xref: /freebsd/contrib/llvm-project/llvm/lib/Target/RISCV/RISCVInstrInfo.cpp (revision 924226fba12cc9a228c73b956e1b7fa24c60b055)
1 //===-- RISCVInstrInfo.cpp - RISCV Instruction Information ------*- C++ -*-===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This file contains the RISCV implementation of the TargetInstrInfo class.
10 //
11 //===----------------------------------------------------------------------===//
12 
13 #include "RISCVInstrInfo.h"
14 #include "MCTargetDesc/RISCVMatInt.h"
15 #include "RISCV.h"
16 #include "RISCVMachineFunctionInfo.h"
17 #include "RISCVSubtarget.h"
18 #include "RISCVTargetMachine.h"
19 #include "llvm/ADT/STLExtras.h"
20 #include "llvm/ADT/SmallVector.h"
21 #include "llvm/Analysis/MemoryLocation.h"
22 #include "llvm/CodeGen/LiveIntervals.h"
23 #include "llvm/CodeGen/LiveVariables.h"
24 #include "llvm/CodeGen/MachineFunctionPass.h"
25 #include "llvm/CodeGen/MachineInstrBuilder.h"
26 #include "llvm/CodeGen/MachineRegisterInfo.h"
27 #include "llvm/CodeGen/RegisterScavenging.h"
28 #include "llvm/MC/MCInstBuilder.h"
29 #include "llvm/MC/TargetRegistry.h"
30 #include "llvm/Support/ErrorHandling.h"
31 
32 using namespace llvm;
33 
34 #define GEN_CHECK_COMPRESS_INSTR
35 #include "RISCVGenCompressInstEmitter.inc"
36 
37 #define GET_INSTRINFO_CTOR_DTOR
38 #define GET_INSTRINFO_NAMED_OPS
39 #include "RISCVGenInstrInfo.inc"
40 
41 static cl::opt<bool> PreferWholeRegisterMove(
42     "riscv-prefer-whole-register-move", cl::init(false), cl::Hidden,
43     cl::desc("Prefer whole register move for vector registers."));
44 
45 namespace llvm {
46 namespace RISCVVPseudosTable {
47 
48 using namespace RISCV;
49 
50 #define GET_RISCVVPseudosTable_IMPL
51 #include "RISCVGenSearchableTables.inc"
52 
53 } // namespace RISCVVPseudosTable
54 } // namespace llvm
55 
56 RISCVInstrInfo::RISCVInstrInfo(RISCVSubtarget &STI)
57     : RISCVGenInstrInfo(RISCV::ADJCALLSTACKDOWN, RISCV::ADJCALLSTACKUP),
58       STI(STI) {}
59 
60 MCInst RISCVInstrInfo::getNop() const {
61   if (STI.getFeatureBits()[RISCV::FeatureStdExtC])
62     return MCInstBuilder(RISCV::C_NOP);
63   return MCInstBuilder(RISCV::ADDI)
64       .addReg(RISCV::X0)
65       .addReg(RISCV::X0)
66       .addImm(0);
67 }
68 
69 unsigned RISCVInstrInfo::isLoadFromStackSlot(const MachineInstr &MI,
70                                              int &FrameIndex) const {
71   switch (MI.getOpcode()) {
72   default:
73     return 0;
74   case RISCV::LB:
75   case RISCV::LBU:
76   case RISCV::LH:
77   case RISCV::LHU:
78   case RISCV::FLH:
79   case RISCV::LW:
80   case RISCV::FLW:
81   case RISCV::LWU:
82   case RISCV::LD:
83   case RISCV::FLD:
84     break;
85   }
86 
87   if (MI.getOperand(1).isFI() && MI.getOperand(2).isImm() &&
88       MI.getOperand(2).getImm() == 0) {
89     FrameIndex = MI.getOperand(1).getIndex();
90     return MI.getOperand(0).getReg();
91   }
92 
93   return 0;
94 }
95 
96 unsigned RISCVInstrInfo::isStoreToStackSlot(const MachineInstr &MI,
97                                             int &FrameIndex) const {
98   switch (MI.getOpcode()) {
99   default:
100     return 0;
101   case RISCV::SB:
102   case RISCV::SH:
103   case RISCV::SW:
104   case RISCV::FSH:
105   case RISCV::FSW:
106   case RISCV::SD:
107   case RISCV::FSD:
108     break;
109   }
110 
111   if (MI.getOperand(1).isFI() && MI.getOperand(2).isImm() &&
112       MI.getOperand(2).getImm() == 0) {
113     FrameIndex = MI.getOperand(1).getIndex();
114     return MI.getOperand(0).getReg();
115   }
116 
117   return 0;
118 }
119 
120 static bool forwardCopyWillClobberTuple(unsigned DstReg, unsigned SrcReg,
121                                         unsigned NumRegs) {
122   return DstReg > SrcReg && (DstReg - SrcReg) < NumRegs;
123 }
124 
125 static bool isConvertibleToVMV_V_V(const RISCVSubtarget &STI,
126                                    const MachineBasicBlock &MBB,
127                                    MachineBasicBlock::const_iterator MBBI,
128                                    MachineBasicBlock::const_iterator &DefMBBI,
129                                    RISCVII::VLMUL LMul) {
130   if (PreferWholeRegisterMove)
131     return false;
132 
133   assert(MBBI->getOpcode() == TargetOpcode::COPY &&
134          "Unexpected COPY instruction.");
135   Register SrcReg = MBBI->getOperand(1).getReg();
136   const TargetRegisterInfo *TRI = STI.getRegisterInfo();
137 
138   bool FoundDef = false;
139   bool FirstVSetVLI = false;
140   unsigned FirstSEW = 0;
141   while (MBBI != MBB.begin()) {
142     --MBBI;
143     if (MBBI->isMetaInstruction())
144       continue;
145 
146     if (MBBI->getOpcode() == RISCV::PseudoVSETVLI ||
147         MBBI->getOpcode() == RISCV::PseudoVSETVLIX0 ||
148         MBBI->getOpcode() == RISCV::PseudoVSETIVLI) {
149       // There is a vsetvli between COPY and source define instruction.
150       // vy = def_vop ...  (producing instruction)
151       // ...
152       // vsetvli
153       // ...
154       // vx = COPY vy
155       if (!FoundDef) {
156         if (!FirstVSetVLI) {
157           FirstVSetVLI = true;
158           unsigned FirstVType = MBBI->getOperand(2).getImm();
159           RISCVII::VLMUL FirstLMul = RISCVVType::getVLMUL(FirstVType);
160           FirstSEW = RISCVVType::getSEW(FirstVType);
161           // The first encountered vsetvli must have the same lmul as the
162           // register class of COPY.
163           if (FirstLMul != LMul)
164             return false;
165         }
166         // Only permit `vsetvli x0, x0, vtype` between COPY and the source
167         // define instruction.
168         if (MBBI->getOperand(0).getReg() != RISCV::X0)
169           return false;
170         if (MBBI->getOperand(1).isImm())
171           return false;
172         if (MBBI->getOperand(1).getReg() != RISCV::X0)
173           return false;
174         continue;
175       }
176 
177       // MBBI is the first vsetvli before the producing instruction.
178       unsigned VType = MBBI->getOperand(2).getImm();
179       // If there is a vsetvli between COPY and the producing instruction.
180       if (FirstVSetVLI) {
181         // If SEW is different, return false.
182         if (RISCVVType::getSEW(VType) != FirstSEW)
183           return false;
184       }
185 
186       // If the vsetvli is tail undisturbed, keep the whole register move.
187       if (!RISCVVType::isTailAgnostic(VType))
188         return false;
189 
190       // The checking is conservative. We only have register classes for
191       // LMUL = 1/2/4/8. We should be able to convert vmv1r.v to vmv.v.v
192       // for fractional LMUL operations. However, we could not use the vsetvli
193       // lmul for widening operations. The result of widening operation is
194       // 2 x LMUL.
195       return LMul == RISCVVType::getVLMUL(VType);
196     } else if (MBBI->isInlineAsm() || MBBI->isCall()) {
197       return false;
198     } else if (MBBI->getNumDefs()) {
199       // Check all the instructions which will change VL.
200       // For example, vleff has implicit def VL.
201       if (MBBI->modifiesRegister(RISCV::VL))
202         return false;
203 
204       // Only converting whole register copies to vmv.v.v when the defining
205       // value appears in the explicit operands.
206       for (const MachineOperand &MO : MBBI->explicit_operands()) {
207         if (!MO.isReg() || !MO.isDef())
208           continue;
209         if (!FoundDef && TRI->isSubRegisterEq(MO.getReg(), SrcReg)) {
210           // We only permit the source of COPY has the same LMUL as the defined
211           // operand.
212           // There are cases we need to keep the whole register copy if the LMUL
213           // is different.
214           // For example,
215           // $x0 = PseudoVSETIVLI 4, 73   // vsetivli zero, 4, e16,m2,ta,m
216           // $v28m4 = PseudoVWADD_VV_M2 $v26m2, $v8m2
217           // # The COPY may be created by vlmul_trunc intrinsic.
218           // $v26m2 = COPY renamable $v28m2, implicit killed $v28m4
219           //
220           // After widening, the valid value will be 4 x e32 elements. If we
221           // convert the COPY to vmv.v.v, it will only copy 4 x e16 elements.
222           // FIXME: The COPY of subregister of Zvlsseg register will not be able
223           // to convert to vmv.v.[v|i] under the constraint.
224           if (MO.getReg() != SrcReg)
225             return false;
226 
227           // In widening reduction instructions with LMUL_1 input vector case,
228           // only checking the LMUL is insufficient due to reduction result is
229           // always LMUL_1.
230           // For example,
231           // $x11 = PseudoVSETIVLI 1, 64 // vsetivli a1, 1, e8, m1, ta, mu
232           // $v8m1 = PseudoVWREDSUM_VS_M1 $v26, $v27
233           // $v26 = COPY killed renamable $v8
234           // After widening, The valid value will be 1 x e16 elements. If we
235           // convert the COPY to vmv.v.v, it will only copy 1 x e8 elements.
236           uint64_t TSFlags = MBBI->getDesc().TSFlags;
237           if (RISCVII::isRVVWideningReduction(TSFlags))
238             return false;
239 
240           // Found the definition.
241           FoundDef = true;
242           DefMBBI = MBBI;
243           // If the producing instruction does not depend on vsetvli, do not
244           // convert COPY to vmv.v.v. For example, VL1R_V or PseudoVRELOAD.
245           if (!RISCVII::hasSEWOp(TSFlags))
246             return false;
247           break;
248         }
249       }
250     }
251   }
252 
253   return false;
254 }
255 
256 void RISCVInstrInfo::copyPhysReg(MachineBasicBlock &MBB,
257                                  MachineBasicBlock::iterator MBBI,
258                                  const DebugLoc &DL, MCRegister DstReg,
259                                  MCRegister SrcReg, bool KillSrc) const {
260   if (RISCV::GPRRegClass.contains(DstReg, SrcReg)) {
261     BuildMI(MBB, MBBI, DL, get(RISCV::ADDI), DstReg)
262         .addReg(SrcReg, getKillRegState(KillSrc))
263         .addImm(0);
264     return;
265   }
266 
267   // FPR->FPR copies and VR->VR copies.
268   unsigned Opc;
269   bool IsScalableVector = true;
270   unsigned NF = 1;
271   RISCVII::VLMUL LMul = RISCVII::LMUL_1;
272   unsigned SubRegIdx = RISCV::sub_vrm1_0;
273   if (RISCV::FPR16RegClass.contains(DstReg, SrcReg)) {
274     Opc = RISCV::FSGNJ_H;
275     IsScalableVector = false;
276   } else if (RISCV::FPR32RegClass.contains(DstReg, SrcReg)) {
277     Opc = RISCV::FSGNJ_S;
278     IsScalableVector = false;
279   } else if (RISCV::FPR64RegClass.contains(DstReg, SrcReg)) {
280     Opc = RISCV::FSGNJ_D;
281     IsScalableVector = false;
282   } else if (RISCV::VRRegClass.contains(DstReg, SrcReg)) {
283     Opc = RISCV::PseudoVMV1R_V;
284     LMul = RISCVII::LMUL_1;
285   } else if (RISCV::VRM2RegClass.contains(DstReg, SrcReg)) {
286     Opc = RISCV::PseudoVMV2R_V;
287     LMul = RISCVII::LMUL_2;
288   } else if (RISCV::VRM4RegClass.contains(DstReg, SrcReg)) {
289     Opc = RISCV::PseudoVMV4R_V;
290     LMul = RISCVII::LMUL_4;
291   } else if (RISCV::VRM8RegClass.contains(DstReg, SrcReg)) {
292     Opc = RISCV::PseudoVMV8R_V;
293     LMul = RISCVII::LMUL_8;
294   } else if (RISCV::VRN2M1RegClass.contains(DstReg, SrcReg)) {
295     Opc = RISCV::PseudoVMV1R_V;
296     SubRegIdx = RISCV::sub_vrm1_0;
297     NF = 2;
298     LMul = RISCVII::LMUL_1;
299   } else if (RISCV::VRN2M2RegClass.contains(DstReg, SrcReg)) {
300     Opc = RISCV::PseudoVMV2R_V;
301     SubRegIdx = RISCV::sub_vrm2_0;
302     NF = 2;
303     LMul = RISCVII::LMUL_2;
304   } else if (RISCV::VRN2M4RegClass.contains(DstReg, SrcReg)) {
305     Opc = RISCV::PseudoVMV4R_V;
306     SubRegIdx = RISCV::sub_vrm4_0;
307     NF = 2;
308     LMul = RISCVII::LMUL_4;
309   } else if (RISCV::VRN3M1RegClass.contains(DstReg, SrcReg)) {
310     Opc = RISCV::PseudoVMV1R_V;
311     SubRegIdx = RISCV::sub_vrm1_0;
312     NF = 3;
313     LMul = RISCVII::LMUL_1;
314   } else if (RISCV::VRN3M2RegClass.contains(DstReg, SrcReg)) {
315     Opc = RISCV::PseudoVMV2R_V;
316     SubRegIdx = RISCV::sub_vrm2_0;
317     NF = 3;
318     LMul = RISCVII::LMUL_2;
319   } else if (RISCV::VRN4M1RegClass.contains(DstReg, SrcReg)) {
320     Opc = RISCV::PseudoVMV1R_V;
321     SubRegIdx = RISCV::sub_vrm1_0;
322     NF = 4;
323     LMul = RISCVII::LMUL_1;
324   } else if (RISCV::VRN4M2RegClass.contains(DstReg, SrcReg)) {
325     Opc = RISCV::PseudoVMV2R_V;
326     SubRegIdx = RISCV::sub_vrm2_0;
327     NF = 4;
328     LMul = RISCVII::LMUL_2;
329   } else if (RISCV::VRN5M1RegClass.contains(DstReg, SrcReg)) {
330     Opc = RISCV::PseudoVMV1R_V;
331     SubRegIdx = RISCV::sub_vrm1_0;
332     NF = 5;
333     LMul = RISCVII::LMUL_1;
334   } else if (RISCV::VRN6M1RegClass.contains(DstReg, SrcReg)) {
335     Opc = RISCV::PseudoVMV1R_V;
336     SubRegIdx = RISCV::sub_vrm1_0;
337     NF = 6;
338     LMul = RISCVII::LMUL_1;
339   } else if (RISCV::VRN7M1RegClass.contains(DstReg, SrcReg)) {
340     Opc = RISCV::PseudoVMV1R_V;
341     SubRegIdx = RISCV::sub_vrm1_0;
342     NF = 7;
343     LMul = RISCVII::LMUL_1;
344   } else if (RISCV::VRN8M1RegClass.contains(DstReg, SrcReg)) {
345     Opc = RISCV::PseudoVMV1R_V;
346     SubRegIdx = RISCV::sub_vrm1_0;
347     NF = 8;
348     LMul = RISCVII::LMUL_1;
349   } else {
350     llvm_unreachable("Impossible reg-to-reg copy");
351   }
352 
353   if (IsScalableVector) {
354     bool UseVMV_V_V = false;
355     MachineBasicBlock::const_iterator DefMBBI;
356     unsigned DefExplicitOpNum;
357     unsigned VIOpc;
358     if (isConvertibleToVMV_V_V(STI, MBB, MBBI, DefMBBI, LMul)) {
359       UseVMV_V_V = true;
360       DefExplicitOpNum = DefMBBI->getNumExplicitOperands();
361       // We only need to handle LMUL = 1/2/4/8 here because we only define
362       // vector register classes for LMUL = 1/2/4/8.
363       switch (LMul) {
364       default:
365         llvm_unreachable("Impossible LMUL for vector register copy.");
366       case RISCVII::LMUL_1:
367         Opc = RISCV::PseudoVMV_V_V_M1;
368         VIOpc = RISCV::PseudoVMV_V_I_M1;
369         break;
370       case RISCVII::LMUL_2:
371         Opc = RISCV::PseudoVMV_V_V_M2;
372         VIOpc = RISCV::PseudoVMV_V_I_M2;
373         break;
374       case RISCVII::LMUL_4:
375         Opc = RISCV::PseudoVMV_V_V_M4;
376         VIOpc = RISCV::PseudoVMV_V_I_M4;
377         break;
378       case RISCVII::LMUL_8:
379         Opc = RISCV::PseudoVMV_V_V_M8;
380         VIOpc = RISCV::PseudoVMV_V_I_M8;
381         break;
382       }
383     }
384 
385     bool UseVMV_V_I = false;
386     if (UseVMV_V_V && (DefMBBI->getOpcode() == VIOpc)) {
387       UseVMV_V_I = true;
388       Opc = VIOpc;
389     }
390 
391     if (NF == 1) {
392       auto MIB = BuildMI(MBB, MBBI, DL, get(Opc), DstReg);
393       if (UseVMV_V_I)
394         MIB = MIB.add(DefMBBI->getOperand(1));
395       else
396         MIB = MIB.addReg(SrcReg, getKillRegState(KillSrc));
397       if (UseVMV_V_V) {
398         // The last two arguments of vector instructions are
399         // AVL, SEW. We also need to append the implicit-use vl and vtype.
400         MIB.add(DefMBBI->getOperand(DefExplicitOpNum - 2)); // AVL
401         MIB.add(DefMBBI->getOperand(DefExplicitOpNum - 1)); // SEW
402         MIB.addReg(RISCV::VL, RegState::Implicit);
403         MIB.addReg(RISCV::VTYPE, RegState::Implicit);
404       }
405     } else {
406       const TargetRegisterInfo *TRI = STI.getRegisterInfo();
407 
408       int I = 0, End = NF, Incr = 1;
409       unsigned SrcEncoding = TRI->getEncodingValue(SrcReg);
410       unsigned DstEncoding = TRI->getEncodingValue(DstReg);
411       unsigned LMulVal;
412       bool Fractional;
413       std::tie(LMulVal, Fractional) = RISCVVType::decodeVLMUL(LMul);
414       assert(!Fractional && "It is impossible be fractional lmul here.");
415       if (forwardCopyWillClobberTuple(DstEncoding, SrcEncoding, NF * LMulVal)) {
416         I = NF - 1;
417         End = -1;
418         Incr = -1;
419       }
420 
421       for (; I != End; I += Incr) {
422         auto MIB = BuildMI(MBB, MBBI, DL, get(Opc),
423                            TRI->getSubReg(DstReg, SubRegIdx + I));
424         if (UseVMV_V_I)
425           MIB = MIB.add(DefMBBI->getOperand(1));
426         else
427           MIB = MIB.addReg(TRI->getSubReg(SrcReg, SubRegIdx + I),
428                            getKillRegState(KillSrc));
429         if (UseVMV_V_V) {
430           MIB.add(DefMBBI->getOperand(DefExplicitOpNum - 2)); // AVL
431           MIB.add(DefMBBI->getOperand(DefExplicitOpNum - 1)); // SEW
432           MIB.addReg(RISCV::VL, RegState::Implicit);
433           MIB.addReg(RISCV::VTYPE, RegState::Implicit);
434         }
435       }
436     }
437   } else {
438     BuildMI(MBB, MBBI, DL, get(Opc), DstReg)
439         .addReg(SrcReg, getKillRegState(KillSrc))
440         .addReg(SrcReg, getKillRegState(KillSrc));
441   }
442 }
443 
444 void RISCVInstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB,
445                                          MachineBasicBlock::iterator I,
446                                          Register SrcReg, bool IsKill, int FI,
447                                          const TargetRegisterClass *RC,
448                                          const TargetRegisterInfo *TRI) const {
449   DebugLoc DL;
450   if (I != MBB.end())
451     DL = I->getDebugLoc();
452 
453   MachineFunction *MF = MBB.getParent();
454   MachineFrameInfo &MFI = MF->getFrameInfo();
455 
456   unsigned Opcode;
457   bool IsScalableVector = true;
458   bool IsZvlsseg = true;
459   if (RISCV::GPRRegClass.hasSubClassEq(RC)) {
460     Opcode = TRI->getRegSizeInBits(RISCV::GPRRegClass) == 32 ?
461              RISCV::SW : RISCV::SD;
462     IsScalableVector = false;
463   } else if (RISCV::FPR16RegClass.hasSubClassEq(RC)) {
464     Opcode = RISCV::FSH;
465     IsScalableVector = false;
466   } else if (RISCV::FPR32RegClass.hasSubClassEq(RC)) {
467     Opcode = RISCV::FSW;
468     IsScalableVector = false;
469   } else if (RISCV::FPR64RegClass.hasSubClassEq(RC)) {
470     Opcode = RISCV::FSD;
471     IsScalableVector = false;
472   } else if (RISCV::VRRegClass.hasSubClassEq(RC)) {
473     Opcode = RISCV::PseudoVSPILL_M1;
474     IsZvlsseg = false;
475   } else if (RISCV::VRM2RegClass.hasSubClassEq(RC)) {
476     Opcode = RISCV::PseudoVSPILL_M2;
477     IsZvlsseg = false;
478   } else if (RISCV::VRM4RegClass.hasSubClassEq(RC)) {
479     Opcode = RISCV::PseudoVSPILL_M4;
480     IsZvlsseg = false;
481   } else if (RISCV::VRM8RegClass.hasSubClassEq(RC)) {
482     Opcode = RISCV::PseudoVSPILL_M8;
483     IsZvlsseg = false;
484   } else if (RISCV::VRN2M1RegClass.hasSubClassEq(RC))
485     Opcode = RISCV::PseudoVSPILL2_M1;
486   else if (RISCV::VRN2M2RegClass.hasSubClassEq(RC))
487     Opcode = RISCV::PseudoVSPILL2_M2;
488   else if (RISCV::VRN2M4RegClass.hasSubClassEq(RC))
489     Opcode = RISCV::PseudoVSPILL2_M4;
490   else if (RISCV::VRN3M1RegClass.hasSubClassEq(RC))
491     Opcode = RISCV::PseudoVSPILL3_M1;
492   else if (RISCV::VRN3M2RegClass.hasSubClassEq(RC))
493     Opcode = RISCV::PseudoVSPILL3_M2;
494   else if (RISCV::VRN4M1RegClass.hasSubClassEq(RC))
495     Opcode = RISCV::PseudoVSPILL4_M1;
496   else if (RISCV::VRN4M2RegClass.hasSubClassEq(RC))
497     Opcode = RISCV::PseudoVSPILL4_M2;
498   else if (RISCV::VRN5M1RegClass.hasSubClassEq(RC))
499     Opcode = RISCV::PseudoVSPILL5_M1;
500   else if (RISCV::VRN6M1RegClass.hasSubClassEq(RC))
501     Opcode = RISCV::PseudoVSPILL6_M1;
502   else if (RISCV::VRN7M1RegClass.hasSubClassEq(RC))
503     Opcode = RISCV::PseudoVSPILL7_M1;
504   else if (RISCV::VRN8M1RegClass.hasSubClassEq(RC))
505     Opcode = RISCV::PseudoVSPILL8_M1;
506   else
507     llvm_unreachable("Can't store this register to stack slot");
508 
509   if (IsScalableVector) {
510     MachineMemOperand *MMO = MF->getMachineMemOperand(
511         MachinePointerInfo::getFixedStack(*MF, FI), MachineMemOperand::MOStore,
512         MemoryLocation::UnknownSize, MFI.getObjectAlign(FI));
513 
514     MFI.setStackID(FI, TargetStackID::ScalableVector);
515     auto MIB = BuildMI(MBB, I, DL, get(Opcode))
516                    .addReg(SrcReg, getKillRegState(IsKill))
517                    .addFrameIndex(FI)
518                    .addMemOperand(MMO);
519     if (IsZvlsseg) {
520       // For spilling/reloading Zvlsseg registers, append the dummy field for
521       // the scaled vector length. The argument will be used when expanding
522       // these pseudo instructions.
523       MIB.addReg(RISCV::X0);
524     }
525   } else {
526     MachineMemOperand *MMO = MF->getMachineMemOperand(
527         MachinePointerInfo::getFixedStack(*MF, FI), MachineMemOperand::MOStore,
528         MFI.getObjectSize(FI), MFI.getObjectAlign(FI));
529 
530     BuildMI(MBB, I, DL, get(Opcode))
531         .addReg(SrcReg, getKillRegState(IsKill))
532         .addFrameIndex(FI)
533         .addImm(0)
534         .addMemOperand(MMO);
535   }
536 }
537 
538 void RISCVInstrInfo::loadRegFromStackSlot(MachineBasicBlock &MBB,
539                                           MachineBasicBlock::iterator I,
540                                           Register DstReg, int FI,
541                                           const TargetRegisterClass *RC,
542                                           const TargetRegisterInfo *TRI) const {
543   DebugLoc DL;
544   if (I != MBB.end())
545     DL = I->getDebugLoc();
546 
547   MachineFunction *MF = MBB.getParent();
548   MachineFrameInfo &MFI = MF->getFrameInfo();
549 
550   unsigned Opcode;
551   bool IsScalableVector = true;
552   bool IsZvlsseg = true;
553   if (RISCV::GPRRegClass.hasSubClassEq(RC)) {
554     Opcode = TRI->getRegSizeInBits(RISCV::GPRRegClass) == 32 ?
555              RISCV::LW : RISCV::LD;
556     IsScalableVector = false;
557   } else if (RISCV::FPR16RegClass.hasSubClassEq(RC)) {
558     Opcode = RISCV::FLH;
559     IsScalableVector = false;
560   } else if (RISCV::FPR32RegClass.hasSubClassEq(RC)) {
561     Opcode = RISCV::FLW;
562     IsScalableVector = false;
563   } else if (RISCV::FPR64RegClass.hasSubClassEq(RC)) {
564     Opcode = RISCV::FLD;
565     IsScalableVector = false;
566   } else if (RISCV::VRRegClass.hasSubClassEq(RC)) {
567     Opcode = RISCV::PseudoVRELOAD_M1;
568     IsZvlsseg = false;
569   } else if (RISCV::VRM2RegClass.hasSubClassEq(RC)) {
570     Opcode = RISCV::PseudoVRELOAD_M2;
571     IsZvlsseg = false;
572   } else if (RISCV::VRM4RegClass.hasSubClassEq(RC)) {
573     Opcode = RISCV::PseudoVRELOAD_M4;
574     IsZvlsseg = false;
575   } else if (RISCV::VRM8RegClass.hasSubClassEq(RC)) {
576     Opcode = RISCV::PseudoVRELOAD_M8;
577     IsZvlsseg = false;
578   } else if (RISCV::VRN2M1RegClass.hasSubClassEq(RC))
579     Opcode = RISCV::PseudoVRELOAD2_M1;
580   else if (RISCV::VRN2M2RegClass.hasSubClassEq(RC))
581     Opcode = RISCV::PseudoVRELOAD2_M2;
582   else if (RISCV::VRN2M4RegClass.hasSubClassEq(RC))
583     Opcode = RISCV::PseudoVRELOAD2_M4;
584   else if (RISCV::VRN3M1RegClass.hasSubClassEq(RC))
585     Opcode = RISCV::PseudoVRELOAD3_M1;
586   else if (RISCV::VRN3M2RegClass.hasSubClassEq(RC))
587     Opcode = RISCV::PseudoVRELOAD3_M2;
588   else if (RISCV::VRN4M1RegClass.hasSubClassEq(RC))
589     Opcode = RISCV::PseudoVRELOAD4_M1;
590   else if (RISCV::VRN4M2RegClass.hasSubClassEq(RC))
591     Opcode = RISCV::PseudoVRELOAD4_M2;
592   else if (RISCV::VRN5M1RegClass.hasSubClassEq(RC))
593     Opcode = RISCV::PseudoVRELOAD5_M1;
594   else if (RISCV::VRN6M1RegClass.hasSubClassEq(RC))
595     Opcode = RISCV::PseudoVRELOAD6_M1;
596   else if (RISCV::VRN7M1RegClass.hasSubClassEq(RC))
597     Opcode = RISCV::PseudoVRELOAD7_M1;
598   else if (RISCV::VRN8M1RegClass.hasSubClassEq(RC))
599     Opcode = RISCV::PseudoVRELOAD8_M1;
600   else
601     llvm_unreachable("Can't load this register from stack slot");
602 
603   if (IsScalableVector) {
604     MachineMemOperand *MMO = MF->getMachineMemOperand(
605         MachinePointerInfo::getFixedStack(*MF, FI), MachineMemOperand::MOLoad,
606         MemoryLocation::UnknownSize, MFI.getObjectAlign(FI));
607 
608     MFI.setStackID(FI, TargetStackID::ScalableVector);
609     auto MIB = BuildMI(MBB, I, DL, get(Opcode), DstReg)
610                    .addFrameIndex(FI)
611                    .addMemOperand(MMO);
612     if (IsZvlsseg) {
613       // For spilling/reloading Zvlsseg registers, append the dummy field for
614       // the scaled vector length. The argument will be used when expanding
615       // these pseudo instructions.
616       MIB.addReg(RISCV::X0);
617     }
618   } else {
619     MachineMemOperand *MMO = MF->getMachineMemOperand(
620         MachinePointerInfo::getFixedStack(*MF, FI), MachineMemOperand::MOLoad,
621         MFI.getObjectSize(FI), MFI.getObjectAlign(FI));
622 
623     BuildMI(MBB, I, DL, get(Opcode), DstReg)
624         .addFrameIndex(FI)
625         .addImm(0)
626         .addMemOperand(MMO);
627   }
628 }
629 
630 void RISCVInstrInfo::movImm(MachineBasicBlock &MBB,
631                             MachineBasicBlock::iterator MBBI,
632                             const DebugLoc &DL, Register DstReg, uint64_t Val,
633                             MachineInstr::MIFlag Flag) const {
634   MachineFunction *MF = MBB.getParent();
635   MachineRegisterInfo &MRI = MF->getRegInfo();
636   Register SrcReg = RISCV::X0;
637   Register Result = MRI.createVirtualRegister(&RISCV::GPRRegClass);
638   unsigned Num = 0;
639 
640   if (!STI.is64Bit() && !isInt<32>(Val))
641     report_fatal_error("Should only materialize 32-bit constants for RV32");
642 
643   RISCVMatInt::InstSeq Seq =
644       RISCVMatInt::generateInstSeq(Val, STI.getFeatureBits());
645   assert(!Seq.empty());
646 
647   for (RISCVMatInt::Inst &Inst : Seq) {
648     // Write the final result to DstReg if it's the last instruction in the Seq.
649     // Otherwise, write the result to the temp register.
650     if (++Num == Seq.size())
651       Result = DstReg;
652 
653     if (Inst.Opc == RISCV::LUI) {
654       BuildMI(MBB, MBBI, DL, get(RISCV::LUI), Result)
655           .addImm(Inst.Imm)
656           .setMIFlag(Flag);
657     } else if (Inst.Opc == RISCV::ADD_UW) {
658       BuildMI(MBB, MBBI, DL, get(RISCV::ADD_UW), Result)
659           .addReg(SrcReg, RegState::Kill)
660           .addReg(RISCV::X0)
661           .setMIFlag(Flag);
662     } else if (Inst.Opc == RISCV::SH1ADD || Inst.Opc == RISCV::SH2ADD ||
663                Inst.Opc == RISCV::SH3ADD) {
664       BuildMI(MBB, MBBI, DL, get(Inst.Opc), Result)
665           .addReg(SrcReg, RegState::Kill)
666           .addReg(SrcReg, RegState::Kill)
667           .setMIFlag(Flag);
668     } else {
669       BuildMI(MBB, MBBI, DL, get(Inst.Opc), Result)
670           .addReg(SrcReg, RegState::Kill)
671           .addImm(Inst.Imm)
672           .setMIFlag(Flag);
673     }
674     // Only the first instruction has X0 as its source.
675     SrcReg = Result;
676   }
677 }
678 
679 static RISCVCC::CondCode getCondFromBranchOpc(unsigned Opc) {
680   switch (Opc) {
681   default:
682     return RISCVCC::COND_INVALID;
683   case RISCV::BEQ:
684     return RISCVCC::COND_EQ;
685   case RISCV::BNE:
686     return RISCVCC::COND_NE;
687   case RISCV::BLT:
688     return RISCVCC::COND_LT;
689   case RISCV::BGE:
690     return RISCVCC::COND_GE;
691   case RISCV::BLTU:
692     return RISCVCC::COND_LTU;
693   case RISCV::BGEU:
694     return RISCVCC::COND_GEU;
695   }
696 }
697 
698 // The contents of values added to Cond are not examined outside of
699 // RISCVInstrInfo, giving us flexibility in what to push to it. For RISCV, we
700 // push BranchOpcode, Reg1, Reg2.
701 static void parseCondBranch(MachineInstr &LastInst, MachineBasicBlock *&Target,
702                             SmallVectorImpl<MachineOperand> &Cond) {
703   // Block ends with fall-through condbranch.
704   assert(LastInst.getDesc().isConditionalBranch() &&
705          "Unknown conditional branch");
706   Target = LastInst.getOperand(2).getMBB();
707   unsigned CC = getCondFromBranchOpc(LastInst.getOpcode());
708   Cond.push_back(MachineOperand::CreateImm(CC));
709   Cond.push_back(LastInst.getOperand(0));
710   Cond.push_back(LastInst.getOperand(1));
711 }
712 
713 const MCInstrDesc &RISCVInstrInfo::getBrCond(RISCVCC::CondCode CC) const {
714   switch (CC) {
715   default:
716     llvm_unreachable("Unknown condition code!");
717   case RISCVCC::COND_EQ:
718     return get(RISCV::BEQ);
719   case RISCVCC::COND_NE:
720     return get(RISCV::BNE);
721   case RISCVCC::COND_LT:
722     return get(RISCV::BLT);
723   case RISCVCC::COND_GE:
724     return get(RISCV::BGE);
725   case RISCVCC::COND_LTU:
726     return get(RISCV::BLTU);
727   case RISCVCC::COND_GEU:
728     return get(RISCV::BGEU);
729   }
730 }
731 
732 RISCVCC::CondCode RISCVCC::getOppositeBranchCondition(RISCVCC::CondCode CC) {
733   switch (CC) {
734   default:
735     llvm_unreachable("Unrecognized conditional branch");
736   case RISCVCC::COND_EQ:
737     return RISCVCC::COND_NE;
738   case RISCVCC::COND_NE:
739     return RISCVCC::COND_EQ;
740   case RISCVCC::COND_LT:
741     return RISCVCC::COND_GE;
742   case RISCVCC::COND_GE:
743     return RISCVCC::COND_LT;
744   case RISCVCC::COND_LTU:
745     return RISCVCC::COND_GEU;
746   case RISCVCC::COND_GEU:
747     return RISCVCC::COND_LTU;
748   }
749 }
750 
751 bool RISCVInstrInfo::analyzeBranch(MachineBasicBlock &MBB,
752                                    MachineBasicBlock *&TBB,
753                                    MachineBasicBlock *&FBB,
754                                    SmallVectorImpl<MachineOperand> &Cond,
755                                    bool AllowModify) const {
756   TBB = FBB = nullptr;
757   Cond.clear();
758 
759   // If the block has no terminators, it just falls into the block after it.
760   MachineBasicBlock::iterator I = MBB.getLastNonDebugInstr();
761   if (I == MBB.end() || !isUnpredicatedTerminator(*I))
762     return false;
763 
764   // Count the number of terminators and find the first unconditional or
765   // indirect branch.
766   MachineBasicBlock::iterator FirstUncondOrIndirectBr = MBB.end();
767   int NumTerminators = 0;
768   for (auto J = I.getReverse(); J != MBB.rend() && isUnpredicatedTerminator(*J);
769        J++) {
770     NumTerminators++;
771     if (J->getDesc().isUnconditionalBranch() ||
772         J->getDesc().isIndirectBranch()) {
773       FirstUncondOrIndirectBr = J.getReverse();
774     }
775   }
776 
777   // If AllowModify is true, we can erase any terminators after
778   // FirstUncondOrIndirectBR.
779   if (AllowModify && FirstUncondOrIndirectBr != MBB.end()) {
780     while (std::next(FirstUncondOrIndirectBr) != MBB.end()) {
781       std::next(FirstUncondOrIndirectBr)->eraseFromParent();
782       NumTerminators--;
783     }
784     I = FirstUncondOrIndirectBr;
785   }
786 
787   // We can't handle blocks that end in an indirect branch.
788   if (I->getDesc().isIndirectBranch())
789     return true;
790 
791   // We can't handle blocks with more than 2 terminators.
792   if (NumTerminators > 2)
793     return true;
794 
795   // Handle a single unconditional branch.
796   if (NumTerminators == 1 && I->getDesc().isUnconditionalBranch()) {
797     TBB = getBranchDestBlock(*I);
798     return false;
799   }
800 
801   // Handle a single conditional branch.
802   if (NumTerminators == 1 && I->getDesc().isConditionalBranch()) {
803     parseCondBranch(*I, TBB, Cond);
804     return false;
805   }
806 
807   // Handle a conditional branch followed by an unconditional branch.
808   if (NumTerminators == 2 && std::prev(I)->getDesc().isConditionalBranch() &&
809       I->getDesc().isUnconditionalBranch()) {
810     parseCondBranch(*std::prev(I), TBB, Cond);
811     FBB = getBranchDestBlock(*I);
812     return false;
813   }
814 
815   // Otherwise, we can't handle this.
816   return true;
817 }
818 
819 unsigned RISCVInstrInfo::removeBranch(MachineBasicBlock &MBB,
820                                       int *BytesRemoved) const {
821   if (BytesRemoved)
822     *BytesRemoved = 0;
823   MachineBasicBlock::iterator I = MBB.getLastNonDebugInstr();
824   if (I == MBB.end())
825     return 0;
826 
827   if (!I->getDesc().isUnconditionalBranch() &&
828       !I->getDesc().isConditionalBranch())
829     return 0;
830 
831   // Remove the branch.
832   if (BytesRemoved)
833     *BytesRemoved += getInstSizeInBytes(*I);
834   I->eraseFromParent();
835 
836   I = MBB.end();
837 
838   if (I == MBB.begin())
839     return 1;
840   --I;
841   if (!I->getDesc().isConditionalBranch())
842     return 1;
843 
844   // Remove the branch.
845   if (BytesRemoved)
846     *BytesRemoved += getInstSizeInBytes(*I);
847   I->eraseFromParent();
848   return 2;
849 }
850 
851 // Inserts a branch into the end of the specific MachineBasicBlock, returning
852 // the number of instructions inserted.
853 unsigned RISCVInstrInfo::insertBranch(
854     MachineBasicBlock &MBB, MachineBasicBlock *TBB, MachineBasicBlock *FBB,
855     ArrayRef<MachineOperand> Cond, const DebugLoc &DL, int *BytesAdded) const {
856   if (BytesAdded)
857     *BytesAdded = 0;
858 
859   // Shouldn't be a fall through.
860   assert(TBB && "insertBranch must not be told to insert a fallthrough");
861   assert((Cond.size() == 3 || Cond.size() == 0) &&
862          "RISCV branch conditions have two components!");
863 
864   // Unconditional branch.
865   if (Cond.empty()) {
866     MachineInstr &MI = *BuildMI(&MBB, DL, get(RISCV::PseudoBR)).addMBB(TBB);
867     if (BytesAdded)
868       *BytesAdded += getInstSizeInBytes(MI);
869     return 1;
870   }
871 
872   // Either a one or two-way conditional branch.
873   auto CC = static_cast<RISCVCC::CondCode>(Cond[0].getImm());
874   MachineInstr &CondMI =
875       *BuildMI(&MBB, DL, getBrCond(CC)).add(Cond[1]).add(Cond[2]).addMBB(TBB);
876   if (BytesAdded)
877     *BytesAdded += getInstSizeInBytes(CondMI);
878 
879   // One-way conditional branch.
880   if (!FBB)
881     return 1;
882 
883   // Two-way conditional branch.
884   MachineInstr &MI = *BuildMI(&MBB, DL, get(RISCV::PseudoBR)).addMBB(FBB);
885   if (BytesAdded)
886     *BytesAdded += getInstSizeInBytes(MI);
887   return 2;
888 }
889 
890 void RISCVInstrInfo::insertIndirectBranch(MachineBasicBlock &MBB,
891                                           MachineBasicBlock &DestBB,
892                                           MachineBasicBlock &RestoreBB,
893                                           const DebugLoc &DL, int64_t BrOffset,
894                                           RegScavenger *RS) const {
895   assert(RS && "RegScavenger required for long branching");
896   assert(MBB.empty() &&
897          "new block should be inserted for expanding unconditional branch");
898   assert(MBB.pred_size() == 1);
899 
900   MachineFunction *MF = MBB.getParent();
901   MachineRegisterInfo &MRI = MF->getRegInfo();
902 
903   if (!isInt<32>(BrOffset))
904     report_fatal_error(
905         "Branch offsets outside of the signed 32-bit range not supported");
906 
907   // FIXME: A virtual register must be used initially, as the register
908   // scavenger won't work with empty blocks (SIInstrInfo::insertIndirectBranch
909   // uses the same workaround).
910   Register ScratchReg = MRI.createVirtualRegister(&RISCV::GPRRegClass);
911   auto II = MBB.end();
912 
913   MachineInstr &MI = *BuildMI(MBB, II, DL, get(RISCV::PseudoJump))
914                           .addReg(ScratchReg, RegState::Define | RegState::Dead)
915                           .addMBB(&DestBB, RISCVII::MO_CALL);
916 
917   RS->enterBasicBlockEnd(MBB);
918   Register Scav = RS->scavengeRegisterBackwards(RISCV::GPRRegClass,
919                                                 MI.getIterator(), false, 0);
920   // TODO: The case when there is no scavenged register needs special handling.
921   assert(Scav != RISCV::NoRegister && "No register is scavenged!");
922   MRI.replaceRegWith(ScratchReg, Scav);
923   MRI.clearVirtRegs();
924   RS->setRegUsed(Scav);
925 }
926 
927 bool RISCVInstrInfo::reverseBranchCondition(
928     SmallVectorImpl<MachineOperand> &Cond) const {
929   assert((Cond.size() == 3) && "Invalid branch condition!");
930   auto CC = static_cast<RISCVCC::CondCode>(Cond[0].getImm());
931   Cond[0].setImm(getOppositeBranchCondition(CC));
932   return false;
933 }
934 
935 MachineBasicBlock *
936 RISCVInstrInfo::getBranchDestBlock(const MachineInstr &MI) const {
937   assert(MI.getDesc().isBranch() && "Unexpected opcode!");
938   // The branch target is always the last operand.
939   int NumOp = MI.getNumExplicitOperands();
940   return MI.getOperand(NumOp - 1).getMBB();
941 }
942 
943 bool RISCVInstrInfo::isBranchOffsetInRange(unsigned BranchOp,
944                                            int64_t BrOffset) const {
945   unsigned XLen = STI.getXLen();
946   // Ideally we could determine the supported branch offset from the
947   // RISCVII::FormMask, but this can't be used for Pseudo instructions like
948   // PseudoBR.
949   switch (BranchOp) {
950   default:
951     llvm_unreachable("Unexpected opcode!");
952   case RISCV::BEQ:
953   case RISCV::BNE:
954   case RISCV::BLT:
955   case RISCV::BGE:
956   case RISCV::BLTU:
957   case RISCV::BGEU:
958     return isIntN(13, BrOffset);
959   case RISCV::JAL:
960   case RISCV::PseudoBR:
961     return isIntN(21, BrOffset);
962   case RISCV::PseudoJump:
963     return isIntN(32, SignExtend64(BrOffset + 0x800, XLen));
964   }
965 }
966 
967 unsigned RISCVInstrInfo::getInstSizeInBytes(const MachineInstr &MI) const {
968   if (MI.isMetaInstruction())
969     return 0;
970 
971   unsigned Opcode = MI.getOpcode();
972 
973   if (Opcode == TargetOpcode::INLINEASM ||
974       Opcode == TargetOpcode::INLINEASM_BR) {
975     const MachineFunction &MF = *MI.getParent()->getParent();
976     const auto &TM = static_cast<const RISCVTargetMachine &>(MF.getTarget());
977     return getInlineAsmLength(MI.getOperand(0).getSymbolName(),
978                               *TM.getMCAsmInfo());
979   }
980 
981   if (MI.getParent() && MI.getParent()->getParent()) {
982     const auto MF = MI.getMF();
983     const auto &TM = static_cast<const RISCVTargetMachine &>(MF->getTarget());
984     const MCRegisterInfo &MRI = *TM.getMCRegisterInfo();
985     const MCSubtargetInfo &STI = *TM.getMCSubtargetInfo();
986     const RISCVSubtarget &ST = MF->getSubtarget<RISCVSubtarget>();
987     if (isCompressibleInst(MI, &ST, MRI, STI))
988       return 2;
989   }
990   return get(Opcode).getSize();
991 }
992 
993 bool RISCVInstrInfo::isAsCheapAsAMove(const MachineInstr &MI) const {
994   const unsigned Opcode = MI.getOpcode();
995   switch (Opcode) {
996   default:
997     break;
998   case RISCV::FSGNJ_D:
999   case RISCV::FSGNJ_S:
1000   case RISCV::FSGNJ_H:
1001     // The canonical floating-point move is fsgnj rd, rs, rs.
1002     return MI.getOperand(1).isReg() && MI.getOperand(2).isReg() &&
1003            MI.getOperand(1).getReg() == MI.getOperand(2).getReg();
1004   case RISCV::ADDI:
1005   case RISCV::ORI:
1006   case RISCV::XORI:
1007     return (MI.getOperand(1).isReg() &&
1008             MI.getOperand(1).getReg() == RISCV::X0) ||
1009            (MI.getOperand(2).isImm() && MI.getOperand(2).getImm() == 0);
1010   }
1011   return MI.isAsCheapAsAMove();
1012 }
1013 
1014 Optional<DestSourcePair>
1015 RISCVInstrInfo::isCopyInstrImpl(const MachineInstr &MI) const {
1016   if (MI.isMoveReg())
1017     return DestSourcePair{MI.getOperand(0), MI.getOperand(1)};
1018   switch (MI.getOpcode()) {
1019   default:
1020     break;
1021   case RISCV::ADDI:
1022     // Operand 1 can be a frameindex but callers expect registers
1023     if (MI.getOperand(1).isReg() && MI.getOperand(2).isImm() &&
1024         MI.getOperand(2).getImm() == 0)
1025       return DestSourcePair{MI.getOperand(0), MI.getOperand(1)};
1026     break;
1027   case RISCV::FSGNJ_D:
1028   case RISCV::FSGNJ_S:
1029   case RISCV::FSGNJ_H:
1030     // The canonical floating-point move is fsgnj rd, rs, rs.
1031     if (MI.getOperand(1).isReg() && MI.getOperand(2).isReg() &&
1032         MI.getOperand(1).getReg() == MI.getOperand(2).getReg())
1033       return DestSourcePair{MI.getOperand(0), MI.getOperand(1)};
1034     break;
1035   }
1036   return None;
1037 }
1038 
1039 bool RISCVInstrInfo::verifyInstruction(const MachineInstr &MI,
1040                                        StringRef &ErrInfo) const {
1041   const MCInstrInfo *MCII = STI.getInstrInfo();
1042   MCInstrDesc const &Desc = MCII->get(MI.getOpcode());
1043 
1044   for (auto &OI : enumerate(Desc.operands())) {
1045     unsigned OpType = OI.value().OperandType;
1046     if (OpType >= RISCVOp::OPERAND_FIRST_RISCV_IMM &&
1047         OpType <= RISCVOp::OPERAND_LAST_RISCV_IMM) {
1048       const MachineOperand &MO = MI.getOperand(OI.index());
1049       if (MO.isImm()) {
1050         int64_t Imm = MO.getImm();
1051         bool Ok;
1052         switch (OpType) {
1053         default:
1054           llvm_unreachable("Unexpected operand type");
1055         case RISCVOp::OPERAND_UIMM2:
1056           Ok = isUInt<2>(Imm);
1057           break;
1058         case RISCVOp::OPERAND_UIMM3:
1059           Ok = isUInt<3>(Imm);
1060           break;
1061         case RISCVOp::OPERAND_UIMM4:
1062           Ok = isUInt<4>(Imm);
1063           break;
1064         case RISCVOp::OPERAND_UIMM5:
1065           Ok = isUInt<5>(Imm);
1066           break;
1067         case RISCVOp::OPERAND_UIMM7:
1068           Ok = isUInt<7>(Imm);
1069           break;
1070         case RISCVOp::OPERAND_UIMM12:
1071           Ok = isUInt<12>(Imm);
1072           break;
1073         case RISCVOp::OPERAND_SIMM12:
1074           Ok = isInt<12>(Imm);
1075           break;
1076         case RISCVOp::OPERAND_UIMM20:
1077           Ok = isUInt<20>(Imm);
1078           break;
1079         case RISCVOp::OPERAND_UIMMLOG2XLEN:
1080           if (STI.getTargetTriple().isArch64Bit())
1081             Ok = isUInt<6>(Imm);
1082           else
1083             Ok = isUInt<5>(Imm);
1084           break;
1085         case RISCVOp::OPERAND_RVKRNUM:
1086           Ok = Imm >= 0 && Imm <= 10;
1087           break;
1088         }
1089         if (!Ok) {
1090           ErrInfo = "Invalid immediate";
1091           return false;
1092         }
1093       }
1094     }
1095   }
1096 
1097   return true;
1098 }
1099 
1100 // Return true if get the base operand, byte offset of an instruction and the
1101 // memory width. Width is the size of memory that is being loaded/stored.
1102 bool RISCVInstrInfo::getMemOperandWithOffsetWidth(
1103     const MachineInstr &LdSt, const MachineOperand *&BaseReg, int64_t &Offset,
1104     unsigned &Width, const TargetRegisterInfo *TRI) const {
1105   if (!LdSt.mayLoadOrStore())
1106     return false;
1107 
1108   // Here we assume the standard RISC-V ISA, which uses a base+offset
1109   // addressing mode. You'll need to relax these conditions to support custom
1110   // load/stores instructions.
1111   if (LdSt.getNumExplicitOperands() != 3)
1112     return false;
1113   if (!LdSt.getOperand(1).isReg() || !LdSt.getOperand(2).isImm())
1114     return false;
1115 
1116   if (!LdSt.hasOneMemOperand())
1117     return false;
1118 
1119   Width = (*LdSt.memoperands_begin())->getSize();
1120   BaseReg = &LdSt.getOperand(1);
1121   Offset = LdSt.getOperand(2).getImm();
1122   return true;
1123 }
1124 
1125 bool RISCVInstrInfo::areMemAccessesTriviallyDisjoint(
1126     const MachineInstr &MIa, const MachineInstr &MIb) const {
1127   assert(MIa.mayLoadOrStore() && "MIa must be a load or store.");
1128   assert(MIb.mayLoadOrStore() && "MIb must be a load or store.");
1129 
1130   if (MIa.hasUnmodeledSideEffects() || MIb.hasUnmodeledSideEffects() ||
1131       MIa.hasOrderedMemoryRef() || MIb.hasOrderedMemoryRef())
1132     return false;
1133 
1134   // Retrieve the base register, offset from the base register and width. Width
1135   // is the size of memory that is being loaded/stored (e.g. 1, 2, 4).  If
1136   // base registers are identical, and the offset of a lower memory access +
1137   // the width doesn't overlap the offset of a higher memory access,
1138   // then the memory accesses are different.
1139   const TargetRegisterInfo *TRI = STI.getRegisterInfo();
1140   const MachineOperand *BaseOpA = nullptr, *BaseOpB = nullptr;
1141   int64_t OffsetA = 0, OffsetB = 0;
1142   unsigned int WidthA = 0, WidthB = 0;
1143   if (getMemOperandWithOffsetWidth(MIa, BaseOpA, OffsetA, WidthA, TRI) &&
1144       getMemOperandWithOffsetWidth(MIb, BaseOpB, OffsetB, WidthB, TRI)) {
1145     if (BaseOpA->isIdenticalTo(*BaseOpB)) {
1146       int LowOffset = std::min(OffsetA, OffsetB);
1147       int HighOffset = std::max(OffsetA, OffsetB);
1148       int LowWidth = (LowOffset == OffsetA) ? WidthA : WidthB;
1149       if (LowOffset + LowWidth <= HighOffset)
1150         return true;
1151     }
1152   }
1153   return false;
1154 }
1155 
1156 std::pair<unsigned, unsigned>
1157 RISCVInstrInfo::decomposeMachineOperandsTargetFlags(unsigned TF) const {
1158   const unsigned Mask = RISCVII::MO_DIRECT_FLAG_MASK;
1159   return std::make_pair(TF & Mask, TF & ~Mask);
1160 }
1161 
1162 ArrayRef<std::pair<unsigned, const char *>>
1163 RISCVInstrInfo::getSerializableDirectMachineOperandTargetFlags() const {
1164   using namespace RISCVII;
1165   static const std::pair<unsigned, const char *> TargetFlags[] = {
1166       {MO_CALL, "riscv-call"},
1167       {MO_PLT, "riscv-plt"},
1168       {MO_LO, "riscv-lo"},
1169       {MO_HI, "riscv-hi"},
1170       {MO_PCREL_LO, "riscv-pcrel-lo"},
1171       {MO_PCREL_HI, "riscv-pcrel-hi"},
1172       {MO_GOT_HI, "riscv-got-hi"},
1173       {MO_TPREL_LO, "riscv-tprel-lo"},
1174       {MO_TPREL_HI, "riscv-tprel-hi"},
1175       {MO_TPREL_ADD, "riscv-tprel-add"},
1176       {MO_TLS_GOT_HI, "riscv-tls-got-hi"},
1177       {MO_TLS_GD_HI, "riscv-tls-gd-hi"}};
1178   return makeArrayRef(TargetFlags);
1179 }
1180 bool RISCVInstrInfo::isFunctionSafeToOutlineFrom(
1181     MachineFunction &MF, bool OutlineFromLinkOnceODRs) const {
1182   const Function &F = MF.getFunction();
1183 
1184   // Can F be deduplicated by the linker? If it can, don't outline from it.
1185   if (!OutlineFromLinkOnceODRs && F.hasLinkOnceODRLinkage())
1186     return false;
1187 
1188   // Don't outline from functions with section markings; the program could
1189   // expect that all the code is in the named section.
1190   if (F.hasSection())
1191     return false;
1192 
1193   // It's safe to outline from MF.
1194   return true;
1195 }
1196 
1197 bool RISCVInstrInfo::isMBBSafeToOutlineFrom(MachineBasicBlock &MBB,
1198                                             unsigned &Flags) const {
1199   // More accurate safety checking is done in getOutliningCandidateInfo.
1200   return TargetInstrInfo::isMBBSafeToOutlineFrom(MBB, Flags);
1201 }
1202 
1203 // Enum values indicating how an outlined call should be constructed.
1204 enum MachineOutlinerConstructionID {
1205   MachineOutlinerDefault
1206 };
1207 
1208 outliner::OutlinedFunction RISCVInstrInfo::getOutliningCandidateInfo(
1209     std::vector<outliner::Candidate> &RepeatedSequenceLocs) const {
1210 
1211   // First we need to filter out candidates where the X5 register (IE t0) can't
1212   // be used to setup the function call.
1213   auto CannotInsertCall = [](outliner::Candidate &C) {
1214     const TargetRegisterInfo *TRI = C.getMF()->getSubtarget().getRegisterInfo();
1215 
1216     C.initLRU(*TRI);
1217     LiveRegUnits LRU = C.LRU;
1218     return !LRU.available(RISCV::X5);
1219   };
1220 
1221   llvm::erase_if(RepeatedSequenceLocs, CannotInsertCall);
1222 
1223   // If the sequence doesn't have enough candidates left, then we're done.
1224   if (RepeatedSequenceLocs.size() < 2)
1225     return outliner::OutlinedFunction();
1226 
1227   unsigned SequenceSize = 0;
1228 
1229   auto I = RepeatedSequenceLocs[0].front();
1230   auto E = std::next(RepeatedSequenceLocs[0].back());
1231   for (; I != E; ++I)
1232     SequenceSize += getInstSizeInBytes(*I);
1233 
1234   // call t0, function = 8 bytes.
1235   unsigned CallOverhead = 8;
1236   for (auto &C : RepeatedSequenceLocs)
1237     C.setCallInfo(MachineOutlinerDefault, CallOverhead);
1238 
1239   // jr t0 = 4 bytes, 2 bytes if compressed instructions are enabled.
1240   unsigned FrameOverhead = 4;
1241   if (RepeatedSequenceLocs[0].getMF()->getSubtarget()
1242           .getFeatureBits()[RISCV::FeatureStdExtC])
1243     FrameOverhead = 2;
1244 
1245   return outliner::OutlinedFunction(RepeatedSequenceLocs, SequenceSize,
1246                                     FrameOverhead, MachineOutlinerDefault);
1247 }
1248 
1249 outliner::InstrType
1250 RISCVInstrInfo::getOutliningType(MachineBasicBlock::iterator &MBBI,
1251                                  unsigned Flags) const {
1252   MachineInstr &MI = *MBBI;
1253   MachineBasicBlock *MBB = MI.getParent();
1254   const TargetRegisterInfo *TRI =
1255       MBB->getParent()->getSubtarget().getRegisterInfo();
1256 
1257   // Positions generally can't safely be outlined.
1258   if (MI.isPosition()) {
1259     // We can manually strip out CFI instructions later.
1260     if (MI.isCFIInstruction())
1261       return outliner::InstrType::Invisible;
1262 
1263     return outliner::InstrType::Illegal;
1264   }
1265 
1266   // Don't trust the user to write safe inline assembly.
1267   if (MI.isInlineAsm())
1268     return outliner::InstrType::Illegal;
1269 
1270   // We can't outline branches to other basic blocks.
1271   if (MI.isTerminator() && !MBB->succ_empty())
1272     return outliner::InstrType::Illegal;
1273 
1274   // We need support for tail calls to outlined functions before return
1275   // statements can be allowed.
1276   if (MI.isReturn())
1277     return outliner::InstrType::Illegal;
1278 
1279   // Don't allow modifying the X5 register which we use for return addresses for
1280   // these outlined functions.
1281   if (MI.modifiesRegister(RISCV::X5, TRI) ||
1282       MI.getDesc().hasImplicitDefOfPhysReg(RISCV::X5))
1283     return outliner::InstrType::Illegal;
1284 
1285   // Make sure the operands don't reference something unsafe.
1286   for (const auto &MO : MI.operands())
1287     if (MO.isMBB() || MO.isBlockAddress() || MO.isCPI() || MO.isJTI())
1288       return outliner::InstrType::Illegal;
1289 
1290   // Don't allow instructions which won't be materialized to impact outlining
1291   // analysis.
1292   if (MI.isMetaInstruction())
1293     return outliner::InstrType::Invisible;
1294 
1295   return outliner::InstrType::Legal;
1296 }
1297 
1298 void RISCVInstrInfo::buildOutlinedFrame(
1299     MachineBasicBlock &MBB, MachineFunction &MF,
1300     const outliner::OutlinedFunction &OF) const {
1301 
1302   // Strip out any CFI instructions
1303   bool Changed = true;
1304   while (Changed) {
1305     Changed = false;
1306     auto I = MBB.begin();
1307     auto E = MBB.end();
1308     for (; I != E; ++I) {
1309       if (I->isCFIInstruction()) {
1310         I->removeFromParent();
1311         Changed = true;
1312         break;
1313       }
1314     }
1315   }
1316 
1317   MBB.addLiveIn(RISCV::X5);
1318 
1319   // Add in a return instruction to the end of the outlined frame.
1320   MBB.insert(MBB.end(), BuildMI(MF, DebugLoc(), get(RISCV::JALR))
1321       .addReg(RISCV::X0, RegState::Define)
1322       .addReg(RISCV::X5)
1323       .addImm(0));
1324 }
1325 
1326 MachineBasicBlock::iterator RISCVInstrInfo::insertOutlinedCall(
1327     Module &M, MachineBasicBlock &MBB, MachineBasicBlock::iterator &It,
1328     MachineFunction &MF, const outliner::Candidate &C) const {
1329 
1330   // Add in a call instruction to the outlined function at the given location.
1331   It = MBB.insert(It,
1332                   BuildMI(MF, DebugLoc(), get(RISCV::PseudoCALLReg), RISCV::X5)
1333                       .addGlobalAddress(M.getNamedValue(MF.getName()), 0,
1334                                         RISCVII::MO_CALL));
1335   return It;
1336 }
1337 
1338 // clang-format off
1339 #define CASE_VFMA_OPCODE_COMMON(OP, TYPE, LMUL)                                \
1340   RISCV::PseudoV##OP##_##TYPE##_##LMUL
1341 
1342 #define CASE_VFMA_OPCODE_LMULS_M1(OP, TYPE)                                    \
1343   CASE_VFMA_OPCODE_COMMON(OP, TYPE, M1):                                       \
1344   case CASE_VFMA_OPCODE_COMMON(OP, TYPE, M2):                                  \
1345   case CASE_VFMA_OPCODE_COMMON(OP, TYPE, M4):                                  \
1346   case CASE_VFMA_OPCODE_COMMON(OP, TYPE, M8)
1347 
1348 #define CASE_VFMA_OPCODE_LMULS_MF2(OP, TYPE)                                   \
1349   CASE_VFMA_OPCODE_COMMON(OP, TYPE, MF2):                                      \
1350   case CASE_VFMA_OPCODE_LMULS_M1(OP, TYPE)
1351 
1352 #define CASE_VFMA_OPCODE_LMULS_MF4(OP, TYPE)                                   \
1353   CASE_VFMA_OPCODE_COMMON(OP, TYPE, MF4):                                      \
1354   case CASE_VFMA_OPCODE_LMULS_MF2(OP, TYPE)
1355 
1356 #define CASE_VFMA_OPCODE_LMULS(OP, TYPE)                                       \
1357   CASE_VFMA_OPCODE_COMMON(OP, TYPE, MF8):                                      \
1358   case CASE_VFMA_OPCODE_LMULS_MF4(OP, TYPE)
1359 
1360 #define CASE_VFMA_SPLATS(OP)                                                   \
1361   CASE_VFMA_OPCODE_LMULS_MF4(OP, VF16):                                        \
1362   case CASE_VFMA_OPCODE_LMULS_MF2(OP, VF32):                                   \
1363   case CASE_VFMA_OPCODE_LMULS_M1(OP, VF64)
1364 // clang-format on
1365 
1366 bool RISCVInstrInfo::findCommutedOpIndices(const MachineInstr &MI,
1367                                            unsigned &SrcOpIdx1,
1368                                            unsigned &SrcOpIdx2) const {
1369   const MCInstrDesc &Desc = MI.getDesc();
1370   if (!Desc.isCommutable())
1371     return false;
1372 
1373   switch (MI.getOpcode()) {
1374   case CASE_VFMA_SPLATS(FMADD):
1375   case CASE_VFMA_SPLATS(FMSUB):
1376   case CASE_VFMA_SPLATS(FMACC):
1377   case CASE_VFMA_SPLATS(FMSAC):
1378   case CASE_VFMA_SPLATS(FNMADD):
1379   case CASE_VFMA_SPLATS(FNMSUB):
1380   case CASE_VFMA_SPLATS(FNMACC):
1381   case CASE_VFMA_SPLATS(FNMSAC):
1382   case CASE_VFMA_OPCODE_LMULS_MF4(FMACC, VV):
1383   case CASE_VFMA_OPCODE_LMULS_MF4(FMSAC, VV):
1384   case CASE_VFMA_OPCODE_LMULS_MF4(FNMACC, VV):
1385   case CASE_VFMA_OPCODE_LMULS_MF4(FNMSAC, VV):
1386   case CASE_VFMA_OPCODE_LMULS(MADD, VX):
1387   case CASE_VFMA_OPCODE_LMULS(NMSUB, VX):
1388   case CASE_VFMA_OPCODE_LMULS(MACC, VX):
1389   case CASE_VFMA_OPCODE_LMULS(NMSAC, VX):
1390   case CASE_VFMA_OPCODE_LMULS(MACC, VV):
1391   case CASE_VFMA_OPCODE_LMULS(NMSAC, VV): {
1392     // If the tail policy is undisturbed we can't commute.
1393     assert(RISCVII::hasVecPolicyOp(MI.getDesc().TSFlags));
1394     if ((MI.getOperand(MI.getNumExplicitOperands() - 1).getImm() & 1) == 0)
1395       return false;
1396 
1397     // For these instructions we can only swap operand 1 and operand 3 by
1398     // changing the opcode.
1399     unsigned CommutableOpIdx1 = 1;
1400     unsigned CommutableOpIdx2 = 3;
1401     if (!fixCommutedOpIndices(SrcOpIdx1, SrcOpIdx2, CommutableOpIdx1,
1402                               CommutableOpIdx2))
1403       return false;
1404     return true;
1405   }
1406   case CASE_VFMA_OPCODE_LMULS_MF4(FMADD, VV):
1407   case CASE_VFMA_OPCODE_LMULS_MF4(FMSUB, VV):
1408   case CASE_VFMA_OPCODE_LMULS_MF4(FNMADD, VV):
1409   case CASE_VFMA_OPCODE_LMULS_MF4(FNMSUB, VV):
1410   case CASE_VFMA_OPCODE_LMULS(MADD, VV):
1411   case CASE_VFMA_OPCODE_LMULS(NMSUB, VV): {
1412     // If the tail policy is undisturbed we can't commute.
1413     assert(RISCVII::hasVecPolicyOp(MI.getDesc().TSFlags));
1414     if ((MI.getOperand(MI.getNumExplicitOperands() - 1).getImm() & 1) == 0)
1415       return false;
1416 
1417     // For these instructions we have more freedom. We can commute with the
1418     // other multiplicand or with the addend/subtrahend/minuend.
1419 
1420     // Any fixed operand must be from source 1, 2 or 3.
1421     if (SrcOpIdx1 != CommuteAnyOperandIndex && SrcOpIdx1 > 3)
1422       return false;
1423     if (SrcOpIdx2 != CommuteAnyOperandIndex && SrcOpIdx2 > 3)
1424       return false;
1425 
1426     // It both ops are fixed one must be the tied source.
1427     if (SrcOpIdx1 != CommuteAnyOperandIndex &&
1428         SrcOpIdx2 != CommuteAnyOperandIndex && SrcOpIdx1 != 1 && SrcOpIdx2 != 1)
1429       return false;
1430 
1431     // Look for two different register operands assumed to be commutable
1432     // regardless of the FMA opcode. The FMA opcode is adjusted later if
1433     // needed.
1434     if (SrcOpIdx1 == CommuteAnyOperandIndex ||
1435         SrcOpIdx2 == CommuteAnyOperandIndex) {
1436       // At least one of operands to be commuted is not specified and
1437       // this method is free to choose appropriate commutable operands.
1438       unsigned CommutableOpIdx1 = SrcOpIdx1;
1439       if (SrcOpIdx1 == SrcOpIdx2) {
1440         // Both of operands are not fixed. Set one of commutable
1441         // operands to the tied source.
1442         CommutableOpIdx1 = 1;
1443       } else if (SrcOpIdx1 == CommuteAnyOperandIndex) {
1444         // Only one of the operands is not fixed.
1445         CommutableOpIdx1 = SrcOpIdx2;
1446       }
1447 
1448       // CommutableOpIdx1 is well defined now. Let's choose another commutable
1449       // operand and assign its index to CommutableOpIdx2.
1450       unsigned CommutableOpIdx2;
1451       if (CommutableOpIdx1 != 1) {
1452         // If we haven't already used the tied source, we must use it now.
1453         CommutableOpIdx2 = 1;
1454       } else {
1455         Register Op1Reg = MI.getOperand(CommutableOpIdx1).getReg();
1456 
1457         // The commuted operands should have different registers.
1458         // Otherwise, the commute transformation does not change anything and
1459         // is useless. We use this as a hint to make our decision.
1460         if (Op1Reg != MI.getOperand(2).getReg())
1461           CommutableOpIdx2 = 2;
1462         else
1463           CommutableOpIdx2 = 3;
1464       }
1465 
1466       // Assign the found pair of commutable indices to SrcOpIdx1 and
1467       // SrcOpIdx2 to return those values.
1468       if (!fixCommutedOpIndices(SrcOpIdx1, SrcOpIdx2, CommutableOpIdx1,
1469                                 CommutableOpIdx2))
1470         return false;
1471     }
1472 
1473     return true;
1474   }
1475   }
1476 
1477   return TargetInstrInfo::findCommutedOpIndices(MI, SrcOpIdx1, SrcOpIdx2);
1478 }
1479 
1480 #define CASE_VFMA_CHANGE_OPCODE_COMMON(OLDOP, NEWOP, TYPE, LMUL)               \
1481   case RISCV::PseudoV##OLDOP##_##TYPE##_##LMUL:                                \
1482     Opc = RISCV::PseudoV##NEWOP##_##TYPE##_##LMUL;                             \
1483     break;
1484 
1485 #define CASE_VFMA_CHANGE_OPCODE_LMULS_M1(OLDOP, NEWOP, TYPE)                   \
1486   CASE_VFMA_CHANGE_OPCODE_COMMON(OLDOP, NEWOP, TYPE, M1)                       \
1487   CASE_VFMA_CHANGE_OPCODE_COMMON(OLDOP, NEWOP, TYPE, M2)                       \
1488   CASE_VFMA_CHANGE_OPCODE_COMMON(OLDOP, NEWOP, TYPE, M4)                       \
1489   CASE_VFMA_CHANGE_OPCODE_COMMON(OLDOP, NEWOP, TYPE, M8)
1490 
1491 #define CASE_VFMA_CHANGE_OPCODE_LMULS_MF2(OLDOP, NEWOP, TYPE)                  \
1492   CASE_VFMA_CHANGE_OPCODE_COMMON(OLDOP, NEWOP, TYPE, MF2)                      \
1493   CASE_VFMA_CHANGE_OPCODE_LMULS_M1(OLDOP, NEWOP, TYPE)
1494 
1495 #define CASE_VFMA_CHANGE_OPCODE_LMULS_MF4(OLDOP, NEWOP, TYPE)                  \
1496   CASE_VFMA_CHANGE_OPCODE_COMMON(OLDOP, NEWOP, TYPE, MF4)                      \
1497   CASE_VFMA_CHANGE_OPCODE_LMULS_MF2(OLDOP, NEWOP, TYPE)
1498 
1499 #define CASE_VFMA_CHANGE_OPCODE_LMULS(OLDOP, NEWOP, TYPE)                      \
1500   CASE_VFMA_CHANGE_OPCODE_COMMON(OLDOP, NEWOP, TYPE, MF8)                      \
1501   CASE_VFMA_CHANGE_OPCODE_LMULS_MF4(OLDOP, NEWOP, TYPE)
1502 
1503 #define CASE_VFMA_CHANGE_OPCODE_SPLATS(OLDOP, NEWOP)                           \
1504   CASE_VFMA_CHANGE_OPCODE_LMULS_MF4(OLDOP, NEWOP, VF16)                        \
1505   CASE_VFMA_CHANGE_OPCODE_LMULS_MF2(OLDOP, NEWOP, VF32)                        \
1506   CASE_VFMA_CHANGE_OPCODE_LMULS_M1(OLDOP, NEWOP, VF64)
1507 
1508 MachineInstr *RISCVInstrInfo::commuteInstructionImpl(MachineInstr &MI,
1509                                                      bool NewMI,
1510                                                      unsigned OpIdx1,
1511                                                      unsigned OpIdx2) const {
1512   auto cloneIfNew = [NewMI](MachineInstr &MI) -> MachineInstr & {
1513     if (NewMI)
1514       return *MI.getParent()->getParent()->CloneMachineInstr(&MI);
1515     return MI;
1516   };
1517 
1518   switch (MI.getOpcode()) {
1519   case CASE_VFMA_SPLATS(FMACC):
1520   case CASE_VFMA_SPLATS(FMADD):
1521   case CASE_VFMA_SPLATS(FMSAC):
1522   case CASE_VFMA_SPLATS(FMSUB):
1523   case CASE_VFMA_SPLATS(FNMACC):
1524   case CASE_VFMA_SPLATS(FNMADD):
1525   case CASE_VFMA_SPLATS(FNMSAC):
1526   case CASE_VFMA_SPLATS(FNMSUB):
1527   case CASE_VFMA_OPCODE_LMULS_MF4(FMACC, VV):
1528   case CASE_VFMA_OPCODE_LMULS_MF4(FMSAC, VV):
1529   case CASE_VFMA_OPCODE_LMULS_MF4(FNMACC, VV):
1530   case CASE_VFMA_OPCODE_LMULS_MF4(FNMSAC, VV):
1531   case CASE_VFMA_OPCODE_LMULS(MADD, VX):
1532   case CASE_VFMA_OPCODE_LMULS(NMSUB, VX):
1533   case CASE_VFMA_OPCODE_LMULS(MACC, VX):
1534   case CASE_VFMA_OPCODE_LMULS(NMSAC, VX):
1535   case CASE_VFMA_OPCODE_LMULS(MACC, VV):
1536   case CASE_VFMA_OPCODE_LMULS(NMSAC, VV): {
1537     // It only make sense to toggle these between clobbering the
1538     // addend/subtrahend/minuend one of the multiplicands.
1539     assert((OpIdx1 == 1 || OpIdx2 == 1) && "Unexpected opcode index");
1540     assert((OpIdx1 == 3 || OpIdx2 == 3) && "Unexpected opcode index");
1541     unsigned Opc;
1542     switch (MI.getOpcode()) {
1543       default:
1544         llvm_unreachable("Unexpected opcode");
1545       CASE_VFMA_CHANGE_OPCODE_SPLATS(FMACC, FMADD)
1546       CASE_VFMA_CHANGE_OPCODE_SPLATS(FMADD, FMACC)
1547       CASE_VFMA_CHANGE_OPCODE_SPLATS(FMSAC, FMSUB)
1548       CASE_VFMA_CHANGE_OPCODE_SPLATS(FMSUB, FMSAC)
1549       CASE_VFMA_CHANGE_OPCODE_SPLATS(FNMACC, FNMADD)
1550       CASE_VFMA_CHANGE_OPCODE_SPLATS(FNMADD, FNMACC)
1551       CASE_VFMA_CHANGE_OPCODE_SPLATS(FNMSAC, FNMSUB)
1552       CASE_VFMA_CHANGE_OPCODE_SPLATS(FNMSUB, FNMSAC)
1553       CASE_VFMA_CHANGE_OPCODE_LMULS_MF4(FMACC, FMADD, VV)
1554       CASE_VFMA_CHANGE_OPCODE_LMULS_MF4(FMSAC, FMSUB, VV)
1555       CASE_VFMA_CHANGE_OPCODE_LMULS_MF4(FNMACC, FNMADD, VV)
1556       CASE_VFMA_CHANGE_OPCODE_LMULS_MF4(FNMSAC, FNMSUB, VV)
1557       CASE_VFMA_CHANGE_OPCODE_LMULS(MACC, MADD, VX)
1558       CASE_VFMA_CHANGE_OPCODE_LMULS(MADD, MACC, VX)
1559       CASE_VFMA_CHANGE_OPCODE_LMULS(NMSAC, NMSUB, VX)
1560       CASE_VFMA_CHANGE_OPCODE_LMULS(NMSUB, NMSAC, VX)
1561       CASE_VFMA_CHANGE_OPCODE_LMULS(MACC, MADD, VV)
1562       CASE_VFMA_CHANGE_OPCODE_LMULS(NMSAC, NMSUB, VV)
1563     }
1564 
1565     auto &WorkingMI = cloneIfNew(MI);
1566     WorkingMI.setDesc(get(Opc));
1567     return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
1568                                                    OpIdx1, OpIdx2);
1569   }
1570   case CASE_VFMA_OPCODE_LMULS_MF4(FMADD, VV):
1571   case CASE_VFMA_OPCODE_LMULS_MF4(FMSUB, VV):
1572   case CASE_VFMA_OPCODE_LMULS_MF4(FNMADD, VV):
1573   case CASE_VFMA_OPCODE_LMULS_MF4(FNMSUB, VV):
1574   case CASE_VFMA_OPCODE_LMULS(MADD, VV):
1575   case CASE_VFMA_OPCODE_LMULS(NMSUB, VV): {
1576     assert((OpIdx1 == 1 || OpIdx2 == 1) && "Unexpected opcode index");
1577     // If one of the operands, is the addend we need to change opcode.
1578     // Otherwise we're just swapping 2 of the multiplicands.
1579     if (OpIdx1 == 3 || OpIdx2 == 3) {
1580       unsigned Opc;
1581       switch (MI.getOpcode()) {
1582         default:
1583           llvm_unreachable("Unexpected opcode");
1584         CASE_VFMA_CHANGE_OPCODE_LMULS_MF4(FMADD, FMACC, VV)
1585         CASE_VFMA_CHANGE_OPCODE_LMULS_MF4(FMSUB, FMSAC, VV)
1586         CASE_VFMA_CHANGE_OPCODE_LMULS_MF4(FNMADD, FNMACC, VV)
1587         CASE_VFMA_CHANGE_OPCODE_LMULS_MF4(FNMSUB, FNMSAC, VV)
1588         CASE_VFMA_CHANGE_OPCODE_LMULS(MADD, MACC, VV)
1589         CASE_VFMA_CHANGE_OPCODE_LMULS(NMSUB, NMSAC, VV)
1590       }
1591 
1592       auto &WorkingMI = cloneIfNew(MI);
1593       WorkingMI.setDesc(get(Opc));
1594       return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
1595                                                      OpIdx1, OpIdx2);
1596     }
1597     // Let the default code handle it.
1598     break;
1599   }
1600   }
1601 
1602   return TargetInstrInfo::commuteInstructionImpl(MI, NewMI, OpIdx1, OpIdx2);
1603 }
1604 
1605 #undef CASE_VFMA_CHANGE_OPCODE_SPLATS
1606 #undef CASE_VFMA_CHANGE_OPCODE_LMULS
1607 #undef CASE_VFMA_CHANGE_OPCODE_COMMON
1608 #undef CASE_VFMA_SPLATS
1609 #undef CASE_VFMA_OPCODE_LMULS
1610 #undef CASE_VFMA_OPCODE_COMMON
1611 
1612 // clang-format off
1613 #define CASE_WIDEOP_OPCODE_COMMON(OP, LMUL)                                    \
1614   RISCV::PseudoV##OP##_##LMUL##_TIED
1615 
1616 #define CASE_WIDEOP_OPCODE_LMULS_MF4(OP)                                       \
1617   CASE_WIDEOP_OPCODE_COMMON(OP, MF4):                                          \
1618   case CASE_WIDEOP_OPCODE_COMMON(OP, MF2):                                     \
1619   case CASE_WIDEOP_OPCODE_COMMON(OP, M1):                                      \
1620   case CASE_WIDEOP_OPCODE_COMMON(OP, M2):                                      \
1621   case CASE_WIDEOP_OPCODE_COMMON(OP, M4)
1622 
1623 #define CASE_WIDEOP_OPCODE_LMULS(OP)                                           \
1624   CASE_WIDEOP_OPCODE_COMMON(OP, MF8):                                          \
1625   case CASE_WIDEOP_OPCODE_LMULS_MF4(OP)
1626 // clang-format on
1627 
1628 #define CASE_WIDEOP_CHANGE_OPCODE_COMMON(OP, LMUL)                             \
1629   case RISCV::PseudoV##OP##_##LMUL##_TIED:                                     \
1630     NewOpc = RISCV::PseudoV##OP##_##LMUL;                                      \
1631     break;
1632 
1633 #define CASE_WIDEOP_CHANGE_OPCODE_LMULS_MF4(OP)                                 \
1634   CASE_WIDEOP_CHANGE_OPCODE_COMMON(OP, MF4)                                    \
1635   CASE_WIDEOP_CHANGE_OPCODE_COMMON(OP, MF2)                                    \
1636   CASE_WIDEOP_CHANGE_OPCODE_COMMON(OP, M1)                                     \
1637   CASE_WIDEOP_CHANGE_OPCODE_COMMON(OP, M2)                                     \
1638   CASE_WIDEOP_CHANGE_OPCODE_COMMON(OP, M4)
1639 
1640 #define CASE_WIDEOP_CHANGE_OPCODE_LMULS(OP)                                    \
1641   CASE_WIDEOP_CHANGE_OPCODE_COMMON(OP, MF8)                                    \
1642   CASE_WIDEOP_CHANGE_OPCODE_LMULS_MF4(OP)
1643 
1644 MachineInstr *RISCVInstrInfo::convertToThreeAddress(MachineInstr &MI,
1645                                                     LiveVariables *LV,
1646                                                     LiveIntervals *LIS) const {
1647   switch (MI.getOpcode()) {
1648   default:
1649     break;
1650   case CASE_WIDEOP_OPCODE_LMULS_MF4(FWADD_WV):
1651   case CASE_WIDEOP_OPCODE_LMULS_MF4(FWSUB_WV):
1652   case CASE_WIDEOP_OPCODE_LMULS(WADD_WV):
1653   case CASE_WIDEOP_OPCODE_LMULS(WADDU_WV):
1654   case CASE_WIDEOP_OPCODE_LMULS(WSUB_WV):
1655   case CASE_WIDEOP_OPCODE_LMULS(WSUBU_WV): {
1656     // clang-format off
1657     unsigned NewOpc;
1658     switch (MI.getOpcode()) {
1659     default:
1660       llvm_unreachable("Unexpected opcode");
1661     CASE_WIDEOP_CHANGE_OPCODE_LMULS_MF4(FWADD_WV)
1662     CASE_WIDEOP_CHANGE_OPCODE_LMULS_MF4(FWSUB_WV)
1663     CASE_WIDEOP_CHANGE_OPCODE_LMULS(WADD_WV)
1664     CASE_WIDEOP_CHANGE_OPCODE_LMULS(WADDU_WV)
1665     CASE_WIDEOP_CHANGE_OPCODE_LMULS(WSUB_WV)
1666     CASE_WIDEOP_CHANGE_OPCODE_LMULS(WSUBU_WV)
1667     }
1668     // clang-format on
1669 
1670     MachineBasicBlock &MBB = *MI.getParent();
1671     MachineInstrBuilder MIB = BuildMI(MBB, MI, MI.getDebugLoc(), get(NewOpc))
1672                                   .add(MI.getOperand(0))
1673                                   .add(MI.getOperand(1))
1674                                   .add(MI.getOperand(2))
1675                                   .add(MI.getOperand(3))
1676                                   .add(MI.getOperand(4));
1677     MIB.copyImplicitOps(MI);
1678 
1679     if (LV) {
1680       unsigned NumOps = MI.getNumOperands();
1681       for (unsigned I = 1; I < NumOps; ++I) {
1682         MachineOperand &Op = MI.getOperand(I);
1683         if (Op.isReg() && Op.isKill())
1684           LV->replaceKillInstruction(Op.getReg(), MI, *MIB);
1685       }
1686     }
1687 
1688     if (LIS) {
1689       SlotIndex Idx = LIS->ReplaceMachineInstrInMaps(MI, *MIB);
1690 
1691       if (MI.getOperand(0).isEarlyClobber()) {
1692         // Use operand 1 was tied to early-clobber def operand 0, so its live
1693         // interval could have ended at an early-clobber slot. Now they are not
1694         // tied we need to update it to the normal register slot.
1695         LiveInterval &LI = LIS->getInterval(MI.getOperand(1).getReg());
1696         LiveRange::Segment *S = LI.getSegmentContaining(Idx);
1697         if (S->end == Idx.getRegSlot(true))
1698           S->end = Idx.getRegSlot();
1699       }
1700     }
1701 
1702     return MIB;
1703   }
1704   }
1705 
1706   return nullptr;
1707 }
1708 
1709 #undef CASE_WIDEOP_CHANGE_OPCODE_LMULS
1710 #undef CASE_WIDEOP_CHANGE_OPCODE_COMMON
1711 #undef CASE_WIDEOP_OPCODE_LMULS
1712 #undef CASE_WIDEOP_OPCODE_COMMON
1713 
1714 Register RISCVInstrInfo::getVLENFactoredAmount(MachineFunction &MF,
1715                                                MachineBasicBlock &MBB,
1716                                                MachineBasicBlock::iterator II,
1717                                                const DebugLoc &DL,
1718                                                int64_t Amount,
1719                                                MachineInstr::MIFlag Flag) const {
1720   assert(Amount > 0 && "There is no need to get VLEN scaled value.");
1721   assert(Amount % 8 == 0 &&
1722          "Reserve the stack by the multiple of one vector size.");
1723 
1724   MachineRegisterInfo &MRI = MF.getRegInfo();
1725   const RISCVInstrInfo *TII = MF.getSubtarget<RISCVSubtarget>().getInstrInfo();
1726   int64_t NumOfVReg = Amount / 8;
1727 
1728   Register VL = MRI.createVirtualRegister(&RISCV::GPRRegClass);
1729   BuildMI(MBB, II, DL, TII->get(RISCV::PseudoReadVLENB), VL)
1730     .setMIFlag(Flag);
1731   assert(isInt<32>(NumOfVReg) &&
1732          "Expect the number of vector registers within 32-bits.");
1733   if (isPowerOf2_32(NumOfVReg)) {
1734     uint32_t ShiftAmount = Log2_32(NumOfVReg);
1735     if (ShiftAmount == 0)
1736       return VL;
1737     BuildMI(MBB, II, DL, TII->get(RISCV::SLLI), VL)
1738         .addReg(VL, RegState::Kill)
1739         .addImm(ShiftAmount)
1740         .setMIFlag(Flag);
1741   } else if (isPowerOf2_32(NumOfVReg - 1)) {
1742     Register ScaledRegister = MRI.createVirtualRegister(&RISCV::GPRRegClass);
1743     uint32_t ShiftAmount = Log2_32(NumOfVReg - 1);
1744     BuildMI(MBB, II, DL, TII->get(RISCV::SLLI), ScaledRegister)
1745         .addReg(VL)
1746         .addImm(ShiftAmount)
1747         .setMIFlag(Flag);
1748     BuildMI(MBB, II, DL, TII->get(RISCV::ADD), VL)
1749         .addReg(ScaledRegister, RegState::Kill)
1750         .addReg(VL, RegState::Kill)
1751         .setMIFlag(Flag);
1752   } else if (isPowerOf2_32(NumOfVReg + 1)) {
1753     Register ScaledRegister = MRI.createVirtualRegister(&RISCV::GPRRegClass);
1754     uint32_t ShiftAmount = Log2_32(NumOfVReg + 1);
1755     BuildMI(MBB, II, DL, TII->get(RISCV::SLLI), ScaledRegister)
1756         .addReg(VL)
1757         .addImm(ShiftAmount)
1758         .setMIFlag(Flag);
1759     BuildMI(MBB, II, DL, TII->get(RISCV::SUB), VL)
1760         .addReg(ScaledRegister, RegState::Kill)
1761         .addReg(VL, RegState::Kill)
1762         .setMIFlag(Flag);
1763   } else {
1764     Register N = MRI.createVirtualRegister(&RISCV::GPRRegClass);
1765     if (!isInt<12>(NumOfVReg))
1766       movImm(MBB, II, DL, N, NumOfVReg);
1767     else {
1768       BuildMI(MBB, II, DL, TII->get(RISCV::ADDI), N)
1769           .addReg(RISCV::X0)
1770           .addImm(NumOfVReg)
1771           .setMIFlag(Flag);
1772     }
1773     if (!MF.getSubtarget<RISCVSubtarget>().hasStdExtM())
1774       MF.getFunction().getContext().diagnose(DiagnosticInfoUnsupported{
1775           MF.getFunction(),
1776           "M-extension must be enabled to calculate the vscaled size/offset."});
1777     BuildMI(MBB, II, DL, TII->get(RISCV::MUL), VL)
1778         .addReg(VL, RegState::Kill)
1779         .addReg(N, RegState::Kill)
1780         .setMIFlag(Flag);
1781   }
1782 
1783   return VL;
1784 }
1785 
1786 static bool isRVVWholeLoadStore(unsigned Opcode) {
1787   switch (Opcode) {
1788   default:
1789     return false;
1790   case RISCV::VS1R_V:
1791   case RISCV::VS2R_V:
1792   case RISCV::VS4R_V:
1793   case RISCV::VS8R_V:
1794   case RISCV::VL1RE8_V:
1795   case RISCV::VL2RE8_V:
1796   case RISCV::VL4RE8_V:
1797   case RISCV::VL8RE8_V:
1798   case RISCV::VL1RE16_V:
1799   case RISCV::VL2RE16_V:
1800   case RISCV::VL4RE16_V:
1801   case RISCV::VL8RE16_V:
1802   case RISCV::VL1RE32_V:
1803   case RISCV::VL2RE32_V:
1804   case RISCV::VL4RE32_V:
1805   case RISCV::VL8RE32_V:
1806   case RISCV::VL1RE64_V:
1807   case RISCV::VL2RE64_V:
1808   case RISCV::VL4RE64_V:
1809   case RISCV::VL8RE64_V:
1810     return true;
1811   }
1812 }
1813 
1814 bool RISCVInstrInfo::isRVVSpill(const MachineInstr &MI, bool CheckFIs) const {
1815   // RVV lacks any support for immediate addressing for stack addresses, so be
1816   // conservative.
1817   unsigned Opcode = MI.getOpcode();
1818   if (!RISCVVPseudosTable::getPseudoInfo(Opcode) &&
1819       !isRVVWholeLoadStore(Opcode) && !isRVVSpillForZvlsseg(Opcode))
1820     return false;
1821   return !CheckFIs || any_of(MI.operands(), [](const MachineOperand &MO) {
1822     return MO.isFI();
1823   });
1824 }
1825 
1826 Optional<std::pair<unsigned, unsigned>>
1827 RISCVInstrInfo::isRVVSpillForZvlsseg(unsigned Opcode) const {
1828   switch (Opcode) {
1829   default:
1830     return None;
1831   case RISCV::PseudoVSPILL2_M1:
1832   case RISCV::PseudoVRELOAD2_M1:
1833     return std::make_pair(2u, 1u);
1834   case RISCV::PseudoVSPILL2_M2:
1835   case RISCV::PseudoVRELOAD2_M2:
1836     return std::make_pair(2u, 2u);
1837   case RISCV::PseudoVSPILL2_M4:
1838   case RISCV::PseudoVRELOAD2_M4:
1839     return std::make_pair(2u, 4u);
1840   case RISCV::PseudoVSPILL3_M1:
1841   case RISCV::PseudoVRELOAD3_M1:
1842     return std::make_pair(3u, 1u);
1843   case RISCV::PseudoVSPILL3_M2:
1844   case RISCV::PseudoVRELOAD3_M2:
1845     return std::make_pair(3u, 2u);
1846   case RISCV::PseudoVSPILL4_M1:
1847   case RISCV::PseudoVRELOAD4_M1:
1848     return std::make_pair(4u, 1u);
1849   case RISCV::PseudoVSPILL4_M2:
1850   case RISCV::PseudoVRELOAD4_M2:
1851     return std::make_pair(4u, 2u);
1852   case RISCV::PseudoVSPILL5_M1:
1853   case RISCV::PseudoVRELOAD5_M1:
1854     return std::make_pair(5u, 1u);
1855   case RISCV::PseudoVSPILL6_M1:
1856   case RISCV::PseudoVRELOAD6_M1:
1857     return std::make_pair(6u, 1u);
1858   case RISCV::PseudoVSPILL7_M1:
1859   case RISCV::PseudoVRELOAD7_M1:
1860     return std::make_pair(7u, 1u);
1861   case RISCV::PseudoVSPILL8_M1:
1862   case RISCV::PseudoVRELOAD8_M1:
1863     return std::make_pair(8u, 1u);
1864   }
1865 }
1866