1 //===-- RISCVISelDAGToDAG.cpp - A dag to dag inst selector for RISCV ------===//
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 // This file defines an instruction selector for the RISCV target.
10 //
11 //===----------------------------------------------------------------------===//
12 
13 #include "RISCVISelDAGToDAG.h"
14 #include "MCTargetDesc/RISCVMCTargetDesc.h"
15 #include "MCTargetDesc/RISCVMatInt.h"
16 #include "RISCVISelLowering.h"
17 #include "RISCVMachineFunctionInfo.h"
18 #include "llvm/CodeGen/MachineFrameInfo.h"
19 #include "llvm/IR/IntrinsicsRISCV.h"
20 #include "llvm/Support/Alignment.h"
21 #include "llvm/Support/Debug.h"
22 #include "llvm/Support/KnownBits.h"
23 #include "llvm/Support/MathExtras.h"
24 #include "llvm/Support/raw_ostream.h"
25 
26 using namespace llvm;
27 
28 #define DEBUG_TYPE "riscv-isel"
29 
30 namespace llvm {
31 namespace RISCV {
32 #define GET_RISCVVSSEGTable_IMPL
33 #define GET_RISCVVLSEGTable_IMPL
34 #define GET_RISCVVLXSEGTable_IMPL
35 #define GET_RISCVVSXSEGTable_IMPL
36 #define GET_RISCVVLETable_IMPL
37 #define GET_RISCVVSETable_IMPL
38 #define GET_RISCVVLXTable_IMPL
39 #define GET_RISCVVSXTable_IMPL
40 #define GET_RISCVMaskedPseudosTable_IMPL
41 #include "RISCVGenSearchableTables.inc"
42 } // namespace RISCV
43 } // namespace llvm
44 
45 void RISCVDAGToDAGISel::PreprocessISelDAG() {
46   SelectionDAG::allnodes_iterator Position = CurDAG->allnodes_end();
47 
48   bool MadeChange = false;
49   while (Position != CurDAG->allnodes_begin()) {
50     SDNode *N = &*--Position;
51     if (N->use_empty())
52       continue;
53 
54     SDValue Result;
55     switch (N->getOpcode()) {
56     case ISD::SPLAT_VECTOR: {
57       // Convert integer SPLAT_VECTOR to VMV_V_X_VL and floating-point
58       // SPLAT_VECTOR to VFMV_V_F_VL to reduce isel burden.
59       MVT VT = N->getSimpleValueType(0);
60       unsigned Opc =
61           VT.isInteger() ? RISCVISD::VMV_V_X_VL : RISCVISD::VFMV_V_F_VL;
62       SDLoc DL(N);
63       SDValue VL = CurDAG->getRegister(RISCV::X0, Subtarget->getXLenVT());
64       Result = CurDAG->getNode(Opc, DL, VT, CurDAG->getUNDEF(VT),
65                                N->getOperand(0), VL);
66       break;
67     }
68     case RISCVISD::SPLAT_VECTOR_SPLIT_I64_VL: {
69       // Lower SPLAT_VECTOR_SPLIT_I64 to two scalar stores and a stride 0 vector
70       // load. Done after lowering and combining so that we have a chance to
71       // optimize this to VMV_V_X_VL when the upper bits aren't needed.
72       assert(N->getNumOperands() == 4 && "Unexpected number of operands");
73       MVT VT = N->getSimpleValueType(0);
74       SDValue Passthru = N->getOperand(0);
75       SDValue Lo = N->getOperand(1);
76       SDValue Hi = N->getOperand(2);
77       SDValue VL = N->getOperand(3);
78       assert(VT.getVectorElementType() == MVT::i64 && VT.isScalableVector() &&
79              Lo.getValueType() == MVT::i32 && Hi.getValueType() == MVT::i32 &&
80              "Unexpected VTs!");
81       MachineFunction &MF = CurDAG->getMachineFunction();
82       RISCVMachineFunctionInfo *FuncInfo =
83           MF.getInfo<RISCVMachineFunctionInfo>();
84       SDLoc DL(N);
85 
86       // We use the same frame index we use for moving two i32s into 64-bit FPR.
87       // This is an analogous operation.
88       int FI = FuncInfo->getMoveF64FrameIndex(MF);
89       MachinePointerInfo MPI = MachinePointerInfo::getFixedStack(MF, FI);
90       const TargetLowering &TLI = CurDAG->getTargetLoweringInfo();
91       SDValue StackSlot =
92           CurDAG->getFrameIndex(FI, TLI.getPointerTy(CurDAG->getDataLayout()));
93 
94       SDValue Chain = CurDAG->getEntryNode();
95       Lo = CurDAG->getStore(Chain, DL, Lo, StackSlot, MPI, Align(8));
96 
97       SDValue OffsetSlot =
98           CurDAG->getMemBasePlusOffset(StackSlot, TypeSize::Fixed(4), DL);
99       Hi = CurDAG->getStore(Chain, DL, Hi, OffsetSlot, MPI.getWithOffset(4),
100                             Align(8));
101 
102       Chain = CurDAG->getNode(ISD::TokenFactor, DL, MVT::Other, Lo, Hi);
103 
104       SDVTList VTs = CurDAG->getVTList({VT, MVT::Other});
105       SDValue IntID =
106           CurDAG->getTargetConstant(Intrinsic::riscv_vlse, DL, MVT::i64);
107       SDValue Ops[] = {Chain,
108                        IntID,
109                        Passthru,
110                        StackSlot,
111                        CurDAG->getRegister(RISCV::X0, MVT::i64),
112                        VL};
113 
114       Result = CurDAG->getMemIntrinsicNode(ISD::INTRINSIC_W_CHAIN, DL, VTs, Ops,
115                                            MVT::i64, MPI, Align(8),
116                                            MachineMemOperand::MOLoad);
117       break;
118     }
119     }
120 
121     if (Result) {
122       LLVM_DEBUG(dbgs() << "RISCV DAG preprocessing replacing:\nOld:    ");
123       LLVM_DEBUG(N->dump(CurDAG));
124       LLVM_DEBUG(dbgs() << "\nNew: ");
125       LLVM_DEBUG(Result->dump(CurDAG));
126       LLVM_DEBUG(dbgs() << "\n");
127 
128       CurDAG->ReplaceAllUsesOfValueWith(SDValue(N, 0), Result);
129       MadeChange = true;
130     }
131   }
132 
133   if (MadeChange)
134     CurDAG->RemoveDeadNodes();
135 }
136 
137 void RISCVDAGToDAGISel::PostprocessISelDAG() {
138   HandleSDNode Dummy(CurDAG->getRoot());
139   SelectionDAG::allnodes_iterator Position = CurDAG->allnodes_end();
140 
141   bool MadeChange = false;
142   while (Position != CurDAG->allnodes_begin()) {
143     SDNode *N = &*--Position;
144     // Skip dead nodes and any non-machine opcodes.
145     if (N->use_empty() || !N->isMachineOpcode())
146       continue;
147 
148     MadeChange |= doPeepholeSExtW(N);
149     MadeChange |= doPeepholeMaskedRVV(N);
150   }
151 
152   CurDAG->setRoot(Dummy.getValue());
153 
154   if (MadeChange)
155     CurDAG->RemoveDeadNodes();
156 }
157 
158 static SDNode *selectImmSeq(SelectionDAG *CurDAG, const SDLoc &DL, const MVT VT,
159                             RISCVMatInt::InstSeq &Seq) {
160   SDNode *Result = nullptr;
161   SDValue SrcReg = CurDAG->getRegister(RISCV::X0, VT);
162   for (RISCVMatInt::Inst &Inst : Seq) {
163     SDValue SDImm = CurDAG->getTargetConstant(Inst.Imm, DL, VT);
164     switch (Inst.getOpndKind()) {
165     case RISCVMatInt::Imm:
166       Result = CurDAG->getMachineNode(Inst.Opc, DL, VT, SDImm);
167       break;
168     case RISCVMatInt::RegX0:
169       Result = CurDAG->getMachineNode(Inst.Opc, DL, VT, SrcReg,
170                                       CurDAG->getRegister(RISCV::X0, VT));
171       break;
172     case RISCVMatInt::RegReg:
173       Result = CurDAG->getMachineNode(Inst.Opc, DL, VT, SrcReg, SrcReg);
174       break;
175     case RISCVMatInt::RegImm:
176       Result = CurDAG->getMachineNode(Inst.Opc, DL, VT, SrcReg, SDImm);
177       break;
178     }
179 
180     // Only the first instruction has X0 as its source.
181     SrcReg = SDValue(Result, 0);
182   }
183 
184   return Result;
185 }
186 
187 static SDNode *selectImm(SelectionDAG *CurDAG, const SDLoc &DL, const MVT VT,
188                          int64_t Imm, const RISCVSubtarget &Subtarget) {
189   RISCVMatInt::InstSeq Seq =
190       RISCVMatInt::generateInstSeq(Imm, Subtarget.getFeatureBits());
191 
192   return selectImmSeq(CurDAG, DL, VT, Seq);
193 }
194 
195 static SDValue createTuple(SelectionDAG &CurDAG, ArrayRef<SDValue> Regs,
196                            unsigned NF, RISCVII::VLMUL LMUL) {
197   static const unsigned M1TupleRegClassIDs[] = {
198       RISCV::VRN2M1RegClassID, RISCV::VRN3M1RegClassID, RISCV::VRN4M1RegClassID,
199       RISCV::VRN5M1RegClassID, RISCV::VRN6M1RegClassID, RISCV::VRN7M1RegClassID,
200       RISCV::VRN8M1RegClassID};
201   static const unsigned M2TupleRegClassIDs[] = {RISCV::VRN2M2RegClassID,
202                                                 RISCV::VRN3M2RegClassID,
203                                                 RISCV::VRN4M2RegClassID};
204 
205   assert(Regs.size() >= 2 && Regs.size() <= 8);
206 
207   unsigned RegClassID;
208   unsigned SubReg0;
209   switch (LMUL) {
210   default:
211     llvm_unreachable("Invalid LMUL.");
212   case RISCVII::VLMUL::LMUL_F8:
213   case RISCVII::VLMUL::LMUL_F4:
214   case RISCVII::VLMUL::LMUL_F2:
215   case RISCVII::VLMUL::LMUL_1:
216     static_assert(RISCV::sub_vrm1_7 == RISCV::sub_vrm1_0 + 7,
217                   "Unexpected subreg numbering");
218     SubReg0 = RISCV::sub_vrm1_0;
219     RegClassID = M1TupleRegClassIDs[NF - 2];
220     break;
221   case RISCVII::VLMUL::LMUL_2:
222     static_assert(RISCV::sub_vrm2_3 == RISCV::sub_vrm2_0 + 3,
223                   "Unexpected subreg numbering");
224     SubReg0 = RISCV::sub_vrm2_0;
225     RegClassID = M2TupleRegClassIDs[NF - 2];
226     break;
227   case RISCVII::VLMUL::LMUL_4:
228     static_assert(RISCV::sub_vrm4_1 == RISCV::sub_vrm4_0 + 1,
229                   "Unexpected subreg numbering");
230     SubReg0 = RISCV::sub_vrm4_0;
231     RegClassID = RISCV::VRN2M4RegClassID;
232     break;
233   }
234 
235   SDLoc DL(Regs[0]);
236   SmallVector<SDValue, 8> Ops;
237 
238   Ops.push_back(CurDAG.getTargetConstant(RegClassID, DL, MVT::i32));
239 
240   for (unsigned I = 0; I < Regs.size(); ++I) {
241     Ops.push_back(Regs[I]);
242     Ops.push_back(CurDAG.getTargetConstant(SubReg0 + I, DL, MVT::i32));
243   }
244   SDNode *N =
245       CurDAG.getMachineNode(TargetOpcode::REG_SEQUENCE, DL, MVT::Untyped, Ops);
246   return SDValue(N, 0);
247 }
248 
249 void RISCVDAGToDAGISel::addVectorLoadStoreOperands(
250     SDNode *Node, unsigned Log2SEW, const SDLoc &DL, unsigned CurOp,
251     bool IsMasked, bool IsStridedOrIndexed, SmallVectorImpl<SDValue> &Operands,
252     bool IsLoad, MVT *IndexVT) {
253   SDValue Chain = Node->getOperand(0);
254   SDValue Glue;
255 
256   Operands.push_back(Node->getOperand(CurOp++)); // Base pointer.
257 
258   if (IsStridedOrIndexed) {
259     Operands.push_back(Node->getOperand(CurOp++)); // Index.
260     if (IndexVT)
261       *IndexVT = Operands.back()->getSimpleValueType(0);
262   }
263 
264   if (IsMasked) {
265     // Mask needs to be copied to V0.
266     SDValue Mask = Node->getOperand(CurOp++);
267     Chain = CurDAG->getCopyToReg(Chain, DL, RISCV::V0, Mask, SDValue());
268     Glue = Chain.getValue(1);
269     Operands.push_back(CurDAG->getRegister(RISCV::V0, Mask.getValueType()));
270   }
271   SDValue VL;
272   selectVLOp(Node->getOperand(CurOp++), VL);
273   Operands.push_back(VL);
274 
275   MVT XLenVT = Subtarget->getXLenVT();
276   SDValue SEWOp = CurDAG->getTargetConstant(Log2SEW, DL, XLenVT);
277   Operands.push_back(SEWOp);
278 
279   // Masked load has the tail policy argument.
280   if (IsMasked && IsLoad) {
281     // Policy must be a constant.
282     uint64_t Policy = Node->getConstantOperandVal(CurOp++);
283     SDValue PolicyOp = CurDAG->getTargetConstant(Policy, DL, XLenVT);
284     Operands.push_back(PolicyOp);
285   }
286 
287   Operands.push_back(Chain); // Chain.
288   if (Glue)
289     Operands.push_back(Glue);
290 }
291 
292 static bool isAllUndef(ArrayRef<SDValue> Values) {
293   return llvm::all_of(Values, [](SDValue V) { return V->isUndef(); });
294 }
295 
296 void RISCVDAGToDAGISel::selectVLSEG(SDNode *Node, bool IsMasked,
297                                     bool IsStrided) {
298   SDLoc DL(Node);
299   unsigned NF = Node->getNumValues() - 1;
300   MVT VT = Node->getSimpleValueType(0);
301   unsigned Log2SEW = Log2_32(VT.getScalarSizeInBits());
302   RISCVII::VLMUL LMUL = RISCVTargetLowering::getLMUL(VT);
303 
304   unsigned CurOp = 2;
305   SmallVector<SDValue, 8> Operands;
306 
307   SmallVector<SDValue, 8> Regs(Node->op_begin() + CurOp,
308                                Node->op_begin() + CurOp + NF);
309   bool IsTU = IsMasked || !isAllUndef(Regs);
310   if (IsTU) {
311     SDValue Merge = createTuple(*CurDAG, Regs, NF, LMUL);
312     Operands.push_back(Merge);
313   }
314   CurOp += NF;
315 
316   addVectorLoadStoreOperands(Node, Log2SEW, DL, CurOp, IsMasked, IsStrided,
317                              Operands, /*IsLoad=*/true);
318 
319   const RISCV::VLSEGPseudo *P =
320       RISCV::getVLSEGPseudo(NF, IsMasked, IsTU, IsStrided, /*FF*/ false, Log2SEW,
321                             static_cast<unsigned>(LMUL));
322   MachineSDNode *Load =
323       CurDAG->getMachineNode(P->Pseudo, DL, MVT::Untyped, MVT::Other, Operands);
324 
325   if (auto *MemOp = dyn_cast<MemSDNode>(Node))
326     CurDAG->setNodeMemRefs(Load, {MemOp->getMemOperand()});
327 
328   SDValue SuperReg = SDValue(Load, 0);
329   for (unsigned I = 0; I < NF; ++I) {
330     unsigned SubRegIdx = RISCVTargetLowering::getSubregIndexByMVT(VT, I);
331     ReplaceUses(SDValue(Node, I),
332                 CurDAG->getTargetExtractSubreg(SubRegIdx, DL, VT, SuperReg));
333   }
334 
335   ReplaceUses(SDValue(Node, NF), SDValue(Load, 1));
336   CurDAG->RemoveDeadNode(Node);
337 }
338 
339 void RISCVDAGToDAGISel::selectVLSEGFF(SDNode *Node, bool IsMasked) {
340   SDLoc DL(Node);
341   unsigned NF = Node->getNumValues() - 2; // Do not count VL and Chain.
342   MVT VT = Node->getSimpleValueType(0);
343   MVT XLenVT = Subtarget->getXLenVT();
344   unsigned Log2SEW = Log2_32(VT.getScalarSizeInBits());
345   RISCVII::VLMUL LMUL = RISCVTargetLowering::getLMUL(VT);
346 
347   unsigned CurOp = 2;
348   SmallVector<SDValue, 7> Operands;
349 
350   SmallVector<SDValue, 8> Regs(Node->op_begin() + CurOp,
351                                Node->op_begin() + CurOp + NF);
352   bool IsTU = IsMasked || !isAllUndef(Regs);
353   if (IsTU) {
354     SDValue MaskedOff = createTuple(*CurDAG, Regs, NF, LMUL);
355     Operands.push_back(MaskedOff);
356   }
357   CurOp += NF;
358 
359   addVectorLoadStoreOperands(Node, Log2SEW, DL, CurOp, IsMasked,
360                              /*IsStridedOrIndexed*/ false, Operands,
361                              /*IsLoad=*/true);
362 
363   const RISCV::VLSEGPseudo *P =
364       RISCV::getVLSEGPseudo(NF, IsMasked, IsTU, /*Strided*/ false, /*FF*/ true,
365                             Log2SEW, static_cast<unsigned>(LMUL));
366   MachineSDNode *Load = CurDAG->getMachineNode(P->Pseudo, DL, MVT::Untyped,
367                                                XLenVT, MVT::Other, Operands);
368 
369   if (auto *MemOp = dyn_cast<MemSDNode>(Node))
370     CurDAG->setNodeMemRefs(Load, {MemOp->getMemOperand()});
371 
372   SDValue SuperReg = SDValue(Load, 0);
373   for (unsigned I = 0; I < NF; ++I) {
374     unsigned SubRegIdx = RISCVTargetLowering::getSubregIndexByMVT(VT, I);
375     ReplaceUses(SDValue(Node, I),
376                 CurDAG->getTargetExtractSubreg(SubRegIdx, DL, VT, SuperReg));
377   }
378 
379   ReplaceUses(SDValue(Node, NF), SDValue(Load, 1));     // VL
380   ReplaceUses(SDValue(Node, NF + 1), SDValue(Load, 2)); // Chain
381   CurDAG->RemoveDeadNode(Node);
382 }
383 
384 void RISCVDAGToDAGISel::selectVLXSEG(SDNode *Node, bool IsMasked,
385                                      bool IsOrdered) {
386   SDLoc DL(Node);
387   unsigned NF = Node->getNumValues() - 1;
388   MVT VT = Node->getSimpleValueType(0);
389   unsigned Log2SEW = Log2_32(VT.getScalarSizeInBits());
390   RISCVII::VLMUL LMUL = RISCVTargetLowering::getLMUL(VT);
391 
392   unsigned CurOp = 2;
393   SmallVector<SDValue, 8> Operands;
394 
395   SmallVector<SDValue, 8> Regs(Node->op_begin() + CurOp,
396                                Node->op_begin() + CurOp + NF);
397   bool IsTU = IsMasked || !isAllUndef(Regs);
398   if (IsTU) {
399     SDValue MaskedOff = createTuple(*CurDAG, Regs, NF, LMUL);
400     Operands.push_back(MaskedOff);
401   }
402   CurOp += NF;
403 
404   MVT IndexVT;
405   addVectorLoadStoreOperands(Node, Log2SEW, DL, CurOp, IsMasked,
406                              /*IsStridedOrIndexed*/ true, Operands,
407                              /*IsLoad=*/true, &IndexVT);
408 
409   assert(VT.getVectorElementCount() == IndexVT.getVectorElementCount() &&
410          "Element count mismatch");
411 
412   RISCVII::VLMUL IndexLMUL = RISCVTargetLowering::getLMUL(IndexVT);
413   unsigned IndexLog2EEW = Log2_32(IndexVT.getScalarSizeInBits());
414   if (IndexLog2EEW == 6 && !Subtarget->is64Bit()) {
415     report_fatal_error("The V extension does not support EEW=64 for index "
416                        "values when XLEN=32");
417   }
418   const RISCV::VLXSEGPseudo *P = RISCV::getVLXSEGPseudo(
419       NF, IsMasked, IsTU, IsOrdered, IndexLog2EEW, static_cast<unsigned>(LMUL),
420       static_cast<unsigned>(IndexLMUL));
421   MachineSDNode *Load =
422       CurDAG->getMachineNode(P->Pseudo, DL, MVT::Untyped, MVT::Other, Operands);
423 
424   if (auto *MemOp = dyn_cast<MemSDNode>(Node))
425     CurDAG->setNodeMemRefs(Load, {MemOp->getMemOperand()});
426 
427   SDValue SuperReg = SDValue(Load, 0);
428   for (unsigned I = 0; I < NF; ++I) {
429     unsigned SubRegIdx = RISCVTargetLowering::getSubregIndexByMVT(VT, I);
430     ReplaceUses(SDValue(Node, I),
431                 CurDAG->getTargetExtractSubreg(SubRegIdx, DL, VT, SuperReg));
432   }
433 
434   ReplaceUses(SDValue(Node, NF), SDValue(Load, 1));
435   CurDAG->RemoveDeadNode(Node);
436 }
437 
438 void RISCVDAGToDAGISel::selectVSSEG(SDNode *Node, bool IsMasked,
439                                     bool IsStrided) {
440   SDLoc DL(Node);
441   unsigned NF = Node->getNumOperands() - 4;
442   if (IsStrided)
443     NF--;
444   if (IsMasked)
445     NF--;
446   MVT VT = Node->getOperand(2)->getSimpleValueType(0);
447   unsigned Log2SEW = Log2_32(VT.getScalarSizeInBits());
448   RISCVII::VLMUL LMUL = RISCVTargetLowering::getLMUL(VT);
449   SmallVector<SDValue, 8> Regs(Node->op_begin() + 2, Node->op_begin() + 2 + NF);
450   SDValue StoreVal = createTuple(*CurDAG, Regs, NF, LMUL);
451 
452   SmallVector<SDValue, 8> Operands;
453   Operands.push_back(StoreVal);
454   unsigned CurOp = 2 + NF;
455 
456   addVectorLoadStoreOperands(Node, Log2SEW, DL, CurOp, IsMasked, IsStrided,
457                              Operands);
458 
459   const RISCV::VSSEGPseudo *P = RISCV::getVSSEGPseudo(
460       NF, IsMasked, IsStrided, Log2SEW, static_cast<unsigned>(LMUL));
461   MachineSDNode *Store =
462       CurDAG->getMachineNode(P->Pseudo, DL, Node->getValueType(0), Operands);
463 
464   if (auto *MemOp = dyn_cast<MemSDNode>(Node))
465     CurDAG->setNodeMemRefs(Store, {MemOp->getMemOperand()});
466 
467   ReplaceNode(Node, Store);
468 }
469 
470 void RISCVDAGToDAGISel::selectVSXSEG(SDNode *Node, bool IsMasked,
471                                      bool IsOrdered) {
472   SDLoc DL(Node);
473   unsigned NF = Node->getNumOperands() - 5;
474   if (IsMasked)
475     --NF;
476   MVT VT = Node->getOperand(2)->getSimpleValueType(0);
477   unsigned Log2SEW = Log2_32(VT.getScalarSizeInBits());
478   RISCVII::VLMUL LMUL = RISCVTargetLowering::getLMUL(VT);
479   SmallVector<SDValue, 8> Regs(Node->op_begin() + 2, Node->op_begin() + 2 + NF);
480   SDValue StoreVal = createTuple(*CurDAG, Regs, NF, LMUL);
481 
482   SmallVector<SDValue, 8> Operands;
483   Operands.push_back(StoreVal);
484   unsigned CurOp = 2 + NF;
485 
486   MVT IndexVT;
487   addVectorLoadStoreOperands(Node, Log2SEW, DL, CurOp, IsMasked,
488                              /*IsStridedOrIndexed*/ true, Operands,
489                              /*IsLoad=*/false, &IndexVT);
490 
491   assert(VT.getVectorElementCount() == IndexVT.getVectorElementCount() &&
492          "Element count mismatch");
493 
494   RISCVII::VLMUL IndexLMUL = RISCVTargetLowering::getLMUL(IndexVT);
495   unsigned IndexLog2EEW = Log2_32(IndexVT.getScalarSizeInBits());
496   if (IndexLog2EEW == 6 && !Subtarget->is64Bit()) {
497     report_fatal_error("The V extension does not support EEW=64 for index "
498                        "values when XLEN=32");
499   }
500   const RISCV::VSXSEGPseudo *P = RISCV::getVSXSEGPseudo(
501       NF, IsMasked, IsOrdered, IndexLog2EEW, static_cast<unsigned>(LMUL),
502       static_cast<unsigned>(IndexLMUL));
503   MachineSDNode *Store =
504       CurDAG->getMachineNode(P->Pseudo, DL, Node->getValueType(0), Operands);
505 
506   if (auto *MemOp = dyn_cast<MemSDNode>(Node))
507     CurDAG->setNodeMemRefs(Store, {MemOp->getMemOperand()});
508 
509   ReplaceNode(Node, Store);
510 }
511 
512 void RISCVDAGToDAGISel::selectVSETVLI(SDNode *Node) {
513   if (!Subtarget->hasVInstructions())
514     return;
515 
516   assert((Node->getOpcode() == ISD::INTRINSIC_W_CHAIN ||
517           Node->getOpcode() == ISD::INTRINSIC_WO_CHAIN) &&
518          "Unexpected opcode");
519 
520   SDLoc DL(Node);
521   MVT XLenVT = Subtarget->getXLenVT();
522 
523   bool HasChain = Node->getOpcode() == ISD::INTRINSIC_W_CHAIN;
524   unsigned IntNoOffset = HasChain ? 1 : 0;
525   unsigned IntNo = Node->getConstantOperandVal(IntNoOffset);
526 
527   assert((IntNo == Intrinsic::riscv_vsetvli ||
528           IntNo == Intrinsic::riscv_vsetvlimax ||
529           IntNo == Intrinsic::riscv_vsetvli_opt ||
530           IntNo == Intrinsic::riscv_vsetvlimax_opt) &&
531          "Unexpected vsetvli intrinsic");
532 
533   bool VLMax = IntNo == Intrinsic::riscv_vsetvlimax ||
534                IntNo == Intrinsic::riscv_vsetvlimax_opt;
535   unsigned Offset = IntNoOffset + (VLMax ? 1 : 2);
536 
537   assert(Node->getNumOperands() == Offset + 2 &&
538          "Unexpected number of operands");
539 
540   unsigned SEW =
541       RISCVVType::decodeVSEW(Node->getConstantOperandVal(Offset) & 0x7);
542   RISCVII::VLMUL VLMul = static_cast<RISCVII::VLMUL>(
543       Node->getConstantOperandVal(Offset + 1) & 0x7);
544 
545   unsigned VTypeI = RISCVVType::encodeVTYPE(VLMul, SEW, /*TailAgnostic*/ true,
546                                             /*MaskAgnostic*/ false);
547   SDValue VTypeIOp = CurDAG->getTargetConstant(VTypeI, DL, XLenVT);
548 
549   SmallVector<EVT, 2> VTs = {XLenVT};
550   if (HasChain)
551     VTs.push_back(MVT::Other);
552 
553   SDValue VLOperand;
554   unsigned Opcode = RISCV::PseudoVSETVLI;
555   if (VLMax) {
556     VLOperand = CurDAG->getRegister(RISCV::X0, XLenVT);
557     Opcode = RISCV::PseudoVSETVLIX0;
558   } else {
559     VLOperand = Node->getOperand(IntNoOffset + 1);
560 
561     if (auto *C = dyn_cast<ConstantSDNode>(VLOperand)) {
562       uint64_t AVL = C->getZExtValue();
563       if (isUInt<5>(AVL)) {
564         SDValue VLImm = CurDAG->getTargetConstant(AVL, DL, XLenVT);
565         SmallVector<SDValue, 3> Ops = {VLImm, VTypeIOp};
566         if (HasChain)
567           Ops.push_back(Node->getOperand(0));
568         ReplaceNode(
569             Node, CurDAG->getMachineNode(RISCV::PseudoVSETIVLI, DL, VTs, Ops));
570         return;
571       }
572     }
573   }
574 
575   SmallVector<SDValue, 3> Ops = {VLOperand, VTypeIOp};
576   if (HasChain)
577     Ops.push_back(Node->getOperand(0));
578 
579   ReplaceNode(Node, CurDAG->getMachineNode(Opcode, DL, VTs, Ops));
580 }
581 
582 void RISCVDAGToDAGISel::Select(SDNode *Node) {
583   // If we have a custom node, we have already selected.
584   if (Node->isMachineOpcode()) {
585     LLVM_DEBUG(dbgs() << "== "; Node->dump(CurDAG); dbgs() << "\n");
586     Node->setNodeId(-1);
587     return;
588   }
589 
590   // Instruction Selection not handled by the auto-generated tablegen selection
591   // should be handled here.
592   unsigned Opcode = Node->getOpcode();
593   MVT XLenVT = Subtarget->getXLenVT();
594   SDLoc DL(Node);
595   MVT VT = Node->getSimpleValueType(0);
596 
597   switch (Opcode) {
598   case ISD::Constant: {
599     auto *ConstNode = cast<ConstantSDNode>(Node);
600     if (VT == XLenVT && ConstNode->isZero()) {
601       SDValue New =
602           CurDAG->getCopyFromReg(CurDAG->getEntryNode(), DL, RISCV::X0, XLenVT);
603       ReplaceNode(Node, New.getNode());
604       return;
605     }
606     int64_t Imm = ConstNode->getSExtValue();
607     // If the upper XLen-16 bits are not used, try to convert this to a simm12
608     // by sign extending bit 15.
609     if (isUInt<16>(Imm) && isInt<12>(SignExtend64<16>(Imm)) &&
610         hasAllHUsers(Node))
611       Imm = SignExtend64<16>(Imm);
612     // If the upper 32-bits are not used try to convert this into a simm32 by
613     // sign extending bit 32.
614     if (!isInt<32>(Imm) && isUInt<32>(Imm) && hasAllWUsers(Node))
615       Imm = SignExtend64<32>(Imm);
616 
617     ReplaceNode(Node, selectImm(CurDAG, DL, VT, Imm, *Subtarget));
618     return;
619   }
620   case ISD::SHL: {
621     auto *N1C = dyn_cast<ConstantSDNode>(Node->getOperand(1));
622     if (!N1C)
623       break;
624     SDValue N0 = Node->getOperand(0);
625     if (N0.getOpcode() != ISD::AND || !N0.hasOneUse() ||
626         !isa<ConstantSDNode>(N0.getOperand(1)))
627       break;
628     unsigned ShAmt = N1C->getZExtValue();
629     uint64_t Mask = N0.getConstantOperandVal(1);
630 
631     // Optimize (shl (and X, C2), C) -> (slli (srliw X, C3), C3+C) where C2 has
632     // 32 leading zeros and C3 trailing zeros.
633     if (ShAmt <= 32 && isShiftedMask_64(Mask)) {
634       unsigned XLen = Subtarget->getXLen();
635       unsigned LeadingZeros = XLen - (64 - countLeadingZeros(Mask));
636       unsigned TrailingZeros = countTrailingZeros(Mask);
637       if (TrailingZeros > 0 && LeadingZeros == 32) {
638         SDNode *SRLIW = CurDAG->getMachineNode(
639             RISCV::SRLIW, DL, VT, N0->getOperand(0),
640             CurDAG->getTargetConstant(TrailingZeros, DL, VT));
641         SDNode *SLLI = CurDAG->getMachineNode(
642             RISCV::SLLI, DL, VT, SDValue(SRLIW, 0),
643             CurDAG->getTargetConstant(TrailingZeros + ShAmt, DL, VT));
644         ReplaceNode(Node, SLLI);
645         return;
646       }
647     }
648     break;
649   }
650   case ISD::SRL: {
651     auto *N1C = dyn_cast<ConstantSDNode>(Node->getOperand(1));
652     if (!N1C)
653       break;
654     SDValue N0 = Node->getOperand(0);
655     if (N0.getOpcode() != ISD::AND || !N0.hasOneUse() ||
656         !isa<ConstantSDNode>(N0.getOperand(1)))
657       break;
658     unsigned ShAmt = N1C->getZExtValue();
659     uint64_t Mask = N0.getConstantOperandVal(1);
660 
661     // Optimize (srl (and X, C2), C) -> (slli (srliw X, C3), C3-C) where C2 has
662     // 32 leading zeros and C3 trailing zeros.
663     if (isShiftedMask_64(Mask)) {
664       unsigned XLen = Subtarget->getXLen();
665       unsigned LeadingZeros = XLen - (64 - countLeadingZeros(Mask));
666       unsigned TrailingZeros = countTrailingZeros(Mask);
667       if (LeadingZeros == 32 && TrailingZeros > ShAmt) {
668         SDNode *SRLIW = CurDAG->getMachineNode(
669             RISCV::SRLIW, DL, VT, N0->getOperand(0),
670             CurDAG->getTargetConstant(TrailingZeros, DL, VT));
671         SDNode *SLLI = CurDAG->getMachineNode(
672             RISCV::SLLI, DL, VT, SDValue(SRLIW, 0),
673             CurDAG->getTargetConstant(TrailingZeros - ShAmt, DL, VT));
674         ReplaceNode(Node, SLLI);
675         return;
676       }
677     }
678 
679     // Optimize (srl (and X, C2), C) ->
680     //          (srli (slli X, (XLen-C3), (XLen-C3) + C)
681     // Where C2 is a mask with C3 trailing ones.
682     // Taking into account that the C2 may have had lower bits unset by
683     // SimplifyDemandedBits. This avoids materializing the C2 immediate.
684     // This pattern occurs when type legalizing right shifts for types with
685     // less than XLen bits.
686     Mask |= maskTrailingOnes<uint64_t>(ShAmt);
687     if (!isMask_64(Mask))
688       break;
689     unsigned TrailingOnes = countTrailingOnes(Mask);
690     // 32 trailing ones should use srliw via tablegen pattern.
691     if (TrailingOnes == 32 || ShAmt >= TrailingOnes)
692       break;
693     unsigned LShAmt = Subtarget->getXLen() - TrailingOnes;
694     SDNode *SLLI =
695         CurDAG->getMachineNode(RISCV::SLLI, DL, VT, N0->getOperand(0),
696                                CurDAG->getTargetConstant(LShAmt, DL, VT));
697     SDNode *SRLI = CurDAG->getMachineNode(
698         RISCV::SRLI, DL, VT, SDValue(SLLI, 0),
699         CurDAG->getTargetConstant(LShAmt + ShAmt, DL, VT));
700     ReplaceNode(Node, SRLI);
701     return;
702   }
703   case ISD::SRA: {
704     // Optimize (sra (sext_inreg X, i16), C) ->
705     //          (srai (slli X, (XLen-16), (XLen-16) + C)
706     // And      (sra (sext_inreg X, i8), C) ->
707     //          (srai (slli X, (XLen-8), (XLen-8) + C)
708     // This can occur when Zbb is enabled, which makes sext_inreg i16/i8 legal.
709     // This transform matches the code we get without Zbb. The shifts are more
710     // compressible, and this can help expose CSE opportunities in the sdiv by
711     // constant optimization.
712     auto *N1C = dyn_cast<ConstantSDNode>(Node->getOperand(1));
713     if (!N1C)
714       break;
715     SDValue N0 = Node->getOperand(0);
716     if (N0.getOpcode() != ISD::SIGN_EXTEND_INREG || !N0.hasOneUse())
717       break;
718     unsigned ShAmt = N1C->getZExtValue();
719     unsigned ExtSize =
720         cast<VTSDNode>(N0.getOperand(1))->getVT().getSizeInBits();
721     // ExtSize of 32 should use sraiw via tablegen pattern.
722     if (ExtSize >= 32 || ShAmt >= ExtSize)
723       break;
724     unsigned LShAmt = Subtarget->getXLen() - ExtSize;
725     SDNode *SLLI =
726         CurDAG->getMachineNode(RISCV::SLLI, DL, VT, N0->getOperand(0),
727                                CurDAG->getTargetConstant(LShAmt, DL, VT));
728     SDNode *SRAI = CurDAG->getMachineNode(
729         RISCV::SRAI, DL, VT, SDValue(SLLI, 0),
730         CurDAG->getTargetConstant(LShAmt + ShAmt, DL, VT));
731     ReplaceNode(Node, SRAI);
732     return;
733   }
734   case ISD::AND: {
735     auto *N1C = dyn_cast<ConstantSDNode>(Node->getOperand(1));
736     if (!N1C)
737       break;
738 
739     SDValue N0 = Node->getOperand(0);
740 
741     bool LeftShift = N0.getOpcode() == ISD::SHL;
742     if (!LeftShift && N0.getOpcode() != ISD::SRL)
743       break;
744 
745     auto *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
746     if (!C)
747       break;
748     unsigned C2 = C->getZExtValue();
749     unsigned XLen = Subtarget->getXLen();
750     assert((C2 > 0 && C2 < XLen) && "Unexpected shift amount!");
751 
752     uint64_t C1 = N1C->getZExtValue();
753 
754     // Keep track of whether this is a c.andi. If we can't use c.andi, the
755     // shift pair might offer more compression opportunities.
756     // TODO: We could check for C extension here, but we don't have many lit
757     // tests with the C extension enabled so not checking gets better coverage.
758     // TODO: What if ANDI faster than shift?
759     bool IsCANDI = isInt<6>(N1C->getSExtValue());
760 
761     // Clear irrelevant bits in the mask.
762     if (LeftShift)
763       C1 &= maskTrailingZeros<uint64_t>(C2);
764     else
765       C1 &= maskTrailingOnes<uint64_t>(XLen - C2);
766 
767     // Some transforms should only be done if the shift has a single use or
768     // the AND would become (srli (slli X, 32), 32)
769     bool OneUseOrZExtW = N0.hasOneUse() || C1 == UINT64_C(0xFFFFFFFF);
770 
771     SDValue X = N0.getOperand(0);
772 
773     // Turn (and (srl x, c2) c1) -> (srli (slli x, c3-c2), c3) if c1 is a mask
774     // with c3 leading zeros.
775     if (!LeftShift && isMask_64(C1)) {
776       unsigned Leading = XLen - (64 - countLeadingZeros(C1));
777       if (C2 < Leading) {
778         // If the number of leading zeros is C2+32 this can be SRLIW.
779         if (C2 + 32 == Leading) {
780           SDNode *SRLIW = CurDAG->getMachineNode(
781               RISCV::SRLIW, DL, VT, X, CurDAG->getTargetConstant(C2, DL, VT));
782           ReplaceNode(Node, SRLIW);
783           return;
784         }
785 
786         // (and (srl (sexti32 Y), c2), c1) -> (srliw (sraiw Y, 31), c3 - 32) if
787         // c1 is a mask with c3 leading zeros and c2 >= 32 and c3-c2==1.
788         //
789         // This pattern occurs when (i32 (srl (sra 31), c3 - 32)) is type
790         // legalized and goes through DAG combine.
791         if (C2 >= 32 && (Leading - C2) == 1 && N0.hasOneUse() &&
792             X.getOpcode() == ISD::SIGN_EXTEND_INREG &&
793             cast<VTSDNode>(X.getOperand(1))->getVT() == MVT::i32) {
794           SDNode *SRAIW =
795               CurDAG->getMachineNode(RISCV::SRAIW, DL, VT, X.getOperand(0),
796                                      CurDAG->getTargetConstant(31, DL, VT));
797           SDNode *SRLIW = CurDAG->getMachineNode(
798               RISCV::SRLIW, DL, VT, SDValue(SRAIW, 0),
799               CurDAG->getTargetConstant(Leading - 32, DL, VT));
800           ReplaceNode(Node, SRLIW);
801           return;
802         }
803 
804         // (srli (slli x, c3-c2), c3).
805         // Skip if we could use (zext.w (sraiw X, C2)).
806         bool Skip = Subtarget->hasStdExtZba() && Leading == 32 &&
807                     X.getOpcode() == ISD::SIGN_EXTEND_INREG &&
808                     cast<VTSDNode>(X.getOperand(1))->getVT() == MVT::i32;
809         // Also Skip if we can use bexti.
810         Skip |= Subtarget->hasStdExtZbs() && Leading == XLen - 1;
811         if (OneUseOrZExtW && !Skip) {
812           SDNode *SLLI = CurDAG->getMachineNode(
813               RISCV::SLLI, DL, VT, X,
814               CurDAG->getTargetConstant(Leading - C2, DL, VT));
815           SDNode *SRLI = CurDAG->getMachineNode(
816               RISCV::SRLI, DL, VT, SDValue(SLLI, 0),
817               CurDAG->getTargetConstant(Leading, DL, VT));
818           ReplaceNode(Node, SRLI);
819           return;
820         }
821       }
822     }
823 
824     // Turn (and (shl x, c2), c1) -> (srli (slli c2+c3), c3) if c1 is a mask
825     // shifted by c2 bits with c3 leading zeros.
826     if (LeftShift && isShiftedMask_64(C1)) {
827       unsigned Leading = XLen - (64 - countLeadingZeros(C1));
828 
829       if (C2 + Leading < XLen &&
830           C1 == (maskTrailingOnes<uint64_t>(XLen - (C2 + Leading)) << C2)) {
831         // Use slli.uw when possible.
832         if ((XLen - (C2 + Leading)) == 32 && Subtarget->hasStdExtZba()) {
833           SDNode *SLLI_UW = CurDAG->getMachineNode(
834               RISCV::SLLI_UW, DL, VT, X, CurDAG->getTargetConstant(C2, DL, VT));
835           ReplaceNode(Node, SLLI_UW);
836           return;
837         }
838 
839         // (srli (slli c2+c3), c3)
840         if (OneUseOrZExtW && !IsCANDI) {
841           SDNode *SLLI = CurDAG->getMachineNode(
842               RISCV::SLLI, DL, VT, X,
843               CurDAG->getTargetConstant(C2 + Leading, DL, VT));
844           SDNode *SRLI = CurDAG->getMachineNode(
845               RISCV::SRLI, DL, VT, SDValue(SLLI, 0),
846               CurDAG->getTargetConstant(Leading, DL, VT));
847           ReplaceNode(Node, SRLI);
848           return;
849         }
850       }
851     }
852 
853     // Turn (and (shr x, c2), c1) -> (slli (srli x, c2+c3), c3) if c1 is a
854     // shifted mask with c2 leading zeros and c3 trailing zeros.
855     if (!LeftShift && isShiftedMask_64(C1)) {
856       unsigned Leading = XLen - (64 - countLeadingZeros(C1));
857       unsigned Trailing = countTrailingZeros(C1);
858       if (Leading == C2 && C2 + Trailing < XLen && OneUseOrZExtW && !IsCANDI) {
859         unsigned SrliOpc = RISCV::SRLI;
860         // If the input is zexti32 we should use SRLIW.
861         if (X.getOpcode() == ISD::AND && isa<ConstantSDNode>(X.getOperand(1)) &&
862             X.getConstantOperandVal(1) == UINT64_C(0xFFFFFFFF)) {
863           SrliOpc = RISCV::SRLIW;
864           X = X.getOperand(0);
865         }
866         SDNode *SRLI = CurDAG->getMachineNode(
867             SrliOpc, DL, VT, X,
868             CurDAG->getTargetConstant(C2 + Trailing, DL, VT));
869         SDNode *SLLI =
870             CurDAG->getMachineNode(RISCV::SLLI, DL, VT, SDValue(SRLI, 0),
871                                    CurDAG->getTargetConstant(Trailing, DL, VT));
872         ReplaceNode(Node, SLLI);
873         return;
874       }
875       // If the leading zero count is C2+32, we can use SRLIW instead of SRLI.
876       if (Leading > 32 && (Leading - 32) == C2 && C2 + Trailing < 32 &&
877           OneUseOrZExtW && !IsCANDI) {
878         SDNode *SRLIW = CurDAG->getMachineNode(
879             RISCV::SRLIW, DL, VT, X,
880             CurDAG->getTargetConstant(C2 + Trailing, DL, VT));
881         SDNode *SLLI =
882             CurDAG->getMachineNode(RISCV::SLLI, DL, VT, SDValue(SRLIW, 0),
883                                    CurDAG->getTargetConstant(Trailing, DL, VT));
884         ReplaceNode(Node, SLLI);
885         return;
886       }
887     }
888 
889     // Turn (and (shl x, c2), c1) -> (slli (srli x, c3-c2), c3) if c1 is a
890     // shifted mask with no leading zeros and c3 trailing zeros.
891     if (LeftShift && isShiftedMask_64(C1)) {
892       unsigned Leading = XLen - (64 - countLeadingZeros(C1));
893       unsigned Trailing = countTrailingZeros(C1);
894       if (Leading == 0 && C2 < Trailing && OneUseOrZExtW && !IsCANDI) {
895         SDNode *SRLI = CurDAG->getMachineNode(
896             RISCV::SRLI, DL, VT, X,
897             CurDAG->getTargetConstant(Trailing - C2, DL, VT));
898         SDNode *SLLI =
899             CurDAG->getMachineNode(RISCV::SLLI, DL, VT, SDValue(SRLI, 0),
900                                    CurDAG->getTargetConstant(Trailing, DL, VT));
901         ReplaceNode(Node, SLLI);
902         return;
903       }
904       // If we have (32-C2) leading zeros, we can use SRLIW instead of SRLI.
905       if (C2 < Trailing && Leading + C2 == 32 && OneUseOrZExtW && !IsCANDI) {
906         SDNode *SRLIW = CurDAG->getMachineNode(
907             RISCV::SRLIW, DL, VT, X,
908             CurDAG->getTargetConstant(Trailing - C2, DL, VT));
909         SDNode *SLLI =
910             CurDAG->getMachineNode(RISCV::SLLI, DL, VT, SDValue(SRLIW, 0),
911                                    CurDAG->getTargetConstant(Trailing, DL, VT));
912         ReplaceNode(Node, SLLI);
913         return;
914       }
915     }
916 
917     break;
918   }
919   case ISD::MUL: {
920     // Special case for calculating (mul (and X, C2), C1) where the full product
921     // fits in XLen bits. We can shift X left by the number of leading zeros in
922     // C2 and shift C1 left by XLen-lzcnt(C2). This will ensure the final
923     // product has XLen trailing zeros, putting it in the output of MULHU. This
924     // can avoid materializing a constant in a register for C2.
925 
926     // RHS should be a constant.
927     auto *N1C = dyn_cast<ConstantSDNode>(Node->getOperand(1));
928     if (!N1C || !N1C->hasOneUse())
929       break;
930 
931     // LHS should be an AND with constant.
932     SDValue N0 = Node->getOperand(0);
933     if (N0.getOpcode() != ISD::AND || !isa<ConstantSDNode>(N0.getOperand(1)))
934       break;
935 
936     uint64_t C2 = cast<ConstantSDNode>(N0.getOperand(1))->getZExtValue();
937 
938     // Constant should be a mask.
939     if (!isMask_64(C2))
940       break;
941 
942     // If this can be an ANDI, ZEXT.H or ZEXT.W, don't do this if the ANDI/ZEXT
943     // has multiple users or the constant is a simm12. This prevents inserting
944     // a shift and still have uses of the AND/ZEXT. Shifting a simm12 will
945     // likely make it more costly to materialize. Otherwise, using a SLLI
946     // might allow it to be compressed.
947     bool IsANDIOrZExt =
948         isInt<12>(C2) ||
949         (C2 == UINT64_C(0xFFFF) &&
950          (Subtarget->hasStdExtZbb() || Subtarget->hasStdExtZbp())) ||
951         (C2 == UINT64_C(0xFFFFFFFF) && Subtarget->hasStdExtZba());
952     if (IsANDIOrZExt && (isInt<12>(N1C->getSExtValue()) || !N0.hasOneUse()))
953       break;
954 
955     // We need to shift left the AND input and C1 by a total of XLen bits.
956 
957     // How far left do we need to shift the AND input?
958     unsigned XLen = Subtarget->getXLen();
959     unsigned LeadingZeros = XLen - (64 - countLeadingZeros(C2));
960 
961     // The constant gets shifted by the remaining amount unless that would
962     // shift bits out.
963     uint64_t C1 = N1C->getZExtValue();
964     unsigned ConstantShift = XLen - LeadingZeros;
965     if (ConstantShift > (XLen - (64 - countLeadingZeros(C1))))
966       break;
967 
968     uint64_t ShiftedC1 = C1 << ConstantShift;
969     // If this RV32, we need to sign extend the constant.
970     if (XLen == 32)
971       ShiftedC1 = SignExtend64<32>(ShiftedC1);
972 
973     // Create (mulhu (slli X, lzcnt(C2)), C1 << (XLen - lzcnt(C2))).
974     SDNode *Imm = selectImm(CurDAG, DL, VT, ShiftedC1, *Subtarget);
975     SDNode *SLLI =
976         CurDAG->getMachineNode(RISCV::SLLI, DL, VT, N0.getOperand(0),
977                                CurDAG->getTargetConstant(LeadingZeros, DL, VT));
978     SDNode *MULHU = CurDAG->getMachineNode(RISCV::MULHU, DL, VT,
979                                            SDValue(SLLI, 0), SDValue(Imm, 0));
980     ReplaceNode(Node, MULHU);
981     return;
982   }
983   case ISD::INTRINSIC_WO_CHAIN: {
984     unsigned IntNo = Node->getConstantOperandVal(0);
985     switch (IntNo) {
986       // By default we do not custom select any intrinsic.
987     default:
988       break;
989     case Intrinsic::riscv_vmsgeu:
990     case Intrinsic::riscv_vmsge: {
991       SDValue Src1 = Node->getOperand(1);
992       SDValue Src2 = Node->getOperand(2);
993       bool IsUnsigned = IntNo == Intrinsic::riscv_vmsgeu;
994       bool IsCmpUnsignedZero = false;
995       // Only custom select scalar second operand.
996       if (Src2.getValueType() != XLenVT)
997         break;
998       // Small constants are handled with patterns.
999       if (auto *C = dyn_cast<ConstantSDNode>(Src2)) {
1000         int64_t CVal = C->getSExtValue();
1001         if (CVal >= -15 && CVal <= 16) {
1002           if (!IsUnsigned || CVal != 0)
1003             break;
1004           IsCmpUnsignedZero = true;
1005         }
1006       }
1007       MVT Src1VT = Src1.getSimpleValueType();
1008       unsigned VMSLTOpcode, VMNANDOpcode, VMSetOpcode;
1009       switch (RISCVTargetLowering::getLMUL(Src1VT)) {
1010       default:
1011         llvm_unreachable("Unexpected LMUL!");
1012 #define CASE_VMSLT_VMNAND_VMSET_OPCODES(lmulenum, suffix, suffix_b)            \
1013   case RISCVII::VLMUL::lmulenum:                                               \
1014     VMSLTOpcode = IsUnsigned ? RISCV::PseudoVMSLTU_VX_##suffix                 \
1015                              : RISCV::PseudoVMSLT_VX_##suffix;                 \
1016     VMNANDOpcode = RISCV::PseudoVMNAND_MM_##suffix;                            \
1017     VMSetOpcode = RISCV::PseudoVMSET_M_##suffix_b;                             \
1018     break;
1019         CASE_VMSLT_VMNAND_VMSET_OPCODES(LMUL_F8, MF8, B1)
1020         CASE_VMSLT_VMNAND_VMSET_OPCODES(LMUL_F4, MF4, B2)
1021         CASE_VMSLT_VMNAND_VMSET_OPCODES(LMUL_F2, MF2, B4)
1022         CASE_VMSLT_VMNAND_VMSET_OPCODES(LMUL_1, M1, B8)
1023         CASE_VMSLT_VMNAND_VMSET_OPCODES(LMUL_2, M2, B16)
1024         CASE_VMSLT_VMNAND_VMSET_OPCODES(LMUL_4, M4, B32)
1025         CASE_VMSLT_VMNAND_VMSET_OPCODES(LMUL_8, M8, B64)
1026 #undef CASE_VMSLT_VMNAND_VMSET_OPCODES
1027       }
1028       SDValue SEW = CurDAG->getTargetConstant(
1029           Log2_32(Src1VT.getScalarSizeInBits()), DL, XLenVT);
1030       SDValue VL;
1031       selectVLOp(Node->getOperand(3), VL);
1032 
1033       // If vmsgeu with 0 immediate, expand it to vmset.
1034       if (IsCmpUnsignedZero) {
1035         ReplaceNode(Node, CurDAG->getMachineNode(VMSetOpcode, DL, VT, VL, SEW));
1036         return;
1037       }
1038 
1039       // Expand to
1040       // vmslt{u}.vx vd, va, x; vmnand.mm vd, vd, vd
1041       SDValue Cmp = SDValue(
1042           CurDAG->getMachineNode(VMSLTOpcode, DL, VT, {Src1, Src2, VL, SEW}),
1043           0);
1044       ReplaceNode(Node, CurDAG->getMachineNode(VMNANDOpcode, DL, VT,
1045                                                {Cmp, Cmp, VL, SEW}));
1046       return;
1047     }
1048     case Intrinsic::riscv_vmsgeu_mask:
1049     case Intrinsic::riscv_vmsge_mask: {
1050       SDValue Src1 = Node->getOperand(2);
1051       SDValue Src2 = Node->getOperand(3);
1052       bool IsUnsigned = IntNo == Intrinsic::riscv_vmsgeu_mask;
1053       bool IsCmpUnsignedZero = false;
1054       // Only custom select scalar second operand.
1055       if (Src2.getValueType() != XLenVT)
1056         break;
1057       // Small constants are handled with patterns.
1058       if (auto *C = dyn_cast<ConstantSDNode>(Src2)) {
1059         int64_t CVal = C->getSExtValue();
1060         if (CVal >= -15 && CVal <= 16) {
1061           if (!IsUnsigned || CVal != 0)
1062             break;
1063           IsCmpUnsignedZero = true;
1064         }
1065       }
1066       MVT Src1VT = Src1.getSimpleValueType();
1067       unsigned VMSLTOpcode, VMSLTMaskOpcode, VMXOROpcode, VMANDNOpcode,
1068           VMOROpcode;
1069       switch (RISCVTargetLowering::getLMUL(Src1VT)) {
1070       default:
1071         llvm_unreachable("Unexpected LMUL!");
1072 #define CASE_VMSLT_OPCODES(lmulenum, suffix, suffix_b)                         \
1073   case RISCVII::VLMUL::lmulenum:                                               \
1074     VMSLTOpcode = IsUnsigned ? RISCV::PseudoVMSLTU_VX_##suffix                 \
1075                              : RISCV::PseudoVMSLT_VX_##suffix;                 \
1076     VMSLTMaskOpcode = IsUnsigned ? RISCV::PseudoVMSLTU_VX_##suffix##_MASK      \
1077                                  : RISCV::PseudoVMSLT_VX_##suffix##_MASK;      \
1078     break;
1079         CASE_VMSLT_OPCODES(LMUL_F8, MF8, B1)
1080         CASE_VMSLT_OPCODES(LMUL_F4, MF4, B2)
1081         CASE_VMSLT_OPCODES(LMUL_F2, MF2, B4)
1082         CASE_VMSLT_OPCODES(LMUL_1, M1, B8)
1083         CASE_VMSLT_OPCODES(LMUL_2, M2, B16)
1084         CASE_VMSLT_OPCODES(LMUL_4, M4, B32)
1085         CASE_VMSLT_OPCODES(LMUL_8, M8, B64)
1086 #undef CASE_VMSLT_OPCODES
1087       }
1088       // Mask operations use the LMUL from the mask type.
1089       switch (RISCVTargetLowering::getLMUL(VT)) {
1090       default:
1091         llvm_unreachable("Unexpected LMUL!");
1092 #define CASE_VMXOR_VMANDN_VMOR_OPCODES(lmulenum, suffix)                       \
1093   case RISCVII::VLMUL::lmulenum:                                               \
1094     VMXOROpcode = RISCV::PseudoVMXOR_MM_##suffix;                              \
1095     VMANDNOpcode = RISCV::PseudoVMANDN_MM_##suffix;                            \
1096     VMOROpcode = RISCV::PseudoVMOR_MM_##suffix;                                \
1097     break;
1098         CASE_VMXOR_VMANDN_VMOR_OPCODES(LMUL_F8, MF8)
1099         CASE_VMXOR_VMANDN_VMOR_OPCODES(LMUL_F4, MF4)
1100         CASE_VMXOR_VMANDN_VMOR_OPCODES(LMUL_F2, MF2)
1101         CASE_VMXOR_VMANDN_VMOR_OPCODES(LMUL_1, M1)
1102         CASE_VMXOR_VMANDN_VMOR_OPCODES(LMUL_2, M2)
1103         CASE_VMXOR_VMANDN_VMOR_OPCODES(LMUL_4, M4)
1104         CASE_VMXOR_VMANDN_VMOR_OPCODES(LMUL_8, M8)
1105 #undef CASE_VMXOR_VMANDN_VMOR_OPCODES
1106       }
1107       SDValue SEW = CurDAG->getTargetConstant(
1108           Log2_32(Src1VT.getScalarSizeInBits()), DL, XLenVT);
1109       SDValue MaskSEW = CurDAG->getTargetConstant(0, DL, XLenVT);
1110       SDValue VL;
1111       selectVLOp(Node->getOperand(5), VL);
1112       SDValue MaskedOff = Node->getOperand(1);
1113       SDValue Mask = Node->getOperand(4);
1114 
1115       // If vmsgeu_mask with 0 immediate, expand it to vmor mask, maskedoff.
1116       if (IsCmpUnsignedZero) {
1117         // We don't need vmor if the MaskedOff and the Mask are the same
1118         // value.
1119         if (Mask == MaskedOff) {
1120           ReplaceUses(Node, Mask.getNode());
1121           return;
1122         }
1123         ReplaceNode(Node,
1124                     CurDAG->getMachineNode(VMOROpcode, DL, VT,
1125                                            {Mask, MaskedOff, VL, MaskSEW}));
1126         return;
1127       }
1128 
1129       // If the MaskedOff value and the Mask are the same value use
1130       // vmslt{u}.vx vt, va, x;  vmandn.mm vd, vd, vt
1131       // This avoids needing to copy v0 to vd before starting the next sequence.
1132       if (Mask == MaskedOff) {
1133         SDValue Cmp = SDValue(
1134             CurDAG->getMachineNode(VMSLTOpcode, DL, VT, {Src1, Src2, VL, SEW}),
1135             0);
1136         ReplaceNode(Node, CurDAG->getMachineNode(VMANDNOpcode, DL, VT,
1137                                                  {Mask, Cmp, VL, MaskSEW}));
1138         return;
1139       }
1140 
1141       // Mask needs to be copied to V0.
1142       SDValue Chain = CurDAG->getCopyToReg(CurDAG->getEntryNode(), DL,
1143                                            RISCV::V0, Mask, SDValue());
1144       SDValue Glue = Chain.getValue(1);
1145       SDValue V0 = CurDAG->getRegister(RISCV::V0, VT);
1146 
1147       // Otherwise use
1148       // vmslt{u}.vx vd, va, x, v0.t; vmxor.mm vd, vd, v0
1149       // The result is mask undisturbed.
1150       // We use the same instructions to emulate mask agnostic behavior, because
1151       // the agnostic result can be either undisturbed or all 1.
1152       SDValue Cmp = SDValue(
1153           CurDAG->getMachineNode(VMSLTMaskOpcode, DL, VT,
1154                                  {MaskedOff, Src1, Src2, V0, VL, SEW, Glue}),
1155           0);
1156       // vmxor.mm vd, vd, v0 is used to update active value.
1157       ReplaceNode(Node, CurDAG->getMachineNode(VMXOROpcode, DL, VT,
1158                                                {Cmp, Mask, VL, MaskSEW}));
1159       return;
1160     }
1161     case Intrinsic::riscv_vsetvli_opt:
1162     case Intrinsic::riscv_vsetvlimax_opt:
1163       return selectVSETVLI(Node);
1164     }
1165     break;
1166   }
1167   case ISD::INTRINSIC_W_CHAIN: {
1168     unsigned IntNo = cast<ConstantSDNode>(Node->getOperand(1))->getZExtValue();
1169     switch (IntNo) {
1170       // By default we do not custom select any intrinsic.
1171     default:
1172       break;
1173     case Intrinsic::riscv_vsetvli:
1174     case Intrinsic::riscv_vsetvlimax:
1175       return selectVSETVLI(Node);
1176     case Intrinsic::riscv_vlseg2:
1177     case Intrinsic::riscv_vlseg3:
1178     case Intrinsic::riscv_vlseg4:
1179     case Intrinsic::riscv_vlseg5:
1180     case Intrinsic::riscv_vlseg6:
1181     case Intrinsic::riscv_vlseg7:
1182     case Intrinsic::riscv_vlseg8: {
1183       selectVLSEG(Node, /*IsMasked*/ false, /*IsStrided*/ false);
1184       return;
1185     }
1186     case Intrinsic::riscv_vlseg2_mask:
1187     case Intrinsic::riscv_vlseg3_mask:
1188     case Intrinsic::riscv_vlseg4_mask:
1189     case Intrinsic::riscv_vlseg5_mask:
1190     case Intrinsic::riscv_vlseg6_mask:
1191     case Intrinsic::riscv_vlseg7_mask:
1192     case Intrinsic::riscv_vlseg8_mask: {
1193       selectVLSEG(Node, /*IsMasked*/ true, /*IsStrided*/ false);
1194       return;
1195     }
1196     case Intrinsic::riscv_vlsseg2:
1197     case Intrinsic::riscv_vlsseg3:
1198     case Intrinsic::riscv_vlsseg4:
1199     case Intrinsic::riscv_vlsseg5:
1200     case Intrinsic::riscv_vlsseg6:
1201     case Intrinsic::riscv_vlsseg7:
1202     case Intrinsic::riscv_vlsseg8: {
1203       selectVLSEG(Node, /*IsMasked*/ false, /*IsStrided*/ true);
1204       return;
1205     }
1206     case Intrinsic::riscv_vlsseg2_mask:
1207     case Intrinsic::riscv_vlsseg3_mask:
1208     case Intrinsic::riscv_vlsseg4_mask:
1209     case Intrinsic::riscv_vlsseg5_mask:
1210     case Intrinsic::riscv_vlsseg6_mask:
1211     case Intrinsic::riscv_vlsseg7_mask:
1212     case Intrinsic::riscv_vlsseg8_mask: {
1213       selectVLSEG(Node, /*IsMasked*/ true, /*IsStrided*/ true);
1214       return;
1215     }
1216     case Intrinsic::riscv_vloxseg2:
1217     case Intrinsic::riscv_vloxseg3:
1218     case Intrinsic::riscv_vloxseg4:
1219     case Intrinsic::riscv_vloxseg5:
1220     case Intrinsic::riscv_vloxseg6:
1221     case Intrinsic::riscv_vloxseg7:
1222     case Intrinsic::riscv_vloxseg8:
1223       selectVLXSEG(Node, /*IsMasked*/ false, /*IsOrdered*/ true);
1224       return;
1225     case Intrinsic::riscv_vluxseg2:
1226     case Intrinsic::riscv_vluxseg3:
1227     case Intrinsic::riscv_vluxseg4:
1228     case Intrinsic::riscv_vluxseg5:
1229     case Intrinsic::riscv_vluxseg6:
1230     case Intrinsic::riscv_vluxseg7:
1231     case Intrinsic::riscv_vluxseg8:
1232       selectVLXSEG(Node, /*IsMasked*/ false, /*IsOrdered*/ false);
1233       return;
1234     case Intrinsic::riscv_vloxseg2_mask:
1235     case Intrinsic::riscv_vloxseg3_mask:
1236     case Intrinsic::riscv_vloxseg4_mask:
1237     case Intrinsic::riscv_vloxseg5_mask:
1238     case Intrinsic::riscv_vloxseg6_mask:
1239     case Intrinsic::riscv_vloxseg7_mask:
1240     case Intrinsic::riscv_vloxseg8_mask:
1241       selectVLXSEG(Node, /*IsMasked*/ true, /*IsOrdered*/ true);
1242       return;
1243     case Intrinsic::riscv_vluxseg2_mask:
1244     case Intrinsic::riscv_vluxseg3_mask:
1245     case Intrinsic::riscv_vluxseg4_mask:
1246     case Intrinsic::riscv_vluxseg5_mask:
1247     case Intrinsic::riscv_vluxseg6_mask:
1248     case Intrinsic::riscv_vluxseg7_mask:
1249     case Intrinsic::riscv_vluxseg8_mask:
1250       selectVLXSEG(Node, /*IsMasked*/ true, /*IsOrdered*/ false);
1251       return;
1252     case Intrinsic::riscv_vlseg8ff:
1253     case Intrinsic::riscv_vlseg7ff:
1254     case Intrinsic::riscv_vlseg6ff:
1255     case Intrinsic::riscv_vlseg5ff:
1256     case Intrinsic::riscv_vlseg4ff:
1257     case Intrinsic::riscv_vlseg3ff:
1258     case Intrinsic::riscv_vlseg2ff: {
1259       selectVLSEGFF(Node, /*IsMasked*/ false);
1260       return;
1261     }
1262     case Intrinsic::riscv_vlseg8ff_mask:
1263     case Intrinsic::riscv_vlseg7ff_mask:
1264     case Intrinsic::riscv_vlseg6ff_mask:
1265     case Intrinsic::riscv_vlseg5ff_mask:
1266     case Intrinsic::riscv_vlseg4ff_mask:
1267     case Intrinsic::riscv_vlseg3ff_mask:
1268     case Intrinsic::riscv_vlseg2ff_mask: {
1269       selectVLSEGFF(Node, /*IsMasked*/ true);
1270       return;
1271     }
1272     case Intrinsic::riscv_vloxei:
1273     case Intrinsic::riscv_vloxei_mask:
1274     case Intrinsic::riscv_vluxei:
1275     case Intrinsic::riscv_vluxei_mask: {
1276       bool IsMasked = IntNo == Intrinsic::riscv_vloxei_mask ||
1277                       IntNo == Intrinsic::riscv_vluxei_mask;
1278       bool IsOrdered = IntNo == Intrinsic::riscv_vloxei ||
1279                        IntNo == Intrinsic::riscv_vloxei_mask;
1280 
1281       MVT VT = Node->getSimpleValueType(0);
1282       unsigned Log2SEW = Log2_32(VT.getScalarSizeInBits());
1283 
1284       unsigned CurOp = 2;
1285       // Masked intrinsic only have TU version pseduo instructions.
1286       bool IsTU = IsMasked || !Node->getOperand(CurOp).isUndef();
1287       SmallVector<SDValue, 8> Operands;
1288       if (IsTU)
1289         Operands.push_back(Node->getOperand(CurOp++));
1290       else
1291         // Skip the undef passthru operand for nomask TA version pseudo
1292         CurOp++;
1293 
1294       MVT IndexVT;
1295       addVectorLoadStoreOperands(Node, Log2SEW, DL, CurOp, IsMasked,
1296                                  /*IsStridedOrIndexed*/ true, Operands,
1297                                  /*IsLoad=*/true, &IndexVT);
1298 
1299       assert(VT.getVectorElementCount() == IndexVT.getVectorElementCount() &&
1300              "Element count mismatch");
1301 
1302       RISCVII::VLMUL LMUL = RISCVTargetLowering::getLMUL(VT);
1303       RISCVII::VLMUL IndexLMUL = RISCVTargetLowering::getLMUL(IndexVT);
1304       unsigned IndexLog2EEW = Log2_32(IndexVT.getScalarSizeInBits());
1305       if (IndexLog2EEW == 6 && !Subtarget->is64Bit()) {
1306         report_fatal_error("The V extension does not support EEW=64 for index "
1307                            "values when XLEN=32");
1308       }
1309       const RISCV::VLX_VSXPseudo *P = RISCV::getVLXPseudo(
1310           IsMasked, IsTU, IsOrdered, IndexLog2EEW, static_cast<unsigned>(LMUL),
1311           static_cast<unsigned>(IndexLMUL));
1312       MachineSDNode *Load =
1313           CurDAG->getMachineNode(P->Pseudo, DL, Node->getVTList(), Operands);
1314 
1315       if (auto *MemOp = dyn_cast<MemSDNode>(Node))
1316         CurDAG->setNodeMemRefs(Load, {MemOp->getMemOperand()});
1317 
1318       ReplaceNode(Node, Load);
1319       return;
1320     }
1321     case Intrinsic::riscv_vlm:
1322     case Intrinsic::riscv_vle:
1323     case Intrinsic::riscv_vle_mask:
1324     case Intrinsic::riscv_vlse:
1325     case Intrinsic::riscv_vlse_mask: {
1326       bool IsMasked = IntNo == Intrinsic::riscv_vle_mask ||
1327                       IntNo == Intrinsic::riscv_vlse_mask;
1328       bool IsStrided =
1329           IntNo == Intrinsic::riscv_vlse || IntNo == Intrinsic::riscv_vlse_mask;
1330 
1331       MVT VT = Node->getSimpleValueType(0);
1332       unsigned Log2SEW = Log2_32(VT.getScalarSizeInBits());
1333 
1334       unsigned CurOp = 2;
1335       // The riscv_vlm intrinsic are always tail agnostic and no passthru operand.
1336       bool HasPassthruOperand = IntNo != Intrinsic::riscv_vlm;
1337       // Masked intrinsic only have TU version pseduo instructions.
1338       bool IsTU = HasPassthruOperand &&
1339                   (IsMasked || !Node->getOperand(CurOp).isUndef());
1340       SmallVector<SDValue, 8> Operands;
1341       if (IsTU)
1342         Operands.push_back(Node->getOperand(CurOp++));
1343       else if (HasPassthruOperand)
1344         // Skip the undef passthru operand for nomask TA version pseudo
1345         CurOp++;
1346 
1347       addVectorLoadStoreOperands(Node, Log2SEW, DL, CurOp, IsMasked, IsStrided,
1348                                  Operands, /*IsLoad=*/true);
1349 
1350       RISCVII::VLMUL LMUL = RISCVTargetLowering::getLMUL(VT);
1351       const RISCV::VLEPseudo *P =
1352           RISCV::getVLEPseudo(IsMasked, IsTU, IsStrided, /*FF*/ false, Log2SEW,
1353                               static_cast<unsigned>(LMUL));
1354       MachineSDNode *Load =
1355           CurDAG->getMachineNode(P->Pseudo, DL, Node->getVTList(), Operands);
1356 
1357       if (auto *MemOp = dyn_cast<MemSDNode>(Node))
1358         CurDAG->setNodeMemRefs(Load, {MemOp->getMemOperand()});
1359 
1360       ReplaceNode(Node, Load);
1361       return;
1362     }
1363     case Intrinsic::riscv_vleff:
1364     case Intrinsic::riscv_vleff_mask: {
1365       bool IsMasked = IntNo == Intrinsic::riscv_vleff_mask;
1366 
1367       MVT VT = Node->getSimpleValueType(0);
1368       unsigned Log2SEW = Log2_32(VT.getScalarSizeInBits());
1369 
1370       unsigned CurOp = 2;
1371       // Masked intrinsic only have TU version pseduo instructions.
1372       bool IsTU = IsMasked || !Node->getOperand(CurOp).isUndef();
1373       SmallVector<SDValue, 7> Operands;
1374       if (IsTU)
1375         Operands.push_back(Node->getOperand(CurOp++));
1376       else
1377         // Skip the undef passthru operand for nomask TA version pseudo
1378         CurOp++;
1379 
1380       addVectorLoadStoreOperands(Node, Log2SEW, DL, CurOp, IsMasked,
1381                                  /*IsStridedOrIndexed*/ false, Operands,
1382                                  /*IsLoad=*/true);
1383 
1384       RISCVII::VLMUL LMUL = RISCVTargetLowering::getLMUL(VT);
1385       const RISCV::VLEPseudo *P =
1386           RISCV::getVLEPseudo(IsMasked, IsTU, /*Strided*/ false, /*FF*/ true,
1387                               Log2SEW, static_cast<unsigned>(LMUL));
1388       MachineSDNode *Load = CurDAG->getMachineNode(
1389           P->Pseudo, DL, Node->getVTList(), Operands);
1390       if (auto *MemOp = dyn_cast<MemSDNode>(Node))
1391         CurDAG->setNodeMemRefs(Load, {MemOp->getMemOperand()});
1392 
1393       ReplaceNode(Node, Load);
1394       return;
1395     }
1396     }
1397     break;
1398   }
1399   case ISD::INTRINSIC_VOID: {
1400     unsigned IntNo = cast<ConstantSDNode>(Node->getOperand(1))->getZExtValue();
1401     switch (IntNo) {
1402     case Intrinsic::riscv_vsseg2:
1403     case Intrinsic::riscv_vsseg3:
1404     case Intrinsic::riscv_vsseg4:
1405     case Intrinsic::riscv_vsseg5:
1406     case Intrinsic::riscv_vsseg6:
1407     case Intrinsic::riscv_vsseg7:
1408     case Intrinsic::riscv_vsseg8: {
1409       selectVSSEG(Node, /*IsMasked*/ false, /*IsStrided*/ false);
1410       return;
1411     }
1412     case Intrinsic::riscv_vsseg2_mask:
1413     case Intrinsic::riscv_vsseg3_mask:
1414     case Intrinsic::riscv_vsseg4_mask:
1415     case Intrinsic::riscv_vsseg5_mask:
1416     case Intrinsic::riscv_vsseg6_mask:
1417     case Intrinsic::riscv_vsseg7_mask:
1418     case Intrinsic::riscv_vsseg8_mask: {
1419       selectVSSEG(Node, /*IsMasked*/ true, /*IsStrided*/ false);
1420       return;
1421     }
1422     case Intrinsic::riscv_vssseg2:
1423     case Intrinsic::riscv_vssseg3:
1424     case Intrinsic::riscv_vssseg4:
1425     case Intrinsic::riscv_vssseg5:
1426     case Intrinsic::riscv_vssseg6:
1427     case Intrinsic::riscv_vssseg7:
1428     case Intrinsic::riscv_vssseg8: {
1429       selectVSSEG(Node, /*IsMasked*/ false, /*IsStrided*/ true);
1430       return;
1431     }
1432     case Intrinsic::riscv_vssseg2_mask:
1433     case Intrinsic::riscv_vssseg3_mask:
1434     case Intrinsic::riscv_vssseg4_mask:
1435     case Intrinsic::riscv_vssseg5_mask:
1436     case Intrinsic::riscv_vssseg6_mask:
1437     case Intrinsic::riscv_vssseg7_mask:
1438     case Intrinsic::riscv_vssseg8_mask: {
1439       selectVSSEG(Node, /*IsMasked*/ true, /*IsStrided*/ true);
1440       return;
1441     }
1442     case Intrinsic::riscv_vsoxseg2:
1443     case Intrinsic::riscv_vsoxseg3:
1444     case Intrinsic::riscv_vsoxseg4:
1445     case Intrinsic::riscv_vsoxseg5:
1446     case Intrinsic::riscv_vsoxseg6:
1447     case Intrinsic::riscv_vsoxseg7:
1448     case Intrinsic::riscv_vsoxseg8:
1449       selectVSXSEG(Node, /*IsMasked*/ false, /*IsOrdered*/ true);
1450       return;
1451     case Intrinsic::riscv_vsuxseg2:
1452     case Intrinsic::riscv_vsuxseg3:
1453     case Intrinsic::riscv_vsuxseg4:
1454     case Intrinsic::riscv_vsuxseg5:
1455     case Intrinsic::riscv_vsuxseg6:
1456     case Intrinsic::riscv_vsuxseg7:
1457     case Intrinsic::riscv_vsuxseg8:
1458       selectVSXSEG(Node, /*IsMasked*/ false, /*IsOrdered*/ false);
1459       return;
1460     case Intrinsic::riscv_vsoxseg2_mask:
1461     case Intrinsic::riscv_vsoxseg3_mask:
1462     case Intrinsic::riscv_vsoxseg4_mask:
1463     case Intrinsic::riscv_vsoxseg5_mask:
1464     case Intrinsic::riscv_vsoxseg6_mask:
1465     case Intrinsic::riscv_vsoxseg7_mask:
1466     case Intrinsic::riscv_vsoxseg8_mask:
1467       selectVSXSEG(Node, /*IsMasked*/ true, /*IsOrdered*/ true);
1468       return;
1469     case Intrinsic::riscv_vsuxseg2_mask:
1470     case Intrinsic::riscv_vsuxseg3_mask:
1471     case Intrinsic::riscv_vsuxseg4_mask:
1472     case Intrinsic::riscv_vsuxseg5_mask:
1473     case Intrinsic::riscv_vsuxseg6_mask:
1474     case Intrinsic::riscv_vsuxseg7_mask:
1475     case Intrinsic::riscv_vsuxseg8_mask:
1476       selectVSXSEG(Node, /*IsMasked*/ true, /*IsOrdered*/ false);
1477       return;
1478     case Intrinsic::riscv_vsoxei:
1479     case Intrinsic::riscv_vsoxei_mask:
1480     case Intrinsic::riscv_vsuxei:
1481     case Intrinsic::riscv_vsuxei_mask: {
1482       bool IsMasked = IntNo == Intrinsic::riscv_vsoxei_mask ||
1483                       IntNo == Intrinsic::riscv_vsuxei_mask;
1484       bool IsOrdered = IntNo == Intrinsic::riscv_vsoxei ||
1485                        IntNo == Intrinsic::riscv_vsoxei_mask;
1486 
1487       MVT VT = Node->getOperand(2)->getSimpleValueType(0);
1488       unsigned Log2SEW = Log2_32(VT.getScalarSizeInBits());
1489 
1490       unsigned CurOp = 2;
1491       SmallVector<SDValue, 8> Operands;
1492       Operands.push_back(Node->getOperand(CurOp++)); // Store value.
1493 
1494       MVT IndexVT;
1495       addVectorLoadStoreOperands(Node, Log2SEW, DL, CurOp, IsMasked,
1496                                  /*IsStridedOrIndexed*/ true, Operands,
1497                                  /*IsLoad=*/false, &IndexVT);
1498 
1499       assert(VT.getVectorElementCount() == IndexVT.getVectorElementCount() &&
1500              "Element count mismatch");
1501 
1502       RISCVII::VLMUL LMUL = RISCVTargetLowering::getLMUL(VT);
1503       RISCVII::VLMUL IndexLMUL = RISCVTargetLowering::getLMUL(IndexVT);
1504       unsigned IndexLog2EEW = Log2_32(IndexVT.getScalarSizeInBits());
1505       if (IndexLog2EEW == 6 && !Subtarget->is64Bit()) {
1506         report_fatal_error("The V extension does not support EEW=64 for index "
1507                            "values when XLEN=32");
1508       }
1509       const RISCV::VLX_VSXPseudo *P = RISCV::getVSXPseudo(
1510           IsMasked, /*TU*/ false, IsOrdered, IndexLog2EEW,
1511           static_cast<unsigned>(LMUL), static_cast<unsigned>(IndexLMUL));
1512       MachineSDNode *Store =
1513           CurDAG->getMachineNode(P->Pseudo, DL, Node->getVTList(), Operands);
1514 
1515       if (auto *MemOp = dyn_cast<MemSDNode>(Node))
1516         CurDAG->setNodeMemRefs(Store, {MemOp->getMemOperand()});
1517 
1518       ReplaceNode(Node, Store);
1519       return;
1520     }
1521     case Intrinsic::riscv_vsm:
1522     case Intrinsic::riscv_vse:
1523     case Intrinsic::riscv_vse_mask:
1524     case Intrinsic::riscv_vsse:
1525     case Intrinsic::riscv_vsse_mask: {
1526       bool IsMasked = IntNo == Intrinsic::riscv_vse_mask ||
1527                       IntNo == Intrinsic::riscv_vsse_mask;
1528       bool IsStrided =
1529           IntNo == Intrinsic::riscv_vsse || IntNo == Intrinsic::riscv_vsse_mask;
1530 
1531       MVT VT = Node->getOperand(2)->getSimpleValueType(0);
1532       unsigned Log2SEW = Log2_32(VT.getScalarSizeInBits());
1533 
1534       unsigned CurOp = 2;
1535       SmallVector<SDValue, 8> Operands;
1536       Operands.push_back(Node->getOperand(CurOp++)); // Store value.
1537 
1538       addVectorLoadStoreOperands(Node, Log2SEW, DL, CurOp, IsMasked, IsStrided,
1539                                  Operands);
1540 
1541       RISCVII::VLMUL LMUL = RISCVTargetLowering::getLMUL(VT);
1542       const RISCV::VSEPseudo *P = RISCV::getVSEPseudo(
1543           IsMasked, IsStrided, Log2SEW, static_cast<unsigned>(LMUL));
1544       MachineSDNode *Store =
1545           CurDAG->getMachineNode(P->Pseudo, DL, Node->getVTList(), Operands);
1546       if (auto *MemOp = dyn_cast<MemSDNode>(Node))
1547         CurDAG->setNodeMemRefs(Store, {MemOp->getMemOperand()});
1548 
1549       ReplaceNode(Node, Store);
1550       return;
1551     }
1552     }
1553     break;
1554   }
1555   case ISD::BITCAST: {
1556     MVT SrcVT = Node->getOperand(0).getSimpleValueType();
1557     // Just drop bitcasts between vectors if both are fixed or both are
1558     // scalable.
1559     if ((VT.isScalableVector() && SrcVT.isScalableVector()) ||
1560         (VT.isFixedLengthVector() && SrcVT.isFixedLengthVector())) {
1561       ReplaceUses(SDValue(Node, 0), Node->getOperand(0));
1562       CurDAG->RemoveDeadNode(Node);
1563       return;
1564     }
1565     break;
1566   }
1567   case ISD::INSERT_SUBVECTOR: {
1568     SDValue V = Node->getOperand(0);
1569     SDValue SubV = Node->getOperand(1);
1570     SDLoc DL(SubV);
1571     auto Idx = Node->getConstantOperandVal(2);
1572     MVT SubVecVT = SubV.getSimpleValueType();
1573 
1574     const RISCVTargetLowering &TLI = *Subtarget->getTargetLowering();
1575     MVT SubVecContainerVT = SubVecVT;
1576     // Establish the correct scalable-vector types for any fixed-length type.
1577     if (SubVecVT.isFixedLengthVector())
1578       SubVecContainerVT = TLI.getContainerForFixedLengthVector(SubVecVT);
1579     if (VT.isFixedLengthVector())
1580       VT = TLI.getContainerForFixedLengthVector(VT);
1581 
1582     const auto *TRI = Subtarget->getRegisterInfo();
1583     unsigned SubRegIdx;
1584     std::tie(SubRegIdx, Idx) =
1585         RISCVTargetLowering::decomposeSubvectorInsertExtractToSubRegs(
1586             VT, SubVecContainerVT, Idx, TRI);
1587 
1588     // If the Idx hasn't been completely eliminated then this is a subvector
1589     // insert which doesn't naturally align to a vector register. These must
1590     // be handled using instructions to manipulate the vector registers.
1591     if (Idx != 0)
1592       break;
1593 
1594     RISCVII::VLMUL SubVecLMUL = RISCVTargetLowering::getLMUL(SubVecContainerVT);
1595     bool IsSubVecPartReg = SubVecLMUL == RISCVII::VLMUL::LMUL_F2 ||
1596                            SubVecLMUL == RISCVII::VLMUL::LMUL_F4 ||
1597                            SubVecLMUL == RISCVII::VLMUL::LMUL_F8;
1598     (void)IsSubVecPartReg; // Silence unused variable warning without asserts.
1599     assert((!IsSubVecPartReg || V.isUndef()) &&
1600            "Expecting lowering to have created legal INSERT_SUBVECTORs when "
1601            "the subvector is smaller than a full-sized register");
1602 
1603     // If we haven't set a SubRegIdx, then we must be going between
1604     // equally-sized LMUL groups (e.g. VR -> VR). This can be done as a copy.
1605     if (SubRegIdx == RISCV::NoSubRegister) {
1606       unsigned InRegClassID = RISCVTargetLowering::getRegClassIDForVecVT(VT);
1607       assert(RISCVTargetLowering::getRegClassIDForVecVT(SubVecContainerVT) ==
1608                  InRegClassID &&
1609              "Unexpected subvector extraction");
1610       SDValue RC = CurDAG->getTargetConstant(InRegClassID, DL, XLenVT);
1611       SDNode *NewNode = CurDAG->getMachineNode(TargetOpcode::COPY_TO_REGCLASS,
1612                                                DL, VT, SubV, RC);
1613       ReplaceNode(Node, NewNode);
1614       return;
1615     }
1616 
1617     SDValue Insert = CurDAG->getTargetInsertSubreg(SubRegIdx, DL, VT, V, SubV);
1618     ReplaceNode(Node, Insert.getNode());
1619     return;
1620   }
1621   case ISD::EXTRACT_SUBVECTOR: {
1622     SDValue V = Node->getOperand(0);
1623     auto Idx = Node->getConstantOperandVal(1);
1624     MVT InVT = V.getSimpleValueType();
1625     SDLoc DL(V);
1626 
1627     const RISCVTargetLowering &TLI = *Subtarget->getTargetLowering();
1628     MVT SubVecContainerVT = VT;
1629     // Establish the correct scalable-vector types for any fixed-length type.
1630     if (VT.isFixedLengthVector())
1631       SubVecContainerVT = TLI.getContainerForFixedLengthVector(VT);
1632     if (InVT.isFixedLengthVector())
1633       InVT = TLI.getContainerForFixedLengthVector(InVT);
1634 
1635     const auto *TRI = Subtarget->getRegisterInfo();
1636     unsigned SubRegIdx;
1637     std::tie(SubRegIdx, Idx) =
1638         RISCVTargetLowering::decomposeSubvectorInsertExtractToSubRegs(
1639             InVT, SubVecContainerVT, Idx, TRI);
1640 
1641     // If the Idx hasn't been completely eliminated then this is a subvector
1642     // extract which doesn't naturally align to a vector register. These must
1643     // be handled using instructions to manipulate the vector registers.
1644     if (Idx != 0)
1645       break;
1646 
1647     // If we haven't set a SubRegIdx, then we must be going between
1648     // equally-sized LMUL types (e.g. VR -> VR). This can be done as a copy.
1649     if (SubRegIdx == RISCV::NoSubRegister) {
1650       unsigned InRegClassID = RISCVTargetLowering::getRegClassIDForVecVT(InVT);
1651       assert(RISCVTargetLowering::getRegClassIDForVecVT(SubVecContainerVT) ==
1652                  InRegClassID &&
1653              "Unexpected subvector extraction");
1654       SDValue RC = CurDAG->getTargetConstant(InRegClassID, DL, XLenVT);
1655       SDNode *NewNode =
1656           CurDAG->getMachineNode(TargetOpcode::COPY_TO_REGCLASS, DL, VT, V, RC);
1657       ReplaceNode(Node, NewNode);
1658       return;
1659     }
1660 
1661     SDValue Extract = CurDAG->getTargetExtractSubreg(SubRegIdx, DL, VT, V);
1662     ReplaceNode(Node, Extract.getNode());
1663     return;
1664   }
1665   case ISD::SPLAT_VECTOR:
1666   case RISCVISD::VMV_S_X_VL:
1667   case RISCVISD::VFMV_S_F_VL:
1668   case RISCVISD::VMV_V_X_VL:
1669   case RISCVISD::VFMV_V_F_VL: {
1670     // Try to match splat of a scalar load to a strided load with stride of x0.
1671     bool IsScalarMove = Node->getOpcode() == RISCVISD::VMV_S_X_VL ||
1672                         Node->getOpcode() == RISCVISD::VFMV_S_F_VL;
1673     bool HasPassthruOperand = Node->getOpcode() != ISD::SPLAT_VECTOR;
1674     if (HasPassthruOperand && !Node->getOperand(0).isUndef())
1675       break;
1676     SDValue Src = HasPassthruOperand ? Node->getOperand(1) : Node->getOperand(0);
1677     auto *Ld = dyn_cast<LoadSDNode>(Src);
1678     if (!Ld)
1679       break;
1680     EVT MemVT = Ld->getMemoryVT();
1681     // The memory VT should be the same size as the element type.
1682     if (MemVT.getStoreSize() != VT.getVectorElementType().getStoreSize())
1683       break;
1684     if (!IsProfitableToFold(Src, Node, Node) ||
1685         !IsLegalToFold(Src, Node, Node, TM.getOptLevel()))
1686       break;
1687 
1688     SDValue VL;
1689     if (Node->getOpcode() == ISD::SPLAT_VECTOR)
1690       VL = CurDAG->getTargetConstant(RISCV::VLMaxSentinel, DL, XLenVT);
1691     else if (IsScalarMove) {
1692       // We could deal with more VL if we update the VSETVLI insert pass to
1693       // avoid introducing more VSETVLI.
1694       if (!isOneConstant(Node->getOperand(2)))
1695         break;
1696       selectVLOp(Node->getOperand(2), VL);
1697     } else
1698       selectVLOp(Node->getOperand(2), VL);
1699 
1700     unsigned Log2SEW = Log2_32(VT.getScalarSizeInBits());
1701     SDValue SEW = CurDAG->getTargetConstant(Log2SEW, DL, XLenVT);
1702 
1703     SDValue Operands[] = {Ld->getBasePtr(),
1704                           CurDAG->getRegister(RISCV::X0, XLenVT), VL, SEW,
1705                           Ld->getChain()};
1706 
1707     RISCVII::VLMUL LMUL = RISCVTargetLowering::getLMUL(VT);
1708     const RISCV::VLEPseudo *P = RISCV::getVLEPseudo(
1709         /*IsMasked*/ false, /*IsTU*/ false, /*IsStrided*/ true, /*FF*/ false,
1710         Log2SEW, static_cast<unsigned>(LMUL));
1711     MachineSDNode *Load =
1712         CurDAG->getMachineNode(P->Pseudo, DL, Node->getVTList(), Operands);
1713 
1714     CurDAG->setNodeMemRefs(Load, {Ld->getMemOperand()});
1715 
1716     ReplaceNode(Node, Load);
1717     return;
1718   }
1719   }
1720 
1721   // Select the default instruction.
1722   SelectCode(Node);
1723 }
1724 
1725 bool RISCVDAGToDAGISel::SelectInlineAsmMemoryOperand(
1726     const SDValue &Op, unsigned ConstraintID, std::vector<SDValue> &OutOps) {
1727   switch (ConstraintID) {
1728   case InlineAsm::Constraint_m:
1729     // We just support simple memory operands that have a single address
1730     // operand and need no special handling.
1731     OutOps.push_back(Op);
1732     return false;
1733   case InlineAsm::Constraint_A:
1734     OutOps.push_back(Op);
1735     return false;
1736   default:
1737     break;
1738   }
1739 
1740   return true;
1741 }
1742 
1743 bool RISCVDAGToDAGISel::SelectAddrFrameIndex(SDValue Addr, SDValue &Base,
1744                                              SDValue &Offset) {
1745   if (auto *FIN = dyn_cast<FrameIndexSDNode>(Addr)) {
1746     Base = CurDAG->getTargetFrameIndex(FIN->getIndex(), Subtarget->getXLenVT());
1747     Offset = CurDAG->getTargetConstant(0, SDLoc(Addr), Subtarget->getXLenVT());
1748     return true;
1749   }
1750 
1751   return false;
1752 }
1753 
1754 // Select a frame index and an optional immediate offset from an ADD or OR.
1755 bool RISCVDAGToDAGISel::SelectFrameAddrRegImm(SDValue Addr, SDValue &Base,
1756                                               SDValue &Offset) {
1757   if (SelectAddrFrameIndex(Addr, Base, Offset))
1758     return true;
1759 
1760   if (!CurDAG->isBaseWithConstantOffset(Addr))
1761     return false;
1762 
1763   if (auto *FIN = dyn_cast<FrameIndexSDNode>(Addr.getOperand(0))) {
1764     int64_t CVal = cast<ConstantSDNode>(Addr.getOperand(1))->getSExtValue();
1765     if (isInt<12>(CVal)) {
1766       Base = CurDAG->getTargetFrameIndex(FIN->getIndex(),
1767                                          Subtarget->getXLenVT());
1768       Offset = CurDAG->getTargetConstant(CVal, SDLoc(Addr),
1769                                          Subtarget->getXLenVT());
1770       return true;
1771     }
1772   }
1773 
1774   return false;
1775 }
1776 
1777 // Fold constant addresses.
1778 static bool selectConstantAddr(SelectionDAG *CurDAG, const SDLoc &DL,
1779                                const MVT VT, const RISCVSubtarget *Subtarget,
1780                                SDValue Addr, SDValue &Base, SDValue &Offset) {
1781   if (!isa<ConstantSDNode>(Addr))
1782     return false;
1783 
1784   int64_t CVal = cast<ConstantSDNode>(Addr)->getSExtValue();
1785 
1786   // If the constant is a simm12, we can fold the whole constant and use X0 as
1787   // the base. If the constant can be materialized with LUI+simm12, use LUI as
1788   // the base. We can't use generateInstSeq because it favors LUI+ADDIW.
1789   int64_t Lo12 = SignExtend64<12>(CVal);
1790   int64_t Hi = (uint64_t)CVal - (uint64_t)Lo12;
1791   if (!Subtarget->is64Bit() || isInt<32>(Hi)) {
1792     if (Hi) {
1793       int64_t Hi20 = (Hi >> 12) & 0xfffff;
1794       Base = SDValue(
1795           CurDAG->getMachineNode(RISCV::LUI, DL, VT,
1796                                  CurDAG->getTargetConstant(Hi20, DL, VT)),
1797           0);
1798     } else {
1799       Base = CurDAG->getRegister(RISCV::X0, VT);
1800     }
1801     Offset = CurDAG->getTargetConstant(Lo12, DL, VT);
1802     return true;
1803   }
1804 
1805   // Ask how constant materialization would handle this constant.
1806   RISCVMatInt::InstSeq Seq =
1807       RISCVMatInt::generateInstSeq(CVal, Subtarget->getFeatureBits());
1808 
1809   // If the last instruction would be an ADDI, we can fold its immediate and
1810   // emit the rest of the sequence as the base.
1811   if (Seq.back().Opc != RISCV::ADDI)
1812     return false;
1813   Lo12 = Seq.back().Imm;
1814 
1815   // Drop the last instruction.
1816   Seq.pop_back();
1817   assert(!Seq.empty() && "Expected more instructions in sequence");
1818 
1819   Base = SDValue(selectImmSeq(CurDAG, DL, VT, Seq), 0);
1820   Offset = CurDAG->getTargetConstant(Lo12, DL, VT);
1821   return true;
1822 }
1823 
1824 // Is this ADD instruction only used as the base pointer of scalar loads and
1825 // stores?
1826 static bool isWorthFoldingAdd(SDValue Add) {
1827   for (auto Use : Add->uses()) {
1828     if (Use->getOpcode() != ISD::LOAD && Use->getOpcode() != ISD::STORE &&
1829         Use->getOpcode() != ISD::ATOMIC_LOAD &&
1830         Use->getOpcode() != ISD::ATOMIC_STORE)
1831       return false;
1832     EVT VT = cast<MemSDNode>(Use)->getMemoryVT();
1833     if (!VT.isScalarInteger() && VT != MVT::f16 && VT != MVT::f32 &&
1834         VT != MVT::f64)
1835       return false;
1836     // Don't allow stores of the value. It must be used as the address.
1837     if (Use->getOpcode() == ISD::STORE &&
1838         cast<StoreSDNode>(Use)->getValue() == Add)
1839       return false;
1840     if (Use->getOpcode() == ISD::ATOMIC_STORE &&
1841         cast<AtomicSDNode>(Use)->getVal() == Add)
1842       return false;
1843   }
1844 
1845   return true;
1846 }
1847 
1848 bool RISCVDAGToDAGISel::SelectAddrRegImm(SDValue Addr, SDValue &Base,
1849                                          SDValue &Offset) {
1850   if (SelectAddrFrameIndex(Addr, Base, Offset))
1851     return true;
1852 
1853   SDLoc DL(Addr);
1854   MVT VT = Addr.getSimpleValueType();
1855 
1856   if (Addr.getOpcode() == RISCVISD::ADD_LO) {
1857     Base = Addr.getOperand(0);
1858     Offset = Addr.getOperand(1);
1859     return true;
1860   }
1861 
1862   if (CurDAG->isBaseWithConstantOffset(Addr)) {
1863     int64_t CVal = cast<ConstantSDNode>(Addr.getOperand(1))->getSExtValue();
1864     if (isInt<12>(CVal)) {
1865       Base = Addr.getOperand(0);
1866       if (Base.getOpcode() == RISCVISD::ADD_LO) {
1867         SDValue LoOperand = Base.getOperand(1);
1868         if (auto *GA = dyn_cast<GlobalAddressSDNode>(LoOperand)) {
1869           // If the Lo in (ADD_LO hi, lo) is a global variable's address
1870           // (its low part, really), then we can rely on the alignment of that
1871           // variable to provide a margin of safety before low part can overflow
1872           // the 12 bits of the load/store offset. Check if CVal falls within
1873           // that margin; if so (low part + CVal) can't overflow.
1874           const DataLayout &DL = CurDAG->getDataLayout();
1875           Align Alignment = commonAlignment(
1876               GA->getGlobal()->getPointerAlignment(DL), GA->getOffset());
1877           if (CVal == 0 || Alignment > CVal) {
1878             int64_t CombinedOffset = CVal + GA->getOffset();
1879             Base = Base.getOperand(0);
1880             Offset = CurDAG->getTargetGlobalAddress(
1881                 GA->getGlobal(), SDLoc(LoOperand), LoOperand.getValueType(),
1882                 CombinedOffset, GA->getTargetFlags());
1883             return true;
1884           }
1885         }
1886       }
1887 
1888       if (auto *FIN = dyn_cast<FrameIndexSDNode>(Base))
1889         Base = CurDAG->getTargetFrameIndex(FIN->getIndex(), VT);
1890       Offset = CurDAG->getTargetConstant(CVal, DL, VT);
1891       return true;
1892     }
1893   }
1894 
1895   // Handle ADD with large immediates.
1896   if (Addr.getOpcode() == ISD::ADD && isa<ConstantSDNode>(Addr.getOperand(1))) {
1897     int64_t CVal = cast<ConstantSDNode>(Addr.getOperand(1))->getSExtValue();
1898     assert(!isInt<12>(CVal) && "simm12 not already handled?");
1899 
1900     // Handle immediates in the range [-4096,-2049] or [2048, 4094]. We can use
1901     // an ADDI for part of the offset and fold the rest into the load/store.
1902     // This mirrors the AddiPair PatFrag in RISCVInstrInfo.td.
1903     if (isInt<12>(CVal / 2) && isInt<12>(CVal - CVal / 2)) {
1904       int64_t Adj = CVal < 0 ? -2048 : 2047;
1905       Base = SDValue(
1906           CurDAG->getMachineNode(RISCV::ADDI, DL, VT, Addr.getOperand(0),
1907                                  CurDAG->getTargetConstant(Adj, DL, VT)),
1908           0);
1909       Offset = CurDAG->getTargetConstant(CVal - Adj, DL, VT);
1910       return true;
1911     }
1912 
1913     // For larger immediates, we might be able to save one instruction from
1914     // constant materialization by folding the Lo12 bits of the immediate into
1915     // the address. We should only do this if the ADD is only used by loads and
1916     // stores that can fold the lo12 bits. Otherwise, the ADD will get iseled
1917     // separately with the full materialized immediate creating extra
1918     // instructions.
1919     if (isWorthFoldingAdd(Addr) &&
1920         selectConstantAddr(CurDAG, DL, VT, Subtarget, Addr.getOperand(1), Base,
1921                            Offset)) {
1922       // Insert an ADD instruction with the materialized Hi52 bits.
1923       Base = SDValue(
1924           CurDAG->getMachineNode(RISCV::ADD, DL, VT, Addr.getOperand(0), Base),
1925           0);
1926       return true;
1927     }
1928   }
1929 
1930   if (selectConstantAddr(CurDAG, DL, VT, Subtarget, Addr, Base, Offset))
1931     return true;
1932 
1933   Base = Addr;
1934   Offset = CurDAG->getTargetConstant(0, DL, VT);
1935   return true;
1936 }
1937 
1938 bool RISCVDAGToDAGISel::selectShiftMask(SDValue N, unsigned ShiftWidth,
1939                                         SDValue &ShAmt) {
1940   // Shift instructions on RISCV only read the lower 5 or 6 bits of the shift
1941   // amount. If there is an AND on the shift amount, we can bypass it if it
1942   // doesn't affect any of those bits.
1943   if (N.getOpcode() == ISD::AND && isa<ConstantSDNode>(N.getOperand(1))) {
1944     const APInt &AndMask = N->getConstantOperandAPInt(1);
1945 
1946     // Since the max shift amount is a power of 2 we can subtract 1 to make a
1947     // mask that covers the bits needed to represent all shift amounts.
1948     assert(isPowerOf2_32(ShiftWidth) && "Unexpected max shift amount!");
1949     APInt ShMask(AndMask.getBitWidth(), ShiftWidth - 1);
1950 
1951     if (ShMask.isSubsetOf(AndMask)) {
1952       ShAmt = N.getOperand(0);
1953       return true;
1954     }
1955 
1956     // SimplifyDemandedBits may have optimized the mask so try restoring any
1957     // bits that are known zero.
1958     KnownBits Known = CurDAG->computeKnownBits(N->getOperand(0));
1959     if (ShMask.isSubsetOf(AndMask | Known.Zero)) {
1960       ShAmt = N.getOperand(0);
1961       return true;
1962     }
1963   } else if (N.getOpcode() == ISD::SUB &&
1964              isa<ConstantSDNode>(N.getOperand(0))) {
1965     uint64_t Imm = N.getConstantOperandVal(0);
1966     // If we are shifting by N-X where N == 0 mod Size, then just shift by -X to
1967     // generate a NEG instead of a SUB of a constant.
1968     if (Imm != 0 && Imm % ShiftWidth == 0) {
1969       SDLoc DL(N);
1970       EVT VT = N.getValueType();
1971       SDValue Zero = CurDAG->getRegister(RISCV::X0, VT);
1972       unsigned NegOpc = VT == MVT::i64 ? RISCV::SUBW : RISCV::SUB;
1973       MachineSDNode *Neg = CurDAG->getMachineNode(NegOpc, DL, VT, Zero,
1974                                                   N.getOperand(1));
1975       ShAmt = SDValue(Neg, 0);
1976       return true;
1977     }
1978   }
1979 
1980   ShAmt = N;
1981   return true;
1982 }
1983 
1984 bool RISCVDAGToDAGISel::selectSExti32(SDValue N, SDValue &Val) {
1985   if (N.getOpcode() == ISD::SIGN_EXTEND_INREG &&
1986       cast<VTSDNode>(N.getOperand(1))->getVT() == MVT::i32) {
1987     Val = N.getOperand(0);
1988     return true;
1989   }
1990   MVT VT = N.getSimpleValueType();
1991   if (CurDAG->ComputeNumSignBits(N) > (VT.getSizeInBits() - 32)) {
1992     Val = N;
1993     return true;
1994   }
1995 
1996   return false;
1997 }
1998 
1999 bool RISCVDAGToDAGISel::selectZExti32(SDValue N, SDValue &Val) {
2000   if (N.getOpcode() == ISD::AND) {
2001     auto *C = dyn_cast<ConstantSDNode>(N.getOperand(1));
2002     if (C && C->getZExtValue() == UINT64_C(0xFFFFFFFF)) {
2003       Val = N.getOperand(0);
2004       return true;
2005     }
2006   }
2007   MVT VT = N.getSimpleValueType();
2008   APInt Mask = APInt::getHighBitsSet(VT.getSizeInBits(), 32);
2009   if (CurDAG->MaskedValueIsZero(N, Mask)) {
2010     Val = N;
2011     return true;
2012   }
2013 
2014   return false;
2015 }
2016 
2017 /// Look for various patterns that can be done with a SHL that can be folded
2018 /// into a SHXADD. \p ShAmt contains 1, 2, or 3 and is set based on which
2019 /// SHXADD we are trying to match.
2020 bool RISCVDAGToDAGISel::selectSHXADDOp(SDValue N, unsigned ShAmt,
2021                                        SDValue &Val) {
2022   if (N.getOpcode() == ISD::AND && isa<ConstantSDNode>(N.getOperand(1))) {
2023     SDValue N0 = N.getOperand(0);
2024 
2025     bool LeftShift = N0.getOpcode() == ISD::SHL;
2026     if ((LeftShift || N0.getOpcode() == ISD::SRL) &&
2027         isa<ConstantSDNode>(N0.getOperand(1))) {
2028       uint64_t Mask = N.getConstantOperandVal(1);
2029       unsigned C2 = N0.getConstantOperandVal(1);
2030 
2031       unsigned XLen = Subtarget->getXLen();
2032       if (LeftShift)
2033         Mask &= maskTrailingZeros<uint64_t>(C2);
2034       else
2035         Mask &= maskTrailingOnes<uint64_t>(XLen - C2);
2036 
2037       // Look for (and (shl y, c2), c1) where c1 is a shifted mask with no
2038       // leading zeros and c3 trailing zeros. We can use an SRLI by c2+c3
2039       // followed by a SHXADD with c3 for the X amount.
2040       if (isShiftedMask_64(Mask)) {
2041         unsigned Leading = XLen - (64 - countLeadingZeros(Mask));
2042         unsigned Trailing = countTrailingZeros(Mask);
2043         if (LeftShift && Leading == 0 && C2 < Trailing && Trailing == ShAmt) {
2044           SDLoc DL(N);
2045           EVT VT = N.getValueType();
2046           Val = SDValue(CurDAG->getMachineNode(
2047                             RISCV::SRLI, DL, VT, N0.getOperand(0),
2048                             CurDAG->getTargetConstant(Trailing - C2, DL, VT)),
2049                         0);
2050           return true;
2051         }
2052         // Look for (and (shr y, c2), c1) where c1 is a shifted mask with c2
2053         // leading zeros and c3 trailing zeros. We can use an SRLI by C3
2054         // followed by a SHXADD using c3 for the X amount.
2055         if (!LeftShift && Leading == C2 && Trailing == ShAmt) {
2056           SDLoc DL(N);
2057           EVT VT = N.getValueType();
2058           Val = SDValue(
2059               CurDAG->getMachineNode(
2060                   RISCV::SRLI, DL, VT, N0.getOperand(0),
2061                   CurDAG->getTargetConstant(Leading + Trailing, DL, VT)),
2062               0);
2063           return true;
2064         }
2065       }
2066     }
2067   }
2068 
2069   bool LeftShift = N.getOpcode() == ISD::SHL;
2070   if ((LeftShift || N.getOpcode() == ISD::SRL) &&
2071       isa<ConstantSDNode>(N.getOperand(1))) {
2072     SDValue N0 = N.getOperand(0);
2073     if (N0.getOpcode() == ISD::AND && N0.hasOneUse() &&
2074         isa<ConstantSDNode>(N0.getOperand(1))) {
2075       uint64_t Mask = N0.getConstantOperandVal(1);
2076       if (isShiftedMask_64(Mask)) {
2077         unsigned C1 = N.getConstantOperandVal(1);
2078         unsigned XLen = Subtarget->getXLen();
2079         unsigned Leading = XLen - (64 - countLeadingZeros(Mask));
2080         unsigned Trailing = countTrailingZeros(Mask);
2081         // Look for (shl (and X, Mask), C1) where Mask has 32 leading zeros and
2082         // C3 trailing zeros. If C1+C3==ShAmt we can use SRLIW+SHXADD.
2083         if (LeftShift && Leading == 32 && Trailing > 0 &&
2084             (Trailing + C1) == ShAmt) {
2085           SDLoc DL(N);
2086           EVT VT = N.getValueType();
2087           Val = SDValue(CurDAG->getMachineNode(
2088                             RISCV::SRLIW, DL, VT, N0.getOperand(0),
2089                             CurDAG->getTargetConstant(Trailing, DL, VT)),
2090                         0);
2091           return true;
2092         }
2093         // Look for (srl (and X, Mask), C1) where Mask has 32 leading zeros and
2094         // C3 trailing zeros. If C3-C1==ShAmt we can use SRLIW+SHXADD.
2095         if (!LeftShift && Leading == 32 && Trailing > C1 &&
2096             (Trailing - C1) == ShAmt) {
2097           SDLoc DL(N);
2098           EVT VT = N.getValueType();
2099           Val = SDValue(CurDAG->getMachineNode(
2100                             RISCV::SRLIW, DL, VT, N0.getOperand(0),
2101                             CurDAG->getTargetConstant(Trailing, DL, VT)),
2102                         0);
2103           return true;
2104         }
2105       }
2106     }
2107   }
2108 
2109   return false;
2110 }
2111 
2112 // Return true if all users of this SDNode* only consume the lower \p Bits.
2113 // This can be used to form W instructions for add/sub/mul/shl even when the
2114 // root isn't a sext_inreg. This can allow the ADDW/SUBW/MULW/SLLIW to CSE if
2115 // SimplifyDemandedBits has made it so some users see a sext_inreg and some
2116 // don't. The sext_inreg+add/sub/mul/shl will get selected, but still leave
2117 // the add/sub/mul/shl to become non-W instructions. By checking the users we
2118 // may be able to use a W instruction and CSE with the other instruction if
2119 // this has happened. We could try to detect that the CSE opportunity exists
2120 // before doing this, but that would be more complicated.
2121 // TODO: Does this need to look through AND/OR/XOR to their users to find more
2122 // opportunities.
2123 bool RISCVDAGToDAGISel::hasAllNBitUsers(SDNode *Node, unsigned Bits) const {
2124   assert((Node->getOpcode() == ISD::ADD || Node->getOpcode() == ISD::SUB ||
2125           Node->getOpcode() == ISD::MUL || Node->getOpcode() == ISD::SHL ||
2126           Node->getOpcode() == ISD::SRL ||
2127           Node->getOpcode() == ISD::SIGN_EXTEND_INREG ||
2128           Node->getOpcode() == RISCVISD::GREV ||
2129           Node->getOpcode() == RISCVISD::GORC ||
2130           isa<ConstantSDNode>(Node)) &&
2131          "Unexpected opcode");
2132 
2133   for (auto UI = Node->use_begin(), UE = Node->use_end(); UI != UE; ++UI) {
2134     SDNode *User = *UI;
2135     // Users of this node should have already been instruction selected
2136     if (!User->isMachineOpcode())
2137       return false;
2138 
2139     // TODO: Add more opcodes?
2140     switch (User->getMachineOpcode()) {
2141     default:
2142       return false;
2143     case RISCV::ADDW:
2144     case RISCV::ADDIW:
2145     case RISCV::SUBW:
2146     case RISCV::MULW:
2147     case RISCV::SLLW:
2148     case RISCV::SLLIW:
2149     case RISCV::SRAW:
2150     case RISCV::SRAIW:
2151     case RISCV::SRLW:
2152     case RISCV::SRLIW:
2153     case RISCV::DIVW:
2154     case RISCV::DIVUW:
2155     case RISCV::REMW:
2156     case RISCV::REMUW:
2157     case RISCV::ROLW:
2158     case RISCV::RORW:
2159     case RISCV::RORIW:
2160     case RISCV::CLZW:
2161     case RISCV::CTZW:
2162     case RISCV::CPOPW:
2163     case RISCV::SLLI_UW:
2164     case RISCV::FMV_W_X:
2165     case RISCV::FCVT_H_W:
2166     case RISCV::FCVT_H_WU:
2167     case RISCV::FCVT_S_W:
2168     case RISCV::FCVT_S_WU:
2169     case RISCV::FCVT_D_W:
2170     case RISCV::FCVT_D_WU:
2171       if (Bits < 32)
2172         return false;
2173       break;
2174     case RISCV::SLLI:
2175       // SLLI only uses the lower (XLen - ShAmt) bits.
2176       if (Bits < Subtarget->getXLen() - User->getConstantOperandVal(1))
2177         return false;
2178       break;
2179     case RISCV::ANDI:
2180       if (Bits < (64 - countLeadingZeros(User->getConstantOperandVal(1))))
2181         return false;
2182       break;
2183     case RISCV::SEXT_B:
2184       if (Bits < 8)
2185         return false;
2186       break;
2187     case RISCV::SEXT_H:
2188     case RISCV::FMV_H_X:
2189     case RISCV::ZEXT_H_RV32:
2190     case RISCV::ZEXT_H_RV64:
2191       if (Bits < 16)
2192         return false;
2193       break;
2194     case RISCV::ADD_UW:
2195     case RISCV::SH1ADD_UW:
2196     case RISCV::SH2ADD_UW:
2197     case RISCV::SH3ADD_UW:
2198       // The first operand to add.uw/shXadd.uw is implicitly zero extended from
2199       // 32 bits.
2200       if (UI.getOperandNo() != 0 || Bits < 32)
2201         return false;
2202       break;
2203     case RISCV::SB:
2204       if (UI.getOperandNo() != 0 || Bits < 8)
2205         return false;
2206       break;
2207     case RISCV::SH:
2208       if (UI.getOperandNo() != 0 || Bits < 16)
2209         return false;
2210       break;
2211     case RISCV::SW:
2212       if (UI.getOperandNo() != 0 || Bits < 32)
2213         return false;
2214       break;
2215     }
2216   }
2217 
2218   return true;
2219 }
2220 
2221 // Select VL as a 5 bit immediate or a value that will become a register. This
2222 // allows us to choose betwen VSETIVLI or VSETVLI later.
2223 bool RISCVDAGToDAGISel::selectVLOp(SDValue N, SDValue &VL) {
2224   auto *C = dyn_cast<ConstantSDNode>(N);
2225   if (C && isUInt<5>(C->getZExtValue())) {
2226     VL = CurDAG->getTargetConstant(C->getZExtValue(), SDLoc(N),
2227                                    N->getValueType(0));
2228   } else if (C && C->isAllOnesValue()) {
2229     // Treat all ones as VLMax.
2230     VL = CurDAG->getTargetConstant(RISCV::VLMaxSentinel, SDLoc(N),
2231                                    N->getValueType(0));
2232   } else if (isa<RegisterSDNode>(N) &&
2233              cast<RegisterSDNode>(N)->getReg() == RISCV::X0) {
2234     // All our VL operands use an operand that allows GPRNoX0 or an immediate
2235     // as the register class. Convert X0 to a special immediate to pass the
2236     // MachineVerifier. This is recognized specially by the vsetvli insertion
2237     // pass.
2238     VL = CurDAG->getTargetConstant(RISCV::VLMaxSentinel, SDLoc(N),
2239                                    N->getValueType(0));
2240   } else {
2241     VL = N;
2242   }
2243 
2244   return true;
2245 }
2246 
2247 bool RISCVDAGToDAGISel::selectVSplat(SDValue N, SDValue &SplatVal) {
2248   if (N.getOpcode() != RISCVISD::VMV_V_X_VL || !N.getOperand(0).isUndef())
2249     return false;
2250   SplatVal = N.getOperand(1);
2251   return true;
2252 }
2253 
2254 using ValidateFn = bool (*)(int64_t);
2255 
2256 static bool selectVSplatSimmHelper(SDValue N, SDValue &SplatVal,
2257                                    SelectionDAG &DAG,
2258                                    const RISCVSubtarget &Subtarget,
2259                                    ValidateFn ValidateImm) {
2260   if (N.getOpcode() != RISCVISD::VMV_V_X_VL || !N.getOperand(0).isUndef() ||
2261       !isa<ConstantSDNode>(N.getOperand(1)))
2262     return false;
2263 
2264   int64_t SplatImm =
2265       cast<ConstantSDNode>(N.getOperand(1))->getSExtValue();
2266 
2267   // The semantics of RISCVISD::VMV_V_X_VL is that when the operand
2268   // type is wider than the resulting vector element type: an implicit
2269   // truncation first takes place. Therefore, perform a manual
2270   // truncation/sign-extension in order to ignore any truncated bits and catch
2271   // any zero-extended immediate.
2272   // For example, we wish to match (i8 -1) -> (XLenVT 255) as a simm5 by first
2273   // sign-extending to (XLenVT -1).
2274   MVT XLenVT = Subtarget.getXLenVT();
2275   assert(XLenVT == N.getOperand(1).getSimpleValueType() &&
2276          "Unexpected splat operand type");
2277   MVT EltVT = N.getSimpleValueType().getVectorElementType();
2278   if (EltVT.bitsLT(XLenVT))
2279     SplatImm = SignExtend64(SplatImm, EltVT.getSizeInBits());
2280 
2281   if (!ValidateImm(SplatImm))
2282     return false;
2283 
2284   SplatVal = DAG.getTargetConstant(SplatImm, SDLoc(N), XLenVT);
2285   return true;
2286 }
2287 
2288 bool RISCVDAGToDAGISel::selectVSplatSimm5(SDValue N, SDValue &SplatVal) {
2289   return selectVSplatSimmHelper(N, SplatVal, *CurDAG, *Subtarget,
2290                                 [](int64_t Imm) { return isInt<5>(Imm); });
2291 }
2292 
2293 bool RISCVDAGToDAGISel::selectVSplatSimm5Plus1(SDValue N, SDValue &SplatVal) {
2294   return selectVSplatSimmHelper(
2295       N, SplatVal, *CurDAG, *Subtarget,
2296       [](int64_t Imm) { return (isInt<5>(Imm) && Imm != -16) || Imm == 16; });
2297 }
2298 
2299 bool RISCVDAGToDAGISel::selectVSplatSimm5Plus1NonZero(SDValue N,
2300                                                       SDValue &SplatVal) {
2301   return selectVSplatSimmHelper(
2302       N, SplatVal, *CurDAG, *Subtarget, [](int64_t Imm) {
2303         return Imm != 0 && ((isInt<5>(Imm) && Imm != -16) || Imm == 16);
2304       });
2305 }
2306 
2307 bool RISCVDAGToDAGISel::selectVSplatUimm5(SDValue N, SDValue &SplatVal) {
2308   if (N.getOpcode() != RISCVISD::VMV_V_X_VL || !N.getOperand(0).isUndef() ||
2309       !isa<ConstantSDNode>(N.getOperand(1)))
2310     return false;
2311 
2312   int64_t SplatImm =
2313       cast<ConstantSDNode>(N.getOperand(1))->getSExtValue();
2314 
2315   if (!isUInt<5>(SplatImm))
2316     return false;
2317 
2318   SplatVal =
2319       CurDAG->getTargetConstant(SplatImm, SDLoc(N), Subtarget->getXLenVT());
2320 
2321   return true;
2322 }
2323 
2324 bool RISCVDAGToDAGISel::selectRVVSimm5(SDValue N, unsigned Width,
2325                                        SDValue &Imm) {
2326   if (auto *C = dyn_cast<ConstantSDNode>(N)) {
2327     int64_t ImmVal = SignExtend64(C->getSExtValue(), Width);
2328 
2329     if (!isInt<5>(ImmVal))
2330       return false;
2331 
2332     Imm = CurDAG->getTargetConstant(ImmVal, SDLoc(N), Subtarget->getXLenVT());
2333     return true;
2334   }
2335 
2336   return false;
2337 }
2338 
2339 // Try to remove sext.w if the input is a W instruction or can be made into
2340 // a W instruction cheaply.
2341 bool RISCVDAGToDAGISel::doPeepholeSExtW(SDNode *N) {
2342   // Look for the sext.w pattern, addiw rd, rs1, 0.
2343   if (N->getMachineOpcode() != RISCV::ADDIW ||
2344       !isNullConstant(N->getOperand(1)))
2345     return false;
2346 
2347   SDValue N0 = N->getOperand(0);
2348   if (!N0.isMachineOpcode())
2349     return false;
2350 
2351   switch (N0.getMachineOpcode()) {
2352   default:
2353     break;
2354   case RISCV::ADD:
2355   case RISCV::ADDI:
2356   case RISCV::SUB:
2357   case RISCV::MUL:
2358   case RISCV::SLLI: {
2359     // Convert sext.w+add/sub/mul to their W instructions. This will create
2360     // a new independent instruction. This improves latency.
2361     unsigned Opc;
2362     switch (N0.getMachineOpcode()) {
2363     default:
2364       llvm_unreachable("Unexpected opcode!");
2365     case RISCV::ADD:  Opc = RISCV::ADDW;  break;
2366     case RISCV::ADDI: Opc = RISCV::ADDIW; break;
2367     case RISCV::SUB:  Opc = RISCV::SUBW;  break;
2368     case RISCV::MUL:  Opc = RISCV::MULW;  break;
2369     case RISCV::SLLI: Opc = RISCV::SLLIW; break;
2370     }
2371 
2372     SDValue N00 = N0.getOperand(0);
2373     SDValue N01 = N0.getOperand(1);
2374 
2375     // Shift amount needs to be uimm5.
2376     if (N0.getMachineOpcode() == RISCV::SLLI &&
2377         !isUInt<5>(cast<ConstantSDNode>(N01)->getSExtValue()))
2378       break;
2379 
2380     SDNode *Result =
2381         CurDAG->getMachineNode(Opc, SDLoc(N), N->getValueType(0),
2382                                N00, N01);
2383     ReplaceUses(N, Result);
2384     return true;
2385   }
2386   case RISCV::ADDW:
2387   case RISCV::ADDIW:
2388   case RISCV::SUBW:
2389   case RISCV::MULW:
2390   case RISCV::SLLIW:
2391   case RISCV::GREVIW:
2392   case RISCV::GORCIW:
2393     // Result is already sign extended just remove the sext.w.
2394     // NOTE: We only handle the nodes that are selected with hasAllWUsers.
2395     ReplaceUses(N, N0.getNode());
2396     return true;
2397   }
2398 
2399   return false;
2400 }
2401 
2402 // Optimize masked RVV pseudo instructions with a known all-ones mask to their
2403 // corresponding "unmasked" pseudo versions. The mask we're interested in will
2404 // take the form of a V0 physical register operand, with a glued
2405 // register-setting instruction.
2406 bool RISCVDAGToDAGISel::doPeepholeMaskedRVV(SDNode *N) {
2407   const RISCV::RISCVMaskedPseudoInfo *I =
2408       RISCV::getMaskedPseudoInfo(N->getMachineOpcode());
2409   if (!I)
2410     return false;
2411 
2412   unsigned MaskOpIdx = I->MaskOpIdx;
2413 
2414   // Check that we're using V0 as a mask register.
2415   if (!isa<RegisterSDNode>(N->getOperand(MaskOpIdx)) ||
2416       cast<RegisterSDNode>(N->getOperand(MaskOpIdx))->getReg() != RISCV::V0)
2417     return false;
2418 
2419   // The glued user defines V0.
2420   const auto *Glued = N->getGluedNode();
2421 
2422   if (!Glued || Glued->getOpcode() != ISD::CopyToReg)
2423     return false;
2424 
2425   // Check that we're defining V0 as a mask register.
2426   if (!isa<RegisterSDNode>(Glued->getOperand(1)) ||
2427       cast<RegisterSDNode>(Glued->getOperand(1))->getReg() != RISCV::V0)
2428     return false;
2429 
2430   // Check the instruction defining V0; it needs to be a VMSET pseudo.
2431   SDValue MaskSetter = Glued->getOperand(2);
2432 
2433   const auto IsVMSet = [](unsigned Opc) {
2434     return Opc == RISCV::PseudoVMSET_M_B1 || Opc == RISCV::PseudoVMSET_M_B16 ||
2435            Opc == RISCV::PseudoVMSET_M_B2 || Opc == RISCV::PseudoVMSET_M_B32 ||
2436            Opc == RISCV::PseudoVMSET_M_B4 || Opc == RISCV::PseudoVMSET_M_B64 ||
2437            Opc == RISCV::PseudoVMSET_M_B8;
2438   };
2439 
2440   // TODO: Check that the VMSET is the expected bitwidth? The pseudo has
2441   // undefined behaviour if it's the wrong bitwidth, so we could choose to
2442   // assume that it's all-ones? Same applies to its VL.
2443   if (!MaskSetter->isMachineOpcode() || !IsVMSet(MaskSetter.getMachineOpcode()))
2444     return false;
2445 
2446   // Retrieve the tail policy operand index, if any.
2447   Optional<unsigned> TailPolicyOpIdx;
2448   const RISCVInstrInfo &TII = *Subtarget->getInstrInfo();
2449   const MCInstrDesc &MaskedMCID = TII.get(N->getMachineOpcode());
2450 
2451   bool IsTA = true;
2452   if (RISCVII::hasVecPolicyOp(MaskedMCID.TSFlags)) {
2453     // The last operand of the pseudo is the policy op, but we might have a
2454     // Glue operand last. We might also have a chain.
2455     TailPolicyOpIdx = N->getNumOperands() - 1;
2456     if (N->getOperand(*TailPolicyOpIdx).getValueType() == MVT::Glue)
2457       (*TailPolicyOpIdx)--;
2458     if (N->getOperand(*TailPolicyOpIdx).getValueType() == MVT::Other)
2459       (*TailPolicyOpIdx)--;
2460 
2461     if (!(N->getConstantOperandVal(*TailPolicyOpIdx) &
2462           RISCVII::TAIL_AGNOSTIC)) {
2463       // Keep the true-masked instruction when there is no unmasked TU
2464       // instruction
2465       if (I->UnmaskedTUPseudo == I->MaskedPseudo && !N->getOperand(0).isUndef())
2466         return false;
2467       // We can't use TA if the tie-operand is not IMPLICIT_DEF
2468       if (!N->getOperand(0).isUndef())
2469         IsTA = false;
2470     }
2471   }
2472 
2473   unsigned Opc = IsTA ? I->UnmaskedPseudo : I->UnmaskedTUPseudo;
2474 
2475   // Check that we're dropping the mask operand and any policy operand
2476   // when we transform to this unmasked pseudo. Additionally, if this insturtion
2477   // is tail agnostic, the unmasked instruction should not have a merge op.
2478   uint64_t TSFlags = TII.get(Opc).TSFlags;
2479   assert((IsTA != RISCVII::hasMergeOp(TSFlags)) &&
2480          RISCVII::hasDummyMaskOp(TSFlags) &&
2481          !RISCVII::hasVecPolicyOp(TSFlags) &&
2482          "Unexpected pseudo to transform to");
2483   (void)TSFlags;
2484 
2485   SmallVector<SDValue, 8> Ops;
2486   // Skip the merge operand at index 0 if IsTA
2487   for (unsigned I = IsTA, E = N->getNumOperands(); I != E; I++) {
2488     // Skip the mask, the policy, and the Glue.
2489     SDValue Op = N->getOperand(I);
2490     if (I == MaskOpIdx || I == TailPolicyOpIdx ||
2491         Op.getValueType() == MVT::Glue)
2492       continue;
2493     Ops.push_back(Op);
2494   }
2495 
2496   // Transitively apply any node glued to our new node.
2497   if (auto *TGlued = Glued->getGluedNode())
2498     Ops.push_back(SDValue(TGlued, TGlued->getNumValues() - 1));
2499 
2500   SDNode *Result = CurDAG->getMachineNode(Opc, SDLoc(N), N->getVTList(), Ops);
2501   ReplaceUses(N, Result);
2502 
2503   return true;
2504 }
2505 
2506 // This pass converts a legalized DAG into a RISCV-specific DAG, ready
2507 // for instruction scheduling.
2508 FunctionPass *llvm::createRISCVISelDag(RISCVTargetMachine &TM,
2509                                        CodeGenOpt::Level OptLevel) {
2510   return new RISCVDAGToDAGISel(TM, OptLevel);
2511 }
2512