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 //===----------------------------------------------------------------------===//
10 /// This file provides a helper that implements much of the TTI interface in
11 /// terms of the target-independent code generator and TargetLowering
14 //===----------------------------------------------------------------------===//
16 #ifndef LLVM_CODEGEN_BASICTTIIMPL_H
17 #define LLVM_CODEGEN_BASICTTIIMPL_H
19 #include "llvm/Analysis/LoopInfo.h"
20 #include "llvm/Analysis/TargetLibraryInfo.h"
21 #include "llvm/Analysis/TargetTransformInfoImpl.h"
22 #include "llvm/Support/CommandLine.h"
23 #include "llvm/Target/TargetLowering.h"
24 #include "llvm/Target/TargetSubtargetInfo.h"
28 extern cl::opt<unsigned> PartialUnrollingThreshold;
30 /// \brief Base class which can be used to help build a TTI implementation.
32 /// This class provides as much implementation of the TTI interface as is
33 /// possible using the target independent parts of the code generator.
35 /// In order to subclass it, your class must implement a getST() method to
36 /// return the subtarget, and a getTLI() method to return the target lowering.
37 /// We need these methods implemented in the derived class so that this class
38 /// doesn't have to duplicate storage for them.
40 class BasicTTIImplBase : public TargetTransformInfoImplCRTPBase<T> {
42 typedef TargetTransformInfoImplCRTPBase<T> BaseT;
43 typedef TargetTransformInfo TTI;
45 /// Estimate a cost of shuffle as a sequence of extract and insert
47 unsigned getPermuteShuffleOverhead(Type *Ty) {
48 assert(Ty->isVectorTy() && "Can only shuffle vectors");
50 // Shuffle cost is equal to the cost of extracting element from its argument
51 // plus the cost of inserting them onto the result vector.
53 // e.g. <4 x float> has a mask of <0,5,2,7> i.e we need to extract from
54 // index 0 of first vector, index 1 of second vector,index 2 of first
55 // vector and finally index 3 of second vector and insert them at index
56 // <0,1,2,3> of result vector.
57 for (int i = 0, e = Ty->getVectorNumElements(); i < e; ++i) {
58 Cost += static_cast<T *>(this)
59 ->getVectorInstrCost(Instruction::InsertElement, Ty, i);
60 Cost += static_cast<T *>(this)
61 ->getVectorInstrCost(Instruction::ExtractElement, Ty, i);
66 /// \brief Local query method delegates up to T which *must* implement this!
67 const TargetSubtargetInfo *getST() const {
68 return static_cast<const T *>(this)->getST();
71 /// \brief Local query method delegates up to T which *must* implement this!
72 const TargetLoweringBase *getTLI() const {
73 return static_cast<const T *>(this)->getTLI();
77 explicit BasicTTIImplBase(const TargetMachine *TM, const DataLayout &DL)
80 using TargetTransformInfoImplBase::DL;
83 /// \name Scalar TTI Implementations
85 bool allowsMisalignedMemoryAccesses(LLVMContext &Context,
86 unsigned BitWidth, unsigned AddressSpace,
87 unsigned Alignment, bool *Fast) const {
88 EVT E = EVT::getIntegerVT(Context, BitWidth);
89 return getTLI()->allowsMisalignedMemoryAccesses(E, AddressSpace, Alignment, Fast);
92 bool hasBranchDivergence() { return false; }
94 bool isSourceOfDivergence(const Value *V) { return false; }
96 unsigned getFlatAddressSpace() {
97 // Return an invalid address space.
101 bool isLegalAddImmediate(int64_t imm) {
102 return getTLI()->isLegalAddImmediate(imm);
105 bool isLegalICmpImmediate(int64_t imm) {
106 return getTLI()->isLegalICmpImmediate(imm);
109 bool isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset,
110 bool HasBaseReg, int64_t Scale,
111 unsigned AddrSpace) {
112 TargetLoweringBase::AddrMode AM;
114 AM.BaseOffs = BaseOffset;
115 AM.HasBaseReg = HasBaseReg;
117 return getTLI()->isLegalAddressingMode(DL, AM, Ty, AddrSpace);
120 bool isLSRCostLess(TTI::LSRCost C1, TTI::LSRCost C2) {
121 return TargetTransformInfoImplBase::isLSRCostLess(C1, C2);
124 int getScalingFactorCost(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset,
125 bool HasBaseReg, int64_t Scale, unsigned AddrSpace) {
126 TargetLoweringBase::AddrMode AM;
128 AM.BaseOffs = BaseOffset;
129 AM.HasBaseReg = HasBaseReg;
131 return getTLI()->getScalingFactorCost(DL, AM, Ty, AddrSpace);
134 bool isFoldableMemAccessOffset(Instruction *I, int64_t Offset) {
135 return getTLI()->isFoldableMemAccessOffset(I, Offset);
138 bool isTruncateFree(Type *Ty1, Type *Ty2) {
139 return getTLI()->isTruncateFree(Ty1, Ty2);
142 bool isProfitableToHoist(Instruction *I) {
143 return getTLI()->isProfitableToHoist(I);
146 bool isTypeLegal(Type *Ty) {
147 EVT VT = getTLI()->getValueType(DL, Ty);
148 return getTLI()->isTypeLegal(VT);
151 int getGEPCost(Type *PointeeType, const Value *Ptr,
152 ArrayRef<const Value *> Operands) {
153 return BaseT::getGEPCost(PointeeType, Ptr, Operands);
156 unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
157 ArrayRef<const Value *> Arguments) {
158 return BaseT::getIntrinsicCost(IID, RetTy, Arguments);
161 unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
162 ArrayRef<Type *> ParamTys) {
163 if (IID == Intrinsic::cttz) {
164 if (getTLI()->isCheapToSpeculateCttz())
165 return TargetTransformInfo::TCC_Basic;
166 return TargetTransformInfo::TCC_Expensive;
169 if (IID == Intrinsic::ctlz) {
170 if (getTLI()->isCheapToSpeculateCtlz())
171 return TargetTransformInfo::TCC_Basic;
172 return TargetTransformInfo::TCC_Expensive;
175 return BaseT::getIntrinsicCost(IID, RetTy, ParamTys);
178 unsigned getEstimatedNumberOfCaseClusters(const SwitchInst &SI,
179 unsigned &JumpTableSize) {
180 /// Try to find the estimated number of clusters. Note that the number of
181 /// clusters identified in this function could be different from the actural
182 /// numbers found in lowering. This function ignore switches that are
183 /// lowered with a mix of jump table / bit test / BTree. This function was
184 /// initially intended to be used when estimating the cost of switch in
185 /// inline cost heuristic, but it's a generic cost model to be used in other
186 /// places (e.g., in loop unrolling).
187 unsigned N = SI.getNumCases();
188 const TargetLoweringBase *TLI = getTLI();
189 const DataLayout &DL = this->getDataLayout();
192 bool IsJTAllowed = TLI->areJTsAllowed(SI.getParent()->getParent());
194 // Early exit if both a jump table and bit test are not allowed.
195 if (N < 1 || (!IsJTAllowed && DL.getPointerSizeInBits() < N))
198 APInt MaxCaseVal = SI.case_begin()->getCaseValue()->getValue();
199 APInt MinCaseVal = MaxCaseVal;
200 for (auto CI : SI.cases()) {
201 const APInt &CaseVal = CI.getCaseValue()->getValue();
202 if (CaseVal.sgt(MaxCaseVal))
203 MaxCaseVal = CaseVal;
204 if (CaseVal.slt(MinCaseVal))
205 MinCaseVal = CaseVal;
208 // Check if suitable for a bit test
209 if (N <= DL.getPointerSizeInBits()) {
210 SmallPtrSet<const BasicBlock *, 4> Dests;
211 for (auto I : SI.cases())
212 Dests.insert(I.getCaseSuccessor());
214 if (TLI->isSuitableForBitTests(Dests.size(), N, MinCaseVal, MaxCaseVal,
219 // Check if suitable for a jump table.
221 if (N < 2 || N < TLI->getMinimumJumpTableEntries())
224 (MaxCaseVal - MinCaseVal).getLimitedValue(UINT64_MAX - 1) + 1;
225 // Check whether a range of clusters is dense enough for a jump table
226 if (TLI->isSuitableForJumpTable(&SI, N, Range)) {
227 JumpTableSize = Range;
234 unsigned getJumpBufAlignment() { return getTLI()->getJumpBufAlignment(); }
236 unsigned getJumpBufSize() { return getTLI()->getJumpBufSize(); }
238 bool shouldBuildLookupTables() {
239 const TargetLoweringBase *TLI = getTLI();
240 return TLI->isOperationLegalOrCustom(ISD::BR_JT, MVT::Other) ||
241 TLI->isOperationLegalOrCustom(ISD::BRIND, MVT::Other);
244 bool haveFastSqrt(Type *Ty) {
245 const TargetLoweringBase *TLI = getTLI();
246 EVT VT = TLI->getValueType(DL, Ty);
247 return TLI->isTypeLegal(VT) &&
248 TLI->isOperationLegalOrCustom(ISD::FSQRT, VT);
251 unsigned getFPOpCost(Type *Ty) {
252 // By default, FP instructions are no more expensive since they are
253 // implemented in HW. Target specific TTI can override this.
254 return TargetTransformInfo::TCC_Basic;
257 unsigned getOperationCost(unsigned Opcode, Type *Ty, Type *OpTy) {
258 const TargetLoweringBase *TLI = getTLI();
261 case Instruction::Trunc: {
262 if (TLI->isTruncateFree(OpTy, Ty))
263 return TargetTransformInfo::TCC_Free;
264 return TargetTransformInfo::TCC_Basic;
266 case Instruction::ZExt: {
267 if (TLI->isZExtFree(OpTy, Ty))
268 return TargetTransformInfo::TCC_Free;
269 return TargetTransformInfo::TCC_Basic;
273 return BaseT::getOperationCost(Opcode, Ty, OpTy);
276 unsigned getInliningThresholdMultiplier() { return 1; }
278 void getUnrollingPreferences(Loop *L, TTI::UnrollingPreferences &UP) {
279 // This unrolling functionality is target independent, but to provide some
280 // motivation for its intended use, for x86:
282 // According to the Intel 64 and IA-32 Architectures Optimization Reference
283 // Manual, Intel Core models and later have a loop stream detector (and
284 // associated uop queue) that can benefit from partial unrolling.
285 // The relevant requirements are:
286 // - The loop must have no more than 4 (8 for Nehalem and later) branches
287 // taken, and none of them may be calls.
288 // - The loop can have no more than 18 (28 for Nehalem and later) uops.
290 // According to the Software Optimization Guide for AMD Family 15h
291 // Processors, models 30h-4fh (Steamroller and later) have a loop predictor
292 // and loop buffer which can benefit from partial unrolling.
293 // The relevant requirements are:
294 // - The loop must have fewer than 16 branches
295 // - The loop must have less than 40 uops in all executed loop branches
297 // The number of taken branches in a loop is hard to estimate here, and
298 // benchmarking has revealed that it is better not to be conservative when
299 // estimating the branch count. As a result, we'll ignore the branch limits
300 // until someone finds a case where it matters in practice.
303 const TargetSubtargetInfo *ST = getST();
304 if (PartialUnrollingThreshold.getNumOccurrences() > 0)
305 MaxOps = PartialUnrollingThreshold;
306 else if (ST->getSchedModel().LoopMicroOpBufferSize > 0)
307 MaxOps = ST->getSchedModel().LoopMicroOpBufferSize;
311 // Scan the loop: don't unroll loops with calls.
312 for (Loop::block_iterator I = L->block_begin(), E = L->block_end(); I != E;
316 for (BasicBlock::iterator J = BB->begin(), JE = BB->end(); J != JE; ++J)
317 if (isa<CallInst>(J) || isa<InvokeInst>(J)) {
318 ImmutableCallSite CS(&*J);
319 if (const Function *F = CS.getCalledFunction()) {
320 if (!static_cast<T *>(this)->isLoweredToCall(F))
328 // Enable runtime and partial unrolling up to the specified size.
329 // Enable using trip count upper bound to unroll loops.
330 UP.Partial = UP.Runtime = UP.UpperBound = true;
331 UP.PartialThreshold = MaxOps;
333 // Avoid unrolling when optimizing for size.
334 UP.OptSizeThreshold = 0;
335 UP.PartialOptSizeThreshold = 0;
337 // Set number of instructions optimized when "back edge"
338 // becomes "fall through" to default value of 2.
344 /// \name Vector TTI Implementations
347 unsigned getNumberOfRegisters(bool Vector) { return Vector ? 0 : 1; }
349 unsigned getRegisterBitWidth(bool Vector) { return 32; }
351 /// Estimate the overhead of scalarizing an instruction. Insert and Extract
352 /// are set if the result needs to be inserted and/or extracted from vectors.
353 unsigned getScalarizationOverhead(Type *Ty, bool Insert, bool Extract) {
354 assert(Ty->isVectorTy() && "Can only scalarize vectors");
357 for (int i = 0, e = Ty->getVectorNumElements(); i < e; ++i) {
359 Cost += static_cast<T *>(this)
360 ->getVectorInstrCost(Instruction::InsertElement, Ty, i);
362 Cost += static_cast<T *>(this)
363 ->getVectorInstrCost(Instruction::ExtractElement, Ty, i);
369 /// Estimate the overhead of scalarizing an instructions unique
370 /// non-constant operands. The types of the arguments are ordinarily
371 /// scalar, in which case the costs are multiplied with VF.
372 unsigned getOperandsScalarizationOverhead(ArrayRef<const Value *> Args,
375 SmallPtrSet<const Value*, 4> UniqueOperands;
376 for (const Value *A : Args) {
377 if (!isa<Constant>(A) && UniqueOperands.insert(A).second) {
378 Type *VecTy = nullptr;
379 if (A->getType()->isVectorTy()) {
380 VecTy = A->getType();
381 // If A is a vector operand, VF should be 1 or correspond to A.
382 assert ((VF == 1 || VF == VecTy->getVectorNumElements()) &&
383 "Vector argument does not match VF");
386 VecTy = VectorType::get(A->getType(), VF);
388 Cost += getScalarizationOverhead(VecTy, false, true);
395 unsigned getScalarizationOverhead(Type *VecTy, ArrayRef<const Value *> Args) {
396 assert (VecTy->isVectorTy());
400 Cost += getScalarizationOverhead(VecTy, true, false);
402 Cost += getOperandsScalarizationOverhead(Args,
403 VecTy->getVectorNumElements());
405 // When no information on arguments is provided, we add the cost
406 // associated with one argument as a heuristic.
407 Cost += getScalarizationOverhead(VecTy, false, true);
412 unsigned getMaxInterleaveFactor(unsigned VF) { return 1; }
414 unsigned getArithmeticInstrCost(
415 unsigned Opcode, Type *Ty,
416 TTI::OperandValueKind Opd1Info = TTI::OK_AnyValue,
417 TTI::OperandValueKind Opd2Info = TTI::OK_AnyValue,
418 TTI::OperandValueProperties Opd1PropInfo = TTI::OP_None,
419 TTI::OperandValueProperties Opd2PropInfo = TTI::OP_None,
420 ArrayRef<const Value *> Args = ArrayRef<const Value *>()) {
421 // Check if any of the operands are vector operands.
422 const TargetLoweringBase *TLI = getTLI();
423 int ISD = TLI->InstructionOpcodeToISD(Opcode);
424 assert(ISD && "Invalid opcode");
426 std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(DL, Ty);
428 bool IsFloat = Ty->getScalarType()->isFloatingPointTy();
429 // Assume that floating point arithmetic operations cost twice as much as
430 // integer operations.
431 unsigned OpCost = (IsFloat ? 2 : 1);
433 if (TLI->isOperationLegalOrPromote(ISD, LT.second)) {
434 // The operation is legal. Assume it costs 1.
435 // TODO: Once we have extract/insert subvector cost we need to use them.
436 return LT.first * OpCost;
439 if (!TLI->isOperationExpand(ISD, LT.second)) {
440 // If the operation is custom lowered, then assume that the code is twice
442 return LT.first * 2 * OpCost;
445 // Else, assume that we need to scalarize this op.
446 // TODO: If one of the types get legalized by splitting, handle this
447 // similarly to what getCastInstrCost() does.
448 if (Ty->isVectorTy()) {
449 unsigned Num = Ty->getVectorNumElements();
450 unsigned Cost = static_cast<T *>(this)
451 ->getArithmeticInstrCost(Opcode, Ty->getScalarType());
452 // Return the cost of multiple scalar invocation plus the cost of
453 // inserting and extracting the values.
454 return getScalarizationOverhead(Ty, Args) + Num * Cost;
457 // We don't know anything about this scalar instruction.
461 unsigned getShuffleCost(TTI::ShuffleKind Kind, Type *Tp, int Index,
463 if (Kind == TTI::SK_Alternate || Kind == TTI::SK_PermuteTwoSrc ||
464 Kind == TTI::SK_PermuteSingleSrc) {
465 return getPermuteShuffleOverhead(Tp);
470 unsigned getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src,
471 const Instruction *I = nullptr) {
472 const TargetLoweringBase *TLI = getTLI();
473 int ISD = TLI->InstructionOpcodeToISD(Opcode);
474 assert(ISD && "Invalid opcode");
475 std::pair<unsigned, MVT> SrcLT = TLI->getTypeLegalizationCost(DL, Src);
476 std::pair<unsigned, MVT> DstLT = TLI->getTypeLegalizationCost(DL, Dst);
478 // Check for NOOP conversions.
479 if (SrcLT.first == DstLT.first &&
480 SrcLT.second.getSizeInBits() == DstLT.second.getSizeInBits()) {
482 // Bitcast between types that are legalized to the same type are free.
483 if (Opcode == Instruction::BitCast || Opcode == Instruction::Trunc)
487 if (Opcode == Instruction::Trunc &&
488 TLI->isTruncateFree(SrcLT.second, DstLT.second))
491 if (Opcode == Instruction::ZExt &&
492 TLI->isZExtFree(SrcLT.second, DstLT.second))
495 if (Opcode == Instruction::AddrSpaceCast &&
496 TLI->isNoopAddrSpaceCast(Src->getPointerAddressSpace(),
497 Dst->getPointerAddressSpace()))
500 // If this is a zext/sext of a load, return 0 if the corresponding
501 // extending load exists on target.
502 if ((Opcode == Instruction::ZExt || Opcode == Instruction::SExt) &&
503 I && isa<LoadInst>(I->getOperand(0))) {
504 EVT ExtVT = EVT::getEVT(Dst);
505 EVT LoadVT = EVT::getEVT(Src);
507 ((Opcode == Instruction::ZExt) ? ISD::ZEXTLOAD : ISD::SEXTLOAD);
508 if (TLI->isLoadExtLegal(LType, ExtVT, LoadVT))
512 // If the cast is marked as legal (or promote) then assume low cost.
513 if (SrcLT.first == DstLT.first &&
514 TLI->isOperationLegalOrPromote(ISD, DstLT.second))
517 // Handle scalar conversions.
518 if (!Src->isVectorTy() && !Dst->isVectorTy()) {
520 // Scalar bitcasts are usually free.
521 if (Opcode == Instruction::BitCast)
524 // Just check the op cost. If the operation is legal then assume it costs
526 if (!TLI->isOperationExpand(ISD, DstLT.second))
529 // Assume that illegal scalar instruction are expensive.
533 // Check vector-to-vector casts.
534 if (Dst->isVectorTy() && Src->isVectorTy()) {
536 // If the cast is between same-sized registers, then the check is simple.
537 if (SrcLT.first == DstLT.first &&
538 SrcLT.second.getSizeInBits() == DstLT.second.getSizeInBits()) {
540 // Assume that Zext is done using AND.
541 if (Opcode == Instruction::ZExt)
544 // Assume that sext is done using SHL and SRA.
545 if (Opcode == Instruction::SExt)
548 // Just check the op cost. If the operation is legal then assume it
550 // 1 and multiply by the type-legalization overhead.
551 if (!TLI->isOperationExpand(ISD, DstLT.second))
552 return SrcLT.first * 1;
555 // If we are legalizing by splitting, query the concrete TTI for the cost
556 // of casting the original vector twice. We also need to factor int the
557 // cost of the split itself. Count that as 1, to be consistent with
558 // TLI->getTypeLegalizationCost().
559 if ((TLI->getTypeAction(Src->getContext(), TLI->getValueType(DL, Src)) ==
560 TargetLowering::TypeSplitVector) ||
561 (TLI->getTypeAction(Dst->getContext(), TLI->getValueType(DL, Dst)) ==
562 TargetLowering::TypeSplitVector)) {
563 Type *SplitDst = VectorType::get(Dst->getVectorElementType(),
564 Dst->getVectorNumElements() / 2);
565 Type *SplitSrc = VectorType::get(Src->getVectorElementType(),
566 Src->getVectorNumElements() / 2);
567 T *TTI = static_cast<T *>(this);
568 return TTI->getVectorSplitCost() +
569 (2 * TTI->getCastInstrCost(Opcode, SplitDst, SplitSrc, I));
572 // In other cases where the source or destination are illegal, assume
573 // the operation will get scalarized.
574 unsigned Num = Dst->getVectorNumElements();
575 unsigned Cost = static_cast<T *>(this)->getCastInstrCost(
576 Opcode, Dst->getScalarType(), Src->getScalarType(), I);
578 // Return the cost of multiple scalar invocation plus the cost of
579 // inserting and extracting the values.
580 return getScalarizationOverhead(Dst, true, true) + Num * Cost;
583 // We already handled vector-to-vector and scalar-to-scalar conversions.
585 // is where we handle bitcast between vectors and scalars. We need to assume
586 // that the conversion is scalarized in one way or another.
587 if (Opcode == Instruction::BitCast)
588 // Illegal bitcasts are done by storing and loading from a stack slot.
589 return (Src->isVectorTy() ? getScalarizationOverhead(Src, false, true)
591 (Dst->isVectorTy() ? getScalarizationOverhead(Dst, true, false)
594 llvm_unreachable("Unhandled cast");
597 unsigned getExtractWithExtendCost(unsigned Opcode, Type *Dst,
598 VectorType *VecTy, unsigned Index) {
599 return static_cast<T *>(this)->getVectorInstrCost(
600 Instruction::ExtractElement, VecTy, Index) +
601 static_cast<T *>(this)->getCastInstrCost(Opcode, Dst,
602 VecTy->getElementType());
605 unsigned getCFInstrCost(unsigned Opcode) {
606 // Branches are assumed to be predicted.
610 unsigned getCmpSelInstrCost(unsigned Opcode, Type *ValTy, Type *CondTy,
611 const Instruction *I) {
612 const TargetLoweringBase *TLI = getTLI();
613 int ISD = TLI->InstructionOpcodeToISD(Opcode);
614 assert(ISD && "Invalid opcode");
616 // Selects on vectors are actually vector selects.
617 if (ISD == ISD::SELECT) {
618 assert(CondTy && "CondTy must exist");
619 if (CondTy->isVectorTy())
622 std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(DL, ValTy);
624 if (!(ValTy->isVectorTy() && !LT.second.isVector()) &&
625 !TLI->isOperationExpand(ISD, LT.second)) {
626 // The operation is legal. Assume it costs 1. Multiply
627 // by the type-legalization overhead.
631 // Otherwise, assume that the cast is scalarized.
632 // TODO: If one of the types get legalized by splitting, handle this
633 // similarly to what getCastInstrCost() does.
634 if (ValTy->isVectorTy()) {
635 unsigned Num = ValTy->getVectorNumElements();
637 CondTy = CondTy->getScalarType();
638 unsigned Cost = static_cast<T *>(this)->getCmpSelInstrCost(
639 Opcode, ValTy->getScalarType(), CondTy, I);
641 // Return the cost of multiple scalar invocation plus the cost of
642 // inserting and extracting the values.
643 return getScalarizationOverhead(ValTy, true, false) + Num * Cost;
646 // Unknown scalar opcode.
650 unsigned getVectorInstrCost(unsigned Opcode, Type *Val, unsigned Index) {
651 std::pair<unsigned, MVT> LT =
652 getTLI()->getTypeLegalizationCost(DL, Val->getScalarType());
657 unsigned getMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
658 unsigned AddressSpace, const Instruction *I = nullptr) {
659 assert(!Src->isVoidTy() && "Invalid type");
660 std::pair<unsigned, MVT> LT = getTLI()->getTypeLegalizationCost(DL, Src);
662 // Assuming that all loads of legal types cost 1.
663 unsigned Cost = LT.first;
665 if (Src->isVectorTy() &&
666 Src->getPrimitiveSizeInBits() < LT.second.getSizeInBits()) {
667 // This is a vector load that legalizes to a larger type than the vector
668 // itself. Unless the corresponding extending load or truncating store is
669 // legal, then this will scalarize.
670 TargetLowering::LegalizeAction LA = TargetLowering::Expand;
671 EVT MemVT = getTLI()->getValueType(DL, Src);
672 if (Opcode == Instruction::Store)
673 LA = getTLI()->getTruncStoreAction(LT.second, MemVT);
675 LA = getTLI()->getLoadExtAction(ISD::EXTLOAD, LT.second, MemVT);
677 if (LA != TargetLowering::Legal && LA != TargetLowering::Custom) {
678 // This is a vector load/store for some illegal type that is scalarized.
679 // We must account for the cost of building or decomposing the vector.
680 Cost += getScalarizationOverhead(Src, Opcode != Instruction::Store,
681 Opcode == Instruction::Store);
688 unsigned getInterleavedMemoryOpCost(unsigned Opcode, Type *VecTy,
690 ArrayRef<unsigned> Indices,
692 unsigned AddressSpace) {
693 VectorType *VT = dyn_cast<VectorType>(VecTy);
694 assert(VT && "Expect a vector type for interleaved memory op");
696 unsigned NumElts = VT->getNumElements();
697 assert(Factor > 1 && NumElts % Factor == 0 && "Invalid interleave factor");
699 unsigned NumSubElts = NumElts / Factor;
700 VectorType *SubVT = VectorType::get(VT->getElementType(), NumSubElts);
702 // Firstly, the cost of load/store operation.
703 unsigned Cost = static_cast<T *>(this)->getMemoryOpCost(
704 Opcode, VecTy, Alignment, AddressSpace);
706 // Legalize the vector type, and get the legalized and unlegalized type
708 MVT VecTyLT = getTLI()->getTypeLegalizationCost(DL, VecTy).second;
710 static_cast<T *>(this)->getDataLayout().getTypeStoreSize(VecTy);
711 unsigned VecTyLTSize = VecTyLT.getStoreSize();
713 // Return the ceiling of dividing A by B.
714 auto ceil = [](unsigned A, unsigned B) { return (A + B - 1) / B; };
716 // Scale the cost of the memory operation by the fraction of legalized
717 // instructions that will actually be used. We shouldn't account for the
718 // cost of dead instructions since they will be removed.
720 // E.g., An interleaved load of factor 8:
721 // %vec = load <16 x i64>, <16 x i64>* %ptr
722 // %v0 = shufflevector %vec, undef, <0, 8>
724 // If <16 x i64> is legalized to 8 v2i64 loads, only 2 of the loads will be
725 // used (those corresponding to elements [0:1] and [8:9] of the unlegalized
726 // type). The other loads are unused.
728 // We only scale the cost of loads since interleaved store groups aren't
729 // allowed to have gaps.
730 if (Opcode == Instruction::Load && VecTySize > VecTyLTSize) {
732 // The number of loads of a legal type it will take to represent a load
733 // of the unlegalized vector type.
734 unsigned NumLegalInsts = ceil(VecTySize, VecTyLTSize);
736 // The number of elements of the unlegalized type that correspond to a
737 // single legal instruction.
738 unsigned NumEltsPerLegalInst = ceil(NumElts, NumLegalInsts);
740 // Determine which legal instructions will be used.
741 BitVector UsedInsts(NumLegalInsts, false);
742 for (unsigned Index : Indices)
743 for (unsigned Elt = 0; Elt < NumSubElts; ++Elt)
744 UsedInsts.set((Index + Elt * Factor) / NumEltsPerLegalInst);
746 // Scale the cost of the load by the fraction of legal instructions that
748 Cost *= UsedInsts.count() / NumLegalInsts;
751 // Then plus the cost of interleave operation.
752 if (Opcode == Instruction::Load) {
753 // The interleave cost is similar to extract sub vectors' elements
754 // from the wide vector, and insert them into sub vectors.
756 // E.g. An interleaved load of factor 2 (with one member of index 0):
757 // %vec = load <8 x i32>, <8 x i32>* %ptr
758 // %v0 = shuffle %vec, undef, <0, 2, 4, 6> ; Index 0
759 // The cost is estimated as extract elements at 0, 2, 4, 6 from the
760 // <8 x i32> vector and insert them into a <4 x i32> vector.
762 assert(Indices.size() <= Factor &&
763 "Interleaved memory op has too many members");
765 for (unsigned Index : Indices) {
766 assert(Index < Factor && "Invalid index for interleaved memory op");
768 // Extract elements from loaded vector for each sub vector.
769 for (unsigned i = 0; i < NumSubElts; i++)
770 Cost += static_cast<T *>(this)->getVectorInstrCost(
771 Instruction::ExtractElement, VT, Index + i * Factor);
774 unsigned InsSubCost = 0;
775 for (unsigned i = 0; i < NumSubElts; i++)
776 InsSubCost += static_cast<T *>(this)->getVectorInstrCost(
777 Instruction::InsertElement, SubVT, i);
779 Cost += Indices.size() * InsSubCost;
781 // The interleave cost is extract all elements from sub vectors, and
782 // insert them into the wide vector.
784 // E.g. An interleaved store of factor 2:
785 // %v0_v1 = shuffle %v0, %v1, <0, 4, 1, 5, 2, 6, 3, 7>
786 // store <8 x i32> %interleaved.vec, <8 x i32>* %ptr
787 // The cost is estimated as extract all elements from both <4 x i32>
788 // vectors and insert into the <8 x i32> vector.
790 unsigned ExtSubCost = 0;
791 for (unsigned i = 0; i < NumSubElts; i++)
792 ExtSubCost += static_cast<T *>(this)->getVectorInstrCost(
793 Instruction::ExtractElement, SubVT, i);
794 Cost += ExtSubCost * Factor;
796 for (unsigned i = 0; i < NumElts; i++)
797 Cost += static_cast<T *>(this)
798 ->getVectorInstrCost(Instruction::InsertElement, VT, i);
804 /// Get intrinsic cost based on arguments.
805 unsigned getIntrinsicInstrCost(Intrinsic::ID IID, Type *RetTy,
806 ArrayRef<Value *> Args, FastMathFlags FMF,
808 unsigned RetVF = (RetTy->isVectorTy() ? RetTy->getVectorNumElements() : 1);
809 assert ((RetVF == 1 || VF == 1) && "VF > 1 and RetVF is a vector type");
813 // Assume that we need to scalarize this intrinsic.
814 SmallVector<Type *, 4> Types;
815 for (Value *Op : Args) {
816 Type *OpTy = Op->getType();
817 assert (VF == 1 || !OpTy->isVectorTy());
818 Types.push_back(VF == 1 ? OpTy : VectorType::get(OpTy, VF));
821 if (VF > 1 && !RetTy->isVoidTy())
822 RetTy = VectorType::get(RetTy, VF);
824 // Compute the scalarization overhead based on Args for a vector
825 // intrinsic. A vectorizer will pass a scalar RetTy and VF > 1, while
826 // CostModel will pass a vector RetTy and VF is 1.
827 unsigned ScalarizationCost = UINT_MAX;
828 if (RetVF > 1 || VF > 1) {
829 ScalarizationCost = 0;
830 if (!RetTy->isVoidTy())
831 ScalarizationCost += getScalarizationOverhead(RetTy, true, false);
832 ScalarizationCost += getOperandsScalarizationOverhead(Args, VF);
835 return static_cast<T *>(this)->
836 getIntrinsicInstrCost(IID, RetTy, Types, FMF, ScalarizationCost);
838 case Intrinsic::masked_scatter: {
839 assert (VF == 1 && "Can't vectorize types here.");
840 Value *Mask = Args[3];
841 bool VarMask = !isa<Constant>(Mask);
842 unsigned Alignment = cast<ConstantInt>(Args[2])->getZExtValue();
844 static_cast<T *>(this)->getGatherScatterOpCost(Instruction::Store,
849 case Intrinsic::masked_gather: {
850 assert (VF == 1 && "Can't vectorize types here.");
851 Value *Mask = Args[2];
852 bool VarMask = !isa<Constant>(Mask);
853 unsigned Alignment = cast<ConstantInt>(Args[1])->getZExtValue();
855 static_cast<T *>(this)->getGatherScatterOpCost(Instruction::Load,
856 RetTy, Args[0], VarMask,
862 /// Get intrinsic cost based on argument types.
863 /// If ScalarizationCostPassed is UINT_MAX, the cost of scalarizing the
864 /// arguments and the return value will be computed based on types.
865 unsigned getIntrinsicInstrCost(Intrinsic::ID IID, Type *RetTy,
866 ArrayRef<Type *> Tys, FastMathFlags FMF,
867 unsigned ScalarizationCostPassed = UINT_MAX) {
868 SmallVector<unsigned, 2> ISDs;
869 unsigned SingleCallCost = 10; // Library call cost. Make it expensive.
872 // Assume that we need to scalarize this intrinsic.
873 unsigned ScalarizationCost = ScalarizationCostPassed;
874 unsigned ScalarCalls = 1;
875 Type *ScalarRetTy = RetTy;
876 if (RetTy->isVectorTy()) {
877 if (ScalarizationCostPassed == UINT_MAX)
878 ScalarizationCost = getScalarizationOverhead(RetTy, true, false);
879 ScalarCalls = std::max(ScalarCalls, RetTy->getVectorNumElements());
880 ScalarRetTy = RetTy->getScalarType();
882 SmallVector<Type *, 4> ScalarTys;
883 for (unsigned i = 0, ie = Tys.size(); i != ie; ++i) {
885 if (Ty->isVectorTy()) {
886 if (ScalarizationCostPassed == UINT_MAX)
887 ScalarizationCost += getScalarizationOverhead(Ty, false, true);
888 ScalarCalls = std::max(ScalarCalls, Ty->getVectorNumElements());
889 Ty = Ty->getScalarType();
891 ScalarTys.push_back(Ty);
893 if (ScalarCalls == 1)
894 return 1; // Return cost of a scalar intrinsic. Assume it to be cheap.
896 unsigned ScalarCost = static_cast<T *>(this)->getIntrinsicInstrCost(
897 IID, ScalarRetTy, ScalarTys, FMF);
899 return ScalarCalls * ScalarCost + ScalarizationCost;
901 // Look for intrinsics that can be lowered directly or turned into a scalar
903 case Intrinsic::sqrt:
904 ISDs.push_back(ISD::FSQRT);
907 ISDs.push_back(ISD::FSIN);
910 ISDs.push_back(ISD::FCOS);
913 ISDs.push_back(ISD::FEXP);
915 case Intrinsic::exp2:
916 ISDs.push_back(ISD::FEXP2);
919 ISDs.push_back(ISD::FLOG);
921 case Intrinsic::log10:
922 ISDs.push_back(ISD::FLOG10);
924 case Intrinsic::log2:
925 ISDs.push_back(ISD::FLOG2);
927 case Intrinsic::fabs:
928 ISDs.push_back(ISD::FABS);
930 case Intrinsic::minnum:
931 ISDs.push_back(ISD::FMINNUM);
933 ISDs.push_back(ISD::FMINNAN);
935 case Intrinsic::maxnum:
936 ISDs.push_back(ISD::FMAXNUM);
938 ISDs.push_back(ISD::FMAXNAN);
940 case Intrinsic::copysign:
941 ISDs.push_back(ISD::FCOPYSIGN);
943 case Intrinsic::floor:
944 ISDs.push_back(ISD::FFLOOR);
946 case Intrinsic::ceil:
947 ISDs.push_back(ISD::FCEIL);
949 case Intrinsic::trunc:
950 ISDs.push_back(ISD::FTRUNC);
952 case Intrinsic::nearbyint:
953 ISDs.push_back(ISD::FNEARBYINT);
955 case Intrinsic::rint:
956 ISDs.push_back(ISD::FRINT);
958 case Intrinsic::round:
959 ISDs.push_back(ISD::FROUND);
962 ISDs.push_back(ISD::FPOW);
965 ISDs.push_back(ISD::FMA);
967 case Intrinsic::fmuladd:
968 ISDs.push_back(ISD::FMA);
970 // FIXME: We should return 0 whenever getIntrinsicCost == TCC_Free.
971 case Intrinsic::lifetime_start:
972 case Intrinsic::lifetime_end:
974 case Intrinsic::masked_store:
975 return static_cast<T *>(this)
976 ->getMaskedMemoryOpCost(Instruction::Store, Tys[0], 0, 0);
977 case Intrinsic::masked_load:
978 return static_cast<T *>(this)
979 ->getMaskedMemoryOpCost(Instruction::Load, RetTy, 0, 0);
980 case Intrinsic::ctpop:
981 ISDs.push_back(ISD::CTPOP);
982 // In case of legalization use TCC_Expensive. This is cheaper than a
983 // library call but still not a cheap instruction.
984 SingleCallCost = TargetTransformInfo::TCC_Expensive;
986 // FIXME: ctlz, cttz, ...
989 const TargetLoweringBase *TLI = getTLI();
990 std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(DL, RetTy);
992 SmallVector<unsigned, 2> LegalCost;
993 SmallVector<unsigned, 2> CustomCost;
994 for (unsigned ISD : ISDs) {
995 if (TLI->isOperationLegalOrPromote(ISD, LT.second)) {
996 if (IID == Intrinsic::fabs && TLI->isFAbsFree(LT.second)) {
1000 // The operation is legal. Assume it costs 1.
1001 // If the type is split to multiple registers, assume that there is some
1002 // overhead to this.
1003 // TODO: Once we have extract/insert subvector cost we need to use them.
1005 LegalCost.push_back(LT.first * 2);
1007 LegalCost.push_back(LT.first * 1);
1008 } else if (!TLI->isOperationExpand(ISD, LT.second)) {
1009 // If the operation is custom lowered then assume
1010 // that the code is twice as expensive.
1011 CustomCost.push_back(LT.first * 2);
1015 auto MinLegalCostI = std::min_element(LegalCost.begin(), LegalCost.end());
1016 if (MinLegalCostI != LegalCost.end())
1017 return *MinLegalCostI;
1019 auto MinCustomCostI = std::min_element(CustomCost.begin(), CustomCost.end());
1020 if (MinCustomCostI != CustomCost.end())
1021 return *MinCustomCostI;
1023 // If we can't lower fmuladd into an FMA estimate the cost as a floating
1024 // point mul followed by an add.
1025 if (IID == Intrinsic::fmuladd)
1026 return static_cast<T *>(this)
1027 ->getArithmeticInstrCost(BinaryOperator::FMul, RetTy) +
1028 static_cast<T *>(this)
1029 ->getArithmeticInstrCost(BinaryOperator::FAdd, RetTy);
1031 // Else, assume that we need to scalarize this intrinsic. For math builtins
1032 // this will emit a costly libcall, adding call overhead and spills. Make it
1034 if (RetTy->isVectorTy()) {
1035 unsigned ScalarizationCost = ((ScalarizationCostPassed != UINT_MAX) ?
1036 ScalarizationCostPassed : getScalarizationOverhead(RetTy, true, false));
1037 unsigned ScalarCalls = RetTy->getVectorNumElements();
1038 SmallVector<Type *, 4> ScalarTys;
1039 for (unsigned i = 0, ie = Tys.size(); i != ie; ++i) {
1041 if (Ty->isVectorTy())
1042 Ty = Ty->getScalarType();
1043 ScalarTys.push_back(Ty);
1045 unsigned ScalarCost = static_cast<T *>(this)->getIntrinsicInstrCost(
1046 IID, RetTy->getScalarType(), ScalarTys, FMF);
1047 for (unsigned i = 0, ie = Tys.size(); i != ie; ++i) {
1048 if (Tys[i]->isVectorTy()) {
1049 if (ScalarizationCostPassed == UINT_MAX)
1050 ScalarizationCost += getScalarizationOverhead(Tys[i], false, true);
1051 ScalarCalls = std::max(ScalarCalls, Tys[i]->getVectorNumElements());
1055 return ScalarCalls * ScalarCost + ScalarizationCost;
1058 // This is going to be turned into a library call, make it expensive.
1059 return SingleCallCost;
1062 /// \brief Compute a cost of the given call instruction.
1064 /// Compute the cost of calling function F with return type RetTy and
1065 /// argument types Tys. F might be nullptr, in this case the cost of an
1066 /// arbitrary call with the specified signature will be returned.
1067 /// This is used, for instance, when we estimate call of a vector
1068 /// counterpart of the given function.
1069 /// \param F Called function, might be nullptr.
1070 /// \param RetTy Return value types.
1071 /// \param Tys Argument types.
1072 /// \returns The cost of Call instruction.
1073 unsigned getCallInstrCost(Function *F, Type *RetTy, ArrayRef<Type *> Tys) {
1077 unsigned getNumberOfParts(Type *Tp) {
1078 std::pair<unsigned, MVT> LT = getTLI()->getTypeLegalizationCost(DL, Tp);
1082 unsigned getAddressComputationCost(Type *Ty, ScalarEvolution *,
1087 /// Try to calculate arithmetic and shuffle op costs for reduction operations.
1088 /// We're assuming that reduction operation are performing the following way:
1089 /// 1. Non-pairwise reduction
1090 /// %val1 = shufflevector<n x t> %val, <n x t> %undef,
1091 /// <n x i32> <i32 n/2, i32 n/2 + 1, ..., i32 n, i32 undef, ..., i32 undef>
1092 /// \----------------v-------------/ \----------v------------/
1093 /// n/2 elements n/2 elements
1094 /// %red1 = op <n x t> %val, <n x t> val1
1095 /// After this operation we have a vector %red1 where only the first n/2
1096 /// elements are meaningful, the second n/2 elements are undefined and can be
1097 /// dropped. All other operations are actually working with the vector of
1098 /// length n/2, not n, though the real vector length is still n.
1099 /// %val2 = shufflevector<n x t> %red1, <n x t> %undef,
1100 /// <n x i32> <i32 n/4, i32 n/4 + 1, ..., i32 n/2, i32 undef, ..., i32 undef>
1101 /// \----------------v-------------/ \----------v------------/
1102 /// n/4 elements 3*n/4 elements
1103 /// %red2 = op <n x t> %red1, <n x t> val2 - working with the vector of
1104 /// length n/2, the resulting vector has length n/4 etc.
1105 /// 2. Pairwise reduction:
1106 /// Everything is the same except for an additional shuffle operation which
1107 /// is used to produce operands for pairwise kind of reductions.
1108 /// %val1 = shufflevector<n x t> %val, <n x t> %undef,
1109 /// <n x i32> <i32 0, i32 2, ..., i32 n-2, i32 undef, ..., i32 undef>
1110 /// \-------------v----------/ \----------v------------/
1111 /// n/2 elements n/2 elements
1112 /// %val2 = shufflevector<n x t> %val, <n x t> %undef,
1113 /// <n x i32> <i32 1, i32 3, ..., i32 n-1, i32 undef, ..., i32 undef>
1114 /// \-------------v----------/ \----------v------------/
1115 /// n/2 elements n/2 elements
1116 /// %red1 = op <n x t> %val1, <n x t> val2
1117 /// Again, the operation is performed on <n x t> vector, but the resulting
1118 /// vector %red1 is <n/2 x t> vector.
1120 /// The cost model should take into account that the actual length of the
1121 /// vector is reduced on each iteration.
1122 unsigned getReductionCost(unsigned Opcode, Type *Ty, bool IsPairwise) {
1123 assert(Ty->isVectorTy() && "Expect a vector type");
1124 Type *ScalarTy = Ty->getVectorElementType();
1125 unsigned NumVecElts = Ty->getVectorNumElements();
1126 unsigned NumReduxLevels = Log2_32(NumVecElts);
1127 unsigned ArithCost = 0;
1128 unsigned ShuffleCost = 0;
1129 auto *ConcreteTTI = static_cast<T *>(this);
1130 std::pair<unsigned, MVT> LT =
1131 ConcreteTTI->getTLI()->getTypeLegalizationCost(DL, Ty);
1132 unsigned LongVectorCount = 0;
1134 LT.second.isVector() ? LT.second.getVectorNumElements() : 1;
1135 while (NumVecElts > MVTLen) {
1137 // Assume the pairwise shuffles add a cost.
1138 ShuffleCost += (IsPairwise + 1) *
1139 ConcreteTTI->getShuffleCost(TTI::SK_ExtractSubvector, Ty,
1141 ArithCost += ConcreteTTI->getArithmeticInstrCost(Opcode, Ty);
1142 Ty = VectorType::get(ScalarTy, NumVecElts);
1145 // The minimal length of the vector is limited by the real length of vector
1146 // operations performed on the current platform. That's why several final
1147 // reduction opertions are perfomed on the vectors with the same
1148 // architecture-dependent length.
1149 ShuffleCost += (NumReduxLevels - LongVectorCount) * (IsPairwise + 1) *
1150 ConcreteTTI->getShuffleCost(TTI::SK_ExtractSubvector, Ty,
1152 ArithCost += (NumReduxLevels - LongVectorCount) *
1153 ConcreteTTI->getArithmeticInstrCost(Opcode, Ty);
1154 return ShuffleCost + ArithCost + getScalarizationOverhead(Ty, false, true);
1157 unsigned getVectorSplitCost() { return 1; }
1162 /// \brief Concrete BasicTTIImpl that can be used if no further customization
1164 class BasicTTIImpl : public BasicTTIImplBase<BasicTTIImpl> {
1165 typedef BasicTTIImplBase<BasicTTIImpl> BaseT;
1166 friend class BasicTTIImplBase<BasicTTIImpl>;
1168 const TargetSubtargetInfo *ST;
1169 const TargetLoweringBase *TLI;
1171 const TargetSubtargetInfo *getST() const { return ST; }
1172 const TargetLoweringBase *getTLI() const { return TLI; }
1175 explicit BasicTTIImpl(const TargetMachine *ST, const Function &F);