1 //===- llvm/Analysis/TargetTransformInfo.cpp ------------------------------===//
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
7 //===----------------------------------------------------------------------===//
9 #include "llvm/Analysis/TargetTransformInfo.h"
10 #include "llvm/Analysis/CFG.h"
11 #include "llvm/Analysis/LoopIterator.h"
12 #include "llvm/Analysis/TargetLibraryInfo.h"
13 #include "llvm/Analysis/TargetTransformInfoImpl.h"
14 #include "llvm/IR/CFG.h"
15 #include "llvm/IR/Dominators.h"
16 #include "llvm/IR/Instruction.h"
17 #include "llvm/IR/Instructions.h"
18 #include "llvm/IR/IntrinsicInst.h"
19 #include "llvm/IR/Module.h"
20 #include "llvm/IR/Operator.h"
21 #include "llvm/IR/PatternMatch.h"
22 #include "llvm/InitializePasses.h"
23 #include "llvm/Support/CommandLine.h"
28 using namespace PatternMatch
;
30 #define DEBUG_TYPE "tti"
32 static cl::opt
<bool> EnableReduxCost("costmodel-reduxcost", cl::init(false),
34 cl::desc("Recognize reduction patterns."));
36 static cl::opt
<unsigned> CacheLineSize(
37 "cache-line-size", cl::init(0), cl::Hidden
,
38 cl::desc("Use this to override the target cache line size when "
39 "specified by the user."));
41 static cl::opt
<unsigned> MinPageSize(
42 "min-page-size", cl::init(0), cl::Hidden
,
43 cl::desc("Use this to override the target's minimum page size."));
45 static cl::opt
<unsigned> PredictableBranchThreshold(
46 "predictable-branch-threshold", cl::init(99), cl::Hidden
,
48 "Use this to override the target's predictable branch threshold (%)."));
51 /// No-op implementation of the TTI interface using the utility base
54 /// This is used when no target specific information is available.
55 struct NoTTIImpl
: TargetTransformInfoImplCRTPBase
<NoTTIImpl
> {
56 explicit NoTTIImpl(const DataLayout
&DL
)
57 : TargetTransformInfoImplCRTPBase
<NoTTIImpl
>(DL
) {}
61 bool HardwareLoopInfo::canAnalyze(LoopInfo
&LI
) {
62 // If the loop has irreducible control flow, it can not be converted to
64 LoopBlocksRPO
RPOT(L
);
66 if (containsIrreducibleCFG
<const BasicBlock
*>(RPOT
, LI
))
71 IntrinsicCostAttributes::IntrinsicCostAttributes(
72 Intrinsic::ID Id
, const CallBase
&CI
, InstructionCost ScalarizationCost
,
74 : II(dyn_cast
<IntrinsicInst
>(&CI
)), RetTy(CI
.getType()), IID(Id
),
75 ScalarizationCost(ScalarizationCost
) {
77 if (const auto *FPMO
= dyn_cast
<FPMathOperator
>(&CI
))
78 FMF
= FPMO
->getFastMathFlags();
81 Arguments
.insert(Arguments
.begin(), CI
.arg_begin(), CI
.arg_end());
82 FunctionType
*FTy
= CI
.getCalledFunction()->getFunctionType();
83 ParamTys
.insert(ParamTys
.begin(), FTy
->param_begin(), FTy
->param_end());
86 IntrinsicCostAttributes::IntrinsicCostAttributes(Intrinsic::ID Id
, Type
*RTy
,
89 const IntrinsicInst
*I
,
90 InstructionCost ScalarCost
)
91 : II(I
), RetTy(RTy
), IID(Id
), FMF(Flags
), ScalarizationCost(ScalarCost
) {
92 ParamTys
.insert(ParamTys
.begin(), Tys
.begin(), Tys
.end());
95 IntrinsicCostAttributes::IntrinsicCostAttributes(Intrinsic::ID Id
, Type
*Ty
,
96 ArrayRef
<const Value
*> Args
)
97 : RetTy(Ty
), IID(Id
) {
99 Arguments
.insert(Arguments
.begin(), Args
.begin(), Args
.end());
100 ParamTys
.reserve(Arguments
.size());
101 for (const Value
*Argument
: Arguments
)
102 ParamTys
.push_back(Argument
->getType());
105 IntrinsicCostAttributes::IntrinsicCostAttributes(Intrinsic::ID Id
, Type
*RTy
,
106 ArrayRef
<const Value
*> Args
,
107 ArrayRef
<Type
*> Tys
,
109 const IntrinsicInst
*I
,
110 InstructionCost ScalarCost
)
111 : II(I
), RetTy(RTy
), IID(Id
), FMF(Flags
), ScalarizationCost(ScalarCost
) {
112 ParamTys
.insert(ParamTys
.begin(), Tys
.begin(), Tys
.end());
113 Arguments
.insert(Arguments
.begin(), Args
.begin(), Args
.end());
116 HardwareLoopInfo::HardwareLoopInfo(Loop
*L
) : L(L
) {
117 // Match default options:
118 // - hardware-loop-counter-bitwidth = 32
119 // - hardware-loop-decrement = 1
120 CountType
= Type::getInt32Ty(L
->getHeader()->getContext());
121 LoopDecrement
= ConstantInt::get(CountType
, 1);
124 bool HardwareLoopInfo::isHardwareLoopCandidate(ScalarEvolution
&SE
,
125 LoopInfo
&LI
, DominatorTree
&DT
,
126 bool ForceNestedLoop
,
127 bool ForceHardwareLoopPHI
) {
128 SmallVector
<BasicBlock
*, 4> ExitingBlocks
;
129 L
->getExitingBlocks(ExitingBlocks
);
131 for (BasicBlock
*BB
: ExitingBlocks
) {
132 // If we pass the updated counter back through a phi, we need to know
133 // which latch the updated value will be coming from.
134 if (!L
->isLoopLatch(BB
)) {
135 if (ForceHardwareLoopPHI
|| CounterInReg
)
139 const SCEV
*EC
= SE
.getExitCount(L
, BB
);
140 if (isa
<SCEVCouldNotCompute
>(EC
))
142 if (const SCEVConstant
*ConstEC
= dyn_cast
<SCEVConstant
>(EC
)) {
143 if (ConstEC
->getValue()->isZero())
145 } else if (!SE
.isLoopInvariant(EC
, L
))
148 if (SE
.getTypeSizeInBits(EC
->getType()) > CountType
->getBitWidth())
151 // If this exiting block is contained in a nested loop, it is not eligible
152 // for insertion of the branch-and-decrement since the inner loop would
153 // end up messing up the value in the CTR.
154 if (!IsNestingLegal
&& LI
.getLoopFor(BB
) != L
&& !ForceNestedLoop
)
157 // We now have a loop-invariant count of loop iterations (which is not the
158 // constant zero) for which we know that this loop will not exit via this
161 // We need to make sure that this block will run on every loop iteration.
162 // For this to be true, we must dominate all blocks with backedges. Such
163 // blocks are in-loop predecessors to the header block.
164 bool NotAlways
= false;
165 for (BasicBlock
*Pred
: predecessors(L
->getHeader())) {
166 if (!L
->contains(Pred
))
169 if (!DT
.dominates(BB
, Pred
)) {
178 // Make sure this blocks ends with a conditional branch.
179 Instruction
*TI
= BB
->getTerminator();
183 if (BranchInst
*BI
= dyn_cast
<BranchInst
>(TI
)) {
184 if (!BI
->isConditional())
191 // Note that this block may not be the loop latch block, even if the loop
192 // has a latch block.
203 TargetTransformInfo::TargetTransformInfo(const DataLayout
&DL
)
204 : TTIImpl(new Model
<NoTTIImpl
>(NoTTIImpl(DL
))) {}
206 TargetTransformInfo::~TargetTransformInfo() = default;
208 TargetTransformInfo::TargetTransformInfo(TargetTransformInfo
&&Arg
)
209 : TTIImpl(std::move(Arg
.TTIImpl
)) {}
211 TargetTransformInfo
&TargetTransformInfo::operator=(TargetTransformInfo
&&RHS
) {
212 TTIImpl
= std::move(RHS
.TTIImpl
);
216 unsigned TargetTransformInfo::getInliningThresholdMultiplier() const {
217 return TTIImpl
->getInliningThresholdMultiplier();
221 TargetTransformInfo::getInliningCostBenefitAnalysisSavingsMultiplier() const {
222 return TTIImpl
->getInliningCostBenefitAnalysisSavingsMultiplier();
226 TargetTransformInfo::getInliningCostBenefitAnalysisProfitableMultiplier()
228 return TTIImpl
->getInliningCostBenefitAnalysisProfitableMultiplier();
232 TargetTransformInfo::adjustInliningThreshold(const CallBase
*CB
) const {
233 return TTIImpl
->adjustInliningThreshold(CB
);
236 unsigned TargetTransformInfo::getCallerAllocaCost(const CallBase
*CB
,
237 const AllocaInst
*AI
) const {
238 return TTIImpl
->getCallerAllocaCost(CB
, AI
);
241 int TargetTransformInfo::getInlinerVectorBonusPercent() const {
242 return TTIImpl
->getInlinerVectorBonusPercent();
245 InstructionCost
TargetTransformInfo::getGEPCost(
246 Type
*PointeeType
, const Value
*Ptr
, ArrayRef
<const Value
*> Operands
,
247 Type
*AccessType
, TTI::TargetCostKind CostKind
) const {
248 return TTIImpl
->getGEPCost(PointeeType
, Ptr
, Operands
, AccessType
, CostKind
);
251 InstructionCost
TargetTransformInfo::getPointersChainCost(
252 ArrayRef
<const Value
*> Ptrs
, const Value
*Base
,
253 const TTI::PointersChainInfo
&Info
, Type
*AccessTy
,
254 TTI::TargetCostKind CostKind
) const {
255 assert((Base
|| !Info
.isSameBase()) &&
256 "If pointers have same base address it has to be provided.");
257 return TTIImpl
->getPointersChainCost(Ptrs
, Base
, Info
, AccessTy
, CostKind
);
260 unsigned TargetTransformInfo::getEstimatedNumberOfCaseClusters(
261 const SwitchInst
&SI
, unsigned &JTSize
, ProfileSummaryInfo
*PSI
,
262 BlockFrequencyInfo
*BFI
) const {
263 return TTIImpl
->getEstimatedNumberOfCaseClusters(SI
, JTSize
, PSI
, BFI
);
267 TargetTransformInfo::getInstructionCost(const User
*U
,
268 ArrayRef
<const Value
*> Operands
,
269 enum TargetCostKind CostKind
) const {
270 InstructionCost Cost
= TTIImpl
->getInstructionCost(U
, Operands
, CostKind
);
271 assert((CostKind
== TTI::TCK_RecipThroughput
|| Cost
>= 0) &&
272 "TTI should not produce negative costs!");
276 BranchProbability
TargetTransformInfo::getPredictableBranchThreshold() const {
277 return PredictableBranchThreshold
.getNumOccurrences() > 0
278 ? BranchProbability(PredictableBranchThreshold
, 100)
279 : TTIImpl
->getPredictableBranchThreshold();
282 InstructionCost
TargetTransformInfo::getBranchMispredictPenalty() const {
283 return TTIImpl
->getBranchMispredictPenalty();
286 bool TargetTransformInfo::hasBranchDivergence(const Function
*F
) const {
287 return TTIImpl
->hasBranchDivergence(F
);
290 bool TargetTransformInfo::isSourceOfDivergence(const Value
*V
) const {
291 return TTIImpl
->isSourceOfDivergence(V
);
294 bool llvm::TargetTransformInfo::isAlwaysUniform(const Value
*V
) const {
295 return TTIImpl
->isAlwaysUniform(V
);
298 bool llvm::TargetTransformInfo::isValidAddrSpaceCast(unsigned FromAS
,
299 unsigned ToAS
) const {
300 return TTIImpl
->isValidAddrSpaceCast(FromAS
, ToAS
);
303 bool llvm::TargetTransformInfo::addrspacesMayAlias(unsigned FromAS
,
304 unsigned ToAS
) const {
305 return TTIImpl
->addrspacesMayAlias(FromAS
, ToAS
);
308 unsigned TargetTransformInfo::getFlatAddressSpace() const {
309 return TTIImpl
->getFlatAddressSpace();
312 bool TargetTransformInfo::collectFlatAddressOperands(
313 SmallVectorImpl
<int> &OpIndexes
, Intrinsic::ID IID
) const {
314 return TTIImpl
->collectFlatAddressOperands(OpIndexes
, IID
);
317 bool TargetTransformInfo::isNoopAddrSpaceCast(unsigned FromAS
,
318 unsigned ToAS
) const {
319 return TTIImpl
->isNoopAddrSpaceCast(FromAS
, ToAS
);
322 bool TargetTransformInfo::canHaveNonUndefGlobalInitializerInAddressSpace(
324 return TTIImpl
->canHaveNonUndefGlobalInitializerInAddressSpace(AS
);
327 unsigned TargetTransformInfo::getAssumedAddrSpace(const Value
*V
) const {
328 return TTIImpl
->getAssumedAddrSpace(V
);
331 bool TargetTransformInfo::isSingleThreaded() const {
332 return TTIImpl
->isSingleThreaded();
335 std::pair
<const Value
*, unsigned>
336 TargetTransformInfo::getPredicatedAddrSpace(const Value
*V
) const {
337 return TTIImpl
->getPredicatedAddrSpace(V
);
340 Value
*TargetTransformInfo::rewriteIntrinsicWithAddressSpace(
341 IntrinsicInst
*II
, Value
*OldV
, Value
*NewV
) const {
342 return TTIImpl
->rewriteIntrinsicWithAddressSpace(II
, OldV
, NewV
);
345 bool TargetTransformInfo::isLoweredToCall(const Function
*F
) const {
346 return TTIImpl
->isLoweredToCall(F
);
349 bool TargetTransformInfo::isHardwareLoopProfitable(
350 Loop
*L
, ScalarEvolution
&SE
, AssumptionCache
&AC
,
351 TargetLibraryInfo
*LibInfo
, HardwareLoopInfo
&HWLoopInfo
) const {
352 return TTIImpl
->isHardwareLoopProfitable(L
, SE
, AC
, LibInfo
, HWLoopInfo
);
355 bool TargetTransformInfo::preferPredicateOverEpilogue(
356 TailFoldingInfo
*TFI
) const {
357 return TTIImpl
->preferPredicateOverEpilogue(TFI
);
360 TailFoldingStyle
TargetTransformInfo::getPreferredTailFoldingStyle(
361 bool IVUpdateMayOverflow
) const {
362 return TTIImpl
->getPreferredTailFoldingStyle(IVUpdateMayOverflow
);
365 std::optional
<Instruction
*>
366 TargetTransformInfo::instCombineIntrinsic(InstCombiner
&IC
,
367 IntrinsicInst
&II
) const {
368 return TTIImpl
->instCombineIntrinsic(IC
, II
);
371 std::optional
<Value
*> TargetTransformInfo::simplifyDemandedUseBitsIntrinsic(
372 InstCombiner
&IC
, IntrinsicInst
&II
, APInt DemandedMask
, KnownBits
&Known
,
373 bool &KnownBitsComputed
) const {
374 return TTIImpl
->simplifyDemandedUseBitsIntrinsic(IC
, II
, DemandedMask
, Known
,
378 std::optional
<Value
*> TargetTransformInfo::simplifyDemandedVectorEltsIntrinsic(
379 InstCombiner
&IC
, IntrinsicInst
&II
, APInt DemandedElts
, APInt
&UndefElts
,
380 APInt
&UndefElts2
, APInt
&UndefElts3
,
381 std::function
<void(Instruction
*, unsigned, APInt
, APInt
&)>
382 SimplifyAndSetOp
) const {
383 return TTIImpl
->simplifyDemandedVectorEltsIntrinsic(
384 IC
, II
, DemandedElts
, UndefElts
, UndefElts2
, UndefElts3
,
388 void TargetTransformInfo::getUnrollingPreferences(
389 Loop
*L
, ScalarEvolution
&SE
, UnrollingPreferences
&UP
,
390 OptimizationRemarkEmitter
*ORE
) const {
391 return TTIImpl
->getUnrollingPreferences(L
, SE
, UP
, ORE
);
394 void TargetTransformInfo::getPeelingPreferences(Loop
*L
, ScalarEvolution
&SE
,
395 PeelingPreferences
&PP
) const {
396 return TTIImpl
->getPeelingPreferences(L
, SE
, PP
);
399 bool TargetTransformInfo::isLegalAddImmediate(int64_t Imm
) const {
400 return TTIImpl
->isLegalAddImmediate(Imm
);
403 bool TargetTransformInfo::isLegalAddScalableImmediate(int64_t Imm
) const {
404 return TTIImpl
->isLegalAddScalableImmediate(Imm
);
407 bool TargetTransformInfo::isLegalICmpImmediate(int64_t Imm
) const {
408 return TTIImpl
->isLegalICmpImmediate(Imm
);
411 bool TargetTransformInfo::isLegalAddressingMode(Type
*Ty
, GlobalValue
*BaseGV
,
413 bool HasBaseReg
, int64_t Scale
,
416 int64_t ScalableOffset
) const {
417 return TTIImpl
->isLegalAddressingMode(Ty
, BaseGV
, BaseOffset
, HasBaseReg
,
418 Scale
, AddrSpace
, I
, ScalableOffset
);
421 bool TargetTransformInfo::isLSRCostLess(const LSRCost
&C1
,
422 const LSRCost
&C2
) const {
423 return TTIImpl
->isLSRCostLess(C1
, C2
);
426 bool TargetTransformInfo::isNumRegsMajorCostOfLSR() const {
427 return TTIImpl
->isNumRegsMajorCostOfLSR();
430 bool TargetTransformInfo::shouldFoldTerminatingConditionAfterLSR() const {
431 return TTIImpl
->shouldFoldTerminatingConditionAfterLSR();
434 bool TargetTransformInfo::shouldDropLSRSolutionIfLessProfitable() const {
435 return TTIImpl
->shouldDropLSRSolutionIfLessProfitable();
438 bool TargetTransformInfo::isProfitableLSRChainElement(Instruction
*I
) const {
439 return TTIImpl
->isProfitableLSRChainElement(I
);
442 bool TargetTransformInfo::canMacroFuseCmp() const {
443 return TTIImpl
->canMacroFuseCmp();
446 bool TargetTransformInfo::canSaveCmp(Loop
*L
, BranchInst
**BI
,
447 ScalarEvolution
*SE
, LoopInfo
*LI
,
448 DominatorTree
*DT
, AssumptionCache
*AC
,
449 TargetLibraryInfo
*LibInfo
) const {
450 return TTIImpl
->canSaveCmp(L
, BI
, SE
, LI
, DT
, AC
, LibInfo
);
453 TTI::AddressingModeKind
454 TargetTransformInfo::getPreferredAddressingMode(const Loop
*L
,
455 ScalarEvolution
*SE
) const {
456 return TTIImpl
->getPreferredAddressingMode(L
, SE
);
459 bool TargetTransformInfo::isLegalMaskedStore(Type
*DataType
,
460 Align Alignment
) const {
461 return TTIImpl
->isLegalMaskedStore(DataType
, Alignment
);
464 bool TargetTransformInfo::isLegalMaskedLoad(Type
*DataType
,
465 Align Alignment
) const {
466 return TTIImpl
->isLegalMaskedLoad(DataType
, Alignment
);
469 bool TargetTransformInfo::isLegalNTStore(Type
*DataType
,
470 Align Alignment
) const {
471 return TTIImpl
->isLegalNTStore(DataType
, Alignment
);
474 bool TargetTransformInfo::isLegalNTLoad(Type
*DataType
, Align Alignment
) const {
475 return TTIImpl
->isLegalNTLoad(DataType
, Alignment
);
478 bool TargetTransformInfo::isLegalBroadcastLoad(Type
*ElementTy
,
479 ElementCount NumElements
) const {
480 return TTIImpl
->isLegalBroadcastLoad(ElementTy
, NumElements
);
483 bool TargetTransformInfo::isLegalMaskedGather(Type
*DataType
,
484 Align Alignment
) const {
485 return TTIImpl
->isLegalMaskedGather(DataType
, Alignment
);
488 bool TargetTransformInfo::isLegalAltInstr(
489 VectorType
*VecTy
, unsigned Opcode0
, unsigned Opcode1
,
490 const SmallBitVector
&OpcodeMask
) const {
491 return TTIImpl
->isLegalAltInstr(VecTy
, Opcode0
, Opcode1
, OpcodeMask
);
494 bool TargetTransformInfo::isLegalMaskedScatter(Type
*DataType
,
495 Align Alignment
) const {
496 return TTIImpl
->isLegalMaskedScatter(DataType
, Alignment
);
499 bool TargetTransformInfo::forceScalarizeMaskedGather(VectorType
*DataType
,
500 Align Alignment
) const {
501 return TTIImpl
->forceScalarizeMaskedGather(DataType
, Alignment
);
504 bool TargetTransformInfo::forceScalarizeMaskedScatter(VectorType
*DataType
,
505 Align Alignment
) const {
506 return TTIImpl
->forceScalarizeMaskedScatter(DataType
, Alignment
);
509 bool TargetTransformInfo::isLegalMaskedCompressStore(Type
*DataType
,
510 Align Alignment
) const {
511 return TTIImpl
->isLegalMaskedCompressStore(DataType
, Alignment
);
514 bool TargetTransformInfo::isLegalMaskedExpandLoad(Type
*DataType
,
515 Align Alignment
) const {
516 return TTIImpl
->isLegalMaskedExpandLoad(DataType
, Alignment
);
519 bool TargetTransformInfo::isLegalStridedLoadStore(Type
*DataType
,
520 Align Alignment
) const {
521 return TTIImpl
->isLegalStridedLoadStore(DataType
, Alignment
);
524 bool TargetTransformInfo::isLegalMaskedVectorHistogram(Type
*AddrType
,
525 Type
*DataType
) const {
526 return TTIImpl
->isLegalMaskedVectorHistogram(AddrType
, DataType
);
529 bool TargetTransformInfo::enableOrderedReductions() const {
530 return TTIImpl
->enableOrderedReductions();
533 bool TargetTransformInfo::hasDivRemOp(Type
*DataType
, bool IsSigned
) const {
534 return TTIImpl
->hasDivRemOp(DataType
, IsSigned
);
537 bool TargetTransformInfo::hasVolatileVariant(Instruction
*I
,
538 unsigned AddrSpace
) const {
539 return TTIImpl
->hasVolatileVariant(I
, AddrSpace
);
542 bool TargetTransformInfo::prefersVectorizedAddressing() const {
543 return TTIImpl
->prefersVectorizedAddressing();
546 InstructionCost
TargetTransformInfo::getScalingFactorCost(
547 Type
*Ty
, GlobalValue
*BaseGV
, StackOffset BaseOffset
, bool HasBaseReg
,
548 int64_t Scale
, unsigned AddrSpace
) const {
549 InstructionCost Cost
= TTIImpl
->getScalingFactorCost(
550 Ty
, BaseGV
, BaseOffset
, HasBaseReg
, Scale
, AddrSpace
);
551 assert(Cost
>= 0 && "TTI should not produce negative costs!");
555 bool TargetTransformInfo::LSRWithInstrQueries() const {
556 return TTIImpl
->LSRWithInstrQueries();
559 bool TargetTransformInfo::isTruncateFree(Type
*Ty1
, Type
*Ty2
) const {
560 return TTIImpl
->isTruncateFree(Ty1
, Ty2
);
563 bool TargetTransformInfo::isProfitableToHoist(Instruction
*I
) const {
564 return TTIImpl
->isProfitableToHoist(I
);
567 bool TargetTransformInfo::useAA() const { return TTIImpl
->useAA(); }
569 bool TargetTransformInfo::isTypeLegal(Type
*Ty
) const {
570 return TTIImpl
->isTypeLegal(Ty
);
573 unsigned TargetTransformInfo::getRegUsageForType(Type
*Ty
) const {
574 return TTIImpl
->getRegUsageForType(Ty
);
577 bool TargetTransformInfo::shouldBuildLookupTables() const {
578 return TTIImpl
->shouldBuildLookupTables();
581 bool TargetTransformInfo::shouldBuildLookupTablesForConstant(
583 return TTIImpl
->shouldBuildLookupTablesForConstant(C
);
586 bool TargetTransformInfo::shouldBuildRelLookupTables() const {
587 return TTIImpl
->shouldBuildRelLookupTables();
590 bool TargetTransformInfo::useColdCCForColdCall(Function
&F
) const {
591 return TTIImpl
->useColdCCForColdCall(F
);
594 InstructionCost
TargetTransformInfo::getScalarizationOverhead(
595 VectorType
*Ty
, const APInt
&DemandedElts
, bool Insert
, bool Extract
,
596 TTI::TargetCostKind CostKind
) const {
597 return TTIImpl
->getScalarizationOverhead(Ty
, DemandedElts
, Insert
, Extract
,
601 InstructionCost
TargetTransformInfo::getOperandsScalarizationOverhead(
602 ArrayRef
<const Value
*> Args
, ArrayRef
<Type
*> Tys
,
603 TTI::TargetCostKind CostKind
) const {
604 return TTIImpl
->getOperandsScalarizationOverhead(Args
, Tys
, CostKind
);
607 bool TargetTransformInfo::supportsEfficientVectorElementLoadStore() const {
608 return TTIImpl
->supportsEfficientVectorElementLoadStore();
611 bool TargetTransformInfo::supportsTailCalls() const {
612 return TTIImpl
->supportsTailCalls();
615 bool TargetTransformInfo::supportsTailCallFor(const CallBase
*CB
) const {
616 return TTIImpl
->supportsTailCallFor(CB
);
619 bool TargetTransformInfo::enableAggressiveInterleaving(
620 bool LoopHasReductions
) const {
621 return TTIImpl
->enableAggressiveInterleaving(LoopHasReductions
);
624 TargetTransformInfo::MemCmpExpansionOptions
625 TargetTransformInfo::enableMemCmpExpansion(bool OptSize
, bool IsZeroCmp
) const {
626 return TTIImpl
->enableMemCmpExpansion(OptSize
, IsZeroCmp
);
629 bool TargetTransformInfo::enableSelectOptimize() const {
630 return TTIImpl
->enableSelectOptimize();
633 bool TargetTransformInfo::shouldTreatInstructionLikeSelect(
634 const Instruction
*I
) const {
635 return TTIImpl
->shouldTreatInstructionLikeSelect(I
);
638 bool TargetTransformInfo::enableInterleavedAccessVectorization() const {
639 return TTIImpl
->enableInterleavedAccessVectorization();
642 bool TargetTransformInfo::enableMaskedInterleavedAccessVectorization() const {
643 return TTIImpl
->enableMaskedInterleavedAccessVectorization();
646 bool TargetTransformInfo::isFPVectorizationPotentiallyUnsafe() const {
647 return TTIImpl
->isFPVectorizationPotentiallyUnsafe();
651 TargetTransformInfo::allowsMisalignedMemoryAccesses(LLVMContext
&Context
,
653 unsigned AddressSpace
,
655 unsigned *Fast
) const {
656 return TTIImpl
->allowsMisalignedMemoryAccesses(Context
, BitWidth
,
657 AddressSpace
, Alignment
, Fast
);
660 TargetTransformInfo::PopcntSupportKind
661 TargetTransformInfo::getPopcntSupport(unsigned IntTyWidthInBit
) const {
662 return TTIImpl
->getPopcntSupport(IntTyWidthInBit
);
665 bool TargetTransformInfo::haveFastSqrt(Type
*Ty
) const {
666 return TTIImpl
->haveFastSqrt(Ty
);
669 bool TargetTransformInfo::isExpensiveToSpeculativelyExecute(
670 const Instruction
*I
) const {
671 return TTIImpl
->isExpensiveToSpeculativelyExecute(I
);
674 bool TargetTransformInfo::isFCmpOrdCheaperThanFCmpZero(Type
*Ty
) const {
675 return TTIImpl
->isFCmpOrdCheaperThanFCmpZero(Ty
);
678 InstructionCost
TargetTransformInfo::getFPOpCost(Type
*Ty
) const {
679 InstructionCost Cost
= TTIImpl
->getFPOpCost(Ty
);
680 assert(Cost
>= 0 && "TTI should not produce negative costs!");
684 InstructionCost
TargetTransformInfo::getIntImmCodeSizeCost(unsigned Opcode
,
688 InstructionCost Cost
= TTIImpl
->getIntImmCodeSizeCost(Opcode
, Idx
, Imm
, Ty
);
689 assert(Cost
>= 0 && "TTI should not produce negative costs!");
694 TargetTransformInfo::getIntImmCost(const APInt
&Imm
, Type
*Ty
,
695 TTI::TargetCostKind CostKind
) const {
696 InstructionCost Cost
= TTIImpl
->getIntImmCost(Imm
, Ty
, CostKind
);
697 assert(Cost
>= 0 && "TTI should not produce negative costs!");
701 InstructionCost
TargetTransformInfo::getIntImmCostInst(
702 unsigned Opcode
, unsigned Idx
, const APInt
&Imm
, Type
*Ty
,
703 TTI::TargetCostKind CostKind
, Instruction
*Inst
) const {
704 InstructionCost Cost
=
705 TTIImpl
->getIntImmCostInst(Opcode
, Idx
, Imm
, Ty
, CostKind
, Inst
);
706 assert(Cost
>= 0 && "TTI should not produce negative costs!");
711 TargetTransformInfo::getIntImmCostIntrin(Intrinsic::ID IID
, unsigned Idx
,
712 const APInt
&Imm
, Type
*Ty
,
713 TTI::TargetCostKind CostKind
) const {
714 InstructionCost Cost
=
715 TTIImpl
->getIntImmCostIntrin(IID
, Idx
, Imm
, Ty
, CostKind
);
716 assert(Cost
>= 0 && "TTI should not produce negative costs!");
720 bool TargetTransformInfo::preferToKeepConstantsAttached(
721 const Instruction
&Inst
, const Function
&Fn
) const {
722 return TTIImpl
->preferToKeepConstantsAttached(Inst
, Fn
);
725 unsigned TargetTransformInfo::getNumberOfRegisters(unsigned ClassID
) const {
726 return TTIImpl
->getNumberOfRegisters(ClassID
);
729 bool TargetTransformInfo::hasConditionalLoadStoreForType(Type
*Ty
) const {
730 return TTIImpl
->hasConditionalLoadStoreForType(Ty
);
733 unsigned TargetTransformInfo::getRegisterClassForType(bool Vector
,
735 return TTIImpl
->getRegisterClassForType(Vector
, Ty
);
738 const char *TargetTransformInfo::getRegisterClassName(unsigned ClassID
) const {
739 return TTIImpl
->getRegisterClassName(ClassID
);
742 TypeSize
TargetTransformInfo::getRegisterBitWidth(
743 TargetTransformInfo::RegisterKind K
) const {
744 return TTIImpl
->getRegisterBitWidth(K
);
747 unsigned TargetTransformInfo::getMinVectorRegisterBitWidth() const {
748 return TTIImpl
->getMinVectorRegisterBitWidth();
751 std::optional
<unsigned> TargetTransformInfo::getMaxVScale() const {
752 return TTIImpl
->getMaxVScale();
755 std::optional
<unsigned> TargetTransformInfo::getVScaleForTuning() const {
756 return TTIImpl
->getVScaleForTuning();
759 bool TargetTransformInfo::isVScaleKnownToBeAPowerOfTwo() const {
760 return TTIImpl
->isVScaleKnownToBeAPowerOfTwo();
763 bool TargetTransformInfo::shouldMaximizeVectorBandwidth(
764 TargetTransformInfo::RegisterKind K
) const {
765 return TTIImpl
->shouldMaximizeVectorBandwidth(K
);
768 ElementCount
TargetTransformInfo::getMinimumVF(unsigned ElemWidth
,
769 bool IsScalable
) const {
770 return TTIImpl
->getMinimumVF(ElemWidth
, IsScalable
);
773 unsigned TargetTransformInfo::getMaximumVF(unsigned ElemWidth
,
774 unsigned Opcode
) const {
775 return TTIImpl
->getMaximumVF(ElemWidth
, Opcode
);
778 unsigned TargetTransformInfo::getStoreMinimumVF(unsigned VF
, Type
*ScalarMemTy
,
779 Type
*ScalarValTy
) const {
780 return TTIImpl
->getStoreMinimumVF(VF
, ScalarMemTy
, ScalarValTy
);
783 bool TargetTransformInfo::shouldConsiderAddressTypePromotion(
784 const Instruction
&I
, bool &AllowPromotionWithoutCommonHeader
) const {
785 return TTIImpl
->shouldConsiderAddressTypePromotion(
786 I
, AllowPromotionWithoutCommonHeader
);
789 unsigned TargetTransformInfo::getCacheLineSize() const {
790 return CacheLineSize
.getNumOccurrences() > 0 ? CacheLineSize
791 : TTIImpl
->getCacheLineSize();
794 std::optional
<unsigned>
795 TargetTransformInfo::getCacheSize(CacheLevel Level
) const {
796 return TTIImpl
->getCacheSize(Level
);
799 std::optional
<unsigned>
800 TargetTransformInfo::getCacheAssociativity(CacheLevel Level
) const {
801 return TTIImpl
->getCacheAssociativity(Level
);
804 std::optional
<unsigned> TargetTransformInfo::getMinPageSize() const {
805 return MinPageSize
.getNumOccurrences() > 0 ? MinPageSize
806 : TTIImpl
->getMinPageSize();
809 unsigned TargetTransformInfo::getPrefetchDistance() const {
810 return TTIImpl
->getPrefetchDistance();
813 unsigned TargetTransformInfo::getMinPrefetchStride(
814 unsigned NumMemAccesses
, unsigned NumStridedMemAccesses
,
815 unsigned NumPrefetches
, bool HasCall
) const {
816 return TTIImpl
->getMinPrefetchStride(NumMemAccesses
, NumStridedMemAccesses
,
817 NumPrefetches
, HasCall
);
820 unsigned TargetTransformInfo::getMaxPrefetchIterationsAhead() const {
821 return TTIImpl
->getMaxPrefetchIterationsAhead();
824 bool TargetTransformInfo::enableWritePrefetching() const {
825 return TTIImpl
->enableWritePrefetching();
828 bool TargetTransformInfo::shouldPrefetchAddressSpace(unsigned AS
) const {
829 return TTIImpl
->shouldPrefetchAddressSpace(AS
);
832 unsigned TargetTransformInfo::getMaxInterleaveFactor(ElementCount VF
) const {
833 return TTIImpl
->getMaxInterleaveFactor(VF
);
836 TargetTransformInfo::OperandValueInfo
837 TargetTransformInfo::getOperandInfo(const Value
*V
) {
838 OperandValueKind OpInfo
= OK_AnyValue
;
839 OperandValueProperties OpProps
= OP_None
;
841 if (isa
<ConstantInt
>(V
) || isa
<ConstantFP
>(V
)) {
842 if (const auto *CI
= dyn_cast
<ConstantInt
>(V
)) {
843 if (CI
->getValue().isPowerOf2())
844 OpProps
= OP_PowerOf2
;
845 else if (CI
->getValue().isNegatedPowerOf2())
846 OpProps
= OP_NegatedPowerOf2
;
848 return {OK_UniformConstantValue
, OpProps
};
851 // A broadcast shuffle creates a uniform value.
852 // TODO: Add support for non-zero index broadcasts.
853 // TODO: Add support for different source vector width.
854 if (const auto *ShuffleInst
= dyn_cast
<ShuffleVectorInst
>(V
))
855 if (ShuffleInst
->isZeroEltSplat())
856 OpInfo
= OK_UniformValue
;
858 const Value
*Splat
= getSplatValue(V
);
860 // Check for a splat of a constant or for a non uniform vector of constants
861 // and check if the constant(s) are all powers of two.
862 if (isa
<ConstantVector
>(V
) || isa
<ConstantDataVector
>(V
)) {
863 OpInfo
= OK_NonUniformConstantValue
;
865 OpInfo
= OK_UniformConstantValue
;
866 if (auto *CI
= dyn_cast
<ConstantInt
>(Splat
)) {
867 if (CI
->getValue().isPowerOf2())
868 OpProps
= OP_PowerOf2
;
869 else if (CI
->getValue().isNegatedPowerOf2())
870 OpProps
= OP_NegatedPowerOf2
;
872 } else if (const auto *CDS
= dyn_cast
<ConstantDataSequential
>(V
)) {
873 bool AllPow2
= true, AllNegPow2
= true;
874 for (unsigned I
= 0, E
= CDS
->getNumElements(); I
!= E
; ++I
) {
875 if (auto *CI
= dyn_cast
<ConstantInt
>(CDS
->getElementAsConstant(I
))) {
876 AllPow2
&= CI
->getValue().isPowerOf2();
877 AllNegPow2
&= CI
->getValue().isNegatedPowerOf2();
878 if (AllPow2
|| AllNegPow2
)
881 AllPow2
= AllNegPow2
= false;
884 OpProps
= AllPow2
? OP_PowerOf2
: OpProps
;
885 OpProps
= AllNegPow2
? OP_NegatedPowerOf2
: OpProps
;
889 // Check for a splat of a uniform value. This is not loop aware, so return
890 // true only for the obviously uniform cases (argument, globalvalue)
891 if (Splat
&& (isa
<Argument
>(Splat
) || isa
<GlobalValue
>(Splat
)))
892 OpInfo
= OK_UniformValue
;
894 return {OpInfo
, OpProps
};
897 InstructionCost
TargetTransformInfo::getArithmeticInstrCost(
898 unsigned Opcode
, Type
*Ty
, TTI::TargetCostKind CostKind
,
899 OperandValueInfo Op1Info
, OperandValueInfo Op2Info
,
900 ArrayRef
<const Value
*> Args
, const Instruction
*CxtI
,
901 const TargetLibraryInfo
*TLibInfo
) const {
903 // Use call cost for frem intructions that have platform specific vector math
904 // functions, as those will be replaced with calls later by SelectionDAG or
905 // ReplaceWithVecLib pass.
906 if (TLibInfo
&& Opcode
== Instruction::FRem
) {
907 VectorType
*VecTy
= dyn_cast
<VectorType
>(Ty
);
910 TLibInfo
->getLibFunc(Instruction::FRem
, Ty
->getScalarType(), Func
) &&
911 TLibInfo
->isFunctionVectorizable(TLibInfo
->getName(Func
),
912 VecTy
->getElementCount()))
913 return getCallInstrCost(nullptr, VecTy
, {VecTy
, VecTy
}, CostKind
);
916 InstructionCost Cost
=
917 TTIImpl
->getArithmeticInstrCost(Opcode
, Ty
, CostKind
,
920 assert(Cost
>= 0 && "TTI should not produce negative costs!");
924 InstructionCost
TargetTransformInfo::getAltInstrCost(
925 VectorType
*VecTy
, unsigned Opcode0
, unsigned Opcode1
,
926 const SmallBitVector
&OpcodeMask
, TTI::TargetCostKind CostKind
) const {
927 InstructionCost Cost
=
928 TTIImpl
->getAltInstrCost(VecTy
, Opcode0
, Opcode1
, OpcodeMask
, CostKind
);
929 assert(Cost
>= 0 && "TTI should not produce negative costs!");
933 InstructionCost
TargetTransformInfo::getShuffleCost(
934 ShuffleKind Kind
, VectorType
*Ty
, ArrayRef
<int> Mask
,
935 TTI::TargetCostKind CostKind
, int Index
, VectorType
*SubTp
,
936 ArrayRef
<const Value
*> Args
, const Instruction
*CxtI
) const {
937 InstructionCost Cost
= TTIImpl
->getShuffleCost(Kind
, Ty
, Mask
, CostKind
,
938 Index
, SubTp
, Args
, CxtI
);
939 assert(Cost
>= 0 && "TTI should not produce negative costs!");
944 TargetTransformInfo::getCastContextHint(const Instruction
*I
) {
946 return CastContextHint::None
;
948 auto getLoadStoreKind
= [](const Value
*V
, unsigned LdStOp
, unsigned MaskedOp
,
949 unsigned GatScatOp
) {
950 const Instruction
*I
= dyn_cast
<Instruction
>(V
);
952 return CastContextHint::None
;
954 if (I
->getOpcode() == LdStOp
)
955 return CastContextHint::Normal
;
957 if (const IntrinsicInst
*II
= dyn_cast
<IntrinsicInst
>(I
)) {
958 if (II
->getIntrinsicID() == MaskedOp
)
959 return TTI::CastContextHint::Masked
;
960 if (II
->getIntrinsicID() == GatScatOp
)
961 return TTI::CastContextHint::GatherScatter
;
964 return TTI::CastContextHint::None
;
967 switch (I
->getOpcode()) {
968 case Instruction::ZExt
:
969 case Instruction::SExt
:
970 case Instruction::FPExt
:
971 return getLoadStoreKind(I
->getOperand(0), Instruction::Load
,
972 Intrinsic::masked_load
, Intrinsic::masked_gather
);
973 case Instruction::Trunc
:
974 case Instruction::FPTrunc
:
976 return getLoadStoreKind(*I
->user_begin(), Instruction::Store
,
977 Intrinsic::masked_store
,
978 Intrinsic::masked_scatter
);
981 return CastContextHint::None
;
984 return TTI::CastContextHint::None
;
987 InstructionCost
TargetTransformInfo::getCastInstrCost(
988 unsigned Opcode
, Type
*Dst
, Type
*Src
, CastContextHint CCH
,
989 TTI::TargetCostKind CostKind
, const Instruction
*I
) const {
990 assert((I
== nullptr || I
->getOpcode() == Opcode
) &&
991 "Opcode should reflect passed instruction.");
992 InstructionCost Cost
=
993 TTIImpl
->getCastInstrCost(Opcode
, Dst
, Src
, CCH
, CostKind
, I
);
994 assert(Cost
>= 0 && "TTI should not produce negative costs!");
998 InstructionCost
TargetTransformInfo::getExtractWithExtendCost(
999 unsigned Opcode
, Type
*Dst
, VectorType
*VecTy
, unsigned Index
) const {
1000 InstructionCost Cost
=
1001 TTIImpl
->getExtractWithExtendCost(Opcode
, Dst
, VecTy
, Index
);
1002 assert(Cost
>= 0 && "TTI should not produce negative costs!");
1006 InstructionCost
TargetTransformInfo::getCFInstrCost(
1007 unsigned Opcode
, TTI::TargetCostKind CostKind
, const Instruction
*I
) const {
1008 assert((I
== nullptr || I
->getOpcode() == Opcode
) &&
1009 "Opcode should reflect passed instruction.");
1010 InstructionCost Cost
= TTIImpl
->getCFInstrCost(Opcode
, CostKind
, I
);
1011 assert(Cost
>= 0 && "TTI should not produce negative costs!");
1015 InstructionCost
TargetTransformInfo::getCmpSelInstrCost(
1016 unsigned Opcode
, Type
*ValTy
, Type
*CondTy
, CmpInst::Predicate VecPred
,
1017 TTI::TargetCostKind CostKind
, const Instruction
*I
) const {
1018 assert((I
== nullptr || I
->getOpcode() == Opcode
) &&
1019 "Opcode should reflect passed instruction.");
1020 InstructionCost Cost
=
1021 TTIImpl
->getCmpSelInstrCost(Opcode
, ValTy
, CondTy
, VecPred
, CostKind
, I
);
1022 assert(Cost
>= 0 && "TTI should not produce negative costs!");
1026 InstructionCost
TargetTransformInfo::getVectorInstrCost(
1027 unsigned Opcode
, Type
*Val
, TTI::TargetCostKind CostKind
, unsigned Index
,
1028 Value
*Op0
, Value
*Op1
) const {
1029 // FIXME: Assert that Opcode is either InsertElement or ExtractElement.
1030 // This is mentioned in the interface description and respected by all
1031 // callers, but never asserted upon.
1032 InstructionCost Cost
=
1033 TTIImpl
->getVectorInstrCost(Opcode
, Val
, CostKind
, Index
, Op0
, Op1
);
1034 assert(Cost
>= 0 && "TTI should not produce negative costs!");
1039 TargetTransformInfo::getVectorInstrCost(const Instruction
&I
, Type
*Val
,
1040 TTI::TargetCostKind CostKind
,
1041 unsigned Index
) const {
1042 // FIXME: Assert that Opcode is either InsertElement or ExtractElement.
1043 // This is mentioned in the interface description and respected by all
1044 // callers, but never asserted upon.
1045 InstructionCost Cost
= TTIImpl
->getVectorInstrCost(I
, Val
, CostKind
, Index
);
1046 assert(Cost
>= 0 && "TTI should not produce negative costs!");
1050 InstructionCost
TargetTransformInfo::getReplicationShuffleCost(
1051 Type
*EltTy
, int ReplicationFactor
, int VF
, const APInt
&DemandedDstElts
,
1052 TTI::TargetCostKind CostKind
) const {
1053 InstructionCost Cost
= TTIImpl
->getReplicationShuffleCost(
1054 EltTy
, ReplicationFactor
, VF
, DemandedDstElts
, CostKind
);
1055 assert(Cost
>= 0 && "TTI should not produce negative costs!");
1059 InstructionCost
TargetTransformInfo::getMemoryOpCost(
1060 unsigned Opcode
, Type
*Src
, Align Alignment
, unsigned AddressSpace
,
1061 TTI::TargetCostKind CostKind
, TTI::OperandValueInfo OpInfo
,
1062 const Instruction
*I
) const {
1063 assert((I
== nullptr || I
->getOpcode() == Opcode
) &&
1064 "Opcode should reflect passed instruction.");
1065 InstructionCost Cost
= TTIImpl
->getMemoryOpCost(
1066 Opcode
, Src
, Alignment
, AddressSpace
, CostKind
, OpInfo
, I
);
1067 assert(Cost
>= 0 && "TTI should not produce negative costs!");
1071 InstructionCost
TargetTransformInfo::getMaskedMemoryOpCost(
1072 unsigned Opcode
, Type
*Src
, Align Alignment
, unsigned AddressSpace
,
1073 TTI::TargetCostKind CostKind
) const {
1074 InstructionCost Cost
= TTIImpl
->getMaskedMemoryOpCost(Opcode
, Src
, Alignment
,
1075 AddressSpace
, CostKind
);
1076 assert(Cost
>= 0 && "TTI should not produce negative costs!");
1080 InstructionCost
TargetTransformInfo::getGatherScatterOpCost(
1081 unsigned Opcode
, Type
*DataTy
, const Value
*Ptr
, bool VariableMask
,
1082 Align Alignment
, TTI::TargetCostKind CostKind
, const Instruction
*I
) const {
1083 InstructionCost Cost
= TTIImpl
->getGatherScatterOpCost(
1084 Opcode
, DataTy
, Ptr
, VariableMask
, Alignment
, CostKind
, I
);
1085 assert((!Cost
.isValid() || Cost
>= 0) &&
1086 "TTI should not produce negative costs!");
1090 InstructionCost
TargetTransformInfo::getStridedMemoryOpCost(
1091 unsigned Opcode
, Type
*DataTy
, const Value
*Ptr
, bool VariableMask
,
1092 Align Alignment
, TTI::TargetCostKind CostKind
, const Instruction
*I
) const {
1093 InstructionCost Cost
= TTIImpl
->getStridedMemoryOpCost(
1094 Opcode
, DataTy
, Ptr
, VariableMask
, Alignment
, CostKind
, I
);
1095 assert(Cost
>= 0 && "TTI should not produce negative costs!");
1099 InstructionCost
TargetTransformInfo::getInterleavedMemoryOpCost(
1100 unsigned Opcode
, Type
*VecTy
, unsigned Factor
, ArrayRef
<unsigned> Indices
,
1101 Align Alignment
, unsigned AddressSpace
, TTI::TargetCostKind CostKind
,
1102 bool UseMaskForCond
, bool UseMaskForGaps
) const {
1103 InstructionCost Cost
= TTIImpl
->getInterleavedMemoryOpCost(
1104 Opcode
, VecTy
, Factor
, Indices
, Alignment
, AddressSpace
, CostKind
,
1105 UseMaskForCond
, UseMaskForGaps
);
1106 assert(Cost
>= 0 && "TTI should not produce negative costs!");
1111 TargetTransformInfo::getIntrinsicInstrCost(const IntrinsicCostAttributes
&ICA
,
1112 TTI::TargetCostKind CostKind
) const {
1113 InstructionCost Cost
= TTIImpl
->getIntrinsicInstrCost(ICA
, CostKind
);
1114 assert(Cost
>= 0 && "TTI should not produce negative costs!");
1119 TargetTransformInfo::getCallInstrCost(Function
*F
, Type
*RetTy
,
1120 ArrayRef
<Type
*> Tys
,
1121 TTI::TargetCostKind CostKind
) const {
1122 InstructionCost Cost
= TTIImpl
->getCallInstrCost(F
, RetTy
, Tys
, CostKind
);
1123 assert(Cost
>= 0 && "TTI should not produce negative costs!");
1127 unsigned TargetTransformInfo::getNumberOfParts(Type
*Tp
) const {
1128 return TTIImpl
->getNumberOfParts(Tp
);
1132 TargetTransformInfo::getAddressComputationCost(Type
*Tp
, ScalarEvolution
*SE
,
1133 const SCEV
*Ptr
) const {
1134 InstructionCost Cost
= TTIImpl
->getAddressComputationCost(Tp
, SE
, Ptr
);
1135 assert(Cost
>= 0 && "TTI should not produce negative costs!");
1139 InstructionCost
TargetTransformInfo::getMemcpyCost(const Instruction
*I
) const {
1140 InstructionCost Cost
= TTIImpl
->getMemcpyCost(I
);
1141 assert(Cost
>= 0 && "TTI should not produce negative costs!");
1145 uint64_t TargetTransformInfo::getMaxMemIntrinsicInlineSizeThreshold() const {
1146 return TTIImpl
->getMaxMemIntrinsicInlineSizeThreshold();
1149 InstructionCost
TargetTransformInfo::getArithmeticReductionCost(
1150 unsigned Opcode
, VectorType
*Ty
, std::optional
<FastMathFlags
> FMF
,
1151 TTI::TargetCostKind CostKind
) const {
1152 InstructionCost Cost
=
1153 TTIImpl
->getArithmeticReductionCost(Opcode
, Ty
, FMF
, CostKind
);
1154 assert(Cost
>= 0 && "TTI should not produce negative costs!");
1158 InstructionCost
TargetTransformInfo::getMinMaxReductionCost(
1159 Intrinsic::ID IID
, VectorType
*Ty
, FastMathFlags FMF
,
1160 TTI::TargetCostKind CostKind
) const {
1161 InstructionCost Cost
=
1162 TTIImpl
->getMinMaxReductionCost(IID
, Ty
, FMF
, CostKind
);
1163 assert(Cost
>= 0 && "TTI should not produce negative costs!");
1167 InstructionCost
TargetTransformInfo::getExtendedReductionCost(
1168 unsigned Opcode
, bool IsUnsigned
, Type
*ResTy
, VectorType
*Ty
,
1169 FastMathFlags FMF
, TTI::TargetCostKind CostKind
) const {
1170 return TTIImpl
->getExtendedReductionCost(Opcode
, IsUnsigned
, ResTy
, Ty
, FMF
,
1174 InstructionCost
TargetTransformInfo::getMulAccReductionCost(
1175 bool IsUnsigned
, Type
*ResTy
, VectorType
*Ty
,
1176 TTI::TargetCostKind CostKind
) const {
1177 return TTIImpl
->getMulAccReductionCost(IsUnsigned
, ResTy
, Ty
, CostKind
);
1181 TargetTransformInfo::getCostOfKeepingLiveOverCall(ArrayRef
<Type
*> Tys
) const {
1182 return TTIImpl
->getCostOfKeepingLiveOverCall(Tys
);
1185 bool TargetTransformInfo::getTgtMemIntrinsic(IntrinsicInst
*Inst
,
1186 MemIntrinsicInfo
&Info
) const {
1187 return TTIImpl
->getTgtMemIntrinsic(Inst
, Info
);
1190 unsigned TargetTransformInfo::getAtomicMemIntrinsicMaxElementSize() const {
1191 return TTIImpl
->getAtomicMemIntrinsicMaxElementSize();
1194 Value
*TargetTransformInfo::getOrCreateResultFromMemIntrinsic(
1195 IntrinsicInst
*Inst
, Type
*ExpectedType
) const {
1196 return TTIImpl
->getOrCreateResultFromMemIntrinsic(Inst
, ExpectedType
);
1199 Type
*TargetTransformInfo::getMemcpyLoopLoweringType(
1200 LLVMContext
&Context
, Value
*Length
, unsigned SrcAddrSpace
,
1201 unsigned DestAddrSpace
, unsigned SrcAlign
, unsigned DestAlign
,
1202 std::optional
<uint32_t> AtomicElementSize
) const {
1203 return TTIImpl
->getMemcpyLoopLoweringType(Context
, Length
, SrcAddrSpace
,
1204 DestAddrSpace
, SrcAlign
, DestAlign
,
1208 void TargetTransformInfo::getMemcpyLoopResidualLoweringType(
1209 SmallVectorImpl
<Type
*> &OpsOut
, LLVMContext
&Context
,
1210 unsigned RemainingBytes
, unsigned SrcAddrSpace
, unsigned DestAddrSpace
,
1211 unsigned SrcAlign
, unsigned DestAlign
,
1212 std::optional
<uint32_t> AtomicCpySize
) const {
1213 TTIImpl
->getMemcpyLoopResidualLoweringType(
1214 OpsOut
, Context
, RemainingBytes
, SrcAddrSpace
, DestAddrSpace
, SrcAlign
,
1215 DestAlign
, AtomicCpySize
);
1218 bool TargetTransformInfo::areInlineCompatible(const Function
*Caller
,
1219 const Function
*Callee
) const {
1220 return TTIImpl
->areInlineCompatible(Caller
, Callee
);
1224 TargetTransformInfo::getInlineCallPenalty(const Function
*F
,
1225 const CallBase
&Call
,
1226 unsigned DefaultCallPenalty
) const {
1227 return TTIImpl
->getInlineCallPenalty(F
, Call
, DefaultCallPenalty
);
1230 bool TargetTransformInfo::areTypesABICompatible(
1231 const Function
*Caller
, const Function
*Callee
,
1232 const ArrayRef
<Type
*> &Types
) const {
1233 return TTIImpl
->areTypesABICompatible(Caller
, Callee
, Types
);
1236 bool TargetTransformInfo::isIndexedLoadLegal(MemIndexedMode Mode
,
1238 return TTIImpl
->isIndexedLoadLegal(Mode
, Ty
);
1241 bool TargetTransformInfo::isIndexedStoreLegal(MemIndexedMode Mode
,
1243 return TTIImpl
->isIndexedStoreLegal(Mode
, Ty
);
1246 unsigned TargetTransformInfo::getLoadStoreVecRegBitWidth(unsigned AS
) const {
1247 return TTIImpl
->getLoadStoreVecRegBitWidth(AS
);
1250 bool TargetTransformInfo::isLegalToVectorizeLoad(LoadInst
*LI
) const {
1251 return TTIImpl
->isLegalToVectorizeLoad(LI
);
1254 bool TargetTransformInfo::isLegalToVectorizeStore(StoreInst
*SI
) const {
1255 return TTIImpl
->isLegalToVectorizeStore(SI
);
1258 bool TargetTransformInfo::isLegalToVectorizeLoadChain(
1259 unsigned ChainSizeInBytes
, Align Alignment
, unsigned AddrSpace
) const {
1260 return TTIImpl
->isLegalToVectorizeLoadChain(ChainSizeInBytes
, Alignment
,
1264 bool TargetTransformInfo::isLegalToVectorizeStoreChain(
1265 unsigned ChainSizeInBytes
, Align Alignment
, unsigned AddrSpace
) const {
1266 return TTIImpl
->isLegalToVectorizeStoreChain(ChainSizeInBytes
, Alignment
,
1270 bool TargetTransformInfo::isLegalToVectorizeReduction(
1271 const RecurrenceDescriptor
&RdxDesc
, ElementCount VF
) const {
1272 return TTIImpl
->isLegalToVectorizeReduction(RdxDesc
, VF
);
1275 bool TargetTransformInfo::isElementTypeLegalForScalableVector(Type
*Ty
) const {
1276 return TTIImpl
->isElementTypeLegalForScalableVector(Ty
);
1279 unsigned TargetTransformInfo::getLoadVectorFactor(unsigned VF
,
1281 unsigned ChainSizeInBytes
,
1282 VectorType
*VecTy
) const {
1283 return TTIImpl
->getLoadVectorFactor(VF
, LoadSize
, ChainSizeInBytes
, VecTy
);
1286 unsigned TargetTransformInfo::getStoreVectorFactor(unsigned VF
,
1288 unsigned ChainSizeInBytes
,
1289 VectorType
*VecTy
) const {
1290 return TTIImpl
->getStoreVectorFactor(VF
, StoreSize
, ChainSizeInBytes
, VecTy
);
1293 bool TargetTransformInfo::preferFixedOverScalableIfEqualCost() const {
1294 return TTIImpl
->preferFixedOverScalableIfEqualCost();
1297 bool TargetTransformInfo::preferInLoopReduction(unsigned Opcode
, Type
*Ty
,
1298 ReductionFlags Flags
) const {
1299 return TTIImpl
->preferInLoopReduction(Opcode
, Ty
, Flags
);
1302 bool TargetTransformInfo::preferPredicatedReductionSelect(
1303 unsigned Opcode
, Type
*Ty
, ReductionFlags Flags
) const {
1304 return TTIImpl
->preferPredicatedReductionSelect(Opcode
, Ty
, Flags
);
1307 bool TargetTransformInfo::preferEpilogueVectorization() const {
1308 return TTIImpl
->preferEpilogueVectorization();
1311 TargetTransformInfo::VPLegalization
1312 TargetTransformInfo::getVPLegalizationStrategy(const VPIntrinsic
&VPI
) const {
1313 return TTIImpl
->getVPLegalizationStrategy(VPI
);
1316 bool TargetTransformInfo::hasArmWideBranch(bool Thumb
) const {
1317 return TTIImpl
->hasArmWideBranch(Thumb
);
1320 unsigned TargetTransformInfo::getMaxNumArgs() const {
1321 return TTIImpl
->getMaxNumArgs();
1324 bool TargetTransformInfo::shouldExpandReduction(const IntrinsicInst
*II
) const {
1325 return TTIImpl
->shouldExpandReduction(II
);
1328 TargetTransformInfo::ReductionShuffle
1329 TargetTransformInfo::getPreferredExpandedReductionShuffle(
1330 const IntrinsicInst
*II
) const {
1331 return TTIImpl
->getPreferredExpandedReductionShuffle(II
);
1334 unsigned TargetTransformInfo::getGISelRematGlobalCost() const {
1335 return TTIImpl
->getGISelRematGlobalCost();
1338 unsigned TargetTransformInfo::getMinTripCountTailFoldingThreshold() const {
1339 return TTIImpl
->getMinTripCountTailFoldingThreshold();
1342 bool TargetTransformInfo::supportsScalableVectors() const {
1343 return TTIImpl
->supportsScalableVectors();
1346 bool TargetTransformInfo::enableScalableVectorization() const {
1347 return TTIImpl
->enableScalableVectorization();
1350 bool TargetTransformInfo::hasActiveVectorLength(unsigned Opcode
, Type
*DataType
,
1351 Align Alignment
) const {
1352 return TTIImpl
->hasActiveVectorLength(Opcode
, DataType
, Alignment
);
1355 TargetTransformInfo::Concept::~Concept() = default;
1357 TargetIRAnalysis::TargetIRAnalysis() : TTICallback(&getDefaultTTI
) {}
1359 TargetIRAnalysis::TargetIRAnalysis(
1360 std::function
<Result(const Function
&)> TTICallback
)
1361 : TTICallback(std::move(TTICallback
)) {}
1363 TargetIRAnalysis::Result
TargetIRAnalysis::run(const Function
&F
,
1364 FunctionAnalysisManager
&) {
1365 return TTICallback(F
);
1368 AnalysisKey
TargetIRAnalysis::Key
;
1370 TargetIRAnalysis::Result
TargetIRAnalysis::getDefaultTTI(const Function
&F
) {
1371 return Result(F
.getDataLayout());
1374 // Register the basic pass.
1375 INITIALIZE_PASS(TargetTransformInfoWrapperPass
, "tti",
1376 "Target Transform Information", false, true)
1377 char TargetTransformInfoWrapperPass::ID
= 0;
1379 void TargetTransformInfoWrapperPass::anchor() {}
1381 TargetTransformInfoWrapperPass::TargetTransformInfoWrapperPass()
1382 : ImmutablePass(ID
) {
1383 initializeTargetTransformInfoWrapperPassPass(
1384 *PassRegistry::getPassRegistry());
1387 TargetTransformInfoWrapperPass::TargetTransformInfoWrapperPass(
1388 TargetIRAnalysis TIRA
)
1389 : ImmutablePass(ID
), TIRA(std::move(TIRA
)) {
1390 initializeTargetTransformInfoWrapperPassPass(
1391 *PassRegistry::getPassRegistry());
1394 TargetTransformInfo
&TargetTransformInfoWrapperPass::getTTI(const Function
&F
) {
1395 FunctionAnalysisManager DummyFAM
;
1396 TTI
= TIRA
.run(F
, DummyFAM
);
1401 llvm::createTargetTransformInfoWrapperPass(TargetIRAnalysis TIRA
) {
1402 return new TargetTransformInfoWrapperPass(std::move(TIRA
));