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/TargetTransformInfoImpl.h"
21 #include "llvm/Support/CommandLine.h"
22 #include "llvm/Target/TargetLowering.h"
23 #include "llvm/Target/TargetSubtargetInfo.h"
24 #include "llvm/Analysis/TargetLibraryInfo.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 int getScalingFactorCost(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset,
121 bool HasBaseReg, int64_t Scale, unsigned AddrSpace) {
122 TargetLoweringBase::AddrMode AM;
124 AM.BaseOffs = BaseOffset;
125 AM.HasBaseReg = HasBaseReg;
127 return getTLI()->getScalingFactorCost(DL, AM, Ty, AddrSpace);
130 bool isFoldableMemAccessOffset(Instruction *I, int64_t Offset) {
131 return getTLI()->isFoldableMemAccessOffset(I, Offset);
134 bool isTruncateFree(Type *Ty1, Type *Ty2) {
135 return getTLI()->isTruncateFree(Ty1, Ty2);
138 bool isProfitableToHoist(Instruction *I) {
139 return getTLI()->isProfitableToHoist(I);
142 bool isTypeLegal(Type *Ty) {
143 EVT VT = getTLI()->getValueType(DL, Ty);
144 return getTLI()->isTypeLegal(VT);
147 int getGEPCost(Type *PointeeType, const Value *Ptr,
148 ArrayRef<const Value *> Operands) {
149 return BaseT::getGEPCost(PointeeType, Ptr, Operands);
152 unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
153 ArrayRef<const Value *> Arguments) {
154 return BaseT::getIntrinsicCost(IID, RetTy, Arguments);
157 unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
158 ArrayRef<Type *> ParamTys) {
159 if (IID == Intrinsic::cttz) {
160 if (getTLI()->isCheapToSpeculateCttz())
161 return TargetTransformInfo::TCC_Basic;
162 return TargetTransformInfo::TCC_Expensive;
165 if (IID == Intrinsic::ctlz) {
166 if (getTLI()->isCheapToSpeculateCtlz())
167 return TargetTransformInfo::TCC_Basic;
168 return TargetTransformInfo::TCC_Expensive;
171 return BaseT::getIntrinsicCost(IID, RetTy, ParamTys);
174 unsigned getEstimatedNumberOfCaseClusters(const SwitchInst &SI,
175 unsigned &JumpTableSize) {
176 /// Try to find the estimated number of clusters. Note that the number of
177 /// clusters identified in this function could be different from the actural
178 /// numbers found in lowering. This function ignore switches that are
179 /// lowered with a mix of jump table / bit test / BTree. This function was
180 /// initially intended to be used when estimating the cost of switch in
181 /// inline cost heuristic, but it's a generic cost model to be used in other
182 /// places (e.g., in loop unrolling).
183 unsigned N = SI.getNumCases();
184 const TargetLoweringBase *TLI = getTLI();
185 const DataLayout &DL = this->getDataLayout();
188 bool IsJTAllowed = TLI->areJTsAllowed(SI.getParent()->getParent());
190 // Early exit if both a jump table and bit test are not allowed.
191 if (N < 1 || (!IsJTAllowed && DL.getPointerSizeInBits() < N))
194 APInt MaxCaseVal = SI.case_begin()->getCaseValue()->getValue();
195 APInt MinCaseVal = MaxCaseVal;
196 for (auto CI : SI.cases()) {
197 const APInt &CaseVal = CI.getCaseValue()->getValue();
198 if (CaseVal.sgt(MaxCaseVal))
199 MaxCaseVal = CaseVal;
200 if (CaseVal.slt(MinCaseVal))
201 MinCaseVal = CaseVal;
204 // Check if suitable for a bit test
205 if (N <= DL.getPointerSizeInBits()) {
206 SmallPtrSet<const BasicBlock *, 4> Dests;
207 for (auto I : SI.cases())
208 Dests.insert(I.getCaseSuccessor());
210 if (TLI->isSuitableForBitTests(Dests.size(), N, MinCaseVal, MaxCaseVal,
215 // Check if suitable for a jump table.
217 if (N < 2 || N < TLI->getMinimumJumpTableEntries())
220 (MaxCaseVal - MinCaseVal).getLimitedValue(UINT64_MAX - 1) + 1;
221 // Check whether a range of clusters is dense enough for a jump table
222 if (TLI->isSuitableForJumpTable(&SI, N, Range)) {
223 JumpTableSize = Range;
230 unsigned getJumpBufAlignment() { return getTLI()->getJumpBufAlignment(); }
232 unsigned getJumpBufSize() { return getTLI()->getJumpBufSize(); }
234 bool shouldBuildLookupTables() {
235 const TargetLoweringBase *TLI = getTLI();
236 return TLI->isOperationLegalOrCustom(ISD::BR_JT, MVT::Other) ||
237 TLI->isOperationLegalOrCustom(ISD::BRIND, MVT::Other);
240 bool haveFastSqrt(Type *Ty) {
241 const TargetLoweringBase *TLI = getTLI();
242 EVT VT = TLI->getValueType(DL, Ty);
243 return TLI->isTypeLegal(VT) &&
244 TLI->isOperationLegalOrCustom(ISD::FSQRT, VT);
247 unsigned getFPOpCost(Type *Ty) {
248 // By default, FP instructions are no more expensive since they are
249 // implemented in HW. Target specific TTI can override this.
250 return TargetTransformInfo::TCC_Basic;
253 unsigned getOperationCost(unsigned Opcode, Type *Ty, Type *OpTy) {
254 const TargetLoweringBase *TLI = getTLI();
257 case Instruction::Trunc: {
258 if (TLI->isTruncateFree(OpTy, Ty))
259 return TargetTransformInfo::TCC_Free;
260 return TargetTransformInfo::TCC_Basic;
262 case Instruction::ZExt: {
263 if (TLI->isZExtFree(OpTy, Ty))
264 return TargetTransformInfo::TCC_Free;
265 return TargetTransformInfo::TCC_Basic;
269 return BaseT::getOperationCost(Opcode, Ty, OpTy);
272 unsigned getInliningThresholdMultiplier() { return 1; }
274 void getUnrollingPreferences(Loop *L, TTI::UnrollingPreferences &UP) {
275 // This unrolling functionality is target independent, but to provide some
276 // motivation for its intended use, for x86:
278 // According to the Intel 64 and IA-32 Architectures Optimization Reference
279 // Manual, Intel Core models and later have a loop stream detector (and
280 // associated uop queue) that can benefit from partial unrolling.
281 // The relevant requirements are:
282 // - The loop must have no more than 4 (8 for Nehalem and later) branches
283 // taken, and none of them may be calls.
284 // - The loop can have no more than 18 (28 for Nehalem and later) uops.
286 // According to the Software Optimization Guide for AMD Family 15h
287 // Processors, models 30h-4fh (Steamroller and later) have a loop predictor
288 // and loop buffer which can benefit from partial unrolling.
289 // The relevant requirements are:
290 // - The loop must have fewer than 16 branches
291 // - The loop must have less than 40 uops in all executed loop branches
293 // The number of taken branches in a loop is hard to estimate here, and
294 // benchmarking has revealed that it is better not to be conservative when
295 // estimating the branch count. As a result, we'll ignore the branch limits
296 // until someone finds a case where it matters in practice.
299 const TargetSubtargetInfo *ST = getST();
300 if (PartialUnrollingThreshold.getNumOccurrences() > 0)
301 MaxOps = PartialUnrollingThreshold;
302 else if (ST->getSchedModel().LoopMicroOpBufferSize > 0)
303 MaxOps = ST->getSchedModel().LoopMicroOpBufferSize;
307 // Scan the loop: don't unroll loops with calls.
308 for (Loop::block_iterator I = L->block_begin(), E = L->block_end(); I != E;
312 for (BasicBlock::iterator J = BB->begin(), JE = BB->end(); J != JE; ++J)
313 if (isa<CallInst>(J) || isa<InvokeInst>(J)) {
314 ImmutableCallSite CS(&*J);
315 if (const Function *F = CS.getCalledFunction()) {
316 if (!static_cast<T *>(this)->isLoweredToCall(F))
324 // Enable runtime and partial unrolling up to the specified size.
325 // Enable using trip count upper bound to unroll loops.
326 UP.Partial = UP.Runtime = UP.UpperBound = true;
327 UP.PartialThreshold = MaxOps;
329 // Avoid unrolling when optimizing for size.
330 UP.OptSizeThreshold = 0;
331 UP.PartialOptSizeThreshold = 0;
333 // Set number of instructions optimized when "back edge"
334 // becomes "fall through" to default value of 2.
340 /// \name Vector TTI Implementations
343 unsigned getNumberOfRegisters(bool Vector) { return Vector ? 0 : 1; }
345 unsigned getRegisterBitWidth(bool Vector) { return 32; }
347 /// Estimate the overhead of scalarizing an instruction. Insert and Extract
348 /// are set if the result needs to be inserted and/or extracted from vectors.
349 unsigned getScalarizationOverhead(Type *Ty, bool Insert, bool Extract) {
350 assert(Ty->isVectorTy() && "Can only scalarize vectors");
353 for (int i = 0, e = Ty->getVectorNumElements(); i < e; ++i) {
355 Cost += static_cast<T *>(this)
356 ->getVectorInstrCost(Instruction::InsertElement, Ty, i);
358 Cost += static_cast<T *>(this)
359 ->getVectorInstrCost(Instruction::ExtractElement, Ty, i);
365 /// Estimate the overhead of scalarizing an instructions unique
366 /// non-constant operands. The types of the arguments are ordinarily
367 /// scalar, in which case the costs are multiplied with VF.
368 unsigned getOperandsScalarizationOverhead(ArrayRef<const Value *> Args,
371 SmallPtrSet<const Value*, 4> UniqueOperands;
372 for (const Value *A : Args) {
373 if (!isa<Constant>(A) && UniqueOperands.insert(A).second) {
374 Type *VecTy = nullptr;
375 if (A->getType()->isVectorTy()) {
376 VecTy = A->getType();
377 // If A is a vector operand, VF should be 1 or correspond to A.
378 assert ((VF == 1 || VF == VecTy->getVectorNumElements()) &&
379 "Vector argument does not match VF");
382 VecTy = VectorType::get(A->getType(), VF);
384 Cost += getScalarizationOverhead(VecTy, false, true);
391 unsigned getScalarizationOverhead(Type *VecTy, ArrayRef<const Value *> Args) {
392 assert (VecTy->isVectorTy());
396 Cost += getScalarizationOverhead(VecTy, true, false);
398 Cost += getOperandsScalarizationOverhead(Args,
399 VecTy->getVectorNumElements());
401 // When no information on arguments is provided, we add the cost
402 // associated with one argument as a heuristic.
403 Cost += getScalarizationOverhead(VecTy, false, true);
408 unsigned getMaxInterleaveFactor(unsigned VF) { return 1; }
410 unsigned getArithmeticInstrCost(
411 unsigned Opcode, Type *Ty,
412 TTI::OperandValueKind Opd1Info = TTI::OK_AnyValue,
413 TTI::OperandValueKind Opd2Info = TTI::OK_AnyValue,
414 TTI::OperandValueProperties Opd1PropInfo = TTI::OP_None,
415 TTI::OperandValueProperties Opd2PropInfo = TTI::OP_None,
416 ArrayRef<const Value *> Args = ArrayRef<const Value *>()) {
417 // Check if any of the operands are vector operands.
418 const TargetLoweringBase *TLI = getTLI();
419 int ISD = TLI->InstructionOpcodeToISD(Opcode);
420 assert(ISD && "Invalid opcode");
422 std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(DL, Ty);
424 bool IsFloat = Ty->getScalarType()->isFloatingPointTy();
425 // Assume that floating point arithmetic operations cost twice as much as
426 // integer operations.
427 unsigned OpCost = (IsFloat ? 2 : 1);
429 if (TLI->isOperationLegalOrPromote(ISD, LT.second)) {
430 // The operation is legal. Assume it costs 1.
431 // TODO: Once we have extract/insert subvector cost we need to use them.
432 return LT.first * OpCost;
435 if (!TLI->isOperationExpand(ISD, LT.second)) {
436 // If the operation is custom lowered, then assume that the code is twice
438 return LT.first * 2 * OpCost;
441 // Else, assume that we need to scalarize this op.
442 // TODO: If one of the types get legalized by splitting, handle this
443 // similarly to what getCastInstrCost() does.
444 if (Ty->isVectorTy()) {
445 unsigned Num = Ty->getVectorNumElements();
446 unsigned Cost = static_cast<T *>(this)
447 ->getArithmeticInstrCost(Opcode, Ty->getScalarType());
448 // Return the cost of multiple scalar invocation plus the cost of
449 // inserting and extracting the values.
450 return getScalarizationOverhead(Ty, Args) + Num * Cost;
453 // We don't know anything about this scalar instruction.
457 unsigned getShuffleCost(TTI::ShuffleKind Kind, Type *Tp, int Index,
459 if (Kind == TTI::SK_Alternate || Kind == TTI::SK_PermuteTwoSrc ||
460 Kind == TTI::SK_PermuteSingleSrc) {
461 return getPermuteShuffleOverhead(Tp);
466 unsigned getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src,
467 const Instruction *I = nullptr) {
468 const TargetLoweringBase *TLI = getTLI();
469 int ISD = TLI->InstructionOpcodeToISD(Opcode);
470 assert(ISD && "Invalid opcode");
471 std::pair<unsigned, MVT> SrcLT = TLI->getTypeLegalizationCost(DL, Src);
472 std::pair<unsigned, MVT> DstLT = TLI->getTypeLegalizationCost(DL, Dst);
474 // Check for NOOP conversions.
475 if (SrcLT.first == DstLT.first &&
476 SrcLT.second.getSizeInBits() == DstLT.second.getSizeInBits()) {
478 // Bitcast between types that are legalized to the same type are free.
479 if (Opcode == Instruction::BitCast || Opcode == Instruction::Trunc)
483 if (Opcode == Instruction::Trunc &&
484 TLI->isTruncateFree(SrcLT.second, DstLT.second))
487 if (Opcode == Instruction::ZExt &&
488 TLI->isZExtFree(SrcLT.second, DstLT.second))
491 if (Opcode == Instruction::AddrSpaceCast &&
492 TLI->isNoopAddrSpaceCast(Src->getPointerAddressSpace(),
493 Dst->getPointerAddressSpace()))
496 // If this is a zext/sext of a load, return 0 if the corresponding
497 // extending load exists on target.
498 if ((Opcode == Instruction::ZExt || Opcode == Instruction::SExt) &&
499 I && isa<LoadInst>(I->getOperand(0))) {
500 EVT ExtVT = EVT::getEVT(Dst);
501 EVT LoadVT = EVT::getEVT(Src);
503 ((Opcode == Instruction::ZExt) ? ISD::ZEXTLOAD : ISD::SEXTLOAD);
504 if (TLI->isLoadExtLegal(LType, ExtVT, LoadVT))
508 // If the cast is marked as legal (or promote) then assume low cost.
509 if (SrcLT.first == DstLT.first &&
510 TLI->isOperationLegalOrPromote(ISD, DstLT.second))
513 // Handle scalar conversions.
514 if (!Src->isVectorTy() && !Dst->isVectorTy()) {
516 // Scalar bitcasts are usually free.
517 if (Opcode == Instruction::BitCast)
520 // Just check the op cost. If the operation is legal then assume it costs
522 if (!TLI->isOperationExpand(ISD, DstLT.second))
525 // Assume that illegal scalar instruction are expensive.
529 // Check vector-to-vector casts.
530 if (Dst->isVectorTy() && Src->isVectorTy()) {
532 // If the cast is between same-sized registers, then the check is simple.
533 if (SrcLT.first == DstLT.first &&
534 SrcLT.second.getSizeInBits() == DstLT.second.getSizeInBits()) {
536 // Assume that Zext is done using AND.
537 if (Opcode == Instruction::ZExt)
540 // Assume that sext is done using SHL and SRA.
541 if (Opcode == Instruction::SExt)
544 // Just check the op cost. If the operation is legal then assume it
546 // 1 and multiply by the type-legalization overhead.
547 if (!TLI->isOperationExpand(ISD, DstLT.second))
548 return SrcLT.first * 1;
551 // If we are legalizing by splitting, query the concrete TTI for the cost
552 // of casting the original vector twice. We also need to factor int the
553 // cost of the split itself. Count that as 1, to be consistent with
554 // TLI->getTypeLegalizationCost().
555 if ((TLI->getTypeAction(Src->getContext(), TLI->getValueType(DL, Src)) ==
556 TargetLowering::TypeSplitVector) ||
557 (TLI->getTypeAction(Dst->getContext(), TLI->getValueType(DL, Dst)) ==
558 TargetLowering::TypeSplitVector)) {
559 Type *SplitDst = VectorType::get(Dst->getVectorElementType(),
560 Dst->getVectorNumElements() / 2);
561 Type *SplitSrc = VectorType::get(Src->getVectorElementType(),
562 Src->getVectorNumElements() / 2);
563 T *TTI = static_cast<T *>(this);
564 return TTI->getVectorSplitCost() +
565 (2 * TTI->getCastInstrCost(Opcode, SplitDst, SplitSrc, I));
568 // In other cases where the source or destination are illegal, assume
569 // the operation will get scalarized.
570 unsigned Num = Dst->getVectorNumElements();
571 unsigned Cost = static_cast<T *>(this)->getCastInstrCost(
572 Opcode, Dst->getScalarType(), Src->getScalarType(), I);
574 // Return the cost of multiple scalar invocation plus the cost of
575 // inserting and extracting the values.
576 return getScalarizationOverhead(Dst, true, true) + Num * Cost;
579 // We already handled vector-to-vector and scalar-to-scalar conversions.
581 // is where we handle bitcast between vectors and scalars. We need to assume
582 // that the conversion is scalarized in one way or another.
583 if (Opcode == Instruction::BitCast)
584 // Illegal bitcasts are done by storing and loading from a stack slot.
585 return (Src->isVectorTy() ? getScalarizationOverhead(Src, false, true)
587 (Dst->isVectorTy() ? getScalarizationOverhead(Dst, true, false)
590 llvm_unreachable("Unhandled cast");
593 unsigned getExtractWithExtendCost(unsigned Opcode, Type *Dst,
594 VectorType *VecTy, unsigned Index) {
595 return static_cast<T *>(this)->getVectorInstrCost(
596 Instruction::ExtractElement, VecTy, Index) +
597 static_cast<T *>(this)->getCastInstrCost(Opcode, Dst,
598 VecTy->getElementType());
601 unsigned getCFInstrCost(unsigned Opcode) {
602 // Branches are assumed to be predicted.
606 unsigned getCmpSelInstrCost(unsigned Opcode, Type *ValTy, Type *CondTy,
607 const Instruction *I) {
608 const TargetLoweringBase *TLI = getTLI();
609 int ISD = TLI->InstructionOpcodeToISD(Opcode);
610 assert(ISD && "Invalid opcode");
612 // Selects on vectors are actually vector selects.
613 if (ISD == ISD::SELECT) {
614 assert(CondTy && "CondTy must exist");
615 if (CondTy->isVectorTy())
618 std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(DL, ValTy);
620 if (!(ValTy->isVectorTy() && !LT.second.isVector()) &&
621 !TLI->isOperationExpand(ISD, LT.second)) {
622 // The operation is legal. Assume it costs 1. Multiply
623 // by the type-legalization overhead.
627 // Otherwise, assume that the cast is scalarized.
628 // TODO: If one of the types get legalized by splitting, handle this
629 // similarly to what getCastInstrCost() does.
630 if (ValTy->isVectorTy()) {
631 unsigned Num = ValTy->getVectorNumElements();
633 CondTy = CondTy->getScalarType();
634 unsigned Cost = static_cast<T *>(this)->getCmpSelInstrCost(
635 Opcode, ValTy->getScalarType(), CondTy, I);
637 // Return the cost of multiple scalar invocation plus the cost of
638 // inserting and extracting the values.
639 return getScalarizationOverhead(ValTy, true, false) + Num * Cost;
642 // Unknown scalar opcode.
646 unsigned getVectorInstrCost(unsigned Opcode, Type *Val, unsigned Index) {
647 std::pair<unsigned, MVT> LT =
648 getTLI()->getTypeLegalizationCost(DL, Val->getScalarType());
653 unsigned getMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
654 unsigned AddressSpace, const Instruction *I = nullptr) {
655 assert(!Src->isVoidTy() && "Invalid type");
656 std::pair<unsigned, MVT> LT = getTLI()->getTypeLegalizationCost(DL, Src);
658 // Assuming that all loads of legal types cost 1.
659 unsigned Cost = LT.first;
661 if (Src->isVectorTy() &&
662 Src->getPrimitiveSizeInBits() < LT.second.getSizeInBits()) {
663 // This is a vector load that legalizes to a larger type than the vector
664 // itself. Unless the corresponding extending load or truncating store is
665 // legal, then this will scalarize.
666 TargetLowering::LegalizeAction LA = TargetLowering::Expand;
667 EVT MemVT = getTLI()->getValueType(DL, Src);
668 if (Opcode == Instruction::Store)
669 LA = getTLI()->getTruncStoreAction(LT.second, MemVT);
671 LA = getTLI()->getLoadExtAction(ISD::EXTLOAD, LT.second, MemVT);
673 if (LA != TargetLowering::Legal && LA != TargetLowering::Custom) {
674 // This is a vector load/store for some illegal type that is scalarized.
675 // We must account for the cost of building or decomposing the vector.
676 Cost += getScalarizationOverhead(Src, Opcode != Instruction::Store,
677 Opcode == Instruction::Store);
684 unsigned getInterleavedMemoryOpCost(unsigned Opcode, Type *VecTy,
686 ArrayRef<unsigned> Indices,
688 unsigned AddressSpace) {
689 VectorType *VT = dyn_cast<VectorType>(VecTy);
690 assert(VT && "Expect a vector type for interleaved memory op");
692 unsigned NumElts = VT->getNumElements();
693 assert(Factor > 1 && NumElts % Factor == 0 && "Invalid interleave factor");
695 unsigned NumSubElts = NumElts / Factor;
696 VectorType *SubVT = VectorType::get(VT->getElementType(), NumSubElts);
698 // Firstly, the cost of load/store operation.
699 unsigned Cost = static_cast<T *>(this)->getMemoryOpCost(
700 Opcode, VecTy, Alignment, AddressSpace);
702 // Legalize the vector type, and get the legalized and unlegalized type
704 MVT VecTyLT = getTLI()->getTypeLegalizationCost(DL, VecTy).second;
706 static_cast<T *>(this)->getDataLayout().getTypeStoreSize(VecTy);
707 unsigned VecTyLTSize = VecTyLT.getStoreSize();
709 // Return the ceiling of dividing A by B.
710 auto ceil = [](unsigned A, unsigned B) { return (A + B - 1) / B; };
712 // Scale the cost of the memory operation by the fraction of legalized
713 // instructions that will actually be used. We shouldn't account for the
714 // cost of dead instructions since they will be removed.
716 // E.g., An interleaved load of factor 8:
717 // %vec = load <16 x i64>, <16 x i64>* %ptr
718 // %v0 = shufflevector %vec, undef, <0, 8>
720 // If <16 x i64> is legalized to 8 v2i64 loads, only 2 of the loads will be
721 // used (those corresponding to elements [0:1] and [8:9] of the unlegalized
722 // type). The other loads are unused.
724 // We only scale the cost of loads since interleaved store groups aren't
725 // allowed to have gaps.
726 if (Opcode == Instruction::Load && VecTySize > VecTyLTSize) {
728 // The number of loads of a legal type it will take to represent a load
729 // of the unlegalized vector type.
730 unsigned NumLegalInsts = ceil(VecTySize, VecTyLTSize);
732 // The number of elements of the unlegalized type that correspond to a
733 // single legal instruction.
734 unsigned NumEltsPerLegalInst = ceil(NumElts, NumLegalInsts);
736 // Determine which legal instructions will be used.
737 BitVector UsedInsts(NumLegalInsts, false);
738 for (unsigned Index : Indices)
739 for (unsigned Elt = 0; Elt < NumSubElts; ++Elt)
740 UsedInsts.set((Index + Elt * Factor) / NumEltsPerLegalInst);
742 // Scale the cost of the load by the fraction of legal instructions that
744 Cost *= UsedInsts.count() / NumLegalInsts;
747 // Then plus the cost of interleave operation.
748 if (Opcode == Instruction::Load) {
749 // The interleave cost is similar to extract sub vectors' elements
750 // from the wide vector, and insert them into sub vectors.
752 // E.g. An interleaved load of factor 2 (with one member of index 0):
753 // %vec = load <8 x i32>, <8 x i32>* %ptr
754 // %v0 = shuffle %vec, undef, <0, 2, 4, 6> ; Index 0
755 // The cost is estimated as extract elements at 0, 2, 4, 6 from the
756 // <8 x i32> vector and insert them into a <4 x i32> vector.
758 assert(Indices.size() <= Factor &&
759 "Interleaved memory op has too many members");
761 for (unsigned Index : Indices) {
762 assert(Index < Factor && "Invalid index for interleaved memory op");
764 // Extract elements from loaded vector for each sub vector.
765 for (unsigned i = 0; i < NumSubElts; i++)
766 Cost += static_cast<T *>(this)->getVectorInstrCost(
767 Instruction::ExtractElement, VT, Index + i * Factor);
770 unsigned InsSubCost = 0;
771 for (unsigned i = 0; i < NumSubElts; i++)
772 InsSubCost += static_cast<T *>(this)->getVectorInstrCost(
773 Instruction::InsertElement, SubVT, i);
775 Cost += Indices.size() * InsSubCost;
777 // The interleave cost is extract all elements from sub vectors, and
778 // insert them into the wide vector.
780 // E.g. An interleaved store of factor 2:
781 // %v0_v1 = shuffle %v0, %v1, <0, 4, 1, 5, 2, 6, 3, 7>
782 // store <8 x i32> %interleaved.vec, <8 x i32>* %ptr
783 // The cost is estimated as extract all elements from both <4 x i32>
784 // vectors and insert into the <8 x i32> vector.
786 unsigned ExtSubCost = 0;
787 for (unsigned i = 0; i < NumSubElts; i++)
788 ExtSubCost += static_cast<T *>(this)->getVectorInstrCost(
789 Instruction::ExtractElement, SubVT, i);
790 Cost += ExtSubCost * Factor;
792 for (unsigned i = 0; i < NumElts; i++)
793 Cost += static_cast<T *>(this)
794 ->getVectorInstrCost(Instruction::InsertElement, VT, i);
800 /// Get intrinsic cost based on arguments.
801 unsigned getIntrinsicInstrCost(Intrinsic::ID IID, Type *RetTy,
802 ArrayRef<Value *> Args, FastMathFlags FMF,
804 unsigned RetVF = (RetTy->isVectorTy() ? RetTy->getVectorNumElements() : 1);
805 assert ((RetVF == 1 || VF == 1) && "VF > 1 and RetVF is a vector type");
809 // Assume that we need to scalarize this intrinsic.
810 SmallVector<Type *, 4> Types;
811 for (Value *Op : Args) {
812 Type *OpTy = Op->getType();
813 assert (VF == 1 || !OpTy->isVectorTy());
814 Types.push_back(VF == 1 ? OpTy : VectorType::get(OpTy, VF));
817 if (VF > 1 && !RetTy->isVoidTy())
818 RetTy = VectorType::get(RetTy, VF);
820 // Compute the scalarization overhead based on Args for a vector
821 // intrinsic. A vectorizer will pass a scalar RetTy and VF > 1, while
822 // CostModel will pass a vector RetTy and VF is 1.
823 unsigned ScalarizationCost = UINT_MAX;
824 if (RetVF > 1 || VF > 1) {
825 ScalarizationCost = 0;
826 if (!RetTy->isVoidTy())
827 ScalarizationCost += getScalarizationOverhead(RetTy, true, false);
828 ScalarizationCost += getOperandsScalarizationOverhead(Args, VF);
831 return static_cast<T *>(this)->
832 getIntrinsicInstrCost(IID, RetTy, Types, FMF, ScalarizationCost);
834 case Intrinsic::masked_scatter: {
835 assert (VF == 1 && "Can't vectorize types here.");
836 Value *Mask = Args[3];
837 bool VarMask = !isa<Constant>(Mask);
838 unsigned Alignment = cast<ConstantInt>(Args[2])->getZExtValue();
840 static_cast<T *>(this)->getGatherScatterOpCost(Instruction::Store,
845 case Intrinsic::masked_gather: {
846 assert (VF == 1 && "Can't vectorize types here.");
847 Value *Mask = Args[2];
848 bool VarMask = !isa<Constant>(Mask);
849 unsigned Alignment = cast<ConstantInt>(Args[1])->getZExtValue();
851 static_cast<T *>(this)->getGatherScatterOpCost(Instruction::Load,
852 RetTy, Args[0], VarMask,
858 /// Get intrinsic cost based on argument types.
859 /// If ScalarizationCostPassed is UINT_MAX, the cost of scalarizing the
860 /// arguments and the return value will be computed based on types.
861 unsigned getIntrinsicInstrCost(Intrinsic::ID IID, Type *RetTy,
862 ArrayRef<Type *> Tys, FastMathFlags FMF,
863 unsigned ScalarizationCostPassed = UINT_MAX) {
864 SmallVector<unsigned, 2> ISDs;
865 unsigned SingleCallCost = 10; // Library call cost. Make it expensive.
868 // Assume that we need to scalarize this intrinsic.
869 unsigned ScalarizationCost = ScalarizationCostPassed;
870 unsigned ScalarCalls = 1;
871 Type *ScalarRetTy = RetTy;
872 if (RetTy->isVectorTy()) {
873 if (ScalarizationCostPassed == UINT_MAX)
874 ScalarizationCost = getScalarizationOverhead(RetTy, true, false);
875 ScalarCalls = std::max(ScalarCalls, RetTy->getVectorNumElements());
876 ScalarRetTy = RetTy->getScalarType();
878 SmallVector<Type *, 4> ScalarTys;
879 for (unsigned i = 0, ie = Tys.size(); i != ie; ++i) {
881 if (Ty->isVectorTy()) {
882 if (ScalarizationCostPassed == UINT_MAX)
883 ScalarizationCost += getScalarizationOverhead(Ty, false, true);
884 ScalarCalls = std::max(ScalarCalls, Ty->getVectorNumElements());
885 Ty = Ty->getScalarType();
887 ScalarTys.push_back(Ty);
889 if (ScalarCalls == 1)
890 return 1; // Return cost of a scalar intrinsic. Assume it to be cheap.
892 unsigned ScalarCost = static_cast<T *>(this)->getIntrinsicInstrCost(
893 IID, ScalarRetTy, ScalarTys, FMF);
895 return ScalarCalls * ScalarCost + ScalarizationCost;
897 // Look for intrinsics that can be lowered directly or turned into a scalar
899 case Intrinsic::sqrt:
900 ISDs.push_back(ISD::FSQRT);
903 ISDs.push_back(ISD::FSIN);
906 ISDs.push_back(ISD::FCOS);
909 ISDs.push_back(ISD::FEXP);
911 case Intrinsic::exp2:
912 ISDs.push_back(ISD::FEXP2);
915 ISDs.push_back(ISD::FLOG);
917 case Intrinsic::log10:
918 ISDs.push_back(ISD::FLOG10);
920 case Intrinsic::log2:
921 ISDs.push_back(ISD::FLOG2);
923 case Intrinsic::fabs:
924 ISDs.push_back(ISD::FABS);
926 case Intrinsic::minnum:
927 ISDs.push_back(ISD::FMINNUM);
929 ISDs.push_back(ISD::FMINNAN);
931 case Intrinsic::maxnum:
932 ISDs.push_back(ISD::FMAXNUM);
934 ISDs.push_back(ISD::FMAXNAN);
936 case Intrinsic::copysign:
937 ISDs.push_back(ISD::FCOPYSIGN);
939 case Intrinsic::floor:
940 ISDs.push_back(ISD::FFLOOR);
942 case Intrinsic::ceil:
943 ISDs.push_back(ISD::FCEIL);
945 case Intrinsic::trunc:
946 ISDs.push_back(ISD::FTRUNC);
948 case Intrinsic::nearbyint:
949 ISDs.push_back(ISD::FNEARBYINT);
951 case Intrinsic::rint:
952 ISDs.push_back(ISD::FRINT);
954 case Intrinsic::round:
955 ISDs.push_back(ISD::FROUND);
958 ISDs.push_back(ISD::FPOW);
961 ISDs.push_back(ISD::FMA);
963 case Intrinsic::fmuladd:
964 ISDs.push_back(ISD::FMA);
966 // FIXME: We should return 0 whenever getIntrinsicCost == TCC_Free.
967 case Intrinsic::lifetime_start:
968 case Intrinsic::lifetime_end:
970 case Intrinsic::masked_store:
971 return static_cast<T *>(this)
972 ->getMaskedMemoryOpCost(Instruction::Store, Tys[0], 0, 0);
973 case Intrinsic::masked_load:
974 return static_cast<T *>(this)
975 ->getMaskedMemoryOpCost(Instruction::Load, RetTy, 0, 0);
976 case Intrinsic::ctpop:
977 ISDs.push_back(ISD::CTPOP);
978 // In case of legalization use TCC_Expensive. This is cheaper than a
979 // library call but still not a cheap instruction.
980 SingleCallCost = TargetTransformInfo::TCC_Expensive;
982 // FIXME: ctlz, cttz, ...
985 const TargetLoweringBase *TLI = getTLI();
986 std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(DL, RetTy);
988 SmallVector<unsigned, 2> LegalCost;
989 SmallVector<unsigned, 2> CustomCost;
990 for (unsigned ISD : ISDs) {
991 if (TLI->isOperationLegalOrPromote(ISD, LT.second)) {
992 if (IID == Intrinsic::fabs && TLI->isFAbsFree(LT.second)) {
996 // The operation is legal. Assume it costs 1.
997 // If the type is split to multiple registers, assume that there is some
999 // TODO: Once we have extract/insert subvector cost we need to use them.
1001 LegalCost.push_back(LT.first * 2);
1003 LegalCost.push_back(LT.first * 1);
1004 } else if (!TLI->isOperationExpand(ISD, LT.second)) {
1005 // If the operation is custom lowered then assume
1006 // that the code is twice as expensive.
1007 CustomCost.push_back(LT.first * 2);
1011 auto MinLegalCostI = std::min_element(LegalCost.begin(), LegalCost.end());
1012 if (MinLegalCostI != LegalCost.end())
1013 return *MinLegalCostI;
1015 auto MinCustomCostI = std::min_element(CustomCost.begin(), CustomCost.end());
1016 if (MinCustomCostI != CustomCost.end())
1017 return *MinCustomCostI;
1019 // If we can't lower fmuladd into an FMA estimate the cost as a floating
1020 // point mul followed by an add.
1021 if (IID == Intrinsic::fmuladd)
1022 return static_cast<T *>(this)
1023 ->getArithmeticInstrCost(BinaryOperator::FMul, RetTy) +
1024 static_cast<T *>(this)
1025 ->getArithmeticInstrCost(BinaryOperator::FAdd, RetTy);
1027 // Else, assume that we need to scalarize this intrinsic. For math builtins
1028 // this will emit a costly libcall, adding call overhead and spills. Make it
1030 if (RetTy->isVectorTy()) {
1031 unsigned ScalarizationCost = ((ScalarizationCostPassed != UINT_MAX) ?
1032 ScalarizationCostPassed : getScalarizationOverhead(RetTy, true, false));
1033 unsigned ScalarCalls = RetTy->getVectorNumElements();
1034 SmallVector<Type *, 4> ScalarTys;
1035 for (unsigned i = 0, ie = Tys.size(); i != ie; ++i) {
1037 if (Ty->isVectorTy())
1038 Ty = Ty->getScalarType();
1039 ScalarTys.push_back(Ty);
1041 unsigned ScalarCost = static_cast<T *>(this)->getIntrinsicInstrCost(
1042 IID, RetTy->getScalarType(), ScalarTys, FMF);
1043 for (unsigned i = 0, ie = Tys.size(); i != ie; ++i) {
1044 if (Tys[i]->isVectorTy()) {
1045 if (ScalarizationCostPassed == UINT_MAX)
1046 ScalarizationCost += getScalarizationOverhead(Tys[i], false, true);
1047 ScalarCalls = std::max(ScalarCalls, Tys[i]->getVectorNumElements());
1051 return ScalarCalls * ScalarCost + ScalarizationCost;
1054 // This is going to be turned into a library call, make it expensive.
1055 return SingleCallCost;
1058 /// \brief Compute a cost of the given call instruction.
1060 /// Compute the cost of calling function F with return type RetTy and
1061 /// argument types Tys. F might be nullptr, in this case the cost of an
1062 /// arbitrary call with the specified signature will be returned.
1063 /// This is used, for instance, when we estimate call of a vector
1064 /// counterpart of the given function.
1065 /// \param F Called function, might be nullptr.
1066 /// \param RetTy Return value types.
1067 /// \param Tys Argument types.
1068 /// \returns The cost of Call instruction.
1069 unsigned getCallInstrCost(Function *F, Type *RetTy, ArrayRef<Type *> Tys) {
1073 unsigned getNumberOfParts(Type *Tp) {
1074 std::pair<unsigned, MVT> LT = getTLI()->getTypeLegalizationCost(DL, Tp);
1078 unsigned getAddressComputationCost(Type *Ty, ScalarEvolution *,
1083 unsigned getReductionCost(unsigned Opcode, Type *Ty, bool IsPairwise) {
1084 assert(Ty->isVectorTy() && "Expect a vector type");
1085 Type *ScalarTy = Ty->getVectorElementType();
1086 unsigned NumVecElts = Ty->getVectorNumElements();
1087 unsigned NumReduxLevels = Log2_32(NumVecElts);
1088 // Try to calculate arithmetic and shuffle op costs for reduction operations.
1089 // We're assuming that reduction operation are performing the following way:
1090 // 1. Non-pairwise reduction
1091 // %val1 = shufflevector<n x t> %val, <n x t> %undef,
1092 // <n x i32> <i32 n/2, i32 n/2 + 1, ..., i32 n, i32 undef, ..., i32 undef>
1093 // \----------------v-------------/ \----------v------------/
1094 // n/2 elements n/2 elements
1095 // %red1 = op <n x t> %val, <n x t> val1
1096 // After this operation we have a vector %red1 with only maningfull the
1097 // first n/2 elements, the second n/2 elements are undefined and can be
1098 // dropped. All other operations are actually working with the vector of
1099 // length n/2, not n. though the real vector length is still n.
1100 // %val2 = shufflevector<n x t> %red1, <n x t> %undef,
1101 // <n x i32> <i32 n/4, i32 n/4 + 1, ..., i32 n/2, i32 undef, ..., i32 undef>
1102 // \----------------v-------------/ \----------v------------/
1103 // n/4 elements 3*n/4 elements
1104 // %red2 = op <n x t> %red1, <n x t> val2 - working with the vector of
1105 // length n/2, the resulting vector has length n/4 etc.
1106 // 2. Pairwise reduction:
1107 // Everything is the same except for an additional shuffle operation which
1108 // is used to produce operands for pairwise kind of reductions.
1109 // %val1 = shufflevector<n x t> %val, <n x t> %undef,
1110 // <n x i32> <i32 0, i32 2, ..., i32 n-2, i32 undef, ..., i32 undef>
1111 // \-------------v----------/ \----------v------------/
1112 // n/2 elements n/2 elements
1113 // %val2 = shufflevector<n x t> %val, <n x t> %undef,
1114 // <n x i32> <i32 1, i32 3, ..., i32 n-1, i32 undef, ..., i32 undef>
1115 // \-------------v----------/ \----------v------------/
1116 // n/2 elements n/2 elements
1117 // %red1 = op <n x t> %val1, <n x t> val2
1118 // Again, the operation is performed on <n x t> vector, but the resulting
1119 // vector %red1 is <n/2 x t> vector.
1121 // The cost model should take into account that the actual length of the
1122 // vector is reduced on each iteration.
1123 unsigned ArithCost = 0;
1124 unsigned ShuffleCost = 0;
1125 auto *ConcreteTTI = static_cast<T *>(this);
1126 std::pair<unsigned, MVT> LT =
1127 ConcreteTTI->getTLI()->getTypeLegalizationCost(DL, Ty);
1128 unsigned LongVectorCount = 0;
1130 LT.second.isVector() ? LT.second.getVectorNumElements() : 1;
1131 while (NumVecElts > MVTLen) {
1133 // Assume the pairwise shuffles add a cost.
1134 ShuffleCost += (IsPairwise + 1) *
1135 ConcreteTTI->getShuffleCost(TTI::SK_ExtractSubvector, Ty,
1137 ArithCost += ConcreteTTI->getArithmeticInstrCost(Opcode, Ty);
1138 Ty = VectorType::get(ScalarTy, NumVecElts);
1141 // The minimal length of the vector is limited by the real length of vector
1142 // operations performed on the current platform. That's why several final
1143 // reduction opertions are perfomed on the vectors with the same
1144 // architecture-dependent length.
1145 ShuffleCost += (NumReduxLevels - LongVectorCount) * (IsPairwise + 1) *
1146 ConcreteTTI->getShuffleCost(TTI::SK_ExtractSubvector, Ty,
1148 ArithCost += (NumReduxLevels - LongVectorCount) *
1149 ConcreteTTI->getArithmeticInstrCost(Opcode, Ty);
1150 return ShuffleCost + ArithCost + getScalarizationOverhead(Ty, false, true);
1153 unsigned getVectorSplitCost() { return 1; }
1158 /// \brief Concrete BasicTTIImpl that can be used if no further customization
1160 class BasicTTIImpl : public BasicTTIImplBase<BasicTTIImpl> {
1161 typedef BasicTTIImplBase<BasicTTIImpl> BaseT;
1162 friend class BasicTTIImplBase<BasicTTIImpl>;
1164 const TargetSubtargetInfo *ST;
1165 const TargetLoweringBase *TLI;
1167 const TargetSubtargetInfo *getST() const { return ST; }
1168 const TargetLoweringBase *getTLI() const { return TLI; }
1171 explicit BasicTTIImpl(const TargetMachine *ST, const Function &F);