LLVM 23.0.0git
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> getMaxVScale() const override;
132 std::optional<unsigned> getVScaleForTuning() const override;
133
136
137 unsigned getRegUsageForType(Type *Ty) const override;
138
139 unsigned getMaximumVF(unsigned ElemWidth, unsigned Opcode) const override;
140
141 bool preferAlternateOpcodeVectorization() const override;
142
143 bool preferEpilogueVectorization(ElementCount Iters) const override {
144 // Epilogue vectorization is usually unprofitable - tail folding or
145 // a smaller VF would have been better. This a blunt hammer - we
146 // should re-examine this once vectorization is better tuned.
147 return false;
148 }
149
150 bool shouldConsiderVectorizationRegPressure() const override { return true; }
151
154 TTI::TargetCostKind CostKind) const override;
155
158
161 const TTI::PointersChainInfo &Info, Type *AccessTy,
162 TTI::TargetCostKind CostKind) const override;
163
166 OptimizationRemarkEmitter *ORE) const override;
167
169 TTI::PeelingPreferences &PP) const override;
170
172 MemIntrinsicInfo &Info) const override;
173
174 unsigned getMinVectorRegisterBitWidth() const override {
175 return ST->useRVVForFixedLengthVectors() ? 16 : 0;
176 }
177
181 VectorType *SubTp, ArrayRef<const Value *> Args = {},
182 const Instruction *CxtI = nullptr) const override;
183
185 getScalarizationOverhead(VectorType *Ty, const APInt &DemandedElts,
186 bool Insert, bool Extract,
188 bool ForPoisonSrc = true, ArrayRef<Value *> VL = {},
190 TTI::VectorInstrContext::None) const override;
191
193 getIntrinsicInstrCost(const IntrinsicCostAttributes &ICA,
194 TTI::TargetCostKind CostKind) const override;
195
197 getAddressComputationCost(Type *PTy, ScalarEvolution *SE, const SCEV *Ptr,
198 TTI::TargetCostKind CostKind) const override;
199
201 unsigned Opcode, Type *VecTy, unsigned Factor, ArrayRef<unsigned> Indices,
202 Align Alignment, unsigned AddressSpace, TTI::TargetCostKind CostKind,
203 bool UseMaskForCond = false, bool UseMaskForGaps = false) const override;
204
205 InstructionCost getGatherScatterOpCost(const MemIntrinsicCostAttributes &MICA,
207
209 getExpandCompressMemoryOpCost(const MemIntrinsicCostAttributes &MICA,
211
212 InstructionCost getStridedMemoryOpCost(const MemIntrinsicCostAttributes &MICA,
214
217
219 getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src,
221 const Instruction *I = nullptr) const override;
222
224 getMinMaxReductionCost(Intrinsic::ID IID, VectorType *Ty, FastMathFlags FMF,
225 TTI::TargetCostKind CostKind) const override;
226
227 std::optional<InstructionCost> getCombinedArithmeticInstructionCost(
228 unsigned ISDOpcode, Type *Ty, TTI::TargetCostKind CostKind,
230 ArrayRef<const Value *> Args, const Instruction *CxtI) const;
231
233 getArithmeticReductionCost(unsigned Opcode, VectorType *Ty,
234 std::optional<FastMathFlags> FMF,
235 TTI::TargetCostKind CostKind) const override;
236
238 getExtendedReductionCost(unsigned Opcode, bool IsUnsigned, Type *ResTy,
239 VectorType *ValTy, std::optional<FastMathFlags> FMF,
240 TTI::TargetCostKind CostKind) const override;
241
243 unsigned Opcode, Type *Src, Align Alignment, unsigned AddressSpace,
246 const Instruction *I = nullptr) const override;
247
249 unsigned Opcode, Type *ValTy, Type *CondTy, CmpInst::Predicate VecPred,
253 const Instruction *I = nullptr) const override;
254
256 const Instruction *I = nullptr) const override;
257
261 unsigned Index, const Value *Op0, const Value *Op1,
263 TTI::VectorInstrContext::None) const override;
264
266 getIndexedVectorInstrCostFromEnd(unsigned Opcode, Type *Val,
268 unsigned Index) const override;
269
271 unsigned Opcode, Type *Ty, TTI::TargetCostKind CostKind,
275 const Instruction *CxtI = nullptr) const override;
276
277 bool isElementTypeLegalForScalableVector(Type *Ty) const override {
278 return TLI->isLegalElementTypeForRVV(TLI->getValueType(DL, Ty));
279 }
280
281 bool isLegalMaskedLoadStore(Type *DataType, Align Alignment) const {
282 if (!ST->hasVInstructions())
283 return false;
284
285 EVT DataTypeVT = TLI->getValueType(DL, DataType);
286
287 // Only support fixed vectors if we know the minimum vector size.
288 if (DataTypeVT.isFixedLengthVector() && !ST->useRVVForFixedLengthVectors())
289 return false;
290
291 EVT ElemType = DataTypeVT.getScalarType();
292 if (!ST->enableUnalignedVectorMem() && Alignment < ElemType.getStoreSize())
293 return false;
294
295 return TLI->isLegalElementTypeForRVV(ElemType);
296 }
297
298 bool isLegalMaskedLoad(Type *DataType, Align Alignment,
299 unsigned /*AddressSpace*/,
300 TTI::MaskKind /*MaskKind*/) const override {
301 return isLegalMaskedLoadStore(DataType, Alignment);
302 }
303 bool isLegalMaskedStore(Type *DataType, Align Alignment,
304 unsigned /*AddressSpace*/,
305 TTI::MaskKind /*MaskKind*/) const override {
306 return isLegalMaskedLoadStore(DataType, Alignment);
307 }
308
309 bool isLegalMaskedGatherScatter(Type *DataType, Align Alignment) const {
310 if (!ST->hasVInstructions())
311 return false;
312
313 EVT DataTypeVT = TLI->getValueType(DL, DataType);
314
315 // Only support fixed vectors if we know the minimum vector size.
316 if (DataTypeVT.isFixedLengthVector() && !ST->useRVVForFixedLengthVectors())
317 return false;
318
319 // We also need to check if the vector of address is valid.
320 EVT PointerTypeVT = EVT(TLI->getPointerTy(DL));
321 if (DataTypeVT.isScalableVector() &&
322 !TLI->isLegalElementTypeForRVV(PointerTypeVT))
323 return false;
324
325 EVT ElemType = DataTypeVT.getScalarType();
326 if (!ST->enableUnalignedVectorMem() && Alignment < ElemType.getStoreSize())
327 return false;
328
329 return TLI->isLegalElementTypeForRVV(ElemType);
330 }
331
332 bool isLegalMaskedGather(Type *DataType, Align Alignment) const override {
333 return isLegalMaskedGatherScatter(DataType, Alignment);
334 }
335 bool isLegalMaskedScatter(Type *DataType, Align Alignment) const override {
336 return isLegalMaskedGatherScatter(DataType, Alignment);
337 }
338
340 Align Alignment) const override {
341 // Scalarize masked gather for RV64 if EEW=64 indices aren't supported.
342 return ST->is64Bit() && !ST->hasVInstructionsI64();
343 }
344
346 Align Alignment) const override {
347 // Scalarize masked scatter for RV64 if EEW=64 indices aren't supported.
348 return ST->is64Bit() && !ST->hasVInstructionsI64();
349 }
350
351 bool isLegalStridedLoadStore(Type *DataType, Align Alignment) const override {
352 EVT DataTypeVT = TLI->getValueType(DL, DataType);
353 return TLI->isLegalStridedLoadStore(DataTypeVT, Alignment);
354 }
355
356 bool isLegalInterleavedAccessType(VectorType *VTy, unsigned Factor,
357 Align Alignment,
358 unsigned AddrSpace) const override {
359 return TLI->isLegalInterleavedAccessType(VTy, Factor, Alignment, AddrSpace,
360 DL);
361 }
362
363 bool isLegalMaskedExpandLoad(Type *DataType, Align Alignment) const override;
364
365 bool isLegalMaskedCompressStore(Type *DataTy, Align Alignment) const override;
366
367 /// \returns How the target needs this vector-predicated operation to be
368 /// transformed.
370 getVPLegalizationStrategy(const VPIntrinsic &PI) const override {
372 static const Intrinsic::ID Supported[] = {
373 Intrinsic::experimental_vp_strided_load,
374 Intrinsic::experimental_vp_strided_store,
375 Intrinsic::experimental_vp_reverse,
376 Intrinsic::experimental_vp_splice,
377 Intrinsic::vp_abs,
378 Intrinsic::vp_add,
379 Intrinsic::vp_and,
380 Intrinsic::vp_ashr,
381 Intrinsic::vp_cttz_elts,
382 Intrinsic::vp_fcmp,
383 Intrinsic::vp_fptrunc,
384 Intrinsic::vp_frem,
385 Intrinsic::vp_fshl,
386 Intrinsic::vp_fshr,
387 Intrinsic::vp_gather,
388 Intrinsic::vp_icmp,
389 Intrinsic::vp_inttoptr,
390 Intrinsic::vp_is_fpclass,
391 Intrinsic::vp_load,
392 Intrinsic::vp_load_ff,
393 Intrinsic::vp_lshr,
394 Intrinsic::vp_merge,
395 Intrinsic::vp_mul,
396 Intrinsic::vp_or,
397 Intrinsic::vp_ptrtoint,
398 Intrinsic::vp_reduce_add,
399 Intrinsic::vp_reduce_and,
400 Intrinsic::vp_reduce_fadd,
401 Intrinsic::vp_reduce_fmax,
402 Intrinsic::vp_reduce_fmaximum,
403 Intrinsic::vp_reduce_fmin,
404 Intrinsic::vp_reduce_fminimum,
405 Intrinsic::vp_reduce_fmul,
406 Intrinsic::vp_reduce_mul,
407 Intrinsic::vp_reduce_or,
408 Intrinsic::vp_reduce_smax,
409 Intrinsic::vp_reduce_smin,
410 Intrinsic::vp_reduce_umax,
411 Intrinsic::vp_reduce_umin,
412 Intrinsic::vp_reduce_xor,
413 Intrinsic::vp_scatter,
414 Intrinsic::vp_sdiv,
415 Intrinsic::vp_select,
416 Intrinsic::vp_sext,
417 Intrinsic::vp_shl,
418 Intrinsic::vp_srem,
419 Intrinsic::vp_store,
420 Intrinsic::vp_sub,
421 Intrinsic::vp_trunc,
422 Intrinsic::vp_udiv,
423 Intrinsic::vp_urem,
424 Intrinsic::vp_xor,
425 Intrinsic::vp_zext};
426 if (!ST->hasVInstructions() ||
427 (PI.getIntrinsicID() == Intrinsic::vp_reduce_mul &&
429 ->getElementType()
430 ->getIntegerBitWidth() != 1) ||
431 !is_contained(Supported, PI.getIntrinsicID()))
434 }
435
437 ElementCount VF) const override {
438 if (!VF.isScalable())
439 return true;
440
441 Type *Ty = RdxDesc.getRecurrenceType();
442 if (!TLI->isLegalElementTypeForRVV(TLI->getValueType(DL, Ty)))
443 return false;
444
445 switch (RdxDesc.getRecurrenceKind()) {
446 case RecurKind::Add:
447 case RecurKind::Sub:
449 case RecurKind::And:
450 case RecurKind::Or:
451 case RecurKind::Xor:
452 case RecurKind::SMin:
453 case RecurKind::SMax:
454 case RecurKind::UMin:
455 case RecurKind::UMax:
456 case RecurKind::FMin:
457 case RecurKind::FMax:
459 return true;
460 case RecurKind::AnyOf:
461 case RecurKind::FAdd:
463 // We can't promote f16/bf16 fadd reductions and scalable vectors can't be
464 // expanded.
465 if (Ty->isBFloatTy() || (Ty->isHalfTy() && !ST->hasVInstructionsF16()))
466 return false;
467 return true;
468 default:
469 return false;
470 }
471 }
472
473 unsigned getMaxInterleaveFactor(ElementCount VF) const override {
474 // Don't interleave if the loop has been vectorized with scalable vectors.
475 if (VF.isScalable())
476 return 1;
477 // If the loop will not be vectorized, don't interleave the loop.
478 // Let regular unroll to unroll the loop.
479 return VF.isScalar() ? 1 : ST->getMaxInterleaveFactor();
480 }
481
482 bool enableInterleavedAccessVectorization() const override { return true; }
483
485 return ST->hasVInstructions();
486 }
487
488 unsigned getMinTripCountTailFoldingThreshold() const override;
489
491 unsigned getNumberOfRegisters(unsigned ClassID) const override {
492 switch (ClassID) {
494 // 31 = 32 GPR - x0 (zero register)
495 // FIXME: Should we exclude fixed registers like SP, TP or GP?
496 return 31;
498 if (ST->hasStdExtF())
499 return 32;
500 return 0;
502 // Although there are 32 vector registers, v0 is special in that it is the
503 // only register that can be used to hold a mask.
504 // FIXME: Should we conservatively return 31 as the number of usable
505 // vector registers?
506 return ST->hasVInstructions() ? 32 : 0;
507 }
508 llvm_unreachable("unknown register class");
509 }
510
512 getPreferredAddressingMode(const Loop *L, ScalarEvolution *SE) const override;
513
515 Type *Ty = nullptr) const override {
516 if (Vector)
518 if (!Ty)
520
521 Type *ScalarTy = Ty->getScalarType();
522 if ((ScalarTy->isHalfTy() && ST->hasStdExtZfhmin()) ||
523 (ScalarTy->isFloatTy() && ST->hasStdExtF()) ||
524 (ScalarTy->isDoubleTy() && ST->hasStdExtD())) {
526 }
527
529 }
530
531 const char *getRegisterClassName(unsigned ClassID) const override {
532 switch (ClassID) {
534 return "RISCV::GPRRC";
536 return "RISCV::FPRRC";
538 return "RISCV::VRRC";
539 }
540 llvm_unreachable("unknown register class");
541 }
542
544 const TargetTransformInfo::LSRCost &C2) const override;
545
547 const Instruction &I,
548 bool &AllowPromotionWithoutCommonHeader) const override;
549 std::optional<unsigned> getMinPageSize() const override { return 4096; }
550 /// Return true if the (vector) instruction I will be lowered to an
551 /// instruction with a scalar splat operand for the given Operand number.
552 bool canSplatOperand(Instruction *I, int Operand) const;
553 /// Return true if a vector instruction will lower to a target instruction
554 /// able to splat the given operand.
555 bool canSplatOperand(unsigned Opcode, int Operand) const;
556
558 SmallVectorImpl<Use *> &Ops) const override;
559
561 enableMemCmpExpansion(bool OptSize, bool IsZeroCmp) const override;
562
563 bool enableSelectOptimize() const override {
564 return ST->enableSelectOptimize();
565 }
566
567 bool shouldTreatInstructionLikeSelect(const Instruction *I) const override;
568
569 bool
571 const Attribute &Attr) const override;
572
573 std::optional<Instruction *>
575};
576
577} // end namespace llvm
578
579#endif // LLVM_LIB_TARGET_RISCV_RISCVTARGETTRANSFORMINFO_H
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
ArrayRef - 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:676
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.
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
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 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
unsigned getMaxInterleaveFactor(ElementCount VF) 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
std::optional< unsigned > getMaxVScale() 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 getShuffleCost(TTI::ShuffleKind Kind, VectorType *DstTy, VectorType *SrcTy, ArrayRef< int > Mask, TTI::TargetCostKind CostKind, int Index, VectorType *SubTp, ArrayRef< const Value * > Args={}, const Instruction *CxtI=nullptr) const override
unsigned getMaximumVF(unsigned ElemWidth, unsigned Opcode) const override
InstructionCost getPointersChainCost(ArrayRef< const Value * > Ptrs, const Value *Base, const TTI::PointersChainInfo &Info, Type *AccessTy, TTI::TargetCostKind CostKind) 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.
VectorInstrContext
Represents a hint about the context in which an insert/extract is used.
@ None
The insert/extract is not used with a load/store.
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.
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.
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:370
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.
FunctionAddr VTableAddr Value
Definition InstrProf.h:137
@ UMin
Unsigned integer min implemented in terms of select(cmp()).
@ Or
Bitwise or logical OR of integers.
@ 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()).
@ FMulAdd
Sum of float products with llvm.fmuladd(a * b + sum).
@ 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()).
@ Sub
Subtraction of integers.
@ Add
Sum of integers.
@ AddChainWithSubs
A chain of adds and subs.
@ FAdd
Sum of floats.
@ 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:403
bool isFixedLengthVector() const
Definition ValueTypes.h:189
EVT getScalarType() const
If this is a vector type, return the element type, otherwise return this.
Definition ValueTypes.h:331
bool isScalableVector() const
Return true if this is a vector type where the runtime length is machine dependent.
Definition ValueTypes.h:182
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.