1 //===- BasicTTIImpl.h -------------------------------------------*- C++ -*-===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
11 /// This file provides a helper that implements much of the TTI interface in
12 /// terms of the target-independent code generator and TargetLowering
15 //===----------------------------------------------------------------------===//
17 #ifndef LLVM_CODEGEN_BASICTTIIMPL_H
18 #define LLVM_CODEGEN_BASICTTIIMPL_H
20 #include "llvm/ADT/APInt.h"
21 #include "llvm/ADT/ArrayRef.h"
22 #include "llvm/ADT/BitVector.h"
23 #include "llvm/ADT/SmallPtrSet.h"
24 #include "llvm/ADT/SmallVector.h"
25 #include "llvm/Analysis/LoopInfo.h"
26 #include "llvm/Analysis/TargetTransformInfo.h"
27 #include "llvm/Analysis/TargetTransformInfoImpl.h"
28 #include "llvm/CodeGen/ISDOpcodes.h"
29 #include "llvm/CodeGen/MachineValueType.h"
30 #include "llvm/CodeGen/TargetLowering.h"
31 #include "llvm/CodeGen/TargetSubtargetInfo.h"
32 #include "llvm/CodeGen/ValueTypes.h"
33 #include "llvm/IR/BasicBlock.h"
34 #include "llvm/IR/CallSite.h"
35 #include "llvm/IR/Constant.h"
36 #include "llvm/IR/Constants.h"
37 #include "llvm/IR/DataLayout.h"
38 #include "llvm/IR/DerivedTypes.h"
39 #include "llvm/IR/InstrTypes.h"
40 #include "llvm/IR/Instruction.h"
41 #include "llvm/IR/Instructions.h"
42 #include "llvm/IR/Intrinsics.h"
43 #include "llvm/IR/Operator.h"
44 #include "llvm/IR/Type.h"
45 #include "llvm/IR/Value.h"
46 #include "llvm/MC/MCSchedule.h"
47 #include "llvm/Support/Casting.h"
48 #include "llvm/Support/CommandLine.h"
49 #include "llvm/Support/ErrorHandling.h"
50 #include "llvm/Support/MathExtras.h"
62 class ScalarEvolution;
66 extern cl::opt<unsigned> PartialUnrollingThreshold;
68 /// \brief Base class which can be used to help build a TTI implementation.
70 /// This class provides as much implementation of the TTI interface as is
71 /// possible using the target independent parts of the code generator.
73 /// In order to subclass it, your class must implement a getST() method to
74 /// return the subtarget, and a getTLI() method to return the target lowering.
75 /// We need these methods implemented in the derived class so that this class
76 /// doesn't have to duplicate storage for them.
78 class BasicTTIImplBase : public TargetTransformInfoImplCRTPBase<T> {
80 using BaseT = TargetTransformInfoImplCRTPBase<T>;
81 using TTI = TargetTransformInfo;
83 /// Estimate a cost of shuffle as a sequence of extract and insert
85 unsigned getPermuteShuffleOverhead(Type *Ty) {
86 assert(Ty->isVectorTy() && "Can only shuffle vectors");
88 // Shuffle cost is equal to the cost of extracting element from its argument
89 // plus the cost of inserting them onto the result vector.
91 // e.g. <4 x float> has a mask of <0,5,2,7> i.e we need to extract from
92 // index 0 of first vector, index 1 of second vector,index 2 of first
93 // vector and finally index 3 of second vector and insert them at index
94 // <0,1,2,3> of result vector.
95 for (int i = 0, e = Ty->getVectorNumElements(); i < e; ++i) {
96 Cost += static_cast<T *>(this)
97 ->getVectorInstrCost(Instruction::InsertElement, Ty, i);
98 Cost += static_cast<T *>(this)
99 ->getVectorInstrCost(Instruction::ExtractElement, Ty, i);
104 /// \brief Local query method delegates up to T which *must* implement this!
105 const TargetSubtargetInfo *getST() const {
106 return static_cast<const T *>(this)->getST();
109 /// \brief Local query method delegates up to T which *must* implement this!
110 const TargetLoweringBase *getTLI() const {
111 return static_cast<const T *>(this)->getTLI();
115 explicit BasicTTIImplBase(const TargetMachine *TM, const DataLayout &DL)
118 using TargetTransformInfoImplBase::DL;
121 /// \name Scalar TTI Implementations
123 bool allowsMisalignedMemoryAccesses(LLVMContext &Context,
124 unsigned BitWidth, unsigned AddressSpace,
125 unsigned Alignment, bool *Fast) const {
126 EVT E = EVT::getIntegerVT(Context, BitWidth);
127 return getTLI()->allowsMisalignedMemoryAccesses(E, AddressSpace, Alignment, Fast);
130 bool hasBranchDivergence() { return false; }
132 bool isSourceOfDivergence(const Value *V) { return false; }
134 bool isAlwaysUniform(const Value *V) { return false; }
136 unsigned getFlatAddressSpace() {
137 // Return an invalid address space.
141 bool isLegalAddImmediate(int64_t imm) {
142 return getTLI()->isLegalAddImmediate(imm);
145 bool isLegalICmpImmediate(int64_t imm) {
146 return getTLI()->isLegalICmpImmediate(imm);
149 bool isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset,
150 bool HasBaseReg, int64_t Scale,
151 unsigned AddrSpace, Instruction *I = nullptr) {
152 TargetLoweringBase::AddrMode AM;
154 AM.BaseOffs = BaseOffset;
155 AM.HasBaseReg = HasBaseReg;
157 return getTLI()->isLegalAddressingMode(DL, AM, Ty, AddrSpace, I);
160 bool isLSRCostLess(TTI::LSRCost C1, TTI::LSRCost C2) {
161 return TargetTransformInfoImplBase::isLSRCostLess(C1, C2);
164 int getScalingFactorCost(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset,
165 bool HasBaseReg, int64_t Scale, unsigned AddrSpace) {
166 TargetLoweringBase::AddrMode AM;
168 AM.BaseOffs = BaseOffset;
169 AM.HasBaseReg = HasBaseReg;
171 return getTLI()->getScalingFactorCost(DL, AM, Ty, AddrSpace);
174 bool isTruncateFree(Type *Ty1, Type *Ty2) {
175 return getTLI()->isTruncateFree(Ty1, Ty2);
178 bool isProfitableToHoist(Instruction *I) {
179 return getTLI()->isProfitableToHoist(I);
182 bool isTypeLegal(Type *Ty) {
183 EVT VT = getTLI()->getValueType(DL, Ty);
184 return getTLI()->isTypeLegal(VT);
187 int getGEPCost(Type *PointeeType, const Value *Ptr,
188 ArrayRef<const Value *> Operands) {
189 return BaseT::getGEPCost(PointeeType, Ptr, Operands);
192 int getExtCost(const Instruction *I, const Value *Src) {
193 if (getTLI()->isExtFree(I))
194 return TargetTransformInfo::TCC_Free;
196 if (isa<ZExtInst>(I) || isa<SExtInst>(I))
197 if (const LoadInst *LI = dyn_cast<LoadInst>(Src))
198 if (getTLI()->isExtLoad(LI, I, DL))
199 return TargetTransformInfo::TCC_Free;
201 return TargetTransformInfo::TCC_Basic;
204 unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
205 ArrayRef<const Value *> Arguments) {
206 return BaseT::getIntrinsicCost(IID, RetTy, Arguments);
209 unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
210 ArrayRef<Type *> ParamTys) {
211 if (IID == Intrinsic::cttz) {
212 if (getTLI()->isCheapToSpeculateCttz())
213 return TargetTransformInfo::TCC_Basic;
214 return TargetTransformInfo::TCC_Expensive;
217 if (IID == Intrinsic::ctlz) {
218 if (getTLI()->isCheapToSpeculateCtlz())
219 return TargetTransformInfo::TCC_Basic;
220 return TargetTransformInfo::TCC_Expensive;
223 return BaseT::getIntrinsicCost(IID, RetTy, ParamTys);
226 unsigned getEstimatedNumberOfCaseClusters(const SwitchInst &SI,
227 unsigned &JumpTableSize) {
228 /// Try to find the estimated number of clusters. Note that the number of
229 /// clusters identified in this function could be different from the actural
230 /// numbers found in lowering. This function ignore switches that are
231 /// lowered with a mix of jump table / bit test / BTree. This function was
232 /// initially intended to be used when estimating the cost of switch in
233 /// inline cost heuristic, but it's a generic cost model to be used in other
234 /// places (e.g., in loop unrolling).
235 unsigned N = SI.getNumCases();
236 const TargetLoweringBase *TLI = getTLI();
237 const DataLayout &DL = this->getDataLayout();
240 bool IsJTAllowed = TLI->areJTsAllowed(SI.getParent()->getParent());
242 // Early exit if both a jump table and bit test are not allowed.
243 if (N < 1 || (!IsJTAllowed && DL.getPointerSizeInBits() < N))
246 APInt MaxCaseVal = SI.case_begin()->getCaseValue()->getValue();
247 APInt MinCaseVal = MaxCaseVal;
248 for (auto CI : SI.cases()) {
249 const APInt &CaseVal = CI.getCaseValue()->getValue();
250 if (CaseVal.sgt(MaxCaseVal))
251 MaxCaseVal = CaseVal;
252 if (CaseVal.slt(MinCaseVal))
253 MinCaseVal = CaseVal;
256 // Check if suitable for a bit test
257 if (N <= DL.getPointerSizeInBits()) {
258 SmallPtrSet<const BasicBlock *, 4> Dests;
259 for (auto I : SI.cases())
260 Dests.insert(I.getCaseSuccessor());
262 if (TLI->isSuitableForBitTests(Dests.size(), N, MinCaseVal, MaxCaseVal,
267 // Check if suitable for a jump table.
269 if (N < 2 || N < TLI->getMinimumJumpTableEntries())
272 (MaxCaseVal - MinCaseVal)
273 .getLimitedValue(std::numeric_limits<uint64_t>::max() - 1) + 1;
274 // Check whether a range of clusters is dense enough for a jump table
275 if (TLI->isSuitableForJumpTable(&SI, N, Range)) {
276 JumpTableSize = Range;
283 unsigned getJumpBufAlignment() { return getTLI()->getJumpBufAlignment(); }
285 unsigned getJumpBufSize() { return getTLI()->getJumpBufSize(); }
287 bool shouldBuildLookupTables() {
288 const TargetLoweringBase *TLI = getTLI();
289 return TLI->isOperationLegalOrCustom(ISD::BR_JT, MVT::Other) ||
290 TLI->isOperationLegalOrCustom(ISD::BRIND, MVT::Other);
293 bool haveFastSqrt(Type *Ty) {
294 const TargetLoweringBase *TLI = getTLI();
295 EVT VT = TLI->getValueType(DL, Ty);
296 return TLI->isTypeLegal(VT) &&
297 TLI->isOperationLegalOrCustom(ISD::FSQRT, VT);
300 bool isFCmpOrdCheaperThanFCmpZero(Type *Ty) {
304 unsigned getFPOpCost(Type *Ty) {
305 // Check whether FADD is available, as a proxy for floating-point in
307 const TargetLoweringBase *TLI = getTLI();
308 EVT VT = TLI->getValueType(DL, Ty);
309 if (TLI->isOperationLegalOrCustomOrPromote(ISD::FADD, VT))
310 return TargetTransformInfo::TCC_Basic;
311 return TargetTransformInfo::TCC_Expensive;
314 unsigned getOperationCost(unsigned Opcode, Type *Ty, Type *OpTy) {
315 const TargetLoweringBase *TLI = getTLI();
318 case Instruction::Trunc:
319 if (TLI->isTruncateFree(OpTy, Ty))
320 return TargetTransformInfo::TCC_Free;
321 return TargetTransformInfo::TCC_Basic;
322 case Instruction::ZExt:
323 if (TLI->isZExtFree(OpTy, Ty))
324 return TargetTransformInfo::TCC_Free;
325 return TargetTransformInfo::TCC_Basic;
328 return BaseT::getOperationCost(Opcode, Ty, OpTy);
331 unsigned getInliningThresholdMultiplier() { return 1; }
333 void getUnrollingPreferences(Loop *L, ScalarEvolution &SE,
334 TTI::UnrollingPreferences &UP) {
335 // This unrolling functionality is target independent, but to provide some
336 // motivation for its intended use, for x86:
338 // According to the Intel 64 and IA-32 Architectures Optimization Reference
339 // Manual, Intel Core models and later have a loop stream detector (and
340 // associated uop queue) that can benefit from partial unrolling.
341 // The relevant requirements are:
342 // - The loop must have no more than 4 (8 for Nehalem and later) branches
343 // taken, and none of them may be calls.
344 // - The loop can have no more than 18 (28 for Nehalem and later) uops.
346 // According to the Software Optimization Guide for AMD Family 15h
347 // Processors, models 30h-4fh (Steamroller and later) have a loop predictor
348 // and loop buffer which can benefit from partial unrolling.
349 // The relevant requirements are:
350 // - The loop must have fewer than 16 branches
351 // - The loop must have less than 40 uops in all executed loop branches
353 // The number of taken branches in a loop is hard to estimate here, and
354 // benchmarking has revealed that it is better not to be conservative when
355 // estimating the branch count. As a result, we'll ignore the branch limits
356 // until someone finds a case where it matters in practice.
359 const TargetSubtargetInfo *ST = getST();
360 if (PartialUnrollingThreshold.getNumOccurrences() > 0)
361 MaxOps = PartialUnrollingThreshold;
362 else if (ST->getSchedModel().LoopMicroOpBufferSize > 0)
363 MaxOps = ST->getSchedModel().LoopMicroOpBufferSize;
367 // Scan the loop: don't unroll loops with calls.
368 for (Loop::block_iterator I = L->block_begin(), E = L->block_end(); I != E;
372 for (BasicBlock::iterator J = BB->begin(), JE = BB->end(); J != JE; ++J)
373 if (isa<CallInst>(J) || isa<InvokeInst>(J)) {
374 ImmutableCallSite CS(&*J);
375 if (const Function *F = CS.getCalledFunction()) {
376 if (!static_cast<T *>(this)->isLoweredToCall(F))
384 // Enable runtime and partial unrolling up to the specified size.
385 // Enable using trip count upper bound to unroll loops.
386 UP.Partial = UP.Runtime = UP.UpperBound = true;
387 UP.PartialThreshold = MaxOps;
389 // Avoid unrolling when optimizing for size.
390 UP.OptSizeThreshold = 0;
391 UP.PartialOptSizeThreshold = 0;
393 // Set number of instructions optimized when "back edge"
394 // becomes "fall through" to default value of 2.
398 int getInstructionLatency(const Instruction *I) {
399 if (isa<LoadInst>(I))
400 return getST()->getSchedModel().DefaultLoadLatency;
402 return BaseT::getInstructionLatency(I);
407 /// \name Vector TTI Implementations
410 unsigned getNumberOfRegisters(bool Vector) { return Vector ? 0 : 1; }
412 unsigned getRegisterBitWidth(bool Vector) const { return 32; }
414 /// Estimate the overhead of scalarizing an instruction. Insert and Extract
415 /// are set if the result needs to be inserted and/or extracted from vectors.
416 unsigned getScalarizationOverhead(Type *Ty, bool Insert, bool Extract) {
417 assert(Ty->isVectorTy() && "Can only scalarize vectors");
420 for (int i = 0, e = Ty->getVectorNumElements(); i < e; ++i) {
422 Cost += static_cast<T *>(this)
423 ->getVectorInstrCost(Instruction::InsertElement, Ty, i);
425 Cost += static_cast<T *>(this)
426 ->getVectorInstrCost(Instruction::ExtractElement, Ty, i);
432 /// Estimate the overhead of scalarizing an instructions unique
433 /// non-constant operands. The types of the arguments are ordinarily
434 /// scalar, in which case the costs are multiplied with VF.
435 unsigned getOperandsScalarizationOverhead(ArrayRef<const Value *> Args,
438 SmallPtrSet<const Value*, 4> UniqueOperands;
439 for (const Value *A : Args) {
440 if (!isa<Constant>(A) && UniqueOperands.insert(A).second) {
441 Type *VecTy = nullptr;
442 if (A->getType()->isVectorTy()) {
443 VecTy = A->getType();
444 // If A is a vector operand, VF should be 1 or correspond to A.
445 assert((VF == 1 || VF == VecTy->getVectorNumElements()) &&
446 "Vector argument does not match VF");
449 VecTy = VectorType::get(A->getType(), VF);
451 Cost += getScalarizationOverhead(VecTy, false, true);
458 unsigned getScalarizationOverhead(Type *VecTy, ArrayRef<const Value *> Args) {
459 assert(VecTy->isVectorTy());
463 Cost += getScalarizationOverhead(VecTy, true, false);
465 Cost += getOperandsScalarizationOverhead(Args,
466 VecTy->getVectorNumElements());
468 // When no information on arguments is provided, we add the cost
469 // associated with one argument as a heuristic.
470 Cost += getScalarizationOverhead(VecTy, false, true);
475 unsigned getMaxInterleaveFactor(unsigned VF) { return 1; }
477 unsigned getArithmeticInstrCost(
478 unsigned Opcode, Type *Ty,
479 TTI::OperandValueKind Opd1Info = TTI::OK_AnyValue,
480 TTI::OperandValueKind Opd2Info = TTI::OK_AnyValue,
481 TTI::OperandValueProperties Opd1PropInfo = TTI::OP_None,
482 TTI::OperandValueProperties Opd2PropInfo = TTI::OP_None,
483 ArrayRef<const Value *> Args = ArrayRef<const Value *>()) {
484 // Check if any of the operands are vector operands.
485 const TargetLoweringBase *TLI = getTLI();
486 int ISD = TLI->InstructionOpcodeToISD(Opcode);
487 assert(ISD && "Invalid opcode");
489 std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(DL, Ty);
491 bool IsFloat = Ty->isFPOrFPVectorTy();
492 // Assume that floating point arithmetic operations cost twice as much as
493 // integer operations.
494 unsigned OpCost = (IsFloat ? 2 : 1);
496 if (TLI->isOperationLegalOrPromote(ISD, LT.second)) {
497 // The operation is legal. Assume it costs 1.
498 // TODO: Once we have extract/insert subvector cost we need to use them.
499 return LT.first * OpCost;
502 if (!TLI->isOperationExpand(ISD, LT.second)) {
503 // If the operation is custom lowered, then assume that the code is twice
505 return LT.first * 2 * OpCost;
508 // Else, assume that we need to scalarize this op.
509 // TODO: If one of the types get legalized by splitting, handle this
510 // similarly to what getCastInstrCost() does.
511 if (Ty->isVectorTy()) {
512 unsigned Num = Ty->getVectorNumElements();
513 unsigned Cost = static_cast<T *>(this)
514 ->getArithmeticInstrCost(Opcode, Ty->getScalarType());
515 // Return the cost of multiple scalar invocation plus the cost of
516 // inserting and extracting the values.
517 return getScalarizationOverhead(Ty, Args) + Num * Cost;
520 // We don't know anything about this scalar instruction.
524 unsigned getShuffleCost(TTI::ShuffleKind Kind, Type *Tp, int Index,
526 if (Kind == TTI::SK_Alternate || Kind == TTI::SK_PermuteTwoSrc ||
527 Kind == TTI::SK_PermuteSingleSrc) {
528 return getPermuteShuffleOverhead(Tp);
533 unsigned getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src,
534 const Instruction *I = nullptr) {
535 const TargetLoweringBase *TLI = getTLI();
536 int ISD = TLI->InstructionOpcodeToISD(Opcode);
537 assert(ISD && "Invalid opcode");
538 std::pair<unsigned, MVT> SrcLT = TLI->getTypeLegalizationCost(DL, Src);
539 std::pair<unsigned, MVT> DstLT = TLI->getTypeLegalizationCost(DL, Dst);
541 // Check for NOOP conversions.
542 if (SrcLT.first == DstLT.first &&
543 SrcLT.second.getSizeInBits() == DstLT.second.getSizeInBits()) {
545 // Bitcast between types that are legalized to the same type are free.
546 if (Opcode == Instruction::BitCast || Opcode == Instruction::Trunc)
550 if (Opcode == Instruction::Trunc &&
551 TLI->isTruncateFree(SrcLT.second, DstLT.second))
554 if (Opcode == Instruction::ZExt &&
555 TLI->isZExtFree(SrcLT.second, DstLT.second))
558 if (Opcode == Instruction::AddrSpaceCast &&
559 TLI->isNoopAddrSpaceCast(Src->getPointerAddressSpace(),
560 Dst->getPointerAddressSpace()))
563 // If this is a zext/sext of a load, return 0 if the corresponding
564 // extending load exists on target.
565 if ((Opcode == Instruction::ZExt || Opcode == Instruction::SExt) &&
566 I && isa<LoadInst>(I->getOperand(0))) {
567 EVT ExtVT = EVT::getEVT(Dst);
568 EVT LoadVT = EVT::getEVT(Src);
570 ((Opcode == Instruction::ZExt) ? ISD::ZEXTLOAD : ISD::SEXTLOAD);
571 if (TLI->isLoadExtLegal(LType, ExtVT, LoadVT))
575 // If the cast is marked as legal (or promote) then assume low cost.
576 if (SrcLT.first == DstLT.first &&
577 TLI->isOperationLegalOrPromote(ISD, DstLT.second))
580 // Handle scalar conversions.
581 if (!Src->isVectorTy() && !Dst->isVectorTy()) {
582 // Scalar bitcasts are usually free.
583 if (Opcode == Instruction::BitCast)
586 // Just check the op cost. If the operation is legal then assume it costs
588 if (!TLI->isOperationExpand(ISD, DstLT.second))
591 // Assume that illegal scalar instruction are expensive.
595 // Check vector-to-vector casts.
596 if (Dst->isVectorTy() && Src->isVectorTy()) {
597 // If the cast is between same-sized registers, then the check is simple.
598 if (SrcLT.first == DstLT.first &&
599 SrcLT.second.getSizeInBits() == DstLT.second.getSizeInBits()) {
601 // Assume that Zext is done using AND.
602 if (Opcode == Instruction::ZExt)
605 // Assume that sext is done using SHL and SRA.
606 if (Opcode == Instruction::SExt)
609 // Just check the op cost. If the operation is legal then assume it
611 // 1 and multiply by the type-legalization overhead.
612 if (!TLI->isOperationExpand(ISD, DstLT.second))
613 return SrcLT.first * 1;
616 // If we are legalizing by splitting, query the concrete TTI for the cost
617 // of casting the original vector twice. We also need to factor int the
618 // cost of the split itself. Count that as 1, to be consistent with
619 // TLI->getTypeLegalizationCost().
620 if ((TLI->getTypeAction(Src->getContext(), TLI->getValueType(DL, Src)) ==
621 TargetLowering::TypeSplitVector) ||
622 (TLI->getTypeAction(Dst->getContext(), TLI->getValueType(DL, Dst)) ==
623 TargetLowering::TypeSplitVector)) {
624 Type *SplitDst = VectorType::get(Dst->getVectorElementType(),
625 Dst->getVectorNumElements() / 2);
626 Type *SplitSrc = VectorType::get(Src->getVectorElementType(),
627 Src->getVectorNumElements() / 2);
628 T *TTI = static_cast<T *>(this);
629 return TTI->getVectorSplitCost() +
630 (2 * TTI->getCastInstrCost(Opcode, SplitDst, SplitSrc, I));
633 // In other cases where the source or destination are illegal, assume
634 // the operation will get scalarized.
635 unsigned Num = Dst->getVectorNumElements();
636 unsigned Cost = static_cast<T *>(this)->getCastInstrCost(
637 Opcode, Dst->getScalarType(), Src->getScalarType(), I);
639 // Return the cost of multiple scalar invocation plus the cost of
640 // inserting and extracting the values.
641 return getScalarizationOverhead(Dst, true, true) + Num * Cost;
644 // We already handled vector-to-vector and scalar-to-scalar conversions.
646 // is where we handle bitcast between vectors and scalars. We need to assume
647 // that the conversion is scalarized in one way or another.
648 if (Opcode == Instruction::BitCast)
649 // Illegal bitcasts are done by storing and loading from a stack slot.
650 return (Src->isVectorTy() ? getScalarizationOverhead(Src, false, true)
652 (Dst->isVectorTy() ? getScalarizationOverhead(Dst, true, false)
655 llvm_unreachable("Unhandled cast");
658 unsigned getExtractWithExtendCost(unsigned Opcode, Type *Dst,
659 VectorType *VecTy, unsigned Index) {
660 return static_cast<T *>(this)->getVectorInstrCost(
661 Instruction::ExtractElement, VecTy, Index) +
662 static_cast<T *>(this)->getCastInstrCost(Opcode, Dst,
663 VecTy->getElementType());
666 unsigned getCFInstrCost(unsigned Opcode) {
667 // Branches are assumed to be predicted.
671 unsigned getCmpSelInstrCost(unsigned Opcode, Type *ValTy, Type *CondTy,
672 const Instruction *I) {
673 const TargetLoweringBase *TLI = getTLI();
674 int ISD = TLI->InstructionOpcodeToISD(Opcode);
675 assert(ISD && "Invalid opcode");
677 // Selects on vectors are actually vector selects.
678 if (ISD == ISD::SELECT) {
679 assert(CondTy && "CondTy must exist");
680 if (CondTy->isVectorTy())
683 std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(DL, ValTy);
685 if (!(ValTy->isVectorTy() && !LT.second.isVector()) &&
686 !TLI->isOperationExpand(ISD, LT.second)) {
687 // The operation is legal. Assume it costs 1. Multiply
688 // by the type-legalization overhead.
692 // Otherwise, assume that the cast is scalarized.
693 // TODO: If one of the types get legalized by splitting, handle this
694 // similarly to what getCastInstrCost() does.
695 if (ValTy->isVectorTy()) {
696 unsigned Num = ValTy->getVectorNumElements();
698 CondTy = CondTy->getScalarType();
699 unsigned Cost = static_cast<T *>(this)->getCmpSelInstrCost(
700 Opcode, ValTy->getScalarType(), CondTy, I);
702 // Return the cost of multiple scalar invocation plus the cost of
703 // inserting and extracting the values.
704 return getScalarizationOverhead(ValTy, true, false) + Num * Cost;
707 // Unknown scalar opcode.
711 unsigned getVectorInstrCost(unsigned Opcode, Type *Val, unsigned Index) {
712 std::pair<unsigned, MVT> LT =
713 getTLI()->getTypeLegalizationCost(DL, Val->getScalarType());
718 unsigned getMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
719 unsigned AddressSpace, const Instruction *I = nullptr) {
720 assert(!Src->isVoidTy() && "Invalid type");
721 std::pair<unsigned, MVT> LT = getTLI()->getTypeLegalizationCost(DL, Src);
723 // Assuming that all loads of legal types cost 1.
724 unsigned Cost = LT.first;
726 if (Src->isVectorTy() &&
727 Src->getPrimitiveSizeInBits() < LT.second.getSizeInBits()) {
728 // This is a vector load that legalizes to a larger type than the vector
729 // itself. Unless the corresponding extending load or truncating store is
730 // legal, then this will scalarize.
731 TargetLowering::LegalizeAction LA = TargetLowering::Expand;
732 EVT MemVT = getTLI()->getValueType(DL, Src);
733 if (Opcode == Instruction::Store)
734 LA = getTLI()->getTruncStoreAction(LT.second, MemVT);
736 LA = getTLI()->getLoadExtAction(ISD::EXTLOAD, LT.second, MemVT);
738 if (LA != TargetLowering::Legal && LA != TargetLowering::Custom) {
739 // This is a vector load/store for some illegal type that is scalarized.
740 // We must account for the cost of building or decomposing the vector.
741 Cost += getScalarizationOverhead(Src, Opcode != Instruction::Store,
742 Opcode == Instruction::Store);
749 unsigned getInterleavedMemoryOpCost(unsigned Opcode, Type *VecTy,
751 ArrayRef<unsigned> Indices,
753 unsigned AddressSpace) {
754 VectorType *VT = dyn_cast<VectorType>(VecTy);
755 assert(VT && "Expect a vector type for interleaved memory op");
757 unsigned NumElts = VT->getNumElements();
758 assert(Factor > 1 && NumElts % Factor == 0 && "Invalid interleave factor");
760 unsigned NumSubElts = NumElts / Factor;
761 VectorType *SubVT = VectorType::get(VT->getElementType(), NumSubElts);
763 // Firstly, the cost of load/store operation.
764 unsigned Cost = static_cast<T *>(this)->getMemoryOpCost(
765 Opcode, VecTy, Alignment, AddressSpace);
767 // Legalize the vector type, and get the legalized and unlegalized type
769 MVT VecTyLT = getTLI()->getTypeLegalizationCost(DL, VecTy).second;
771 static_cast<T *>(this)->getDataLayout().getTypeStoreSize(VecTy);
772 unsigned VecTyLTSize = VecTyLT.getStoreSize();
774 // Return the ceiling of dividing A by B.
775 auto ceil = [](unsigned A, unsigned B) { return (A + B - 1) / B; };
777 // Scale the cost of the memory operation by the fraction of legalized
778 // instructions that will actually be used. We shouldn't account for the
779 // cost of dead instructions since they will be removed.
781 // E.g., An interleaved load of factor 8:
782 // %vec = load <16 x i64>, <16 x i64>* %ptr
783 // %v0 = shufflevector %vec, undef, <0, 8>
785 // If <16 x i64> is legalized to 8 v2i64 loads, only 2 of the loads will be
786 // used (those corresponding to elements [0:1] and [8:9] of the unlegalized
787 // type). The other loads are unused.
789 // We only scale the cost of loads since interleaved store groups aren't
790 // allowed to have gaps.
791 if (Opcode == Instruction::Load && VecTySize > VecTyLTSize) {
792 // The number of loads of a legal type it will take to represent a load
793 // of the unlegalized vector type.
794 unsigned NumLegalInsts = ceil(VecTySize, VecTyLTSize);
796 // The number of elements of the unlegalized type that correspond to a
797 // single legal instruction.
798 unsigned NumEltsPerLegalInst = ceil(NumElts, NumLegalInsts);
800 // Determine which legal instructions will be used.
801 BitVector UsedInsts(NumLegalInsts, false);
802 for (unsigned Index : Indices)
803 for (unsigned Elt = 0; Elt < NumSubElts; ++Elt)
804 UsedInsts.set((Index + Elt * Factor) / NumEltsPerLegalInst);
806 // Scale the cost of the load by the fraction of legal instructions that
808 Cost *= UsedInsts.count() / NumLegalInsts;
811 // Then plus the cost of interleave operation.
812 if (Opcode == Instruction::Load) {
813 // The interleave cost is similar to extract sub vectors' elements
814 // from the wide vector, and insert them into sub vectors.
816 // E.g. An interleaved load of factor 2 (with one member of index 0):
817 // %vec = load <8 x i32>, <8 x i32>* %ptr
818 // %v0 = shuffle %vec, undef, <0, 2, 4, 6> ; Index 0
819 // The cost is estimated as extract elements at 0, 2, 4, 6 from the
820 // <8 x i32> vector and insert them into a <4 x i32> vector.
822 assert(Indices.size() <= Factor &&
823 "Interleaved memory op has too many members");
825 for (unsigned Index : Indices) {
826 assert(Index < Factor && "Invalid index for interleaved memory op");
828 // Extract elements from loaded vector for each sub vector.
829 for (unsigned i = 0; i < NumSubElts; i++)
830 Cost += static_cast<T *>(this)->getVectorInstrCost(
831 Instruction::ExtractElement, VT, Index + i * Factor);
834 unsigned InsSubCost = 0;
835 for (unsigned i = 0; i < NumSubElts; i++)
836 InsSubCost += static_cast<T *>(this)->getVectorInstrCost(
837 Instruction::InsertElement, SubVT, i);
839 Cost += Indices.size() * InsSubCost;
841 // The interleave cost is extract all elements from sub vectors, and
842 // insert them into the wide vector.
844 // E.g. An interleaved store of factor 2:
845 // %v0_v1 = shuffle %v0, %v1, <0, 4, 1, 5, 2, 6, 3, 7>
846 // store <8 x i32> %interleaved.vec, <8 x i32>* %ptr
847 // The cost is estimated as extract all elements from both <4 x i32>
848 // vectors and insert into the <8 x i32> vector.
850 unsigned ExtSubCost = 0;
851 for (unsigned i = 0; i < NumSubElts; i++)
852 ExtSubCost += static_cast<T *>(this)->getVectorInstrCost(
853 Instruction::ExtractElement, SubVT, i);
854 Cost += ExtSubCost * Factor;
856 for (unsigned i = 0; i < NumElts; i++)
857 Cost += static_cast<T *>(this)
858 ->getVectorInstrCost(Instruction::InsertElement, VT, i);
864 /// Get intrinsic cost based on arguments.
865 unsigned getIntrinsicInstrCost(Intrinsic::ID IID, Type *RetTy,
866 ArrayRef<Value *> Args, FastMathFlags FMF,
868 unsigned RetVF = (RetTy->isVectorTy() ? RetTy->getVectorNumElements() : 1);
869 assert((RetVF == 1 || VF == 1) && "VF > 1 and RetVF is a vector type");
873 // Assume that we need to scalarize this intrinsic.
874 SmallVector<Type *, 4> Types;
875 for (Value *Op : Args) {
876 Type *OpTy = Op->getType();
877 assert(VF == 1 || !OpTy->isVectorTy());
878 Types.push_back(VF == 1 ? OpTy : VectorType::get(OpTy, VF));
881 if (VF > 1 && !RetTy->isVoidTy())
882 RetTy = VectorType::get(RetTy, VF);
884 // Compute the scalarization overhead based on Args for a vector
885 // intrinsic. A vectorizer will pass a scalar RetTy and VF > 1, while
886 // CostModel will pass a vector RetTy and VF is 1.
887 unsigned ScalarizationCost = std::numeric_limits<unsigned>::max();
888 if (RetVF > 1 || VF > 1) {
889 ScalarizationCost = 0;
890 if (!RetTy->isVoidTy())
891 ScalarizationCost += getScalarizationOverhead(RetTy, true, false);
892 ScalarizationCost += getOperandsScalarizationOverhead(Args, VF);
895 return static_cast<T *>(this)->
896 getIntrinsicInstrCost(IID, RetTy, Types, FMF, ScalarizationCost);
898 case Intrinsic::masked_scatter: {
899 assert(VF == 1 && "Can't vectorize types here.");
900 Value *Mask = Args[3];
901 bool VarMask = !isa<Constant>(Mask);
902 unsigned Alignment = cast<ConstantInt>(Args[2])->getZExtValue();
904 static_cast<T *>(this)->getGatherScatterOpCost(Instruction::Store,
909 case Intrinsic::masked_gather: {
910 assert(VF == 1 && "Can't vectorize types here.");
911 Value *Mask = Args[2];
912 bool VarMask = !isa<Constant>(Mask);
913 unsigned Alignment = cast<ConstantInt>(Args[1])->getZExtValue();
915 static_cast<T *>(this)->getGatherScatterOpCost(Instruction::Load,
916 RetTy, Args[0], VarMask,
922 /// Get intrinsic cost based on argument types.
923 /// If ScalarizationCostPassed is std::numeric_limits<unsigned>::max(), the
924 /// cost of scalarizing the arguments and the return value will be computed
926 unsigned getIntrinsicInstrCost(
927 Intrinsic::ID IID, Type *RetTy, ArrayRef<Type *> Tys, FastMathFlags FMF,
928 unsigned ScalarizationCostPassed = std::numeric_limits<unsigned>::max()) {
929 SmallVector<unsigned, 2> ISDs;
930 unsigned SingleCallCost = 10; // Library call cost. Make it expensive.
933 // Assume that we need to scalarize this intrinsic.
934 unsigned ScalarizationCost = ScalarizationCostPassed;
935 unsigned ScalarCalls = 1;
936 Type *ScalarRetTy = RetTy;
937 if (RetTy->isVectorTy()) {
938 if (ScalarizationCostPassed == std::numeric_limits<unsigned>::max())
939 ScalarizationCost = getScalarizationOverhead(RetTy, true, false);
940 ScalarCalls = std::max(ScalarCalls, RetTy->getVectorNumElements());
941 ScalarRetTy = RetTy->getScalarType();
943 SmallVector<Type *, 4> ScalarTys;
944 for (unsigned i = 0, ie = Tys.size(); i != ie; ++i) {
946 if (Ty->isVectorTy()) {
947 if (ScalarizationCostPassed == std::numeric_limits<unsigned>::max())
948 ScalarizationCost += getScalarizationOverhead(Ty, false, true);
949 ScalarCalls = std::max(ScalarCalls, Ty->getVectorNumElements());
950 Ty = Ty->getScalarType();
952 ScalarTys.push_back(Ty);
954 if (ScalarCalls == 1)
955 return 1; // Return cost of a scalar intrinsic. Assume it to be cheap.
957 unsigned ScalarCost = static_cast<T *>(this)->getIntrinsicInstrCost(
958 IID, ScalarRetTy, ScalarTys, FMF);
960 return ScalarCalls * ScalarCost + ScalarizationCost;
962 // Look for intrinsics that can be lowered directly or turned into a scalar
964 case Intrinsic::sqrt:
965 ISDs.push_back(ISD::FSQRT);
968 ISDs.push_back(ISD::FSIN);
971 ISDs.push_back(ISD::FCOS);
974 ISDs.push_back(ISD::FEXP);
976 case Intrinsic::exp2:
977 ISDs.push_back(ISD::FEXP2);
980 ISDs.push_back(ISD::FLOG);
982 case Intrinsic::log10:
983 ISDs.push_back(ISD::FLOG10);
985 case Intrinsic::log2:
986 ISDs.push_back(ISD::FLOG2);
988 case Intrinsic::fabs:
989 ISDs.push_back(ISD::FABS);
991 case Intrinsic::minnum:
992 ISDs.push_back(ISD::FMINNUM);
994 ISDs.push_back(ISD::FMINNAN);
996 case Intrinsic::maxnum:
997 ISDs.push_back(ISD::FMAXNUM);
999 ISDs.push_back(ISD::FMAXNAN);
1001 case Intrinsic::copysign:
1002 ISDs.push_back(ISD::FCOPYSIGN);
1004 case Intrinsic::floor:
1005 ISDs.push_back(ISD::FFLOOR);
1007 case Intrinsic::ceil:
1008 ISDs.push_back(ISD::FCEIL);
1010 case Intrinsic::trunc:
1011 ISDs.push_back(ISD::FTRUNC);
1013 case Intrinsic::nearbyint:
1014 ISDs.push_back(ISD::FNEARBYINT);
1016 case Intrinsic::rint:
1017 ISDs.push_back(ISD::FRINT);
1019 case Intrinsic::round:
1020 ISDs.push_back(ISD::FROUND);
1022 case Intrinsic::pow:
1023 ISDs.push_back(ISD::FPOW);
1025 case Intrinsic::fma:
1026 ISDs.push_back(ISD::FMA);
1028 case Intrinsic::fmuladd:
1029 ISDs.push_back(ISD::FMA);
1031 // FIXME: We should return 0 whenever getIntrinsicCost == TCC_Free.
1032 case Intrinsic::lifetime_start:
1033 case Intrinsic::lifetime_end:
1034 case Intrinsic::sideeffect:
1036 case Intrinsic::masked_store:
1037 return static_cast<T *>(this)
1038 ->getMaskedMemoryOpCost(Instruction::Store, Tys[0], 0, 0);
1039 case Intrinsic::masked_load:
1040 return static_cast<T *>(this)
1041 ->getMaskedMemoryOpCost(Instruction::Load, RetTy, 0, 0);
1042 case Intrinsic::ctpop:
1043 ISDs.push_back(ISD::CTPOP);
1044 // In case of legalization use TCC_Expensive. This is cheaper than a
1045 // library call but still not a cheap instruction.
1046 SingleCallCost = TargetTransformInfo::TCC_Expensive;
1048 // FIXME: ctlz, cttz, ...
1051 const TargetLoweringBase *TLI = getTLI();
1052 std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(DL, RetTy);
1054 SmallVector<unsigned, 2> LegalCost;
1055 SmallVector<unsigned, 2> CustomCost;
1056 for (unsigned ISD : ISDs) {
1057 if (TLI->isOperationLegalOrPromote(ISD, LT.second)) {
1058 if (IID == Intrinsic::fabs && TLI->isFAbsFree(LT.second)) {
1062 // The operation is legal. Assume it costs 1.
1063 // If the type is split to multiple registers, assume that there is some
1064 // overhead to this.
1065 // TODO: Once we have extract/insert subvector cost we need to use them.
1067 LegalCost.push_back(LT.first * 2);
1069 LegalCost.push_back(LT.first * 1);
1070 } else if (!TLI->isOperationExpand(ISD, LT.second)) {
1071 // If the operation is custom lowered then assume
1072 // that the code is twice as expensive.
1073 CustomCost.push_back(LT.first * 2);
1077 auto MinLegalCostI = std::min_element(LegalCost.begin(), LegalCost.end());
1078 if (MinLegalCostI != LegalCost.end())
1079 return *MinLegalCostI;
1081 auto MinCustomCostI = std::min_element(CustomCost.begin(), CustomCost.end());
1082 if (MinCustomCostI != CustomCost.end())
1083 return *MinCustomCostI;
1085 // If we can't lower fmuladd into an FMA estimate the cost as a floating
1086 // point mul followed by an add.
1087 if (IID == Intrinsic::fmuladd)
1088 return static_cast<T *>(this)
1089 ->getArithmeticInstrCost(BinaryOperator::FMul, RetTy) +
1090 static_cast<T *>(this)
1091 ->getArithmeticInstrCost(BinaryOperator::FAdd, RetTy);
1093 // Else, assume that we need to scalarize this intrinsic. For math builtins
1094 // this will emit a costly libcall, adding call overhead and spills. Make it
1096 if (RetTy->isVectorTy()) {
1097 unsigned ScalarizationCost =
1098 ((ScalarizationCostPassed != std::numeric_limits<unsigned>::max())
1099 ? ScalarizationCostPassed
1100 : getScalarizationOverhead(RetTy, true, false));
1101 unsigned ScalarCalls = RetTy->getVectorNumElements();
1102 SmallVector<Type *, 4> ScalarTys;
1103 for (unsigned i = 0, ie = Tys.size(); i != ie; ++i) {
1105 if (Ty->isVectorTy())
1106 Ty = Ty->getScalarType();
1107 ScalarTys.push_back(Ty);
1109 unsigned ScalarCost = static_cast<T *>(this)->getIntrinsicInstrCost(
1110 IID, RetTy->getScalarType(), ScalarTys, FMF);
1111 for (unsigned i = 0, ie = Tys.size(); i != ie; ++i) {
1112 if (Tys[i]->isVectorTy()) {
1113 if (ScalarizationCostPassed == std::numeric_limits<unsigned>::max())
1114 ScalarizationCost += getScalarizationOverhead(Tys[i], false, true);
1115 ScalarCalls = std::max(ScalarCalls, Tys[i]->getVectorNumElements());
1119 return ScalarCalls * ScalarCost + ScalarizationCost;
1122 // This is going to be turned into a library call, make it expensive.
1123 return SingleCallCost;
1126 /// \brief Compute a cost of the given call instruction.
1128 /// Compute the cost of calling function F with return type RetTy and
1129 /// argument types Tys. F might be nullptr, in this case the cost of an
1130 /// arbitrary call with the specified signature will be returned.
1131 /// This is used, for instance, when we estimate call of a vector
1132 /// counterpart of the given function.
1133 /// \param F Called function, might be nullptr.
1134 /// \param RetTy Return value types.
1135 /// \param Tys Argument types.
1136 /// \returns The cost of Call instruction.
1137 unsigned getCallInstrCost(Function *F, Type *RetTy, ArrayRef<Type *> Tys) {
1141 unsigned getNumberOfParts(Type *Tp) {
1142 std::pair<unsigned, MVT> LT = getTLI()->getTypeLegalizationCost(DL, Tp);
1146 unsigned getAddressComputationCost(Type *Ty, ScalarEvolution *,
1151 /// Try to calculate arithmetic and shuffle op costs for reduction operations.
1152 /// We're assuming that reduction operation are performing the following way:
1153 /// 1. Non-pairwise reduction
1154 /// %val1 = shufflevector<n x t> %val, <n x t> %undef,
1155 /// <n x i32> <i32 n/2, i32 n/2 + 1, ..., i32 n, i32 undef, ..., i32 undef>
1156 /// \----------------v-------------/ \----------v------------/
1157 /// n/2 elements n/2 elements
1158 /// %red1 = op <n x t> %val, <n x t> val1
1159 /// After this operation we have a vector %red1 where only the first n/2
1160 /// elements are meaningful, the second n/2 elements are undefined and can be
1161 /// dropped. All other operations are actually working with the vector of
1162 /// length n/2, not n, though the real vector length is still n.
1163 /// %val2 = shufflevector<n x t> %red1, <n x t> %undef,
1164 /// <n x i32> <i32 n/4, i32 n/4 + 1, ..., i32 n/2, i32 undef, ..., i32 undef>
1165 /// \----------------v-------------/ \----------v------------/
1166 /// n/4 elements 3*n/4 elements
1167 /// %red2 = op <n x t> %red1, <n x t> val2 - working with the vector of
1168 /// length n/2, the resulting vector has length n/4 etc.
1169 /// 2. Pairwise reduction:
1170 /// Everything is the same except for an additional shuffle operation which
1171 /// is used to produce operands for pairwise kind of reductions.
1172 /// %val1 = shufflevector<n x t> %val, <n x t> %undef,
1173 /// <n x i32> <i32 0, i32 2, ..., i32 n-2, i32 undef, ..., i32 undef>
1174 /// \-------------v----------/ \----------v------------/
1175 /// n/2 elements n/2 elements
1176 /// %val2 = shufflevector<n x t> %val, <n x t> %undef,
1177 /// <n x i32> <i32 1, i32 3, ..., i32 n-1, i32 undef, ..., i32 undef>
1178 /// \-------------v----------/ \----------v------------/
1179 /// n/2 elements n/2 elements
1180 /// %red1 = op <n x t> %val1, <n x t> val2
1181 /// Again, the operation is performed on <n x t> vector, but the resulting
1182 /// vector %red1 is <n/2 x t> vector.
1184 /// The cost model should take into account that the actual length of the
1185 /// vector is reduced on each iteration.
1186 unsigned getArithmeticReductionCost(unsigned Opcode, Type *Ty,
1188 assert(Ty->isVectorTy() && "Expect a vector type");
1189 Type *ScalarTy = Ty->getVectorElementType();
1190 unsigned NumVecElts = Ty->getVectorNumElements();
1191 unsigned NumReduxLevels = Log2_32(NumVecElts);
1192 unsigned ArithCost = 0;
1193 unsigned ShuffleCost = 0;
1194 auto *ConcreteTTI = static_cast<T *>(this);
1195 std::pair<unsigned, MVT> LT =
1196 ConcreteTTI->getTLI()->getTypeLegalizationCost(DL, Ty);
1197 unsigned LongVectorCount = 0;
1199 LT.second.isVector() ? LT.second.getVectorNumElements() : 1;
1200 while (NumVecElts > MVTLen) {
1202 // Assume the pairwise shuffles add a cost.
1203 ShuffleCost += (IsPairwise + 1) *
1204 ConcreteTTI->getShuffleCost(TTI::SK_ExtractSubvector, Ty,
1206 ArithCost += ConcreteTTI->getArithmeticInstrCost(Opcode, Ty);
1207 Ty = VectorType::get(ScalarTy, NumVecElts);
1210 // The minimal length of the vector is limited by the real length of vector
1211 // operations performed on the current platform. That's why several final
1212 // reduction operations are performed on the vectors with the same
1213 // architecture-dependent length.
1214 ShuffleCost += (NumReduxLevels - LongVectorCount) * (IsPairwise + 1) *
1215 ConcreteTTI->getShuffleCost(TTI::SK_ExtractSubvector, Ty,
1217 ArithCost += (NumReduxLevels - LongVectorCount) *
1218 ConcreteTTI->getArithmeticInstrCost(Opcode, Ty);
1219 return ShuffleCost + ArithCost + getScalarizationOverhead(Ty, false, true);
1222 /// Try to calculate op costs for min/max reduction operations.
1223 /// \param CondTy Conditional type for the Select instruction.
1224 unsigned getMinMaxReductionCost(Type *Ty, Type *CondTy, bool IsPairwise,
1226 assert(Ty->isVectorTy() && "Expect a vector type");
1227 Type *ScalarTy = Ty->getVectorElementType();
1228 Type *ScalarCondTy = CondTy->getVectorElementType();
1229 unsigned NumVecElts = Ty->getVectorNumElements();
1230 unsigned NumReduxLevels = Log2_32(NumVecElts);
1232 if (Ty->isFPOrFPVectorTy()) {
1233 CmpOpcode = Instruction::FCmp;
1235 assert(Ty->isIntOrIntVectorTy() &&
1236 "expecting floating point or integer type for min/max reduction");
1237 CmpOpcode = Instruction::ICmp;
1239 unsigned MinMaxCost = 0;
1240 unsigned ShuffleCost = 0;
1241 auto *ConcreteTTI = static_cast<T *>(this);
1242 std::pair<unsigned, MVT> LT =
1243 ConcreteTTI->getTLI()->getTypeLegalizationCost(DL, Ty);
1244 unsigned LongVectorCount = 0;
1246 LT.second.isVector() ? LT.second.getVectorNumElements() : 1;
1247 while (NumVecElts > MVTLen) {
1249 // Assume the pairwise shuffles add a cost.
1250 ShuffleCost += (IsPairwise + 1) *
1251 ConcreteTTI->getShuffleCost(TTI::SK_ExtractSubvector, Ty,
1254 ConcreteTTI->getCmpSelInstrCost(CmpOpcode, Ty, CondTy, nullptr) +
1255 ConcreteTTI->getCmpSelInstrCost(Instruction::Select, Ty, CondTy,
1257 Ty = VectorType::get(ScalarTy, NumVecElts);
1258 CondTy = VectorType::get(ScalarCondTy, NumVecElts);
1261 // The minimal length of the vector is limited by the real length of vector
1262 // operations performed on the current platform. That's why several final
1263 // reduction opertions are perfomed on the vectors with the same
1264 // architecture-dependent length.
1265 ShuffleCost += (NumReduxLevels - LongVectorCount) * (IsPairwise + 1) *
1266 ConcreteTTI->getShuffleCost(TTI::SK_ExtractSubvector, Ty,
1269 (NumReduxLevels - LongVectorCount) *
1270 (ConcreteTTI->getCmpSelInstrCost(CmpOpcode, Ty, CondTy, nullptr) +
1271 ConcreteTTI->getCmpSelInstrCost(Instruction::Select, Ty, CondTy,
1273 // Need 3 extractelement instructions for scalarization + an additional
1274 // scalar select instruction.
1275 return ShuffleCost + MinMaxCost +
1276 3 * getScalarizationOverhead(Ty, /*Insert=*/false,
1278 ConcreteTTI->getCmpSelInstrCost(Instruction::Select, ScalarTy,
1279 ScalarCondTy, nullptr);
1282 unsigned getVectorSplitCost() { return 1; }
1287 /// \brief Concrete BasicTTIImpl that can be used if no further customization
1289 class BasicTTIImpl : public BasicTTIImplBase<BasicTTIImpl> {
1290 using BaseT = BasicTTIImplBase<BasicTTIImpl>;
1292 friend class BasicTTIImplBase<BasicTTIImpl>;
1294 const TargetSubtargetInfo *ST;
1295 const TargetLoweringBase *TLI;
1297 const TargetSubtargetInfo *getST() const { return ST; }
1298 const TargetLoweringBase *getTLI() const { return TLI; }
1301 explicit BasicTTIImpl(const TargetMachine *ST, const Function &F);
1304 } // end namespace llvm
1306 #endif // LLVM_CODEGEN_BASICTTIIMPL_H