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 getJumpBufAlignment() { return getTLI()->getJumpBufAlignment(); }
176 unsigned getJumpBufSize() { return getTLI()->getJumpBufSize(); }
178 bool shouldBuildLookupTables() {
179 const TargetLoweringBase *TLI = getTLI();
180 return TLI->isOperationLegalOrCustom(ISD::BR_JT, MVT::Other) ||
181 TLI->isOperationLegalOrCustom(ISD::BRIND, MVT::Other);
184 bool haveFastSqrt(Type *Ty) {
185 const TargetLoweringBase *TLI = getTLI();
186 EVT VT = TLI->getValueType(DL, Ty);
187 return TLI->isTypeLegal(VT) &&
188 TLI->isOperationLegalOrCustom(ISD::FSQRT, VT);
191 unsigned getFPOpCost(Type *Ty) {
192 // By default, FP instructions are no more expensive since they are
193 // implemented in HW. Target specific TTI can override this.
194 return TargetTransformInfo::TCC_Basic;
197 unsigned getOperationCost(unsigned Opcode, Type *Ty, Type *OpTy) {
198 const TargetLoweringBase *TLI = getTLI();
201 case Instruction::Trunc: {
202 if (TLI->isTruncateFree(OpTy, Ty))
203 return TargetTransformInfo::TCC_Free;
204 return TargetTransformInfo::TCC_Basic;
206 case Instruction::ZExt: {
207 if (TLI->isZExtFree(OpTy, Ty))
208 return TargetTransformInfo::TCC_Free;
209 return TargetTransformInfo::TCC_Basic;
213 return BaseT::getOperationCost(Opcode, Ty, OpTy);
216 unsigned getInliningThresholdMultiplier() { return 1; }
218 void getUnrollingPreferences(Loop *L, TTI::UnrollingPreferences &UP) {
219 // This unrolling functionality is target independent, but to provide some
220 // motivation for its intended use, for x86:
222 // According to the Intel 64 and IA-32 Architectures Optimization Reference
223 // Manual, Intel Core models and later have a loop stream detector (and
224 // associated uop queue) that can benefit from partial unrolling.
225 // The relevant requirements are:
226 // - The loop must have no more than 4 (8 for Nehalem and later) branches
227 // taken, and none of them may be calls.
228 // - The loop can have no more than 18 (28 for Nehalem and later) uops.
230 // According to the Software Optimization Guide for AMD Family 15h
231 // Processors, models 30h-4fh (Steamroller and later) have a loop predictor
232 // and loop buffer which can benefit from partial unrolling.
233 // The relevant requirements are:
234 // - The loop must have fewer than 16 branches
235 // - The loop must have less than 40 uops in all executed loop branches
237 // The number of taken branches in a loop is hard to estimate here, and
238 // benchmarking has revealed that it is better not to be conservative when
239 // estimating the branch count. As a result, we'll ignore the branch limits
240 // until someone finds a case where it matters in practice.
243 const TargetSubtargetInfo *ST = getST();
244 if (PartialUnrollingThreshold.getNumOccurrences() > 0)
245 MaxOps = PartialUnrollingThreshold;
246 else if (ST->getSchedModel().LoopMicroOpBufferSize > 0)
247 MaxOps = ST->getSchedModel().LoopMicroOpBufferSize;
251 // Scan the loop: don't unroll loops with calls.
252 for (Loop::block_iterator I = L->block_begin(), E = L->block_end(); I != E;
256 for (BasicBlock::iterator J = BB->begin(), JE = BB->end(); J != JE; ++J)
257 if (isa<CallInst>(J) || isa<InvokeInst>(J)) {
258 ImmutableCallSite CS(&*J);
259 if (const Function *F = CS.getCalledFunction()) {
260 if (!static_cast<T *>(this)->isLoweredToCall(F))
268 // Enable runtime and partial unrolling up to the specified size.
269 // Enable using trip count upper bound to unroll loops.
270 UP.Partial = UP.Runtime = UP.UpperBound = true;
271 UP.PartialThreshold = MaxOps;
273 // Avoid unrolling when optimizing for size.
274 UP.OptSizeThreshold = 0;
275 UP.PartialOptSizeThreshold = 0;
277 // Set number of instructions optimized when "back edge"
278 // becomes "fall through" to default value of 2.
284 /// \name Vector TTI Implementations
287 unsigned getNumberOfRegisters(bool Vector) { return Vector ? 0 : 1; }
289 unsigned getRegisterBitWidth(bool Vector) { return 32; }
291 /// Estimate the overhead of scalarizing an instruction. Insert and Extract
292 /// are set if the result needs to be inserted and/or extracted from vectors.
293 unsigned getScalarizationOverhead(Type *Ty, bool Insert, bool Extract) {
294 assert(Ty->isVectorTy() && "Can only scalarize vectors");
297 for (int i = 0, e = Ty->getVectorNumElements(); i < e; ++i) {
299 Cost += static_cast<T *>(this)
300 ->getVectorInstrCost(Instruction::InsertElement, Ty, i);
302 Cost += static_cast<T *>(this)
303 ->getVectorInstrCost(Instruction::ExtractElement, Ty, i);
309 /// Estimate the overhead of scalarizing an instructions unique
310 /// non-constant operands. The types of the arguments are ordinarily
311 /// scalar, in which case the costs are multiplied with VF.
312 unsigned getOperandsScalarizationOverhead(ArrayRef<const Value *> Args,
315 SmallPtrSet<const Value*, 4> UniqueOperands;
316 for (const Value *A : Args) {
317 if (!isa<Constant>(A) && UniqueOperands.insert(A).second) {
318 Type *VecTy = nullptr;
319 if (A->getType()->isVectorTy()) {
320 VecTy = A->getType();
321 // If A is a vector operand, VF should be 1 or correspond to A.
322 assert ((VF == 1 || VF == VecTy->getVectorNumElements()) &&
323 "Vector argument does not match VF");
326 VecTy = VectorType::get(A->getType(), VF);
328 Cost += getScalarizationOverhead(VecTy, false, true);
335 unsigned getScalarizationOverhead(Type *VecTy, ArrayRef<const Value *> Args) {
336 assert (VecTy->isVectorTy());
340 Cost += getScalarizationOverhead(VecTy, true, false);
342 Cost += getOperandsScalarizationOverhead(Args,
343 VecTy->getVectorNumElements());
345 // When no information on arguments is provided, we add the cost
346 // associated with one argument as a heuristic.
347 Cost += getScalarizationOverhead(VecTy, false, true);
352 unsigned getMaxInterleaveFactor(unsigned VF) { return 1; }
354 unsigned getArithmeticInstrCost(
355 unsigned Opcode, Type *Ty,
356 TTI::OperandValueKind Opd1Info = TTI::OK_AnyValue,
357 TTI::OperandValueKind Opd2Info = TTI::OK_AnyValue,
358 TTI::OperandValueProperties Opd1PropInfo = TTI::OP_None,
359 TTI::OperandValueProperties Opd2PropInfo = TTI::OP_None,
360 ArrayRef<const Value *> Args = ArrayRef<const Value *>()) {
361 // Check if any of the operands are vector operands.
362 const TargetLoweringBase *TLI = getTLI();
363 int ISD = TLI->InstructionOpcodeToISD(Opcode);
364 assert(ISD && "Invalid opcode");
366 std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(DL, Ty);
368 bool IsFloat = Ty->getScalarType()->isFloatingPointTy();
369 // Assume that floating point arithmetic operations cost twice as much as
370 // integer operations.
371 unsigned OpCost = (IsFloat ? 2 : 1);
373 if (TLI->isOperationLegalOrPromote(ISD, LT.second)) {
374 // The operation is legal. Assume it costs 1.
375 // TODO: Once we have extract/insert subvector cost we need to use them.
376 return LT.first * OpCost;
379 if (!TLI->isOperationExpand(ISD, LT.second)) {
380 // If the operation is custom lowered, then assume that the code is twice
382 return LT.first * 2 * OpCost;
385 // Else, assume that we need to scalarize this op.
386 // TODO: If one of the types get legalized by splitting, handle this
387 // similarly to what getCastInstrCost() does.
388 if (Ty->isVectorTy()) {
389 unsigned Num = Ty->getVectorNumElements();
390 unsigned Cost = static_cast<T *>(this)
391 ->getArithmeticInstrCost(Opcode, Ty->getScalarType());
392 // Return the cost of multiple scalar invocation plus the cost of
393 // inserting and extracting the values.
394 return getScalarizationOverhead(Ty, Args) + Num * Cost;
397 // We don't know anything about this scalar instruction.
401 unsigned getShuffleCost(TTI::ShuffleKind Kind, Type *Tp, int Index,
403 if (Kind == TTI::SK_Alternate || Kind == TTI::SK_PermuteTwoSrc ||
404 Kind == TTI::SK_PermuteSingleSrc) {
405 return getPermuteShuffleOverhead(Tp);
410 unsigned getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src,
411 const Instruction *I = nullptr) {
412 const TargetLoweringBase *TLI = getTLI();
413 int ISD = TLI->InstructionOpcodeToISD(Opcode);
414 assert(ISD && "Invalid opcode");
415 std::pair<unsigned, MVT> SrcLT = TLI->getTypeLegalizationCost(DL, Src);
416 std::pair<unsigned, MVT> DstLT = TLI->getTypeLegalizationCost(DL, Dst);
418 // Check for NOOP conversions.
419 if (SrcLT.first == DstLT.first &&
420 SrcLT.second.getSizeInBits() == DstLT.second.getSizeInBits()) {
422 // Bitcast between types that are legalized to the same type are free.
423 if (Opcode == Instruction::BitCast || Opcode == Instruction::Trunc)
427 if (Opcode == Instruction::Trunc &&
428 TLI->isTruncateFree(SrcLT.second, DstLT.second))
431 if (Opcode == Instruction::ZExt &&
432 TLI->isZExtFree(SrcLT.second, DstLT.second))
435 if (Opcode == Instruction::AddrSpaceCast &&
436 TLI->isNoopAddrSpaceCast(Src->getPointerAddressSpace(),
437 Dst->getPointerAddressSpace()))
440 // If this is a zext/sext of a load, return 0 if the corresponding
441 // extending load exists on target.
442 if ((Opcode == Instruction::ZExt || Opcode == Instruction::SExt) &&
443 I && isa<LoadInst>(I->getOperand(0))) {
444 EVT ExtVT = EVT::getEVT(Dst);
445 EVT LoadVT = EVT::getEVT(Src);
447 ((Opcode == Instruction::ZExt) ? ISD::ZEXTLOAD : ISD::SEXTLOAD);
448 if (TLI->isLoadExtLegal(LType, ExtVT, LoadVT))
452 // If the cast is marked as legal (or promote) then assume low cost.
453 if (SrcLT.first == DstLT.first &&
454 TLI->isOperationLegalOrPromote(ISD, DstLT.second))
457 // Handle scalar conversions.
458 if (!Src->isVectorTy() && !Dst->isVectorTy()) {
460 // Scalar bitcasts are usually free.
461 if (Opcode == Instruction::BitCast)
464 // Just check the op cost. If the operation is legal then assume it costs
466 if (!TLI->isOperationExpand(ISD, DstLT.second))
469 // Assume that illegal scalar instruction are expensive.
473 // Check vector-to-vector casts.
474 if (Dst->isVectorTy() && Src->isVectorTy()) {
476 // If the cast is between same-sized registers, then the check is simple.
477 if (SrcLT.first == DstLT.first &&
478 SrcLT.second.getSizeInBits() == DstLT.second.getSizeInBits()) {
480 // Assume that Zext is done using AND.
481 if (Opcode == Instruction::ZExt)
484 // Assume that sext is done using SHL and SRA.
485 if (Opcode == Instruction::SExt)
488 // Just check the op cost. If the operation is legal then assume it
490 // 1 and multiply by the type-legalization overhead.
491 if (!TLI->isOperationExpand(ISD, DstLT.second))
492 return SrcLT.first * 1;
495 // If we are legalizing by splitting, query the concrete TTI for the cost
496 // of casting the original vector twice. We also need to factor int the
497 // cost of the split itself. Count that as 1, to be consistent with
498 // TLI->getTypeLegalizationCost().
499 if ((TLI->getTypeAction(Src->getContext(), TLI->getValueType(DL, Src)) ==
500 TargetLowering::TypeSplitVector) ||
501 (TLI->getTypeAction(Dst->getContext(), TLI->getValueType(DL, Dst)) ==
502 TargetLowering::TypeSplitVector)) {
503 Type *SplitDst = VectorType::get(Dst->getVectorElementType(),
504 Dst->getVectorNumElements() / 2);
505 Type *SplitSrc = VectorType::get(Src->getVectorElementType(),
506 Src->getVectorNumElements() / 2);
507 T *TTI = static_cast<T *>(this);
508 return TTI->getVectorSplitCost() +
509 (2 * TTI->getCastInstrCost(Opcode, SplitDst, SplitSrc, I));
512 // In other cases where the source or destination are illegal, assume
513 // the operation will get scalarized.
514 unsigned Num = Dst->getVectorNumElements();
515 unsigned Cost = static_cast<T *>(this)->getCastInstrCost(
516 Opcode, Dst->getScalarType(), Src->getScalarType(), I);
518 // Return the cost of multiple scalar invocation plus the cost of
519 // inserting and extracting the values.
520 return getScalarizationOverhead(Dst, true, true) + Num * Cost;
523 // We already handled vector-to-vector and scalar-to-scalar conversions.
525 // is where we handle bitcast between vectors and scalars. We need to assume
526 // that the conversion is scalarized in one way or another.
527 if (Opcode == Instruction::BitCast)
528 // Illegal bitcasts are done by storing and loading from a stack slot.
529 return (Src->isVectorTy() ? getScalarizationOverhead(Src, false, true)
531 (Dst->isVectorTy() ? getScalarizationOverhead(Dst, true, false)
534 llvm_unreachable("Unhandled cast");
537 unsigned getExtractWithExtendCost(unsigned Opcode, Type *Dst,
538 VectorType *VecTy, unsigned Index) {
539 return static_cast<T *>(this)->getVectorInstrCost(
540 Instruction::ExtractElement, VecTy, Index) +
541 static_cast<T *>(this)->getCastInstrCost(Opcode, Dst,
542 VecTy->getElementType());
545 unsigned getCFInstrCost(unsigned Opcode) {
546 // Branches are assumed to be predicted.
550 unsigned getCmpSelInstrCost(unsigned Opcode, Type *ValTy, Type *CondTy,
551 const Instruction *I) {
552 const TargetLoweringBase *TLI = getTLI();
553 int ISD = TLI->InstructionOpcodeToISD(Opcode);
554 assert(ISD && "Invalid opcode");
556 // Selects on vectors are actually vector selects.
557 if (ISD == ISD::SELECT) {
558 assert(CondTy && "CondTy must exist");
559 if (CondTy->isVectorTy())
562 std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(DL, ValTy);
564 if (!(ValTy->isVectorTy() && !LT.second.isVector()) &&
565 !TLI->isOperationExpand(ISD, LT.second)) {
566 // The operation is legal. Assume it costs 1. Multiply
567 // by the type-legalization overhead.
571 // Otherwise, assume that the cast is scalarized.
572 // TODO: If one of the types get legalized by splitting, handle this
573 // similarly to what getCastInstrCost() does.
574 if (ValTy->isVectorTy()) {
575 unsigned Num = ValTy->getVectorNumElements();
577 CondTy = CondTy->getScalarType();
578 unsigned Cost = static_cast<T *>(this)->getCmpSelInstrCost(
579 Opcode, ValTy->getScalarType(), CondTy, I);
581 // Return the cost of multiple scalar invocation plus the cost of
582 // inserting and extracting the values.
583 return getScalarizationOverhead(ValTy, true, false) + Num * Cost;
586 // Unknown scalar opcode.
590 unsigned getVectorInstrCost(unsigned Opcode, Type *Val, unsigned Index) {
591 std::pair<unsigned, MVT> LT =
592 getTLI()->getTypeLegalizationCost(DL, Val->getScalarType());
597 unsigned getMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
598 unsigned AddressSpace, const Instruction *I = nullptr) {
599 assert(!Src->isVoidTy() && "Invalid type");
600 std::pair<unsigned, MVT> LT = getTLI()->getTypeLegalizationCost(DL, Src);
602 // Assuming that all loads of legal types cost 1.
603 unsigned Cost = LT.first;
605 if (Src->isVectorTy() &&
606 Src->getPrimitiveSizeInBits() < LT.second.getSizeInBits()) {
607 // This is a vector load that legalizes to a larger type than the vector
608 // itself. Unless the corresponding extending load or truncating store is
609 // legal, then this will scalarize.
610 TargetLowering::LegalizeAction LA = TargetLowering::Expand;
611 EVT MemVT = getTLI()->getValueType(DL, Src);
612 if (Opcode == Instruction::Store)
613 LA = getTLI()->getTruncStoreAction(LT.second, MemVT);
615 LA = getTLI()->getLoadExtAction(ISD::EXTLOAD, LT.second, MemVT);
617 if (LA != TargetLowering::Legal && LA != TargetLowering::Custom) {
618 // This is a vector load/store for some illegal type that is scalarized.
619 // We must account for the cost of building or decomposing the vector.
620 Cost += getScalarizationOverhead(Src, Opcode != Instruction::Store,
621 Opcode == Instruction::Store);
628 unsigned getInterleavedMemoryOpCost(unsigned Opcode, Type *VecTy,
630 ArrayRef<unsigned> Indices,
632 unsigned AddressSpace) {
633 VectorType *VT = dyn_cast<VectorType>(VecTy);
634 assert(VT && "Expect a vector type for interleaved memory op");
636 unsigned NumElts = VT->getNumElements();
637 assert(Factor > 1 && NumElts % Factor == 0 && "Invalid interleave factor");
639 unsigned NumSubElts = NumElts / Factor;
640 VectorType *SubVT = VectorType::get(VT->getElementType(), NumSubElts);
642 // Firstly, the cost of load/store operation.
643 unsigned Cost = static_cast<T *>(this)->getMemoryOpCost(
644 Opcode, VecTy, Alignment, AddressSpace);
646 // Legalize the vector type, and get the legalized and unlegalized type
648 MVT VecTyLT = getTLI()->getTypeLegalizationCost(DL, VecTy).second;
650 static_cast<T *>(this)->getDataLayout().getTypeStoreSize(VecTy);
651 unsigned VecTyLTSize = VecTyLT.getStoreSize();
653 // Return the ceiling of dividing A by B.
654 auto ceil = [](unsigned A, unsigned B) { return (A + B - 1) / B; };
656 // Scale the cost of the memory operation by the fraction of legalized
657 // instructions that will actually be used. We shouldn't account for the
658 // cost of dead instructions since they will be removed.
660 // E.g., An interleaved load of factor 8:
661 // %vec = load <16 x i64>, <16 x i64>* %ptr
662 // %v0 = shufflevector %vec, undef, <0, 8>
664 // If <16 x i64> is legalized to 8 v2i64 loads, only 2 of the loads will be
665 // used (those corresponding to elements [0:1] and [8:9] of the unlegalized
666 // type). The other loads are unused.
668 // We only scale the cost of loads since interleaved store groups aren't
669 // allowed to have gaps.
670 if (Opcode == Instruction::Load && VecTySize > VecTyLTSize) {
672 // The number of loads of a legal type it will take to represent a load
673 // of the unlegalized vector type.
674 unsigned NumLegalInsts = ceil(VecTySize, VecTyLTSize);
676 // The number of elements of the unlegalized type that correspond to a
677 // single legal instruction.
678 unsigned NumEltsPerLegalInst = ceil(NumElts, NumLegalInsts);
680 // Determine which legal instructions will be used.
681 BitVector UsedInsts(NumLegalInsts, false);
682 for (unsigned Index : Indices)
683 for (unsigned Elt = 0; Elt < NumSubElts; ++Elt)
684 UsedInsts.set((Index + Elt * Factor) / NumEltsPerLegalInst);
686 // Scale the cost of the load by the fraction of legal instructions that
688 Cost *= UsedInsts.count() / NumLegalInsts;
691 // Then plus the cost of interleave operation.
692 if (Opcode == Instruction::Load) {
693 // The interleave cost is similar to extract sub vectors' elements
694 // from the wide vector, and insert them into sub vectors.
696 // E.g. An interleaved load of factor 2 (with one member of index 0):
697 // %vec = load <8 x i32>, <8 x i32>* %ptr
698 // %v0 = shuffle %vec, undef, <0, 2, 4, 6> ; Index 0
699 // The cost is estimated as extract elements at 0, 2, 4, 6 from the
700 // <8 x i32> vector and insert them into a <4 x i32> vector.
702 assert(Indices.size() <= Factor &&
703 "Interleaved memory op has too many members");
705 for (unsigned Index : Indices) {
706 assert(Index < Factor && "Invalid index for interleaved memory op");
708 // Extract elements from loaded vector for each sub vector.
709 for (unsigned i = 0; i < NumSubElts; i++)
710 Cost += static_cast<T *>(this)->getVectorInstrCost(
711 Instruction::ExtractElement, VT, Index + i * Factor);
714 unsigned InsSubCost = 0;
715 for (unsigned i = 0; i < NumSubElts; i++)
716 InsSubCost += static_cast<T *>(this)->getVectorInstrCost(
717 Instruction::InsertElement, SubVT, i);
719 Cost += Indices.size() * InsSubCost;
721 // The interleave cost is extract all elements from sub vectors, and
722 // insert them into the wide vector.
724 // E.g. An interleaved store of factor 2:
725 // %v0_v1 = shuffle %v0, %v1, <0, 4, 1, 5, 2, 6, 3, 7>
726 // store <8 x i32> %interleaved.vec, <8 x i32>* %ptr
727 // The cost is estimated as extract all elements from both <4 x i32>
728 // vectors and insert into the <8 x i32> vector.
730 unsigned ExtSubCost = 0;
731 for (unsigned i = 0; i < NumSubElts; i++)
732 ExtSubCost += static_cast<T *>(this)->getVectorInstrCost(
733 Instruction::ExtractElement, SubVT, i);
734 Cost += ExtSubCost * Factor;
736 for (unsigned i = 0; i < NumElts; i++)
737 Cost += static_cast<T *>(this)
738 ->getVectorInstrCost(Instruction::InsertElement, VT, i);
744 /// Get intrinsic cost based on arguments.
745 unsigned getIntrinsicInstrCost(Intrinsic::ID IID, Type *RetTy,
746 ArrayRef<Value *> Args, FastMathFlags FMF,
748 unsigned RetVF = (RetTy->isVectorTy() ? RetTy->getVectorNumElements() : 1);
749 assert ((RetVF == 1 || VF == 1) && "VF > 1 and RetVF is a vector type");
753 // Assume that we need to scalarize this intrinsic.
754 SmallVector<Type *, 4> Types;
755 for (Value *Op : Args) {
756 Type *OpTy = Op->getType();
757 assert (VF == 1 || !OpTy->isVectorTy());
758 Types.push_back(VF == 1 ? OpTy : VectorType::get(OpTy, VF));
761 if (VF > 1 && !RetTy->isVoidTy())
762 RetTy = VectorType::get(RetTy, VF);
764 // Compute the scalarization overhead based on Args for a vector
765 // intrinsic. A vectorizer will pass a scalar RetTy and VF > 1, while
766 // CostModel will pass a vector RetTy and VF is 1.
767 unsigned ScalarizationCost = UINT_MAX;
768 if (RetVF > 1 || VF > 1) {
769 ScalarizationCost = 0;
770 if (!RetTy->isVoidTy())
771 ScalarizationCost += getScalarizationOverhead(RetTy, true, false);
772 ScalarizationCost += getOperandsScalarizationOverhead(Args, VF);
775 return static_cast<T *>(this)->
776 getIntrinsicInstrCost(IID, RetTy, Types, FMF, ScalarizationCost);
778 case Intrinsic::masked_scatter: {
779 assert (VF == 1 && "Can't vectorize types here.");
780 Value *Mask = Args[3];
781 bool VarMask = !isa<Constant>(Mask);
782 unsigned Alignment = cast<ConstantInt>(Args[2])->getZExtValue();
784 static_cast<T *>(this)->getGatherScatterOpCost(Instruction::Store,
789 case Intrinsic::masked_gather: {
790 assert (VF == 1 && "Can't vectorize types here.");
791 Value *Mask = Args[2];
792 bool VarMask = !isa<Constant>(Mask);
793 unsigned Alignment = cast<ConstantInt>(Args[1])->getZExtValue();
795 static_cast<T *>(this)->getGatherScatterOpCost(Instruction::Load,
796 RetTy, Args[0], VarMask,
802 /// Get intrinsic cost based on argument types.
803 /// If ScalarizationCostPassed is UINT_MAX, the cost of scalarizing the
804 /// arguments and the return value will be computed based on types.
805 unsigned getIntrinsicInstrCost(Intrinsic::ID IID, Type *RetTy,
806 ArrayRef<Type *> Tys, FastMathFlags FMF,
807 unsigned ScalarizationCostPassed = UINT_MAX) {
808 SmallVector<unsigned, 2> ISDs;
809 unsigned SingleCallCost = 10; // Library call cost. Make it expensive.
812 // Assume that we need to scalarize this intrinsic.
813 unsigned ScalarizationCost = ScalarizationCostPassed;
814 unsigned ScalarCalls = 1;
815 Type *ScalarRetTy = RetTy;
816 if (RetTy->isVectorTy()) {
817 if (ScalarizationCostPassed == UINT_MAX)
818 ScalarizationCost = getScalarizationOverhead(RetTy, true, false);
819 ScalarCalls = std::max(ScalarCalls, RetTy->getVectorNumElements());
820 ScalarRetTy = RetTy->getScalarType();
822 SmallVector<Type *, 4> ScalarTys;
823 for (unsigned i = 0, ie = Tys.size(); i != ie; ++i) {
825 if (Ty->isVectorTy()) {
826 if (ScalarizationCostPassed == UINT_MAX)
827 ScalarizationCost += getScalarizationOverhead(Ty, false, true);
828 ScalarCalls = std::max(ScalarCalls, Ty->getVectorNumElements());
829 Ty = Ty->getScalarType();
831 ScalarTys.push_back(Ty);
833 if (ScalarCalls == 1)
834 return 1; // Return cost of a scalar intrinsic. Assume it to be cheap.
836 unsigned ScalarCost = static_cast<T *>(this)->getIntrinsicInstrCost(
837 IID, ScalarRetTy, ScalarTys, FMF);
839 return ScalarCalls * ScalarCost + ScalarizationCost;
841 // Look for intrinsics that can be lowered directly or turned into a scalar
843 case Intrinsic::sqrt:
844 ISDs.push_back(ISD::FSQRT);
847 ISDs.push_back(ISD::FSIN);
850 ISDs.push_back(ISD::FCOS);
853 ISDs.push_back(ISD::FEXP);
855 case Intrinsic::exp2:
856 ISDs.push_back(ISD::FEXP2);
859 ISDs.push_back(ISD::FLOG);
861 case Intrinsic::log10:
862 ISDs.push_back(ISD::FLOG10);
864 case Intrinsic::log2:
865 ISDs.push_back(ISD::FLOG2);
867 case Intrinsic::fabs:
868 ISDs.push_back(ISD::FABS);
870 case Intrinsic::minnum:
871 ISDs.push_back(ISD::FMINNUM);
873 ISDs.push_back(ISD::FMINNAN);
875 case Intrinsic::maxnum:
876 ISDs.push_back(ISD::FMAXNUM);
878 ISDs.push_back(ISD::FMAXNAN);
880 case Intrinsic::copysign:
881 ISDs.push_back(ISD::FCOPYSIGN);
883 case Intrinsic::floor:
884 ISDs.push_back(ISD::FFLOOR);
886 case Intrinsic::ceil:
887 ISDs.push_back(ISD::FCEIL);
889 case Intrinsic::trunc:
890 ISDs.push_back(ISD::FTRUNC);
892 case Intrinsic::nearbyint:
893 ISDs.push_back(ISD::FNEARBYINT);
895 case Intrinsic::rint:
896 ISDs.push_back(ISD::FRINT);
898 case Intrinsic::round:
899 ISDs.push_back(ISD::FROUND);
902 ISDs.push_back(ISD::FPOW);
905 ISDs.push_back(ISD::FMA);
907 case Intrinsic::fmuladd:
908 ISDs.push_back(ISD::FMA);
910 // FIXME: We should return 0 whenever getIntrinsicCost == TCC_Free.
911 case Intrinsic::lifetime_start:
912 case Intrinsic::lifetime_end:
914 case Intrinsic::masked_store:
915 return static_cast<T *>(this)
916 ->getMaskedMemoryOpCost(Instruction::Store, Tys[0], 0, 0);
917 case Intrinsic::masked_load:
918 return static_cast<T *>(this)
919 ->getMaskedMemoryOpCost(Instruction::Load, RetTy, 0, 0);
920 case Intrinsic::ctpop:
921 ISDs.push_back(ISD::CTPOP);
922 // In case of legalization use TCC_Expensive. This is cheaper than a
923 // library call but still not a cheap instruction.
924 SingleCallCost = TargetTransformInfo::TCC_Expensive;
926 // FIXME: ctlz, cttz, ...
929 const TargetLoweringBase *TLI = getTLI();
930 std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(DL, RetTy);
932 SmallVector<unsigned, 2> LegalCost;
933 SmallVector<unsigned, 2> CustomCost;
934 for (unsigned ISD : ISDs) {
935 if (TLI->isOperationLegalOrPromote(ISD, LT.second)) {
936 if (IID == Intrinsic::fabs && TLI->isFAbsFree(LT.second)) {
940 // The operation is legal. Assume it costs 1.
941 // If the type is split to multiple registers, assume that there is some
943 // TODO: Once we have extract/insert subvector cost we need to use them.
945 LegalCost.push_back(LT.first * 2);
947 LegalCost.push_back(LT.first * 1);
948 } else if (!TLI->isOperationExpand(ISD, LT.second)) {
949 // If the operation is custom lowered then assume
950 // that the code is twice as expensive.
951 CustomCost.push_back(LT.first * 2);
955 auto MinLegalCostI = std::min_element(LegalCost.begin(), LegalCost.end());
956 if (MinLegalCostI != LegalCost.end())
957 return *MinLegalCostI;
959 auto MinCustomCostI = std::min_element(CustomCost.begin(), CustomCost.end());
960 if (MinCustomCostI != CustomCost.end())
961 return *MinCustomCostI;
963 // If we can't lower fmuladd into an FMA estimate the cost as a floating
964 // point mul followed by an add.
965 if (IID == Intrinsic::fmuladd)
966 return static_cast<T *>(this)
967 ->getArithmeticInstrCost(BinaryOperator::FMul, RetTy) +
968 static_cast<T *>(this)
969 ->getArithmeticInstrCost(BinaryOperator::FAdd, RetTy);
971 // Else, assume that we need to scalarize this intrinsic. For math builtins
972 // this will emit a costly libcall, adding call overhead and spills. Make it
974 if (RetTy->isVectorTy()) {
975 unsigned ScalarizationCost = ((ScalarizationCostPassed != UINT_MAX) ?
976 ScalarizationCostPassed : getScalarizationOverhead(RetTy, true, false));
977 unsigned ScalarCalls = RetTy->getVectorNumElements();
978 SmallVector<Type *, 4> ScalarTys;
979 for (unsigned i = 0, ie = Tys.size(); i != ie; ++i) {
981 if (Ty->isVectorTy())
982 Ty = Ty->getScalarType();
983 ScalarTys.push_back(Ty);
985 unsigned ScalarCost = static_cast<T *>(this)->getIntrinsicInstrCost(
986 IID, RetTy->getScalarType(), ScalarTys, FMF);
987 for (unsigned i = 0, ie = Tys.size(); i != ie; ++i) {
988 if (Tys[i]->isVectorTy()) {
989 if (ScalarizationCostPassed == UINT_MAX)
990 ScalarizationCost += getScalarizationOverhead(Tys[i], false, true);
991 ScalarCalls = std::max(ScalarCalls, Tys[i]->getVectorNumElements());
995 return ScalarCalls * ScalarCost + ScalarizationCost;
998 // This is going to be turned into a library call, make it expensive.
999 return SingleCallCost;
1002 /// \brief Compute a cost of the given call instruction.
1004 /// Compute the cost of calling function F with return type RetTy and
1005 /// argument types Tys. F might be nullptr, in this case the cost of an
1006 /// arbitrary call with the specified signature will be returned.
1007 /// This is used, for instance, when we estimate call of a vector
1008 /// counterpart of the given function.
1009 /// \param F Called function, might be nullptr.
1010 /// \param RetTy Return value types.
1011 /// \param Tys Argument types.
1012 /// \returns The cost of Call instruction.
1013 unsigned getCallInstrCost(Function *F, Type *RetTy, ArrayRef<Type *> Tys) {
1017 unsigned getNumberOfParts(Type *Tp) {
1018 std::pair<unsigned, MVT> LT = getTLI()->getTypeLegalizationCost(DL, Tp);
1022 unsigned getAddressComputationCost(Type *Ty, ScalarEvolution *,
1027 unsigned getReductionCost(unsigned Opcode, Type *Ty, bool IsPairwise) {
1028 assert(Ty->isVectorTy() && "Expect a vector type");
1029 Type *ScalarTy = Ty->getVectorElementType();
1030 unsigned NumVecElts = Ty->getVectorNumElements();
1031 unsigned NumReduxLevels = Log2_32(NumVecElts);
1032 // Try to calculate arithmetic and shuffle op costs for reduction operations.
1033 // We're assuming that reduction operation are performing the following way:
1034 // 1. Non-pairwise reduction
1035 // %val1 = shufflevector<n x t> %val, <n x t> %undef,
1036 // <n x i32> <i32 n/2, i32 n/2 + 1, ..., i32 n, i32 undef, ..., i32 undef>
1037 // \----------------v-------------/ \----------v------------/
1038 // n/2 elements n/2 elements
1039 // %red1 = op <n x t> %val, <n x t> val1
1040 // After this operation we have a vector %red1 with only maningfull the
1041 // first n/2 elements, the second n/2 elements are undefined and can be
1042 // dropped. All other operations are actually working with the vector of
1043 // length n/2, not n. though the real vector length is still n.
1044 // %val2 = shufflevector<n x t> %red1, <n x t> %undef,
1045 // <n x i32> <i32 n/4, i32 n/4 + 1, ..., i32 n/2, i32 undef, ..., i32 undef>
1046 // \----------------v-------------/ \----------v------------/
1047 // n/4 elements 3*n/4 elements
1048 // %red2 = op <n x t> %red1, <n x t> val2 - working with the vector of
1049 // length n/2, the resulting vector has length n/4 etc.
1050 // 2. Pairwise reduction:
1051 // Everything is the same except for an additional shuffle operation which
1052 // is used to produce operands for pairwise kind of reductions.
1053 // %val1 = shufflevector<n x t> %val, <n x t> %undef,
1054 // <n x i32> <i32 0, i32 2, ..., i32 n-2, i32 undef, ..., i32 undef>
1055 // \-------------v----------/ \----------v------------/
1056 // n/2 elements n/2 elements
1057 // %val2 = shufflevector<n x t> %val, <n x t> %undef,
1058 // <n x i32> <i32 1, i32 3, ..., i32 n-1, i32 undef, ..., i32 undef>
1059 // \-------------v----------/ \----------v------------/
1060 // n/2 elements n/2 elements
1061 // %red1 = op <n x t> %val1, <n x t> val2
1062 // Again, the operation is performed on <n x t> vector, but the resulting
1063 // vector %red1 is <n/2 x t> vector.
1065 // The cost model should take into account that the actual length of the
1066 // vector is reduced on each iteration.
1067 unsigned ArithCost = 0;
1068 unsigned ShuffleCost = 0;
1069 auto *ConcreteTTI = static_cast<T *>(this);
1070 std::pair<unsigned, MVT> LT =
1071 ConcreteTTI->getTLI()->getTypeLegalizationCost(DL, Ty);
1072 unsigned LongVectorCount = 0;
1074 LT.second.isVector() ? LT.second.getVectorNumElements() : 1;
1075 while (NumVecElts > MVTLen) {
1077 // Assume the pairwise shuffles add a cost.
1078 ShuffleCost += (IsPairwise + 1) *
1079 ConcreteTTI->getShuffleCost(TTI::SK_ExtractSubvector, Ty,
1081 ArithCost += ConcreteTTI->getArithmeticInstrCost(Opcode, Ty);
1082 Ty = VectorType::get(ScalarTy, NumVecElts);
1085 // The minimal length of the vector is limited by the real length of vector
1086 // operations performed on the current platform. That's why several final
1087 // reduction opertions are perfomed on the vectors with the same
1088 // architecture-dependent length.
1089 ShuffleCost += (NumReduxLevels - LongVectorCount) * (IsPairwise + 1) *
1090 ConcreteTTI->getShuffleCost(TTI::SK_ExtractSubvector, Ty,
1092 ArithCost += (NumReduxLevels - LongVectorCount) *
1093 ConcreteTTI->getArithmeticInstrCost(Opcode, Ty);
1094 return ShuffleCost + ArithCost + getScalarizationOverhead(Ty, false, true);
1097 unsigned getVectorSplitCost() { return 1; }
1102 /// \brief Concrete BasicTTIImpl that can be used if no further customization
1104 class BasicTTIImpl : public BasicTTIImplBase<BasicTTIImpl> {
1105 typedef BasicTTIImplBase<BasicTTIImpl> BaseT;
1106 friend class BasicTTIImplBase<BasicTTIImpl>;
1108 const TargetSubtargetInfo *ST;
1109 const TargetLoweringBase *TLI;
1111 const TargetSubtargetInfo *getST() const { return ST; }
1112 const TargetLoweringBase *getTLI() const { return TLI; }
1115 explicit BasicTTIImpl(const TargetMachine *ST, const Function &F);