI, bool IsLeft,
1072                                      unsigned Level) {
1073   // We don't need a shuffle if we just want to have element 0 in position 0 of
1074   // the vector.
1075   if (!SI && Level == 0 && IsLeft)
1076     return true;
1077   else if (!SI)
1078     return false;
1079 
1080   SmallVector<int, 32> Mask(
1081       cast<FixedVectorType>(SI->getType())->getNumElements(), -1);
1082 
1083   // Build a mask of 0, 2, ... (left) or 1, 3, ... (right) depending on whether
1084   // we look at the left or right side.
1085   for (unsigned i = 0, e = (1 << Level), val = !IsLeft; i != e; ++i, val += 2)
1086     Mask[i] = val;
1087 
1088   ArrayRef<int> ActualMask = SI->getShuffleMask();
1089   return Mask == ActualMask;
1090 }
1091 
1092 static Optional<TTI::ReductionData> getReductionData(Instruction *I) {
1093   Value *L, *R;
1094   if (m_BinOp(m_Value(L), m_Value(R)).match(I))
1095     return TTI::ReductionData(TTI::RK_Arithmetic, I->getOpcode(), L, R);
1096   if (auto *SI = dyn_cast<SelectInst>(I)) {
1097     if (m_SMin(m_Value(L), m_Value(R)).match(SI) ||
1098         m_SMax(m_Value(L), m_Value(R)).match(SI) ||
1099         m_OrdFMin(m_Value(L), m_Value(R)).match(SI) ||
1100         m_OrdFMax(m_Value(L), m_Value(R)).match(SI) ||
1101         m_UnordFMin(m_Value(L), m_Value(R)).match(SI) ||
1102         m_UnordFMax(m_Value(L), m_Value(R)).match(SI)) {
1103       auto *CI = cast<CmpInst>(SI->getCondition());
1104       return TTI::ReductionData(TTI::RK_MinMax, CI->getOpcode(), L, R);
1105     }
1106     if (m_UMin(m_Value(L), m_Value(R)).match(SI) ||
1107         m_UMax(m_Value(L), m_Value(R)).match(SI)) {
1108       auto *CI = cast<CmpInst>(SI->getCondition());
1109       return TTI::ReductionData(TTI::RK_UnsignedMinMax, CI->getOpcode(), L, R);
1110     }
1111   }
1112   return llvm::None;
1113 }
1114 
1115 static TTI::ReductionKind matchPairwiseReductionAtLevel(Instruction *I,
1116                                                         unsigned Level,
1117                                                         unsigned NumLevels) {
1118   // Match one level of pairwise operations.
1119   // %rdx.shuf.0.0 = shufflevector <4 x float> %rdx, <4 x float> undef,
1120   //       <4 x i32> <i32 0, i32 2 , i32 undef, i32 undef>
1121   // %rdx.shuf.0.1 = shufflevector <4 x float> %rdx, <4 x float> undef,
1122   //       <4 x i32> <i32 1, i32 3, i32 undef, i32 undef>
1123   // %bin.rdx.0 = fadd <4 x float> %rdx.shuf.0.0, %rdx.shuf.0.1
1124   if (!I)
1125     return TTI::RK_None;
1126 
1127   assert(I->getType()->isVectorTy() && "Expecting a vector type");
1128 
1129   Optional<TTI::ReductionData> RD = getReductionData(I);
1130   if (!RD)
1131     return TTI::RK_None;
1132 
1133   ShuffleVectorInst *LS = dyn_cast<ShuffleVectorInst>(RD->LHS);
1134   if (!LS && Level)
1135     return TTI::RK_None;
1136   ShuffleVectorInst *RS = dyn_cast<ShuffleVectorInst>(RD->RHS);
1137   if (!RS && Level)
1138     return TTI::RK_None;
1139 
1140   // On level 0 we can omit one shufflevector instruction.
1141   if (!Level && !RS && !LS)
1142     return TTI::RK_None;
1143 
1144   // Shuffle inputs must match.
1145   Value *NextLevelOpL = LS ? LS->getOperand(0) : nullptr;
1146   Value *NextLevelOpR = RS ? RS->getOperand(0) : nullptr;
1147   Value *NextLevelOp = nullptr;
1148   if (NextLevelOpR && NextLevelOpL) {
1149     // If we have two shuffles their operands must match.
1150     if (NextLevelOpL != NextLevelOpR)
1151       return TTI::RK_None;
1152 
1153     NextLevelOp = NextLevelOpL;
1154   } else if (Level == 0 && (NextLevelOpR || NextLevelOpL)) {
1155     // On the first level we can omit the shufflevector <0, undef,...>. So the
1156     // input to the other shufflevector <1, undef> must match with one of the
1157     // inputs to the current binary operation.
1158     // Example:
1159     //  %NextLevelOpL = shufflevector %R, <1, undef ...>
1160     //  %BinOp        = fadd          %NextLevelOpL, %R
1161     if (NextLevelOpL && NextLevelOpL != RD->RHS)
1162       return TTI::RK_None;
1163     else if (NextLevelOpR && NextLevelOpR != RD->LHS)
1164       return TTI::RK_None;
1165 
1166     NextLevelOp = NextLevelOpL ? RD->RHS : RD->LHS;
1167   } else
1168     return TTI::RK_None;
1169 
1170   // Check that the next levels binary operation exists and matches with the
1171   // current one.
1172   if (Level + 1 != NumLevels) {
1173     if (!isa<Instruction>(NextLevelOp))
1174       return TTI::RK_None;
1175     Optional<TTI::ReductionData> NextLevelRD =
1176         getReductionData(cast<Instruction>(NextLevelOp));
1177     if (!NextLevelRD || !RD->hasSameData(*NextLevelRD))
1178       return TTI::RK_None;
1179   }
1180 
1181   // Shuffle mask for pairwise operation must match.
1182   if (matchPairwiseShuffleMask(LS, /*IsLeft=*/true, Level)) {
1183     if (!matchPairwiseShuffleMask(RS, /*IsLeft=*/false, Level))
1184       return TTI::RK_None;
1185   } else if (matchPairwiseShuffleMask(RS, /*IsLeft=*/true, Level)) {
1186     if (!matchPairwiseShuffleMask(LS, /*IsLeft=*/false, Level))
1187       return TTI::RK_None;
1188   } else {
1189     return TTI::RK_None;
1190   }
1191 
1192   if (++Level == NumLevels)
1193     return RD->Kind;
1194 
1195   // Match next level.
1196   return matchPairwiseReductionAtLevel(dyn_cast<Instruction>(NextLevelOp), Level,
1197                                        NumLevels);
1198 }
1199 
1200 TTI::ReductionKind TTI::matchPairwiseReduction(
1201   const ExtractElementInst *ReduxRoot, unsigned &Opcode, VectorType *&Ty) {
1202   if (!EnableReduxCost)
1203     return TTI::RK_None;
1204 
1205   // Need to extract the first element.
1206   ConstantInt *CI = dyn_cast<ConstantInt>(ReduxRoot->getOperand(1));
1207   unsigned Idx = ~0u;
1208   if (CI)
1209     Idx = CI->getZExtValue();
1210   if (Idx != 0)
1211     return TTI::RK_None;
1212 
1213   auto *RdxStart = dyn_cast<Instruction>(ReduxRoot->getOperand(0));
1214   if (!RdxStart)
1215     return TTI::RK_None;
1216   Optional<TTI::ReductionData> RD = getReductionData(RdxStart);
1217   if (!RD)
1218     return TTI::RK_None;
1219 
1220   auto *VecTy = cast<FixedVectorType>(RdxStart->getType());
1221   unsigned NumVecElems = VecTy->getNumElements();
1222   if (!isPowerOf2_32(NumVecElems))
1223     return TTI::RK_None;
1224 
1225   // We look for a sequence of shuffle,shuffle,add triples like the following
1226   // that builds a pairwise reduction tree.
1227   //
1228   //  (X0, X1, X2, X3)
1229   //   (X0 + X1, X2 + X3, undef, undef)
1230   //    ((X0 + X1) + (X2 + X3), undef, undef, undef)
1231   //
1232   // %rdx.shuf.0.0 = shufflevector <4 x float> %rdx, <4 x float> undef,
1233   //       <4 x i32> <i32 0, i32 2 , i32 undef, i32 undef>
1234   // %rdx.shuf.0.1 = shufflevector <4 x float> %rdx, <4 x float> undef,
1235   //       <4 x i32> <i32 1, i32 3, i32 undef, i32 undef>
1236   // %bin.rdx.0 = fadd <4 x float> %rdx.shuf.0.0, %rdx.shuf.0.1
1237   // %rdx.shuf.1.0 = shufflevector <4 x float> %bin.rdx.0, <4 x float> undef,
1238   //       <4 x i32> <i32 0, i32 undef, i32 undef, i32 undef>
1239   // %rdx.shuf.1.1 = shufflevector <4 x float> %bin.rdx.0, <4 x float> undef,
1240   //       <4 x i32> <i32 1, i32 undef, i32 undef, i32 undef>
1241   // %bin.rdx8 = fadd <4 x float> %rdx.shuf.1.0, %rdx.shuf.1.1
1242   // %r = extractelement <4 x float> %bin.rdx8, i32 0
1243   if (matchPairwiseReductionAtLevel(RdxStart, 0, Log2_32(NumVecElems)) ==
1244       TTI::RK_None)
1245     return TTI::RK_None;
1246 
1247   Opcode = RD->Opcode;
1248   Ty = VecTy;
1249 
1250   return RD->Kind;
1251 }
1252 
1253 static std::pair<Value *, ShuffleVectorInst *>
1254 getShuffleAndOtherOprd(Value *L, Value *R) {
1255   ShuffleVectorInst *S = nullptr;
1256 
1257   if ((S = dyn_cast<ShuffleVectorInst>(L)))
1258     return std::make_pair(R, S);
1259 
1260   S = dyn_cast<ShuffleVectorInst>(R);
1261   return std::make_pair(L, S);
1262 }
1263 
1264 TTI::ReductionKind TTI::matchVectorSplittingReduction(
1265   const ExtractElementInst *ReduxRoot, unsigned &Opcode, VectorType *&Ty) {
1266 
1267   if (!EnableReduxCost)
1268     return TTI::RK_None;
1269 
1270   // Need to extract the first element.
1271   ConstantInt *CI = dyn_cast<ConstantInt>(ReduxRoot->getOperand(1));
1272   unsigned Idx = ~0u;
1273   if (CI)
1274     Idx = CI->getZExtValue();
1275   if (Idx != 0)
1276     return TTI::RK_None;
1277 
1278   auto *RdxStart = dyn_cast<Instruction>(ReduxRoot->getOperand(0));
1279   if (!RdxStart)
1280     return TTI::RK_None;
1281   Optional<TTI::ReductionData> RD = getReductionData(RdxStart);
1282   if (!RD)
1283     return TTI::RK_None;
1284 
1285   auto *VecTy = cast<FixedVectorType>(ReduxRoot->getOperand(0)->getType());
1286   unsigned NumVecElems = VecTy->getNumElements();
1287   if (!isPowerOf2_32(NumVecElems))
1288     return TTI::RK_None;
1289 
1290   // We look for a sequence of shuffles and adds like the following matching one
1291   // fadd, shuffle vector pair at a time.
1292   //
1293   // %rdx.shuf = shufflevector <4 x float> %rdx, <4 x float> undef,
1294   //                           <4 x i32> <i32 2, i32 3, i32 undef, i32 undef>
1295   // %bin.rdx = fadd <4 x float> %rdx, %rdx.shuf
1296   // %rdx.shuf7 = shufflevector <4 x float> %bin.rdx, <4 x float> undef,
1297   //                          <4 x i32> <i32 1, i32 undef, i32 undef, i32 undef>
1298   // %bin.rdx8 = fadd <4 x float> %bin.rdx, %rdx.shuf7
1299   // %r = extractelement <4 x float> %bin.rdx8, i32 0
1300 
1301   unsigned MaskStart = 1;
1302   Instruction *RdxOp = RdxStart;
1303   SmallVector<int, 32> ShuffleMask(NumVecElems, 0);
1304   unsigned NumVecElemsRemain = NumVecElems;
1305   while (NumVecElemsRemain - 1) {
1306     // Check for the right reduction operation.
1307     if (!RdxOp)
1308       return TTI::RK_None;
1309     Optional<TTI::ReductionData> RDLevel = getReductionData(RdxOp);
1310     if (!RDLevel || !RDLevel->hasSameData(*RD))
1311       return TTI::RK_None;
1312 
1313     Value *NextRdxOp;
1314     ShuffleVectorInst *Shuffle;
1315     std::tie(NextRdxOp, Shuffle) =
1316         getShuffleAndOtherOprd(RDLevel->LHS, RDLevel->RHS);
1317 
1318     // Check the current reduction operation and the shuffle use the same value.
1319     if (Shuffle == nullptr)
1320       return TTI::RK_None;
1321     if (Shuffle->getOperand(0) != NextRdxOp)
1322       return TTI::RK_None;
1323 
1324     // Check that shuffle masks matches.
1325     for (unsigned j = 0; j != MaskStart; ++j)
1326       ShuffleMask[j] = MaskStart + j;
1327     // Fill the rest of the mask with -1 for undef.
1328     std::fill(&ShuffleMask[MaskStart], ShuffleMask.end(), -1);
1329 
1330     ArrayRef<int> Mask = Shuffle->getShuffleMask();
1331     if (ShuffleMask != Mask)
1332       return TTI::RK_None;
1333 
1334     RdxOp = dyn_cast<Instruction>(NextRdxOp);
1335     NumVecElemsRemain /= 2;
1336     MaskStart *= 2;
1337   }
1338 
1339   Opcode = RD->Opcode;
1340   Ty = VecTy;
1341   return RD->Kind;
1342 }
1343 
1344 TTI::ReductionKind
1345 TTI::matchVectorReduction(const ExtractElementInst *Root, unsigned &Opcode,
1346                           VectorType *&Ty, bool &IsPairwise) {
1347   TTI::ReductionKind RdxKind = matchVectorSplittingReduction(Root, Opcode, Ty);
1348   if (RdxKind != TTI::ReductionKind::RK_None) {
1349     IsPairwise = false;
1350     return RdxKind;
1351   }
1352   IsPairwise = true;
1353   return matchPairwiseReduction(Root, Opcode, Ty);
1354 }
1355 
1356 int TargetTransformInfo::getInstructionThroughput(const Instruction *I) const {
1357   TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput;
1358 
1359   switch (I->getOpcode()) {
1360   case Instruction::GetElementPtr:
1361   case Instruction::Ret:
1362   case Instruction::PHI:
1363   case Instruction::Br:
1364   case Instruction::Add:
1365   case Instruction::FAdd:
1366   case Instruction::Sub:
1367   case Instruction::FSub:
1368   case Instruction::Mul:
1369   case Instruction::FMul:
1370   case Instruction::UDiv:
1371   case Instruction::SDiv:
1372   case Instruction::FDiv:
1373   case Instruction::URem:
1374   case Instruction::SRem:
1375   case Instruction::FRem:
1376   case Instruction::Shl:
1377   case Instruction::LShr:
1378   case Instruction::AShr:
1379   case Instruction::And:
1380   case Instruction::Or:
1381   case Instruction::Xor:
1382   case Instruction::FNeg:
1383   case Instruction::Select:
1384   case Instruction::ICmp:
1385   case Instruction::FCmp:
1386   case Instruction::Store:
1387   case Instruction::Load:
1388   case Instruction::ZExt:
1389   case Instruction::SExt:
1390   case Instruction::FPToUI:
1391   case Instruction::FPToSI:
1392   case Instruction::FPExt:
1393   case Instruction::PtrToInt:
1394   case Instruction::IntToPtr:
1395   case Instruction::SIToFP:
1396   case Instruction::UIToFP:
1397   case Instruction::Trunc:
1398   case Instruction::FPTrunc:
1399   case Instruction::BitCast:
1400   case Instruction::AddrSpaceCast:
1401   case Instruction::ExtractElement:
1402   case Instruction::InsertElement:
1403   case Instruction::ExtractValue:
1404   case Instruction::ShuffleVector:
1405   case Instruction::Call:
1406     return getUserCost(I, CostKind);
1407   default:
1408     // We don't have any information on this instruction.
1409     return -1;
1410   }
1411 }
1412 
1413 TargetTransformInfo::Concept::~Concept() {}
1414 
1415 TargetIRAnalysis::TargetIRAnalysis() : TTICallback(&getDefaultTTI) {}
1416 
1417 TargetIRAnalysis::TargetIRAnalysis(
1418     std::function<Result(const Function &)> TTICallback)
1419     : TTICallback(std::move(TTICallback)) {}
1420 
1421 TargetIRAnalysis::Result TargetIRAnalysis::run(const Function &F,
1422                                                FunctionAnalysisManager &) {
1423   return TTICallback(F);
1424 }
1425 
1426 AnalysisKey TargetIRAnalysis::Key;
1427 
1428 TargetIRAnalysis::Result TargetIRAnalysis::getDefaultTTI(const Function &F) {
1429   return Result(F.getParent()->getDataLayout());
1430 }
1431 
1432 // Register the basic pass.
1433 INITIALIZE_PASS(TargetTransformInfoWrapperPass, "tti",
1434                 "Target Transform Information", false, true)
1435 char TargetTransformInfoWrapperPass::ID = 0;
1436 
1437 void TargetTransformInfoWrapperPass::anchor() {}
1438 
1439 TargetTransformInfoWrapperPass::TargetTransformInfoWrapperPass()
1440     : ImmutablePass(ID) {
1441   initializeTargetTransformInfoWrapperPassPass(
1442       *PassRegistry::getPassRegistry());
1443 }
1444 
1445 TargetTransformInfoWrapperPass::TargetTransformInfoWrapperPass(
1446     TargetIRAnalysis TIRA)
1447     : ImmutablePass(ID), TIRA(std::move(TIRA)) {
1448   initializeTargetTransformInfoWrapperPassPass(
1449       *PassRegistry::getPassRegistry());
1450 }
1451 
1452 TargetTransformInfo &TargetTransformInfoWrapperPass::getTTI(const Function &F) {
1453   FunctionAnalysisManager DummyFAM;
1454   TTI = TIRA.run(F, DummyFAM);
1455   return *TTI;
1456 }
1457 
1458 ImmutablePass *
1459 llvm::createTargetTransformInfoWrapperPass(TargetIRAnalysis TIRA) {
1460   return new TargetTransformInfoWrapperPass(std::move(TIRA));
1461 }
1462