llvm/lib/Target/AMDGPU/AMDGPULowerVGPREncoding.cpp

//===- AMDGPULowerVGPREncoding.cpp - lower VGPRs above v255 ---------------===//
//
// 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
/// Lower VGPRs above first 256 on gfx1250.
///
/// The pass scans used VGPRs and inserts S_SET_VGPR_MSB instructions to switch
/// VGPR addressing mode. The mode change is effective until the next change.
/// This instruction provides high bits of a VGPR address for four of the
/// operands: vdst, src0, src1, and src2, or other 4 operands depending on the
/// instruction encoding. If bits are set they are added as MSB to the
/// corresponding operand VGPR number.
///
/// There is no need to replace actual register operands because encoding of the
/// high and low VGPRs is the same. I.e. v0 has the encoding 0x100, so does
/// v256. v1 has the encoding 0x101 and v257 has the same encoding. So high
/// VGPRs will survive until actual encoding and will result in a same actual
/// bit encoding.
///
/// As a result the pass only inserts S_SET_VGPR_MSB to provide an actual offset
/// to a VGPR address of the subseqent instructions. The InstPrinter will take
/// care of the printing a low VGPR instead of a high one. In prinicple this
/// shall be viable to print actual high VGPR numbers, but that would disagree
/// with a disasm printing and create a situation where asm text is not
/// deterministic.
///
/// This pass creates a convention where non-fall through basic blocks shall
/// start with all 4 MSBs zero. Otherwise a disassembly would not be readable.
/// An optimization here is possible but deemed not desirable because of the
/// readbility concerns.
///
/// Consequentially the ABI is set to expect all 4 MSBs to be zero on entry.
/// The pass must run very late in the pipeline to make sure no changes to VGPR
/// operands will be made after it.
//
//===----------------------------------------------------------------------===//

#include "AMDGPULowerVGPREncoding.h"
#include "AMDGPU.h"
#include "GCNSubtarget.h"
#include "SIDefines.h"
#include "SIInstrInfo.h"
#include "llvm/ADT/bit.h"
#include "llvm/CodeGen/MachineBasicBlock.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/MathExtras.h"

using namespace llvm;

#define DEBUG_TYPE "amdgpu-lower-vgpr-encoding"

namespace {

class AMDGPULowerVGPREncoding {
  static constexpr unsigned OpNum = 4;
  static constexpr unsigned BitsPerField = 2;
  static constexpr unsigned NumFields = 4;
  static constexpr unsigned ModeWidth = NumFields * BitsPerField;
  static constexpr unsigned ModeMask = (1 << ModeWidth) - 1;
  static constexpr unsigned VGPRMSBShift =
      llvm::countr_zero_constexpr<unsigned>(AMDGPU::Hwreg::DST_VGPR_MSB);

  struct OpMode {
    // No MSBs set means they are not required to be of a particular value.
    std::optional<unsigned> MSBits;

    bool update(const OpMode &New, bool &Rewritten) {
      bool Updated = false;
      if (New.MSBits) {
        if (*New.MSBits != MSBits.value_or(0)) {
          Updated = true;
          Rewritten |= MSBits.has_value();
        }
        MSBits = New.MSBits;
      }
      return Updated;
    }
  };

  struct ModeTy {
    OpMode Ops[OpNum];

    bool update(const ModeTy &New, bool &Rewritten) {
      bool Updated = false;
      for (unsigned I : seq(OpNum))
        Updated |= Ops[I].update(New.Ops[I], Rewritten);
      return Updated;
    }

    unsigned encode() const {
      // Layout: [src0 msb, src1 msb, src2 msb, dst msb].
      unsigned V = 0;
      for (const auto &[I, Op] : enumerate(Ops))
        V |= Op.MSBits.value_or(0) << (I * 2);
      return V;
    }

    void print(raw_ostream &OS) const {
      static const char *FieldNames[] = {"src0", "src1", "src2", "dst"};
      OS << '{';
      for (const auto &[I, Op] : enumerate(Ops)) {
        if (I)
          OS << ", ";
        OS << FieldNames[I] << '=';
        if (Op.MSBits)
          OS << *Op.MSBits;
        else
          OS << '?';
      }
      OS << '}';
    }

    // Check if this mode is compatible with required \p NewMode without
    // modification.
    bool isCompatible(const ModeTy NewMode) const {
      for (unsigned I : seq(OpNum)) {
        if (!NewMode.Ops[I].MSBits.has_value())
          continue;
        if (Ops[I].MSBits.value_or(0) != NewMode.Ops[I].MSBits.value_or(0))
          return false;
      }
      return true;
    }
  };

public:
  bool run(MachineFunction &MF);

private:
  const SIInstrInfo *TII;
  const SIRegisterInfo *TRI;

  // Current basic block.
  MachineBasicBlock *MBB;

  /// Most recent s_set_* instruction.
  MachineInstr *MostRecentModeSet;

  /// Current mode bits.
  ModeTy CurrentMode;

  /// Number of current hard clause instructions.
  unsigned ClauseLen;

  /// Number of hard clause instructions remaining.
  unsigned ClauseRemaining;

  /// Clause group breaks.
  unsigned ClauseBreaks;

  /// Last hard clause instruction.
  MachineInstr *Clause;

  // Remember whether XCNT is known to be zero because of an S_SET_VGPR_MSB
  // instruction that we inserted, which implicitly waits for XCNT==0.
  bool XCntIsZero;

  /// Insert mode change before \p I. \returns true if mode was changed.
  bool setMode(ModeTy NewMode, MachineBasicBlock::instr_iterator I);

  /// Reset mode to default.
  void resetMode(MachineBasicBlock::instr_iterator I) {
    ModeTy Mode;
    for (OpMode &Op : Mode.Ops)
      Op.MSBits = 0;
    setMode(Mode, I);
  }

  /// If \p MO references VGPRs, return the MSBs. Otherwise, return nullopt.
  std::optional<unsigned> getMSBs(const MachineOperand &MO) const;

  /// Handle single \p MI. \return true if changed.
  bool runOnMachineInstr(MachineInstr &MI);

  /// Compute the mode for a single \p MI given \p Ops operands
  /// bit mapping. Optionally takes second array \p Ops2 for VOPD.
  /// If provided and an operand from \p Ops is not a VGPR, then \p Ops2
  /// is checked.
  void computeMode(ModeTy &NewMode, const MachineInstr &MI,
                   const AMDGPU::OpName Ops[OpNum],
                   const AMDGPU::OpName *Ops2 = nullptr);

  /// Check if an instruction \p I is within a clause and returns a suitable
  /// iterator to insert mode change. It may also modify the S_CLAUSE
  /// instruction to extend it or drop the clause if it cannot be adjusted.
  MachineBasicBlock::instr_iterator
  handleClause(MachineBasicBlock::instr_iterator I);

  /// Check if an instruction \p I is immediately after another program state
  /// instruction which it cannot coissue with. If so, insert before that
  /// instruction to encourage more coissuing.
  MachineBasicBlock::instr_iterator
  handleCoissue(MachineBasicBlock::instr_iterator I);

  /// S_SET_VGPR_MSB immediately after S_SETREG_IMM32_B32 targeting MODE is
  /// silently dropped on GFX1250. When set, the next S_SET_VGPR_MSB insertion
  /// must be preceded by S_NOP to avoid the hazard.
  bool needNopBeforeSetVGPRMSB(MachineBasicBlock::instr_iterator I);

  /// Handle S_SETREG_IMM32_B32 targeting MODE register. On certain hardware,
  /// this instruction clobbers VGPR MSB bits[12:19], so we need to restore
  /// the current mode. \returns true if the instruction was modified or a
  /// new one was inserted.
  bool handleSetregMode(MachineInstr &MI);

  /// Update bits[12:19] of the imm operand in S_SETREG_IMM32_B32 to contain
  /// the VGPR MSB mode value. \returns true if the immediate was changed.
  bool updateSetregModeImm(MachineInstr &MI, int64_t ModeValue);
};

bool AMDGPULowerVGPREncoding::setMode(ModeTy NewMode,
                                      MachineBasicBlock::instr_iterator I) {
  LLVM_DEBUG({
    dbgs() << "  setMode: NewMode=";
    NewMode.print(dbgs());
    dbgs() << " CurrentMode=";
    CurrentMode.print(dbgs());
    dbgs() << " MostRecentModeSet=" << (MostRecentModeSet ? "yes" : "null");
    if (I != MBB->instr_end())
      dbgs() << " before: " << *I;
    else
      dbgs() << " at end\n";
  });

  // Record previous mode into high 8 bits of the immediate.
  int64_t OldModeBits = CurrentMode.encode() << ModeWidth;

  bool Rewritten = false;
  if (!CurrentMode.update(NewMode, Rewritten)) {
    LLVM_DEBUG(dbgs() << "    -> no change needed\n");
    return false;
  }

  LLVM_DEBUG(dbgs() << "    Rewritten=" << Rewritten << " after update\n");

  if (MostRecentModeSet && !Rewritten) {
    // Update MostRecentModeSet with the new mode.
    MachineOperand &Op = MostRecentModeSet->getOperand(0);
    // Carry old mode bits from the existing instruction.
    int64_t OldModeBits = Op.getImm() & (ModeMask << ModeWidth);
    Op.setImm(CurrentMode.encode() | OldModeBits);
    LLVM_DEBUG(dbgs() << "    -> piggybacked onto S_SET_VGPR_MSB: "
                      << *MostRecentModeSet);
    return true;
  }

  MachineBasicBlock::instr_iterator InsertPt = handleClause(I);
  InsertPt = handleCoissue(InsertPt);
  // Case 2 match in handleSetregMode: the setreg's imm[12:19] matched
  // current MSBs, but the next VALU needs different MSBs, so this
  // S_SET_VGPR_MSB would land right after the setreg. Insert S_NOP to
  // prevent it from being silently dropped.
  if (needNopBeforeSetVGPRMSB(InsertPt))
    BuildMI(*MBB, InsertPt, {}, TII->get(AMDGPU::S_NOP)).addImm(0);
  MostRecentModeSet =
      BuildMI(*MBB, InsertPt, {}, TII->get(AMDGPU::S_SET_VGPR_MSB))
          .addImm(NewMode.encode() | OldModeBits);
  LLVM_DEBUG(dbgs() << "    -> inserted new S_SET_VGPR_MSB: "
                    << *MostRecentModeSet);

  // If we inserted S_SET_VGPR_MSB early then XCNT should remain zero from the
  // insertion point to the current instruction. Remove any redundant
  // S_WAIT_XCNT instructions in that range.
  for (MachineInstr &MI : make_early_inc_range(make_range(InsertPt, I))) {
    assert(!SIInstrInfo::isVMEM(MI) && !SIInstrInfo::isSMRD(MI));
    if (MI.getOpcode() == AMDGPU::S_WAIT_XCNT)
      MI.eraseFromBundle();
  }
  XCntIsZero = true;

  CurrentMode = NewMode;
  return true;
}

std::optional<unsigned>
AMDGPULowerVGPREncoding::getMSBs(const MachineOperand &MO) const {
  if (!MO.isReg())
    return std::nullopt;

  MCRegister Reg = MO.getReg();
  const TargetRegisterClass *RC = TRI->getPhysRegBaseClass(Reg);
  if (!RC || !TRI->isVGPRClass(RC))
    return std::nullopt;

  unsigned Idx = TRI->getHWRegIndex(Reg);
  return Idx >> 8;
}

void AMDGPULowerVGPREncoding::computeMode(ModeTy &NewMode,
                                          const MachineInstr &MI,
                                          const AMDGPU::OpName Ops[OpNum],
                                          const AMDGPU::OpName *Ops2) {
  NewMode = {};

  for (unsigned I = 0; I < OpNum; ++I) {
    const MachineOperand *Op = TII->getNamedOperand(MI, Ops[I]);

    std::optional<unsigned> MSBits;
    if (Op)
      MSBits = getMSBs(*Op);

#if !defined(NDEBUG)
    if (MSBits.has_value() && Ops2) {
      const MachineOperand *Op2 = TII->getNamedOperand(MI, Ops2[I]);
      if (Op2) {
        std::optional<unsigned> MSBits2;
        MSBits2 = getMSBs(*Op2);
        if (MSBits2.has_value() && MSBits != MSBits2)
          llvm_unreachable("Invalid VOPD pair was created");
      }
    }
#endif

    if (!MSBits.has_value() && Ops2) {
      Op = TII->getNamedOperand(MI, Ops2[I]);
      if (Op)
        MSBits = getMSBs(*Op);
    }

    if (!MSBits.has_value())
      continue;

    // Skip tied uses of src2 of VOP2, these will be handled along with defs and
    // only vdst bit affects these operands. We cannot skip tied uses of VOP3,
    // these uses are real even if must match the vdst.
    if (Ops[I] == AMDGPU::OpName::src2 && !Op->isDef() && Op->isTied() &&
        (SIInstrInfo::isVOP2(MI) ||
         (SIInstrInfo::isVOP3(MI) &&
          TII->hasVALU32BitEncoding(MI.getOpcode()))))
      continue;

    NewMode.Ops[I].MSBits = MSBits.value();
  }
}

bool AMDGPULowerVGPREncoding::runOnMachineInstr(MachineInstr &MI) {
  auto Ops = AMDGPU::getVGPRLoweringOperandTables(MI.getDesc());
  if (Ops.first) {
    ModeTy NewMode;
    computeMode(NewMode, MI, Ops.first, Ops.second);
    LLVM_DEBUG({
      dbgs() << "  runOnMachineInstr: ";
      MI.print(dbgs());
      dbgs() << "    computed NewMode=";
      NewMode.print(dbgs());
      dbgs() << " compatible=" << CurrentMode.isCompatible(NewMode) << '\n';
    });
    if (!CurrentMode.isCompatible(NewMode) && MI.isCommutable() &&
        TII->commuteInstruction(MI)) {
      ModeTy NewModeCommuted;
      computeMode(NewModeCommuted, MI, Ops.first, Ops.second);
      LLVM_DEBUG({
        dbgs() << "    commuted NewMode=";
        NewModeCommuted.print(dbgs());
        dbgs() << " compatible=" << CurrentMode.isCompatible(NewModeCommuted)
               << '\n';
      });
      if (CurrentMode.isCompatible(NewModeCommuted)) {
        // Update CurrentMode with mode bits the commuted instruction relies on.
        // This prevents later instructions from piggybacking and corrupting
        // those bits (e.g., a nullopt src treated as 0 could be overwritten).
        bool Unused = false;
        CurrentMode.update(NewModeCommuted, Unused);
        // MI was modified by the commute above.
        return true;
      }
      // Commute back.
      if (!TII->commuteInstruction(MI))
        llvm_unreachable("Failed to restore commuted instruction.");
    }
    return setMode(NewMode, MI.getIterator());
  }
  assert(!TII->hasVGPRUses(MI) || MI.isMetaInstruction() || MI.isPseudo());
  return false;
}

MachineBasicBlock::instr_iterator
AMDGPULowerVGPREncoding::handleClause(MachineBasicBlock::instr_iterator I) {
  if (!ClauseRemaining)
    return I;

  // A clause cannot start with a special instruction, place it right before
  // the clause.
  if (ClauseRemaining == ClauseLen) {
    I = Clause->getPrevNode()->getIterator();
    assert(I->isBundle());
    return I;
  }

  // If a clause defines breaks each group cannot start with a mode change.
  // just drop the clause.
  if (ClauseBreaks) {
    Clause->eraseFromBundle();
    ClauseRemaining = 0;
    return I;
  }

  // Otherwise adjust a number of instructions in the clause if it fits.
  // If it does not clause will just become shorter. Since the length
  // recorded in the clause is one less, increment the length after the
  // update. Note that SIMM16[5:0] must be 1-62, not 0 or 63.
  if (ClauseLen < 63)
    Clause->getOperand(0).setImm(ClauseLen | (ClauseBreaks << 8));

  ++ClauseLen;

  return I;
}

MachineBasicBlock::instr_iterator
AMDGPULowerVGPREncoding::handleCoissue(MachineBasicBlock::instr_iterator I) {
  // "Program State instructions" are instructions which are used to control
  // operation of the GPU rather than performing arithmetic. Such instructions
  // have different coissuing rules w.r.t s_set_vgpr_msb.
  auto isProgramStateInstr = [this](MachineInstr *MI) {
    unsigned Opc = MI->getOpcode();
    return TII->isBarrier(Opc) || TII->isWaitcnt(Opc) ||
           Opc == AMDGPU::S_DELAY_ALU;
  };

  while (I != MBB->begin()) {
    auto Prev = std::prev(I);
    if (!isProgramStateInstr(&*Prev))
      return I;
    I = Prev;
  }

  return I;
}

/// Returns whether \p MI is a S_SETREG_IMM32_B32(MODE).
static bool isSetregMode(const MachineInstr &MI, const SIInstrInfo &TII) {
  if (MI.getOpcode() != AMDGPU::S_SETREG_IMM32_B32)
    return false;

  const MachineOperand *SIMM16Op =
      TII.getNamedOperand(MI, AMDGPU::OpName::simm16);
  auto [HwRegId, _Offset, _Size] =
      AMDGPU::Hwreg::HwregEncoding::decode(SIMM16Op->getImm());
  return HwRegId == AMDGPU::Hwreg::ID_MODE;
}

bool AMDGPULowerVGPREncoding::needNopBeforeSetVGPRMSB(
    MachineBasicBlock::instr_iterator I) {
  MachineBasicBlock *CurrentMBB = MBB;
  while (true) {
    // Walk the block backward until we hit a non-meta instruction or the
    // beginning of the block.
    while (I != CurrentMBB->instr_begin()) {
      I = std::prev(I);
      if (isSetregMode(*I, *TII))
        return true;
      if (!I->isMetaInstruction())
        return false;
    }

    // Look for a potential fallthrough predecessor block. When it ends with a
    // S_SETREG_IMM32_B32(MODE) we need to insert a S_NOP too. We assume that an
    // explicit jump to the current block from the block that would otherwise
    // have naturally fallen through to it will remain in the final assembly.
    CurrentMBB = CurrentMBB->getPrevNode();
    if (!CurrentMBB || !CurrentMBB->canFallThrough())
      return false;
    I = CurrentMBB->instr_end();
  }
  return false;
}

/// Convert mode value from S_SET_VGPR_MSB format to MODE register format.
/// S_SET_VGPR_MSB uses: (src0[0-1], src1[2-3], src2[4-5], dst[6-7])
/// MODE register uses:  (dst[0-1], src0[2-3], src1[4-5], src2[6-7])
/// This is a left rotation by 2 bits on an 8-bit value.
static int64_t convertModeToSetregFormat(int64_t Mode) {
  assert(isUInt<8>(Mode) && "Mode expected to be 8-bit");
  return llvm::rotl<uint8_t>(static_cast<uint8_t>(Mode), /*R=*/2);
}

bool AMDGPULowerVGPREncoding::updateSetregModeImm(MachineInstr &MI,
                                                  int64_t ModeValue) {
  assert(MI.getOpcode() == AMDGPU::S_SETREG_IMM32_B32);

  // Convert from S_SET_VGPR_MSB format to MODE register format
  int64_t SetregMode = convertModeToSetregFormat(ModeValue);

  MachineOperand *ImmOp = TII->getNamedOperand(MI, AMDGPU::OpName::imm);
  int64_t OldImm = ImmOp->getImm();
  // Note that Offset is ignored for mode bits here.
  int64_t NewImm = (OldImm & ~int64_t(AMDGPU::Hwreg::VGPR_MSB_MASK)) |
                   (SetregMode << VGPRMSBShift);
  ImmOp->setImm(NewImm);
  return NewImm != OldImm;
}

bool AMDGPULowerVGPREncoding::handleSetregMode(MachineInstr &MI) {
  using namespace AMDGPU::Hwreg;

  assert(MI.getOpcode() == AMDGPU::S_SETREG_IMM32_B32 &&
         "only S_SETREG_IMM32_B32 needs to be handled");

  LLVM_DEBUG(dbgs() << "  handleSetregMode: " << MI);

  MachineOperand *SIMM16Op = TII->getNamedOperand(MI, AMDGPU::OpName::simm16);
  assert(SIMM16Op && "SIMM16Op must be present");

  auto [HwRegId, Offset, Size] = HwregEncoding::decode(SIMM16Op->getImm());
  LLVM_DEBUG(dbgs() << "    HwRegId=" << HwRegId << " Offset=" << Offset
                    << " Size=" << Size << '\n');
  if (HwRegId != ID_MODE) {
    LLVM_DEBUG(dbgs() << "    -> not ID_MODE, skipping\n");
    return false;
  }

  // MostRecentModeSet is clobbered by SETREG and not relevant anymore.
  MostRecentModeSet = nullptr;

  int64_t ModeValue = CurrentMode.encode();
  LLVM_DEBUG({
    dbgs() << "    CurrentMode=";
    CurrentMode.print(dbgs());
    dbgs() << " encoded=0x" << Twine::utohexstr(ModeValue)
           << " VGPRMSBShift=" << VGPRMSBShift << '\n';
  });

  // Case 1: Size <= 12 - the original instruction uses imm32[0:Size-1], so
  // imm32[12:19] is unused, or Offset is zero and it is safe to set
  // imm32[12:19] to the correct VGPR MSBs.
  if (!Offset || Size <= VGPRMSBShift) {
    // Set imm32[12:19] to the correct VGPR MSBs.
    LLVM_DEBUG(dbgs() << "    Case 1: Size(" << Size << ") <= VGPRMSBShift("
                      << VGPRMSBShift << "), update mode bits[12:19]\n");
    bool Changed = updateSetregModeImm(MI, ModeValue);
    LLVM_DEBUG(dbgs() << "    -> " << MI);
    return Changed;
  }

  // Case 2: Size > 12 - the original instruction uses bits beyond 11, so we
  // cannot arbitrarily modify imm32[12:19]. Check if it already matches VGPR
  // MSBs. Note: imm32[12:19] is in MODE register format, while ModeValue is
  // in S_SET_VGPR_MSB format, so we need to convert before comparing.
  MachineOperand *ImmOp = TII->getNamedOperand(MI, AMDGPU::OpName::imm);
  assert(ImmOp && "ImmOp must be present");
  int64_t ImmBits12To19 = (ImmOp->getImm() & VGPR_MSB_MASK) >> VGPRMSBShift;
  int64_t SetregModeValue = convertModeToSetregFormat(ModeValue);
  LLVM_DEBUG(dbgs() << "    Case 2: Size(" << Size << ") > VGPRMSBShift, "
                    << "ImmBits12To19=0x" << Twine::utohexstr(ImmBits12To19)
                    << " SetregModeValue=0x"
                    << Twine::utohexstr(SetregModeValue) << '\n');
  if (ImmBits12To19 == SetregModeValue) {
    LLVM_DEBUG(dbgs() << "    -> bits[12:19] already correct\n");
    return false;
  }

  // imm32[12:19] doesn't match VGPR MSBs - insert s_set_vgpr_msb after
  // the original instruction to restore the correct value. Insert S_NOP
  // to avoid the GFX1250 hazard where S_SET_VGPR_MSB immediately after
  // S_SETREG_IMM32_B32(MODE) is silently dropped.
  MachineBasicBlock::iterator InsertPt = std::next(MI.getIterator());
  BuildMI(*MBB, InsertPt, MI.getDebugLoc(), TII->get(AMDGPU::S_NOP)).addImm(0);
  MostRecentModeSet = BuildMI(*MBB, InsertPt, MI.getDebugLoc(),
                              TII->get(AMDGPU::S_SET_VGPR_MSB))
                          .addImm(ModeValue | (ModeValue << ModeWidth));
  LLVM_DEBUG(dbgs() << "    -> inserted S_SET_VGPR_MSB after setreg: "
                    << *MostRecentModeSet);
  return true;
}

bool AMDGPULowerVGPREncoding::run(MachineFunction &MF) {
  const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>();
  if (!ST.has1024AddressableVGPRs())
    return false;

  TII = ST.getInstrInfo();
  TRI = ST.getRegisterInfo();

  LLVM_DEBUG(dbgs() << "*** AMDGPULowerVGPREncoding on " << MF.getName()
                    << " ***\n");

  bool Changed = false;
  ClauseLen = ClauseRemaining = 0;
  CurrentMode = {};
  for (auto &MBB : MF) {
    MostRecentModeSet = nullptr;
    XCntIsZero = false;
    this->MBB = &MBB;

    LLVM_DEBUG(dbgs() << "BB#" << MBB.getNumber() << ' ' << MBB.getName()
                      << ":\n");

    for (auto &MI : llvm::make_early_inc_range(MBB.instrs())) {
      if (MI.isMetaInstruction())
        continue;

      if (MI.isTerminator() || MI.isCall()) {
        LLVM_DEBUG(dbgs() << "  terminator/call: " << MI);
        if (MI.getOpcode() == AMDGPU::S_ENDPGM ||
            MI.getOpcode() == AMDGPU::S_ENDPGM_SAVED)
          CurrentMode = {};
        else
          resetMode(MI.getIterator());
        continue;
      }

      if (MI.isInlineAsm()) {
        LLVM_DEBUG(dbgs() << "  inline asm: " << MI);
        if (TII->hasVGPRUses(MI))
          resetMode(MI.getIterator());
        continue;
      }

      if (MI.getOpcode() == AMDGPU::S_CLAUSE) {
        assert(!ClauseRemaining && "Nested clauses are not supported");
        ClauseLen = MI.getOperand(0).getImm();
        ClauseBreaks = (ClauseLen >> 8) & 15;
        ClauseLen = ClauseRemaining = (ClauseLen & 63) + 1;
        Clause = &MI;
        LLVM_DEBUG(dbgs() << "  clause: len=" << ClauseLen
                          << " breaks=" << ClauseBreaks << '\n');
        continue;
      }

      if (MI.getOpcode() == AMDGPU::S_SETREG_IMM32_B32 &&
          ST.hasSetregVGPRMSBFixup()) {
        Changed |= handleSetregMode(MI);
        continue;
      }

      // If XCNT is known to be zero then any S_WAIT_XCNT instruction is
      // redundant and can be removed.
      if (MI.getOpcode() == AMDGPU::S_WAIT_XCNT && XCntIsZero) {
        MI.eraseFromBundle();
        Changed = true;
        continue;
      }

      Changed |= runOnMachineInstr(MI);

      // Any VMEM or SMEM instruction may increment XCNT.
      if (SIInstrInfo::isVMEM(MI) || SIInstrInfo::isSMRD(MI))
        XCntIsZero = false;

      if (ClauseRemaining)
        --ClauseRemaining;
    }

    // Reset the mode if we are falling through.
    LLVM_DEBUG(dbgs() << "  end of BB, resetting mode\n");
    resetMode(MBB.instr_end());
  }

  return Changed;
}

class AMDGPULowerVGPREncodingLegacy : public MachineFunctionPass {
public:
  static char ID;

  AMDGPULowerVGPREncodingLegacy() : MachineFunctionPass(ID) {}

  bool runOnMachineFunction(MachineFunction &MF) override {
    return AMDGPULowerVGPREncoding().run(MF);
  }

  void getAnalysisUsage(AnalysisUsage &AU) const override {
    AU.setPreservesCFG();
    MachineFunctionPass::getAnalysisUsage(AU);
  }
};

} // namespace

char AMDGPULowerVGPREncodingLegacy::ID = 0;

char &llvm::AMDGPULowerVGPREncodingLegacyID = AMDGPULowerVGPREncodingLegacy::ID;

INITIALIZE_PASS(AMDGPULowerVGPREncodingLegacy, DEBUG_TYPE,
                "AMDGPU Lower VGPR Encoding", false, false)

PreservedAnalyses
AMDGPULowerVGPREncodingPass::run(MachineFunction &MF,
                                 MachineFunctionAnalysisManager &MFAM) {
  if (!AMDGPULowerVGPREncoding().run(MF))
    return PreservedAnalyses::all();

  return getMachineFunctionPassPreservedAnalyses().preserveSet<CFGAnalyses>();
}