Target/AMDGPU/AMDGPUInstructionSelector.cpp

//===- AMDGPUInstructionSelector.cpp ----------------------------*- 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
//
//===----------------------------------------------------------------------===//
/// \file
/// This file implements the targeting of the InstructionSelector class for
/// AMDGPU.
/// \todo This should be generated by TableGen.
//===----------------------------------------------------------------------===//

#include "AMDGPUInstructionSelector.h"
#include "AMDGPUInstrInfo.h"
#include "AMDGPUGlobalISelUtils.h"
#include "AMDGPURegisterBankInfo.h"
#include "AMDGPURegisterInfo.h"
#include "AMDGPUSubtarget.h"
#include "AMDGPUTargetMachine.h"
#include "MCTargetDesc/AMDGPUMCTargetDesc.h"
#include "SIMachineFunctionInfo.h"
#include "llvm/CodeGen/GlobalISel/GISelKnownBits.h"
#include "llvm/CodeGen/GlobalISel/InstructionSelector.h"
#include "llvm/CodeGen/GlobalISel/InstructionSelectorImpl.h"
#include "llvm/CodeGen/GlobalISel/MachineIRBuilder.h"
#include "llvm/CodeGen/GlobalISel/MIPatternMatch.h"
#include "llvm/CodeGen/GlobalISel/Utils.h"
#include "llvm/CodeGen/MachineBasicBlock.h"
#include "llvm/CodeGen/MachineFunction.h"
#include "llvm/CodeGen/MachineInstr.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/IR/Type.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/raw_ostream.h"

#define DEBUG_TYPE "amdgpu-isel"

using namespace llvm;
using namespace MIPatternMatch;

#define GET_GLOBALISEL_IMPL
#define AMDGPUSubtarget GCNSubtarget
#include "AMDGPUGenGlobalISel.inc"
#undef GET_GLOBALISEL_IMPL
#undef AMDGPUSubtarget

AMDGPUInstructionSelector::AMDGPUInstructionSelector(
    const GCNSubtarget &STI, const AMDGPURegisterBankInfo &RBI,
    const AMDGPUTargetMachine &TM)
    : InstructionSelector(), TII(*STI.getInstrInfo()),
      TRI(*STI.getRegisterInfo()), RBI(RBI), TM(TM),
      STI(STI),
      EnableLateStructurizeCFG(AMDGPUTargetMachine::EnableLateStructurizeCFG),
#define GET_GLOBALISEL_PREDICATES_INIT
#include "AMDGPUGenGlobalISel.inc"
#undef GET_GLOBALISEL_PREDICATES_INIT
#define GET_GLOBALISEL_TEMPORARIES_INIT
#include "AMDGPUGenGlobalISel.inc"
#undef GET_GLOBALISEL_TEMPORARIES_INIT
{
}

const char *AMDGPUInstructionSelector::getName() { return DEBUG_TYPE; }

void AMDGPUInstructionSelector::setupMF(MachineFunction &MF, GISelKnownBits &KB,
                                        CodeGenCoverage &CoverageInfo) {
  MRI = &MF.getRegInfo();
  InstructionSelector::setupMF(MF, KB, CoverageInfo);
}

bool AMDGPUInstructionSelector::isVCC(Register Reg,
                                      const MachineRegisterInfo &MRI) const {
  if (Register::isPhysicalRegister(Reg))
    return Reg == TRI.getVCC();

  auto &RegClassOrBank = MRI.getRegClassOrRegBank(Reg);
  const TargetRegisterClass *RC =
      RegClassOrBank.dyn_cast<const TargetRegisterClass*>();
  if (RC) {
    const LLT Ty = MRI.getType(Reg);
    return RC->hasSuperClassEq(TRI.getBoolRC()) &&
           Ty.isValid() && Ty.getSizeInBits() == 1;
  }

  const RegisterBank *RB = RegClassOrBank.get<const RegisterBank *>();
  return RB->getID() == AMDGPU::VCCRegBankID;
}

bool AMDGPUInstructionSelector::selectCOPY(MachineInstr &I) const {
  const DebugLoc &DL = I.getDebugLoc();
  MachineBasicBlock *BB = I.getParent();
  I.setDesc(TII.get(TargetOpcode::COPY));

  const MachineOperand &Src = I.getOperand(1);
  MachineOperand &Dst = I.getOperand(0);
  Register DstReg = Dst.getReg();
  Register SrcReg = Src.getReg();

  if (isVCC(DstReg, *MRI)) {
    if (SrcReg == AMDGPU::SCC) {
      const TargetRegisterClass *RC
        = TRI.getConstrainedRegClassForOperand(Dst, *MRI);
      if (!RC)
        return true;
      return RBI.constrainGenericRegister(DstReg, *RC, *MRI);
    }

    if (!isVCC(SrcReg, *MRI)) {
      // TODO: Should probably leave the copy and let copyPhysReg expand it.
      if (!RBI.constrainGenericRegister(DstReg, *TRI.getBoolRC(), *MRI))
        return false;

      const TargetRegisterClass *SrcRC
        = TRI.getConstrainedRegClassForOperand(Src, *MRI);

      Register MaskedReg = MRI->createVirtualRegister(SrcRC);

      // We can't trust the high bits at this point, so clear them.

      // TODO: Skip masking high bits if def is known boolean.

      unsigned AndOpc = TRI.isSGPRClass(SrcRC) ?
        AMDGPU::S_AND_B32 : AMDGPU::V_AND_B32_e32;
      BuildMI(*BB, &I, DL, TII.get(AndOpc), MaskedReg)
        .addImm(1)
        .addReg(SrcReg);
      BuildMI(*BB, &I, DL, TII.get(AMDGPU::V_CMP_NE_U32_e64), DstReg)
        .addImm(0)
        .addReg(MaskedReg);

      if (!MRI->getRegClassOrNull(SrcReg))
        MRI->setRegClass(SrcReg, SrcRC);
      I.eraseFromParent();
      return true;
    }

    const TargetRegisterClass *RC =
      TRI.getConstrainedRegClassForOperand(Dst, *MRI);
    if (RC && !RBI.constrainGenericRegister(DstReg, *RC, *MRI))
      return false;

    // Don't constrain the source register to a class so the def instruction
    // handles it (unless it's undef).
    //
    // FIXME: This is a hack. When selecting the def, we neeed to know
    // specifically know that the result is VCCRegBank, and not just an SGPR
    // with size 1. An SReg_32 with size 1 is ambiguous with wave32.
    if (Src.isUndef()) {
      const TargetRegisterClass *SrcRC =
        TRI.getConstrainedRegClassForOperand(Src, *MRI);
      if (SrcRC && !RBI.constrainGenericRegister(SrcReg, *SrcRC, *MRI))
        return false;
    }

    return true;
  }

  for (const MachineOperand &MO : I.operands()) {
    if (Register::isPhysicalRegister(MO.getReg()))
      continue;

    const TargetRegisterClass *RC =
            TRI.getConstrainedRegClassForOperand(MO, *MRI);
    if (!RC)
      continue;
    RBI.constrainGenericRegister(MO.getReg(), *RC, *MRI);
  }
  return true;
}

bool AMDGPUInstructionSelector::selectPHI(MachineInstr &I) const {
  const Register DefReg = I.getOperand(0).getReg();
  const LLT DefTy = MRI->getType(DefReg);

  // TODO: Verify this doesn't have insane operands (i.e. VGPR to SGPR copy)

  const RegClassOrRegBank &RegClassOrBank =
    MRI->getRegClassOrRegBank(DefReg);

  const TargetRegisterClass *DefRC
    = RegClassOrBank.dyn_cast<const TargetRegisterClass *>();
  if (!DefRC) {
    if (!DefTy.isValid()) {
      LLVM_DEBUG(dbgs() << "PHI operand has no type, not a gvreg?\n");
      return false;
    }

    const RegisterBank &RB = *RegClassOrBank.get<const RegisterBank *>();
    DefRC = TRI.getRegClassForTypeOnBank(DefTy, RB, *MRI);
    if (!DefRC) {
      LLVM_DEBUG(dbgs() << "PHI operand has unexpected size/bank\n");
      return false;
    }
  }

  // TODO: Verify that all registers have the same bank
  I.setDesc(TII.get(TargetOpcode::PHI));
  return RBI.constrainGenericRegister(DefReg, *DefRC, *MRI);
}

MachineOperand
AMDGPUInstructionSelector::getSubOperand64(MachineOperand &MO,
                                           const TargetRegisterClass &SubRC,
                                           unsigned SubIdx) const {

  MachineInstr *MI = MO.getParent();
  MachineBasicBlock *BB = MO.getParent()->getParent();
  Register DstReg = MRI->createVirtualRegister(&SubRC);

  if (MO.isReg()) {
    unsigned ComposedSubIdx = TRI.composeSubRegIndices(MO.getSubReg(), SubIdx);
    Register Reg = MO.getReg();
    BuildMI(*BB, MI, MI->getDebugLoc(), TII.get(AMDGPU::COPY), DstReg)
            .addReg(Reg, 0, ComposedSubIdx);

    return MachineOperand::CreateReg(DstReg, MO.isDef(), MO.isImplicit(),
                                     MO.isKill(), MO.isDead(), MO.isUndef(),
                                     MO.isEarlyClobber(), 0, MO.isDebug(),
                                     MO.isInternalRead());
  }

  assert(MO.isImm());

  APInt Imm(64, MO.getImm());

  switch (SubIdx) {
  default:
    llvm_unreachable("do not know to split immediate with this sub index.");
  case AMDGPU::sub0:
    return MachineOperand::CreateImm(Imm.getLoBits(32).getSExtValue());
  case AMDGPU::sub1:
    return MachineOperand::CreateImm(Imm.getHiBits(32).getSExtValue());
  }
}

static unsigned getLogicalBitOpcode(unsigned Opc, bool Is64) {
  switch (Opc) {
  case AMDGPU::G_AND:
    return Is64 ? AMDGPU::S_AND_B64 : AMDGPU::S_AND_B32;
  case AMDGPU::G_OR:
    return Is64 ? AMDGPU::S_OR_B64 : AMDGPU::S_OR_B32;
  case AMDGPU::G_XOR:
    return Is64 ? AMDGPU::S_XOR_B64 : AMDGPU::S_XOR_B32;
  default:
    llvm_unreachable("not a bit op");
  }
}

bool AMDGPUInstructionSelector::selectG_AND_OR_XOR(MachineInstr &I) const {
  MachineOperand &Dst = I.getOperand(0);
  MachineOperand &Src0 = I.getOperand(1);
  MachineOperand &Src1 = I.getOperand(2);
  Register DstReg = Dst.getReg();
  unsigned Size = RBI.getSizeInBits(DstReg, *MRI, TRI);

  const RegisterBank *DstRB = RBI.getRegBank(DstReg, *MRI, TRI);
  if (DstRB->getID() == AMDGPU::VCCRegBankID) {
    const TargetRegisterClass *RC = TRI.getBoolRC();
    unsigned InstOpc = getLogicalBitOpcode(I.getOpcode(),
                                           RC == &AMDGPU::SReg_64RegClass);
    I.setDesc(TII.get(InstOpc));

    // FIXME: Hack to avoid turning the register bank into a register class.
    // The selector for G_ICMP relies on seeing the register bank for the result
    // is VCC. In wave32 if we constrain the registers to SReg_32 here, it will
    // be ambiguous whether it's a scalar or vector bool.
    if (Src0.isUndef() && !MRI->getRegClassOrNull(Src0.getReg()))
      MRI->setRegClass(Src0.getReg(), RC);
    if (Src1.isUndef() && !MRI->getRegClassOrNull(Src1.getReg()))
      MRI->setRegClass(Src1.getReg(), RC);

    return RBI.constrainGenericRegister(DstReg, *RC, *MRI);
  }

  // TODO: Should this allow an SCC bank result, and produce a copy from SCC for
  // the result?
  if (DstRB->getID() == AMDGPU::SGPRRegBankID) {
    unsigned InstOpc = getLogicalBitOpcode(I.getOpcode(), Size > 32);
    I.setDesc(TII.get(InstOpc));
    // Dead implicit-def of scc
    I.addOperand(MachineOperand::CreateReg(AMDGPU::SCC, true, // isDef
                                           true, // isImp
                                           false, // isKill
                                           true)); // isDead
    return constrainSelectedInstRegOperands(I, TII, TRI, RBI);
  }

  return false;
}

bool AMDGPUInstructionSelector::selectG_ADD_SUB(MachineInstr &I) const {
  MachineBasicBlock *BB = I.getParent();
  MachineFunction *MF = BB->getParent();
  Register DstReg = I.getOperand(0).getReg();
  const DebugLoc &DL = I.getDebugLoc();
  unsigned Size = RBI.getSizeInBits(DstReg, *MRI, TRI);
  const RegisterBank *DstRB = RBI.getRegBank(DstReg, *MRI, TRI);
  const bool IsSALU = DstRB->getID() == AMDGPU::SGPRRegBankID;
  const bool Sub = I.getOpcode() == TargetOpcode::G_SUB;

  if (Size == 32) {
    if (IsSALU) {
      const unsigned Opc = Sub ? AMDGPU::S_SUB_U32 : AMDGPU::S_ADD_U32;
      MachineInstr *Add =
        BuildMI(*BB, &I, DL, TII.get(Opc), DstReg)
        .add(I.getOperand(1))
        .add(I.getOperand(2));
      I.eraseFromParent();
      return constrainSelectedInstRegOperands(*Add, TII, TRI, RBI);
    }

    if (STI.hasAddNoCarry()) {
      const unsigned Opc = Sub ? AMDGPU::V_SUB_U32_e64 : AMDGPU::V_ADD_U32_e64;
      I.setDesc(TII.get(Opc));
      I.addOperand(*MF, MachineOperand::CreateImm(0));
      I.addOperand(*MF, MachineOperand::CreateReg(AMDGPU::EXEC, false, true));
      return constrainSelectedInstRegOperands(I, TII, TRI, RBI);
    }

    const unsigned Opc = Sub ? AMDGPU::V_SUB_I32_e64 : AMDGPU::V_ADD_I32_e64;

    Register UnusedCarry = MRI->createVirtualRegister(TRI.getWaveMaskRegClass());
    MachineInstr *Add
      = BuildMI(*BB, &I, DL, TII.get(Opc), DstReg)
      .addDef(UnusedCarry, RegState::Dead)
      .add(I.getOperand(1))
      .add(I.getOperand(2))
      .addImm(0);
    I.eraseFromParent();
    return constrainSelectedInstRegOperands(*Add, TII, TRI, RBI);
  }

  assert(!Sub && "illegal sub should not reach here");

  const TargetRegisterClass &RC
    = IsSALU ? AMDGPU::SReg_64_XEXECRegClass : AMDGPU::VReg_64RegClass;
  const TargetRegisterClass &HalfRC
    = IsSALU ? AMDGPU::SReg_32RegClass : AMDGPU::VGPR_32RegClass;

  MachineOperand Lo1(getSubOperand64(I.getOperand(1), HalfRC, AMDGPU::sub0));
  MachineOperand Lo2(getSubOperand64(I.getOperand(2), HalfRC, AMDGPU::sub0));
  MachineOperand Hi1(getSubOperand64(I.getOperand(1), HalfRC, AMDGPU::sub1));
  MachineOperand Hi2(getSubOperand64(I.getOperand(2), HalfRC, AMDGPU::sub1));

  Register DstLo = MRI->createVirtualRegister(&HalfRC);
  Register DstHi = MRI->createVirtualRegister(&HalfRC);

  if (IsSALU) {
    BuildMI(*BB, &I, DL, TII.get(AMDGPU::S_ADD_U32), DstLo)
      .add(Lo1)
      .add(Lo2);
    BuildMI(*BB, &I, DL, TII.get(AMDGPU::S_ADDC_U32), DstHi)
      .add(Hi1)
      .add(Hi2);
  } else {
    const TargetRegisterClass *CarryRC = TRI.getWaveMaskRegClass();
    Register CarryReg = MRI->createVirtualRegister(CarryRC);
    BuildMI(*BB, &I, DL, TII.get(AMDGPU::V_ADD_I32_e64), DstLo)
      .addDef(CarryReg)
      .add(Lo1)
      .add(Lo2)
      .addImm(0);
    MachineInstr *Addc = BuildMI(*BB, &I, DL, TII.get(AMDGPU::V_ADDC_U32_e64), DstHi)
      .addDef(MRI->createVirtualRegister(CarryRC), RegState::Dead)
      .add(Hi1)
      .add(Hi2)
      .addReg(CarryReg, RegState::Kill)
      .addImm(0);

    if (!constrainSelectedInstRegOperands(*Addc, TII, TRI, RBI))
      return false;
  }

  BuildMI(*BB, &I, DL, TII.get(AMDGPU::REG_SEQUENCE), DstReg)
    .addReg(DstLo)
    .addImm(AMDGPU::sub0)
    .addReg(DstHi)
    .addImm(AMDGPU::sub1);


  if (!RBI.constrainGenericRegister(DstReg, RC, *MRI))
    return false;

  I.eraseFromParent();
  return true;
}

bool AMDGPUInstructionSelector::selectG_UADDO_USUBO_UADDE_USUBE(
  MachineInstr &I) const {
  MachineBasicBlock *BB = I.getParent();
  MachineFunction *MF = BB->getParent();
  const DebugLoc &DL = I.getDebugLoc();
  Register Dst0Reg = I.getOperand(0).getReg();
  Register Dst1Reg = I.getOperand(1).getReg();
  const bool IsAdd = I.getOpcode() == AMDGPU::G_UADDO ||
                     I.getOpcode() == AMDGPU::G_UADDE;
  const bool HasCarryIn = I.getOpcode() == AMDGPU::G_UADDE ||
                          I.getOpcode() == AMDGPU::G_USUBE;

  if (isVCC(Dst1Reg, *MRI)) {
      // The name of the opcodes are misleading. v_add_i32/v_sub_i32 have unsigned
      // carry out despite the _i32 name. These were renamed in VI to _U32.
      // FIXME: We should probably rename the opcodes here.
    unsigned NoCarryOpc = IsAdd ? AMDGPU::V_ADD_I32_e64 : AMDGPU::V_SUB_I32_e64;
    unsigned CarryOpc = IsAdd ? AMDGPU::V_ADDC_U32_e64 : AMDGPU::V_SUBB_U32_e64;
    I.setDesc(TII.get(HasCarryIn ? CarryOpc : NoCarryOpc));
    I.addOperand(*MF, MachineOperand::CreateReg(AMDGPU::EXEC, false, true));
    I.addOperand(*MF, MachineOperand::CreateImm(0));
    return constrainSelectedInstRegOperands(I, TII, TRI, RBI);
  }

  Register Src0Reg = I.getOperand(2).getReg();
  Register Src1Reg = I.getOperand(3).getReg();

  if (HasCarryIn) {
    BuildMI(*BB, &I, DL, TII.get(AMDGPU::COPY), AMDGPU::SCC)
      .addReg(I.getOperand(4).getReg());
  }

  unsigned NoCarryOpc = IsAdd ? AMDGPU::S_ADD_U32 : AMDGPU::S_SUB_U32;
  unsigned CarryOpc = IsAdd ? AMDGPU::S_ADDC_U32 : AMDGPU::S_SUBB_U32;

  BuildMI(*BB, &I, DL, TII.get(HasCarryIn ? CarryOpc : NoCarryOpc), Dst0Reg)
    .add(I.getOperand(2))
    .add(I.getOperand(3));
  BuildMI(*BB, &I, DL, TII.get(AMDGPU::COPY), Dst1Reg)
    .addReg(AMDGPU::SCC);

  if (!MRI->getRegClassOrNull(Dst1Reg))
    MRI->setRegClass(Dst1Reg, &AMDGPU::SReg_32RegClass);

  if (!RBI.constrainGenericRegister(Dst0Reg, AMDGPU::SReg_32RegClass, *MRI) ||
      !RBI.constrainGenericRegister(Src0Reg, AMDGPU::SReg_32RegClass, *MRI) ||
      !RBI.constrainGenericRegister(Src1Reg, AMDGPU::SReg_32RegClass, *MRI))
    return false;

  if (HasCarryIn &&
      !RBI.constrainGenericRegister(I.getOperand(4).getReg(),
                                    AMDGPU::SReg_32RegClass, *MRI))
    return false;

  I.eraseFromParent();
  return true;
}

bool AMDGPUInstructionSelector::selectG_EXTRACT(MachineInstr &I) const {
  MachineBasicBlock *BB = I.getParent();
  Register DstReg = I.getOperand(0).getReg();
  Register SrcReg = I.getOperand(1).getReg();
  LLT DstTy = MRI->getType(DstReg);
  LLT SrcTy = MRI->getType(SrcReg);
  const unsigned SrcSize = SrcTy.getSizeInBits();
  const unsigned DstSize = DstTy.getSizeInBits();

  // TODO: Should handle any multiple of 32 offset.
  unsigned Offset = I.getOperand(2).getImm();
  if (Offset % DstSize != 0)
    return false;

  const RegisterBank *SrcBank = RBI.getRegBank(SrcReg, *MRI, TRI);
  const TargetRegisterClass *SrcRC =
    TRI.getRegClassForSizeOnBank(SrcSize, *SrcBank, *MRI);
  if (!SrcRC)
    return false;

  ArrayRef<int16_t> SubRegs = TRI.getRegSplitParts(SrcRC, DstSize / 8);

  const DebugLoc &DL = I.getDebugLoc();
  MachineInstr *Copy = BuildMI(*BB, &I, DL, TII.get(TargetOpcode::COPY), DstReg)
                               .addReg(SrcReg, 0, SubRegs[Offset / DstSize]);

  for (const MachineOperand &MO : Copy->operands()) {
    const TargetRegisterClass *RC =
            TRI.getConstrainedRegClassForOperand(MO, *MRI);
    if (!RC)
      continue;
    RBI.constrainGenericRegister(MO.getReg(), *RC, *MRI);
  }
  I.eraseFromParent();
  return true;
}

bool AMDGPUInstructionSelector::selectG_MERGE_VALUES(MachineInstr &MI) const {
  MachineBasicBlock *BB = MI.getParent();
  Register DstReg = MI.getOperand(0).getReg();
  LLT DstTy = MRI->getType(DstReg);
  LLT SrcTy = MRI->getType(MI.getOperand(1).getReg());

  const unsigned SrcSize = SrcTy.getSizeInBits();
  if (SrcSize < 32)
    return selectImpl(MI, *CoverageInfo);

  const DebugLoc &DL = MI.getDebugLoc();
  const RegisterBank *DstBank = RBI.getRegBank(DstReg, *MRI, TRI);
  const unsigned DstSize = DstTy.getSizeInBits();
  const TargetRegisterClass *DstRC =
    TRI.getRegClassForSizeOnBank(DstSize, *DstBank, *MRI);
  if (!DstRC)
    return false;

  ArrayRef<int16_t> SubRegs = TRI.getRegSplitParts(DstRC, SrcSize / 8);
  MachineInstrBuilder MIB =
    BuildMI(*BB, &MI, DL, TII.get(TargetOpcode::REG_SEQUENCE), DstReg);
  for (int I = 0, E = MI.getNumOperands() - 1; I != E; ++I) {
    MachineOperand &Src = MI.getOperand(I + 1);
    MIB.addReg(Src.getReg(), getUndefRegState(Src.isUndef()));
    MIB.addImm(SubRegs[I]);

    const TargetRegisterClass *SrcRC
      = TRI.getConstrainedRegClassForOperand(Src, *MRI);
    if (SrcRC && !RBI.constrainGenericRegister(Src.getReg(), *SrcRC, *MRI))
      return false;
  }

  if (!RBI.constrainGenericRegister(DstReg, *DstRC, *MRI))
    return false;

  MI.eraseFromParent();
  return true;
}

bool AMDGPUInstructionSelector::selectG_UNMERGE_VALUES(MachineInstr &MI) const {
  MachineBasicBlock *BB = MI.getParent();
  const int NumDst = MI.getNumOperands() - 1;

  MachineOperand &Src = MI.getOperand(NumDst);

  Register SrcReg = Src.getReg();
  Register DstReg0 = MI.getOperand(0).getReg();
  LLT DstTy = MRI->getType(DstReg0);
  LLT SrcTy = MRI->getType(SrcReg);

  const unsigned DstSize = DstTy.getSizeInBits();
  const unsigned SrcSize = SrcTy.getSizeInBits();
  const DebugLoc &DL = MI.getDebugLoc();
  const RegisterBank *SrcBank = RBI.getRegBank(SrcReg, *MRI, TRI);

  const TargetRegisterClass *SrcRC =
    TRI.getRegClassForSizeOnBank(SrcSize, *SrcBank, *MRI);
  if (!SrcRC || !RBI.constrainGenericRegister(SrcReg, *SrcRC, *MRI))
    return false;

  const unsigned SrcFlags = getUndefRegState(Src.isUndef());

  // Note we could have mixed SGPR and VGPR destination banks for an SGPR
  // source, and this relies on the fact that the same subregister indices are
  // used for both.
  ArrayRef<int16_t> SubRegs = TRI.getRegSplitParts(SrcRC, DstSize / 8);
  for (int I = 0, E = NumDst; I != E; ++I) {
    MachineOperand &Dst = MI.getOperand(I);
    BuildMI(*BB, &MI, DL, TII.get(TargetOpcode::COPY), Dst.getReg())
      .addReg(SrcReg, SrcFlags, SubRegs[I]);

    const TargetRegisterClass *DstRC =
      TRI.getConstrainedRegClassForOperand(Dst, *MRI);
    if (DstRC && !RBI.constrainGenericRegister(Dst.getReg(), *DstRC, *MRI))
      return false;
  }

  MI.eraseFromParent();
  return true;
}

bool AMDGPUInstructionSelector::selectG_PTR_ADD(MachineInstr &I) const {
  return selectG_ADD_SUB(I);
}

bool AMDGPUInstructionSelector::selectG_IMPLICIT_DEF(MachineInstr &I) const {
  const MachineOperand &MO = I.getOperand(0);

  // FIXME: Interface for getConstrainedRegClassForOperand needs work. The
  // regbank check here is to know why getConstrainedRegClassForOperand failed.
  const TargetRegisterClass *RC = TRI.getConstrainedRegClassForOperand(MO, *MRI);
  if ((!RC && !MRI->getRegBankOrNull(MO.getReg())) ||
      (RC && RBI.constrainGenericRegister(MO.getReg(), *RC, *MRI))) {
    I.setDesc(TII.get(TargetOpcode::IMPLICIT_DEF));
    return true;
  }

  return false;
}

bool AMDGPUInstructionSelector::selectG_INSERT(MachineInstr &I) const {
  MachineBasicBlock *BB = I.getParent();

  Register DstReg = I.getOperand(0).getReg();
  Register Src0Reg = I.getOperand(1).getReg();
  Register Src1Reg = I.getOperand(2).getReg();
  LLT Src1Ty = MRI->getType(Src1Reg);

  unsigned DstSize = MRI->getType(DstReg).getSizeInBits();
  unsigned InsSize = Src1Ty.getSizeInBits();

  int64_t Offset = I.getOperand(3).getImm();
  if (Offset % 32 != 0)
    return false;

  unsigned SubReg = TRI.getSubRegFromChannel(Offset / 32, InsSize / 32);
  if (SubReg == AMDGPU::NoSubRegister)
    return false;

  const RegisterBank *DstBank = RBI.getRegBank(DstReg, *MRI, TRI);
  const TargetRegisterClass *DstRC =
    TRI.getRegClassForSizeOnBank(DstSize, *DstBank, *MRI);
  if (!DstRC)
    return false;

  const RegisterBank *Src0Bank = RBI.getRegBank(Src0Reg, *MRI, TRI);
  const RegisterBank *Src1Bank = RBI.getRegBank(Src1Reg, *MRI, TRI);
  const TargetRegisterClass *Src0RC =
    TRI.getRegClassForSizeOnBank(DstSize, *Src0Bank, *MRI);
  const TargetRegisterClass *Src1RC =
    TRI.getRegClassForSizeOnBank(InsSize, *Src1Bank, *MRI);

  // Deal with weird cases where the class only partially supports the subreg
  // index.
  Src0RC = TRI.getSubClassWithSubReg(Src0RC, SubReg);
  if (!Src0RC)
    return false;

  if (!RBI.constrainGenericRegister(DstReg, *DstRC, *MRI) ||
      !RBI.constrainGenericRegister(Src0Reg, *Src0RC, *MRI) ||
      !RBI.constrainGenericRegister(Src1Reg, *Src1RC, *MRI))
    return false;

  const DebugLoc &DL = I.getDebugLoc();
  BuildMI(*BB, &I, DL, TII.get(TargetOpcode::INSERT_SUBREG), DstReg)
    .addReg(Src0Reg)
    .addReg(Src1Reg)
    .addImm(SubReg);

  I.eraseFromParent();
  return true;
}

bool AMDGPUInstructionSelector::selectInterpP1F16(MachineInstr &MI) const {
  if (STI.getLDSBankCount() != 16)
    return selectImpl(MI, *CoverageInfo);

  Register Dst = MI.getOperand(0).getReg();
  Register Src0 = MI.getOperand(2).getReg();
  Register M0Val = MI.getOperand(6).getReg();
  if (!RBI.constrainGenericRegister(M0Val, AMDGPU::SReg_32RegClass, *MRI) ||
      !RBI.constrainGenericRegister(Dst, AMDGPU::VGPR_32RegClass, *MRI) ||
      !RBI.constrainGenericRegister(Src0, AMDGPU::VGPR_32RegClass, *MRI))
    return false;

  // This requires 2 instructions. It is possible to write a pattern to support
  // this, but the generated isel emitter doesn't correctly deal with multiple
  // output instructions using the same physical register input. The copy to m0
  // is incorrectly placed before the second instruction.
  //
  // TODO: Match source modifiers.

  Register InterpMov = MRI->createVirtualRegister(&AMDGPU::VGPR_32RegClass);
  const DebugLoc &DL = MI.getDebugLoc();
  MachineBasicBlock *MBB = MI.getParent();

  BuildMI(*MBB, &MI, DL, TII.get(AMDGPU::COPY), AMDGPU::M0)
    .addReg(M0Val);
  BuildMI(*MBB, &MI, DL, TII.get(AMDGPU::V_INTERP_MOV_F32), InterpMov)
    .addImm(2)
    .addImm(MI.getOperand(4).getImm())  // $attr
    .addImm(MI.getOperand(3).getImm()); // $attrchan

  BuildMI(*MBB, &MI, DL, TII.get(AMDGPU::V_INTERP_P1LV_F16), Dst)
    .addImm(0)                          // $src0_modifiers
    .addReg(Src0)                       // $src0
    .addImm(MI.getOperand(4).getImm())  // $attr
    .addImm(MI.getOperand(3).getImm())  // $attrchan
    .addImm(0)                          // $src2_modifiers
    .addReg(InterpMov)                  // $src2 - 2 f16 values selected by high
    .addImm(MI.getOperand(5).getImm())  // $high
    .addImm(0)                          // $clamp
    .addImm(0);                         // $omod

  MI.eraseFromParent();
  return true;
}

bool AMDGPUInstructionSelector::selectG_INTRINSIC(MachineInstr &I) const {
  unsigned IntrinsicID = I.getIntrinsicID();
  switch (IntrinsicID) {
  case Intrinsic::amdgcn_if_break: {
    MachineBasicBlock *BB = I.getParent();

    // FIXME: Manually selecting to avoid dealiing with the SReg_1 trick
    // SelectionDAG uses for wave32 vs wave64.
    BuildMI(*BB, &I, I.getDebugLoc(), TII.get(AMDGPU::SI_IF_BREAK))
      .add(I.getOperand(0))
      .add(I.getOperand(2))
      .add(I.getOperand(3));

    Register DstReg = I.getOperand(0).getReg();
    Register Src0Reg = I.getOperand(2).getReg();
    Register Src1Reg = I.getOperand(3).getReg();

    I.eraseFromParent();

    for (Register Reg : { DstReg, Src0Reg, Src1Reg })
      MRI->setRegClass(Reg, TRI.getWaveMaskRegClass());

    return true;
  }
  case Intrinsic::amdgcn_interp_p1_f16:
    return selectInterpP1F16(I);
  default:
    return selectImpl(I, *CoverageInfo);
  }
}

static int getV_CMPOpcode(CmpInst::Predicate P, unsigned Size) {
  if (Size != 32 && Size != 64)
    return -1;
  switch (P) {
  default:
    llvm_unreachable("Unknown condition code!");
  case CmpInst::ICMP_NE:
    return Size == 32 ? AMDGPU::V_CMP_NE_U32_e64 : AMDGPU::V_CMP_NE_U64_e64;
  case CmpInst::ICMP_EQ:
    return Size == 32 ? AMDGPU::V_CMP_EQ_U32_e64 : AMDGPU::V_CMP_EQ_U64_e64;
  case CmpInst::ICMP_SGT:
    return Size == 32 ? AMDGPU::V_CMP_GT_I32_e64 : AMDGPU::V_CMP_GT_I64_e64;
  case CmpInst::ICMP_SGE:
    return Size == 32 ? AMDGPU::V_CMP_GE_I32_e64 : AMDGPU::V_CMP_GE_I64_e64;
  case CmpInst::ICMP_SLT:
    return Size == 32 ? AMDGPU::V_CMP_LT_I32_e64 : AMDGPU::V_CMP_LT_I64_e64;
  case CmpInst::ICMP_SLE:
    return Size == 32 ? AMDGPU::V_CMP_LE_I32_e64 : AMDGPU::V_CMP_LE_I64_e64;
  case CmpInst::ICMP_UGT:
    return Size == 32 ? AMDGPU::V_CMP_GT_U32_e64 : AMDGPU::V_CMP_GT_U64_e64;
  case CmpInst::ICMP_UGE:
    return Size == 32 ? AMDGPU::V_CMP_GE_U32_e64 : AMDGPU::V_CMP_GE_U64_e64;
  case CmpInst::ICMP_ULT:
    return Size == 32 ? AMDGPU::V_CMP_LT_U32_e64 : AMDGPU::V_CMP_LT_U64_e64;
  case CmpInst::ICMP_ULE:
    return Size == 32 ? AMDGPU::V_CMP_LE_U32_e64 : AMDGPU::V_CMP_LE_U64_e64;
  }
}

int AMDGPUInstructionSelector::getS_CMPOpcode(CmpInst::Predicate P,
                                              unsigned Size) const {
  if (Size == 64) {
    if (!STI.hasScalarCompareEq64())
      return -1;

    switch (P) {
    case CmpInst::ICMP_NE:
      return AMDGPU::S_CMP_LG_U64;
    case CmpInst::ICMP_EQ:
      return AMDGPU::S_CMP_EQ_U64;
    default:
      return -1;
    }
  }

  if (Size != 32)
    return -1;

  switch (P) {
  case CmpInst::ICMP_NE:
    return AMDGPU::S_CMP_LG_U32;
  case CmpInst::ICMP_EQ:
    return AMDGPU::S_CMP_EQ_U32;
  case CmpInst::ICMP_SGT:
    return AMDGPU::S_CMP_GT_I32;
  case CmpInst::ICMP_SGE:
    return AMDGPU::S_CMP_GE_I32;
  case CmpInst::ICMP_SLT:
    return AMDGPU::S_CMP_LT_I32;
  case CmpInst::ICMP_SLE:
    return AMDGPU::S_CMP_LE_I32;
  case CmpInst::ICMP_UGT:
    return AMDGPU::S_CMP_GT_U32;
  case CmpInst::ICMP_UGE:
    return AMDGPU::S_CMP_GE_U32;
  case CmpInst::ICMP_ULT:
    return AMDGPU::S_CMP_LT_U32;
  case CmpInst::ICMP_ULE:
    return AMDGPU::S_CMP_LE_U32;
  default:
    llvm_unreachable("Unknown condition code!");
  }
}

bool AMDGPUInstructionSelector::selectG_ICMP(MachineInstr &I) const {
  MachineBasicBlock *BB = I.getParent();
  const DebugLoc &DL = I.getDebugLoc();

  Register SrcReg = I.getOperand(2).getReg();
  unsigned Size = RBI.getSizeInBits(SrcReg, *MRI, TRI);

  auto Pred = (CmpInst::Predicate)I.getOperand(1).getPredicate();

  Register CCReg = I.getOperand(0).getReg();
  if (!isVCC(CCReg, *MRI)) {
    int Opcode = getS_CMPOpcode(Pred, Size);
    if (Opcode == -1)
      return false;
    MachineInstr *ICmp = BuildMI(*BB, &I, DL, TII.get(Opcode))
            .add(I.getOperand(2))
            .add(I.getOperand(3));
    BuildMI(*BB, &I, DL, TII.get(AMDGPU::COPY), CCReg)
      .addReg(AMDGPU::SCC);
    bool Ret =
        constrainSelectedInstRegOperands(*ICmp, TII, TRI, RBI) &&
        RBI.constrainGenericRegister(CCReg, AMDGPU::SReg_32RegClass, *MRI);
    I.eraseFromParent();
    return Ret;
  }

  int Opcode = getV_CMPOpcode(Pred, Size);
  if (Opcode == -1)
    return false;

  MachineInstr *ICmp = BuildMI(*BB, &I, DL, TII.get(Opcode),
            I.getOperand(0).getReg())
            .add(I.getOperand(2))
            .add(I.getOperand(3));
  RBI.constrainGenericRegister(ICmp->getOperand(0).getReg(),
                               *TRI.getBoolRC(), *MRI);
  bool Ret = constrainSelectedInstRegOperands(*ICmp, TII, TRI, RBI);
  I.eraseFromParent();
  return Ret;
}

static bool isZero(Register Reg, MachineRegisterInfo &MRI) {
  int64_t C;
  if (mi_match(Reg, MRI, m_ICst(C)) && C == 0)
    return true;

  // FIXME: matcher should ignore copies
  return mi_match(Reg, MRI, m_Copy(m_ICst(C))) && C == 0;
}

static unsigned extractGLC(unsigned AuxiliaryData) {
  return AuxiliaryData & 1;
}

static unsigned extractSLC(unsigned AuxiliaryData) {
  return (AuxiliaryData >> 1) & 1;
}

static unsigned extractDLC(unsigned AuxiliaryData) {
  return (AuxiliaryData >> 2) & 1;
}

static unsigned extractSWZ(unsigned AuxiliaryData) {
  return (AuxiliaryData >> 3) & 1;
}

static unsigned getBufferStoreOpcode(LLT Ty,
                                     const unsigned MemSize,
                                     const bool Offen) {
  const int Size = Ty.getSizeInBits();
  switch (8 * MemSize) {
  case 8:
    return Offen ? AMDGPU::BUFFER_STORE_BYTE_OFFEN_exact :
                   AMDGPU::BUFFER_STORE_BYTE_OFFSET_exact;
  case 16:
    return Offen ? AMDGPU::BUFFER_STORE_SHORT_OFFEN_exact :
                   AMDGPU::BUFFER_STORE_SHORT_OFFSET_exact;
  default:
    unsigned Opc = Offen ? AMDGPU::BUFFER_STORE_DWORD_OFFEN_exact :
                           AMDGPU::BUFFER_STORE_DWORD_OFFSET_exact;
    if (Size > 32)
      Opc = AMDGPU::getMUBUFOpcode(Opc, Size / 32);
    return Opc;
  }
}

static unsigned getBufferStoreFormatOpcode(LLT Ty,
                                           const unsigned MemSize,
                                           const bool Offen) {
  bool IsD16Packed = Ty.getScalarSizeInBits() == 16;
  bool IsD16Unpacked = 8 * MemSize < Ty.getSizeInBits();
  int NumElts = Ty.isVector() ? Ty.getNumElements() : 1;

  if (IsD16Packed) {
    switch (NumElts) {
    case 1:
      return Offen ? AMDGPU::BUFFER_STORE_FORMAT_D16_X_OFFEN_exact :
                     AMDGPU::BUFFER_STORE_FORMAT_D16_X_OFFSET_exact;
    case 2:
      return Offen ? AMDGPU::BUFFER_STORE_FORMAT_D16_XY_OFFEN_exact :
                     AMDGPU::BUFFER_STORE_FORMAT_D16_XY_OFFSET_exact;
    case 3:
      return Offen ? AMDGPU::BUFFER_STORE_FORMAT_D16_XYZ_OFFEN_exact :
                     AMDGPU::BUFFER_STORE_FORMAT_D16_XYZ_OFFSET_exact;
    case 4:
      return Offen ? AMDGPU::BUFFER_STORE_FORMAT_D16_XYZW_OFFEN_exact :
                     AMDGPU::BUFFER_STORE_FORMAT_D16_XYZW_OFFSET_exact;
    default:
      return -1;
    }
  }

  if (IsD16Unpacked) {
    switch (NumElts) {
    case 1:
      return Offen ? AMDGPU::BUFFER_STORE_FORMAT_D16_X_OFFEN_exact :
                     AMDGPU::BUFFER_STORE_FORMAT_D16_X_OFFSET_exact;
    case 2:
      return Offen ? AMDGPU::BUFFER_STORE_FORMAT_D16_XY_gfx80_OFFEN_exact :
                     AMDGPU::BUFFER_STORE_FORMAT_D16_XY_gfx80_OFFSET_exact;
    case 3:
      return Offen ? AMDGPU::BUFFER_STORE_FORMAT_D16_XYZ_gfx80_OFFEN_exact :
                     AMDGPU::BUFFER_STORE_FORMAT_D16_XYZ_gfx80_OFFSET_exact;
    case 4:
      return Offen ? AMDGPU::BUFFER_STORE_FORMAT_D16_XYZW_gfx80_OFFEN_exact :
                     AMDGPU::BUFFER_STORE_FORMAT_D16_XYZW_gfx80_OFFSET_exact;
    default:
      return -1;
    }
  }

  switch (NumElts) {
  case 1:
    return Offen ? AMDGPU::BUFFER_STORE_FORMAT_X_OFFEN_exact :
                   AMDGPU::BUFFER_STORE_FORMAT_X_OFFSET_exact;
  case 2:
    return Offen ? AMDGPU::BUFFER_STORE_FORMAT_XY_OFFEN_exact :
                  AMDGPU::BUFFER_STORE_FORMAT_XY_OFFSET_exact;
  case 3:
    return Offen ? AMDGPU::BUFFER_STORE_FORMAT_XYZ_OFFEN_exact :
                   AMDGPU::BUFFER_STORE_FORMAT_XYZ_OFFSET_exact;
  case 4:
    return Offen ? AMDGPU::BUFFER_STORE_FORMAT_XYZW_OFFEN_exact :
                   AMDGPU::BUFFER_STORE_FORMAT_XYZW_OFFSET_exact;
  default:
    return -1;
  }

  llvm_unreachable("unhandled buffer store");
}

// TODO: Move this to combiner
// Returns base register, imm offset, total constant offset.
std::tuple<Register, unsigned, unsigned>
AMDGPUInstructionSelector::splitBufferOffsets(MachineIRBuilder &B,
                                              Register OrigOffset) const {
  const unsigned MaxImm = 4095;
  Register BaseReg;
  unsigned TotalConstOffset;
  MachineInstr *OffsetDef;

  std::tie(BaseReg, TotalConstOffset, OffsetDef)
    = AMDGPU::getBaseWithConstantOffset(*MRI, OrigOffset);

  unsigned ImmOffset = TotalConstOffset;

  // If the immediate value is too big for the immoffset field, put the value
  // and -4096 into the immoffset field so that the value that is copied/added
  // for the voffset field is a multiple of 4096, and it stands more chance
  // of being CSEd with the copy/add for another similar load/store.f
  // However, do not do that rounding down to a multiple of 4096 if that is a
  // negative number, as it appears to be illegal to have a negative offset
  // in the vgpr, even if adding the immediate offset makes it positive.
  unsigned Overflow = ImmOffset & ~MaxImm;
  ImmOffset -= Overflow;
  if ((int32_t)Overflow < 0) {
    Overflow += ImmOffset;
    ImmOffset = 0;
  }

  if (Overflow != 0) {
    // In case this is in a waterfall loop, insert offset code at the def point
    // of the offset, not inside the loop.
    MachineBasicBlock::iterator OldInsPt = B.getInsertPt();
    MachineBasicBlock &OldMBB = B.getMBB();
    B.setInstr(*OffsetDef);

    if (!BaseReg) {
      BaseReg = MRI->createVirtualRegister(&AMDGPU::VGPR_32RegClass);
      B.buildInstr(AMDGPU::V_MOV_B32_e32)
        .addDef(BaseReg)
        .addImm(Overflow);
    } else {
      Register OverflowVal = MRI->createVirtualRegister(&AMDGPU::VGPR_32RegClass);
      B.buildInstr(AMDGPU::V_MOV_B32_e32)
        .addDef(OverflowVal)
        .addImm(Overflow);

      Register NewBaseReg = MRI->createVirtualRegister(&AMDGPU::VGPR_32RegClass);
      TII.getAddNoCarry(B.getMBB(), B.getInsertPt(), B.getDebugLoc(), NewBaseReg)
        .addReg(BaseReg)
        .addReg(OverflowVal, RegState::Kill)
        .addImm(0);
      BaseReg = NewBaseReg;
    }

    B.setInsertPt(OldMBB, OldInsPt);
  }

  return std::make_tuple(BaseReg, ImmOffset, TotalConstOffset);
}

bool AMDGPUInstructionSelector::selectStoreIntrinsic(MachineInstr &MI,
                                                     bool IsFormat) const {
  MachineIRBuilder B(MI);
  MachineFunction &MF = B.getMF();
  Register VData = MI.getOperand(1).getReg();
  LLT Ty = MRI->getType(VData);

  int Size = Ty.getSizeInBits();
  if (Size % 32 != 0)
    return false;

  // FIXME: Verifier should enforce 1 MMO for these intrinsics.
  MachineMemOperand *MMO = *MI.memoperands_begin();
  const int MemSize = MMO->getSize();

  Register RSrc = MI.getOperand(2).getReg();
  Register VOffset = MI.getOperand(3).getReg();
  Register SOffset = MI.getOperand(4).getReg();
  unsigned AuxiliaryData = MI.getOperand(5).getImm();
  unsigned ImmOffset;
  unsigned TotalOffset;

  std::tie(VOffset, ImmOffset, TotalOffset) = splitBufferOffsets(B, VOffset);
  if (TotalOffset != 0)
    MMO = MF.getMachineMemOperand(MMO, TotalOffset, MemSize);

  const bool Offen = !isZero(VOffset, *MRI);

  int Opc = IsFormat ? getBufferStoreFormatOpcode(Ty, MemSize, Offen) :
    getBufferStoreOpcode(Ty, MemSize, Offen);
  if (Opc == -1)
    return false;

  MachineInstrBuilder MIB = B.buildInstr(Opc)
    .addUse(VData);

  if (Offen)
    MIB.addUse(VOffset);

  MIB.addUse(RSrc)
     .addUse(SOffset)
     .addImm(ImmOffset)
     .addImm(extractGLC(AuxiliaryData))
     .addImm(extractSLC(AuxiliaryData))
     .addImm(0) // tfe: FIXME: Remove from inst
     .addImm(extractDLC(AuxiliaryData))
     .addImm(extractSWZ(AuxiliaryData))
     .addMemOperand(MMO);

  MI.eraseFromParent();

  return constrainSelectedInstRegOperands(*MIB, TII, TRI, RBI);
}

static unsigned getDSShaderTypeValue(const MachineFunction &MF) {
  switch (MF.getFunction().getCallingConv()) {
  case CallingConv::AMDGPU_PS:
    return 1;
  case CallingConv::AMDGPU_VS:
    return 2;
  case CallingConv::AMDGPU_GS:
    return 3;
  case CallingConv::AMDGPU_HS:
  case CallingConv::AMDGPU_LS:
  case CallingConv::AMDGPU_ES:
    report_fatal_error("ds_ordered_count unsupported for this calling conv");
  case CallingConv::AMDGPU_CS:
  case CallingConv::AMDGPU_KERNEL:
  case CallingConv::C:
  case CallingConv::Fast:
  default:
    // Assume other calling conventions are various compute callable functions
    return 0;
  }
}

bool AMDGPUInstructionSelector::selectDSOrderedIntrinsic(
  MachineInstr &MI, Intrinsic::ID IntrID) const {
  MachineBasicBlock *MBB = MI.getParent();
  MachineFunction *MF = MBB->getParent();
  const DebugLoc &DL = MI.getDebugLoc();

  unsigned IndexOperand = MI.getOperand(7).getImm();
  bool WaveRelease = MI.getOperand(8).getImm() != 0;
  bool WaveDone = MI.getOperand(9).getImm() != 0;

  if (WaveDone && !WaveRelease)
    report_fatal_error("ds_ordered_count: wave_done requires wave_release");

  unsigned OrderedCountIndex = IndexOperand & 0x3f;
  IndexOperand &= ~0x3f;
  unsigned CountDw = 0;

  if (STI.getGeneration() >= AMDGPUSubtarget::GFX10) {
    CountDw = (IndexOperand >> 24) & 0xf;
    IndexOperand &= ~(0xf << 24);

    if (CountDw < 1 || CountDw > 4) {
      report_fatal_error(
        "ds_ordered_count: dword count must be between 1 and 4");
    }
  }

  if (IndexOperand)
    report_fatal_error("ds_ordered_count: bad index operand");

  unsigned Instruction = IntrID == Intrinsic::amdgcn_ds_ordered_add ? 0 : 1;
  unsigned ShaderType = getDSShaderTypeValue(*MF);

  unsigned Offset0 = OrderedCountIndex << 2;
  unsigned Offset1 = WaveRelease | (WaveDone << 1) | (ShaderType << 2) |
                     (Instruction << 4);

  if (STI.getGeneration() >= AMDGPUSubtarget::GFX10)
    Offset1 |= (CountDw - 1) << 6;

  unsigned Offset = Offset0 | (Offset1 << 8);

  Register M0Val = MI.getOperand(2).getReg();
  BuildMI(*MBB, &MI, DL, TII.get(AMDGPU::COPY), AMDGPU::M0)
    .addReg(M0Val);

  Register DstReg = MI.getOperand(0).getReg();
  Register ValReg = MI.getOperand(3).getReg();
  MachineInstrBuilder DS =
    BuildMI(*MBB, &MI, DL, TII.get(AMDGPU::DS_ORDERED_COUNT), DstReg)
      .addReg(ValReg)
      .addImm(Offset)
      .cloneMemRefs(MI);

  if (!RBI.constrainGenericRegister(M0Val, AMDGPU::SReg_32RegClass, *MRI))
    return false;

  bool Ret = constrainSelectedInstRegOperands(*DS, TII, TRI, RBI);
  MI.eraseFromParent();
  return Ret;
}

static unsigned gwsIntrinToOpcode(unsigned IntrID) {
  switch (IntrID) {
  case Intrinsic::amdgcn_ds_gws_init:
    return AMDGPU::DS_GWS_INIT;
  case Intrinsic::amdgcn_ds_gws_barrier:
    return AMDGPU::DS_GWS_BARRIER;
  case Intrinsic::amdgcn_ds_gws_sema_v:
    return AMDGPU::DS_GWS_SEMA_V;
  case Intrinsic::amdgcn_ds_gws_sema_br:
    return AMDGPU::DS_GWS_SEMA_BR;
  case Intrinsic::amdgcn_ds_gws_sema_p:
    return AMDGPU::DS_GWS_SEMA_P;
  case Intrinsic::amdgcn_ds_gws_sema_release_all:
    return AMDGPU::DS_GWS_SEMA_RELEASE_ALL;
  default:
    llvm_unreachable("not a gws intrinsic");
  }
}

bool AMDGPUInstructionSelector::selectDSGWSIntrinsic(MachineInstr &MI,
                                                     Intrinsic::ID IID) const {
  if (IID == Intrinsic::amdgcn_ds_gws_sema_release_all &&
      !STI.hasGWSSemaReleaseAll())
    return false;

  // intrinsic ID, vsrc, offset
  const bool HasVSrc = MI.getNumOperands() == 3;
  assert(HasVSrc || MI.getNumOperands() == 2);

  Register BaseOffset = MI.getOperand(HasVSrc ? 2 : 1).getReg();
  const RegisterBank *OffsetRB = RBI.getRegBank(BaseOffset, *MRI, TRI);
  if (OffsetRB->getID() != AMDGPU::SGPRRegBankID)
    return false;

  MachineInstr *OffsetDef = getDefIgnoringCopies(BaseOffset, *MRI);
  assert(OffsetDef);

  unsigned ImmOffset;

  MachineBasicBlock *MBB = MI.getParent();
  const DebugLoc &DL = MI.getDebugLoc();

  MachineInstr *Readfirstlane = nullptr;

  // If we legalized the VGPR input, strip out the readfirstlane to analyze the
  // incoming offset, in case there's an add of a constant. We'll have to put it
  // back later.
  if (OffsetDef->getOpcode() == AMDGPU::V_READFIRSTLANE_B32) {
    Readfirstlane = OffsetDef;
    BaseOffset = OffsetDef->getOperand(1).getReg();
    OffsetDef = getDefIgnoringCopies(BaseOffset, *MRI);
  }

  if (OffsetDef->getOpcode() == AMDGPU::G_CONSTANT) {
    // If we have a constant offset, try to use the 0 in m0 as the base.
    // TODO: Look into changing the default m0 initialization value. If the
    // default -1 only set the low 16-bits, we could leave it as-is and add 1 to
    // the immediate offset.

    ImmOffset = OffsetDef->getOperand(1).getCImm()->getZExtValue();
    BuildMI(*MBB, &MI, DL, TII.get(AMDGPU::S_MOV_B32), AMDGPU::M0)
      .addImm(0);
  } else {
    std::tie(BaseOffset, ImmOffset, OffsetDef)
      = AMDGPU::getBaseWithConstantOffset(*MRI, BaseOffset);

    if (Readfirstlane) {
      // We have the constant offset now, so put the readfirstlane back on the
      // variable component.
      if (!RBI.constrainGenericRegister(BaseOffset, AMDGPU::VGPR_32RegClass, *MRI))
        return false;

      Readfirstlane->getOperand(1).setReg(BaseOffset);
      BaseOffset = Readfirstlane->getOperand(0).getReg();
    } else {
      if (!RBI.constrainGenericRegister(BaseOffset,
                                        AMDGPU::SReg_32RegClass, *MRI))
        return false;
    }

    Register M0Base = MRI->createVirtualRegister(&AMDGPU::SReg_32RegClass);
    BuildMI(*MBB, &MI, DL, TII.get(AMDGPU::S_LSHL_B32), M0Base)
      .addReg(BaseOffset)
      .addImm(16);

    BuildMI(*MBB, &MI, DL, TII.get(AMDGPU::COPY), AMDGPU::M0)
      .addReg(M0Base);
  }

  // The resource id offset is computed as (<isa opaque base> + M0[21:16] +
  // offset field) % 64. Some versions of the programming guide omit the m0
  // part, or claim it's from offset 0.
  auto MIB = BuildMI(*MBB, &MI, DL, TII.get(gwsIntrinToOpcode(IID)));

  if (HasVSrc) {
    Register VSrc = MI.getOperand(1).getReg();
    MIB.addReg(VSrc);
    if (!RBI.constrainGenericRegister(VSrc, AMDGPU::VGPR_32RegClass, *MRI))
      return false;
  }

  MIB.addImm(ImmOffset)
     .addImm(-1) // $gds
     .cloneMemRefs(MI);

  MI.eraseFromParent();
  return true;
}

bool AMDGPUInstructionSelector::selectDSAppendConsume(MachineInstr &MI,
                                                      bool IsAppend) const {
  Register PtrBase = MI.getOperand(2).getReg();
  LLT PtrTy = MRI->getType(PtrBase);
  bool IsGDS = PtrTy.getAddressSpace() == AMDGPUAS::REGION_ADDRESS;

  unsigned Offset;
  std::tie(PtrBase, Offset) = selectDS1Addr1OffsetImpl(MI.getOperand(2));

  // TODO: Should this try to look through readfirstlane like GWS?
  if (!isDSOffsetLegal(PtrBase, Offset, 16)) {
    PtrBase = MI.getOperand(2).getReg();
    Offset = 0;
  }

  MachineBasicBlock *MBB = MI.getParent();
  const DebugLoc &DL = MI.getDebugLoc();
  const unsigned Opc = IsAppend ? AMDGPU::DS_APPEND : AMDGPU::DS_CONSUME;

  BuildMI(*MBB, &MI, DL, TII.get(AMDGPU::COPY), AMDGPU::M0)
    .addReg(PtrBase);
  BuildMI(*MBB, &MI, DL, TII.get(Opc), MI.getOperand(0).getReg())
    .addImm(Offset)
    .addImm(IsGDS ? -1 : 0)
    .cloneMemRefs(MI);

  MI.eraseFromParent();
  return true;
}

bool AMDGPUInstructionSelector::selectG_INTRINSIC_W_SIDE_EFFECTS(
    MachineInstr &I) const {
  MachineBasicBlock *BB = I.getParent();
  unsigned IntrinsicID = I.getIntrinsicID();
  switch (IntrinsicID) {
  case Intrinsic::amdgcn_end_cf: {
    // FIXME: Manually selecting to avoid dealiing with the SReg_1 trick
    // SelectionDAG uses for wave32 vs wave64.
    BuildMI(*BB, &I, I.getDebugLoc(),
            TII.get(AMDGPU::SI_END_CF))
      .add(I.getOperand(1));

    Register Reg = I.getOperand(1).getReg();
    I.eraseFromParent();

    if (!MRI->getRegClassOrNull(Reg))
      MRI->setRegClass(Reg, TRI.getWaveMaskRegClass());
    return true;
  }
  case Intrinsic::amdgcn_raw_buffer_store:
    return selectStoreIntrinsic(I, false);
  case Intrinsic::amdgcn_raw_buffer_store_format:
    return selectStoreIntrinsic(I, true);
  case Intrinsic::amdgcn_ds_ordered_add:
  case Intrinsic::amdgcn_ds_ordered_swap:
    return selectDSOrderedIntrinsic(I, IntrinsicID);
  case Intrinsic::amdgcn_ds_gws_init:
  case Intrinsic::amdgcn_ds_gws_barrier:
  case Intrinsic::amdgcn_ds_gws_sema_v:
  case Intrinsic::amdgcn_ds_gws_sema_br:
  case Intrinsic::amdgcn_ds_gws_sema_p:
  case Intrinsic::amdgcn_ds_gws_sema_release_all:
    return selectDSGWSIntrinsic(I, IntrinsicID);
  case Intrinsic::amdgcn_ds_append:
    return selectDSAppendConsume(I, true);
  case Intrinsic::amdgcn_ds_consume:
    return selectDSAppendConsume(I, false);
  default:
    return selectImpl(I, *CoverageInfo);
  }
}

bool AMDGPUInstructionSelector::selectG_SELECT(MachineInstr &I) const {
  MachineBasicBlock *BB = I.getParent();
  const DebugLoc &DL = I.getDebugLoc();

  Register DstReg = I.getOperand(0).getReg();
  unsigned Size = RBI.getSizeInBits(DstReg, *MRI, TRI);
  assert(Size <= 32 || Size == 64);
  const MachineOperand &CCOp = I.getOperand(1);
  Register CCReg = CCOp.getReg();
  if (!isVCC(CCReg, *MRI)) {
    unsigned SelectOpcode = Size == 64 ? AMDGPU::S_CSELECT_B64 :
                                         AMDGPU::S_CSELECT_B32;
    MachineInstr *CopySCC = BuildMI(*BB, &I, DL, TII.get(AMDGPU::COPY), AMDGPU::SCC)
            .addReg(CCReg);

    // The generic constrainSelectedInstRegOperands doesn't work for the scc register
    // bank, because it does not cover the register class that we used to represent
    // for it.  So we need to manually set the register class here.
    if (!MRI->getRegClassOrNull(CCReg))
        MRI->setRegClass(CCReg, TRI.getConstrainedRegClassForOperand(CCOp, *MRI));
    MachineInstr *Select = BuildMI(*BB, &I, DL, TII.get(SelectOpcode), DstReg)
            .add(I.getOperand(2))
            .add(I.getOperand(3));

    bool Ret = constrainSelectedInstRegOperands(*Select, TII, TRI, RBI) |
               constrainSelectedInstRegOperands(*CopySCC, TII, TRI, RBI);
    I.eraseFromParent();
    return Ret;
  }

  // Wide VGPR select should have been split in RegBankSelect.
  if (Size > 32)
    return false;

  MachineInstr *Select =
      BuildMI(*BB, &I, DL, TII.get(AMDGPU::V_CNDMASK_B32_e64), DstReg)
              .addImm(0)
              .add(I.getOperand(3))
              .addImm(0)
              .add(I.getOperand(2))
              .add(I.getOperand(1));

  bool Ret = constrainSelectedInstRegOperands(*Select, TII, TRI, RBI);
  I.eraseFromParent();
  return Ret;
}

bool AMDGPUInstructionSelector::selectG_STORE(MachineInstr &I) const {
  initM0(I);
  return selectImpl(I, *CoverageInfo);
}

static int sizeToSubRegIndex(unsigned Size) {
  switch (Size) {
  case 32:
    return AMDGPU::sub0;
  case 64:
    return AMDGPU::sub0_sub1;
  case 96:
    return AMDGPU::sub0_sub1_sub2;
  case 128:
    return AMDGPU::sub0_sub1_sub2_sub3;
  case 256:
    return AMDGPU::sub0_sub1_sub2_sub3_sub4_sub5_sub6_sub7;
  default:
    if (Size < 32)
      return AMDGPU::sub0;
    if (Size > 256)
      return -1;
    return sizeToSubRegIndex(PowerOf2Ceil(Size));
  }
}

bool AMDGPUInstructionSelector::selectG_TRUNC(MachineInstr &I) const {
  Register DstReg = I.getOperand(0).getReg();
  Register SrcReg = I.getOperand(1).getReg();
  const LLT DstTy = MRI->getType(DstReg);
  const LLT SrcTy = MRI->getType(SrcReg);
  if (!DstTy.isScalar())
    return false;

  const LLT S1 = LLT::scalar(1);

  const RegisterBank *SrcRB = RBI.getRegBank(SrcReg, *MRI, TRI);
  const RegisterBank *DstRB;
  if (DstTy == S1) {
    // This is a special case. We don't treat s1 for legalization artifacts as
    // vcc booleans.
    DstRB = SrcRB;
  } else {
    DstRB = RBI.getRegBank(DstReg, *MRI, TRI);
    if (SrcRB != DstRB)
      return false;
  }

  unsigned DstSize = DstTy.getSizeInBits();
  unsigned SrcSize = SrcTy.getSizeInBits();

  const TargetRegisterClass *SrcRC
    = TRI.getRegClassForSizeOnBank(SrcSize, *SrcRB, *MRI);
  const TargetRegisterClass *DstRC
    = TRI.getRegClassForSizeOnBank(DstSize, *DstRB, *MRI);

  if (SrcSize > 32) {
    int SubRegIdx = sizeToSubRegIndex(DstSize);
    if (SubRegIdx == -1)
      return false;

    // Deal with weird cases where the class only partially supports the subreg
    // index.
    SrcRC = TRI.getSubClassWithSubReg(SrcRC, SubRegIdx);
    if (!SrcRC)
      return false;

    I.getOperand(1).setSubReg(SubRegIdx);
  }

  if (!RBI.constrainGenericRegister(SrcReg, *SrcRC, *MRI) ||
      !RBI.constrainGenericRegister(DstReg, *DstRC, *MRI)) {
    LLVM_DEBUG(dbgs() << "Failed to constrain G_TRUNC\n");
    return false;
  }

  I.setDesc(TII.get(TargetOpcode::COPY));
  return true;
}

/// \returns true if a bitmask for \p Size bits will be an inline immediate.
static bool shouldUseAndMask(unsigned Size, unsigned &Mask) {
  Mask = maskTrailingOnes<unsigned>(Size);
  int SignedMask = static_cast<int>(Mask);
  return SignedMask >= -16 && SignedMask <= 64;
}

// Like RegisterBankInfo::getRegBank, but don't assume vcc for s1.
const RegisterBank *AMDGPUInstructionSelector::getArtifactRegBank(
  Register Reg, const MachineRegisterInfo &MRI,
  const TargetRegisterInfo &TRI) const {
  const RegClassOrRegBank &RegClassOrBank = MRI.getRegClassOrRegBank(Reg);
  if (auto *RB = RegClassOrBank.dyn_cast<const RegisterBank *>())
    return RB;

  // Ignore the type, since we don't use vcc in artifacts.
  if (auto *RC = RegClassOrBank.dyn_cast<const TargetRegisterClass *>())
    return &RBI.getRegBankFromRegClass(*RC, LLT());
  return nullptr;
}

bool AMDGPUInstructionSelector::selectG_SZA_EXT(MachineInstr &I) const {
  bool Signed = I.getOpcode() == AMDGPU::G_SEXT;
  const DebugLoc &DL = I.getDebugLoc();
  MachineBasicBlock &MBB = *I.getParent();
  const Register DstReg = I.getOperand(0).getReg();
  const Register SrcReg = I.getOperand(1).getReg();

  const LLT DstTy = MRI->getType(DstReg);
  const LLT SrcTy = MRI->getType(SrcReg);
  const unsigned SrcSize = SrcTy.getSizeInBits();
  const unsigned DstSize = DstTy.getSizeInBits();
  if (!DstTy.isScalar())
    return false;

  if (I.getOpcode() == AMDGPU::G_ANYEXT)
    return selectCOPY(I);

  // Artifact casts should never use vcc.
  const RegisterBank *SrcBank = getArtifactRegBank(SrcReg, *MRI, TRI);

  if (SrcBank->getID() == AMDGPU::VGPRRegBankID && DstSize <= 32) {
    // 64-bit should have been split up in RegBankSelect

    // Try to use an and with a mask if it will save code size.
    unsigned Mask;
    if (!Signed && shouldUseAndMask(SrcSize, Mask)) {
      MachineInstr *ExtI =
      BuildMI(MBB, I, DL, TII.get(AMDGPU::V_AND_B32_e32), DstReg)
        .addImm(Mask)
        .addReg(SrcReg);
      I.eraseFromParent();
      return constrainSelectedInstRegOperands(*ExtI, TII, TRI, RBI);
    }

    const unsigned BFE = Signed ? AMDGPU::V_BFE_I32 : AMDGPU::V_BFE_U32;
    MachineInstr *ExtI =
      BuildMI(MBB, I, DL, TII.get(BFE), DstReg)
      .addReg(SrcReg)
      .addImm(0) // Offset
      .addImm(SrcSize); // Width
    I.eraseFromParent();
    return constrainSelectedInstRegOperands(*ExtI, TII, TRI, RBI);
  }

  if (SrcBank->getID() == AMDGPU::SGPRRegBankID && DstSize <= 64) {
    if (!RBI.constrainGenericRegister(SrcReg, AMDGPU::SReg_32RegClass, *MRI))
      return false;

    if (Signed && DstSize == 32 && (SrcSize == 8 || SrcSize == 16)) {
      const unsigned SextOpc = SrcSize == 8 ?
        AMDGPU::S_SEXT_I32_I8 : AMDGPU::S_SEXT_I32_I16;
      BuildMI(MBB, I, DL, TII.get(SextOpc), DstReg)
        .addReg(SrcReg);
      I.eraseFromParent();
      return RBI.constrainGenericRegister(DstReg, AMDGPU::SReg_32RegClass, *MRI);
    }

    const unsigned BFE64 = Signed ? AMDGPU::S_BFE_I64 : AMDGPU::S_BFE_U64;
    const unsigned BFE32 = Signed ? AMDGPU::S_BFE_I32 : AMDGPU::S_BFE_U32;

    // Scalar BFE is encoded as S1[5:0] = offset, S1[22:16]= width.
    if (DstSize > 32 && SrcSize <= 32) {
      // We need a 64-bit register source, but the high bits don't matter.
      Register ExtReg = MRI->createVirtualRegister(&AMDGPU::SReg_64RegClass);
      Register UndefReg = MRI->createVirtualRegister(&AMDGPU::SReg_32RegClass);
      BuildMI(MBB, I, DL, TII.get(AMDGPU::IMPLICIT_DEF), UndefReg);
      BuildMI(MBB, I, DL, TII.get(AMDGPU::REG_SEQUENCE), ExtReg)
        .addReg(SrcReg)
        .addImm(AMDGPU::sub0)
        .addReg(UndefReg)
        .addImm(AMDGPU::sub1);

      BuildMI(MBB, I, DL, TII.get(BFE64), DstReg)
        .addReg(ExtReg)
        .addImm(SrcSize << 16);

      I.eraseFromParent();
      return RBI.constrainGenericRegister(DstReg, AMDGPU::SReg_64RegClass, *MRI);
    }

    unsigned Mask;
    if (!Signed && shouldUseAndMask(SrcSize, Mask)) {
      BuildMI(MBB, I, DL, TII.get(AMDGPU::S_AND_B32), DstReg)
        .addReg(SrcReg)
        .addImm(Mask);
    } else {
      BuildMI(MBB, I, DL, TII.get(BFE32), DstReg)
        .addReg(SrcReg)
        .addImm(SrcSize << 16);
    }

    I.eraseFromParent();
    return RBI.constrainGenericRegister(DstReg, AMDGPU::SReg_32RegClass, *MRI);
  }

  return false;
}

bool AMDGPUInstructionSelector::selectG_CONSTANT(MachineInstr &I) const {
  MachineBasicBlock *BB = I.getParent();
  MachineOperand &ImmOp = I.getOperand(1);

  // The AMDGPU backend only supports Imm operands and not CImm or FPImm.
  if (ImmOp.isFPImm()) {
    const APInt &Imm = ImmOp.getFPImm()->getValueAPF().bitcastToAPInt();
    ImmOp.ChangeToImmediate(Imm.getZExtValue());
  } else if (ImmOp.isCImm()) {
    ImmOp.ChangeToImmediate(ImmOp.getCImm()->getZExtValue());
  }

  Register DstReg = I.getOperand(0).getReg();
  unsigned Size;
  bool IsSgpr;
  const RegisterBank *RB = MRI->getRegBankOrNull(I.getOperand(0).getReg());
  if (RB) {
    IsSgpr = RB->getID() == AMDGPU::SGPRRegBankID;
    Size = MRI->getType(DstReg).getSizeInBits();
  } else {
    const TargetRegisterClass *RC = TRI.getRegClassForReg(*MRI, DstReg);
    IsSgpr = TRI.isSGPRClass(RC);
    Size = TRI.getRegSizeInBits(*RC);
  }

  if (Size != 32 && Size != 64)
    return false;

  unsigned Opcode = IsSgpr ? AMDGPU::S_MOV_B32 : AMDGPU::V_MOV_B32_e32;
  if (Size == 32) {
    I.setDesc(TII.get(Opcode));
    I.addImplicitDefUseOperands(*MF);
    return constrainSelectedInstRegOperands(I, TII, TRI, RBI);
  }

  const DebugLoc &DL = I.getDebugLoc();

  APInt Imm(Size, I.getOperand(1).getImm());

  MachineInstr *ResInst;
  if (IsSgpr && TII.isInlineConstant(Imm)) {
    ResInst = BuildMI(*BB, &I, DL, TII.get(AMDGPU::S_MOV_B64), DstReg)
      .addImm(I.getOperand(1).getImm());
  } else {
    const TargetRegisterClass *RC = IsSgpr ?
      &AMDGPU::SReg_32RegClass : &AMDGPU::VGPR_32RegClass;
    Register LoReg = MRI->createVirtualRegister(RC);
    Register HiReg = MRI->createVirtualRegister(RC);

    BuildMI(*BB, &I, DL, TII.get(Opcode), LoReg)
      .addImm(Imm.trunc(32).getZExtValue());

    BuildMI(*BB, &I, DL, TII.get(Opcode), HiReg)
      .addImm(Imm.ashr(32).getZExtValue());

    ResInst = BuildMI(*BB, &I, DL, TII.get(AMDGPU::REG_SEQUENCE), DstReg)
      .addReg(LoReg)
      .addImm(AMDGPU::sub0)
      .addReg(HiReg)
      .addImm(AMDGPU::sub1);
  }

  // We can't call constrainSelectedInstRegOperands here, because it doesn't
  // work for target independent opcodes
  I.eraseFromParent();
  const TargetRegisterClass *DstRC =
    TRI.getConstrainedRegClassForOperand(ResInst->getOperand(0), *MRI);
  if (!DstRC)
    return true;
  return RBI.constrainGenericRegister(DstReg, *DstRC, *MRI);
}

static bool isConstant(const MachineInstr &MI) {
  return MI.getOpcode() == TargetOpcode::G_CONSTANT;
}

void AMDGPUInstructionSelector::getAddrModeInfo(const MachineInstr &Load,
    const MachineRegisterInfo &MRI, SmallVectorImpl<GEPInfo> &AddrInfo) const {

  const MachineInstr *PtrMI = MRI.getUniqueVRegDef(Load.getOperand(1).getReg());

  assert(PtrMI);

  if (PtrMI->getOpcode() != TargetOpcode::G_PTR_ADD)
    return;

  GEPInfo GEPInfo(*PtrMI);

  for (unsigned i = 1; i != 3; ++i) {
    const MachineOperand &GEPOp = PtrMI->getOperand(i);
    const MachineInstr *OpDef = MRI.getUniqueVRegDef(GEPOp.getReg());
    assert(OpDef);
    if (i == 2 && isConstant(*OpDef)) {
      // TODO: Could handle constant base + variable offset, but a combine
      // probably should have commuted it.
      assert(GEPInfo.Imm == 0);
      GEPInfo.Imm = OpDef->getOperand(1).getCImm()->getSExtValue();
      continue;
    }
    const RegisterBank *OpBank = RBI.getRegBank(GEPOp.getReg(), MRI, TRI);
    if (OpBank->getID() == AMDGPU::SGPRRegBankID)
      GEPInfo.SgprParts.push_back(GEPOp.getReg());
    else
      GEPInfo.VgprParts.push_back(GEPOp.getReg());
  }

  AddrInfo.push_back(GEPInfo);
  getAddrModeInfo(*PtrMI, MRI, AddrInfo);
}

bool AMDGPUInstructionSelector::isInstrUniform(const MachineInstr &MI) const {
  if (!MI.hasOneMemOperand())
    return false;

  const MachineMemOperand *MMO = *MI.memoperands_begin();
  const Value *Ptr = MMO->getValue();

  // UndefValue means this is a load of a kernel input.  These are uniform.
  // Sometimes LDS instructions have constant pointers.
  // If Ptr is null, then that means this mem operand contains a
  // PseudoSourceValue like GOT.
  if (!Ptr || isa<UndefValue>(Ptr) || isa<Argument>(Ptr) ||
      isa<Constant>(Ptr) || isa<GlobalValue>(Ptr))
    return true;

  if (MMO->getAddrSpace() == AMDGPUAS::CONSTANT_ADDRESS_32BIT)
    return true;

  const Instruction *I = dyn_cast<Instruction>(Ptr);
  return I && I->getMetadata("amdgpu.uniform");
}

bool AMDGPUInstructionSelector::hasVgprParts(ArrayRef<GEPInfo> AddrInfo) const {
  for (const GEPInfo &GEPInfo : AddrInfo) {
    if (!GEPInfo.VgprParts.empty())
      return true;
  }
  return false;
}

void AMDGPUInstructionSelector::initM0(MachineInstr &I) const {
  MachineBasicBlock *BB = I.getParent();

  const LLT PtrTy = MRI->getType(I.getOperand(1).getReg());
  unsigned AS = PtrTy.getAddressSpace();
  if ((AS == AMDGPUAS::LOCAL_ADDRESS || AS == AMDGPUAS::REGION_ADDRESS) &&
      STI.ldsRequiresM0Init()) {
    // If DS instructions require M0 initializtion, insert it before selecting.
    BuildMI(*BB, &I, I.getDebugLoc(), TII.get(AMDGPU::S_MOV_B32), AMDGPU::M0)
      .addImm(-1);
  }
}

bool AMDGPUInstructionSelector::selectG_LOAD_ATOMICRMW(MachineInstr &I) const {
  initM0(I);
  return selectImpl(I, *CoverageInfo);
}

bool AMDGPUInstructionSelector::selectG_BRCOND(MachineInstr &I) const {
  MachineBasicBlock *BB = I.getParent();
  MachineOperand &CondOp = I.getOperand(0);
  Register CondReg = CondOp.getReg();
  const DebugLoc &DL = I.getDebugLoc();

  unsigned BrOpcode;
  Register CondPhysReg;
  const TargetRegisterClass *ConstrainRC;

  // In SelectionDAG, we inspect the IR block for uniformity metadata to decide
  // whether the branch is uniform when selecting the instruction. In
  // GlobalISel, we should push that decision into RegBankSelect. Assume for now
  // RegBankSelect knows what it's doing if the branch condition is scc, even
  // though it currently does not.
  if (!isVCC(CondReg, *MRI)) {
    if (MRI->getType(CondReg) != LLT::scalar(32))
      return false;

    CondPhysReg = AMDGPU::SCC;
    BrOpcode = AMDGPU::S_CBRANCH_SCC1;
    // FIXME: Hack for isSCC tests
    ConstrainRC = &AMDGPU::SGPR_32RegClass;
  } else {
    // FIXME: Do we have to insert an and with exec here, like in SelectionDAG?
    // We sort of know that a VCC producer based on the register bank, that ands
    // inactive lanes with 0. What if there was a logical operation with vcc
    // producers in different blocks/with different exec masks?
    // FIXME: Should scc->vcc copies and with exec?
    CondPhysReg = TRI.getVCC();
    BrOpcode = AMDGPU::S_CBRANCH_VCCNZ;
    ConstrainRC = TRI.getBoolRC();
  }

  if (!MRI->getRegClassOrNull(CondReg))
    MRI->setRegClass(CondReg, ConstrainRC);

  BuildMI(*BB, &I, DL, TII.get(AMDGPU::COPY), CondPhysReg)
    .addReg(CondReg);
  BuildMI(*BB, &I, DL, TII.get(BrOpcode))
    .addMBB(I.getOperand(1).getMBB());

  I.eraseFromParent();
  return true;
}

bool AMDGPUInstructionSelector::selectG_FRAME_INDEX(MachineInstr &I) const {
  Register DstReg = I.getOperand(0).getReg();
  const RegisterBank *DstRB = RBI.getRegBank(DstReg, *MRI, TRI);
  const bool IsVGPR = DstRB->getID() == AMDGPU::VGPRRegBankID;
  I.setDesc(TII.get(IsVGPR ? AMDGPU::V_MOV_B32_e32 : AMDGPU::S_MOV_B32));
  if (IsVGPR)
    I.addOperand(*MF, MachineOperand::CreateReg(AMDGPU::EXEC, false, true));

  return RBI.constrainGenericRegister(
    DstReg, IsVGPR ? AMDGPU::VGPR_32RegClass : AMDGPU::SReg_32RegClass, *MRI);
}

bool AMDGPUInstructionSelector::selectG_PTR_MASK(MachineInstr &I) const {
  uint64_t Align = I.getOperand(2).getImm();
  const uint64_t Mask = ~((UINT64_C(1) << Align) - 1);

  MachineBasicBlock *BB = I.getParent();

  Register DstReg = I.getOperand(0).getReg();
  Register SrcReg = I.getOperand(1).getReg();

  const RegisterBank *DstRB = RBI.getRegBank(DstReg, *MRI, TRI);
  const RegisterBank *SrcRB = RBI.getRegBank(SrcReg, *MRI, TRI);
  const bool IsVGPR = DstRB->getID() == AMDGPU::VGPRRegBankID;
  unsigned NewOpc = IsVGPR ? AMDGPU::V_AND_B32_e64 : AMDGPU::S_AND_B32;
  unsigned MovOpc = IsVGPR ? AMDGPU::V_MOV_B32_e32 : AMDGPU::S_MOV_B32;
  const TargetRegisterClass &RegRC
    = IsVGPR ? AMDGPU::VGPR_32RegClass : AMDGPU::SReg_32RegClass;

  LLT Ty = MRI->getType(DstReg);

  const TargetRegisterClass *DstRC = TRI.getRegClassForTypeOnBank(Ty, *DstRB,
                                                                  *MRI);
  const TargetRegisterClass *SrcRC = TRI.getRegClassForTypeOnBank(Ty, *SrcRB,
                                                                  *MRI);
  if (!RBI.constrainGenericRegister(DstReg, *DstRC, *MRI) ||
      !RBI.constrainGenericRegister(SrcReg, *SrcRC, *MRI))
    return false;

  const DebugLoc &DL = I.getDebugLoc();
  Register ImmReg = MRI->createVirtualRegister(&RegRC);
  BuildMI(*BB, &I, DL, TII.get(MovOpc), ImmReg)
    .addImm(Mask);

  if (Ty.getSizeInBits() == 32) {
    BuildMI(*BB, &I, DL, TII.get(NewOpc), DstReg)
      .addReg(SrcReg)
      .addReg(ImmReg);
    I.eraseFromParent();
    return true;
  }

  Register HiReg = MRI->createVirtualRegister(&RegRC);
  Register LoReg = MRI->createVirtualRegister(&RegRC);
  Register MaskLo = MRI->createVirtualRegister(&RegRC);

  BuildMI(*BB, &I, DL, TII.get(AMDGPU::COPY), LoReg)
    .addReg(SrcReg, 0, AMDGPU::sub0);
  BuildMI(*BB, &I, DL, TII.get(AMDGPU::COPY), HiReg)
    .addReg(SrcReg, 0, AMDGPU::sub1);

  BuildMI(*BB, &I, DL, TII.get(NewOpc), MaskLo)
    .addReg(LoReg)
    .addReg(ImmReg);
  BuildMI(*BB, &I, DL, TII.get(AMDGPU::REG_SEQUENCE), DstReg)
    .addReg(MaskLo)
    .addImm(AMDGPU::sub0)
    .addReg(HiReg)
    .addImm(AMDGPU::sub1);
  I.eraseFromParent();
  return true;
}

/// Return the register to use for the index value, and the subregister to use
/// for the indirectly accessed register.
static std::pair<Register, unsigned>
computeIndirectRegIndex(MachineRegisterInfo &MRI,
                        const SIRegisterInfo &TRI,
                        const TargetRegisterClass *SuperRC,
                        Register IdxReg,
                        unsigned EltSize) {
  Register IdxBaseReg;
  int Offset;
  MachineInstr *Unused;

  std::tie(IdxBaseReg, Offset, Unused)
    = AMDGPU::getBaseWithConstantOffset(MRI, IdxReg);

  ArrayRef<int16_t> SubRegs = TRI.getRegSplitParts(SuperRC, EltSize);

  // Skip out of bounds offsets, or else we would end up using an undefined
  // register.
  if (static_cast<unsigned>(Offset) >= SubRegs.size())
    return std::make_pair(IdxReg, SubRegs[0]);
  return std::make_pair(IdxBaseReg, SubRegs[Offset]);
}

bool AMDGPUInstructionSelector::selectG_EXTRACT_VECTOR_ELT(
  MachineInstr &MI) const {
  Register DstReg = MI.getOperand(0).getReg();
  Register SrcReg = MI.getOperand(1).getReg();
  Register IdxReg = MI.getOperand(2).getReg();

  LLT DstTy = MRI->getType(DstReg);
  LLT SrcTy = MRI->getType(SrcReg);

  const RegisterBank *DstRB = RBI.getRegBank(DstReg, *MRI, TRI);
  const RegisterBank *SrcRB = RBI.getRegBank(SrcReg, *MRI, TRI);
  const RegisterBank *IdxRB = RBI.getRegBank(IdxReg, *MRI, TRI);

  // The index must be scalar. If it wasn't RegBankSelect should have moved this
  // into a waterfall loop.
  if (IdxRB->getID() != AMDGPU::SGPRRegBankID)
    return false;

  const TargetRegisterClass *SrcRC = TRI.getRegClassForTypeOnBank(SrcTy, *SrcRB,
                                                                  *MRI);
  const TargetRegisterClass *DstRC = TRI.getRegClassForTypeOnBank(DstTy, *DstRB,
                                                                  *MRI);
  if (!RBI.constrainGenericRegister(SrcReg, *SrcRC, *MRI) ||
      !RBI.constrainGenericRegister(DstReg, *DstRC, *MRI) ||
      !RBI.constrainGenericRegister(IdxReg, AMDGPU::SReg_32RegClass, *MRI))
    return false;

  MachineBasicBlock *BB = MI.getParent();
  const DebugLoc &DL = MI.getDebugLoc();
  const bool Is64 = DstTy.getSizeInBits() == 64;

  unsigned SubReg;
  std::tie(IdxReg, SubReg) = computeIndirectRegIndex(*MRI, TRI, SrcRC, IdxReg,
                                                     DstTy.getSizeInBits() / 8);

  if (SrcRB->getID() == AMDGPU::SGPRRegBankID) {
    if (DstTy.getSizeInBits() != 32 && !Is64)
      return false;

    BuildMI(*BB, &MI, DL, TII.get(AMDGPU::COPY), AMDGPU::M0)
      .addReg(IdxReg);

    unsigned Opc = Is64 ? AMDGPU::S_MOVRELS_B64 : AMDGPU::S_MOVRELS_B32;
    BuildMI(*BB, &MI, DL, TII.get(Opc), DstReg)
      .addReg(SrcReg, 0, SubReg)
      .addReg(SrcReg, RegState::Implicit);
    MI.eraseFromParent();
    return true;
  }

  if (SrcRB->getID() != AMDGPU::VGPRRegBankID || DstTy.getSizeInBits() != 32)
    return false;

  if (!STI.useVGPRIndexMode()) {
    BuildMI(*BB, &MI, DL, TII.get(AMDGPU::COPY), AMDGPU::M0)
      .addReg(IdxReg);
    BuildMI(*BB, &MI, DL, TII.get(AMDGPU::V_MOVRELS_B32_e32), DstReg)
      .addReg(SrcReg, RegState::Undef, SubReg)
      .addReg(SrcReg, RegState::Implicit);
    MI.eraseFromParent();
    return true;
  }

  BuildMI(*BB, MI, DL, TII.get(AMDGPU::S_SET_GPR_IDX_ON))
    .addReg(IdxReg)
    .addImm(AMDGPU::VGPRIndexMode::SRC0_ENABLE);
  BuildMI(*BB, MI, DL, TII.get(AMDGPU::V_MOV_B32_e32), DstReg)
    .addReg(SrcReg, RegState::Undef, SubReg)
    .addReg(SrcReg, RegState::Implicit)
    .addReg(AMDGPU::M0, RegState::Implicit);
  BuildMI(*BB, MI, DL, TII.get(AMDGPU::S_SET_GPR_IDX_OFF));

  MI.eraseFromParent();
  return true;
}

// TODO: Fold insert_vector_elt (extract_vector_elt) into movrelsd
bool AMDGPUInstructionSelector::selectG_INSERT_VECTOR_ELT(
  MachineInstr &MI) const {
  Register DstReg = MI.getOperand(0).getReg();
  Register VecReg = MI.getOperand(1).getReg();
  Register ValReg = MI.getOperand(2).getReg();
  Register IdxReg = MI.getOperand(3).getReg();

  LLT VecTy = MRI->getType(DstReg);
  LLT ValTy = MRI->getType(ValReg);
  unsigned VecSize = VecTy.getSizeInBits();
  unsigned ValSize = ValTy.getSizeInBits();

  const RegisterBank *VecRB = RBI.getRegBank(VecReg, *MRI, TRI);
  const RegisterBank *ValRB = RBI.getRegBank(ValReg, *MRI, TRI);
  const RegisterBank *IdxRB = RBI.getRegBank(IdxReg, *MRI, TRI);

  assert(VecTy.getElementType() == ValTy);

  // The index must be scalar. If it wasn't RegBankSelect should have moved this
  // into a waterfall loop.
  if (IdxRB->getID() != AMDGPU::SGPRRegBankID)
    return false;

  const TargetRegisterClass *VecRC = TRI.getRegClassForTypeOnBank(VecTy, *VecRB,
                                                                  *MRI);
  const TargetRegisterClass *ValRC = TRI.getRegClassForTypeOnBank(ValTy, *ValRB,
                                                                  *MRI);

  if (!RBI.constrainGenericRegister(VecReg, *VecRC, *MRI) ||
      !RBI.constrainGenericRegister(DstReg, *VecRC, *MRI) ||
      !RBI.constrainGenericRegister(ValReg, *ValRC, *MRI) ||
      !RBI.constrainGenericRegister(IdxReg, AMDGPU::SReg_32RegClass, *MRI))
    return false;

  if (VecRB->getID() == AMDGPU::VGPRRegBankID && ValSize != 32)
    return false;

  unsigned SubReg;
  std::tie(IdxReg, SubReg) = computeIndirectRegIndex(*MRI, TRI, VecRC, IdxReg,
                                                     ValSize / 8);

  const bool IndexMode = VecRB->getID() == AMDGPU::VGPRRegBankID &&
                         STI.useVGPRIndexMode();

  MachineBasicBlock *BB = MI.getParent();
  const DebugLoc &DL = MI.getDebugLoc();

  if (IndexMode) {
    BuildMI(*BB, MI, DL, TII.get(AMDGPU::S_SET_GPR_IDX_ON))
      .addReg(IdxReg)
      .addImm(AMDGPU::VGPRIndexMode::DST_ENABLE);
  } else {
    BuildMI(*BB, &MI, DL, TII.get(AMDGPU::COPY), AMDGPU::M0)
      .addReg(IdxReg);
  }

  const MCInstrDesc &RegWriteOp
    = TII.getIndirectRegWritePseudo(VecSize, ValSize,
                                    VecRB->getID() == AMDGPU::SGPRRegBankID);
  BuildMI(*BB, MI, DL, RegWriteOp, DstReg)
    .addReg(VecReg)
    .addReg(ValReg)
    .addImm(SubReg);

  if (IndexMode)
    BuildMI(*BB, MI, DL, TII.get(AMDGPU::S_SET_GPR_IDX_OFF));

  MI.eraseFromParent();
  return true;
}

bool AMDGPUInstructionSelector::select(MachineInstr &I) {
  if (I.isPHI())
    return selectPHI(I);

  if (!I.isPreISelOpcode()) {
    if (I.isCopy())
      return selectCOPY(I);
    return true;
  }

  switch (I.getOpcode()) {
  case TargetOpcode::G_AND:
  case TargetOpcode::G_OR:
  case TargetOpcode::G_XOR:
    if (selectG_AND_OR_XOR(I))
      return true;
    return selectImpl(I, *CoverageInfo);
  case TargetOpcode::G_ADD:
  case TargetOpcode::G_SUB:
    if (selectImpl(I, *CoverageInfo))
      return true;
    return selectG_ADD_SUB(I);
  case TargetOpcode::G_UADDO:
  case TargetOpcode::G_USUBO:
  case TargetOpcode::G_UADDE:
  case TargetOpcode::G_USUBE:
    return selectG_UADDO_USUBO_UADDE_USUBE(I);
  case TargetOpcode::G_INTTOPTR:
  case TargetOpcode::G_BITCAST:
  case TargetOpcode::G_PTRTOINT:
    return selectCOPY(I);
  case TargetOpcode::G_CONSTANT:
  case TargetOpcode::G_FCONSTANT:
    return selectG_CONSTANT(I);
  case TargetOpcode::G_EXTRACT:
    return selectG_EXTRACT(I);
  case TargetOpcode::G_MERGE_VALUES:
  case TargetOpcode::G_BUILD_VECTOR:
  case TargetOpcode::G_CONCAT_VECTORS:
    return selectG_MERGE_VALUES(I);
  case TargetOpcode::G_UNMERGE_VALUES:
    return selectG_UNMERGE_VALUES(I);
  case TargetOpcode::G_PTR_ADD:
    return selectG_PTR_ADD(I);
  case TargetOpcode::G_IMPLICIT_DEF:
    return selectG_IMPLICIT_DEF(I);
  case TargetOpcode::G_INSERT:
    return selectG_INSERT(I);
  case TargetOpcode::G_INTRINSIC:
    return selectG_INTRINSIC(I);
  case TargetOpcode::G_INTRINSIC_W_SIDE_EFFECTS:
    return selectG_INTRINSIC_W_SIDE_EFFECTS(I);
  case TargetOpcode::G_ICMP:
    if (selectG_ICMP(I))
      return true;
    return selectImpl(I, *CoverageInfo);
  case TargetOpcode::G_LOAD:
  case TargetOpcode::G_ATOMIC_CMPXCHG:
  case TargetOpcode::G_ATOMICRMW_XCHG:
  case TargetOpcode::G_ATOMICRMW_ADD:
  case TargetOpcode::G_ATOMICRMW_SUB:
  case TargetOpcode::G_ATOMICRMW_AND:
  case TargetOpcode::G_ATOMICRMW_OR:
  case TargetOpcode::G_ATOMICRMW_XOR:
  case TargetOpcode::G_ATOMICRMW_MIN:
  case TargetOpcode::G_ATOMICRMW_MAX:
  case TargetOpcode::G_ATOMICRMW_UMIN:
  case TargetOpcode::G_ATOMICRMW_UMAX:
  case TargetOpcode::G_ATOMICRMW_FADD:
    return selectG_LOAD_ATOMICRMW(I);
  case TargetOpcode::G_SELECT:
    return selectG_SELECT(I);
  case TargetOpcode::G_STORE:
    return selectG_STORE(I);
  case TargetOpcode::G_TRUNC:
    return selectG_TRUNC(I);
  case TargetOpcode::G_SEXT:
  case TargetOpcode::G_ZEXT:
  case TargetOpcode::G_ANYEXT:
    if (selectImpl(I, *CoverageInfo))
      return true;
    return selectG_SZA_EXT(I);
  case TargetOpcode::G_BRCOND:
    return selectG_BRCOND(I);
  case TargetOpcode::G_FRAME_INDEX:
    return selectG_FRAME_INDEX(I);
  case TargetOpcode::G_PTR_MASK:
    return selectG_PTR_MASK(I);
  case TargetOpcode::G_EXTRACT_VECTOR_ELT:
    return selectG_EXTRACT_VECTOR_ELT(I);
  case TargetOpcode::G_INSERT_VECTOR_ELT:
    return selectG_INSERT_VECTOR_ELT(I);
  case AMDGPU::G_AMDGPU_ATOMIC_INC:
  case AMDGPU::G_AMDGPU_ATOMIC_DEC:
    initM0(I);
    return selectImpl(I, *CoverageInfo);
  default:
    return selectImpl(I, *CoverageInfo);
  }
  return false;
}

InstructionSelector::ComplexRendererFns
AMDGPUInstructionSelector::selectVCSRC(MachineOperand &Root) const {
  return {{
      [=](MachineInstrBuilder &MIB) { MIB.add(Root); }
  }};

}

std::pair<Register, unsigned>
AMDGPUInstructionSelector::selectVOP3ModsImpl(
  Register Src) const {
  unsigned Mods = 0;
  MachineInstr *MI = MRI->getVRegDef(Src);

  if (MI && MI->getOpcode() == AMDGPU::G_FNEG) {
    Src = MI->getOperand(1).getReg();
    Mods |= SISrcMods::NEG;
    MI = MRI->getVRegDef(Src);
  }

  if (MI && MI->getOpcode() == AMDGPU::G_FABS) {
    Src = MI->getOperand(1).getReg();
    Mods |= SISrcMods::ABS;
  }

  return std::make_pair(Src, Mods);
}

///
/// This will select either an SGPR or VGPR operand and will save us from
/// having to write an extra tablegen pattern.
InstructionSelector::ComplexRendererFns
AMDGPUInstructionSelector::selectVSRC0(MachineOperand &Root) const {
  return {{
      [=](MachineInstrBuilder &MIB) { MIB.add(Root); }
  }};
}

InstructionSelector::ComplexRendererFns
AMDGPUInstructionSelector::selectVOP3Mods0(MachineOperand &Root) const {
  Register Src;
  unsigned Mods;
  std::tie(Src, Mods) = selectVOP3ModsImpl(Root.getReg());

  return {{
      [=](MachineInstrBuilder &MIB) { MIB.addReg(Src); },
      [=](MachineInstrBuilder &MIB) { MIB.addImm(Mods); }, // src0_mods
      [=](MachineInstrBuilder &MIB) { MIB.addImm(0); },    // clamp
      [=](MachineInstrBuilder &MIB) { MIB.addImm(0); }     // omod
  }};
}

InstructionSelector::ComplexRendererFns
AMDGPUInstructionSelector::selectVOP3OMods(MachineOperand &Root) const {
  return {{
      [=](MachineInstrBuilder &MIB) { MIB.add(Root); },
      [=](MachineInstrBuilder &MIB) { MIB.addImm(0); }, // clamp
      [=](MachineInstrBuilder &MIB) { MIB.addImm(0); }  // omod
  }};
}

InstructionSelector::ComplexRendererFns
AMDGPUInstructionSelector::selectVOP3Mods(MachineOperand &Root) const {
  Register Src;
  unsigned Mods;
  std::tie(Src, Mods) = selectVOP3ModsImpl(Root.getReg());

  return {{
      [=](MachineInstrBuilder &MIB) { MIB.addReg(Src); },
      [=](MachineInstrBuilder &MIB) { MIB.addImm(Mods); }  // src_mods
  }};
}

InstructionSelector::ComplexRendererFns
AMDGPUInstructionSelector::selectVOP3Mods_nnan(MachineOperand &Root) const {
  Register Src;
  unsigned Mods;
  std::tie(Src, Mods) = selectVOP3ModsImpl(Root.getReg());
  if (!TM.Options.NoNaNsFPMath && !isKnownNeverNaN(Src, *MRI))
    return None;

  return {{
      [=](MachineInstrBuilder &MIB) { MIB.addReg(Src); },
      [=](MachineInstrBuilder &MIB) { MIB.addImm(Mods); }  // src_mods
  }};
}

InstructionSelector::ComplexRendererFns
AMDGPUInstructionSelector::selectVOP3OpSelMods0(MachineOperand &Root) const {
  // FIXME: Handle clamp and op_sel
  return {{
      [=](MachineInstrBuilder &MIB) { MIB.addReg(Root.getReg()); },
      [=](MachineInstrBuilder &MIB) { MIB.addImm(0); }, // src_mods
      [=](MachineInstrBuilder &MIB) { MIB.addImm(0); }  // clamp
  }};
}

InstructionSelector::ComplexRendererFns
AMDGPUInstructionSelector::selectVOP3OpSelMods(MachineOperand &Root) const {
  // FIXME: Handle op_sel
  return {{
      [=](MachineInstrBuilder &MIB) { MIB.addReg(Root.getReg()); },
      [=](MachineInstrBuilder &MIB) { MIB.addImm(0); } // src_mods
  }};
}

InstructionSelector::ComplexRendererFns
AMDGPUInstructionSelector::selectSmrdImm(MachineOperand &Root) const {
  SmallVector<GEPInfo, 4> AddrInfo;
  getAddrModeInfo(*Root.getParent(), *MRI, AddrInfo);

  if (AddrInfo.empty() || AddrInfo[0].SgprParts.size() != 1)
    return None;

  const GEPInfo &GEPInfo = AddrInfo[0];

  if (!AMDGPU::isLegalSMRDImmOffset(STI, GEPInfo.Imm))
    return None;

  unsigned PtrReg = GEPInfo.SgprParts[0];
  int64_t EncodedImm = AMDGPU::getSMRDEncodedOffset(STI, GEPInfo.Imm);
  return {{
    [=](MachineInstrBuilder &MIB) { MIB.addReg(PtrReg); },
    [=](MachineInstrBuilder &MIB) { MIB.addImm(EncodedImm); }
  }};
}

InstructionSelector::ComplexRendererFns
AMDGPUInstructionSelector::selectSmrdImm32(MachineOperand &Root) const {
  SmallVector<GEPInfo, 4> AddrInfo;
  getAddrModeInfo(*Root.getParent(), *MRI, AddrInfo);

  if (AddrInfo.empty() || AddrInfo[0].SgprParts.size() != 1)
    return None;

  const GEPInfo &GEPInfo = AddrInfo[0];
  unsigned PtrReg = GEPInfo.SgprParts[0];
  int64_t EncodedImm = AMDGPU::getSMRDEncodedOffset(STI, GEPInfo.Imm);
  if (!isUInt<32>(EncodedImm))
    return None;

  return {{
    [=](MachineInstrBuilder &MIB) { MIB.addReg(PtrReg); },
    [=](MachineInstrBuilder &MIB) { MIB.addImm(EncodedImm); }
  }};
}

InstructionSelector::ComplexRendererFns
AMDGPUInstructionSelector::selectSmrdSgpr(MachineOperand &Root) const {
  MachineInstr *MI = Root.getParent();
  MachineBasicBlock *MBB = MI->getParent();

  SmallVector<GEPInfo, 4> AddrInfo;
  getAddrModeInfo(*MI, *MRI, AddrInfo);

  // FIXME: We should shrink the GEP if the offset is known to be <= 32-bits,
  // then we can select all ptr + 32-bit offsets not just immediate offsets.
  if (AddrInfo.empty() || AddrInfo[0].SgprParts.size() != 1)
    return None;

  const GEPInfo &GEPInfo = AddrInfo[0];
  if (!GEPInfo.Imm || !isUInt<32>(GEPInfo.Imm))
    return None;

  // If we make it this far we have a load with an 32-bit immediate offset.
  // It is OK to select this using a sgpr offset, because we have already
  // failed trying to select this load into one of the _IMM variants since
  // the _IMM Patterns are considered before the _SGPR patterns.
  unsigned PtrReg = GEPInfo.SgprParts[0];
  Register OffsetReg = MRI->createVirtualRegister(&AMDGPU::SReg_32RegClass);
  BuildMI(*MBB, MI, MI->getDebugLoc(), TII.get(AMDGPU::S_MOV_B32), OffsetReg)
          .addImm(GEPInfo.Imm);
  return {{
    [=](MachineInstrBuilder &MIB) { MIB.addReg(PtrReg); },
    [=](MachineInstrBuilder &MIB) { MIB.addReg(OffsetReg); }
  }};
}

template <bool Signed>
InstructionSelector::ComplexRendererFns
AMDGPUInstructionSelector::selectFlatOffsetImpl(MachineOperand &Root) const {
  MachineInstr *MI = Root.getParent();

  InstructionSelector::ComplexRendererFns Default = {{
      [=](MachineInstrBuilder &MIB) { MIB.addReg(Root.getReg()); },
      [=](MachineInstrBuilder &MIB) { MIB.addImm(0); },  // offset
      [=](MachineInstrBuilder &MIB) { MIB.addImm(0); }  // slc
    }};

  if (!STI.hasFlatInstOffsets())
    return Default;

  const MachineInstr *OpDef = MRI->getVRegDef(Root.getReg());
  if (!OpDef || OpDef->getOpcode() != AMDGPU::G_PTR_ADD)
    return Default;

  Optional<int64_t> Offset =
    getConstantVRegVal(OpDef->getOperand(2).getReg(), *MRI);
  if (!Offset.hasValue())
    return Default;

  unsigned AddrSpace = (*MI->memoperands_begin())->getAddrSpace();
  if (!TII.isLegalFLATOffset(Offset.getValue(), AddrSpace, Signed))
    return Default;

  Register BasePtr = OpDef->getOperand(1).getReg();

  return {{
      [=](MachineInstrBuilder &MIB) { MIB.addReg(BasePtr); },
      [=](MachineInstrBuilder &MIB) { MIB.addImm(Offset.getValue()); },
      [=](MachineInstrBuilder &MIB) { MIB.addImm(0); }  // slc
    }};
}

InstructionSelector::ComplexRendererFns
AMDGPUInstructionSelector::selectFlatOffset(MachineOperand &Root) const {
  return selectFlatOffsetImpl<false>(Root);
}

InstructionSelector::ComplexRendererFns
AMDGPUInstructionSelector::selectFlatOffsetSigned(MachineOperand &Root) const {
  return selectFlatOffsetImpl<true>(Root);
}

static bool isStackPtrRelative(const MachinePointerInfo &PtrInfo) {
  auto PSV = PtrInfo.V.dyn_cast<const PseudoSourceValue *>();
  return PSV && PSV->isStack();
}

InstructionSelector::ComplexRendererFns
AMDGPUInstructionSelector::selectMUBUFScratchOffen(MachineOperand &Root) const {
  MachineInstr *MI = Root.getParent();
  MachineBasicBlock *MBB = MI->getParent();
  MachineFunction *MF = MBB->getParent();
  const SIMachineFunctionInfo *Info = MF->getInfo<SIMachineFunctionInfo>();

  int64_t Offset = 0;
  if (mi_match(Root.getReg(), *MRI, m_ICst(Offset))) {
    Register HighBits = MRI->createVirtualRegister(&AMDGPU::VGPR_32RegClass);

    // TODO: Should this be inside the render function? The iterator seems to
    // move.
    BuildMI(*MBB, MI, MI->getDebugLoc(), TII.get(AMDGPU::V_MOV_B32_e32),
            HighBits)
      .addImm(Offset & ~4095);

    return {{[=](MachineInstrBuilder &MIB) { // rsrc
               MIB.addReg(Info->getScratchRSrcReg());
             },
             [=](MachineInstrBuilder &MIB) { // vaddr
               MIB.addReg(HighBits);
             },
             [=](MachineInstrBuilder &MIB) { // soffset
               const MachineMemOperand *MMO = *MI->memoperands_begin();
               const MachinePointerInfo &PtrInfo = MMO->getPointerInfo();

               Register SOffsetReg = isStackPtrRelative(PtrInfo)
                                         ? Info->getStackPtrOffsetReg()
                                         : Info->getScratchWaveOffsetReg();
               MIB.addReg(SOffsetReg);
             },
             [=](MachineInstrBuilder &MIB) { // offset
               MIB.addImm(Offset & 4095);
             }}};
  }

  assert(Offset == 0);

  // Try to fold a frame index directly into the MUBUF vaddr field, and any
  // offsets.
  Optional<int> FI;
  Register VAddr = Root.getReg();
  if (const MachineInstr *RootDef = MRI->getVRegDef(Root.getReg())) {
    if (isBaseWithConstantOffset(Root, *MRI)) {
      const MachineOperand &LHS = RootDef->getOperand(1);
      const MachineOperand &RHS = RootDef->getOperand(2);
      const MachineInstr *LHSDef = MRI->getVRegDef(LHS.getReg());
      const MachineInstr *RHSDef = MRI->getVRegDef(RHS.getReg());
      if (LHSDef && RHSDef) {
        int64_t PossibleOffset =
            RHSDef->getOperand(1).getCImm()->getSExtValue();
        if (SIInstrInfo::isLegalMUBUFImmOffset(PossibleOffset) &&
            (!STI.privateMemoryResourceIsRangeChecked() ||
             KnownBits->signBitIsZero(LHS.getReg()))) {
          if (LHSDef->getOpcode() == AMDGPU::G_FRAME_INDEX)
            FI = LHSDef->getOperand(1).getIndex();
          else
            VAddr = LHS.getReg();
          Offset = PossibleOffset;
        }
      }
    } else if (RootDef->getOpcode() == AMDGPU::G_FRAME_INDEX) {
      FI = RootDef->getOperand(1).getIndex();
    }
  }

  // If we don't know this private access is a local stack object, it needs to
  // be relative to the entry point's scratch wave offset register.
  // TODO: Should split large offsets that don't fit like above.
  // TODO: Don't use scratch wave offset just because the offset didn't fit.
  Register SOffset = FI.hasValue() ? Info->getStackPtrOffsetReg()
                                   : Info->getScratchWaveOffsetReg();

  return {{[=](MachineInstrBuilder &MIB) { // rsrc
             MIB.addReg(Info->getScratchRSrcReg());
           },
           [=](MachineInstrBuilder &MIB) { // vaddr
             if (FI.hasValue())
               MIB.addFrameIndex(FI.getValue());
             else
               MIB.addReg(VAddr);
           },
           [=](MachineInstrBuilder &MIB) { // soffset
             MIB.addReg(SOffset);
           },
           [=](MachineInstrBuilder &MIB) { // offset
             MIB.addImm(Offset);
           }}};
}

bool AMDGPUInstructionSelector::isDSOffsetLegal(Register Base,
                                                int64_t Offset,
                                                unsigned OffsetBits) const {
  if ((OffsetBits == 16 && !isUInt<16>(Offset)) ||
      (OffsetBits == 8 && !isUInt<8>(Offset)))
    return false;

  if (STI.hasUsableDSOffset() || STI.unsafeDSOffsetFoldingEnabled())
    return true;

  // On Southern Islands instruction with a negative base value and an offset
  // don't seem to work.
  return KnownBits->signBitIsZero(Base);
}

InstructionSelector::ComplexRendererFns
AMDGPUInstructionSelector::selectMUBUFScratchOffset(
    MachineOperand &Root) const {
  MachineInstr *MI = Root.getParent();
  MachineBasicBlock *MBB = MI->getParent();

  int64_t Offset = 0;
  if (!mi_match(Root.getReg(), *MRI, m_ICst(Offset)) ||
      !SIInstrInfo::isLegalMUBUFImmOffset(Offset))
    return {};

  const MachineFunction *MF = MBB->getParent();
  const SIMachineFunctionInfo *Info = MF->getInfo<SIMachineFunctionInfo>();
  const MachineMemOperand *MMO = *MI->memoperands_begin();
  const MachinePointerInfo &PtrInfo = MMO->getPointerInfo();

  Register SOffsetReg = isStackPtrRelative(PtrInfo)
                            ? Info->getStackPtrOffsetReg()
                            : Info->getScratchWaveOffsetReg();
  return {{
      [=](MachineInstrBuilder &MIB) {
        MIB.addReg(Info->getScratchRSrcReg());
      },                                                         // rsrc
      [=](MachineInstrBuilder &MIB) { MIB.addReg(SOffsetReg); }, // soffset
      [=](MachineInstrBuilder &MIB) { MIB.addImm(Offset); }      // offset
  }};
}

std::pair<Register, unsigned>
AMDGPUInstructionSelector::selectDS1Addr1OffsetImpl(MachineOperand &Root) const {
  const MachineInstr *RootDef = MRI->getVRegDef(Root.getReg());
  if (!RootDef)
    return std::make_pair(Root.getReg(), 0);

  int64_t ConstAddr = 0;
  if (isBaseWithConstantOffset(Root, *MRI)) {
    const MachineOperand &LHS = RootDef->getOperand(1);
    const MachineOperand &RHS = RootDef->getOperand(2);
    const MachineInstr *LHSDef = MRI->getVRegDef(LHS.getReg());
    const MachineInstr *RHSDef = MRI->getVRegDef(RHS.getReg());
    if (LHSDef && RHSDef) {
      int64_t PossibleOffset =
        RHSDef->getOperand(1).getCImm()->getSExtValue();
      if (isDSOffsetLegal(LHS.getReg(), PossibleOffset, 16)) {
        // (add n0, c0)
        return std::make_pair(LHS.getReg(), PossibleOffset);
      }
    }
  } else if (RootDef->getOpcode() == AMDGPU::G_SUB) {
    // TODO


  } else if (mi_match(Root.getReg(), *MRI, m_ICst(ConstAddr))) {
    // TODO

  }

  return std::make_pair(Root.getReg(), 0);
}

InstructionSelector::ComplexRendererFns
AMDGPUInstructionSelector::selectDS1Addr1Offset(MachineOperand &Root) const {

  Register Reg;
  unsigned Offset;
  std::tie(Reg, Offset) = selectDS1Addr1OffsetImpl(Root);
  return {{
      [=](MachineInstrBuilder &MIB) { MIB.addReg(Reg); },
      [=](MachineInstrBuilder &MIB) { MIB.addImm(Offset); }
    }};
}

void AMDGPUInstructionSelector::renderTruncImm32(MachineInstrBuilder &MIB,
                                                 const MachineInstr &MI,
                                                 int OpIdx) const {
  assert(MI.getOpcode() == TargetOpcode::G_CONSTANT && OpIdx == -1 &&
         "Expected G_CONSTANT");
  Optional<int64_t> CstVal = getConstantVRegVal(MI.getOperand(0).getReg(), *MRI);
  assert(CstVal && "Expected constant value");
  MIB.addImm(CstVal.getValue());
}

void AMDGPUInstructionSelector::renderNegateImm(MachineInstrBuilder &MIB,
                                                const MachineInstr &MI,
                                                int OpIdx) const {
  assert(MI.getOpcode() == TargetOpcode::G_CONSTANT && OpIdx == -1 &&
         "Expected G_CONSTANT");
  MIB.addImm(-MI.getOperand(1).getCImm()->getSExtValue());
}

void AMDGPUInstructionSelector::renderBitcastImm(MachineInstrBuilder &MIB,
                                                 const MachineInstr &MI,
                                                 int OpIdx) const {
  assert(OpIdx == -1);

  const MachineOperand &Op = MI.getOperand(1);
  if (MI.getOpcode() == TargetOpcode::G_FCONSTANT)
    MIB.addImm(Op.getFPImm()->getValueAPF().bitcastToAPInt().getZExtValue());
  else {
    assert(MI.getOpcode() == TargetOpcode::G_CONSTANT && "Expected G_CONSTANT");
    MIB.addImm(Op.getCImm()->getSExtValue());
  }
}

void AMDGPUInstructionSelector::renderPopcntImm(MachineInstrBuilder &MIB,
                                                const MachineInstr &MI,
                                                int OpIdx) const {
  assert(MI.getOpcode() == TargetOpcode::G_CONSTANT && OpIdx == -1 &&
         "Expected G_CONSTANT");
  MIB.addImm(MI.getOperand(1).getCImm()->getValue().countPopulation());
}

/// This only really exists to satisfy DAG type checking machinery, so is a
/// no-op here.
void AMDGPUInstructionSelector::renderTruncTImm32(MachineInstrBuilder &MIB,
                                                  const MachineInstr &MI,
                                                  int OpIdx) const {
  MIB.addImm(MI.getOperand(OpIdx).getImm());
}

void AMDGPUInstructionSelector::renderTruncTImm16(MachineInstrBuilder &MIB,
                                                  const MachineInstr &MI,
                                                  int OpIdx) const {
  MIB.addImm(MI.getOperand(OpIdx).getImm());
}

void AMDGPUInstructionSelector::renderTruncTImm1(MachineInstrBuilder &MIB,
                                                 const MachineInstr &MI,
                                                 int OpIdx) const {
  MIB.addImm(MI.getOperand(OpIdx).getImm());
}

bool AMDGPUInstructionSelector::isInlineImmediate16(int64_t Imm) const {
  return AMDGPU::isInlinableLiteral16(Imm, STI.hasInv2PiInlineImm());
}

bool AMDGPUInstructionSelector::isInlineImmediate32(int64_t Imm) const {
  return AMDGPU::isInlinableLiteral32(Imm, STI.hasInv2PiInlineImm());
}

bool AMDGPUInstructionSelector::isInlineImmediate64(int64_t Imm) const {
  return AMDGPU::isInlinableLiteral64(Imm, STI.hasInv2PiInlineImm());
}

bool AMDGPUInstructionSelector::isInlineImmediate(const APFloat &Imm) const {
  return TII.isInlineConstant(Imm);
}