LLVM 24.0.0git
RISCVTargetTransformInfo.h
Go to the documentation of this file.
1//===- RISCVTargetTransformInfo.h - RISC-V specific TTI ---------*- C++ -*-===//
2//
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
6//
7//===----------------------------------------------------------------------===//
8/// \file
9/// This file defines a TargetTransformInfoImplBase conforming object specific
10/// to the RISC-V target machine. It uses the target's detailed information to
11/// provide more precise answers to certain TTI queries, while letting the
12/// target independent and default TTI implementations handle the rest.
13///
14//===----------------------------------------------------------------------===//
15
16#ifndef LLVM_LIB_TARGET_RISCV_RISCVTARGETTRANSFORMINFO_H
17#define LLVM_LIB_TARGET_RISCV_RISCVTARGETTRANSFORMINFO_H
18
19#include "RISCVSubtarget.h"
20#include "RISCVTargetMachine.h"
23#include "llvm/IR/Function.h"
24#include <optional>
25
26namespace llvm {
27
28class RISCVTTIImpl final : public BasicTTIImplBase<RISCVTTIImpl> {
30 using TTI = TargetTransformInfo;
31
32 friend BaseT;
33
34 const RISCVSubtarget *ST;
35 const RISCVTargetLowering *TLI;
36
37 const RISCVSubtarget *getST() const { return ST; }
38 const RISCVTargetLowering *getTLI() const { return TLI; }
39
40 /// This function returns an estimate for VL to be used in VL based terms
41 /// of the cost model. For fixed length vectors, this is simply the
42 /// vector length. For scalable vectors, we return results consistent
43 /// with getVScaleForTuning under the assumption that clients are also
44 /// using that when comparing costs between scalar and vector representation.
45 /// This does unfortunately mean that we can both undershoot and overshot
46 /// the true cost significantly if getVScaleForTuning is wildly off for the
47 /// actual target hardware.
48 unsigned getEstimatedVLFor(VectorType *Ty) const;
49
50 /// This function calculates the costs for one or more RVV opcodes based
51 /// on the vtype and the cost kind.
52 /// \param Opcodes A list of opcodes of the RVV instruction to evaluate.
53 /// \param VT The MVT of vtype associated with the RVV instructions.
54 /// For widening/narrowing instructions where the result and source types
55 /// differ, it is important to check the spec to determine whether the vtype
56 /// refers to the result or source type.
57 /// \param CostKind The type of cost to compute.
58 InstructionCost getRISCVInstructionCost(ArrayRef<unsigned> OpCodes, MVT VT,
60
61 // Return the cost of generating a PC relative address
63 getStaticDataAddrGenerationCost(const TTI::TargetCostKind CostKind) const;
64
65 /// Return the cost of accessing a constant pool entry of the specified
66 /// type.
67 InstructionCost getConstantPoolLoadCost(Type *Ty,
69
70 /// If this shuffle can be lowered as a masked slide pair (at worst),
71 /// return a cost for it.
72 InstructionCost getSlideCost(FixedVectorType *Tp, ArrayRef<int> Mask,
74
75public:
76 explicit RISCVTTIImpl(const RISCVTargetMachine *TM, const Function &F)
77 : BaseT(TM, F.getDataLayout()), ST(TM->getSubtargetImpl(F)),
78 TLI(ST->getTargetLowering()) {}
79
80 /// Return the cost of materializing an immediate for a value operand of
81 /// a store instruction.
84
86 TTI::TargetCostKind CostKind) const override;
87 InstructionCost getIntImmCostInst(unsigned Opcode, unsigned Idx,
88 const APInt &Imm, Type *Ty,
90 Instruction *Inst = nullptr) const override;
92 getIntImmCostIntrin(Intrinsic::ID IID, unsigned Idx, const APInt &Imm,
93 Type *Ty, TTI::TargetCostKind CostKind) const override;
94
95 /// \name EVL Support for predicated vectorization.
96 /// Whether the target supports the %evl parameter of VP intrinsic efficiently
97 /// in hardware. (see LLVM Language Reference - "Vector Predication
98 /// Intrinsics",
99 /// https://llvm.org/docs/LangRef.html#vector-predication-intrinsics and
100 /// "IR-level VP intrinsics",
101 /// https://llvm.org/docs/Proposals/VectorPredication.html#ir-level-vp-intrinsics).
102 bool hasActiveVectorLength() const override;
103
105 getPopcntSupport(unsigned TyWidth) const override;
106
108 unsigned Opcode, Type *InputTypeA, Type *InputTypeB, Type *AccumType,
110 TTI::PartialReductionExtendKind OpBExtend, std::optional<unsigned> BinOp,
112 std::optional<FastMathFlags> FMF) const override;
113
114 bool shouldExpandReduction(const IntrinsicInst *II) const override;
115 bool supportsScalableVectors() const override {
116 // VLEN=32 support is incomplete.
117 return ST->hasVInstructions() &&
118 (ST->getRealMinVLen() >= RISCV::RVVBitsPerBlock);
119 }
120 bool enableOrderedReductions() const override { return true; }
121 bool enableScalableVectorization() const override {
122 return ST->hasVInstructions();
123 }
125 return ST->hasVInstructions();
126 }
128 return ST->hasVInstructions() ? TailFoldingStyle::DataWithEVL
130 }
131 std::optional<unsigned> getVScaleForTuning() const override;
132
135
136 unsigned getRegUsageForType(Type *Ty) const override;
137
138 unsigned getMaximumVF(unsigned ElemWidth, unsigned Opcode) const override;
139
140 bool preferAlternateOpcodeVectorization() const override;
141
142 bool preferEpilogueVectorization(ElementCount Iters) const override {
143 // Epilogue vectorization is usually unprofitable - tail folding or
144 // a smaller VF would have been better. This a blunt hammer - we
145 // should re-examine this once vectorization is better tuned.
146 return false;
147 }
148
149 bool shouldConsiderVectorizationRegPressure() const override { return true; }
150
153 TTI::TargetCostKind CostKind) const override;
154
157
160 const TTI::PointersChainInfo &Info, Type *AccessTy,
161 const TTI::TargetCostKind CostKind) const override;
162
165 OptimizationRemarkEmitter *ORE) const override;
166
168 TTI::PeelingPreferences &PP) const override;
169
171 MemIntrinsicInfo &Info) const override;
172
173 unsigned getMinVectorRegisterBitWidth() const override {
174 return ST->useRVVForFixedLengthVectors() ? 16 : 0;
175 }
176
180 VectorType *SubTp, ArrayRef<const Value *> Args = {},
181 const Instruction *CxtI = nullptr) const override;
182
184 getScalarizationOverhead(VectorType *Ty, const APInt &DemandedElts,
185 bool Insert, bool Extract,
187 bool ForPoisonSrc = true, ArrayRef<Value *> VL = {},
189 TTI::VectorInstrContext::None) const override;
190
192 getIntrinsicInstrCost(const IntrinsicCostAttributes &ICA,
193 TTI::TargetCostKind CostKind) const override;
194
196 getAddressComputationCost(Type *PTy, ScalarEvolution *SE, const SCEV *Ptr,
197 TTI::TargetCostKind CostKind) const override;
198
200 unsigned Opcode, Type *VecTy, unsigned Factor, ArrayRef<unsigned> Indices,
201 Align Alignment, unsigned AddressSpace, TTI::TargetCostKind CostKind,
202 bool UseMaskForCond = false, bool UseMaskForGaps = false) const override;
203
204 InstructionCost getGatherScatterOpCost(const MemIntrinsicCostAttributes &MICA,
206
208 getExpandCompressMemoryOpCost(const MemIntrinsicCostAttributes &MICA,
210
211 InstructionCost getStridedMemoryOpCost(const MemIntrinsicCostAttributes &MICA,
213
216
218 getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src,
220 const Instruction *I = nullptr) const override;
221
223 getMinMaxReductionCost(Intrinsic::ID IID, VectorType *Ty, FastMathFlags FMF,
224 TTI::TargetCostKind CostKind) const override;
225
226 std::optional<InstructionCost> getCombinedArithmeticInstructionCost(
227 unsigned ISDOpcode, Type *Ty, TTI::TargetCostKind CostKind,
229 ArrayRef<const Value *> Args, const Instruction *CxtI) const;
230
232 getArithmeticReductionCost(unsigned Opcode, VectorType *Ty,
233 std::optional<FastMathFlags> FMF,
234 TTI::TargetCostKind CostKind) const override;
235
237 getExtendedReductionCost(unsigned Opcode, bool IsUnsigned, Type *ResTy,
238 VectorType *ValTy, std::optional<FastMathFlags> FMF,
239 TTI::TargetCostKind CostKind) const override;
240
242 unsigned Opcode, Type *Src, Align Alignment, unsigned AddressSpace,
245 const Instruction *I = nullptr) const override;
246
248 unsigned Opcode, Type *ValTy, Type *CondTy, CmpInst::Predicate VecPred,
252 const Instruction *I = nullptr) const override;
253
255 const Instruction *I = nullptr) const override;
256
260 unsigned Index, const Value *Op0, const Value *Op1,
262 TTI::VectorInstrContext::None) const override;
263
265 getIndexedVectorInstrCostFromEnd(unsigned Opcode, Type *Val,
267 unsigned Index) const override;
268
270 unsigned Opcode, Type *Ty, TTI::TargetCostKind CostKind,
274 const Instruction *CxtI = nullptr) const override;
275
276 bool isElementTypeLegalForScalableVector(Type *Ty) const override {
277 return TLI->isLegalElementTypeForRVV(TLI->getValueType(DL, Ty));
278 }
279
280 bool isLegalMaskedLoadStore(Type *DataType, Align Alignment) const {
281 if (!ST->hasVInstructions())
282 return false;
283
284 EVT DataTypeVT = TLI->getValueType(DL, DataType);
285
286 // Only support fixed vectors if we know the minimum vector size.
287 if (DataTypeVT.isFixedLengthVector() && !ST->useRVVForFixedLengthVectors())
288 return false;
289
290 EVT ElemType = DataTypeVT.getScalarType();
291 if (!ST->enableUnalignedVectorMem() && Alignment < ElemType.getStoreSize())
292 return false;
293
294 return TLI->isLegalElementTypeForRVV(ElemType);
295 }
296
297 bool isLegalMaskedLoad(Type *DataType, Align Alignment,
298 unsigned /*AddressSpace*/,
299 TTI::MaskKind /*MaskKind*/) const override {
300 return isLegalMaskedLoadStore(DataType, Alignment);
301 }
302 bool isLegalMaskedStore(Type *DataType, Align Alignment,
303 unsigned /*AddressSpace*/,
304 TTI::MaskKind /*MaskKind*/) const override {
305 return isLegalMaskedLoadStore(DataType, Alignment);
306 }
307
308 bool isLegalMaskedGatherScatter(Type *DataType, Align Alignment) const {
309 if (!ST->hasVInstructions())
310 return false;
311
312 EVT DataTypeVT = TLI->getValueType(DL, DataType);
313
314 // Only support fixed vectors if we know the minimum vector size.
315 if (DataTypeVT.isFixedLengthVector() && !ST->useRVVForFixedLengthVectors())
316 return false;
317
318 // We also need to check if the vector of address is valid.
319 EVT PointerTypeVT = EVT(TLI->getPointerTy(DL));
320 if (DataTypeVT.isScalableVector() &&
321 !TLI->isLegalElementTypeForRVV(PointerTypeVT))
322 return false;
323
324 EVT ElemType = DataTypeVT.getScalarType();
325 if (!ST->enableUnalignedVectorMem() && Alignment < ElemType.getStoreSize())
326 return false;
327
328 return TLI->isLegalElementTypeForRVV(ElemType);
329 }
330
331 bool isLegalMaskedGather(Type *DataType, Align Alignment) const override {
332 return isLegalMaskedGatherScatter(DataType, Alignment);
333 }
334 bool isLegalMaskedScatter(Type *DataType, Align Alignment) const override {
335 return isLegalMaskedGatherScatter(DataType, Alignment);
336 }
337
339 Align Alignment) const override {
340 // Scalarize masked gather for RV64 if EEW=64 indices aren't supported.
341 return ST->is64Bit() && !ST->hasVInstructionsI64();
342 }
343
345 Align Alignment) const override {
346 // Scalarize masked scatter for RV64 if EEW=64 indices aren't supported.
347 return ST->is64Bit() && !ST->hasVInstructionsI64();
348 }
349
350 bool isLegalStridedLoadStore(Type *DataType, Align Alignment) const override {
351 EVT DataTypeVT = TLI->getValueType(DL, DataType);
352 return TLI->isLegalStridedLoadStore(DataTypeVT, Alignment);
353 }
354
355 bool isLegalInterleavedAccessType(VectorType *VTy, unsigned Factor,
356 Align Alignment,
357 unsigned AddrSpace) const override {
358 return TLI->isLegalInterleavedAccessType(VTy, Factor, Alignment, AddrSpace,
359 DL);
360 }
361
362 bool isLegalMaskedExpandLoad(Type *DataType, Align Alignment) const override;
363
364 bool isLegalMaskedCompressStore(Type *DataTy, Align Alignment) const override;
365
366 bool isLegalBroadcastLoad(Type *ElementTy,
367 ElementCount NumElements) const override;
368
369 /// \returns How the target needs this vector-predicated operation to be
370 /// transformed.
372 getVPLegalizationStrategy(const VPIntrinsic &PI) const override {
374 static const Intrinsic::ID Supported[] = {
375 Intrinsic::experimental_vp_strided_load,
376 Intrinsic::experimental_vp_strided_store,
377 Intrinsic::experimental_vp_reverse,
378 Intrinsic::experimental_vp_splice,
379 Intrinsic::vp_cttz_elts,
380 Intrinsic::vp_gather,
381 Intrinsic::vp_load,
382 Intrinsic::vp_load_ff,
383 Intrinsic::vp_merge,
384 Intrinsic::vp_reduce_add,
385 Intrinsic::vp_reduce_and,
386 Intrinsic::vp_reduce_fadd,
387 Intrinsic::vp_reduce_fmax,
388 Intrinsic::vp_reduce_fmaximum,
389 Intrinsic::vp_reduce_fmin,
390 Intrinsic::vp_reduce_fminimum,
391 Intrinsic::vp_reduce_fmul,
392 Intrinsic::vp_reduce_mul,
393 Intrinsic::vp_reduce_or,
394 Intrinsic::vp_reduce_smax,
395 Intrinsic::vp_reduce_smin,
396 Intrinsic::vp_reduce_umax,
397 Intrinsic::vp_reduce_umin,
398 Intrinsic::vp_reduce_xor,
399 Intrinsic::vp_scatter,
400 Intrinsic::vp_sdiv,
401 Intrinsic::vp_srem,
402 Intrinsic::vp_store,
403 Intrinsic::vp_udiv,
404 Intrinsic::vp_urem};
405 if (!ST->hasVInstructions() ||
406 (PI.getIntrinsicID() == Intrinsic::vp_reduce_mul &&
408 ->getElementType()
409 ->getIntegerBitWidth() != 1) ||
410 !is_contained(Supported, PI.getIntrinsicID()))
413 }
414
416 ElementCount VF) const override {
417 if (!VF.isScalable())
418 return true;
419
420 Type *Ty = RdxDesc.getRecurrenceType();
421 if (!TLI->isLegalElementTypeForRVV(TLI->getValueType(DL, Ty)))
422 return false;
423
424 switch (RdxDesc.getRecurrenceKind()) {
425 case RecurKind::Add:
426 case RecurKind::Sub:
428 case RecurKind::And:
429 case RecurKind::Or:
430 case RecurKind::Xor:
431 case RecurKind::SMin:
432 case RecurKind::SMax:
433 case RecurKind::UMin:
434 case RecurKind::UMax:
435 case RecurKind::FMin:
436 case RecurKind::FMax:
439 return true;
440 case RecurKind::AnyOf:
441 case RecurKind::FAdd:
442 case RecurKind::FSub:
444 // We can't promote f16/bf16 fadd reductions and scalable vectors can't be
445 // expanded.
446 if (Ty->isBFloatTy() || (Ty->isHalfTy() && !ST->hasVInstructionsF16()))
447 return false;
448 return true;
449 case RecurKind::Mul:
450 case RecurKind::FMul:
458 return false;
459 case RecurKind::None:
460 llvm_unreachable("Unknown reduction kind.");
461 }
462 }
463
465 bool HasUnorderedReductions) const override {
466 // Don't interleave if the loop has been vectorized with scalable vectors.
467 if (VF.isScalable())
468 return 1;
469 // If the loop will not be vectorized, don't interleave the loop.
470 // Let regular unroll to unroll the loop.
471 return VF.isScalar() ? 1 : ST->getMaxInterleaveFactor();
472 }
473
474 bool enableInterleavedAccessVectorization() const override { return true; }
475
477 return ST->hasVInstructions();
478 }
479
480 unsigned getMinTripCountTailFoldingThreshold() const override;
481
483 unsigned getNumberOfRegisters(unsigned ClassID) const override {
484 switch (ClassID) {
486 // 31 = 32 GPR - x0 (zero register)
487 // FIXME: Should we exclude fixed registers like SP, TP or GP?
488 return 31;
490 if (ST->hasStdExtF())
491 return 32;
492 return 0;
494 // Although there are 32 vector registers, v0 is special in that it is the
495 // only register that can be used to hold a mask.
496 // FIXME: Should we conservatively return 31 as the number of usable
497 // vector registers?
498 return ST->hasVInstructions() ? 32 : 0;
499 }
500 llvm_unreachable("unknown register class");
501 }
502
504 getPreferredAddressingMode(const Loop *L, ScalarEvolution *SE) const override;
505
507 Type *Ty = nullptr) const override {
508 if (Vector)
510 if (!Ty)
512
513 Type *ScalarTy = Ty->getScalarType();
514 if ((ScalarTy->isHalfTy() && ST->hasStdExtZfhmin()) ||
515 (ScalarTy->isFloatTy() && ST->hasStdExtF()) ||
516 (ScalarTy->isDoubleTy() && ST->hasStdExtD())) {
518 }
519
521 }
522
523 const char *getRegisterClassName(unsigned ClassID) const override {
524 switch (ClassID) {
526 return "RISCV::GPRRC";
528 return "RISCV::FPRRC";
530 return "RISCV::VRRC";
531 }
532 llvm_unreachable("unknown register class");
533 }
534
536 const TargetTransformInfo::LSRCost &C2) const override;
537
539 const Instruction &I,
540 bool &AllowPromotionWithoutCommonHeader) const override;
541 std::optional<unsigned> getMinPageSize() const override { return 4096; }
542 /// Return true if the (vector) instruction I will be lowered to an
543 /// instruction with a scalar splat operand for the given Operand number.
544 bool canSplatOperand(Instruction *I, int Operand) const;
545 /// Return true if a vector instruction will lower to a target instruction
546 /// able to splat the given operand.
547 bool canSplatOperand(unsigned Opcode, int Operand) const;
548
550 SmallVectorImpl<Use *> &Ops) const override;
551
553 enableMemCmpExpansion(bool OptSize, bool IsZeroCmp) const override;
554
555 bool enableSelectOptimize() const override {
556 return ST->enableSelectOptimize();
557 }
558
559 bool shouldTreatInstructionLikeSelect(const Instruction *I) const override;
560
561 bool
563 const Attribute &Attr) const override;
564
565 std::optional<Instruction *>
567};
568
569} // end namespace llvm
570
571#endif // LLVM_LIB_TARGET_RISCV_RISCVTARGETTRANSFORMINFO_H
unsigned Imm
This file provides a helper that implements much of the TTI interface in terms of the target-independ...
static cl::opt< OutputCostKind > CostKind("cost-kind", cl::desc("Target cost kind"), cl::init(OutputCostKind::RecipThroughput), cl::values(clEnumValN(OutputCostKind::RecipThroughput, "throughput", "Reciprocal throughput"), clEnumValN(OutputCostKind::Latency, "latency", "Instruction latency"), clEnumValN(OutputCostKind::CodeSize, "code-size", "Code size"), clEnumValN(OutputCostKind::SizeAndLatency, "size-latency", "Code size and latency"), clEnumValN(OutputCostKind::All, "all", "Print all cost kinds")))
TargetTransformInfo::VPLegalization VPLegalization
const AbstractManglingParser< Derived, Alloc >::OperatorInfo AbstractManglingParser< Derived, Alloc >::Ops[]
#define F(x, y, z)
Definition MD5.cpp:54
#define I(x, y, z)
Definition MD5.cpp:57
uint64_t IntrinsicInst * II
This pass exposes codegen information to IR-level passes.
Class for arbitrary precision integers.
Definition APInt.h:78
Represent a constant reference to an array (0 or more elements consecutively in memory),...
Definition ArrayRef.h:40
Functions, function parameters, and return types can have attributes to indicate how they should be t...
Definition Attributes.h:105
InstructionCost getVectorInstrCost(unsigned Opcode, Type *Val, TTI::TargetCostKind CostKind, unsigned Index, const Value *Op0, const Value *Op1, TTI::VectorInstrContext VIC=TTI::VectorInstrContext::None) const override
BasicTTIImplBase(const TargetMachine *TM, const DataLayout &DL)
Value * getArgOperand(unsigned i) const
Predicate
This enumeration lists the possible predicates for CmpInst subclasses.
Definition InstrTypes.h:740
constexpr bool isScalar() const
Exactly one element.
Definition TypeSize.h:320
Class to represent fixed width SIMD vectors.
The core instruction combiner logic.
A wrapper class for inspecting calls to intrinsic functions.
Intrinsic::ID getIntrinsicID() const
Return the intrinsic ID of this intrinsic.
Represents a single loop in the control flow graph.
Definition LoopInfo.h:40
Machine Value Type.
Information for memory intrinsic cost model.
The optimization diagnostic interface.
unsigned getMaxInterleaveFactor(ElementCount VF, bool HasUnorderedReductions) const override
InstructionCost getExtendedReductionCost(unsigned Opcode, bool IsUnsigned, Type *ResTy, VectorType *ValTy, std::optional< FastMathFlags > FMF, TTI::TargetCostKind CostKind) const override
bool supportsScalableVectors() const override
InstructionCost getCFInstrCost(unsigned Opcode, TTI::TargetCostKind CostKind, const Instruction *I=nullptr) const override
InstructionCost getArithmeticInstrCost(unsigned Opcode, Type *Ty, TTI::TargetCostKind CostKind, TTI::OperandValueInfo Op1Info={TTI::OK_AnyValue, TTI::OP_None}, TTI::OperandValueInfo Op2Info={TTI::OK_AnyValue, TTI::OP_None}, ArrayRef< const Value * > Args={}, const Instruction *CxtI=nullptr) const override
bool shouldCopyAttributeWhenOutliningFrom(const Function *Caller, const Attribute &Attr) const override
InstructionCost getVectorInstrCost(unsigned Opcode, Type *Val, TTI::TargetCostKind CostKind, unsigned Index, const Value *Op0, const Value *Op1, TTI::VectorInstrContext VIC=TTI::VectorInstrContext::None) const override
bool isLegalMaskedExpandLoad(Type *DataType, Align Alignment) const override
InstructionCost getStridedMemoryOpCost(const MemIntrinsicCostAttributes &MICA, TTI::TargetCostKind CostKind) const
InstructionCost getShuffleCost(TTI::ShuffleKind Kind, VectorType *DstTy, VectorType *SrcTy, TTI::TargetCostKind CostKind, ArrayRef< int > Mask, int Index, VectorType *SubTp, ArrayRef< const Value * > Args={}, const Instruction *CxtI=nullptr) const override
bool isLegalMaskedLoadStore(Type *DataType, Align Alignment) const
InstructionCost getIntImmCostIntrin(Intrinsic::ID IID, unsigned Idx, const APInt &Imm, Type *Ty, TTI::TargetCostKind CostKind) const override
unsigned getMinTripCountTailFoldingThreshold() const override
unsigned getRegisterClassForType(bool Vector, Type *Ty=nullptr) const override
TTI::AddressingModeKind getPreferredAddressingMode(const Loop *L, ScalarEvolution *SE) const override
InstructionCost getAddressComputationCost(Type *PTy, ScalarEvolution *SE, const SCEV *Ptr, TTI::TargetCostKind CostKind) const override
bool preferTailFoldingOverEpilogue(TailFoldingInfo *TFI) const override
InstructionCost getStoreImmCost(Type *VecTy, TTI::OperandValueInfo OpInfo, TTI::TargetCostKind CostKind) const
Return the cost of materializing an immediate for a value operand of a store instruction.
bool isLegalMaskedStore(Type *DataType, Align Alignment, unsigned, TTI::MaskKind) const override
bool getTgtMemIntrinsic(IntrinsicInst *Inst, MemIntrinsicInfo &Info) const override
bool isElementTypeLegalForScalableVector(Type *Ty) const override
bool enableMaskedInterleavedAccessVectorization() const override
bool isLegalInterleavedAccessType(VectorType *VTy, unsigned Factor, Align Alignment, unsigned AddrSpace) const override
std::optional< unsigned > getMinPageSize() const override
InstructionCost getCostOfKeepingLiveOverCall(ArrayRef< Type * > Tys) const override
std::optional< InstructionCost > getCombinedArithmeticInstructionCost(unsigned ISDOpcode, Type *Ty, TTI::TargetCostKind CostKind, TTI::OperandValueInfo Opd1Info, TTI::OperandValueInfo Opd2Info, ArrayRef< const Value * > Args, const Instruction *CxtI) const
Check to see if this instruction is expected to be combined to a simpler operation during/before lowe...
bool enableSelectOptimize() const override
bool hasActiveVectorLength() const override
InstructionCost getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src, TTI::CastContextHint CCH, TTI::TargetCostKind CostKind, const Instruction *I=nullptr) const override
bool shouldConsiderVectorizationRegPressure() const override
InstructionCost getCmpSelInstrCost(unsigned Opcode, Type *ValTy, Type *CondTy, CmpInst::Predicate VecPred, TTI::TargetCostKind CostKind, TTI::OperandValueInfo Op1Info={TTI::OK_AnyValue, TTI::OP_None}, TTI::OperandValueInfo Op2Info={TTI::OK_AnyValue, TTI::OP_None}, const Instruction *I=nullptr) const override
InstructionCost getIndexedVectorInstrCostFromEnd(unsigned Opcode, Type *Val, TTI::TargetCostKind CostKind, unsigned Index) const override
void getUnrollingPreferences(Loop *L, ScalarEvolution &SE, TTI::UnrollingPreferences &UP, OptimizationRemarkEmitter *ORE) const override
bool isLegalBroadcastLoad(Type *ElementTy, ElementCount NumElements) const override
bool forceScalarizeMaskedScatter(VectorType *VTy, Align Alignment) const override
const char * getRegisterClassName(unsigned ClassID) const override
InstructionCost getIntImmCostInst(unsigned Opcode, unsigned Idx, const APInt &Imm, Type *Ty, TTI::TargetCostKind CostKind, Instruction *Inst=nullptr) const override
InstructionCost getMinMaxReductionCost(Intrinsic::ID IID, VectorType *Ty, FastMathFlags FMF, TTI::TargetCostKind CostKind) const override
Try to calculate op costs for min/max reduction operations.
bool canSplatOperand(Instruction *I, int Operand) const
Return true if the (vector) instruction I will be lowered to an instruction with a scalar splat opera...
bool enableInterleavedAccessVectorization() const override
TailFoldingStyle getPreferredTailFoldingStyle() const override
bool isLegalMaskedGatherScatter(Type *DataType, Align Alignment) const
bool isLSRCostLess(const TargetTransformInfo::LSRCost &C1, const TargetTransformInfo::LSRCost &C2) const override
bool isLegalStridedLoadStore(Type *DataType, Align Alignment) const override
unsigned getRegUsageForType(Type *Ty) const override
InstructionCost getInterleavedMemoryOpCost(unsigned Opcode, Type *VecTy, unsigned Factor, ArrayRef< unsigned > Indices, Align Alignment, unsigned AddressSpace, TTI::TargetCostKind CostKind, bool UseMaskForCond=false, bool UseMaskForGaps=false) const override
unsigned getMinVectorRegisterBitWidth() const override
bool isLegalMaskedScatter(Type *DataType, Align Alignment) const override
bool isLegalMaskedCompressStore(Type *DataTy, Align Alignment) const override
InstructionCost getGatherScatterOpCost(const MemIntrinsicCostAttributes &MICA, TTI::TargetCostKind CostKind) const
bool enableOrderedReductions() const override
InstructionCost getPartialReductionCost(unsigned Opcode, Type *InputTypeA, Type *InputTypeB, Type *AccumType, ElementCount VF, TTI::PartialReductionExtendKind OpAExtend, TTI::PartialReductionExtendKind OpBExtend, std::optional< unsigned > BinOp, TTI::TargetCostKind CostKind, std::optional< FastMathFlags > FMF) const override
bool shouldTreatInstructionLikeSelect(const Instruction *I) const override
InstructionCost getExpandCompressMemoryOpCost(const MemIntrinsicCostAttributes &MICA, TTI::TargetCostKind CostKind) const
RISCVTTIImpl(const RISCVTargetMachine *TM, const Function &F)
TargetTransformInfo::VPLegalization getVPLegalizationStrategy(const VPIntrinsic &PI) const override
bool preferEpilogueVectorization(ElementCount Iters) const override
bool preferAlternateOpcodeVectorization() const override
bool isProfitableToSinkOperands(Instruction *I, SmallVectorImpl< Use * > &Ops) const override
Check if sinking I's operands to I's basic block is profitable, because the operands can be folded in...
unsigned getNumberOfRegisters(unsigned ClassID) const override
bool shouldExpandReduction(const IntrinsicInst *II) const override
std::optional< unsigned > getVScaleForTuning() const override
InstructionCost getMemIntrinsicInstrCost(const MemIntrinsicCostAttributes &MICA, TTI::TargetCostKind CostKind) const override
Get memory intrinsic cost based on arguments.
bool isLegalMaskedGather(Type *DataType, Align Alignment) const override
bool isLegalMaskedLoad(Type *DataType, Align Alignment, unsigned, TTI::MaskKind) const override
InstructionCost getPointersChainCost(ArrayRef< const Value * > Ptrs, const Value *Base, const TTI::PointersChainInfo &Info, Type *AccessTy, const TTI::TargetCostKind CostKind) const override
unsigned getMaximumVF(unsigned ElemWidth, unsigned Opcode) const override
TTI::MemCmpExpansionOptions enableMemCmpExpansion(bool OptSize, bool IsZeroCmp) const override
InstructionCost getScalarizationOverhead(VectorType *Ty, const APInt &DemandedElts, bool Insert, bool Extract, TTI::TargetCostKind CostKind, bool ForPoisonSrc=true, ArrayRef< Value * > VL={}, TTI::VectorInstrContext VIC=TTI::VectorInstrContext::None) const override
Estimate the overhead of scalarizing an instruction.
InstructionCost getMemoryOpCost(unsigned Opcode, Type *Src, Align Alignment, unsigned AddressSpace, TTI::TargetCostKind CostKind, TTI::OperandValueInfo OpdInfo={TTI::OK_AnyValue, TTI::OP_None}, const Instruction *I=nullptr) const override
bool isLegalToVectorizeReduction(const RecurrenceDescriptor &RdxDesc, ElementCount VF) const override
bool enableScalableVectorization() const override
InstructionCost getIntrinsicInstrCost(const IntrinsicCostAttributes &ICA, TTI::TargetCostKind CostKind) const override
Get intrinsic cost based on arguments.
InstructionCost getMaskedMemoryOpCost(const MemIntrinsicCostAttributes &MICA, TTI::TargetCostKind CostKind) const
InstructionCost getArithmeticReductionCost(unsigned Opcode, VectorType *Ty, std::optional< FastMathFlags > FMF, TTI::TargetCostKind CostKind) const override
TypeSize getRegisterBitWidth(TargetTransformInfo::RegisterKind K) const override
void getPeelingPreferences(Loop *L, ScalarEvolution &SE, TTI::PeelingPreferences &PP) const override
std::optional< Instruction * > instCombineIntrinsic(InstCombiner &IC, IntrinsicInst &II) const override
bool shouldConsiderAddressTypePromotion(const Instruction &I, bool &AllowPromotionWithoutCommonHeader) const override
See if I should be considered for address type promotion.
InstructionCost getIntImmCost(const APInt &Imm, Type *Ty, TTI::TargetCostKind CostKind) const override
bool forceScalarizeMaskedGather(VectorType *VTy, Align Alignment) const override
TargetTransformInfo::PopcntSupportKind getPopcntSupport(unsigned TyWidth) const override
The RecurrenceDescriptor is used to identify recurrences variables in a loop.
Type * getRecurrenceType() const
Returns the type of the recurrence.
RecurKind getRecurrenceKind() const
The main scalar evolution driver.
This class consists of common code factored out of the SmallVector class to reduce code duplication b...
virtual const DataLayout & getDataLayout() const
This pass provides access to the codegen interfaces that are needed for IR-level transformations.
MaskKind
Some targets only support masked load/store with a constant mask.
TargetCostKind
The kind of cost model.
PopcntSupportKind
Flags indicating the kind of support for population count.
llvm::VectorInstrContext VectorInstrContext
AddressingModeKind
Which addressing mode Loop Strength Reduction will try to generate.
ShuffleKind
The various kinds of shuffle patterns for vector queries.
CastContextHint
Represents a hint about the context in which a cast is used.
@ None
The cast is not used with a load/store of any kind.
The instances of the Type class are immutable: once they are created, they are never changed.
Definition Type.h:46
bool isFloatTy() const
Return true if this is 'float', a 32-bit IEEE fp type.
Definition Type.h:155
Type * getScalarType() const
If this is a vector type, return the element type, otherwise return 'this'.
Definition Type.h:368
bool isHalfTy() const
Return true if this is 'half', a 16-bit IEEE fp type.
Definition Type.h:144
bool isDoubleTy() const
Return true if this is 'double', a 64-bit IEEE fp type.
Definition Type.h:158
This is the common base class for vector predication intrinsics.
LLVM Value Representation.
Definition Value.h:75
Type * getType() const
All values are typed, get the type of this value.
Definition Value.h:255
Base class of all SIMD vector types.
constexpr bool isScalable() const
Returns whether the quantity is scaled by a runtime quantity (vscale).
Definition TypeSize.h:168
#define llvm_unreachable(msg)
Marks that the current location is not supposed to be reachable.
constexpr char Args[]
Key for Kernel::Metadata::mArgs.
static constexpr unsigned RVVBitsPerBlock
friend class Instruction
Iterator for Instructions in a `BasicBlock.
Definition BasicBlock.h:73
This is an optimization pass for GlobalISel generic memory operations.
RelativeUniformCounterPtr ValuesPtrExpr VTableAddr Value
Definition InstrProf.h:143
@ UMin
Unsigned integer min implemented in terms of select(cmp()).
@ FMinimumNum
FP min with llvm.minimumnum semantics.
@ FindIV
FindIV reduction with select(icmp(),x,y) where one of (x,y) is a loop induction variable (increasing ...
@ Or
Bitwise or logical OR of integers.
@ FMinimum
FP min with llvm.minimum semantics.
@ FMaxNum
FP max with llvm.maxnum semantics including NaNs.
@ Mul
Product of integers.
@ FSub
Subtraction of floats.
@ FAddChainWithSubs
A chain of fadds and fsubs.
@ None
Not a recurrence.
@ AnyOf
AnyOf reduction with select(cmp(),x,y) where one of (x,y) is loop invariant, and both x and y are int...
@ Xor
Bitwise or logical XOR of integers.
@ FindLast
FindLast reduction with select(cmp(),x,y) where x and y.
@ FMax
FP max implemented in terms of select(cmp()).
@ FMaximum
FP max with llvm.maximum semantics.
@ FMulAdd
Sum of float products with llvm.fmuladd(a * b + sum).
@ FMul
Product of floats.
@ SMax
Signed integer max implemented in terms of select(cmp()).
@ And
Bitwise or logical AND of integers.
@ SMin
Signed integer min implemented in terms of select(cmp()).
@ FMin
FP min implemented in terms of select(cmp()).
@ FMinNum
FP min with llvm.minnum semantics including NaNs.
@ Sub
Subtraction of integers.
@ Add
Sum of integers.
@ AddChainWithSubs
A chain of adds and subs.
@ FAdd
Sum of floats.
@ FMaximumNum
FP max with llvm.maximumnum semantics.
@ UMax
Unsigned integer max implemented in terms of select(cmp()).
ArrayRef(const T &OneElt) -> ArrayRef< T >
decltype(auto) cast(const From &Val)
cast<X> - Return the argument parameter cast to the specified type.
Definition Casting.h:559
bool is_contained(R &&Range, const E &Element)
Returns true if Element is found in Range.
Definition STLExtras.h:1947
@ None
Don't use tail folding.
@ DataWithEVL
Use predicated EVL instructions for tail-folding.
This struct is a compact representation of a valid (non-zero power of two) alignment.
Definition Alignment.h:39
Extended Value Type.
Definition ValueTypes.h:35
TypeSize getStoreSize() const
Return the number of bytes overwritten by a store of the specified value type.
Definition ValueTypes.h:418
bool isFixedLengthVector() const
Definition ValueTypes.h:199
EVT getScalarType() const
If this is a vector type, return the element type, otherwise return this.
Definition ValueTypes.h:346
bool isScalableVector() const
Return true if this is a vector type where the runtime length is machine dependent.
Definition ValueTypes.h:187
Information about a load/store intrinsic defined by the target.
Returns options for expansion of memcmp. IsZeroCmp is.
Describe known properties for a set of pointers.
Parameters that control the generic loop unrolling transformation.