AMDGPUAsmParser.cpp

Go to the documentation of this file.

//===- AMDGPUAsmParser.cpp - Parse SI asm to MCInst instructions ----------===//
//
// 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
//
//===----------------------------------------------------------------------===//
 
#include "AMDKernelCodeT.h"
#include "MCTargetDesc/AMDGPUMCTargetDesc.h"
#include "MCTargetDesc/AMDGPUTargetStreamer.h"
#include "SIDefines.h"
#include "SIInstrInfo.h"
#include "SIRegisterInfo.h"
#include "TargetInfo/AMDGPUTargetInfo.h"
#include "Utils/AMDGPUAsmUtils.h"
#include "Utils/AMDGPUBaseInfo.h"
#include "Utils/AMDKernelCodeTUtils.h"
#include "llvm/ADT/APFloat.h"
#include "llvm/ADT/SmallBitVector.h"
#include "llvm/ADT/StringSet.h"
#include "llvm/ADT/Twine.h"
#include "llvm/BinaryFormat/ELF.h"
#include "llvm/MC/MCAsmInfo.h"
#include "llvm/MC/MCContext.h"
#include "llvm/MC/MCExpr.h"
#include "llvm/MC/MCInst.h"
#include "llvm/MC/MCInstrDesc.h"
#include "llvm/MC/MCParser/MCAsmLexer.h"
#include "llvm/MC/MCParser/MCAsmParser.h"
#include "llvm/MC/MCParser/MCParsedAsmOperand.h"
#include "llvm/MC/MCParser/MCTargetAsmParser.h"
#include "llvm/MC/MCSymbol.h"
#include "llvm/MC/TargetRegistry.h"
#include "llvm/Support/AMDGPUMetadata.h"
#include "llvm/Support/AMDHSAKernelDescriptor.h"
#include "llvm/Support/Casting.h"
#include "llvm/Support/MachineValueType.h"
#include "llvm/Support/MathExtras.h"
#include "llvm/TargetParser/TargetParser.h"
#include <optional>
 
using namespace llvm;
using namespace llvm::AMDGPU;
using namespace llvm::amdhsa;
 
namespace {
 
class AMDGPUAsmParser;
 
enum RegisterKind { IS_UNKNOWN, IS_VGPR, IS_SGPR, IS_AGPR, IS_TTMP, IS_SPECIAL };
 
//===----------------------------------------------------------------------===//
// Operand
//===----------------------------------------------------------------------===//
 
class AMDGPUOperand : public MCParsedAsmOperand {
  enum KindTy {
    Token,
    Immediate,
    Register,
    Expression
  } Kind;
 
  SMLoc StartLoc, EndLoc;
  const AMDGPUAsmParser *AsmParser;
 
public:
  AMDGPUOperand(KindTy Kind_, const AMDGPUAsmParser *AsmParser_)
      : Kind(Kind_), AsmParser(AsmParser_) {}
 
  using Ptr = std::unique_ptr<AMDGPUOperand>;
 
  struct Modifiers {
    bool Abs = false;
    bool Neg = false;
    bool Sext = false;
 
    bool hasFPModifiers() const { return Abs || Neg; }
    bool hasIntModifiers() const { return Sext; }
    bool hasModifiers() const { return hasFPModifiers() || hasIntModifiers(); }
 
    int64_t getFPModifiersOperand() const {
      int64_t Operand = 0;
      Operand |= Abs ? SISrcMods::ABS : 0u;
      Operand |= Neg ? SISrcMods::NEG : 0u;
      return Operand;
    }
 
    int64_t getIntModifiersOperand() const {
      int64_t Operand = 0;
      Operand |= Sext ? SISrcMods::SEXT : 0u;
      return Operand;
    }
 
    int64_t getModifiersOperand() const {
      assert(!(hasFPModifiers() && hasIntModifiers())
           && "fp and int modifiers should not be used simultaneously");
      if (hasFPModifiers()) {
        return getFPModifiersOperand();
      } else if (hasIntModifiers()) {
        return getIntModifiersOperand();
      } else {
        return 0;
      }
    }
 
    friend raw_ostream &operator <<(raw_ostream &OS, AMDGPUOperand::Modifiers Mods);
  };
 
  enum ImmTy {
    ImmTyNone,
    ImmTyGDS,
    ImmTyLDS,
    ImmTyOffen,
    ImmTyIdxen,
    ImmTyAddr64,
    ImmTyOffset,
    ImmTyInstOffset,
    ImmTyOffset0,
    ImmTyOffset1,
    ImmTySMEMOffsetMod,
    ImmTyCPol,
    ImmTySWZ,
    ImmTyTFE,
    ImmTyD16,
    ImmTyClampSI,
    ImmTyOModSI,
    ImmTySdwaDstSel,
    ImmTySdwaSrc0Sel,
    ImmTySdwaSrc1Sel,
    ImmTySdwaDstUnused,
    ImmTyDMask,
    ImmTyDim,
    ImmTyUNorm,
    ImmTyDA,
    ImmTyR128A16,
    ImmTyA16,
    ImmTyLWE,
    ImmTyExpTgt,
    ImmTyExpCompr,
    ImmTyExpVM,
    ImmTyFORMAT,
    ImmTyHwreg,
    ImmTyOff,
    ImmTySendMsg,
    ImmTyInterpSlot,
    ImmTyInterpAttr,
    ImmTyAttrChan,
    ImmTyOpSel,
    ImmTyOpSelHi,
    ImmTyNegLo,
    ImmTyNegHi,
    ImmTyDPP8,
    ImmTyDppCtrl,
    ImmTyDppRowMask,
    ImmTyDppBankMask,
    ImmTyDppBoundCtrl,
    ImmTyDppFi,
    ImmTySwizzle,
    ImmTyGprIdxMode,
    ImmTyHigh,
    ImmTyBLGP,
    ImmTyCBSZ,
    ImmTyABID,
    ImmTyEndpgm,
    ImmTyWaitVDST,
    ImmTyWaitEXP,
  };
 
  // Immediate operand kind.
  // It helps to identify the location of an offending operand after an error.
  // Note that regular literals and mandatory literals (KImm) must be handled
  // differently. When looking for an offending operand, we should usually
  // ignore mandatory literals because they are part of the instruction and
  // cannot be changed. Report location of mandatory operands only for VOPD,
  // when both OpX and OpY have a KImm and there are no other literals.
  enum ImmKindTy {
    ImmKindTyNone,
    ImmKindTyLiteral,
    ImmKindTyMandatoryLiteral,
    ImmKindTyConst,
  };
 
private:
  struct TokOp {
    const char *Data;
    unsigned Length;
  };
 
  struct ImmOp {
    int64_t Val;
    ImmTy Type;
    bool IsFPImm;
    mutable ImmKindTy Kind;
    Modifiers Mods;
  };
 
  struct RegOp {
    unsigned RegNo;
    Modifiers Mods;
  };
 
  union {
    TokOp Tok;
    ImmOp Imm;
    RegOp Reg;
    const MCExpr *Expr;
  };
 
public:
  bool isToken() const override { return Kind == Token; }
 
  bool isSymbolRefExpr() const {
    return isExpr() && Expr && isa<MCSymbolRefExpr>(Expr);
  }
 
  bool isImm() const override {
    return Kind == Immediate;
  }
 
  void setImmKindNone() const {
    assert(isImm());
    Imm.Kind = ImmKindTyNone;
  }
 
  void setImmKindLiteral() const {
    assert(isImm());
    Imm.Kind = ImmKindTyLiteral;
  }
 
  void setImmKindMandatoryLiteral() const {
    assert(isImm());
    Imm.Kind = ImmKindTyMandatoryLiteral;
  }
 
  void setImmKindConst() const {
    assert(isImm());
    Imm.Kind = ImmKindTyConst;
  }
 
  bool IsImmKindLiteral() const {
    return isImm() && Imm.Kind == ImmKindTyLiteral;
  }
 
  bool IsImmKindMandatoryLiteral() const {
    return isImm() && Imm.Kind == ImmKindTyMandatoryLiteral;
  }
 
  bool isImmKindConst() const {
    return isImm() && Imm.Kind == ImmKindTyConst;
  }
 
  bool isInlinableImm(MVT type) const;
  bool isLiteralImm(MVT type) const;
 
  bool isRegKind() const {
    return Kind == Register;
  }
 
  bool isReg() const override {
    return isRegKind() && !hasModifiers();
  }
 
  bool isRegOrInline(unsigned RCID, MVT type) const {
    return isRegClass(RCID) || isInlinableImm(type);
  }
 
  bool isRegOrImmWithInputMods(unsigned RCID, MVT type) const {
    return isRegOrInline(RCID, type) || isLiteralImm(type);
  }
 
  bool isRegOrImmWithInt16InputMods() const {
    return isRegOrImmWithInputMods(AMDGPU::VS_32RegClassID, MVT::i16);
  }
 
  bool isRegOrImmWithInt32InputMods() const {
    return isRegOrImmWithInputMods(AMDGPU::VS_32RegClassID, MVT::i32);
  }
 
  bool isRegOrInlineImmWithInt16InputMods() const {
    return isRegOrInline(AMDGPU::VS_32RegClassID, MVT::i16);
  }
 
  bool isRegOrInlineImmWithInt32InputMods() const {
    return isRegOrInline(AMDGPU::VS_32RegClassID, MVT::i32);
  }
 
  bool isRegOrImmWithInt64InputMods() const {
    return isRegOrImmWithInputMods(AMDGPU::VS_64RegClassID, MVT::i64);
  }
 
  bool isRegOrImmWithFP16InputMods() const {
    return isRegOrImmWithInputMods(AMDGPU::VS_32RegClassID, MVT::f16);
  }
 
  bool isRegOrImmWithFP32InputMods() const {
    return isRegOrImmWithInputMods(AMDGPU::VS_32RegClassID, MVT::f32);
  }
 
  bool isRegOrImmWithFP64InputMods() const {
    return isRegOrImmWithInputMods(AMDGPU::VS_64RegClassID, MVT::f64);
  }
 
  bool isRegOrInlineImmWithFP16InputMods() const {
    return isRegOrInline(AMDGPU::VS_32RegClassID, MVT::f16);
  }
 
  bool isRegOrInlineImmWithFP32InputMods() const {
    return isRegOrInline(AMDGPU::VS_32RegClassID, MVT::f32);
  }
 
 
  bool isVReg() const {
    return isRegClass(AMDGPU::VGPR_32RegClassID) ||
           isRegClass(AMDGPU::VReg_64RegClassID) ||
           isRegClass(AMDGPU::VReg_96RegClassID) ||
           isRegClass(AMDGPU::VReg_128RegClassID) ||
           isRegClass(AMDGPU::VReg_160RegClassID) ||
           isRegClass(AMDGPU::VReg_192RegClassID) ||
           isRegClass(AMDGPU::VReg_256RegClassID) ||
           isRegClass(AMDGPU::VReg_512RegClassID) ||
           isRegClass(AMDGPU::VReg_1024RegClassID);
  }
 
  bool isVReg32() const {
    return isRegClass(AMDGPU::VGPR_32RegClassID);
  }
 
  bool isVReg32OrOff() const {
    return isOff() || isVReg32();
  }
 
  bool isNull() const {
    return isRegKind() && getReg() == AMDGPU::SGPR_NULL;
  }
 
  bool isVRegWithInputMods() const;
  bool isT16VRegWithInputMods() const;
 
  bool isSDWAOperand(MVT type) const;
  bool isSDWAFP16Operand() const;
  bool isSDWAFP32Operand() const;
  bool isSDWAInt16Operand() const;
  bool isSDWAInt32Operand() const;
 
  bool isImmTy(ImmTy ImmT) const {
    return isImm() && Imm.Type == ImmT;
  }
 
  bool isImmLiteral() const { return isImmTy(ImmTyNone); }
 
  bool isImmModifier() const {
    return isImm() && Imm.Type != ImmTyNone;
  }
 
  bool isClampSI() const { return isImmTy(ImmTyClampSI); }
  bool isOModSI() const { return isImmTy(ImmTyOModSI); }
  bool isDMask() const { return isImmTy(ImmTyDMask); }
  bool isDim() const { return isImmTy(ImmTyDim); }
  bool isUNorm() const { return isImmTy(ImmTyUNorm); }
  bool isDA() const { return isImmTy(ImmTyDA); }
  bool isR128A16() const { return isImmTy(ImmTyR128A16); }
  bool isA16() const { return isImmTy(ImmTyA16); }
  bool isLWE() const { return isImmTy(ImmTyLWE); }
  bool isOff() const { return isImmTy(ImmTyOff); }
  bool isExpTgt() const { return isImmTy(ImmTyExpTgt); }
  bool isExpVM() const { return isImmTy(ImmTyExpVM); }
  bool isExpCompr() const { return isImmTy(ImmTyExpCompr); }
  bool isOffen() const { return isImmTy(ImmTyOffen); }
  bool isIdxen() const { return isImmTy(ImmTyIdxen); }
  bool isAddr64() const { return isImmTy(ImmTyAddr64); }
  bool isOffset() const { return isImmTy(ImmTyOffset) && isUInt<16>(getImm()); }
  bool isOffset0() const { return isImmTy(ImmTyOffset0) && isUInt<8>(getImm()); }
  bool isOffset1() const { return isImmTy(ImmTyOffset1) && isUInt<8>(getImm()); }
  bool isSMEMOffsetMod() const { return isImmTy(ImmTySMEMOffsetMod); }
  bool isFlatOffset() const { return isImmTy(ImmTyOffset) || isImmTy(ImmTyInstOffset); }
  bool isGDS() const { return isImmTy(ImmTyGDS); }
  bool isLDS() const { return isImmTy(ImmTyLDS); }
  bool isCPol() const { return isImmTy(ImmTyCPol); }
  bool isSWZ() const { return isImmTy(ImmTySWZ); }
  bool isTFE() const { return isImmTy(ImmTyTFE); }
  bool isD16() const { return isImmTy(ImmTyD16); }
  bool isFORMAT() const { return isImmTy(ImmTyFORMAT) && isUInt<7>(getImm()); }
  bool isBankMask() const { return isImmTy(ImmTyDppBankMask); }
  bool isRowMask() const { return isImmTy(ImmTyDppRowMask); }
  bool isDppBoundCtrl() const { return isImmTy(ImmTyDppBoundCtrl); }
  bool isFI() const { return isImmTy(ImmTyDppFi); }
  bool isSDWADstSel() const { return isImmTy(ImmTySdwaDstSel); }
  bool isSDWASrc0Sel() const { return isImmTy(ImmTySdwaSrc0Sel); }
  bool isSDWASrc1Sel() const { return isImmTy(ImmTySdwaSrc1Sel); }
  bool isSDWADstUnused() const { return isImmTy(ImmTySdwaDstUnused); }
  bool isInterpSlot() const { return isImmTy(ImmTyInterpSlot); }
  bool isInterpAttr() const { return isImmTy(ImmTyInterpAttr); }
  bool isAttrChan() const { return isImmTy(ImmTyAttrChan); }
  bool isOpSel() const { return isImmTy(ImmTyOpSel); }
  bool isOpSelHi() const { return isImmTy(ImmTyOpSelHi); }
  bool isNegLo() const { return isImmTy(ImmTyNegLo); }
  bool isNegHi() const { return isImmTy(ImmTyNegHi); }
  bool isHigh() const { return isImmTy(ImmTyHigh); }
 
  bool isRegOrImm() const {
    return isReg() || isImm();
  }
 
  bool isRegClass(unsigned RCID) const;
 
  bool isInlineValue() const;
 
  bool isRegOrInlineNoMods(unsigned RCID, MVT type) const {
    return isRegOrInline(RCID, type) && !hasModifiers();
  }
 
  bool isSCSrcB16() const {
    return isRegOrInlineNoMods(AMDGPU::SReg_32RegClassID, MVT::i16);
  }
 
  bool isSCSrcV2B16() const {
    return isSCSrcB16();
  }
 
  bool isSCSrcB32() const {
    return isRegOrInlineNoMods(AMDGPU::SReg_32RegClassID, MVT::i32);
  }
 
  bool isSCSrcB64() const {
    return isRegOrInlineNoMods(AMDGPU::SReg_64RegClassID, MVT::i64);
  }
 
  bool isBoolReg() const;
 
  bool isSCSrcF16() const {
    return isRegOrInlineNoMods(AMDGPU::SReg_32RegClassID, MVT::f16);
  }
 
  bool isSCSrcV2F16() const {
    return isSCSrcF16();
  }
 
  bool isSCSrcF32() const {
    return isRegOrInlineNoMods(AMDGPU::SReg_32RegClassID, MVT::f32);
  }
 
  bool isSCSrcF64() const {
    return isRegOrInlineNoMods(AMDGPU::SReg_64RegClassID, MVT::f64);
  }
 
  bool isSSrcB32() const {
    return isSCSrcB32() || isLiteralImm(MVT::i32) || isExpr();
  }
 
  bool isSSrcB16() const {
    return isSCSrcB16() || isLiteralImm(MVT::i16);
  }
 
  bool isSSrcV2B16() const {
    llvm_unreachable("cannot happen");
    return isSSrcB16();
  }
 
  bool isSSrcB64() const {
    // TODO: Find out how SALU supports extension of 32-bit literals to 64 bits.
    // See isVSrc64().
    return isSCSrcB64() || isLiteralImm(MVT::i64);
  }
 
  bool isSSrcF32() const {
    return isSCSrcB32() || isLiteralImm(MVT::f32) || isExpr();
  }
 
  bool isSSrcF64() const {
    return isSCSrcB64() || isLiteralImm(MVT::f64);
  }
 
  bool isSSrcF16() const {
    return isSCSrcB16() || isLiteralImm(MVT::f16);
  }
 
  bool isSSrcV2F16() const {
    llvm_unreachable("cannot happen");
    return isSSrcF16();
  }
 
  bool isSSrcV2FP32() const {
    llvm_unreachable("cannot happen");
    return isSSrcF32();
  }
 
  bool isSCSrcV2FP32() const {
    llvm_unreachable("cannot happen");
    return isSCSrcF32();
  }
 
  bool isSSrcV2INT32() const {
    llvm_unreachable("cannot happen");
    return isSSrcB32();
  }
 
  bool isSCSrcV2INT32() const {
    llvm_unreachable("cannot happen");
    return isSCSrcB32();
  }
 
  bool isSSrcOrLdsB32() const {
    return isRegOrInlineNoMods(AMDGPU::SRegOrLds_32RegClassID, MVT::i32) ||
           isLiteralImm(MVT::i32) || isExpr();
  }
 
  bool isVCSrcB32() const {
    return isRegOrInlineNoMods(AMDGPU::VS_32RegClassID, MVT::i32);
  }
 
  bool isVCSrcB64() const {
    return isRegOrInlineNoMods(AMDGPU::VS_64RegClassID, MVT::i64);
  }
 
  bool isVCSrcTB16_Lo128() const {
    return isRegOrInlineNoMods(AMDGPU::VS_32_Lo128RegClassID, MVT::i16);
  }
 
  bool isVCSrcB16() const {
    return isRegOrInlineNoMods(AMDGPU::VS_32RegClassID, MVT::i16);
  }
 
  bool isVCSrcV2B16() const {
    return isVCSrcB16();
  }
 
  bool isVCSrcF32() const {
    return isRegOrInlineNoMods(AMDGPU::VS_32RegClassID, MVT::f32);
  }
 
  bool isVCSrcF64() const {
    return isRegOrInlineNoMods(AMDGPU::VS_64RegClassID, MVT::f64);
  }
 
  bool isVCSrcTF16_Lo128() const {
    return isRegOrInlineNoMods(AMDGPU::VS_32_Lo128RegClassID, MVT::f16);
  }
 
  bool isVCSrcF16() const {
    return isRegOrInlineNoMods(AMDGPU::VS_32RegClassID, MVT::f16);
  }
 
  bool isVCSrcV2F16() const {
    return isVCSrcF16();
  }
 
  bool isVSrcB32() const {
    return isVCSrcF32() || isLiteralImm(MVT::i32) || isExpr();
  }
 
  bool isVSrcB64() const {
    return isVCSrcF64() || isLiteralImm(MVT::i64);
  }
 
  bool isVSrcTB16_Lo128() const {
    return isVCSrcTB16_Lo128() || isLiteralImm(MVT::i16);
  }
 
  bool isVSrcB16() const {
    return isVCSrcB16() || isLiteralImm(MVT::i16);
  }
 
  bool isVSrcV2B16() const {
    return isVSrcB16() || isLiteralImm(MVT::v2i16);
  }
 
  bool isVCSrcV2FP32() const {
    return isVCSrcF64();
  }
 
  bool isVSrcV2FP32() const {
    return isVSrcF64() || isLiteralImm(MVT::v2f32);
  }
 
  bool isVCSrcV2INT32() const {
    return isVCSrcB64();
  }
 
  bool isVSrcV2INT32() const {
    return isVSrcB64() || isLiteralImm(MVT::v2i32);
  }
 
  bool isVSrcF32() const {
    return isVCSrcF32() || isLiteralImm(MVT::f32) || isExpr();
  }
 
  bool isVSrcF64() const {
    return isVCSrcF64() || isLiteralImm(MVT::f64);
  }
 
  bool isVSrcTF16_Lo128() const {
    return isVCSrcTF16_Lo128() || isLiteralImm(MVT::f16);
  }
 
  bool isVSrcF16() const {
    return isVCSrcF16() || isLiteralImm(MVT::f16);
  }
 
  bool isVSrcV2F16() const {
    return isVSrcF16() || isLiteralImm(MVT::v2f16);
  }
 
  bool isVISrcB32() const {
    return isRegOrInlineNoMods(AMDGPU::VGPR_32RegClassID, MVT::i32);
  }
 
  bool isVISrcB16() const {
    return isRegOrInlineNoMods(AMDGPU::VGPR_32RegClassID, MVT::i16);
  }
 
  bool isVISrcV2B16() const {
    return isVISrcB16();
  }
 
  bool isVISrcF32() const {
    return isRegOrInlineNoMods(AMDGPU::VGPR_32RegClassID, MVT::f32);
  }
 
  bool isVISrcF16() const {
    return isRegOrInlineNoMods(AMDGPU::VGPR_32RegClassID, MVT::f16);
  }
 
  bool isVISrcV2F16() const {
    return isVISrcF16() || isVISrcB32();
  }
 
  bool isVISrc_64B64() const {
    return isRegOrInlineNoMods(AMDGPU::VReg_64RegClassID, MVT::i64);
  }
 
  bool isVISrc_64F64() const {
    return isRegOrInlineNoMods(AMDGPU::VReg_64RegClassID, MVT::f64);
  }
 
  bool isVISrc_64V2FP32() const {
    return isRegOrInlineNoMods(AMDGPU::VReg_64RegClassID, MVT::f32);
  }
 
  bool isVISrc_64V2INT32() const {
    return isRegOrInlineNoMods(AMDGPU::VReg_64RegClassID, MVT::i32);
  }
 
  bool isVISrc_256B64() const {
    return isRegOrInlineNoMods(AMDGPU::VReg_256RegClassID, MVT::i64);
  }
 
  bool isVISrc_256F64() const {
    return isRegOrInlineNoMods(AMDGPU::VReg_256RegClassID, MVT::f64);
  }
 
  bool isVISrc_128B16() const {
    return isRegOrInlineNoMods(AMDGPU::VReg_128RegClassID, MVT::i16);
  }
 
  bool isVISrc_128V2B16() const {
    return isVISrc_128B16();
  }
 
  bool isVISrc_128B32() const {
    return isRegOrInlineNoMods(AMDGPU::VReg_128RegClassID, MVT::i32);
  }
 
  bool isVISrc_128F32() const {
    return isRegOrInlineNoMods(AMDGPU::VReg_128RegClassID, MVT::f32);
  }
 
  bool isVISrc_256V2FP32() const {
    return isRegOrInlineNoMods(AMDGPU::VReg_256RegClassID, MVT::f32);
  }
 
  bool isVISrc_256V2INT32() const {
    return isRegOrInlineNoMods(AMDGPU::VReg_256RegClassID, MVT::i32);
  }
 
  bool isVISrc_512B32() const {
    return isRegOrInlineNoMods(AMDGPU::VReg_512RegClassID, MVT::i32);
  }
 
  bool isVISrc_512B16() const {
    return isRegOrInlineNoMods(AMDGPU::VReg_512RegClassID, MVT::i16);
  }
 
  bool isVISrc_512V2B16() const {
    return isVISrc_512B16();
  }
 
  bool isVISrc_512F32() const {
    return isRegOrInlineNoMods(AMDGPU::VReg_512RegClassID, MVT::f32);
  }
 
  bool isVISrc_512F16() const {
    return isRegOrInlineNoMods(AMDGPU::VReg_512RegClassID, MVT::f16);
  }
 
  bool isVISrc_512V2F16() const {
    return isVISrc_512F16() || isVISrc_512B32();
  }
 
  bool isVISrc_1024B32() const {
    return isRegOrInlineNoMods(AMDGPU::VReg_1024RegClassID, MVT::i32);
  }
 
  bool isVISrc_1024B16() const {
    return isRegOrInlineNoMods(AMDGPU::VReg_1024RegClassID, MVT::i16);
  }
 
  bool isVISrc_1024V2B16() const {
    return isVISrc_1024B16();
  }
 
  bool isVISrc_1024F32() const {
    return isRegOrInlineNoMods(AMDGPU::VReg_1024RegClassID, MVT::f32);
  }
 
  bool isVISrc_1024F16() const {
    return isRegOrInlineNoMods(AMDGPU::VReg_1024RegClassID, MVT::f16);
  }
 
  bool isVISrc_1024V2F16() const {
    return isVISrc_1024F16() || isVISrc_1024B32();
  }
 
  bool isAISrcB32() const {
    return isRegOrInlineNoMods(AMDGPU::AGPR_32RegClassID, MVT::i32);
  }
 
  bool isAISrcB16() const {
    return isRegOrInlineNoMods(AMDGPU::AGPR_32RegClassID, MVT::i16);
  }
 
  bool isAISrcV2B16() const {
    return isAISrcB16();
  }
 
  bool isAISrcF32() const {
    return isRegOrInlineNoMods(AMDGPU::AGPR_32RegClassID, MVT::f32);
  }
 
  bool isAISrcF16() const {
    return isRegOrInlineNoMods(AMDGPU::AGPR_32RegClassID, MVT::f16);
  }
 
  bool isAISrcV2F16() const {
    return isAISrcF16() || isAISrcB32();
  }
 
  bool isAISrc_64B64() const {
    return isRegOrInlineNoMods(AMDGPU::AReg_64RegClassID, MVT::i64);
  }
 
  bool isAISrc_64F64() const {
    return isRegOrInlineNoMods(AMDGPU::AReg_64RegClassID, MVT::f64);
  }
 
  bool isAISrc_128B32() const {
    return isRegOrInlineNoMods(AMDGPU::AReg_128RegClassID, MVT::i32);
  }
 
  bool isAISrc_128B16() const {
    return isRegOrInlineNoMods(AMDGPU::AReg_128RegClassID, MVT::i16);
  }
 
  bool isAISrc_128V2B16() const {
    return isAISrc_128B16();
  }
 
  bool isAISrc_128F32() const {
    return isRegOrInlineNoMods(AMDGPU::AReg_128RegClassID, MVT::f32);
  }
 
  bool isAISrc_128F16() const {
    return isRegOrInlineNoMods(AMDGPU::AReg_128RegClassID, MVT::f16);
  }
 
  bool isAISrc_128V2F16() const {
    return isAISrc_128F16() || isAISrc_128B32();
  }
 
  bool isVISrc_128F16() const {
    return isRegOrInlineNoMods(AMDGPU::VReg_128RegClassID, MVT::f16);
  }
 
  bool isVISrc_128V2F16() const {
    return isVISrc_128F16() || isVISrc_128B32();
  }
 
  bool isAISrc_256B64() const {
    return isRegOrInlineNoMods(AMDGPU::AReg_256RegClassID, MVT::i64);
  }
 
  bool isAISrc_256F64() const {
    return isRegOrInlineNoMods(AMDGPU::AReg_256RegClassID, MVT::f64);
  }
 
  bool isAISrc_512B32() const {
    return isRegOrInlineNoMods(AMDGPU::AReg_512RegClassID, MVT::i32);
  }
 
  bool isAISrc_512B16() const {
    return isRegOrInlineNoMods(AMDGPU::AReg_512RegClassID, MVT::i16);
  }
 
  bool isAISrc_512V2B16() const {
    return isAISrc_512B16();
  }
 
  bool isAISrc_512F32() const {
    return isRegOrInlineNoMods(AMDGPU::AReg_512RegClassID, MVT::f32);
  }
 
  bool isAISrc_512F16() const {
    return isRegOrInlineNoMods(AMDGPU::AReg_512RegClassID, MVT::f16);
  }
 
  bool isAISrc_512V2F16() const {
    return isAISrc_512F16() || isAISrc_512B32();
  }
 
  bool isAISrc_1024B32() const {
    return isRegOrInlineNoMods(AMDGPU::AReg_1024RegClassID, MVT::i32);
  }
 
  bool isAISrc_1024B16() const {
    return isRegOrInlineNoMods(AMDGPU::AReg_1024RegClassID, MVT::i16);
  }
 
  bool isAISrc_1024V2B16() const {
    return isAISrc_1024B16();
  }
 
  bool isAISrc_1024F32() const {
    return isRegOrInlineNoMods(AMDGPU::AReg_1024RegClassID, MVT::f32);
  }
 
  bool isAISrc_1024F16() const {
    return isRegOrInlineNoMods(AMDGPU::AReg_1024RegClassID, MVT::f16);
  }
 
  bool isAISrc_1024V2F16() const {
    return isAISrc_1024F16() || isAISrc_1024B32();
  }
 
  bool isKImmFP32() const {
    return isLiteralImm(MVT::f32);
  }
 
  bool isKImmFP16() const {
    return isLiteralImm(MVT::f16);
  }
 
  bool isMem() const override {
    return false;
  }
 
  bool isExpr() const {
    return Kind == Expression;
  }
 
  bool isSoppBrTarget() const {
    return isExpr() || isImm();
  }
 
  bool isSWaitCnt() const;
  bool isDepCtr() const;
  bool isSDelayAlu() const;
  bool isHwreg() const;
  bool isSendMsg() const;
  bool isSwizzle() const;
  bool isSMRDOffset8() const;
  bool isSMEMOffset() const;
  bool isSMRDLiteralOffset() const;
  bool isDPP8() const;
  bool isDPPCtrl() const;
  bool isBLGP() const;
  bool isCBSZ() const;
  bool isABID() const;
  bool isGPRIdxMode() const;
  bool isS16Imm() const;
  bool isU16Imm() const;
  bool isEndpgm() const;
  bool isWaitVDST() const;
  bool isWaitEXP() const;
 
  StringRef getToken() const {
    assert(isToken());
    return StringRef(Tok.Data, Tok.Length);
  }
 
  int64_t getImm() const {
    assert(isImm());
    return Imm.Val;
  }
 
  void setImm(int64_t Val) {
    assert(isImm());
    Imm.Val = Val;
  }
 
  ImmTy getImmTy() const {
    assert(isImm());
    return Imm.Type;
  }
 
  unsigned getReg() const override {
    assert(isRegKind());
    return Reg.RegNo;
  }
 
  SMLoc getStartLoc() const override {
    return StartLoc;
  }
 
  SMLoc getEndLoc() const override {
    return EndLoc;
  }
 
  SMRange getLocRange() const {
    return SMRange(StartLoc, EndLoc);
  }
 
  Modifiers getModifiers() const {
    assert(isRegKind() || isImmTy(ImmTyNone));
    return isRegKind() ? Reg.Mods : Imm.Mods;
  }
 
  void setModifiers(Modifiers Mods) {
    assert(isRegKind() || isImmTy(ImmTyNone));
    if (isRegKind())
      Reg.Mods = Mods;
    else
      Imm.Mods = Mods;
  }
 
  bool hasModifiers() const {
    return getModifiers().hasModifiers();
  }
 
  bool hasFPModifiers() const {
    return getModifiers().hasFPModifiers();
  }
 
  bool hasIntModifiers() const {
    return getModifiers().hasIntModifiers();
  }
 
  uint64_t applyInputFPModifiers(uint64_t Val, unsigned Size) const;
 
  void addImmOperands(MCInst &Inst, unsigned N, bool ApplyModifiers = true) const;
 
  void addLiteralImmOperand(MCInst &Inst, int64_t Val, bool ApplyModifiers) const;
 
  template <unsigned Bitwidth>
  void addKImmFPOperands(MCInst &Inst, unsigned N) const;
 
  void addKImmFP16Operands(MCInst &Inst, unsigned N) const {
    addKImmFPOperands<16>(Inst, N);
  }
 
  void addKImmFP32Operands(MCInst &Inst, unsigned N) const {
    addKImmFPOperands<32>(Inst, N);
  }
 
  void addRegOperands(MCInst &Inst, unsigned N) const;
 
  void addBoolRegOperands(MCInst &Inst, unsigned N) const {
    addRegOperands(Inst, N);
  }
 
  void addRegOrImmOperands(MCInst &Inst, unsigned N) const {
    if (isRegKind())
      addRegOperands(Inst, N);
    else if (isExpr())
      Inst.addOperand(MCOperand::createExpr(Expr));
    else
      addImmOperands(Inst, N);
  }
 
  void addRegOrImmWithInputModsOperands(MCInst &Inst, unsigned N) const {
    Modifiers Mods = getModifiers();
    Inst.addOperand(MCOperand::createImm(Mods.getModifiersOperand()));
    if (isRegKind()) {
      addRegOperands(Inst, N);
    } else {
      addImmOperands(Inst, N, false);
    }
  }
 
  void addRegOrImmWithFPInputModsOperands(MCInst &Inst, unsigned N) const {
    assert(!hasIntModifiers());
    addRegOrImmWithInputModsOperands(Inst, N);
  }
 
  void addRegOrImmWithIntInputModsOperands(MCInst &Inst, unsigned N) const {
    assert(!hasFPModifiers());
    addRegOrImmWithInputModsOperands(Inst, N);
  }
 
  void addRegWithInputModsOperands(MCInst &Inst, unsigned N) const {
    Modifiers Mods = getModifiers();
    Inst.addOperand(MCOperand::createImm(Mods.getModifiersOperand()));
    assert(isRegKind());
    addRegOperands(Inst, N);
  }
 
  void addRegWithFPInputModsOperands(MCInst &Inst, unsigned N) const {
    assert(!hasIntModifiers());
    addRegWithInputModsOperands(Inst, N);
  }
 
  void addRegWithIntInputModsOperands(MCInst &Inst, unsigned N) const {
    assert(!hasFPModifiers());
    addRegWithInputModsOperands(Inst, N);
  }
 
  void addSoppBrTargetOperands(MCInst &Inst, unsigned N) const {
    if (isImm())
      addImmOperands(Inst, N);
    else {
      assert(isExpr());
      Inst.addOperand(MCOperand::createExpr(Expr));
    }
  }
 
  static void printImmTy(raw_ostream& OS, ImmTy Type) {
    switch (Type) {
    case ImmTyNone: OS << "None"; break;
    case ImmTyGDS: OS << "GDS"; break;
    case ImmTyLDS: OS << "LDS"; break;
    case ImmTyOffen: OS << "Offen"; break;
    case ImmTyIdxen: OS << "Idxen"; break;
    case ImmTyAddr64: OS << "Addr64"; break;
    case ImmTyOffset: OS << "Offset"; break;
    case ImmTyInstOffset: OS << "InstOffset"; break;
    case ImmTyOffset0: OS << "Offset0"; break;
    case ImmTyOffset1: OS << "Offset1"; break;
    case ImmTySMEMOffsetMod: OS << "SMEMOffsetMod"; break;
    case ImmTyCPol: OS << "CPol"; break;
    case ImmTySWZ: OS << "SWZ"; break;
    case ImmTyTFE: OS << "TFE"; break;
    case ImmTyD16: OS << "D16"; break;
    case ImmTyFORMAT: OS << "FORMAT"; break;
    case ImmTyClampSI: OS << "ClampSI"; break;
    case ImmTyOModSI: OS << "OModSI"; break;
    case ImmTyDPP8: OS << "DPP8"; break;
    case ImmTyDppCtrl: OS << "DppCtrl"; break;
    case ImmTyDppRowMask: OS << "DppRowMask"; break;
    case ImmTyDppBankMask: OS << "DppBankMask"; break;
    case ImmTyDppBoundCtrl: OS << "DppBoundCtrl"; break;
    case ImmTyDppFi: OS << "FI"; break;
    case ImmTySdwaDstSel: OS << "SdwaDstSel"; break;
    case ImmTySdwaSrc0Sel: OS << "SdwaSrc0Sel"; break;
    case ImmTySdwaSrc1Sel: OS << "SdwaSrc1Sel"; break;
    case ImmTySdwaDstUnused: OS << "SdwaDstUnused"; break;
    case ImmTyDMask: OS << "DMask"; break;
    case ImmTyDim: OS << "Dim"; break;
    case ImmTyUNorm: OS << "UNorm"; break;
    case ImmTyDA: OS << "DA"; break;
    case ImmTyR128A16: OS << "R128A16"; break;
    case ImmTyA16: OS << "A16"; break;
    case ImmTyLWE: OS << "LWE"; break;
    case ImmTyOff: OS << "Off"; break;
    case ImmTyExpTgt: OS << "ExpTgt"; break;
    case ImmTyExpCompr: OS << "ExpCompr"; break;
    case ImmTyExpVM: OS << "ExpVM"; break;
    case ImmTyHwreg: OS << "Hwreg"; break;
    case ImmTySendMsg: OS << "SendMsg"; break;
    case ImmTyInterpSlot: OS << "InterpSlot"; break;
    case ImmTyInterpAttr: OS << "InterpAttr"; break;
    case ImmTyAttrChan: OS << "AttrChan"; break;
    case ImmTyOpSel: OS << "OpSel"; break;
    case ImmTyOpSelHi: OS << "OpSelHi"; break;
    case ImmTyNegLo: OS << "NegLo"; break;
    case ImmTyNegHi: OS << "NegHi"; break;
    case ImmTySwizzle: OS << "Swizzle"; break;
    case ImmTyGprIdxMode: OS << "GprIdxMode"; break;
    case ImmTyHigh: OS << "High"; break;
    case ImmTyBLGP: OS << "BLGP"; break;
    case ImmTyCBSZ: OS << "CBSZ"; break;
    case ImmTyABID: OS << "ABID"; break;
    case ImmTyEndpgm: OS << "Endpgm"; break;
    case ImmTyWaitVDST: OS << "WaitVDST"; break;
    case ImmTyWaitEXP: OS << "WaitEXP"; break;
    }
  }
 
  void print(raw_ostream &OS) const override {
    switch (Kind) {
    case Register:
      OS << "<register " << getReg() << " mods: " << Reg.Mods << '>';
      break;
    case Immediate:
      OS << '<' << getImm();
      if (getImmTy() != ImmTyNone) {
        OS << " type: "; printImmTy(OS, getImmTy());
      }
      OS << " mods: " << Imm.Mods << '>';
      break;
    case Token:
      OS << '\'' << getToken() << '\'';
      break;
    case Expression:
      OS << "<expr " << *Expr << '>';
      break;
    }
  }
 
  static AMDGPUOperand::Ptr CreateImm(const AMDGPUAsmParser *AsmParser,
                                      int64_t Val, SMLoc Loc,
                                      ImmTy Type = ImmTyNone,
                                      bool IsFPImm = false) {
    auto Op = std::make_unique<AMDGPUOperand>(Immediate, AsmParser);
    Op->Imm.Val = Val;
    Op->Imm.IsFPImm = IsFPImm;
    Op->Imm.Kind = ImmKindTyNone;
    Op->Imm.Type = Type;
    Op->Imm.Mods = Modifiers();
    Op->StartLoc = Loc;
    Op->EndLoc = Loc;
    return Op;
  }
 
  static AMDGPUOperand::Ptr CreateToken(const AMDGPUAsmParser *AsmParser,
                                        StringRef Str, SMLoc Loc,
                                        bool HasExplicitEncodingSize = true) {
    auto Res = std::make_unique<AMDGPUOperand>(Token, AsmParser);
    Res->Tok.Data = Str.data();
    Res->Tok.Length = Str.size();
    Res->StartLoc = Loc;
    Res->EndLoc = Loc;
    return Res;
  }
 
  static AMDGPUOperand::Ptr CreateReg(const AMDGPUAsmParser *AsmParser,
                                      unsigned RegNo, SMLoc S,
                                      SMLoc E) {
    auto Op = std::make_unique<AMDGPUOperand>(Register, AsmParser);
    Op->Reg.RegNo = RegNo;
    Op->Reg.Mods = Modifiers();
    Op->StartLoc = S;
    Op->EndLoc = E;
    return Op;
  }
 
  static AMDGPUOperand::Ptr CreateExpr(const AMDGPUAsmParser *AsmParser,
                                       const class MCExpr *Expr, SMLoc S) {
    auto Op = std::make_unique<AMDGPUOperand>(Expression, AsmParser);
    Op->Expr = Expr;
    Op->StartLoc = S;
    Op->EndLoc = S;
    return Op;
  }
};
 
raw_ostream &operator <<(raw_ostream &OS, AMDGPUOperand::Modifiers Mods) {
  OS << "abs:" << Mods.Abs << " neg: " << Mods.Neg << " sext:" << Mods.Sext;
  return OS;
}
 
//===----------------------------------------------------------------------===//
// AsmParser
//===----------------------------------------------------------------------===//
 
// Holds info related to the current kernel, e.g. count of SGPRs used.
// Kernel scope begins at .amdgpu_hsa_kernel directive, ends at next
// .amdgpu_hsa_kernel or at EOF.
class KernelScopeInfo {
  int SgprIndexUnusedMin = -1;
  int VgprIndexUnusedMin = -1;
  int AgprIndexUnusedMin = -1;
  MCContext *Ctx = nullptr;
  MCSubtargetInfo const *MSTI = nullptr;
 
  void usesSgprAt(int i) {
    if (i >= SgprIndexUnusedMin) {
      SgprIndexUnusedMin = ++i;
      if (Ctx) {
        MCSymbol* const Sym =
          Ctx->getOrCreateSymbol(Twine(".kernel.sgpr_count"));
        Sym->setVariableValue(MCConstantExpr::create(SgprIndexUnusedMin, *Ctx));
      }
    }
  }
 
  void usesVgprAt(int i) {
    if (i >= VgprIndexUnusedMin) {
      VgprIndexUnusedMin = ++i;
      if (Ctx) {
        MCSymbol* const Sym =
          Ctx->getOrCreateSymbol(Twine(".kernel.vgpr_count"));
        int totalVGPR = getTotalNumVGPRs(isGFX90A(*MSTI), AgprIndexUnusedMin,
                                         VgprIndexUnusedMin);
        Sym->setVariableValue(MCConstantExpr::create(totalVGPR, *Ctx));
      }
    }
  }
 
  void usesAgprAt(int i) {
    // Instruction will error in AMDGPUAsmParser::MatchAndEmitInstruction
    if (!hasMAIInsts(*MSTI))
      return;
 
    if (i >= AgprIndexUnusedMin) {
      AgprIndexUnusedMin = ++i;
      if (Ctx) {
        MCSymbol* const Sym =
          Ctx->getOrCreateSymbol(Twine(".kernel.agpr_count"));
        Sym->setVariableValue(MCConstantExpr::create(AgprIndexUnusedMin, *Ctx));
 
        // Also update vgpr_count (dependent on agpr_count for gfx908/gfx90a)
        MCSymbol* const vSym =
          Ctx->getOrCreateSymbol(Twine(".kernel.vgpr_count"));
        int totalVGPR = getTotalNumVGPRs(isGFX90A(*MSTI), AgprIndexUnusedMin,
                                         VgprIndexUnusedMin);
        vSym->setVariableValue(MCConstantExpr::create(totalVGPR, *Ctx));
      }
    }
  }
 
public:
  KernelScopeInfo() = default;
 
  void initialize(MCContext &Context) {
    Ctx = &Context;
    MSTI = Ctx->getSubtargetInfo();
 
    usesSgprAt(SgprIndexUnusedMin = -1);
    usesVgprAt(VgprIndexUnusedMin = -1);
    if (hasMAIInsts(*MSTI)) {
      usesAgprAt(AgprIndexUnusedMin = -1);
    }
  }
 
  void usesRegister(RegisterKind RegKind, unsigned DwordRegIndex,
                    unsigned RegWidth) {
    switch (RegKind) {
    case IS_SGPR:
      usesSgprAt(DwordRegIndex + divideCeil(RegWidth, 32) - 1);
      break;
    case IS_AGPR:
      usesAgprAt(DwordRegIndex + divideCeil(RegWidth, 32) - 1);
      break;
    case IS_VGPR:
      usesVgprAt(DwordRegIndex + divideCeil(RegWidth, 32) - 1);
      break;
    default:
      break;
    }
  }
};
 
class AMDGPUAsmParser : public MCTargetAsmParser {
  MCAsmParser &Parser;
 
  unsigned ForcedEncodingSize = 0;
  bool ForcedDPP = false;
  bool ForcedSDWA = false;
  KernelScopeInfo KernelScope;
 
  /// @name Auto-generated Match Functions
  /// {
 
#define GET_ASSEMBLER_HEADER
#include "AMDGPUGenAsmMatcher.inc"
 
  /// }
 
private:
  bool ParseAsAbsoluteExpression(uint32_t &Ret);
  bool OutOfRangeError(SMRange Range);
  /// Calculate VGPR/SGPR blocks required for given target, reserved
  /// registers, and user-specified NextFreeXGPR values.
  ///
  /// \param Features [in] Target features, used for bug corrections.
  /// \param VCCUsed [in] Whether VCC special SGPR is reserved.
  /// \param FlatScrUsed [in] Whether FLAT_SCRATCH special SGPR is reserved.
  /// \param XNACKUsed [in] Whether XNACK_MASK special SGPR is reserved.
  /// \param EnableWavefrontSize32 [in] Value of ENABLE_WAVEFRONT_SIZE32 kernel
  /// descriptor field, if valid.
  /// \param NextFreeVGPR [in] Max VGPR number referenced, plus one.
  /// \param VGPRRange [in] Token range, used for VGPR diagnostics.
  /// \param NextFreeSGPR [in] Max SGPR number referenced, plus one.
  /// \param SGPRRange [in] Token range, used for SGPR diagnostics.
  /// \param VGPRBlocks [out] Result VGPR block count.
  /// \param SGPRBlocks [out] Result SGPR block count.
  bool calculateGPRBlocks(const FeatureBitset &Features, bool VCCUsed,
                          bool FlatScrUsed, bool XNACKUsed,
                          std::optional<bool> EnableWavefrontSize32,
                          unsigned NextFreeVGPR, SMRange VGPRRange,
                          unsigned NextFreeSGPR, SMRange SGPRRange,
                          unsigned &VGPRBlocks, unsigned &SGPRBlocks);
  bool ParseDirectiveAMDGCNTarget();
  bool ParseDirectiveAMDHSAKernel();
  bool ParseDirectiveMajorMinor(uint32_t &Major, uint32_t &Minor);
  bool ParseDirectiveHSACodeObjectVersion();
  bool ParseDirectiveHSACodeObjectISA();
  bool ParseAMDKernelCodeTValue(StringRef ID, amd_kernel_code_t &Header);
  bool ParseDirectiveAMDKernelCodeT();
  // TODO: Possibly make subtargetHasRegister const.
  bool subtargetHasRegister(const MCRegisterInfo &MRI, unsigned RegNo);
  bool ParseDirectiveAMDGPUHsaKernel();
 
  bool ParseDirectiveISAVersion();
  bool ParseDirectiveHSAMetadata();
  bool ParseDirectivePALMetadataBegin();
  bool ParseDirectivePALMetadata();
  bool ParseDirectiveAMDGPULDS();
 
  /// Common code to parse out a block of text (typically YAML) between start and
  /// end directives.
  bool ParseToEndDirective(const char *AssemblerDirectiveBegin,
                           const char *AssemblerDirectiveEnd,
                           std::string &CollectString);
 
  bool AddNextRegisterToList(unsigned& Reg, unsigned& RegWidth,
                             RegisterKind RegKind, unsigned Reg1, SMLoc Loc);
  bool ParseAMDGPURegister(RegisterKind &RegKind, unsigned &Reg,
                           unsigned &RegNum, unsigned &RegWidth,
                           bool RestoreOnFailure = false);
  bool ParseAMDGPURegister(RegisterKind &RegKind, unsigned &Reg,
                           unsigned &RegNum, unsigned &RegWidth,
                           SmallVectorImpl<AsmToken> &Tokens);
  unsigned ParseRegularReg(RegisterKind &RegKind, unsigned &RegNum,
                           unsigned &RegWidth,
                           SmallVectorImpl<AsmToken> &Tokens);
  unsigned ParseSpecialReg(RegisterKind &RegKind, unsigned &RegNum,
                           unsigned &RegWidth,
                           SmallVectorImpl<AsmToken> &Tokens);
  unsigned ParseRegList(RegisterKind &RegKind, unsigned &RegNum,
                        unsigned &RegWidth, SmallVectorImpl<AsmToken> &Tokens);
  bool ParseRegRange(unsigned& Num, unsigned& Width);
  unsigned getRegularReg(RegisterKind RegKind,
                         unsigned RegNum,
                         unsigned RegWidth,
                         SMLoc Loc);
 
  bool isRegister();
  bool isRegister(const AsmToken &Token, const AsmToken &NextToken) const;
  std::optional<StringRef> getGprCountSymbolName(RegisterKind RegKind);
  void initializeGprCountSymbol(RegisterKind RegKind);
  bool updateGprCountSymbols(RegisterKind RegKind, unsigned DwordRegIndex,
                             unsigned RegWidth);
  void cvtMubufImpl(MCInst &Inst, const OperandVector &Operands,
                    bool IsAtomic);
  void cvtDSImpl(MCInst &Inst, const OperandVector &Operands,
                 bool IsGdsHardcoded);
 
public:
  enum AMDGPUMatchResultTy {
    Match_PreferE32 = FIRST_TARGET_MATCH_RESULT_TY
  };
  enum OperandMode {
    OperandMode_Default,
    OperandMode_NSA,
  };
 
  using OptionalImmIndexMap = std::map<AMDGPUOperand::ImmTy, unsigned>;
 
  AMDGPUAsmParser(const MCSubtargetInfo &STI, MCAsmParser &_Parser,
               const MCInstrInfo &MII,
               const MCTargetOptions &Options)
      : MCTargetAsmParser(Options, STI, MII), Parser(_Parser) {
    MCAsmParserExtension::Initialize(Parser);
 
    if (getFeatureBits().none()) {
      // Set default features.
      copySTI().ToggleFeature("southern-islands");
    }
 
    setAvailableFeatures(ComputeAvailableFeatures(getFeatureBits()));
 
    {
      // TODO: make those pre-defined variables read-only.
      // Currently there is none suitable machinery in the core llvm-mc for this.
      // MCSymbol::isRedefinable is intended for another purpose, and
      // AsmParser::parseDirectiveSet() cannot be specialized for specific target.
      AMDGPU::IsaVersion ISA = AMDGPU::getIsaVersion(getSTI().getCPU());
      MCContext &Ctx = getContext();
      if (ISA.Major >= 6 && isHsaAbiVersion3AndAbove(&getSTI())) {
        MCSymbol *Sym =
            Ctx.getOrCreateSymbol(Twine(".amdgcn.gfx_generation_number"));
        Sym->setVariableValue(MCConstantExpr::create(ISA.Major, Ctx));
        Sym = Ctx.getOrCreateSymbol(Twine(".amdgcn.gfx_generation_minor"));
        Sym->setVariableValue(MCConstantExpr::create(ISA.Minor, Ctx));
        Sym = Ctx.getOrCreateSymbol(Twine(".amdgcn.gfx_generation_stepping"));
        Sym->setVariableValue(MCConstantExpr::create(ISA.Stepping, Ctx));
      } else {
        MCSymbol *Sym =
            Ctx.getOrCreateSymbol(Twine(".option.machine_version_major"));
        Sym->setVariableValue(MCConstantExpr::create(ISA.Major, Ctx));
        Sym = Ctx.getOrCreateSymbol(Twine(".option.machine_version_minor"));
        Sym->setVariableValue(MCConstantExpr::create(ISA.Minor, Ctx));
        Sym = Ctx.getOrCreateSymbol(Twine(".option.machine_version_stepping"));
        Sym->setVariableValue(MCConstantExpr::create(ISA.Stepping, Ctx));
      }
      if (ISA.Major >= 6 && isHsaAbiVersion3AndAbove(&getSTI())) {
        initializeGprCountSymbol(IS_VGPR);
        initializeGprCountSymbol(IS_SGPR);
      } else
        KernelScope.initialize(getContext());
    }
  }
 
  bool hasMIMG_R128() const {
    return AMDGPU::hasMIMG_R128(getSTI());
  }
 
  bool hasPackedD16() const {
    return AMDGPU::hasPackedD16(getSTI());
  }
 
  bool hasA16() const { return AMDGPU::hasA16(getSTI()); }
 
  bool hasG16() const { return AMDGPU::hasG16(getSTI()); }
 
  bool isSI() const {
    return AMDGPU::isSI(getSTI());
  }
 
  bool isCI() const {
    return AMDGPU::isCI(getSTI());
  }
 
  bool isVI() const {
    return AMDGPU::isVI(getSTI());
  }
 
  bool isGFX9() const {
    return AMDGPU::isGFX9(getSTI());
  }
 
  // TODO: isGFX90A is also true for GFX940. We need to clean it.
  bool isGFX90A() const {
    return AMDGPU::isGFX90A(getSTI());
  }
 
  bool isGFX940() const {
    return AMDGPU::isGFX940(getSTI());
  }
 
  bool isGFX9Plus() const {
    return AMDGPU::isGFX9Plus(getSTI());
  }
 
  bool isGFX10() const {
    return AMDGPU::isGFX10(getSTI());
  }
 
  bool isGFX10Plus() const { return AMDGPU::isGFX10Plus(getSTI()); }
 
  bool isGFX11() const {
    return AMDGPU::isGFX11(getSTI());
  }
 
  bool isGFX11Plus() const {
    return AMDGPU::isGFX11Plus(getSTI());
  }
 
  bool isGFX10_AEncoding() const { return AMDGPU::isGFX10_AEncoding(getSTI()); }
 
  bool isGFX10_BEncoding() const {
    return AMDGPU::isGFX10_BEncoding(getSTI());
  }
 
  bool hasInv2PiInlineImm() const {
    return getFeatureBits()[AMDGPU::FeatureInv2PiInlineImm];
  }
 
  bool hasFlatOffsets() const {
    return getFeatureBits()[AMDGPU::FeatureFlatInstOffsets];
  }
 
  bool hasArchitectedFlatScratch() const {
    return getFeatureBits()[AMDGPU::FeatureArchitectedFlatScratch];
  }
 
  bool hasSGPR102_SGPR103() const {
    return !isVI() && !isGFX9();
  }
 
  bool hasSGPR104_SGPR105() const { return isGFX10Plus(); }
 
  bool hasIntClamp() const {
    return getFeatureBits()[AMDGPU::FeatureIntClamp];
  }
 
  bool hasPartialNSAEncoding() const {
    return getFeatureBits()[AMDGPU::FeaturePartialNSAEncoding];
  }
 
  unsigned getNSAMaxSize() const {
    return AMDGPU::getNSAMaxSize(getSTI());
  }
 
  AMDGPUTargetStreamer &getTargetStreamer() {
    MCTargetStreamer &TS = *getParser().getStreamer().getTargetStreamer();
    return static_cast<AMDGPUTargetStreamer &>(TS);
  }
 
  const MCRegisterInfo *getMRI() const {
    // We need this const_cast because for some reason getContext() is not const
    // in MCAsmParser.
    return const_cast<AMDGPUAsmParser*>(this)->getContext().getRegisterInfo();
  }
 
  const MCInstrInfo *getMII() const {
    return &MII;
  }
 
  const FeatureBitset &getFeatureBits() const {
    return getSTI().getFeatureBits();
  }
 
  void setForcedEncodingSize(unsigned Size) { ForcedEncodingSize = Size; }
  void setForcedDPP(bool ForceDPP_) { ForcedDPP = ForceDPP_; }
  void setForcedSDWA(bool ForceSDWA_) { ForcedSDWA = ForceSDWA_; }
 
  unsigned getForcedEncodingSize() const { return ForcedEncodingSize; }
  bool isForcedVOP3() const { return ForcedEncodingSize == 64; }
  bool isForcedDPP() const { return ForcedDPP; }
  bool isForcedSDWA() const { return ForcedSDWA; }
  ArrayRef<unsigned> getMatchedVariants() const;
  StringRef getMatchedVariantName() const;
 
  std::unique_ptr<AMDGPUOperand> parseRegister(bool RestoreOnFailure = false);
  bool ParseRegister(MCRegister &RegNo, SMLoc &StartLoc, SMLoc &EndLoc,
                     bool RestoreOnFailure);
  bool parseRegister(MCRegister &RegNo, SMLoc &StartLoc,
                     SMLoc &EndLoc) override;
  OperandMatchResultTy tryParseRegister(MCRegister &RegNo, SMLoc &StartLoc,
                                        SMLoc &EndLoc) override;
  unsigned checkTargetMatchPredicate(MCInst &Inst) override;
  unsigned validateTargetOperandClass(MCParsedAsmOperand &Op,
                                      unsigned Kind) override;
  bool MatchAndEmitInstruction(SMLoc IDLoc, unsigned &Opcode,
                               OperandVector &Operands, MCStreamer &Out,
                               uint64_t &ErrorInfo,
                               bool MatchingInlineAsm) override;
  bool ParseDirective(AsmToken DirectiveID) override;
  OperandMatchResultTy parseOperand(OperandVector &Operands, StringRef Mnemonic,
                                    OperandMode Mode = OperandMode_Default);
  StringRef parseMnemonicSuffix(StringRef Name);
  bool ParseInstruction(ParseInstructionInfo &Info, StringRef Name,
                        SMLoc NameLoc, OperandVector &Operands) override;
  //bool ProcessInstruction(MCInst &Inst);
 
  OperandMatchResultTy parseTokenOp(StringRef Name, OperandVector &Operands);
 
  OperandMatchResultTy parseIntWithPrefix(const char *Prefix, int64_t &Int);
 
  OperandMatchResultTy
  parseIntWithPrefix(const char *Prefix, OperandVector &Operands,
                     AMDGPUOperand::ImmTy ImmTy = AMDGPUOperand::ImmTyNone,
                     bool (*ConvertResult)(int64_t &) = nullptr);
 
  OperandMatchResultTy
  parseOperandArrayWithPrefix(const char *Prefix,
                              OperandVector &Operands,
                              AMDGPUOperand::ImmTy ImmTy = AMDGPUOperand::ImmTyNone,
                              bool (*ConvertResult)(int64_t&) = nullptr);
 
  OperandMatchResultTy
  parseNamedBit(StringRef Name, OperandVector &Operands,
                AMDGPUOperand::ImmTy ImmTy = AMDGPUOperand::ImmTyNone);
  unsigned getCPolKind(StringRef Id, StringRef Mnemo, bool &Disabling) const;
  OperandMatchResultTy parseCPol(OperandVector &Operands);
  OperandMatchResultTy parseStringWithPrefix(StringRef Prefix,
                                             StringRef &Value,
                                             SMLoc &StringLoc);
 
  bool isModifier();
  bool isOperandModifier(const AsmToken &Token, const AsmToken &NextToken) const;
  bool isRegOrOperandModifier(const AsmToken &Token, const AsmToken &NextToken) const;
  bool isNamedOperandModifier(const AsmToken &Token, const AsmToken &NextToken) const;
  bool isOpcodeModifierWithVal(const AsmToken &Token, const AsmToken &NextToken) const;
  bool parseSP3NegModifier();
  OperandMatchResultTy parseImm(OperandVector &Operands, bool HasSP3AbsModifier = false);
  OperandMatchResultTy parseReg(OperandVector &Operands);
  OperandMatchResultTy parseRegOrImm(OperandVector &Operands, bool HasSP3AbsMod = false);
  OperandMatchResultTy parseRegOrImmWithFPInputMods(OperandVector &Operands, bool AllowImm = true);
  OperandMatchResultTy parseRegOrImmWithIntInputMods(OperandVector &Operands, bool AllowImm = true);
  OperandMatchResultTy parseRegWithFPInputMods(OperandVector &Operands);
  OperandMatchResultTy parseRegWithIntInputMods(OperandVector &Operands);
  OperandMatchResultTy parseVReg32OrOff(OperandVector &Operands);
  OperandMatchResultTy parseDfmtNfmt(int64_t &Format);
  OperandMatchResultTy parseUfmt(int64_t &Format);
  OperandMatchResultTy parseSymbolicSplitFormat(StringRef FormatStr, SMLoc Loc, int64_t &Format);
  OperandMatchResultTy parseSymbolicUnifiedFormat(StringRef FormatStr, SMLoc Loc, int64_t &Format);
  OperandMatchResultTy parseFORMAT(OperandVector &Operands);
  OperandMatchResultTy parseSymbolicOrNumericFormat(int64_t &Format);
  OperandMatchResultTy parseNumericFormat(int64_t &Format);
  OperandMatchResultTy parseFlatOffset(OperandVector &Operands);
  OperandMatchResultTy parseR128A16(OperandVector &Operands);
  bool tryParseFmt(const char *Pref, int64_t MaxVal, int64_t &Val);
  bool matchDfmtNfmt(int64_t &Dfmt, int64_t &Nfmt, StringRef FormatStr, SMLoc Loc);
 
  void cvtDSOffset01(MCInst &Inst, const OperandVector &Operands);
  void cvtDS(MCInst &Inst, const OperandVector &Operands) { cvtDSImpl(Inst, Operands, false); }
  void cvtDSGds(MCInst &Inst, const OperandVector &Operands) { cvtDSImpl(Inst, Operands, true); }
  void cvtExp(MCInst &Inst, const OperandVector &Operands);
 
  bool parseCnt(int64_t &IntVal);
  OperandMatchResultTy parseSWaitCntOps(OperandVector &Operands);
 
  bool parseDepCtr(int64_t &IntVal, unsigned &Mask);
  void depCtrError(SMLoc Loc, int ErrorId, StringRef DepCtrName);
  OperandMatchResultTy parseDepCtrOps(OperandVector &Operands);
 
  bool parseDelay(int64_t &Delay);
  OperandMatchResultTy parseSDelayAluOps(OperandVector &Operands);
 
  OperandMatchResultTy parseHwreg(OperandVector &Operands);
 
private:
  struct OperandInfoTy {
    SMLoc Loc;
    int64_t Id;
    bool IsSymbolic = false;
    bool IsDefined = false;
 
    OperandInfoTy(int64_t Id_) : Id(Id_) {}
  };
 
  bool parseSendMsgBody(OperandInfoTy &Msg, OperandInfoTy &Op, OperandInfoTy &Stream);
  bool validateSendMsg(const OperandInfoTy &Msg,
                       const OperandInfoTy &Op,
                       const OperandInfoTy &Stream);
 
  bool parseHwregBody(OperandInfoTy &HwReg,
                      OperandInfoTy &Offset,
                      OperandInfoTy &Width);
  bool validateHwreg(const OperandInfoTy &HwReg,
                     const OperandInfoTy &Offset,
                     const OperandInfoTy &Width);
 
  SMLoc getFlatOffsetLoc(const OperandVector &Operands) const;
  SMLoc getSMEMOffsetLoc(const OperandVector &Operands) const;
  SMLoc getBLGPLoc(const OperandVector &Operands) const;
 
  SMLoc getOperandLoc(std::function<bool(const AMDGPUOperand&)> Test,
                      const OperandVector &Operands) const;
  SMLoc getImmLoc(AMDGPUOperand::ImmTy Type, const OperandVector &Operands) const;
  SMLoc getRegLoc(unsigned Reg, const OperandVector &Operands) const;
  SMLoc getLitLoc(const OperandVector &Operands,
                  bool SearchMandatoryLiterals = false) const;
  SMLoc getMandatoryLitLoc(const OperandVector &Operands) const;
  SMLoc getConstLoc(const OperandVector &Operands) const;
  SMLoc getInstLoc(const OperandVector &Operands) const;
 
  bool validateInstruction(const MCInst &Inst, const SMLoc &IDLoc, const OperandVector &Operands);
  bool validateFlatOffset(const MCInst &Inst, const OperandVector &Operands);
  bool validateSMEMOffset(const MCInst &Inst, const OperandVector &Operands);
  bool validateSOPLiteral(const MCInst &Inst) const;
  bool validateConstantBusLimitations(const MCInst &Inst, const OperandVector &Operands);
  bool validateVOPDRegBankConstraints(const MCInst &Inst,
                                      const OperandVector &Operands);
  bool validateIntClampSupported(const MCInst &Inst);
  bool validateMIMGAtomicDMask(const MCInst &Inst);
  bool validateMIMGGatherDMask(const MCInst &Inst);
  bool validateMovrels(const MCInst &Inst, const OperandVector &Operands);
  bool validateMIMGDataSize(const MCInst &Inst, const SMLoc &IDLoc);
  bool validateMIMGAddrSize(const MCInst &Inst);
  bool validateMIMGD16(const MCInst &Inst);
  bool validateMIMGMSAA(const MCInst &Inst);
  bool validateOpSel(const MCInst &Inst);
  bool validateDPP(const MCInst &Inst, const OperandVector &Operands);
  bool validateVccOperand(unsigned Reg) const;
  bool validateVOPLiteral(const MCInst &Inst, const OperandVector &Operands);
  bool validateMAIAccWrite(const MCInst &Inst, const OperandVector &Operands);
  bool validateMAISrc2(const MCInst &Inst, const OperandVector &Operands);
  bool validateMFMA(const MCInst &Inst, const OperandVector &Operands);
  bool validateAGPRLdSt(const MCInst &Inst) const;
  bool validateVGPRAlign(const MCInst &Inst) const;
  bool validateBLGP(const MCInst &Inst, const OperandVector &Operands);
  bool validateGWS(const MCInst &Inst, const OperandVector &Operands);
  bool validateDivScale(const MCInst &Inst);
  bool validateWaitCnt(const MCInst &Inst, const OperandVector &Operands);
  bool validateCoherencyBits(const MCInst &Inst, const OperandVector &Operands,
                             const SMLoc &IDLoc);
  bool validateExeczVcczOperands(const OperandVector &Operands);
  bool validateTFE(const MCInst &Inst, const OperandVector &Operands);
  std::optional<StringRef> validateLdsDirect(const MCInst &Inst);
  unsigned getConstantBusLimit(unsigned Opcode) const;
  bool usesConstantBus(const MCInst &Inst, unsigned OpIdx);
  bool isInlineConstant(const MCInst &Inst, unsigned OpIdx) const;
  unsigned findImplicitSGPRReadInVOP(const MCInst &Inst) const;
 
  bool isSupportedMnemo(StringRef Mnemo,
                        const FeatureBitset &FBS);
  bool isSupportedMnemo(StringRef Mnemo,
                        const FeatureBitset &FBS,
                        ArrayRef<unsigned> Variants);
  bool checkUnsupportedInstruction(StringRef Name, const SMLoc &IDLoc);
 
  bool isId(const StringRef Id) const;
  bool isId(const AsmToken &Token, const StringRef Id) const;
  bool isToken(const AsmToken::TokenKind Kind) const;
  StringRef getId() const;
  bool trySkipId(const StringRef Id);
  bool trySkipId(const StringRef Pref, const StringRef Id);
  bool trySkipId(const StringRef Id, const AsmToken::TokenKind Kind);
  bool trySkipToken(const AsmToken::TokenKind Kind);
  bool skipToken(const AsmToken::TokenKind Kind, const StringRef ErrMsg);
  bool parseString(StringRef &Val, const StringRef ErrMsg = "expected a string");
  bool parseId(StringRef &Val, const StringRef ErrMsg = "");
 
  void peekTokens(MutableArrayRef<AsmToken> Tokens);
  AsmToken::TokenKind getTokenKind() const;
  bool parseExpr(int64_t &Imm, StringRef Expected = "");
  bool parseExpr(OperandVector &Operands);
  StringRef getTokenStr() const;
  AsmToken peekToken(bool ShouldSkipSpace = true);
  AsmToken getToken() const;
  SMLoc getLoc() const;
  void lex();
 
public:
  void onBeginOfFile() override;
 
  OperandMatchResultTy parseCustomOperand(OperandVector &Operands,
                                          unsigned MCK);
 
  OperandMatchResultTy parseExpTgt(OperandVector &Operands);
  OperandMatchResultTy parseSendMsgOp(OperandVector &Operands);
  OperandMatchResultTy parseInterpSlot(OperandVector &Operands);
  OperandMatchResultTy parseInterpAttr(OperandVector &Operands);
  OperandMatchResultTy parseSOppBrTarget(OperandVector &Operands);
  OperandMatchResultTy parseBoolReg(OperandVector &Operands);
 
  bool parseSwizzleOperand(int64_t &Op,
                           const unsigned MinVal,
                           const unsigned MaxVal,
                           const StringRef ErrMsg,
                           SMLoc &Loc);
  bool parseSwizzleOperands(const unsigned OpNum, int64_t* Op,
                            const unsigned MinVal,
                            const unsigned MaxVal,
                            const StringRef ErrMsg);
  OperandMatchResultTy parseSwizzleOp(OperandVector &Operands);
  bool parseSwizzleOffset(int64_t &Imm);
  bool parseSwizzleMacro(int64_t &Imm);
  bool parseSwizzleQuadPerm(int64_t &Imm);
  bool parseSwizzleBitmaskPerm(int64_t &Imm);
  bool parseSwizzleBroadcast(int64_t &Imm);
  bool parseSwizzleSwap(int64_t &Imm);
  bool parseSwizzleReverse(int64_t &Imm);
 
  OperandMatchResultTy parseGPRIdxMode(OperandVector &Operands);
  int64_t parseGPRIdxMacro();
 
  void cvtMubuf(MCInst &Inst, const OperandVector &Operands) { cvtMubufImpl(Inst, Operands, false); }
  void cvtMubufAtomic(MCInst &Inst, const OperandVector &Operands) { cvtMubufImpl(Inst, Operands, true); }
  void cvtMtbuf(MCInst &Inst, const OperandVector &Operands);
 
  AMDGPUOperand::Ptr defaultCPol() const;
 
  AMDGPUOperand::Ptr defaultSMRDOffset8() const;
  AMDGPUOperand::Ptr defaultSMEMOffset() const;
  AMDGPUOperand::Ptr defaultSMEMOffsetMod() const;
  AMDGPUOperand::Ptr defaultSMRDLiteralOffset() const;
  AMDGPUOperand::Ptr defaultFlatOffset() const;
 
  OperandMatchResultTy parseOModOperand(OperandVector &Operands);
 
  void cvtVOP3(MCInst &Inst, const OperandVector &Operands,
               OptionalImmIndexMap &OptionalIdx);
  void cvtVOP3OpSel(MCInst &Inst, const OperandVector &Operands);
  void cvtVOP3(MCInst &Inst, const OperandVector &Operands);
  void cvtVOP3P(MCInst &Inst, const OperandVector &Operands);
  void cvtVOPD(MCInst &Inst, const OperandVector &Operands);
  void cvtVOP3OpSel(MCInst &Inst, const OperandVector &Operands,
                    OptionalImmIndexMap &OptionalIdx);
  void cvtVOP3P(MCInst &Inst, const OperandVector &Operands,
                OptionalImmIndexMap &OptionalIdx);
 
  void cvtVOP3Interp(MCInst &Inst, const OperandVector &Operands);
  void cvtVINTERP(MCInst &Inst, const OperandVector &Operands);
 
  void cvtMIMG(MCInst &Inst, const OperandVector &Operands,
               bool IsAtomic = false);
  void cvtMIMGAtomic(MCInst &Inst, const OperandVector &Operands);
  void cvtIntersectRay(MCInst &Inst, const OperandVector &Operands);
 
  void cvtSMEMAtomic(MCInst &Inst, const OperandVector &Operands);
 
  bool parseDimId(unsigned &Encoding);
  OperandMatchResultTy parseDim(OperandVector &Operands);
  OperandMatchResultTy parseDPP8(OperandVector &Operands);
  OperandMatchResultTy parseDPPCtrl(OperandVector &Operands);
  bool isSupportedDPPCtrl(StringRef Ctrl, const OperandVector &Operands);
  int64_t parseDPPCtrlSel(StringRef Ctrl);
  int64_t parseDPPCtrlPerm();
  AMDGPUOperand::Ptr defaultRowMask() const;
  AMDGPUOperand::Ptr defaultBankMask() const;
  AMDGPUOperand::Ptr defaultDppBoundCtrl() const;
  AMDGPUOperand::Ptr defaultFI() const;
  void cvtDPP(MCInst &Inst, const OperandVector &Operands, bool IsDPP8 = false);
  void cvtDPP8(MCInst &Inst, const OperandVector &Operands) {
    cvtDPP(Inst, Operands, true);
  }
  void cvtVOP3DPP(MCInst &Inst, const OperandVector &Operands,
                  bool IsDPP8 = false);
  void cvtVOP3DPP8(MCInst &Inst, const OperandVector &Operands) {
    cvtVOP3DPP(Inst, Operands, true);
  }
 
  OperandMatchResultTy parseSDWASel(OperandVector &Operands, StringRef Prefix,
                                    AMDGPUOperand::ImmTy Type);
  OperandMatchResultTy parseSDWADstUnused(OperandVector &Operands);
  void cvtSdwaVOP1(MCInst &Inst, const OperandVector &Operands);
  void cvtSdwaVOP2(MCInst &Inst, const OperandVector &Operands);
  void cvtSdwaVOP2b(MCInst &Inst, const OperandVector &Operands);
  void cvtSdwaVOP2e(MCInst &Inst, const OperandVector &Operands);
  void cvtSdwaVOPC(MCInst &Inst, const OperandVector &Operands);
  void cvtSDWA(MCInst &Inst, const OperandVector &Operands,
               uint64_t BasicInstType,
               bool SkipDstVcc = false,
               bool SkipSrcVcc = false);
 
  AMDGPUOperand::Ptr defaultBLGP() const;
  AMDGPUOperand::Ptr defaultCBSZ() const;
  AMDGPUOperand::Ptr defaultABID() const;
 
  OperandMatchResultTy parseEndpgmOp(OperandVector &Operands);
  AMDGPUOperand::Ptr defaultEndpgmImmOperands() const;
 
  AMDGPUOperand::Ptr defaultWaitVDST() const;
  AMDGPUOperand::Ptr defaultWaitEXP() const;
  OperandMatchResultTy parseVOPD(OperandVector &Operands);
};
 
} // end anonymous namespace
 
// May be called with integer type with equivalent bitwidth.
static const fltSemantics *getFltSemantics(unsigned Size) {
  switch (Size) {
  case 4:
    return &APFloat::IEEEsingle();
  case 8:
    return &APFloat::IEEEdouble();
  case 2:
    return &APFloat::IEEEhalf();
  default:
    llvm_unreachable("unsupported fp type");
  }
}
 
static const fltSemantics *getFltSemantics(MVT VT) {
  return getFltSemantics(VT.getSizeInBits() / 8);
}
 
static const fltSemantics *getOpFltSemantics(uint8_t OperandType) {
  switch (OperandType) {
  case AMDGPU::OPERAND_REG_IMM_INT32:
  case AMDGPU::OPERAND_REG_IMM_FP32:
  case AMDGPU::OPERAND_REG_IMM_FP32_DEFERRED:
  case AMDGPU::OPERAND_REG_INLINE_C_INT32:
  case AMDGPU::OPERAND_REG_INLINE_C_FP32:
  case AMDGPU::OPERAND_REG_INLINE_AC_INT32:
  case AMDGPU::OPERAND_REG_INLINE_AC_FP32:
  case AMDGPU::OPERAND_REG_INLINE_C_V2FP32:
  case AMDGPU::OPERAND_REG_IMM_V2FP32:
  case AMDGPU::OPERAND_REG_INLINE_C_V2INT32:
  case AMDGPU::OPERAND_REG_IMM_V2INT32:
  case AMDGPU::OPERAND_KIMM32:
    return &APFloat::IEEEsingle();
  case AMDGPU::OPERAND_REG_IMM_INT64:
  case AMDGPU::OPERAND_REG_IMM_FP64:
  case AMDGPU::OPERAND_REG_INLINE_C_INT64:
  case AMDGPU::OPERAND_REG_INLINE_C_FP64:
  case AMDGPU::OPERAND_REG_INLINE_AC_FP64:
    return &APFloat::IEEEdouble();
  case AMDGPU::OPERAND_REG_IMM_INT16:
  case AMDGPU::OPERAND_REG_IMM_FP16:
  case AMDGPU::OPERAND_REG_IMM_FP16_DEFERRED:
  case AMDGPU::OPERAND_REG_INLINE_C_INT16:
  case AMDGPU::OPERAND_REG_INLINE_C_FP16:
  case AMDGPU::OPERAND_REG_INLINE_C_V2INT16:
  case AMDGPU::OPERAND_REG_INLINE_C_V2FP16:
  case AMDGPU::OPERAND_REG_INLINE_AC_INT16:
  case AMDGPU::OPERAND_REG_INLINE_AC_FP16:
  case AMDGPU::OPERAND_REG_INLINE_AC_V2INT16:
  case AMDGPU::OPERAND_REG_INLINE_AC_V2FP16:
  case AMDGPU::OPERAND_REG_IMM_V2INT16:
  case AMDGPU::OPERAND_REG_IMM_V2FP16:
  case AMDGPU::OPERAND_KIMM16:
    return &APFloat::IEEEhalf();
  default:
    llvm_unreachable("unsupported fp type");
  }
}
 
//===----------------------------------------------------------------------===//
// Operand
//===----------------------------------------------------------------------===//
 
static bool canLosslesslyConvertToFPType(APFloat &FPLiteral, MVT VT) {
  bool Lost;
 
  // Convert literal to single precision
  APFloat::opStatus Status = FPLiteral.convert(*getFltSemantics(VT),
                                               APFloat::rmNearestTiesToEven,
                                               &Lost);
  // We allow precision lost but not overflow or underflow
  if (Status != APFloat::opOK &&
      Lost &&
      ((Status & APFloat::opOverflow)  != 0 ||
       (Status & APFloat::opUnderflow) != 0)) {
    return false;
  }
 
  return true;
}
 
static bool isSafeTruncation(int64_t Val, unsigned Size) {
  return isUIntN(Size, Val) || isIntN(Size, Val);
}
 
static bool isInlineableLiteralOp16(int64_t Val, MVT VT, bool HasInv2Pi) {
  if (VT.getScalarType() == MVT::i16) {
    // FP immediate values are broken.
    return isInlinableIntLiteral(Val);
  }
 
  // f16/v2f16 operands work correctly for all values.
  return AMDGPU::isInlinableLiteral16(Val, HasInv2Pi);
}
 
bool AMDGPUOperand::isInlinableImm(MVT type) const {
 
  // This is a hack to enable named inline values like
  // shared_base with both 32-bit and 64-bit operands.
  // Note that these values are defined as
  // 32-bit operands only.
  if (isInlineValue()) {
    return true;
  }
 
  if (!isImmTy(ImmTyNone)) {
    // Only plain immediates are inlinable (e.g. "clamp" attribute is not)
    return false;
  }
  // TODO: We should avoid using host float here. It would be better to
  // check the float bit values which is what a few other places do.
  // We've had bot failures before due to weird NaN support on mips hosts.
 
  APInt Literal(64, Imm.Val);
 
  if (Imm.IsFPImm) { // We got fp literal token
    if (type == MVT::f64 || type == MVT::i64) { // Expected 64-bit operand
      return AMDGPU::isInlinableLiteral64(Imm.Val,
                                          AsmParser->hasInv2PiInlineImm());
    }
 
    APFloat FPLiteral(APFloat::IEEEdouble(), APInt(64, Imm.Val));
    if (!canLosslesslyConvertToFPType(FPLiteral, type))
      return false;
 
    if (type.getScalarSizeInBits() == 16) {
      return isInlineableLiteralOp16(
        static_cast<int16_t>(FPLiteral.bitcastToAPInt().getZExtValue()),
        type, AsmParser->hasInv2PiInlineImm());
    }
 
    // Check if single precision literal is inlinable
    return AMDGPU::isInlinableLiteral32(
      static_cast<int32_t>(FPLiteral.bitcastToAPInt().getZExtValue()),
      AsmParser->hasInv2PiInlineImm());
  }
 
  // We got int literal token.
  if (type == MVT::f64 || type == MVT::i64) { // Expected 64-bit operand
    return AMDGPU::isInlinableLiteral64(Imm.Val,
                                        AsmParser->hasInv2PiInlineImm());
  }
 
  if (!isSafeTruncation(Imm.Val, type.getScalarSizeInBits())) {
    return false;
  }
 
  if (type.getScalarSizeInBits() == 16) {
    return isInlineableLiteralOp16(
      static_cast<int16_t>(Literal.getLoBits(16).getSExtValue()),
      type, AsmParser->hasInv2PiInlineImm());
  }
 
  return AMDGPU::isInlinableLiteral32(
    static_cast<int32_t>(Literal.getLoBits(32).getZExtValue()),
    AsmParser->hasInv2PiInlineImm());
}
 
bool AMDGPUOperand::isLiteralImm(MVT type) const {
  // Check that this immediate can be added as literal
  if (!isImmTy(ImmTyNone)) {
    return false;
  }
 
  if (!Imm.IsFPImm) {
    // We got int literal token.
 
    if (type == MVT::f64 && hasFPModifiers()) {
      // Cannot apply fp modifiers to int literals preserving the same semantics
      // for VOP1/2/C and VOP3 because of integer truncation. To avoid ambiguity,
      // disable these cases.
      return false;
    }
 
    unsigned Size = type.getSizeInBits();
    if (Size == 64)
      Size = 32;
 
    // FIXME: 64-bit operands can zero extend, sign extend, or pad zeroes for FP
    // types.
    return isSafeTruncation(Imm.Val, Size);
  }
 
  // We got fp literal token
  if (type == MVT::f64) { // Expected 64-bit fp operand
    // We would set low 64-bits of literal to zeroes but we accept this literals
    return true;
  }
 
  if (type == MVT::i64) { // Expected 64-bit int operand
    // We don't allow fp literals in 64-bit integer instructions. It is
    // unclear how we should encode them.
    return false;
  }
 
  // We allow fp literals with f16x2 operands assuming that the specified
  // literal goes into the lower half and the upper half is zero. We also
  // require that the literal may be losslessly converted to f16.
  MVT ExpectedType = (type == MVT::v2f16)? MVT::f16 :
                     (type == MVT::v2i16)? MVT::i16 :
                     (type == MVT::v2f32)? MVT::f32 : type;
 
  APFloat FPLiteral(APFloat::IEEEdouble(), APInt(64, Imm.Val));
  return canLosslesslyConvertToFPType(FPLiteral, ExpectedType);
}
 
bool AMDGPUOperand::isRegClass(unsigned RCID) const {
  return isRegKind() && AsmParser->getMRI()->getRegClass(RCID).contains(getReg());
}
 
bool AMDGPUOperand::isVRegWithInputMods() const {
  return isRegClass(AMDGPU::VGPR_32RegClassID) ||
         // GFX90A allows DPP on 64-bit operands.
         (isRegClass(AMDGPU::VReg_64RegClassID) &&
          AsmParser->getFeatureBits()[AMDGPU::Feature64BitDPP]);
}
 
bool AMDGPUOperand::isT16VRegWithInputMods() const {
  return isRegClass(AMDGPU::VGPR_32_Lo128RegClassID);
}
 
bool AMDGPUOperand::isSDWAOperand(MVT type) const {
  if (AsmParser->isVI())
    return isVReg32();
  else if (AsmParser->isGFX9Plus())
    return isRegClass(AMDGPU::VS_32RegClassID) || isInlinableImm(type);
  else
    return false;
}
 
bool AMDGPUOperand::isSDWAFP16Operand() const {
  return isSDWAOperand(MVT::f16);
}
 
bool AMDGPUOperand::isSDWAFP32Operand() const {
  return isSDWAOperand(MVT::f32);
}
 
bool AMDGPUOperand::isSDWAInt16Operand() const {
  return isSDWAOperand(MVT::i16);
}
 
bool AMDGPUOperand::isSDWAInt32Operand() const {
  return isSDWAOperand(MVT::i32);
}
 
bool AMDGPUOperand::isBoolReg() const {
  auto FB = AsmParser->getFeatureBits();
  return isReg() && ((FB[AMDGPU::FeatureWavefrontSize64] && isSCSrcB64()) ||
                     (FB[AMDGPU::FeatureWavefrontSize32] && isSCSrcB32()));
}
 
uint64_t AMDGPUOperand::applyInputFPModifiers(uint64_t Val, unsigned Size) const
{
  assert(isImmTy(ImmTyNone) && Imm.Mods.hasFPModifiers());
  assert(Size == 2 || Size == 4 || Size == 8);
 
  const uint64_t FpSignMask = (1ULL << (Size * 8 - 1));
 
  if (Imm.Mods.Abs) {
    Val &= ~FpSignMask;
  }
  if (Imm.Mods.Neg) {
    Val ^= FpSignMask;
  }
 
  return Val;
}
 
void AMDGPUOperand::addImmOperands(MCInst &Inst, unsigned N, bool ApplyModifiers) const {
  if (AMDGPU::isSISrcOperand(AsmParser->getMII()->get(Inst.getOpcode()),
                             Inst.getNumOperands())) {
    addLiteralImmOperand(Inst, Imm.Val,
                         ApplyModifiers &
                         isImmTy(ImmTyNone) && Imm.Mods.hasFPModifiers());
  } else {
    assert(!isImmTy(ImmTyNone) || !hasModifiers());
    Inst.addOperand(MCOperand::createImm(Imm.Val));
    setImmKindNone();
  }
}
 
void AMDGPUOperand::addLiteralImmOperand(MCInst &Inst, int64_t Val, bool ApplyModifiers) const {
  const auto& InstDesc = AsmParser->getMII()->get(Inst.getOpcode());
  auto OpNum = Inst.getNumOperands();
  // Check that this operand accepts literals
  assert(AMDGPU::isSISrcOperand(InstDesc, OpNum));
 
  if (ApplyModifiers) {
    assert(AMDGPU::isSISrcFPOperand(InstDesc, OpNum));
    const unsigned Size = Imm.IsFPImm ? sizeof(double) : getOperandSize(InstDesc, OpNum);
    Val = applyInputFPModifiers(Val, Size);
  }
 
  APInt Literal(64, Val);
  uint8_t OpTy = InstDesc.operands()[OpNum].OperandType;
 
  if (Imm.IsFPImm) { // We got fp literal token
    switch (OpTy) {
    case AMDGPU::OPERAND_REG_IMM_INT64:
    case AMDGPU::OPERAND_REG_IMM_FP64:
    case AMDGPU::OPERAND_REG_INLINE_C_INT64:
    case AMDGPU::OPERAND_REG_INLINE_C_FP64:
    case AMDGPU::OPERAND_REG_INLINE_AC_FP64:
      if (AMDGPU::isInlinableLiteral64(Literal.getZExtValue(),
                                       AsmParser->hasInv2PiInlineImm())) {
        Inst.addOperand(MCOperand::createImm(Literal.getZExtValue()));
        setImmKindConst();
        return;
      }
 
      // Non-inlineable
      if (AMDGPU::isSISrcFPOperand(InstDesc, OpNum)) { // Expected 64-bit fp operand
        // For fp operands we check if low 32 bits are zeros
        if (Literal.getLoBits(32) != 0) {
          const_cast<AMDGPUAsmParser *>(AsmParser)->Warning(Inst.getLoc(),
          "Can't encode literal as exact 64-bit floating-point operand. "
          "Low 32-bits will be set to zero");
        }
 
        Inst.addOperand(MCOperand::createImm(Literal.lshr(32).getZExtValue()));
        setImmKindLiteral();
        return;
      }
 
      // We don't allow fp literals in 64-bit integer instructions. It is
      // unclear how we should encode them. This case should be checked earlier
      // in predicate methods (isLiteralImm())
      llvm_unreachable("fp literal in 64-bit integer instruction.");
 
    case AMDGPU::OPERAND_REG_IMM_INT32:
    case AMDGPU::OPERAND_REG_IMM_FP32:
    case AMDGPU::OPERAND_REG_IMM_FP32_DEFERRED:
    case AMDGPU::OPERAND_REG_INLINE_C_INT32:
    case AMDGPU::OPERAND_REG_INLINE_C_FP32:
    case AMDGPU::OPERAND_REG_INLINE_AC_INT32:
    case AMDGPU::OPERAND_REG_INLINE_AC_FP32:
    case AMDGPU::OPERAND_REG_IMM_INT16:
    case AMDGPU::OPERAND_REG_IMM_FP16:
    case AMDGPU::OPERAND_REG_IMM_FP16_DEFERRED:
    case AMDGPU::OPERAND_REG_INLINE_C_INT16:
    case AMDGPU::OPERAND_REG_INLINE_C_FP16:
    case AMDGPU::OPERAND_REG_INLINE_C_V2INT16:
    case AMDGPU::OPERAND_REG_INLINE_C_V2FP16:
    case AMDGPU::OPERAND_REG_INLINE_AC_INT16:
    case AMDGPU::OPERAND_REG_INLINE_AC_FP16:
    case AMDGPU::OPERAND_REG_INLINE_AC_V2INT16:
    case AMDGPU::OPERAND_REG_INLINE_AC_V2FP16:
    case AMDGPU::OPERAND_REG_IMM_V2INT16:
    case AMDGPU::OPERAND_REG_IMM_V2FP16:
    case AMDGPU::OPERAND_REG_INLINE_C_V2FP32:
    case AMDGPU::OPERAND_REG_IMM_V2FP32:
    case AMDGPU::OPERAND_REG_INLINE_C_V2INT32:
    case AMDGPU::OPERAND_REG_IMM_V2INT32:
    case AMDGPU::OPERAND_KIMM32:
    case AMDGPU::OPERAND_KIMM16: {
      bool lost;
      APFloat FPLiteral(APFloat::IEEEdouble(), Literal);
      // Convert literal to single precision
      FPLiteral.convert(*getOpFltSemantics(OpTy),
                        APFloat::rmNearestTiesToEven, &lost);
      // We allow precision lost but not overflow or underflow. This should be
      // checked earlier in isLiteralImm()
 
      uint64_t ImmVal = FPLiteral.bitcastToAPInt().getZExtValue();
      Inst.addOperand(MCOperand::createImm(ImmVal));
      if (OpTy == AMDGPU::OPERAND_KIMM32 || OpTy == AMDGPU::OPERAND_KIMM16) {
        setImmKindMandatoryLiteral();
      } else {
        setImmKindLiteral();
      }
      return;
    }
    default:
      llvm_unreachable("invalid operand size");
    }
 
    return;
  }
 
  // We got int literal token.
  // Only sign extend inline immediates.
  switch (OpTy) {
  case AMDGPU::OPERAND_REG_IMM_INT32:
  case AMDGPU::OPERAND_REG_IMM_FP32:
  case AMDGPU::OPERAND_REG_IMM_FP32_DEFERRED:
  case AMDGPU::OPERAND_REG_INLINE_C_INT32:
  case AMDGPU::OPERAND_REG_INLINE_C_FP32:
  case AMDGPU::OPERAND_REG_INLINE_AC_INT32:
  case AMDGPU::OPERAND_REG_INLINE_AC_FP32:
  case AMDGPU::OPERAND_REG_IMM_V2INT16:
  case AMDGPU::OPERAND_REG_IMM_V2FP16:
  case AMDGPU::OPERAND_REG_IMM_V2FP32:
  case AMDGPU::OPERAND_REG_INLINE_C_V2FP32:
  case AMDGPU::OPERAND_REG_IMM_V2INT32:
  case AMDGPU::OPERAND_REG_INLINE_C_V2INT32:
    if (isSafeTruncation(Val, 32) &&
        AMDGPU::isInlinableLiteral32(static_cast<int32_t>(Val),
                                     AsmParser->hasInv2PiInlineImm())) {
      Inst.addOperand(MCOperand::createImm(Val));
      setImmKindConst();
      return;
    }
 
    Inst.addOperand(MCOperand::createImm(Val & 0xffffffff));
    setImmKindLiteral();
    return;
 
  case AMDGPU::OPERAND_REG_IMM_INT64:
  case AMDGPU::OPERAND_REG_IMM_FP64:
  case AMDGPU::OPERAND_REG_INLINE_C_INT64:
  case AMDGPU::OPERAND_REG_INLINE_C_FP64:
  case AMDGPU::OPERAND_REG_INLINE_AC_FP64:
    if (AMDGPU::isInlinableLiteral64(Val, AsmParser->hasInv2PiInlineImm())) {
      Inst.addOperand(MCOperand::createImm(Val));
      setImmKindConst();
      return;
    }
 
    Inst.addOperand(MCOperand::createImm(Lo_32(Val)));
    setImmKindLiteral();
    return;
 
  case AMDGPU::OPERAND_REG_IMM_INT16:
  case AMDGPU::OPERAND_REG_IMM_FP16:
  case AMDGPU::OPERAND_REG_IMM_FP16_DEFERRED:
  case AMDGPU::OPERAND_REG_INLINE_C_INT16:
  case AMDGPU::OPERAND_REG_INLINE_C_FP16:
  case AMDGPU::OPERAND_REG_INLINE_AC_INT16:
  case AMDGPU::OPERAND_REG_INLINE_AC_FP16:
    if (isSafeTruncation(Val, 16) &&
        AMDGPU::isInlinableLiteral16(static_cast<int16_t>(Val),
                                     AsmParser->hasInv2PiInlineImm())) {
      Inst.addOperand(MCOperand::createImm(Val));
      setImmKindConst();
      return;
    }
 
    Inst.addOperand(MCOperand::createImm(Val & 0xffff));
    setImmKindLiteral();
    return;
 
  case AMDGPU::OPERAND_REG_INLINE_C_V2INT16:
  case AMDGPU::OPERAND_REG_INLINE_C_V2FP16:
  case AMDGPU::OPERAND_REG_INLINE_AC_V2INT16:
  case AMDGPU::OPERAND_REG_INLINE_AC_V2FP16: {
    assert(isSafeTruncation(Val, 16));
    assert(AMDGPU::isInlinableLiteral16(static_cast<int16_t>(Val),
                                        AsmParser->hasInv2PiInlineImm()));
 
    Inst.addOperand(MCOperand::createImm(Val));
    return;
  }
  case AMDGPU::OPERAND_KIMM32:
    Inst.addOperand(MCOperand::createImm(Literal.getLoBits(32).getZExtValue()));
    setImmKindMandatoryLiteral();
    return;
  case AMDGPU::OPERAND_KIMM16:
    Inst.addOperand(MCOperand::createImm(Literal.getLoBits(16).getZExtValue()));
    setImmKindMandatoryLiteral();
    return;
  default:
    llvm_unreachable("invalid operand size");
  }
}
 
template <unsigned Bitwidth>
void AMDGPUOperand::addKImmFPOperands(MCInst &Inst, unsigned N) const {
  APInt Literal(64, Imm.Val);
  setImmKindMandatoryLiteral();
 
  if (!Imm.IsFPImm) {
    // We got int literal token.
    Inst.addOperand(MCOperand::createImm(Literal.getLoBits(Bitwidth).getZExtValue()));
    return;
  }
 
  bool Lost;
  APFloat FPLiteral(APFloat::IEEEdouble(), Literal);
  FPLiteral.convert(*getFltSemantics(Bitwidth / 8),
                    APFloat::rmNearestTiesToEven, &Lost);
  Inst.addOperand(MCOperand::createImm(FPLiteral.bitcastToAPInt().getZExtValue()));
}
 
void AMDGPUOperand::addRegOperands(MCInst &Inst, unsigned N) const {
  Inst.addOperand(MCOperand::createReg(AMDGPU::getMCReg(getReg(), AsmParser->getSTI())));
}
 
bool AMDGPUOperand::isInlineValue() const {
  return isRegKind() && ::isInlineValue(getReg());
}
 
//===----------------------------------------------------------------------===//
// AsmParser
//===----------------------------------------------------------------------===//
 
static int getRegClass(RegisterKind Is, unsigned RegWidth) {
  if (Is == IS_VGPR) {
    switch (RegWidth) {
      default: return -1;
      case 32:
        return AMDGPU::VGPR_32RegClassID;
      case 64:
        return AMDGPU::VReg_64RegClassID;
      case 96:
        return AMDGPU::VReg_96RegClassID;
      case 128:
        return AMDGPU::VReg_128RegClassID;
      case 160:
        return AMDGPU::VReg_160RegClassID;
      case 192:
        return AMDGPU::VReg_192RegClassID;
      case 224:
        return AMDGPU::VReg_224RegClassID;
      case 256:
        return AMDGPU::VReg_256RegClassID;
      case 288:
        return AMDGPU::VReg_288RegClassID;
      case 320:
        return AMDGPU::VReg_320RegClassID;
      case 352:
        return AMDGPU::VReg_352RegClassID;
      case 384:
        return AMDGPU::VReg_384RegClassID;
      case 512:
        return AMDGPU::VReg_512RegClassID;
      case 1024:
        return AMDGPU::VReg_1024RegClassID;
    }
  } else if (Is == IS_TTMP) {
    switch (RegWidth) {
      default: return -1;
      case 32:
        return AMDGPU::TTMP_32RegClassID;
      case 64:
        return AMDGPU::TTMP_64RegClassID;
      case 128:
        return AMDGPU::TTMP_128RegClassID;
      case 256:
        return AMDGPU::TTMP_256RegClassID;
      case 512:
        return AMDGPU::TTMP_512RegClassID;
    }
  } else if (Is == IS_SGPR) {
    switch (RegWidth) {
      default: return -1;
      case 32:
        return AMDGPU::SGPR_32RegClassID;
      case 64:
        return AMDGPU::SGPR_64RegClassID;
      case 96:
        return AMDGPU::SGPR_96RegClassID;
      case 128:
        return AMDGPU::SGPR_128RegClassID;
      case 160:
        return AMDGPU::SGPR_160RegClassID;
      case 192:
        return AMDGPU::SGPR_192RegClassID;
      case 224:
        return AMDGPU::SGPR_224RegClassID;
      case 256:
        return AMDGPU::SGPR_256RegClassID;
      case 288:
        return AMDGPU::SGPR_288RegClassID;
      case 320:
        return AMDGPU::SGPR_320RegClassID;
      case 352:
        return AMDGPU::SGPR_352RegClassID;
      case 384:
        return AMDGPU::SGPR_384RegClassID;
      case 512:
        return AMDGPU::SGPR_512RegClassID;
    }
  } else if (Is == IS_AGPR) {
    switch (RegWidth) {
      default: return -1;
      case 32:
        return AMDGPU::AGPR_32RegClassID;
      case 64:
        return AMDGPU::AReg_64RegClassID;
      case 96:
        return AMDGPU::AReg_96RegClassID;
      case 128:
        return AMDGPU::AReg_128RegClassID;
      case 160:
        return AMDGPU::AReg_160RegClassID;
      case 192:
        return AMDGPU::AReg_192RegClassID;
      case 224:
        return AMDGPU::AReg_224RegClassID;
      case 256:
        return AMDGPU::AReg_256RegClassID;
      case 288:
        return AMDGPU::AReg_288RegClassID;
      case 320:
        return AMDGPU::AReg_320RegClassID;
      case 352:
        return AMDGPU::AReg_352RegClassID;
      case 384:
        return AMDGPU::AReg_384RegClassID;
      case 512:
        return AMDGPU::AReg_512RegClassID;
      case 1024:
        return AMDGPU::AReg_1024RegClassID;
    }
  }
  return -1;
}
 
static unsigned getSpecialRegForName(StringRef RegName) {
  return StringSwitch<unsigned>(RegName)
    .Case("exec", AMDGPU::EXEC)
    .Case("vcc", AMDGPU::VCC)
    .Case("flat_scratch", AMDGPU::FLAT_SCR)
    .Case("xnack_mask", AMDGPU::XNACK_MASK)
    .Case("shared_base", AMDGPU::SRC_SHARED_BASE)
    .Case("src_shared_base", AMDGPU::SRC_SHARED_BASE)
    .Case("shared_limit", AMDGPU::SRC_SHARED_LIMIT)
    .Case("src_shared_limit", AMDGPU::SRC_SHARED_LIMIT)
    .Case("private_base", AMDGPU::SRC_PRIVATE_BASE)
    .Case("src_private_base", AMDGPU::SRC_PRIVATE_BASE)
    .Case("private_limit", AMDGPU::SRC_PRIVATE_LIMIT)
    .Case("src_private_limit", AMDGPU::SRC_PRIVATE_LIMIT)
    .Case("pops_exiting_wave_id", AMDGPU::SRC_POPS_EXITING_WAVE_ID)
    .Case("src_pops_exiting_wave_id", AMDGPU::SRC_POPS_EXITING_WAVE_ID)
    .Case("lds_direct", AMDGPU::LDS_DIRECT)
    .Case("src_lds_direct", AMDGPU::LDS_DIRECT)
    .Case("m0", AMDGPU::M0)
    .Case("vccz", AMDGPU::SRC_VCCZ)
    .Case("src_vccz", AMDGPU::SRC_VCCZ)
    .Case("execz", AMDGPU::SRC_EXECZ)
    .Case("src_execz", AMDGPU::SRC_EXECZ)
    .Case("scc", AMDGPU::SRC_SCC)
    .Case("src_scc", AMDGPU::SRC_SCC)
    .Case("tba", AMDGPU::TBA)
    .Case("tma", AMDGPU::TMA)
    .Case("flat_scratch_lo", AMDGPU::FLAT_SCR_LO)
    .Case("flat_scratch_hi", AMDGPU::FLAT_SCR_HI)
    .Case("xnack_mask_lo", AMDGPU::XNACK_MASK_LO)
    .Case("xnack_mask_hi", AMDGPU::XNACK_MASK_HI)
    .Case("vcc_lo", AMDGPU::VCC_LO)
    .Case("vcc_hi", AMDGPU::VCC_HI)
    .Case("exec_lo", AMDGPU::EXEC_LO)
    .Case("exec_hi", AMDGPU::EXEC_HI)
    .Case("tma_lo", AMDGPU::TMA_LO)
    .Case("tma_hi", AMDGPU::TMA_HI)
    .Case("tba_lo", AMDGPU::TBA_LO)
    .Case("tba_hi", AMDGPU::TBA_HI)
    .Case("pc", AMDGPU::PC_REG)
    .Case("null", AMDGPU::SGPR_NULL)
    .Default(AMDGPU::NoRegister);
}
 
bool AMDGPUAsmParser::ParseRegister(MCRegister &RegNo, SMLoc &StartLoc,
                                    SMLoc &EndLoc, bool RestoreOnFailure) {
  auto R = parseRegister();
  if (!R) return true;
  assert(R->isReg());
  RegNo = R->getReg();
  StartLoc = R->getStartLoc();
  EndLoc = R->getEndLoc();
  return false;
}
 
bool AMDGPUAsmParser::parseRegister(MCRegister &RegNo, SMLoc &StartLoc,
                                    SMLoc &EndLoc) {
  return ParseRegister(RegNo, StartLoc, EndLoc, /*RestoreOnFailure=*/false);
}
 
OperandMatchResultTy AMDGPUAsmParser::tryParseRegister(MCRegister &RegNo,
                                                       SMLoc &StartLoc,
                                                       SMLoc &EndLoc) {
  bool Result =
      ParseRegister(RegNo, StartLoc, EndLoc, /*RestoreOnFailure=*/true);
  bool PendingErrors = getParser().hasPendingError();
  getParser().clearPendingErrors();
  if (PendingErrors)
    return MatchOperand_ParseFail;
  if (Result)
    return MatchOperand_NoMatch;
  return MatchOperand_Success;
}
 
bool AMDGPUAsmParser::AddNextRegisterToList(unsigned &Reg, unsigned &RegWidth,
                                            RegisterKind RegKind, unsigned Reg1,
                                            SMLoc Loc) {
  switch (RegKind) {
  case IS_SPECIAL:
    if (Reg == AMDGPU::EXEC_LO && Reg1 == AMDGPU::EXEC_HI) {
      Reg = AMDGPU::EXEC;
      RegWidth = 64;
      return true;
    }
    if (Reg == AMDGPU::FLAT_SCR_LO && Reg1 == AMDGPU::FLAT_SCR_HI) {
      Reg = AMDGPU::FLAT_SCR;
      RegWidth = 64;
      return true;
    }
    if (Reg == AMDGPU::XNACK_MASK_LO && Reg1 == AMDGPU::XNACK_MASK_HI) {
      Reg = AMDGPU::XNACK_MASK;
      RegWidth = 64;
      return true;
    }
    if (Reg == AMDGPU::VCC_LO && Reg1 == AMDGPU::VCC_HI) {
      Reg = AMDGPU::VCC;
      RegWidth = 64;
      return true;
    }
    if (Reg == AMDGPU::TBA_LO && Reg1 == AMDGPU::TBA_HI) {
      Reg = AMDGPU::TBA;
      RegWidth = 64;
      return true;
    }
    if (Reg == AMDGPU::TMA_LO && Reg1 == AMDGPU::TMA_HI) {
      Reg = AMDGPU::TMA;
      RegWidth = 64;
      return true;
    }
    Error(Loc, "register does not fit in the list");
    return false;
  case IS_VGPR:
  case IS_SGPR:
  case IS_AGPR:
  case IS_TTMP:
    if (Reg1 != Reg + RegWidth / 32) {
      Error(Loc, "registers in a list must have consecutive indices");
      return false;
    }
    RegWidth += 32;
    return true;
  default:
    llvm_unreachable("unexpected register kind");
  }
}
 
struct RegInfo {
  StringLiteral Name;
  RegisterKind Kind;
};
 
static constexpr RegInfo RegularRegisters[] = {
  {{"v"},    IS_VGPR},
  {{"s"},    IS_SGPR},
  {{"ttmp"}, IS_TTMP},
  {{"acc"},  IS_AGPR},
  {{"a"},    IS_AGPR},
};
 
static bool isRegularReg(RegisterKind Kind) {
  return Kind == IS_VGPR ||
         Kind == IS_SGPR ||
         Kind == IS_TTMP ||
         Kind == IS_AGPR;
}
 
static const RegInfo* getRegularRegInfo(StringRef Str) {
  for (const RegInfo &Reg : RegularRegisters)
    if (Str.startswith(Reg.Name))
      return &Reg;
  return nullptr;
}
 
static bool getRegNum(StringRef Str, unsigned& Num) {
  return !Str.getAsInteger(10, Num);
}
 
bool
AMDGPUAsmParser::isRegister(const AsmToken &Token,
                            const AsmToken &NextToken) const {
 
  // A list of consecutive registers: [s0,s1,s2,s3]
  if (Token.is(AsmToken::LBrac))
    return true;
 
  if (!Token.is(AsmToken::Identifier))
    return false;
 
  // A single register like s0 or a range of registers like s[0:1]
 
  StringRef Str = Token.getString();
  const RegInfo *Reg = getRegularRegInfo(Str);
  if (Reg) {
    StringRef RegName = Reg->Name;
    StringRef RegSuffix = Str.substr(RegName.size());
    if (!RegSuffix.empty()) {
      unsigned Num;
      // A single register with an index: rXX
      if (getRegNum(RegSuffix, Num))
        return true;
    } else {
      // A range of registers: r[XX:YY].
      if (NextToken.is(AsmToken::LBrac))
        return true;
    }
  }
 
  return getSpecialRegForName(Str) != AMDGPU::NoRegister;
}
 
bool
AMDGPUAsmParser::isRegister()
{
  return isRegister(getToken(), peekToken());
}
 
unsigned
AMDGPUAsmParser::getRegularReg(RegisterKind RegKind,
                               unsigned RegNum,
                               unsigned RegWidth,
                               SMLoc Loc) {
 
  assert(isRegularReg(RegKind));
 
  unsigned AlignSize = 1;
  if (RegKind == IS_SGPR || RegKind == IS_TTMP) {
    // SGPR and TTMP registers must be aligned.
    // Max required alignment is 4 dwords.
    AlignSize = std::min(RegWidth / 32, 4u);
  }
 
  if (RegNum % AlignSize != 0) {
    Error(Loc, "invalid register alignment");
    return AMDGPU::NoRegister;
  }
 
  unsigned RegIdx = RegNum / AlignSize;
  int RCID = getRegClass(RegKind, RegWidth);
  if (RCID == -1) {
    Error(Loc, "invalid or unsupported register size");
    return AMDGPU::NoRegister;
  }
 
  const MCRegisterInfo *TRI = getContext().getRegisterInfo();
  const MCRegisterClass RC = TRI->getRegClass(RCID);
  if (RegIdx >= RC.getNumRegs()) {
    Error(Loc, "register index is out of range");
    return AMDGPU::NoRegister;
  }
 
  return RC.getRegister(RegIdx);
}
 
bool AMDGPUAsmParser::ParseRegRange(unsigned &Num, unsigned &RegWidth) {
  int64_t RegLo, RegHi;
  if (!skipToken(AsmToken::LBrac, "missing register index"))
    return false;
 
  SMLoc FirstIdxLoc = getLoc();
  SMLoc SecondIdxLoc;
 
  if (!parseExpr(RegLo))
    return false;
 
  if (trySkipToken(AsmToken::Colon)) {
    SecondIdxLoc = getLoc();
    if (!parseExpr(RegHi))
      return false;
  } else {
    RegHi = RegLo;
  }
 
  if (!skipToken(AsmToken::RBrac, "expected a closing square bracket"))
    return false;
 
  if (!isUInt<32>(RegLo)) {
    Error(FirstIdxLoc, "invalid register index");
    return false;
  }
 
  if (!isUInt<32>(RegHi)) {
    Error(SecondIdxLoc, "invalid register index");
    return false;
  }
 
  if (RegLo > RegHi) {
    Error(FirstIdxLoc, "first register index should not exceed second index");
    return false;
  }
 
  Num = static_cast<unsigned>(RegLo);
  RegWidth = 32 * ((RegHi - RegLo) + 1);
  return true;
}
 
unsigned AMDGPUAsmParser::ParseSpecialReg(RegisterKind &RegKind,
                                          unsigned &RegNum, unsigned &RegWidth,
                                          SmallVectorImpl<AsmToken> &Tokens) {
  assert(isToken(AsmToken::Identifier));
  unsigned Reg = getSpecialRegForName(getTokenStr());
  if (Reg) {
    RegNum = 0;
    RegWidth = 32;
    RegKind = IS_SPECIAL;
    Tokens.push_back(getToken());
    lex(); // skip register name
  }
  return Reg;
}
 
unsigned AMDGPUAsmParser::ParseRegularReg(RegisterKind &RegKind,
                                          unsigned &RegNum, unsigned &RegWidth,
                                          SmallVectorImpl<AsmToken> &Tokens) {
  assert(isToken(AsmToken::Identifier));
  StringRef RegName = getTokenStr();
  auto Loc = getLoc();
 
  const RegInfo *RI = getRegularRegInfo(RegName);
  if (!RI) {
    Error(Loc, "invalid register name");
    return AMDGPU::NoRegister;
  }
 
  Tokens.push_back(getToken());
  lex(); // skip register name
 
  RegKind = RI->Kind;
  StringRef RegSuffix = RegName.substr(RI->Name.size());
  if (!RegSuffix.empty()) {
    // Single 32-bit register: vXX.
    if (!getRegNum(RegSuffix, RegNum)) {
      Error(Loc, "invalid register index");
      return AMDGPU::NoRegister;
    }
    RegWidth = 32;
  } else {
    // Range of registers: v[XX:YY]. ":YY" is optional.
    if (!ParseRegRange(RegNum, RegWidth))
      return AMDGPU::NoRegister;
  }
 
  return getRegularReg(RegKind, RegNum, RegWidth, Loc);
}
 
unsigned AMDGPUAsmParser::ParseRegList(RegisterKind &RegKind, unsigned &RegNum,
                                       unsigned &RegWidth,
                                       SmallVectorImpl<AsmToken> &Tokens) {
  unsigned Reg = AMDGPU::NoRegister;
  auto ListLoc = getLoc();
 
  if (!skipToken(AsmToken::LBrac,
                 "expected a register or a list of registers")) {
    return AMDGPU::NoRegister;
  }
 
  // List of consecutive registers, e.g.: [s0,s1,s2,s3]
 
  auto Loc = getLoc();
  if (!ParseAMDGPURegister(RegKind, Reg, RegNum, RegWidth))
    return AMDGPU::NoRegister;
  if (RegWidth != 32) {
    Error(Loc, "expected a single 32-bit register");
    return AMDGPU::NoRegister;
  }
 
  for (; trySkipToken(AsmToken::Comma); ) {
    RegisterKind NextRegKind;
    unsigned NextReg, NextRegNum, NextRegWidth;
    Loc = getLoc();
 
    if (!ParseAMDGPURegister(NextRegKind, NextReg,
                             NextRegNum, NextRegWidth,
                             Tokens)) {
      return AMDGPU::NoRegister;
    }
    if (NextRegWidth != 32) {
      Error(Loc, "expected a single 32-bit register");
      return AMDGPU::NoRegister;
    }
    if (NextRegKind != RegKind) {
      Error(Loc, "registers in a list must be of the same kind");
      return AMDGPU::NoRegister;
    }
    if (!AddNextRegisterToList(Reg, RegWidth, RegKind, NextReg, Loc))
      return AMDGPU::NoRegister;
  }
 
  if (!skipToken(AsmToken::RBrac,
                 "expected a comma or a closing square bracket")) {
    return AMDGPU::NoRegister;
  }
 
  if (isRegularReg(RegKind))
    Reg = getRegularReg(RegKind, RegNum, RegWidth, ListLoc);
 
  return Reg;
}
 
bool AMDGPUAsmParser::ParseAMDGPURegister(RegisterKind &RegKind, unsigned &Reg,
                                          unsigned &RegNum, unsigned &RegWidth,
                                          SmallVectorImpl<AsmToken> &Tokens) {
  auto Loc = getLoc();
  Reg = AMDGPU::NoRegister;
 
  if (isToken(AsmToken::Identifier)) {
    Reg = ParseSpecialReg(RegKind, RegNum, RegWidth, Tokens);
    if (Reg == AMDGPU::NoRegister)
      Reg = ParseRegularReg(RegKind, RegNum, RegWidth, Tokens);
  } else {
    Reg = ParseRegList(RegKind, RegNum, RegWidth, Tokens);
  }
 
  const MCRegisterInfo *TRI = getContext().getRegisterInfo();
  if (Reg == AMDGPU::NoRegister) {
    assert(Parser.hasPendingError());
    return false;
  }
 
  if (!subtargetHasRegister(*TRI, Reg)) {
    if (Reg == AMDGPU::SGPR_NULL) {
      Error(Loc, "'null' operand is not supported on this GPU");
    } else {
      Error(Loc, "register not available on this GPU");
    }
    return false;
  }
 
  return true;
}
 
bool AMDGPUAsmParser::ParseAMDGPURegister(RegisterKind &RegKind, unsigned &Reg,
                                          unsigned &RegNum, unsigned &RegWidth,
                                          bool RestoreOnFailure /*=false*/) {
  Reg = AMDGPU::NoRegister;
 
  SmallVector<AsmToken, 1> Tokens;
  if (ParseAMDGPURegister(RegKind, Reg, RegNum, RegWidth, Tokens)) {
    if (RestoreOnFailure) {
      while (!Tokens.empty()) {
        getLexer().UnLex(Tokens.pop_back_val());
      }
    }
    return true;
  }
  return false;
}
 
std::optional<StringRef>
AMDGPUAsmParser::getGprCountSymbolName(RegisterKind RegKind) {
  switch (RegKind) {
  case IS_VGPR:
    return StringRef(".amdgcn.next_free_vgpr");
  case IS_SGPR:
    return StringRef(".amdgcn.next_free_sgpr");
  default:
    return std::nullopt;
  }
}
 
void AMDGPUAsmParser::initializeGprCountSymbol(RegisterKind RegKind) {
  auto SymbolName = getGprCountSymbolName(RegKind);
  assert(SymbolName && "initializing invalid register kind");
  MCSymbol *Sym = getContext().getOrCreateSymbol(*SymbolName);
  Sym->setVariableValue(MCConstantExpr::create(0, getContext()));
}
 
bool AMDGPUAsmParser::updateGprCountSymbols(RegisterKind RegKind,
                                            unsigned DwordRegIndex,
                                            unsigned RegWidth) {
  // Symbols are only defined for GCN targets
  if (AMDGPU::getIsaVersion(getSTI().getCPU()).Major < 6)
    return true;
 
  auto SymbolName = getGprCountSymbolName(RegKind);
  if (!SymbolName)
    return true;
  MCSymbol *Sym = getContext().getOrCreateSymbol(*SymbolName);
 
  int64_t NewMax = DwordRegIndex + divideCeil(RegWidth, 32) - 1;
  int64_t OldCount;
 
  if (!Sym->isVariable())
    return !Error(getLoc(),
                  ".amdgcn.next_free_{v,s}gpr symbols must be variable");
  if (!Sym->getVariableValue(false)->evaluateAsAbsolute(OldCount))
    return !Error(
        getLoc(),
        ".amdgcn.next_free_{v,s}gpr symbols must be absolute expressions");
 
  if (OldCount <= NewMax)
    Sym->setVariableValue(MCConstantExpr::create(NewMax + 1, getContext()));
 
  return true;
}
 
std::unique_ptr<AMDGPUOperand>
AMDGPUAsmParser::parseRegister(bool RestoreOnFailure) {
  const auto &Tok = getToken();
  SMLoc StartLoc = Tok.getLoc();
  SMLoc EndLoc = Tok.getEndLoc();
  RegisterKind RegKind;
  unsigned Reg, RegNum, RegWidth;
 
  if (!ParseAMDGPURegister(RegKind, Reg, RegNum, RegWidth)) {
    return nullptr;
  }
  if (isHsaAbiVersion3AndAbove(&getSTI())) {
    if (!updateGprCountSymbols(RegKind, RegNum, RegWidth))
      return nullptr;
  } else
    KernelScope.usesRegister(RegKind, RegNum, RegWidth);
  return AMDGPUOperand::CreateReg(this, Reg, StartLoc, EndLoc);
}
 
OperandMatchResultTy
AMDGPUAsmParser::parseImm(OperandVector &Operands, bool HasSP3AbsModifier) {
  // TODO: add syntactic sugar for 1/(2*PI)
 
  if (isRegister())
    return MatchOperand_NoMatch;
  assert(!isModifier());
 
  const auto& Tok = getToken();
  const auto& NextTok = peekToken();
  bool IsReal = Tok.is(AsmToken::Real);
  SMLoc S = getLoc();
  bool Negate = false;
 
  if (!IsReal && Tok.is(AsmToken::Minus) && NextTok.is(AsmToken::Real)) {
    lex();
    IsReal = true;
    Negate = true;
  }
 
  if (IsReal) {
    // Floating-point expressions are not supported.
    // Can only allow floating-point literals with an
    // optional sign.
 
    StringRef Num = getTokenStr();
    lex();
 
    APFloat RealVal(APFloat::IEEEdouble());
    auto roundMode = APFloat::rmNearestTiesToEven;
    if (errorToBool(RealVal.convertFromString(Num, roundMode).takeError())) {
      return MatchOperand_ParseFail;
    }
    if (Negate)
      RealVal.changeSign();
 
    Operands.push_back(
      AMDGPUOperand::CreateImm(this, RealVal.bitcastToAPInt().getZExtValue(), S,
                               AMDGPUOperand::ImmTyNone, true));
 
    return MatchOperand_Success;
 
  } else {
    int64_t IntVal;
    const MCExpr *Expr;
    SMLoc S = getLoc();
 
    if (HasSP3AbsModifier) {
      // This is a workaround for handling expressions
      // as arguments of SP3 'abs' modifier, for example:
      //     |1.0|
      //     |-1|
      //     |1+x|
      // This syntax is not compatible with syntax of standard
      // MC expressions (due to the trailing '|').
      SMLoc EndLoc;
      if (getParser().parsePrimaryExpr(Expr, EndLoc, nullptr))
        return MatchOperand_ParseFail;
    } else {
      if (Parser.parseExpression(Expr))
        return MatchOperand_ParseFail;
    }
 
    if (Expr->evaluateAsAbsolute(IntVal)) {
      Operands.push_back(AMDGPUOperand::CreateImm(this, IntVal, S));
    } else {
      Operands.push_back(AMDGPUOperand::CreateExpr(this, Expr, S));
    }
 
    return MatchOperand_Success;
  }
 
  return MatchOperand_NoMatch;
}
 
OperandMatchResultTy
AMDGPUAsmParser::parseReg(OperandVector &Operands) {
  if (!isRegister())
    return MatchOperand_NoMatch;
 
  if (auto R = parseRegister()) {
    assert(R->isReg());
    Operands.push_back(std::move(R));
    return MatchOperand_Success;
  }
  return MatchOperand_ParseFail;
}
 
OperandMatchResultTy
AMDGPUAsmParser::parseRegOrImm(OperandVector &Operands, bool HasSP3AbsMod) {
  auto res = parseReg(Operands);
  if (res != MatchOperand_NoMatch) {
    return res;
  } else if (isModifier()) {
    return MatchOperand_NoMatch;
  } else {
    return parseImm(Operands, HasSP3AbsMod);
  }
}
 
bool
AMDGPUAsmParser::isNamedOperandModifier(const AsmToken &Token, const AsmToken &NextToken) const {
  if (Token.is(AsmToken::Identifier) && NextToken.is(AsmToken::LParen)) {
    const auto &str = Token.getString();
    return str == "abs" || str == "neg" || str == "sext";
  }
  return false;
}
 
bool
AMDGPUAsmParser::isOpcodeModifierWithVal(const AsmToken &Token, const AsmToken &NextToken) const {
  return Token.is(AsmToken::Identifier) && NextToken.is(AsmToken::Colon);
}
 
bool
AMDGPUAsmParser::isOperandModifier(const AsmToken &Token, const AsmToken &NextToken) const {
  return isNamedOperandModifier(Token, NextToken) || Token.is(AsmToken::Pipe);
}
 
bool
AMDGPUAsmParser::isRegOrOperandModifier(const AsmToken &Token, const AsmToken &NextToken) const {
  return isRegister(Token, NextToken) || isOperandModifier(Token, NextToken);
}
 
// Check if this is an operand modifier or an opcode modifier
// which may look like an expression but it is not. We should
// avoid parsing these modifiers as expressions. Currently
// recognized sequences are:
//   |...|
//   abs(...)
//   neg(...)
//   sext(...)
//   -reg
//   -|...|
//   -abs(...)
//   name:...
//
bool
AMDGPUAsmParser::isModifier() {
 
  AsmToken Tok = getToken();
  AsmToken NextToken[2];
  peekTokens(NextToken);
 
  return isOperandModifier(Tok, NextToken[0]) ||
         (Tok.is(AsmToken::Minus) && isRegOrOperandModifier(NextToken[0], NextToken[1])) ||
         isOpcodeModifierWithVal(Tok, NextToken[0]);
}
 
// Check if the current token is an SP3 'neg' modifier.
// Currently this modifier is allowed in the following context:
//
// 1. Before a register, e.g. "-v0", "-v[...]" or "-[v0,v1]".
// 2. Before an 'abs' modifier: -abs(...)
// 3. Before an SP3 'abs' modifier: -|...|
//
// In all other cases "-" is handled as a part
// of an expression that follows the sign.
//
// Note: When "-" is followed by an integer literal,
// this is interpreted as integer negation rather
// than a floating-point NEG modifier applied to N.
// Beside being contr-intuitive, such use of floating-point
// NEG modifier would have resulted in different meaning
// of integer literals used with VOP1/2/C and VOP3,
// for example:
//    v_exp_f32_e32 v5, -1 // VOP1: src0 = 0xFFFFFFFF
//    v_exp_f32_e64 v5, -1 // VOP3: src0 = 0x80000001
// Negative fp literals with preceding "-" are
// handled likewise for uniformity
//
bool
AMDGPUAsmParser::parseSP3NegModifier() {
 
  AsmToken NextToken[2];
  peekTokens(NextToken);
 
  if (isToken(AsmToken::Minus) &&
      (isRegister(NextToken[0], NextToken[1]) ||
       NextToken[0].is(AsmToken::Pipe) ||
       isId(NextToken[0], "abs"))) {
    lex();
    return true;
  }
 
  return false;
}
 
OperandMatchResultTy
AMDGPUAsmParser::parseRegOrImmWithFPInputMods(OperandVector &Operands,
                                              bool AllowImm) {
  bool Neg, SP3Neg;
  bool Abs, SP3Abs;
  SMLoc Loc;
 
  // Disable ambiguous constructs like '--1' etc. Should use neg(-1) instead.
  if (isToken(AsmToken::Minus) && peekToken().is(AsmToken::Minus)) {
    Error(getLoc(), "invalid syntax, expected 'neg' modifier");
    return MatchOperand_ParseFail;
  }
 
  SP3Neg = parseSP3NegModifier();
 
  Loc = getLoc();
  Neg = trySkipId("neg");
  if (Neg && SP3Neg) {
    Error(Loc, "expected register or immediate");
    return MatchOperand_ParseFail;
  }
  if (Neg && !skipToken(AsmToken::LParen, "expected left paren after neg"))
    return MatchOperand_ParseFail;
 
  Abs = trySkipId("abs");
  if (Abs && !skipToken(AsmToken::LParen, "expected left paren after abs"))
    return MatchOperand_ParseFail;
 
  Loc = getLoc();
  SP3Abs = trySkipToken(AsmToken::Pipe);
  if (Abs && SP3Abs) {
    Error(Loc, "expected register or immediate");
    return MatchOperand_ParseFail;
  }
 
  OperandMatchResultTy Res;
  if (AllowImm) {
    Res = parseRegOrImm(Operands, SP3Abs);
  } else {
    Res = parseReg(Operands);
  }
  if (Res != MatchOperand_Success) {
    return (SP3Neg || Neg || SP3Abs || Abs)? MatchOperand_ParseFail : Res;
  }
 
  if (SP3Abs && !skipToken(AsmToken::Pipe, "expected vertical bar"))
    return MatchOperand_ParseFail;
  if (Abs && !skipToken(AsmToken::RParen, "expected closing parentheses"))
    return MatchOperand_ParseFail;
  if (Neg && !skipToken(AsmToken::RParen, "expected closing parentheses"))
    return MatchOperand_ParseFail;
 
  AMDGPUOperand::Modifiers Mods;
  Mods.Abs = Abs || SP3Abs;
  Mods.Neg = Neg || SP3Neg;
 
  if (Mods.hasFPModifiers()) {
    AMDGPUOperand &Op = static_cast<AMDGPUOperand &>(*Operands.back());
    if (Op.isExpr()) {
      Error(Op.getStartLoc(), "expected an absolute expression");
      return MatchOperand_ParseFail;
    }
    Op.setModifiers(Mods);
  }
  return MatchOperand_Success;
}
 
OperandMatchResultTy
AMDGPUAsmParser::parseRegOrImmWithIntInputMods(OperandVector &Operands,
                                               bool AllowImm) {
  bool Sext = trySkipId("sext");
  if (Sext && !skipToken(AsmToken::LParen, "expected left paren after sext"))
    return MatchOperand_ParseFail;
 
  OperandMatchResultTy Res;
  if (AllowImm) {
    Res = parseRegOrImm(Operands);
  } else {
    Res = parseReg(Operands);
  }
  if (Res != MatchOperand_Success) {
    return Sext? MatchOperand_ParseFail : Res;
  }
 
  if (Sext && !skipToken(AsmToken::RParen, "expected closing parentheses"))
    return MatchOperand_ParseFail;
 
  AMDGPUOperand::Modifiers Mods;
  Mods.Sext = Sext;
 
  if (Mods.hasIntModifiers()) {
    AMDGPUOperand &Op = static_cast<AMDGPUOperand &>(*Operands.back());
    if (Op.isExpr()) {
      Error(Op.getStartLoc(), "expected an absolute expression");
      return MatchOperand_ParseFail;
    }
    Op.setModifiers(Mods);
  }
 
  return MatchOperand_Success;
}
 
OperandMatchResultTy
AMDGPUAsmParser::parseRegWithFPInputMods(OperandVector &Operands) {
  return parseRegOrImmWithFPInputMods(Operands, false);
}
 
OperandMatchResultTy
AMDGPUAsmParser::parseRegWithIntInputMods(OperandVector &Operands) {
  return parseRegOrImmWithIntInputMods(Operands, false);
}
 
OperandMatchResultTy AMDGPUAsmParser::parseVReg32OrOff(OperandVector &Operands) {
  auto Loc = getLoc();
  if (trySkipId("off")) {
    Operands.push_back(AMDGPUOperand::CreateImm(this, 0, Loc,
                                                AMDGPUOperand::ImmTyOff, false));
    return MatchOperand_Success;
  }
 
  if (!isRegister())
    return MatchOperand_NoMatch;
 
  std::unique_ptr<AMDGPUOperand> Reg = parseRegister();
  if (Reg) {
    Operands.push_back(std::move(Reg));
    return MatchOperand_Success;
  }
 
  return MatchOperand_ParseFail;
 
}
 
unsigned AMDGPUAsmParser::checkTargetMatchPredicate(MCInst &Inst) {
  uint64_t TSFlags = MII.get(Inst.getOpcode()).TSFlags;
 
  if ((getForcedEncodingSize() == 32 && (TSFlags & SIInstrFlags::VOP3)) ||
      (getForcedEncodingSize() == 64 && !(TSFlags & SIInstrFlags::VOP3)) ||
      (isForcedDPP() && !(TSFlags & SIInstrFlags::DPP)) ||
      (isForcedSDWA() && !(TSFlags & SIInstrFlags::SDWA)) )
    return Match_InvalidOperand;
 
  if ((TSFlags & SIInstrFlags::VOP3) &&
      (TSFlags & SIInstrFlags::VOPAsmPrefer32Bit) &&
      getForcedEncodingSize() != 64)
    return Match_PreferE32;
 
  if (Inst.getOpcode() == AMDGPU::V_MAC_F32_sdwa_vi ||
      Inst.getOpcode() == AMDGPU::V_MAC_F16_sdwa_vi) {
    // v_mac_f32/16 allow only dst_sel == DWORD;
    auto OpNum =
        AMDGPU::getNamedOperandIdx(Inst.getOpcode(), AMDGPU::OpName::dst_sel);
    const auto &Op = Inst.getOperand(OpNum);
    if (!Op.isImm() || Op.getImm() != AMDGPU::SDWA::SdwaSel::DWORD) {
      return Match_InvalidOperand;
    }
  }
 
  return Match_Success;
}
 
static ArrayRef<unsigned> getAllVariants() {
  static const unsigned Variants[] = {
    AMDGPUAsmVariants::DEFAULT, AMDGPUAsmVariants::VOP3,
    AMDGPUAsmVariants::SDWA, AMDGPUAsmVariants::SDWA9,
    AMDGPUAsmVariants::DPP, AMDGPUAsmVariants::VOP3_DPP
  };
 
  return ArrayRef(Variants);
}
 
// What asm variants we should check
ArrayRef<unsigned> AMDGPUAsmParser::getMatchedVariants() const {
  if (isForcedDPP() && isForcedVOP3()) {
    static const unsigned Variants[] = {AMDGPUAsmVariants::VOP3_DPP};
    return ArrayRef(Variants);
  }
  if (getForcedEncodingSize() == 32) {
    static const unsigned Variants[] = {AMDGPUAsmVariants::DEFAULT};
    return ArrayRef(Variants);
  }
 
  if (isForcedVOP3()) {
    static const unsigned Variants[] = {AMDGPUAsmVariants::VOP3};
    return ArrayRef(Variants);
  }
 
  if (isForcedSDWA()) {
    static const unsigned Variants[] = {AMDGPUAsmVariants::SDWA,
                                        AMDGPUAsmVariants::SDWA9};
    return ArrayRef(Variants);
  }
 
  if (isForcedDPP()) {
    static const unsigned Variants[] = {AMDGPUAsmVariants::DPP};
    return ArrayRef(Variants);
  }
 
  return getAllVariants();
}
 
StringRef AMDGPUAsmParser::getMatchedVariantName() const {
  if (isForcedDPP() && isForcedVOP3())
    return "e64_dpp";
 
  if (getForcedEncodingSize() == 32)
    return "e32";
 
  if (isForcedVOP3())
    return "e64";
 
  if (isForcedSDWA())
    return "sdwa";
 
  if (isForcedDPP())
    return "dpp";
 
  return "";
}
 
unsigned AMDGPUAsmParser::findImplicitSGPRReadInVOP(const MCInst &Inst) const {
  const MCInstrDesc &Desc = MII.get(Inst.getOpcode());
  for (MCPhysReg Reg : Desc.implicit_uses()) {
    switch (Reg) {
    case AMDGPU::FLAT_SCR:
    case AMDGPU::VCC:
    case AMDGPU::VCC_LO:
    case AMDGPU::VCC_HI:
    case AMDGPU::M0:
      return Reg;
    default:
      break;
    }
  }
  return AMDGPU::NoRegister;
}
 
// NB: This code is correct only when used to check constant
// bus limitations because GFX7 support no f16 inline constants.
// Note that there are no cases when a GFX7 opcode violates
// constant bus limitations due to the use of an f16 constant.
bool AMDGPUAsmParser::isInlineConstant(const MCInst &Inst,
                                       unsigned OpIdx) const {
  const MCInstrDesc &Desc = MII.get(Inst.getOpcode());
 
  if (!AMDGPU::isSISrcOperand(Desc, OpIdx) ||
      AMDGPU::isKImmOperand(Desc, OpIdx)) {
    return false;
  }
 
  const MCOperand &MO = Inst.getOperand(OpIdx);
 
  int64_t Val = MO.getImm();
  auto OpSize = AMDGPU::getOperandSize(Desc, OpIdx);
 
  switch (OpSize) { // expected operand size
  case 8:
    return AMDGPU::isInlinableLiteral64(Val, hasInv2PiInlineImm());
  case 4:
    return AMDGPU::isInlinableLiteral32(Val, hasInv2PiInlineImm());
  case 2: {
    const unsigned OperandType = Desc.operands()[OpIdx].OperandType;
    if (OperandType == AMDGPU::OPERAND_REG_IMM_INT16 ||
        OperandType == AMDGPU::OPERAND_REG_INLINE_C_INT16 ||
        OperandType == AMDGPU::OPERAND_REG_INLINE_AC_INT16)
      return AMDGPU::isInlinableIntLiteral(Val);
 
    if (OperandType == AMDGPU::OPERAND_REG_INLINE_C_V2INT16 ||
        OperandType == AMDGPU::OPERAND_REG_INLINE_AC_V2INT16 ||
        OperandType == AMDGPU::OPERAND_REG_IMM_V2INT16)
      return AMDGPU::isInlinableIntLiteralV216(Val);
 
    if (OperandType == AMDGPU::OPERAND_REG_INLINE_C_V2FP16 ||
        OperandType == AMDGPU::OPERAND_REG_INLINE_AC_V2FP16 ||
        OperandType == AMDGPU::OPERAND_REG_IMM_V2FP16)
      return AMDGPU::isInlinableLiteralV216(Val, hasInv2PiInlineImm());
 
    return AMDGPU::isInlinableLiteral16(Val, hasInv2PiInlineImm());
  }
  default:
    llvm_unreachable("invalid operand size");
  }
}
 
unsigned AMDGPUAsmParser::getConstantBusLimit(unsigned Opcode) const {
  if (!isGFX10Plus())
    return 1;
 
  switch (Opcode) {
  // 64-bit shift instructions can use only one scalar value input
  case AMDGPU::V_LSHLREV_B64_e64:
  case AMDGPU::V_LSHLREV_B64_gfx10:
  case AMDGPU::V_LSHLREV_B64_e64_gfx11:
  case AMDGPU::V_LSHRREV_B64_e64:
  case AMDGPU::V_LSHRREV_B64_gfx10:
  case AMDGPU::V_LSHRREV_B64_e64_gfx11:
  case AMDGPU::V_ASHRREV_I64_e64:
  case AMDGPU::V_ASHRREV_I64_gfx10:
  case AMDGPU::V_ASHRREV_I64_e64_gfx11:
  case AMDGPU::V_LSHL_B64_e64:
  case AMDGPU::V_LSHR_B64_e64:
  case AMDGPU::V_ASHR_I64_e64:
    return 1;
  default:
    return 2;
  }
}
 
constexpr unsigned MAX_SRC_OPERANDS_NUM = 6;
using OperandIndices = SmallVector<int16_t, MAX_SRC_OPERANDS_NUM>;
 
// Get regular operand indices in the same order as specified
// in the instruction (but append mandatory literals to the end).
static OperandIndices getSrcOperandIndices(unsigned Opcode,
                                           bool AddMandatoryLiterals = false) {
 
  int16_t ImmIdx =
      AddMandatoryLiterals ? getNamedOperandIdx(Opcode, OpName::imm) : -1;
 
  if (isVOPD(Opcode)) {
    int16_t ImmDeferredIdx =
        AddMandatoryLiterals ? getNamedOperandIdx(Opcode, OpName::immDeferred)
                             : -1;
 
    return {getNamedOperandIdx(Opcode, OpName::src0X),
            getNamedOperandIdx(Opcode, OpName::vsrc1X),
            getNamedOperandIdx(Opcode, OpName::src0Y),
            getNamedOperandIdx(Opcode, OpName::vsrc1Y),
            ImmDeferredIdx,
            ImmIdx};
  }
 
  return {getNamedOperandIdx(Opcode, OpName::src0),
          getNamedOperandIdx(Opcode, OpName::src1),
          getNamedOperandIdx(Opcode, OpName::src2), ImmIdx};
}
 
bool AMDGPUAsmParser::usesConstantBus(const MCInst &Inst, unsigned OpIdx) {
  const MCOperand &MO = Inst.getOperand(OpIdx);
  if (MO.isImm()) {
    return !isInlineConstant(Inst, OpIdx);
  } else if (MO.isReg()) {
    auto Reg = MO.getReg();
    const MCRegisterInfo *TRI = getContext().getRegisterInfo();
    auto PReg = mc2PseudoReg(Reg);
    return isSGPR(PReg, TRI) && PReg != SGPR_NULL;
  } else {
    return true;
  }
}
 
bool AMDGPUAsmParser::validateConstantBusLimitations(
    const MCInst &Inst, const OperandVector &Operands) {
  const unsigned Opcode = Inst.getOpcode();
  const MCInstrDesc &Desc = MII.get(Opcode);
  unsigned LastSGPR = AMDGPU::NoRegister;
  unsigned ConstantBusUseCount = 0;
  unsigned NumLiterals = 0;
  unsigned LiteralSize;
 
  if (!(Desc.TSFlags &
        (SIInstrFlags::VOPC | SIInstrFlags::VOP1 | SIInstrFlags::VOP2 |
         SIInstrFlags::VOP3 | SIInstrFlags::VOP3P | SIInstrFlags::SDWA)) &&
      !isVOPD(Opcode))
    return true;
 
  // Check special imm operands (used by madmk, etc)
  if (AMDGPU::hasNamedOperand(Opcode, AMDGPU::OpName::imm)) {
    ++NumLiterals;
    LiteralSize = 4;
  }
 
  SmallDenseSet<unsigned> SGPRsUsed;
  unsigned SGPRUsed = findImplicitSGPRReadInVOP(Inst);
  if (SGPRUsed != AMDGPU::NoRegister) {
    SGPRsUsed.insert(SGPRUsed);
    ++ConstantBusUseCount;
  }
 
  OperandIndices OpIndices = getSrcOperandIndices(Opcode);
 
  for (int OpIdx : OpIndices) {
    if (OpIdx == -1)
      continue;
 
    const MCOperand &MO = Inst.getOperand(OpIdx);
    if (usesConstantBus(Inst, OpIdx)) {
      if (MO.isReg()) {
        LastSGPR = mc2PseudoReg(MO.getReg());
        // Pairs of registers with a partial intersections like these
        //   s0, s[0:1]
        //   flat_scratch_lo, flat_scratch
        //   flat_scratch_lo, flat_scratch_hi
        // are theoretically valid but they are disabled anyway.
        // Note that this code mimics SIInstrInfo::verifyInstruction
        if (SGPRsUsed.insert(LastSGPR).second) {
          ++ConstantBusUseCount;
        }
      } else { // Expression or a literal
 
        if (Desc.operands()[OpIdx].OperandType == MCOI::OPERAND_IMMEDIATE)
          continue; // special operand like VINTERP attr_chan
 
        // An instruction may use only one literal.
        // This has been validated on the previous step.
        // See validateVOPLiteral.
        // This literal may be used as more than one operand.
        // If all these operands are of the same size,
        // this literal counts as one scalar value.
        // Otherwise it counts as 2 scalar values.
        // See "GFX10 Shader Programming", section 3.6.2.3.
 
        unsigned Size = AMDGPU::getOperandSize(Desc, OpIdx);
        if (Size < 4)
          Size = 4;
 
        if (NumLiterals == 0) {
          NumLiterals = 1;
          LiteralSize = Size;
        } else if (LiteralSize != Size) {
          NumLiterals = 2;
        }
      }
    }
  }
  ConstantBusUseCount += NumLiterals;
 
  if (ConstantBusUseCount <= getConstantBusLimit(Opcode))
    return true;
 
  SMLoc LitLoc = getLitLoc(Operands);
  SMLoc RegLoc = getRegLoc(LastSGPR, Operands);
  SMLoc Loc = (LitLoc.getPointer() < RegLoc.getPointer()) ? RegLoc : LitLoc;
  Error(Loc, "invalid operand (violates constant bus restrictions)");
  return false;
}
 
bool AMDGPUAsmParser::validateVOPDRegBankConstraints(
    const MCInst &Inst, const OperandVector &Operands) {
 
  const unsigned Opcode = Inst.getOpcode();
  if (!isVOPD(Opcode))
    return true;
 
  const MCRegisterInfo *TRI = getContext().getRegisterInfo();
 
  auto getVRegIdx = [&](unsigned, unsigned OperandIdx) {
    const MCOperand &Opr = Inst.getOperand(OperandIdx);
    return (Opr.isReg() && !isSGPR(mc2PseudoReg(Opr.getReg()), TRI))
               ? Opr.getReg()
               : MCRegister::NoRegister;
  };
 
  const auto &InstInfo = getVOPDInstInfo(Opcode, &MII);
  auto InvalidCompOprIdx = InstInfo.getInvalidCompOperandIndex(getVRegIdx);
  if (!InvalidCompOprIdx)
    return true;
 
  auto CompOprIdx = *InvalidCompOprIdx;
  auto ParsedIdx =
      std::max(InstInfo[VOPD::X].getIndexInParsedOperands(CompOprIdx),
               InstInfo[VOPD::Y].getIndexInParsedOperands(CompOprIdx));
  assert(ParsedIdx > 0 && ParsedIdx < Operands.size());
 
  auto Loc = ((AMDGPUOperand &)*Operands[ParsedIdx]).getStartLoc();
  if (CompOprIdx == VOPD::Component::DST) {
    Error(Loc, "one dst register must be even and the other odd");
  } else {
    auto CompSrcIdx = CompOprIdx - VOPD::Component::DST_NUM;
    Error(Loc, Twine("src") + Twine(CompSrcIdx) +
                   " operands must use different VGPR banks");
  }
 
  return false;
}
 
bool AMDGPUAsmParser::validateIntClampSupported(const MCInst &Inst) {
 
  const unsigned Opc = Inst.getOpcode();
  const MCInstrDesc &Desc = MII.get(Opc);
 
  if ((Desc.TSFlags & SIInstrFlags::IntClamp) != 0 && !hasIntClamp()) {
    int ClampIdx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::clamp);
    assert(ClampIdx != -1);
    return Inst.getOperand(ClampIdx).getImm() == 0;
  }
 
  return true;
}
 
bool AMDGPUAsmParser::validateMIMGDataSize(const MCInst &Inst,
                                           const SMLoc &IDLoc) {
 
  const unsigned Opc = Inst.getOpcode();
  const MCInstrDesc &Desc = MII.get(Opc);
 
  if ((Desc.TSFlags & SIInstrFlags::MIMG) == 0)
    return true;
 
  int VDataIdx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::vdata);
  int DMaskIdx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::dmask);
  int TFEIdx   = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::tfe);
 
  assert(VDataIdx != -1);
 
  if ((DMaskIdx == -1 || TFEIdx == -1) && isGFX10_AEncoding()) // intersect_ray
    return true;
 
  unsigned VDataSize = AMDGPU::getRegOperandSize(getMRI(), Desc, VDataIdx);
  unsigned TFESize = (TFEIdx != -1 && Inst.getOperand(TFEIdx).getImm()) ? 1 : 0;
  unsigned DMask = Inst.getOperand(DMaskIdx).getImm() & 0xf;
  if (DMask == 0)
    DMask = 1;
 
  bool IsPackedD16 = false;
  unsigned DataSize =
      (Desc.TSFlags & SIInstrFlags::Gather4) ? 4 : llvm::popcount(DMask);
  if (hasPackedD16()) {
    int D16Idx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::d16);
    IsPackedD16 = D16Idx >= 0;
    if (IsPackedD16 && Inst.getOperand(D16Idx).getImm())
      DataSize = (DataSize + 1) / 2;
  }
 
  if ((VDataSize / 4) == DataSize + TFESize)
    return true;
 
  StringRef Modifiers;
  if (isGFX90A())
    Modifiers = IsPackedD16 ? "dmask and d16" : "dmask";
  else
    Modifiers = IsPackedD16 ? "dmask, d16 and tfe" : "dmask and tfe";
 
  Error(IDLoc, Twine("image data size does not match ") + Modifiers);
  return false;
}
 
bool AMDGPUAsmParser::validateMIMGAddrSize(const MCInst &Inst) {
  const unsigned Opc = Inst.getOpcode();
  const MCInstrDesc &Desc = MII.get(Opc);
 
  if ((Desc.TSFlags & SIInstrFlags::MIMG) == 0 || !isGFX10Plus())
    return true;
 
  const AMDGPU::MIMGInfo *Info = AMDGPU::getMIMGInfo(Opc);
 
  const AMDGPU::MIMGBaseOpcodeInfo *BaseOpcode =
      AMDGPU::getMIMGBaseOpcodeInfo(Info->BaseOpcode);
  int VAddr0Idx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::vaddr0);
  int SrsrcIdx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::srsrc);
  int DimIdx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::dim);
  int A16Idx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::a16);
 
  assert(VAddr0Idx != -1);
  assert(SrsrcIdx != -1);
  assert(SrsrcIdx > VAddr0Idx);
 
  if (DimIdx == -1)
    return true; // intersect_ray
 
  unsigned Dim = Inst.getOperand(DimIdx).getImm();
  const AMDGPU::MIMGDimInfo *DimInfo = AMDGPU::getMIMGDimInfoByEncoding(Dim);
  bool IsNSA = SrsrcIdx - VAddr0Idx > 1;
  unsigned ActualAddrSize =
      IsNSA ? SrsrcIdx - VAddr0Idx
            : AMDGPU::getRegOperandSize(getMRI(), Desc, VAddr0Idx) / 4;
  bool IsA16 = (A16Idx != -1 && Inst.getOperand(A16Idx).getImm());
 
  unsigned ExpectedAddrSize =
      AMDGPU::getAddrSizeMIMGOp(BaseOpcode, DimInfo, IsA16, hasG16());
 
  if (IsNSA) {
    if (hasPartialNSAEncoding() && ExpectedAddrSize > getNSAMaxSize()) {
      int VAddrLastIdx = SrsrcIdx - 1;
      unsigned VAddrLastSize =
          AMDGPU::getRegOperandSize(getMRI(), Desc, VAddrLastIdx) / 4;
 
      return VAddrLastIdx - VAddr0Idx + VAddrLastSize == ExpectedAddrSize;
    }
  } else {
    if (ExpectedAddrSize > 12)
      ExpectedAddrSize = 16;
 
    // Allow oversized 8 VGPR vaddr when only 5/6/7 VGPRs are required.
    // This provides backward compatibility for assembly created
    // before 160b/192b/224b types were directly supported.
    if (ActualAddrSize == 8 && (ExpectedAddrSize >= 5 && ExpectedAddrSize <= 7))
      return true;
  }
 
  return ActualAddrSize == ExpectedAddrSize;
}
 
bool AMDGPUAsmParser::validateMIMGAtomicDMask(const MCInst &Inst) {
 
  const unsigned Opc = Inst.getOpcode();
  const MCInstrDesc &Desc = MII.get(Opc);
 
  if ((Desc.TSFlags & SIInstrFlags::MIMG) == 0)
    return true;
  if (!Desc.mayLoad() || !Desc.mayStore())
    return true; // Not atomic
 
  int DMaskIdx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::dmask);
  unsigned DMask = Inst.getOperand(DMaskIdx).getImm() & 0xf;
 
  // This is an incomplete check because image_atomic_cmpswap
  // may only use 0x3 and 0xf while other atomic operations
  // may use 0x1 and 0x3. However these limitations are
  // verified when we check that dmask matches dst size.
  return DMask == 0x1 || DMask == 0x3 || DMask == 0xf;
}
 
bool AMDGPUAsmParser::validateMIMGGatherDMask(const MCInst &Inst) {
 
  const unsigned Opc = Inst.getOpcode();
  const MCInstrDesc &Desc = MII.get(Opc);
 
  if ((Desc.TSFlags & SIInstrFlags::Gather4) == 0)
    return true;
 
  int DMaskIdx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::dmask);
  unsigned DMask = Inst.getOperand(DMaskIdx).getImm() & 0xf;
 
  // GATHER4 instructions use dmask in a different fashion compared to
  // other MIMG instructions. The only useful DMASK values are
  // 1=red, 2=green, 4=blue, 8=alpha. (e.g. 1 returns
  // (red,red,red,red) etc.) The ISA document doesn't mention
  // this.
  return DMask == 0x1 || DMask == 0x2 || DMask == 0x4 || DMask == 0x8;
}
 
bool AMDGPUAsmParser::validateMIMGMSAA(const MCInst &Inst) {
  const unsigned Opc = Inst.getOpcode();
  const MCInstrDesc &Desc = MII.get(Opc);
 
  if ((Desc.TSFlags & SIInstrFlags::MIMG) == 0)
    return true;
 
  const AMDGPU::MIMGInfo *Info = AMDGPU::getMIMGInfo(Opc);
  const AMDGPU::MIMGBaseOpcodeInfo *BaseOpcode =
      AMDGPU::getMIMGBaseOpcodeInfo(Info->BaseOpcode);
 
  if (!BaseOpcode->MSAA)
    return true;
 
  int DimIdx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::dim);
  assert(DimIdx != -1);
 
  unsigned Dim = Inst.getOperand(DimIdx).getImm();
  const AMDGPU::MIMGDimInfo *DimInfo = AMDGPU::getMIMGDimInfoByEncoding(Dim);
 
  return DimInfo->MSAA;
}
 
static bool IsMovrelsSDWAOpcode(const unsigned Opcode)
{
  switch (Opcode) {
  case AMDGPU::V_MOVRELS_B32_sdwa_gfx10:
  case AMDGPU::V_MOVRELSD_B32_sdwa_gfx10:
  case AMDGPU::V_MOVRELSD_2_B32_sdwa_gfx10:
    return true;
  default:
    return false;
  }
}
 
// movrels* opcodes should only allow VGPRS as src0.
// This is specified in .td description for vop1/vop3,
// but sdwa is handled differently. See isSDWAOperand.
bool AMDGPUAsmParser::validateMovrels(const MCInst &Inst,
                                      const OperandVector &Operands) {
 
  const unsigned Opc = Inst.getOpcode();
  const MCInstrDesc &Desc = MII.get(Opc);
 
  if ((Desc.TSFlags & SIInstrFlags::SDWA) == 0 || !IsMovrelsSDWAOpcode(Opc))
    return true;
 
  const int Src0Idx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src0);
  assert(Src0Idx != -1);
 
  SMLoc ErrLoc;
  const MCOperand &Src0 = Inst.getOperand(Src0Idx);
  if (Src0.isReg()) {
    auto Reg = mc2PseudoReg(Src0.getReg());
    const MCRegisterInfo *TRI = getContext().getRegisterInfo();
    if (!isSGPR(Reg, TRI))
      return true;
    ErrLoc = getRegLoc(Reg, Operands);
  } else {
    ErrLoc = getConstLoc(Operands);
  }
 
  Error(ErrLoc, "source operand must be a VGPR");
  return false;
}
 
bool AMDGPUAsmParser::validateMAIAccWrite(const MCInst &Inst,
                                          const OperandVector &Operands) {
 
  const unsigned Opc = Inst.getOpcode();
 
  if (Opc != AMDGPU::V_ACCVGPR_WRITE_B32_vi)
    return true;
 
  const int Src0Idx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src0);
  assert(Src0Idx != -1);
 
  const MCOperand &Src0 = Inst.getOperand(Src0Idx);
  if (!Src0.isReg())
    return true;
 
  auto Reg = mc2PseudoReg(Src0.getReg());
  const MCRegisterInfo *TRI = getContext().getRegisterInfo();
  if (!isGFX90A() && isSGPR(Reg, TRI)) {
    Error(getRegLoc(Reg, Operands),
          "source operand must be either a VGPR or an inline constant");
    return false;
  }
 
  return true;
}
 
bool AMDGPUAsmParser::validateMAISrc2(const MCInst &Inst,
                                      const OperandVector &Operands) {
  unsigned Opcode = Inst.getOpcode();
  const MCInstrDesc &Desc = MII.get(Opcode);
 
  if (!(Desc.TSFlags & SIInstrFlags::IsMAI) ||
      !getFeatureBits()[FeatureMFMAInlineLiteralBug])
    return true;
 
  const int Src2Idx = getNamedOperandIdx(Opcode, OpName::src2);
  if (Src2Idx == -1)
    return true;
 
  if (Inst.getOperand(Src2Idx).isImm() && isInlineConstant(Inst, Src2Idx)) {
    Error(getConstLoc(Operands),
          "inline constants are not allowed for this operand");
    return false;
  }
 
  return true;
}
 
bool AMDGPUAsmParser::validateMFMA(const MCInst &Inst,
                                   const OperandVector &Operands) {
  const unsigned Opc = Inst.getOpcode();
  const MCInstrDesc &Desc = MII.get(Opc);
 
  if ((Desc.TSFlags & SIInstrFlags::IsMAI) == 0)
    return true;
 
  const int Src2Idx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src2);
  if (Src2Idx == -1)
    return true;
 
  const MCOperand &Src2 = Inst.getOperand(Src2Idx);
  if (!Src2.isReg())
    return true;
 
  MCRegister Src2Reg = Src2.getReg();
  MCRegister DstReg = Inst.getOperand(0).getReg();
  if (Src2Reg == DstReg)
    return true;
 
  const MCRegisterInfo *TRI = getContext().getRegisterInfo();
  if (TRI->getRegClass(Desc.operands()[0].RegClass).getSizeInBits() <= 128)
    return true;
 
  if (TRI->regsOverlap(Src2Reg, DstReg)) {
    Error(getRegLoc(mc2PseudoReg(Src2Reg), Operands),
          "source 2 operand must not partially overlap with dst");
    return false;
  }
 
  return true;
}
 
bool AMDGPUAsmParser::validateDivScale(const MCInst &Inst) {
  switch (Inst.getOpcode()) {
  default:
    return true;
  case V_DIV_SCALE_F32_gfx6_gfx7:
  case V_DIV_SCALE_F32_vi:
  case V_DIV_SCALE_F32_gfx10:
  case V_DIV_SCALE_F64_gfx6_gfx7:
  case V_DIV_SCALE_F64_vi:
  case V_DIV_SCALE_F64_gfx10:
    break;
  }
 
  // TODO: Check that src0 = src1 or src2.
 
  for (auto Name : {AMDGPU::OpName::src0_modifiers,
                    AMDGPU::OpName::src2_modifiers,
                    AMDGPU::OpName::src2_modifiers}) {
    if (Inst.getOperand(AMDGPU::getNamedOperandIdx(Inst.getOpcode(), Name))
            .getImm() &
        SISrcMods::ABS) {
      return false;
    }
  }
 
  return true;
}
 
bool AMDGPUAsmParser::validateMIMGD16(const MCInst &Inst) {
 
  const unsigned Opc = Inst.getOpcode();
  const MCInstrDesc &Desc = MII.get(Opc);
 
  if ((Desc.TSFlags & SIInstrFlags::MIMG) == 0)
    return true;
 
  int D16Idx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::d16);
  if (D16Idx >= 0 && Inst.getOperand(D16Idx).getImm()) {
    if (isCI() || isSI())
      return false;
  }
 
  return true;
}
 
static bool IsRevOpcode(const unsigned Opcode)
{
  switch (Opcode) {
  case AMDGPU::V_SUBREV_F32_e32:
  case AMDGPU::V_SUBREV_F32_e64:
  case AMDGPU::V_SUBREV_F32_e32_gfx10:
  case AMDGPU::V_SUBREV_F32_e32_gfx6_gfx7:
  case AMDGPU::V_SUBREV_F32_e32_vi:
  case AMDGPU::V_SUBREV_F32_e64_gfx10:
  case AMDGPU::V_SUBREV_F32_e64_gfx6_gfx7:
  case AMDGPU::V_SUBREV_F32_e64_vi:
 
  case AMDGPU::V_SUBREV_CO_U32_e32:
  case AMDGPU::V_SUBREV_CO_U32_e64:
  case AMDGPU::V_SUBREV_I32_e32_gfx6_gfx7:
  case AMDGPU::V_SUBREV_I32_e64_gfx6_gfx7:
 
  case AMDGPU::V_SUBBREV_U32_e32:
  case AMDGPU::V_SUBBREV_U32_e64:
  case AMDGPU::V_SUBBREV_U32_e32_gfx6_gfx7:
  case AMDGPU::V_SUBBREV_U32_e32_vi:
  case AMDGPU::V_SUBBREV_U32_e64_gfx6_gfx7:
  case AMDGPU::V_SUBBREV_U32_e64_vi:
 
  case AMDGPU::V_SUBREV_U32_e32:
  case AMDGPU::V_SUBREV_U32_e64:
  case AMDGPU::V_SUBREV_U32_e32_gfx9:
  case AMDGPU::V_SUBREV_U32_e32_vi:
  case AMDGPU::V_SUBREV_U32_e64_gfx9:
  case AMDGPU::V_SUBREV_U32_e64_vi:
 
  case AMDGPU::V_SUBREV_F16_e32:
  case AMDGPU::V_SUBREV_F16_e64:
  case AMDGPU::V_SUBREV_F16_e32_gfx10:
  case AMDGPU::V_SUBREV_F16_e32_vi:
  case AMDGPU::V_SUBREV_F16_e64_gfx10:
  case AMDGPU::V_SUBREV_F16_e64_vi:
 
  case AMDGPU::V_SUBREV_U16_e32:
  case AMDGPU::V_SUBREV_U16_e64:
  case AMDGPU::V_SUBREV_U16_e32_vi:
  case AMDGPU::V_SUBREV_U16_e64_vi:
 
  case AMDGPU::V_SUBREV_CO_U32_e32_gfx9:
  case AMDGPU::V_SUBREV_CO_U32_e64_gfx10:
  case AMDGPU::V_SUBREV_CO_U32_e64_gfx9:
 
  case AMDGPU::V_SUBBREV_CO_U32_e32_gfx9:
  case AMDGPU::V_SUBBREV_CO_U32_e64_gfx9:
 
  case AMDGPU::V_SUBREV_NC_U32_e32_gfx10:
  case AMDGPU::V_SUBREV_NC_U32_e64_gfx10:
 
  case AMDGPU::V_SUBREV_CO_CI_U32_e32_gfx10:
  case AMDGPU::V_SUBREV_CO_CI_U32_e64_gfx10:
 
  case AMDGPU::V_LSHRREV_B32_e32:
  case AMDGPU::V_LSHRREV_B32_e64:
  case AMDGPU::V_LSHRREV_B32_e32_gfx6_gfx7:
  case AMDGPU::V_LSHRREV_B32_e64_gfx6_gfx7:
  case AMDGPU::V_LSHRREV_B32_e32_vi:
  case AMDGPU::V_LSHRREV_B32_e64_vi:
  case AMDGPU::V_LSHRREV_B32_e32_gfx10:
  case AMDGPU::V_LSHRREV_B32_e64_gfx10:
 
  case AMDGPU::V_ASHRREV_I32_e32:
  case AMDGPU::V_ASHRREV_I32_e64:
  case AMDGPU::V_ASHRREV_I32_e32_gfx10:
  case AMDGPU::V_ASHRREV_I32_e32_gfx6_gfx7:
  case AMDGPU::V_ASHRREV_I32_e32_vi:
  case AMDGPU::V_ASHRREV_I32_e64_gfx10:
  case AMDGPU::V_ASHRREV_I32_e64_gfx6_gfx7:
  case AMDGPU::V_ASHRREV_I32_e64_vi:
 
  case AMDGPU::V_LSHLREV_B32_e32:
  case AMDGPU::V_LSHLREV_B32_e64:
  case AMDGPU::V_LSHLREV_B32_e32_gfx10:
  case AMDGPU::V_LSHLREV_B32_e32_gfx6_gfx7:
  case AMDGPU::V_LSHLREV_B32_e32_vi:
  case AMDGPU::V_LSHLREV_B32_e64_gfx10:
  case AMDGPU::V_LSHLREV_B32_e64_gfx6_gfx7:
  case AMDGPU::V_LSHLREV_B32_e64_vi:
 
  case AMDGPU::V_LSHLREV_B16_e32:
  case AMDGPU::V_LSHLREV_B16_e64:
  case AMDGPU::V_LSHLREV_B16_e32_vi:
  case AMDGPU::V_LSHLREV_B16_e64_vi:
  case AMDGPU::V_LSHLREV_B16_gfx10:
 
  case AMDGPU::V_LSHRREV_B16_e32:
  case AMDGPU::V_LSHRREV_B16_e64:
  case AMDGPU::V_LSHRREV_B16_e32_vi:
  case AMDGPU::V_LSHRREV_B16_e64_vi:
  case AMDGPU::V_LSHRREV_B16_gfx10:
 
  case AMDGPU::V_ASHRREV_I16_e32:
  case AMDGPU::V_ASHRREV_I16_e64:
  case AMDGPU::V_ASHRREV_I16_e32_vi:
  case AMDGPU::V_ASHRREV_I16_e64_vi:
  case AMDGPU::V_ASHRREV_I16_gfx10:
 
  case AMDGPU::V_LSHLREV_B64_e64:
  case AMDGPU::V_LSHLREV_B64_gfx10:
  case AMDGPU::V_LSHLREV_B64_vi:
 
  case AMDGPU::V_LSHRREV_B64_e64:
  case AMDGPU::V_LSHRREV_B64_gfx10:
  case AMDGPU::V_LSHRREV_B64_vi:
 
  case AMDGPU::V_ASHRREV_I64_e64:
  case AMDGPU::V_ASHRREV_I64_gfx10:
  case AMDGPU::V_ASHRREV_I64_vi:
 
  case AMDGPU::V_PK_LSHLREV_B16:
  case AMDGPU::V_PK_LSHLREV_B16_gfx10:
  case AMDGPU::V_PK_LSHLREV_B16_vi:
 
  case AMDGPU::V_PK_LSHRREV_B16:
  case AMDGPU::V_PK_LSHRREV_B16_gfx10:
  case AMDGPU::V_PK_LSHRREV_B16_vi:
  case AMDGPU::V_PK_ASHRREV_I16:
  case AMDGPU::V_PK_ASHRREV_I16_gfx10:
  case AMDGPU::V_PK_ASHRREV_I16_vi:
    return true;
  default:
    return false;
  }
}
 
std::optional<StringRef>
AMDGPUAsmParser::validateLdsDirect(const MCInst &Inst) {
 
  using namespace SIInstrFlags;
  const unsigned Opcode = Inst.getOpcode();
  const MCInstrDesc &Desc = MII.get(Opcode);
 
  // lds_direct register is defined so that it can be used
  // with 9-bit operands only. Ignore encodings which do not accept these.
  const auto Enc = VOP1 | VOP2 | VOP3 | VOPC | VOP3P | SIInstrFlags::SDWA;
  if ((Desc.TSFlags & Enc) == 0)
    return std::nullopt;
 
  for (auto SrcName : {OpName::src0, OpName::src1, OpName::src2}) {
    auto SrcIdx = getNamedOperandIdx(Opcode, SrcName);
    if (SrcIdx == -1)
      break;
    const auto &Src = Inst.getOperand(SrcIdx);
    if (Src.isReg() && Src.getReg() == LDS_DIRECT) {
 
      if (isGFX90A() || isGFX11Plus())
        return StringRef("lds_direct is not supported on this GPU");
 
      if (IsRevOpcode(Opcode) || (Desc.TSFlags & SIInstrFlags::SDWA))
        return StringRef("lds_direct cannot be used with this instruction");
 
      if (SrcName != OpName::src0)
        return StringRef("lds_direct may be used as src0 only");
    }
  }
 
  return std::nullopt;
}
 
SMLoc AMDGPUAsmParser::getFlatOffsetLoc(const OperandVector &Operands) const {
  for (unsigned i = 1, e = Operands.size(); i != e; ++i) {
    AMDGPUOperand &Op = ((AMDGPUOperand &)*Operands[i]);
    if (Op.isFlatOffset())
      return Op.getStartLoc();
  }
  return getLoc();
}
 
bool AMDGPUAsmParser::validateFlatOffset(const MCInst &Inst,
                                         const OperandVector &Operands) {
  uint64_t TSFlags = MII.get(Inst.getOpcode()).TSFlags;
  if ((TSFlags & SIInstrFlags::FLAT) == 0)
    return true;
 
  auto Opcode = Inst.getOpcode();
  auto OpNum = AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::offset);
  assert(OpNum != -1);
 
  const auto &Op = Inst.getOperand(OpNum);
  if (!hasFlatOffsets() && Op.getImm() != 0) {
    Error(getFlatOffsetLoc(Operands),
          "flat offset modifier is not supported on this GPU");
    return false;
  }
 
  // For FLAT segment the offset must be positive;
  // MSB is ignored and forced to zero.
  unsigned OffsetSize = AMDGPU::getNumFlatOffsetBits(getSTI());
  bool AllowNegative =
      TSFlags & (SIInstrFlags::FlatGlobal | SIInstrFlags::FlatScratch);
  if (!isIntN(OffsetSize, Op.getImm()) || (!AllowNegative && Op.getImm() < 0)) {
    Error(getFlatOffsetLoc(Operands),
          Twine("expected a ") +
              (AllowNegative ? Twine(OffsetSize) + "-bit signed offset"
                             : Twine(OffsetSize - 1) + "-bit unsigned offset"));
    return false;
  }
 
  return true;
}
 
SMLoc AMDGPUAsmParser::getSMEMOffsetLoc(const OperandVector &Operands) const {
  // Start with second operand because SMEM Offset cannot be dst or src0.
  for (unsigned i = 2, e = Operands.size(); i != e; ++i) {
    AMDGPUOperand &Op = ((AMDGPUOperand &)*Operands[i]);
    if (Op.isSMEMOffset() || Op.isSMEMOffsetMod())
      return Op.getStartLoc();
  }
  return getLoc();
}
 
bool AMDGPUAsmParser::validateSMEMOffset(const MCInst &Inst,
                                         const OperandVector &Operands) {
  if (isCI() || isSI())
    return true;
 
  uint64_t TSFlags = MII.get(Inst.getOpcode()).TSFlags;
  if ((TSFlags & SIInstrFlags::SMRD) == 0)
    return true;
 
  auto Opcode = Inst.getOpcode();
  auto OpNum = AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::offset);
  if (OpNum == -1)
    return true;
 
  const auto &Op = Inst.getOperand(OpNum);
  if (!Op.isImm())
    return true;
 
  uint64_t Offset = Op.getImm();
  bool IsBuffer = AMDGPU::getSMEMIsBuffer(Opcode);
  if (AMDGPU::isLegalSMRDEncodedUnsignedOffset(getSTI(), Offset) ||
      AMDGPU::isLegalSMRDEncodedSignedOffset(getSTI(), Offset, IsBuffer))
    return true;
 
  Error(getSMEMOffsetLoc(Operands),
        (isVI() || IsBuffer) ? "expected a 20-bit unsigned offset" :
                               "expected a 21-bit signed offset");
 
  return false;
}
 
bool AMDGPUAsmParser::validateSOPLiteral(const MCInst &Inst) const {
  unsigned Opcode = Inst.getOpcode();
  const MCInstrDesc &Desc = MII.get(Opcode);
  if (!(Desc.TSFlags & (SIInstrFlags::SOP2 | SIInstrFlags::SOPC)))
    return true;
 
  const int Src0Idx = AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::src0);
  const int Src1Idx = AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::src1);
 
  const int OpIndices[] = { Src0Idx, Src1Idx };
 
  unsigned NumExprs = 0;
  unsigned NumLiterals = 0;
  uint32_t LiteralValue;
 
  for (int OpIdx : OpIndices) {
    if (OpIdx == -1) break;
 
    const MCOperand &MO = Inst.getOperand(OpIdx);
    // Exclude special imm operands (like that used by s_set_gpr_idx_on)
    if (AMDGPU::isSISrcOperand(Desc, OpIdx)) {
      if (MO.isImm() && !isInlineConstant(Inst, OpIdx)) {
        uint32_t Value = static_cast<uint32_t>(MO.getImm());
        if (NumLiterals == 0 || LiteralValue != Value) {
          LiteralValue = Value;
          ++NumLiterals;
        }
      } else if (MO.isExpr()) {
        ++NumExprs;
      }
    }
  }
 
  return NumLiterals + NumExprs <= 1;
}
 
bool AMDGPUAsmParser::validateOpSel(const MCInst &Inst) {
  const unsigned Opc = Inst.getOpcode();
  if (isPermlane16(Opc)) {
    int OpSelIdx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::op_sel);
    unsigned OpSel = Inst.getOperand(OpSelIdx).getImm();
 
    if (OpSel & ~3)
      return false;
  }
 
  uint64_t TSFlags = MII.get(Opc).TSFlags;
 
  if (isGFX940() && (TSFlags & SIInstrFlags::IsDOT)) {
    int OpSelIdx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::op_sel);
    if (OpSelIdx != -1) {
      if (Inst.getOperand(OpSelIdx).getImm() != 0)
        return false;
    }
    int OpSelHiIdx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::op_sel_hi);
    if (OpSelHiIdx != -1) {
      if (Inst.getOperand(OpSelHiIdx).getImm() != -1)
        return false;
    }
  }
 
  // op_sel[0:1] must be 0 for v_dot2_bf16_bf16 and v_dot2_f16_f16 (VOP3 Dot).
  if (isGFX11Plus() && (TSFlags & SIInstrFlags::IsDOT) &&
      (TSFlags & SIInstrFlags::VOP3) && !(TSFlags & SIInstrFlags::VOP3P)) {
    int OpSelIdx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::op_sel);
    unsigned OpSel = Inst.getOperand(OpSelIdx).getImm();
    if (OpSel & 3)
      return false;
  }
 
  return true;
}
 
bool AMDGPUAsmParser::validateDPP(const MCInst &Inst,
                                  const OperandVector &Operands) {
  const unsigned Opc = Inst.getOpcode();
  int DppCtrlIdx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::dpp_ctrl);
  if (DppCtrlIdx < 0)
    return true;
  unsigned DppCtrl = Inst.getOperand(DppCtrlIdx).getImm();
 
  if (!AMDGPU::isLegal64BitDPPControl(DppCtrl)) {
    // DPP64 is supported for row_newbcast only.
    int Src0Idx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src0);
    if (Src0Idx >= 0 &&
        getMRI()->getSubReg(Inst.getOperand(Src0Idx).getReg(), AMDGPU::sub1)) {
      SMLoc S = getImmLoc(AMDGPUOperand::ImmTyDppCtrl, Operands);
      Error(S, "64 bit dpp only supports row_newbcast");
      return false;
    }
  }
 
  return true;
}
 
// Check if VCC register matches wavefront size
bool AMDGPUAsmParser::validateVccOperand(unsigned Reg) const {
  auto FB = getFeatureBits();
  return (FB[AMDGPU::FeatureWavefrontSize64] && Reg == AMDGPU::VCC) ||
    (FB[AMDGPU::FeatureWavefrontSize32] && Reg == AMDGPU::VCC_LO);
}
 
// One unique literal can be used. VOP3 literal is only allowed in GFX10+
bool AMDGPUAsmParser::validateVOPLiteral(const MCInst &Inst,
                                         const OperandVector &Operands) {
  unsigned Opcode = Inst.getOpcode();
  const MCInstrDesc &Desc = MII.get(Opcode);
  bool HasMandatoryLiteral = getNamedOperandIdx(Opcode, OpName::imm) != -1;
  if (!(Desc.TSFlags & (SIInstrFlags::VOP3 | SIInstrFlags::VOP3P)) &&
      !HasMandatoryLiteral && !isVOPD(Opcode))
    return true;
 
  OperandIndices OpIndices = getSrcOperandIndices(Opcode, HasMandatoryLiteral);
 
  unsigned NumExprs = 0;
  unsigned NumLiterals = 0;
  uint32_t LiteralValue;
 
  for (int OpIdx : OpIndices) {
    if (OpIdx == -1)
      continue;
 
    const MCOperand &MO = Inst.getOperand(OpIdx);
    if (!MO.isImm() && !MO.isExpr())
      continue;
    if (!isSISrcOperand(Desc, OpIdx))
      continue;
 
    if (MO.isImm() && !isInlineConstant(Inst, OpIdx)) {
      uint32_t Value = static_cast<uint32_t>(MO.getImm());
      if (NumLiterals == 0 || LiteralValue != Value) {
        LiteralValue = Value;
        ++NumLiterals;
      }
    } else if (MO.isExpr()) {
      ++NumExprs;
    }
  }
  NumLiterals += NumExprs;
 
  if (!NumLiterals)
    return true;
 
  if (!HasMandatoryLiteral && !getFeatureBits()[FeatureVOP3Literal]) {
    Error(getLitLoc(Operands), "literal operands are not supported");
    return false;
  }
 
  if (NumLiterals > 1) {
    Error(getLitLoc(Operands, true), "only one unique literal operand is allowed");
    return false;
  }
 
  return true;
}
 
// Returns -1 if not a register, 0 if VGPR and 1 if AGPR.
static int IsAGPROperand(const MCInst &Inst, uint16_t NameIdx,
                         const MCRegisterInfo *MRI) {
  int OpIdx = AMDGPU::getNamedOperandIdx(Inst.getOpcode(), NameIdx);
  if (OpIdx < 0)
    return -1;
 
  const MCOperand &Op = Inst.getOperand(OpIdx);
  if (!Op.isReg())
    return -1;
 
  unsigned Sub = MRI->getSubReg(Op.getReg(), AMDGPU::sub0);
  auto Reg = Sub ? Sub : Op.getReg();
  const MCRegisterClass &AGPR32 = MRI->getRegClass(AMDGPU::AGPR_32RegClassID);
  return AGPR32.contains(Reg) ? 1 : 0;
}
 
bool AMDGPUAsmParser::validateAGPRLdSt(const MCInst &Inst) const {
  uint64_t TSFlags = MII.get(Inst.getOpcode()).TSFlags;
  if ((TSFlags & (SIInstrFlags::FLAT | SIInstrFlags::MUBUF |
                  SIInstrFlags::MTBUF | SIInstrFlags::MIMG |
                  SIInstrFlags::DS)) == 0)
    return true;
 
  uint16_t DataNameIdx = (TSFlags & SIInstrFlags::DS) ? AMDGPU::OpName::data0
                                                      : AMDGPU::OpName::vdata;
 
  const MCRegisterInfo *MRI = getMRI();
  int DstAreg = IsAGPROperand(Inst, AMDGPU::OpName::vdst, MRI);
  int DataAreg = IsAGPROperand(Inst, DataNameIdx, MRI);
 
  if ((TSFlags & SIInstrFlags::DS) && DataAreg >= 0) {
    int Data2Areg = IsAGPROperand(Inst, AMDGPU::OpName::data1, MRI);
    if (Data2Areg >= 0 && Data2Areg != DataAreg)
      return false;
  }
 
  auto FB = getFeatureBits();
  if (FB[AMDGPU::FeatureGFX90AInsts]) {
    if (DataAreg < 0 || DstAreg < 0)
      return true;
    return DstAreg == DataAreg;
  }
 
  return DstAreg < 1 && DataAreg < 1;
}
 
bool AMDGPUAsmParser::validateVGPRAlign(const MCInst &Inst) const {
  auto FB = getFeatureBits();
  if (!FB[AMDGPU::FeatureGFX90AInsts])
    return true;
 
  const MCRegisterInfo *MRI = getMRI();
  const MCRegisterClass &VGPR32 = MRI->getRegClass(AMDGPU::VGPR_32RegClassID);
  const MCRegisterClass &AGPR32 = MRI->getRegClass(AMDGPU::AGPR_32RegClassID);
  for (unsigned I = 0, E = Inst.getNumOperands(); I != E; ++I) {
    const MCOperand &Op = Inst.getOperand(I);
    if (!Op.isReg())
      continue;
 
    unsigned Sub = MRI->getSubReg(Op.getReg(), AMDGPU::sub0);
    if (!Sub)
      continue;
 
    if (VGPR32.contains(Sub) && ((Sub - AMDGPU::VGPR0) & 1))
      return false;
    if (AGPR32.contains(Sub) && ((Sub - AMDGPU::AGPR0) & 1))
      return false;
  }
 
  return true;
}
 
SMLoc AMDGPUAsmParser::getBLGPLoc(const OperandVector &Operands) const {
  for (unsigned i = 1, e = Operands.size(); i != e; ++i) {
    AMDGPUOperand &Op = ((AMDGPUOperand &)*Operands[i]);
    if (Op.isBLGP())
      return Op.getStartLoc();
  }
  return SMLoc();
}
 
bool AMDGPUAsmParser::validateBLGP(const MCInst &Inst,
                                   const OperandVector &Operands) {
  unsigned Opc = Inst.getOpcode();
  int BlgpIdx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::blgp);
  if (BlgpIdx == -1)
    return true;
  SMLoc BLGPLoc = getBLGPLoc(Operands);
  if (!BLGPLoc.isValid())
    return true;
  bool IsNeg = StringRef(BLGPLoc.getPointer()).startswith("neg:");
  auto FB = getFeatureBits();
  bool UsesNeg = false;
  if (FB[AMDGPU::FeatureGFX940Insts]) {
    switch (Opc) {
    case AMDGPU::V_MFMA_F64_16X16X4F64_gfx940_acd:
    case AMDGPU::V_MFMA_F64_16X16X4F64_gfx940_vcd:
    case AMDGPU::V_MFMA_F64_4X4X4F64_gfx940_acd:
    case AMDGPU::V_MFMA_F64_4X4X4F64_gfx940_vcd:
      UsesNeg = true;
    }
  }
 
  if (IsNeg == UsesNeg)
    return true;
 
  Error(BLGPLoc,
        UsesNeg ? "invalid modifier: blgp is not supported"
                : "invalid modifier: neg is not supported");
 
  return false;
}
 
bool AMDGPUAsmParser::validateWaitCnt(const MCInst &Inst,
                                      const OperandVector &Operands) {
  if (!isGFX11Plus())
    return true;
 
  unsigned Opc = Inst.getOpcode();
  if (Opc != AMDGPU::S_WAITCNT_EXPCNT_gfx11 &&
      Opc != AMDGPU::S_WAITCNT_LGKMCNT_gfx11 &&
      Opc != AMDGPU::S_WAITCNT_VMCNT_gfx11 &&
      Opc != AMDGPU::S_WAITCNT_VSCNT_gfx11)
    return true;
 
  int Src0Idx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::sdst);
  assert(Src0Idx >= 0 && Inst.getOperand(Src0Idx).isReg());
  auto Reg = mc2PseudoReg(Inst.getOperand(Src0Idx).getReg());
  if (Reg == AMDGPU::SGPR_NULL)
    return true;
 
  SMLoc RegLoc = getRegLoc(Reg, Operands);
  Error(RegLoc, "src0 must be null");
  return false;
}
 
// gfx90a has an undocumented limitation:
// DS_GWS opcodes must use even aligned registers.
bool AMDGPUAsmParser::validateGWS(const MCInst &Inst,
                                  const OperandVector &Operands) {
  if (!getFeatureBits()[AMDGPU::FeatureGFX90AInsts])
    return true;
 
  int Opc = Inst.getOpcode();
  if (Opc != AMDGPU::DS_GWS_INIT_vi && Opc != AMDGPU::DS_GWS_BARRIER_vi &&
      Opc != AMDGPU::DS_GWS_SEMA_BR_vi)
    return true;
 
  const MCRegisterInfo *MRI = getMRI();
  const MCRegisterClass &VGPR32 = MRI->getRegClass(AMDGPU::VGPR_32RegClassID);
  int Data0Pos =
      AMDGPU::getNamedOperandIdx(Inst.getOpcode(), AMDGPU::OpName::data0);
  assert(Data0Pos != -1);
  auto Reg = Inst.getOperand(Data0Pos).getReg();
  auto RegIdx = Reg - (VGPR32.contains(Reg) ? AMDGPU::VGPR0 : AMDGPU::AGPR0);
  if (RegIdx & 1) {
    SMLoc RegLoc = getRegLoc(Reg, Operands);
    Error(RegLoc, "vgpr must be even aligned");
    return false;
  }
 
  return true;
}
 
bool AMDGPUAsmParser::validateCoherencyBits(const MCInst &Inst,
                                            const OperandVector &Operands,
                                            const SMLoc &IDLoc) {
  int CPolPos = AMDGPU::getNamedOperandIdx(Inst.getOpcode(),
                                           AMDGPU::OpName::cpol);
  if (CPolPos == -1)
    return true;
 
  unsigned CPol = Inst.getOperand(CPolPos).getImm();
 
  uint64_t TSFlags = MII.get(Inst.getOpcode()).TSFlags;
  if (TSFlags & SIInstrFlags::SMRD) {
    if (CPol && (isSI() || isCI())) {
      SMLoc S = getImmLoc(AMDGPUOperand::ImmTyCPol, Operands);
      Error(S, "cache policy is not supported for SMRD instructions");
      return false;
    }
    if (CPol & ~(AMDGPU::CPol::GLC | AMDGPU::CPol::DLC)) {
      Error(IDLoc, "invalid cache policy for SMEM instruction");
      return false;
    }
  }
 
  if (isGFX90A() && !isGFX940() && (CPol & CPol::SCC)) {
    SMLoc S = getImmLoc(AMDGPUOperand::ImmTyCPol, Operands);
    StringRef CStr(S.getPointer());
    S = SMLoc::getFromPointer(&CStr.data()[CStr.find("scc")]);
    Error(S, "scc is not supported on this GPU");
    return false;
  }
 
  if (!(TSFlags & (SIInstrFlags::IsAtomicNoRet | SIInstrFlags::IsAtomicRet)))
    return true;
 
  if (TSFlags & SIInstrFlags::IsAtomicRet) {
    if (!(TSFlags & SIInstrFlags::MIMG) && !(CPol & CPol::GLC)) {
      Error(IDLoc, isGFX940() ? "instruction must use sc0"
                              : "instruction must use glc");
      return false;
    }
  } else {
    if (CPol & CPol::GLC) {
      SMLoc S = getImmLoc(AMDGPUOperand::ImmTyCPol, Operands);
      StringRef CStr(S.getPointer());
      S = SMLoc::getFromPointer(
          &CStr.data()[CStr.find(isGFX940() ? "sc0" : "glc")]);
      Error(S, isGFX940() ? "instruction must not use sc0"
                          : "instruction must not use glc");
      return false;
    }
  }
 
  return true;
}
 
bool AMDGPUAsmParser::validateExeczVcczOperands(const OperandVector &Operands) {
  if (!isGFX11Plus())
    return true;
  for (auto &Operand : Operands) {
    if (!Operand->isReg())
      continue;
    unsigned Reg = Operand->getReg();
    if (Reg == SRC_EXECZ || Reg == SRC_VCCZ) {
      Error(getRegLoc(Reg, Operands),
            "execz and vccz are not supported on this GPU");
      return false;
    }
  }
  return true;
}
 
bool AMDGPUAsmParser::validateTFE(const MCInst &Inst,
                                  const OperandVector &Operands) {
  const MCInstrDesc &Desc = MII.get(Inst.getOpcode());
  if (Desc.mayStore() &&
      (Desc.TSFlags & (SIInstrFlags::MUBUF | SIInstrFlags::MTBUF))) {
    SMLoc Loc = getImmLoc(AMDGPUOperand::ImmTyTFE, Operands);
    if (Loc != getInstLoc(Operands)) {
      Error(Loc, "TFE modifier has no meaning for store instructions");
      return false;
    }
  }
 
  return true;
}
 
bool AMDGPUAsmParser::validateInstruction(const MCInst &Inst,
                                          const SMLoc &IDLoc,
                                          const OperandVector &Operands) {
  if (auto ErrMsg = validateLdsDirect(Inst)) {
    Error(getRegLoc(LDS_DIRECT, Operands), *ErrMsg);
    return false;
  }
  if (!validateSOPLiteral(Inst)) {
    Error(getLitLoc(Operands),
      "only one unique literal operand is allowed");
    return false;
  }
  if (!validateVOPLiteral(Inst, Operands)) {
    return false;
  }
  if (!validateConstantBusLimitations(Inst, Operands)) {
    return false;
  }
  if (!validateVOPDRegBankConstraints(Inst, Operands)) {
    return false;
  }
  if (!validateIntClampSupported(Inst)) {
    Error(getImmLoc(AMDGPUOperand::ImmTyClampSI, Operands),
      "integer clamping is not supported on this GPU");
    return false;
  }
  if (!validateOpSel(Inst)) {
    Error(getImmLoc(AMDGPUOperand::ImmTyOpSel, Operands),
      "invalid op_sel operand");
    return false;
  }
  if (!validateDPP(Inst, Operands)) {
    return false;
  }
  // For MUBUF/MTBUF d16 is a part of opcode, so there is nothing to validate.
  if (!validateMIMGD16(Inst)) {
    Error(getImmLoc(AMDGPUOperand::ImmTyD16, Operands),
      "d16 modifier is not supported on this GPU");
    return false;
  }
  if (!validateMIMGMSAA(Inst)) {
    Error(getImmLoc(AMDGPUOperand::ImmTyDim, Operands),
          "invalid dim; must be MSAA type");
    return false;
  }
  if (!validateMIMGDataSize(Inst, IDLoc)) {
    return false;
  }
  if (!validateMIMGAddrSize(Inst)) {
    Error(IDLoc,
      "image address size does not match dim and a16");
    return false;
  }
  if (!validateMIMGAtomicDMask(Inst)) {
    Error(getImmLoc(AMDGPUOperand::ImmTyDMask, Operands),
      "invalid atomic image dmask");
    return false;
  }
  if (!validateMIMGGatherDMask(Inst)) {
    Error(getImmLoc(AMDGPUOperand::ImmTyDMask, Operands),
      "invalid image_gather dmask: only one bit must be set");
    return false;
  }
  if (!validateMovrels(Inst, Operands)) {
    return false;
  }
  if (!validateFlatOffset(Inst, Operands)) {
    return false;
  }
  if (!validateSMEMOffset(Inst, Operands)) {
    return false;
  }
  if (!validateMAIAccWrite(Inst, Operands)) {
    return false;
  }
  if (!validateMAISrc2(Inst, Operands)) {
    return false;
  }
  if (!validateMFMA(Inst, Operands)) {
    return false;
  }
  if (!validateCoherencyBits(Inst, Operands, IDLoc)) {
    return false;
  }
 
  if (!validateAGPRLdSt(Inst)) {
    Error(IDLoc, getFeatureBits()[AMDGPU::FeatureGFX90AInsts]
    ? "invalid register class: data and dst should be all VGPR or AGPR"
    : "invalid register class: agpr loads and stores not supported on this GPU"
    );
    return false;
  }
  if (!validateVGPRAlign(Inst)) {
    Error(IDLoc,
      "invalid register class: vgpr tuples must be 64 bit aligned");
    return false;
  }
  if (!validateGWS(Inst, Operands)) {
    return false;
  }
 
  if (!validateBLGP(Inst, Operands)) {
    return false;
  }
 
  if (!validateDivScale(Inst)) {
    Error(IDLoc, "ABS not allowed in VOP3B instructions");
    return false;
  }
  if (!validateWaitCnt(Inst, Operands)) {
    return false;
  }
  if (!validateExeczVcczOperands(Operands)) {
    return false;
  }
  if (!validateTFE(Inst, Operands)) {
    return false;
  }
 
  return true;
}
 
static std::string AMDGPUMnemonicSpellCheck(StringRef S,
                                            const FeatureBitset &FBS,
                                            unsigned VariantID = 0);
 
static bool AMDGPUCheckMnemonic(StringRef Mnemonic,
                                const FeatureBitset &AvailableFeatures,
                                unsigned VariantID);
 
bool AMDGPUAsmParser::isSupportedMnemo(StringRef Mnemo,
                                       const FeatureBitset &FBS) {
  return isSupportedMnemo(Mnemo, FBS, getAllVariants());
}
 
bool AMDGPUAsmParser::isSupportedMnemo(StringRef Mnemo,
                                       const FeatureBitset &FBS,
                                       ArrayRef<unsigned> Variants) {
  for (auto Variant : Variants) {
    if (AMDGPUCheckMnemonic(Mnemo, FBS, Variant))
      return true;
  }
 
  return false;
}
 
bool AMDGPUAsmParser::checkUnsupportedInstruction(StringRef Mnemo,
                                                  const SMLoc &IDLoc) {
  FeatureBitset FBS = ComputeAvailableFeatures(getFeatureBits());
 
  // Check if requested instruction variant is supported.
  if (isSupportedMnemo(Mnemo, FBS, getMatchedVariants()))
    return false;
 
  // This instruction is not supported.
  // Clear any other pending errors because they are no longer relevant.
  getParser().clearPendingErrors();
 
  // Requested instruction variant is not supported.
  // Check if any other variants are supported.
  StringRef VariantName = getMatchedVariantName();
  if (!VariantName.empty() && isSupportedMnemo(Mnemo, FBS)) {
    return Error(IDLoc,
                 Twine(VariantName,
                       " variant of this instruction is not supported"));
  }
 
  // Check if this instruction may be used with a different wavesize.
  if (isGFX10Plus() && getFeatureBits()[AMDGPU::FeatureWavefrontSize64] &&
      !getFeatureBits()[AMDGPU::FeatureWavefrontSize32]) {
 
    FeatureBitset FeaturesWS32 = getFeatureBits();
    FeaturesWS32.flip(AMDGPU::FeatureWavefrontSize64)
        .flip(AMDGPU::FeatureWavefrontSize32);
    FeatureBitset AvailableFeaturesWS32 =
        ComputeAvailableFeatures(FeaturesWS32);
 
    if (isSupportedMnemo(Mnemo, AvailableFeaturesWS32, getMatchedVariants()))
      return Error(IDLoc, "instruction requires wavesize=32");
  }
 
  // Finally check if this instruction is supported on any other GPU.
  if (isSupportedMnemo(Mnemo, FeatureBitset().set())) {
    return Error(IDLoc, "instruction not supported on this GPU");
  }
 
  // Instruction not supported on any GPU. Probably a typo.
  std::string Suggestion = AMDGPUMnemonicSpellCheck(Mnemo, FBS);
  return Error(IDLoc, "invalid instruction" + Suggestion);
}
 
static bool isInvalidVOPDY(const OperandVector &Operands,
                           uint64_t InvalidOprIdx) {
  assert(InvalidOprIdx < Operands.size());
  const auto &Op = ((AMDGPUOperand &)*Operands[InvalidOprIdx]);
  if (Op.isToken() && InvalidOprIdx > 1) {
    const auto &PrevOp = ((AMDGPUOperand &)*Operands[InvalidOprIdx - 1]);
    return PrevOp.isToken() && PrevOp.getToken() == "::";
  }
  return false;
}
 
bool AMDGPUAsmParser::MatchAndEmitInstruction(SMLoc IDLoc, unsigned &Opcode,
                                              OperandVector &Operands,
                                              MCStreamer &Out,
                                              uint64_t &ErrorInfo,
                                              bool MatchingInlineAsm) {
  MCInst Inst;
  unsigned Result = Match_Success;
  for (auto Variant : getMatchedVariants()) {
    uint64_t EI;
    auto R = MatchInstructionImpl(Operands, Inst, EI, MatchingInlineAsm,
                                  Variant);
    // We order match statuses from least to most specific. We use most specific
    // status as resulting
    // Match_MnemonicFail < Match_InvalidOperand < Match_MissingFeature < Match_PreferE32
    if ((R == Match_Success) ||
        (R == Match_PreferE32) ||
        (R == Match_MissingFeature && Result != Match_PreferE32) ||
        (R == Match_InvalidOperand && Result != Match_MissingFeature
                                   && Result != Match_PreferE32) ||
        (R == Match_MnemonicFail   && Result != Match_InvalidOperand
                                   && Result != Match_MissingFeature
                                   && Result != Match_PreferE32)) {
      Result = R;
      ErrorInfo = EI;
    }
    if (R == Match_Success)
      break;
  }
 
  if (Result == Match_Success) {
    if (!validateInstruction(Inst, IDLoc, Operands)) {
      return true;
    }
    Inst.setLoc(IDLoc);
    Out.emitInstruction(Inst, getSTI());
    return false;
  }
 
  StringRef Mnemo = ((AMDGPUOperand &)*Operands[0]).getToken();
  if (checkUnsupportedInstruction(Mnemo, IDLoc)) {
    return true;
  }
 
  switch (Result) {
  default: break;
  case Match_MissingFeature:
    // It has been verified that the specified instruction
    // mnemonic is valid. A match was found but it requires
    // features which are not supported on this GPU.
    return Error(IDLoc, "operands are not valid for this GPU or mode");
 
  case Match_InvalidOperand: {
    SMLoc ErrorLoc = IDLoc;
    if (ErrorInfo != ~0ULL) {
      if (ErrorInfo >= Operands.size()) {
        return Error(IDLoc, "too few operands for instruction");
      }
      ErrorLoc = ((AMDGPUOperand &)*Operands[ErrorInfo]).getStartLoc();
      if (ErrorLoc == SMLoc())
        ErrorLoc = IDLoc;
 
      if (isInvalidVOPDY(Operands, ErrorInfo))
        return Error(ErrorLoc, "invalid VOPDY instruction");
    }
    return Error(ErrorLoc, "invalid operand for instruction");
  }
 
  case Match_PreferE32:
    return Error(IDLoc, "internal error: instruction without _e64 suffix "
                        "should be encoded as e32");
  case Match_MnemonicFail:
    llvm_unreachable("Invalid instructions should have been handled already");
  }
  llvm_unreachable("Implement any new match types added!");
}
 
bool AMDGPUAsmParser::ParseAsAbsoluteExpression(uint32_t &Ret) {
  int64_t Tmp = -1;
  if (!isToken(AsmToken::Integer) && !isToken(AsmToken::Identifier)) {
    return true;
  }
  if (getParser().parseAbsoluteExpression(Tmp)) {
    return true;
  }
  Ret = static_cast<uint32_t>(Tmp);
  return false;
}
 
bool AMDGPUAsmParser::ParseDirectiveMajorMinor(uint32_t &Major,
                                               uint32_t &Minor) {
  if (ParseAsAbsoluteExpression(Major))
    return TokError("invalid major version");
 
  if (!trySkipToken(AsmToken::Comma))
    return TokError("minor version number required, comma expected");
 
  if (ParseAsAbsoluteExpression(Minor))
    return TokError("invalid minor version");
 
  return false;
}
 
bool AMDGPUAsmParser::ParseDirectiveAMDGCNTarget() {
  if (getSTI().getTargetTriple().getArch() != Triple::amdgcn)
    return TokError("directive only supported for amdgcn architecture");
 
  std::string TargetIDDirective;
  SMLoc TargetStart = getTok().getLoc();
  if (getParser().parseEscapedString(TargetIDDirective))
    return true;
 
  SMRange TargetRange = SMRange(TargetStart, getTok().getLoc());
  if (getTargetStreamer().getTargetID()->toString() != TargetIDDirective)
    return getParser().Error(TargetRange.Start,
        (Twine(".amdgcn_target directive's target id ") +
         Twine(TargetIDDirective) +
         Twine(" does not match the specified target id ") +
         Twine(getTargetStreamer().getTargetID()->toString())).str());
 
  return false;
}
 
bool AMDGPUAsmParser::OutOfRangeError(SMRange Range) {
  return Error(Range.Start, "value out of range", Range);
}
 
bool AMDGPUAsmParser::calculateGPRBlocks(
    const FeatureBitset &Features, bool VCCUsed, bool FlatScrUsed,
    bool XNACKUsed, std::optional<bool> EnableWavefrontSize32,
    unsigned NextFreeVGPR, SMRange VGPRRange, unsigned NextFreeSGPR,
    SMRange SGPRRange, unsigned &VGPRBlocks, unsigned &SGPRBlocks) {
  // TODO(scott.linder): These calculations are duplicated from
  // AMDGPUAsmPrinter::getSIProgramInfo and could be unified.
  IsaVersion Version = getIsaVersion(getSTI().getCPU());
 
  unsigned NumVGPRs = NextFreeVGPR;
  unsigned NumSGPRs = NextFreeSGPR;
 
  if (Version.Major >= 10)
    NumSGPRs = 0;
  else {
    unsigned MaxAddressableNumSGPRs =
        IsaInfo::getAddressableNumSGPRs(&getSTI());
 
    if (Version.Major >= 8 && !Features.test(FeatureSGPRInitBug) &&
        NumSGPRs > MaxAddressableNumSGPRs)
      return OutOfRangeError(SGPRRange);
 
    NumSGPRs +=
        IsaInfo::getNumExtraSGPRs(&getSTI(), VCCUsed, FlatScrUsed, XNACKUsed);
 
    if ((Version.Major <= 7 || Features.test(FeatureSGPRInitBug)) &&
        NumSGPRs > MaxAddressableNumSGPRs)
      return OutOfRangeError(SGPRRange);
 
    if (Features.test(FeatureSGPRInitBug))
      NumSGPRs = IsaInfo::FIXED_NUM_SGPRS_FOR_INIT_BUG;
  }
 
  VGPRBlocks =
      IsaInfo::getNumVGPRBlocks(&getSTI(), NumVGPRs, EnableWavefrontSize32);
  SGPRBlocks = IsaInfo::getNumSGPRBlocks(&getSTI(), NumSGPRs);
 
  return false;
}
 
bool AMDGPUAsmParser::ParseDirectiveAMDHSAKernel() {
  if (getSTI().getTargetTriple().getArch() != Triple::amdgcn)
    return TokError("directive only supported for amdgcn architecture");
 
  if (getSTI().getTargetTriple().getOS() != Triple::AMDHSA)
    return TokError("directive only supported for amdhsa OS");
 
  StringRef KernelName;
  if (getParser().parseIdentifier(KernelName))
    return true;
 
  kernel_descriptor_t KD = getDefaultAmdhsaKernelDescriptor(&getSTI());
 
  StringSet<> Seen;
 
  IsaVersion IVersion = getIsaVersion(getSTI().getCPU());
 
  SMRange VGPRRange;
  uint64_t NextFreeVGPR = 0;
  uint64_t AccumOffset = 0;
  uint64_t SharedVGPRCount = 0;
  SMRange SGPRRange;
  uint64_t NextFreeSGPR = 0;
 
  // Count the number of user SGPRs implied from the enabled feature bits.
  unsigned ImpliedUserSGPRCount = 0;
 
  // Track if the asm explicitly contains the directive for the user SGPR
  // count.
  std::optional<unsigned> ExplicitUserSGPRCount;
  bool ReserveVCC = true;
  bool ReserveFlatScr = true;
  std::optional<bool> EnableWavefrontSize32;
 
  while (true) {
    while (trySkipToken(AsmToken::EndOfStatement));
 
    StringRef ID;
    SMRange IDRange = getTok().getLocRange();
    if (!parseId(ID, "expected .amdhsa_ directive or .end_amdhsa_kernel"))
      return true;
 
    if (ID == ".end_amdhsa_kernel")
      break;
 
    if (!Seen.insert(ID).second)
      return TokError(".amdhsa_ directives cannot be repeated");
 
    SMLoc ValStart = getLoc();
    int64_t IVal;
    if (getParser().parseAbsoluteExpression(IVal))
      return true;
    SMLoc ValEnd = getLoc();
    SMRange ValRange = SMRange(ValStart, ValEnd);
 
    if (IVal < 0)
      return OutOfRangeError(ValRange);
 
    uint64_t Val = IVal;
 
#define PARSE_BITS_ENTRY(FIELD, ENTRY, VALUE, RANGE)                           \
  if (!isUInt<ENTRY##_WIDTH>(VALUE))                                           \
    return OutOfRangeError(RANGE);                                             \
  AMDHSA_BITS_SET(FIELD, ENTRY, VALUE);
 
    if (ID == ".amdhsa_group_segment_fixed_size") {
      if (!isUInt<sizeof(KD.group_segment_fixed_size) * CHAR_BIT>(Val))
        return OutOfRangeError(ValRange);
      KD.group_segment_fixed_size = Val;
    } else if (ID == ".amdhsa_private_segment_fixed_size") {
      if (!isUInt<sizeof(KD.private_segment_fixed_size) * CHAR_BIT>(Val))
        return OutOfRangeError(ValRange);
      KD.private_segment_fixed_size = Val;
    } else if (ID == ".amdhsa_kernarg_size") {
      if (!isUInt<sizeof(KD.kernarg_size) * CHAR_BIT>(Val))
        return OutOfRangeError(ValRange);
      KD.kernarg_size = Val;
    } else if (ID == ".amdhsa_user_sgpr_count") {
      ExplicitUserSGPRCount = Val;
    } else if (ID == ".amdhsa_user_sgpr_private_segment_buffer") {
      if (hasArchitectedFlatScratch())
        return Error(IDRange.Start,
                     "directive is not supported with architected flat scratch",
                     IDRange);
      PARSE_BITS_ENTRY(KD.kernel_code_properties,
                       KERNEL_CODE_PROPERTY_ENABLE_SGPR_PRIVATE_SEGMENT_BUFFER,
                       Val, ValRange);
      if (Val)
        ImpliedUserSGPRCount += 4;
    } else if (ID == ".amdhsa_user_sgpr_dispatch_ptr") {
      PARSE_BITS_ENTRY(KD.kernel_code_properties,
                       KERNEL_CODE_PROPERTY_ENABLE_SGPR_DISPATCH_PTR, Val,
                       ValRange);
      if (Val)
        ImpliedUserSGPRCount += 2;
    } else if (ID == ".amdhsa_user_sgpr_queue_ptr") {
      PARSE_BITS_ENTRY(KD.kernel_code_properties,
                       KERNEL_CODE_PROPERTY_ENABLE_SGPR_QUEUE_PTR, Val,
                       ValRange);
      if (Val)
        ImpliedUserSGPRCount += 2;
    } else if (ID == ".amdhsa_user_sgpr_kernarg_segment_ptr") {
      PARSE_BITS_ENTRY(KD.kernel_code_properties,
                       KERNEL_CODE_PROPERTY_ENABLE_SGPR_KERNARG_SEGMENT_PTR,
                       Val, ValRange);
      if (Val)
        ImpliedUserSGPRCount += 2;
    } else if (ID == ".amdhsa_user_sgpr_dispatch_id") {
      PARSE_BITS_ENTRY(KD.kernel_code_properties,
                       KERNEL_CODE_PROPERTY_ENABLE_SGPR_DISPATCH_ID, Val,
                       ValRange);
      if (Val)
        ImpliedUserSGPRCount += 2;
    } else if (ID == ".amdhsa_user_sgpr_flat_scratch_init") {
      if (hasArchitectedFlatScratch())
        return Error(IDRange.Start,
                     "directive is not supported with architected flat scratch",
                     IDRange);
      PARSE_BITS_ENTRY(KD.kernel_code_properties,
                       KERNEL_CODE_PROPERTY_ENABLE_SGPR_FLAT_SCRATCH_INIT, Val,
                       ValRange);
      if (Val)
        ImpliedUserSGPRCount += 2;
    } else if (ID == ".amdhsa_user_sgpr_private_segment_size") {
      PARSE_BITS_ENTRY(KD.kernel_code_properties,
                       KERNEL_CODE_PROPERTY_ENABLE_SGPR_PRIVATE_SEGMENT_SIZE,
                       Val, ValRange);
      if (Val)
        ImpliedUserSGPRCount += 1;
    } else if (ID == ".amdhsa_wavefront_size32") {
      if (IVersion.Major < 10)
        return Error(IDRange.Start, "directive requires gfx10+", IDRange);
      EnableWavefrontSize32 = Val;
      PARSE_BITS_ENTRY(KD.kernel_code_properties,
                       KERNEL_CODE_PROPERTY_ENABLE_WAVEFRONT_SIZE32,
                       Val, ValRange);
    } else if (ID == ".amdhsa_uses_dynamic_stack") {
      PARSE_BITS_ENTRY(KD.kernel_code_properties,
                       KERNEL_CODE_PROPERTY_USES_DYNAMIC_STACK, Val, ValRange);
    } else if (ID == ".amdhsa_system_sgpr_private_segment_wavefront_offset") {
      if (hasArchitectedFlatScratch())
        return Error(IDRange.Start,
                     "directive is not supported with architected flat scratch",
                     IDRange);
      PARSE_BITS_ENTRY(KD.compute_pgm_rsrc2,
                       COMPUTE_PGM_RSRC2_ENABLE_PRIVATE_SEGMENT, Val, ValRange);
    } else if (ID == ".amdhsa_enable_private_segment") {
      if (!hasArchitectedFlatScratch())
        return Error(
            IDRange.Start,
            "directive is not supported without architected flat scratch",
            IDRange);
      PARSE_BITS_ENTRY(KD.compute_pgm_rsrc2,
                       COMPUTE_PGM_RSRC2_ENABLE_PRIVATE_SEGMENT, Val, ValRange);
    } else if (ID == ".amdhsa_system_sgpr_workgroup_id_x") {
      PARSE_BITS_ENTRY(KD.compute_pgm_rsrc2,
                       COMPUTE_PGM_RSRC2_ENABLE_SGPR_WORKGROUP_ID_X, Val,
                       ValRange);
    } else if (ID == ".amdhsa_system_sgpr_workgroup_id_y") {
      PARSE_BITS_ENTRY(KD.compute_pgm_rsrc2,
                       COMPUTE_PGM_RSRC2_ENABLE_SGPR_WORKGROUP_ID_Y, Val,
                       ValRange);
    } else if (ID == ".amdhsa_system_sgpr_workgroup_id_z") {
      PARSE_BITS_ENTRY(KD.compute_pgm_rsrc2,
                       COMPUTE_PGM_RSRC2_ENABLE_SGPR_WORKGROUP_ID_Z, Val,
                       ValRange);
    } else if (ID == ".amdhsa_system_sgpr_workgroup_info") {
      PARSE_BITS_ENTRY(KD.compute_pgm_rsrc2,
                       COMPUTE_PGM_RSRC2_ENABLE_SGPR_WORKGROUP_INFO, Val,
                       ValRange);
    } else if (ID == ".amdhsa_system_vgpr_workitem_id") {
      PARSE_BITS_ENTRY(KD.compute_pgm_rsrc2,
                       COMPUTE_PGM_RSRC2_ENABLE_VGPR_WORKITEM_ID, Val,
                       ValRange);
    } else if (ID == ".amdhsa_next_free_vgpr") {
      VGPRRange = ValRange;
      NextFreeVGPR = Val;
    } else if (ID == ".amdhsa_next_free_sgpr") {
      SGPRRange = ValRange;
      NextFreeSGPR = Val;
    } else if (ID == ".amdhsa_accum_offset") {
      if (!isGFX90A())
        return Error(IDRange.Start, "directive requires gfx90a+", IDRange);
      AccumOffset = Val;
    } else if (ID == ".amdhsa_reserve_vcc") {
      if (!isUInt<1>(Val))
        return OutOfRangeError(ValRange);
      ReserveVCC = Val;
    } else if (ID == ".amdhsa_reserve_flat_scratch") {
      if (IVersion.Major < 7)
        return Error(IDRange.Start, "directive requires gfx7+", IDRange);
      if (hasArchitectedFlatScratch())
        return Error(IDRange.Start,
                     "directive is not supported with architected flat scratch",
                     IDRange);
      if (!isUInt<1>(Val))
        return OutOfRangeError(ValRange);
      ReserveFlatScr = Val;
    } else if (ID == ".amdhsa_reserve_xnack_mask") {
      if (IVersion.Major < 8)
        return Error(IDRange.Start, "directive requires gfx8+", IDRange);
      if (!isUInt<1>(Val))
        return OutOfRangeError(ValRange);
      if (Val != getTargetStreamer().getTargetID()->isXnackOnOrAny())
        return getParser().Error(IDRange.Start, ".amdhsa_reserve_xnack_mask does not match target id",
                                 IDRange);
    } else if (ID == ".amdhsa_float_round_mode_32") {
      PARSE_BITS_ENTRY(KD.compute_pgm_rsrc1,
                       COMPUTE_PGM_RSRC1_FLOAT_ROUND_MODE_32, Val, ValRange);
    } else if (ID == ".amdhsa_float_round_mode_16_64") {
      PARSE_BITS_ENTRY(KD.compute_pgm_rsrc1,
                       COMPUTE_PGM_RSRC1_FLOAT_ROUND_MODE_16_64, Val, ValRange);
    } else if (ID == ".amdhsa_float_denorm_mode_32") {
      PARSE_BITS_ENTRY(KD.compute_pgm_rsrc1,
                       COMPUTE_PGM_RSRC1_FLOAT_DENORM_MODE_32, Val, ValRange);
    } else if (ID == ".amdhsa_float_denorm_mode_16_64") {
      PARSE_BITS_ENTRY(KD.compute_pgm_rsrc1,
                       COMPUTE_PGM_RSRC1_FLOAT_DENORM_MODE_16_64, Val,
                       ValRange);
    } else if (ID == ".amdhsa_dx10_clamp") {
      PARSE_BITS_ENTRY(KD.compute_pgm_rsrc1,
                       COMPUTE_PGM_RSRC1_ENABLE_DX10_CLAMP, Val, ValRange);
    } else if (ID == ".amdhsa_ieee_mode") {
      PARSE_BITS_ENTRY(KD.compute_pgm_rsrc1, COMPUTE_PGM_RSRC1_ENABLE_IEEE_MODE,
                       Val, ValRange);
    } else if (ID == ".amdhsa_fp16_overflow") {
      if (IVersion.Major < 9)
        return Error(IDRange.Start, "directive requires gfx9+", IDRange);
      PARSE_BITS_ENTRY(KD.compute_pgm_rsrc1, COMPUTE_PGM_RSRC1_FP16_OVFL, Val,
                       ValRange);
    } else if (ID == ".amdhsa_tg_split") {
      if (!isGFX90A())
        return Error(IDRange.Start, "directive requires gfx90a+", IDRange);
      PARSE_BITS_ENTRY(KD.compute_pgm_rsrc3, COMPUTE_PGM_RSRC3_GFX90A_TG_SPLIT, Val,
                       ValRange);
    } else if (ID == ".amdhsa_workgroup_processor_mode") {
      if (IVersion.Major < 10)
        return Error(IDRange.Start, "directive requires gfx10+", IDRange);
      PARSE_BITS_ENTRY(KD.compute_pgm_rsrc1, COMPUTE_PGM_RSRC1_WGP_MODE, Val,
                       ValRange);
    } else if (ID == ".amdhsa_memory_ordered") {
      if (IVersion.Major < 10)
        return Error(IDRange.Start, "directive requires gfx10+", IDRange);
      PARSE_BITS_ENTRY(KD.compute_pgm_rsrc1, COMPUTE_PGM_RSRC1_MEM_ORDERED, Val,
                       ValRange);
    } else if (ID == ".amdhsa_forward_progress") {
      if (IVersion.Major < 10)
        return Error(IDRange.Start, "directive requires gfx10+", IDRange);
      PARSE_BITS_ENTRY(KD.compute_pgm_rsrc1, COMPUTE_PGM_RSRC1_FWD_PROGRESS, Val,
                       ValRange);
    } else if (ID == ".amdhsa_shared_vgpr_count") {
      if (IVersion.Major < 10)
        return Error(IDRange.Start, "directive requires gfx10+", IDRange);
      SharedVGPRCount = Val;
      PARSE_BITS_ENTRY(KD.compute_pgm_rsrc3,
                       COMPUTE_PGM_RSRC3_GFX10_PLUS_SHARED_VGPR_COUNT, Val,
                       ValRange);
    } else if (ID == ".amdhsa_exception_fp_ieee_invalid_op") {
      PARSE_BITS_ENTRY(
          KD.compute_pgm_rsrc2,
          COMPUTE_PGM_RSRC2_ENABLE_EXCEPTION_IEEE_754_FP_INVALID_OPERATION, Val,
          ValRange);
    } else if (ID == ".amdhsa_exception_fp_denorm_src") {
      PARSE_BITS_ENTRY(KD.compute_pgm_rsrc2,
                       COMPUTE_PGM_RSRC2_ENABLE_EXCEPTION_FP_DENORMAL_SOURCE,
                       Val, ValRange);
    } else if (ID == ".amdhsa_exception_fp_ieee_div_zero") {
      PARSE_BITS_ENTRY(
          KD.compute_pgm_rsrc2,
          COMPUTE_PGM_RSRC2_ENABLE_EXCEPTION_IEEE_754_FP_DIVISION_BY_ZERO, Val,
          ValRange);
    } else if (ID == ".amdhsa_exception_fp_ieee_overflow") {
      PARSE_BITS_ENTRY(KD.compute_pgm_rsrc2,
                       COMPUTE_PGM_RSRC2_ENABLE_EXCEPTION_IEEE_754_FP_OVERFLOW,
                       Val, ValRange);
    } else if (ID == ".amdhsa_exception_fp_ieee_underflow") {
      PARSE_BITS_ENTRY(KD.compute_pgm_rsrc2,
                       COMPUTE_PGM_RSRC2_ENABLE_EXCEPTION_IEEE_754_FP_UNDERFLOW,
                       Val, ValRange);
    } else if (ID == ".amdhsa_exception_fp_ieee_inexact") {
      PARSE_BITS_ENTRY(KD.compute_pgm_rsrc2,
                       COMPUTE_PGM_RSRC2_ENABLE_EXCEPTION_IEEE_754_FP_INEXACT,
                       Val, ValRange);
    } else if (ID == ".amdhsa_exception_int_div_zero") {
      PARSE_BITS_ENTRY(KD.compute_pgm_rsrc2,
                       COMPUTE_PGM_RSRC2_ENABLE_EXCEPTION_INT_DIVIDE_BY_ZERO,
                       Val, ValRange);
    } else {
      return Error(IDRange.Start, "unknown .amdhsa_kernel directive", IDRange);
    }
 
#undef PARSE_BITS_ENTRY
  }
 
  if (!Seen.contains(".amdhsa_next_free_vgpr"))
    return TokError(".amdhsa_next_free_vgpr directive is required");
 
  if (!Seen.contains(".amdhsa_next_free_sgpr"))
    return TokError(".amdhsa_next_free_sgpr directive is required");
 
  unsigned VGPRBlocks;
  unsigned SGPRBlocks;
  if (calculateGPRBlocks(getFeatureBits(), ReserveVCC, Reserv