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/Dominators.h"
15 #include "llvm/IR/Instruction.h"
16 #include "llvm/IR/Instructions.h"
17 #include "llvm/IR/IntrinsicInst.h"
18 #include "llvm/IR/Module.h"
19 #include "llvm/IR/Operator.h"
20 #include "llvm/IR/PatternMatch.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 namespace {
41 /// No-op implementation of the TTI interface using the utility base
42 /// classes.
43 ///
44 /// This is used when no target specific information is available.
45 struct NoTTIImpl : TargetTransformInfoImplCRTPBase<NoTTIImpl> {
NoTTIImpl__anon76921ad20111::NoTTIImpl46 explicit NoTTIImpl(const DataLayout &DL)
47 : TargetTransformInfoImplCRTPBase<NoTTIImpl>(DL) {}
48 };
49 } // namespace
50
canAnalyze(LoopInfo & LI)51 bool HardwareLoopInfo::canAnalyze(LoopInfo &LI) {
52 // If the loop has irreducible control flow, it can not be converted to
53 // Hardware loop.
54 LoopBlocksRPO RPOT(L);
55 RPOT.perform(&LI);
56 if (containsIrreducibleCFG<const BasicBlock *>(RPOT, LI))
57 return false;
58 return true;
59 }
60
IntrinsicCostAttributes(Intrinsic::ID Id,const CallBase & CI,InstructionCost ScalarizationCost,bool TypeBasedOnly)61 IntrinsicCostAttributes::IntrinsicCostAttributes(
62 Intrinsic::ID Id, const CallBase &CI, InstructionCost ScalarizationCost,
63 bool TypeBasedOnly)
64 : II(dyn_cast<IntrinsicInst>(&CI)), RetTy(CI.getType()), IID(Id),
65 ScalarizationCost(ScalarizationCost) {
66
67 if (const auto *FPMO = dyn_cast<FPMathOperator>(&CI))
68 FMF = FPMO->getFastMathFlags();
69
70 if (!TypeBasedOnly)
71 Arguments.insert(Arguments.begin(), CI.arg_begin(), CI.arg_end());
72 FunctionType *FTy = CI.getCalledFunction()->getFunctionType();
73 ParamTys.insert(ParamTys.begin(), FTy->param_begin(), FTy->param_end());
74 }
75
IntrinsicCostAttributes(Intrinsic::ID Id,Type * RTy,ArrayRef<Type * > Tys,FastMathFlags Flags,const IntrinsicInst * I,InstructionCost ScalarCost)76 IntrinsicCostAttributes::IntrinsicCostAttributes(Intrinsic::ID Id, Type *RTy,
77 ArrayRef<Type *> Tys,
78 FastMathFlags Flags,
79 const IntrinsicInst *I,
80 InstructionCost ScalarCost)
81 : II(I), RetTy(RTy), IID(Id), FMF(Flags), ScalarizationCost(ScalarCost) {
82 ParamTys.insert(ParamTys.begin(), Tys.begin(), Tys.end());
83 }
84
IntrinsicCostAttributes(Intrinsic::ID Id,Type * Ty,ArrayRef<const Value * > Args)85 IntrinsicCostAttributes::IntrinsicCostAttributes(Intrinsic::ID Id, Type *Ty,
86 ArrayRef<const Value *> Args)
87 : RetTy(Ty), IID(Id) {
88
89 Arguments.insert(Arguments.begin(), Args.begin(), Args.end());
90 ParamTys.reserve(Arguments.size());
91 for (unsigned Idx = 0, Size = Arguments.size(); Idx != Size; ++Idx)
92 ParamTys.push_back(Arguments[Idx]->getType());
93 }
94
IntrinsicCostAttributes(Intrinsic::ID Id,Type * RTy,ArrayRef<const Value * > Args,ArrayRef<Type * > Tys,FastMathFlags Flags,const IntrinsicInst * I,InstructionCost ScalarCost)95 IntrinsicCostAttributes::IntrinsicCostAttributes(Intrinsic::ID Id, Type *RTy,
96 ArrayRef<const Value *> Args,
97 ArrayRef<Type *> Tys,
98 FastMathFlags Flags,
99 const IntrinsicInst *I,
100 InstructionCost ScalarCost)
101 : II(I), RetTy(RTy), IID(Id), FMF(Flags), ScalarizationCost(ScalarCost) {
102 ParamTys.insert(ParamTys.begin(), Tys.begin(), Tys.end());
103 Arguments.insert(Arguments.begin(), Args.begin(), Args.end());
104 }
105
isHardwareLoopCandidate(ScalarEvolution & SE,LoopInfo & LI,DominatorTree & DT,bool ForceNestedLoop,bool ForceHardwareLoopPHI)106 bool HardwareLoopInfo::isHardwareLoopCandidate(ScalarEvolution &SE,
107 LoopInfo &LI, DominatorTree &DT,
108 bool ForceNestedLoop,
109 bool ForceHardwareLoopPHI) {
110 SmallVector<BasicBlock *, 4> ExitingBlocks;
111 L->getExitingBlocks(ExitingBlocks);
112
113 for (BasicBlock *BB : ExitingBlocks) {
114 // If we pass the updated counter back through a phi, we need to know
115 // which latch the updated value will be coming from.
116 if (!L->isLoopLatch(BB)) {
117 if (ForceHardwareLoopPHI || CounterInReg)
118 continue;
119 }
120
121 const SCEV *EC = SE.getExitCount(L, BB);
122 if (isa<SCEVCouldNotCompute>(EC))
123 continue;
124 if (const SCEVConstant *ConstEC = dyn_cast<SCEVConstant>(EC)) {
125 if (ConstEC->getValue()->isZero())
126 continue;
127 } else if (!SE.isLoopInvariant(EC, L))
128 continue;
129
130 if (SE.getTypeSizeInBits(EC->getType()) > CountType->getBitWidth())
131 continue;
132
133 // If this exiting block is contained in a nested loop, it is not eligible
134 // for insertion of the branch-and-decrement since the inner loop would
135 // end up messing up the value in the CTR.
136 if (!IsNestingLegal && LI.getLoopFor(BB) != L && !ForceNestedLoop)
137 continue;
138
139 // We now have a loop-invariant count of loop iterations (which is not the
140 // constant zero) for which we know that this loop will not exit via this
141 // existing block.
142
143 // We need to make sure that this block will run on every loop iteration.
144 // For this to be true, we must dominate all blocks with backedges. Such
145 // blocks are in-loop predecessors to the header block.
146 bool NotAlways = false;
147 for (BasicBlock *Pred : predecessors(L->getHeader())) {
148 if (!L->contains(Pred))
149 continue;
150
151 if (!DT.dominates(BB, Pred)) {
152 NotAlways = true;
153 break;
154 }
155 }
156
157 if (NotAlways)
158 continue;
159
160 // Make sure this blocks ends with a conditional branch.
161 Instruction *TI = BB->getTerminator();
162 if (!TI)
163 continue;
164
165 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
166 if (!BI->isConditional())
167 continue;
168
169 ExitBranch = BI;
170 } else
171 continue;
172
173 // Note that this block may not be the loop latch block, even if the loop
174 // has a latch block.
175 ExitBlock = BB;
176 ExitCount = EC;
177 break;
178 }
179
180 if (!ExitBlock)
181 return false;
182 return true;
183 }
184
TargetTransformInfo(const DataLayout & DL)185 TargetTransformInfo::TargetTransformInfo(const DataLayout &DL)
186 : TTIImpl(new Model<NoTTIImpl>(NoTTIImpl(DL))) {}
187
188 TargetTransformInfo::~TargetTransformInfo() = default;
189
TargetTransformInfo(TargetTransformInfo && Arg)190 TargetTransformInfo::TargetTransformInfo(TargetTransformInfo &&Arg)
191 : TTIImpl(std::move(Arg.TTIImpl)) {}
192
operator =(TargetTransformInfo && RHS)193 TargetTransformInfo &TargetTransformInfo::operator=(TargetTransformInfo &&RHS) {
194 TTIImpl = std::move(RHS.TTIImpl);
195 return *this;
196 }
197
getInliningThresholdMultiplier() const198 unsigned TargetTransformInfo::getInliningThresholdMultiplier() const {
199 return TTIImpl->getInliningThresholdMultiplier();
200 }
201
202 unsigned
adjustInliningThreshold(const CallBase * CB) const203 TargetTransformInfo::adjustInliningThreshold(const CallBase *CB) const {
204 return TTIImpl->adjustInliningThreshold(CB);
205 }
206
getInlinerVectorBonusPercent() const207 int TargetTransformInfo::getInlinerVectorBonusPercent() const {
208 return TTIImpl->getInlinerVectorBonusPercent();
209 }
210
211 InstructionCost
getGEPCost(Type * PointeeType,const Value * Ptr,ArrayRef<const Value * > Operands,TTI::TargetCostKind CostKind) const212 TargetTransformInfo::getGEPCost(Type *PointeeType, const Value *Ptr,
213 ArrayRef<const Value *> Operands,
214 TTI::TargetCostKind CostKind) const {
215 return TTIImpl->getGEPCost(PointeeType, Ptr, Operands, CostKind);
216 }
217
getEstimatedNumberOfCaseClusters(const SwitchInst & SI,unsigned & JTSize,ProfileSummaryInfo * PSI,BlockFrequencyInfo * BFI) const218 unsigned TargetTransformInfo::getEstimatedNumberOfCaseClusters(
219 const SwitchInst &SI, unsigned &JTSize, ProfileSummaryInfo *PSI,
220 BlockFrequencyInfo *BFI) const {
221 return TTIImpl->getEstimatedNumberOfCaseClusters(SI, JTSize, PSI, BFI);
222 }
223
224 InstructionCost
getInstructionCost(const User * U,ArrayRef<const Value * > Operands,enum TargetCostKind CostKind) const225 TargetTransformInfo::getInstructionCost(const User *U,
226 ArrayRef<const Value *> Operands,
227 enum TargetCostKind CostKind) const {
228 InstructionCost Cost = TTIImpl->getInstructionCost(U, Operands, CostKind);
229 assert((CostKind == TTI::TCK_RecipThroughput || Cost >= 0) &&
230 "TTI should not produce negative costs!");
231 return Cost;
232 }
233
getPredictableBranchThreshold() const234 BranchProbability TargetTransformInfo::getPredictableBranchThreshold() const {
235 return TTIImpl->getPredictableBranchThreshold();
236 }
237
hasBranchDivergence() const238 bool TargetTransformInfo::hasBranchDivergence() const {
239 return TTIImpl->hasBranchDivergence();
240 }
241
useGPUDivergenceAnalysis() const242 bool TargetTransformInfo::useGPUDivergenceAnalysis() const {
243 return TTIImpl->useGPUDivergenceAnalysis();
244 }
245
isSourceOfDivergence(const Value * V) const246 bool TargetTransformInfo::isSourceOfDivergence(const Value *V) const {
247 return TTIImpl->isSourceOfDivergence(V);
248 }
249
isAlwaysUniform(const Value * V) const250 bool llvm::TargetTransformInfo::isAlwaysUniform(const Value *V) const {
251 return TTIImpl->isAlwaysUniform(V);
252 }
253
getFlatAddressSpace() const254 unsigned TargetTransformInfo::getFlatAddressSpace() const {
255 return TTIImpl->getFlatAddressSpace();
256 }
257
collectFlatAddressOperands(SmallVectorImpl<int> & OpIndexes,Intrinsic::ID IID) const258 bool TargetTransformInfo::collectFlatAddressOperands(
259 SmallVectorImpl<int> &OpIndexes, Intrinsic::ID IID) const {
260 return TTIImpl->collectFlatAddressOperands(OpIndexes, IID);
261 }
262
isNoopAddrSpaceCast(unsigned FromAS,unsigned ToAS) const263 bool TargetTransformInfo::isNoopAddrSpaceCast(unsigned FromAS,
264 unsigned ToAS) const {
265 return TTIImpl->isNoopAddrSpaceCast(FromAS, ToAS);
266 }
267
canHaveNonUndefGlobalInitializerInAddressSpace(unsigned AS) const268 bool TargetTransformInfo::canHaveNonUndefGlobalInitializerInAddressSpace(
269 unsigned AS) const {
270 return TTIImpl->canHaveNonUndefGlobalInitializerInAddressSpace(AS);
271 }
272
getAssumedAddrSpace(const Value * V) const273 unsigned TargetTransformInfo::getAssumedAddrSpace(const Value *V) const {
274 return TTIImpl->getAssumedAddrSpace(V);
275 }
276
isSingleThreaded() const277 bool TargetTransformInfo::isSingleThreaded() const {
278 return TTIImpl->isSingleThreaded();
279 }
280
281 std::pair<const Value *, unsigned>
getPredicatedAddrSpace(const Value * V) const282 TargetTransformInfo::getPredicatedAddrSpace(const Value *V) const {
283 return TTIImpl->getPredicatedAddrSpace(V);
284 }
285
rewriteIntrinsicWithAddressSpace(IntrinsicInst * II,Value * OldV,Value * NewV) const286 Value *TargetTransformInfo::rewriteIntrinsicWithAddressSpace(
287 IntrinsicInst *II, Value *OldV, Value *NewV) const {
288 return TTIImpl->rewriteIntrinsicWithAddressSpace(II, OldV, NewV);
289 }
290
isLoweredToCall(const Function * F) const291 bool TargetTransformInfo::isLoweredToCall(const Function *F) const {
292 return TTIImpl->isLoweredToCall(F);
293 }
294
isHardwareLoopProfitable(Loop * L,ScalarEvolution & SE,AssumptionCache & AC,TargetLibraryInfo * LibInfo,HardwareLoopInfo & HWLoopInfo) const295 bool TargetTransformInfo::isHardwareLoopProfitable(
296 Loop *L, ScalarEvolution &SE, AssumptionCache &AC,
297 TargetLibraryInfo *LibInfo, HardwareLoopInfo &HWLoopInfo) const {
298 return TTIImpl->isHardwareLoopProfitable(L, SE, AC, LibInfo, HWLoopInfo);
299 }
300
preferPredicateOverEpilogue(Loop * L,LoopInfo * LI,ScalarEvolution & SE,AssumptionCache & AC,TargetLibraryInfo * TLI,DominatorTree * DT,LoopVectorizationLegality * LVL,InterleavedAccessInfo * IAI) const301 bool TargetTransformInfo::preferPredicateOverEpilogue(
302 Loop *L, LoopInfo *LI, ScalarEvolution &SE, AssumptionCache &AC,
303 TargetLibraryInfo *TLI, DominatorTree *DT, LoopVectorizationLegality *LVL,
304 InterleavedAccessInfo *IAI) const {
305 return TTIImpl->preferPredicateOverEpilogue(L, LI, SE, AC, TLI, DT, LVL, IAI);
306 }
307
emitGetActiveLaneMask() const308 PredicationStyle TargetTransformInfo::emitGetActiveLaneMask() const {
309 return TTIImpl->emitGetActiveLaneMask();
310 }
311
312 std::optional<Instruction *>
instCombineIntrinsic(InstCombiner & IC,IntrinsicInst & II) const313 TargetTransformInfo::instCombineIntrinsic(InstCombiner &IC,
314 IntrinsicInst &II) const {
315 return TTIImpl->instCombineIntrinsic(IC, II);
316 }
317
simplifyDemandedUseBitsIntrinsic(InstCombiner & IC,IntrinsicInst & II,APInt DemandedMask,KnownBits & Known,bool & KnownBitsComputed) const318 std::optional<Value *> TargetTransformInfo::simplifyDemandedUseBitsIntrinsic(
319 InstCombiner &IC, IntrinsicInst &II, APInt DemandedMask, KnownBits &Known,
320 bool &KnownBitsComputed) const {
321 return TTIImpl->simplifyDemandedUseBitsIntrinsic(IC, II, DemandedMask, Known,
322 KnownBitsComputed);
323 }
324
simplifyDemandedVectorEltsIntrinsic(InstCombiner & IC,IntrinsicInst & II,APInt DemandedElts,APInt & UndefElts,APInt & UndefElts2,APInt & UndefElts3,std::function<void (Instruction *,unsigned,APInt,APInt &)> SimplifyAndSetOp) const325 std::optional<Value *> TargetTransformInfo::simplifyDemandedVectorEltsIntrinsic(
326 InstCombiner &IC, IntrinsicInst &II, APInt DemandedElts, APInt &UndefElts,
327 APInt &UndefElts2, APInt &UndefElts3,
328 std::function<void(Instruction *, unsigned, APInt, APInt &)>
329 SimplifyAndSetOp) const {
330 return TTIImpl->simplifyDemandedVectorEltsIntrinsic(
331 IC, II, DemandedElts, UndefElts, UndefElts2, UndefElts3,
332 SimplifyAndSetOp);
333 }
334
getUnrollingPreferences(Loop * L,ScalarEvolution & SE,UnrollingPreferences & UP,OptimizationRemarkEmitter * ORE) const335 void TargetTransformInfo::getUnrollingPreferences(
336 Loop *L, ScalarEvolution &SE, UnrollingPreferences &UP,
337 OptimizationRemarkEmitter *ORE) const {
338 return TTIImpl->getUnrollingPreferences(L, SE, UP, ORE);
339 }
340
getPeelingPreferences(Loop * L,ScalarEvolution & SE,PeelingPreferences & PP) const341 void TargetTransformInfo::getPeelingPreferences(Loop *L, ScalarEvolution &SE,
342 PeelingPreferences &PP) const {
343 return TTIImpl->getPeelingPreferences(L, SE, PP);
344 }
345
isLegalAddImmediate(int64_t Imm) const346 bool TargetTransformInfo::isLegalAddImmediate(int64_t Imm) const {
347 return TTIImpl->isLegalAddImmediate(Imm);
348 }
349
isLegalICmpImmediate(int64_t Imm) const350 bool TargetTransformInfo::isLegalICmpImmediate(int64_t Imm) const {
351 return TTIImpl->isLegalICmpImmediate(Imm);
352 }
353
isLegalAddressingMode(Type * Ty,GlobalValue * BaseGV,int64_t BaseOffset,bool HasBaseReg,int64_t Scale,unsigned AddrSpace,Instruction * I) const354 bool TargetTransformInfo::isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV,
355 int64_t BaseOffset,
356 bool HasBaseReg, int64_t Scale,
357 unsigned AddrSpace,
358 Instruction *I) const {
359 return TTIImpl->isLegalAddressingMode(Ty, BaseGV, BaseOffset, HasBaseReg,
360 Scale, AddrSpace, I);
361 }
362
isLSRCostLess(const LSRCost & C1,const LSRCost & C2) const363 bool TargetTransformInfo::isLSRCostLess(const LSRCost &C1,
364 const LSRCost &C2) const {
365 return TTIImpl->isLSRCostLess(C1, C2);
366 }
367
isNumRegsMajorCostOfLSR() const368 bool TargetTransformInfo::isNumRegsMajorCostOfLSR() const {
369 return TTIImpl->isNumRegsMajorCostOfLSR();
370 }
371
isProfitableLSRChainElement(Instruction * I) const372 bool TargetTransformInfo::isProfitableLSRChainElement(Instruction *I) const {
373 return TTIImpl->isProfitableLSRChainElement(I);
374 }
375
canMacroFuseCmp() const376 bool TargetTransformInfo::canMacroFuseCmp() const {
377 return TTIImpl->canMacroFuseCmp();
378 }
379
canSaveCmp(Loop * L,BranchInst ** BI,ScalarEvolution * SE,LoopInfo * LI,DominatorTree * DT,AssumptionCache * AC,TargetLibraryInfo * LibInfo) const380 bool TargetTransformInfo::canSaveCmp(Loop *L, BranchInst **BI,
381 ScalarEvolution *SE, LoopInfo *LI,
382 DominatorTree *DT, AssumptionCache *AC,
383 TargetLibraryInfo *LibInfo) const {
384 return TTIImpl->canSaveCmp(L, BI, SE, LI, DT, AC, LibInfo);
385 }
386
387 TTI::AddressingModeKind
getPreferredAddressingMode(const Loop * L,ScalarEvolution * SE) const388 TargetTransformInfo::getPreferredAddressingMode(const Loop *L,
389 ScalarEvolution *SE) const {
390 return TTIImpl->getPreferredAddressingMode(L, SE);
391 }
392
isLegalMaskedStore(Type * DataType,Align Alignment) const393 bool TargetTransformInfo::isLegalMaskedStore(Type *DataType,
394 Align Alignment) const {
395 return TTIImpl->isLegalMaskedStore(DataType, Alignment);
396 }
397
isLegalMaskedLoad(Type * DataType,Align Alignment) const398 bool TargetTransformInfo::isLegalMaskedLoad(Type *DataType,
399 Align Alignment) const {
400 return TTIImpl->isLegalMaskedLoad(DataType, Alignment);
401 }
402
isLegalNTStore(Type * DataType,Align Alignment) const403 bool TargetTransformInfo::isLegalNTStore(Type *DataType,
404 Align Alignment) const {
405 return TTIImpl->isLegalNTStore(DataType, Alignment);
406 }
407
isLegalNTLoad(Type * DataType,Align Alignment) const408 bool TargetTransformInfo::isLegalNTLoad(Type *DataType, Align Alignment) const {
409 return TTIImpl->isLegalNTLoad(DataType, Alignment);
410 }
411
isLegalBroadcastLoad(Type * ElementTy,ElementCount NumElements) const412 bool TargetTransformInfo::isLegalBroadcastLoad(Type *ElementTy,
413 ElementCount NumElements) const {
414 return TTIImpl->isLegalBroadcastLoad(ElementTy, NumElements);
415 }
416
isLegalMaskedGather(Type * DataType,Align Alignment) const417 bool TargetTransformInfo::isLegalMaskedGather(Type *DataType,
418 Align Alignment) const {
419 return TTIImpl->isLegalMaskedGather(DataType, Alignment);
420 }
421
isLegalAltInstr(VectorType * VecTy,unsigned Opcode0,unsigned Opcode1,const SmallBitVector & OpcodeMask) const422 bool TargetTransformInfo::isLegalAltInstr(
423 VectorType *VecTy, unsigned Opcode0, unsigned Opcode1,
424 const SmallBitVector &OpcodeMask) const {
425 return TTIImpl->isLegalAltInstr(VecTy, Opcode0, Opcode1, OpcodeMask);
426 }
427
isLegalMaskedScatter(Type * DataType,Align Alignment) const428 bool TargetTransformInfo::isLegalMaskedScatter(Type *DataType,
429 Align Alignment) const {
430 return TTIImpl->isLegalMaskedScatter(DataType, Alignment);
431 }
432
forceScalarizeMaskedGather(VectorType * DataType,Align Alignment) const433 bool TargetTransformInfo::forceScalarizeMaskedGather(VectorType *DataType,
434 Align Alignment) const {
435 return TTIImpl->forceScalarizeMaskedGather(DataType, Alignment);
436 }
437
forceScalarizeMaskedScatter(VectorType * DataType,Align Alignment) const438 bool TargetTransformInfo::forceScalarizeMaskedScatter(VectorType *DataType,
439 Align Alignment) const {
440 return TTIImpl->forceScalarizeMaskedScatter(DataType, Alignment);
441 }
442
isLegalMaskedCompressStore(Type * DataType) const443 bool TargetTransformInfo::isLegalMaskedCompressStore(Type *DataType) const {
444 return TTIImpl->isLegalMaskedCompressStore(DataType);
445 }
446
isLegalMaskedExpandLoad(Type * DataType) const447 bool TargetTransformInfo::isLegalMaskedExpandLoad(Type *DataType) const {
448 return TTIImpl->isLegalMaskedExpandLoad(DataType);
449 }
450
enableOrderedReductions() const451 bool TargetTransformInfo::enableOrderedReductions() const {
452 return TTIImpl->enableOrderedReductions();
453 }
454
hasDivRemOp(Type * DataType,bool IsSigned) const455 bool TargetTransformInfo::hasDivRemOp(Type *DataType, bool IsSigned) const {
456 return TTIImpl->hasDivRemOp(DataType, IsSigned);
457 }
458
hasVolatileVariant(Instruction * I,unsigned AddrSpace) const459 bool TargetTransformInfo::hasVolatileVariant(Instruction *I,
460 unsigned AddrSpace) const {
461 return TTIImpl->hasVolatileVariant(I, AddrSpace);
462 }
463
prefersVectorizedAddressing() const464 bool TargetTransformInfo::prefersVectorizedAddressing() const {
465 return TTIImpl->prefersVectorizedAddressing();
466 }
467
getScalingFactorCost(Type * Ty,GlobalValue * BaseGV,int64_t BaseOffset,bool HasBaseReg,int64_t Scale,unsigned AddrSpace) const468 InstructionCost TargetTransformInfo::getScalingFactorCost(
469 Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset, bool HasBaseReg,
470 int64_t Scale, unsigned AddrSpace) const {
471 InstructionCost Cost = TTIImpl->getScalingFactorCost(
472 Ty, BaseGV, BaseOffset, HasBaseReg, Scale, AddrSpace);
473 assert(Cost >= 0 && "TTI should not produce negative costs!");
474 return Cost;
475 }
476
LSRWithInstrQueries() const477 bool TargetTransformInfo::LSRWithInstrQueries() const {
478 return TTIImpl->LSRWithInstrQueries();
479 }
480
isTruncateFree(Type * Ty1,Type * Ty2) const481 bool TargetTransformInfo::isTruncateFree(Type *Ty1, Type *Ty2) const {
482 return TTIImpl->isTruncateFree(Ty1, Ty2);
483 }
484
isProfitableToHoist(Instruction * I) const485 bool TargetTransformInfo::isProfitableToHoist(Instruction *I) const {
486 return TTIImpl->isProfitableToHoist(I);
487 }
488
useAA() const489 bool TargetTransformInfo::useAA() const { return TTIImpl->useAA(); }
490
isTypeLegal(Type * Ty) const491 bool TargetTransformInfo::isTypeLegal(Type *Ty) const {
492 return TTIImpl->isTypeLegal(Ty);
493 }
494
getRegUsageForType(Type * Ty) const495 unsigned TargetTransformInfo::getRegUsageForType(Type *Ty) const {
496 return TTIImpl->getRegUsageForType(Ty);
497 }
498
shouldBuildLookupTables() const499 bool TargetTransformInfo::shouldBuildLookupTables() const {
500 return TTIImpl->shouldBuildLookupTables();
501 }
502
shouldBuildLookupTablesForConstant(Constant * C) const503 bool TargetTransformInfo::shouldBuildLookupTablesForConstant(
504 Constant *C) const {
505 return TTIImpl->shouldBuildLookupTablesForConstant(C);
506 }
507
shouldBuildRelLookupTables() const508 bool TargetTransformInfo::shouldBuildRelLookupTables() const {
509 return TTIImpl->shouldBuildRelLookupTables();
510 }
511
useColdCCForColdCall(Function & F) const512 bool TargetTransformInfo::useColdCCForColdCall(Function &F) const {
513 return TTIImpl->useColdCCForColdCall(F);
514 }
515
getScalarizationOverhead(VectorType * Ty,const APInt & DemandedElts,bool Insert,bool Extract,TTI::TargetCostKind CostKind) const516 InstructionCost TargetTransformInfo::getScalarizationOverhead(
517 VectorType *Ty, const APInt &DemandedElts, bool Insert, bool Extract,
518 TTI::TargetCostKind CostKind) const {
519 return TTIImpl->getScalarizationOverhead(Ty, DemandedElts, Insert, Extract,
520 CostKind);
521 }
522
getOperandsScalarizationOverhead(ArrayRef<const Value * > Args,ArrayRef<Type * > Tys,TTI::TargetCostKind CostKind) const523 InstructionCost TargetTransformInfo::getOperandsScalarizationOverhead(
524 ArrayRef<const Value *> Args, ArrayRef<Type *> Tys,
525 TTI::TargetCostKind CostKind) const {
526 return TTIImpl->getOperandsScalarizationOverhead(Args, Tys, CostKind);
527 }
528
supportsEfficientVectorElementLoadStore() const529 bool TargetTransformInfo::supportsEfficientVectorElementLoadStore() const {
530 return TTIImpl->supportsEfficientVectorElementLoadStore();
531 }
532
supportsTailCalls() const533 bool TargetTransformInfo::supportsTailCalls() const {
534 return TTIImpl->supportsTailCalls();
535 }
536
supportsTailCallFor(const CallBase * CB) const537 bool TargetTransformInfo::supportsTailCallFor(const CallBase *CB) const {
538 return TTIImpl->supportsTailCallFor(CB);
539 }
540
enableAggressiveInterleaving(bool LoopHasReductions) const541 bool TargetTransformInfo::enableAggressiveInterleaving(
542 bool LoopHasReductions) const {
543 return TTIImpl->enableAggressiveInterleaving(LoopHasReductions);
544 }
545
546 TargetTransformInfo::MemCmpExpansionOptions
enableMemCmpExpansion(bool OptSize,bool IsZeroCmp) const547 TargetTransformInfo::enableMemCmpExpansion(bool OptSize, bool IsZeroCmp) const {
548 return TTIImpl->enableMemCmpExpansion(OptSize, IsZeroCmp);
549 }
550
enableSelectOptimize() const551 bool TargetTransformInfo::enableSelectOptimize() const {
552 return TTIImpl->enableSelectOptimize();
553 }
554
enableInterleavedAccessVectorization() const555 bool TargetTransformInfo::enableInterleavedAccessVectorization() const {
556 return TTIImpl->enableInterleavedAccessVectorization();
557 }
558
enableMaskedInterleavedAccessVectorization() const559 bool TargetTransformInfo::enableMaskedInterleavedAccessVectorization() const {
560 return TTIImpl->enableMaskedInterleavedAccessVectorization();
561 }
562
isFPVectorizationPotentiallyUnsafe() const563 bool TargetTransformInfo::isFPVectorizationPotentiallyUnsafe() const {
564 return TTIImpl->isFPVectorizationPotentiallyUnsafe();
565 }
566
567 bool
allowsMisalignedMemoryAccesses(LLVMContext & Context,unsigned BitWidth,unsigned AddressSpace,Align Alignment,unsigned * Fast) const568 TargetTransformInfo::allowsMisalignedMemoryAccesses(LLVMContext &Context,
569 unsigned BitWidth,
570 unsigned AddressSpace,
571 Align Alignment,
572 unsigned *Fast) const {
573 return TTIImpl->allowsMisalignedMemoryAccesses(Context, BitWidth,
574 AddressSpace, Alignment, Fast);
575 }
576
577 TargetTransformInfo::PopcntSupportKind
getPopcntSupport(unsigned IntTyWidthInBit) const578 TargetTransformInfo::getPopcntSupport(unsigned IntTyWidthInBit) const {
579 return TTIImpl->getPopcntSupport(IntTyWidthInBit);
580 }
581
haveFastSqrt(Type * Ty) const582 bool TargetTransformInfo::haveFastSqrt(Type *Ty) const {
583 return TTIImpl->haveFastSqrt(Ty);
584 }
585
isExpensiveToSpeculativelyExecute(const Instruction * I) const586 bool TargetTransformInfo::isExpensiveToSpeculativelyExecute(
587 const Instruction *I) const {
588 return TTIImpl->isExpensiveToSpeculativelyExecute(I);
589 }
590
isFCmpOrdCheaperThanFCmpZero(Type * Ty) const591 bool TargetTransformInfo::isFCmpOrdCheaperThanFCmpZero(Type *Ty) const {
592 return TTIImpl->isFCmpOrdCheaperThanFCmpZero(Ty);
593 }
594
getFPOpCost(Type * Ty) const595 InstructionCost TargetTransformInfo::getFPOpCost(Type *Ty) const {
596 InstructionCost Cost = TTIImpl->getFPOpCost(Ty);
597 assert(Cost >= 0 && "TTI should not produce negative costs!");
598 return Cost;
599 }
600
getIntImmCodeSizeCost(unsigned Opcode,unsigned Idx,const APInt & Imm,Type * Ty) const601 InstructionCost TargetTransformInfo::getIntImmCodeSizeCost(unsigned Opcode,
602 unsigned Idx,
603 const APInt &Imm,
604 Type *Ty) const {
605 InstructionCost Cost = TTIImpl->getIntImmCodeSizeCost(Opcode, Idx, Imm, Ty);
606 assert(Cost >= 0 && "TTI should not produce negative costs!");
607 return Cost;
608 }
609
610 InstructionCost
getIntImmCost(const APInt & Imm,Type * Ty,TTI::TargetCostKind CostKind) const611 TargetTransformInfo::getIntImmCost(const APInt &Imm, Type *Ty,
612 TTI::TargetCostKind CostKind) const {
613 InstructionCost Cost = TTIImpl->getIntImmCost(Imm, Ty, CostKind);
614 assert(Cost >= 0 && "TTI should not produce negative costs!");
615 return Cost;
616 }
617
getIntImmCostInst(unsigned Opcode,unsigned Idx,const APInt & Imm,Type * Ty,TTI::TargetCostKind CostKind,Instruction * Inst) const618 InstructionCost TargetTransformInfo::getIntImmCostInst(
619 unsigned Opcode, unsigned Idx, const APInt &Imm, Type *Ty,
620 TTI::TargetCostKind CostKind, Instruction *Inst) const {
621 InstructionCost Cost =
622 TTIImpl->getIntImmCostInst(Opcode, Idx, Imm, Ty, CostKind, Inst);
623 assert(Cost >= 0 && "TTI should not produce negative costs!");
624 return Cost;
625 }
626
627 InstructionCost
getIntImmCostIntrin(Intrinsic::ID IID,unsigned Idx,const APInt & Imm,Type * Ty,TTI::TargetCostKind CostKind) const628 TargetTransformInfo::getIntImmCostIntrin(Intrinsic::ID IID, unsigned Idx,
629 const APInt &Imm, Type *Ty,
630 TTI::TargetCostKind CostKind) const {
631 InstructionCost Cost =
632 TTIImpl->getIntImmCostIntrin(IID, Idx, Imm, Ty, CostKind);
633 assert(Cost >= 0 && "TTI should not produce negative costs!");
634 return Cost;
635 }
636
getNumberOfRegisters(unsigned ClassID) const637 unsigned TargetTransformInfo::getNumberOfRegisters(unsigned ClassID) const {
638 return TTIImpl->getNumberOfRegisters(ClassID);
639 }
640
getRegisterClassForType(bool Vector,Type * Ty) const641 unsigned TargetTransformInfo::getRegisterClassForType(bool Vector,
642 Type *Ty) const {
643 return TTIImpl->getRegisterClassForType(Vector, Ty);
644 }
645
getRegisterClassName(unsigned ClassID) const646 const char *TargetTransformInfo::getRegisterClassName(unsigned ClassID) const {
647 return TTIImpl->getRegisterClassName(ClassID);
648 }
649
getRegisterBitWidth(TargetTransformInfo::RegisterKind K) const650 TypeSize TargetTransformInfo::getRegisterBitWidth(
651 TargetTransformInfo::RegisterKind K) const {
652 return TTIImpl->getRegisterBitWidth(K);
653 }
654
getMinVectorRegisterBitWidth() const655 unsigned TargetTransformInfo::getMinVectorRegisterBitWidth() const {
656 return TTIImpl->getMinVectorRegisterBitWidth();
657 }
658
getMaxVScale() const659 std::optional<unsigned> TargetTransformInfo::getMaxVScale() const {
660 return TTIImpl->getMaxVScale();
661 }
662
getVScaleForTuning() const663 std::optional<unsigned> TargetTransformInfo::getVScaleForTuning() const {
664 return TTIImpl->getVScaleForTuning();
665 }
666
shouldMaximizeVectorBandwidth(TargetTransformInfo::RegisterKind K) const667 bool TargetTransformInfo::shouldMaximizeVectorBandwidth(
668 TargetTransformInfo::RegisterKind K) const {
669 return TTIImpl->shouldMaximizeVectorBandwidth(K);
670 }
671
getMinimumVF(unsigned ElemWidth,bool IsScalable) const672 ElementCount TargetTransformInfo::getMinimumVF(unsigned ElemWidth,
673 bool IsScalable) const {
674 return TTIImpl->getMinimumVF(ElemWidth, IsScalable);
675 }
676
getMaximumVF(unsigned ElemWidth,unsigned Opcode) const677 unsigned TargetTransformInfo::getMaximumVF(unsigned ElemWidth,
678 unsigned Opcode) const {
679 return TTIImpl->getMaximumVF(ElemWidth, Opcode);
680 }
681
getStoreMinimumVF(unsigned VF,Type * ScalarMemTy,Type * ScalarValTy) const682 unsigned TargetTransformInfo::getStoreMinimumVF(unsigned VF, Type *ScalarMemTy,
683 Type *ScalarValTy) const {
684 return TTIImpl->getStoreMinimumVF(VF, ScalarMemTy, ScalarValTy);
685 }
686
shouldConsiderAddressTypePromotion(const Instruction & I,bool & AllowPromotionWithoutCommonHeader) const687 bool TargetTransformInfo::shouldConsiderAddressTypePromotion(
688 const Instruction &I, bool &AllowPromotionWithoutCommonHeader) const {
689 return TTIImpl->shouldConsiderAddressTypePromotion(
690 I, AllowPromotionWithoutCommonHeader);
691 }
692
getCacheLineSize() const693 unsigned TargetTransformInfo::getCacheLineSize() const {
694 return CacheLineSize.getNumOccurrences() > 0 ? CacheLineSize
695 : TTIImpl->getCacheLineSize();
696 }
697
698 std::optional<unsigned>
getCacheSize(CacheLevel Level) const699 TargetTransformInfo::getCacheSize(CacheLevel Level) const {
700 return TTIImpl->getCacheSize(Level);
701 }
702
703 std::optional<unsigned>
getCacheAssociativity(CacheLevel Level) const704 TargetTransformInfo::getCacheAssociativity(CacheLevel Level) const {
705 return TTIImpl->getCacheAssociativity(Level);
706 }
707
getPrefetchDistance() const708 unsigned TargetTransformInfo::getPrefetchDistance() const {
709 return TTIImpl->getPrefetchDistance();
710 }
711
getMinPrefetchStride(unsigned NumMemAccesses,unsigned NumStridedMemAccesses,unsigned NumPrefetches,bool HasCall) const712 unsigned TargetTransformInfo::getMinPrefetchStride(
713 unsigned NumMemAccesses, unsigned NumStridedMemAccesses,
714 unsigned NumPrefetches, bool HasCall) const {
715 return TTIImpl->getMinPrefetchStride(NumMemAccesses, NumStridedMemAccesses,
716 NumPrefetches, HasCall);
717 }
718
getMaxPrefetchIterationsAhead() const719 unsigned TargetTransformInfo::getMaxPrefetchIterationsAhead() const {
720 return TTIImpl->getMaxPrefetchIterationsAhead();
721 }
722
enableWritePrefetching() const723 bool TargetTransformInfo::enableWritePrefetching() const {
724 return TTIImpl->enableWritePrefetching();
725 }
726
shouldPrefetchAddressSpace(unsigned AS) const727 bool TargetTransformInfo::shouldPrefetchAddressSpace(unsigned AS) const {
728 return TTIImpl->shouldPrefetchAddressSpace(AS);
729 }
730
getMaxInterleaveFactor(unsigned VF) const731 unsigned TargetTransformInfo::getMaxInterleaveFactor(unsigned VF) const {
732 return TTIImpl->getMaxInterleaveFactor(VF);
733 }
734
735 TargetTransformInfo::OperandValueInfo
getOperandInfo(const Value * V)736 TargetTransformInfo::getOperandInfo(const Value *V) {
737 OperandValueKind OpInfo = OK_AnyValue;
738 OperandValueProperties OpProps = OP_None;
739
740 if (isa<ConstantInt>(V) || isa<ConstantFP>(V)) {
741 if (const auto *CI = dyn_cast<ConstantInt>(V)) {
742 if (CI->getValue().isPowerOf2())
743 OpProps = OP_PowerOf2;
744 else if (CI->getValue().isNegatedPowerOf2())
745 OpProps = OP_NegatedPowerOf2;
746 }
747 return {OK_UniformConstantValue, OpProps};
748 }
749
750 // A broadcast shuffle creates a uniform value.
751 // TODO: Add support for non-zero index broadcasts.
752 // TODO: Add support for different source vector width.
753 if (const auto *ShuffleInst = dyn_cast<ShuffleVectorInst>(V))
754 if (ShuffleInst->isZeroEltSplat())
755 OpInfo = OK_UniformValue;
756
757 const Value *Splat = getSplatValue(V);
758
759 // Check for a splat of a constant or for a non uniform vector of constants
760 // and check if the constant(s) are all powers of two.
761 if (isa<ConstantVector>(V) || isa<ConstantDataVector>(V)) {
762 OpInfo = OK_NonUniformConstantValue;
763 if (Splat) {
764 OpInfo = OK_UniformConstantValue;
765 if (auto *CI = dyn_cast<ConstantInt>(Splat)) {
766 if (CI->getValue().isPowerOf2())
767 OpProps = OP_PowerOf2;
768 else if (CI->getValue().isNegatedPowerOf2())
769 OpProps = OP_NegatedPowerOf2;
770 }
771 } else if (const auto *CDS = dyn_cast<ConstantDataSequential>(V)) {
772 bool AllPow2 = true, AllNegPow2 = true;
773 for (unsigned I = 0, E = CDS->getNumElements(); I != E; ++I) {
774 if (auto *CI = dyn_cast<ConstantInt>(CDS->getElementAsConstant(I))) {
775 AllPow2 &= CI->getValue().isPowerOf2();
776 AllNegPow2 &= CI->getValue().isNegatedPowerOf2();
777 if (AllPow2 || AllNegPow2)
778 continue;
779 }
780 AllPow2 = AllNegPow2 = false;
781 break;
782 }
783 OpProps = AllPow2 ? OP_PowerOf2 : OpProps;
784 OpProps = AllNegPow2 ? OP_NegatedPowerOf2 : OpProps;
785 }
786 }
787
788 // Check for a splat of a uniform value. This is not loop aware, so return
789 // true only for the obviously uniform cases (argument, globalvalue)
790 if (Splat && (isa<Argument>(Splat) || isa<GlobalValue>(Splat)))
791 OpInfo = OK_UniformValue;
792
793 return {OpInfo, OpProps};
794 }
795
getArithmeticInstrCost(unsigned Opcode,Type * Ty,TTI::TargetCostKind CostKind,OperandValueInfo Op1Info,OperandValueInfo Op2Info,ArrayRef<const Value * > Args,const Instruction * CxtI) const796 InstructionCost TargetTransformInfo::getArithmeticInstrCost(
797 unsigned Opcode, Type *Ty, TTI::TargetCostKind CostKind,
798 OperandValueInfo Op1Info, OperandValueInfo Op2Info,
799 ArrayRef<const Value *> Args, const Instruction *CxtI) const {
800 InstructionCost Cost =
801 TTIImpl->getArithmeticInstrCost(Opcode, Ty, CostKind,
802 Op1Info, Op2Info,
803 Args, CxtI);
804 assert(Cost >= 0 && "TTI should not produce negative costs!");
805 return Cost;
806 }
807
getShuffleCost(ShuffleKind Kind,VectorType * Ty,ArrayRef<int> Mask,TTI::TargetCostKind CostKind,int Index,VectorType * SubTp,ArrayRef<const Value * > Args) const808 InstructionCost TargetTransformInfo::getShuffleCost(
809 ShuffleKind Kind, VectorType *Ty, ArrayRef<int> Mask,
810 TTI::TargetCostKind CostKind, int Index, VectorType *SubTp,
811 ArrayRef<const Value *> Args) const {
812 InstructionCost Cost =
813 TTIImpl->getShuffleCost(Kind, Ty, Mask, CostKind, Index, SubTp, Args);
814 assert(Cost >= 0 && "TTI should not produce negative costs!");
815 return Cost;
816 }
817
818 TTI::CastContextHint
getCastContextHint(const Instruction * I)819 TargetTransformInfo::getCastContextHint(const Instruction *I) {
820 if (!I)
821 return CastContextHint::None;
822
823 auto getLoadStoreKind = [](const Value *V, unsigned LdStOp, unsigned MaskedOp,
824 unsigned GatScatOp) {
825 const Instruction *I = dyn_cast<Instruction>(V);
826 if (!I)
827 return CastContextHint::None;
828
829 if (I->getOpcode() == LdStOp)
830 return CastContextHint::Normal;
831
832 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
833 if (II->getIntrinsicID() == MaskedOp)
834 return TTI::CastContextHint::Masked;
835 if (II->getIntrinsicID() == GatScatOp)
836 return TTI::CastContextHint::GatherScatter;
837 }
838
839 return TTI::CastContextHint::None;
840 };
841
842 switch (I->getOpcode()) {
843 case Instruction::ZExt:
844 case Instruction::SExt:
845 case Instruction::FPExt:
846 return getLoadStoreKind(I->getOperand(0), Instruction::Load,
847 Intrinsic::masked_load, Intrinsic::masked_gather);
848 case Instruction::Trunc:
849 case Instruction::FPTrunc:
850 if (I->hasOneUse())
851 return getLoadStoreKind(*I->user_begin(), Instruction::Store,
852 Intrinsic::masked_store,
853 Intrinsic::masked_scatter);
854 break;
855 default:
856 return CastContextHint::None;
857 }
858
859 return TTI::CastContextHint::None;
860 }
861
getCastInstrCost(unsigned Opcode,Type * Dst,Type * Src,CastContextHint CCH,TTI::TargetCostKind CostKind,const Instruction * I) const862 InstructionCost TargetTransformInfo::getCastInstrCost(
863 unsigned Opcode, Type *Dst, Type *Src, CastContextHint CCH,
864 TTI::TargetCostKind CostKind, const Instruction *I) const {
865 assert((I == nullptr || I->getOpcode() == Opcode) &&
866 "Opcode should reflect passed instruction.");
867 InstructionCost Cost =
868 TTIImpl->getCastInstrCost(Opcode, Dst, Src, CCH, CostKind, I);
869 assert(Cost >= 0 && "TTI should not produce negative costs!");
870 return Cost;
871 }
872
getExtractWithExtendCost(unsigned Opcode,Type * Dst,VectorType * VecTy,unsigned Index) const873 InstructionCost TargetTransformInfo::getExtractWithExtendCost(
874 unsigned Opcode, Type *Dst, VectorType *VecTy, unsigned Index) const {
875 InstructionCost Cost =
876 TTIImpl->getExtractWithExtendCost(Opcode, Dst, VecTy, Index);
877 assert(Cost >= 0 && "TTI should not produce negative costs!");
878 return Cost;
879 }
880
getCFInstrCost(unsigned Opcode,TTI::TargetCostKind CostKind,const Instruction * I) const881 InstructionCost TargetTransformInfo::getCFInstrCost(
882 unsigned Opcode, TTI::TargetCostKind CostKind, const Instruction *I) const {
883 assert((I == nullptr || I->getOpcode() == Opcode) &&
884 "Opcode should reflect passed instruction.");
885 InstructionCost Cost = TTIImpl->getCFInstrCost(Opcode, CostKind, I);
886 assert(Cost >= 0 && "TTI should not produce negative costs!");
887 return Cost;
888 }
889
getCmpSelInstrCost(unsigned Opcode,Type * ValTy,Type * CondTy,CmpInst::Predicate VecPred,TTI::TargetCostKind CostKind,const Instruction * I) const890 InstructionCost TargetTransformInfo::getCmpSelInstrCost(
891 unsigned Opcode, Type *ValTy, Type *CondTy, CmpInst::Predicate VecPred,
892 TTI::TargetCostKind CostKind, const Instruction *I) const {
893 assert((I == nullptr || I->getOpcode() == Opcode) &&
894 "Opcode should reflect passed instruction.");
895 InstructionCost Cost =
896 TTIImpl->getCmpSelInstrCost(Opcode, ValTy, CondTy, VecPred, CostKind, I);
897 assert(Cost >= 0 && "TTI should not produce negative costs!");
898 return Cost;
899 }
900
getVectorInstrCost(unsigned Opcode,Type * Val,TTI::TargetCostKind CostKind,unsigned Index,Value * Op0,Value * Op1) const901 InstructionCost TargetTransformInfo::getVectorInstrCost(
902 unsigned Opcode, Type *Val, TTI::TargetCostKind CostKind, unsigned Index,
903 Value *Op0, Value *Op1) const {
904 // FIXME: Assert that Opcode is either InsertElement or ExtractElement.
905 // This is mentioned in the interface description and respected by all
906 // callers, but never asserted upon.
907 InstructionCost Cost =
908 TTIImpl->getVectorInstrCost(Opcode, Val, CostKind, Index, Op0, Op1);
909 assert(Cost >= 0 && "TTI should not produce negative costs!");
910 return Cost;
911 }
912
913 InstructionCost
getVectorInstrCost(const Instruction & I,Type * Val,TTI::TargetCostKind CostKind,unsigned Index) const914 TargetTransformInfo::getVectorInstrCost(const Instruction &I, Type *Val,
915 TTI::TargetCostKind CostKind,
916 unsigned Index) const {
917 // FIXME: Assert that Opcode is either InsertElement or ExtractElement.
918 // This is mentioned in the interface description and respected by all
919 // callers, but never asserted upon.
920 InstructionCost Cost = TTIImpl->getVectorInstrCost(I, Val, CostKind, Index);
921 assert(Cost >= 0 && "TTI should not produce negative costs!");
922 return Cost;
923 }
924
getReplicationShuffleCost(Type * EltTy,int ReplicationFactor,int VF,const APInt & DemandedDstElts,TTI::TargetCostKind CostKind)925 InstructionCost TargetTransformInfo::getReplicationShuffleCost(
926 Type *EltTy, int ReplicationFactor, int VF, const APInt &DemandedDstElts,
927 TTI::TargetCostKind CostKind) {
928 InstructionCost Cost = TTIImpl->getReplicationShuffleCost(
929 EltTy, ReplicationFactor, VF, DemandedDstElts, CostKind);
930 assert(Cost >= 0 && "TTI should not produce negative costs!");
931 return Cost;
932 }
933
getMemoryOpCost(unsigned Opcode,Type * Src,Align Alignment,unsigned AddressSpace,TTI::TargetCostKind CostKind,TTI::OperandValueInfo OpInfo,const Instruction * I) const934 InstructionCost TargetTransformInfo::getMemoryOpCost(
935 unsigned Opcode, Type *Src, Align Alignment, unsigned AddressSpace,
936 TTI::TargetCostKind CostKind, TTI::OperandValueInfo OpInfo,
937 const Instruction *I) const {
938 assert((I == nullptr || I->getOpcode() == Opcode) &&
939 "Opcode should reflect passed instruction.");
940 InstructionCost Cost = TTIImpl->getMemoryOpCost(
941 Opcode, Src, Alignment, AddressSpace, CostKind, OpInfo, I);
942 assert(Cost >= 0 && "TTI should not produce negative costs!");
943 return Cost;
944 }
945
getMaskedMemoryOpCost(unsigned Opcode,Type * Src,Align Alignment,unsigned AddressSpace,TTI::TargetCostKind CostKind) const946 InstructionCost TargetTransformInfo::getMaskedMemoryOpCost(
947 unsigned Opcode, Type *Src, Align Alignment, unsigned AddressSpace,
948 TTI::TargetCostKind CostKind) const {
949 InstructionCost Cost = TTIImpl->getMaskedMemoryOpCost(Opcode, Src, Alignment,
950 AddressSpace, CostKind);
951 assert(Cost >= 0 && "TTI should not produce negative costs!");
952 return Cost;
953 }
954
getGatherScatterOpCost(unsigned Opcode,Type * DataTy,const Value * Ptr,bool VariableMask,Align Alignment,TTI::TargetCostKind CostKind,const Instruction * I) const955 InstructionCost TargetTransformInfo::getGatherScatterOpCost(
956 unsigned Opcode, Type *DataTy, const Value *Ptr, bool VariableMask,
957 Align Alignment, TTI::TargetCostKind CostKind, const Instruction *I) const {
958 InstructionCost Cost = TTIImpl->getGatherScatterOpCost(
959 Opcode, DataTy, Ptr, VariableMask, Alignment, CostKind, I);
960 assert(Cost >= 0 && "TTI should not produce negative costs!");
961 return Cost;
962 }
963
getInterleavedMemoryOpCost(unsigned Opcode,Type * VecTy,unsigned Factor,ArrayRef<unsigned> Indices,Align Alignment,unsigned AddressSpace,TTI::TargetCostKind CostKind,bool UseMaskForCond,bool UseMaskForGaps) const964 InstructionCost TargetTransformInfo::getInterleavedMemoryOpCost(
965 unsigned Opcode, Type *VecTy, unsigned Factor, ArrayRef<unsigned> Indices,
966 Align Alignment, unsigned AddressSpace, TTI::TargetCostKind CostKind,
967 bool UseMaskForCond, bool UseMaskForGaps) const {
968 InstructionCost Cost = TTIImpl->getInterleavedMemoryOpCost(
969 Opcode, VecTy, Factor, Indices, Alignment, AddressSpace, CostKind,
970 UseMaskForCond, UseMaskForGaps);
971 assert(Cost >= 0 && "TTI should not produce negative costs!");
972 return Cost;
973 }
974
975 InstructionCost
getIntrinsicInstrCost(const IntrinsicCostAttributes & ICA,TTI::TargetCostKind CostKind) const976 TargetTransformInfo::getIntrinsicInstrCost(const IntrinsicCostAttributes &ICA,
977 TTI::TargetCostKind CostKind) const {
978 InstructionCost Cost = TTIImpl->getIntrinsicInstrCost(ICA, CostKind);
979 assert(Cost >= 0 && "TTI should not produce negative costs!");
980 return Cost;
981 }
982
983 InstructionCost
getCallInstrCost(Function * F,Type * RetTy,ArrayRef<Type * > Tys,TTI::TargetCostKind CostKind) const984 TargetTransformInfo::getCallInstrCost(Function *F, Type *RetTy,
985 ArrayRef<Type *> Tys,
986 TTI::TargetCostKind CostKind) const {
987 InstructionCost Cost = TTIImpl->getCallInstrCost(F, RetTy, Tys, CostKind);
988 assert(Cost >= 0 && "TTI should not produce negative costs!");
989 return Cost;
990 }
991
getNumberOfParts(Type * Tp) const992 unsigned TargetTransformInfo::getNumberOfParts(Type *Tp) const {
993 return TTIImpl->getNumberOfParts(Tp);
994 }
995
996 InstructionCost
getAddressComputationCost(Type * Tp,ScalarEvolution * SE,const SCEV * Ptr) const997 TargetTransformInfo::getAddressComputationCost(Type *Tp, ScalarEvolution *SE,
998 const SCEV *Ptr) const {
999 InstructionCost Cost = TTIImpl->getAddressComputationCost(Tp, SE, Ptr);
1000 assert(Cost >= 0 && "TTI should not produce negative costs!");
1001 return Cost;
1002 }
1003
getMemcpyCost(const Instruction * I) const1004 InstructionCost TargetTransformInfo::getMemcpyCost(const Instruction *I) const {
1005 InstructionCost Cost = TTIImpl->getMemcpyCost(I);
1006 assert(Cost >= 0 && "TTI should not produce negative costs!");
1007 return Cost;
1008 }
1009
getArithmeticReductionCost(unsigned Opcode,VectorType * Ty,std::optional<FastMathFlags> FMF,TTI::TargetCostKind CostKind) const1010 InstructionCost TargetTransformInfo::getArithmeticReductionCost(
1011 unsigned Opcode, VectorType *Ty, std::optional<FastMathFlags> FMF,
1012 TTI::TargetCostKind CostKind) const {
1013 InstructionCost Cost =
1014 TTIImpl->getArithmeticReductionCost(Opcode, Ty, FMF, CostKind);
1015 assert(Cost >= 0 && "TTI should not produce negative costs!");
1016 return Cost;
1017 }
1018
getMinMaxReductionCost(VectorType * Ty,VectorType * CondTy,bool IsUnsigned,TTI::TargetCostKind CostKind) const1019 InstructionCost TargetTransformInfo::getMinMaxReductionCost(
1020 VectorType *Ty, VectorType *CondTy, bool IsUnsigned,
1021 TTI::TargetCostKind CostKind) const {
1022 InstructionCost Cost =
1023 TTIImpl->getMinMaxReductionCost(Ty, CondTy, IsUnsigned, CostKind);
1024 assert(Cost >= 0 && "TTI should not produce negative costs!");
1025 return Cost;
1026 }
1027
getExtendedReductionCost(unsigned Opcode,bool IsUnsigned,Type * ResTy,VectorType * Ty,std::optional<FastMathFlags> FMF,TTI::TargetCostKind CostKind) const1028 InstructionCost TargetTransformInfo::getExtendedReductionCost(
1029 unsigned Opcode, bool IsUnsigned, Type *ResTy, VectorType *Ty,
1030 std::optional<FastMathFlags> FMF, TTI::TargetCostKind CostKind) const {
1031 return TTIImpl->getExtendedReductionCost(Opcode, IsUnsigned, ResTy, Ty, FMF,
1032 CostKind);
1033 }
1034
getMulAccReductionCost(bool IsUnsigned,Type * ResTy,VectorType * Ty,TTI::TargetCostKind CostKind) const1035 InstructionCost TargetTransformInfo::getMulAccReductionCost(
1036 bool IsUnsigned, Type *ResTy, VectorType *Ty,
1037 TTI::TargetCostKind CostKind) const {
1038 return TTIImpl->getMulAccReductionCost(IsUnsigned, ResTy, Ty, CostKind);
1039 }
1040
1041 InstructionCost
getCostOfKeepingLiveOverCall(ArrayRef<Type * > Tys) const1042 TargetTransformInfo::getCostOfKeepingLiveOverCall(ArrayRef<Type *> Tys) const {
1043 return TTIImpl->getCostOfKeepingLiveOverCall(Tys);
1044 }
1045
getTgtMemIntrinsic(IntrinsicInst * Inst,MemIntrinsicInfo & Info) const1046 bool TargetTransformInfo::getTgtMemIntrinsic(IntrinsicInst *Inst,
1047 MemIntrinsicInfo &Info) const {
1048 return TTIImpl->getTgtMemIntrinsic(Inst, Info);
1049 }
1050
getAtomicMemIntrinsicMaxElementSize() const1051 unsigned TargetTransformInfo::getAtomicMemIntrinsicMaxElementSize() const {
1052 return TTIImpl->getAtomicMemIntrinsicMaxElementSize();
1053 }
1054
getOrCreateResultFromMemIntrinsic(IntrinsicInst * Inst,Type * ExpectedType) const1055 Value *TargetTransformInfo::getOrCreateResultFromMemIntrinsic(
1056 IntrinsicInst *Inst, Type *ExpectedType) const {
1057 return TTIImpl->getOrCreateResultFromMemIntrinsic(Inst, ExpectedType);
1058 }
1059
getMemcpyLoopLoweringType(LLVMContext & Context,Value * Length,unsigned SrcAddrSpace,unsigned DestAddrSpace,unsigned SrcAlign,unsigned DestAlign,std::optional<uint32_t> AtomicElementSize) const1060 Type *TargetTransformInfo::getMemcpyLoopLoweringType(
1061 LLVMContext &Context, Value *Length, unsigned SrcAddrSpace,
1062 unsigned DestAddrSpace, unsigned SrcAlign, unsigned DestAlign,
1063 std::optional<uint32_t> AtomicElementSize) const {
1064 return TTIImpl->getMemcpyLoopLoweringType(Context, Length, SrcAddrSpace,
1065 DestAddrSpace, SrcAlign, DestAlign,
1066 AtomicElementSize);
1067 }
1068
getMemcpyLoopResidualLoweringType(SmallVectorImpl<Type * > & OpsOut,LLVMContext & Context,unsigned RemainingBytes,unsigned SrcAddrSpace,unsigned DestAddrSpace,unsigned SrcAlign,unsigned DestAlign,std::optional<uint32_t> AtomicCpySize) const1069 void TargetTransformInfo::getMemcpyLoopResidualLoweringType(
1070 SmallVectorImpl<Type *> &OpsOut, LLVMContext &Context,
1071 unsigned RemainingBytes, unsigned SrcAddrSpace, unsigned DestAddrSpace,
1072 unsigned SrcAlign, unsigned DestAlign,
1073 std::optional<uint32_t> AtomicCpySize) const {
1074 TTIImpl->getMemcpyLoopResidualLoweringType(
1075 OpsOut, Context, RemainingBytes, SrcAddrSpace, DestAddrSpace, SrcAlign,
1076 DestAlign, AtomicCpySize);
1077 }
1078
areInlineCompatible(const Function * Caller,const Function * Callee) const1079 bool TargetTransformInfo::areInlineCompatible(const Function *Caller,
1080 const Function *Callee) const {
1081 return TTIImpl->areInlineCompatible(Caller, Callee);
1082 }
1083
areTypesABICompatible(const Function * Caller,const Function * Callee,const ArrayRef<Type * > & Types) const1084 bool TargetTransformInfo::areTypesABICompatible(
1085 const Function *Caller, const Function *Callee,
1086 const ArrayRef<Type *> &Types) const {
1087 return TTIImpl->areTypesABICompatible(Caller, Callee, Types);
1088 }
1089
isIndexedLoadLegal(MemIndexedMode Mode,Type * Ty) const1090 bool TargetTransformInfo::isIndexedLoadLegal(MemIndexedMode Mode,
1091 Type *Ty) const {
1092 return TTIImpl->isIndexedLoadLegal(Mode, Ty);
1093 }
1094
isIndexedStoreLegal(MemIndexedMode Mode,Type * Ty) const1095 bool TargetTransformInfo::isIndexedStoreLegal(MemIndexedMode Mode,
1096 Type *Ty) const {
1097 return TTIImpl->isIndexedStoreLegal(Mode, Ty);
1098 }
1099
getLoadStoreVecRegBitWidth(unsigned AS) const1100 unsigned TargetTransformInfo::getLoadStoreVecRegBitWidth(unsigned AS) const {
1101 return TTIImpl->getLoadStoreVecRegBitWidth(AS);
1102 }
1103
isLegalToVectorizeLoad(LoadInst * LI) const1104 bool TargetTransformInfo::isLegalToVectorizeLoad(LoadInst *LI) const {
1105 return TTIImpl->isLegalToVectorizeLoad(LI);
1106 }
1107
isLegalToVectorizeStore(StoreInst * SI) const1108 bool TargetTransformInfo::isLegalToVectorizeStore(StoreInst *SI) const {
1109 return TTIImpl->isLegalToVectorizeStore(SI);
1110 }
1111
isLegalToVectorizeLoadChain(unsigned ChainSizeInBytes,Align Alignment,unsigned AddrSpace) const1112 bool TargetTransformInfo::isLegalToVectorizeLoadChain(
1113 unsigned ChainSizeInBytes, Align Alignment, unsigned AddrSpace) const {
1114 return TTIImpl->isLegalToVectorizeLoadChain(ChainSizeInBytes, Alignment,
1115 AddrSpace);
1116 }
1117
isLegalToVectorizeStoreChain(unsigned ChainSizeInBytes,Align Alignment,unsigned AddrSpace) const1118 bool TargetTransformInfo::isLegalToVectorizeStoreChain(
1119 unsigned ChainSizeInBytes, Align Alignment, unsigned AddrSpace) const {
1120 return TTIImpl->isLegalToVectorizeStoreChain(ChainSizeInBytes, Alignment,
1121 AddrSpace);
1122 }
1123
isLegalToVectorizeReduction(const RecurrenceDescriptor & RdxDesc,ElementCount VF) const1124 bool TargetTransformInfo::isLegalToVectorizeReduction(
1125 const RecurrenceDescriptor &RdxDesc, ElementCount VF) const {
1126 return TTIImpl->isLegalToVectorizeReduction(RdxDesc, VF);
1127 }
1128
isElementTypeLegalForScalableVector(Type * Ty) const1129 bool TargetTransformInfo::isElementTypeLegalForScalableVector(Type *Ty) const {
1130 return TTIImpl->isElementTypeLegalForScalableVector(Ty);
1131 }
1132
getLoadVectorFactor(unsigned VF,unsigned LoadSize,unsigned ChainSizeInBytes,VectorType * VecTy) const1133 unsigned TargetTransformInfo::getLoadVectorFactor(unsigned VF,
1134 unsigned LoadSize,
1135 unsigned ChainSizeInBytes,
1136 VectorType *VecTy) const {
1137 return TTIImpl->getLoadVectorFactor(VF, LoadSize, ChainSizeInBytes, VecTy);
1138 }
1139
getStoreVectorFactor(unsigned VF,unsigned StoreSize,unsigned ChainSizeInBytes,VectorType * VecTy) const1140 unsigned TargetTransformInfo::getStoreVectorFactor(unsigned VF,
1141 unsigned StoreSize,
1142 unsigned ChainSizeInBytes,
1143 VectorType *VecTy) const {
1144 return TTIImpl->getStoreVectorFactor(VF, StoreSize, ChainSizeInBytes, VecTy);
1145 }
1146
preferInLoopReduction(unsigned Opcode,Type * Ty,ReductionFlags Flags) const1147 bool TargetTransformInfo::preferInLoopReduction(unsigned Opcode, Type *Ty,
1148 ReductionFlags Flags) const {
1149 return TTIImpl->preferInLoopReduction(Opcode, Ty, Flags);
1150 }
1151
preferPredicatedReductionSelect(unsigned Opcode,Type * Ty,ReductionFlags Flags) const1152 bool TargetTransformInfo::preferPredicatedReductionSelect(
1153 unsigned Opcode, Type *Ty, ReductionFlags Flags) const {
1154 return TTIImpl->preferPredicatedReductionSelect(Opcode, Ty, Flags);
1155 }
1156
preferEpilogueVectorization() const1157 bool TargetTransformInfo::preferEpilogueVectorization() const {
1158 return TTIImpl->preferEpilogueVectorization();
1159 }
1160
1161 TargetTransformInfo::VPLegalization
getVPLegalizationStrategy(const VPIntrinsic & VPI) const1162 TargetTransformInfo::getVPLegalizationStrategy(const VPIntrinsic &VPI) const {
1163 return TTIImpl->getVPLegalizationStrategy(VPI);
1164 }
1165
shouldExpandReduction(const IntrinsicInst * II) const1166 bool TargetTransformInfo::shouldExpandReduction(const IntrinsicInst *II) const {
1167 return TTIImpl->shouldExpandReduction(II);
1168 }
1169
getGISelRematGlobalCost() const1170 unsigned TargetTransformInfo::getGISelRematGlobalCost() const {
1171 return TTIImpl->getGISelRematGlobalCost();
1172 }
1173
getMinTripCountTailFoldingThreshold() const1174 unsigned TargetTransformInfo::getMinTripCountTailFoldingThreshold() const {
1175 return TTIImpl->getMinTripCountTailFoldingThreshold();
1176 }
1177
supportsScalableVectors() const1178 bool TargetTransformInfo::supportsScalableVectors() const {
1179 return TTIImpl->supportsScalableVectors();
1180 }
1181
enableScalableVectorization() const1182 bool TargetTransformInfo::enableScalableVectorization() const {
1183 return TTIImpl->enableScalableVectorization();
1184 }
1185
hasActiveVectorLength(unsigned Opcode,Type * DataType,Align Alignment) const1186 bool TargetTransformInfo::hasActiveVectorLength(unsigned Opcode, Type *DataType,
1187 Align Alignment) const {
1188 return TTIImpl->hasActiveVectorLength(Opcode, DataType, Alignment);
1189 }
1190
1191 TargetTransformInfo::Concept::~Concept() = default;
1192
TargetIRAnalysis()1193 TargetIRAnalysis::TargetIRAnalysis() : TTICallback(&getDefaultTTI) {}
1194
TargetIRAnalysis(std::function<Result (const Function &)> TTICallback)1195 TargetIRAnalysis::TargetIRAnalysis(
1196 std::function<Result(const Function &)> TTICallback)
1197 : TTICallback(std::move(TTICallback)) {}
1198
run(const Function & F,FunctionAnalysisManager &)1199 TargetIRAnalysis::Result TargetIRAnalysis::run(const Function &F,
1200 FunctionAnalysisManager &) {
1201 return TTICallback(F);
1202 }
1203
1204 AnalysisKey TargetIRAnalysis::Key;
1205
getDefaultTTI(const Function & F)1206 TargetIRAnalysis::Result TargetIRAnalysis::getDefaultTTI(const Function &F) {
1207 return Result(F.getParent()->getDataLayout());
1208 }
1209
1210 // Register the basic pass.
1211 INITIALIZE_PASS(TargetTransformInfoWrapperPass, "tti",
1212 "Target Transform Information", false, true)
1213 char TargetTransformInfoWrapperPass::ID = 0;
1214
anchor()1215 void TargetTransformInfoWrapperPass::anchor() {}
1216
TargetTransformInfoWrapperPass()1217 TargetTransformInfoWrapperPass::TargetTransformInfoWrapperPass()
1218 : ImmutablePass(ID) {
1219 initializeTargetTransformInfoWrapperPassPass(
1220 *PassRegistry::getPassRegistry());
1221 }
1222
TargetTransformInfoWrapperPass(TargetIRAnalysis TIRA)1223 TargetTransformInfoWrapperPass::TargetTransformInfoWrapperPass(
1224 TargetIRAnalysis TIRA)
1225 : ImmutablePass(ID), TIRA(std::move(TIRA)) {
1226 initializeTargetTransformInfoWrapperPassPass(
1227 *PassRegistry::getPassRegistry());
1228 }
1229
getTTI(const Function & F)1230 TargetTransformInfo &TargetTransformInfoWrapperPass::getTTI(const Function &F) {
1231 FunctionAnalysisManager DummyFAM;
1232 TTI = TIRA.run(F, DummyFAM);
1233 return *TTI;
1234 }
1235
1236 ImmutablePass *
createTargetTransformInfoWrapperPass(TargetIRAnalysis TIRA)1237 llvm::createTargetTransformInfoWrapperPass(TargetIRAnalysis TIRA) {
1238 return new TargetTransformInfoWrapperPass(std::move(TIRA));
1239 }
1240