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   for (SelectionDAG::allnodes_iterator I = CurDAG->allnodes_begin(),
47                                        E = CurDAG->allnodes_end();
48        I != E;) {
49     SDNode *N = &*I++; // Preincrement iterator to avoid invalidation issues.
50 
51     // Convert integer SPLAT_VECTOR to VMV_V_X_VL and floating-point
52     // SPLAT_VECTOR to VFMV_V_F_VL to reduce isel burden.
53     if (N->getOpcode() == ISD::SPLAT_VECTOR) {
54       MVT VT = N->getSimpleValueType(0);
55       unsigned Opc =
56           VT.isInteger() ? RISCVISD::VMV_V_X_VL : RISCVISD::VFMV_V_F_VL;
57       SDLoc DL(N);
58       SDValue VL = CurDAG->getRegister(RISCV::X0, Subtarget->getXLenVT());
59       SDValue Result = CurDAG->getNode(Opc, DL, VT, CurDAG->getUNDEF(VT),
60                                        N->getOperand(0), VL);
61 
62       --I;
63       CurDAG->ReplaceAllUsesOfValueWith(SDValue(N, 0), Result);
64       ++I;
65       CurDAG->DeleteNode(N);
66       continue;
67     }
68 
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     if (N->getOpcode() != RISCVISD::SPLAT_VECTOR_SPLIT_I64_VL)
73       continue;
74 
75     assert(N->getNumOperands() == 4 && "Unexpected number of operands");
76     MVT VT = N->getSimpleValueType(0);
77     SDValue Passthru = N->getOperand(0);
78     SDValue Lo = N->getOperand(1);
79     SDValue Hi = N->getOperand(2);
80     SDValue VL = N->getOperand(3);
81     assert(VT.getVectorElementType() == MVT::i64 && VT.isScalableVector() &&
82            Lo.getValueType() == MVT::i32 && Hi.getValueType() == MVT::i32 &&
83            "Unexpected VTs!");
84     MachineFunction &MF = CurDAG->getMachineFunction();
85     RISCVMachineFunctionInfo *FuncInfo = MF.getInfo<RISCVMachineFunctionInfo>();
86     SDLoc DL(N);
87 
88     // We use the same frame index we use for moving two i32s into 64-bit FPR.
89     // This is an analogous operation.
90     int FI = FuncInfo->getMoveF64FrameIndex(MF);
91     MachinePointerInfo MPI = MachinePointerInfo::getFixedStack(MF, FI);
92     const TargetLowering &TLI = CurDAG->getTargetLoweringInfo();
93     SDValue StackSlot =
94         CurDAG->getFrameIndex(FI, TLI.getPointerTy(CurDAG->getDataLayout()));
95 
96     SDValue Chain = CurDAG->getEntryNode();
97     Lo = CurDAG->getStore(Chain, DL, Lo, StackSlot, MPI, Align(8));
98 
99     SDValue OffsetSlot =
100         CurDAG->getMemBasePlusOffset(StackSlot, TypeSize::Fixed(4), DL);
101     Hi = CurDAG->getStore(Chain, DL, Hi, OffsetSlot, MPI.getWithOffset(4),
102                           Align(8));
103 
104     Chain = CurDAG->getNode(ISD::TokenFactor, DL, MVT::Other, Lo, Hi);
105 
106     SDVTList VTs = CurDAG->getVTList({VT, MVT::Other});
107     SDValue IntID =
108         CurDAG->getTargetConstant(Intrinsic::riscv_vlse, DL, MVT::i64);
109     SDValue Ops[] = {Chain,
110                      IntID,
111                      Passthru,
112                      StackSlot,
113                      CurDAG->getRegister(RISCV::X0, MVT::i64),
114                      VL};
115 
116     SDValue Result = CurDAG->getMemIntrinsicNode(
117         ISD::INTRINSIC_W_CHAIN, DL, VTs, Ops, MVT::i64, MPI, Align(8),
118         MachineMemOperand::MOLoad);
119 
120     // We're about to replace all uses of the SPLAT_VECTOR_SPLIT_I64 with the
121     // vlse we created.  This will cause general havok on the dag because
122     // anything below the conversion could be folded into other existing nodes.
123     // To avoid invalidating 'I', back it up to the convert node.
124     --I;
125     CurDAG->ReplaceAllUsesOfValueWith(SDValue(N, 0), Result);
126 
127     // Now that we did that, the node is dead.  Increment the iterator to the
128     // next node to process, then delete N.
129     ++I;
130     CurDAG->DeleteNode(N);
131   }
132 }
133 
134 void RISCVDAGToDAGISel::PostprocessISelDAG() {
135   HandleSDNode Dummy(CurDAG->getRoot());
136   SelectionDAG::allnodes_iterator Position = CurDAG->allnodes_end();
137 
138   bool MadeChange = false;
139   while (Position != CurDAG->allnodes_begin()) {
140     SDNode *N = &*--Position;
141     // Skip dead nodes and any non-machine opcodes.
142     if (N->use_empty() || !N->isMachineOpcode())
143       continue;
144 
145     MadeChange |= doPeepholeSExtW(N);
146     MadeChange |= doPeepholeLoadStoreADDI(N);
147     MadeChange |= doPeepholeMaskedRVV(N);
148   }
149 
150   CurDAG->setRoot(Dummy.getValue());
151 
152   if (MadeChange)
153     CurDAG->RemoveDeadNodes();
154 }
155 
156 // Returns true if N is a MachineSDNode that has a reg and simm12 memory
157 // operand. The indices of the base pointer and offset are returned in BaseOpIdx
158 // and OffsetOpIdx.
159 static bool hasMemOffset(SDNode *N, unsigned &BaseOpIdx,
160                          unsigned &OffsetOpIdx) {
161   switch (N->getMachineOpcode()) {
162   case RISCV::LB:
163   case RISCV::LH:
164   case RISCV::LW:
165   case RISCV::LBU:
166   case RISCV::LHU:
167   case RISCV::LWU:
168   case RISCV::LD:
169   case RISCV::FLH:
170   case RISCV::FLW:
171   case RISCV::FLD:
172     BaseOpIdx = 0;
173     OffsetOpIdx = 1;
174     return true;
175   case RISCV::SB:
176   case RISCV::SH:
177   case RISCV::SW:
178   case RISCV::SD:
179   case RISCV::FSH:
180   case RISCV::FSW:
181   case RISCV::FSD:
182     BaseOpIdx = 1;
183     OffsetOpIdx = 2;
184     return true;
185   }
186 
187   return false;
188 }
189 
190 static SDNode *selectImmWithConstantPool(SelectionDAG *CurDAG, const SDLoc &DL,
191                                          const MVT VT, int64_t Imm,
192                                          const RISCVSubtarget &Subtarget) {
193   assert(VT == MVT::i64 && "Expecting MVT::i64");
194   const RISCVTargetLowering *TLI = Subtarget.getTargetLowering();
195   ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(CurDAG->getConstantPool(
196       ConstantInt::get(EVT(VT).getTypeForEVT(*CurDAG->getContext()), Imm), VT));
197   SDValue Addr = TLI->getAddr(CP, *CurDAG);
198   SDValue Offset = CurDAG->getTargetConstant(0, DL, VT);
199   // Since there is no data race, the chain can be the entry node.
200   SDNode *Load = CurDAG->getMachineNode(RISCV::LD, DL, VT, Addr, Offset,
201                                         CurDAG->getEntryNode());
202   MachineFunction &MF = CurDAG->getMachineFunction();
203   MachineMemOperand *MemOp = MF.getMachineMemOperand(
204       MachinePointerInfo::getConstantPool(MF), MachineMemOperand::MOLoad,
205       LLT(VT), CP->getAlign());
206   CurDAG->setNodeMemRefs(cast<MachineSDNode>(Load), {MemOp});
207   return Load;
208 }
209 
210 static SDNode *selectImm(SelectionDAG *CurDAG, const SDLoc &DL, const MVT VT,
211                          int64_t Imm, const RISCVSubtarget &Subtarget) {
212   MVT XLenVT = Subtarget.getXLenVT();
213   RISCVMatInt::InstSeq Seq =
214       RISCVMatInt::generateInstSeq(Imm, Subtarget.getFeatureBits());
215 
216   // If Imm is expensive to build, then we put it into constant pool.
217   if (Subtarget.useConstantPoolForLargeInts() &&
218       Seq.size() > Subtarget.getMaxBuildIntsCost())
219     return selectImmWithConstantPool(CurDAG, DL, VT, Imm, Subtarget);
220 
221   SDNode *Result = nullptr;
222   SDValue SrcReg = CurDAG->getRegister(RISCV::X0, XLenVT);
223   for (RISCVMatInt::Inst &Inst : Seq) {
224     SDValue SDImm = CurDAG->getTargetConstant(Inst.Imm, DL, XLenVT);
225     switch (Inst.getOpndKind()) {
226     case RISCVMatInt::Imm:
227       Result = CurDAG->getMachineNode(Inst.Opc, DL, XLenVT, SDImm);
228       break;
229     case RISCVMatInt::RegX0:
230       Result = CurDAG->getMachineNode(Inst.Opc, DL, XLenVT, SrcReg,
231                                       CurDAG->getRegister(RISCV::X0, XLenVT));
232       break;
233     case RISCVMatInt::RegReg:
234       Result = CurDAG->getMachineNode(Inst.Opc, DL, XLenVT, SrcReg, SrcReg);
235       break;
236     case RISCVMatInt::RegImm:
237       Result = CurDAG->getMachineNode(Inst.Opc, DL, XLenVT, SrcReg, SDImm);
238       break;
239     }
240 
241     // Only the first instruction has X0 as its source.
242     SrcReg = SDValue(Result, 0);
243   }
244 
245   return Result;
246 }
247 
248 static SDValue createTupleImpl(SelectionDAG &CurDAG, ArrayRef<SDValue> Regs,
249                                unsigned RegClassID, unsigned SubReg0) {
250   assert(Regs.size() >= 2 && Regs.size() <= 8);
251 
252   SDLoc DL(Regs[0]);
253   SmallVector<SDValue, 8> Ops;
254 
255   Ops.push_back(CurDAG.getTargetConstant(RegClassID, DL, MVT::i32));
256 
257   for (unsigned I = 0; I < Regs.size(); ++I) {
258     Ops.push_back(Regs[I]);
259     Ops.push_back(CurDAG.getTargetConstant(SubReg0 + I, DL, MVT::i32));
260   }
261   SDNode *N =
262       CurDAG.getMachineNode(TargetOpcode::REG_SEQUENCE, DL, MVT::Untyped, Ops);
263   return SDValue(N, 0);
264 }
265 
266 static SDValue createM1Tuple(SelectionDAG &CurDAG, ArrayRef<SDValue> Regs,
267                              unsigned NF) {
268   static const unsigned RegClassIDs[] = {
269       RISCV::VRN2M1RegClassID, RISCV::VRN3M1RegClassID, RISCV::VRN4M1RegClassID,
270       RISCV::VRN5M1RegClassID, RISCV::VRN6M1RegClassID, RISCV::VRN7M1RegClassID,
271       RISCV::VRN8M1RegClassID};
272 
273   return createTupleImpl(CurDAG, Regs, RegClassIDs[NF - 2], RISCV::sub_vrm1_0);
274 }
275 
276 static SDValue createM2Tuple(SelectionDAG &CurDAG, ArrayRef<SDValue> Regs,
277                              unsigned NF) {
278   static const unsigned RegClassIDs[] = {RISCV::VRN2M2RegClassID,
279                                          RISCV::VRN3M2RegClassID,
280                                          RISCV::VRN4M2RegClassID};
281 
282   return createTupleImpl(CurDAG, Regs, RegClassIDs[NF - 2], RISCV::sub_vrm2_0);
283 }
284 
285 static SDValue createM4Tuple(SelectionDAG &CurDAG, ArrayRef<SDValue> Regs,
286                              unsigned NF) {
287   return createTupleImpl(CurDAG, Regs, RISCV::VRN2M4RegClassID,
288                          RISCV::sub_vrm4_0);
289 }
290 
291 static SDValue createTuple(SelectionDAG &CurDAG, ArrayRef<SDValue> Regs,
292                            unsigned NF, RISCVII::VLMUL LMUL) {
293   switch (LMUL) {
294   default:
295     llvm_unreachable("Invalid LMUL.");
296   case RISCVII::VLMUL::LMUL_F8:
297   case RISCVII::VLMUL::LMUL_F4:
298   case RISCVII::VLMUL::LMUL_F2:
299   case RISCVII::VLMUL::LMUL_1:
300     return createM1Tuple(CurDAG, Regs, NF);
301   case RISCVII::VLMUL::LMUL_2:
302     return createM2Tuple(CurDAG, Regs, NF);
303   case RISCVII::VLMUL::LMUL_4:
304     return createM4Tuple(CurDAG, Regs, NF);
305   }
306 }
307 
308 void RISCVDAGToDAGISel::addVectorLoadStoreOperands(
309     SDNode *Node, unsigned Log2SEW, const SDLoc &DL, unsigned CurOp,
310     bool IsMasked, bool IsStridedOrIndexed, SmallVectorImpl<SDValue> &Operands,
311     bool IsLoad, MVT *IndexVT) {
312   SDValue Chain = Node->getOperand(0);
313   SDValue Glue;
314 
315   SDValue Base;
316   SelectBaseAddr(Node->getOperand(CurOp++), Base);
317   Operands.push_back(Base); // Base pointer.
318 
319   if (IsStridedOrIndexed) {
320     Operands.push_back(Node->getOperand(CurOp++)); // Index.
321     if (IndexVT)
322       *IndexVT = Operands.back()->getSimpleValueType(0);
323   }
324 
325   if (IsMasked) {
326     // Mask needs to be copied to V0.
327     SDValue Mask = Node->getOperand(CurOp++);
328     Chain = CurDAG->getCopyToReg(Chain, DL, RISCV::V0, Mask, SDValue());
329     Glue = Chain.getValue(1);
330     Operands.push_back(CurDAG->getRegister(RISCV::V0, Mask.getValueType()));
331   }
332   SDValue VL;
333   selectVLOp(Node->getOperand(CurOp++), VL);
334   Operands.push_back(VL);
335 
336   MVT XLenVT = Subtarget->getXLenVT();
337   SDValue SEWOp = CurDAG->getTargetConstant(Log2SEW, DL, XLenVT);
338   Operands.push_back(SEWOp);
339 
340   // Masked load has the tail policy argument.
341   if (IsMasked && IsLoad) {
342     // Policy must be a constant.
343     uint64_t Policy = Node->getConstantOperandVal(CurOp++);
344     SDValue PolicyOp = CurDAG->getTargetConstant(Policy, DL, XLenVT);
345     Operands.push_back(PolicyOp);
346   }
347 
348   Operands.push_back(Chain); // Chain.
349   if (Glue)
350     Operands.push_back(Glue);
351 }
352 
353 static bool isAllUndef(ArrayRef<SDValue> Values) {
354   return llvm::all_of(Values, [](SDValue V) { return V->isUndef(); });
355 }
356 
357 void RISCVDAGToDAGISel::selectVLSEG(SDNode *Node, bool IsMasked,
358                                     bool IsStrided) {
359   SDLoc DL(Node);
360   unsigned NF = Node->getNumValues() - 1;
361   MVT VT = Node->getSimpleValueType(0);
362   unsigned Log2SEW = Log2_32(VT.getScalarSizeInBits());
363   RISCVII::VLMUL LMUL = RISCVTargetLowering::getLMUL(VT);
364 
365   unsigned CurOp = 2;
366   SmallVector<SDValue, 8> Operands;
367 
368   SmallVector<SDValue, 8> Regs(Node->op_begin() + CurOp,
369                                Node->op_begin() + CurOp + NF);
370   bool IsTU = IsMasked || !isAllUndef(Regs);
371   if (IsTU) {
372     SDValue Merge = createTuple(*CurDAG, Regs, NF, LMUL);
373     Operands.push_back(Merge);
374   }
375   CurOp += NF;
376 
377   addVectorLoadStoreOperands(Node, Log2SEW, DL, CurOp, IsMasked, IsStrided,
378                              Operands, /*IsLoad=*/true);
379 
380   const RISCV::VLSEGPseudo *P =
381       RISCV::getVLSEGPseudo(NF, IsMasked, IsTU, IsStrided, /*FF*/ false, Log2SEW,
382                             static_cast<unsigned>(LMUL));
383   MachineSDNode *Load =
384       CurDAG->getMachineNode(P->Pseudo, DL, MVT::Untyped, MVT::Other, Operands);
385 
386   if (auto *MemOp = dyn_cast<MemSDNode>(Node))
387     CurDAG->setNodeMemRefs(Load, {MemOp->getMemOperand()});
388 
389   SDValue SuperReg = SDValue(Load, 0);
390   for (unsigned I = 0; I < NF; ++I) {
391     unsigned SubRegIdx = RISCVTargetLowering::getSubregIndexByMVT(VT, I);
392     ReplaceUses(SDValue(Node, I),
393                 CurDAG->getTargetExtractSubreg(SubRegIdx, DL, VT, SuperReg));
394   }
395 
396   ReplaceUses(SDValue(Node, NF), SDValue(Load, 1));
397   CurDAG->RemoveDeadNode(Node);
398 }
399 
400 void RISCVDAGToDAGISel::selectVLSEGFF(SDNode *Node, bool IsMasked) {
401   SDLoc DL(Node);
402   unsigned NF = Node->getNumValues() - 2; // Do not count VL and Chain.
403   MVT VT = Node->getSimpleValueType(0);
404   MVT XLenVT = Subtarget->getXLenVT();
405   unsigned SEW = VT.getScalarSizeInBits();
406   unsigned Log2SEW = Log2_32(SEW);
407   RISCVII::VLMUL LMUL = RISCVTargetLowering::getLMUL(VT);
408 
409   unsigned CurOp = 2;
410   SmallVector<SDValue, 7> Operands;
411 
412   SmallVector<SDValue, 8> Regs(Node->op_begin() + CurOp,
413                                Node->op_begin() + CurOp + NF);
414   bool IsTU = IsMasked || !isAllUndef(Regs);
415   if (IsTU) {
416     SDValue MaskedOff = createTuple(*CurDAG, Regs, NF, LMUL);
417     Operands.push_back(MaskedOff);
418   }
419   CurOp += NF;
420 
421   addVectorLoadStoreOperands(Node, Log2SEW, DL, CurOp, IsMasked,
422                              /*IsStridedOrIndexed*/ false, Operands,
423                              /*IsLoad=*/true);
424 
425   const RISCV::VLSEGPseudo *P =
426       RISCV::getVLSEGPseudo(NF, IsMasked, IsTU, /*Strided*/ false, /*FF*/ true,
427                             Log2SEW, static_cast<unsigned>(LMUL));
428   MachineSDNode *Load = CurDAG->getMachineNode(P->Pseudo, DL, MVT::Untyped,
429                                                MVT::Other, MVT::Glue, Operands);
430   bool TailAgnostic = true;
431   bool MaskAgnostic = false;
432   if (IsMasked) {
433     uint64_t Policy = Node->getConstantOperandVal(Node->getNumOperands() - 1);
434     TailAgnostic = Policy & RISCVII::TAIL_AGNOSTIC;
435     MaskAgnostic = Policy & RISCVII::MASK_AGNOSTIC;
436   }
437   unsigned VType =
438       RISCVVType::encodeVTYPE(LMUL, SEW, TailAgnostic, MaskAgnostic);
439   SDValue VTypeOp = CurDAG->getTargetConstant(VType, DL, XLenVT);
440   SDNode *ReadVL = CurDAG->getMachineNode(RISCV::PseudoReadVL, DL, XLenVT,
441                                           VTypeOp, /*Glue*/ SDValue(Load, 2));
442 
443   if (auto *MemOp = dyn_cast<MemSDNode>(Node))
444     CurDAG->setNodeMemRefs(Load, {MemOp->getMemOperand()});
445 
446   SDValue SuperReg = SDValue(Load, 0);
447   for (unsigned I = 0; I < NF; ++I) {
448     unsigned SubRegIdx = RISCVTargetLowering::getSubregIndexByMVT(VT, I);
449     ReplaceUses(SDValue(Node, I),
450                 CurDAG->getTargetExtractSubreg(SubRegIdx, DL, VT, SuperReg));
451   }
452 
453   ReplaceUses(SDValue(Node, NF), SDValue(ReadVL, 0));   // VL
454   ReplaceUses(SDValue(Node, NF + 1), SDValue(Load, 1)); // Chain
455   CurDAG->RemoveDeadNode(Node);
456 }
457 
458 void RISCVDAGToDAGISel::selectVLXSEG(SDNode *Node, bool IsMasked,
459                                      bool IsOrdered) {
460   SDLoc DL(Node);
461   unsigned NF = Node->getNumValues() - 1;
462   MVT VT = Node->getSimpleValueType(0);
463   unsigned Log2SEW = Log2_32(VT.getScalarSizeInBits());
464   RISCVII::VLMUL LMUL = RISCVTargetLowering::getLMUL(VT);
465 
466   unsigned CurOp = 2;
467   SmallVector<SDValue, 8> Operands;
468 
469   SmallVector<SDValue, 8> Regs(Node->op_begin() + CurOp,
470                                Node->op_begin() + CurOp + NF);
471   bool IsTU = IsMasked || !isAllUndef(Regs);
472   if (IsTU) {
473     SDValue MaskedOff = createTuple(*CurDAG, Regs, NF, LMUL);
474     Operands.push_back(MaskedOff);
475   }
476   CurOp += NF;
477 
478   MVT IndexVT;
479   addVectorLoadStoreOperands(Node, Log2SEW, DL, CurOp, IsMasked,
480                              /*IsStridedOrIndexed*/ true, Operands,
481                              /*IsLoad=*/true, &IndexVT);
482 
483   assert(VT.getVectorElementCount() == IndexVT.getVectorElementCount() &&
484          "Element count mismatch");
485 
486   RISCVII::VLMUL IndexLMUL = RISCVTargetLowering::getLMUL(IndexVT);
487   unsigned IndexLog2EEW = Log2_32(IndexVT.getScalarSizeInBits());
488   if (IndexLog2EEW == 6 && !Subtarget->is64Bit()) {
489     report_fatal_error("The V extension does not support EEW=64 for index "
490                        "values when XLEN=32");
491   }
492   const RISCV::VLXSEGPseudo *P = RISCV::getVLXSEGPseudo(
493       NF, IsMasked, IsTU, IsOrdered, IndexLog2EEW, static_cast<unsigned>(LMUL),
494       static_cast<unsigned>(IndexLMUL));
495   MachineSDNode *Load =
496       CurDAG->getMachineNode(P->Pseudo, DL, MVT::Untyped, MVT::Other, Operands);
497 
498   if (auto *MemOp = dyn_cast<MemSDNode>(Node))
499     CurDAG->setNodeMemRefs(Load, {MemOp->getMemOperand()});
500 
501   SDValue SuperReg = SDValue(Load, 0);
502   for (unsigned I = 0; I < NF; ++I) {
503     unsigned SubRegIdx = RISCVTargetLowering::getSubregIndexByMVT(VT, I);
504     ReplaceUses(SDValue(Node, I),
505                 CurDAG->getTargetExtractSubreg(SubRegIdx, DL, VT, SuperReg));
506   }
507 
508   ReplaceUses(SDValue(Node, NF), SDValue(Load, 1));
509   CurDAG->RemoveDeadNode(Node);
510 }
511 
512 void RISCVDAGToDAGISel::selectVSSEG(SDNode *Node, bool IsMasked,
513                                     bool IsStrided) {
514   SDLoc DL(Node);
515   unsigned NF = Node->getNumOperands() - 4;
516   if (IsStrided)
517     NF--;
518   if (IsMasked)
519     NF--;
520   MVT VT = Node->getOperand(2)->getSimpleValueType(0);
521   unsigned Log2SEW = Log2_32(VT.getScalarSizeInBits());
522   RISCVII::VLMUL LMUL = RISCVTargetLowering::getLMUL(VT);
523   SmallVector<SDValue, 8> Regs(Node->op_begin() + 2, Node->op_begin() + 2 + NF);
524   SDValue StoreVal = createTuple(*CurDAG, Regs, NF, LMUL);
525 
526   SmallVector<SDValue, 8> Operands;
527   Operands.push_back(StoreVal);
528   unsigned CurOp = 2 + NF;
529 
530   addVectorLoadStoreOperands(Node, Log2SEW, DL, CurOp, IsMasked, IsStrided,
531                              Operands);
532 
533   const RISCV::VSSEGPseudo *P = RISCV::getVSSEGPseudo(
534       NF, IsMasked, IsStrided, Log2SEW, static_cast<unsigned>(LMUL));
535   MachineSDNode *Store =
536       CurDAG->getMachineNode(P->Pseudo, DL, Node->getValueType(0), Operands);
537 
538   if (auto *MemOp = dyn_cast<MemSDNode>(Node))
539     CurDAG->setNodeMemRefs(Store, {MemOp->getMemOperand()});
540 
541   ReplaceNode(Node, Store);
542 }
543 
544 void RISCVDAGToDAGISel::selectVSXSEG(SDNode *Node, bool IsMasked,
545                                      bool IsOrdered) {
546   SDLoc DL(Node);
547   unsigned NF = Node->getNumOperands() - 5;
548   if (IsMasked)
549     --NF;
550   MVT VT = Node->getOperand(2)->getSimpleValueType(0);
551   unsigned Log2SEW = Log2_32(VT.getScalarSizeInBits());
552   RISCVII::VLMUL LMUL = RISCVTargetLowering::getLMUL(VT);
553   SmallVector<SDValue, 8> Regs(Node->op_begin() + 2, Node->op_begin() + 2 + NF);
554   SDValue StoreVal = createTuple(*CurDAG, Regs, NF, LMUL);
555 
556   SmallVector<SDValue, 8> Operands;
557   Operands.push_back(StoreVal);
558   unsigned CurOp = 2 + NF;
559 
560   MVT IndexVT;
561   addVectorLoadStoreOperands(Node, Log2SEW, DL, CurOp, IsMasked,
562                              /*IsStridedOrIndexed*/ true, Operands,
563                              /*IsLoad=*/false, &IndexVT);
564 
565   assert(VT.getVectorElementCount() == IndexVT.getVectorElementCount() &&
566          "Element count mismatch");
567 
568   RISCVII::VLMUL IndexLMUL = RISCVTargetLowering::getLMUL(IndexVT);
569   unsigned IndexLog2EEW = Log2_32(IndexVT.getScalarSizeInBits());
570   if (IndexLog2EEW == 6 && !Subtarget->is64Bit()) {
571     report_fatal_error("The V extension does not support EEW=64 for index "
572                        "values when XLEN=32");
573   }
574   const RISCV::VSXSEGPseudo *P = RISCV::getVSXSEGPseudo(
575       NF, IsMasked, IsOrdered, IndexLog2EEW, static_cast<unsigned>(LMUL),
576       static_cast<unsigned>(IndexLMUL));
577   MachineSDNode *Store =
578       CurDAG->getMachineNode(P->Pseudo, DL, Node->getValueType(0), Operands);
579 
580   if (auto *MemOp = dyn_cast<MemSDNode>(Node))
581     CurDAG->setNodeMemRefs(Store, {MemOp->getMemOperand()});
582 
583   ReplaceNode(Node, Store);
584 }
585 
586 void RISCVDAGToDAGISel::selectVSETVLI(SDNode *Node) {
587   if (!Subtarget->hasVInstructions())
588     return;
589 
590   assert((Node->getOpcode() == ISD::INTRINSIC_W_CHAIN ||
591           Node->getOpcode() == ISD::INTRINSIC_WO_CHAIN) &&
592          "Unexpected opcode");
593 
594   SDLoc DL(Node);
595   MVT XLenVT = Subtarget->getXLenVT();
596 
597   bool HasChain = Node->getOpcode() == ISD::INTRINSIC_W_CHAIN;
598   unsigned IntNoOffset = HasChain ? 1 : 0;
599   unsigned IntNo = Node->getConstantOperandVal(IntNoOffset);
600 
601   assert((IntNo == Intrinsic::riscv_vsetvli ||
602           IntNo == Intrinsic::riscv_vsetvlimax ||
603           IntNo == Intrinsic::riscv_vsetvli_opt ||
604           IntNo == Intrinsic::riscv_vsetvlimax_opt) &&
605          "Unexpected vsetvli intrinsic");
606 
607   bool VLMax = IntNo == Intrinsic::riscv_vsetvlimax ||
608                IntNo == Intrinsic::riscv_vsetvlimax_opt;
609   unsigned Offset = IntNoOffset + (VLMax ? 1 : 2);
610 
611   assert(Node->getNumOperands() == Offset + 2 &&
612          "Unexpected number of operands");
613 
614   unsigned SEW =
615       RISCVVType::decodeVSEW(Node->getConstantOperandVal(Offset) & 0x7);
616   RISCVII::VLMUL VLMul = static_cast<RISCVII::VLMUL>(
617       Node->getConstantOperandVal(Offset + 1) & 0x7);
618 
619   unsigned VTypeI = RISCVVType::encodeVTYPE(VLMul, SEW, /*TailAgnostic*/ true,
620                                             /*MaskAgnostic*/ false);
621   SDValue VTypeIOp = CurDAG->getTargetConstant(VTypeI, DL, XLenVT);
622 
623   SmallVector<EVT, 2> VTs = {XLenVT};
624   if (HasChain)
625     VTs.push_back(MVT::Other);
626 
627   SDValue VLOperand;
628   unsigned Opcode = RISCV::PseudoVSETVLI;
629   if (VLMax) {
630     VLOperand = CurDAG->getRegister(RISCV::X0, XLenVT);
631     Opcode = RISCV::PseudoVSETVLIX0;
632   } else {
633     VLOperand = Node->getOperand(IntNoOffset + 1);
634 
635     if (auto *C = dyn_cast<ConstantSDNode>(VLOperand)) {
636       uint64_t AVL = C->getZExtValue();
637       if (isUInt<5>(AVL)) {
638         SDValue VLImm = CurDAG->getTargetConstant(AVL, DL, XLenVT);
639         SmallVector<SDValue, 3> Ops = {VLImm, VTypeIOp};
640         if (HasChain)
641           Ops.push_back(Node->getOperand(0));
642         ReplaceNode(
643             Node, CurDAG->getMachineNode(RISCV::PseudoVSETIVLI, DL, VTs, Ops));
644         return;
645       }
646     }
647   }
648 
649   SmallVector<SDValue, 3> Ops = {VLOperand, VTypeIOp};
650   if (HasChain)
651     Ops.push_back(Node->getOperand(0));
652 
653   ReplaceNode(Node, CurDAG->getMachineNode(Opcode, DL, VTs, Ops));
654 }
655 
656 void RISCVDAGToDAGISel::Select(SDNode *Node) {
657   // If we have a custom node, we have already selected.
658   if (Node->isMachineOpcode()) {
659     LLVM_DEBUG(dbgs() << "== "; Node->dump(CurDAG); dbgs() << "\n");
660     Node->setNodeId(-1);
661     return;
662   }
663 
664   // Instruction Selection not handled by the auto-generated tablegen selection
665   // should be handled here.
666   unsigned Opcode = Node->getOpcode();
667   MVT XLenVT = Subtarget->getXLenVT();
668   SDLoc DL(Node);
669   MVT VT = Node->getSimpleValueType(0);
670 
671   switch (Opcode) {
672   case ISD::Constant: {
673     auto *ConstNode = cast<ConstantSDNode>(Node);
674     if (VT == XLenVT && ConstNode->isZero()) {
675       SDValue New =
676           CurDAG->getCopyFromReg(CurDAG->getEntryNode(), DL, RISCV::X0, XLenVT);
677       ReplaceNode(Node, New.getNode());
678       return;
679     }
680     int64_t Imm = ConstNode->getSExtValue();
681     // If the upper XLen-16 bits are not used, try to convert this to a simm12
682     // by sign extending bit 15.
683     if (isUInt<16>(Imm) && isInt<12>(SignExtend64<16>(Imm)) &&
684         hasAllHUsers(Node))
685       Imm = SignExtend64<16>(Imm);
686     // If the upper 32-bits are not used try to convert this into a simm32 by
687     // sign extending bit 32.
688     if (!isInt<32>(Imm) && isUInt<32>(Imm) && hasAllWUsers(Node))
689       Imm = SignExtend64<32>(Imm);
690 
691     ReplaceNode(Node, selectImm(CurDAG, DL, VT, Imm, *Subtarget));
692     return;
693   }
694   case ISD::FrameIndex: {
695     SDValue Imm = CurDAG->getTargetConstant(0, DL, XLenVT);
696     int FI = cast<FrameIndexSDNode>(Node)->getIndex();
697     SDValue TFI = CurDAG->getTargetFrameIndex(FI, VT);
698     ReplaceNode(Node, CurDAG->getMachineNode(RISCV::ADDI, DL, VT, TFI, Imm));
699     return;
700   }
701   case ISD::ADD: {
702     // Try to select ADD + immediate used as memory addresses to
703     // (ADDI (ADD X, Imm-Lo12), Lo12) if it will allow the ADDI to be removed by
704     // doPeepholeLoadStoreADDI.
705 
706     // LHS should be an immediate.
707     auto *N1C = dyn_cast<ConstantSDNode>(Node->getOperand(1));
708     if (!N1C)
709       break;
710 
711     int64_t Offset = N1C->getSExtValue();
712     int64_t Lo12 = SignExtend64<12>(Offset);
713 
714     // Don't do this if the lower 12 bits are 0 or we could use ADDI directly.
715     if (Lo12 == 0 || isInt<12>(Offset))
716       break;
717 
718     // Don't do this if we can use a pair of ADDIs.
719     if (isInt<12>(Offset / 2) && isInt<12>(Offset - Offset / 2))
720       break;
721 
722     bool AllPointerUses = true;
723     for (auto UI = Node->use_begin(), UE = Node->use_end(); UI != UE; ++UI) {
724       SDNode *User = *UI;
725 
726       // Is this user a memory instruction that uses a register and immediate
727       // that has this ADD as its pointer.
728       unsigned BaseOpIdx, OffsetOpIdx;
729       if (!User->isMachineOpcode() ||
730           !hasMemOffset(User, BaseOpIdx, OffsetOpIdx) ||
731           UI.getOperandNo() != BaseOpIdx) {
732         AllPointerUses = false;
733         break;
734       }
735 
736       // If the memory instruction already has an offset, make sure the combined
737       // offset is foldable.
738       int64_t MemOffs =
739           cast<ConstantSDNode>(User->getOperand(OffsetOpIdx))->getSExtValue();
740       MemOffs += Lo12;
741       if (!isInt<12>(MemOffs)) {
742         AllPointerUses = false;
743         break;
744       }
745     }
746 
747     if (!AllPointerUses)
748       break;
749 
750     Offset -= Lo12;
751     // Restore sign bits for RV32.
752     if (!Subtarget->is64Bit())
753       Offset = SignExtend64<32>(Offset);
754 
755     // Emit (ADDI (ADD X, Hi), Lo)
756     SDNode *Imm = selectImm(CurDAG, DL, VT, Offset, *Subtarget);
757     SDNode *ADD = CurDAG->getMachineNode(RISCV::ADD, DL, VT,
758                                          Node->getOperand(0), SDValue(Imm, 0));
759     SDNode *ADDI =
760         CurDAG->getMachineNode(RISCV::ADDI, DL, VT, SDValue(ADD, 0),
761                                CurDAG->getTargetConstant(Lo12, DL, VT));
762     ReplaceNode(Node, ADDI);
763     return;
764   }
765   case ISD::SRL: {
766     // Optimize (srl (and X, C2), C) ->
767     //          (srli (slli X, (XLen-C3), (XLen-C3) + C)
768     // Where C2 is a mask with C3 trailing ones.
769     // Taking into account that the C2 may have had lower bits unset by
770     // SimplifyDemandedBits. This avoids materializing the C2 immediate.
771     // This pattern occurs when type legalizing right shifts for types with
772     // less than XLen bits.
773     auto *N1C = dyn_cast<ConstantSDNode>(Node->getOperand(1));
774     if (!N1C)
775       break;
776     SDValue N0 = Node->getOperand(0);
777     if (N0.getOpcode() != ISD::AND || !N0.hasOneUse() ||
778         !isa<ConstantSDNode>(N0.getOperand(1)))
779       break;
780     unsigned ShAmt = N1C->getZExtValue();
781     uint64_t Mask = N0.getConstantOperandVal(1);
782     Mask |= maskTrailingOnes<uint64_t>(ShAmt);
783     if (!isMask_64(Mask))
784       break;
785     unsigned TrailingOnes = countTrailingOnes(Mask);
786     // 32 trailing ones should use srliw via tablegen pattern.
787     if (TrailingOnes == 32 || ShAmt >= TrailingOnes)
788       break;
789     unsigned LShAmt = Subtarget->getXLen() - TrailingOnes;
790     SDNode *SLLI =
791         CurDAG->getMachineNode(RISCV::SLLI, DL, VT, N0->getOperand(0),
792                                CurDAG->getTargetConstant(LShAmt, DL, VT));
793     SDNode *SRLI = CurDAG->getMachineNode(
794         RISCV::SRLI, DL, VT, SDValue(SLLI, 0),
795         CurDAG->getTargetConstant(LShAmt + ShAmt, DL, VT));
796     ReplaceNode(Node, SRLI);
797     return;
798   }
799   case ISD::SRA: {
800     // Optimize (sra (sext_inreg X, i16), C) ->
801     //          (srai (slli X, (XLen-16), (XLen-16) + C)
802     // And      (sra (sext_inreg X, i8), C) ->
803     //          (srai (slli X, (XLen-8), (XLen-8) + C)
804     // This can occur when Zbb is enabled, which makes sext_inreg i16/i8 legal.
805     // This transform matches the code we get without Zbb. The shifts are more
806     // compressible, and this can help expose CSE opportunities in the sdiv by
807     // constant optimization.
808     auto *N1C = dyn_cast<ConstantSDNode>(Node->getOperand(1));
809     if (!N1C)
810       break;
811     SDValue N0 = Node->getOperand(0);
812     if (N0.getOpcode() != ISD::SIGN_EXTEND_INREG || !N0.hasOneUse())
813       break;
814     unsigned ShAmt = N1C->getZExtValue();
815     unsigned ExtSize =
816         cast<VTSDNode>(N0.getOperand(1))->getVT().getSizeInBits();
817     // ExtSize of 32 should use sraiw via tablegen pattern.
818     if (ExtSize >= 32 || ShAmt >= ExtSize)
819       break;
820     unsigned LShAmt = Subtarget->getXLen() - ExtSize;
821     SDNode *SLLI =
822         CurDAG->getMachineNode(RISCV::SLLI, DL, VT, N0->getOperand(0),
823                                CurDAG->getTargetConstant(LShAmt, DL, VT));
824     SDNode *SRAI = CurDAG->getMachineNode(
825         RISCV::SRAI, DL, VT, SDValue(SLLI, 0),
826         CurDAG->getTargetConstant(LShAmt + ShAmt, DL, VT));
827     ReplaceNode(Node, SRAI);
828     return;
829   }
830   case ISD::AND: {
831     auto *N1C = dyn_cast<ConstantSDNode>(Node->getOperand(1));
832     if (!N1C)
833       break;
834 
835     SDValue N0 = Node->getOperand(0);
836 
837     bool LeftShift = N0.getOpcode() == ISD::SHL;
838     if (!LeftShift && N0.getOpcode() != ISD::SRL)
839       break;
840 
841     auto *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
842     if (!C)
843       break;
844     uint64_t C2 = C->getZExtValue();
845     unsigned XLen = Subtarget->getXLen();
846     if (!C2 || C2 >= XLen)
847       break;
848 
849     uint64_t C1 = N1C->getZExtValue();
850 
851     // Keep track of whether this is a c.andi. If we can't use c.andi, the
852     // shift pair might offer more compression opportunities.
853     // TODO: We could check for C extension here, but we don't have many lit
854     // tests with the C extension enabled so not checking gets better coverage.
855     // TODO: What if ANDI faster than shift?
856     bool IsCANDI = isInt<6>(N1C->getSExtValue());
857 
858     // Clear irrelevant bits in the mask.
859     if (LeftShift)
860       C1 &= maskTrailingZeros<uint64_t>(C2);
861     else
862       C1 &= maskTrailingOnes<uint64_t>(XLen - C2);
863 
864     // Some transforms should only be done if the shift has a single use or
865     // the AND would become (srli (slli X, 32), 32)
866     bool OneUseOrZExtW = N0.hasOneUse() || C1 == UINT64_C(0xFFFFFFFF);
867 
868     SDValue X = N0.getOperand(0);
869 
870     // Turn (and (srl x, c2) c1) -> (srli (slli x, c3-c2), c3) if c1 is a mask
871     // with c3 leading zeros.
872     if (!LeftShift && isMask_64(C1)) {
873       uint64_t C3 = XLen - (64 - countLeadingZeros(C1));
874       if (C2 < C3) {
875         // If the number of leading zeros is C2+32 this can be SRLIW.
876         if (C2 + 32 == C3) {
877           SDNode *SRLIW =
878               CurDAG->getMachineNode(RISCV::SRLIW, DL, XLenVT, X,
879                                      CurDAG->getTargetConstant(C2, DL, XLenVT));
880           ReplaceNode(Node, SRLIW);
881           return;
882         }
883 
884         // (and (srl (sexti32 Y), c2), c1) -> (srliw (sraiw Y, 31), c3 - 32) if
885         // c1 is a mask with c3 leading zeros and c2 >= 32 and c3-c2==1.
886         //
887         // This pattern occurs when (i32 (srl (sra 31), c3 - 32)) is type
888         // legalized and goes through DAG combine.
889         if (C2 >= 32 && (C3 - C2) == 1 && N0.hasOneUse() &&
890             X.getOpcode() == ISD::SIGN_EXTEND_INREG &&
891             cast<VTSDNode>(X.getOperand(1))->getVT() == MVT::i32) {
892           SDNode *SRAIW =
893               CurDAG->getMachineNode(RISCV::SRAIW, DL, XLenVT, X.getOperand(0),
894                                      CurDAG->getTargetConstant(31, DL, XLenVT));
895           SDNode *SRLIW = CurDAG->getMachineNode(
896               RISCV::SRLIW, DL, XLenVT, SDValue(SRAIW, 0),
897               CurDAG->getTargetConstant(C3 - 32, DL, XLenVT));
898           ReplaceNode(Node, SRLIW);
899           return;
900         }
901 
902         // (srli (slli x, c3-c2), c3).
903         // Skip if we could use (zext.w (sraiw X, C2)).
904         bool Skip = Subtarget->hasStdExtZba() && C3 == 32 &&
905                     X.getOpcode() == ISD::SIGN_EXTEND_INREG &&
906                     cast<VTSDNode>(X.getOperand(1))->getVT() == MVT::i32;
907         // Also Skip if we can use bexti.
908         Skip |= Subtarget->hasStdExtZbs() && C3 == XLen - 1;
909         if (OneUseOrZExtW && !Skip) {
910           SDNode *SLLI = CurDAG->getMachineNode(
911               RISCV::SLLI, DL, XLenVT, X,
912               CurDAG->getTargetConstant(C3 - C2, DL, XLenVT));
913           SDNode *SRLI =
914               CurDAG->getMachineNode(RISCV::SRLI, DL, XLenVT, SDValue(SLLI, 0),
915                                      CurDAG->getTargetConstant(C3, DL, XLenVT));
916           ReplaceNode(Node, SRLI);
917           return;
918         }
919       }
920     }
921 
922     // Turn (and (shl x, c2), c1) -> (srli (slli c2+c3), c3) if c1 is a mask
923     // shifted by c2 bits with c3 leading zeros.
924     if (LeftShift && isShiftedMask_64(C1)) {
925       uint64_t C3 = XLen - (64 - countLeadingZeros(C1));
926 
927       if (C2 + C3 < XLen &&
928           C1 == (maskTrailingOnes<uint64_t>(XLen - (C2 + C3)) << C2)) {
929         // Use slli.uw when possible.
930         if ((XLen - (C2 + C3)) == 32 && Subtarget->hasStdExtZba()) {
931           SDNode *SLLI_UW =
932               CurDAG->getMachineNode(RISCV::SLLI_UW, DL, XLenVT, X,
933                                      CurDAG->getTargetConstant(C2, DL, XLenVT));
934           ReplaceNode(Node, SLLI_UW);
935           return;
936         }
937 
938         // (srli (slli c2+c3), c3)
939         if (OneUseOrZExtW && !IsCANDI) {
940           SDNode *SLLI = CurDAG->getMachineNode(
941               RISCV::SLLI, DL, XLenVT, X,
942               CurDAG->getTargetConstant(C2 + C3, DL, XLenVT));
943           SDNode *SRLI =
944               CurDAG->getMachineNode(RISCV::SRLI, DL, XLenVT, SDValue(SLLI, 0),
945                                      CurDAG->getTargetConstant(C3, DL, XLenVT));
946           ReplaceNode(Node, SRLI);
947           return;
948         }
949       }
950     }
951 
952     // Turn (and (shr x, c2), c1) -> (slli (srli x, c2+c3), c3) if c1 is a
953     // shifted mask with c2 leading zeros and c3 trailing zeros.
954     if (!LeftShift && isShiftedMask_64(C1)) {
955       uint64_t Leading = XLen - (64 - countLeadingZeros(C1));
956       uint64_t C3 = countTrailingZeros(C1);
957       if (Leading == C2 && C2 + C3 < XLen && OneUseOrZExtW && !IsCANDI) {
958         unsigned SrliOpc = RISCV::SRLI;
959         // If the input is zexti32 we should use SRLIW.
960         if (X.getOpcode() == ISD::AND && isa<ConstantSDNode>(X.getOperand(1)) &&
961             X.getConstantOperandVal(1) == UINT64_C(0xFFFFFFFF)) {
962           SrliOpc = RISCV::SRLIW;
963           X = X.getOperand(0);
964         }
965         SDNode *SRLI = CurDAG->getMachineNode(
966             SrliOpc, DL, XLenVT, X,
967             CurDAG->getTargetConstant(C2 + C3, DL, XLenVT));
968         SDNode *SLLI =
969             CurDAG->getMachineNode(RISCV::SLLI, DL, XLenVT, SDValue(SRLI, 0),
970                                    CurDAG->getTargetConstant(C3, DL, XLenVT));
971         ReplaceNode(Node, SLLI);
972         return;
973       }
974       // If the leading zero count is C2+32, we can use SRLIW instead of SRLI.
975       if (Leading > 32 && (Leading - 32) == C2 && C2 + C3 < 32 &&
976           OneUseOrZExtW && !IsCANDI) {
977         SDNode *SRLIW = CurDAG->getMachineNode(
978             RISCV::SRLIW, DL, XLenVT, X,
979             CurDAG->getTargetConstant(C2 + C3, DL, XLenVT));
980         SDNode *SLLI =
981             CurDAG->getMachineNode(RISCV::SLLI, DL, XLenVT, SDValue(SRLIW, 0),
982                                    CurDAG->getTargetConstant(C3, DL, XLenVT));
983         ReplaceNode(Node, SLLI);
984         return;
985       }
986     }
987 
988     // Turn (and (shl x, c2), c1) -> (slli (srli x, c3-c2), c3) if c1 is a
989     // shifted mask with no leading zeros and c3 trailing zeros.
990     if (LeftShift && isShiftedMask_64(C1)) {
991       uint64_t Leading = XLen - (64 - countLeadingZeros(C1));
992       uint64_t C3 = countTrailingZeros(C1);
993       if (Leading == 0 && C2 < C3 && OneUseOrZExtW && !IsCANDI) {
994         SDNode *SRLI = CurDAG->getMachineNode(
995             RISCV::SRLI, DL, XLenVT, X,
996             CurDAG->getTargetConstant(C3 - C2, DL, XLenVT));
997         SDNode *SLLI =
998             CurDAG->getMachineNode(RISCV::SLLI, DL, XLenVT, SDValue(SRLI, 0),
999                                    CurDAG->getTargetConstant(C3, DL, XLenVT));
1000         ReplaceNode(Node, SLLI);
1001         return;
1002       }
1003       // If we have (32-C2) leading zeros, we can use SRLIW instead of SRLI.
1004       if (C2 < C3 && Leading + C2 == 32 && OneUseOrZExtW && !IsCANDI) {
1005         SDNode *SRLIW = CurDAG->getMachineNode(
1006             RISCV::SRLIW, DL, XLenVT, X,
1007             CurDAG->getTargetConstant(C3 - C2, DL, XLenVT));
1008         SDNode *SLLI =
1009             CurDAG->getMachineNode(RISCV::SLLI, DL, XLenVT, SDValue(SRLIW, 0),
1010                                    CurDAG->getTargetConstant(C3, DL, XLenVT));
1011         ReplaceNode(Node, SLLI);
1012         return;
1013       }
1014     }
1015 
1016     break;
1017   }
1018   case ISD::MUL: {
1019     // Special case for calculating (mul (and X, C2), C1) where the full product
1020     // fits in XLen bits. We can shift X left by the number of leading zeros in
1021     // C2 and shift C1 left by XLen-lzcnt(C2). This will ensure the final
1022     // product has XLen trailing zeros, putting it in the output of MULHU. This
1023     // can avoid materializing a constant in a register for C2.
1024 
1025     // RHS should be a constant.
1026     auto *N1C = dyn_cast<ConstantSDNode>(Node->getOperand(1));
1027     if (!N1C || !N1C->hasOneUse())
1028       break;
1029 
1030     // LHS should be an AND with constant.
1031     SDValue N0 = Node->getOperand(0);
1032     if (N0.getOpcode() != ISD::AND || !isa<ConstantSDNode>(N0.getOperand(1)))
1033       break;
1034 
1035     uint64_t C2 = cast<ConstantSDNode>(N0.getOperand(1))->getZExtValue();
1036 
1037     // Constant should be a mask.
1038     if (!isMask_64(C2))
1039       break;
1040 
1041     // This should be the only use of the AND unless we will use
1042     // (SRLI (SLLI X, 32), 32). We don't use a shift pair for other AND
1043     // constants.
1044     if (!N0.hasOneUse() && C2 != UINT64_C(0xFFFFFFFF))
1045       break;
1046 
1047     // If this can be an ANDI, ZEXT.H or ZEXT.W we don't need to do this
1048     // optimization.
1049     if (isInt<12>(C2) ||
1050         (C2 == UINT64_C(0xFFFF) &&
1051          (Subtarget->hasStdExtZbb() || Subtarget->hasStdExtZbp())) ||
1052         (C2 == UINT64_C(0xFFFFFFFF) && Subtarget->hasStdExtZba()))
1053       break;
1054 
1055     // We need to shift left the AND input and C1 by a total of XLen bits.
1056 
1057     // How far left do we need to shift the AND input?
1058     unsigned XLen = Subtarget->getXLen();
1059     unsigned LeadingZeros = XLen - (64 - countLeadingZeros(C2));
1060 
1061     // The constant gets shifted by the remaining amount unless that would
1062     // shift bits out.
1063     uint64_t C1 = N1C->getZExtValue();
1064     unsigned ConstantShift = XLen - LeadingZeros;
1065     if (ConstantShift > (XLen - (64 - countLeadingZeros(C1))))
1066       break;
1067 
1068     uint64_t ShiftedC1 = C1 << ConstantShift;
1069     // If this RV32, we need to sign extend the constant.
1070     if (XLen == 32)
1071       ShiftedC1 = SignExtend64<32>(ShiftedC1);
1072 
1073     // Create (mulhu (slli X, lzcnt(C2)), C1 << (XLen - lzcnt(C2))).
1074     SDNode *Imm = selectImm(CurDAG, DL, VT, ShiftedC1, *Subtarget);
1075     SDNode *SLLI =
1076         CurDAG->getMachineNode(RISCV::SLLI, DL, VT, N0.getOperand(0),
1077                                CurDAG->getTargetConstant(LeadingZeros, DL, VT));
1078     SDNode *MULHU = CurDAG->getMachineNode(RISCV::MULHU, DL, VT,
1079                                            SDValue(SLLI, 0), SDValue(Imm, 0));
1080     ReplaceNode(Node, MULHU);
1081     return;
1082   }
1083   case ISD::INTRINSIC_WO_CHAIN: {
1084     unsigned IntNo = Node->getConstantOperandVal(0);
1085     switch (IntNo) {
1086       // By default we do not custom select any intrinsic.
1087     default:
1088       break;
1089     case Intrinsic::riscv_vmsgeu:
1090     case Intrinsic::riscv_vmsge: {
1091       SDValue Src1 = Node->getOperand(1);
1092       SDValue Src2 = Node->getOperand(2);
1093       bool IsUnsigned = IntNo == Intrinsic::riscv_vmsgeu;
1094       bool IsCmpUnsignedZero = false;
1095       // Only custom select scalar second operand.
1096       if (Src2.getValueType() != XLenVT)
1097         break;
1098       // Small constants are handled with patterns.
1099       if (auto *C = dyn_cast<ConstantSDNode>(Src2)) {
1100         int64_t CVal = C->getSExtValue();
1101         if (CVal >= -15 && CVal <= 16) {
1102           if (!IsUnsigned || CVal != 0)
1103             break;
1104           IsCmpUnsignedZero = true;
1105         }
1106       }
1107       MVT Src1VT = Src1.getSimpleValueType();
1108       unsigned VMSLTOpcode, VMNANDOpcode, VMSetOpcode;
1109       switch (RISCVTargetLowering::getLMUL(Src1VT)) {
1110       default:
1111         llvm_unreachable("Unexpected LMUL!");
1112 #define CASE_VMSLT_VMNAND_VMSET_OPCODES(lmulenum, suffix, suffix_b)            \
1113   case RISCVII::VLMUL::lmulenum:                                               \
1114     VMSLTOpcode = IsUnsigned ? RISCV::PseudoVMSLTU_VX_##suffix                 \
1115                              : RISCV::PseudoVMSLT_VX_##suffix;                 \
1116     VMNANDOpcode = RISCV::PseudoVMNAND_MM_##suffix;                            \
1117     VMSetOpcode = RISCV::PseudoVMSET_M_##suffix_b;                             \
1118     break;
1119         CASE_VMSLT_VMNAND_VMSET_OPCODES(LMUL_F8, MF8, B1)
1120         CASE_VMSLT_VMNAND_VMSET_OPCODES(LMUL_F4, MF4, B2)
1121         CASE_VMSLT_VMNAND_VMSET_OPCODES(LMUL_F2, MF2, B4)
1122         CASE_VMSLT_VMNAND_VMSET_OPCODES(LMUL_1, M1, B8)
1123         CASE_VMSLT_VMNAND_VMSET_OPCODES(LMUL_2, M2, B16)
1124         CASE_VMSLT_VMNAND_VMSET_OPCODES(LMUL_4, M4, B32)
1125         CASE_VMSLT_VMNAND_VMSET_OPCODES(LMUL_8, M8, B64)
1126 #undef CASE_VMSLT_VMNAND_VMSET_OPCODES
1127       }
1128       SDValue SEW = CurDAG->getTargetConstant(
1129           Log2_32(Src1VT.getScalarSizeInBits()), DL, XLenVT);
1130       SDValue VL;
1131       selectVLOp(Node->getOperand(3), VL);
1132 
1133       // If vmsgeu with 0 immediate, expand it to vmset.
1134       if (IsCmpUnsignedZero) {
1135         ReplaceNode(Node, CurDAG->getMachineNode(VMSetOpcode, DL, VT, VL, SEW));
1136         return;
1137       }
1138 
1139       // Expand to
1140       // vmslt{u}.vx vd, va, x; vmnand.mm vd, vd, vd
1141       SDValue Cmp = SDValue(
1142           CurDAG->getMachineNode(VMSLTOpcode, DL, VT, {Src1, Src2, VL, SEW}),
1143           0);
1144       ReplaceNode(Node, CurDAG->getMachineNode(VMNANDOpcode, DL, VT,
1145                                                {Cmp, Cmp, VL, SEW}));
1146       return;
1147     }
1148     case Intrinsic::riscv_vmsgeu_mask:
1149     case Intrinsic::riscv_vmsge_mask: {
1150       SDValue Src1 = Node->getOperand(2);
1151       SDValue Src2 = Node->getOperand(3);
1152       bool IsUnsigned = IntNo == Intrinsic::riscv_vmsgeu_mask;
1153       bool IsCmpUnsignedZero = false;
1154       // Only custom select scalar second operand.
1155       if (Src2.getValueType() != XLenVT)
1156         break;
1157       // Small constants are handled with patterns.
1158       if (auto *C = dyn_cast<ConstantSDNode>(Src2)) {
1159         int64_t CVal = C->getSExtValue();
1160         if (CVal >= -15 && CVal <= 16) {
1161           if (!IsUnsigned || CVal != 0)
1162             break;
1163           IsCmpUnsignedZero = true;
1164         }
1165       }
1166       MVT Src1VT = Src1.getSimpleValueType();
1167       unsigned VMSLTOpcode, VMSLTMaskOpcode, VMXOROpcode, VMANDNOpcode,
1168           VMOROpcode;
1169       switch (RISCVTargetLowering::getLMUL(Src1VT)) {
1170       default:
1171         llvm_unreachable("Unexpected LMUL!");
1172 #define CASE_VMSLT_OPCODES(lmulenum, suffix, suffix_b)                         \
1173   case RISCVII::VLMUL::lmulenum:                                               \
1174     VMSLTOpcode = IsUnsigned ? RISCV::PseudoVMSLTU_VX_##suffix                 \
1175                              : RISCV::PseudoVMSLT_VX_##suffix;                 \
1176     VMSLTMaskOpcode = IsUnsigned ? RISCV::PseudoVMSLTU_VX_##suffix##_MASK      \
1177                                  : RISCV::PseudoVMSLT_VX_##suffix##_MASK;      \
1178     break;
1179         CASE_VMSLT_OPCODES(LMUL_F8, MF8, B1)
1180         CASE_VMSLT_OPCODES(LMUL_F4, MF4, B2)
1181         CASE_VMSLT_OPCODES(LMUL_F2, MF2, B4)
1182         CASE_VMSLT_OPCODES(LMUL_1, M1, B8)
1183         CASE_VMSLT_OPCODES(LMUL_2, M2, B16)
1184         CASE_VMSLT_OPCODES(LMUL_4, M4, B32)
1185         CASE_VMSLT_OPCODES(LMUL_8, M8, B64)
1186 #undef CASE_VMSLT_OPCODES
1187       }
1188       // Mask operations use the LMUL from the mask type.
1189       switch (RISCVTargetLowering::getLMUL(VT)) {
1190       default:
1191         llvm_unreachable("Unexpected LMUL!");
1192 #define CASE_VMXOR_VMANDN_VMOR_OPCODES(lmulenum, suffix)                       \
1193   case RISCVII::VLMUL::lmulenum:                                               \
1194     VMXOROpcode = RISCV::PseudoVMXOR_MM_##suffix;                              \
1195     VMANDNOpcode = RISCV::PseudoVMANDN_MM_##suffix;                            \
1196     VMOROpcode = RISCV::PseudoVMOR_MM_##suffix;                                \
1197     break;
1198         CASE_VMXOR_VMANDN_VMOR_OPCODES(LMUL_F8, MF8)
1199         CASE_VMXOR_VMANDN_VMOR_OPCODES(LMUL_F4, MF4)
1200         CASE_VMXOR_VMANDN_VMOR_OPCODES(LMUL_F2, MF2)
1201         CASE_VMXOR_VMANDN_VMOR_OPCODES(LMUL_1, M1)
1202         CASE_VMXOR_VMANDN_VMOR_OPCODES(LMUL_2, M2)
1203         CASE_VMXOR_VMANDN_VMOR_OPCODES(LMUL_4, M4)
1204         CASE_VMXOR_VMANDN_VMOR_OPCODES(LMUL_8, M8)
1205 #undef CASE_VMXOR_VMANDN_VMOR_OPCODES
1206       }
1207       SDValue SEW = CurDAG->getTargetConstant(
1208           Log2_32(Src1VT.getScalarSizeInBits()), DL, XLenVT);
1209       SDValue MaskSEW = CurDAG->getTargetConstant(0, DL, XLenVT);
1210       SDValue VL;
1211       selectVLOp(Node->getOperand(5), VL);
1212       SDValue MaskedOff = Node->getOperand(1);
1213       SDValue Mask = Node->getOperand(4);
1214 
1215       // If vmsgeu_mask with 0 immediate, expand it to vmor mask, maskedoff.
1216       if (IsCmpUnsignedZero) {
1217         // We don't need vmor if the MaskedOff and the Mask are the same
1218         // value.
1219         if (Mask == MaskedOff) {
1220           ReplaceUses(Node, Mask.getNode());
1221           return;
1222         }
1223         ReplaceNode(Node,
1224                     CurDAG->getMachineNode(VMOROpcode, DL, VT,
1225                                            {Mask, MaskedOff, VL, MaskSEW}));
1226         return;
1227       }
1228 
1229       // If the MaskedOff value and the Mask are the same value use
1230       // vmslt{u}.vx vt, va, x;  vmandn.mm vd, vd, vt
1231       // This avoids needing to copy v0 to vd before starting the next sequence.
1232       if (Mask == MaskedOff) {
1233         SDValue Cmp = SDValue(
1234             CurDAG->getMachineNode(VMSLTOpcode, DL, VT, {Src1, Src2, VL, SEW}),
1235             0);
1236         ReplaceNode(Node, CurDAG->getMachineNode(VMANDNOpcode, DL, VT,
1237                                                  {Mask, Cmp, VL, MaskSEW}));
1238         return;
1239       }
1240 
1241       // Mask needs to be copied to V0.
1242       SDValue Chain = CurDAG->getCopyToReg(CurDAG->getEntryNode(), DL,
1243                                            RISCV::V0, Mask, SDValue());
1244       SDValue Glue = Chain.getValue(1);
1245       SDValue V0 = CurDAG->getRegister(RISCV::V0, VT);
1246 
1247       // Otherwise use
1248       // vmslt{u}.vx vd, va, x, v0.t; vmxor.mm vd, vd, v0
1249       // The result is mask undisturbed.
1250       // We use the same instructions to emulate mask agnostic behavior, because
1251       // the agnostic result can be either undisturbed or all 1.
1252       SDValue Cmp = SDValue(
1253           CurDAG->getMachineNode(VMSLTMaskOpcode, DL, VT,
1254                                  {MaskedOff, Src1, Src2, V0, VL, SEW, Glue}),
1255           0);
1256       // vmxor.mm vd, vd, v0 is used to update active value.
1257       ReplaceNode(Node, CurDAG->getMachineNode(VMXOROpcode, DL, VT,
1258                                                {Cmp, Mask, VL, MaskSEW}));
1259       return;
1260     }
1261     case Intrinsic::riscv_vsetvli_opt:
1262     case Intrinsic::riscv_vsetvlimax_opt:
1263       return selectVSETVLI(Node);
1264     }
1265     break;
1266   }
1267   case ISD::INTRINSIC_W_CHAIN: {
1268     unsigned IntNo = cast<ConstantSDNode>(Node->getOperand(1))->getZExtValue();
1269     switch (IntNo) {
1270       // By default we do not custom select any intrinsic.
1271     default:
1272       break;
1273     case Intrinsic::riscv_vsetvli:
1274     case Intrinsic::riscv_vsetvlimax:
1275       return selectVSETVLI(Node);
1276     case Intrinsic::riscv_vlseg2:
1277     case Intrinsic::riscv_vlseg3:
1278     case Intrinsic::riscv_vlseg4:
1279     case Intrinsic::riscv_vlseg5:
1280     case Intrinsic::riscv_vlseg6:
1281     case Intrinsic::riscv_vlseg7:
1282     case Intrinsic::riscv_vlseg8: {
1283       selectVLSEG(Node, /*IsMasked*/ false, /*IsStrided*/ false);
1284       return;
1285     }
1286     case Intrinsic::riscv_vlseg2_mask:
1287     case Intrinsic::riscv_vlseg3_mask:
1288     case Intrinsic::riscv_vlseg4_mask:
1289     case Intrinsic::riscv_vlseg5_mask:
1290     case Intrinsic::riscv_vlseg6_mask:
1291     case Intrinsic::riscv_vlseg7_mask:
1292     case Intrinsic::riscv_vlseg8_mask: {
1293       selectVLSEG(Node, /*IsMasked*/ true, /*IsStrided*/ false);
1294       return;
1295     }
1296     case Intrinsic::riscv_vlsseg2:
1297     case Intrinsic::riscv_vlsseg3:
1298     case Intrinsic::riscv_vlsseg4:
1299     case Intrinsic::riscv_vlsseg5:
1300     case Intrinsic::riscv_vlsseg6:
1301     case Intrinsic::riscv_vlsseg7:
1302     case Intrinsic::riscv_vlsseg8: {
1303       selectVLSEG(Node, /*IsMasked*/ false, /*IsStrided*/ true);
1304       return;
1305     }
1306     case Intrinsic::riscv_vlsseg2_mask:
1307     case Intrinsic::riscv_vlsseg3_mask:
1308     case Intrinsic::riscv_vlsseg4_mask:
1309     case Intrinsic::riscv_vlsseg5_mask:
1310     case Intrinsic::riscv_vlsseg6_mask:
1311     case Intrinsic::riscv_vlsseg7_mask:
1312     case Intrinsic::riscv_vlsseg8_mask: {
1313       selectVLSEG(Node, /*IsMasked*/ true, /*IsStrided*/ true);
1314       return;
1315     }
1316     case Intrinsic::riscv_vloxseg2:
1317     case Intrinsic::riscv_vloxseg3:
1318     case Intrinsic::riscv_vloxseg4:
1319     case Intrinsic::riscv_vloxseg5:
1320     case Intrinsic::riscv_vloxseg6:
1321     case Intrinsic::riscv_vloxseg7:
1322     case Intrinsic::riscv_vloxseg8:
1323       selectVLXSEG(Node, /*IsMasked*/ false, /*IsOrdered*/ true);
1324       return;
1325     case Intrinsic::riscv_vluxseg2:
1326     case Intrinsic::riscv_vluxseg3:
1327     case Intrinsic::riscv_vluxseg4:
1328     case Intrinsic::riscv_vluxseg5:
1329     case Intrinsic::riscv_vluxseg6:
1330     case Intrinsic::riscv_vluxseg7:
1331     case Intrinsic::riscv_vluxseg8:
1332       selectVLXSEG(Node, /*IsMasked*/ false, /*IsOrdered*/ false);
1333       return;
1334     case Intrinsic::riscv_vloxseg2_mask:
1335     case Intrinsic::riscv_vloxseg3_mask:
1336     case Intrinsic::riscv_vloxseg4_mask:
1337     case Intrinsic::riscv_vloxseg5_mask:
1338     case Intrinsic::riscv_vloxseg6_mask:
1339     case Intrinsic::riscv_vloxseg7_mask:
1340     case Intrinsic::riscv_vloxseg8_mask:
1341       selectVLXSEG(Node, /*IsMasked*/ true, /*IsOrdered*/ true);
1342       return;
1343     case Intrinsic::riscv_vluxseg2_mask:
1344     case Intrinsic::riscv_vluxseg3_mask:
1345     case Intrinsic::riscv_vluxseg4_mask:
1346     case Intrinsic::riscv_vluxseg5_mask:
1347     case Intrinsic::riscv_vluxseg6_mask:
1348     case Intrinsic::riscv_vluxseg7_mask:
1349     case Intrinsic::riscv_vluxseg8_mask:
1350       selectVLXSEG(Node, /*IsMasked*/ true, /*IsOrdered*/ false);
1351       return;
1352     case Intrinsic::riscv_vlseg8ff:
1353     case Intrinsic::riscv_vlseg7ff:
1354     case Intrinsic::riscv_vlseg6ff:
1355     case Intrinsic::riscv_vlseg5ff:
1356     case Intrinsic::riscv_vlseg4ff:
1357     case Intrinsic::riscv_vlseg3ff:
1358     case Intrinsic::riscv_vlseg2ff: {
1359       selectVLSEGFF(Node, /*IsMasked*/ false);
1360       return;
1361     }
1362     case Intrinsic::riscv_vlseg8ff_mask:
1363     case Intrinsic::riscv_vlseg7ff_mask:
1364     case Intrinsic::riscv_vlseg6ff_mask:
1365     case Intrinsic::riscv_vlseg5ff_mask:
1366     case Intrinsic::riscv_vlseg4ff_mask:
1367     case Intrinsic::riscv_vlseg3ff_mask:
1368     case Intrinsic::riscv_vlseg2ff_mask: {
1369       selectVLSEGFF(Node, /*IsMasked*/ true);
1370       return;
1371     }
1372     case Intrinsic::riscv_vloxei:
1373     case Intrinsic::riscv_vloxei_mask:
1374     case Intrinsic::riscv_vluxei:
1375     case Intrinsic::riscv_vluxei_mask: {
1376       bool IsMasked = IntNo == Intrinsic::riscv_vloxei_mask ||
1377                       IntNo == Intrinsic::riscv_vluxei_mask;
1378       bool IsOrdered = IntNo == Intrinsic::riscv_vloxei ||
1379                        IntNo == Intrinsic::riscv_vloxei_mask;
1380 
1381       MVT VT = Node->getSimpleValueType(0);
1382       unsigned Log2SEW = Log2_32(VT.getScalarSizeInBits());
1383 
1384       unsigned CurOp = 2;
1385       // Masked intrinsic only have TU version pseduo instructions.
1386       bool IsTU = IsMasked || (!IsMasked && !Node->getOperand(CurOp).isUndef());
1387       SmallVector<SDValue, 8> Operands;
1388       if (IsTU)
1389         Operands.push_back(Node->getOperand(CurOp++));
1390       else
1391         // Skip the undef passthru operand for nomask TA version pseudo
1392         CurOp++;
1393 
1394       MVT IndexVT;
1395       addVectorLoadStoreOperands(Node, Log2SEW, DL, CurOp, IsMasked,
1396                                  /*IsStridedOrIndexed*/ true, Operands,
1397                                  /*IsLoad=*/true, &IndexVT);
1398 
1399       assert(VT.getVectorElementCount() == IndexVT.getVectorElementCount() &&
1400              "Element count mismatch");
1401 
1402       RISCVII::VLMUL LMUL = RISCVTargetLowering::getLMUL(VT);
1403       RISCVII::VLMUL IndexLMUL = RISCVTargetLowering::getLMUL(IndexVT);
1404       unsigned IndexLog2EEW = Log2_32(IndexVT.getScalarSizeInBits());
1405       if (IndexLog2EEW == 6 && !Subtarget->is64Bit()) {
1406         report_fatal_error("The V extension does not support EEW=64 for index "
1407                            "values when XLEN=32");
1408       }
1409       const RISCV::VLX_VSXPseudo *P = RISCV::getVLXPseudo(
1410           IsMasked, IsTU, IsOrdered, IndexLog2EEW, static_cast<unsigned>(LMUL),
1411           static_cast<unsigned>(IndexLMUL));
1412       MachineSDNode *Load =
1413           CurDAG->getMachineNode(P->Pseudo, DL, Node->getVTList(), Operands);
1414 
1415       if (auto *MemOp = dyn_cast<MemSDNode>(Node))
1416         CurDAG->setNodeMemRefs(Load, {MemOp->getMemOperand()});
1417 
1418       ReplaceNode(Node, Load);
1419       return;
1420     }
1421     case Intrinsic::riscv_vlm:
1422     case Intrinsic::riscv_vle:
1423     case Intrinsic::riscv_vle_mask:
1424     case Intrinsic::riscv_vlse:
1425     case Intrinsic::riscv_vlse_mask: {
1426       bool IsMasked = IntNo == Intrinsic::riscv_vle_mask ||
1427                       IntNo == Intrinsic::riscv_vlse_mask;
1428       bool IsStrided =
1429           IntNo == Intrinsic::riscv_vlse || IntNo == Intrinsic::riscv_vlse_mask;
1430 
1431       MVT VT = Node->getSimpleValueType(0);
1432       unsigned Log2SEW = Log2_32(VT.getScalarSizeInBits());
1433 
1434       unsigned CurOp = 2;
1435       // The riscv_vlm intrinsic are always tail agnostic and no passthru operand.
1436       bool HasPassthruOperand = IntNo != Intrinsic::riscv_vlm;
1437       // Masked intrinsic only have TU version pseduo instructions.
1438       bool IsTU =
1439           HasPassthruOperand &&
1440           ((!IsMasked && !Node->getOperand(CurOp).isUndef()) || IsMasked);
1441       SmallVector<SDValue, 8> Operands;
1442       if (IsTU)
1443         Operands.push_back(Node->getOperand(CurOp++));
1444       else if (HasPassthruOperand)
1445         // Skip the undef passthru operand for nomask TA version pseudo
1446         CurOp++;
1447 
1448       addVectorLoadStoreOperands(Node, Log2SEW, DL, CurOp, IsMasked, IsStrided,
1449                                  Operands, /*IsLoad=*/true);
1450 
1451       RISCVII::VLMUL LMUL = RISCVTargetLowering::getLMUL(VT);
1452       const RISCV::VLEPseudo *P =
1453           RISCV::getVLEPseudo(IsMasked, IsTU, IsStrided, /*FF*/ false, Log2SEW,
1454                               static_cast<unsigned>(LMUL));
1455       MachineSDNode *Load =
1456           CurDAG->getMachineNode(P->Pseudo, DL, Node->getVTList(), Operands);
1457 
1458       if (auto *MemOp = dyn_cast<MemSDNode>(Node))
1459         CurDAG->setNodeMemRefs(Load, {MemOp->getMemOperand()});
1460 
1461       ReplaceNode(Node, Load);
1462       return;
1463     }
1464     case Intrinsic::riscv_vleff:
1465     case Intrinsic::riscv_vleff_mask: {
1466       bool IsMasked = IntNo == Intrinsic::riscv_vleff_mask;
1467 
1468       MVT VT = Node->getSimpleValueType(0);
1469       unsigned SEW = VT.getScalarSizeInBits();
1470       unsigned Log2SEW = Log2_32(SEW);
1471 
1472       unsigned CurOp = 2;
1473       // Masked intrinsic only have TU version pseduo instructions.
1474       bool IsTU = IsMasked || (!IsMasked && !Node->getOperand(CurOp).isUndef());
1475       SmallVector<SDValue, 7> Operands;
1476       if (IsTU)
1477         Operands.push_back(Node->getOperand(CurOp++));
1478       else
1479         // Skip the undef passthru operand for nomask TA version pseudo
1480         CurOp++;
1481 
1482       addVectorLoadStoreOperands(Node, Log2SEW, DL, CurOp, IsMasked,
1483                                  /*IsStridedOrIndexed*/ false, Operands,
1484                                  /*IsLoad=*/true);
1485 
1486       RISCVII::VLMUL LMUL = RISCVTargetLowering::getLMUL(VT);
1487       const RISCV::VLEPseudo *P =
1488           RISCV::getVLEPseudo(IsMasked, IsTU, /*Strided*/ false, /*FF*/ true,
1489                               Log2SEW, static_cast<unsigned>(LMUL));
1490       MachineSDNode *Load =
1491           CurDAG->getMachineNode(P->Pseudo, DL, Node->getValueType(0),
1492                                  MVT::Other, MVT::Glue, Operands);
1493       bool TailAgnostic = !IsTU;
1494       bool MaskAgnostic = false;
1495       if (IsMasked) {
1496         uint64_t Policy =
1497             Node->getConstantOperandVal(Node->getNumOperands() - 1);
1498         TailAgnostic = Policy & RISCVII::TAIL_AGNOSTIC;
1499         MaskAgnostic = Policy & RISCVII::MASK_AGNOSTIC;
1500       }
1501       unsigned VType =
1502           RISCVVType::encodeVTYPE(LMUL, SEW, TailAgnostic, MaskAgnostic);
1503       SDValue VTypeOp = CurDAG->getTargetConstant(VType, DL, XLenVT);
1504       SDNode *ReadVL =
1505           CurDAG->getMachineNode(RISCV::PseudoReadVL, DL, XLenVT, VTypeOp,
1506                                  /*Glue*/ SDValue(Load, 2));
1507 
1508       if (auto *MemOp = dyn_cast<MemSDNode>(Node))
1509         CurDAG->setNodeMemRefs(Load, {MemOp->getMemOperand()});
1510 
1511       ReplaceUses(SDValue(Node, 0), SDValue(Load, 0));
1512       ReplaceUses(SDValue(Node, 1), SDValue(ReadVL, 0)); // VL
1513       ReplaceUses(SDValue(Node, 2), SDValue(Load, 1));   // Chain
1514       CurDAG->RemoveDeadNode(Node);
1515       return;
1516     }
1517     }
1518     break;
1519   }
1520   case ISD::INTRINSIC_VOID: {
1521     unsigned IntNo = cast<ConstantSDNode>(Node->getOperand(1))->getZExtValue();
1522     switch (IntNo) {
1523     case Intrinsic::riscv_vsseg2:
1524     case Intrinsic::riscv_vsseg3:
1525     case Intrinsic::riscv_vsseg4:
1526     case Intrinsic::riscv_vsseg5:
1527     case Intrinsic::riscv_vsseg6:
1528     case Intrinsic::riscv_vsseg7:
1529     case Intrinsic::riscv_vsseg8: {
1530       selectVSSEG(Node, /*IsMasked*/ false, /*IsStrided*/ false);
1531       return;
1532     }
1533     case Intrinsic::riscv_vsseg2_mask:
1534     case Intrinsic::riscv_vsseg3_mask:
1535     case Intrinsic::riscv_vsseg4_mask:
1536     case Intrinsic::riscv_vsseg5_mask:
1537     case Intrinsic::riscv_vsseg6_mask:
1538     case Intrinsic::riscv_vsseg7_mask:
1539     case Intrinsic::riscv_vsseg8_mask: {
1540       selectVSSEG(Node, /*IsMasked*/ true, /*IsStrided*/ false);
1541       return;
1542     }
1543     case Intrinsic::riscv_vssseg2:
1544     case Intrinsic::riscv_vssseg3:
1545     case Intrinsic::riscv_vssseg4:
1546     case Intrinsic::riscv_vssseg5:
1547     case Intrinsic::riscv_vssseg6:
1548     case Intrinsic::riscv_vssseg7:
1549     case Intrinsic::riscv_vssseg8: {
1550       selectVSSEG(Node, /*IsMasked*/ false, /*IsStrided*/ true);
1551       return;
1552     }
1553     case Intrinsic::riscv_vssseg2_mask:
1554     case Intrinsic::riscv_vssseg3_mask:
1555     case Intrinsic::riscv_vssseg4_mask:
1556     case Intrinsic::riscv_vssseg5_mask:
1557     case Intrinsic::riscv_vssseg6_mask:
1558     case Intrinsic::riscv_vssseg7_mask:
1559     case Intrinsic::riscv_vssseg8_mask: {
1560       selectVSSEG(Node, /*IsMasked*/ true, /*IsStrided*/ true);
1561       return;
1562     }
1563     case Intrinsic::riscv_vsoxseg2:
1564     case Intrinsic::riscv_vsoxseg3:
1565     case Intrinsic::riscv_vsoxseg4:
1566     case Intrinsic::riscv_vsoxseg5:
1567     case Intrinsic::riscv_vsoxseg6:
1568     case Intrinsic::riscv_vsoxseg7:
1569     case Intrinsic::riscv_vsoxseg8:
1570       selectVSXSEG(Node, /*IsMasked*/ false, /*IsOrdered*/ true);
1571       return;
1572     case Intrinsic::riscv_vsuxseg2:
1573     case Intrinsic::riscv_vsuxseg3:
1574     case Intrinsic::riscv_vsuxseg4:
1575     case Intrinsic::riscv_vsuxseg5:
1576     case Intrinsic::riscv_vsuxseg6:
1577     case Intrinsic::riscv_vsuxseg7:
1578     case Intrinsic::riscv_vsuxseg8:
1579       selectVSXSEG(Node, /*IsMasked*/ false, /*IsOrdered*/ false);
1580       return;
1581     case Intrinsic::riscv_vsoxseg2_mask:
1582     case Intrinsic::riscv_vsoxseg3_mask:
1583     case Intrinsic::riscv_vsoxseg4_mask:
1584     case Intrinsic::riscv_vsoxseg5_mask:
1585     case Intrinsic::riscv_vsoxseg6_mask:
1586     case Intrinsic::riscv_vsoxseg7_mask:
1587     case Intrinsic::riscv_vsoxseg8_mask:
1588       selectVSXSEG(Node, /*IsMasked*/ true, /*IsOrdered*/ true);
1589       return;
1590     case Intrinsic::riscv_vsuxseg2_mask:
1591     case Intrinsic::riscv_vsuxseg3_mask:
1592     case Intrinsic::riscv_vsuxseg4_mask:
1593     case Intrinsic::riscv_vsuxseg5_mask:
1594     case Intrinsic::riscv_vsuxseg6_mask:
1595     case Intrinsic::riscv_vsuxseg7_mask:
1596     case Intrinsic::riscv_vsuxseg8_mask:
1597       selectVSXSEG(Node, /*IsMasked*/ true, /*IsOrdered*/ false);
1598       return;
1599     case Intrinsic::riscv_vsoxei:
1600     case Intrinsic::riscv_vsoxei_mask:
1601     case Intrinsic::riscv_vsuxei:
1602     case Intrinsic::riscv_vsuxei_mask: {
1603       bool IsMasked = IntNo == Intrinsic::riscv_vsoxei_mask ||
1604                       IntNo == Intrinsic::riscv_vsuxei_mask;
1605       bool IsOrdered = IntNo == Intrinsic::riscv_vsoxei ||
1606                        IntNo == Intrinsic::riscv_vsoxei_mask;
1607 
1608       MVT VT = Node->getOperand(2)->getSimpleValueType(0);
1609       unsigned Log2SEW = Log2_32(VT.getScalarSizeInBits());
1610 
1611       unsigned CurOp = 2;
1612       SmallVector<SDValue, 8> Operands;
1613       Operands.push_back(Node->getOperand(CurOp++)); // Store value.
1614 
1615       MVT IndexVT;
1616       addVectorLoadStoreOperands(Node, Log2SEW, DL, CurOp, IsMasked,
1617                                  /*IsStridedOrIndexed*/ true, Operands,
1618                                  /*IsLoad=*/false, &IndexVT);
1619 
1620       assert(VT.getVectorElementCount() == IndexVT.getVectorElementCount() &&
1621              "Element count mismatch");
1622 
1623       RISCVII::VLMUL LMUL = RISCVTargetLowering::getLMUL(VT);
1624       RISCVII::VLMUL IndexLMUL = RISCVTargetLowering::getLMUL(IndexVT);
1625       unsigned IndexLog2EEW = Log2_32(IndexVT.getScalarSizeInBits());
1626       if (IndexLog2EEW == 6 && !Subtarget->is64Bit()) {
1627         report_fatal_error("The V extension does not support EEW=64 for index "
1628                            "values when XLEN=32");
1629       }
1630       const RISCV::VLX_VSXPseudo *P = RISCV::getVSXPseudo(
1631           IsMasked, /*TU*/ false, IsOrdered, IndexLog2EEW,
1632           static_cast<unsigned>(LMUL), static_cast<unsigned>(IndexLMUL));
1633       MachineSDNode *Store =
1634           CurDAG->getMachineNode(P->Pseudo, DL, Node->getVTList(), Operands);
1635 
1636       if (auto *MemOp = dyn_cast<MemSDNode>(Node))
1637         CurDAG->setNodeMemRefs(Store, {MemOp->getMemOperand()});
1638 
1639       ReplaceNode(Node, Store);
1640       return;
1641     }
1642     case Intrinsic::riscv_vsm:
1643     case Intrinsic::riscv_vse:
1644     case Intrinsic::riscv_vse_mask:
1645     case Intrinsic::riscv_vsse:
1646     case Intrinsic::riscv_vsse_mask: {
1647       bool IsMasked = IntNo == Intrinsic::riscv_vse_mask ||
1648                       IntNo == Intrinsic::riscv_vsse_mask;
1649       bool IsStrided =
1650           IntNo == Intrinsic::riscv_vsse || IntNo == Intrinsic::riscv_vsse_mask;
1651 
1652       MVT VT = Node->getOperand(2)->getSimpleValueType(0);
1653       unsigned Log2SEW = Log2_32(VT.getScalarSizeInBits());
1654 
1655       unsigned CurOp = 2;
1656       SmallVector<SDValue, 8> Operands;
1657       Operands.push_back(Node->getOperand(CurOp++)); // Store value.
1658 
1659       addVectorLoadStoreOperands(Node, Log2SEW, DL, CurOp, IsMasked, IsStrided,
1660                                  Operands);
1661 
1662       RISCVII::VLMUL LMUL = RISCVTargetLowering::getLMUL(VT);
1663       const RISCV::VSEPseudo *P = RISCV::getVSEPseudo(
1664           IsMasked, IsStrided, Log2SEW, static_cast<unsigned>(LMUL));
1665       MachineSDNode *Store =
1666           CurDAG->getMachineNode(P->Pseudo, DL, Node->getVTList(), Operands);
1667       if (auto *MemOp = dyn_cast<MemSDNode>(Node))
1668         CurDAG->setNodeMemRefs(Store, {MemOp->getMemOperand()});
1669 
1670       ReplaceNode(Node, Store);
1671       return;
1672     }
1673     }
1674     break;
1675   }
1676   case ISD::BITCAST: {
1677     MVT SrcVT = Node->getOperand(0).getSimpleValueType();
1678     // Just drop bitcasts between vectors if both are fixed or both are
1679     // scalable.
1680     if ((VT.isScalableVector() && SrcVT.isScalableVector()) ||
1681         (VT.isFixedLengthVector() && SrcVT.isFixedLengthVector())) {
1682       ReplaceUses(SDValue(Node, 0), Node->getOperand(0));
1683       CurDAG->RemoveDeadNode(Node);
1684       return;
1685     }
1686     break;
1687   }
1688   case ISD::INSERT_SUBVECTOR: {
1689     SDValue V = Node->getOperand(0);
1690     SDValue SubV = Node->getOperand(1);
1691     SDLoc DL(SubV);
1692     auto Idx = Node->getConstantOperandVal(2);
1693     MVT SubVecVT = SubV.getSimpleValueType();
1694 
1695     const RISCVTargetLowering &TLI = *Subtarget->getTargetLowering();
1696     MVT SubVecContainerVT = SubVecVT;
1697     // Establish the correct scalable-vector types for any fixed-length type.
1698     if (SubVecVT.isFixedLengthVector())
1699       SubVecContainerVT = TLI.getContainerForFixedLengthVector(SubVecVT);
1700     if (VT.isFixedLengthVector())
1701       VT = TLI.getContainerForFixedLengthVector(VT);
1702 
1703     const auto *TRI = Subtarget->getRegisterInfo();
1704     unsigned SubRegIdx;
1705     std::tie(SubRegIdx, Idx) =
1706         RISCVTargetLowering::decomposeSubvectorInsertExtractToSubRegs(
1707             VT, SubVecContainerVT, Idx, TRI);
1708 
1709     // If the Idx hasn't been completely eliminated then this is a subvector
1710     // insert which doesn't naturally align to a vector register. These must
1711     // be handled using instructions to manipulate the vector registers.
1712     if (Idx != 0)
1713       break;
1714 
1715     RISCVII::VLMUL SubVecLMUL = RISCVTargetLowering::getLMUL(SubVecContainerVT);
1716     bool IsSubVecPartReg = SubVecLMUL == RISCVII::VLMUL::LMUL_F2 ||
1717                            SubVecLMUL == RISCVII::VLMUL::LMUL_F4 ||
1718                            SubVecLMUL == RISCVII::VLMUL::LMUL_F8;
1719     (void)IsSubVecPartReg; // Silence unused variable warning without asserts.
1720     assert((!IsSubVecPartReg || V.isUndef()) &&
1721            "Expecting lowering to have created legal INSERT_SUBVECTORs when "
1722            "the subvector is smaller than a full-sized register");
1723 
1724     // If we haven't set a SubRegIdx, then we must be going between
1725     // equally-sized LMUL groups (e.g. VR -> VR). This can be done as a copy.
1726     if (SubRegIdx == RISCV::NoSubRegister) {
1727       unsigned InRegClassID = RISCVTargetLowering::getRegClassIDForVecVT(VT);
1728       assert(RISCVTargetLowering::getRegClassIDForVecVT(SubVecContainerVT) ==
1729                  InRegClassID &&
1730              "Unexpected subvector extraction");
1731       SDValue RC = CurDAG->getTargetConstant(InRegClassID, DL, XLenVT);
1732       SDNode *NewNode = CurDAG->getMachineNode(TargetOpcode::COPY_TO_REGCLASS,
1733                                                DL, VT, SubV, RC);
1734       ReplaceNode(Node, NewNode);
1735       return;
1736     }
1737 
1738     SDValue Insert = CurDAG->getTargetInsertSubreg(SubRegIdx, DL, VT, V, SubV);
1739     ReplaceNode(Node, Insert.getNode());
1740     return;
1741   }
1742   case ISD::EXTRACT_SUBVECTOR: {
1743     SDValue V = Node->getOperand(0);
1744     auto Idx = Node->getConstantOperandVal(1);
1745     MVT InVT = V.getSimpleValueType();
1746     SDLoc DL(V);
1747 
1748     const RISCVTargetLowering &TLI = *Subtarget->getTargetLowering();
1749     MVT SubVecContainerVT = VT;
1750     // Establish the correct scalable-vector types for any fixed-length type.
1751     if (VT.isFixedLengthVector())
1752       SubVecContainerVT = TLI.getContainerForFixedLengthVector(VT);
1753     if (InVT.isFixedLengthVector())
1754       InVT = TLI.getContainerForFixedLengthVector(InVT);
1755 
1756     const auto *TRI = Subtarget->getRegisterInfo();
1757     unsigned SubRegIdx;
1758     std::tie(SubRegIdx, Idx) =
1759         RISCVTargetLowering::decomposeSubvectorInsertExtractToSubRegs(
1760             InVT, SubVecContainerVT, Idx, TRI);
1761 
1762     // If the Idx hasn't been completely eliminated then this is a subvector
1763     // extract which doesn't naturally align to a vector register. These must
1764     // be handled using instructions to manipulate the vector registers.
1765     if (Idx != 0)
1766       break;
1767 
1768     // If we haven't set a SubRegIdx, then we must be going between
1769     // equally-sized LMUL types (e.g. VR -> VR). This can be done as a copy.
1770     if (SubRegIdx == RISCV::NoSubRegister) {
1771       unsigned InRegClassID = RISCVTargetLowering::getRegClassIDForVecVT(InVT);
1772       assert(RISCVTargetLowering::getRegClassIDForVecVT(SubVecContainerVT) ==
1773                  InRegClassID &&
1774              "Unexpected subvector extraction");
1775       SDValue RC = CurDAG->getTargetConstant(InRegClassID, DL, XLenVT);
1776       SDNode *NewNode =
1777           CurDAG->getMachineNode(TargetOpcode::COPY_TO_REGCLASS, DL, VT, V, RC);
1778       ReplaceNode(Node, NewNode);
1779       return;
1780     }
1781 
1782     SDValue Extract = CurDAG->getTargetExtractSubreg(SubRegIdx, DL, VT, V);
1783     ReplaceNode(Node, Extract.getNode());
1784     return;
1785   }
1786   case ISD::SPLAT_VECTOR:
1787   case RISCVISD::VMV_S_X_VL:
1788   case RISCVISD::VFMV_S_F_VL:
1789   case RISCVISD::VMV_V_X_VL:
1790   case RISCVISD::VFMV_V_F_VL: {
1791     // Try to match splat of a scalar load to a strided load with stride of x0.
1792     bool IsScalarMove = Node->getOpcode() == RISCVISD::VMV_S_X_VL ||
1793                         Node->getOpcode() == RISCVISD::VFMV_S_F_VL;
1794     bool HasPassthruOperand = Node->getOpcode() != ISD::SPLAT_VECTOR;
1795     if (HasPassthruOperand && !Node->getOperand(0).isUndef())
1796       break;
1797     SDValue Src = HasPassthruOperand ? Node->getOperand(1) : Node->getOperand(0);
1798     auto *Ld = dyn_cast<LoadSDNode>(Src);
1799     if (!Ld)
1800       break;
1801     EVT MemVT = Ld->getMemoryVT();
1802     // The memory VT should be the same size as the element type.
1803     if (MemVT.getStoreSize() != VT.getVectorElementType().getStoreSize())
1804       break;
1805     if (!IsProfitableToFold(Src, Node, Node) ||
1806         !IsLegalToFold(Src, Node, Node, TM.getOptLevel()))
1807       break;
1808 
1809     SDValue VL;
1810     if (Node->getOpcode() == ISD::SPLAT_VECTOR)
1811       VL = CurDAG->getTargetConstant(RISCV::VLMaxSentinel, DL, XLenVT);
1812     else if (IsScalarMove) {
1813       // We could deal with more VL if we update the VSETVLI insert pass to
1814       // avoid introducing more VSETVLI.
1815       if (!isOneConstant(Node->getOperand(2)))
1816         break;
1817       selectVLOp(Node->getOperand(2), VL);
1818     } else
1819       selectVLOp(Node->getOperand(2), VL);
1820 
1821     unsigned Log2SEW = Log2_32(VT.getScalarSizeInBits());
1822     SDValue SEW = CurDAG->getTargetConstant(Log2SEW, DL, XLenVT);
1823 
1824     SDValue Operands[] = {Ld->getBasePtr(),
1825                           CurDAG->getRegister(RISCV::X0, XLenVT), VL, SEW,
1826                           Ld->getChain()};
1827 
1828     RISCVII::VLMUL LMUL = RISCVTargetLowering::getLMUL(VT);
1829     const RISCV::VLEPseudo *P = RISCV::getVLEPseudo(
1830         /*IsMasked*/ false, /*IsTU*/ false, /*IsStrided*/ true, /*FF*/ false,
1831         Log2SEW, static_cast<unsigned>(LMUL));
1832     MachineSDNode *Load =
1833         CurDAG->getMachineNode(P->Pseudo, DL, Node->getVTList(), Operands);
1834 
1835     CurDAG->setNodeMemRefs(Load, {Ld->getMemOperand()});
1836 
1837     ReplaceNode(Node, Load);
1838     return;
1839   }
1840   }
1841 
1842   // Select the default instruction.
1843   SelectCode(Node);
1844 }
1845 
1846 bool RISCVDAGToDAGISel::SelectInlineAsmMemoryOperand(
1847     const SDValue &Op, unsigned ConstraintID, std::vector<SDValue> &OutOps) {
1848   switch (ConstraintID) {
1849   case InlineAsm::Constraint_m:
1850     // We just support simple memory operands that have a single address
1851     // operand and need no special handling.
1852     OutOps.push_back(Op);
1853     return false;
1854   case InlineAsm::Constraint_A:
1855     OutOps.push_back(Op);
1856     return false;
1857   default:
1858     break;
1859   }
1860 
1861   return true;
1862 }
1863 
1864 bool RISCVDAGToDAGISel::SelectAddrFI(SDValue Addr, SDValue &Base) {
1865   if (auto *FIN = dyn_cast<FrameIndexSDNode>(Addr)) {
1866     Base = CurDAG->getTargetFrameIndex(FIN->getIndex(), Subtarget->getXLenVT());
1867     return true;
1868   }
1869   return false;
1870 }
1871 
1872 bool RISCVDAGToDAGISel::SelectBaseAddr(SDValue Addr, SDValue &Base) {
1873   // If this is FrameIndex, select it directly. Otherwise just let it get
1874   // selected to a register independently.
1875   if (auto *FIN = dyn_cast<FrameIndexSDNode>(Addr))
1876     Base = CurDAG->getTargetFrameIndex(FIN->getIndex(), Subtarget->getXLenVT());
1877   else
1878     Base = Addr;
1879   return true;
1880 }
1881 
1882 bool RISCVDAGToDAGISel::selectShiftMask(SDValue N, unsigned ShiftWidth,
1883                                         SDValue &ShAmt) {
1884   // Shift instructions on RISCV only read the lower 5 or 6 bits of the shift
1885   // amount. If there is an AND on the shift amount, we can bypass it if it
1886   // doesn't affect any of those bits.
1887   if (N.getOpcode() == ISD::AND && isa<ConstantSDNode>(N.getOperand(1))) {
1888     const APInt &AndMask = N->getConstantOperandAPInt(1);
1889 
1890     // Since the max shift amount is a power of 2 we can subtract 1 to make a
1891     // mask that covers the bits needed to represent all shift amounts.
1892     assert(isPowerOf2_32(ShiftWidth) && "Unexpected max shift amount!");
1893     APInt ShMask(AndMask.getBitWidth(), ShiftWidth - 1);
1894 
1895     if (ShMask.isSubsetOf(AndMask)) {
1896       ShAmt = N.getOperand(0);
1897       return true;
1898     }
1899 
1900     // SimplifyDemandedBits may have optimized the mask so try restoring any
1901     // bits that are known zero.
1902     KnownBits Known = CurDAG->computeKnownBits(N->getOperand(0));
1903     if (ShMask.isSubsetOf(AndMask | Known.Zero)) {
1904       ShAmt = N.getOperand(0);
1905       return true;
1906     }
1907   } else if (N.getOpcode() == ISD::SUB &&
1908              isa<ConstantSDNode>(N.getOperand(0))) {
1909     uint64_t Imm = N.getConstantOperandVal(0);
1910     // If we are shifting by N-X where N == 0 mod Size, then just shift by -X to
1911     // generate a NEG instead of a SUB of a constant.
1912     if (Imm != 0 && Imm % ShiftWidth == 0) {
1913       SDLoc DL(N);
1914       EVT VT = N.getValueType();
1915       SDValue Zero =
1916           CurDAG->getCopyFromReg(CurDAG->getEntryNode(), DL, RISCV::X0, VT);
1917       unsigned NegOpc = VT == MVT::i64 ? RISCV::SUBW : RISCV::SUB;
1918       MachineSDNode *Neg = CurDAG->getMachineNode(NegOpc, DL, VT, Zero,
1919                                                   N.getOperand(1));
1920       ShAmt = SDValue(Neg, 0);
1921       return true;
1922     }
1923   }
1924 
1925   ShAmt = N;
1926   return true;
1927 }
1928 
1929 bool RISCVDAGToDAGISel::selectSExti32(SDValue N, SDValue &Val) {
1930   if (N.getOpcode() == ISD::SIGN_EXTEND_INREG &&
1931       cast<VTSDNode>(N.getOperand(1))->getVT() == MVT::i32) {
1932     Val = N.getOperand(0);
1933     return true;
1934   }
1935   MVT VT = N.getSimpleValueType();
1936   if (CurDAG->ComputeNumSignBits(N) > (VT.getSizeInBits() - 32)) {
1937     Val = N;
1938     return true;
1939   }
1940 
1941   return false;
1942 }
1943 
1944 bool RISCVDAGToDAGISel::selectZExti32(SDValue N, SDValue &Val) {
1945   if (N.getOpcode() == ISD::AND) {
1946     auto *C = dyn_cast<ConstantSDNode>(N.getOperand(1));
1947     if (C && C->getZExtValue() == UINT64_C(0xFFFFFFFF)) {
1948       Val = N.getOperand(0);
1949       return true;
1950     }
1951   }
1952   MVT VT = N.getSimpleValueType();
1953   APInt Mask = APInt::getHighBitsSet(VT.getSizeInBits(), 32);
1954   if (CurDAG->MaskedValueIsZero(N, Mask)) {
1955     Val = N;
1956     return true;
1957   }
1958 
1959   return false;
1960 }
1961 
1962 // Return true if all users of this SDNode* only consume the lower \p Bits.
1963 // This can be used to form W instructions for add/sub/mul/shl even when the
1964 // root isn't a sext_inreg. This can allow the ADDW/SUBW/MULW/SLLIW to CSE if
1965 // SimplifyDemandedBits has made it so some users see a sext_inreg and some
1966 // don't. The sext_inreg+add/sub/mul/shl will get selected, but still leave
1967 // the add/sub/mul/shl to become non-W instructions. By checking the users we
1968 // may be able to use a W instruction and CSE with the other instruction if
1969 // this has happened. We could try to detect that the CSE opportunity exists
1970 // before doing this, but that would be more complicated.
1971 // TODO: Does this need to look through AND/OR/XOR to their users to find more
1972 // opportunities.
1973 bool RISCVDAGToDAGISel::hasAllNBitUsers(SDNode *Node, unsigned Bits) const {
1974   assert((Node->getOpcode() == ISD::ADD || Node->getOpcode() == ISD::SUB ||
1975           Node->getOpcode() == ISD::MUL || Node->getOpcode() == ISD::SHL ||
1976           Node->getOpcode() == ISD::SRL ||
1977           Node->getOpcode() == ISD::SIGN_EXTEND_INREG ||
1978           Node->getOpcode() == RISCVISD::GREV ||
1979           Node->getOpcode() == RISCVISD::GORC ||
1980           isa<ConstantSDNode>(Node)) &&
1981          "Unexpected opcode");
1982 
1983   for (auto UI = Node->use_begin(), UE = Node->use_end(); UI != UE; ++UI) {
1984     SDNode *User = *UI;
1985     // Users of this node should have already been instruction selected
1986     if (!User->isMachineOpcode())
1987       return false;
1988 
1989     // TODO: Add more opcodes?
1990     switch (User->getMachineOpcode()) {
1991     default:
1992       return false;
1993     case RISCV::ADDW:
1994     case RISCV::ADDIW:
1995     case RISCV::SUBW:
1996     case RISCV::MULW:
1997     case RISCV::SLLW:
1998     case RISCV::SLLIW:
1999     case RISCV::SRAW:
2000     case RISCV::SRAIW:
2001     case RISCV::SRLW:
2002     case RISCV::SRLIW:
2003     case RISCV::DIVW:
2004     case RISCV::DIVUW:
2005     case RISCV::REMW:
2006     case RISCV::REMUW:
2007     case RISCV::ROLW:
2008     case RISCV::RORW:
2009     case RISCV::RORIW:
2010     case RISCV::CLZW:
2011     case RISCV::CTZW:
2012     case RISCV::CPOPW:
2013     case RISCV::SLLI_UW:
2014     case RISCV::FMV_W_X:
2015     case RISCV::FCVT_H_W:
2016     case RISCV::FCVT_H_WU:
2017     case RISCV::FCVT_S_W:
2018     case RISCV::FCVT_S_WU:
2019     case RISCV::FCVT_D_W:
2020     case RISCV::FCVT_D_WU:
2021       if (Bits < 32)
2022         return false;
2023       break;
2024     case RISCV::SLLI:
2025       // SLLI only uses the lower (XLen - ShAmt) bits.
2026       if (Bits < Subtarget->getXLen() - User->getConstantOperandVal(1))
2027         return false;
2028       break;
2029     case RISCV::ANDI:
2030       if (Bits < (64 - countLeadingZeros(User->getConstantOperandVal(1))))
2031         return false;
2032       break;
2033     case RISCV::SEXT_B:
2034       if (Bits < 8)
2035         return false;
2036       break;
2037     case RISCV::SEXT_H:
2038     case RISCV::FMV_H_X:
2039     case RISCV::ZEXT_H_RV32:
2040     case RISCV::ZEXT_H_RV64:
2041       if (Bits < 16)
2042         return false;
2043       break;
2044     case RISCV::ADD_UW:
2045     case RISCV::SH1ADD_UW:
2046     case RISCV::SH2ADD_UW:
2047     case RISCV::SH3ADD_UW:
2048       // The first operand to add.uw/shXadd.uw is implicitly zero extended from
2049       // 32 bits.
2050       if (UI.getOperandNo() != 0 || Bits < 32)
2051         return false;
2052       break;
2053     case RISCV::SB:
2054       if (UI.getOperandNo() != 0 || Bits < 8)
2055         return false;
2056       break;
2057     case RISCV::SH:
2058       if (UI.getOperandNo() != 0 || Bits < 16)
2059         return false;
2060       break;
2061     case RISCV::SW:
2062       if (UI.getOperandNo() != 0 || Bits < 32)
2063         return false;
2064       break;
2065     }
2066   }
2067 
2068   return true;
2069 }
2070 
2071 // Select VL as a 5 bit immediate or a value that will become a register. This
2072 // allows us to choose betwen VSETIVLI or VSETVLI later.
2073 bool RISCVDAGToDAGISel::selectVLOp(SDValue N, SDValue &VL) {
2074   auto *C = dyn_cast<ConstantSDNode>(N);
2075   if (C && isUInt<5>(C->getZExtValue())) {
2076     VL = CurDAG->getTargetConstant(C->getZExtValue(), SDLoc(N),
2077                                    N->getValueType(0));
2078   } else if (C && C->isAllOnesValue()) {
2079     // Treat all ones as VLMax.
2080     VL = CurDAG->getTargetConstant(RISCV::VLMaxSentinel, SDLoc(N),
2081                                    N->getValueType(0));
2082   } else if (isa<RegisterSDNode>(N) &&
2083              cast<RegisterSDNode>(N)->getReg() == RISCV::X0) {
2084     // All our VL operands use an operand that allows GPRNoX0 or an immediate
2085     // as the register class. Convert X0 to a special immediate to pass the
2086     // MachineVerifier. This is recognized specially by the vsetvli insertion
2087     // pass.
2088     VL = CurDAG->getTargetConstant(RISCV::VLMaxSentinel, SDLoc(N),
2089                                    N->getValueType(0));
2090   } else {
2091     VL = N;
2092   }
2093 
2094   return true;
2095 }
2096 
2097 bool RISCVDAGToDAGISel::selectVSplat(SDValue N, SDValue &SplatVal) {
2098   if (N.getOpcode() != RISCVISD::VMV_V_X_VL || !N.getOperand(0).isUndef())
2099     return false;
2100   SplatVal = N.getOperand(1);
2101   return true;
2102 }
2103 
2104 using ValidateFn = bool (*)(int64_t);
2105 
2106 static bool selectVSplatSimmHelper(SDValue N, SDValue &SplatVal,
2107                                    SelectionDAG &DAG,
2108                                    const RISCVSubtarget &Subtarget,
2109                                    ValidateFn ValidateImm) {
2110   if (N.getOpcode() != RISCVISD::VMV_V_X_VL || !N.getOperand(0).isUndef() ||
2111       !isa<ConstantSDNode>(N.getOperand(1)))
2112     return false;
2113 
2114   int64_t SplatImm =
2115       cast<ConstantSDNode>(N.getOperand(1))->getSExtValue();
2116 
2117   // The semantics of RISCVISD::VMV_V_X_VL is that when the operand
2118   // type is wider than the resulting vector element type: an implicit
2119   // truncation first takes place. Therefore, perform a manual
2120   // truncation/sign-extension in order to ignore any truncated bits and catch
2121   // any zero-extended immediate.
2122   // For example, we wish to match (i8 -1) -> (XLenVT 255) as a simm5 by first
2123   // sign-extending to (XLenVT -1).
2124   MVT XLenVT = Subtarget.getXLenVT();
2125   assert(XLenVT == N.getOperand(1).getSimpleValueType() &&
2126          "Unexpected splat operand type");
2127   MVT EltVT = N.getSimpleValueType().getVectorElementType();
2128   if (EltVT.bitsLT(XLenVT))
2129     SplatImm = SignExtend64(SplatImm, EltVT.getSizeInBits());
2130 
2131   if (!ValidateImm(SplatImm))
2132     return false;
2133 
2134   SplatVal = DAG.getTargetConstant(SplatImm, SDLoc(N), XLenVT);
2135   return true;
2136 }
2137 
2138 bool RISCVDAGToDAGISel::selectVSplatSimm5(SDValue N, SDValue &SplatVal) {
2139   return selectVSplatSimmHelper(N, SplatVal, *CurDAG, *Subtarget,
2140                                 [](int64_t Imm) { return isInt<5>(Imm); });
2141 }
2142 
2143 bool RISCVDAGToDAGISel::selectVSplatSimm5Plus1(SDValue N, SDValue &SplatVal) {
2144   return selectVSplatSimmHelper(
2145       N, SplatVal, *CurDAG, *Subtarget,
2146       [](int64_t Imm) { return (isInt<5>(Imm) && Imm != -16) || Imm == 16; });
2147 }
2148 
2149 bool RISCVDAGToDAGISel::selectVSplatSimm5Plus1NonZero(SDValue N,
2150                                                       SDValue &SplatVal) {
2151   return selectVSplatSimmHelper(
2152       N, SplatVal, *CurDAG, *Subtarget, [](int64_t Imm) {
2153         return Imm != 0 && ((isInt<5>(Imm) && Imm != -16) || Imm == 16);
2154       });
2155 }
2156 
2157 bool RISCVDAGToDAGISel::selectVSplatUimm5(SDValue N, SDValue &SplatVal) {
2158   if (N.getOpcode() != RISCVISD::VMV_V_X_VL || !N.getOperand(0).isUndef() ||
2159       !isa<ConstantSDNode>(N.getOperand(1)))
2160     return false;
2161 
2162   int64_t SplatImm =
2163       cast<ConstantSDNode>(N.getOperand(1))->getSExtValue();
2164 
2165   if (!isUInt<5>(SplatImm))
2166     return false;
2167 
2168   SplatVal =
2169       CurDAG->getTargetConstant(SplatImm, SDLoc(N), Subtarget->getXLenVT());
2170 
2171   return true;
2172 }
2173 
2174 bool RISCVDAGToDAGISel::selectRVVSimm5(SDValue N, unsigned Width,
2175                                        SDValue &Imm) {
2176   if (auto *C = dyn_cast<ConstantSDNode>(N)) {
2177     int64_t ImmVal = SignExtend64(C->getSExtValue(), Width);
2178 
2179     if (!isInt<5>(ImmVal))
2180       return false;
2181 
2182     Imm = CurDAG->getTargetConstant(ImmVal, SDLoc(N), Subtarget->getXLenVT());
2183     return true;
2184   }
2185 
2186   return false;
2187 }
2188 
2189 // Merge an ADDI into the offset of a load/store instruction where possible.
2190 // (load (addi base, off1), off2) -> (load base, off1+off2)
2191 // (store val, (addi base, off1), off2) -> (store val, base, off1+off2)
2192 // (load (add base, (addi src, off1)), off2)
2193 //    -> (load (add base, src), off1+off2)
2194 // (store val, (add base, (addi src, off1)), off2)
2195 //    -> (store val, (add base, src), off1+off2)
2196 // This is possible when off1+off2 fits a 12-bit immediate.
2197 bool RISCVDAGToDAGISel::doPeepholeLoadStoreADDI(SDNode *N) {
2198   unsigned OffsetOpIdx, BaseOpIdx;
2199   if (!hasMemOffset(N, BaseOpIdx, OffsetOpIdx))
2200     return false;
2201 
2202   if (!isa<ConstantSDNode>(N->getOperand(OffsetOpIdx)))
2203     return false;
2204 
2205   SDValue Base = N->getOperand(BaseOpIdx);
2206 
2207   if (!Base.isMachineOpcode())
2208     return false;
2209 
2210   // If the base is an ADDI, we can merge it in to the load/store.
2211   if (Base.getMachineOpcode() != RISCV::ADDI)
2212     return false;
2213 
2214   SDValue ImmOperand = Base.getOperand(1);
2215   uint64_t Offset2 = N->getConstantOperandVal(OffsetOpIdx);
2216 
2217   if (auto *Const = dyn_cast<ConstantSDNode>(ImmOperand)) {
2218     int64_t Offset1 = Const->getSExtValue();
2219     int64_t CombinedOffset = Offset1 + Offset2;
2220     if (!isInt<12>(CombinedOffset))
2221       return false;
2222     ImmOperand = CurDAG->getTargetConstant(CombinedOffset, SDLoc(ImmOperand),
2223                                            ImmOperand.getValueType());
2224   } else if (auto *GA = dyn_cast<GlobalAddressSDNode>(ImmOperand)) {
2225     // If the off1 in (addi base, off1) is a global variable's address (its
2226     // low part, really), then we can rely on the alignment of that variable
2227     // to provide a margin of safety before off1 can overflow the 12 bits.
2228     // Check if off2 falls within that margin; if so off1+off2 can't overflow.
2229     const DataLayout &DL = CurDAG->getDataLayout();
2230     Align Alignment = GA->getGlobal()->getPointerAlignment(DL);
2231     if (Offset2 != 0 && Alignment <= Offset2)
2232       return false;
2233     int64_t Offset1 = GA->getOffset();
2234     int64_t CombinedOffset = Offset1 + Offset2;
2235     ImmOperand = CurDAG->getTargetGlobalAddress(
2236         GA->getGlobal(), SDLoc(ImmOperand), ImmOperand.getValueType(),
2237         CombinedOffset, GA->getTargetFlags());
2238   } else if (auto *CP = dyn_cast<ConstantPoolSDNode>(ImmOperand)) {
2239     // Ditto.
2240     Align Alignment = CP->getAlign();
2241     if (Offset2 != 0 && Alignment <= Offset2)
2242       return false;
2243     int64_t Offset1 = CP->getOffset();
2244     int64_t CombinedOffset = Offset1 + Offset2;
2245     ImmOperand = CurDAG->getTargetConstantPool(
2246         CP->getConstVal(), ImmOperand.getValueType(), CP->getAlign(),
2247         CombinedOffset, CP->getTargetFlags());
2248   } else {
2249     return false;
2250   }
2251 
2252   LLVM_DEBUG(dbgs() << "Folding add-immediate into mem-op:\nBase:    ");
2253   LLVM_DEBUG(Base->dump(CurDAG));
2254   LLVM_DEBUG(dbgs() << "\nN: ");
2255   LLVM_DEBUG(N->dump(CurDAG));
2256   LLVM_DEBUG(dbgs() << "\n");
2257 
2258   // Modify the offset operand of the load/store.
2259   if (BaseOpIdx == 0) { // Load
2260     N = CurDAG->UpdateNodeOperands(N, Base.getOperand(0), ImmOperand,
2261                                    N->getOperand(2));
2262   } else { // Store
2263     N = CurDAG->UpdateNodeOperands(N, N->getOperand(0), Base.getOperand(0),
2264                                    ImmOperand, N->getOperand(3));
2265   }
2266 
2267   return true;
2268 }
2269 
2270 // Try to remove sext.w if the input is a W instruction or can be made into
2271 // a W instruction cheaply.
2272 bool RISCVDAGToDAGISel::doPeepholeSExtW(SDNode *N) {
2273   // Look for the sext.w pattern, addiw rd, rs1, 0.
2274   if (N->getMachineOpcode() != RISCV::ADDIW ||
2275       !isNullConstant(N->getOperand(1)))
2276     return false;
2277 
2278   SDValue N0 = N->getOperand(0);
2279   if (!N0.isMachineOpcode())
2280     return false;
2281 
2282   switch (N0.getMachineOpcode()) {
2283   default:
2284     break;
2285   case RISCV::ADD:
2286   case RISCV::ADDI:
2287   case RISCV::SUB:
2288   case RISCV::MUL:
2289   case RISCV::SLLI: {
2290     // Convert sext.w+add/sub/mul to their W instructions. This will create
2291     // a new independent instruction. This improves latency.
2292     unsigned Opc;
2293     switch (N0.getMachineOpcode()) {
2294     default:
2295       llvm_unreachable("Unexpected opcode!");
2296     case RISCV::ADD:  Opc = RISCV::ADDW;  break;
2297     case RISCV::ADDI: Opc = RISCV::ADDIW; break;
2298     case RISCV::SUB:  Opc = RISCV::SUBW;  break;
2299     case RISCV::MUL:  Opc = RISCV::MULW;  break;
2300     case RISCV::SLLI: Opc = RISCV::SLLIW; break;
2301     }
2302 
2303     SDValue N00 = N0.getOperand(0);
2304     SDValue N01 = N0.getOperand(1);
2305 
2306     // Shift amount needs to be uimm5.
2307     if (N0.getMachineOpcode() == RISCV::SLLI &&
2308         !isUInt<5>(cast<ConstantSDNode>(N01)->getSExtValue()))
2309       break;
2310 
2311     SDNode *Result =
2312         CurDAG->getMachineNode(Opc, SDLoc(N), N->getValueType(0),
2313                                N00, N01);
2314     ReplaceUses(N, Result);
2315     return true;
2316   }
2317   case RISCV::ADDW:
2318   case RISCV::ADDIW:
2319   case RISCV::SUBW:
2320   case RISCV::MULW:
2321   case RISCV::SLLIW:
2322   case RISCV::GREVIW:
2323   case RISCV::GORCIW:
2324     // Result is already sign extended just remove the sext.w.
2325     // NOTE: We only handle the nodes that are selected with hasAllWUsers.
2326     ReplaceUses(N, N0.getNode());
2327     return true;
2328   }
2329 
2330   return false;
2331 }
2332 
2333 // Optimize masked RVV pseudo instructions with a known all-ones mask to their
2334 // corresponding "unmasked" pseudo versions. The mask we're interested in will
2335 // take the form of a V0 physical register operand, with a glued
2336 // register-setting instruction.
2337 bool RISCVDAGToDAGISel::doPeepholeMaskedRVV(SDNode *N) {
2338   const RISCV::RISCVMaskedPseudoInfo *I =
2339       RISCV::getMaskedPseudoInfo(N->getMachineOpcode());
2340   if (!I)
2341     return false;
2342 
2343   unsigned MaskOpIdx = I->MaskOpIdx;
2344 
2345   // Check that we're using V0 as a mask register.
2346   if (!isa<RegisterSDNode>(N->getOperand(MaskOpIdx)) ||
2347       cast<RegisterSDNode>(N->getOperand(MaskOpIdx))->getReg() != RISCV::V0)
2348     return false;
2349 
2350   // The glued user defines V0.
2351   const auto *Glued = N->getGluedNode();
2352 
2353   if (!Glued || Glued->getOpcode() != ISD::CopyToReg)
2354     return false;
2355 
2356   // Check that we're defining V0 as a mask register.
2357   if (!isa<RegisterSDNode>(Glued->getOperand(1)) ||
2358       cast<RegisterSDNode>(Glued->getOperand(1))->getReg() != RISCV::V0)
2359     return false;
2360 
2361   // Check the instruction defining V0; it needs to be a VMSET pseudo.
2362   SDValue MaskSetter = Glued->getOperand(2);
2363 
2364   const auto IsVMSet = [](unsigned Opc) {
2365     return Opc == RISCV::PseudoVMSET_M_B1 || Opc == RISCV::PseudoVMSET_M_B16 ||
2366            Opc == RISCV::PseudoVMSET_M_B2 || Opc == RISCV::PseudoVMSET_M_B32 ||
2367            Opc == RISCV::PseudoVMSET_M_B4 || Opc == RISCV::PseudoVMSET_M_B64 ||
2368            Opc == RISCV::PseudoVMSET_M_B8;
2369   };
2370 
2371   // TODO: Check that the VMSET is the expected bitwidth? The pseudo has
2372   // undefined behaviour if it's the wrong bitwidth, so we could choose to
2373   // assume that it's all-ones? Same applies to its VL.
2374   if (!MaskSetter->isMachineOpcode() || !IsVMSet(MaskSetter.getMachineOpcode()))
2375     return false;
2376 
2377   // Retrieve the tail policy operand index, if any.
2378   Optional<unsigned> TailPolicyOpIdx;
2379   const RISCVInstrInfo *TII = static_cast<const RISCVInstrInfo *>(
2380       CurDAG->getSubtarget().getInstrInfo());
2381 
2382   const MCInstrDesc &MaskedMCID = TII->get(N->getMachineOpcode());
2383 
2384   bool IsTA = true;
2385   if (RISCVII::hasVecPolicyOp(MaskedMCID.TSFlags)) {
2386     // The last operand of the pseudo is the policy op, but we might have a
2387     // Glue operand last. We might also have a chain.
2388     TailPolicyOpIdx = N->getNumOperands() - 1;
2389     if (N->getOperand(*TailPolicyOpIdx).getValueType() == MVT::Glue)
2390       (*TailPolicyOpIdx)--;
2391     if (N->getOperand(*TailPolicyOpIdx).getValueType() == MVT::Other)
2392       (*TailPolicyOpIdx)--;
2393 
2394     if (!(N->getConstantOperandVal(*TailPolicyOpIdx) &
2395           RISCVII::TAIL_AGNOSTIC)) {
2396       // Keep the true-masked instruction when there is no unmasked TU
2397       // instruction
2398       if (I->UnmaskedTUPseudo == I->MaskedPseudo && !N->getOperand(0).isUndef())
2399         return false;
2400       // We can't use TA if the tie-operand is not IMPLICIT_DEF
2401       if (!N->getOperand(0).isUndef())
2402         IsTA = false;
2403     }
2404   }
2405 
2406   if (IsTA) {
2407     uint64_t TSFlags = TII->get(I->UnmaskedPseudo).TSFlags;
2408 
2409     // Check that we're dropping the merge operand, the mask operand, and any
2410     // policy operand when we transform to this unmasked pseudo.
2411     assert(!RISCVII::hasMergeOp(TSFlags) && RISCVII::hasDummyMaskOp(TSFlags) &&
2412            !RISCVII::hasVecPolicyOp(TSFlags) &&
2413            "Unexpected pseudo to transform to");
2414     (void)TSFlags;
2415   } else {
2416     uint64_t TSFlags = TII->get(I->UnmaskedTUPseudo).TSFlags;
2417 
2418     // Check that we're dropping the mask operand, and any policy operand
2419     // when we transform to this unmasked tu pseudo.
2420     assert(RISCVII::hasMergeOp(TSFlags) && RISCVII::hasDummyMaskOp(TSFlags) &&
2421            !RISCVII::hasVecPolicyOp(TSFlags) &&
2422            "Unexpected pseudo to transform to");
2423     (void)TSFlags;
2424   }
2425 
2426   unsigned Opc = IsTA ? I->UnmaskedPseudo : I->UnmaskedTUPseudo;
2427   SmallVector<SDValue, 8> Ops;
2428   // Skip the merge operand at index 0 if IsTA
2429   for (unsigned I = IsTA, E = N->getNumOperands(); I != E; I++) {
2430     // Skip the mask, the policy, and the Glue.
2431     SDValue Op = N->getOperand(I);
2432     if (I == MaskOpIdx || I == TailPolicyOpIdx ||
2433         Op.getValueType() == MVT::Glue)
2434       continue;
2435     Ops.push_back(Op);
2436   }
2437 
2438   // Transitively apply any node glued to our new node.
2439   if (auto *TGlued = Glued->getGluedNode())
2440     Ops.push_back(SDValue(TGlued, TGlued->getNumValues() - 1));
2441 
2442   SDNode *Result = CurDAG->getMachineNode(Opc, SDLoc(N), N->getVTList(), Ops);
2443   ReplaceUses(N, Result);
2444 
2445   return true;
2446 }
2447 
2448 // This pass converts a legalized DAG into a RISCV-specific DAG, ready
2449 // for instruction scheduling.
2450 FunctionPass *llvm::createRISCVISelDag(RISCVTargetMachine &TM,
2451                                        CodeGenOpt::Level OptLevel) {
2452   return new RISCVDAGToDAGISel(TM, OptLevel);
2453 }
2454