1 //===- BasicTTIImpl.h -------------------------------------------*- C++ -*-===//
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
7 //===----------------------------------------------------------------------===//
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/ADT/APInt.h"
20 #include "llvm/ADT/ArrayRef.h"
21 #include "llvm/ADT/BitVector.h"
22 #include "llvm/ADT/SmallPtrSet.h"
23 #include "llvm/ADT/SmallVector.h"
24 #include "llvm/Analysis/LoopInfo.h"
25 #include "llvm/Analysis/TargetTransformInfo.h"
26 #include "llvm/Analysis/TargetTransformInfoImpl.h"
27 #include "llvm/CodeGen/ISDOpcodes.h"
28 #include "llvm/CodeGen/TargetLowering.h"
29 #include "llvm/CodeGen/TargetSubtargetInfo.h"
30 #include "llvm/CodeGen/ValueTypes.h"
31 #include "llvm/IR/BasicBlock.h"
32 #include "llvm/IR/CallSite.h"
33 #include "llvm/IR/Constant.h"
34 #include "llvm/IR/Constants.h"
35 #include "llvm/IR/DataLayout.h"
36 #include "llvm/IR/DerivedTypes.h"
37 #include "llvm/IR/InstrTypes.h"
38 #include "llvm/IR/Instruction.h"
39 #include "llvm/IR/Instructions.h"
40 #include "llvm/IR/Intrinsics.h"
41 #include "llvm/IR/Operator.h"
42 #include "llvm/IR/Type.h"
43 #include "llvm/IR/Value.h"
44 #include "llvm/MC/MCSchedule.h"
45 #include "llvm/Support/Casting.h"
46 #include "llvm/Support/CommandLine.h"
47 #include "llvm/Support/ErrorHandling.h"
48 #include "llvm/Support/MachineValueType.h"
49 #include "llvm/Support/MathExtras.h"
61 class ScalarEvolution;
65 extern cl::opt<unsigned> PartialUnrollingThreshold;
67 /// Base class which can be used to help build a TTI implementation.
69 /// This class provides as much implementation of the TTI interface as is
70 /// possible using the target independent parts of the code generator.
72 /// In order to subclass it, your class must implement a getST() method to
73 /// return the subtarget, and a getTLI() method to return the target lowering.
74 /// We need these methods implemented in the derived class so that this class
75 /// doesn't have to duplicate storage for them.
77 class BasicTTIImplBase : public TargetTransformInfoImplCRTPBase<T> {
79 using BaseT = TargetTransformInfoImplCRTPBase<T>;
80 using TTI = TargetTransformInfo;
82 /// Estimate a cost of Broadcast as an extract and sequence of insert
84 unsigned getBroadcastShuffleOverhead(Type *Ty) {
85 assert(Ty->isVectorTy() && "Can only shuffle vectors");
87 // Broadcast cost is equal to the cost of extracting the zero'th element
88 // plus the cost of inserting it into every element of the result vector.
89 Cost += static_cast<T *>(this)->getVectorInstrCost(
90 Instruction::ExtractElement, Ty, 0);
92 for (int i = 0, e = Ty->getVectorNumElements(); i < e; ++i) {
93 Cost += static_cast<T *>(this)->getVectorInstrCost(
94 Instruction::InsertElement, Ty, i);
99 /// Estimate a cost of shuffle as a sequence of extract and insert
101 unsigned getPermuteShuffleOverhead(Type *Ty) {
102 assert(Ty->isVectorTy() && "Can only shuffle vectors");
104 // Shuffle cost is equal to the cost of extracting element from its argument
105 // plus the cost of inserting them onto the result vector.
107 // e.g. <4 x float> has a mask of <0,5,2,7> i.e we need to extract from
108 // index 0 of first vector, index 1 of second vector,index 2 of first
109 // vector and finally index 3 of second vector and insert them at index
110 // <0,1,2,3> of result vector.
111 for (int i = 0, e = Ty->getVectorNumElements(); i < e; ++i) {
112 Cost += static_cast<T *>(this)
113 ->getVectorInstrCost(Instruction::InsertElement, Ty, i);
114 Cost += static_cast<T *>(this)
115 ->getVectorInstrCost(Instruction::ExtractElement, Ty, i);
120 /// Estimate a cost of subvector extraction as a sequence of extract and
121 /// insert operations.
122 unsigned getExtractSubvectorOverhead(Type *Ty, int Index, Type *SubTy) {
123 assert(Ty && Ty->isVectorTy() && SubTy && SubTy->isVectorTy() &&
124 "Can only extract subvectors from vectors");
125 int NumSubElts = SubTy->getVectorNumElements();
126 assert((Index + NumSubElts) <= (int)Ty->getVectorNumElements() &&
127 "SK_ExtractSubvector index out of range");
130 // Subvector extraction cost is equal to the cost of extracting element from
131 // the source type plus the cost of inserting them into the result vector
133 for (int i = 0; i != NumSubElts; ++i) {
134 Cost += static_cast<T *>(this)->getVectorInstrCost(
135 Instruction::ExtractElement, Ty, i + Index);
136 Cost += static_cast<T *>(this)->getVectorInstrCost(
137 Instruction::InsertElement, SubTy, i);
142 /// Estimate a cost of subvector insertion as a sequence of extract and
143 /// insert operations.
144 unsigned getInsertSubvectorOverhead(Type *Ty, int Index, Type *SubTy) {
145 assert(Ty && Ty->isVectorTy() && SubTy && SubTy->isVectorTy() &&
146 "Can only insert subvectors into vectors");
147 int NumSubElts = SubTy->getVectorNumElements();
148 assert((Index + NumSubElts) <= (int)Ty->getVectorNumElements() &&
149 "SK_InsertSubvector index out of range");
152 // Subvector insertion cost is equal to the cost of extracting element from
153 // the source type plus the cost of inserting them into the result vector
155 for (int i = 0; i != NumSubElts; ++i) {
156 Cost += static_cast<T *>(this)->getVectorInstrCost(
157 Instruction::ExtractElement, SubTy, i);
158 Cost += static_cast<T *>(this)->getVectorInstrCost(
159 Instruction::InsertElement, Ty, i + Index);
164 /// Local query method delegates up to T which *must* implement this!
165 const TargetSubtargetInfo *getST() const {
166 return static_cast<const T *>(this)->getST();
169 /// Local query method delegates up to T which *must* implement this!
170 const TargetLoweringBase *getTLI() const {
171 return static_cast<const T *>(this)->getTLI();
174 static ISD::MemIndexedMode getISDIndexedMode(TTI::MemIndexedMode M) {
176 case TTI::MIM_Unindexed:
177 return ISD::UNINDEXED;
178 case TTI::MIM_PreInc:
180 case TTI::MIM_PreDec:
182 case TTI::MIM_PostInc:
183 return ISD::POST_INC;
184 case TTI::MIM_PostDec:
185 return ISD::POST_DEC;
187 llvm_unreachable("Unexpected MemIndexedMode");
191 explicit BasicTTIImplBase(const TargetMachine *TM, const DataLayout &DL)
194 using TargetTransformInfoImplBase::DL;
197 /// \name Scalar TTI Implementations
199 bool allowsMisalignedMemoryAccesses(LLVMContext &Context, unsigned BitWidth,
200 unsigned AddressSpace, unsigned Alignment,
202 EVT E = EVT::getIntegerVT(Context, BitWidth);
203 return getTLI()->allowsMisalignedMemoryAccesses(
204 E, AddressSpace, Alignment, MachineMemOperand::MONone, Fast);
207 bool hasBranchDivergence() { return false; }
209 bool isSourceOfDivergence(const Value *V) { return false; }
211 bool isAlwaysUniform(const Value *V) { return false; }
213 unsigned getFlatAddressSpace() {
214 // Return an invalid address space.
218 bool isLegalAddImmediate(int64_t imm) {
219 return getTLI()->isLegalAddImmediate(imm);
222 bool isLegalICmpImmediate(int64_t imm) {
223 return getTLI()->isLegalICmpImmediate(imm);
226 bool isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset,
227 bool HasBaseReg, int64_t Scale,
228 unsigned AddrSpace, Instruction *I = nullptr) {
229 TargetLoweringBase::AddrMode AM;
231 AM.BaseOffs = BaseOffset;
232 AM.HasBaseReg = HasBaseReg;
234 return getTLI()->isLegalAddressingMode(DL, AM, Ty, AddrSpace, I);
237 bool isIndexedLoadLegal(TTI::MemIndexedMode M, Type *Ty,
238 const DataLayout &DL) const {
239 EVT VT = getTLI()->getValueType(DL, Ty);
240 return getTLI()->isIndexedLoadLegal(getISDIndexedMode(M), VT);
243 bool isIndexedStoreLegal(TTI::MemIndexedMode M, Type *Ty,
244 const DataLayout &DL) const {
245 EVT VT = getTLI()->getValueType(DL, Ty);
246 return getTLI()->isIndexedStoreLegal(getISDIndexedMode(M), VT);
249 bool isLSRCostLess(TTI::LSRCost C1, TTI::LSRCost C2) {
250 return TargetTransformInfoImplBase::isLSRCostLess(C1, C2);
253 int getScalingFactorCost(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset,
254 bool HasBaseReg, int64_t Scale, unsigned AddrSpace) {
255 TargetLoweringBase::AddrMode AM;
257 AM.BaseOffs = BaseOffset;
258 AM.HasBaseReg = HasBaseReg;
260 return getTLI()->getScalingFactorCost(DL, AM, Ty, AddrSpace);
263 bool isTruncateFree(Type *Ty1, Type *Ty2) {
264 return getTLI()->isTruncateFree(Ty1, Ty2);
267 bool isProfitableToHoist(Instruction *I) {
268 return getTLI()->isProfitableToHoist(I);
271 bool useAA() const { return getST()->useAA(); }
273 bool isTypeLegal(Type *Ty) {
274 EVT VT = getTLI()->getValueType(DL, Ty);
275 return getTLI()->isTypeLegal(VT);
278 int getGEPCost(Type *PointeeType, const Value *Ptr,
279 ArrayRef<const Value *> Operands) {
280 return BaseT::getGEPCost(PointeeType, Ptr, Operands);
283 int getExtCost(const Instruction *I, const Value *Src) {
284 if (getTLI()->isExtFree(I))
285 return TargetTransformInfo::TCC_Free;
287 if (isa<ZExtInst>(I) || isa<SExtInst>(I))
288 if (const LoadInst *LI = dyn_cast<LoadInst>(Src))
289 if (getTLI()->isExtLoad(LI, I, DL))
290 return TargetTransformInfo::TCC_Free;
292 return TargetTransformInfo::TCC_Basic;
295 unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
296 ArrayRef<const Value *> Arguments, const User *U) {
297 return BaseT::getIntrinsicCost(IID, RetTy, Arguments, U);
300 unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
301 ArrayRef<Type *> ParamTys, const User *U) {
302 if (IID == Intrinsic::cttz) {
303 if (getTLI()->isCheapToSpeculateCttz())
304 return TargetTransformInfo::TCC_Basic;
305 return TargetTransformInfo::TCC_Expensive;
308 if (IID == Intrinsic::ctlz) {
309 if (getTLI()->isCheapToSpeculateCtlz())
310 return TargetTransformInfo::TCC_Basic;
311 return TargetTransformInfo::TCC_Expensive;
314 return BaseT::getIntrinsicCost(IID, RetTy, ParamTys, U);
317 unsigned getEstimatedNumberOfCaseClusters(const SwitchInst &SI,
318 unsigned &JumpTableSize) {
319 /// Try to find the estimated number of clusters. Note that the number of
320 /// clusters identified in this function could be different from the actural
321 /// numbers found in lowering. This function ignore switches that are
322 /// lowered with a mix of jump table / bit test / BTree. This function was
323 /// initially intended to be used when estimating the cost of switch in
324 /// inline cost heuristic, but it's a generic cost model to be used in other
325 /// places (e.g., in loop unrolling).
326 unsigned N = SI.getNumCases();
327 const TargetLoweringBase *TLI = getTLI();
328 const DataLayout &DL = this->getDataLayout();
331 bool IsJTAllowed = TLI->areJTsAllowed(SI.getParent()->getParent());
333 // Early exit if both a jump table and bit test are not allowed.
334 if (N < 1 || (!IsJTAllowed && DL.getIndexSizeInBits(0u) < N))
337 APInt MaxCaseVal = SI.case_begin()->getCaseValue()->getValue();
338 APInt MinCaseVal = MaxCaseVal;
339 for (auto CI : SI.cases()) {
340 const APInt &CaseVal = CI.getCaseValue()->getValue();
341 if (CaseVal.sgt(MaxCaseVal))
342 MaxCaseVal = CaseVal;
343 if (CaseVal.slt(MinCaseVal))
344 MinCaseVal = CaseVal;
347 // Check if suitable for a bit test
348 if (N <= DL.getIndexSizeInBits(0u)) {
349 SmallPtrSet<const BasicBlock *, 4> Dests;
350 for (auto I : SI.cases())
351 Dests.insert(I.getCaseSuccessor());
353 if (TLI->isSuitableForBitTests(Dests.size(), N, MinCaseVal, MaxCaseVal,
358 // Check if suitable for a jump table.
360 if (N < 2 || N < TLI->getMinimumJumpTableEntries())
363 (MaxCaseVal - MinCaseVal)
364 .getLimitedValue(std::numeric_limits<uint64_t>::max() - 1) + 1;
365 // Check whether a range of clusters is dense enough for a jump table
366 if (TLI->isSuitableForJumpTable(&SI, N, Range)) {
367 JumpTableSize = Range;
374 unsigned getJumpBufAlignment() { return getTLI()->getJumpBufAlignment(); }
376 unsigned getJumpBufSize() { return getTLI()->getJumpBufSize(); }
378 bool shouldBuildLookupTables() {
379 const TargetLoweringBase *TLI = getTLI();
380 return TLI->isOperationLegalOrCustom(ISD::BR_JT, MVT::Other) ||
381 TLI->isOperationLegalOrCustom(ISD::BRIND, MVT::Other);
384 bool haveFastSqrt(Type *Ty) {
385 const TargetLoweringBase *TLI = getTLI();
386 EVT VT = TLI->getValueType(DL, Ty);
387 return TLI->isTypeLegal(VT) &&
388 TLI->isOperationLegalOrCustom(ISD::FSQRT, VT);
391 bool isFCmpOrdCheaperThanFCmpZero(Type *Ty) {
395 unsigned getFPOpCost(Type *Ty) {
396 // Check whether FADD is available, as a proxy for floating-point in
398 const TargetLoweringBase *TLI = getTLI();
399 EVT VT = TLI->getValueType(DL, Ty);
400 if (TLI->isOperationLegalOrCustomOrPromote(ISD::FADD, VT))
401 return TargetTransformInfo::TCC_Basic;
402 return TargetTransformInfo::TCC_Expensive;
405 unsigned getOperationCost(unsigned Opcode, Type *Ty, Type *OpTy) {
406 const TargetLoweringBase *TLI = getTLI();
409 case Instruction::Trunc:
410 if (TLI->isTruncateFree(OpTy, Ty))
411 return TargetTransformInfo::TCC_Free;
412 return TargetTransformInfo::TCC_Basic;
413 case Instruction::ZExt:
414 if (TLI->isZExtFree(OpTy, Ty))
415 return TargetTransformInfo::TCC_Free;
416 return TargetTransformInfo::TCC_Basic;
418 case Instruction::AddrSpaceCast:
419 if (TLI->isFreeAddrSpaceCast(OpTy->getPointerAddressSpace(),
420 Ty->getPointerAddressSpace()))
421 return TargetTransformInfo::TCC_Free;
422 return TargetTransformInfo::TCC_Basic;
425 return BaseT::getOperationCost(Opcode, Ty, OpTy);
428 unsigned getInliningThresholdMultiplier() { return 1; }
430 int getInlinerVectorBonusPercent() { return 150; }
432 void getUnrollingPreferences(Loop *L, ScalarEvolution &SE,
433 TTI::UnrollingPreferences &UP) {
434 // This unrolling functionality is target independent, but to provide some
435 // motivation for its intended use, for x86:
437 // According to the Intel 64 and IA-32 Architectures Optimization Reference
438 // Manual, Intel Core models and later have a loop stream detector (and
439 // associated uop queue) that can benefit from partial unrolling.
440 // The relevant requirements are:
441 // - The loop must have no more than 4 (8 for Nehalem and later) branches
442 // taken, and none of them may be calls.
443 // - The loop can have no more than 18 (28 for Nehalem and later) uops.
445 // According to the Software Optimization Guide for AMD Family 15h
446 // Processors, models 30h-4fh (Steamroller and later) have a loop predictor
447 // and loop buffer which can benefit from partial unrolling.
448 // The relevant requirements are:
449 // - The loop must have fewer than 16 branches
450 // - The loop must have less than 40 uops in all executed loop branches
452 // The number of taken branches in a loop is hard to estimate here, and
453 // benchmarking has revealed that it is better not to be conservative when
454 // estimating the branch count. As a result, we'll ignore the branch limits
455 // until someone finds a case where it matters in practice.
458 const TargetSubtargetInfo *ST = getST();
459 if (PartialUnrollingThreshold.getNumOccurrences() > 0)
460 MaxOps = PartialUnrollingThreshold;
461 else if (ST->getSchedModel().LoopMicroOpBufferSize > 0)
462 MaxOps = ST->getSchedModel().LoopMicroOpBufferSize;
466 // Scan the loop: don't unroll loops with calls.
467 for (Loop::block_iterator I = L->block_begin(), E = L->block_end(); I != E;
471 for (BasicBlock::iterator J = BB->begin(), JE = BB->end(); J != JE; ++J)
472 if (isa<CallInst>(J) || isa<InvokeInst>(J)) {
473 ImmutableCallSite CS(&*J);
474 if (const Function *F = CS.getCalledFunction()) {
475 if (!static_cast<T *>(this)->isLoweredToCall(F))
483 // Enable runtime and partial unrolling up to the specified size.
484 // Enable using trip count upper bound to unroll loops.
485 UP.Partial = UP.Runtime = UP.UpperBound = true;
486 UP.PartialThreshold = MaxOps;
488 // Avoid unrolling when optimizing for size.
489 UP.OptSizeThreshold = 0;
490 UP.PartialOptSizeThreshold = 0;
492 // Set number of instructions optimized when "back edge"
493 // becomes "fall through" to default value of 2.
497 bool isHardwareLoopProfitable(Loop *L, ScalarEvolution &SE,
499 TargetLibraryInfo *LibInfo,
500 HardwareLoopInfo &HWLoopInfo) {
501 return BaseT::isHardwareLoopProfitable(L, SE, AC, LibInfo, HWLoopInfo);
504 int getInstructionLatency(const Instruction *I) {
505 if (isa<LoadInst>(I))
506 return getST()->getSchedModel().DefaultLoadLatency;
508 return BaseT::getInstructionLatency(I);
513 /// \name Vector TTI Implementations
516 unsigned getNumberOfRegisters(bool Vector) { return Vector ? 0 : 1; }
518 unsigned getRegisterBitWidth(bool Vector) const { return 32; }
520 /// Estimate the overhead of scalarizing an instruction. Insert and Extract
521 /// are set if the result needs to be inserted and/or extracted from vectors.
522 unsigned getScalarizationOverhead(Type *Ty, bool Insert, bool Extract) {
523 assert(Ty->isVectorTy() && "Can only scalarize vectors");
526 for (int i = 0, e = Ty->getVectorNumElements(); i < e; ++i) {
528 Cost += static_cast<T *>(this)
529 ->getVectorInstrCost(Instruction::InsertElement, Ty, i);
531 Cost += static_cast<T *>(this)
532 ->getVectorInstrCost(Instruction::ExtractElement, Ty, i);
538 /// Estimate the overhead of scalarizing an instructions unique
539 /// non-constant operands. The types of the arguments are ordinarily
540 /// scalar, in which case the costs are multiplied with VF.
541 unsigned getOperandsScalarizationOverhead(ArrayRef<const Value *> Args,
544 SmallPtrSet<const Value*, 4> UniqueOperands;
545 for (const Value *A : Args) {
546 if (!isa<Constant>(A) && UniqueOperands.insert(A).second) {
547 Type *VecTy = nullptr;
548 if (A->getType()->isVectorTy()) {
549 VecTy = A->getType();
550 // If A is a vector operand, VF should be 1 or correspond to A.
551 assert((VF == 1 || VF == VecTy->getVectorNumElements()) &&
552 "Vector argument does not match VF");
555 VecTy = VectorType::get(A->getType(), VF);
557 Cost += getScalarizationOverhead(VecTy, false, true);
564 unsigned getScalarizationOverhead(Type *VecTy, ArrayRef<const Value *> Args) {
565 assert(VecTy->isVectorTy());
569 Cost += getScalarizationOverhead(VecTy, true, false);
571 Cost += getOperandsScalarizationOverhead(Args,
572 VecTy->getVectorNumElements());
574 // When no information on arguments is provided, we add the cost
575 // associated with one argument as a heuristic.
576 Cost += getScalarizationOverhead(VecTy, false, true);
581 unsigned getMaxInterleaveFactor(unsigned VF) { return 1; }
583 unsigned getArithmeticInstrCost(
584 unsigned Opcode, Type *Ty,
585 TTI::OperandValueKind Opd1Info = TTI::OK_AnyValue,
586 TTI::OperandValueKind Opd2Info = TTI::OK_AnyValue,
587 TTI::OperandValueProperties Opd1PropInfo = TTI::OP_None,
588 TTI::OperandValueProperties Opd2PropInfo = TTI::OP_None,
589 ArrayRef<const Value *> Args = ArrayRef<const Value *>()) {
590 // Check if any of the operands are vector operands.
591 const TargetLoweringBase *TLI = getTLI();
592 int ISD = TLI->InstructionOpcodeToISD(Opcode);
593 assert(ISD && "Invalid opcode");
595 std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(DL, Ty);
597 bool IsFloat = Ty->isFPOrFPVectorTy();
598 // Assume that floating point arithmetic operations cost twice as much as
599 // integer operations.
600 unsigned OpCost = (IsFloat ? 2 : 1);
602 if (TLI->isOperationLegalOrPromote(ISD, LT.second)) {
603 // The operation is legal. Assume it costs 1.
604 // TODO: Once we have extract/insert subvector cost we need to use them.
605 return LT.first * OpCost;
608 if (!TLI->isOperationExpand(ISD, LT.second)) {
609 // If the operation is custom lowered, then assume that the code is twice
611 return LT.first * 2 * OpCost;
614 // Else, assume that we need to scalarize this op.
615 // TODO: If one of the types get legalized by splitting, handle this
616 // similarly to what getCastInstrCost() does.
617 if (Ty->isVectorTy()) {
618 unsigned Num = Ty->getVectorNumElements();
619 unsigned Cost = static_cast<T *>(this)
620 ->getArithmeticInstrCost(Opcode, Ty->getScalarType());
621 // Return the cost of multiple scalar invocation plus the cost of
622 // inserting and extracting the values.
623 return getScalarizationOverhead(Ty, Args) + Num * Cost;
626 // We don't know anything about this scalar instruction.
630 unsigned getShuffleCost(TTI::ShuffleKind Kind, Type *Tp, int Index,
633 case TTI::SK_Broadcast:
634 return getBroadcastShuffleOverhead(Tp);
636 case TTI::SK_Reverse:
637 case TTI::SK_Transpose:
638 case TTI::SK_PermuteSingleSrc:
639 case TTI::SK_PermuteTwoSrc:
640 return getPermuteShuffleOverhead(Tp);
641 case TTI::SK_ExtractSubvector:
642 return getExtractSubvectorOverhead(Tp, Index, SubTp);
643 case TTI::SK_InsertSubvector:
644 return getInsertSubvectorOverhead(Tp, Index, SubTp);
646 llvm_unreachable("Unknown TTI::ShuffleKind");
649 unsigned getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src,
650 const Instruction *I = nullptr) {
651 const TargetLoweringBase *TLI = getTLI();
652 int ISD = TLI->InstructionOpcodeToISD(Opcode);
653 assert(ISD && "Invalid opcode");
654 std::pair<unsigned, MVT> SrcLT = TLI->getTypeLegalizationCost(DL, Src);
655 std::pair<unsigned, MVT> DstLT = TLI->getTypeLegalizationCost(DL, Dst);
657 // Check for NOOP conversions.
658 if (SrcLT.first == DstLT.first &&
659 SrcLT.second.getSizeInBits() == DstLT.second.getSizeInBits()) {
661 // Bitcast between types that are legalized to the same type are free.
662 if (Opcode == Instruction::BitCast || Opcode == Instruction::Trunc)
666 if (Opcode == Instruction::Trunc &&
667 TLI->isTruncateFree(SrcLT.second, DstLT.second))
670 if (Opcode == Instruction::ZExt &&
671 TLI->isZExtFree(SrcLT.second, DstLT.second))
674 if (Opcode == Instruction::AddrSpaceCast &&
675 TLI->isFreeAddrSpaceCast(Src->getPointerAddressSpace(),
676 Dst->getPointerAddressSpace()))
679 // If this is a zext/sext of a load, return 0 if the corresponding
680 // extending load exists on target.
681 if ((Opcode == Instruction::ZExt || Opcode == Instruction::SExt) &&
682 I && isa<LoadInst>(I->getOperand(0))) {
683 EVT ExtVT = EVT::getEVT(Dst);
684 EVT LoadVT = EVT::getEVT(Src);
686 ((Opcode == Instruction::ZExt) ? ISD::ZEXTLOAD : ISD::SEXTLOAD);
687 if (TLI->isLoadExtLegal(LType, ExtVT, LoadVT))
691 // If the cast is marked as legal (or promote) then assume low cost.
692 if (SrcLT.first == DstLT.first &&
693 TLI->isOperationLegalOrPromote(ISD, DstLT.second))
696 // Handle scalar conversions.
697 if (!Src->isVectorTy() && !Dst->isVectorTy()) {
698 // Scalar bitcasts are usually free.
699 if (Opcode == Instruction::BitCast)
702 // Just check the op cost. If the operation is legal then assume it costs
704 if (!TLI->isOperationExpand(ISD, DstLT.second))
707 // Assume that illegal scalar instruction are expensive.
711 // Check vector-to-vector casts.
712 if (Dst->isVectorTy() && Src->isVectorTy()) {
713 // If the cast is between same-sized registers, then the check is simple.
714 if (SrcLT.first == DstLT.first &&
715 SrcLT.second.getSizeInBits() == DstLT.second.getSizeInBits()) {
717 // Assume that Zext is done using AND.
718 if (Opcode == Instruction::ZExt)
721 // Assume that sext is done using SHL and SRA.
722 if (Opcode == Instruction::SExt)
725 // Just check the op cost. If the operation is legal then assume it
727 // 1 and multiply by the type-legalization overhead.
728 if (!TLI->isOperationExpand(ISD, DstLT.second))
729 return SrcLT.first * 1;
732 // If we are legalizing by splitting, query the concrete TTI for the cost
733 // of casting the original vector twice. We also need to factor in the
734 // cost of the split itself. Count that as 1, to be consistent with
735 // TLI->getTypeLegalizationCost().
736 if ((TLI->getTypeAction(Src->getContext(), TLI->getValueType(DL, Src)) ==
737 TargetLowering::TypeSplitVector) ||
738 (TLI->getTypeAction(Dst->getContext(), TLI->getValueType(DL, Dst)) ==
739 TargetLowering::TypeSplitVector)) {
740 Type *SplitDst = VectorType::get(Dst->getVectorElementType(),
741 Dst->getVectorNumElements() / 2);
742 Type *SplitSrc = VectorType::get(Src->getVectorElementType(),
743 Src->getVectorNumElements() / 2);
744 T *TTI = static_cast<T *>(this);
745 return TTI->getVectorSplitCost() +
746 (2 * TTI->getCastInstrCost(Opcode, SplitDst, SplitSrc, I));
749 // In other cases where the source or destination are illegal, assume
750 // the operation will get scalarized.
751 unsigned Num = Dst->getVectorNumElements();
752 unsigned Cost = static_cast<T *>(this)->getCastInstrCost(
753 Opcode, Dst->getScalarType(), Src->getScalarType(), I);
755 // Return the cost of multiple scalar invocation plus the cost of
756 // inserting and extracting the values.
757 return getScalarizationOverhead(Dst, true, true) + Num * Cost;
760 // We already handled vector-to-vector and scalar-to-scalar conversions.
762 // is where we handle bitcast between vectors and scalars. We need to assume
763 // that the conversion is scalarized in one way or another.
764 if (Opcode == Instruction::BitCast)
765 // Illegal bitcasts are done by storing and loading from a stack slot.
766 return (Src->isVectorTy() ? getScalarizationOverhead(Src, false, true)
768 (Dst->isVectorTy() ? getScalarizationOverhead(Dst, true, false)
771 llvm_unreachable("Unhandled cast");
774 unsigned getExtractWithExtendCost(unsigned Opcode, Type *Dst,
775 VectorType *VecTy, unsigned Index) {
776 return static_cast<T *>(this)->getVectorInstrCost(
777 Instruction::ExtractElement, VecTy, Index) +
778 static_cast<T *>(this)->getCastInstrCost(Opcode, Dst,
779 VecTy->getElementType());
782 unsigned getCFInstrCost(unsigned Opcode) {
783 // Branches are assumed to be predicted.
787 unsigned getCmpSelInstrCost(unsigned Opcode, Type *ValTy, Type *CondTy,
788 const Instruction *I) {
789 const TargetLoweringBase *TLI = getTLI();
790 int ISD = TLI->InstructionOpcodeToISD(Opcode);
791 assert(ISD && "Invalid opcode");
793 // Selects on vectors are actually vector selects.
794 if (ISD == ISD::SELECT) {
795 assert(CondTy && "CondTy must exist");
796 if (CondTy->isVectorTy())
799 std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(DL, ValTy);
801 if (!(ValTy->isVectorTy() && !LT.second.isVector()) &&
802 !TLI->isOperationExpand(ISD, LT.second)) {
803 // The operation is legal. Assume it costs 1. Multiply
804 // by the type-legalization overhead.
808 // Otherwise, assume that the cast is scalarized.
809 // TODO: If one of the types get legalized by splitting, handle this
810 // similarly to what getCastInstrCost() does.
811 if (ValTy->isVectorTy()) {
812 unsigned Num = ValTy->getVectorNumElements();
814 CondTy = CondTy->getScalarType();
815 unsigned Cost = static_cast<T *>(this)->getCmpSelInstrCost(
816 Opcode, ValTy->getScalarType(), CondTy, I);
818 // Return the cost of multiple scalar invocation plus the cost of
819 // inserting and extracting the values.
820 return getScalarizationOverhead(ValTy, true, false) + Num * Cost;
823 // Unknown scalar opcode.
827 unsigned getVectorInstrCost(unsigned Opcode, Type *Val, unsigned Index) {
828 std::pair<unsigned, MVT> LT =
829 getTLI()->getTypeLegalizationCost(DL, Val->getScalarType());
834 unsigned getMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
835 unsigned AddressSpace, const Instruction *I = nullptr) {
836 assert(!Src->isVoidTy() && "Invalid type");
837 std::pair<unsigned, MVT> LT = getTLI()->getTypeLegalizationCost(DL, Src);
839 // Assuming that all loads of legal types cost 1.
840 unsigned Cost = LT.first;
842 if (Src->isVectorTy() &&
843 Src->getPrimitiveSizeInBits() < LT.second.getSizeInBits()) {
844 // This is a vector load that legalizes to a larger type than the vector
845 // itself. Unless the corresponding extending load or truncating store is
846 // legal, then this will scalarize.
847 TargetLowering::LegalizeAction LA = TargetLowering::Expand;
848 EVT MemVT = getTLI()->getValueType(DL, Src);
849 if (Opcode == Instruction::Store)
850 LA = getTLI()->getTruncStoreAction(LT.second, MemVT);
852 LA = getTLI()->getLoadExtAction(ISD::EXTLOAD, LT.second, MemVT);
854 if (LA != TargetLowering::Legal && LA != TargetLowering::Custom) {
855 // This is a vector load/store for some illegal type that is scalarized.
856 // We must account for the cost of building or decomposing the vector.
857 Cost += getScalarizationOverhead(Src, Opcode != Instruction::Store,
858 Opcode == Instruction::Store);
865 unsigned getInterleavedMemoryOpCost(unsigned Opcode, Type *VecTy,
867 ArrayRef<unsigned> Indices,
868 unsigned Alignment, unsigned AddressSpace,
869 bool UseMaskForCond = false,
870 bool UseMaskForGaps = false) {
871 VectorType *VT = dyn_cast<VectorType>(VecTy);
872 assert(VT && "Expect a vector type for interleaved memory op");
874 unsigned NumElts = VT->getNumElements();
875 assert(Factor > 1 && NumElts % Factor == 0 && "Invalid interleave factor");
877 unsigned NumSubElts = NumElts / Factor;
878 VectorType *SubVT = VectorType::get(VT->getElementType(), NumSubElts);
880 // Firstly, the cost of load/store operation.
882 if (UseMaskForCond || UseMaskForGaps)
883 Cost = static_cast<T *>(this)->getMaskedMemoryOpCost(
884 Opcode, VecTy, Alignment, AddressSpace);
886 Cost = static_cast<T *>(this)->getMemoryOpCost(Opcode, VecTy, Alignment,
889 // Legalize the vector type, and get the legalized and unlegalized type
891 MVT VecTyLT = getTLI()->getTypeLegalizationCost(DL, VecTy).second;
893 static_cast<T *>(this)->getDataLayout().getTypeStoreSize(VecTy);
894 unsigned VecTyLTSize = VecTyLT.getStoreSize();
896 // Return the ceiling of dividing A by B.
897 auto ceil = [](unsigned A, unsigned B) { return (A + B - 1) / B; };
899 // Scale the cost of the memory operation by the fraction of legalized
900 // instructions that will actually be used. We shouldn't account for the
901 // cost of dead instructions since they will be removed.
903 // E.g., An interleaved load of factor 8:
904 // %vec = load <16 x i64>, <16 x i64>* %ptr
905 // %v0 = shufflevector %vec, undef, <0, 8>
907 // If <16 x i64> is legalized to 8 v2i64 loads, only 2 of the loads will be
908 // used (those corresponding to elements [0:1] and [8:9] of the unlegalized
909 // type). The other loads are unused.
911 // We only scale the cost of loads since interleaved store groups aren't
912 // allowed to have gaps.
913 if (Opcode == Instruction::Load && VecTySize > VecTyLTSize) {
914 // The number of loads of a legal type it will take to represent a load
915 // of the unlegalized vector type.
916 unsigned NumLegalInsts = ceil(VecTySize, VecTyLTSize);
918 // The number of elements of the unlegalized type that correspond to a
919 // single legal instruction.
920 unsigned NumEltsPerLegalInst = ceil(NumElts, NumLegalInsts);
922 // Determine which legal instructions will be used.
923 BitVector UsedInsts(NumLegalInsts, false);
924 for (unsigned Index : Indices)
925 for (unsigned Elt = 0; Elt < NumSubElts; ++Elt)
926 UsedInsts.set((Index + Elt * Factor) / NumEltsPerLegalInst);
928 // Scale the cost of the load by the fraction of legal instructions that
930 Cost *= UsedInsts.count() / NumLegalInsts;
933 // Then plus the cost of interleave operation.
934 if (Opcode == Instruction::Load) {
935 // The interleave cost is similar to extract sub vectors' elements
936 // from the wide vector, and insert them into sub vectors.
938 // E.g. An interleaved load of factor 2 (with one member of index 0):
939 // %vec = load <8 x i32>, <8 x i32>* %ptr
940 // %v0 = shuffle %vec, undef, <0, 2, 4, 6> ; Index 0
941 // The cost is estimated as extract elements at 0, 2, 4, 6 from the
942 // <8 x i32> vector and insert them into a <4 x i32> vector.
944 assert(Indices.size() <= Factor &&
945 "Interleaved memory op has too many members");
947 for (unsigned Index : Indices) {
948 assert(Index < Factor && "Invalid index for interleaved memory op");
950 // Extract elements from loaded vector for each sub vector.
951 for (unsigned i = 0; i < NumSubElts; i++)
952 Cost += static_cast<T *>(this)->getVectorInstrCost(
953 Instruction::ExtractElement, VT, Index + i * Factor);
956 unsigned InsSubCost = 0;
957 for (unsigned i = 0; i < NumSubElts; i++)
958 InsSubCost += static_cast<T *>(this)->getVectorInstrCost(
959 Instruction::InsertElement, SubVT, i);
961 Cost += Indices.size() * InsSubCost;
963 // The interleave cost is extract all elements from sub vectors, and
964 // insert them into the wide vector.
966 // E.g. An interleaved store of factor 2:
967 // %v0_v1 = shuffle %v0, %v1, <0, 4, 1, 5, 2, 6, 3, 7>
968 // store <8 x i32> %interleaved.vec, <8 x i32>* %ptr
969 // The cost is estimated as extract all elements from both <4 x i32>
970 // vectors and insert into the <8 x i32> vector.
972 unsigned ExtSubCost = 0;
973 for (unsigned i = 0; i < NumSubElts; i++)
974 ExtSubCost += static_cast<T *>(this)->getVectorInstrCost(
975 Instruction::ExtractElement, SubVT, i);
976 Cost += ExtSubCost * Factor;
978 for (unsigned i = 0; i < NumElts; i++)
979 Cost += static_cast<T *>(this)
980 ->getVectorInstrCost(Instruction::InsertElement, VT, i);
986 Type *I8Type = Type::getInt8Ty(VT->getContext());
987 VectorType *MaskVT = VectorType::get(I8Type, NumElts);
988 SubVT = VectorType::get(I8Type, NumSubElts);
990 // The Mask shuffling cost is extract all the elements of the Mask
991 // and insert each of them Factor times into the wide vector:
993 // E.g. an interleaved group with factor 3:
994 // %mask = icmp ult <8 x i32> %vec1, %vec2
995 // %interleaved.mask = shufflevector <8 x i1> %mask, <8 x i1> undef,
996 // <24 x i32> <0,0,0,1,1,1,2,2,2,3,3,3,4,4,4,5,5,5,6,6,6,7,7,7>
997 // The cost is estimated as extract all mask elements from the <8xi1> mask
998 // vector and insert them factor times into the <24xi1> shuffled mask
1000 for (unsigned i = 0; i < NumSubElts; i++)
1001 Cost += static_cast<T *>(this)->getVectorInstrCost(
1002 Instruction::ExtractElement, SubVT, i);
1004 for (unsigned i = 0; i < NumElts; i++)
1005 Cost += static_cast<T *>(this)->getVectorInstrCost(
1006 Instruction::InsertElement, MaskVT, i);
1008 // The Gaps mask is invariant and created outside the loop, therefore the
1009 // cost of creating it is not accounted for here. However if we have both
1010 // a MaskForGaps and some other mask that guards the execution of the
1011 // memory access, we need to account for the cost of And-ing the two masks
1014 Cost += static_cast<T *>(this)->getArithmeticInstrCost(
1015 BinaryOperator::And, MaskVT);
1020 /// Get intrinsic cost based on arguments.
1021 unsigned getIntrinsicInstrCost(Intrinsic::ID IID, Type *RetTy,
1022 ArrayRef<Value *> Args, FastMathFlags FMF,
1024 unsigned RetVF = (RetTy->isVectorTy() ? RetTy->getVectorNumElements() : 1);
1025 assert((RetVF == 1 || VF == 1) && "VF > 1 and RetVF is a vector type");
1026 auto *ConcreteTTI = static_cast<T *>(this);
1030 // Assume that we need to scalarize this intrinsic.
1031 SmallVector<Type *, 4> Types;
1032 for (Value *Op : Args) {
1033 Type *OpTy = Op->getType();
1034 assert(VF == 1 || !OpTy->isVectorTy());
1035 Types.push_back(VF == 1 ? OpTy : VectorType::get(OpTy, VF));
1038 if (VF > 1 && !RetTy->isVoidTy())
1039 RetTy = VectorType::get(RetTy, VF);
1041 // Compute the scalarization overhead based on Args for a vector
1042 // intrinsic. A vectorizer will pass a scalar RetTy and VF > 1, while
1043 // CostModel will pass a vector RetTy and VF is 1.
1044 unsigned ScalarizationCost = std::numeric_limits<unsigned>::max();
1045 if (RetVF > 1 || VF > 1) {
1046 ScalarizationCost = 0;
1047 if (!RetTy->isVoidTy())
1048 ScalarizationCost += getScalarizationOverhead(RetTy, true, false);
1049 ScalarizationCost += getOperandsScalarizationOverhead(Args, VF);
1052 return ConcreteTTI->getIntrinsicInstrCost(IID, RetTy, Types, FMF,
1055 case Intrinsic::masked_scatter: {
1056 assert(VF == 1 && "Can't vectorize types here.");
1057 Value *Mask = Args[3];
1058 bool VarMask = !isa<Constant>(Mask);
1059 unsigned Alignment = cast<ConstantInt>(Args[2])->getZExtValue();
1060 return ConcreteTTI->getGatherScatterOpCost(
1061 Instruction::Store, Args[0]->getType(), Args[1], VarMask, Alignment);
1063 case Intrinsic::masked_gather: {
1064 assert(VF == 1 && "Can't vectorize types here.");
1065 Value *Mask = Args[2];
1066 bool VarMask = !isa<Constant>(Mask);
1067 unsigned Alignment = cast<ConstantInt>(Args[1])->getZExtValue();
1068 return ConcreteTTI->getGatherScatterOpCost(Instruction::Load, RetTy,
1069 Args[0], VarMask, Alignment);
1071 case Intrinsic::experimental_vector_reduce_add:
1072 case Intrinsic::experimental_vector_reduce_mul:
1073 case Intrinsic::experimental_vector_reduce_and:
1074 case Intrinsic::experimental_vector_reduce_or:
1075 case Intrinsic::experimental_vector_reduce_xor:
1076 case Intrinsic::experimental_vector_reduce_v2_fadd:
1077 case Intrinsic::experimental_vector_reduce_v2_fmul:
1078 case Intrinsic::experimental_vector_reduce_smax:
1079 case Intrinsic::experimental_vector_reduce_smin:
1080 case Intrinsic::experimental_vector_reduce_fmax:
1081 case Intrinsic::experimental_vector_reduce_fmin:
1082 case Intrinsic::experimental_vector_reduce_umax:
1083 case Intrinsic::experimental_vector_reduce_umin:
1084 return getIntrinsicInstrCost(IID, RetTy, Args[0]->getType(), FMF);
1085 case Intrinsic::fshl:
1086 case Intrinsic::fshr: {
1090 TTI::OperandValueProperties OpPropsX, OpPropsY, OpPropsZ, OpPropsBW;
1091 TTI::OperandValueKind OpKindX = TTI::getOperandInfo(X, OpPropsX);
1092 TTI::OperandValueKind OpKindY = TTI::getOperandInfo(Y, OpPropsY);
1093 TTI::OperandValueKind OpKindZ = TTI::getOperandInfo(Z, OpPropsZ);
1094 TTI::OperandValueKind OpKindBW = TTI::OK_UniformConstantValue;
1095 OpPropsBW = isPowerOf2_32(RetTy->getScalarSizeInBits()) ? TTI::OP_PowerOf2
1097 // fshl: (X << (Z % BW)) | (Y >> (BW - (Z % BW)))
1098 // fshr: (X << (BW - (Z % BW))) | (Y >> (Z % BW))
1100 Cost += ConcreteTTI->getArithmeticInstrCost(BinaryOperator::Or, RetTy);
1101 Cost += ConcreteTTI->getArithmeticInstrCost(BinaryOperator::Sub, RetTy);
1102 Cost += ConcreteTTI->getArithmeticInstrCost(BinaryOperator::Shl, RetTy,
1103 OpKindX, OpKindZ, OpPropsX);
1104 Cost += ConcreteTTI->getArithmeticInstrCost(BinaryOperator::LShr, RetTy,
1105 OpKindY, OpKindZ, OpPropsY);
1106 // Non-constant shift amounts requires a modulo.
1107 if (OpKindZ != TTI::OK_UniformConstantValue &&
1108 OpKindZ != TTI::OK_NonUniformConstantValue)
1109 Cost += ConcreteTTI->getArithmeticInstrCost(BinaryOperator::URem, RetTy,
1110 OpKindZ, OpKindBW, OpPropsZ,
1112 // For non-rotates (X != Y) we must add shift-by-zero handling costs.
1114 Type *CondTy = Type::getInt1Ty(RetTy->getContext());
1116 CondTy = VectorType::get(CondTy, RetVF);
1117 Cost += ConcreteTTI->getCmpSelInstrCost(BinaryOperator::ICmp, RetTy,
1119 Cost += ConcreteTTI->getCmpSelInstrCost(BinaryOperator::Select, RetTy,
1127 /// Get intrinsic cost based on argument types.
1128 /// If ScalarizationCostPassed is std::numeric_limits<unsigned>::max(), the
1129 /// cost of scalarizing the arguments and the return value will be computed
1131 unsigned getIntrinsicInstrCost(
1132 Intrinsic::ID IID, Type *RetTy, ArrayRef<Type *> Tys, FastMathFlags FMF,
1133 unsigned ScalarizationCostPassed = std::numeric_limits<unsigned>::max()) {
1134 unsigned RetVF = (RetTy->isVectorTy() ? RetTy->getVectorNumElements() : 1);
1135 auto *ConcreteTTI = static_cast<T *>(this);
1137 SmallVector<unsigned, 2> ISDs;
1138 unsigned SingleCallCost = 10; // Library call cost. Make it expensive.
1141 // Assume that we need to scalarize this intrinsic.
1142 unsigned ScalarizationCost = ScalarizationCostPassed;
1143 unsigned ScalarCalls = 1;
1144 Type *ScalarRetTy = RetTy;
1145 if (RetTy->isVectorTy()) {
1146 if (ScalarizationCostPassed == std::numeric_limits<unsigned>::max())
1147 ScalarizationCost = getScalarizationOverhead(RetTy, true, false);
1148 ScalarCalls = std::max(ScalarCalls, RetTy->getVectorNumElements());
1149 ScalarRetTy = RetTy->getScalarType();
1151 SmallVector<Type *, 4> ScalarTys;
1152 for (unsigned i = 0, ie = Tys.size(); i != ie; ++i) {
1154 if (Ty->isVectorTy()) {
1155 if (ScalarizationCostPassed == std::numeric_limits<unsigned>::max())
1156 ScalarizationCost += getScalarizationOverhead(Ty, false, true);
1157 ScalarCalls = std::max(ScalarCalls, Ty->getVectorNumElements());
1158 Ty = Ty->getScalarType();
1160 ScalarTys.push_back(Ty);
1162 if (ScalarCalls == 1)
1163 return 1; // Return cost of a scalar intrinsic. Assume it to be cheap.
1165 unsigned ScalarCost =
1166 ConcreteTTI->getIntrinsicInstrCost(IID, ScalarRetTy, ScalarTys, FMF);
1168 return ScalarCalls * ScalarCost + ScalarizationCost;
1170 // Look for intrinsics that can be lowered directly or turned into a scalar
1172 case Intrinsic::sqrt:
1173 ISDs.push_back(ISD::FSQRT);
1175 case Intrinsic::sin:
1176 ISDs.push_back(ISD::FSIN);
1178 case Intrinsic::cos:
1179 ISDs.push_back(ISD::FCOS);
1181 case Intrinsic::exp:
1182 ISDs.push_back(ISD::FEXP);
1184 case Intrinsic::exp2:
1185 ISDs.push_back(ISD::FEXP2);
1187 case Intrinsic::log:
1188 ISDs.push_back(ISD::FLOG);
1190 case Intrinsic::log10:
1191 ISDs.push_back(ISD::FLOG10);
1193 case Intrinsic::log2:
1194 ISDs.push_back(ISD::FLOG2);
1196 case Intrinsic::fabs:
1197 ISDs.push_back(ISD::FABS);
1199 case Intrinsic::canonicalize:
1200 ISDs.push_back(ISD::FCANONICALIZE);
1202 case Intrinsic::minnum:
1203 ISDs.push_back(ISD::FMINNUM);
1205 ISDs.push_back(ISD::FMINIMUM);
1207 case Intrinsic::maxnum:
1208 ISDs.push_back(ISD::FMAXNUM);
1210 ISDs.push_back(ISD::FMAXIMUM);
1212 case Intrinsic::copysign:
1213 ISDs.push_back(ISD::FCOPYSIGN);
1215 case Intrinsic::floor:
1216 ISDs.push_back(ISD::FFLOOR);
1218 case Intrinsic::ceil:
1219 ISDs.push_back(ISD::FCEIL);
1221 case Intrinsic::trunc:
1222 ISDs.push_back(ISD::FTRUNC);
1224 case Intrinsic::nearbyint:
1225 ISDs.push_back(ISD::FNEARBYINT);
1227 case Intrinsic::rint:
1228 ISDs.push_back(ISD::FRINT);
1230 case Intrinsic::round:
1231 ISDs.push_back(ISD::FROUND);
1233 case Intrinsic::pow:
1234 ISDs.push_back(ISD::FPOW);
1236 case Intrinsic::fma:
1237 ISDs.push_back(ISD::FMA);
1239 case Intrinsic::fmuladd:
1240 ISDs.push_back(ISD::FMA);
1242 // FIXME: We should return 0 whenever getIntrinsicCost == TCC_Free.
1243 case Intrinsic::lifetime_start:
1244 case Intrinsic::lifetime_end:
1245 case Intrinsic::sideeffect:
1247 case Intrinsic::masked_store:
1248 return ConcreteTTI->getMaskedMemoryOpCost(Instruction::Store, Tys[0], 0,
1250 case Intrinsic::masked_load:
1251 return ConcreteTTI->getMaskedMemoryOpCost(Instruction::Load, RetTy, 0, 0);
1252 case Intrinsic::experimental_vector_reduce_add:
1253 return ConcreteTTI->getArithmeticReductionCost(Instruction::Add, Tys[0],
1254 /*IsPairwiseForm=*/false);
1255 case Intrinsic::experimental_vector_reduce_mul:
1256 return ConcreteTTI->getArithmeticReductionCost(Instruction::Mul, Tys[0],
1257 /*IsPairwiseForm=*/false);
1258 case Intrinsic::experimental_vector_reduce_and:
1259 return ConcreteTTI->getArithmeticReductionCost(Instruction::And, Tys[0],
1260 /*IsPairwiseForm=*/false);
1261 case Intrinsic::experimental_vector_reduce_or:
1262 return ConcreteTTI->getArithmeticReductionCost(Instruction::Or, Tys[0],
1263 /*IsPairwiseForm=*/false);
1264 case Intrinsic::experimental_vector_reduce_xor:
1265 return ConcreteTTI->getArithmeticReductionCost(Instruction::Xor, Tys[0],
1266 /*IsPairwiseForm=*/false);
1267 case Intrinsic::experimental_vector_reduce_v2_fadd:
1268 return ConcreteTTI->getArithmeticReductionCost(
1269 Instruction::FAdd, Tys[0],
1270 /*IsPairwiseForm=*/false); // FIXME: Add new flag for cost of strict
1272 case Intrinsic::experimental_vector_reduce_v2_fmul:
1273 return ConcreteTTI->getArithmeticReductionCost(
1274 Instruction::FMul, Tys[0],
1275 /*IsPairwiseForm=*/false); // FIXME: Add new flag for cost of strict
1277 case Intrinsic::experimental_vector_reduce_smax:
1278 case Intrinsic::experimental_vector_reduce_smin:
1279 case Intrinsic::experimental_vector_reduce_fmax:
1280 case Intrinsic::experimental_vector_reduce_fmin:
1281 return ConcreteTTI->getMinMaxReductionCost(
1282 Tys[0], CmpInst::makeCmpResultType(Tys[0]), /*IsPairwiseForm=*/false,
1283 /*IsUnsigned=*/true);
1284 case Intrinsic::experimental_vector_reduce_umax:
1285 case Intrinsic::experimental_vector_reduce_umin:
1286 return ConcreteTTI->getMinMaxReductionCost(
1287 Tys[0], CmpInst::makeCmpResultType(Tys[0]), /*IsPairwiseForm=*/false,
1288 /*IsUnsigned=*/false);
1289 case Intrinsic::sadd_sat:
1290 case Intrinsic::ssub_sat: {
1291 Type *CondTy = Type::getInt1Ty(RetTy->getContext());
1293 CondTy = VectorType::get(CondTy, RetVF);
1295 Type *OpTy = StructType::create({RetTy, CondTy});
1296 Intrinsic::ID OverflowOp = IID == Intrinsic::sadd_sat
1297 ? Intrinsic::sadd_with_overflow
1298 : Intrinsic::ssub_with_overflow;
1300 // SatMax -> Overflow && SumDiff < 0
1301 // SatMin -> Overflow && SumDiff >= 0
1303 Cost += ConcreteTTI->getIntrinsicInstrCost(
1304 OverflowOp, OpTy, {RetTy, RetTy}, FMF, ScalarizationCostPassed);
1305 Cost += ConcreteTTI->getCmpSelInstrCost(BinaryOperator::ICmp, RetTy,
1307 Cost += 2 * ConcreteTTI->getCmpSelInstrCost(BinaryOperator::Select, RetTy,
1311 case Intrinsic::uadd_sat:
1312 case Intrinsic::usub_sat: {
1313 Type *CondTy = Type::getInt1Ty(RetTy->getContext());
1315 CondTy = VectorType::get(CondTy, RetVF);
1317 Type *OpTy = StructType::create({RetTy, CondTy});
1318 Intrinsic::ID OverflowOp = IID == Intrinsic::uadd_sat
1319 ? Intrinsic::uadd_with_overflow
1320 : Intrinsic::usub_with_overflow;
1323 Cost += ConcreteTTI->getIntrinsicInstrCost(
1324 OverflowOp, OpTy, {RetTy, RetTy}, FMF, ScalarizationCostPassed);
1325 Cost += ConcreteTTI->getCmpSelInstrCost(BinaryOperator::Select, RetTy,
1329 case Intrinsic::smul_fix:
1330 case Intrinsic::umul_fix: {
1331 unsigned ExtSize = RetTy->getScalarSizeInBits() * 2;
1332 Type *ExtTy = Type::getIntNTy(RetTy->getContext(), ExtSize);
1334 ExtTy = VectorType::get(ExtTy, RetVF);
1337 IID == Intrinsic::smul_fix ? Instruction::SExt : Instruction::ZExt;
1340 Cost += 2 * ConcreteTTI->getCastInstrCost(ExtOp, ExtTy, RetTy);
1341 Cost += ConcreteTTI->getArithmeticInstrCost(Instruction::Mul, ExtTy);
1343 2 * ConcreteTTI->getCastInstrCost(Instruction::Trunc, RetTy, ExtTy);
1344 Cost += ConcreteTTI->getArithmeticInstrCost(Instruction::LShr, RetTy,
1346 TTI::OK_UniformConstantValue);
1347 Cost += ConcreteTTI->getArithmeticInstrCost(Instruction::Shl, RetTy,
1349 TTI::OK_UniformConstantValue);
1350 Cost += ConcreteTTI->getArithmeticInstrCost(Instruction::Or, RetTy);
1353 case Intrinsic::sadd_with_overflow:
1354 case Intrinsic::ssub_with_overflow: {
1355 Type *SumTy = RetTy->getContainedType(0);
1356 Type *OverflowTy = RetTy->getContainedType(1);
1357 unsigned Opcode = IID == Intrinsic::sadd_with_overflow
1358 ? BinaryOperator::Add
1359 : BinaryOperator::Sub;
1361 // LHSSign -> LHS >= 0
1362 // RHSSign -> RHS >= 0
1363 // SumSign -> Sum >= 0
1366 // Overflow -> (LHSSign == RHSSign) && (LHSSign != SumSign)
1368 // Overflow -> (LHSSign != RHSSign) && (LHSSign != SumSign)
1370 Cost += ConcreteTTI->getArithmeticInstrCost(Opcode, SumTy);
1371 Cost += 3 * ConcreteTTI->getCmpSelInstrCost(BinaryOperator::ICmp, SumTy,
1372 OverflowTy, nullptr);
1373 Cost += 2 * ConcreteTTI->getCmpSelInstrCost(
1374 BinaryOperator::ICmp, OverflowTy, OverflowTy, nullptr);
1376 ConcreteTTI->getArithmeticInstrCost(BinaryOperator::And, OverflowTy);
1379 case Intrinsic::uadd_with_overflow:
1380 case Intrinsic::usub_with_overflow: {
1381 Type *SumTy = RetTy->getContainedType(0);
1382 Type *OverflowTy = RetTy->getContainedType(1);
1383 unsigned Opcode = IID == Intrinsic::uadd_with_overflow
1384 ? BinaryOperator::Add
1385 : BinaryOperator::Sub;
1388 Cost += ConcreteTTI->getArithmeticInstrCost(Opcode, SumTy);
1389 Cost += ConcreteTTI->getCmpSelInstrCost(BinaryOperator::ICmp, SumTy,
1390 OverflowTy, nullptr);
1393 case Intrinsic::smul_with_overflow:
1394 case Intrinsic::umul_with_overflow: {
1395 Type *MulTy = RetTy->getContainedType(0);
1396 Type *OverflowTy = RetTy->getContainedType(1);
1397 unsigned ExtSize = MulTy->getScalarSizeInBits() * 2;
1398 Type *ExtTy = Type::getIntNTy(RetTy->getContext(), ExtSize);
1399 if (MulTy->isVectorTy())
1400 ExtTy = VectorType::get(ExtTy, MulTy->getVectorNumElements() );
1403 IID == Intrinsic::smul_fix ? Instruction::SExt : Instruction::ZExt;
1406 Cost += 2 * ConcreteTTI->getCastInstrCost(ExtOp, ExtTy, MulTy);
1407 Cost += ConcreteTTI->getArithmeticInstrCost(Instruction::Mul, ExtTy);
1409 2 * ConcreteTTI->getCastInstrCost(Instruction::Trunc, MulTy, ExtTy);
1410 Cost += ConcreteTTI->getArithmeticInstrCost(Instruction::LShr, MulTy,
1412 TTI::OK_UniformConstantValue);
1414 if (IID == Intrinsic::smul_with_overflow)
1415 Cost += ConcreteTTI->getArithmeticInstrCost(
1416 Instruction::AShr, MulTy, TTI::OK_AnyValue,
1417 TTI::OK_UniformConstantValue);
1419 Cost += ConcreteTTI->getCmpSelInstrCost(BinaryOperator::ICmp, MulTy,
1420 OverflowTy, nullptr);
1423 case Intrinsic::ctpop:
1424 ISDs.push_back(ISD::CTPOP);
1425 // In case of legalization use TCC_Expensive. This is cheaper than a
1426 // library call but still not a cheap instruction.
1427 SingleCallCost = TargetTransformInfo::TCC_Expensive;
1429 // FIXME: ctlz, cttz, ...
1432 const TargetLoweringBase *TLI = getTLI();
1433 std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(DL, RetTy);
1435 SmallVector<unsigned, 2> LegalCost;
1436 SmallVector<unsigned, 2> CustomCost;
1437 for (unsigned ISD : ISDs) {
1438 if (TLI->isOperationLegalOrPromote(ISD, LT.second)) {
1439 if (IID == Intrinsic::fabs && LT.second.isFloatingPoint() &&
1440 TLI->isFAbsFree(LT.second)) {
1444 // The operation is legal. Assume it costs 1.
1445 // If the type is split to multiple registers, assume that there is some
1446 // overhead to this.
1447 // TODO: Once we have extract/insert subvector cost we need to use them.
1449 LegalCost.push_back(LT.first * 2);
1451 LegalCost.push_back(LT.first * 1);
1452 } else if (!TLI->isOperationExpand(ISD, LT.second)) {
1453 // If the operation is custom lowered then assume
1454 // that the code is twice as expensive.
1455 CustomCost.push_back(LT.first * 2);
1459 auto MinLegalCostI = std::min_element(LegalCost.begin(), LegalCost.end());
1460 if (MinLegalCostI != LegalCost.end())
1461 return *MinLegalCostI;
1463 auto MinCustomCostI =
1464 std::min_element(CustomCost.begin(), CustomCost.end());
1465 if (MinCustomCostI != CustomCost.end())
1466 return *MinCustomCostI;
1468 // If we can't lower fmuladd into an FMA estimate the cost as a floating
1469 // point mul followed by an add.
1470 if (IID == Intrinsic::fmuladd)
1471 return ConcreteTTI->getArithmeticInstrCost(BinaryOperator::FMul, RetTy) +
1472 ConcreteTTI->getArithmeticInstrCost(BinaryOperator::FAdd, RetTy);
1474 // Else, assume that we need to scalarize this intrinsic. For math builtins
1475 // this will emit a costly libcall, adding call overhead and spills. Make it
1477 if (RetTy->isVectorTy()) {
1478 unsigned ScalarizationCost =
1479 ((ScalarizationCostPassed != std::numeric_limits<unsigned>::max())
1480 ? ScalarizationCostPassed
1481 : getScalarizationOverhead(RetTy, true, false));
1482 unsigned ScalarCalls = RetTy->getVectorNumElements();
1483 SmallVector<Type *, 4> ScalarTys;
1484 for (unsigned i = 0, ie = Tys.size(); i != ie; ++i) {
1486 if (Ty->isVectorTy())
1487 Ty = Ty->getScalarType();
1488 ScalarTys.push_back(Ty);
1490 unsigned ScalarCost = ConcreteTTI->getIntrinsicInstrCost(
1491 IID, RetTy->getScalarType(), ScalarTys, FMF);
1492 for (unsigned i = 0, ie = Tys.size(); i != ie; ++i) {
1493 if (Tys[i]->isVectorTy()) {
1494 if (ScalarizationCostPassed == std::numeric_limits<unsigned>::max())
1495 ScalarizationCost += getScalarizationOverhead(Tys[i], false, true);
1496 ScalarCalls = std::max(ScalarCalls, Tys[i]->getVectorNumElements());
1500 return ScalarCalls * ScalarCost + ScalarizationCost;
1503 // This is going to be turned into a library call, make it expensive.
1504 return SingleCallCost;
1507 /// Compute a cost of the given call instruction.
1509 /// Compute the cost of calling function F with return type RetTy and
1510 /// argument types Tys. F might be nullptr, in this case the cost of an
1511 /// arbitrary call with the specified signature will be returned.
1512 /// This is used, for instance, when we estimate call of a vector
1513 /// counterpart of the given function.
1514 /// \param F Called function, might be nullptr.
1515 /// \param RetTy Return value types.
1516 /// \param Tys Argument types.
1517 /// \returns The cost of Call instruction.
1518 unsigned getCallInstrCost(Function *F, Type *RetTy, ArrayRef<Type *> Tys) {
1522 unsigned getNumberOfParts(Type *Tp) {
1523 std::pair<unsigned, MVT> LT = getTLI()->getTypeLegalizationCost(DL, Tp);
1527 unsigned getAddressComputationCost(Type *Ty, ScalarEvolution *,
1532 /// Try to calculate arithmetic and shuffle op costs for reduction operations.
1533 /// We're assuming that reduction operation are performing the following way:
1534 /// 1. Non-pairwise reduction
1535 /// %val1 = shufflevector<n x t> %val, <n x t> %undef,
1536 /// <n x i32> <i32 n/2, i32 n/2 + 1, ..., i32 n, i32 undef, ..., i32 undef>
1537 /// \----------------v-------------/ \----------v------------/
1538 /// n/2 elements n/2 elements
1539 /// %red1 = op <n x t> %val, <n x t> val1
1540 /// After this operation we have a vector %red1 where only the first n/2
1541 /// elements are meaningful, the second n/2 elements are undefined and can be
1542 /// dropped. All other operations are actually working with the vector of
1543 /// length n/2, not n, though the real vector length is still n.
1544 /// %val2 = shufflevector<n x t> %red1, <n x t> %undef,
1545 /// <n x i32> <i32 n/4, i32 n/4 + 1, ..., i32 n/2, i32 undef, ..., i32 undef>
1546 /// \----------------v-------------/ \----------v------------/
1547 /// n/4 elements 3*n/4 elements
1548 /// %red2 = op <n x t> %red1, <n x t> val2 - working with the vector of
1549 /// length n/2, the resulting vector has length n/4 etc.
1550 /// 2. Pairwise reduction:
1551 /// Everything is the same except for an additional shuffle operation which
1552 /// is used to produce operands for pairwise kind of reductions.
1553 /// %val1 = shufflevector<n x t> %val, <n x t> %undef,
1554 /// <n x i32> <i32 0, i32 2, ..., i32 n-2, i32 undef, ..., i32 undef>
1555 /// \-------------v----------/ \----------v------------/
1556 /// n/2 elements n/2 elements
1557 /// %val2 = shufflevector<n x t> %val, <n x t> %undef,
1558 /// <n x i32> <i32 1, i32 3, ..., i32 n-1, i32 undef, ..., i32 undef>
1559 /// \-------------v----------/ \----------v------------/
1560 /// n/2 elements n/2 elements
1561 /// %red1 = op <n x t> %val1, <n x t> val2
1562 /// Again, the operation is performed on <n x t> vector, but the resulting
1563 /// vector %red1 is <n/2 x t> vector.
1565 /// The cost model should take into account that the actual length of the
1566 /// vector is reduced on each iteration.
1567 unsigned getArithmeticReductionCost(unsigned Opcode, Type *Ty,
1569 assert(Ty->isVectorTy() && "Expect a vector type");
1570 Type *ScalarTy = Ty->getVectorElementType();
1571 unsigned NumVecElts = Ty->getVectorNumElements();
1572 unsigned NumReduxLevels = Log2_32(NumVecElts);
1573 unsigned ArithCost = 0;
1574 unsigned ShuffleCost = 0;
1575 auto *ConcreteTTI = static_cast<T *>(this);
1576 std::pair<unsigned, MVT> LT =
1577 ConcreteTTI->getTLI()->getTypeLegalizationCost(DL, Ty);
1578 unsigned LongVectorCount = 0;
1580 LT.second.isVector() ? LT.second.getVectorNumElements() : 1;
1581 while (NumVecElts > MVTLen) {
1583 Type *SubTy = VectorType::get(ScalarTy, NumVecElts);
1584 // Assume the pairwise shuffles add a cost.
1585 ShuffleCost += (IsPairwise + 1) *
1586 ConcreteTTI->getShuffleCost(TTI::SK_ExtractSubvector, Ty,
1588 ArithCost += ConcreteTTI->getArithmeticInstrCost(Opcode, SubTy);
1593 NumReduxLevels -= LongVectorCount;
1595 // The minimal length of the vector is limited by the real length of vector
1596 // operations performed on the current platform. That's why several final
1597 // reduction operations are performed on the vectors with the same
1598 // architecture-dependent length.
1600 // Non pairwise reductions need one shuffle per reduction level. Pairwise
1601 // reductions need two shuffles on every level, but the last one. On that
1602 // level one of the shuffles is <0, u, u, ...> which is identity.
1603 unsigned NumShuffles = NumReduxLevels;
1604 if (IsPairwise && NumReduxLevels >= 1)
1605 NumShuffles += NumReduxLevels - 1;
1606 ShuffleCost += NumShuffles *
1607 ConcreteTTI->getShuffleCost(TTI::SK_PermuteSingleSrc, Ty,
1609 ArithCost += NumReduxLevels *
1610 ConcreteTTI->getArithmeticInstrCost(Opcode, Ty);
1611 return ShuffleCost + ArithCost +
1612 ConcreteTTI->getVectorInstrCost(Instruction::ExtractElement, Ty, 0);
1615 /// Try to calculate op costs for min/max reduction operations.
1616 /// \param CondTy Conditional type for the Select instruction.
1617 unsigned getMinMaxReductionCost(Type *Ty, Type *CondTy, bool IsPairwise,
1619 assert(Ty->isVectorTy() && "Expect a vector type");
1620 Type *ScalarTy = Ty->getVectorElementType();
1621 Type *ScalarCondTy = CondTy->getVectorElementType();
1622 unsigned NumVecElts = Ty->getVectorNumElements();
1623 unsigned NumReduxLevels = Log2_32(NumVecElts);
1625 if (Ty->isFPOrFPVectorTy()) {
1626 CmpOpcode = Instruction::FCmp;
1628 assert(Ty->isIntOrIntVectorTy() &&
1629 "expecting floating point or integer type for min/max reduction");
1630 CmpOpcode = Instruction::ICmp;
1632 unsigned MinMaxCost = 0;
1633 unsigned ShuffleCost = 0;
1634 auto *ConcreteTTI = static_cast<T *>(this);
1635 std::pair<unsigned, MVT> LT =
1636 ConcreteTTI->getTLI()->getTypeLegalizationCost(DL, Ty);
1637 unsigned LongVectorCount = 0;
1639 LT.second.isVector() ? LT.second.getVectorNumElements() : 1;
1640 while (NumVecElts > MVTLen) {
1642 Type *SubTy = VectorType::get(ScalarTy, NumVecElts);
1643 CondTy = VectorType::get(ScalarCondTy, NumVecElts);
1645 // Assume the pairwise shuffles add a cost.
1646 ShuffleCost += (IsPairwise + 1) *
1647 ConcreteTTI->getShuffleCost(TTI::SK_ExtractSubvector, Ty,
1650 ConcreteTTI->getCmpSelInstrCost(CmpOpcode, SubTy, CondTy, nullptr) +
1651 ConcreteTTI->getCmpSelInstrCost(Instruction::Select, SubTy, CondTy,
1657 NumReduxLevels -= LongVectorCount;
1659 // The minimal length of the vector is limited by the real length of vector
1660 // operations performed on the current platform. That's why several final
1661 // reduction opertions are perfomed on the vectors with the same
1662 // architecture-dependent length.
1664 // Non pairwise reductions need one shuffle per reduction level. Pairwise
1665 // reductions need two shuffles on every level, but the last one. On that
1666 // level one of the shuffles is <0, u, u, ...> which is identity.
1667 unsigned NumShuffles = NumReduxLevels;
1668 if (IsPairwise && NumReduxLevels >= 1)
1669 NumShuffles += NumReduxLevels - 1;
1670 ShuffleCost += NumShuffles *
1671 ConcreteTTI->getShuffleCost(TTI::SK_PermuteSingleSrc, Ty,
1675 (ConcreteTTI->getCmpSelInstrCost(CmpOpcode, Ty, CondTy, nullptr) +
1676 ConcreteTTI->getCmpSelInstrCost(Instruction::Select, Ty, CondTy,
1678 // The last min/max should be in vector registers and we counted it above.
1679 // So just need a single extractelement.
1680 return ShuffleCost + MinMaxCost +
1681 ConcreteTTI->getVectorInstrCost(Instruction::ExtractElement, Ty, 0);
1684 unsigned getVectorSplitCost() { return 1; }
1689 /// Concrete BasicTTIImpl that can be used if no further customization
1691 class BasicTTIImpl : public BasicTTIImplBase<BasicTTIImpl> {
1692 using BaseT = BasicTTIImplBase<BasicTTIImpl>;
1694 friend class BasicTTIImplBase<BasicTTIImpl>;
1696 const TargetSubtargetInfo *ST;
1697 const TargetLoweringBase *TLI;
1699 const TargetSubtargetInfo *getST() const { return ST; }
1700 const TargetLoweringBase *getTLI() const { return TLI; }
1703 explicit BasicTTIImpl(const TargetMachine *TM, const Function &F);
1706 } // end namespace llvm
1708 #endif // LLVM_CODEGEN_BASICTTIIMPL_H