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 bool isAlwaysUniform(const Value *V) { return false; }
98 unsigned getFlatAddressSpace() {
99 // Return an invalid address space.
103 bool isLegalAddImmediate(int64_t imm) {
104 return getTLI()->isLegalAddImmediate(imm);
107 bool isLegalICmpImmediate(int64_t imm) {
108 return getTLI()->isLegalICmpImmediate(imm);
111 bool isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset,
112 bool HasBaseReg, int64_t Scale,
113 unsigned AddrSpace) {
114 TargetLoweringBase::AddrMode AM;
116 AM.BaseOffs = BaseOffset;
117 AM.HasBaseReg = HasBaseReg;
119 return getTLI()->isLegalAddressingMode(DL, AM, Ty, AddrSpace);
122 bool isLSRCostLess(TTI::LSRCost C1, TTI::LSRCost C2) {
123 return TargetTransformInfoImplBase::isLSRCostLess(C1, C2);
126 int getScalingFactorCost(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset,
127 bool HasBaseReg, int64_t Scale, unsigned AddrSpace) {
128 TargetLoweringBase::AddrMode AM;
130 AM.BaseOffs = BaseOffset;
131 AM.HasBaseReg = HasBaseReg;
133 return getTLI()->getScalingFactorCost(DL, AM, Ty, AddrSpace);
136 bool isFoldableMemAccessOffset(Instruction *I, int64_t Offset) {
137 return getTLI()->isFoldableMemAccessOffset(I, Offset);
140 bool isTruncateFree(Type *Ty1, Type *Ty2) {
141 return getTLI()->isTruncateFree(Ty1, Ty2);
144 bool isProfitableToHoist(Instruction *I) {
145 return getTLI()->isProfitableToHoist(I);
148 bool isTypeLegal(Type *Ty) {
149 EVT VT = getTLI()->getValueType(DL, Ty);
150 return getTLI()->isTypeLegal(VT);
153 int getGEPCost(Type *PointeeType, const Value *Ptr,
154 ArrayRef<const Value *> Operands) {
155 return BaseT::getGEPCost(PointeeType, Ptr, Operands);
158 unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
159 ArrayRef<const Value *> Arguments) {
160 return BaseT::getIntrinsicCost(IID, RetTy, Arguments);
163 unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
164 ArrayRef<Type *> ParamTys) {
165 if (IID == Intrinsic::cttz) {
166 if (getTLI()->isCheapToSpeculateCttz())
167 return TargetTransformInfo::TCC_Basic;
168 return TargetTransformInfo::TCC_Expensive;
171 if (IID == Intrinsic::ctlz) {
172 if (getTLI()->isCheapToSpeculateCtlz())
173 return TargetTransformInfo::TCC_Basic;
174 return TargetTransformInfo::TCC_Expensive;
177 return BaseT::getIntrinsicCost(IID, RetTy, ParamTys);
180 unsigned getEstimatedNumberOfCaseClusters(const SwitchInst &SI,
181 unsigned &JumpTableSize) {
182 /// Try to find the estimated number of clusters. Note that the number of
183 /// clusters identified in this function could be different from the actural
184 /// numbers found in lowering. This function ignore switches that are
185 /// lowered with a mix of jump table / bit test / BTree. This function was
186 /// initially intended to be used when estimating the cost of switch in
187 /// inline cost heuristic, but it's a generic cost model to be used in other
188 /// places (e.g., in loop unrolling).
189 unsigned N = SI.getNumCases();
190 const TargetLoweringBase *TLI = getTLI();
191 const DataLayout &DL = this->getDataLayout();
194 bool IsJTAllowed = TLI->areJTsAllowed(SI.getParent()->getParent());
196 // Early exit if both a jump table and bit test are not allowed.
197 if (N < 1 || (!IsJTAllowed && DL.getPointerSizeInBits() < N))
200 APInt MaxCaseVal = SI.case_begin()->getCaseValue()->getValue();
201 APInt MinCaseVal = MaxCaseVal;
202 for (auto CI : SI.cases()) {
203 const APInt &CaseVal = CI.getCaseValue()->getValue();
204 if (CaseVal.sgt(MaxCaseVal))
205 MaxCaseVal = CaseVal;
206 if (CaseVal.slt(MinCaseVal))
207 MinCaseVal = CaseVal;
210 // Check if suitable for a bit test
211 if (N <= DL.getPointerSizeInBits()) {
212 SmallPtrSet<const BasicBlock *, 4> Dests;
213 for (auto I : SI.cases())
214 Dests.insert(I.getCaseSuccessor());
216 if (TLI->isSuitableForBitTests(Dests.size(), N, MinCaseVal, MaxCaseVal,
221 // Check if suitable for a jump table.
223 if (N < 2 || N < TLI->getMinimumJumpTableEntries())
226 (MaxCaseVal - MinCaseVal).getLimitedValue(UINT64_MAX - 1) + 1;
227 // Check whether a range of clusters is dense enough for a jump table
228 if (TLI->isSuitableForJumpTable(&SI, N, Range)) {
229 JumpTableSize = Range;
236 unsigned getJumpBufAlignment() { return getTLI()->getJumpBufAlignment(); }
238 unsigned getJumpBufSize() { return getTLI()->getJumpBufSize(); }
240 bool shouldBuildLookupTables() {
241 const TargetLoweringBase *TLI = getTLI();
242 return TLI->isOperationLegalOrCustom(ISD::BR_JT, MVT::Other) ||
243 TLI->isOperationLegalOrCustom(ISD::BRIND, MVT::Other);
246 bool haveFastSqrt(Type *Ty) {
247 const TargetLoweringBase *TLI = getTLI();
248 EVT VT = TLI->getValueType(DL, Ty);
249 return TLI->isTypeLegal(VT) &&
250 TLI->isOperationLegalOrCustom(ISD::FSQRT, VT);
253 unsigned getFPOpCost(Type *Ty) {
254 // By default, FP instructions are no more expensive since they are
255 // implemented in HW. Target specific TTI can override this.
256 return TargetTransformInfo::TCC_Basic;
259 unsigned getOperationCost(unsigned Opcode, Type *Ty, Type *OpTy) {
260 const TargetLoweringBase *TLI = getTLI();
263 case Instruction::Trunc: {
264 if (TLI->isTruncateFree(OpTy, Ty))
265 return TargetTransformInfo::TCC_Free;
266 return TargetTransformInfo::TCC_Basic;
268 case Instruction::ZExt: {
269 if (TLI->isZExtFree(OpTy, Ty))
270 return TargetTransformInfo::TCC_Free;
271 return TargetTransformInfo::TCC_Basic;
275 return BaseT::getOperationCost(Opcode, Ty, OpTy);
278 unsigned getInliningThresholdMultiplier() { return 1; }
280 void getUnrollingPreferences(Loop *L, TTI::UnrollingPreferences &UP) {
281 // This unrolling functionality is target independent, but to provide some
282 // motivation for its intended use, for x86:
284 // According to the Intel 64 and IA-32 Architectures Optimization Reference
285 // Manual, Intel Core models and later have a loop stream detector (and
286 // associated uop queue) that can benefit from partial unrolling.
287 // The relevant requirements are:
288 // - The loop must have no more than 4 (8 for Nehalem and later) branches
289 // taken, and none of them may be calls.
290 // - The loop can have no more than 18 (28 for Nehalem and later) uops.
292 // According to the Software Optimization Guide for AMD Family 15h
293 // Processors, models 30h-4fh (Steamroller and later) have a loop predictor
294 // and loop buffer which can benefit from partial unrolling.
295 // The relevant requirements are:
296 // - The loop must have fewer than 16 branches
297 // - The loop must have less than 40 uops in all executed loop branches
299 // The number of taken branches in a loop is hard to estimate here, and
300 // benchmarking has revealed that it is better not to be conservative when
301 // estimating the branch count. As a result, we'll ignore the branch limits
302 // until someone finds a case where it matters in practice.
305 const TargetSubtargetInfo *ST = getST();
306 if (PartialUnrollingThreshold.getNumOccurrences() > 0)
307 MaxOps = PartialUnrollingThreshold;
308 else if (ST->getSchedModel().LoopMicroOpBufferSize > 0)
309 MaxOps = ST->getSchedModel().LoopMicroOpBufferSize;
313 // Scan the loop: don't unroll loops with calls.
314 for (Loop::block_iterator I = L->block_begin(), E = L->block_end(); I != E;
318 for (BasicBlock::iterator J = BB->begin(), JE = BB->end(); J != JE; ++J)
319 if (isa<CallInst>(J) || isa<InvokeInst>(J)) {
320 ImmutableCallSite CS(&*J);
321 if (const Function *F = CS.getCalledFunction()) {
322 if (!static_cast<T *>(this)->isLoweredToCall(F))
330 // Enable runtime and partial unrolling up to the specified size.
331 // Enable using trip count upper bound to unroll loops.
332 UP.Partial = UP.Runtime = UP.UpperBound = true;
333 UP.PartialThreshold = MaxOps;
335 // Avoid unrolling when optimizing for size.
336 UP.OptSizeThreshold = 0;
337 UP.PartialOptSizeThreshold = 0;
339 // Set number of instructions optimized when "back edge"
340 // becomes "fall through" to default value of 2.
346 /// \name Vector TTI Implementations
349 unsigned getNumberOfRegisters(bool Vector) { return Vector ? 0 : 1; }
351 unsigned getRegisterBitWidth(bool Vector) const { return 32; }
353 /// Estimate the overhead of scalarizing an instruction. Insert and Extract
354 /// are set if the result needs to be inserted and/or extracted from vectors.
355 unsigned getScalarizationOverhead(Type *Ty, bool Insert, bool Extract) {
356 assert(Ty->isVectorTy() && "Can only scalarize vectors");
359 for (int i = 0, e = Ty->getVectorNumElements(); i < e; ++i) {
361 Cost += static_cast<T *>(this)
362 ->getVectorInstrCost(Instruction::InsertElement, Ty, i);
364 Cost += static_cast<T *>(this)
365 ->getVectorInstrCost(Instruction::ExtractElement, Ty, i);
371 /// Estimate the overhead of scalarizing an instructions unique
372 /// non-constant operands. The types of the arguments are ordinarily
373 /// scalar, in which case the costs are multiplied with VF.
374 unsigned getOperandsScalarizationOverhead(ArrayRef<const Value *> Args,
377 SmallPtrSet<const Value*, 4> UniqueOperands;
378 for (const Value *A : Args) {
379 if (!isa<Constant>(A) && UniqueOperands.insert(A).second) {
380 Type *VecTy = nullptr;
381 if (A->getType()->isVectorTy()) {
382 VecTy = A->getType();
383 // If A is a vector operand, VF should be 1 or correspond to A.
384 assert ((VF == 1 || VF == VecTy->getVectorNumElements()) &&
385 "Vector argument does not match VF");
388 VecTy = VectorType::get(A->getType(), VF);
390 Cost += getScalarizationOverhead(VecTy, false, true);
397 unsigned getScalarizationOverhead(Type *VecTy, ArrayRef<const Value *> Args) {
398 assert (VecTy->isVectorTy());
402 Cost += getScalarizationOverhead(VecTy, true, false);
404 Cost += getOperandsScalarizationOverhead(Args,
405 VecTy->getVectorNumElements());
407 // When no information on arguments is provided, we add the cost
408 // associated with one argument as a heuristic.
409 Cost += getScalarizationOverhead(VecTy, false, true);
414 unsigned getMaxInterleaveFactor(unsigned VF) { return 1; }
416 unsigned getArithmeticInstrCost(
417 unsigned Opcode, Type *Ty,
418 TTI::OperandValueKind Opd1Info = TTI::OK_AnyValue,
419 TTI::OperandValueKind Opd2Info = TTI::OK_AnyValue,
420 TTI::OperandValueProperties Opd1PropInfo = TTI::OP_None,
421 TTI::OperandValueProperties Opd2PropInfo = TTI::OP_None,
422 ArrayRef<const Value *> Args = ArrayRef<const Value *>()) {
423 // Check if any of the operands are vector operands.
424 const TargetLoweringBase *TLI = getTLI();
425 int ISD = TLI->InstructionOpcodeToISD(Opcode);
426 assert(ISD && "Invalid opcode");
428 std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(DL, Ty);
430 bool IsFloat = Ty->getScalarType()->isFloatingPointTy();
431 // Assume that floating point arithmetic operations cost twice as much as
432 // integer operations.
433 unsigned OpCost = (IsFloat ? 2 : 1);
435 if (TLI->isOperationLegalOrPromote(ISD, LT.second)) {
436 // The operation is legal. Assume it costs 1.
437 // TODO: Once we have extract/insert subvector cost we need to use them.
438 return LT.first * OpCost;
441 if (!TLI->isOperationExpand(ISD, LT.second)) {
442 // If the operation is custom lowered, then assume that the code is twice
444 return LT.first * 2 * OpCost;
447 // Else, assume that we need to scalarize this op.
448 // TODO: If one of the types get legalized by splitting, handle this
449 // similarly to what getCastInstrCost() does.
450 if (Ty->isVectorTy()) {
451 unsigned Num = Ty->getVectorNumElements();
452 unsigned Cost = static_cast<T *>(this)
453 ->getArithmeticInstrCost(Opcode, Ty->getScalarType());
454 // Return the cost of multiple scalar invocation plus the cost of
455 // inserting and extracting the values.
456 return getScalarizationOverhead(Ty, Args) + Num * Cost;
459 // We don't know anything about this scalar instruction.
463 unsigned getShuffleCost(TTI::ShuffleKind Kind, Type *Tp, int Index,
465 if (Kind == TTI::SK_Alternate || Kind == TTI::SK_PermuteTwoSrc ||
466 Kind == TTI::SK_PermuteSingleSrc) {
467 return getPermuteShuffleOverhead(Tp);
472 unsigned getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src,
473 const Instruction *I = nullptr) {
474 const TargetLoweringBase *TLI = getTLI();
475 int ISD = TLI->InstructionOpcodeToISD(Opcode);
476 assert(ISD && "Invalid opcode");
477 std::pair<unsigned, MVT> SrcLT = TLI->getTypeLegalizationCost(DL, Src);
478 std::pair<unsigned, MVT> DstLT = TLI->getTypeLegalizationCost(DL, Dst);
480 // Check for NOOP conversions.
481 if (SrcLT.first == DstLT.first &&
482 SrcLT.second.getSizeInBits() == DstLT.second.getSizeInBits()) {
484 // Bitcast between types that are legalized to the same type are free.
485 if (Opcode == Instruction::BitCast || Opcode == Instruction::Trunc)
489 if (Opcode == Instruction::Trunc &&
490 TLI->isTruncateFree(SrcLT.second, DstLT.second))
493 if (Opcode == Instruction::ZExt &&
494 TLI->isZExtFree(SrcLT.second, DstLT.second))
497 if (Opcode == Instruction::AddrSpaceCast &&
498 TLI->isNoopAddrSpaceCast(Src->getPointerAddressSpace(),
499 Dst->getPointerAddressSpace()))
502 // If this is a zext/sext of a load, return 0 if the corresponding
503 // extending load exists on target.
504 if ((Opcode == Instruction::ZExt || Opcode == Instruction::SExt) &&
505 I && isa<LoadInst>(I->getOperand(0))) {
506 EVT ExtVT = EVT::getEVT(Dst);
507 EVT LoadVT = EVT::getEVT(Src);
509 ((Opcode == Instruction::ZExt) ? ISD::ZEXTLOAD : ISD::SEXTLOAD);
510 if (TLI->isLoadExtLegal(LType, ExtVT, LoadVT))
514 // If the cast is marked as legal (or promote) then assume low cost.
515 if (SrcLT.first == DstLT.first &&
516 TLI->isOperationLegalOrPromote(ISD, DstLT.second))
519 // Handle scalar conversions.
520 if (!Src->isVectorTy() && !Dst->isVectorTy()) {
522 // Scalar bitcasts are usually free.
523 if (Opcode == Instruction::BitCast)
526 // Just check the op cost. If the operation is legal then assume it costs
528 if (!TLI->isOperationExpand(ISD, DstLT.second))
531 // Assume that illegal scalar instruction are expensive.
535 // Check vector-to-vector casts.
536 if (Dst->isVectorTy() && Src->isVectorTy()) {
538 // If the cast is between same-sized registers, then the check is simple.
539 if (SrcLT.first == DstLT.first &&
540 SrcLT.second.getSizeInBits() == DstLT.second.getSizeInBits()) {
542 // Assume that Zext is done using AND.
543 if (Opcode == Instruction::ZExt)
546 // Assume that sext is done using SHL and SRA.
547 if (Opcode == Instruction::SExt)
550 // Just check the op cost. If the operation is legal then assume it
552 // 1 and multiply by the type-legalization overhead.
553 if (!TLI->isOperationExpand(ISD, DstLT.second))
554 return SrcLT.first * 1;
557 // If we are legalizing by splitting, query the concrete TTI for the cost
558 // of casting the original vector twice. We also need to factor int the
559 // cost of the split itself. Count that as 1, to be consistent with
560 // TLI->getTypeLegalizationCost().
561 if ((TLI->getTypeAction(Src->getContext(), TLI->getValueType(DL, Src)) ==
562 TargetLowering::TypeSplitVector) ||
563 (TLI->getTypeAction(Dst->getContext(), TLI->getValueType(DL, Dst)) ==
564 TargetLowering::TypeSplitVector)) {
565 Type *SplitDst = VectorType::get(Dst->getVectorElementType(),
566 Dst->getVectorNumElements() / 2);
567 Type *SplitSrc = VectorType::get(Src->getVectorElementType(),
568 Src->getVectorNumElements() / 2);
569 T *TTI = static_cast<T *>(this);
570 return TTI->getVectorSplitCost() +
571 (2 * TTI->getCastInstrCost(Opcode, SplitDst, SplitSrc, I));
574 // In other cases where the source or destination are illegal, assume
575 // the operation will get scalarized.
576 unsigned Num = Dst->getVectorNumElements();
577 unsigned Cost = static_cast<T *>(this)->getCastInstrCost(
578 Opcode, Dst->getScalarType(), Src->getScalarType(), I);
580 // Return the cost of multiple scalar invocation plus the cost of
581 // inserting and extracting the values.
582 return getScalarizationOverhead(Dst, true, true) + Num * Cost;
585 // We already handled vector-to-vector and scalar-to-scalar conversions.
587 // is where we handle bitcast between vectors and scalars. We need to assume
588 // that the conversion is scalarized in one way or another.
589 if (Opcode == Instruction::BitCast)
590 // Illegal bitcasts are done by storing and loading from a stack slot.
591 return (Src->isVectorTy() ? getScalarizationOverhead(Src, false, true)
593 (Dst->isVectorTy() ? getScalarizationOverhead(Dst, true, false)
596 llvm_unreachable("Unhandled cast");
599 unsigned getExtractWithExtendCost(unsigned Opcode, Type *Dst,
600 VectorType *VecTy, unsigned Index) {
601 return static_cast<T *>(this)->getVectorInstrCost(
602 Instruction::ExtractElement, VecTy, Index) +
603 static_cast<T *>(this)->getCastInstrCost(Opcode, Dst,
604 VecTy->getElementType());
607 unsigned getCFInstrCost(unsigned Opcode) {
608 // Branches are assumed to be predicted.
612 unsigned getCmpSelInstrCost(unsigned Opcode, Type *ValTy, Type *CondTy,
613 const Instruction *I) {
614 const TargetLoweringBase *TLI = getTLI();
615 int ISD = TLI->InstructionOpcodeToISD(Opcode);
616 assert(ISD && "Invalid opcode");
618 // Selects on vectors are actually vector selects.
619 if (ISD == ISD::SELECT) {
620 assert(CondTy && "CondTy must exist");
621 if (CondTy->isVectorTy())
624 std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(DL, ValTy);
626 if (!(ValTy->isVectorTy() && !LT.second.isVector()) &&
627 !TLI->isOperationExpand(ISD, LT.second)) {
628 // The operation is legal. Assume it costs 1. Multiply
629 // by the type-legalization overhead.
633 // Otherwise, assume that the cast is scalarized.
634 // TODO: If one of the types get legalized by splitting, handle this
635 // similarly to what getCastInstrCost() does.
636 if (ValTy->isVectorTy()) {
637 unsigned Num = ValTy->getVectorNumElements();
639 CondTy = CondTy->getScalarType();
640 unsigned Cost = static_cast<T *>(this)->getCmpSelInstrCost(
641 Opcode, ValTy->getScalarType(), CondTy, I);
643 // Return the cost of multiple scalar invocation plus the cost of
644 // inserting and extracting the values.
645 return getScalarizationOverhead(ValTy, true, false) + Num * Cost;
648 // Unknown scalar opcode.
652 unsigned getVectorInstrCost(unsigned Opcode, Type *Val, unsigned Index) {
653 std::pair<unsigned, MVT> LT =
654 getTLI()->getTypeLegalizationCost(DL, Val->getScalarType());
659 unsigned getMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
660 unsigned AddressSpace, const Instruction *I = nullptr) {
661 assert(!Src->isVoidTy() && "Invalid type");
662 std::pair<unsigned, MVT> LT = getTLI()->getTypeLegalizationCost(DL, Src);
664 // Assuming that all loads of legal types cost 1.
665 unsigned Cost = LT.first;
667 if (Src->isVectorTy() &&
668 Src->getPrimitiveSizeInBits() < LT.second.getSizeInBits()) {
669 // This is a vector load that legalizes to a larger type than the vector
670 // itself. Unless the corresponding extending load or truncating store is
671 // legal, then this will scalarize.
672 TargetLowering::LegalizeAction LA = TargetLowering::Expand;
673 EVT MemVT = getTLI()->getValueType(DL, Src);
674 if (Opcode == Instruction::Store)
675 LA = getTLI()->getTruncStoreAction(LT.second, MemVT);
677 LA = getTLI()->getLoadExtAction(ISD::EXTLOAD, LT.second, MemVT);
679 if (LA != TargetLowering::Legal && LA != TargetLowering::Custom) {
680 // This is a vector load/store for some illegal type that is scalarized.
681 // We must account for the cost of building or decomposing the vector.
682 Cost += getScalarizationOverhead(Src, Opcode != Instruction::Store,
683 Opcode == Instruction::Store);
690 unsigned getInterleavedMemoryOpCost(unsigned Opcode, Type *VecTy,
692 ArrayRef<unsigned> Indices,
694 unsigned AddressSpace) {
695 VectorType *VT = dyn_cast<VectorType>(VecTy);
696 assert(VT && "Expect a vector type for interleaved memory op");
698 unsigned NumElts = VT->getNumElements();
699 assert(Factor > 1 && NumElts % Factor == 0 && "Invalid interleave factor");
701 unsigned NumSubElts = NumElts / Factor;
702 VectorType *SubVT = VectorType::get(VT->getElementType(), NumSubElts);
704 // Firstly, the cost of load/store operation.
705 unsigned Cost = static_cast<T *>(this)->getMemoryOpCost(
706 Opcode, VecTy, Alignment, AddressSpace);
708 // Legalize the vector type, and get the legalized and unlegalized type
710 MVT VecTyLT = getTLI()->getTypeLegalizationCost(DL, VecTy).second;
712 static_cast<T *>(this)->getDataLayout().getTypeStoreSize(VecTy);
713 unsigned VecTyLTSize = VecTyLT.getStoreSize();
715 // Return the ceiling of dividing A by B.
716 auto ceil = [](unsigned A, unsigned B) { return (A + B - 1) / B; };
718 // Scale the cost of the memory operation by the fraction of legalized
719 // instructions that will actually be used. We shouldn't account for the
720 // cost of dead instructions since they will be removed.
722 // E.g., An interleaved load of factor 8:
723 // %vec = load <16 x i64>, <16 x i64>* %ptr
724 // %v0 = shufflevector %vec, undef, <0, 8>
726 // If <16 x i64> is legalized to 8 v2i64 loads, only 2 of the loads will be
727 // used (those corresponding to elements [0:1] and [8:9] of the unlegalized
728 // type). The other loads are unused.
730 // We only scale the cost of loads since interleaved store groups aren't
731 // allowed to have gaps.
732 if (Opcode == Instruction::Load && VecTySize > VecTyLTSize) {
734 // The number of loads of a legal type it will take to represent a load
735 // of the unlegalized vector type.
736 unsigned NumLegalInsts = ceil(VecTySize, VecTyLTSize);
738 // The number of elements of the unlegalized type that correspond to a
739 // single legal instruction.
740 unsigned NumEltsPerLegalInst = ceil(NumElts, NumLegalInsts);
742 // Determine which legal instructions will be used.
743 BitVector UsedInsts(NumLegalInsts, false);
744 for (unsigned Index : Indices)
745 for (unsigned Elt = 0; Elt < NumSubElts; ++Elt)
746 UsedInsts.set((Index + Elt * Factor) / NumEltsPerLegalInst);
748 // Scale the cost of the load by the fraction of legal instructions that
750 Cost *= UsedInsts.count() / NumLegalInsts;
753 // Then plus the cost of interleave operation.
754 if (Opcode == Instruction::Load) {
755 // The interleave cost is similar to extract sub vectors' elements
756 // from the wide vector, and insert them into sub vectors.
758 // E.g. An interleaved load of factor 2 (with one member of index 0):
759 // %vec = load <8 x i32>, <8 x i32>* %ptr
760 // %v0 = shuffle %vec, undef, <0, 2, 4, 6> ; Index 0
761 // The cost is estimated as extract elements at 0, 2, 4, 6 from the
762 // <8 x i32> vector and insert them into a <4 x i32> vector.
764 assert(Indices.size() <= Factor &&
765 "Interleaved memory op has too many members");
767 for (unsigned Index : Indices) {
768 assert(Index < Factor && "Invalid index for interleaved memory op");
770 // Extract elements from loaded vector for each sub vector.
771 for (unsigned i = 0; i < NumSubElts; i++)
772 Cost += static_cast<T *>(this)->getVectorInstrCost(
773 Instruction::ExtractElement, VT, Index + i * Factor);
776 unsigned InsSubCost = 0;
777 for (unsigned i = 0; i < NumSubElts; i++)
778 InsSubCost += static_cast<T *>(this)->getVectorInstrCost(
779 Instruction::InsertElement, SubVT, i);
781 Cost += Indices.size() * InsSubCost;
783 // The interleave cost is extract all elements from sub vectors, and
784 // insert them into the wide vector.
786 // E.g. An interleaved store of factor 2:
787 // %v0_v1 = shuffle %v0, %v1, <0, 4, 1, 5, 2, 6, 3, 7>
788 // store <8 x i32> %interleaved.vec, <8 x i32>* %ptr
789 // The cost is estimated as extract all elements from both <4 x i32>
790 // vectors and insert into the <8 x i32> vector.
792 unsigned ExtSubCost = 0;
793 for (unsigned i = 0; i < NumSubElts; i++)
794 ExtSubCost += static_cast<T *>(this)->getVectorInstrCost(
795 Instruction::ExtractElement, SubVT, i);
796 Cost += ExtSubCost * Factor;
798 for (unsigned i = 0; i < NumElts; i++)
799 Cost += static_cast<T *>(this)
800 ->getVectorInstrCost(Instruction::InsertElement, VT, i);
806 /// Get intrinsic cost based on arguments.
807 unsigned getIntrinsicInstrCost(Intrinsic::ID IID, Type *RetTy,
808 ArrayRef<Value *> Args, FastMathFlags FMF,
810 unsigned RetVF = (RetTy->isVectorTy() ? RetTy->getVectorNumElements() : 1);
811 assert ((RetVF == 1 || VF == 1) && "VF > 1 and RetVF is a vector type");
815 // Assume that we need to scalarize this intrinsic.
816 SmallVector<Type *, 4> Types;
817 for (Value *Op : Args) {
818 Type *OpTy = Op->getType();
819 assert (VF == 1 || !OpTy->isVectorTy());
820 Types.push_back(VF == 1 ? OpTy : VectorType::get(OpTy, VF));
823 if (VF > 1 && !RetTy->isVoidTy())
824 RetTy = VectorType::get(RetTy, VF);
826 // Compute the scalarization overhead based on Args for a vector
827 // intrinsic. A vectorizer will pass a scalar RetTy and VF > 1, while
828 // CostModel will pass a vector RetTy and VF is 1.
829 unsigned ScalarizationCost = UINT_MAX;
830 if (RetVF > 1 || VF > 1) {
831 ScalarizationCost = 0;
832 if (!RetTy->isVoidTy())
833 ScalarizationCost += getScalarizationOverhead(RetTy, true, false);
834 ScalarizationCost += getOperandsScalarizationOverhead(Args, VF);
837 return static_cast<T *>(this)->
838 getIntrinsicInstrCost(IID, RetTy, Types, FMF, ScalarizationCost);
840 case Intrinsic::masked_scatter: {
841 assert (VF == 1 && "Can't vectorize types here.");
842 Value *Mask = Args[3];
843 bool VarMask = !isa<Constant>(Mask);
844 unsigned Alignment = cast<ConstantInt>(Args[2])->getZExtValue();
846 static_cast<T *>(this)->getGatherScatterOpCost(Instruction::Store,
851 case Intrinsic::masked_gather: {
852 assert (VF == 1 && "Can't vectorize types here.");
853 Value *Mask = Args[2];
854 bool VarMask = !isa<Constant>(Mask);
855 unsigned Alignment = cast<ConstantInt>(Args[1])->getZExtValue();
857 static_cast<T *>(this)->getGatherScatterOpCost(Instruction::Load,
858 RetTy, Args[0], VarMask,
864 /// Get intrinsic cost based on argument types.
865 /// If ScalarizationCostPassed is UINT_MAX, the cost of scalarizing the
866 /// arguments and the return value will be computed based on types.
867 unsigned getIntrinsicInstrCost(Intrinsic::ID IID, Type *RetTy,
868 ArrayRef<Type *> Tys, FastMathFlags FMF,
869 unsigned ScalarizationCostPassed = UINT_MAX) {
870 SmallVector<unsigned, 2> ISDs;
871 unsigned SingleCallCost = 10; // Library call cost. Make it expensive.
874 // Assume that we need to scalarize this intrinsic.
875 unsigned ScalarizationCost = ScalarizationCostPassed;
876 unsigned ScalarCalls = 1;
877 Type *ScalarRetTy = RetTy;
878 if (RetTy->isVectorTy()) {
879 if (ScalarizationCostPassed == UINT_MAX)
880 ScalarizationCost = getScalarizationOverhead(RetTy, true, false);
881 ScalarCalls = std::max(ScalarCalls, RetTy->getVectorNumElements());
882 ScalarRetTy = RetTy->getScalarType();
884 SmallVector<Type *, 4> ScalarTys;
885 for (unsigned i = 0, ie = Tys.size(); i != ie; ++i) {
887 if (Ty->isVectorTy()) {
888 if (ScalarizationCostPassed == UINT_MAX)
889 ScalarizationCost += getScalarizationOverhead(Ty, false, true);
890 ScalarCalls = std::max(ScalarCalls, Ty->getVectorNumElements());
891 Ty = Ty->getScalarType();
893 ScalarTys.push_back(Ty);
895 if (ScalarCalls == 1)
896 return 1; // Return cost of a scalar intrinsic. Assume it to be cheap.
898 unsigned ScalarCost = static_cast<T *>(this)->getIntrinsicInstrCost(
899 IID, ScalarRetTy, ScalarTys, FMF);
901 return ScalarCalls * ScalarCost + ScalarizationCost;
903 // Look for intrinsics that can be lowered directly or turned into a scalar
905 case Intrinsic::sqrt:
906 ISDs.push_back(ISD::FSQRT);
909 ISDs.push_back(ISD::FSIN);
912 ISDs.push_back(ISD::FCOS);
915 ISDs.push_back(ISD::FEXP);
917 case Intrinsic::exp2:
918 ISDs.push_back(ISD::FEXP2);
921 ISDs.push_back(ISD::FLOG);
923 case Intrinsic::log10:
924 ISDs.push_back(ISD::FLOG10);
926 case Intrinsic::log2:
927 ISDs.push_back(ISD::FLOG2);
929 case Intrinsic::fabs:
930 ISDs.push_back(ISD::FABS);
932 case Intrinsic::minnum:
933 ISDs.push_back(ISD::FMINNUM);
935 ISDs.push_back(ISD::FMINNAN);
937 case Intrinsic::maxnum:
938 ISDs.push_back(ISD::FMAXNUM);
940 ISDs.push_back(ISD::FMAXNAN);
942 case Intrinsic::copysign:
943 ISDs.push_back(ISD::FCOPYSIGN);
945 case Intrinsic::floor:
946 ISDs.push_back(ISD::FFLOOR);
948 case Intrinsic::ceil:
949 ISDs.push_back(ISD::FCEIL);
951 case Intrinsic::trunc:
952 ISDs.push_back(ISD::FTRUNC);
954 case Intrinsic::nearbyint:
955 ISDs.push_back(ISD::FNEARBYINT);
957 case Intrinsic::rint:
958 ISDs.push_back(ISD::FRINT);
960 case Intrinsic::round:
961 ISDs.push_back(ISD::FROUND);
964 ISDs.push_back(ISD::FPOW);
967 ISDs.push_back(ISD::FMA);
969 case Intrinsic::fmuladd:
970 ISDs.push_back(ISD::FMA);
972 // FIXME: We should return 0 whenever getIntrinsicCost == TCC_Free.
973 case Intrinsic::lifetime_start:
974 case Intrinsic::lifetime_end:
976 case Intrinsic::masked_store:
977 return static_cast<T *>(this)
978 ->getMaskedMemoryOpCost(Instruction::Store, Tys[0], 0, 0);
979 case Intrinsic::masked_load:
980 return static_cast<T *>(this)
981 ->getMaskedMemoryOpCost(Instruction::Load, RetTy, 0, 0);
982 case Intrinsic::ctpop:
983 ISDs.push_back(ISD::CTPOP);
984 // In case of legalization use TCC_Expensive. This is cheaper than a
985 // library call but still not a cheap instruction.
986 SingleCallCost = TargetTransformInfo::TCC_Expensive;
988 // FIXME: ctlz, cttz, ...
991 const TargetLoweringBase *TLI = getTLI();
992 std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(DL, RetTy);
994 SmallVector<unsigned, 2> LegalCost;
995 SmallVector<unsigned, 2> CustomCost;
996 for (unsigned ISD : ISDs) {
997 if (TLI->isOperationLegalOrPromote(ISD, LT.second)) {
998 if (IID == Intrinsic::fabs && TLI->isFAbsFree(LT.second)) {
1002 // The operation is legal. Assume it costs 1.
1003 // If the type is split to multiple registers, assume that there is some
1004 // overhead to this.
1005 // TODO: Once we have extract/insert subvector cost we need to use them.
1007 LegalCost.push_back(LT.first * 2);
1009 LegalCost.push_back(LT.first * 1);
1010 } else if (!TLI->isOperationExpand(ISD, LT.second)) {
1011 // If the operation is custom lowered then assume
1012 // that the code is twice as expensive.
1013 CustomCost.push_back(LT.first * 2);
1017 auto MinLegalCostI = std::min_element(LegalCost.begin(), LegalCost.end());
1018 if (MinLegalCostI != LegalCost.end())
1019 return *MinLegalCostI;
1021 auto MinCustomCostI = std::min_element(CustomCost.begin(), CustomCost.end());
1022 if (MinCustomCostI != CustomCost.end())
1023 return *MinCustomCostI;
1025 // If we can't lower fmuladd into an FMA estimate the cost as a floating
1026 // point mul followed by an add.
1027 if (IID == Intrinsic::fmuladd)
1028 return static_cast<T *>(this)
1029 ->getArithmeticInstrCost(BinaryOperator::FMul, RetTy) +
1030 static_cast<T *>(this)
1031 ->getArithmeticInstrCost(BinaryOperator::FAdd, RetTy);
1033 // Else, assume that we need to scalarize this intrinsic. For math builtins
1034 // this will emit a costly libcall, adding call overhead and spills. Make it
1036 if (RetTy->isVectorTy()) {
1037 unsigned ScalarizationCost = ((ScalarizationCostPassed != UINT_MAX) ?
1038 ScalarizationCostPassed : getScalarizationOverhead(RetTy, true, false));
1039 unsigned ScalarCalls = RetTy->getVectorNumElements();
1040 SmallVector<Type *, 4> ScalarTys;
1041 for (unsigned i = 0, ie = Tys.size(); i != ie; ++i) {
1043 if (Ty->isVectorTy())
1044 Ty = Ty->getScalarType();
1045 ScalarTys.push_back(Ty);
1047 unsigned ScalarCost = static_cast<T *>(this)->getIntrinsicInstrCost(
1048 IID, RetTy->getScalarType(), ScalarTys, FMF);
1049 for (unsigned i = 0, ie = Tys.size(); i != ie; ++i) {
1050 if (Tys[i]->isVectorTy()) {
1051 if (ScalarizationCostPassed == UINT_MAX)
1052 ScalarizationCost += getScalarizationOverhead(Tys[i], false, true);
1053 ScalarCalls = std::max(ScalarCalls, Tys[i]->getVectorNumElements());
1057 return ScalarCalls * ScalarCost + ScalarizationCost;
1060 // This is going to be turned into a library call, make it expensive.
1061 return SingleCallCost;
1064 /// \brief Compute a cost of the given call instruction.
1066 /// Compute the cost of calling function F with return type RetTy and
1067 /// argument types Tys. F might be nullptr, in this case the cost of an
1068 /// arbitrary call with the specified signature will be returned.
1069 /// This is used, for instance, when we estimate call of a vector
1070 /// counterpart of the given function.
1071 /// \param F Called function, might be nullptr.
1072 /// \param RetTy Return value types.
1073 /// \param Tys Argument types.
1074 /// \returns The cost of Call instruction.
1075 unsigned getCallInstrCost(Function *F, Type *RetTy, ArrayRef<Type *> Tys) {
1079 unsigned getNumberOfParts(Type *Tp) {
1080 std::pair<unsigned, MVT> LT = getTLI()->getTypeLegalizationCost(DL, Tp);
1084 unsigned getAddressComputationCost(Type *Ty, ScalarEvolution *,
1089 /// Try to calculate arithmetic and shuffle op costs for reduction operations.
1090 /// We're assuming that reduction operation are performing the following way:
1091 /// 1. Non-pairwise reduction
1092 /// %val1 = shufflevector<n x t> %val, <n x t> %undef,
1093 /// <n x i32> <i32 n/2, i32 n/2 + 1, ..., i32 n, i32 undef, ..., i32 undef>
1094 /// \----------------v-------------/ \----------v------------/
1095 /// n/2 elements n/2 elements
1096 /// %red1 = op <n x t> %val, <n x t> val1
1097 /// After this operation we have a vector %red1 where only the first n/2
1098 /// elements are meaningful, the second n/2 elements are undefined and can be
1099 /// dropped. All other operations are actually working with the vector of
1100 /// length n/2, not n, though the real vector length is still n.
1101 /// %val2 = shufflevector<n x t> %red1, <n x t> %undef,
1102 /// <n x i32> <i32 n/4, i32 n/4 + 1, ..., i32 n/2, i32 undef, ..., i32 undef>
1103 /// \----------------v-------------/ \----------v------------/
1104 /// n/4 elements 3*n/4 elements
1105 /// %red2 = op <n x t> %red1, <n x t> val2 - working with the vector of
1106 /// length n/2, the resulting vector has length n/4 etc.
1107 /// 2. Pairwise reduction:
1108 /// Everything is the same except for an additional shuffle operation which
1109 /// is used to produce operands for pairwise kind of reductions.
1110 /// %val1 = shufflevector<n x t> %val, <n x t> %undef,
1111 /// <n x i32> <i32 0, i32 2, ..., i32 n-2, i32 undef, ..., i32 undef>
1112 /// \-------------v----------/ \----------v------------/
1113 /// n/2 elements n/2 elements
1114 /// %val2 = shufflevector<n x t> %val, <n x t> %undef,
1115 /// <n x i32> <i32 1, i32 3, ..., i32 n-1, i32 undef, ..., i32 undef>
1116 /// \-------------v----------/ \----------v------------/
1117 /// n/2 elements n/2 elements
1118 /// %red1 = op <n x t> %val1, <n x t> val2
1119 /// Again, the operation is performed on <n x t> vector, but the resulting
1120 /// vector %red1 is <n/2 x t> vector.
1122 /// The cost model should take into account that the actual length of the
1123 /// vector is reduced on each iteration.
1124 unsigned getReductionCost(unsigned Opcode, Type *Ty, bool IsPairwise) {
1125 assert(Ty->isVectorTy() && "Expect a vector type");
1126 Type *ScalarTy = Ty->getVectorElementType();
1127 unsigned NumVecElts = Ty->getVectorNumElements();
1128 unsigned NumReduxLevels = Log2_32(NumVecElts);
1129 unsigned ArithCost = 0;
1130 unsigned ShuffleCost = 0;
1131 auto *ConcreteTTI = static_cast<T *>(this);
1132 std::pair<unsigned, MVT> LT =
1133 ConcreteTTI->getTLI()->getTypeLegalizationCost(DL, Ty);
1134 unsigned LongVectorCount = 0;
1136 LT.second.isVector() ? LT.second.getVectorNumElements() : 1;
1137 while (NumVecElts > MVTLen) {
1139 // Assume the pairwise shuffles add a cost.
1140 ShuffleCost += (IsPairwise + 1) *
1141 ConcreteTTI->getShuffleCost(TTI::SK_ExtractSubvector, Ty,
1143 ArithCost += ConcreteTTI->getArithmeticInstrCost(Opcode, Ty);
1144 Ty = VectorType::get(ScalarTy, NumVecElts);
1147 // The minimal length of the vector is limited by the real length of vector
1148 // operations performed on the current platform. That's why several final
1149 // reduction opertions are perfomed on the vectors with the same
1150 // architecture-dependent length.
1151 ShuffleCost += (NumReduxLevels - LongVectorCount) * (IsPairwise + 1) *
1152 ConcreteTTI->getShuffleCost(TTI::SK_ExtractSubvector, Ty,
1154 ArithCost += (NumReduxLevels - LongVectorCount) *
1155 ConcreteTTI->getArithmeticInstrCost(Opcode, Ty);
1156 return ShuffleCost + ArithCost + getScalarizationOverhead(Ty, false, true);
1159 unsigned getVectorSplitCost() { return 1; }
1164 /// \brief Concrete BasicTTIImpl that can be used if no further customization
1166 class BasicTTIImpl : public BasicTTIImplBase<BasicTTIImpl> {
1167 typedef BasicTTIImplBase<BasicTTIImpl> BaseT;
1168 friend class BasicTTIImplBase<BasicTTIImpl>;
1170 const TargetSubtargetInfo *ST;
1171 const TargetLoweringBase *TLI;
1173 const TargetSubtargetInfo *getST() const { return ST; }
1174 const TargetLoweringBase *getTLI() const { return TLI; }
1177 explicit BasicTTIImpl(const TargetMachine *ST, const Function &F);