xref: /freebsd/contrib/llvm-project/llvm/lib/Target/RISCV/RISCVTargetTransformInfo.cpp (revision 63f537551380d2dab29fa402ad1269feae17e594)
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/Analysis/TargetTransformInfo.h"
12 #include "llvm/CodeGen/BasicTTIImpl.h"
13 #include "llvm/CodeGen/CostTable.h"
14 #include "llvm/CodeGen/TargetLowering.h"
15 #include <cmath>
16 #include <optional>
17 using namespace llvm;
18 
19 #define DEBUG_TYPE "riscvtti"
20 
21 static cl::opt<unsigned> RVVRegisterWidthLMUL(
22     "riscv-v-register-bit-width-lmul",
23     cl::desc(
24         "The LMUL to use for getRegisterBitWidth queries. Affects LMUL used "
25         "by autovectorized code. Fractional LMULs are not supported."),
26     cl::init(1), cl::Hidden);
27 
28 static cl::opt<unsigned> SLPMaxVF(
29     "riscv-v-slp-max-vf",
30     cl::desc(
31         "Result used for getMaximumVF query which is used exclusively by "
32         "SLP vectorizer.  Defaults to 1 which disables SLP."),
33     cl::init(1), cl::Hidden);
34 
35 InstructionCost RISCVTTIImpl::getLMULCost(MVT VT) {
36   // TODO: Here assume reciprocal throughput is 1 for LMUL_1, it is
37   // implementation-defined.
38   if (!VT.isVector())
39     return InstructionCost::getInvalid();
40   unsigned Cost;
41   if (VT.isScalableVector()) {
42     unsigned LMul;
43     bool Fractional;
44     std::tie(LMul, Fractional) =
45         RISCVVType::decodeVLMUL(RISCVTargetLowering::getLMUL(VT));
46     if (Fractional)
47       Cost = 1;
48     else
49       Cost = LMul;
50   } else {
51     Cost = VT.getSizeInBits() / ST->getRealMinVLen();
52   }
53   return std::max<unsigned>(Cost, 1);
54 }
55 
56 InstructionCost RISCVTTIImpl::getIntImmCost(const APInt &Imm, Type *Ty,
57                                             TTI::TargetCostKind CostKind) {
58   assert(Ty->isIntegerTy() &&
59          "getIntImmCost can only estimate cost of materialising integers");
60 
61   // We have a Zero register, so 0 is always free.
62   if (Imm == 0)
63     return TTI::TCC_Free;
64 
65   // Otherwise, we check how many instructions it will take to materialise.
66   const DataLayout &DL = getDataLayout();
67   return RISCVMatInt::getIntMatCost(Imm, DL.getTypeSizeInBits(Ty),
68                                     getST()->getFeatureBits());
69 }
70 
71 // Look for patterns of shift followed by AND that can be turned into a pair of
72 // shifts. We won't need to materialize an immediate for the AND so these can
73 // be considered free.
74 static bool canUseShiftPair(Instruction *Inst, const APInt &Imm) {
75   uint64_t Mask = Imm.getZExtValue();
76   auto *BO = dyn_cast<BinaryOperator>(Inst->getOperand(0));
77   if (!BO || !BO->hasOneUse())
78     return false;
79 
80   if (BO->getOpcode() != Instruction::Shl)
81     return false;
82 
83   if (!isa<ConstantInt>(BO->getOperand(1)))
84     return false;
85 
86   unsigned ShAmt = cast<ConstantInt>(BO->getOperand(1))->getZExtValue();
87   // (and (shl x, c2), c1) will be matched to (srli (slli x, c2+c3), c3) if c1
88   // is a mask shifted by c2 bits with c3 leading zeros.
89   if (isShiftedMask_64(Mask)) {
90     unsigned Trailing = countTrailingZeros(Mask);
91     if (ShAmt == Trailing)
92       return true;
93   }
94 
95   return false;
96 }
97 
98 InstructionCost RISCVTTIImpl::getIntImmCostInst(unsigned Opcode, unsigned Idx,
99                                                 const APInt &Imm, Type *Ty,
100                                                 TTI::TargetCostKind CostKind,
101                                                 Instruction *Inst) {
102   assert(Ty->isIntegerTy() &&
103          "getIntImmCost can only estimate cost of materialising integers");
104 
105   // We have a Zero register, so 0 is always free.
106   if (Imm == 0)
107     return TTI::TCC_Free;
108 
109   // Some instructions in RISC-V can take a 12-bit immediate. Some of these are
110   // commutative, in others the immediate comes from a specific argument index.
111   bool Takes12BitImm = false;
112   unsigned ImmArgIdx = ~0U;
113 
114   switch (Opcode) {
115   case Instruction::GetElementPtr:
116     // Never hoist any arguments to a GetElementPtr. CodeGenPrepare will
117     // split up large offsets in GEP into better parts than ConstantHoisting
118     // can.
119     return TTI::TCC_Free;
120   case Instruction::And:
121     // zext.h
122     if (Imm == UINT64_C(0xffff) && ST->hasStdExtZbb())
123       return TTI::TCC_Free;
124     // zext.w
125     if (Imm == UINT64_C(0xffffffff) && ST->hasStdExtZba())
126       return TTI::TCC_Free;
127     // bclri
128     if (ST->hasStdExtZbs() && (~Imm).isPowerOf2())
129       return TTI::TCC_Free;
130     if (Inst && Idx == 1 && Imm.getBitWidth() <= ST->getXLen() &&
131         canUseShiftPair(Inst, Imm))
132       return TTI::TCC_Free;
133     Takes12BitImm = true;
134     break;
135   case Instruction::Add:
136     Takes12BitImm = true;
137     break;
138   case Instruction::Or:
139   case Instruction::Xor:
140     // bseti/binvi
141     if (ST->hasStdExtZbs() && Imm.isPowerOf2())
142       return TTI::TCC_Free;
143     Takes12BitImm = true;
144     break;
145   case Instruction::Mul:
146     // Negated power of 2 is a shift and a negate.
147     if (Imm.isNegatedPowerOf2())
148       return TTI::TCC_Free;
149     // FIXME: There is no MULI instruction.
150     Takes12BitImm = true;
151     break;
152   case Instruction::Sub:
153   case Instruction::Shl:
154   case Instruction::LShr:
155   case Instruction::AShr:
156     Takes12BitImm = true;
157     ImmArgIdx = 1;
158     break;
159   default:
160     break;
161   }
162 
163   if (Takes12BitImm) {
164     // Check immediate is the correct argument...
165     if (Instruction::isCommutative(Opcode) || Idx == ImmArgIdx) {
166       // ... and fits into the 12-bit immediate.
167       if (Imm.getMinSignedBits() <= 64 &&
168           getTLI()->isLegalAddImmediate(Imm.getSExtValue())) {
169         return TTI::TCC_Free;
170       }
171     }
172 
173     // Otherwise, use the full materialisation cost.
174     return getIntImmCost(Imm, Ty, CostKind);
175   }
176 
177   // By default, prevent hoisting.
178   return TTI::TCC_Free;
179 }
180 
181 InstructionCost
182 RISCVTTIImpl::getIntImmCostIntrin(Intrinsic::ID IID, unsigned Idx,
183                                   const APInt &Imm, Type *Ty,
184                                   TTI::TargetCostKind CostKind) {
185   // Prevent hoisting in unknown cases.
186   return TTI::TCC_Free;
187 }
188 
189 TargetTransformInfo::PopcntSupportKind
190 RISCVTTIImpl::getPopcntSupport(unsigned TyWidth) {
191   assert(isPowerOf2_32(TyWidth) && "Ty width must be power of 2");
192   return ST->hasStdExtZbb() ? TTI::PSK_FastHardware : TTI::PSK_Software;
193 }
194 
195 bool RISCVTTIImpl::shouldExpandReduction(const IntrinsicInst *II) const {
196   // Currently, the ExpandReductions pass can't expand scalable-vector
197   // reductions, but we still request expansion as RVV doesn't support certain
198   // reductions and the SelectionDAG can't legalize them either.
199   switch (II->getIntrinsicID()) {
200   default:
201     return false;
202   // These reductions have no equivalent in RVV
203   case Intrinsic::vector_reduce_mul:
204   case Intrinsic::vector_reduce_fmul:
205     return true;
206   }
207 }
208 
209 std::optional<unsigned> RISCVTTIImpl::getMaxVScale() const {
210   if (ST->hasVInstructions())
211     return ST->getRealMaxVLen() / RISCV::RVVBitsPerBlock;
212   return BaseT::getMaxVScale();
213 }
214 
215 std::optional<unsigned> RISCVTTIImpl::getVScaleForTuning() const {
216   if (ST->hasVInstructions())
217     if (unsigned MinVLen = ST->getRealMinVLen();
218         MinVLen >= RISCV::RVVBitsPerBlock)
219       return MinVLen / RISCV::RVVBitsPerBlock;
220   return BaseT::getVScaleForTuning();
221 }
222 
223 TypeSize
224 RISCVTTIImpl::getRegisterBitWidth(TargetTransformInfo::RegisterKind K) const {
225   unsigned LMUL = PowerOf2Floor(
226       std::max<unsigned>(std::min<unsigned>(RVVRegisterWidthLMUL, 8), 1));
227   switch (K) {
228   case TargetTransformInfo::RGK_Scalar:
229     return TypeSize::getFixed(ST->getXLen());
230   case TargetTransformInfo::RGK_FixedWidthVector:
231     return TypeSize::getFixed(
232         ST->useRVVForFixedLengthVectors() ? LMUL * ST->getRealMinVLen() : 0);
233   case TargetTransformInfo::RGK_ScalableVector:
234     return TypeSize::getScalable(
235         (ST->hasVInstructions() &&
236          ST->getRealMinVLen() >= RISCV::RVVBitsPerBlock)
237             ? LMUL * RISCV::RVVBitsPerBlock
238             : 0);
239   }
240 
241   llvm_unreachable("Unsupported register kind");
242 }
243 
244 InstructionCost RISCVTTIImpl::getSpliceCost(VectorType *Tp, int Index) {
245   std::pair<InstructionCost, MVT> LT = getTypeLegalizationCost(Tp);
246 
247   unsigned Cost = 2; // vslidedown+vslideup.
248   // TODO: Multiplying by LT.first implies this legalizes into multiple copies
249   // of similar code, but I think we expand through memory.
250   return Cost * LT.first * getLMULCost(LT.second);
251 }
252 
253 InstructionCost RISCVTTIImpl::getShuffleCost(TTI::ShuffleKind Kind,
254                                              VectorType *Tp, ArrayRef<int> Mask,
255                                              TTI::TargetCostKind CostKind,
256                                              int Index, VectorType *SubTp,
257                                              ArrayRef<const Value *> Args) {
258   if (isa<ScalableVectorType>(Tp)) {
259     std::pair<InstructionCost, MVT> LT = getTypeLegalizationCost(Tp);
260     switch (Kind) {
261     default:
262       // Fallthrough to generic handling.
263       // TODO: Most of these cases will return getInvalid in generic code, and
264       // must be implemented here.
265       break;
266     case TTI::SK_Broadcast: {
267       return LT.first * 1;
268     }
269     case TTI::SK_Splice:
270       return getSpliceCost(Tp, Index);
271     case TTI::SK_Reverse:
272       // Most of the cost here is producing the vrgather index register
273       // Example sequence:
274       //   csrr a0, vlenb
275       //   srli a0, a0, 3
276       //   addi a0, a0, -1
277       //   vsetvli a1, zero, e8, mf8, ta, mu (ignored)
278       //   vid.v v9
279       //   vrsub.vx v10, v9, a0
280       //   vrgather.vv v9, v8, v10
281       if (Tp->getElementType()->isIntegerTy(1))
282         // Mask operation additionally required extend and truncate
283         return LT.first * 9;
284       return LT.first * 6;
285     }
286   }
287 
288   if (isa<FixedVectorType>(Tp) && Kind == TargetTransformInfo::SK_Broadcast) {
289     std::pair<InstructionCost, MVT> LT = getTypeLegalizationCost(Tp);
290     bool HasScalar = (Args.size() > 0) && (Operator::getOpcode(Args[0]) ==
291                                            Instruction::InsertElement);
292     if (LT.second.getScalarSizeInBits() == 1) {
293       if (HasScalar) {
294         // Example sequence:
295         //   andi a0, a0, 1
296         //   vsetivli zero, 2, e8, mf8, ta, ma (ignored)
297         //   vmv.v.x v8, a0
298         //   vmsne.vi v0, v8, 0
299         return LT.first * getLMULCost(LT.second) * 3;
300       }
301       // Example sequence:
302       //   vsetivli  zero, 2, e8, mf8, ta, mu (ignored)
303       //   vmv.v.i v8, 0
304       //   vmerge.vim      v8, v8, 1, v0
305       //   vmv.x.s a0, v8
306       //   andi    a0, a0, 1
307       //   vmv.v.x v8, a0
308       //   vmsne.vi  v0, v8, 0
309 
310       return LT.first * getLMULCost(LT.second) * 6;
311     }
312 
313     if (HasScalar) {
314       // Example sequence:
315       //   vmv.v.x v8, a0
316       return LT.first * getLMULCost(LT.second);
317     }
318 
319     // Example sequence:
320     //   vrgather.vi     v9, v8, 0
321     // TODO: vrgather could be slower than vmv.v.x. It is
322     // implementation-dependent.
323     return LT.first * getLMULCost(LT.second);
324   }
325 
326   return BaseT::getShuffleCost(Kind, Tp, Mask, CostKind, Index, SubTp);
327 }
328 
329 InstructionCost
330 RISCVTTIImpl::getMaskedMemoryOpCost(unsigned Opcode, Type *Src, Align Alignment,
331                                     unsigned AddressSpace,
332                                     TTI::TargetCostKind CostKind) {
333   if (!isLegalMaskedLoadStore(Src, Alignment) ||
334       CostKind != TTI::TCK_RecipThroughput)
335     return BaseT::getMaskedMemoryOpCost(Opcode, Src, Alignment, AddressSpace,
336                                         CostKind);
337 
338   return getMemoryOpCost(Opcode, Src, Alignment, AddressSpace, CostKind);
339 }
340 
341 InstructionCost RISCVTTIImpl::getGatherScatterOpCost(
342     unsigned Opcode, Type *DataTy, const Value *Ptr, bool VariableMask,
343     Align Alignment, TTI::TargetCostKind CostKind, const Instruction *I) {
344   if (CostKind != TTI::TCK_RecipThroughput)
345     return BaseT::getGatherScatterOpCost(Opcode, DataTy, Ptr, VariableMask,
346                                          Alignment, CostKind, I);
347 
348   if ((Opcode == Instruction::Load &&
349        !isLegalMaskedGather(DataTy, Align(Alignment))) ||
350       (Opcode == Instruction::Store &&
351        !isLegalMaskedScatter(DataTy, Align(Alignment))))
352     return BaseT::getGatherScatterOpCost(Opcode, DataTy, Ptr, VariableMask,
353                                          Alignment, CostKind, I);
354 
355   // Cost is proportional to the number of memory operations implied.  For
356   // scalable vectors, we use an estimate on that number since we don't
357   // know exactly what VL will be.
358   auto &VTy = *cast<VectorType>(DataTy);
359   InstructionCost MemOpCost =
360       getMemoryOpCost(Opcode, VTy.getElementType(), Alignment, 0, CostKind,
361                       {TTI::OK_AnyValue, TTI::OP_None}, I);
362   unsigned NumLoads = getEstimatedVLFor(&VTy);
363   return NumLoads * MemOpCost;
364 }
365 
366 // Currently, these represent both throughput and codesize costs
367 // for the respective intrinsics.  The costs in this table are simply
368 // instruction counts with the following adjustments made:
369 // * One vsetvli is considered free.
370 static const CostTblEntry VectorIntrinsicCostTable[]{
371     {Intrinsic::floor, MVT::v2f32, 9},
372     {Intrinsic::floor, MVT::v4f32, 9},
373     {Intrinsic::floor, MVT::v8f32, 9},
374     {Intrinsic::floor, MVT::v16f32, 9},
375     {Intrinsic::floor, MVT::nxv1f32, 9},
376     {Intrinsic::floor, MVT::nxv2f32, 9},
377     {Intrinsic::floor, MVT::nxv4f32, 9},
378     {Intrinsic::floor, MVT::nxv8f32, 9},
379     {Intrinsic::floor, MVT::nxv16f32, 9},
380     {Intrinsic::floor, MVT::v2f64, 9},
381     {Intrinsic::floor, MVT::v4f64, 9},
382     {Intrinsic::floor, MVT::v8f64, 9},
383     {Intrinsic::floor, MVT::v16f64, 9},
384     {Intrinsic::floor, MVT::nxv1f64, 9},
385     {Intrinsic::floor, MVT::nxv2f64, 9},
386     {Intrinsic::floor, MVT::nxv4f64, 9},
387     {Intrinsic::floor, MVT::nxv8f64, 9},
388     {Intrinsic::ceil, MVT::v2f32, 9},
389     {Intrinsic::ceil, MVT::v4f32, 9},
390     {Intrinsic::ceil, MVT::v8f32, 9},
391     {Intrinsic::ceil, MVT::v16f32, 9},
392     {Intrinsic::ceil, MVT::nxv1f32, 9},
393     {Intrinsic::ceil, MVT::nxv2f32, 9},
394     {Intrinsic::ceil, MVT::nxv4f32, 9},
395     {Intrinsic::ceil, MVT::nxv8f32, 9},
396     {Intrinsic::ceil, MVT::nxv16f32, 9},
397     {Intrinsic::ceil, MVT::v2f64, 9},
398     {Intrinsic::ceil, MVT::v4f64, 9},
399     {Intrinsic::ceil, MVT::v8f64, 9},
400     {Intrinsic::ceil, MVT::v16f64, 9},
401     {Intrinsic::ceil, MVT::nxv1f64, 9},
402     {Intrinsic::ceil, MVT::nxv2f64, 9},
403     {Intrinsic::ceil, MVT::nxv4f64, 9},
404     {Intrinsic::ceil, MVT::nxv8f64, 9},
405     {Intrinsic::trunc, MVT::v2f32, 7},
406     {Intrinsic::trunc, MVT::v4f32, 7},
407     {Intrinsic::trunc, MVT::v8f32, 7},
408     {Intrinsic::trunc, MVT::v16f32, 7},
409     {Intrinsic::trunc, MVT::nxv1f32, 7},
410     {Intrinsic::trunc, MVT::nxv2f32, 7},
411     {Intrinsic::trunc, MVT::nxv4f32, 7},
412     {Intrinsic::trunc, MVT::nxv8f32, 7},
413     {Intrinsic::trunc, MVT::nxv16f32, 7},
414     {Intrinsic::trunc, MVT::v2f64, 7},
415     {Intrinsic::trunc, MVT::v4f64, 7},
416     {Intrinsic::trunc, MVT::v8f64, 7},
417     {Intrinsic::trunc, MVT::v16f64, 7},
418     {Intrinsic::trunc, MVT::nxv1f64, 7},
419     {Intrinsic::trunc, MVT::nxv2f64, 7},
420     {Intrinsic::trunc, MVT::nxv4f64, 7},
421     {Intrinsic::trunc, MVT::nxv8f64, 7},
422     {Intrinsic::round, MVT::v2f32, 9},
423     {Intrinsic::round, MVT::v4f32, 9},
424     {Intrinsic::round, MVT::v8f32, 9},
425     {Intrinsic::round, MVT::v16f32, 9},
426     {Intrinsic::round, MVT::nxv1f32, 9},
427     {Intrinsic::round, MVT::nxv2f32, 9},
428     {Intrinsic::round, MVT::nxv4f32, 9},
429     {Intrinsic::round, MVT::nxv8f32, 9},
430     {Intrinsic::round, MVT::nxv16f32, 9},
431     {Intrinsic::round, MVT::v2f64, 9},
432     {Intrinsic::round, MVT::v4f64, 9},
433     {Intrinsic::round, MVT::v8f64, 9},
434     {Intrinsic::round, MVT::v16f64, 9},
435     {Intrinsic::round, MVT::nxv1f64, 9},
436     {Intrinsic::round, MVT::nxv2f64, 9},
437     {Intrinsic::round, MVT::nxv4f64, 9},
438     {Intrinsic::round, MVT::nxv8f64, 9},
439     {Intrinsic::roundeven, MVT::v2f32, 9},
440     {Intrinsic::roundeven, MVT::v4f32, 9},
441     {Intrinsic::roundeven, MVT::v8f32, 9},
442     {Intrinsic::roundeven, MVT::v16f32, 9},
443     {Intrinsic::roundeven, MVT::nxv1f32, 9},
444     {Intrinsic::roundeven, MVT::nxv2f32, 9},
445     {Intrinsic::roundeven, MVT::nxv4f32, 9},
446     {Intrinsic::roundeven, MVT::nxv8f32, 9},
447     {Intrinsic::roundeven, MVT::nxv16f32, 9},
448     {Intrinsic::roundeven, MVT::v2f64, 9},
449     {Intrinsic::roundeven, MVT::v4f64, 9},
450     {Intrinsic::roundeven, MVT::v8f64, 9},
451     {Intrinsic::roundeven, MVT::v16f64, 9},
452     {Intrinsic::roundeven, MVT::nxv1f64, 9},
453     {Intrinsic::roundeven, MVT::nxv2f64, 9},
454     {Intrinsic::roundeven, MVT::nxv4f64, 9},
455     {Intrinsic::roundeven, MVT::nxv8f64, 9},
456     {Intrinsic::bswap, MVT::v2i16, 3},
457     {Intrinsic::bswap, MVT::v4i16, 3},
458     {Intrinsic::bswap, MVT::v8i16, 3},
459     {Intrinsic::bswap, MVT::v16i16, 3},
460     {Intrinsic::bswap, MVT::nxv1i16, 3},
461     {Intrinsic::bswap, MVT::nxv2i16, 3},
462     {Intrinsic::bswap, MVT::nxv4i16, 3},
463     {Intrinsic::bswap, MVT::nxv8i16, 3},
464     {Intrinsic::bswap, MVT::nxv16i16, 3},
465     {Intrinsic::bswap, MVT::v2i32, 12},
466     {Intrinsic::bswap, MVT::v4i32, 12},
467     {Intrinsic::bswap, MVT::v8i32, 12},
468     {Intrinsic::bswap, MVT::v16i32, 12},
469     {Intrinsic::bswap, MVT::nxv1i32, 12},
470     {Intrinsic::bswap, MVT::nxv2i32, 12},
471     {Intrinsic::bswap, MVT::nxv4i32, 12},
472     {Intrinsic::bswap, MVT::nxv8i32, 12},
473     {Intrinsic::bswap, MVT::nxv16i32, 12},
474     {Intrinsic::bswap, MVT::v2i64, 31},
475     {Intrinsic::bswap, MVT::v4i64, 31},
476     {Intrinsic::bswap, MVT::v8i64, 31},
477     {Intrinsic::bswap, MVT::v16i64, 31},
478     {Intrinsic::bswap, MVT::nxv1i64, 31},
479     {Intrinsic::bswap, MVT::nxv2i64, 31},
480     {Intrinsic::bswap, MVT::nxv4i64, 31},
481     {Intrinsic::bswap, MVT::nxv8i64, 31},
482     {Intrinsic::vp_bswap, MVT::v2i16, 3},
483     {Intrinsic::vp_bswap, MVT::v4i16, 3},
484     {Intrinsic::vp_bswap, MVT::v8i16, 3},
485     {Intrinsic::vp_bswap, MVT::v16i16, 3},
486     {Intrinsic::vp_bswap, MVT::nxv1i16, 3},
487     {Intrinsic::vp_bswap, MVT::nxv2i16, 3},
488     {Intrinsic::vp_bswap, MVT::nxv4i16, 3},
489     {Intrinsic::vp_bswap, MVT::nxv8i16, 3},
490     {Intrinsic::vp_bswap, MVT::nxv16i16, 3},
491     {Intrinsic::vp_bswap, MVT::v2i32, 12},
492     {Intrinsic::vp_bswap, MVT::v4i32, 12},
493     {Intrinsic::vp_bswap, MVT::v8i32, 12},
494     {Intrinsic::vp_bswap, MVT::v16i32, 12},
495     {Intrinsic::vp_bswap, MVT::nxv1i32, 12},
496     {Intrinsic::vp_bswap, MVT::nxv2i32, 12},
497     {Intrinsic::vp_bswap, MVT::nxv4i32, 12},
498     {Intrinsic::vp_bswap, MVT::nxv8i32, 12},
499     {Intrinsic::vp_bswap, MVT::nxv16i32, 12},
500     {Intrinsic::vp_bswap, MVT::v2i64, 31},
501     {Intrinsic::vp_bswap, MVT::v4i64, 31},
502     {Intrinsic::vp_bswap, MVT::v8i64, 31},
503     {Intrinsic::vp_bswap, MVT::v16i64, 31},
504     {Intrinsic::vp_bswap, MVT::nxv1i64, 31},
505     {Intrinsic::vp_bswap, MVT::nxv2i64, 31},
506     {Intrinsic::vp_bswap, MVT::nxv4i64, 31},
507     {Intrinsic::vp_bswap, MVT::nxv8i64, 31},
508     {Intrinsic::vp_fshl, MVT::v2i8, 7},
509     {Intrinsic::vp_fshl, MVT::v4i8, 7},
510     {Intrinsic::vp_fshl, MVT::v8i8, 7},
511     {Intrinsic::vp_fshl, MVT::v16i8, 7},
512     {Intrinsic::vp_fshl, MVT::nxv1i8, 7},
513     {Intrinsic::vp_fshl, MVT::nxv2i8, 7},
514     {Intrinsic::vp_fshl, MVT::nxv4i8, 7},
515     {Intrinsic::vp_fshl, MVT::nxv8i8, 7},
516     {Intrinsic::vp_fshl, MVT::nxv16i8, 7},
517     {Intrinsic::vp_fshl, MVT::nxv32i8, 7},
518     {Intrinsic::vp_fshl, MVT::nxv64i8, 7},
519     {Intrinsic::vp_fshl, MVT::v2i16, 7},
520     {Intrinsic::vp_fshl, MVT::v4i16, 7},
521     {Intrinsic::vp_fshl, MVT::v8i16, 7},
522     {Intrinsic::vp_fshl, MVT::v16i16, 7},
523     {Intrinsic::vp_fshl, MVT::nxv1i16, 7},
524     {Intrinsic::vp_fshl, MVT::nxv2i16, 7},
525     {Intrinsic::vp_fshl, MVT::nxv4i16, 7},
526     {Intrinsic::vp_fshl, MVT::nxv8i16, 7},
527     {Intrinsic::vp_fshl, MVT::nxv16i16, 7},
528     {Intrinsic::vp_fshl, MVT::nxv32i16, 7},
529     {Intrinsic::vp_fshl, MVT::v2i32, 7},
530     {Intrinsic::vp_fshl, MVT::v4i32, 7},
531     {Intrinsic::vp_fshl, MVT::v8i32, 7},
532     {Intrinsic::vp_fshl, MVT::v16i32, 7},
533     {Intrinsic::vp_fshl, MVT::nxv1i32, 7},
534     {Intrinsic::vp_fshl, MVT::nxv2i32, 7},
535     {Intrinsic::vp_fshl, MVT::nxv4i32, 7},
536     {Intrinsic::vp_fshl, MVT::nxv8i32, 7},
537     {Intrinsic::vp_fshl, MVT::nxv16i32, 7},
538     {Intrinsic::vp_fshl, MVT::v2i64, 7},
539     {Intrinsic::vp_fshl, MVT::v4i64, 7},
540     {Intrinsic::vp_fshl, MVT::v8i64, 7},
541     {Intrinsic::vp_fshl, MVT::v16i64, 7},
542     {Intrinsic::vp_fshl, MVT::nxv1i64, 7},
543     {Intrinsic::vp_fshl, MVT::nxv2i64, 7},
544     {Intrinsic::vp_fshl, MVT::nxv4i64, 7},
545     {Intrinsic::vp_fshl, MVT::nxv8i64, 7},
546     {Intrinsic::vp_fshr, MVT::v2i8, 7},
547     {Intrinsic::vp_fshr, MVT::v4i8, 7},
548     {Intrinsic::vp_fshr, MVT::v8i8, 7},
549     {Intrinsic::vp_fshr, MVT::v16i8, 7},
550     {Intrinsic::vp_fshr, MVT::nxv1i8, 7},
551     {Intrinsic::vp_fshr, MVT::nxv2i8, 7},
552     {Intrinsic::vp_fshr, MVT::nxv4i8, 7},
553     {Intrinsic::vp_fshr, MVT::nxv8i8, 7},
554     {Intrinsic::vp_fshr, MVT::nxv16i8, 7},
555     {Intrinsic::vp_fshr, MVT::nxv32i8, 7},
556     {Intrinsic::vp_fshr, MVT::nxv64i8, 7},
557     {Intrinsic::vp_fshr, MVT::v2i16, 7},
558     {Intrinsic::vp_fshr, MVT::v4i16, 7},
559     {Intrinsic::vp_fshr, MVT::v8i16, 7},
560     {Intrinsic::vp_fshr, MVT::v16i16, 7},
561     {Intrinsic::vp_fshr, MVT::nxv1i16, 7},
562     {Intrinsic::vp_fshr, MVT::nxv2i16, 7},
563     {Intrinsic::vp_fshr, MVT::nxv4i16, 7},
564     {Intrinsic::vp_fshr, MVT::nxv8i16, 7},
565     {Intrinsic::vp_fshr, MVT::nxv16i16, 7},
566     {Intrinsic::vp_fshr, MVT::nxv32i16, 7},
567     {Intrinsic::vp_fshr, MVT::v2i32, 7},
568     {Intrinsic::vp_fshr, MVT::v4i32, 7},
569     {Intrinsic::vp_fshr, MVT::v8i32, 7},
570     {Intrinsic::vp_fshr, MVT::v16i32, 7},
571     {Intrinsic::vp_fshr, MVT::nxv1i32, 7},
572     {Intrinsic::vp_fshr, MVT::nxv2i32, 7},
573     {Intrinsic::vp_fshr, MVT::nxv4i32, 7},
574     {Intrinsic::vp_fshr, MVT::nxv8i32, 7},
575     {Intrinsic::vp_fshr, MVT::nxv16i32, 7},
576     {Intrinsic::vp_fshr, MVT::v2i64, 7},
577     {Intrinsic::vp_fshr, MVT::v4i64, 7},
578     {Intrinsic::vp_fshr, MVT::v8i64, 7},
579     {Intrinsic::vp_fshr, MVT::v16i64, 7},
580     {Intrinsic::vp_fshr, MVT::nxv1i64, 7},
581     {Intrinsic::vp_fshr, MVT::nxv2i64, 7},
582     {Intrinsic::vp_fshr, MVT::nxv4i64, 7},
583     {Intrinsic::vp_fshr, MVT::nxv8i64, 7},
584     {Intrinsic::bitreverse, MVT::v2i8, 17},
585     {Intrinsic::bitreverse, MVT::v4i8, 17},
586     {Intrinsic::bitreverse, MVT::v8i8, 17},
587     {Intrinsic::bitreverse, MVT::v16i8, 17},
588     {Intrinsic::bitreverse, MVT::nxv1i8, 17},
589     {Intrinsic::bitreverse, MVT::nxv2i8, 17},
590     {Intrinsic::bitreverse, MVT::nxv4i8, 17},
591     {Intrinsic::bitreverse, MVT::nxv8i8, 17},
592     {Intrinsic::bitreverse, MVT::nxv16i8, 17},
593     {Intrinsic::bitreverse, MVT::v2i16, 24},
594     {Intrinsic::bitreverse, MVT::v4i16, 24},
595     {Intrinsic::bitreverse, MVT::v8i16, 24},
596     {Intrinsic::bitreverse, MVT::v16i16, 24},
597     {Intrinsic::bitreverse, MVT::nxv1i16, 24},
598     {Intrinsic::bitreverse, MVT::nxv2i16, 24},
599     {Intrinsic::bitreverse, MVT::nxv4i16, 24},
600     {Intrinsic::bitreverse, MVT::nxv8i16, 24},
601     {Intrinsic::bitreverse, MVT::nxv16i16, 24},
602     {Intrinsic::bitreverse, MVT::v2i32, 33},
603     {Intrinsic::bitreverse, MVT::v4i32, 33},
604     {Intrinsic::bitreverse, MVT::v8i32, 33},
605     {Intrinsic::bitreverse, MVT::v16i32, 33},
606     {Intrinsic::bitreverse, MVT::nxv1i32, 33},
607     {Intrinsic::bitreverse, MVT::nxv2i32, 33},
608     {Intrinsic::bitreverse, MVT::nxv4i32, 33},
609     {Intrinsic::bitreverse, MVT::nxv8i32, 33},
610     {Intrinsic::bitreverse, MVT::nxv16i32, 33},
611     {Intrinsic::bitreverse, MVT::v2i64, 52},
612     {Intrinsic::bitreverse, MVT::v4i64, 52},
613     {Intrinsic::bitreverse, MVT::v8i64, 52},
614     {Intrinsic::bitreverse, MVT::v16i64, 52},
615     {Intrinsic::bitreverse, MVT::nxv1i64, 52},
616     {Intrinsic::bitreverse, MVT::nxv2i64, 52},
617     {Intrinsic::bitreverse, MVT::nxv4i64, 52},
618     {Intrinsic::bitreverse, MVT::nxv8i64, 52},
619     {Intrinsic::vp_bitreverse, MVT::v2i8, 17},
620     {Intrinsic::vp_bitreverse, MVT::v4i8, 17},
621     {Intrinsic::vp_bitreverse, MVT::v8i8, 17},
622     {Intrinsic::vp_bitreverse, MVT::v16i8, 17},
623     {Intrinsic::vp_bitreverse, MVT::nxv1i8, 17},
624     {Intrinsic::vp_bitreverse, MVT::nxv2i8, 17},
625     {Intrinsic::vp_bitreverse, MVT::nxv4i8, 17},
626     {Intrinsic::vp_bitreverse, MVT::nxv8i8, 17},
627     {Intrinsic::vp_bitreverse, MVT::nxv16i8, 17},
628     {Intrinsic::vp_bitreverse, MVT::v2i16, 24},
629     {Intrinsic::vp_bitreverse, MVT::v4i16, 24},
630     {Intrinsic::vp_bitreverse, MVT::v8i16, 24},
631     {Intrinsic::vp_bitreverse, MVT::v16i16, 24},
632     {Intrinsic::vp_bitreverse, MVT::nxv1i16, 24},
633     {Intrinsic::vp_bitreverse, MVT::nxv2i16, 24},
634     {Intrinsic::vp_bitreverse, MVT::nxv4i16, 24},
635     {Intrinsic::vp_bitreverse, MVT::nxv8i16, 24},
636     {Intrinsic::vp_bitreverse, MVT::nxv16i16, 24},
637     {Intrinsic::vp_bitreverse, MVT::v2i32, 33},
638     {Intrinsic::vp_bitreverse, MVT::v4i32, 33},
639     {Intrinsic::vp_bitreverse, MVT::v8i32, 33},
640     {Intrinsic::vp_bitreverse, MVT::v16i32, 33},
641     {Intrinsic::vp_bitreverse, MVT::nxv1i32, 33},
642     {Intrinsic::vp_bitreverse, MVT::nxv2i32, 33},
643     {Intrinsic::vp_bitreverse, MVT::nxv4i32, 33},
644     {Intrinsic::vp_bitreverse, MVT::nxv8i32, 33},
645     {Intrinsic::vp_bitreverse, MVT::nxv16i32, 33},
646     {Intrinsic::vp_bitreverse, MVT::v2i64, 52},
647     {Intrinsic::vp_bitreverse, MVT::v4i64, 52},
648     {Intrinsic::vp_bitreverse, MVT::v8i64, 52},
649     {Intrinsic::vp_bitreverse, MVT::v16i64, 52},
650     {Intrinsic::vp_bitreverse, MVT::nxv1i64, 52},
651     {Intrinsic::vp_bitreverse, MVT::nxv2i64, 52},
652     {Intrinsic::vp_bitreverse, MVT::nxv4i64, 52},
653     {Intrinsic::vp_bitreverse, MVT::nxv8i64, 52},
654     {Intrinsic::ctpop, MVT::v2i8, 12},
655     {Intrinsic::ctpop, MVT::v4i8, 12},
656     {Intrinsic::ctpop, MVT::v8i8, 12},
657     {Intrinsic::ctpop, MVT::v16i8, 12},
658     {Intrinsic::ctpop, MVT::nxv1i8, 12},
659     {Intrinsic::ctpop, MVT::nxv2i8, 12},
660     {Intrinsic::ctpop, MVT::nxv4i8, 12},
661     {Intrinsic::ctpop, MVT::nxv8i8, 12},
662     {Intrinsic::ctpop, MVT::nxv16i8, 12},
663     {Intrinsic::ctpop, MVT::v2i16, 19},
664     {Intrinsic::ctpop, MVT::v4i16, 19},
665     {Intrinsic::ctpop, MVT::v8i16, 19},
666     {Intrinsic::ctpop, MVT::v16i16, 19},
667     {Intrinsic::ctpop, MVT::nxv1i16, 19},
668     {Intrinsic::ctpop, MVT::nxv2i16, 19},
669     {Intrinsic::ctpop, MVT::nxv4i16, 19},
670     {Intrinsic::ctpop, MVT::nxv8i16, 19},
671     {Intrinsic::ctpop, MVT::nxv16i16, 19},
672     {Intrinsic::ctpop, MVT::v2i32, 20},
673     {Intrinsic::ctpop, MVT::v4i32, 20},
674     {Intrinsic::ctpop, MVT::v8i32, 20},
675     {Intrinsic::ctpop, MVT::v16i32, 20},
676     {Intrinsic::ctpop, MVT::nxv1i32, 20},
677     {Intrinsic::ctpop, MVT::nxv2i32, 20},
678     {Intrinsic::ctpop, MVT::nxv4i32, 20},
679     {Intrinsic::ctpop, MVT::nxv8i32, 20},
680     {Intrinsic::ctpop, MVT::nxv16i32, 20},
681     {Intrinsic::ctpop, MVT::v2i64, 21},
682     {Intrinsic::ctpop, MVT::v4i64, 21},
683     {Intrinsic::ctpop, MVT::v8i64, 21},
684     {Intrinsic::ctpop, MVT::v16i64, 21},
685     {Intrinsic::ctpop, MVT::nxv1i64, 21},
686     {Intrinsic::ctpop, MVT::nxv2i64, 21},
687     {Intrinsic::ctpop, MVT::nxv4i64, 21},
688     {Intrinsic::ctpop, MVT::nxv8i64, 21},
689     {Intrinsic::vp_ctpop, MVT::v2i8, 12},
690     {Intrinsic::vp_ctpop, MVT::v4i8, 12},
691     {Intrinsic::vp_ctpop, MVT::v8i8, 12},
692     {Intrinsic::vp_ctpop, MVT::v16i8, 12},
693     {Intrinsic::vp_ctpop, MVT::nxv1i8, 12},
694     {Intrinsic::vp_ctpop, MVT::nxv2i8, 12},
695     {Intrinsic::vp_ctpop, MVT::nxv4i8, 12},
696     {Intrinsic::vp_ctpop, MVT::nxv8i8, 12},
697     {Intrinsic::vp_ctpop, MVT::nxv16i8, 12},
698     {Intrinsic::vp_ctpop, MVT::v2i16, 19},
699     {Intrinsic::vp_ctpop, MVT::v4i16, 19},
700     {Intrinsic::vp_ctpop, MVT::v8i16, 19},
701     {Intrinsic::vp_ctpop, MVT::v16i16, 19},
702     {Intrinsic::vp_ctpop, MVT::nxv1i16, 19},
703     {Intrinsic::vp_ctpop, MVT::nxv2i16, 19},
704     {Intrinsic::vp_ctpop, MVT::nxv4i16, 19},
705     {Intrinsic::vp_ctpop, MVT::nxv8i16, 19},
706     {Intrinsic::vp_ctpop, MVT::nxv16i16, 19},
707     {Intrinsic::vp_ctpop, MVT::v2i32, 20},
708     {Intrinsic::vp_ctpop, MVT::v4i32, 20},
709     {Intrinsic::vp_ctpop, MVT::v8i32, 20},
710     {Intrinsic::vp_ctpop, MVT::v16i32, 20},
711     {Intrinsic::vp_ctpop, MVT::nxv1i32, 20},
712     {Intrinsic::vp_ctpop, MVT::nxv2i32, 20},
713     {Intrinsic::vp_ctpop, MVT::nxv4i32, 20},
714     {Intrinsic::vp_ctpop, MVT::nxv8i32, 20},
715     {Intrinsic::vp_ctpop, MVT::nxv16i32, 20},
716     {Intrinsic::vp_ctpop, MVT::v2i64, 21},
717     {Intrinsic::vp_ctpop, MVT::v4i64, 21},
718     {Intrinsic::vp_ctpop, MVT::v8i64, 21},
719     {Intrinsic::vp_ctpop, MVT::v16i64, 21},
720     {Intrinsic::vp_ctpop, MVT::nxv1i64, 21},
721     {Intrinsic::vp_ctpop, MVT::nxv2i64, 21},
722     {Intrinsic::vp_ctpop, MVT::nxv4i64, 21},
723     {Intrinsic::vp_ctpop, MVT::nxv8i64, 21},
724     {Intrinsic::vp_ctlz, MVT::v2i8, 19},
725     {Intrinsic::vp_ctlz, MVT::v4i8, 19},
726     {Intrinsic::vp_ctlz, MVT::v8i8, 19},
727     {Intrinsic::vp_ctlz, MVT::v16i8, 19},
728     {Intrinsic::vp_ctlz, MVT::nxv1i8, 19},
729     {Intrinsic::vp_ctlz, MVT::nxv2i8, 19},
730     {Intrinsic::vp_ctlz, MVT::nxv4i8, 19},
731     {Intrinsic::vp_ctlz, MVT::nxv8i8, 19},
732     {Intrinsic::vp_ctlz, MVT::nxv16i8, 19},
733     {Intrinsic::vp_ctlz, MVT::nxv32i8, 19},
734     {Intrinsic::vp_ctlz, MVT::nxv64i8, 19},
735     {Intrinsic::vp_ctlz, MVT::v2i16, 28},
736     {Intrinsic::vp_ctlz, MVT::v4i16, 28},
737     {Intrinsic::vp_ctlz, MVT::v8i16, 28},
738     {Intrinsic::vp_ctlz, MVT::v16i16, 28},
739     {Intrinsic::vp_ctlz, MVT::nxv1i16, 28},
740     {Intrinsic::vp_ctlz, MVT::nxv2i16, 28},
741     {Intrinsic::vp_ctlz, MVT::nxv4i16, 28},
742     {Intrinsic::vp_ctlz, MVT::nxv8i16, 28},
743     {Intrinsic::vp_ctlz, MVT::nxv16i16, 28},
744     {Intrinsic::vp_ctlz, MVT::nxv32i16, 28},
745     {Intrinsic::vp_ctlz, MVT::v2i32, 31},
746     {Intrinsic::vp_ctlz, MVT::v4i32, 31},
747     {Intrinsic::vp_ctlz, MVT::v8i32, 31},
748     {Intrinsic::vp_ctlz, MVT::v16i32, 31},
749     {Intrinsic::vp_ctlz, MVT::nxv1i32, 31},
750     {Intrinsic::vp_ctlz, MVT::nxv2i32, 31},
751     {Intrinsic::vp_ctlz, MVT::nxv4i32, 31},
752     {Intrinsic::vp_ctlz, MVT::nxv8i32, 31},
753     {Intrinsic::vp_ctlz, MVT::nxv16i32, 31},
754     {Intrinsic::vp_ctlz, MVT::v2i64, 35},
755     {Intrinsic::vp_ctlz, MVT::v4i64, 35},
756     {Intrinsic::vp_ctlz, MVT::v8i64, 35},
757     {Intrinsic::vp_ctlz, MVT::v16i64, 35},
758     {Intrinsic::vp_ctlz, MVT::nxv1i64, 35},
759     {Intrinsic::vp_ctlz, MVT::nxv2i64, 35},
760     {Intrinsic::vp_ctlz, MVT::nxv4i64, 35},
761     {Intrinsic::vp_ctlz, MVT::nxv8i64, 35},
762     {Intrinsic::vp_cttz, MVT::v2i8, 16},
763     {Intrinsic::vp_cttz, MVT::v4i8, 16},
764     {Intrinsic::vp_cttz, MVT::v8i8, 16},
765     {Intrinsic::vp_cttz, MVT::v16i8, 16},
766     {Intrinsic::vp_cttz, MVT::nxv1i8, 16},
767     {Intrinsic::vp_cttz, MVT::nxv2i8, 16},
768     {Intrinsic::vp_cttz, MVT::nxv4i8, 16},
769     {Intrinsic::vp_cttz, MVT::nxv8i8, 16},
770     {Intrinsic::vp_cttz, MVT::nxv16i8, 16},
771     {Intrinsic::vp_cttz, MVT::nxv32i8, 16},
772     {Intrinsic::vp_cttz, MVT::nxv64i8, 16},
773     {Intrinsic::vp_cttz, MVT::v2i16, 23},
774     {Intrinsic::vp_cttz, MVT::v4i16, 23},
775     {Intrinsic::vp_cttz, MVT::v8i16, 23},
776     {Intrinsic::vp_cttz, MVT::v16i16, 23},
777     {Intrinsic::vp_cttz, MVT::nxv1i16, 23},
778     {Intrinsic::vp_cttz, MVT::nxv2i16, 23},
779     {Intrinsic::vp_cttz, MVT::nxv4i16, 23},
780     {Intrinsic::vp_cttz, MVT::nxv8i16, 23},
781     {Intrinsic::vp_cttz, MVT::nxv16i16, 23},
782     {Intrinsic::vp_cttz, MVT::nxv32i16, 23},
783     {Intrinsic::vp_cttz, MVT::v2i32, 24},
784     {Intrinsic::vp_cttz, MVT::v4i32, 24},
785     {Intrinsic::vp_cttz, MVT::v8i32, 24},
786     {Intrinsic::vp_cttz, MVT::v16i32, 24},
787     {Intrinsic::vp_cttz, MVT::nxv1i32, 24},
788     {Intrinsic::vp_cttz, MVT::nxv2i32, 24},
789     {Intrinsic::vp_cttz, MVT::nxv4i32, 24},
790     {Intrinsic::vp_cttz, MVT::nxv8i32, 24},
791     {Intrinsic::vp_cttz, MVT::nxv16i32, 24},
792     {Intrinsic::vp_cttz, MVT::v2i64, 25},
793     {Intrinsic::vp_cttz, MVT::v4i64, 25},
794     {Intrinsic::vp_cttz, MVT::v8i64, 25},
795     {Intrinsic::vp_cttz, MVT::v16i64, 25},
796     {Intrinsic::vp_cttz, MVT::nxv1i64, 25},
797     {Intrinsic::vp_cttz, MVT::nxv2i64, 25},
798     {Intrinsic::vp_cttz, MVT::nxv4i64, 25},
799     {Intrinsic::vp_cttz, MVT::nxv8i64, 25},
800 };
801 
802 static unsigned getISDForVPIntrinsicID(Intrinsic::ID ID) {
803   switch (ID) {
804 #define HELPER_MAP_VPID_TO_VPSD(VPID, VPSD)                                    \
805   case Intrinsic::VPID:                                                        \
806     return ISD::VPSD;
807 #include "llvm/IR/VPIntrinsics.def"
808 #undef HELPER_MAP_VPID_TO_VPSD
809   }
810   return ISD::DELETED_NODE;
811 }
812 
813 InstructionCost
814 RISCVTTIImpl::getIntrinsicInstrCost(const IntrinsicCostAttributes &ICA,
815                                     TTI::TargetCostKind CostKind) {
816   auto *RetTy = ICA.getReturnType();
817   switch (ICA.getID()) {
818   case Intrinsic::ceil:
819   case Intrinsic::floor:
820   case Intrinsic::trunc:
821   case Intrinsic::rint:
822   case Intrinsic::round:
823   case Intrinsic::roundeven: {
824     // These all use the same code.
825     auto LT = getTypeLegalizationCost(RetTy);
826     if (!LT.second.isVector() && TLI->isOperationCustom(ISD::FCEIL, LT.second))
827       return LT.first * 8;
828     break;
829   }
830   case Intrinsic::umin:
831   case Intrinsic::umax:
832   case Intrinsic::smin:
833   case Intrinsic::smax: {
834     auto LT = getTypeLegalizationCost(RetTy);
835     if ((ST->hasVInstructions() && LT.second.isVector()) ||
836         (LT.second.isScalarInteger() && ST->hasStdExtZbb()))
837       return LT.first;
838     break;
839   }
840   case Intrinsic::sadd_sat:
841   case Intrinsic::ssub_sat:
842   case Intrinsic::uadd_sat:
843   case Intrinsic::usub_sat: {
844     auto LT = getTypeLegalizationCost(RetTy);
845     if (ST->hasVInstructions() && LT.second.isVector())
846       return LT.first;
847     break;
848   }
849   case Intrinsic::abs: {
850     auto LT = getTypeLegalizationCost(RetTy);
851     if (ST->hasVInstructions() && LT.second.isVector()) {
852       // vrsub.vi v10, v8, 0
853       // vmax.vv v8, v8, v10
854       return LT.first * 2;
855     }
856     break;
857   }
858   case Intrinsic::fabs:
859   case Intrinsic::sqrt: {
860     auto LT = getTypeLegalizationCost(RetTy);
861     if (ST->hasVInstructions() && LT.second.isVector())
862       return LT.first;
863     break;
864   }
865   // TODO: add more intrinsic
866   case Intrinsic::experimental_stepvector: {
867     unsigned Cost = 1; // vid
868     auto LT = getTypeLegalizationCost(RetTy);
869     return Cost + (LT.first - 1);
870   }
871   case Intrinsic::vp_rint: {
872     // RISC-V target uses at least 5 instructions to lower rounding intrinsics.
873     unsigned Cost = 5;
874     auto LT = getTypeLegalizationCost(RetTy);
875     if (TLI->isOperationCustom(ISD::VP_FRINT, LT.second))
876       return Cost * LT.first;
877     break;
878   }
879   case Intrinsic::vp_nearbyint: {
880     // More one read and one write for fflags than vp_rint.
881     unsigned Cost = 7;
882     auto LT = getTypeLegalizationCost(RetTy);
883     if (TLI->isOperationCustom(ISD::VP_FRINT, LT.second))
884       return Cost * LT.first;
885     break;
886   }
887   case Intrinsic::vp_ceil:
888   case Intrinsic::vp_floor:
889   case Intrinsic::vp_round:
890   case Intrinsic::vp_roundeven:
891   case Intrinsic::vp_roundtozero: {
892     // Rounding with static rounding mode needs two more instructions to
893     // swap/write FRM than vp_rint.
894     unsigned Cost = 7;
895     auto LT = getTypeLegalizationCost(RetTy);
896     unsigned VPISD = getISDForVPIntrinsicID(ICA.getID());
897     if (TLI->isOperationCustom(VPISD, LT.second))
898       return Cost * LT.first;
899     break;
900   }
901   }
902 
903   if (ST->hasVInstructions() && RetTy->isVectorTy()) {
904     auto LT = getTypeLegalizationCost(RetTy);
905     if (const auto *Entry = CostTableLookup(VectorIntrinsicCostTable,
906                                             ICA.getID(), LT.second))
907       return LT.first * Entry->Cost;
908   }
909 
910   return BaseT::getIntrinsicInstrCost(ICA, CostKind);
911 }
912 
913 InstructionCost RISCVTTIImpl::getCastInstrCost(unsigned Opcode, Type *Dst,
914                                                Type *Src,
915                                                TTI::CastContextHint CCH,
916                                                TTI::TargetCostKind CostKind,
917                                                const Instruction *I) {
918   if (isa<VectorType>(Dst) && isa<VectorType>(Src)) {
919     // FIXME: Need to compute legalizing cost for illegal types.
920     if (!isTypeLegal(Src) || !isTypeLegal(Dst))
921       return BaseT::getCastInstrCost(Opcode, Dst, Src, CCH, CostKind, I);
922 
923     // Skip if element size of Dst or Src is bigger than ELEN.
924     if (Src->getScalarSizeInBits() > ST->getELEN() ||
925         Dst->getScalarSizeInBits() > ST->getELEN())
926       return BaseT::getCastInstrCost(Opcode, Dst, Src, CCH, CostKind, I);
927 
928     int ISD = TLI->InstructionOpcodeToISD(Opcode);
929     assert(ISD && "Invalid opcode");
930 
931     // FIXME: Need to consider vsetvli and lmul.
932     int PowDiff = (int)Log2_32(Dst->getScalarSizeInBits()) -
933                   (int)Log2_32(Src->getScalarSizeInBits());
934     switch (ISD) {
935     case ISD::SIGN_EXTEND:
936     case ISD::ZERO_EXTEND:
937       if (Src->getScalarSizeInBits() == 1) {
938         // We do not use vsext/vzext to extend from mask vector.
939         // Instead we use the following instructions to extend from mask vector:
940         // vmv.v.i v8, 0
941         // vmerge.vim v8, v8, -1, v0
942         return 2;
943       }
944       return 1;
945     case ISD::TRUNCATE:
946       if (Dst->getScalarSizeInBits() == 1) {
947         // We do not use several vncvt to truncate to mask vector. So we could
948         // not use PowDiff to calculate it.
949         // Instead we use the following instructions to truncate to mask vector:
950         // vand.vi v8, v8, 1
951         // vmsne.vi v0, v8, 0
952         return 2;
953       }
954       [[fallthrough]];
955     case ISD::FP_EXTEND:
956     case ISD::FP_ROUND:
957       // Counts of narrow/widen instructions.
958       return std::abs(PowDiff);
959     case ISD::FP_TO_SINT:
960     case ISD::FP_TO_UINT:
961     case ISD::SINT_TO_FP:
962     case ISD::UINT_TO_FP:
963       if (Src->getScalarSizeInBits() == 1 || Dst->getScalarSizeInBits() == 1) {
964         // The cost of convert from or to mask vector is different from other
965         // cases. We could not use PowDiff to calculate it.
966         // For mask vector to fp, we should use the following instructions:
967         // vmv.v.i v8, 0
968         // vmerge.vim v8, v8, -1, v0
969         // vfcvt.f.x.v v8, v8
970 
971         // And for fp vector to mask, we use:
972         // vfncvt.rtz.x.f.w v9, v8
973         // vand.vi v8, v9, 1
974         // vmsne.vi v0, v8, 0
975         return 3;
976       }
977       if (std::abs(PowDiff) <= 1)
978         return 1;
979       // Backend could lower (v[sz]ext i8 to double) to vfcvt(v[sz]ext.f8 i8),
980       // so it only need two conversion.
981       if (Src->isIntOrIntVectorTy())
982         return 2;
983       // Counts of narrow/widen instructions.
984       return std::abs(PowDiff);
985     }
986   }
987   return BaseT::getCastInstrCost(Opcode, Dst, Src, CCH, CostKind, I);
988 }
989 
990 unsigned RISCVTTIImpl::getEstimatedVLFor(VectorType *Ty) {
991   if (isa<ScalableVectorType>(Ty)) {
992     const unsigned EltSize = DL.getTypeSizeInBits(Ty->getElementType());
993     const unsigned MinSize = DL.getTypeSizeInBits(Ty).getKnownMinValue();
994     const unsigned VectorBits = *getVScaleForTuning() * RISCV::RVVBitsPerBlock;
995     return RISCVTargetLowering::computeVLMAX(VectorBits, EltSize, MinSize);
996   }
997   return cast<FixedVectorType>(Ty)->getNumElements();
998 }
999 
1000 InstructionCost
1001 RISCVTTIImpl::getMinMaxReductionCost(VectorType *Ty, VectorType *CondTy,
1002                                      bool IsUnsigned,
1003                                      TTI::TargetCostKind CostKind) {
1004   if (isa<FixedVectorType>(Ty) && !ST->useRVVForFixedLengthVectors())
1005     return BaseT::getMinMaxReductionCost(Ty, CondTy, IsUnsigned, CostKind);
1006 
1007   // Skip if scalar size of Ty is bigger than ELEN.
1008   if (Ty->getScalarSizeInBits() > ST->getELEN())
1009     return BaseT::getMinMaxReductionCost(Ty, CondTy, IsUnsigned, CostKind);
1010 
1011   std::pair<InstructionCost, MVT> LT = getTypeLegalizationCost(Ty);
1012   if (Ty->getElementType()->isIntegerTy(1))
1013     // vcpop sequences, see vreduction-mask.ll.  umax, smin actually only
1014     // cost 2, but we don't have enough info here so we slightly over cost.
1015     return (LT.first - 1) + 3;
1016 
1017   // IR Reduction is composed by two vmv and one rvv reduction instruction.
1018   InstructionCost BaseCost = 2;
1019   unsigned VL = getEstimatedVLFor(Ty);
1020   return (LT.first - 1) + BaseCost + Log2_32_Ceil(VL);
1021 }
1022 
1023 InstructionCost
1024 RISCVTTIImpl::getArithmeticReductionCost(unsigned Opcode, VectorType *Ty,
1025                                          std::optional<FastMathFlags> FMF,
1026                                          TTI::TargetCostKind CostKind) {
1027   if (isa<FixedVectorType>(Ty) && !ST->useRVVForFixedLengthVectors())
1028     return BaseT::getArithmeticReductionCost(Opcode, Ty, FMF, CostKind);
1029 
1030   // Skip if scalar size of Ty is bigger than ELEN.
1031   if (Ty->getScalarSizeInBits() > ST->getELEN())
1032     return BaseT::getArithmeticReductionCost(Opcode, Ty, FMF, CostKind);
1033 
1034   int ISD = TLI->InstructionOpcodeToISD(Opcode);
1035   assert(ISD && "Invalid opcode");
1036 
1037   if (ISD != ISD::ADD && ISD != ISD::OR && ISD != ISD::XOR && ISD != ISD::AND &&
1038       ISD != ISD::FADD)
1039     return BaseT::getArithmeticReductionCost(Opcode, Ty, FMF, CostKind);
1040 
1041   std::pair<InstructionCost, MVT> LT = getTypeLegalizationCost(Ty);
1042   if (Ty->getElementType()->isIntegerTy(1))
1043     // vcpop sequences, see vreduction-mask.ll
1044     return (LT.first - 1) + (ISD == ISD::AND ? 3 : 2);
1045 
1046   // IR Reduction is composed by two vmv and one rvv reduction instruction.
1047   InstructionCost BaseCost = 2;
1048   unsigned VL = getEstimatedVLFor(Ty);
1049   if (TTI::requiresOrderedReduction(FMF))
1050     return (LT.first - 1) + BaseCost + VL;
1051   return (LT.first - 1) + BaseCost + Log2_32_Ceil(VL);
1052 }
1053 
1054 InstructionCost RISCVTTIImpl::getExtendedReductionCost(
1055     unsigned Opcode, bool IsUnsigned, Type *ResTy, VectorType *ValTy,
1056     std::optional<FastMathFlags> FMF, TTI::TargetCostKind CostKind) {
1057   if (isa<FixedVectorType>(ValTy) && !ST->useRVVForFixedLengthVectors())
1058     return BaseT::getExtendedReductionCost(Opcode, IsUnsigned, ResTy, ValTy,
1059                                            FMF, CostKind);
1060 
1061   // Skip if scalar size of ResTy is bigger than ELEN.
1062   if (ResTy->getScalarSizeInBits() > ST->getELEN())
1063     return BaseT::getExtendedReductionCost(Opcode, IsUnsigned, ResTy, ValTy,
1064                                            FMF, CostKind);
1065 
1066   if (Opcode != Instruction::Add && Opcode != Instruction::FAdd)
1067     return BaseT::getExtendedReductionCost(Opcode, IsUnsigned, ResTy, ValTy,
1068                                            FMF, CostKind);
1069 
1070   std::pair<InstructionCost, MVT> LT = getTypeLegalizationCost(ValTy);
1071 
1072   if (ResTy->getScalarSizeInBits() != 2 * LT.second.getScalarSizeInBits())
1073     return BaseT::getExtendedReductionCost(Opcode, IsUnsigned, ResTy, ValTy,
1074                                            FMF, CostKind);
1075 
1076   return (LT.first - 1) +
1077          getArithmeticReductionCost(Opcode, ValTy, FMF, CostKind);
1078 }
1079 
1080 InstructionCost RISCVTTIImpl::getStoreImmCost(Type *Ty,
1081                                               TTI::OperandValueInfo OpInfo,
1082                                               TTI::TargetCostKind CostKind) {
1083   assert(OpInfo.isConstant() && "non constant operand?");
1084   if (!isa<VectorType>(Ty))
1085     // FIXME: We need to account for immediate materialization here, but doing
1086     // a decent job requires more knowledge about the immediate than we
1087     // currently have here.
1088     return 0;
1089 
1090   if (OpInfo.isUniform())
1091     // vmv.x.i, vmv.v.x, or vfmv.v.f
1092     // We ignore the cost of the scalar constant materialization to be consistent
1093     // with how we treat scalar constants themselves just above.
1094     return 1;
1095 
1096   // Add a cost of address generation + the cost of the vector load. The
1097   // address is expected to be a PC relative offset to a constant pool entry
1098   // using auipc/addi.
1099   return 2 + getMemoryOpCost(Instruction::Load, Ty, DL.getABITypeAlign(Ty),
1100                              /*AddressSpace=*/0, CostKind);
1101 }
1102 
1103 
1104 InstructionCost RISCVTTIImpl::getMemoryOpCost(unsigned Opcode, Type *Src,
1105                                               MaybeAlign Alignment,
1106                                               unsigned AddressSpace,
1107                                               TTI::TargetCostKind CostKind,
1108                                               TTI::OperandValueInfo OpInfo,
1109                                               const Instruction *I) {
1110   InstructionCost Cost = 0;
1111   if (Opcode == Instruction::Store && OpInfo.isConstant())
1112     Cost += getStoreImmCost(Src, OpInfo, CostKind);
1113   return Cost + BaseT::getMemoryOpCost(Opcode, Src, Alignment, AddressSpace,
1114                                        CostKind, OpInfo, I);
1115 }
1116 
1117 InstructionCost RISCVTTIImpl::getCmpSelInstrCost(unsigned Opcode, Type *ValTy,
1118                                                  Type *CondTy,
1119                                                  CmpInst::Predicate VecPred,
1120                                                  TTI::TargetCostKind CostKind,
1121                                                  const Instruction *I) {
1122   if (CostKind != TTI::TCK_RecipThroughput)
1123     return BaseT::getCmpSelInstrCost(Opcode, ValTy, CondTy, VecPred, CostKind,
1124                                      I);
1125 
1126   if (isa<FixedVectorType>(ValTy) && !ST->useRVVForFixedLengthVectors())
1127     return BaseT::getCmpSelInstrCost(Opcode, ValTy, CondTy, VecPred, CostKind,
1128                                      I);
1129 
1130   // Skip if scalar size of ValTy is bigger than ELEN.
1131   if (ValTy->isVectorTy() && ValTy->getScalarSizeInBits() > ST->getELEN())
1132     return BaseT::getCmpSelInstrCost(Opcode, ValTy, CondTy, VecPred, CostKind,
1133                                      I);
1134 
1135   if (Opcode == Instruction::Select && ValTy->isVectorTy()) {
1136     std::pair<InstructionCost, MVT> LT = getTypeLegalizationCost(ValTy);
1137     if (CondTy->isVectorTy()) {
1138       if (ValTy->getScalarSizeInBits() == 1) {
1139         // vmandn.mm v8, v8, v9
1140         // vmand.mm v9, v0, v9
1141         // vmor.mm v0, v9, v8
1142         return LT.first * 3;
1143       }
1144       // vselect and max/min are supported natively.
1145       return LT.first * 1;
1146     }
1147 
1148     if (ValTy->getScalarSizeInBits() == 1) {
1149       //  vmv.v.x v9, a0
1150       //  vmsne.vi v9, v9, 0
1151       //  vmandn.mm v8, v8, v9
1152       //  vmand.mm v9, v0, v9
1153       //  vmor.mm v0, v9, v8
1154       return LT.first * 5;
1155     }
1156 
1157     // vmv.v.x v10, a0
1158     // vmsne.vi v0, v10, 0
1159     // vmerge.vvm v8, v9, v8, v0
1160     return LT.first * 3;
1161   }
1162 
1163   if ((Opcode == Instruction::ICmp || Opcode == Instruction::FCmp) &&
1164       ValTy->isVectorTy()) {
1165     std::pair<InstructionCost, MVT> LT = getTypeLegalizationCost(ValTy);
1166 
1167     // Support natively.
1168     if (CmpInst::isIntPredicate(VecPred))
1169       return LT.first * 1;
1170 
1171     // If we do not support the input floating point vector type, use the base
1172     // one which will calculate as:
1173     // ScalarizeCost + Num * Cost for fixed vector,
1174     // InvalidCost for scalable vector.
1175     if ((ValTy->getScalarSizeInBits() == 16 && !ST->hasVInstructionsF16()) ||
1176         (ValTy->getScalarSizeInBits() == 32 && !ST->hasVInstructionsF32()) ||
1177         (ValTy->getScalarSizeInBits() == 64 && !ST->hasVInstructionsF64()))
1178       return BaseT::getCmpSelInstrCost(Opcode, ValTy, CondTy, VecPred, CostKind,
1179                                        I);
1180     switch (VecPred) {
1181       // Support natively.
1182     case CmpInst::FCMP_OEQ:
1183     case CmpInst::FCMP_OGT:
1184     case CmpInst::FCMP_OGE:
1185     case CmpInst::FCMP_OLT:
1186     case CmpInst::FCMP_OLE:
1187     case CmpInst::FCMP_UNE:
1188       return LT.first * 1;
1189     // TODO: Other comparisons?
1190     default:
1191       break;
1192     }
1193   }
1194 
1195   // TODO: Add cost for scalar type.
1196 
1197   return BaseT::getCmpSelInstrCost(Opcode, ValTy, CondTy, VecPred, CostKind, I);
1198 }
1199 
1200 InstructionCost RISCVTTIImpl::getVectorInstrCost(unsigned Opcode, Type *Val,
1201                                                  TTI::TargetCostKind CostKind,
1202                                                  unsigned Index, Value *Op0,
1203                                                  Value *Op1) {
1204   assert(Val->isVectorTy() && "This must be a vector type");
1205 
1206   if (Opcode != Instruction::ExtractElement &&
1207       Opcode != Instruction::InsertElement)
1208     return BaseT::getVectorInstrCost(Opcode, Val, CostKind, Index, Op0, Op1);
1209 
1210   // Legalize the type.
1211   std::pair<InstructionCost, MVT> LT = getTypeLegalizationCost(Val);
1212 
1213   // This type is legalized to a scalar type.
1214   if (!LT.second.isVector())
1215     return 0;
1216 
1217   // For unsupported scalable vector.
1218   if (LT.second.isScalableVector() && !LT.first.isValid())
1219     return LT.first;
1220 
1221   if (!isTypeLegal(Val))
1222     return BaseT::getVectorInstrCost(Opcode, Val, CostKind, Index, Op0, Op1);
1223 
1224   // In RVV, we could use vslidedown + vmv.x.s to extract element from vector
1225   // and vslideup + vmv.s.x to insert element to vector.
1226   unsigned BaseCost = 1;
1227   // When insertelement we should add the index with 1 as the input of vslideup.
1228   unsigned SlideCost = Opcode == Instruction::InsertElement ? 2 : 1;
1229 
1230   if (Index != -1U) {
1231     // The type may be split. For fixed-width vectors we can normalize the
1232     // index to the new type.
1233     if (LT.second.isFixedLengthVector()) {
1234       unsigned Width = LT.second.getVectorNumElements();
1235       Index = Index % Width;
1236     }
1237 
1238     // We could extract/insert the first element without vslidedown/vslideup.
1239     if (Index == 0)
1240       SlideCost = 0;
1241     else if (Opcode == Instruction::InsertElement)
1242       SlideCost = 1; // With a constant index, we do not need to use addi.
1243   }
1244 
1245   // Mask vector extract/insert element is different from normal case.
1246   if (Val->getScalarSizeInBits() == 1) {
1247     // For extractelement, we need the following instructions:
1248     // vmv.v.i v8, 0
1249     // vmerge.vim v8, v8, 1, v0
1250     // vsetivli zero, 1, e8, m2, ta, mu (not count)
1251     // vslidedown.vx v8, v8, a0
1252     // vmv.x.s a0, v8
1253 
1254     // For insertelement, we need the following instructions:
1255     // vsetvli a2, zero, e8, m1, ta, mu (not count)
1256     // vmv.s.x v8, a0
1257     // vmv.v.i v9, 0
1258     // vmerge.vim v9, v9, 1, v0
1259     // addi a0, a1, 1
1260     // vsetvli zero, a0, e8, m1, tu, mu (not count)
1261     // vslideup.vx v9, v8, a1
1262     // vsetvli a0, zero, e8, m1, ta, mu (not count)
1263     // vand.vi v8, v9, 1
1264     // vmsne.vi v0, v8, 0
1265 
1266     // TODO: should we count these special vsetvlis?
1267     BaseCost = Opcode == Instruction::InsertElement ? 5 : 3;
1268   }
1269   // Extract i64 in the target that has XLEN=32 need more instruction.
1270   if (Val->getScalarType()->isIntegerTy() &&
1271       ST->getXLen() < Val->getScalarSizeInBits()) {
1272     // For extractelement, we need the following instructions:
1273     // vsetivli zero, 1, e64, m1, ta, mu (not count)
1274     // vslidedown.vx v8, v8, a0
1275     // vmv.x.s a0, v8
1276     // li a1, 32
1277     // vsrl.vx v8, v8, a1
1278     // vmv.x.s a1, v8
1279 
1280     // For insertelement, we need the following instructions:
1281     // vsetivli zero, 2, e32, m4, ta, mu (not count)
1282     // vmv.v.i v12, 0
1283     // vslide1up.vx v16, v12, a1
1284     // vslide1up.vx v12, v16, a0
1285     // addi a0, a2, 1
1286     // vsetvli zero, a0, e64, m4, tu, mu (not count)
1287     // vslideup.vx v8, v12, a2
1288 
1289     // TODO: should we count these special vsetvlis?
1290     BaseCost = Opcode == Instruction::InsertElement ? 3 : 4;
1291   }
1292   return BaseCost + SlideCost;
1293 }
1294 
1295 InstructionCost RISCVTTIImpl::getArithmeticInstrCost(
1296     unsigned Opcode, Type *Ty, TTI::TargetCostKind CostKind,
1297     TTI::OperandValueInfo Op1Info, TTI::OperandValueInfo Op2Info,
1298     ArrayRef<const Value *> Args, const Instruction *CxtI) {
1299 
1300   // TODO: Handle more cost kinds.
1301   if (CostKind != TTI::TCK_RecipThroughput)
1302     return BaseT::getArithmeticInstrCost(Opcode, Ty, CostKind, Op1Info, Op2Info,
1303                                          Args, CxtI);
1304 
1305   if (isa<FixedVectorType>(Ty) && !ST->useRVVForFixedLengthVectors())
1306     return BaseT::getArithmeticInstrCost(Opcode, Ty, CostKind, Op1Info, Op2Info,
1307                                          Args, CxtI);
1308 
1309   // Skip if scalar size of Ty is bigger than ELEN.
1310   if (isa<VectorType>(Ty) && Ty->getScalarSizeInBits() > ST->getELEN())
1311     return BaseT::getArithmeticInstrCost(Opcode, Ty, CostKind, Op1Info, Op2Info,
1312                                          Args, CxtI);
1313 
1314   // Legalize the type.
1315   std::pair<InstructionCost, MVT> LT = getTypeLegalizationCost(Ty);
1316 
1317   // TODO: Handle scalar type.
1318   if (!LT.second.isVector())
1319     return BaseT::getArithmeticInstrCost(Opcode, Ty, CostKind, Op1Info, Op2Info,
1320                                          Args, CxtI);
1321 
1322 
1323   auto getConstantMatCost =
1324     [&](unsigned Operand, TTI::OperandValueInfo OpInfo) -> InstructionCost {
1325     if (OpInfo.isUniform() && TLI->canSplatOperand(Opcode, Operand))
1326       // Two sub-cases:
1327       // * Has a 5 bit immediate operand which can be splatted.
1328       // * Has a larger immediate which must be materialized in scalar register
1329       // We return 0 for both as we currently ignore the cost of materializing
1330       // scalar constants in GPRs.
1331       return 0;
1332 
1333     // Add a cost of address generation + the cost of the vector load. The
1334     // address is expected to be a PC relative offset to a constant pool entry
1335     // using auipc/addi.
1336     return 2 + getMemoryOpCost(Instruction::Load, Ty, DL.getABITypeAlign(Ty),
1337                                /*AddressSpace=*/0, CostKind);
1338   };
1339 
1340   // Add the cost of materializing any constant vectors required.
1341   InstructionCost ConstantMatCost = 0;
1342   if (Op1Info.isConstant())
1343     ConstantMatCost += getConstantMatCost(0, Op1Info);
1344   if (Op2Info.isConstant())
1345     ConstantMatCost += getConstantMatCost(1, Op2Info);
1346 
1347   switch (TLI->InstructionOpcodeToISD(Opcode)) {
1348   case ISD::ADD:
1349   case ISD::SUB:
1350   case ISD::AND:
1351   case ISD::OR:
1352   case ISD::XOR:
1353   case ISD::SHL:
1354   case ISD::SRL:
1355   case ISD::SRA:
1356   case ISD::MUL:
1357   case ISD::MULHS:
1358   case ISD::MULHU:
1359   case ISD::FADD:
1360   case ISD::FSUB:
1361   case ISD::FMUL:
1362   case ISD::FNEG: {
1363     return ConstantMatCost + getLMULCost(LT.second) * LT.first * 1;
1364   }
1365   default:
1366     return ConstantMatCost +
1367            BaseT::getArithmeticInstrCost(Opcode, Ty, CostKind, Op1Info, Op2Info,
1368                                          Args, CxtI);
1369   }
1370 }
1371 
1372 void RISCVTTIImpl::getUnrollingPreferences(Loop *L, ScalarEvolution &SE,
1373                                            TTI::UnrollingPreferences &UP,
1374                                            OptimizationRemarkEmitter *ORE) {
1375   // TODO: More tuning on benchmarks and metrics with changes as needed
1376   //       would apply to all settings below to enable performance.
1377 
1378 
1379   if (ST->enableDefaultUnroll())
1380     return BasicTTIImplBase::getUnrollingPreferences(L, SE, UP, ORE);
1381 
1382   // Enable Upper bound unrolling universally, not dependant upon the conditions
1383   // below.
1384   UP.UpperBound = true;
1385 
1386   // Disable loop unrolling for Oz and Os.
1387   UP.OptSizeThreshold = 0;
1388   UP.PartialOptSizeThreshold = 0;
1389   if (L->getHeader()->getParent()->hasOptSize())
1390     return;
1391 
1392   SmallVector<BasicBlock *, 4> ExitingBlocks;
1393   L->getExitingBlocks(ExitingBlocks);
1394   LLVM_DEBUG(dbgs() << "Loop has:\n"
1395                     << "Blocks: " << L->getNumBlocks() << "\n"
1396                     << "Exit blocks: " << ExitingBlocks.size() << "\n");
1397 
1398   // Only allow another exit other than the latch. This acts as an early exit
1399   // as it mirrors the profitability calculation of the runtime unroller.
1400   if (ExitingBlocks.size() > 2)
1401     return;
1402 
1403   // Limit the CFG of the loop body for targets with a branch predictor.
1404   // Allowing 4 blocks permits if-then-else diamonds in the body.
1405   if (L->getNumBlocks() > 4)
1406     return;
1407 
1408   // Don't unroll vectorized loops, including the remainder loop
1409   if (getBooleanLoopAttribute(L, "llvm.loop.isvectorized"))
1410     return;
1411 
1412   // Scan the loop: don't unroll loops with calls as this could prevent
1413   // inlining.
1414   InstructionCost Cost = 0;
1415   for (auto *BB : L->getBlocks()) {
1416     for (auto &I : *BB) {
1417       // Initial setting - Don't unroll loops containing vectorized
1418       // instructions.
1419       if (I.getType()->isVectorTy())
1420         return;
1421 
1422       if (isa<CallInst>(I) || isa<InvokeInst>(I)) {
1423         if (const Function *F = cast<CallBase>(I).getCalledFunction()) {
1424           if (!isLoweredToCall(F))
1425             continue;
1426         }
1427         return;
1428       }
1429 
1430       SmallVector<const Value *> Operands(I.operand_values());
1431       Cost += getInstructionCost(&I, Operands,
1432                                  TargetTransformInfo::TCK_SizeAndLatency);
1433     }
1434   }
1435 
1436   LLVM_DEBUG(dbgs() << "Cost of loop: " << Cost << "\n");
1437 
1438   UP.Partial = true;
1439   UP.Runtime = true;
1440   UP.UnrollRemainder = true;
1441   UP.UnrollAndJam = true;
1442   UP.UnrollAndJamInnerLoopThreshold = 60;
1443 
1444   // Force unrolling small loops can be very useful because of the branch
1445   // taken cost of the backedge.
1446   if (Cost < 12)
1447     UP.Force = true;
1448 }
1449 
1450 void RISCVTTIImpl::getPeelingPreferences(Loop *L, ScalarEvolution &SE,
1451                                          TTI::PeelingPreferences &PP) {
1452   BaseT::getPeelingPreferences(L, SE, PP);
1453 }
1454 
1455 unsigned RISCVTTIImpl::getRegUsageForType(Type *Ty) {
1456   TypeSize Size = DL.getTypeSizeInBits(Ty);
1457   if (Ty->isVectorTy()) {
1458     if (Size.isScalable() && ST->hasVInstructions())
1459       return divideCeil(Size.getKnownMinValue(), RISCV::RVVBitsPerBlock);
1460 
1461     if (ST->useRVVForFixedLengthVectors())
1462       return divideCeil(Size, ST->getRealMinVLen());
1463   }
1464 
1465   return BaseT::getRegUsageForType(Ty);
1466 }
1467 
1468 unsigned RISCVTTIImpl::getMaximumVF(unsigned ElemWidth, unsigned Opcode) const {
1469   // This interface is currently only used by SLP.  Returning 1 (which is the
1470   // default value for SLPMaxVF) disables SLP. We currently have a cost modeling
1471   // problem w/ constant materialization which causes SLP to perform majorly
1472   // unprofitable transformations.
1473   // TODO: Figure out constant materialization cost modeling and remove.
1474   return SLPMaxVF;
1475 }
1476 
1477 bool RISCVTTIImpl::isLSRCostLess(const TargetTransformInfo::LSRCost &C1,
1478                                  const TargetTransformInfo::LSRCost &C2) {
1479   // RISCV specific here are "instruction number 1st priority".
1480   return std::tie(C1.Insns, C1.NumRegs, C1.AddRecCost,
1481                   C1.NumIVMuls, C1.NumBaseAdds,
1482                   C1.ScaleCost, C1.ImmCost, C1.SetupCost) <
1483          std::tie(C2.Insns, C2.NumRegs, C2.AddRecCost,
1484                   C2.NumIVMuls, C2.NumBaseAdds,
1485                   C2.ScaleCost, C2.ImmCost, C2.SetupCost);
1486 }
1487