1//===-- RISCVTargetTransformInfo.cpp - RISC-V specific TTI ----------------===//
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
9#include "RISCVTargetTransformInfo.h"
10#include "MCTargetDesc/RISCVMatInt.h"
11#include "llvm/ADT/STLExtras.h"
12#include "llvm/Analysis/TargetTransformInfo.h"
13#include "llvm/CodeGen/BasicTTIImpl.h"
14#include "llvm/CodeGen/CostTable.h"
15#include "llvm/CodeGen/TargetLowering.h"
16#include "llvm/CodeGen/ValueTypes.h"
17#include "llvm/IR/Instructions.h"
18#include "llvm/IR/IntrinsicsRISCV.h"
19#include "llvm/IR/PatternMatch.h"
20#include "llvm/Transforms/InstCombine/InstCombiner.h"
21#include <cmath>
22#include <optional>
23using namespace llvm;
24using namespace llvm::PatternMatch;
25
26#define DEBUG_TYPE "riscvtti"
27
28static cl::opt<unsigned> RVVRegisterWidthLMUL(
29 "riscv-v-register-bit-width-lmul",
30 cl::desc(
31 "The LMUL to use for getRegisterBitWidth queries. Affects LMUL used "
32 "by autovectorized code. Fractional LMULs are not supported."),
33 cl::init(Val: 2), cl::Hidden);
34
35static cl::opt<unsigned> SLPMaxVF(
36 "riscv-v-slp-max-vf",
37 cl::desc(
38 "Overrides result used for getMaximumVF query which is used "
39 "exclusively by SLP vectorizer."),
40 cl::Hidden);
41
42static cl::opt<unsigned>
43 RVVMinTripCount("riscv-v-min-trip-count",
44 cl::desc("Set the lower bound of a trip count to decide on "
45 "vectorization while tail-folding."),
46 cl::init(Val: 5), cl::Hidden);
47
48static cl::opt<bool> EnableOrLikeSelectOpt("enable-riscv-or-like-select",
49 cl::init(Val: true), cl::Hidden);
50
51InstructionCost
52RISCVTTIImpl::getRISCVInstructionCost(ArrayRef<unsigned> OpCodes, MVT VT,
53 TTI::TargetCostKind CostKind) const {
54 // Check if the type is valid for all CostKind
55 if (!VT.isVector())
56 return InstructionCost::getInvalid();
57 size_t NumInstr = OpCodes.size();
58 if (CostKind == TTI::TCK_CodeSize)
59 return NumInstr;
60 InstructionCost LMULCost = TLI->getLMULCost(VT);
61 if ((CostKind != TTI::TCK_RecipThroughput) && (CostKind != TTI::TCK_Latency))
62 return LMULCost * NumInstr;
63 InstructionCost Cost = 0;
64 for (auto Op : OpCodes) {
65 switch (Op) {
66 case RISCV::VRGATHER_VI:
67 Cost += TLI->getVRGatherVICost(VT);
68 break;
69 case RISCV::VRGATHER_VV:
70 Cost += TLI->getVRGatherVVCost(VT);
71 break;
72 case RISCV::VSLIDEUP_VI:
73 case RISCV::VSLIDEDOWN_VI:
74 Cost += TLI->getVSlideVICost(VT);
75 break;
76 case RISCV::VSLIDEUP_VX:
77 case RISCV::VSLIDEDOWN_VX:
78 Cost += TLI->getVSlideVXCost(VT);
79 break;
80 case RISCV::VREDMAX_VS:
81 case RISCV::VREDMIN_VS:
82 case RISCV::VREDMAXU_VS:
83 case RISCV::VREDMINU_VS:
84 case RISCV::VREDSUM_VS:
85 case RISCV::VREDAND_VS:
86 case RISCV::VREDOR_VS:
87 case RISCV::VREDXOR_VS:
88 case RISCV::VFREDMAX_VS:
89 case RISCV::VFREDMIN_VS:
90 case RISCV::VFREDUSUM_VS: {
91 unsigned VL = VT.getVectorMinNumElements();
92 if (!VT.isFixedLengthVector())
93 VL *= *getVScaleForTuning();
94 Cost += Log2_32_Ceil(Value: VL);
95 break;
96 }
97 case RISCV::VFREDOSUM_VS: {
98 unsigned VL = VT.getVectorMinNumElements();
99 if (!VT.isFixedLengthVector())
100 VL *= *getVScaleForTuning();
101 Cost += VL;
102 break;
103 }
104 case RISCV::VMV_X_S:
105 case RISCV::VMV_S_X:
106 case RISCV::VFMV_F_S:
107 case RISCV::VFMV_S_F:
108 case RISCV::VMOR_MM:
109 case RISCV::VMXOR_MM:
110 case RISCV::VMAND_MM:
111 case RISCV::VMANDN_MM:
112 case RISCV::VMNAND_MM:
113 case RISCV::VCPOP_M:
114 case RISCV::VFIRST_M:
115 Cost += 1;
116 break;
117 case RISCV::VDIV_VV:
118 case RISCV::VREM_VV:
119 Cost += LMULCost * TTI::TCC_Expensive;
120 break;
121 default:
122 Cost += LMULCost;
123 }
124 }
125 return Cost;
126}
127
128static InstructionCost getIntImmCostImpl(const DataLayout &DL,
129 const RISCVSubtarget *ST,
130 const APInt &Imm, Type *Ty,
131 TTI::TargetCostKind CostKind,
132 bool FreeZeroes) {
133 assert(Ty->isIntegerTy() &&
134 "getIntImmCost can only estimate cost of materialising integers");
135
136 // We have a Zero register, so 0 is always free.
137 if (Imm == 0)
138 return TTI::TCC_Free;
139
140 // Otherwise, we check how many instructions it will take to materialise.
141 return RISCVMatInt::getIntMatCost(Val: Imm, Size: DL.getTypeSizeInBits(Ty), STI: *ST,
142 /*CompressionCost=*/false, FreeZeroes);
143}
144
145InstructionCost
146RISCVTTIImpl::getIntImmCost(const APInt &Imm, Type *Ty,
147 TTI::TargetCostKind CostKind) const {
148 return getIntImmCostImpl(DL: getDataLayout(), ST: getST(), Imm, Ty, CostKind, FreeZeroes: false);
149}
150
151// Look for patterns of shift followed by AND that can be turned into a pair of
152// shifts. We won't need to materialize an immediate for the AND so these can
153// be considered free.
154static bool canUseShiftPair(Instruction *Inst, const APInt &Imm) {
155 uint64_t Mask = Imm.getZExtValue();
156 auto *BO = dyn_cast<BinaryOperator>(Val: Inst->getOperand(i: 0));
157 if (!BO || !BO->hasOneUse())
158 return false;
159
160 if (BO->getOpcode() != Instruction::Shl)
161 return false;
162
163 if (!isa<ConstantInt>(Val: BO->getOperand(i_nocapture: 1)))
164 return false;
165
166 unsigned ShAmt = cast<ConstantInt>(Val: BO->getOperand(i_nocapture: 1))->getZExtValue();
167 // (and (shl x, c2), c1) will be matched to (srli (slli x, c2+c3), c3) if c1
168 // is a mask shifted by c2 bits with c3 leading zeros.
169 if (isShiftedMask_64(Value: Mask)) {
170 unsigned Trailing = llvm::countr_zero(Val: Mask);
171 if (ShAmt == Trailing)
172 return true;
173 }
174
175 return false;
176}
177
178// If this is i64 AND is part of (X & -(1 << C1) & 0xffffffff) == C2 << C1),
179// DAGCombiner can convert this to (sraiw X, C1) == sext(C2) for RV64. On RV32,
180// the type will be split so only the lower 32 bits need to be compared using
181// (srai/srli X, C) == C2.
182static bool canUseShiftCmp(Instruction *Inst, const APInt &Imm) {
183 if (!Inst->hasOneUse())
184 return false;
185
186 // Look for equality comparison.
187 auto *Cmp = dyn_cast<ICmpInst>(Val: *Inst->user_begin());
188 if (!Cmp || !Cmp->isEquality())
189 return false;
190
191 // Right hand side of comparison should be a constant.
192 auto *C = dyn_cast<ConstantInt>(Val: Cmp->getOperand(i_nocapture: 1));
193 if (!C)
194 return false;
195
196 uint64_t Mask = Imm.getZExtValue();
197
198 // Mask should be of the form -(1 << C) in the lower 32 bits.
199 if (!isUInt<32>(x: Mask) || !isPowerOf2_32(Value: -uint32_t(Mask)))
200 return false;
201
202 // Comparison constant should be a subset of Mask.
203 uint64_t CmpC = C->getZExtValue();
204 if ((CmpC & Mask) != CmpC)
205 return false;
206
207 // We'll need to sign extend the comparison constant and shift it right. Make
208 // sure the new constant can use addi/xori+seqz/snez.
209 unsigned ShiftBits = llvm::countr_zero(Val: Mask);
210 int64_t NewCmpC = SignExtend64<32>(x: CmpC) >> ShiftBits;
211 return NewCmpC >= -2048 && NewCmpC <= 2048;
212}
213
214InstructionCost RISCVTTIImpl::getIntImmCostInst(unsigned Opcode, unsigned Idx,
215 const APInt &Imm, Type *Ty,
216 TTI::TargetCostKind CostKind,
217 Instruction *Inst) const {
218 assert(Ty->isIntegerTy() &&
219 "getIntImmCost can only estimate cost of materialising integers");
220
221 // We have a Zero register, so 0 is always free.
222 if (Imm == 0)
223 return TTI::TCC_Free;
224
225 // Some instructions in RISC-V can take a 12-bit immediate. Some of these are
226 // commutative, in others the immediate comes from a specific argument index.
227 bool Takes12BitImm = false;
228 unsigned ImmArgIdx = ~0U;
229
230 switch (Opcode) {
231 case Instruction::GetElementPtr:
232 // Never hoist any arguments to a GetElementPtr. CodeGenPrepare will
233 // split up large offsets in GEP into better parts than ConstantHoisting
234 // can.
235 return TTI::TCC_Free;
236 case Instruction::Store: {
237 // Use the materialization cost regardless of if it's the address or the
238 // value that is constant, except for if the store is misaligned and
239 // misaligned accesses are not legal (experience shows constant hoisting
240 // can sometimes be harmful in such cases).
241 if (Idx == 1 || !Inst)
242 return getIntImmCostImpl(DL: getDataLayout(), ST: getST(), Imm, Ty, CostKind,
243 /*FreeZeroes=*/true);
244
245 StoreInst *ST = cast<StoreInst>(Val: Inst);
246 if (!getTLI()->allowsMemoryAccessForAlignment(
247 Context&: Ty->getContext(), DL, VT: getTLI()->getValueType(DL, Ty),
248 AddrSpace: ST->getPointerAddressSpace(), Alignment: ST->getAlign()))
249 return TTI::TCC_Free;
250
251 return getIntImmCostImpl(DL: getDataLayout(), ST: getST(), Imm, Ty, CostKind,
252 /*FreeZeroes=*/true);
253 }
254 case Instruction::Load:
255 // If the address is a constant, use the materialization cost.
256 return getIntImmCost(Imm, Ty, CostKind);
257 case Instruction::And:
258 // zext.h
259 if (Imm == UINT64_C(0xffff) && ST->hasStdExtZbb())
260 return TTI::TCC_Free;
261 // zext.w
262 if (Imm == UINT64_C(0xffffffff) &&
263 ((ST->hasStdExtZba() && ST->isRV64()) || ST->isRV32()))
264 return TTI::TCC_Free;
265 // bclri
266 if (ST->hasStdExtZbs() && (~Imm).isPowerOf2())
267 return TTI::TCC_Free;
268 if (Inst && Idx == 1 && Imm.getBitWidth() <= ST->getXLen() &&
269 canUseShiftPair(Inst, Imm))
270 return TTI::TCC_Free;
271 if (Inst && Idx == 1 && Imm.getBitWidth() == 64 &&
272 canUseShiftCmp(Inst, Imm))
273 return TTI::TCC_Free;
274 Takes12BitImm = true;
275 break;
276 case Instruction::Add:
277 Takes12BitImm = true;
278 break;
279 case Instruction::Or:
280 case Instruction::Xor:
281 // bseti/binvi
282 if (ST->hasStdExtZbs() && Imm.isPowerOf2())
283 return TTI::TCC_Free;
284 Takes12BitImm = true;
285 break;
286 case Instruction::Mul:
287 // Power of 2 is a shift. Negated power of 2 is a shift and a negate.
288 if (Imm.isPowerOf2() || Imm.isNegatedPowerOf2())
289 return TTI::TCC_Free;
290 // One more or less than a power of 2 can use SLLI+ADD/SUB.
291 if ((Imm + 1).isPowerOf2() || (Imm - 1).isPowerOf2())
292 return TTI::TCC_Free;
293 // FIXME: There is no MULI instruction.
294 Takes12BitImm = true;
295 break;
296 case Instruction::Sub:
297 case Instruction::Shl:
298 case Instruction::LShr:
299 case Instruction::AShr:
300 Takes12BitImm = true;
301 ImmArgIdx = 1;
302 break;
303 default:
304 break;
305 }
306
307 if (Takes12BitImm) {
308 // Check immediate is the correct argument...
309 if (Instruction::isCommutative(Opcode) || Idx == ImmArgIdx) {
310 // ... and fits into the 12-bit immediate.
311 if (Imm.getSignificantBits() <= 64 &&
312 getTLI()->isLegalAddImmediate(Imm: Imm.getSExtValue())) {
313 return TTI::TCC_Free;
314 }
315 }
316
317 // Otherwise, use the full materialisation cost.
318 return getIntImmCost(Imm, Ty, CostKind);
319 }
320
321 // By default, prevent hoisting.
322 return TTI::TCC_Free;
323}
324
325InstructionCost
326RISCVTTIImpl::getIntImmCostIntrin(Intrinsic::ID IID, unsigned Idx,
327 const APInt &Imm, Type *Ty,
328 TTI::TargetCostKind CostKind) const {
329 // Prevent hoisting in unknown cases.
330 return TTI::TCC_Free;
331}
332
333bool RISCVTTIImpl::hasActiveVectorLength() const {
334 return ST->hasVInstructions();
335}
336
337TargetTransformInfo::PopcntSupportKind
338RISCVTTIImpl::getPopcntSupport(unsigned TyWidth) const {
339 assert(isPowerOf2_32(TyWidth) && "Ty width must be power of 2");
340 return ST->hasCPOPLike() ? TTI::PSK_FastHardware : TTI::PSK_Software;
341}
342
343InstructionCost RISCVTTIImpl::getPartialReductionCost(
344 unsigned Opcode, Type *InputTypeA, Type *InputTypeB, Type *AccumType,
345 ElementCount VF, TTI::PartialReductionExtendKind OpAExtend,
346 TTI::PartialReductionExtendKind OpBExtend, std::optional<unsigned> BinOp,
347 TTI::TargetCostKind CostKind, std::optional<FastMathFlags> FMF) const {
348 if (Opcode == Instruction::FAdd)
349 return InstructionCost::getInvalid();
350
351 // zve32x is broken for partial_reduce_umla, but let's make sure we
352 // don't generate them.
353 if (!ST->hasStdExtZvdot4a8i() || ST->getELen() < 64 ||
354 Opcode != Instruction::Add || !BinOp || *BinOp != Instruction::Mul ||
355 InputTypeA != InputTypeB || !InputTypeA->isIntegerTy(Bitwidth: 8) ||
356 !AccumType->isIntegerTy(Bitwidth: 32) || !VF.isKnownMultipleOf(RHS: 4))
357 return InstructionCost::getInvalid();
358
359 Type *Tp = VectorType::get(ElementType: AccumType, EC: VF.divideCoefficientBy(RHS: 4));
360 std::pair<InstructionCost, MVT> LT = getTypeLegalizationCost(Ty: Tp);
361 // Note: Asuming all vdota4* variants are equal cost
362 return LT.first *
363 getRISCVInstructionCost(OpCodes: RISCV::VDOTA4_VV, VT: LT.second, CostKind);
364}
365
366bool RISCVTTIImpl::shouldExpandReduction(const IntrinsicInst *II) const {
367 // Currently, the ExpandReductions pass can't expand scalable-vector
368 // reductions, but we still request expansion as RVV doesn't support certain
369 // reductions and the SelectionDAG can't legalize them either.
370 switch (II->getIntrinsicID()) {
371 default:
372 return false;
373 // These reductions have no equivalent in RVV
374 case Intrinsic::vector_reduce_mul:
375 case Intrinsic::vector_reduce_fmul:
376 return true;
377 }
378}
379
380std::optional<unsigned> RISCVTTIImpl::getMaxVScale() const {
381 if (ST->hasVInstructions())
382 return ST->getRealMaxVLen() / RISCV::RVVBitsPerBlock;
383 return BaseT::getMaxVScale();
384}
385
386std::optional<unsigned> RISCVTTIImpl::getVScaleForTuning() const {
387 if (ST->hasVInstructions())
388 if (unsigned MinVLen = ST->getRealMinVLen();
389 MinVLen >= RISCV::RVVBitsPerBlock)
390 return MinVLen / RISCV::RVVBitsPerBlock;
391 return BaseT::getVScaleForTuning();
392}
393
394TypeSize
395RISCVTTIImpl::getRegisterBitWidth(TargetTransformInfo::RegisterKind K) const {
396 unsigned LMUL =
397 llvm::bit_floor(Value: std::clamp<unsigned>(val: RVVRegisterWidthLMUL, lo: 1, hi: 8));
398 switch (K) {
399 case TargetTransformInfo::RGK_Scalar:
400 return TypeSize::getFixed(ExactSize: ST->getXLen());
401 case TargetTransformInfo::RGK_FixedWidthVector:
402 return TypeSize::getFixed(
403 ExactSize: ST->useRVVForFixedLengthVectors() ? LMUL * ST->getRealMinVLen() : 0);
404 case TargetTransformInfo::RGK_ScalableVector:
405 return TypeSize::getScalable(
406 MinimumSize: (ST->hasVInstructions() &&
407 ST->getRealMinVLen() >= RISCV::RVVBitsPerBlock)
408 ? LMUL * RISCV::RVVBitsPerBlock
409 : 0);
410 }
411
412 llvm_unreachable("Unsupported register kind");
413}
414
415InstructionCost RISCVTTIImpl::getStaticDataAddrGenerationCost(
416 const TTI::TargetCostKind CostKind) const {
417 switch (CostKind) {
418 case TTI::TCK_CodeSize:
419 case TTI::TCK_SizeAndLatency:
420 // Always 2 instructions
421 return 2;
422 case TTI::TCK_Latency:
423 case TTI::TCK_RecipThroughput:
424 // Depending on the memory model the address generation will
425 // require AUIPC + ADDI (medany) or LUI + ADDI (medlow). Don't
426 // have a way of getting this information here, so conservatively
427 // require both.
428 // In practice, these are generally implemented together.
429 return (ST->hasAUIPCADDIFusion() && ST->hasLUIADDIFusion()) ? 1 : 2;
430 }
431 llvm_unreachable("Unsupported cost kind");
432}
433
434InstructionCost
435RISCVTTIImpl::getConstantPoolLoadCost(Type *Ty,
436 TTI::TargetCostKind CostKind) const {
437 // Add a cost of address generation + the cost of the load. The address
438 // is expected to be a PC relative offset to a constant pool entry
439 // using auipc/addi.
440 return getStaticDataAddrGenerationCost(CostKind) +
441 getMemoryOpCost(Opcode: Instruction::Load, Src: Ty, Alignment: DL.getABITypeAlign(Ty),
442 /*AddressSpace=*/0, CostKind);
443}
444
445static bool isRepeatedConcatMask(ArrayRef<int> Mask, int &SubVectorSize) {
446 unsigned Size = Mask.size();
447 if (!isPowerOf2_32(Value: Size))
448 return false;
449 for (unsigned I = 0; I != Size; ++I) {
450 if (static_cast<unsigned>(Mask[I]) == I)
451 continue;
452 if (Mask[I] != 0)
453 return false;
454 if (Size % I != 0)
455 return false;
456 for (unsigned J = I + 1; J != Size; ++J)
457 // Check the pattern is repeated.
458 if (static_cast<unsigned>(Mask[J]) != J % I)
459 return false;
460 SubVectorSize = I;
461 return true;
462 }
463 // That means Mask is <0, 1, 2, 3>. This is not a concatenation.
464 return false;
465}
466
467static VectorType *getVRGatherIndexType(MVT DataVT, const RISCVSubtarget &ST,
468 LLVMContext &C) {
469 assert((DataVT.getScalarSizeInBits() != 8 ||
470 DataVT.getVectorNumElements() <= 256) && "unhandled case in lowering");
471 MVT IndexVT = DataVT.changeTypeToInteger();
472 if (IndexVT.getScalarType().bitsGT(VT: ST.getXLenVT()))
473 IndexVT = IndexVT.changeVectorElementType(EltVT: MVT::i16);
474 return cast<VectorType>(Val: EVT(IndexVT).getTypeForEVT(Context&: C));
475}
476
477/// Attempt to approximate the cost of a shuffle which will require splitting
478/// during legalization. Note that processShuffleMasks is not an exact proxy
479/// for the algorithm used in LegalizeVectorTypes, but hopefully it's a
480/// reasonably close upperbound.
481static InstructionCost costShuffleViaSplitting(const RISCVTTIImpl &TTI,
482 MVT LegalVT, VectorType *Tp,
483 ArrayRef<int> Mask,
484 TTI::TargetCostKind CostKind) {
485 assert(LegalVT.isFixedLengthVector() && !Mask.empty() &&
486 "Expected fixed vector type and non-empty mask");
487 unsigned LegalNumElts = LegalVT.getVectorNumElements();
488 // Number of destination vectors after legalization:
489 unsigned NumOfDests = divideCeil(Numerator: Mask.size(), Denominator: LegalNumElts);
490 // We are going to permute multiple sources and the result will be in
491 // multiple destinations. Providing an accurate cost only for splits where
492 // the element type remains the same.
493 if (NumOfDests <= 1 ||
494 LegalVT.getVectorElementType().getSizeInBits() !=
495 Tp->getElementType()->getPrimitiveSizeInBits() ||
496 LegalNumElts >= Tp->getElementCount().getFixedValue())
497 return InstructionCost::getInvalid();
498
499 unsigned VecTySize = TTI.getDataLayout().getTypeStoreSize(Ty: Tp);
500 unsigned LegalVTSize = LegalVT.getStoreSize();
501 // Number of source vectors after legalization:
502 unsigned NumOfSrcs = divideCeil(Numerator: VecTySize, Denominator: LegalVTSize);
503
504 auto *SingleOpTy = FixedVectorType::get(ElementType: Tp->getElementType(), NumElts: LegalNumElts);
505
506 unsigned NormalizedVF = LegalNumElts * std::max(a: NumOfSrcs, b: NumOfDests);
507 unsigned NumOfSrcRegs = NormalizedVF / LegalNumElts;
508 unsigned NumOfDestRegs = NormalizedVF / LegalNumElts;
509 SmallVector<int> NormalizedMask(NormalizedVF, PoisonMaskElem);
510 assert(NormalizedVF >= Mask.size() &&
511 "Normalized mask expected to be not shorter than original mask.");
512 copy(Range&: Mask, Out: NormalizedMask.begin());
513 InstructionCost Cost = 0;
514 SmallDenseSet<std::pair<ArrayRef<int>, unsigned>> ReusedSingleSrcShuffles;
515 processShuffleMasks(
516 Mask: NormalizedMask, NumOfSrcRegs, NumOfDestRegs, NumOfUsedRegs: NumOfDestRegs, NoInputAction: []() {},
517 SingleInputAction: [&](ArrayRef<int> RegMask, unsigned SrcReg, unsigned DestReg) {
518 if (ShuffleVectorInst::isIdentityMask(Mask: RegMask, NumSrcElts: RegMask.size()))
519 return;
520 if (!ReusedSingleSrcShuffles.insert(V: std::make_pair(x&: RegMask, y&: SrcReg))
521 .second)
522 return;
523 Cost += TTI.getShuffleCost(
524 Kind: TTI::SK_PermuteSingleSrc,
525 DstTy: FixedVectorType::get(ElementType: SingleOpTy->getElementType(), NumElts: RegMask.size()),
526 SrcTy: SingleOpTy, Mask: RegMask, CostKind, Index: 0, SubTp: nullptr);
527 },
528 ManyInputsAction: [&](ArrayRef<int> RegMask, unsigned Idx1, unsigned Idx2, bool NewReg) {
529 Cost += TTI.getShuffleCost(
530 Kind: TTI::SK_PermuteTwoSrc,
531 DstTy: FixedVectorType::get(ElementType: SingleOpTy->getElementType(), NumElts: RegMask.size()),
532 SrcTy: SingleOpTy, Mask: RegMask, CostKind, Index: 0, SubTp: nullptr);
533 });
534 return Cost;
535}
536
537/// Try to perform better estimation of the permutation.
538/// 1. Split the source/destination vectors into real registers.
539/// 2. Do the mask analysis to identify which real registers are
540/// permuted. If more than 1 source registers are used for the
541/// destination register building, the cost for this destination register
542/// is (Number_of_source_register - 1) * Cost_PermuteTwoSrc. If only one
543/// source register is used, build mask and calculate the cost as a cost
544/// of PermuteSingleSrc.
545/// Also, for the single register permute we try to identify if the
546/// destination register is just a copy of the source register or the
547/// copy of the previous destination register (the cost is
548/// TTI::TCC_Basic). If the source register is just reused, the cost for
549/// this operation is 0.
550static InstructionCost
551costShuffleViaVRegSplitting(const RISCVTTIImpl &TTI, MVT LegalVT,
552 std::optional<unsigned> VLen, VectorType *Tp,
553 ArrayRef<int> Mask, TTI::TargetCostKind CostKind) {
554 assert(LegalVT.isFixedLengthVector());
555 if (!VLen || Mask.empty())
556 return InstructionCost::getInvalid();
557 MVT ElemVT = LegalVT.getVectorElementType();
558 unsigned ElemsPerVReg = *VLen / ElemVT.getFixedSizeInBits();
559 LegalVT = TTI.getTypeLegalizationCost(
560 Ty: FixedVectorType::get(ElementType: Tp->getElementType(), NumElts: ElemsPerVReg))
561 .second;
562 // Number of destination vectors after legalization:
563 InstructionCost NumOfDests =
564 divideCeil(Numerator: Mask.size(), Denominator: LegalVT.getVectorNumElements());
565 if (NumOfDests <= 1 ||
566 LegalVT.getVectorElementType().getSizeInBits() !=
567 Tp->getElementType()->getPrimitiveSizeInBits() ||
568 LegalVT.getVectorNumElements() >= Tp->getElementCount().getFixedValue())
569 return InstructionCost::getInvalid();
570
571 unsigned VecTySize = TTI.getDataLayout().getTypeStoreSize(Ty: Tp);
572 unsigned LegalVTSize = LegalVT.getStoreSize();
573 // Number of source vectors after legalization:
574 unsigned NumOfSrcs = divideCeil(Numerator: VecTySize, Denominator: LegalVTSize);
575
576 auto *SingleOpTy = FixedVectorType::get(ElementType: Tp->getElementType(),
577 NumElts: LegalVT.getVectorNumElements());
578
579 unsigned E = NumOfDests.getValue();
580 unsigned NormalizedVF =
581 LegalVT.getVectorNumElements() * std::max(a: NumOfSrcs, b: E);
582 unsigned NumOfSrcRegs = NormalizedVF / LegalVT.getVectorNumElements();
583 unsigned NumOfDestRegs = NormalizedVF / LegalVT.getVectorNumElements();
584 SmallVector<int> NormalizedMask(NormalizedVF, PoisonMaskElem);
585 assert(NormalizedVF >= Mask.size() &&
586 "Normalized mask expected to be not shorter than original mask.");
587 copy(Range&: Mask, Out: NormalizedMask.begin());
588 InstructionCost Cost = 0;
589 int NumShuffles = 0;
590 SmallDenseSet<std::pair<ArrayRef<int>, unsigned>> ReusedSingleSrcShuffles;
591 processShuffleMasks(
592 Mask: NormalizedMask, NumOfSrcRegs, NumOfDestRegs, NumOfUsedRegs: NumOfDestRegs, NoInputAction: []() {},
593 SingleInputAction: [&](ArrayRef<int> RegMask, unsigned SrcReg, unsigned DestReg) {
594 if (ShuffleVectorInst::isIdentityMask(Mask: RegMask, NumSrcElts: RegMask.size()))
595 return;
596 if (!ReusedSingleSrcShuffles.insert(V: std::make_pair(x&: RegMask, y&: SrcReg))
597 .second)
598 return;
599 ++NumShuffles;
600 Cost += TTI.getShuffleCost(Kind: TTI::SK_PermuteSingleSrc, DstTy: SingleOpTy,
601 SrcTy: SingleOpTy, Mask: RegMask, CostKind, Index: 0, SubTp: nullptr);
602 },
603 ManyInputsAction: [&](ArrayRef<int> RegMask, unsigned Idx1, unsigned Idx2, bool NewReg) {
604 Cost += TTI.getShuffleCost(Kind: TTI::SK_PermuteTwoSrc, DstTy: SingleOpTy,
605 SrcTy: SingleOpTy, Mask: RegMask, CostKind, Index: 0, SubTp: nullptr);
606 NumShuffles += 2;
607 });
608 // Note: check that we do not emit too many shuffles here to prevent code
609 // size explosion.
610 // TODO: investigate, if it can be improved by extra analysis of the masks
611 // to check if the code is more profitable.
612 if ((NumOfDestRegs > 2 && NumShuffles <= static_cast<int>(NumOfDestRegs)) ||
613 (NumOfDestRegs <= 2 && NumShuffles < 4))
614 return Cost;
615 return InstructionCost::getInvalid();
616}
617
618InstructionCost RISCVTTIImpl::getSlideCost(FixedVectorType *Tp,
619 ArrayRef<int> Mask,
620 TTI::TargetCostKind CostKind) const {
621 // Avoid missing masks and length changing shuffles
622 if (Mask.size() <= 2 || Mask.size() != Tp->getNumElements())
623 return InstructionCost::getInvalid();
624
625 int NumElts = Tp->getNumElements();
626 std::pair<InstructionCost, MVT> LT = getTypeLegalizationCost(Ty: Tp);
627 // Avoid scalarization cases
628 if (!LT.second.isFixedLengthVector())
629 return InstructionCost::getInvalid();
630
631 // Requires moving elements between parts, which requires additional
632 // unmodeled instructions.
633 if (LT.first != 1)
634 return InstructionCost::getInvalid();
635
636 auto GetSlideOpcode = [&](int SlideAmt) {
637 assert(SlideAmt != 0);
638 bool IsVI = isUInt<5>(x: std::abs(x: SlideAmt));
639 if (SlideAmt < 0)
640 return IsVI ? RISCV::VSLIDEDOWN_VI : RISCV::VSLIDEDOWN_VX;
641 return IsVI ? RISCV::VSLIDEUP_VI : RISCV::VSLIDEUP_VX;
642 };
643
644 std::array<std::pair<int, int>, 2> SrcInfo;
645 if (!isMaskedSlidePair(Mask, NumElts, SrcInfo))
646 return InstructionCost::getInvalid();
647
648 if (SrcInfo[1].second == 0)
649 std::swap(x&: SrcInfo[0], y&: SrcInfo[1]);
650
651 InstructionCost FirstSlideCost = 0;
652 if (SrcInfo[0].second != 0) {
653 unsigned Opcode = GetSlideOpcode(SrcInfo[0].second);
654 FirstSlideCost = getRISCVInstructionCost(OpCodes: Opcode, VT: LT.second, CostKind);
655 }
656
657 if (SrcInfo[1].first == -1)
658 return FirstSlideCost;
659
660 InstructionCost SecondSlideCost = 0;
661 if (SrcInfo[1].second != 0) {
662 unsigned Opcode = GetSlideOpcode(SrcInfo[1].second);
663 SecondSlideCost = getRISCVInstructionCost(OpCodes: Opcode, VT: LT.second, CostKind);
664 } else {
665 SecondSlideCost =
666 getRISCVInstructionCost(OpCodes: RISCV::VMERGE_VVM, VT: LT.second, CostKind);
667 }
668
669 auto EC = Tp->getElementCount();
670 VectorType *MaskTy =
671 VectorType::get(ElementType: IntegerType::getInt1Ty(C&: Tp->getContext()), EC);
672 InstructionCost MaskCost = getConstantPoolLoadCost(Ty: MaskTy, CostKind);
673 return FirstSlideCost + SecondSlideCost + MaskCost;
674}
675
676InstructionCost
677RISCVTTIImpl::getShuffleCost(TTI::ShuffleKind Kind, VectorType *DstTy,
678 VectorType *SrcTy, ArrayRef<int> Mask,
679 TTI::TargetCostKind CostKind, int Index,
680 VectorType *SubTp, ArrayRef<const Value *> Args,
681 const Instruction *CxtI) const {
682 assert((Mask.empty() || DstTy->isScalableTy() ||
683 Mask.size() == DstTy->getElementCount().getKnownMinValue()) &&
684 "Expected the Mask to match the return size if given");
685 assert(SrcTy->getScalarType() == DstTy->getScalarType() &&
686 "Expected the same scalar types");
687
688 Kind = improveShuffleKindFromMask(Kind, Mask, SrcTy, Index, SubTy&: SubTp);
689
690 // TODO: Add proper cost model for P extension fixed vectors (e.g., v4i16)
691 // For now, skip all fixed vector cost analysis when P extension is available
692 // to avoid crashes in getMinRVVVectorSizeInBits()
693 if (ST->hasStdExtP() && isa<FixedVectorType>(Val: SrcTy))
694 return 1;
695
696 std::pair<InstructionCost, MVT> LT = getTypeLegalizationCost(Ty: SrcTy);
697
698 // First, handle cases where having a fixed length vector enables us to
699 // give a more accurate cost than falling back to generic scalable codegen.
700 // TODO: Each of these cases hints at a modeling gap around scalable vectors.
701 if (auto *FVTp = dyn_cast<FixedVectorType>(Val: SrcTy);
702 FVTp && ST->hasVInstructions() && LT.second.isFixedLengthVector()) {
703 InstructionCost VRegSplittingCost = costShuffleViaVRegSplitting(
704 TTI: *this, LegalVT: LT.second, VLen: ST->getRealVLen(),
705 Tp: Kind == TTI::SK_InsertSubvector ? DstTy : SrcTy, Mask, CostKind);
706 if (VRegSplittingCost.isValid())
707 return VRegSplittingCost;
708 switch (Kind) {
709 default:
710 break;
711 case TTI::SK_PermuteSingleSrc: {
712 if (Mask.size() >= 2) {
713 MVT EltTp = LT.second.getVectorElementType();
714 // If the size of the element is < ELEN then shuffles of interleaves and
715 // deinterleaves of 2 vectors can be lowered into the following
716 // sequences
717 if (EltTp.getScalarSizeInBits() < ST->getELen()) {
718 // Example sequence:
719 // vsetivli zero, 4, e8, mf4, ta, ma (ignored)
720 // vwaddu.vv v10, v8, v9
721 // li a0, -1 (ignored)
722 // vwmaccu.vx v10, a0, v9
723 if (ShuffleVectorInst::isInterleaveMask(Mask, Factor: 2, NumInputElts: Mask.size()))
724 return 2 * LT.first * TLI->getLMULCost(VT: LT.second);
725
726 if (Mask[0] == 0 || Mask[0] == 1) {
727 auto DeinterleaveMask = createStrideMask(Start: Mask[0], Stride: 2, VF: Mask.size());
728 // Example sequence:
729 // vnsrl.wi v10, v8, 0
730 if (equal(LRange&: DeinterleaveMask, RRange&: Mask))
731 return LT.first * getRISCVInstructionCost(OpCodes: RISCV::VNSRL_WI,
732 VT: LT.second, CostKind);
733 }
734 }
735 int SubVectorSize;
736 if (LT.second.getScalarSizeInBits() != 1 &&
737 isRepeatedConcatMask(Mask, SubVectorSize)) {
738 InstructionCost Cost = 0;
739 unsigned NumSlides = Log2_32(Value: Mask.size() / SubVectorSize);
740 // The cost of extraction from a subvector is 0 if the index is 0.
741 for (unsigned I = 0; I != NumSlides; ++I) {
742 unsigned InsertIndex = SubVectorSize * (1 << I);
743 FixedVectorType *SubTp =
744 FixedVectorType::get(ElementType: SrcTy->getElementType(), NumElts: InsertIndex);
745 FixedVectorType *DestTp =
746 FixedVectorType::getDoubleElementsVectorType(VTy: SubTp);
747 std::pair<InstructionCost, MVT> DestLT =
748 getTypeLegalizationCost(Ty: DestTp);
749 // Add the cost of whole vector register move because the
750 // destination vector register group for vslideup cannot overlap the
751 // source.
752 Cost += DestLT.first * TLI->getLMULCost(VT: DestLT.second);
753 Cost += getShuffleCost(Kind: TTI::SK_InsertSubvector, DstTy: DestTp, SrcTy: DestTp, Mask: {},
754 CostKind, Index: InsertIndex, SubTp);
755 }
756 return Cost;
757 }
758 }
759
760 if (InstructionCost SlideCost = getSlideCost(Tp: FVTp, Mask, CostKind);
761 SlideCost.isValid())
762 return SlideCost;
763
764 // vrgather + cost of generating the mask constant.
765 // We model this for an unknown mask with a single vrgather.
766 if (LT.first == 1 && (LT.second.getScalarSizeInBits() != 8 ||
767 LT.second.getVectorNumElements() <= 256)) {
768 VectorType *IdxTy =
769 getVRGatherIndexType(DataVT: LT.second, ST: *ST, C&: SrcTy->getContext());
770 InstructionCost IndexCost = getConstantPoolLoadCost(Ty: IdxTy, CostKind);
771 return IndexCost +
772 getRISCVInstructionCost(OpCodes: RISCV::VRGATHER_VV, VT: LT.second, CostKind);
773 }
774 break;
775 }
776 case TTI::SK_Transpose:
777 case TTI::SK_PermuteTwoSrc: {
778
779 if (InstructionCost SlideCost = getSlideCost(Tp: FVTp, Mask, CostKind);
780 SlideCost.isValid())
781 return SlideCost;
782
783 // 2 x (vrgather + cost of generating the mask constant) + cost of mask
784 // register for the second vrgather. We model this for an unknown
785 // (shuffle) mask.
786 if (LT.first == 1 && (LT.second.getScalarSizeInBits() != 8 ||
787 LT.second.getVectorNumElements() <= 256)) {
788 auto &C = SrcTy->getContext();
789 auto EC = SrcTy->getElementCount();
790 VectorType *IdxTy = getVRGatherIndexType(DataVT: LT.second, ST: *ST, C);
791 VectorType *MaskTy = VectorType::get(ElementType: IntegerType::getInt1Ty(C), EC);
792 InstructionCost IndexCost = getConstantPoolLoadCost(Ty: IdxTy, CostKind);
793 InstructionCost MaskCost = getConstantPoolLoadCost(Ty: MaskTy, CostKind);
794 return 2 * IndexCost +
795 getRISCVInstructionCost(OpCodes: {RISCV::VRGATHER_VV, RISCV::VRGATHER_VV},
796 VT: LT.second, CostKind) +
797 MaskCost;
798 }
799 break;
800 }
801 }
802
803 auto shouldSplit = [](TTI::ShuffleKind Kind) {
804 switch (Kind) {
805 default:
806 return false;
807 case TTI::SK_PermuteSingleSrc:
808 case TTI::SK_Transpose:
809 case TTI::SK_PermuteTwoSrc:
810 return true;
811 }
812 };
813
814 if (!Mask.empty() && LT.first.isValid() && LT.first != 1 &&
815 shouldSplit(Kind)) {
816 InstructionCost SplitCost =
817 costShuffleViaSplitting(TTI: *this, LegalVT: LT.second, Tp: FVTp, Mask, CostKind);
818 if (SplitCost.isValid())
819 return SplitCost;
820 }
821 }
822
823 // Handle scalable vectors (and fixed vectors legalized to scalable vectors).
824 switch (Kind) {
825 default:
826 // Fallthrough to generic handling.
827 // TODO: Most of these cases will return getInvalid in generic code, and
828 // must be implemented here.
829 break;
830 case TTI::SK_ExtractSubvector:
831 // Extract at zero is always a subregister extract
832 if (Index == 0)
833 return TTI::TCC_Free;
834
835 // If we're extracting a subvector of at most m1 size at a sub-register
836 // boundary - which unfortunately we need exact vlen to identify - this is
837 // a subregister extract at worst and thus won't require a vslidedown.
838 // TODO: Extend for aligned m2, m4 subvector extracts
839 // TODO: Extend for misalgined (but contained) extracts
840 // TODO: Extend for scalable subvector types
841 if (std::pair<InstructionCost, MVT> SubLT = getTypeLegalizationCost(Ty: SubTp);
842 SubLT.second.isValid() && SubLT.second.isFixedLengthVector()) {
843 if (std::optional<unsigned> VLen = ST->getRealVLen();
844 VLen && SubLT.second.getScalarSizeInBits() * Index % *VLen == 0 &&
845 SubLT.second.getSizeInBits() <= *VLen)
846 return TTI::TCC_Free;
847 }
848
849 // Example sequence:
850 // vsetivli zero, 4, e8, mf2, tu, ma (ignored)
851 // vslidedown.vi v8, v9, 2
852 return LT.first *
853 getRISCVInstructionCost(OpCodes: RISCV::VSLIDEDOWN_VI, VT: LT.second, CostKind);
854 case TTI::SK_InsertSubvector:
855 // Example sequence:
856 // vsetivli zero, 4, e8, mf2, tu, ma (ignored)
857 // vslideup.vi v8, v9, 2
858 LT = getTypeLegalizationCost(Ty: DstTy);
859 return LT.first *
860 getRISCVInstructionCost(OpCodes: RISCV::VSLIDEUP_VI, VT: LT.second, CostKind);
861 case TTI::SK_Select: {
862 // Example sequence:
863 // li a0, 90
864 // vsetivli zero, 8, e8, mf2, ta, ma (ignored)
865 // vmv.s.x v0, a0
866 // vmerge.vvm v8, v9, v8, v0
867 // We use 2 for the cost of the mask materialization as this is the true
868 // cost for small masks and most shuffles are small. At worst, this cost
869 // should be a very small constant for the constant pool load. As such,
870 // we may bias towards large selects slightly more than truly warranted.
871 return LT.first *
872 (1 + getRISCVInstructionCost(OpCodes: {RISCV::VMV_S_X, RISCV::VMERGE_VVM},
873 VT: LT.second, CostKind));
874 }
875 case TTI::SK_Broadcast: {
876 bool HasScalar = (Args.size() > 0) && (Operator::getOpcode(V: Args[0]) ==
877 Instruction::InsertElement);
878 if (LT.second.getScalarSizeInBits() == 1) {
879 if (HasScalar) {
880 // Example sequence:
881 // andi a0, a0, 1
882 // vsetivli zero, 2, e8, mf8, ta, ma (ignored)
883 // vmv.v.x v8, a0
884 // vmsne.vi v0, v8, 0
885 return LT.first *
886 (1 + getRISCVInstructionCost(OpCodes: {RISCV::VMV_V_X, RISCV::VMSNE_VI},
887 VT: LT.second, CostKind));
888 }
889 // Example sequence:
890 // vsetivli zero, 2, e8, mf8, ta, mu (ignored)
891 // vmv.v.i v8, 0
892 // vmerge.vim v8, v8, 1, v0
893 // vmv.x.s a0, v8
894 // andi a0, a0, 1
895 // vmv.v.x v8, a0
896 // vmsne.vi v0, v8, 0
897
898 return LT.first *
899 (1 + getRISCVInstructionCost(OpCodes: {RISCV::VMV_V_I, RISCV::VMERGE_VIM,
900 RISCV::VMV_X_S, RISCV::VMV_V_X,
901 RISCV::VMSNE_VI},
902 VT: LT.second, CostKind));
903 }
904
905 if (HasScalar) {
906 // Example sequence:
907 // vmv.v.x v8, a0
908 return LT.first *
909 getRISCVInstructionCost(OpCodes: RISCV::VMV_V_X, VT: LT.second, CostKind);
910 }
911
912 // Example sequence:
913 // vrgather.vi v9, v8, 0
914 return LT.first *
915 getRISCVInstructionCost(OpCodes: RISCV::VRGATHER_VI, VT: LT.second, CostKind);
916 }
917 case TTI::SK_Splice: {
918 // vslidedown+vslideup.
919 // TODO: Multiplying by LT.first implies this legalizes into multiple copies
920 // of similar code, but I think we expand through memory.
921 unsigned Opcodes[2] = {RISCV::VSLIDEDOWN_VX, RISCV::VSLIDEUP_VX};
922 if (Index >= 0 && Index < 32)
923 Opcodes[0] = RISCV::VSLIDEDOWN_VI;
924 else if (Index < 0 && Index > -32)
925 Opcodes[1] = RISCV::VSLIDEUP_VI;
926 return LT.first * getRISCVInstructionCost(OpCodes: Opcodes, VT: LT.second, CostKind);
927 }
928 case TTI::SK_Reverse: {
929
930 if (!LT.second.isVector())
931 return InstructionCost::getInvalid();
932
933 // TODO: Cases to improve here:
934 // * Illegal vector types
935 // * i64 on RV32
936 if (SrcTy->getElementType()->isIntegerTy(Bitwidth: 1)) {
937 VectorType *WideTy =
938 VectorType::get(ElementType: IntegerType::get(C&: SrcTy->getContext(), NumBits: 8),
939 EC: cast<VectorType>(Val: SrcTy)->getElementCount());
940 return getCastInstrCost(Opcode: Instruction::ZExt, Dst: WideTy, Src: SrcTy,
941 CCH: TTI::CastContextHint::None, CostKind) +
942 getShuffleCost(Kind: TTI::SK_Reverse, DstTy: WideTy, SrcTy: WideTy, Mask: {}, CostKind, Index: 0,
943 SubTp: nullptr) +
944 getCastInstrCost(Opcode: Instruction::Trunc, Dst: SrcTy, Src: WideTy,
945 CCH: TTI::CastContextHint::None, CostKind);
946 }
947
948 MVT ContainerVT = LT.second;
949 if (LT.second.isFixedLengthVector())
950 ContainerVT = TLI->getContainerForFixedLengthVector(VT: LT.second);
951 MVT M1VT = RISCVTargetLowering::getM1VT(VT: ContainerVT);
952 if (ContainerVT.bitsLE(VT: M1VT)) {
953 // Example sequence:
954 // csrr a0, vlenb
955 // srli a0, a0, 3
956 // addi a0, a0, -1
957 // vsetvli a1, zero, e8, mf8, ta, mu (ignored)
958 // vid.v v9
959 // vrsub.vx v10, v9, a0
960 // vrgather.vv v9, v8, v10
961 InstructionCost LenCost = 3;
962 if (LT.second.isFixedLengthVector())
963 // vrsub.vi has a 5 bit immediate field, otherwise an li suffices
964 LenCost = isInt<5>(x: LT.second.getVectorNumElements() - 1) ? 0 : 1;
965 unsigned Opcodes[] = {RISCV::VID_V, RISCV::VRSUB_VX, RISCV::VRGATHER_VV};
966 if (LT.second.isFixedLengthVector() &&
967 isInt<5>(x: LT.second.getVectorNumElements() - 1))
968 Opcodes[1] = RISCV::VRSUB_VI;
969 InstructionCost GatherCost =
970 getRISCVInstructionCost(OpCodes: Opcodes, VT: LT.second, CostKind);
971 return LT.first * (LenCost + GatherCost);
972 }
973
974 // At high LMUL, we split into a series of M1 reverses (see
975 // lowerVECTOR_REVERSE) and then do a single slide at the end to eliminate
976 // the resulting gap at the bottom (for fixed vectors only). The important
977 // bit is that the cost scales linearly, not quadratically with LMUL.
978 unsigned M1Opcodes[] = {RISCV::VID_V, RISCV::VRSUB_VX};
979 InstructionCost FixedCost =
980 getRISCVInstructionCost(OpCodes: M1Opcodes, VT: M1VT, CostKind) + 3;
981 unsigned Ratio =
982 ContainerVT.getVectorMinNumElements() / M1VT.getVectorMinNumElements();
983 InstructionCost GatherCost =
984 getRISCVInstructionCost(OpCodes: {RISCV::VRGATHER_VV}, VT: M1VT, CostKind) * Ratio;
985 InstructionCost SlideCost = !LT.second.isFixedLengthVector() ? 0 :
986 getRISCVInstructionCost(OpCodes: {RISCV::VSLIDEDOWN_VX}, VT: LT.second, CostKind);
987 return FixedCost + LT.first * (GatherCost + SlideCost);
988 }
989 }
990 return BaseT::getShuffleCost(Kind, DstTy, SrcTy, Mask, CostKind, Index,
991 SubTp);
992}
993
994static unsigned isM1OrSmaller(MVT VT) {
995 RISCVVType::VLMUL LMUL = RISCVTargetLowering::getLMUL(VT);
996 return (LMUL == RISCVVType::VLMUL::LMUL_F8 ||
997 LMUL == RISCVVType::VLMUL::LMUL_F4 ||
998 LMUL == RISCVVType::VLMUL::LMUL_F2 ||
999 LMUL == RISCVVType::VLMUL::LMUL_1);
1000}
1001
1002InstructionCost RISCVTTIImpl::getScalarizationOverhead(
1003 VectorType *Ty, const APInt &DemandedElts, bool Insert, bool Extract,
1004 TTI::TargetCostKind CostKind, bool ForPoisonSrc, ArrayRef<Value *> VL,
1005 TTI::VectorInstrContext VIC) const {
1006 if (isa<ScalableVectorType>(Val: Ty))
1007 return InstructionCost::getInvalid();
1008
1009 // TODO: Add proper cost model for P extension fixed vectors (e.g., v4i16)
1010 // For now, skip all fixed vector cost analysis when P extension is available
1011 // to avoid crashes in getMinRVVVectorSizeInBits()
1012 if (ST->hasStdExtP() && isa<FixedVectorType>(Val: Ty)) {
1013 return 1; // Treat as single instruction cost for now
1014 }
1015
1016 // A build_vector (which is m1 sized or smaller) can be done in no
1017 // worse than one vslide1down.vx per element in the type. We could
1018 // in theory do an explode_vector in the inverse manner, but our
1019 // lowering today does not have a first class node for this pattern.
1020 InstructionCost Cost = BaseT::getScalarizationOverhead(
1021 InTy: Ty, DemandedElts, Insert, Extract, CostKind);
1022 std::pair<InstructionCost, MVT> LT = getTypeLegalizationCost(Ty);
1023 if (Insert && !Extract && LT.first.isValid() && LT.second.isVector()) {
1024 if (Ty->getScalarSizeInBits() == 1) {
1025 auto *WideVecTy = cast<VectorType>(Val: Ty->getWithNewBitWidth(NewBitWidth: 8));
1026 // Note: Implicit scalar anyextend is assumed to be free since the i1
1027 // must be stored in a GPR.
1028 return getScalarizationOverhead(Ty: WideVecTy, DemandedElts, Insert, Extract,
1029 CostKind) +
1030 getCastInstrCost(Opcode: Instruction::Trunc, Dst: Ty, Src: WideVecTy,
1031 CCH: TTI::CastContextHint::None, CostKind, I: nullptr);
1032 }
1033
1034 assert(LT.second.isFixedLengthVector());
1035 MVT ContainerVT = TLI->getContainerForFixedLengthVector(VT: LT.second);
1036 if (isM1OrSmaller(VT: ContainerVT)) {
1037 InstructionCost BV =
1038 cast<FixedVectorType>(Val: Ty)->getNumElements() *
1039 getRISCVInstructionCost(OpCodes: RISCV::VSLIDE1DOWN_VX, VT: LT.second, CostKind);
1040 if (BV < Cost)
1041 Cost = BV;
1042 }
1043 }
1044 return Cost;
1045}
1046
1047InstructionCost
1048RISCVTTIImpl::getMemIntrinsicInstrCost(const MemIntrinsicCostAttributes &MICA,
1049 TTI::TargetCostKind CostKind) const {
1050 Type *DataTy = MICA.getDataType();
1051 Align Alignment = MICA.getAlignment();
1052 switch (MICA.getID()) {
1053 case Intrinsic::vp_load_ff: {
1054 EVT DataTypeVT = TLI->getValueType(DL, Ty: DataTy);
1055 if (!TLI->isLegalFirstFaultLoad(DataType: DataTypeVT, Alignment))
1056 return BaseT::getMemIntrinsicInstrCost(MICA, CostKind);
1057
1058 unsigned AS = MICA.getAddressSpace();
1059 return getMemoryOpCost(Opcode: Instruction::Load, Src: DataTy, Alignment, AddressSpace: AS, CostKind,
1060 OpdInfo: {.Kind: TTI::OK_AnyValue, .Properties: TTI::OP_None}, I: nullptr);
1061 }
1062 case Intrinsic::experimental_vp_strided_load:
1063 case Intrinsic::experimental_vp_strided_store:
1064 return getStridedMemoryOpCost(MICA, CostKind);
1065 case Intrinsic::masked_compressstore:
1066 case Intrinsic::masked_expandload:
1067 return getExpandCompressMemoryOpCost(MICA, CostKind);
1068 case Intrinsic::vp_scatter:
1069 case Intrinsic::vp_gather:
1070 case Intrinsic::masked_scatter:
1071 case Intrinsic::masked_gather:
1072 return getGatherScatterOpCost(MICA, CostKind);
1073 case Intrinsic::vp_load:
1074 case Intrinsic::vp_store:
1075 case Intrinsic::masked_load:
1076 case Intrinsic::masked_store:
1077 return getMaskedMemoryOpCost(MICA, CostKind);
1078 }
1079 return BaseT::getMemIntrinsicInstrCost(MICA, CostKind);
1080}
1081
1082InstructionCost
1083RISCVTTIImpl::getMaskedMemoryOpCost(const MemIntrinsicCostAttributes &MICA,
1084 TTI::TargetCostKind CostKind) const {
1085 unsigned Opcode = MICA.getID() == Intrinsic::masked_load ? Instruction::Load
1086 : Instruction::Store;
1087 Type *Src = MICA.getDataType();
1088 Align Alignment = MICA.getAlignment();
1089 unsigned AddressSpace = MICA.getAddressSpace();
1090
1091 if (!isLegalMaskedLoadStore(DataType: Src, Alignment) ||
1092 CostKind != TTI::TCK_RecipThroughput)
1093 return BaseT::getMemIntrinsicInstrCost(MICA, CostKind);
1094
1095 return getMemoryOpCost(Opcode, Src, Alignment, AddressSpace, CostKind);
1096}
1097
1098InstructionCost RISCVTTIImpl::getInterleavedMemoryOpCost(
1099 unsigned Opcode, Type *VecTy, unsigned Factor, ArrayRef<unsigned> Indices,
1100 Align Alignment, unsigned AddressSpace, TTI::TargetCostKind CostKind,
1101 bool UseMaskForCond, bool UseMaskForGaps) const {
1102
1103 // The interleaved memory access pass will lower (de)interleave ops combined
1104 // with an adjacent appropriate memory to vlseg/vsseg intrinsics. vlseg/vsseg
1105 // only support masking per-iteration (i.e. condition), not per-segment (i.e.
1106 // gap).
1107 if (!UseMaskForGaps && Factor <= TLI->getMaxSupportedInterleaveFactor()) {
1108 auto *VTy = cast<VectorType>(Val: VecTy);
1109 std::pair<InstructionCost, MVT> LT = getTypeLegalizationCost(Ty: VTy);
1110 // Need to make sure type has't been scalarized
1111 if (LT.second.isVector()) {
1112 auto *SubVecTy =
1113 VectorType::get(ElementType: VTy->getElementType(),
1114 EC: VTy->getElementCount().divideCoefficientBy(RHS: Factor));
1115 if (VTy->getElementCount().isKnownMultipleOf(RHS: Factor) &&
1116 TLI->isLegalInterleavedAccessType(VTy: SubVecTy, Factor, Alignment,
1117 AddrSpace: AddressSpace, DL)) {
1118
1119 // Some processors optimize segment loads/stores as one wide memory op +
1120 // Factor * LMUL shuffle ops.
1121 if (ST->hasOptimizedSegmentLoadStore(NF: Factor)) {
1122 InstructionCost Cost =
1123 getMemoryOpCost(Opcode, Src: VTy, Alignment, AddressSpace, CostKind);
1124 MVT SubVecVT = getTLI()->getValueType(DL, Ty: SubVecTy).getSimpleVT();
1125 Cost += Factor * TLI->getLMULCost(VT: SubVecVT);
1126 return LT.first * Cost;
1127 }
1128
1129 // Otherwise, the cost is proportional to the number of elements (VL *
1130 // Factor ops).
1131 InstructionCost MemOpCost =
1132 getMemoryOpCost(Opcode, Src: VTy->getElementType(), Alignment, AddressSpace: 0,
1133 CostKind, OpdInfo: {.Kind: TTI::OK_AnyValue, .Properties: TTI::OP_None});
1134 unsigned NumLoads = getEstimatedVLFor(Ty: VTy);
1135 return NumLoads * MemOpCost;
1136 }
1137 }
1138 }
1139
1140 // TODO: Return the cost of interleaved accesses for scalable vector when
1141 // unable to convert to segment accesses instructions.
1142 if (isa<ScalableVectorType>(Val: VecTy))
1143 return InstructionCost::getInvalid();
1144
1145 auto *FVTy = cast<FixedVectorType>(Val: VecTy);
1146 InstructionCost MemCost =
1147 getMemoryOpCost(Opcode, Src: VecTy, Alignment, AddressSpace, CostKind);
1148 unsigned VF = FVTy->getNumElements() / Factor;
1149
1150 // An interleaved load will look like this for Factor=3:
1151 // %wide.vec = load <12 x i32>, ptr %3, align 4
1152 // %strided.vec = shufflevector %wide.vec, poison, <4 x i32> <stride mask>
1153 // %strided.vec1 = shufflevector %wide.vec, poison, <4 x i32> <stride mask>
1154 // %strided.vec2 = shufflevector %wide.vec, poison, <4 x i32> <stride mask>
1155 if (Opcode == Instruction::Load) {
1156 InstructionCost Cost = MemCost;
1157 for (unsigned Index : Indices) {
1158 FixedVectorType *VecTy =
1159 FixedVectorType::get(ElementType: FVTy->getElementType(), NumElts: VF * Factor);
1160 auto Mask = createStrideMask(Start: Index, Stride: Factor, VF);
1161 Mask.resize(N: VF * Factor, NV: -1);
1162 InstructionCost ShuffleCost =
1163 getShuffleCost(Kind: TTI::ShuffleKind::SK_PermuteSingleSrc, DstTy: VecTy, SrcTy: VecTy,
1164 Mask, CostKind, Index: 0, SubTp: nullptr, Args: {});
1165 Cost += ShuffleCost;
1166 }
1167 return Cost;
1168 }
1169
1170 // TODO: Model for NF > 2
1171 // We'll need to enhance getShuffleCost to model shuffles that are just
1172 // inserts and extracts into subvectors, since they won't have the full cost
1173 // of a vrgather.
1174 // An interleaved store for 3 vectors of 4 lanes will look like
1175 // %11 = shufflevector <4 x i32> %4, <4 x i32> %6, <8 x i32> <0...7>
1176 // %12 = shufflevector <4 x i32> %9, <4 x i32> poison, <8 x i32> <0...3>
1177 // %13 = shufflevector <8 x i32> %11, <8 x i32> %12, <12 x i32> <0...11>
1178 // %interleaved.vec = shufflevector %13, poison, <12 x i32> <interleave mask>
1179 // store <12 x i32> %interleaved.vec, ptr %10, align 4
1180 if (Factor != 2)
1181 return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices,
1182 Alignment, AddressSpace, CostKind,
1183 UseMaskForCond, UseMaskForGaps);
1184
1185 assert(Opcode == Instruction::Store && "Opcode must be a store");
1186 // For an interleaving store of 2 vectors, we perform one large interleaving
1187 // shuffle that goes into the wide store
1188 auto Mask = createInterleaveMask(VF, NumVecs: Factor);
1189 InstructionCost ShuffleCost =
1190 getShuffleCost(Kind: TTI::ShuffleKind::SK_PermuteSingleSrc, DstTy: FVTy, SrcTy: FVTy, Mask,
1191 CostKind, Index: 0, SubTp: nullptr, Args: {});
1192 return MemCost + ShuffleCost;
1193}
1194
1195InstructionCost
1196RISCVTTIImpl::getGatherScatterOpCost(const MemIntrinsicCostAttributes &MICA,
1197 TTI::TargetCostKind CostKind) const {
1198
1199 bool IsLoad = MICA.getID() == Intrinsic::masked_gather ||
1200 MICA.getID() == Intrinsic::vp_gather;
1201 unsigned Opcode = IsLoad ? Instruction::Load : Instruction::Store;
1202 Type *DataTy = MICA.getDataType();
1203 Align Alignment = MICA.getAlignment();
1204 if (CostKind != TTI::TCK_RecipThroughput)
1205 return BaseT::getMemIntrinsicInstrCost(MICA, CostKind);
1206
1207 if ((Opcode == Instruction::Load &&
1208 !isLegalMaskedGather(DataType: DataTy, Alignment: Align(Alignment))) ||
1209 (Opcode == Instruction::Store &&
1210 !isLegalMaskedScatter(DataType: DataTy, Alignment: Align(Alignment))))
1211 return BaseT::getMemIntrinsicInstrCost(MICA, CostKind);
1212
1213 // Cost is proportional to the number of memory operations implied. For
1214 // scalable vectors, we use an estimate on that number since we don't
1215 // know exactly what VL will be.
1216 auto &VTy = *cast<VectorType>(Val: DataTy);
1217 unsigned NumLoads = getEstimatedVLFor(Ty: &VTy);
1218 return NumLoads * TTI::TCC_Basic;
1219}
1220
1221InstructionCost RISCVTTIImpl::getExpandCompressMemoryOpCost(
1222 const MemIntrinsicCostAttributes &MICA,
1223 TTI::TargetCostKind CostKind) const {
1224 unsigned Opcode = MICA.getID() == Intrinsic::masked_expandload
1225 ? Instruction::Load
1226 : Instruction::Store;
1227 Type *DataTy = MICA.getDataType();
1228 bool VariableMask = MICA.getVariableMask();
1229 Align Alignment = MICA.getAlignment();
1230 bool IsLegal = (Opcode == Instruction::Store &&
1231 isLegalMaskedCompressStore(DataTy, Alignment)) ||
1232 (Opcode == Instruction::Load &&
1233 isLegalMaskedExpandLoad(DataType: DataTy, Alignment));
1234 if (!IsLegal || CostKind != TTI::TCK_RecipThroughput)
1235 return BaseT::getMemIntrinsicInstrCost(MICA, CostKind);
1236 // Example compressstore sequence:
1237 // vsetivli zero, 8, e32, m2, ta, ma (ignored)
1238 // vcompress.vm v10, v8, v0
1239 // vcpop.m a1, v0
1240 // vsetvli zero, a1, e32, m2, ta, ma
1241 // vse32.v v10, (a0)
1242 // Example expandload sequence:
1243 // vsetivli zero, 8, e8, mf2, ta, ma (ignored)
1244 // vcpop.m a1, v0
1245 // vsetvli zero, a1, e32, m2, ta, ma
1246 // vle32.v v10, (a0)
1247 // vsetivli zero, 8, e32, m2, ta, ma
1248 // viota.m v12, v0
1249 // vrgather.vv v8, v10, v12, v0.t
1250 auto MemOpCost =
1251 getMemoryOpCost(Opcode, Src: DataTy, Alignment, /*AddressSpace*/ 0, CostKind);
1252 auto LT = getTypeLegalizationCost(Ty: DataTy);
1253 SmallVector<unsigned, 4> Opcodes{RISCV::VSETVLI};
1254 if (VariableMask)
1255 Opcodes.push_back(Elt: RISCV::VCPOP_M);
1256 if (Opcode == Instruction::Store)
1257 Opcodes.append(IL: {RISCV::VCOMPRESS_VM});
1258 else
1259 Opcodes.append(IL: {RISCV::VSETIVLI, RISCV::VIOTA_M, RISCV::VRGATHER_VV});
1260 return MemOpCost +
1261 LT.first * getRISCVInstructionCost(OpCodes: Opcodes, VT: LT.second, CostKind);
1262}
1263
1264InstructionCost
1265RISCVTTIImpl::getStridedMemoryOpCost(const MemIntrinsicCostAttributes &MICA,
1266 TTI::TargetCostKind CostKind) const {
1267
1268 unsigned Opcode = MICA.getID() == Intrinsic::experimental_vp_strided_load
1269 ? Instruction::Load
1270 : Instruction::Store;
1271
1272 Type *DataTy = MICA.getDataType();
1273 Align Alignment = MICA.getAlignment();
1274 const Instruction *I = MICA.getInst();
1275
1276 if (!isLegalStridedLoadStore(DataType: DataTy, Alignment))
1277 return BaseT::getMemIntrinsicInstrCost(MICA, CostKind);
1278
1279 if (CostKind == TTI::TCK_CodeSize)
1280 return TTI::TCC_Basic;
1281
1282 // Cost is proportional to the number of memory operations implied. For
1283 // scalable vectors, we use an estimate on that number since we don't
1284 // know exactly what VL will be.
1285 // FIXME: This will overcost for i64 on rv32 with +zve64x.
1286 auto &VTy = *cast<VectorType>(Val: DataTy);
1287 InstructionCost MemOpCost =
1288 getMemoryOpCost(Opcode, Src: VTy.getElementType(), Alignment, AddressSpace: 0, CostKind,
1289 OpdInfo: {.Kind: TTI::OK_AnyValue, .Properties: TTI::OP_None}, I);
1290 unsigned NumLoads = getEstimatedVLFor(Ty: &VTy);
1291 return NumLoads * MemOpCost;
1292}
1293
1294InstructionCost
1295RISCVTTIImpl::getCostOfKeepingLiveOverCall(ArrayRef<Type *> Tys) const {
1296 // FIXME: This is a property of the default vector convention, not
1297 // all possible calling conventions. Fixing that will require
1298 // some TTI API and SLP rework.
1299 InstructionCost Cost = 0;
1300 TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput;
1301 for (auto *Ty : Tys) {
1302 if (!Ty->isVectorTy())
1303 continue;
1304 Align A = DL.getPrefTypeAlign(Ty);
1305 Cost += getMemoryOpCost(Opcode: Instruction::Store, Src: Ty, Alignment: A, AddressSpace: 0, CostKind) +
1306 getMemoryOpCost(Opcode: Instruction::Load, Src: Ty, Alignment: A, AddressSpace: 0, CostKind);
1307 }
1308 return Cost;
1309}
1310
1311// Currently, these represent both throughput and codesize costs
1312// for the respective intrinsics. The costs in this table are simply
1313// instruction counts with the following adjustments made:
1314// * One vsetvli is considered free.
1315static const CostTblEntry VectorIntrinsicCostTable[]{
1316 {.ISD: Intrinsic::floor, .Type: MVT::f32, .Cost: 9},
1317 {.ISD: Intrinsic::floor, .Type: MVT::f64, .Cost: 9},
1318 {.ISD: Intrinsic::ceil, .Type: MVT::f32, .Cost: 9},
1319 {.ISD: Intrinsic::ceil, .Type: MVT::f64, .Cost: 9},
1320 {.ISD: Intrinsic::trunc, .Type: MVT::f32, .Cost: 7},
1321 {.ISD: Intrinsic::trunc, .Type: MVT::f64, .Cost: 7},
1322 {.ISD: Intrinsic::round, .Type: MVT::f32, .Cost: 9},
1323 {.ISD: Intrinsic::round, .Type: MVT::f64, .Cost: 9},
1324 {.ISD: Intrinsic::roundeven, .Type: MVT::f32, .Cost: 9},
1325 {.ISD: Intrinsic::roundeven, .Type: MVT::f64, .Cost: 9},
1326 {.ISD: Intrinsic::rint, .Type: MVT::f32, .Cost: 7},
1327 {.ISD: Intrinsic::rint, .Type: MVT::f64, .Cost: 7},
1328 {.ISD: Intrinsic::nearbyint, .Type: MVT::f32, .Cost: 9},
1329 {.ISD: Intrinsic::nearbyint, .Type: MVT::f64, .Cost: 9},
1330 {.ISD: Intrinsic::bswap, .Type: MVT::i16, .Cost: 3},
1331 {.ISD: Intrinsic::bswap, .Type: MVT::i32, .Cost: 12},
1332 {.ISD: Intrinsic::bswap, .Type: MVT::i64, .Cost: 31},
1333 {.ISD: Intrinsic::vp_bswap, .Type: MVT::i16, .Cost: 3},
1334 {.ISD: Intrinsic::vp_bswap, .Type: MVT::i32, .Cost: 12},
1335 {.ISD: Intrinsic::vp_bswap, .Type: MVT::i64, .Cost: 31},
1336 {.ISD: Intrinsic::vp_fshl, .Type: MVT::i8, .Cost: 7},
1337 {.ISD: Intrinsic::vp_fshl, .Type: MVT::i16, .Cost: 7},
1338 {.ISD: Intrinsic::vp_fshl, .Type: MVT::i32, .Cost: 7},
1339 {.ISD: Intrinsic::vp_fshl, .Type: MVT::i64, .Cost: 7},
1340 {.ISD: Intrinsic::vp_fshr, .Type: MVT::i8, .Cost: 7},
1341 {.ISD: Intrinsic::vp_fshr, .Type: MVT::i16, .Cost: 7},
1342 {.ISD: Intrinsic::vp_fshr, .Type: MVT::i32, .Cost: 7},
1343 {.ISD: Intrinsic::vp_fshr, .Type: MVT::i64, .Cost: 7},
1344 {.ISD: Intrinsic::bitreverse, .Type: MVT::i8, .Cost: 17},
1345 {.ISD: Intrinsic::bitreverse, .Type: MVT::i16, .Cost: 24},
1346 {.ISD: Intrinsic::bitreverse, .Type: MVT::i32, .Cost: 33},
1347 {.ISD: Intrinsic::bitreverse, .Type: MVT::i64, .Cost: 52},
1348 {.ISD: Intrinsic::vp_bitreverse, .Type: MVT::i8, .Cost: 17},
1349 {.ISD: Intrinsic::vp_bitreverse, .Type: MVT::i16, .Cost: 24},
1350 {.ISD: Intrinsic::vp_bitreverse, .Type: MVT::i32, .Cost: 33},
1351 {.ISD: Intrinsic::vp_bitreverse, .Type: MVT::i64, .Cost: 52},
1352 {.ISD: Intrinsic::ctpop, .Type: MVT::i8, .Cost: 12},
1353 {.ISD: Intrinsic::ctpop, .Type: MVT::i16, .Cost: 19},
1354 {.ISD: Intrinsic::ctpop, .Type: MVT::i32, .Cost: 20},
1355 {.ISD: Intrinsic::ctpop, .Type: MVT::i64, .Cost: 21},
1356 {.ISD: Intrinsic::ctlz, .Type: MVT::i8, .Cost: 19},
1357 {.ISD: Intrinsic::ctlz, .Type: MVT::i16, .Cost: 28},
1358 {.ISD: Intrinsic::ctlz, .Type: MVT::i32, .Cost: 31},
1359 {.ISD: Intrinsic::ctlz, .Type: MVT::i64, .Cost: 35},
1360 {.ISD: Intrinsic::cttz, .Type: MVT::i8, .Cost: 16},
1361 {.ISD: Intrinsic::cttz, .Type: MVT::i16, .Cost: 23},
1362 {.ISD: Intrinsic::cttz, .Type: MVT::i32, .Cost: 24},
1363 {.ISD: Intrinsic::cttz, .Type: MVT::i64, .Cost: 25},
1364 {.ISD: Intrinsic::vp_ctpop, .Type: MVT::i8, .Cost: 12},
1365 {.ISD: Intrinsic::vp_ctpop, .Type: MVT::i16, .Cost: 19},
1366 {.ISD: Intrinsic::vp_ctpop, .Type: MVT::i32, .Cost: 20},
1367 {.ISD: Intrinsic::vp_ctpop, .Type: MVT::i64, .Cost: 21},
1368 {.ISD: Intrinsic::vp_ctlz, .Type: MVT::i8, .Cost: 19},
1369 {.ISD: Intrinsic::vp_ctlz, .Type: MVT::i16, .Cost: 28},
1370 {.ISD: Intrinsic::vp_ctlz, .Type: MVT::i32, .Cost: 31},
1371 {.ISD: Intrinsic::vp_ctlz, .Type: MVT::i64, .Cost: 35},
1372 {.ISD: Intrinsic::vp_cttz, .Type: MVT::i8, .Cost: 16},
1373 {.ISD: Intrinsic::vp_cttz, .Type: MVT::i16, .Cost: 23},
1374 {.ISD: Intrinsic::vp_cttz, .Type: MVT::i32, .Cost: 24},
1375 {.ISD: Intrinsic::vp_cttz, .Type: MVT::i64, .Cost: 25},
1376};
1377
1378InstructionCost
1379RISCVTTIImpl::getIntrinsicInstrCost(const IntrinsicCostAttributes &ICA,
1380 TTI::TargetCostKind CostKind) const {
1381 auto *RetTy = ICA.getReturnType();
1382 switch (ICA.getID()) {
1383 case Intrinsic::lrint:
1384 case Intrinsic::llrint:
1385 case Intrinsic::lround:
1386 case Intrinsic::llround: {
1387 auto LT = getTypeLegalizationCost(Ty: RetTy);
1388 Type *SrcTy = ICA.getArgTypes().front();
1389 auto SrcLT = getTypeLegalizationCost(Ty: SrcTy);
1390 if (ST->hasVInstructions() && LT.second.isVector()) {
1391 SmallVector<unsigned, 2> Ops;
1392 unsigned SrcEltSz = DL.getTypeSizeInBits(Ty: SrcTy->getScalarType());
1393 unsigned DstEltSz = DL.getTypeSizeInBits(Ty: RetTy->getScalarType());
1394 if (LT.second.getVectorElementType() == MVT::bf16) {
1395 if (!ST->hasVInstructionsBF16Minimal())
1396 return InstructionCost::getInvalid();
1397 if (DstEltSz == 32)
1398 Ops = {RISCV::VFWCVTBF16_F_F_V, RISCV::VFCVT_X_F_V};
1399 else
1400 Ops = {RISCV::VFWCVTBF16_F_F_V, RISCV::VFWCVT_X_F_V};
1401 } else if (LT.second.getVectorElementType() == MVT::f16 &&
1402 !ST->hasVInstructionsF16()) {
1403 if (!ST->hasVInstructionsF16Minimal())
1404 return InstructionCost::getInvalid();
1405 if (DstEltSz == 32)
1406 Ops = {RISCV::VFWCVT_F_F_V, RISCV::VFCVT_X_F_V};
1407 else
1408 Ops = {RISCV::VFWCVT_F_F_V, RISCV::VFWCVT_X_F_V};
1409
1410 } else if (SrcEltSz > DstEltSz) {
1411 Ops = {RISCV::VFNCVT_X_F_W};
1412 } else if (SrcEltSz < DstEltSz) {
1413 Ops = {RISCV::VFWCVT_X_F_V};
1414 } else {
1415 Ops = {RISCV::VFCVT_X_F_V};
1416 }
1417
1418 // We need to use the source LMUL in the case of a narrowing op, and the
1419 // destination LMUL otherwise.
1420 if (SrcEltSz > DstEltSz)
1421 return SrcLT.first *
1422 getRISCVInstructionCost(OpCodes: Ops, VT: SrcLT.second, CostKind);
1423 return LT.first * getRISCVInstructionCost(OpCodes: Ops, VT: LT.second, CostKind);
1424 }
1425 break;
1426 }
1427 case Intrinsic::ceil:
1428 case Intrinsic::floor:
1429 case Intrinsic::trunc:
1430 case Intrinsic::rint:
1431 case Intrinsic::round:
1432 case Intrinsic::roundeven: {
1433 // These all use the same code.
1434 auto LT = getTypeLegalizationCost(Ty: RetTy);
1435 if (!LT.second.isVector() && TLI->isOperationCustom(Op: ISD::FCEIL, VT: LT.second))
1436 return LT.first * 8;
1437 break;
1438 }
1439 case Intrinsic::umin:
1440 case Intrinsic::umax:
1441 case Intrinsic::smin:
1442 case Intrinsic::smax: {
1443 auto LT = getTypeLegalizationCost(Ty: RetTy);
1444 if (LT.second.isScalarInteger() && ST->hasStdExtZbb())
1445 return LT.first;
1446
1447 if (ST->hasVInstructions() && LT.second.isVector()) {
1448 unsigned Op;
1449 switch (ICA.getID()) {
1450 case Intrinsic::umin:
1451 Op = RISCV::VMINU_VV;
1452 break;
1453 case Intrinsic::umax:
1454 Op = RISCV::VMAXU_VV;
1455 break;
1456 case Intrinsic::smin:
1457 Op = RISCV::VMIN_VV;
1458 break;
1459 case Intrinsic::smax:
1460 Op = RISCV::VMAX_VV;
1461 break;
1462 }
1463 return LT.first * getRISCVInstructionCost(OpCodes: Op, VT: LT.second, CostKind);
1464 }
1465 break;
1466 }
1467 case Intrinsic::sadd_sat:
1468 case Intrinsic::ssub_sat:
1469 case Intrinsic::uadd_sat:
1470 case Intrinsic::usub_sat: {
1471 auto LT = getTypeLegalizationCost(Ty: RetTy);
1472 if (ST->hasVInstructions() && LT.second.isVector()) {
1473 unsigned Op;
1474 switch (ICA.getID()) {
1475 case Intrinsic::sadd_sat:
1476 Op = RISCV::VSADD_VV;
1477 break;
1478 case Intrinsic::ssub_sat:
1479 Op = RISCV::VSSUB_VV;
1480 break;
1481 case Intrinsic::uadd_sat:
1482 Op = RISCV::VSADDU_VV;
1483 break;
1484 case Intrinsic::usub_sat:
1485 Op = RISCV::VSSUBU_VV;
1486 break;
1487 }
1488 return LT.first * getRISCVInstructionCost(OpCodes: Op, VT: LT.second, CostKind);
1489 }
1490 break;
1491 }
1492 case Intrinsic::fma:
1493 case Intrinsic::fmuladd: {
1494 // TODO: handle promotion with f16/bf16 with zvfhmin/zvfbfmin
1495 auto LT = getTypeLegalizationCost(Ty: RetTy);
1496 if (ST->hasVInstructions() && LT.second.isVector())
1497 return LT.first *
1498 getRISCVInstructionCost(OpCodes: RISCV::VFMADD_VV, VT: LT.second, CostKind);
1499 break;
1500 }
1501 case Intrinsic::fabs: {
1502 auto LT = getTypeLegalizationCost(Ty: RetTy);
1503 if (ST->hasVInstructions() && LT.second.isVector()) {
1504 // lui a0, 8
1505 // addi a0, a0, -1
1506 // vsetvli a1, zero, e16, m1, ta, ma
1507 // vand.vx v8, v8, a0
1508 // f16 with zvfhmin and bf16 with zvfhbmin
1509 if (LT.second.getVectorElementType() == MVT::bf16 ||
1510 (LT.second.getVectorElementType() == MVT::f16 &&
1511 !ST->hasVInstructionsF16()))
1512 return LT.first * getRISCVInstructionCost(OpCodes: RISCV::VAND_VX, VT: LT.second,
1513 CostKind) +
1514 2;
1515 else
1516 return LT.first *
1517 getRISCVInstructionCost(OpCodes: RISCV::VFSGNJX_VV, VT: LT.second, CostKind);
1518 }
1519 break;
1520 }
1521 case Intrinsic::sqrt: {
1522 auto LT = getTypeLegalizationCost(Ty: RetTy);
1523 if (ST->hasVInstructions() && LT.second.isVector()) {
1524 SmallVector<unsigned, 4> ConvOp;
1525 SmallVector<unsigned, 2> FsqrtOp;
1526 MVT ConvType = LT.second;
1527 MVT FsqrtType = LT.second;
1528 // f16 with zvfhmin and bf16 with zvfbfmin and the type of nxv32[b]f16
1529 // will be spilt.
1530 if (LT.second.getVectorElementType() == MVT::bf16) {
1531 if (LT.second == MVT::nxv32bf16) {
1532 ConvOp = {RISCV::VFWCVTBF16_F_F_V, RISCV::VFWCVTBF16_F_F_V,
1533 RISCV::VFNCVTBF16_F_F_W, RISCV::VFNCVTBF16_F_F_W};
1534 FsqrtOp = {RISCV::VFSQRT_V, RISCV::VFSQRT_V};
1535 ConvType = MVT::nxv16f16;
1536 FsqrtType = MVT::nxv16f32;
1537 } else {
1538 ConvOp = {RISCV::VFWCVTBF16_F_F_V, RISCV::VFNCVTBF16_F_F_W};
1539 FsqrtOp = {RISCV::VFSQRT_V};
1540 FsqrtType = TLI->getTypeToPromoteTo(Op: ISD::FSQRT, VT: FsqrtType);
1541 }
1542 } else if (LT.second.getVectorElementType() == MVT::f16 &&
1543 !ST->hasVInstructionsF16()) {
1544 if (LT.second == MVT::nxv32f16) {
1545 ConvOp = {RISCV::VFWCVT_F_F_V, RISCV::VFWCVT_F_F_V,
1546 RISCV::VFNCVT_F_F_W, RISCV::VFNCVT_F_F_W};
1547 FsqrtOp = {RISCV::VFSQRT_V, RISCV::VFSQRT_V};
1548 ConvType = MVT::nxv16f16;
1549 FsqrtType = MVT::nxv16f32;
1550 } else {
1551 ConvOp = {RISCV::VFWCVT_F_F_V, RISCV::VFNCVT_F_F_W};
1552 FsqrtOp = {RISCV::VFSQRT_V};
1553 FsqrtType = TLI->getTypeToPromoteTo(Op: ISD::FSQRT, VT: FsqrtType);
1554 }
1555 } else {
1556 FsqrtOp = {RISCV::VFSQRT_V};
1557 }
1558
1559 return LT.first * (getRISCVInstructionCost(OpCodes: FsqrtOp, VT: FsqrtType, CostKind) +
1560 getRISCVInstructionCost(OpCodes: ConvOp, VT: ConvType, CostKind));
1561 }
1562 break;
1563 }
1564 case Intrinsic::cttz:
1565 case Intrinsic::ctlz:
1566 case Intrinsic::ctpop: {
1567 auto LT = getTypeLegalizationCost(Ty: RetTy);
1568 if (ST->hasStdExtZvbb() && LT.second.isVector()) {
1569 unsigned Op;
1570 switch (ICA.getID()) {
1571 case Intrinsic::cttz:
1572 Op = RISCV::VCTZ_V;
1573 break;
1574 case Intrinsic::ctlz:
1575 Op = RISCV::VCLZ_V;
1576 break;
1577 case Intrinsic::ctpop:
1578 Op = RISCV::VCPOP_V;
1579 break;
1580 }
1581 return LT.first * getRISCVInstructionCost(OpCodes: Op, VT: LT.second, CostKind);
1582 }
1583 break;
1584 }
1585 case Intrinsic::abs: {
1586 auto LT = getTypeLegalizationCost(Ty: RetTy);
1587 if (ST->hasVInstructions() && LT.second.isVector()) {
1588 // vabs.v v10, v8
1589 if (ST->hasStdExtZvabd())
1590 return LT.first *
1591 getRISCVInstructionCost(OpCodes: {RISCV::VABS_V}, VT: LT.second, CostKind);
1592
1593 // vrsub.vi v10, v8, 0
1594 // vmax.vv v8, v8, v10
1595 return LT.first *
1596 getRISCVInstructionCost(OpCodes: {RISCV::VRSUB_VI, RISCV::VMAX_VV},
1597 VT: LT.second, CostKind);
1598 }
1599 break;
1600 }
1601 case Intrinsic::fshl:
1602 case Intrinsic::fshr: {
1603 if (ICA.getArgs().empty())
1604 break;
1605
1606 // Funnel-shifts are ROTL/ROTR when the first and second operand are equal.
1607 // When Zbb/Zbkb is enabled we can use a single ROL(W)/ROR(I)(W)
1608 // instruction.
1609 if ((ST->hasStdExtZbb() || ST->hasStdExtZbkb()) && RetTy->isIntegerTy() &&
1610 ICA.getArgs()[0] == ICA.getArgs()[1] &&
1611 (RetTy->getIntegerBitWidth() == 32 ||
1612 RetTy->getIntegerBitWidth() == 64) &&
1613 RetTy->getIntegerBitWidth() <= ST->getXLen()) {
1614 return 1;
1615 }
1616 break;
1617 }
1618 case Intrinsic::get_active_lane_mask: {
1619 if (ST->hasVInstructions()) {
1620 Type *ExpRetTy = VectorType::get(
1621 ElementType: ICA.getArgTypes()[0], EC: cast<VectorType>(Val: RetTy)->getElementCount());
1622 auto LT = getTypeLegalizationCost(Ty: ExpRetTy);
1623
1624 // vid.v v8 // considered hoisted
1625 // vsaddu.vx v8, v8, a0
1626 // vmsltu.vx v0, v8, a1
1627 return LT.first *
1628 getRISCVInstructionCost(OpCodes: {RISCV::VSADDU_VX, RISCV::VMSLTU_VX},
1629 VT: LT.second, CostKind);
1630 }
1631 break;
1632 }
1633 // TODO: add more intrinsic
1634 case Intrinsic::stepvector: {
1635 auto LT = getTypeLegalizationCost(Ty: RetTy);
1636 // Legalisation of illegal types involves an `index' instruction plus
1637 // (LT.first - 1) vector adds.
1638 if (ST->hasVInstructions())
1639 return getRISCVInstructionCost(OpCodes: RISCV::VID_V, VT: LT.second, CostKind) +
1640 (LT.first - 1) *
1641 getRISCVInstructionCost(OpCodes: RISCV::VADD_VX, VT: LT.second, CostKind);
1642 return 1 + (LT.first - 1);
1643 }
1644 case Intrinsic::vector_splice_left:
1645 case Intrinsic::vector_splice_right: {
1646 auto LT = getTypeLegalizationCost(Ty: RetTy);
1647 // Constant offsets fall through to getShuffleCost.
1648 if (!ICA.isTypeBasedOnly() && isa<ConstantInt>(Val: ICA.getArgs()[2]))
1649 break;
1650 if (ST->hasVInstructions() && LT.second.isVector()) {
1651 return LT.first *
1652 getRISCVInstructionCost(OpCodes: {RISCV::VSLIDEDOWN_VX, RISCV::VSLIDEUP_VX},
1653 VT: LT.second, CostKind);
1654 }
1655 break;
1656 }
1657 case Intrinsic::experimental_cttz_elts: {
1658 Type *ArgTy = ICA.getArgTypes()[0];
1659 EVT ArgType = TLI->getValueType(DL, Ty: ArgTy, AllowUnknown: true);
1660 if (getTLI()->shouldExpandCttzElements(VT: ArgType))
1661 break;
1662 InstructionCost Cost = getRISCVInstructionCost(
1663 OpCodes: RISCV::VFIRST_M, VT: getTypeLegalizationCost(Ty: ArgTy).second, CostKind);
1664
1665 // If zero_is_poison is false, then we will generate additional
1666 // cmp + select instructions to convert -1 to EVL.
1667 Type *BoolTy = Type::getInt1Ty(C&: RetTy->getContext());
1668 if (ICA.getArgs().size() > 1 &&
1669 cast<ConstantInt>(Val: ICA.getArgs()[1])->isZero())
1670 Cost += getCmpSelInstrCost(Opcode: Instruction::ICmp, ValTy: BoolTy, CondTy: RetTy,
1671 VecPred: CmpInst::ICMP_SLT, CostKind) +
1672 getCmpSelInstrCost(Opcode: Instruction::Select, ValTy: RetTy, CondTy: BoolTy,
1673 VecPred: CmpInst::BAD_ICMP_PREDICATE, CostKind);
1674
1675 return Cost;
1676 }
1677 case Intrinsic::experimental_vp_splice: {
1678 // To support type-based query from vectorizer, set the index to 0.
1679 // Note that index only change the cost from vslide.vx to vslide.vi and in
1680 // current implementations they have same costs.
1681 return getShuffleCost(Kind: TTI::SK_Splice, DstTy: cast<VectorType>(Val: ICA.getReturnType()),
1682 SrcTy: cast<VectorType>(Val: ICA.getArgTypes()[0]), Mask: {}, CostKind,
1683 Index: 0, SubTp: cast<VectorType>(Val: ICA.getReturnType()));
1684 }
1685 case Intrinsic::fptoui_sat:
1686 case Intrinsic::fptosi_sat: {
1687 InstructionCost Cost = 0;
1688 bool IsSigned = ICA.getID() == Intrinsic::fptosi_sat;
1689 Type *SrcTy = ICA.getArgTypes()[0];
1690
1691 auto SrcLT = getTypeLegalizationCost(Ty: SrcTy);
1692 auto DstLT = getTypeLegalizationCost(Ty: RetTy);
1693 if (!SrcTy->isVectorTy())
1694 break;
1695
1696 if (!SrcLT.first.isValid() || !DstLT.first.isValid())
1697 return InstructionCost::getInvalid();
1698
1699 Cost +=
1700 getCastInstrCost(Opcode: IsSigned ? Instruction::FPToSI : Instruction::FPToUI,
1701 Dst: RetTy, Src: SrcTy, CCH: TTI::CastContextHint::None, CostKind);
1702
1703 // Handle NaN.
1704 // vmfne v0, v8, v8 # If v8[i] is NaN set v0[i] to 1.
1705 // vmerge.vim v8, v8, 0, v0 # Convert NaN to 0.
1706 Type *CondTy = RetTy->getWithNewBitWidth(NewBitWidth: 1);
1707 Cost += getCmpSelInstrCost(Opcode: BinaryOperator::FCmp, ValTy: SrcTy, CondTy,
1708 VecPred: CmpInst::FCMP_UNO, CostKind);
1709 Cost += getCmpSelInstrCost(Opcode: BinaryOperator::Select, ValTy: RetTy, CondTy,
1710 VecPred: CmpInst::FCMP_UNO, CostKind);
1711 return Cost;
1712 }
1713 case Intrinsic::experimental_vector_extract_last_active: {
1714 auto *ValTy = cast<VectorType>(Val: ICA.getArgTypes()[0]);
1715 auto *MaskTy = cast<VectorType>(Val: ICA.getArgTypes()[1]);
1716
1717 auto ValLT = getTypeLegalizationCost(Ty: ValTy);
1718 auto MaskLT = getTypeLegalizationCost(Ty: MaskTy);
1719
1720 // TODO: Return cheaper cost when the entire lane is inactive.
1721 // The expected asm sequence is:
1722 // vcpop.m a0, v0
1723 // beqz a0, exit # Return passthru when the entire lane is inactive.
1724 // vid v10, v0.t
1725 // vredmaxu.vs v10, v10, v10
1726 // vmv.x.s a0, v10
1727 // zext.b a0, a0
1728 // vslidedown.vx v8, v8, a0
1729 // vmv.x.s a0, v8
1730 // exit:
1731 // ...
1732
1733 // Find a suitable type for a stepvector.
1734 ConstantRange VScaleRange(APInt(64, 1), APInt::getZero(numBits: 64));
1735 unsigned EltWidth = getTLI()->getBitWidthForCttzElements(
1736 RetVT: TLI->getVectorIdxTy(DL: getDataLayout()), EC: MaskTy->getElementCount(),
1737 /*ZeroIsPoison=*/true, VScaleRange: &VScaleRange);
1738 EltWidth = std::max(a: EltWidth, b: MaskTy->getScalarSizeInBits());
1739 Type *StepTy = Type::getIntNTy(C&: MaskTy->getContext(), N: EltWidth);
1740 auto *StepVecTy = VectorType::get(ElementType: StepTy, EC: ValTy->getElementCount());
1741 auto StepLT = getTypeLegalizationCost(Ty: StepVecTy);
1742
1743 // Currently expandVectorFindLastActive cannot handle step vector split.
1744 // So return invalid when the type needs split.
1745 // FIXME: Remove this if expandVectorFindLastActive supports split vector.
1746 if (StepLT.first > 1)
1747 return InstructionCost::getInvalid();
1748
1749 InstructionCost Cost = 0;
1750 unsigned Opcodes[] = {RISCV::VID_V, RISCV::VREDMAXU_VS, RISCV::VMV_X_S};
1751
1752 Cost += MaskLT.first *
1753 getRISCVInstructionCost(OpCodes: RISCV::VCPOP_M, VT: MaskLT.second, CostKind);
1754 Cost += getCFInstrCost(Opcode: Instruction::CondBr, CostKind, I: nullptr);
1755 Cost += StepLT.first *
1756 getRISCVInstructionCost(OpCodes: Opcodes, VT: StepLT.second, CostKind);
1757 Cost += getCastInstrCost(Opcode: Instruction::ZExt,
1758 Dst: Type::getInt64Ty(C&: ValTy->getContext()), Src: StepTy,
1759 CCH: TTI::CastContextHint::None, CostKind, I: nullptr);
1760 Cost += ValLT.first *
1761 getRISCVInstructionCost(OpCodes: {RISCV::VSLIDEDOWN_VI, RISCV::VMV_X_S},
1762 VT: ValLT.second, CostKind);
1763 return Cost;
1764 }
1765 }
1766
1767 if (ST->hasVInstructions() && RetTy->isVectorTy()) {
1768 if (auto LT = getTypeLegalizationCost(Ty: RetTy);
1769 LT.second.isVector()) {
1770 MVT EltTy = LT.second.getVectorElementType();
1771 if (const auto *Entry = CostTableLookup(Table: VectorIntrinsicCostTable,
1772 ISD: ICA.getID(), Ty: EltTy))
1773 return LT.first * Entry->Cost;
1774 }
1775 }
1776
1777 return BaseT::getIntrinsicInstrCost(ICA, CostKind);
1778}
1779
1780InstructionCost
1781RISCVTTIImpl::getAddressComputationCost(Type *PtrTy, ScalarEvolution *SE,
1782 const SCEV *Ptr,
1783 TTI::TargetCostKind CostKind) const {
1784 // Address computations for vector indexed load/store likely require an offset
1785 // and/or scaling.
1786 if (ST->hasVInstructions() && PtrTy->isVectorTy())
1787 return getArithmeticInstrCost(Opcode: Instruction::Add, Ty: PtrTy, CostKind);
1788
1789 return BaseT::getAddressComputationCost(PtrTy, SE, Ptr, CostKind);
1790}
1791
1792InstructionCost RISCVTTIImpl::getCastInstrCost(unsigned Opcode, Type *Dst,
1793 Type *Src,
1794 TTI::CastContextHint CCH,
1795 TTI::TargetCostKind CostKind,
1796 const Instruction *I) const {
1797 bool IsVectorType = isa<VectorType>(Val: Dst) && isa<VectorType>(Val: Src);
1798 if (!IsVectorType)
1799 return BaseT::getCastInstrCost(Opcode, Dst, Src, CCH, CostKind, I);
1800
1801 // TODO: Add proper cost model for P extension fixed vectors (e.g., v4i16)
1802 // For now, skip all fixed vector cost analysis when P extension is available
1803 // to avoid crashes in getMinRVVVectorSizeInBits()
1804 if (ST->hasStdExtP() &&
1805 (isa<FixedVectorType>(Val: Dst) || isa<FixedVectorType>(Val: Src))) {
1806 return 1; // Treat as single instruction cost for now
1807 }
1808
1809 // FIXME: Need to compute legalizing cost for illegal types. The current
1810 // code handles only legal types and those which can be trivially
1811 // promoted to legal.
1812 if (!ST->hasVInstructions() || Src->getScalarSizeInBits() > ST->getELen() ||
1813 Dst->getScalarSizeInBits() > ST->getELen())
1814 return BaseT::getCastInstrCost(Opcode, Dst, Src, CCH, CostKind, I);
1815
1816 int ISD = TLI->InstructionOpcodeToISD(Opcode);
1817 assert(ISD && "Invalid opcode");
1818 std::pair<InstructionCost, MVT> SrcLT = getTypeLegalizationCost(Ty: Src);
1819 std::pair<InstructionCost, MVT> DstLT = getTypeLegalizationCost(Ty: Dst);
1820
1821 // Handle i1 source and dest cases *before* calling logic in BasicTTI.
1822 // The shared implementation doesn't model vector widening during legalization
1823 // and instead assumes scalarization. In order to scalarize an <N x i1>
1824 // vector, we need to extend/trunc to/from i8. If we don't special case
1825 // this, we can get an infinite recursion cycle.
1826 switch (ISD) {
1827 default:
1828 break;
1829 case ISD::SIGN_EXTEND:
1830 case ISD::ZERO_EXTEND:
1831 if (Src->getScalarSizeInBits() == 1) {
1832 // We do not use vsext/vzext to extend from mask vector.
1833 // Instead we use the following instructions to extend from mask vector:
1834 // vmv.v.i v8, 0
1835 // vmerge.vim v8, v8, -1, v0 (repeated per split)
1836 return getRISCVInstructionCost(OpCodes: RISCV::VMV_V_I, VT: DstLT.second, CostKind) +
1837 DstLT.first * getRISCVInstructionCost(OpCodes: RISCV::VMERGE_VIM,
1838 VT: DstLT.second, CostKind) +
1839 DstLT.first - 1;
1840 }
1841 break;
1842 case ISD::TRUNCATE:
1843 if (Dst->getScalarSizeInBits() == 1) {
1844 // We do not use several vncvt to truncate to mask vector. So we could
1845 // not use PowDiff to calculate it.
1846 // Instead we use the following instructions to truncate to mask vector:
1847 // vand.vi v8, v8, 1
1848 // vmsne.vi v0, v8, 0
1849 return SrcLT.first *
1850 getRISCVInstructionCost(OpCodes: {RISCV::VAND_VI, RISCV::VMSNE_VI},
1851 VT: SrcLT.second, CostKind) +
1852 SrcLT.first - 1;
1853 }
1854 break;
1855 };
1856
1857 // Our actual lowering for the case where a wider legal type is available
1858 // uses promotion to the wider type. This is reflected in the result of
1859 // getTypeLegalizationCost, but BasicTTI assumes the widened cases are
1860 // scalarized if the legalized Src and Dst are not equal sized.
1861 const DataLayout &DL = this->getDataLayout();
1862 if (!SrcLT.second.isVector() || !DstLT.second.isVector() ||
1863 !SrcLT.first.isValid() || !DstLT.first.isValid() ||
1864 !TypeSize::isKnownLE(LHS: DL.getTypeSizeInBits(Ty: Src),
1865 RHS: SrcLT.second.getSizeInBits()) ||
1866 !TypeSize::isKnownLE(LHS: DL.getTypeSizeInBits(Ty: Dst),
1867 RHS: DstLT.second.getSizeInBits()) ||
1868 SrcLT.first > 1 || DstLT.first > 1)
1869 return BaseT::getCastInstrCost(Opcode, Dst, Src, CCH, CostKind, I);
1870
1871 // The split cost is handled by the base getCastInstrCost
1872 assert((SrcLT.first == 1) && (DstLT.first == 1) && "Illegal type");
1873
1874 int PowDiff = (int)Log2_32(Value: DstLT.second.getScalarSizeInBits()) -
1875 (int)Log2_32(Value: SrcLT.second.getScalarSizeInBits());
1876 switch (ISD) {
1877 case ISD::SIGN_EXTEND:
1878 case ISD::ZERO_EXTEND: {
1879 if ((PowDiff < 1) || (PowDiff > 3))
1880 return BaseT::getCastInstrCost(Opcode, Dst, Src, CCH, CostKind, I);
1881 unsigned SExtOp[] = {RISCV::VSEXT_VF2, RISCV::VSEXT_VF4, RISCV::VSEXT_VF8};
1882 unsigned ZExtOp[] = {RISCV::VZEXT_VF2, RISCV::VZEXT_VF4, RISCV::VZEXT_VF8};
1883 unsigned Op =
1884 (ISD == ISD::SIGN_EXTEND) ? SExtOp[PowDiff - 1] : ZExtOp[PowDiff - 1];
1885 return getRISCVInstructionCost(OpCodes: Op, VT: DstLT.second, CostKind);
1886 }
1887 case ISD::TRUNCATE:
1888 case ISD::FP_EXTEND:
1889 case ISD::FP_ROUND: {
1890 // Counts of narrow/widen instructions.
1891 unsigned SrcEltSize = SrcLT.second.getScalarSizeInBits();
1892 unsigned DstEltSize = DstLT.second.getScalarSizeInBits();
1893
1894 unsigned Op = (ISD == ISD::TRUNCATE) ? RISCV::VNSRL_WI
1895 : (ISD == ISD::FP_EXTEND) ? RISCV::VFWCVT_F_F_V
1896 : RISCV::VFNCVT_F_F_W;
1897 InstructionCost Cost = 0;
1898 for (; SrcEltSize != DstEltSize;) {
1899 MVT ElementMVT = (ISD == ISD::TRUNCATE)
1900 ? MVT::getIntegerVT(BitWidth: DstEltSize)
1901 : MVT::getFloatingPointVT(BitWidth: DstEltSize);
1902 MVT DstMVT = DstLT.second.changeVectorElementType(EltVT: ElementMVT);
1903 DstEltSize =
1904 (DstEltSize > SrcEltSize) ? DstEltSize >> 1 : DstEltSize << 1;
1905 Cost += getRISCVInstructionCost(OpCodes: Op, VT: DstMVT, CostKind);
1906 }
1907 return Cost;
1908 }
1909 case ISD::FP_TO_SINT:
1910 case ISD::FP_TO_UINT: {
1911 unsigned IsSigned = ISD == ISD::FP_TO_SINT;
1912 unsigned FCVT = IsSigned ? RISCV::VFCVT_RTZ_X_F_V : RISCV::VFCVT_RTZ_XU_F_V;
1913 unsigned FWCVT =
1914 IsSigned ? RISCV::VFWCVT_RTZ_X_F_V : RISCV::VFWCVT_RTZ_XU_F_V;
1915 unsigned FNCVT =
1916 IsSigned ? RISCV::VFNCVT_RTZ_X_F_W : RISCV::VFNCVT_RTZ_XU_F_W;
1917 unsigned SrcEltSize = Src->getScalarSizeInBits();
1918 unsigned DstEltSize = Dst->getScalarSizeInBits();
1919 InstructionCost Cost = 0;
1920 if ((SrcEltSize == 16) &&
1921 (!ST->hasVInstructionsF16() || ((DstEltSize / 2) > SrcEltSize))) {
1922 // If the target only supports zvfhmin or it is fp16-to-i64 conversion
1923 // pre-widening to f32 and then convert f32 to integer
1924 VectorType *VecF32Ty =
1925 VectorType::get(ElementType: Type::getFloatTy(C&: Dst->getContext()),
1926 EC: cast<VectorType>(Val: Dst)->getElementCount());
1927 std::pair<InstructionCost, MVT> VecF32LT =
1928 getTypeLegalizationCost(Ty: VecF32Ty);
1929 Cost +=
1930 VecF32LT.first * getRISCVInstructionCost(OpCodes: RISCV::VFWCVT_F_F_V,
1931 VT: VecF32LT.second, CostKind);
1932 Cost += getCastInstrCost(Opcode, Dst, Src: VecF32Ty, CCH, CostKind, I);
1933 return Cost;
1934 }
1935 if (DstEltSize == SrcEltSize)
1936 Cost += getRISCVInstructionCost(OpCodes: FCVT, VT: DstLT.second, CostKind);
1937 else if (DstEltSize > SrcEltSize)
1938 Cost += getRISCVInstructionCost(OpCodes: FWCVT, VT: DstLT.second, CostKind);
1939 else { // (SrcEltSize > DstEltSize)
1940 // First do a narrowing conversion to an integer half the size, then
1941 // truncate if needed.
1942 MVT ElementVT = MVT::getIntegerVT(BitWidth: SrcEltSize / 2);
1943 MVT VecVT = DstLT.second.changeVectorElementType(EltVT: ElementVT);
1944 Cost += getRISCVInstructionCost(OpCodes: FNCVT, VT: VecVT, CostKind);
1945 if ((SrcEltSize / 2) > DstEltSize) {
1946 Type *VecTy = EVT(VecVT).getTypeForEVT(Context&: Dst->getContext());
1947 Cost +=
1948 getCastInstrCost(Opcode: Instruction::Trunc, Dst, Src: VecTy, CCH, CostKind, I);
1949 }
1950 }
1951 return Cost;
1952 }
1953 case ISD::SINT_TO_FP:
1954 case ISD::UINT_TO_FP: {
1955 unsigned IsSigned = ISD == ISD::SINT_TO_FP;
1956 unsigned FCVT = IsSigned ? RISCV::VFCVT_F_X_V : RISCV::VFCVT_F_XU_V;
1957 unsigned FWCVT = IsSigned ? RISCV::VFWCVT_F_X_V : RISCV::VFWCVT_F_XU_V;
1958 unsigned FNCVT = IsSigned ? RISCV::VFNCVT_F_X_W : RISCV::VFNCVT_F_XU_W;
1959 unsigned SrcEltSize = Src->getScalarSizeInBits();
1960 unsigned DstEltSize = Dst->getScalarSizeInBits();
1961
1962 InstructionCost Cost = 0;
1963 if ((DstEltSize == 16) &&
1964 (!ST->hasVInstructionsF16() || ((SrcEltSize / 2) > DstEltSize))) {
1965 // If the target only supports zvfhmin or it is i64-to-fp16 conversion
1966 // it is converted to f32 and then converted to f16
1967 VectorType *VecF32Ty =
1968 VectorType::get(ElementType: Type::getFloatTy(C&: Dst->getContext()),
1969 EC: cast<VectorType>(Val: Dst)->getElementCount());
1970 std::pair<InstructionCost, MVT> VecF32LT =
1971 getTypeLegalizationCost(Ty: VecF32Ty);
1972 Cost += getCastInstrCost(Opcode, Dst: VecF32Ty, Src, CCH, CostKind, I);
1973 Cost += VecF32LT.first * getRISCVInstructionCost(OpCodes: RISCV::VFNCVT_F_F_W,
1974 VT: DstLT.second, CostKind);
1975 return Cost;
1976 }
1977
1978 if (DstEltSize == SrcEltSize)
1979 Cost += getRISCVInstructionCost(OpCodes: FCVT, VT: DstLT.second, CostKind);
1980 else if (DstEltSize > SrcEltSize) {
1981 if ((DstEltSize / 2) > SrcEltSize) {
1982 VectorType *VecTy =
1983 VectorType::get(ElementType: IntegerType::get(C&: Dst->getContext(), NumBits: DstEltSize / 2),
1984 EC: cast<VectorType>(Val: Dst)->getElementCount());
1985 unsigned Op = IsSigned ? Instruction::SExt : Instruction::ZExt;
1986 Cost += getCastInstrCost(Opcode: Op, Dst: VecTy, Src, CCH, CostKind, I);
1987 }
1988 Cost += getRISCVInstructionCost(OpCodes: FWCVT, VT: DstLT.second, CostKind);
1989 } else
1990 Cost += getRISCVInstructionCost(OpCodes: FNCVT, VT: DstLT.second, CostKind);
1991 return Cost;
1992 }
1993 }
1994 return BaseT::getCastInstrCost(Opcode, Dst, Src, CCH, CostKind, I);
1995}
1996
1997unsigned RISCVTTIImpl::getEstimatedVLFor(VectorType *Ty) const {
1998 if (isa<ScalableVectorType>(Val: Ty)) {
1999 const unsigned EltSize = DL.getTypeSizeInBits(Ty: Ty->getElementType());
2000 const unsigned MinSize = DL.getTypeSizeInBits(Ty).getKnownMinValue();
2001 const unsigned VectorBits = *getVScaleForTuning() * RISCV::RVVBitsPerBlock;
2002 return RISCVTargetLowering::computeVLMAX(VectorBits, EltSize, MinSize);
2003 }
2004 return cast<FixedVectorType>(Val: Ty)->getNumElements();
2005}
2006
2007InstructionCost
2008RISCVTTIImpl::getMinMaxReductionCost(Intrinsic::ID IID, VectorType *Ty,
2009 FastMathFlags FMF,
2010 TTI::TargetCostKind CostKind) const {
2011 if (isa<FixedVectorType>(Val: Ty) && !ST->useRVVForFixedLengthVectors())
2012 return BaseT::getMinMaxReductionCost(IID, Ty, FMF, CostKind);
2013
2014 // Skip if scalar size of Ty is bigger than ELEN.
2015 if (Ty->getScalarSizeInBits() > ST->getELen())
2016 return BaseT::getMinMaxReductionCost(IID, Ty, FMF, CostKind);
2017
2018 std::pair<InstructionCost, MVT> LT = getTypeLegalizationCost(Ty);
2019 if (Ty->getElementType()->isIntegerTy(Bitwidth: 1)) {
2020 // SelectionDAGBuilder does following transforms:
2021 // vector_reduce_{smin,umax}(<n x i1>) --> vector_reduce_or(<n x i1>)
2022 // vector_reduce_{smax,umin}(<n x i1>) --> vector_reduce_and(<n x i1>)
2023 if (IID == Intrinsic::umax || IID == Intrinsic::smin)
2024 return getArithmeticReductionCost(Opcode: Instruction::Or, Ty, FMF, CostKind);
2025 else
2026 return getArithmeticReductionCost(Opcode: Instruction::And, Ty, FMF, CostKind);
2027 }
2028
2029 if (IID == Intrinsic::maximum || IID == Intrinsic::minimum) {
2030 SmallVector<unsigned, 3> Opcodes;
2031 InstructionCost ExtraCost = 0;
2032 switch (IID) {
2033 case Intrinsic::maximum:
2034 if (FMF.noNaNs()) {
2035 Opcodes = {RISCV::VFREDMAX_VS, RISCV::VFMV_F_S};
2036 } else {
2037 Opcodes = {RISCV::VMFNE_VV, RISCV::VCPOP_M, RISCV::VFREDMAX_VS,
2038 RISCV::VFMV_F_S};
2039 // Cost of Canonical Nan + branch
2040 // lui a0, 523264
2041 // fmv.w.x fa0, a0
2042 Type *DstTy = Ty->getScalarType();
2043 const unsigned EltTyBits = DstTy->getScalarSizeInBits();
2044 Type *SrcTy = IntegerType::getIntNTy(C&: DstTy->getContext(), N: EltTyBits);
2045 ExtraCost = 1 +
2046 getCastInstrCost(Opcode: Instruction::UIToFP, Dst: DstTy, Src: SrcTy,
2047 CCH: TTI::CastContextHint::None, CostKind) +
2048 getCFInstrCost(Opcode: Instruction::CondBr, CostKind);
2049 }
2050 break;
2051
2052 case Intrinsic::minimum:
2053 if (FMF.noNaNs()) {
2054 Opcodes = {RISCV::VFREDMIN_VS, RISCV::VFMV_F_S};
2055 } else {
2056 Opcodes = {RISCV::VMFNE_VV, RISCV::VCPOP_M, RISCV::VFREDMIN_VS,
2057 RISCV::VFMV_F_S};
2058 // Cost of Canonical Nan + branch
2059 // lui a0, 523264
2060 // fmv.w.x fa0, a0
2061 Type *DstTy = Ty->getScalarType();
2062 const unsigned EltTyBits = DL.getTypeSizeInBits(Ty: DstTy);
2063 Type *SrcTy = IntegerType::getIntNTy(C&: DstTy->getContext(), N: EltTyBits);
2064 ExtraCost = 1 +
2065 getCastInstrCost(Opcode: Instruction::UIToFP, Dst: DstTy, Src: SrcTy,
2066 CCH: TTI::CastContextHint::None, CostKind) +
2067 getCFInstrCost(Opcode: Instruction::CondBr, CostKind);
2068 }
2069 break;
2070 }
2071 return ExtraCost + getRISCVInstructionCost(OpCodes: Opcodes, VT: LT.second, CostKind);
2072 }
2073
2074 // IR Reduction is composed by one rvv reduction instruction and vmv
2075 unsigned SplitOp;
2076 SmallVector<unsigned, 3> Opcodes;
2077 switch (IID) {
2078 default:
2079 llvm_unreachable("Unsupported intrinsic");
2080 case Intrinsic::smax:
2081 SplitOp = RISCV::VMAX_VV;
2082 Opcodes = {RISCV::VREDMAX_VS, RISCV::VMV_X_S};
2083 break;
2084 case Intrinsic::smin:
2085 SplitOp = RISCV::VMIN_VV;
2086 Opcodes = {RISCV::VREDMIN_VS, RISCV::VMV_X_S};
2087 break;
2088 case Intrinsic::umax:
2089 SplitOp = RISCV::VMAXU_VV;
2090 Opcodes = {RISCV::VREDMAXU_VS, RISCV::VMV_X_S};
2091 break;
2092 case Intrinsic::umin:
2093 SplitOp = RISCV::VMINU_VV;
2094 Opcodes = {RISCV::VREDMINU_VS, RISCV::VMV_X_S};
2095 break;
2096 case Intrinsic::maxnum:
2097 SplitOp = RISCV::VFMAX_VV;
2098 Opcodes = {RISCV::VFREDMAX_VS, RISCV::VFMV_F_S};
2099 break;
2100 case Intrinsic::minnum:
2101 SplitOp = RISCV::VFMIN_VV;
2102 Opcodes = {RISCV::VFREDMIN_VS, RISCV::VFMV_F_S};
2103 break;
2104 }
2105 // Add a cost for data larger than LMUL8
2106 InstructionCost SplitCost =
2107 (LT.first > 1) ? (LT.first - 1) *
2108 getRISCVInstructionCost(OpCodes: SplitOp, VT: LT.second, CostKind)
2109 : 0;
2110 return SplitCost + getRISCVInstructionCost(OpCodes: Opcodes, VT: LT.second, CostKind);
2111}
2112
2113InstructionCost
2114RISCVTTIImpl::getArithmeticReductionCost(unsigned Opcode, VectorType *Ty,
2115 std::optional<FastMathFlags> FMF,
2116 TTI::TargetCostKind CostKind) const {
2117 if (isa<FixedVectorType>(Val: Ty) && !ST->useRVVForFixedLengthVectors())
2118 return BaseT::getArithmeticReductionCost(Opcode, Ty, FMF, CostKind);
2119
2120 // Skip if scalar size of Ty is bigger than ELEN.
2121 if (Ty->getScalarSizeInBits() > ST->getELen())
2122 return BaseT::getArithmeticReductionCost(Opcode, Ty, FMF, CostKind);
2123
2124 int ISD = TLI->InstructionOpcodeToISD(Opcode);
2125 assert(ISD && "Invalid opcode");
2126
2127 if (ISD != ISD::ADD && ISD != ISD::OR && ISD != ISD::XOR && ISD != ISD::AND &&
2128 ISD != ISD::FADD)
2129 return BaseT::getArithmeticReductionCost(Opcode, Ty, FMF, CostKind);
2130
2131 std::pair<InstructionCost, MVT> LT = getTypeLegalizationCost(Ty);
2132 Type *ElementTy = Ty->getElementType();
2133 if (ElementTy->isIntegerTy(Bitwidth: 1)) {
2134 // Example sequences:
2135 // vfirst.m a0, v0
2136 // seqz a0, a0
2137 if (LT.second == MVT::v1i1)
2138 return getRISCVInstructionCost(OpCodes: RISCV::VFIRST_M, VT: LT.second, CostKind) +
2139 getCmpSelInstrCost(Opcode: Instruction::ICmp, ValTy: ElementTy, CondTy: ElementTy,
2140 VecPred: CmpInst::ICMP_EQ, CostKind);
2141
2142 if (ISD == ISD::AND) {
2143 // Example sequences:
2144 // vmand.mm v8, v9, v8 ; needed every time type is split
2145 // vmnot.m v8, v0 ; alias for vmnand
2146 // vcpop.m a0, v8
2147 // seqz a0, a0
2148
2149 // See the discussion: https://github.com/llvm/llvm-project/pull/119160
2150 // For LMUL <= 8, there is no splitting,
2151 // the sequences are vmnot, vcpop and seqz.
2152 // When LMUL > 8 and split = 1,
2153 // the sequences are vmnand, vcpop and seqz.
2154 // When LMUL > 8 and split > 1,
2155 // the sequences are (LT.first-2) * vmand, vmnand, vcpop and seqz.
2156 return ((LT.first > 2) ? (LT.first - 2) : 0) *
2157 getRISCVInstructionCost(OpCodes: RISCV::VMAND_MM, VT: LT.second, CostKind) +
2158 getRISCVInstructionCost(OpCodes: RISCV::VMNAND_MM, VT: LT.second, CostKind) +
2159 getRISCVInstructionCost(OpCodes: RISCV::VCPOP_M, VT: LT.second, CostKind) +
2160 getCmpSelInstrCost(Opcode: Instruction::ICmp, ValTy: ElementTy, CondTy: ElementTy,
2161 VecPred: CmpInst::ICMP_EQ, CostKind);
2162 } else if (ISD == ISD::XOR || ISD == ISD::ADD) {
2163 // Example sequences:
2164 // vsetvli a0, zero, e8, mf8, ta, ma
2165 // vmxor.mm v8, v0, v8 ; needed every time type is split
2166 // vcpop.m a0, v8
2167 // andi a0, a0, 1
2168 return (LT.first - 1) *
2169 getRISCVInstructionCost(OpCodes: RISCV::VMXOR_MM, VT: LT.second, CostKind) +
2170 getRISCVInstructionCost(OpCodes: RISCV::VCPOP_M, VT: LT.second, CostKind) + 1;
2171 } else {
2172 assert(ISD == ISD::OR);
2173 // Example sequences:
2174 // vsetvli a0, zero, e8, mf8, ta, ma
2175 // vmor.mm v8, v9, v8 ; needed every time type is split
2176 // vcpop.m a0, v0
2177 // snez a0, a0
2178 return (LT.first - 1) *
2179 getRISCVInstructionCost(OpCodes: RISCV::VMOR_MM, VT: LT.second, CostKind) +
2180 getRISCVInstructionCost(OpCodes: RISCV::VCPOP_M, VT: LT.second, CostKind) +
2181 getCmpSelInstrCost(Opcode: Instruction::ICmp, ValTy: ElementTy, CondTy: ElementTy,
2182 VecPred: CmpInst::ICMP_NE, CostKind);
2183 }
2184 }
2185
2186 // IR Reduction of or/and is composed by one vmv and one rvv reduction
2187 // instruction, and others is composed by two vmv and one rvv reduction
2188 // instruction
2189 unsigned SplitOp;
2190 SmallVector<unsigned, 3> Opcodes;
2191 switch (ISD) {
2192 case ISD::ADD:
2193 SplitOp = RISCV::VADD_VV;
2194 Opcodes = {RISCV::VMV_S_X, RISCV::VREDSUM_VS, RISCV::VMV_X_S};
2195 break;
2196 case ISD::OR:
2197 SplitOp = RISCV::VOR_VV;
2198 Opcodes = {RISCV::VREDOR_VS, RISCV::VMV_X_S};
2199 break;
2200 case ISD::XOR:
2201 SplitOp = RISCV::VXOR_VV;
2202 Opcodes = {RISCV::VMV_S_X, RISCV::VREDXOR_VS, RISCV::VMV_X_S};
2203 break;
2204 case ISD::AND:
2205 SplitOp = RISCV::VAND_VV;
2206 Opcodes = {RISCV::VREDAND_VS, RISCV::VMV_X_S};
2207 break;
2208 case ISD::FADD:
2209 // We can't promote f16/bf16 fadd reductions.
2210 if ((LT.second.getScalarType() == MVT::f16 && !ST->hasVInstructionsF16()) ||
2211 LT.second.getScalarType() == MVT::bf16)
2212 return BaseT::getArithmeticReductionCost(Opcode, Ty, FMF, CostKind);
2213 if (TTI::requiresOrderedReduction(FMF)) {
2214 Opcodes.push_back(Elt: RISCV::VFMV_S_F);
2215 for (unsigned i = 0; i < LT.first.getValue(); i++)
2216 Opcodes.push_back(Elt: RISCV::VFREDOSUM_VS);
2217 Opcodes.push_back(Elt: RISCV::VFMV_F_S);
2218 return getRISCVInstructionCost(OpCodes: Opcodes, VT: LT.second, CostKind);
2219 }
2220 SplitOp = RISCV::VFADD_VV;
2221 Opcodes = {RISCV::VFMV_S_F, RISCV::VFREDUSUM_VS, RISCV::VFMV_F_S};
2222 break;
2223 }
2224 // Add a cost for data larger than LMUL8
2225 InstructionCost SplitCost =
2226 (LT.first > 1) ? (LT.first - 1) *
2227 getRISCVInstructionCost(OpCodes: SplitOp, VT: LT.second, CostKind)
2228 : 0;
2229 return SplitCost + getRISCVInstructionCost(OpCodes: Opcodes, VT: LT.second, CostKind);
2230}
2231
2232InstructionCost RISCVTTIImpl::getExtendedReductionCost(
2233 unsigned Opcode, bool IsUnsigned, Type *ResTy, VectorType *ValTy,
2234 std::optional<FastMathFlags> FMF, TTI::TargetCostKind CostKind) const {
2235 if (isa<FixedVectorType>(Val: ValTy) && !ST->useRVVForFixedLengthVectors())
2236 return BaseT::getExtendedReductionCost(Opcode, IsUnsigned, ResTy, Ty: ValTy,
2237 FMF, CostKind);
2238
2239 // Skip if scalar size of ResTy is bigger than ELEN.
2240 if (ResTy->getScalarSizeInBits() > ST->getELen())
2241 return BaseT::getExtendedReductionCost(Opcode, IsUnsigned, ResTy, Ty: ValTy,
2242 FMF, CostKind);
2243
2244 if (Opcode != Instruction::Add && Opcode != Instruction::FAdd)
2245 return BaseT::getExtendedReductionCost(Opcode, IsUnsigned, ResTy, Ty: ValTy,
2246 FMF, CostKind);
2247
2248 std::pair<InstructionCost, MVT> LT = getTypeLegalizationCost(Ty: ValTy);
2249
2250 if (IsUnsigned && Opcode == Instruction::Add &&
2251 LT.second.isFixedLengthVector() && LT.second.getScalarType() == MVT::i1) {
2252 // Represent vector_reduce_add(ZExt(<n x i1>)) as
2253 // ZExtOrTrunc(ctpop(bitcast <n x i1> to in)).
2254 return LT.first *
2255 getRISCVInstructionCost(OpCodes: RISCV::VCPOP_M, VT: LT.second, CostKind);
2256 }
2257
2258 if (ResTy->getScalarSizeInBits() != 2 * LT.second.getScalarSizeInBits())
2259 return BaseT::getExtendedReductionCost(Opcode, IsUnsigned, ResTy, Ty: ValTy,
2260 FMF, CostKind);
2261
2262 return (LT.first - 1) +
2263 getArithmeticReductionCost(Opcode, Ty: ValTy, FMF, CostKind);
2264}
2265
2266InstructionCost
2267RISCVTTIImpl::getStoreImmCost(Type *Ty, TTI::OperandValueInfo OpInfo,
2268 TTI::TargetCostKind CostKind) const {
2269 assert(OpInfo.isConstant() && "non constant operand?");
2270 if (!isa<VectorType>(Val: Ty))
2271 // FIXME: We need to account for immediate materialization here, but doing
2272 // a decent job requires more knowledge about the immediate than we
2273 // currently have here.
2274 return 0;
2275
2276 if (OpInfo.isUniform())
2277 // vmv.v.i, vmv.v.x, or vfmv.v.f
2278 // We ignore the cost of the scalar constant materialization to be consistent
2279 // with how we treat scalar constants themselves just above.
2280 return 1;
2281
2282 return getConstantPoolLoadCost(Ty, CostKind);
2283}
2284
2285InstructionCost RISCVTTIImpl::getMemoryOpCost(unsigned Opcode, Type *Src,
2286 Align Alignment,
2287 unsigned AddressSpace,
2288 TTI::TargetCostKind CostKind,
2289 TTI::OperandValueInfo OpInfo,
2290 const Instruction *I) const {
2291 EVT VT = TLI->getValueType(DL, Ty: Src, AllowUnknown: true);
2292 // Type legalization can't handle structs
2293 if (VT == MVT::Other)
2294 return BaseT::getMemoryOpCost(Opcode, Src, Alignment, AddressSpace,
2295 CostKind, OpInfo, I);
2296
2297 InstructionCost Cost = 0;
2298 if (Opcode == Instruction::Store && OpInfo.isConstant())
2299 Cost += getStoreImmCost(Ty: Src, OpInfo, CostKind);
2300
2301 std::pair<InstructionCost, MVT> LT = getTypeLegalizationCost(Ty: Src);
2302
2303 InstructionCost BaseCost = [&]() {
2304 InstructionCost Cost = LT.first;
2305 if (CostKind != TTI::TCK_RecipThroughput)
2306 return Cost;
2307
2308 // Our actual lowering for the case where a wider legal type is available
2309 // uses the a VL predicated load on the wider type. This is reflected in
2310 // the result of getTypeLegalizationCost, but BasicTTI assumes the
2311 // widened cases are scalarized.
2312 const DataLayout &DL = this->getDataLayout();
2313 if (Src->isVectorTy() && LT.second.isVector() &&
2314 TypeSize::isKnownLT(LHS: DL.getTypeStoreSizeInBits(Ty: Src),
2315 RHS: LT.second.getSizeInBits()))
2316 return Cost;
2317
2318 return BaseT::getMemoryOpCost(Opcode, Src, Alignment, AddressSpace,
2319 CostKind, OpInfo, I);
2320 }();
2321
2322 // Assume memory ops cost scale with the number of vector registers
2323 // possible accessed by the instruction. Note that BasicTTI already
2324 // handles the LT.first term for us.
2325 if (ST->hasVInstructions() && LT.second.isVector() &&
2326 CostKind != TTI::TCK_CodeSize)
2327 BaseCost *= TLI->getLMULCost(VT: LT.second);
2328 return Cost + BaseCost;
2329}
2330
2331InstructionCost RISCVTTIImpl::getCmpSelInstrCost(
2332 unsigned Opcode, Type *ValTy, Type *CondTy, CmpInst::Predicate VecPred,
2333 TTI::TargetCostKind CostKind, TTI::OperandValueInfo Op1Info,
2334 TTI::OperandValueInfo Op2Info, const Instruction *I) const {
2335 if (CostKind != TTI::TCK_RecipThroughput)
2336 return BaseT::getCmpSelInstrCost(Opcode, ValTy, CondTy, VecPred, CostKind,
2337 Op1Info, Op2Info, I);
2338
2339 if (isa<FixedVectorType>(Val: ValTy) && !ST->useRVVForFixedLengthVectors())
2340 return BaseT::getCmpSelInstrCost(Opcode, ValTy, CondTy, VecPred, CostKind,
2341 Op1Info, Op2Info, I);
2342
2343 // Skip if scalar size of ValTy is bigger than ELEN.
2344 if (ValTy->isVectorTy() && ValTy->getScalarSizeInBits() > ST->getELen())
2345 return BaseT::getCmpSelInstrCost(Opcode, ValTy, CondTy, VecPred, CostKind,
2346 Op1Info, Op2Info, I);
2347
2348 auto GetConstantMatCost =
2349 [&](TTI::OperandValueInfo OpInfo) -> InstructionCost {
2350 if (OpInfo.isUniform())
2351 // We return 0 we currently ignore the cost of materializing scalar
2352 // constants in GPRs.
2353 return 0;
2354
2355 return getConstantPoolLoadCost(Ty: ValTy, CostKind);
2356 };
2357
2358 InstructionCost ConstantMatCost;
2359 if (Op1Info.isConstant())
2360 ConstantMatCost += GetConstantMatCost(Op1Info);
2361 if (Op2Info.isConstant())
2362 ConstantMatCost += GetConstantMatCost(Op2Info);
2363
2364 std::pair<InstructionCost, MVT> LT = getTypeLegalizationCost(Ty: ValTy);
2365 if (Opcode == Instruction::Select && ValTy->isVectorTy()) {
2366 if (CondTy->isVectorTy()) {
2367 if (ValTy->getScalarSizeInBits() == 1) {
2368 // vmandn.mm v8, v8, v9
2369 // vmand.mm v9, v0, v9
2370 // vmor.mm v0, v9, v8
2371 return ConstantMatCost +
2372 LT.first *
2373 getRISCVInstructionCost(
2374 OpCodes: {RISCV::VMANDN_MM, RISCV::VMAND_MM, RISCV::VMOR_MM},
2375 VT: LT.second, CostKind);
2376 }
2377 // vselect and max/min are supported natively.
2378 return ConstantMatCost +
2379 LT.first * getRISCVInstructionCost(OpCodes: RISCV::VMERGE_VVM, VT: LT.second,
2380 CostKind);
2381 }
2382
2383 if (ValTy->getScalarSizeInBits() == 1) {
2384 // vmv.v.x v9, a0
2385 // vmsne.vi v9, v9, 0
2386 // vmandn.mm v8, v8, v9
2387 // vmand.mm v9, v0, v9
2388 // vmor.mm v0, v9, v8
2389 MVT InterimVT = LT.second.changeVectorElementType(EltVT: MVT::i8);
2390 return ConstantMatCost +
2391 LT.first *
2392 getRISCVInstructionCost(OpCodes: {RISCV::VMV_V_X, RISCV::VMSNE_VI},
2393 VT: InterimVT, CostKind) +
2394 LT.first * getRISCVInstructionCost(
2395 OpCodes: {RISCV::VMANDN_MM, RISCV::VMAND_MM, RISCV::VMOR_MM},
2396 VT: LT.second, CostKind);
2397 }
2398
2399 // vmv.v.x v10, a0
2400 // vmsne.vi v0, v10, 0
2401 // vmerge.vvm v8, v9, v8, v0
2402 return ConstantMatCost +
2403 LT.first * getRISCVInstructionCost(
2404 OpCodes: {RISCV::VMV_V_X, RISCV::VMSNE_VI, RISCV::VMERGE_VVM},
2405 VT: LT.second, CostKind);
2406 }
2407
2408 if ((Opcode == Instruction::ICmp) && ValTy->isVectorTy() &&
2409 CmpInst::isIntPredicate(P: VecPred)) {
2410 // Use VMSLT_VV to represent VMSEQ, VMSNE, VMSLTU, VMSLEU, VMSLT, VMSLE
2411 // provided they incur the same cost across all implementations
2412 return ConstantMatCost + LT.first * getRISCVInstructionCost(OpCodes: RISCV::VMSLT_VV,
2413 VT: LT.second,
2414 CostKind);
2415 }
2416
2417 if ((Opcode == Instruction::FCmp) && ValTy->isVectorTy() &&
2418 CmpInst::isFPPredicate(P: VecPred)) {
2419
2420 // Use VMXOR_MM and VMXNOR_MM to generate all true/false mask
2421 if ((VecPred == CmpInst::FCMP_FALSE) || (VecPred == CmpInst::FCMP_TRUE))
2422 return ConstantMatCost +
2423 getRISCVInstructionCost(OpCodes: RISCV::VMXOR_MM, VT: LT.second, CostKind);
2424
2425 // If we do not support the input floating point vector type, use the base
2426 // one which will calculate as:
2427 // ScalarizeCost + Num * Cost for fixed vector,
2428 // InvalidCost for scalable vector.
2429 if ((ValTy->getScalarSizeInBits() == 16 && !ST->hasVInstructionsF16()) ||
2430 (ValTy->getScalarSizeInBits() == 32 && !ST->hasVInstructionsF32()) ||
2431 (ValTy->getScalarSizeInBits() == 64 && !ST->hasVInstructionsF64()))
2432 return BaseT::getCmpSelInstrCost(Opcode, ValTy, CondTy, VecPred, CostKind,
2433 Op1Info, Op2Info, I);
2434
2435 // Assuming vector fp compare and mask instructions are all the same cost
2436 // until a need arises to differentiate them.
2437 switch (VecPred) {
2438 case CmpInst::FCMP_ONE: // vmflt.vv + vmflt.vv + vmor.mm
2439 case CmpInst::FCMP_ORD: // vmfeq.vv + vmfeq.vv + vmand.mm
2440 case CmpInst::FCMP_UNO: // vmfne.vv + vmfne.vv + vmor.mm
2441 case CmpInst::FCMP_UEQ: // vmflt.vv + vmflt.vv + vmnor.mm
2442 return ConstantMatCost +
2443 LT.first * getRISCVInstructionCost(
2444 OpCodes: {RISCV::VMFLT_VV, RISCV::VMFLT_VV, RISCV::VMOR_MM},
2445 VT: LT.second, CostKind);
2446
2447 case CmpInst::FCMP_UGT: // vmfle.vv + vmnot.m
2448 case CmpInst::FCMP_UGE: // vmflt.vv + vmnot.m
2449 case CmpInst::FCMP_ULT: // vmfle.vv + vmnot.m
2450 case CmpInst::FCMP_ULE: // vmflt.vv + vmnot.m
2451 return ConstantMatCost +
2452 LT.first *
2453 getRISCVInstructionCost(OpCodes: {RISCV::VMFLT_VV, RISCV::VMNAND_MM},
2454 VT: LT.second, CostKind);
2455
2456 case CmpInst::FCMP_OEQ: // vmfeq.vv
2457 case CmpInst::FCMP_OGT: // vmflt.vv
2458 case CmpInst::FCMP_OGE: // vmfle.vv
2459 case CmpInst::FCMP_OLT: // vmflt.vv
2460 case CmpInst::FCMP_OLE: // vmfle.vv
2461 case CmpInst::FCMP_UNE: // vmfne.vv
2462 return ConstantMatCost +
2463 LT.first *
2464 getRISCVInstructionCost(OpCodes: RISCV::VMFLT_VV, VT: LT.second, CostKind);
2465 default:
2466 break;
2467 }
2468 }
2469
2470 // With ShortForwardBranchOpt or ConditionalMoveFusion, scalar icmp + select
2471 // instructions will lower to SELECT_CC and lower to PseudoCCMOVGPR which will
2472 // generate a conditional branch + mv. The cost of scalar (icmp + select) will
2473 // be (0 + select instr cost).
2474 if (ST->hasConditionalMoveFusion() && I && isa<ICmpInst>(Val: I) &&
2475 ValTy->isIntegerTy() && !I->user_empty()) {
2476 if (all_of(Range: I->users(), P: [&](const User *U) {
2477 return match(V: U, P: m_Select(C: m_Specific(V: I), L: m_Value(), R: m_Value())) &&
2478 U->getType()->isIntegerTy() &&
2479 !isa<ConstantData>(Val: U->getOperand(i: 1)) &&
2480 !isa<ConstantData>(Val: U->getOperand(i: 2));
2481 }))
2482 return 0;
2483 }
2484
2485 // TODO: Add cost for scalar type.
2486
2487 return BaseT::getCmpSelInstrCost(Opcode, ValTy, CondTy, VecPred, CostKind,
2488 Op1Info, Op2Info, I);
2489}
2490
2491InstructionCost RISCVTTIImpl::getCFInstrCost(unsigned Opcode,
2492 TTI::TargetCostKind CostKind,
2493 const Instruction *I) const {
2494 if (CostKind != TTI::TCK_RecipThroughput)
2495 return Opcode == Instruction::PHI ? 0 : 1;
2496 // Branches are assumed to be predicted.
2497 return 0;
2498}
2499
2500InstructionCost RISCVTTIImpl::getVectorInstrCost(
2501 unsigned Opcode, Type *Val, TTI::TargetCostKind CostKind, unsigned Index,
2502 const Value *Op0, const Value *Op1, TTI::VectorInstrContext VIC) const {
2503 assert(Val->isVectorTy() && "This must be a vector type");
2504
2505 // TODO: Add proper cost model for P extension fixed vectors (e.g., v4i16)
2506 // For now, skip all fixed vector cost analysis when P extension is available
2507 // to avoid crashes in getMinRVVVectorSizeInBits()
2508 if (ST->hasStdExtP() && isa<FixedVectorType>(Val)) {
2509 return 1; // Treat as single instruction cost for now
2510 }
2511
2512 if (Opcode != Instruction::ExtractElement &&
2513 Opcode != Instruction::InsertElement)
2514 return BaseT::getVectorInstrCost(Opcode, Val, CostKind, Index, Op0, Op1,
2515 VIC);
2516
2517 // Legalize the type.
2518 std::pair<InstructionCost, MVT> LT = getTypeLegalizationCost(Ty: Val);
2519
2520 // This type is legalized to a scalar type.
2521 if (!LT.second.isVector()) {
2522 auto *FixedVecTy = cast<FixedVectorType>(Val);
2523 // If Index is a known constant, cost is zero.
2524 if (Index != -1U)
2525 return 0;
2526 // Extract/InsertElement with non-constant index is very costly when
2527 // scalarized; estimate cost of loads/stores sequence via the stack:
2528 // ExtractElement cost: store vector to stack, load scalar;
2529 // InsertElement cost: store vector to stack, store scalar, load vector.
2530 Type *ElemTy = FixedVecTy->getElementType();
2531 auto NumElems = FixedVecTy->getNumElements();
2532 auto Align = DL.getPrefTypeAlign(Ty: ElemTy);
2533 InstructionCost LoadCost =
2534 getMemoryOpCost(Opcode: Instruction::Load, Src: ElemTy, Alignment: Align, AddressSpace: 0, CostKind);
2535 InstructionCost StoreCost =
2536 getMemoryOpCost(Opcode: Instruction::Store, Src: ElemTy, Alignment: Align, AddressSpace: 0, CostKind);
2537 return Opcode == Instruction::ExtractElement
2538 ? StoreCost * NumElems + LoadCost
2539 : (StoreCost + LoadCost) * NumElems + StoreCost;
2540 }
2541
2542 // For unsupported scalable vector.
2543 if (LT.second.isScalableVector() && !LT.first.isValid())
2544 return LT.first;
2545
2546 // Mask vector extract/insert is expanded via e8.
2547 if (Val->getScalarSizeInBits() == 1) {
2548 VectorType *WideTy =
2549 VectorType::get(ElementType: IntegerType::get(C&: Val->getContext(), NumBits: 8),
2550 EC: cast<VectorType>(Val)->getElementCount());
2551 if (Opcode == Instruction::ExtractElement) {
2552 InstructionCost ExtendCost
2553 = getCastInstrCost(Opcode: Instruction::ZExt, Dst: WideTy, Src: Val,
2554 CCH: TTI::CastContextHint::None, CostKind);
2555 InstructionCost ExtractCost
2556 = getVectorInstrCost(Opcode, Val: WideTy, CostKind, Index, Op0: nullptr, Op1: nullptr);
2557 return ExtendCost + ExtractCost;
2558 }
2559 InstructionCost ExtendCost
2560 = getCastInstrCost(Opcode: Instruction::ZExt, Dst: WideTy, Src: Val,
2561 CCH: TTI::CastContextHint::None, CostKind);
2562 InstructionCost InsertCost
2563 = getVectorInstrCost(Opcode, Val: WideTy, CostKind, Index, Op0: nullptr, Op1: nullptr);
2564 InstructionCost TruncCost
2565 = getCastInstrCost(Opcode: Instruction::Trunc, Dst: Val, Src: WideTy,
2566 CCH: TTI::CastContextHint::None, CostKind);
2567 return ExtendCost + InsertCost + TruncCost;
2568 }
2569
2570
2571 // In RVV, we could use vslidedown + vmv.x.s to extract element from vector
2572 // and vslideup + vmv.s.x to insert element to vector.
2573 unsigned BaseCost = 1;
2574 // When insertelement we should add the index with 1 as the input of vslideup.
2575 unsigned SlideCost = Opcode == Instruction::InsertElement ? 2 : 1;
2576
2577 if (Index != -1U) {
2578 // The type may be split. For fixed-width vectors we can normalize the
2579 // index to the new type.
2580 if (LT.second.isFixedLengthVector()) {
2581 unsigned Width = LT.second.getVectorNumElements();
2582 Index = Index % Width;
2583 }
2584
2585 // If exact VLEN is known, we will insert/extract into the appropriate
2586 // subvector with no additional subvector insert/extract cost.
2587 if (auto VLEN = ST->getRealVLen()) {
2588 unsigned EltSize = LT.second.getScalarSizeInBits();
2589 unsigned M1Max = *VLEN / EltSize;
2590 Index = Index % M1Max;
2591 }
2592
2593 if (Index == 0)
2594 // We can extract/insert the first element without vslidedown/vslideup.
2595 SlideCost = 0;
2596 else if (Opcode == Instruction::InsertElement)
2597 SlideCost = 1; // With a constant index, we do not need to use addi.
2598 }
2599
2600 // When the vector needs to split into multiple register groups and the index
2601 // exceeds single vector register group, we need to insert/extract the element
2602 // via stack.
2603 if (LT.first > 1 &&
2604 ((Index == -1U) || (Index >= LT.second.getVectorMinNumElements() &&
2605 LT.second.isScalableVector()))) {
2606 Type *ScalarType = Val->getScalarType();
2607 Align VecAlign = DL.getPrefTypeAlign(Ty: Val);
2608 Align SclAlign = DL.getPrefTypeAlign(Ty: ScalarType);
2609 // Extra addi for unknown index.
2610 InstructionCost IdxCost = Index == -1U ? 1 : 0;
2611
2612 // Store all split vectors into stack and load the target element.
2613 if (Opcode == Instruction::ExtractElement)
2614 return getMemoryOpCost(Opcode: Instruction::Store, Src: Val, Alignment: VecAlign, AddressSpace: 0, CostKind) +
2615 getMemoryOpCost(Opcode: Instruction::Load, Src: ScalarType, Alignment: SclAlign, AddressSpace: 0,
2616 CostKind) +
2617 IdxCost;
2618
2619 // Store all split vectors into stack and store the target element and load
2620 // vectors back.
2621 return getMemoryOpCost(Opcode: Instruction::Store, Src: Val, Alignment: VecAlign, AddressSpace: 0, CostKind) +
2622 getMemoryOpCost(Opcode: Instruction::Load, Src: Val, Alignment: VecAlign, AddressSpace: 0, CostKind) +
2623 getMemoryOpCost(Opcode: Instruction::Store, Src: ScalarType, Alignment: SclAlign, AddressSpace: 0,
2624 CostKind) +
2625 IdxCost;
2626 }
2627
2628 // Extract i64 in the target that has XLEN=32 need more instruction.
2629 if (Val->getScalarType()->isIntegerTy() &&
2630 ST->getXLen() < Val->getScalarSizeInBits()) {
2631 // For extractelement, we need the following instructions:
2632 // vsetivli zero, 1, e64, m1, ta, mu (not count)
2633 // vslidedown.vx v8, v8, a0
2634 // vmv.x.s a0, v8
2635 // li a1, 32
2636 // vsrl.vx v8, v8, a1
2637 // vmv.x.s a1, v8
2638
2639 // For insertelement, we need the following instructions:
2640 // vsetivli zero, 2, e32, m4, ta, mu (not count)
2641 // vmv.v.i v12, 0
2642 // vslide1up.vx v16, v12, a1
2643 // vslide1up.vx v12, v16, a0
2644 // addi a0, a2, 1
2645 // vsetvli zero, a0, e64, m4, tu, mu (not count)
2646 // vslideup.vx v8, v12, a2
2647
2648 // TODO: should we count these special vsetvlis?
2649 BaseCost = Opcode == Instruction::InsertElement ? 3 : 4;
2650 }
2651 return BaseCost + SlideCost;
2652}
2653
2654InstructionCost
2655RISCVTTIImpl::getIndexedVectorInstrCostFromEnd(unsigned Opcode, Type *Val,
2656 TTI::TargetCostKind CostKind,
2657 unsigned Index) const {
2658 if (isa<FixedVectorType>(Val))
2659 return BaseT::getIndexedVectorInstrCostFromEnd(Opcode, Val, CostKind,
2660 Index);
2661
2662 // TODO: This code replicates what LoopVectorize.cpp used to do when asking
2663 // for the cost of extracting the last lane of a scalable vector. It probably
2664 // needs a more accurate cost.
2665 ElementCount EC = cast<VectorType>(Val)->getElementCount();
2666 assert(Index < EC.getKnownMinValue() && "Unexpected reverse index");
2667 return getVectorInstrCost(Opcode, Val, CostKind,
2668 Index: EC.getKnownMinValue() - 1 - Index, Op0: nullptr,
2669 Op1: nullptr);
2670}
2671
2672InstructionCost RISCVTTIImpl::getArithmeticInstrCost(
2673 unsigned Opcode, Type *Ty, TTI::TargetCostKind CostKind,
2674 TTI::OperandValueInfo Op1Info, TTI::OperandValueInfo Op2Info,
2675 ArrayRef<const Value *> Args, const Instruction *CxtI) const {
2676
2677 // TODO: Handle more cost kinds.
2678 if (CostKind != TTI::TCK_RecipThroughput)
2679 return BaseT::getArithmeticInstrCost(Opcode, Ty, CostKind, Opd1Info: Op1Info, Opd2Info: Op2Info,
2680 Args, CxtI);
2681
2682 if (isa<FixedVectorType>(Val: Ty) && !ST->useRVVForFixedLengthVectors())
2683 return BaseT::getArithmeticInstrCost(Opcode, Ty, CostKind, Opd1Info: Op1Info, Opd2Info: Op2Info,
2684 Args, CxtI);
2685
2686 // Skip if scalar size of Ty is bigger than ELEN.
2687 if (isa<VectorType>(Val: Ty) && Ty->getScalarSizeInBits() > ST->getELen())
2688 return BaseT::getArithmeticInstrCost(Opcode, Ty, CostKind, Opd1Info: Op1Info, Opd2Info: Op2Info,
2689 Args, CxtI);
2690
2691 // Legalize the type.
2692 std::pair<InstructionCost, MVT> LT = getTypeLegalizationCost(Ty);
2693 unsigned ISDOpcode = TLI->InstructionOpcodeToISD(Opcode);
2694
2695 // TODO: Handle scalar type.
2696 if (!LT.second.isVector()) {
2697 static const CostTblEntry DivTbl[]{
2698 {.ISD: ISD::UDIV, .Type: MVT::i32, .Cost: TTI::TCC_Expensive},
2699 {.ISD: ISD::UDIV, .Type: MVT::i64, .Cost: TTI::TCC_Expensive},
2700 {.ISD: ISD::SDIV, .Type: MVT::i32, .Cost: TTI::TCC_Expensive},
2701 {.ISD: ISD::SDIV, .Type: MVT::i64, .Cost: TTI::TCC_Expensive},
2702 {.ISD: ISD::UREM, .Type: MVT::i32, .Cost: TTI::TCC_Expensive},
2703 {.ISD: ISD::UREM, .Type: MVT::i64, .Cost: TTI::TCC_Expensive},
2704 {.ISD: ISD::SREM, .Type: MVT::i32, .Cost: TTI::TCC_Expensive},
2705 {.ISD: ISD::SREM, .Type: MVT::i64, .Cost: TTI::TCC_Expensive}};
2706 if (TLI->isOperationLegalOrPromote(Op: ISDOpcode, VT: LT.second))
2707 if (const auto *Entry = CostTableLookup(Table: DivTbl, ISD: ISDOpcode, Ty: LT.second))
2708 return Entry->Cost * LT.first;
2709
2710 return BaseT::getArithmeticInstrCost(Opcode, Ty, CostKind, Opd1Info: Op1Info, Opd2Info: Op2Info,
2711 Args, CxtI);
2712 }
2713
2714 // f16 with zvfhmin and bf16 will be promoted to f32.
2715 // FIXME: nxv32[b]f16 will be custom lowered and split.
2716 InstructionCost CastCost = 0;
2717 if ((LT.second.getVectorElementType() == MVT::f16 ||
2718 LT.second.getVectorElementType() == MVT::bf16) &&
2719 TLI->getOperationAction(Op: ISDOpcode, VT: LT.second) ==
2720 TargetLoweringBase::LegalizeAction::Promote) {
2721 MVT PromotedVT = TLI->getTypeToPromoteTo(Op: ISDOpcode, VT: LT.second);
2722 Type *PromotedTy = EVT(PromotedVT).getTypeForEVT(Context&: Ty->getContext());
2723 Type *LegalTy = EVT(LT.second).getTypeForEVT(Context&: Ty->getContext());
2724 // Add cost of extending arguments
2725 CastCost += LT.first * Args.size() *
2726 getCastInstrCost(Opcode: Instruction::FPExt, Dst: PromotedTy, Src: LegalTy,
2727 CCH: TTI::CastContextHint::None, CostKind);
2728 // Add cost of truncating result
2729 CastCost +=
2730 LT.first * getCastInstrCost(Opcode: Instruction::FPTrunc, Dst: LegalTy, Src: PromotedTy,
2731 CCH: TTI::CastContextHint::None, CostKind);
2732 // Compute cost of op in promoted type
2733 LT.second = PromotedVT;
2734 }
2735
2736 auto getConstantMatCost =
2737 [&](unsigned Operand, TTI::OperandValueInfo OpInfo) -> InstructionCost {
2738 if (OpInfo.isUniform() && canSplatOperand(Opcode, Operand))
2739 // Two sub-cases:
2740 // * Has a 5 bit immediate operand which can be splatted.
2741 // * Has a larger immediate which must be materialized in scalar register
2742 // We return 0 for both as we currently ignore the cost of materializing
2743 // scalar constants in GPRs.
2744 return 0;
2745
2746 return getConstantPoolLoadCost(Ty, CostKind);
2747 };
2748
2749 // Add the cost of materializing any constant vectors required.
2750 InstructionCost ConstantMatCost = 0;
2751 if (Op1Info.isConstant())
2752 ConstantMatCost += getConstantMatCost(0, Op1Info);
2753 if (Op2Info.isConstant())
2754 ConstantMatCost += getConstantMatCost(1, Op2Info);
2755
2756 unsigned Op;
2757 switch (ISDOpcode) {
2758 case ISD::ADD:
2759 case ISD::SUB:
2760 Op = RISCV::VADD_VV;
2761 break;
2762 case ISD::SHL:
2763 case ISD::SRL:
2764 case ISD::SRA:
2765 Op = RISCV::VSLL_VV;
2766 break;
2767 case ISD::AND:
2768 case ISD::OR:
2769 case ISD::XOR:
2770 Op = (Ty->getScalarSizeInBits() == 1) ? RISCV::VMAND_MM : RISCV::VAND_VV;
2771 break;
2772 case ISD::MUL:
2773 case ISD::MULHS:
2774 case ISD::MULHU:
2775 Op = RISCV::VMUL_VV;
2776 break;
2777 case ISD::SDIV:
2778 case ISD::UDIV:
2779 Op = RISCV::VDIV_VV;
2780 break;
2781 case ISD::SREM:
2782 case ISD::UREM:
2783 Op = RISCV::VREM_VV;
2784 break;
2785 case ISD::FADD:
2786 case ISD::FSUB:
2787 Op = RISCV::VFADD_VV;
2788 break;
2789 case ISD::FMUL:
2790 Op = RISCV::VFMUL_VV;
2791 break;
2792 case ISD::FDIV:
2793 Op = RISCV::VFDIV_VV;
2794 break;
2795 case ISD::FNEG:
2796 Op = RISCV::VFSGNJN_VV;
2797 break;
2798 default:
2799 // Assuming all other instructions have the same cost until a need arises to
2800 // differentiate them.
2801 return CastCost + ConstantMatCost +
2802 BaseT::getArithmeticInstrCost(Opcode, Ty, CostKind, Opd1Info: Op1Info, Opd2Info: Op2Info,
2803 Args, CxtI);
2804 }
2805
2806 InstructionCost InstrCost = getRISCVInstructionCost(OpCodes: Op, VT: LT.second, CostKind);
2807 // We use BasicTTIImpl to calculate scalar costs, which assumes floating point
2808 // ops are twice as expensive as integer ops. Do the same for vectors so
2809 // scalar floating point ops aren't cheaper than their vector equivalents.
2810 if (Ty->isFPOrFPVectorTy())
2811 InstrCost *= 2;
2812 return CastCost + ConstantMatCost + LT.first * InstrCost;
2813}
2814
2815// TODO: Deduplicate from TargetTransformInfoImplCRTPBase.
2816InstructionCost RISCVTTIImpl::getPointersChainCost(
2817 ArrayRef<const Value *> Ptrs, const Value *Base,
2818 const TTI::PointersChainInfo &Info, Type *AccessTy,
2819 TTI::TargetCostKind CostKind) const {
2820 InstructionCost Cost = TTI::TCC_Free;
2821 // In the basic model we take into account GEP instructions only
2822 // (although here can come alloca instruction, a value, constants and/or
2823 // constant expressions, PHIs, bitcasts ... whatever allowed to be used as a
2824 // pointer). Typically, if Base is a not a GEP-instruction and all the
2825 // pointers are relative to the same base address, all the rest are
2826 // either GEP instructions, PHIs, bitcasts or constants. When we have same
2827 // base, we just calculate cost of each non-Base GEP as an ADD operation if
2828 // any their index is a non-const.
2829 // If no known dependencies between the pointers cost is calculated as a sum
2830 // of costs of GEP instructions.
2831 for (auto [I, V] : enumerate(First&: Ptrs)) {
2832 const auto *GEP = dyn_cast<GetElementPtrInst>(Val: V);
2833 if (!GEP)
2834 continue;
2835 if (Info.isSameBase() && V != Base) {
2836 if (GEP->hasAllConstantIndices())
2837 continue;
2838 // If the chain is unit-stride and BaseReg + stride*i is a legal
2839 // addressing mode, then presume the base GEP is sitting around in a
2840 // register somewhere and check if we can fold the offset relative to
2841 // it.
2842 unsigned Stride = DL.getTypeStoreSize(Ty: AccessTy);
2843 if (Info.isUnitStride() &&
2844 isLegalAddressingMode(Ty: AccessTy,
2845 /* BaseGV */ nullptr,
2846 /* BaseOffset */ Stride * I,
2847 /* HasBaseReg */ true,
2848 /* Scale */ 0,
2849 AddrSpace: GEP->getType()->getPointerAddressSpace()))
2850 continue;
2851 Cost += getArithmeticInstrCost(Opcode: Instruction::Add, Ty: GEP->getType(), CostKind,
2852 Op1Info: {.Kind: TTI::OK_AnyValue, .Properties: TTI::OP_None},
2853 Op2Info: {.Kind: TTI::OK_AnyValue, .Properties: TTI::OP_None}, Args: {});
2854 } else {
2855 SmallVector<const Value *> Indices(GEP->indices());
2856 Cost += getGEPCost(PointeeType: GEP->getSourceElementType(), Ptr: GEP->getPointerOperand(),
2857 Operands: Indices, AccessType: AccessTy, CostKind);
2858 }
2859 }
2860 return Cost;
2861}
2862
2863void RISCVTTIImpl::getUnrollingPreferences(
2864 Loop *L, ScalarEvolution &SE, TTI::UnrollingPreferences &UP,
2865 OptimizationRemarkEmitter *ORE) const {
2866 // TODO: More tuning on benchmarks and metrics with changes as needed
2867 // would apply to all settings below to enable performance.
2868
2869
2870 if (ST->enableDefaultUnroll())
2871 return BasicTTIImplBase::getUnrollingPreferences(L, SE, UP, ORE);
2872
2873 // Enable Upper bound unrolling universally, not dependent upon the conditions
2874 // below.
2875 UP.UpperBound = true;
2876
2877 // Disable loop unrolling for Oz and Os.
2878 UP.OptSizeThreshold = 0;
2879 UP.PartialOptSizeThreshold = 0;
2880 if (L->getHeader()->getParent()->hasOptSize())
2881 return;
2882
2883 SmallVector<BasicBlock *, 4> ExitingBlocks;
2884 L->getExitingBlocks(ExitingBlocks);
2885 LLVM_DEBUG(dbgs() << "Loop has:\n"
2886 << "Blocks: " << L->getNumBlocks() << "\n"
2887 << "Exit blocks: " << ExitingBlocks.size() << "\n");
2888
2889 // Only allow another exit other than the latch. This acts as an early exit
2890 // as it mirrors the profitability calculation of the runtime unroller.
2891 if (ExitingBlocks.size() > 2)
2892 return;
2893
2894 // Limit the CFG of the loop body for targets with a branch predictor.
2895 // Allowing 4 blocks permits if-then-else diamonds in the body.
2896 if (L->getNumBlocks() > 4)
2897 return;
2898
2899 // Scan the loop: don't unroll loops with calls as this could prevent
2900 // inlining. Don't unroll auto-vectorized loops either, though do allow
2901 // unrolling of the scalar remainder.
2902 bool IsVectorized = getBooleanLoopAttribute(TheLoop: L, Name: "llvm.loop.isvectorized");
2903 InstructionCost Cost = 0;
2904 for (auto *BB : L->getBlocks()) {
2905 for (auto &I : *BB) {
2906 // Both auto-vectorized loops and the scalar remainder have the
2907 // isvectorized attribute, so differentiate between them by the presence
2908 // of vector instructions.
2909 if (IsVectorized && (I.getType()->isVectorTy() ||
2910 llvm::any_of(Range: I.operand_values(), P: [](Value *V) {
2911 return V->getType()->isVectorTy();
2912 })))
2913 return;
2914
2915 if (isa<CallInst>(Val: I) || isa<InvokeInst>(Val: I)) {
2916 if (const Function *F = cast<CallBase>(Val&: I).getCalledFunction()) {
2917 if (!isLoweredToCall(F))
2918 continue;
2919 }
2920 return;
2921 }
2922
2923 SmallVector<const Value *> Operands(I.operand_values());
2924 Cost += getInstructionCost(U: &I, Operands,
2925 CostKind: TargetTransformInfo::TCK_SizeAndLatency);
2926 }
2927 }
2928
2929 LLVM_DEBUG(dbgs() << "Cost of loop: " << Cost << "\n");
2930
2931 UP.Partial = true;
2932 UP.Runtime = true;
2933 UP.UnrollRemainder = true;
2934 UP.UnrollAndJam = true;
2935
2936 // Force unrolling small loops can be very useful because of the branch
2937 // taken cost of the backedge.
2938 if (Cost < 12)
2939 UP.Force = true;
2940}
2941
2942void RISCVTTIImpl::getPeelingPreferences(Loop *L, ScalarEvolution &SE,
2943 TTI::PeelingPreferences &PP) const {
2944 BaseT::getPeelingPreferences(L, SE, PP);
2945}
2946
2947bool RISCVTTIImpl::getTgtMemIntrinsic(IntrinsicInst *Inst,
2948 MemIntrinsicInfo &Info) const {
2949 const DataLayout &DL = getDataLayout();
2950 Intrinsic::ID IID = Inst->getIntrinsicID();
2951 LLVMContext &C = Inst->getContext();
2952 bool HasMask = false;
2953
2954 auto getSegNum = [](const IntrinsicInst *II, unsigned PtrOperandNo,
2955 bool IsWrite) -> int64_t {
2956 if (auto *TarExtTy =
2957 dyn_cast<TargetExtType>(Val: II->getArgOperand(i: 0)->getType()))
2958 return TarExtTy->getIntParameter(i: 0);
2959
2960 return 1;
2961 };
2962
2963 switch (IID) {
2964 case Intrinsic::riscv_vle_mask:
2965 case Intrinsic::riscv_vse_mask:
2966 case Intrinsic::riscv_vlseg2_mask:
2967 case Intrinsic::riscv_vlseg3_mask:
2968 case Intrinsic::riscv_vlseg4_mask:
2969 case Intrinsic::riscv_vlseg5_mask:
2970 case Intrinsic::riscv_vlseg6_mask:
2971 case Intrinsic::riscv_vlseg7_mask:
2972 case Intrinsic::riscv_vlseg8_mask:
2973 case Intrinsic::riscv_vsseg2_mask:
2974 case Intrinsic::riscv_vsseg3_mask:
2975 case Intrinsic::riscv_vsseg4_mask:
2976 case Intrinsic::riscv_vsseg5_mask:
2977 case Intrinsic::riscv_vsseg6_mask:
2978 case Intrinsic::riscv_vsseg7_mask:
2979 case Intrinsic::riscv_vsseg8_mask:
2980 HasMask = true;
2981 [[fallthrough]];
2982 case Intrinsic::riscv_vle:
2983 case Intrinsic::riscv_vse:
2984 case Intrinsic::riscv_vlseg2:
2985 case Intrinsic::riscv_vlseg3:
2986 case Intrinsic::riscv_vlseg4:
2987 case Intrinsic::riscv_vlseg5:
2988 case Intrinsic::riscv_vlseg6:
2989 case Intrinsic::riscv_vlseg7:
2990 case Intrinsic::riscv_vlseg8:
2991 case Intrinsic::riscv_vsseg2:
2992 case Intrinsic::riscv_vsseg3:
2993 case Intrinsic::riscv_vsseg4:
2994 case Intrinsic::riscv_vsseg5:
2995 case Intrinsic::riscv_vsseg6:
2996 case Intrinsic::riscv_vsseg7:
2997 case Intrinsic::riscv_vsseg8: {
2998 // Intrinsic interface:
2999 // riscv_vle(merge, ptr, vl)
3000 // riscv_vle_mask(merge, ptr, mask, vl, policy)
3001 // riscv_vse(val, ptr, vl)
3002 // riscv_vse_mask(val, ptr, mask, vl, policy)
3003 // riscv_vlseg#(merge, ptr, vl, sew)
3004 // riscv_vlseg#_mask(merge, ptr, mask, vl, policy, sew)
3005 // riscv_vsseg#(val, ptr, vl, sew)
3006 // riscv_vsseg#_mask(val, ptr, mask, vl, sew)
3007 bool IsWrite = Inst->getType()->isVoidTy();
3008 Type *Ty = IsWrite ? Inst->getArgOperand(i: 0)->getType() : Inst->getType();
3009 // The results of segment loads are TargetExtType.
3010 if (auto *TarExtTy = dyn_cast<TargetExtType>(Val: Ty)) {
3011 unsigned SEW =
3012 1 << cast<ConstantInt>(Val: Inst->getArgOperand(i: Inst->arg_size() - 1))
3013 ->getZExtValue();
3014 Ty = TarExtTy->getTypeParameter(i: 0U);
3015 Ty = ScalableVectorType::get(
3016 ElementType: IntegerType::get(C, NumBits: SEW),
3017 MinNumElts: cast<ScalableVectorType>(Val: Ty)->getMinNumElements() * 8 / SEW);
3018 }
3019 const auto *RVVIInfo = RISCVVIntrinsicsTable::getRISCVVIntrinsicInfo(IntrinsicID: IID);
3020 unsigned VLIndex = RVVIInfo->VLOperand;
3021 unsigned PtrOperandNo = VLIndex - 1 - HasMask;
3022 MaybeAlign Alignment =
3023 Inst->getArgOperand(i: PtrOperandNo)->getPointerAlignment(DL);
3024 Type *MaskType = Ty->getWithNewType(EltTy: Type::getInt1Ty(C));
3025 Value *Mask = ConstantInt::getTrue(Ty: MaskType);
3026 if (HasMask)
3027 Mask = Inst->getArgOperand(i: VLIndex - 1);
3028 Value *EVL = Inst->getArgOperand(i: VLIndex);
3029 unsigned SegNum = getSegNum(Inst, PtrOperandNo, IsWrite);
3030 // RVV uses contiguous elements as a segment.
3031 if (SegNum > 1) {
3032 unsigned ElemSize = Ty->getScalarSizeInBits();
3033 auto *SegTy = IntegerType::get(C, NumBits: ElemSize * SegNum);
3034 Ty = VectorType::get(ElementType: SegTy, Other: cast<VectorType>(Val: Ty));
3035 }
3036 Info.InterestingOperands.emplace_back(Args&: Inst, Args&: PtrOperandNo, Args&: IsWrite, Args&: Ty,
3037 Args&: Alignment, Args&: Mask, Args&: EVL);
3038 return true;
3039 }
3040 case Intrinsic::riscv_vlse_mask:
3041 case Intrinsic::riscv_vsse_mask:
3042 case Intrinsic::riscv_vlsseg2_mask:
3043 case Intrinsic::riscv_vlsseg3_mask:
3044 case Intrinsic::riscv_vlsseg4_mask:
3045 case Intrinsic::riscv_vlsseg5_mask:
3046 case Intrinsic::riscv_vlsseg6_mask:
3047 case Intrinsic::riscv_vlsseg7_mask:
3048 case Intrinsic::riscv_vlsseg8_mask:
3049 case Intrinsic::riscv_vssseg2_mask:
3050 case Intrinsic::riscv_vssseg3_mask:
3051 case Intrinsic::riscv_vssseg4_mask:
3052 case Intrinsic::riscv_vssseg5_mask:
3053 case Intrinsic::riscv_vssseg6_mask:
3054 case Intrinsic::riscv_vssseg7_mask:
3055 case Intrinsic::riscv_vssseg8_mask:
3056 HasMask = true;
3057 [[fallthrough]];
3058 case Intrinsic::riscv_vlse:
3059 case Intrinsic::riscv_vsse:
3060 case Intrinsic::riscv_vlsseg2:
3061 case Intrinsic::riscv_vlsseg3:
3062 case Intrinsic::riscv_vlsseg4:
3063 case Intrinsic::riscv_vlsseg5:
3064 case Intrinsic::riscv_vlsseg6:
3065 case Intrinsic::riscv_vlsseg7:
3066 case Intrinsic::riscv_vlsseg8:
3067 case Intrinsic::riscv_vssseg2:
3068 case Intrinsic::riscv_vssseg3:
3069 case Intrinsic::riscv_vssseg4:
3070 case Intrinsic::riscv_vssseg5:
3071 case Intrinsic::riscv_vssseg6:
3072 case Intrinsic::riscv_vssseg7:
3073 case Intrinsic::riscv_vssseg8: {
3074 // Intrinsic interface:
3075 // riscv_vlse(merge, ptr, stride, vl)
3076 // riscv_vlse_mask(merge, ptr, stride, mask, vl, policy)
3077 // riscv_vsse(val, ptr, stride, vl)
3078 // riscv_vsse_mask(val, ptr, stride, mask, vl, policy)
3079 // riscv_vlsseg#(merge, ptr, offset, vl, sew)
3080 // riscv_vlsseg#_mask(merge, ptr, offset, mask, vl, policy, sew)
3081 // riscv_vssseg#(val, ptr, offset, vl, sew)
3082 // riscv_vssseg#_mask(val, ptr, offset, mask, vl, sew)
3083 bool IsWrite = Inst->getType()->isVoidTy();
3084 Type *Ty = IsWrite ? Inst->getArgOperand(i: 0)->getType() : Inst->getType();
3085 // The results of segment loads are TargetExtType.
3086 if (auto *TarExtTy = dyn_cast<TargetExtType>(Val: Ty)) {
3087 unsigned SEW =
3088 1 << cast<ConstantInt>(Val: Inst->getArgOperand(i: Inst->arg_size() - 1))
3089 ->getZExtValue();
3090 Ty = TarExtTy->getTypeParameter(i: 0U);
3091 Ty = ScalableVectorType::get(
3092 ElementType: IntegerType::get(C, NumBits: SEW),
3093 MinNumElts: cast<ScalableVectorType>(Val: Ty)->getMinNumElements() * 8 / SEW);
3094 }
3095 const auto *RVVIInfo = RISCVVIntrinsicsTable::getRISCVVIntrinsicInfo(IntrinsicID: IID);
3096 unsigned VLIndex = RVVIInfo->VLOperand;
3097 unsigned PtrOperandNo = VLIndex - 2 - HasMask;
3098 MaybeAlign Alignment =
3099 Inst->getArgOperand(i: PtrOperandNo)->getPointerAlignment(DL);
3100
3101 Value *Stride = Inst->getArgOperand(i: PtrOperandNo + 1);
3102 // Use the pointer alignment as the element alignment if the stride is a
3103 // multiple of the pointer alignment. Otherwise, the element alignment
3104 // should be the greatest common divisor of pointer alignment and stride.
3105 // For simplicity, just consider unalignment for elements.
3106 unsigned PointerAlign = Alignment.valueOrOne().value();
3107 if (!isa<ConstantInt>(Val: Stride) ||
3108 cast<ConstantInt>(Val: Stride)->getZExtValue() % PointerAlign != 0)
3109 Alignment = Align(1);
3110
3111 Type *MaskType = Ty->getWithNewType(EltTy: Type::getInt1Ty(C));
3112 Value *Mask = ConstantInt::getTrue(Ty: MaskType);
3113 if (HasMask)
3114 Mask = Inst->getArgOperand(i: VLIndex - 1);
3115 Value *EVL = Inst->getArgOperand(i: VLIndex);
3116 unsigned SegNum = getSegNum(Inst, PtrOperandNo, IsWrite);
3117 // RVV uses contiguous elements as a segment.
3118 if (SegNum > 1) {
3119 unsigned ElemSize = Ty->getScalarSizeInBits();
3120 auto *SegTy = IntegerType::get(C, NumBits: ElemSize * SegNum);
3121 Ty = VectorType::get(ElementType: SegTy, Other: cast<VectorType>(Val: Ty));
3122 }
3123 Info.InterestingOperands.emplace_back(Args&: Inst, Args&: PtrOperandNo, Args&: IsWrite, Args&: Ty,
3124 Args&: Alignment, Args&: Mask, Args&: EVL, Args&: Stride);
3125 return true;
3126 }
3127 case Intrinsic::riscv_vloxei_mask:
3128 case Intrinsic::riscv_vluxei_mask:
3129 case Intrinsic::riscv_vsoxei_mask:
3130 case Intrinsic::riscv_vsuxei_mask:
3131 case Intrinsic::riscv_vloxseg2_mask:
3132 case Intrinsic::riscv_vloxseg3_mask:
3133 case Intrinsic::riscv_vloxseg4_mask:
3134 case Intrinsic::riscv_vloxseg5_mask:
3135 case Intrinsic::riscv_vloxseg6_mask:
3136 case Intrinsic::riscv_vloxseg7_mask:
3137 case Intrinsic::riscv_vloxseg8_mask:
3138 case Intrinsic::riscv_vluxseg2_mask:
3139 case Intrinsic::riscv_vluxseg3_mask:
3140 case Intrinsic::riscv_vluxseg4_mask:
3141 case Intrinsic::riscv_vluxseg5_mask:
3142 case Intrinsic::riscv_vluxseg6_mask:
3143 case Intrinsic::riscv_vluxseg7_mask:
3144 case Intrinsic::riscv_vluxseg8_mask:
3145 case Intrinsic::riscv_vsoxseg2_mask:
3146 case Intrinsic::riscv_vsoxseg3_mask:
3147 case Intrinsic::riscv_vsoxseg4_mask:
3148 case Intrinsic::riscv_vsoxseg5_mask:
3149 case Intrinsic::riscv_vsoxseg6_mask:
3150 case Intrinsic::riscv_vsoxseg7_mask:
3151 case Intrinsic::riscv_vsoxseg8_mask:
3152 case Intrinsic::riscv_vsuxseg2_mask:
3153 case Intrinsic::riscv_vsuxseg3_mask:
3154 case Intrinsic::riscv_vsuxseg4_mask:
3155 case Intrinsic::riscv_vsuxseg5_mask:
3156 case Intrinsic::riscv_vsuxseg6_mask:
3157 case Intrinsic::riscv_vsuxseg7_mask:
3158 case Intrinsic::riscv_vsuxseg8_mask:
3159 HasMask = true;
3160 [[fallthrough]];
3161 case Intrinsic::riscv_vloxei:
3162 case Intrinsic::riscv_vluxei:
3163 case Intrinsic::riscv_vsoxei:
3164 case Intrinsic::riscv_vsuxei:
3165 case Intrinsic::riscv_vloxseg2:
3166 case Intrinsic::riscv_vloxseg3:
3167 case Intrinsic::riscv_vloxseg4:
3168 case Intrinsic::riscv_vloxseg5:
3169 case Intrinsic::riscv_vloxseg6:
3170 case Intrinsic::riscv_vloxseg7:
3171 case Intrinsic::riscv_vloxseg8:
3172 case Intrinsic::riscv_vluxseg2:
3173 case Intrinsic::riscv_vluxseg3:
3174 case Intrinsic::riscv_vluxseg4:
3175 case Intrinsic::riscv_vluxseg5:
3176 case Intrinsic::riscv_vluxseg6:
3177 case Intrinsic::riscv_vluxseg7:
3178 case Intrinsic::riscv_vluxseg8:
3179 case Intrinsic::riscv_vsoxseg2:
3180 case Intrinsic::riscv_vsoxseg3:
3181 case Intrinsic::riscv_vsoxseg4:
3182 case Intrinsic::riscv_vsoxseg5:
3183 case Intrinsic::riscv_vsoxseg6:
3184 case Intrinsic::riscv_vsoxseg7:
3185 case Intrinsic::riscv_vsoxseg8:
3186 case Intrinsic::riscv_vsuxseg2:
3187 case Intrinsic::riscv_vsuxseg3:
3188 case Intrinsic::riscv_vsuxseg4:
3189 case Intrinsic::riscv_vsuxseg5:
3190 case Intrinsic::riscv_vsuxseg6:
3191 case Intrinsic::riscv_vsuxseg7:
3192 case Intrinsic::riscv_vsuxseg8: {
3193 // Intrinsic interface (only listed ordered version):
3194 // riscv_vloxei(merge, ptr, index, vl)
3195 // riscv_vloxei_mask(merge, ptr, index, mask, vl, policy)
3196 // riscv_vsoxei(val, ptr, index, vl)
3197 // riscv_vsoxei_mask(val, ptr, index, mask, vl, policy)
3198 // riscv_vloxseg#(merge, ptr, index, vl, sew)
3199 // riscv_vloxseg#_mask(merge, ptr, index, mask, vl, policy, sew)
3200 // riscv_vsoxseg#(val, ptr, index, vl, sew)
3201 // riscv_vsoxseg#_mask(val, ptr, index, mask, vl, sew)
3202 bool IsWrite = Inst->getType()->isVoidTy();
3203 Type *Ty = IsWrite ? Inst->getArgOperand(i: 0)->getType() : Inst->getType();
3204 // The results of segment loads are TargetExtType.
3205 if (auto *TarExtTy = dyn_cast<TargetExtType>(Val: Ty)) {
3206 unsigned SEW =
3207 1 << cast<ConstantInt>(Val: Inst->getArgOperand(i: Inst->arg_size() - 1))
3208 ->getZExtValue();
3209 Ty = TarExtTy->getTypeParameter(i: 0U);
3210 Ty = ScalableVectorType::get(
3211 ElementType: IntegerType::get(C, NumBits: SEW),
3212 MinNumElts: cast<ScalableVectorType>(Val: Ty)->getMinNumElements() * 8 / SEW);
3213 }
3214 const auto *RVVIInfo = RISCVVIntrinsicsTable::getRISCVVIntrinsicInfo(IntrinsicID: IID);
3215 unsigned VLIndex = RVVIInfo->VLOperand;
3216 unsigned PtrOperandNo = VLIndex - 2 - HasMask;
3217 Value *Mask;
3218 if (HasMask) {
3219 Mask = Inst->getArgOperand(i: VLIndex - 1);
3220 } else {
3221 // Mask cannot be nullptr here: vector GEP produces <vscale x N x ptr>,
3222 // and casting that to scalar i64 triggers a vector/scalar mismatch
3223 // assertion in CreatePointerCast. Use an all-true mask so ASan lowers it
3224 // via extractelement instead.
3225 Type *MaskType = Ty->getWithNewType(EltTy: Type::getInt1Ty(C));
3226 Mask = ConstantInt::getTrue(Ty: MaskType);
3227 }
3228 Value *EVL = Inst->getArgOperand(i: VLIndex);
3229 unsigned SegNum = getSegNum(Inst, PtrOperandNo, IsWrite);
3230 // RVV uses contiguous elements as a segment.
3231 if (SegNum > 1) {
3232 unsigned ElemSize = Ty->getScalarSizeInBits();
3233 auto *SegTy = IntegerType::get(C, NumBits: ElemSize * SegNum);
3234 Ty = VectorType::get(ElementType: SegTy, Other: cast<VectorType>(Val: Ty));
3235 }
3236 Value *OffsetOp = Inst->getArgOperand(i: PtrOperandNo + 1);
3237 Info.InterestingOperands.emplace_back(Args&: Inst, Args&: PtrOperandNo, Args&: IsWrite, Args&: Ty,
3238 Args: Align(1), Args&: Mask, Args&: EVL,
3239 /* Stride */ Args: nullptr, Args&: OffsetOp);
3240 return true;
3241 }
3242 }
3243 return false;
3244}
3245
3246unsigned RISCVTTIImpl::getRegUsageForType(Type *Ty) const {
3247 if (Ty->isVectorTy()) {
3248 // f16 with only zvfhmin and bf16 will be promoted to f32
3249 Type *EltTy = cast<VectorType>(Val: Ty)->getElementType();
3250 if ((EltTy->isHalfTy() && !ST->hasVInstructionsF16()) ||
3251 EltTy->isBFloatTy())
3252 Ty = VectorType::get(ElementType: Type::getFloatTy(C&: Ty->getContext()),
3253 Other: cast<VectorType>(Val: Ty));
3254
3255 TypeSize Size = DL.getTypeSizeInBits(Ty);
3256 if (Size.isScalable() && ST->hasVInstructions())
3257 return divideCeil(Numerator: Size.getKnownMinValue(), Denominator: RISCV::RVVBitsPerBlock);
3258
3259 if (ST->useRVVForFixedLengthVectors())
3260 return divideCeil(Numerator: Size, Denominator: ST->getRealMinVLen());
3261 }
3262
3263 return BaseT::getRegUsageForType(Ty);
3264}
3265
3266unsigned RISCVTTIImpl::getMaximumVF(unsigned ElemWidth, unsigned Opcode) const {
3267 if (SLPMaxVF.getNumOccurrences())
3268 return SLPMaxVF;
3269
3270 // Return how many elements can fit in getRegisterBitwidth. This is the
3271 // same routine as used in LoopVectorizer. We should probably be
3272 // accounting for whether we actually have instructions with the right
3273 // lane type, but we don't have enough information to do that without
3274 // some additional plumbing which hasn't been justified yet.
3275 TypeSize RegWidth =
3276 getRegisterBitWidth(K: TargetTransformInfo::RGK_FixedWidthVector);
3277 // If no vector registers, or absurd element widths, disable
3278 // vectorization by returning 1.
3279 return std::max<unsigned>(a: 1U, b: RegWidth.getFixedValue() / ElemWidth);
3280}
3281
3282unsigned RISCVTTIImpl::getMinTripCountTailFoldingThreshold() const {
3283 return RVVMinTripCount;
3284}
3285
3286bool RISCVTTIImpl::preferAlternateOpcodeVectorization() const {
3287 return ST->enableUnalignedVectorMem();
3288}
3289
3290TTI::AddressingModeKind
3291RISCVTTIImpl::getPreferredAddressingMode(const Loop *L,
3292 ScalarEvolution *SE) const {
3293 if (ST->hasVendorXCVmem() && !ST->is64Bit())
3294 return TTI::AMK_PostIndexed;
3295
3296 return BasicTTIImplBase::getPreferredAddressingMode(L, SE);
3297}
3298
3299bool RISCVTTIImpl::isLSRCostLess(const TargetTransformInfo::LSRCost &C1,
3300 const TargetTransformInfo::LSRCost &C2) const {
3301 // RISC-V specific here are "instruction number 1st priority".
3302 // If we need to emit adds inside the loop to add up base registers, then
3303 // we need at least one extra temporary register.
3304 unsigned C1NumRegs = C1.NumRegs + (C1.NumBaseAdds != 0);
3305 unsigned C2NumRegs = C2.NumRegs + (C2.NumBaseAdds != 0);
3306 return std::tie(args: C1.Insns, args&: C1NumRegs, args: C1.AddRecCost,
3307 args: C1.NumIVMuls, args: C1.NumBaseAdds,
3308 args: C1.ScaleCost, args: C1.ImmCost, args: C1.SetupCost) <
3309 std::tie(args: C2.Insns, args&: C2NumRegs, args: C2.AddRecCost,
3310 args: C2.NumIVMuls, args: C2.NumBaseAdds,
3311 args: C2.ScaleCost, args: C2.ImmCost, args: C2.SetupCost);
3312}
3313
3314bool RISCVTTIImpl::isLegalMaskedExpandLoad(Type *DataTy,
3315 Align Alignment) const {
3316 auto *VTy = dyn_cast<VectorType>(Val: DataTy);
3317 if (!VTy || VTy->isScalableTy())
3318 return false;
3319
3320 if (!isLegalMaskedLoadStore(DataType: DataTy, Alignment))
3321 return false;
3322
3323 // FIXME: If it is an i8 vector and the element count exceeds 256, we should
3324 // scalarize these types with LMUL >= maximum fixed-length LMUL.
3325 if (VTy->getElementType()->isIntegerTy(Bitwidth: 8))
3326 if (VTy->getElementCount().getFixedValue() > 256)
3327 return VTy->getPrimitiveSizeInBits() / ST->getRealMinVLen() <
3328 ST->getMaxLMULForFixedLengthVectors();
3329 return true;
3330}
3331
3332bool RISCVTTIImpl::isLegalMaskedCompressStore(Type *DataTy,
3333 Align Alignment) const {
3334 auto *VTy = dyn_cast<VectorType>(Val: DataTy);
3335 if (!VTy || VTy->isScalableTy())
3336 return false;
3337
3338 if (!isLegalMaskedLoadStore(DataType: DataTy, Alignment))
3339 return false;
3340 return true;
3341}
3342
3343/// See if \p I should be considered for address type promotion. We check if \p
3344/// I is a sext with right type and used in memory accesses. If it used in a
3345/// "complex" getelementptr, we allow it to be promoted without finding other
3346/// sext instructions that sign extended the same initial value. A getelementptr
3347/// is considered as "complex" if it has more than 2 operands.
3348bool RISCVTTIImpl::shouldConsiderAddressTypePromotion(
3349 const Instruction &I, bool &AllowPromotionWithoutCommonHeader) const {
3350 bool Considerable = false;
3351 AllowPromotionWithoutCommonHeader = false;
3352 if (!isa<SExtInst>(Val: &I))
3353 return false;
3354 Type *ConsideredSExtType =
3355 Type::getInt64Ty(C&: I.getParent()->getParent()->getContext());
3356 if (I.getType() != ConsideredSExtType)
3357 return false;
3358 // See if the sext is the one with the right type and used in at least one
3359 // GetElementPtrInst.
3360 for (const User *U : I.users()) {
3361 if (const GetElementPtrInst *GEPInst = dyn_cast<GetElementPtrInst>(Val: U)) {
3362 Considerable = true;
3363 // A getelementptr is considered as "complex" if it has more than 2
3364 // operands. We will promote a SExt used in such complex GEP as we
3365 // expect some computation to be merged if they are done on 64 bits.
3366 if (GEPInst->getNumOperands() > 2) {
3367 AllowPromotionWithoutCommonHeader = true;
3368 break;
3369 }
3370 }
3371 }
3372 return Considerable;
3373}
3374
3375bool RISCVTTIImpl::canSplatOperand(unsigned Opcode, int Operand) const {
3376 switch (Opcode) {
3377 case Instruction::Add:
3378 case Instruction::Sub:
3379 case Instruction::Mul:
3380 case Instruction::And:
3381 case Instruction::Or:
3382 case Instruction::Xor:
3383 case Instruction::FAdd:
3384 case Instruction::FSub:
3385 case Instruction::FMul:
3386 case Instruction::FDiv:
3387 case Instruction::ICmp:
3388 case Instruction::FCmp:
3389 return true;
3390 case Instruction::Shl:
3391 case Instruction::LShr:
3392 case Instruction::AShr:
3393 case Instruction::UDiv:
3394 case Instruction::SDiv:
3395 case Instruction::URem:
3396 case Instruction::SRem:
3397 case Instruction::Select:
3398 return Operand == 1;
3399 default:
3400 return false;
3401 }
3402}
3403
3404bool RISCVTTIImpl::canSplatOperand(Instruction *I, int Operand) const {
3405 if (!I->getType()->isVectorTy() || !ST->hasVInstructions())
3406 return false;
3407
3408 if (canSplatOperand(Opcode: I->getOpcode(), Operand))
3409 return true;
3410
3411 auto *II = dyn_cast<IntrinsicInst>(Val: I);
3412 if (!II)
3413 return false;
3414
3415 switch (II->getIntrinsicID()) {
3416 case Intrinsic::fma:
3417 case Intrinsic::vp_fma:
3418 case Intrinsic::fmuladd:
3419 case Intrinsic::vp_fmuladd:
3420 return Operand == 0 || Operand == 1;
3421 case Intrinsic::vp_shl:
3422 case Intrinsic::vp_lshr:
3423 case Intrinsic::vp_ashr:
3424 case Intrinsic::vp_udiv:
3425 case Intrinsic::vp_sdiv:
3426 case Intrinsic::vp_urem:
3427 case Intrinsic::vp_srem:
3428 case Intrinsic::ssub_sat:
3429 case Intrinsic::vp_ssub_sat:
3430 case Intrinsic::usub_sat:
3431 case Intrinsic::vp_usub_sat:
3432 case Intrinsic::vp_select:
3433 return Operand == 1;
3434 // These intrinsics are commutative.
3435 case Intrinsic::vp_add:
3436 case Intrinsic::vp_mul:
3437 case Intrinsic::vp_and:
3438 case Intrinsic::vp_or:
3439 case Intrinsic::vp_xor:
3440 case Intrinsic::vp_fadd:
3441 case Intrinsic::vp_fmul:
3442 case Intrinsic::vp_icmp:
3443 case Intrinsic::vp_fcmp:
3444 case Intrinsic::smin:
3445 case Intrinsic::vp_smin:
3446 case Intrinsic::umin:
3447 case Intrinsic::vp_umin:
3448 case Intrinsic::smax:
3449 case Intrinsic::vp_smax:
3450 case Intrinsic::umax:
3451 case Intrinsic::vp_umax:
3452 case Intrinsic::sadd_sat:
3453 case Intrinsic::vp_sadd_sat:
3454 case Intrinsic::uadd_sat:
3455 case Intrinsic::vp_uadd_sat:
3456 // These intrinsics have 'vr' versions.
3457 case Intrinsic::vp_sub:
3458 case Intrinsic::vp_fsub:
3459 case Intrinsic::vp_fdiv:
3460 return Operand == 0 || Operand == 1;
3461 default:
3462 return false;
3463 }
3464}
3465
3466/// Check if sinking \p I's operands to I's basic block is profitable, because
3467/// the operands can be folded into a target instruction, e.g.
3468/// splats of scalars can fold into vector instructions.
3469bool RISCVTTIImpl::isProfitableToSinkOperands(
3470 Instruction *I, SmallVectorImpl<Use *> &Ops) const {
3471 using namespace llvm::PatternMatch;
3472
3473 if (I->isBitwiseLogicOp()) {
3474 if (!I->getType()->isVectorTy()) {
3475 if (ST->hasStdExtZbb() || ST->hasStdExtZbkb()) {
3476 for (auto &Op : I->operands()) {
3477 // (and/or/xor X, (not Y)) -> (andn/orn/xnor X, Y)
3478 if (match(V: Op.get(), P: m_Not(V: m_Value()))) {
3479 Ops.push_back(Elt: &Op);
3480 return true;
3481 }
3482 }
3483 }
3484 } else if (I->getOpcode() == Instruction::And && ST->hasStdExtZvkb()) {
3485 for (auto &Op : I->operands()) {
3486 // (and X, (not Y)) -> (vandn.vv X, Y)
3487 if (match(V: Op.get(), P: m_Not(V: m_Value()))) {
3488 Ops.push_back(Elt: &Op);
3489 return true;
3490 }
3491 // (and X, (splat (not Y))) -> (vandn.vx X, Y)
3492 if (match(V: Op.get(), P: m_Shuffle(v1: m_InsertElt(Val: m_Value(), Elt: m_Not(V: m_Value()),
3493 Idx: m_ZeroInt()),
3494 v2: m_Value(), mask: m_ZeroMask()))) {
3495 Use &InsertElt = cast<Instruction>(Val&: Op)->getOperandUse(i: 0);
3496 Use &Not = cast<Instruction>(Val&: InsertElt)->getOperandUse(i: 1);
3497 Ops.push_back(Elt: &Not);
3498 Ops.push_back(Elt: &InsertElt);
3499 Ops.push_back(Elt: &Op);
3500 return true;
3501 }
3502 }
3503 }
3504 }
3505
3506 if (!I->getType()->isVectorTy() || !ST->hasVInstructions())
3507 return false;
3508
3509 // Don't sink splat operands if the target prefers it. Some targets requires
3510 // S2V transfer buffers and we can run out of them copying the same value
3511 // repeatedly.
3512 // FIXME: It could still be worth doing if it would improve vector register
3513 // pressure and prevent a vector spill.
3514 if (!ST->sinkSplatOperands())
3515 return false;
3516
3517 for (auto OpIdx : enumerate(First: I->operands())) {
3518 if (!canSplatOperand(I, Operand: OpIdx.index()))
3519 continue;
3520
3521 Instruction *Op = dyn_cast<Instruction>(Val: OpIdx.value().get());
3522 // Make sure we are not already sinking this operand
3523 if (!Op || any_of(Range&: Ops, P: [&](Use *U) { return U->get() == Op; }))
3524 continue;
3525
3526 // We are looking for a splat that can be sunk.
3527 if (!match(V: Op, P: m_Shuffle(v1: m_InsertElt(Val: m_Value(), Elt: m_Value(), Idx: m_ZeroInt()),
3528 v2: m_Value(), mask: m_ZeroMask())))
3529 continue;
3530
3531 // Don't sink i1 splats.
3532 if (cast<VectorType>(Val: Op->getType())->getElementType()->isIntegerTy(Bitwidth: 1))
3533 continue;
3534
3535 // All uses of the shuffle should be sunk to avoid duplicating it across gpr
3536 // and vector registers
3537 for (Use &U : Op->uses()) {
3538 Instruction *Insn = cast<Instruction>(Val: U.getUser());
3539 if (!canSplatOperand(I: Insn, Operand: U.getOperandNo()))
3540 return false;
3541 }
3542
3543 // Sink any fpexts since they might be used in a widening fp pattern.
3544 Use *InsertEltUse = &Op->getOperandUse(i: 0);
3545 auto *InsertElt = cast<InsertElementInst>(Val: InsertEltUse);
3546 if (isa<FPExtInst>(Val: InsertElt->getOperand(i_nocapture: 1)))
3547 Ops.push_back(Elt: &InsertElt->getOperandUse(i: 1));
3548 Ops.push_back(Elt: InsertEltUse);
3549 Ops.push_back(Elt: &OpIdx.value());
3550 }
3551 return true;
3552}
3553
3554RISCVTTIImpl::TTI::MemCmpExpansionOptions
3555RISCVTTIImpl::enableMemCmpExpansion(bool OptSize, bool IsZeroCmp) const {
3556 TTI::MemCmpExpansionOptions Options;
3557 // TODO: Enable expansion when unaligned access is not supported after we fix
3558 // issues in ExpandMemcmp.
3559 if (!ST->enableUnalignedScalarMem())
3560 return Options;
3561
3562 if (!ST->hasStdExtZbb() && !ST->hasStdExtZbkb() && !IsZeroCmp)
3563 return Options;
3564
3565 Options.AllowOverlappingLoads = true;
3566 Options.MaxNumLoads = TLI->getMaxExpandSizeMemcmp(OptSize);
3567 Options.NumLoadsPerBlock = Options.MaxNumLoads;
3568 if (ST->is64Bit()) {
3569 Options.LoadSizes = {8, 4, 2, 1};
3570 Options.AllowedTailExpansions = {3, 5, 6};
3571 } else {
3572 Options.LoadSizes = {4, 2, 1};
3573 Options.AllowedTailExpansions = {3};
3574 }
3575
3576 if (IsZeroCmp && ST->hasVInstructions()) {
3577 unsigned VLenB = ST->getRealMinVLen() / 8;
3578 // The minimum size should be `XLen / 8 + 1`, and the maxinum size should be
3579 // `VLenB * MaxLMUL` so that it fits in a single register group.
3580 unsigned MinSize = ST->getXLen() / 8 + 1;
3581 unsigned MaxSize = VLenB * ST->getMaxLMULForFixedLengthVectors();
3582 for (unsigned Size = MinSize; Size <= MaxSize; Size++)
3583 Options.LoadSizes.insert(I: Options.LoadSizes.begin(), Elt: Size);
3584 }
3585 return Options;
3586}
3587
3588bool RISCVTTIImpl::shouldTreatInstructionLikeSelect(
3589 const Instruction *I) const {
3590 if (EnableOrLikeSelectOpt) {
3591 // For the binary operators (e.g. or) we need to be more careful than
3592 // selects, here we only transform them if they are already at a natural
3593 // break point in the code - the end of a block with an unconditional
3594 // terminator.
3595 if (I->getOpcode() == Instruction::Or &&
3596 isa<UncondBrInst>(Val: I->getNextNode()))
3597 return true;
3598
3599 if (I->getOpcode() == Instruction::Add ||
3600 I->getOpcode() == Instruction::Sub)
3601 return true;
3602 }
3603 return BaseT::shouldTreatInstructionLikeSelect(I);
3604}
3605
3606bool RISCVTTIImpl::shouldCopyAttributeWhenOutliningFrom(
3607 const Function *Caller, const Attribute &Attr) const {
3608 // "interrupt" controls the prolog/epilog of interrupt handlers (and includes
3609 // restrictions on their signatures). We can outline from the bodies of these
3610 // handlers, but when we do we need to make sure we don't mark the outlined
3611 // function as an interrupt handler too.
3612 if (Attr.isStringAttribute() && Attr.getKindAsString() == "interrupt")
3613 return false;
3614
3615 return BaseT::shouldCopyAttributeWhenOutliningFrom(Caller, Attr);
3616}
3617
3618std::optional<Instruction *>
3619RISCVTTIImpl::instCombineIntrinsic(InstCombiner &IC, IntrinsicInst &II) const {
3620 // If all operands of a vmv.v.x are constant, fold a bitcast(vmv.v.x) to scale
3621 // the vmv.v.x, enabling removal of the bitcast. The transform helps avoid
3622 // creating redundant masks.
3623 const DataLayout &DL = IC.getDataLayout();
3624 if (II.user_empty())
3625 return {};
3626 auto *TargetVecTy = dyn_cast<ScalableVectorType>(Val: II.user_back()->getType());
3627 if (!TargetVecTy)
3628 return {};
3629 const APInt *Scalar;
3630 uint64_t VL;
3631 if (!match(V: &II, P: m_Intrinsic<Intrinsic::riscv_vmv_v_x>(
3632 Op0: m_Poison(), Op1: m_APInt(Res&: Scalar), Op2: m_ConstantInt(V&: VL))) ||
3633 !all_of(Range: II.users(), P: [TargetVecTy](User *U) {
3634 return U->getType() == TargetVecTy && match(V: U, P: m_BitCast(Op: m_Value()));
3635 }))
3636 return {};
3637 auto *SourceVecTy = cast<ScalableVectorType>(Val: II.getType());
3638 unsigned TargetEltBW = DL.getTypeSizeInBits(Ty: TargetVecTy->getElementType());
3639 unsigned SourceEltBW = DL.getTypeSizeInBits(Ty: SourceVecTy->getElementType());
3640 if (TargetEltBW % SourceEltBW)
3641 return {};
3642 unsigned TargetScale = TargetEltBW / SourceEltBW;
3643 if (VL % TargetScale)
3644 return {};
3645 Type *VLTy = II.getOperand(i_nocapture: 2)->getType();
3646 ElementCount SourceEC = SourceVecTy->getElementCount();
3647 unsigned NewEltBW = SourceEltBW * TargetScale;
3648 if (!SourceEC.isKnownMultipleOf(RHS: TargetScale) ||
3649 !DL.fitsInLegalInteger(Width: NewEltBW))
3650 return {};
3651 auto *NewEltTy = IntegerType::get(C&: II.getContext(), NumBits: NewEltBW);
3652 if (!TLI->isLegalElementTypeForRVV(ScalarTy: TLI->getValueType(DL, Ty: NewEltTy)))
3653 return {};
3654 ElementCount NewEC = SourceEC.divideCoefficientBy(RHS: TargetScale);
3655 Type *RetTy = VectorType::get(ElementType: NewEltTy, EC: NewEC);
3656 assert(SourceVecTy->canLosslesslyBitCastTo(RetTy) &&
3657 "Lossless bitcast between types expected");
3658 APInt NewScalar = APInt::getSplat(NewLen: NewEltBW, V: *Scalar);
3659 return IC.replaceInstUsesWith(
3660 I&: II,
3661 V: IC.Builder.CreateBitCast(
3662 V: IC.Builder.CreateIntrinsic(
3663 RetTy, ID: Intrinsic::riscv_vmv_v_x,
3664 Args: {PoisonValue::get(T: RetTy), ConstantInt::get(Ty: NewEltTy, V: NewScalar),
3665 ConstantInt::get(Ty: VLTy, V: VL / TargetScale)}),
3666 DestTy: SourceVecTy));
3667}
3668