RISCVRegisterInfo.cpp

Go to the documentation of this file.

//===-- RISCVRegisterInfo.cpp - RISC-V Register Information -----*- C++ -*-===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//
//
// This file contains the RISC-V implementation of the TargetRegisterInfo class.
//
//===----------------------------------------------------------------------===//
 
#include "RISCVRegisterInfo.h"
#include "RISCV.h"
#include "RISCVSubtarget.h"
#include "llvm/ADT/SmallSet.h"
#include "llvm/BinaryFormat/Dwarf.h"
#include "llvm/CodeGen/MachineFrameInfo.h"
#include "llvm/CodeGen/MachineFunction.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/CodeGen/RegisterScavenging.h"
#include "llvm/CodeGen/TargetFrameLowering.h"
#include "llvm/CodeGen/TargetInstrInfo.h"
#include "llvm/IR/DebugInfoMetadata.h"
#include "llvm/Support/ErrorHandling.h"
 
#define GET_REGINFO_TARGET_DESC
#include "RISCVGenRegisterInfo.inc"
 
using namespace llvm;
 
static_assert(RISCV::X1 == RISCV::X0 + 1, "Register list not consecutive");
static_assert(RISCV::X31 == RISCV::X0 + 31, "Register list not consecutive");
static_assert(RISCV::F1_H == RISCV::F0_H + 1, "Register list not consecutive");
static_assert(RISCV::F31_H == RISCV::F0_H + 31,
              "Register list not consecutive");
static_assert(RISCV::F1_F == RISCV::F0_F + 1, "Register list not consecutive");
static_assert(RISCV::F31_F == RISCV::F0_F + 31,
              "Register list not consecutive");
static_assert(RISCV::F1_D == RISCV::F0_D + 1, "Register list not consecutive");
static_assert(RISCV::F31_D == RISCV::F0_D + 31,
              "Register list not consecutive");
static_assert(RISCV::F1_Q == RISCV::F0_Q + 1, "Register list not consecutive");
static_assert(RISCV::F31_Q == RISCV::F0_Q + 31,
              "Register list not consecutive");
static_assert(RISCV::V1 == RISCV::V0 + 1, "Register list not consecutive");
static_assert(RISCV::V31 == RISCV::V0 + 31, "Register list not consecutive");
 
RISCVRegisterInfo::RISCVRegisterInfo(unsigned HwMode,
                                     const std::bitset<32> &UserReservedXRegs)
    : RISCVGenRegisterInfo(RISCV::X1, /*DwarfFlavour*/ 0, /*EHFlavor*/ 0,
                           /*PC*/ 0, HwMode),
      UserReservedRegs(getNumRegs()) {
  // Expand the reserve-xN features to include aliases.
  for (unsigned I = 1; I != 32; ++I) {
    if (!UserReservedXRegs[I])
      continue;
    for (MCRegAliasIterator AI(RISCV::X0 + I, this, /*IncludeSelf=*/true);
         AI.isValid(); ++AI)
      UserReservedRegs.set((*AI).id());
  }
}
 
const MCPhysReg *
RISCVRegisterInfo::getIPRACSRegs(const MachineFunction *MF) const {
  return CSR_IPRA_SaveList;
}
 
const MCPhysReg *
RISCVRegisterInfo::getCalleeSavedRegs(const MachineFunction *MF) const {
  auto &Subtarget = MF->getSubtarget<RISCVSubtarget>();
  if (MF->getFunction().getCallingConv() == CallingConv::GHC)
    return CSR_NoRegs_SaveList;
  if (MF->getFunction().getCallingConv() == CallingConv::PreserveMost)
    return Subtarget.hasStdExtE() ? CSR_RT_MostRegs_RVE_SaveList
                                  : CSR_RT_MostRegs_SaveList;
  if (MF->getFunction().hasFnAttribute("interrupt")) {
    if (Subtarget.hasVInstructions()) {
      if (Subtarget.hasStdExtD())
        return Subtarget.hasStdExtE() ? CSR_XLEN_F64_V_Interrupt_RVE_SaveList
                                      : CSR_XLEN_F64_V_Interrupt_SaveList;
      if (Subtarget.hasStdExtF())
        return Subtarget.hasStdExtE() ? CSR_XLEN_F32_V_Interrupt_RVE_SaveList
                                      : CSR_XLEN_F32_V_Interrupt_SaveList;
      return Subtarget.hasStdExtE() ? CSR_XLEN_V_Interrupt_RVE_SaveList
                                    : CSR_XLEN_V_Interrupt_SaveList;
    }
    if (Subtarget.hasStdExtD())
      return Subtarget.hasStdExtE() ? CSR_XLEN_F64_Interrupt_RVE_SaveList
                                    : CSR_XLEN_F64_Interrupt_SaveList;
    if (Subtarget.hasStdExtF())
      return Subtarget.hasStdExtE() ? CSR_XLEN_F32_Interrupt_RVE_SaveList
                                    : CSR_XLEN_F32_Interrupt_SaveList;
    return Subtarget.hasStdExtE() ? CSR_Interrupt_RVE_SaveList
                                  : CSR_Interrupt_SaveList;
  }
 
  bool HasVectorCSR =
      MF->getFunction().getCallingConv() == CallingConv::RISCV_VectorCall &&
      Subtarget.hasVInstructions();
 
  RISCVABI::ABI ABI = Subtarget.getTargetABI();
  assert(ABI != RISCVABI::ABI::Unknown && !RISCVABI::isPurecap(ABI) &&
         "Unrecognized ABI");
  if (RISCVABI::isRVE(ABI))
    return CSR_ILP32E_LP64E_SaveList;
  if (RISCVABI::isSoftFP(ABI))
    return HasVectorCSR ? CSR_ILP32_LP64_V_SaveList : CSR_ILP32_LP64_SaveList;
  if (RISCVABI::isSingleFP(ABI))
    return HasVectorCSR ? CSR_ILP32F_LP64F_V_SaveList
                        : CSR_ILP32F_LP64F_SaveList;
  assert(RISCVABI::isDoubleFP(ABI) && "Unrecognized ABI");
  return HasVectorCSR ? CSR_ILP32D_LP64D_V_SaveList : CSR_ILP32D_LP64D_SaveList;
}
 
const TargetRegisterClass *RISCVRegisterInfo::getConstrainedRegClassForReg(
    Register Reg, const MachineRegisterInfo &MRI) const {
  const RISCVSubtarget &STI = MRI.getMF().getSubtarget<RISCVSubtarget>();
 
  const RegClassOrRegBank &RCOrRB = MRI.getRegClassOrRegBank(Reg);
  if (const RegisterBank *RB = dyn_cast<const RegisterBank *>(RCOrRB))
    return getRegClassForTypeOnBank(MRI.getType(Reg), *RB, STI.is64Bit());
 
  if (const auto *RC = dyn_cast<const TargetRegisterClass *>(RCOrRB)) {
    return getAllocatableClass(RC);
  }
 
  return nullptr;
}
 
const TargetRegisterClass *
RISCVRegisterInfo::getRegClassForTypeOnBank(LLT Ty, const RegisterBank &RB,
                                            bool Is64Bit) const {
  if (RB.getID() == RISCV::GPRBRegBankID) {
    if (Ty.getSizeInBits() <= 32 || (Is64Bit && Ty.getSizeInBits() == 64))
      return &RISCV::GPRRegClass;
  }
 
  if (RB.getID() == RISCV::FPRBRegBankID) {
    if (Ty.getSizeInBits() == 16)
      return &RISCV::FPR16RegClass;
    if (Ty.getSizeInBits() == 32)
      return &RISCV::FPR32RegClass;
    if (Ty.getSizeInBits() == 64)
      return &RISCV::FPR64RegClass;
  }
 
  if (RB.getID() == RISCV::VRBRegBankID) {
    if (Ty.getSizeInBits().getKnownMinValue() <= 64)
      return &RISCV::VRRegClass;
 
    if (Ty.getSizeInBits().getKnownMinValue() == 128)
      return &RISCV::VRM2RegClass;
 
    if (Ty.getSizeInBits().getKnownMinValue() == 256)
      return &RISCV::VRM4RegClass;
 
    if (Ty.getSizeInBits().getKnownMinValue() == 512)
      return &RISCV::VRM8RegClass;
  }
 
  return nullptr;
}
 
BitVector RISCVRegisterInfo::getReservedRegs(const MachineFunction &MF) const {
  const RISCVFrameLowering *TFI = getFrameLowering(MF);
  BitVector Reserved(getNumRegs());
  auto &Subtarget = MF.getSubtarget<RISCVSubtarget>();
 
  // Mark any GPRs requested to be reserved as such.
  const std::bitset<32> &UserReservedXRegs = Subtarget.getUserReservedXRegs();
  for (unsigned I = 1; I != 32; ++I) {
    if (!UserReservedXRegs[I])
      continue;
    markSuperRegs(Reserved, RISCV::X0 + I);
  }
 
  // Use markSuperRegs to ensure any register aliases are also reserved
  markSuperRegs(Reserved, RISCV::X0_H); // x0
  markSuperRegs(Reserved, RISCV::X2_H); // sp
  markSuperRegs(Reserved, RISCV::X3_H); // gp
  markSuperRegs(Reserved, RISCV::X4_H); // tp
  if (TFI->hasFP(MF))
    markSuperRegs(Reserved, RISCV::X8_H); // fp
  // Reserve the base register if we need to realign the stack and allocate
  // variable-sized objects at runtime.
  if (TFI->hasBP(MF))
    markSuperRegs(Reserved, RISCVABI::getBPReg()); // bp
 
  // There are only 16 GPRs for RVE.
  if (Subtarget.hasStdExtE())
    for (MCPhysReg Reg = RISCV::X16_H; Reg <= RISCV::X31_H; Reg++)
      markSuperRegs(Reserved, Reg);
 
  // V registers for code generation. We handle them manually.
  markSuperRegs(Reserved, RISCV::VL);
  markSuperRegs(Reserved, RISCV::VTYPE);
  markSuperRegs(Reserved, RISCV::VXSAT);
  markSuperRegs(Reserved, RISCV::VXRM);
  markSuperRegs(Reserved, RISCV::VLENB);
 
  // Floating point environment registers.
  markSuperRegs(Reserved, RISCV::FRM);
  markSuperRegs(Reserved, RISCV::FFLAGS);
 
  // SiFive VCIX state registers.
  markSuperRegs(Reserved, RISCV::SF_VCIX_STATE);
 
  if (MF.getFunction().getCallingConv() == CallingConv::GRAAL) {
    if (Subtarget.hasStdExtE())
      reportFatalUsageError("Graal reserved registers do not exist in RVE");
    markSuperRegs(Reserved, RISCV::X23_H);
    markSuperRegs(Reserved, RISCV::X27_H);
  }
 
  // Shadow stack pointer.
  markSuperRegs(Reserved, RISCV::SSP);
 
  // XSfmmbase
  for (MCPhysReg Reg = RISCV::T0; Reg <= RISCV::T15; Reg++)
    markSuperRegs(Reserved, Reg);
 
  assert(checkAllSuperRegsMarked(Reserved));
  return Reserved;
}
 
bool RISCVRegisterInfo::isAsmClobberable(const MachineFunction &MF,
                                         MCRegister PhysReg) const {
  return !MF.getSubtarget().isRegisterReservedByUser(PhysReg);
}
 
const uint32_t *RISCVRegisterInfo::getNoPreservedMask() const {
  return CSR_NoRegs_RegMask;
}
 
void RISCVRegisterInfo::adjustReg(MachineBasicBlock &MBB,
                                  MachineBasicBlock::iterator II,
                                  const DebugLoc &DL, Register DestReg,
                                  Register SrcReg, StackOffset Offset,
                                  MachineInstr::MIFlag Flag,
                                  MaybeAlign RequiredAlign) const {
 
  if (DestReg == SrcReg && !Offset.getFixed() && !Offset.getScalable())
    return;
 
  MachineFunction &MF = *MBB.getParent();
  MachineRegisterInfo &MRI = MF.getRegInfo();
  const RISCVSubtarget &ST = MF.getSubtarget<RISCVSubtarget>();
  const RISCVInstrInfo *TII = ST.getInstrInfo();
 
  // Optimize compile time offset case
  if (Offset.getScalable()) {
    if (auto VLEN = ST.getRealVLen()) {
      // 1. Multiply the number of v-slots by the (constant) length of register
      const int64_t VLENB = *VLEN / 8;
      assert(Offset.getScalable() % RISCV::RVVBytesPerBlock == 0 &&
             "Reserve the stack by the multiple of one vector size.");
      const int64_t NumOfVReg = Offset.getScalable() / 8;
      const int64_t FixedOffset = NumOfVReg * VLENB;
      if (!isInt<32>(FixedOffset)) {
        // This check might also need to be updated to 64bit.
        // However mulImm() still assumes 32bit. For now only support fixed
        // 64bit frame offsets, since scalable offsets would require the number
        // of spilled registers to exceed 2^31, which is unlikely.
        reportFatalUsageError("Scalable frame size outside of the signed "
                              "32-bit range not supported");
      }
      Offset = StackOffset::getFixed(FixedOffset + Offset.getFixed());
    }
  }
 
  bool KillSrcReg = false;
 
  if (Offset.getScalable()) {
    unsigned ScalableAdjOpc = RISCV::ADD;
    int64_t ScalableValue = Offset.getScalable();
    if (ScalableValue < 0) {
      ScalableValue = -ScalableValue;
      ScalableAdjOpc = RISCV::SUB;
    }
    // Get vlenb and multiply vlen with the number of vector registers.
    Register ScratchReg = DestReg;
    if (DestReg == SrcReg)
      ScratchReg = MRI.createVirtualRegister(&RISCV::GPRRegClass);
 
    assert(ScalableValue > 0 && "There is no need to get VLEN scaled value.");
    assert(ScalableValue % RISCV::RVVBytesPerBlock == 0 &&
           "Reserve the stack by the multiple of one vector size.");
    assert(isInt<32>(ScalableValue / RISCV::RVVBytesPerBlock) &&
           "Expect the number of vector registers within 32-bits.");
    uint32_t NumOfVReg = ScalableValue / RISCV::RVVBytesPerBlock;
    // Only use vsetvli rather than vlenb if adjusting in the prologue or
    // epilogue, otherwise it may disturb the VTYPE and VL status.
    bool IsPrologueOrEpilogue =
        Flag == MachineInstr::FrameSetup || Flag == MachineInstr::FrameDestroy;
    bool UseVsetvliRatherThanVlenb =
        IsPrologueOrEpilogue && ST.preferVsetvliOverReadVLENB();
    if (UseVsetvliRatherThanVlenb && (NumOfVReg == 1 || NumOfVReg == 2 ||
                                      NumOfVReg == 4 || NumOfVReg == 8)) {
      BuildMI(MBB, II, DL, TII->get(RISCV::PseudoReadVLENBViaVSETVLIX0),
              ScratchReg)
          .addImm(NumOfVReg)
          .setMIFlag(Flag);
      BuildMI(MBB, II, DL, TII->get(ScalableAdjOpc), DestReg)
          .addReg(SrcReg)
          .addReg(ScratchReg, RegState::Kill)
          .setMIFlag(Flag);
    } else {
      if (UseVsetvliRatherThanVlenb)
        BuildMI(MBB, II, DL, TII->get(RISCV::PseudoReadVLENBViaVSETVLIX0),
                ScratchReg)
            .addImm(1)
            .setMIFlag(Flag);
      else
        BuildMI(MBB, II, DL, TII->get(RISCV::PseudoReadVLENB), ScratchReg)
            .setMIFlag(Flag);
 
      if (ScalableAdjOpc == RISCV::ADD && ST.hasStdExtZba() &&
          (NumOfVReg == 2 || NumOfVReg == 4 || NumOfVReg == 8)) {
        unsigned Opc = NumOfVReg == 2
                           ? RISCV::SH1ADD
                           : (NumOfVReg == 4 ? RISCV::SH2ADD : RISCV::SH3ADD);
        BuildMI(MBB, II, DL, TII->get(Opc), DestReg)
            .addReg(ScratchReg, RegState::Kill)
            .addReg(SrcReg)
            .setMIFlag(Flag);
      } else {
        TII->mulImm(MF, MBB, II, DL, ScratchReg, NumOfVReg, Flag);
        BuildMI(MBB, II, DL, TII->get(ScalableAdjOpc), DestReg)
            .addReg(SrcReg)
            .addReg(ScratchReg, RegState::Kill)
            .setMIFlag(Flag);
      }
    }
    SrcReg = DestReg;
    KillSrcReg = true;
  }
 
  int64_t Val = Offset.getFixed();
  if (DestReg == SrcReg && Val == 0)
    return;
 
  const uint64_t Align = RequiredAlign.valueOrOne().value();
 
  if (isInt<12>(Val)) {
    BuildMI(MBB, II, DL, TII->get(RISCV::ADDI), DestReg)
        .addReg(SrcReg, getKillRegState(KillSrcReg))
        .addImm(Val)
        .setMIFlag(Flag);
    return;
  }
 
  // Use the QC_E_ADDI instruction from the Xqcilia extension that can take a
  // signed 26-bit immediate.
  if (ST.hasVendorXqcilia() && isInt<26>(Val)) {
    // The one case where using this instruction is sub-optimal is if Val can be
    // materialized with a single compressible LUI and following add/sub is also
    // compressible. Avoid doing this if that is the case.
    int Hi20 = (Val & 0xFFFFF000) >> 12;
    bool IsCompressLUI =
        ((Val & 0xFFF) == 0) && (Hi20 != 0) &&
        (isUInt<5>(Hi20) || (Hi20 >= 0xfffe0 && Hi20 <= 0xfffff));
    bool IsCompressAddSub =
        (SrcReg == DestReg) &&
        ((Val > 0 && RISCV::GPRNoX0RegClass.contains(SrcReg)) ||
         (Val < 0 && RISCV::GPRCRegClass.contains(SrcReg)));
 
    if (!(IsCompressLUI && IsCompressAddSub)) {
      BuildMI(MBB, II, DL, TII->get(RISCV::QC_E_ADDI), DestReg)
          .addReg(SrcReg, getKillRegState(KillSrcReg))
          .addImm(Val)
          .setMIFlag(Flag);
      return;
    }
  }
 
  // Try to split the offset across two ADDIs. We need to keep the intermediate
  // result aligned after each ADDI.  We need to determine the maximum value we
  // can put in each ADDI. In the negative direction, we can use -2048 which is
  // always sufficiently aligned. In the positive direction, we need to find the
  // largest 12-bit immediate that is aligned.  Exclude -4096 since it can be
  // created with LUI.
  assert(Align < 2048 && "Required alignment too large");
  int64_t MaxPosAdjStep = 2048 - Align;
  if (Val > -4096 && Val <= (2 * MaxPosAdjStep)) {
    int64_t FirstAdj = Val < 0 ? -2048 : MaxPosAdjStep;
    Val -= FirstAdj;
    BuildMI(MBB, II, DL, TII->get(RISCV::ADDI), DestReg)
        .addReg(SrcReg, getKillRegState(KillSrcReg))
        .addImm(FirstAdj)
        .setMIFlag(Flag);
    BuildMI(MBB, II, DL, TII->get(RISCV::ADDI), DestReg)
        .addReg(DestReg, RegState::Kill)
        .addImm(Val)
        .setMIFlag(Flag);
    return;
  }
 
  // Use shNadd if doing so lets us materialize a 12 bit immediate with a single
  // instruction.  This saves 1 instruction over the full lui/addi+add fallback
  // path.  We avoid anything which can be done with a single lui as it might
  // be compressible.  Note that the sh1add case is fully covered by the 2x addi
  // case just above and is thus omitted.
  if (ST.hasStdExtZba() && (Val & 0xFFF) != 0) {
    unsigned Opc = 0;
    if (isShiftedInt<12, 3>(Val)) {
      Opc = RISCV::SH3ADD;
      Val = Val >> 3;
    } else if (isShiftedInt<12, 2>(Val)) {
      Opc = RISCV::SH2ADD;
      Val = Val >> 2;
    }
    if (Opc) {
      Register ScratchReg = MRI.createVirtualRegister(&RISCV::GPRRegClass);
      TII->movImm(MBB, II, DL, ScratchReg, Val, Flag);
      BuildMI(MBB, II, DL, TII->get(Opc), DestReg)
          .addReg(ScratchReg, RegState::Kill)
          .addReg(SrcReg, getKillRegState(KillSrcReg))
          .setMIFlag(Flag);
      return;
    }
  }
 
  // Emit a PseudoAddUpperImm instead of LUI+ADD when the offset is a multiple
  // of 4096 and the source is the frame register. The frame register is
  // invariant after PEI, so MachineLateInstrsCleanup can CSE identical pseudos.
  // The pseudo is later expanded back to LUI+ADD.
  if (Flag == MachineInstr::NoFlags && !KillSrcReg && DestReg != SrcReg &&
      SrcReg == getFrameRegister(MF) && isShiftedInt<20, 12>(Val)) {
    BuildMI(MBB, II, DL, TII->get(RISCV::PseudoAddUpperImm), DestReg)
        .addReg(SrcReg)
        .addImm(static_cast<uint32_t>(Val) >> 12);
    return;
  }
 
  unsigned Opc = RISCV::ADD;
  if (Val < 0) {
    Val = -Val;
    Opc = RISCV::SUB;
  }
 
  Register ScratchReg = MRI.createVirtualRegister(&RISCV::GPRRegClass);
  TII->movImm(MBB, II, DL, ScratchReg, Val, Flag);
  BuildMI(MBB, II, DL, TII->get(Opc), DestReg)
      .addReg(SrcReg, getKillRegState(KillSrcReg))
      .addReg(ScratchReg, RegState::Kill)
      .setMIFlag(Flag);
}
 
static std::tuple<RISCVVType::VLMUL, const TargetRegisterClass &, unsigned>
getSpillReloadInfo(unsigned NumRemaining, uint16_t RegEncoding, bool IsSpill) {
  if (NumRemaining >= 8 && RegEncoding % 8 == 0)
    return {RISCVVType::LMUL_8, RISCV::VRM8RegClass,
            IsSpill ? RISCV::VS8R_V : RISCV::VL8RE8_V};
  if (NumRemaining >= 4 && RegEncoding % 4 == 0)
    return {RISCVVType::LMUL_4, RISCV::VRM4RegClass,
            IsSpill ? RISCV::VS4R_V : RISCV::VL4RE8_V};
  if (NumRemaining >= 2 && RegEncoding % 2 == 0)
    return {RISCVVType::LMUL_2, RISCV::VRM2RegClass,
            IsSpill ? RISCV::VS2R_V : RISCV::VL2RE8_V};
  return {RISCVVType::LMUL_1, RISCV::VRRegClass,
          IsSpill ? RISCV::VS1R_V : RISCV::VL1RE8_V};
}
 
// Split a VSPILLx_Mx/VSPILLx_Mx pseudo into multiple whole register stores
// separated by LMUL*VLENB bytes.
void RISCVRegisterInfo::lowerSegmentSpillReload(MachineBasicBlock::iterator II,
                                                bool IsSpill) const {
  DebugLoc DL = II->getDebugLoc();
  MachineBasicBlock &MBB = *II->getParent();
  MachineFunction &MF = *MBB.getParent();
  MachineRegisterInfo &MRI = MF.getRegInfo();
  const RISCVSubtarget &STI = MF.getSubtarget<RISCVSubtarget>();
  const TargetInstrInfo *TII = STI.getInstrInfo();
  const TargetRegisterInfo *TRI = STI.getRegisterInfo();
 
  auto ZvlssegInfo = RISCV::isRVVSpillForZvlsseg(II->getOpcode());
  unsigned NF = ZvlssegInfo->first;
  unsigned LMUL = ZvlssegInfo->second;
  unsigned NumRegs = NF * LMUL;
  assert(NumRegs <= 8 && "Invalid NF/LMUL combinations.");
 
  Register Reg = II->getOperand(0).getReg();
  uint16_t RegEncoding = TRI->getEncodingValue(Reg);
  Register Base = II->getOperand(1).getReg();
  bool IsBaseKill = II->getOperand(1).isKill();
  Register NewBase = MRI.createVirtualRegister(&RISCV::GPRRegClass);
 
  auto *OldMMO = *(II->memoperands_begin());
  LocationSize OldLoc = OldMMO->getSize();
  assert(OldLoc.isPrecise() && OldLoc.getValue().isKnownMultipleOf(NF));
  TypeSize VRegSize = OldLoc.getValue().divideCoefficientBy(NumRegs);
 
  Register VLENB = 0;
  unsigned VLENBShift = 0;
  unsigned PrevHandledNum = 0;
  unsigned I = 0;
  while (I != NumRegs) {
    auto [LMulHandled, RegClass, Opcode] =
        getSpillReloadInfo(NumRegs - I, RegEncoding, IsSpill);
    auto [RegNumHandled, _] = RISCVVType::decodeVLMUL(LMulHandled);
    bool IsLast = I + RegNumHandled == NumRegs;
    if (PrevHandledNum) {
      Register Step;
      // Optimize for constant VLEN.
      if (auto VLEN = STI.getRealVLen()) {
        int64_t Offset = *VLEN / 8 * PrevHandledNum;
        Step = MRI.createVirtualRegister(&RISCV::GPRRegClass);
        STI.getInstrInfo()->movImm(MBB, II, DL, Step, Offset);
      } else {
        if (!VLENB) {
          VLENB = MRI.createVirtualRegister(&RISCV::GPRRegClass);
          BuildMI(MBB, II, DL, TII->get(RISCV::PseudoReadVLENB), VLENB);
        }
        uint32_t ShiftAmount = Log2_32(PrevHandledNum);
        // To avoid using an extra register, we shift the VLENB register and
        // remember how much it has been shifted. We can then use relative
        // shifts to adjust to the desired shift amount.
        if (VLENBShift > ShiftAmount) {
          BuildMI(MBB, II, DL, TII->get(RISCV::SRLI), VLENB)
              .addReg(VLENB, RegState::Kill)
              .addImm(VLENBShift - ShiftAmount);
        } else if (VLENBShift < ShiftAmount) {
          BuildMI(MBB, II, DL, TII->get(RISCV::SLLI), VLENB)
              .addReg(VLENB, RegState::Kill)
              .addImm(ShiftAmount - VLENBShift);
        }
        VLENBShift = ShiftAmount;
        Step = VLENB;
      }
 
      BuildMI(MBB, II, DL, TII->get(RISCV::ADD), NewBase)
          .addReg(Base, getKillRegState(I != 0 || IsBaseKill))
          .addReg(Step, getKillRegState(Step != VLENB || IsLast));
      Base = NewBase;
    }
 
    MCRegister ActualReg = findVRegWithEncoding(RegClass, RegEncoding);
    MachineInstrBuilder MIB =
        BuildMI(MBB, II, DL, TII->get(Opcode))
            .addReg(ActualReg, getDefRegState(!IsSpill))
            .addReg(Base, getKillRegState(IsLast))
            .addMemOperand(MF.getMachineMemOperand(OldMMO, OldMMO->getOffset(),
                                                   VRegSize * RegNumHandled));
 
    // Adding implicit-use of super register to describe we are using part of
    // super register, that prevents machine verifier complaining when part of
    // subreg is undef, see comment in MachineVerifier::checkLiveness for more
    // detail.
    if (IsSpill)
      MIB.addReg(Reg, RegState::Implicit);
 
    PrevHandledNum = RegNumHandled;
    RegEncoding += RegNumHandled;
    I += RegNumHandled;
  }
  II->eraseFromParent();
}
 
static unsigned getXqciloWideOpcode(unsigned Opc) {
  switch (Opc) {
  case RISCV::LW:
    return RISCV::QC_E_LW;
  case RISCV::SW:
    return RISCV::QC_E_SW;
  case RISCV::LB:
    return RISCV::QC_E_LB;
  case RISCV::LBU:
    return RISCV::QC_E_LBU;
  case RISCV::LH:
    return RISCV::QC_E_LH;
  case RISCV::LHU:
    return RISCV::QC_E_LHU;
  case RISCV::SB:
    return RISCV::QC_E_SB;
  case RISCV::SH:
    return RISCV::QC_E_SH;
  default:
    return 0;
  }
}
 
bool RISCVRegisterInfo::eliminateFrameIndex(MachineBasicBlock::iterator II,
                                            int SPAdj, unsigned FIOperandNum,
                                            RegScavenger *RS) const {
  assert(SPAdj == 0 && "Unexpected non-zero SPAdj value");
 
  MachineInstr &MI = *II;
  MachineFunction &MF = *MI.getParent()->getParent();
  MachineRegisterInfo &MRI = MF.getRegInfo();
  const RISCVSubtarget &ST = MF.getSubtarget<RISCVSubtarget>();
  const RISCVInstrInfo *TII = ST.getInstrInfo();
  bool Is64Bit = ST.is64Bit();
  DebugLoc DL = MI.getDebugLoc();
 
  int FrameIndex = MI.getOperand(FIOperandNum).getIndex();
  Register FrameReg;
  StackOffset Offset =
      getFrameLowering(MF)->getFrameIndexReference(MF, FrameIndex, FrameReg);
  bool IsRVVSpill = RISCV::isRVVSpill(MI);
  if (!IsRVVSpill)
    Offset += StackOffset::getFixed(MI.getOperand(FIOperandNum + 1).getImm());
 
  if (!Is64Bit && !isInt<32>(Offset.getFixed())) {
    reportFatalUsageError("Frame offsets outside of the signed 32-bit range "
                          "not supported on RV32");
  }
 
  if (!IsRVVSpill) {
    int64_t Val = Offset.getFixed();
    int64_t Lo12 = SignExtend64<12>(Val);
    int64_t Lo26 = SignExtend64<26>(Val);
    unsigned Opc = MI.getOpcode();
 
    if (Opc == RISCV::ADDI && !isInt<12>(Val)) {
      // We chose to emit the canonical immediate sequence rather than folding
      // the offset into the using add under the theory that doing so doesn't
      // save dynamic instruction count and some target may fuse the canonical
      // 32 bit immediate sequence.  We still need to clear the portion of the
      // offset encoded in the immediate.
      MI.getOperand(FIOperandNum + 1).ChangeToImmediate(0);
    } else if ((Opc == RISCV::PREFETCH_I || Opc == RISCV::PREFETCH_R ||
                Opc == RISCV::PREFETCH_W) &&
               (Lo12 & 0b11111) != 0) {
      // Prefetch instructions require the offset to be 32 byte aligned.
      MI.getOperand(FIOperandNum + 1).ChangeToImmediate(0);
    } else if (Opc == RISCV::MIPS_PREF && !isUInt<9>(Val)) {
      // MIPS Prefetch instructions require the offset to be 9 bits encoded.
      MI.getOperand(FIOperandNum + 1).ChangeToImmediate(0);
    } else if ((Opc == RISCV::PseudoRV32ZdinxLD ||
                Opc == RISCV::PseudoRV32ZdinxSD ||
                Opc == RISCV::PseudoLD_RV32_OPT ||
                Opc == RISCV::PseudoSD_RV32_OPT) &&
               Lo12 >= 2044) {
      // This instruction will/might be split into 2 instructions. The second
      // instruction will add 4 to the immediate. If that would overflow 12
      // bits, we can't fold the offset.
      MI.getOperand(FIOperandNum + 1).ChangeToImmediate(0);
    } else if (unsigned WideOpc = getXqciloWideOpcode(Opc);
               !isInt<12>(Val) && ST.hasVendorXqcilo() && WideOpc) {
      // The resolved frame offset exceeds simm12 but the instruction is a
      // standard load/store (LW/SW/etc). Promote to the wide Xqcilo equivalent
      // so the full 26-bit offset folds directly, avoiding a separate
      // base-adjust instruction. This runs post-RA and does not affect
      // register allocation decisions.
      MI.setDesc(TII->get(WideOpc));
      MI.getOperand(FIOperandNum + 1).ChangeToImmediate(Lo26);
      Offset = StackOffset::get((uint64_t)Val - (uint64_t)Lo26,
                                Offset.getScalable());
    } else if (Opc == RISCV::QC_E_ADDI || RISCVInstrInfo::isBaseQCLoad(MI) ||
               RISCVInstrInfo::isBaseQCStore(MI)) {
      MI.getOperand(FIOperandNum + 1).ChangeToImmediate(Lo26);
      Offset = StackOffset::get((uint64_t)Val - (uint64_t)Lo26,
                                Offset.getScalable());
    } else {
      // We can encode an add with 12 bit signed immediate in the immediate
      // operand of our user instruction.  As a result, the remaining
      // offset can by construction, at worst, a LUI and a ADD.
      MI.getOperand(FIOperandNum + 1).ChangeToImmediate(Lo12);
      Offset = StackOffset::get((uint64_t)Val - (uint64_t)Lo12,
                                Offset.getScalable());
    }
  }
 
  if (Offset.getScalable() || Offset.getFixed()) {
    Register DestReg;
    if (MI.getOpcode() == RISCV::ADDI)
      DestReg = MI.getOperand(0).getReg();
    else
      DestReg = MRI.createVirtualRegister(&RISCV::GPRRegClass);
    adjustReg(*II->getParent(), II, DL, DestReg, FrameReg, Offset,
              MachineInstr::NoFlags, std::nullopt);
    MI.getOperand(FIOperandNum).ChangeToRegister(DestReg, /*IsDef*/false,
                                                 /*IsImp*/false,
                                                 /*IsKill*/true);
  } else {
    MI.getOperand(FIOperandNum).ChangeToRegister(FrameReg, /*IsDef*/false,
                                                 /*IsImp*/false,
                                                 /*IsKill*/false);
  }
 
  // If after materializing the adjustment, we have a pointless ADDI, remove it
  if (MI.getOpcode() == RISCV::ADDI &&
      MI.getOperand(0).getReg() == MI.getOperand(1).getReg() &&
      MI.getOperand(2).getImm() == 0) {
    MI.eraseFromParent();
    return true;
  }
 
  // Handle spill/fill of synthetic register classes for segment operations to
  // ensure correctness in the edge case one gets spilled.
  switch (MI.getOpcode()) {
  case RISCV::PseudoVSPILL2_M1:
  case RISCV::PseudoVSPILL2_M2:
  case RISCV::PseudoVSPILL2_M4:
  case RISCV::PseudoVSPILL3_M1:
  case RISCV::PseudoVSPILL3_M2:
  case RISCV::PseudoVSPILL4_M1:
  case RISCV::PseudoVSPILL4_M2:
  case RISCV::PseudoVSPILL5_M1:
  case RISCV::PseudoVSPILL6_M1:
  case RISCV::PseudoVSPILL7_M1:
  case RISCV::PseudoVSPILL8_M1:
    lowerSegmentSpillReload(II, /*IsSpill=*/true);
    return true;
  case RISCV::PseudoVRELOAD2_M1:
  case RISCV::PseudoVRELOAD2_M2:
  case RISCV::PseudoVRELOAD2_M4:
  case RISCV::PseudoVRELOAD3_M1:
  case RISCV::PseudoVRELOAD3_M2:
  case RISCV::PseudoVRELOAD4_M1:
  case RISCV::PseudoVRELOAD4_M2:
  case RISCV::PseudoVRELOAD5_M1:
  case RISCV::PseudoVRELOAD6_M1:
  case RISCV::PseudoVRELOAD7_M1:
  case RISCV::PseudoVRELOAD8_M1:
    lowerSegmentSpillReload(II, /*IsSpill=*/false);
    return true;
  }
 
  return false;
}
 
bool RISCVRegisterInfo::requiresVirtualBaseRegisters(
    const MachineFunction &MF) const {
  return true;
}
 
// Returns true if the instruction's frame index reference would be better
// served by a base register other than FP or SP.
// Used by LocalStackSlotAllocation pass to determine which frame index
// references it should create new base registers for.
bool RISCVRegisterInfo::needsFrameBaseReg(MachineInstr *MI,
                                          int64_t Offset) const {
  unsigned FIOperandNum = 0;
  for (; !MI->getOperand(FIOperandNum).isFI(); FIOperandNum++)
    assert(FIOperandNum < MI->getNumOperands() &&
           "Instr doesn't have FrameIndex operand");
 
  // For RISC-V, The machine instructions that include a FrameIndex operand
  // are load/store, ADDI instructions.
  unsigned MIFrm = RISCVII::getFormat(MI->getDesc().TSFlags);
  if (MIFrm != RISCVII::InstFormatI && MIFrm != RISCVII::InstFormatS)
    return false;
  // We only generate virtual base registers for loads and stores, so
  // return false for everything else.
  if (!MI->mayLoad() && !MI->mayStore())
    return false;
 
  const MachineFunction &MF = *MI->getMF();
  const MachineFrameInfo &MFI = MF.getFrameInfo();
  const RISCVFrameLowering *TFI = getFrameLowering(MF);
  const MachineRegisterInfo &MRI = MF.getRegInfo();
 
  if (TFI->hasFP(MF) && !shouldRealignStack(MF)) {
    auto &Subtarget = MF.getSubtarget<RISCVSubtarget>();
    // Estimate the stack size used to store callee saved registers(
    // excludes reserved registers).
    unsigned CalleeSavedSize = 0;
    for (const MCPhysReg *R = MRI.getCalleeSavedRegs(); MCPhysReg Reg = *R;
         ++R) {
      if (Subtarget.isRegisterReservedByUser(Reg))
        continue;
 
      if (RISCV::GPRRegClass.contains(Reg))
        CalleeSavedSize += getSpillSize(RISCV::GPRRegClass);
      else if (RISCV::FPR64RegClass.contains(Reg))
        CalleeSavedSize += getSpillSize(RISCV::FPR64RegClass);
      else if (RISCV::FPR32RegClass.contains(Reg))
        CalleeSavedSize += getSpillSize(RISCV::FPR32RegClass);
      // Ignore vector registers.
    }
 
    int64_t MaxFPOffset = Offset - CalleeSavedSize;
    if (isFrameOffsetLegal(MI, RISCV::X8, MaxFPOffset))
      return false;
 
    // If the FP-relative offset doesn't fit, fall through to check the
    // SP-relative offset. getFrameIndexReference may select SP over FP when
    // the SP offset fits in the compressed instruction immediate range, so a
    // base register might not be needed.
  }
 
  // Assume 128 bytes spill slots size to estimate the maximum possible
  // offset relative to the stack pointer.
  // FIXME: The 128 is copied from ARM. We should run some statistics and pick a
  // real one for RISC-V.
  int64_t MaxSPOffset = Offset + 128;
  MaxSPOffset += MFI.getLocalFrameSize();
  return !isFrameOffsetLegal(MI, RISCV::X2, MaxSPOffset);
}
 
// Determine whether a given base register plus offset immediate is
// encodable to resolve a frame index.
bool RISCVRegisterInfo::isFrameOffsetLegal(const MachineInstr *MI,
                                           Register BaseReg,
                                           int64_t Offset) const {
  unsigned FIOperandNum = 0;
  while (!MI->getOperand(FIOperandNum).isFI()) {
    FIOperandNum++;
    assert(FIOperandNum < MI->getNumOperands() &&
           "Instr does not have a FrameIndex operand!");
  }
 
  Offset += getFrameIndexInstrOffset(MI, FIOperandNum);
  return isInt<12>(Offset);
}
 
// Insert defining instruction(s) for a pointer to FrameIdx before
// insertion point I.
// Return materialized frame pointer.
Register RISCVRegisterInfo::materializeFrameBaseRegister(MachineBasicBlock *MBB,
                                                         int FrameIdx,
                                                         int64_t Offset) const {
  MachineBasicBlock::iterator MBBI = MBB->begin();
  DebugLoc DL;
  if (MBBI != MBB->end())
    DL = MBBI->getDebugLoc();
  MachineFunction *MF = MBB->getParent();
  MachineRegisterInfo &MFI = MF->getRegInfo();
  const TargetInstrInfo *TII = MF->getSubtarget().getInstrInfo();
 
  Register BaseReg = MFI.createVirtualRegister(&RISCV::GPRRegClass);
  BuildMI(*MBB, MBBI, DL, TII->get(RISCV::ADDI), BaseReg)
      .addFrameIndex(FrameIdx)
      .addImm(Offset);
  return BaseReg;
}
 
// Resolve a frame index operand of an instruction to reference the
// indicated base register plus offset instead.
void RISCVRegisterInfo::resolveFrameIndex(MachineInstr &MI, Register BaseReg,
                                          int64_t Offset) const {
  unsigned FIOperandNum = 0;
  while (!MI.getOperand(FIOperandNum).isFI()) {
    FIOperandNum++;
    assert(FIOperandNum < MI.getNumOperands() &&
           "Instr does not have a FrameIndex operand!");
  }
 
  Offset += getFrameIndexInstrOffset(&MI, FIOperandNum);
  // FrameIndex Operands are always represented as a
  // register followed by an immediate.
  MI.getOperand(FIOperandNum).ChangeToRegister(BaseReg, false);
  MI.getOperand(FIOperandNum + 1).ChangeToImmediate(Offset);
}
 
// Get the offset from the referenced frame index in the instruction,
// if there is one.
int64_t RISCVRegisterInfo::getFrameIndexInstrOffset(const MachineInstr *MI,
                                                    int Idx) const {
  assert((RISCVII::getFormat(MI->getDesc().TSFlags) == RISCVII::InstFormatI ||
          RISCVII::getFormat(MI->getDesc().TSFlags) == RISCVII::InstFormatS) &&
         "The MI must be I or S format.");
  assert(MI->getOperand(Idx).isFI() && "The Idx'th operand of MI is not a "
                                       "FrameIndex operand");
  return MI->getOperand(Idx + 1).getImm();
}
 
Register RISCVRegisterInfo::getFrameRegister(const MachineFunction &MF) const {
  const TargetFrameLowering *TFI = getFrameLowering(MF);
  return TFI->hasFP(MF) ? RISCV::X8 : RISCV::X2;
}
 
bool RISCVRegisterInfo::isArgumentRegister(const MachineFunction &MF,
                                           MCRegister Reg) const {
  auto const &STI = MF.getSubtarget<RISCVSubtarget>();
  const RISCVRegisterInfo *TRI = STI.getRegisterInfo();
 
  if (TRI->isGeneralPurposeRegister(MF, Reg))
    return llvm::is_contained(RISCV::getArgGPRs(STI), Reg);
 
  if (TRI->isFPRegister(Reg))
    return llvm::is_contained(RISCV::getArgFPRs(STI), Reg);
 
  if (RISCV::VRRegClass.contains(Reg))
    return llvm::is_contained(RISCV::getArgVRs(STI), Reg);
 
  return false;
}
 
StringRef RISCVRegisterInfo::getRegAsmName(MCRegister Reg) const {
  if (Reg == RISCV::SF_VCIX_STATE)
    return "sf.vcix_state";
  return TargetRegisterInfo::getRegAsmName(Reg);
}
 
const uint32_t *
RISCVRegisterInfo::getCallPreservedMask(const MachineFunction & MF,
                                        CallingConv::ID CC) const {
  auto &Subtarget = MF.getSubtarget<RISCVSubtarget>();
 
  if (CC == CallingConv::GHC)
    return CSR_NoRegs_RegMask;
  RISCVABI::ABI ABI = Subtarget.getTargetABI();
  assert(ABI != RISCVABI::ABI::Unknown && !RISCVABI::isPurecap(ABI) &&
         "Unrecognized ABI");
  if (CC == CallingConv::PreserveMost) {
    if (RISCVABI::isRVE(ABI))
      return CSR_RT_MostRegs_RVE_RegMask;
    return CSR_RT_MostRegs_RegMask;
  }
  if (RISCVABI::isRVE(ABI))
    return CSR_ILP32E_LP64E_RegMask;
  bool IsVectorCall = CC == CallingConv::RISCV_VectorCall;
  if (RISCVABI::isSoftFP(ABI))
    return IsVectorCall ? CSR_ILP32_LP64_V_RegMask : CSR_ILP32_LP64_RegMask;
  if (RISCVABI::isSingleFP(ABI))
    return IsVectorCall ? CSR_ILP32F_LP64F_V_RegMask : CSR_ILP32F_LP64F_RegMask;
  assert(RISCVABI::isDoubleFP(ABI) && "Unrecognized ABI");
  return IsVectorCall ? CSR_ILP32D_LP64D_V_RegMask : CSR_ILP32D_LP64D_RegMask;
}
 
const TargetRegisterClass *
RISCVRegisterInfo::getLargestLegalSuperClass(const TargetRegisterClass *RC,
                                             const MachineFunction &) const {
  if (RC == &RISCV::VMV0RegClass)
    return &RISCV::VRRegClass;
  if (RC == &RISCV::VRNoV0RegClass)
    return &RISCV::VRRegClass;
  if (RC == &RISCV::VRM2NoV0RegClass)
    return &RISCV::VRM2RegClass;
  if (RC == &RISCV::VRM4NoV0RegClass)
    return &RISCV::VRM4RegClass;
  if (RC == &RISCV::VRM8NoV0RegClass)
    return &RISCV::VRM8RegClass;
  return RC;
}
 
void RISCVRegisterInfo::getOffsetOpcodes(const StackOffset &Offset,
                                         SmallVectorImpl<uint64_t> &Ops) const {
  // VLENB is the length of a vector register in bytes. We use <vscale x 8 x i8>
  // to represent one vector register. The dwarf offset is
  // VLENB * scalable_offset / 8.
  assert(Offset.getScalable() % 8 == 0 && "Invalid frame offset");
 
  // Add fixed-sized offset using existing DIExpression interface.
  DIExpression::appendOffset(Ops, Offset.getFixed());
 
  unsigned VLENB = getDwarfRegNum(RISCV::VLENB, true);
  int64_t VLENBSized = Offset.getScalable() / 8;
  if (VLENBSized > 0) {
    Ops.push_back(dwarf::DW_OP_constu);
    Ops.push_back(VLENBSized);
    Ops.append({dwarf::DW_OP_bregx, VLENB, 0ULL});
    Ops.push_back(dwarf::DW_OP_mul);
    Ops.push_back(dwarf::DW_OP_plus);
  } else if (VLENBSized < 0) {
    Ops.push_back(dwarf::DW_OP_constu);
    Ops.push_back(-VLENBSized);
    Ops.append({dwarf::DW_OP_bregx, VLENB, 0ULL});
    Ops.push_back(dwarf::DW_OP_mul);
    Ops.push_back(dwarf::DW_OP_minus);
  }
}
 
std::pair<Register, Register>
RISCVRegisterInfo::getGPRPairSubRegs(MCRegister Pair) const {
  if (Pair == RISCV::X0_Pair)
    return {RISCV::X0, RISCV::X0};
  return {getSubReg(Pair, RISCV::sub_gpr_even),
          getSubReg(Pair, RISCV::sub_gpr_odd)};
}
 
unsigned
RISCVRegisterInfo::getRegisterCostTableIndex(const MachineFunction &MF) const {
  // Set CostPerUse to 1 only when optimizing for size and RVC exists.
  const auto &ST = MF.getSubtarget<RISCVSubtarget>();
  return MF.getFunction().hasOptSize() && ST.hasStdExtZca() &&
                 ST.getCLOpts().cost_per_use
             ? 1
             : 0;
}
 
float RISCVRegisterInfo::getSpillWeightScaleFactor(
    const TargetRegisterClass *RC) const {
  return getRegClassWeight(RC).RegWeight;
}
 
// Add two address hints to improve chances of being able to use a compressed
// instruction.
bool RISCVRegisterInfo::getRegAllocationHints(
    Register VirtReg, ArrayRef<MCPhysReg> Order,
    SmallSetVector<MCPhysReg, 16> &Hints, const MachineFunction &MF,
    const VirtRegMap *VRM, const LiveRegMatrix *Matrix) const {
  const MachineRegisterInfo *MRI = &MF.getRegInfo();
  auto &Subtarget = MF.getSubtarget<RISCVSubtarget>();
 
  // Handle RegPairEven/RegPairOdd hints for Zilsd register pairs
  std::pair<unsigned, Register> Hint = MRI->getRegAllocationHint(VirtReg);
  unsigned HintType = Hint.first;
  Register Partner = Hint.second;
 
  MCRegister TargetReg;
  if (HintType == RISCVRI::RegPairEven || HintType == RISCVRI::RegPairOdd) {
    // Check if we want the even or odd register of a consecutive pair
    bool WantOdd = (HintType == RISCVRI::RegPairOdd);
 
    // First priority: Check if partner is already allocated
    if (Partner.isVirtual() && VRM && VRM->hasPhys(Partner)) {
      MCRegister PartnerPhys = VRM->getPhys(Partner);
      // Calculate the exact register we need for consecutive pairing
      TargetReg = PartnerPhys.id() + (WantOdd ? 1 : -1);
 
      // Verify it's valid and available
      if (RISCV::GPRRegClass.contains(TargetReg) &&
          is_contained(Order, TargetReg))
        Hints.insert(TargetReg.id());
    }
 
    // Second priority: Try to find consecutive register pairs in the allocation
    // order
    for (MCPhysReg PhysReg : Order) {
      // Don't add the hint if we already added above.
      if (TargetReg == PhysReg)
        continue;
 
      unsigned RegNum = getEncodingValue(PhysReg);
      // Check if this register matches the even/odd requirement
      bool IsOdd = (RegNum % 2 != 0);
 
      // Don't provide hints that are paired to a reserved register.
      MCRegister Paired = PhysReg + (IsOdd ? -1 : 1);
      if (WantOdd == IsOdd && !MRI->isReserved(Paired))
        Hints.insert(PhysReg);
    }
  }
 
  bool BaseImplRetVal = TargetRegisterInfo::getRegAllocationHints(
      VirtReg, Order, Hints, MF, VRM, Matrix);
 
  if (!VRM || !MF.getSubtarget<RISCVSubtarget>().getCLOpts().regalloc_hints)
    return BaseImplRetVal;
 
  // Add any two address hints after any copy hints.
  SmallSet<Register, 4> TwoAddrHints;
 
  auto tryAddHint = [&](const MachineOperand &VRRegMO, const MachineOperand &MO,
                        bool NeedGPRC) -> void {
    Register Reg = MO.getReg();
    Register PhysReg = Reg.isPhysical() ? Reg : Register(VRM->getPhys(Reg));
    // TODO: Support GPRPair subregisters? Need to be careful with even/odd
    // registers. If the virtual register is an odd register of a pair and the
    // physical register is even (or vice versa), we should not add the hint.
    if (PhysReg && (!NeedGPRC || RISCV::GPRCRegClass.contains(PhysReg)) &&
        !MO.getSubReg() && !VRRegMO.getSubReg()) {
      if (!MRI->isReserved(PhysReg) && !Hints.contains(PhysReg))
        TwoAddrHints.insert(PhysReg);
    }
  };
 
  // This is all of the compressible binary instructions. If an instruction
  // needs GPRC register class operands \p NeedGPRC will be set to true.
  auto isCompressible = [&Subtarget](const MachineInstr &MI, bool &NeedGPRC) {
    NeedGPRC = false;
    switch (MI.getOpcode()) {
    default:
      return false;
    case RISCV::AND:
    case RISCV::OR:
    case RISCV::XOR:
    case RISCV::SUB:
    case RISCV::ADDW:
    case RISCV::SUBW:
      NeedGPRC = true;
      return true;
    case RISCV::ANDI: {
      NeedGPRC = true;
      if (!MI.getOperand(2).isImm())
        return false;
      int64_t Imm = MI.getOperand(2).getImm();
      if (isInt<6>(Imm))
        return true;
      // c.zext.b
      return Subtarget.hasStdExtZcb() && Imm == 255;
    }
    case RISCV::SRAI:
    case RISCV::SRLI:
      NeedGPRC = true;
      return true;
    case RISCV::ADD:
    case RISCV::SLLI:
      return true;
    case RISCV::ADDI:
    case RISCV::ADDIW:
      return MI.getOperand(2).isImm() && isInt<6>(MI.getOperand(2).getImm());
    case RISCV::MUL:
      // c.mul
      NeedGPRC = true;
      return Subtarget.hasStdExtZcb();
    case RISCV::SEXT_B:
    case RISCV::SEXT_H:
    case RISCV::ZEXT_H_RV32:
    case RISCV::ZEXT_H_RV64:
      // c.sext.b, c.sext.h, c.zext.h
      NeedGPRC = true;
      return Subtarget.hasStdExtZcb() && Subtarget.hasStdExtZbb();
    case RISCV::ADD_UW:
      // c.zext.w
      NeedGPRC = true;
      return Subtarget.hasStdExtZcb() && MI.getOperand(2).isReg() &&
             MI.getOperand(2).getReg() == RISCV::X0;
    case RISCV::XORI:
      // c.not
      NeedGPRC = true;
      return Subtarget.hasStdExtZcb() && MI.getOperand(2).isImm() &&
             MI.getOperand(2).getImm() == -1;
    case RISCV::QC_EXTU:
      return MI.getOperand(2).getImm() >= 6 && MI.getOperand(3).getImm() == 0;
    case RISCV::BSETI:
    case RISCV::BEXTI:
      // qc.c.bseti, qc.c.bexti
      NeedGPRC = true;
      return Subtarget.hasVendorXqcibm() && MI.getOperand(2).getImm() != 0;
    }
  };
 
  // Returns true if this operand is compressible. For non-registers it always
  // returns true. Immediate range was already checked in isCompressible.
  // For registers, it checks if the register is a GPRC register. reg-reg
  // instructions that require GPRC need all register operands to be GPRC.
  auto isCompressibleOpnd = [&](const MachineOperand &MO) {
    if (!MO.isReg())
      return true;
    Register Reg = MO.getReg();
    Register PhysReg = Reg.isPhysical() ? Reg : Register(VRM->getPhys(Reg));
    return PhysReg && RISCV::GPRCRegClass.contains(PhysReg);
  };
 
  for (auto &MO : MRI->reg_nodbg_operands(VirtReg)) {
    const MachineInstr &MI = *MO.getParent();
    unsigned OpIdx = MO.getOperandNo();
    bool NeedGPRC;
    if (isCompressible(MI, NeedGPRC)) {
      if (OpIdx == 0 && MI.getOperand(1).isReg()) {
        if (!NeedGPRC || MI.getNumExplicitOperands() < 3 ||
            MI.getOpcode() == RISCV::ADD_UW ||
            isCompressibleOpnd(MI.getOperand(2)))
          tryAddHint(MO, MI.getOperand(1), NeedGPRC);
        if (MI.isCommutable() && MI.getOperand(2).isReg() &&
            (!NeedGPRC || isCompressibleOpnd(MI.getOperand(1))))
          tryAddHint(MO, MI.getOperand(2), NeedGPRC);
      } else if (OpIdx == 1 && (!NeedGPRC || MI.getNumExplicitOperands() < 3 ||
                                isCompressibleOpnd(MI.getOperand(2)))) {
        tryAddHint(MO, MI.getOperand(0), NeedGPRC);
      } else if (MI.isCommutable() && OpIdx == 2 &&
                 (!NeedGPRC || isCompressibleOpnd(MI.getOperand(1)))) {
        tryAddHint(MO, MI.getOperand(0), NeedGPRC);
      }
    }
 
    // Add a hint if it would allow auipc/lui+addi(w) fusion.  We do this even
    // without the fusions explicitly enabled as the impact is rarely negative
    // and some cores do implement this fusion.
    if ((MI.getOpcode() == RISCV::ADDIW || MI.getOpcode() == RISCV::ADDI) &&
        MI.getOperand(1).isReg()) {
      const MachineBasicBlock &MBB = *MI.getParent();
      MachineBasicBlock::const_iterator I = MI.getIterator();
      // Is the previous instruction a LUI or AUIPC that can be fused?
      if (I != MBB.begin()) {
        I = skipDebugInstructionsBackward(std::prev(I), MBB.begin());
        if ((I->getOpcode() == RISCV::LUI || I->getOpcode() == RISCV::AUIPC) &&
            I->getOperand(0).getReg() == MI.getOperand(1).getReg()) {
          if (OpIdx == 0)
            tryAddHint(MO, MI.getOperand(1), /*NeedGPRC=*/false);
          else
            tryAddHint(MO, MI.getOperand(0), /*NeedGPRC=*/false);
        }
      }
    }
  }
 
  for (MCPhysReg OrderReg : Order)
    if (TwoAddrHints.count(OrderReg))
      Hints.insert(OrderReg);
 
  return BaseImplRetVal;
}
 
void RISCVRegisterInfo::updateRegAllocHint(Register Reg, Register NewReg,
                                           MachineFunction &MF) const {
  MachineRegisterInfo *MRI = &MF.getRegInfo();
  std::pair<unsigned, Register> Hint = MRI->getRegAllocationHint(Reg);
 
  // Handle RegPairEven/RegPairOdd hints for Zilsd register pairs
  if ((Hint.first == RISCVRI::RegPairOdd ||
       Hint.first == RISCVRI::RegPairEven) &&
      Hint.second.isVirtual()) {
    // If 'Reg' is one of the even/odd register pair and it's now changed
    // (e.g. coalesced) into a different register, the other register of the
    // pair allocation hint must be updated to reflect the relationship change.
    Register Partner = Hint.second;
    std::pair<unsigned, Register> PartnerHint =
        MRI->getRegAllocationHint(Partner);
 
    // Make sure partner still points to us
    if (PartnerHint.second == Reg) {
      // Update partner to point to NewReg instead of Reg
      MRI->setRegAllocationHint(Partner, PartnerHint.first, NewReg);
 
      // If NewReg is virtual, set up the reciprocal hint
      // NewReg takes over Reg's role, so it gets the SAME hint type as Reg
      if (NewReg.isVirtual())
        MRI->setRegAllocationHint(NewReg, Hint.first, Partner);
    }
  }
}
 
MCRegister
RISCVRegisterInfo::findVRegWithEncoding(const TargetRegisterClass &RegClass,
                                        uint16_t Encoding) const {
  MCRegister Reg = RISCV::V0 + Encoding;
  if (RISCVRI::getLMul(RegClass.TSFlags) == RISCVVType::LMUL_1)
    return Reg;
  return getMatchingSuperReg(Reg, RISCV::sub_vrm1_0, &RegClass);
}