llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.h

//===- AMDGPUBaseInfo.h - Top level definitions for AMDGPU ------*- C++ -*-===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//

#ifndef LLVM_LIB_TARGET_AMDGPU_UTILS_AMDGPUBASEINFO_H
#define LLVM_LIB_TARGET_AMDGPU_UTILS_AMDGPUBASEINFO_H

#include "AMDGPUSubtarget.h"
#include "SIDefines.h"
#include "llvm/ADT/APFloat.h"
#include "llvm/ADT/StringExtras.h"
#include "llvm/ADT/StringTable.h"
#include "llvm/IR/CallingConv.h"
#include "llvm/IR/InstrTypes.h"
#include "llvm/IR/Module.h"
#include "llvm/Support/Alignment.h"
#include "llvm/TargetParser/AMDGPUTargetParser.h"
#include <array>
#include <functional>
#include <optional>
#include <utility>

// Pull in OpName enum definition and getNamedOperandIdx() declaration.
#define GET_INSTRINFO_OPERAND_ENUM
#include "AMDGPUGenInstrInfo.inc"

struct amd_kernel_code_t;

namespace llvm {

struct Align;
class Argument;
class Function;
class GlobalValue;
class MachineInstr;
class MCInstrInfo;
class MCRegisterClass;
class MCRegisterInfo;
class MCSubtargetInfo;
class MDNode;
class StringRef;
class Triple;
class raw_ostream;

namespace AMDGPU {

struct AMDGPUMCKernelCodeT;
struct IsaVersion;

/// Generic target versions emitted by this version of LLVM.
///
/// These numbers are incremented every time a codegen breaking change occurs
/// within a generic family.
namespace GenericVersion {
static constexpr unsigned GFX9 = 1;
static constexpr unsigned GFX9_4 = 1;
static constexpr unsigned GFX10_1 = 1;
static constexpr unsigned GFX10_3 = 1;
static constexpr unsigned GFX11 = 1;
static constexpr unsigned GFX11_7 = 1;
static constexpr unsigned GFX12 = 1;
static constexpr unsigned GFX12_5 = 1;
static constexpr unsigned GFX13 = 1;
} // namespace GenericVersion

enum { AMDHSA_COV4 = 4, AMDHSA_COV5 = 5, AMDHSA_COV6 = 6 };

enum class FPType { None, FP4, FP8 };

/// \returns True if \p STI is AMDHSA.
bool isHsaAbi(const MCSubtargetInfo &STI);

/// \returns Code object version from the IR module flag.
unsigned getAMDHSACodeObjectVersion(const Module &M);

/// \returns Code object version from ELF's e_ident[EI_ABIVERSION].
unsigned getAMDHSACodeObjectVersion(unsigned ABIVersion);

/// \returns The default HSA code object version. This should only be used when
/// we lack a more accurate CodeObjectVersion value (e.g. from the IR module
/// flag or a .amdhsa_code_object_version directive)
unsigned getDefaultAMDHSACodeObjectVersion();

/// \returns ABIVersion suitable for use in ELF's e_ident[EI_ABIVERSION]. \param
/// CodeObjectVersion is a value returned by getAMDHSACodeObjectVersion().
uint8_t getELFABIVersion(const Triple &OS, unsigned CodeObjectVersion);

/// \returns The offset of the multigrid_sync_arg argument from implicitarg_ptr
unsigned getMultigridSyncArgImplicitArgPosition(unsigned COV);

/// \returns The offset of the hostcall pointer argument from implicitarg_ptr
unsigned getHostcallImplicitArgPosition(unsigned COV);

unsigned getDefaultQueueImplicitArgPosition(unsigned COV);
unsigned getCompletionActionImplicitArgPosition(unsigned COV);

struct GcnBufferFormatInfo {
  unsigned Format;
  unsigned BitsPerComp;
  unsigned NumComponents;
  unsigned NumFormat;
  unsigned DataFormat;
};

struct MAIInstInfo {
  uint32_t Opcode;
  bool is_dgemm;
  bool is_gfx940_xdl;
};

struct MFMA_F8F6F4_Info {
  unsigned Opcode;
  unsigned F8F8Opcode;
  uint8_t NumRegsSrcA;
  uint8_t NumRegsSrcB;
};

/// Normalized WMMA or SWMMAC family used to select co-execution rules.
enum class WMMAVariant {
  Unknown = 0,
  IU8_16x16x64,
  F8F6F4_16x16x128,
  F8F6F4_16x16x128_BothF4,
  FP8BF8_16x16x64,
  F16BF16_16x16x32,
  FP8BF8_16x16x128,
  F4_32x16x128,
};

struct True16D16Info {
  unsigned T16Op;
  unsigned HiOp;
  unsigned LoOp;
};

struct WMMAInstInfo {
  uint32_t Opcode;
  bool is_wmma_xdl;
  bool HasMatrixScale;
  WMMAVariant CoExecVariant;
};

#define GET_MIMGBaseOpcode_DECL
#define GET_MIMGDim_DECL
#define GET_MIMGEncoding_DECL
#define GET_MIMGLZMapping_DECL
#define GET_MIMGMIPMapping_DECL
#define GET_MIMGBiASMapping_DECL
#define GET_MAIInstInfoTable_DECL
#define GET_isMFMA_F8F6F4Table_DECL
#define GET_isCvtScaleF32_F32F16ToF8F4Table_DECL
#define GET_True16D16Table_DECL
#define GET_WMMAInstInfoTable_DECL
#include "AMDGPUGenSearchableTables.inc"

using TargetIDSetting = AMDGPU::TargetIDSetting;
using TargetID = AMDGPU::TargetID;

/// Construct TargetID from MCSubtargetInfo. \p FeatureString is used to
/// determine explicitly requested xnack/sramecc settings.
TargetID createAMDGPUTargetID(const MCSubtargetInfo &STI,
                              StringRef FeatureString);

namespace IsaInfo {

enum {
  FIXED_NUM_SGPRS_FOR_INIT_BUG = AMDGPU::FIXED_NUM_SGPRS_FOR_INIT_BUG,
  TRAP_NUM_SGPRS = 16
};

/// Returns true if \p Lhs and \p Rhs are incompatible (both specific but
/// different).
inline bool targetIDSettingsConflict(TargetIDSetting Lhs, TargetIDSetting Rhs) {
  return Lhs != TargetIDSetting::Any && Rhs != TargetIDSetting::Any &&
         Lhs != Rhs;
}

/// \returns Instruction cache line size in bytes for given subtarget \p STI.
unsigned getInstCacheLineSize(const MCSubtargetInfo &STI);

/// \returns Wavefront size for given subtarget \p STI.
unsigned getWavefrontSize(const MCSubtargetInfo &STI);

/// \returns Local memory size in bytes for given subtarget \p STI.
unsigned getLocalMemorySize(const MCSubtargetInfo &STI);

/// \returns Maximum addressable local memory size in bytes for given subtarget
/// \p STI.
unsigned getAddressableLocalMemorySize(const MCSubtargetInfo &STI);

/// \returns Maximum number of work groups per compute unit for given subtarget
/// \p STI and limited by given \p FlatWorkGroupSize.
unsigned getMaxWorkGroupsPerCU(const MCSubtargetInfo &STI,
                               unsigned FlatWorkGroupSize);

/// \returns Number of waves per execution unit required to support the given \p
/// FlatWorkGroupSize.
unsigned getWavesPerEUForWorkGroup(const MCSubtargetInfo &STI,
                                   unsigned FlatWorkGroupSize);

/// \returns Number of waves per work group for given subtarget \p STI and
/// \p FlatWorkGroupSize.
unsigned getWavesPerWorkGroup(const MCSubtargetInfo &STI,
                              unsigned FlatWorkGroupSize);

/// \returns SGPR encoding granularity for given subtarget \p STI.
unsigned getSGPREncodingGranule(const MCSubtargetInfo &STI);

/// \returns Minimum number of SGPRs that meets the given number of waves per
/// execution unit requirement for given subtarget \p STI.
unsigned getMinNumSGPRs(const MCSubtargetInfo &STI, unsigned WavesPerEU);

/// \returns Maximum number of SGPRs that meets the given number of waves per
/// execution unit requirement for given subtarget \p STI.
unsigned getMaxNumSGPRs(const MCSubtargetInfo &STI, unsigned WavesPerEU,
                        bool Addressable);

/// \returns Number of extra SGPRs implicitly required by given subtarget \p
/// STI when the given special registers are used.
unsigned getNumExtraSGPRs(const MCSubtargetInfo &STI, bool VCCUsed,
                          bool FlatScrUsed, bool XNACKUsed);

/// \returns Number of SGPR blocks needed for given subtarget \p STI when
/// \p NumSGPRs are used. \p NumSGPRs should already include any special
/// register counts.
unsigned getNumSGPRBlocks(const MCSubtargetInfo &STI, unsigned NumSGPRs);

/// \returns VGPR allocation granularity for given subtarget \p STI.
///
/// For subtargets which support it, \p EnableWavefrontSize32 should match
/// the ENABLE_WAVEFRONT_SIZE32 kernel descriptor field.
unsigned
getVGPRAllocGranule(const MCSubtargetInfo &STI, unsigned DynamicVGPRBlockSize,
                    std::optional<bool> EnableWavefrontSize32 = std::nullopt);

/// \returns VGPR encoding granularity for given subtarget \p STI.
///
/// For subtargets which support it, \p EnableWavefrontSize32 should match
/// the ENABLE_WAVEFRONT_SIZE32 kernel descriptor field.
unsigned getVGPREncodingGranule(
    const MCSubtargetInfo &STI,
    std::optional<bool> EnableWavefrontSize32 = std::nullopt);

/// For subtargets with a unified VGPR file and mixed ArchVGPR/AGPR usage,
/// returns the allocation granule for ArchVGPRs.
unsigned getArchVGPRAllocGranule();

/// Maximum number of VGPR blocks that can be allocated in dynamic VGPR mode.
static constexpr unsigned MaxDynamicVGPRBlocks = 8;

/// \returns Addressable number of architectural VGPRs for a given subtarget \p
/// STI.
unsigned getAddressableNumArchVGPRs(const MCSubtargetInfo &STI);

/// \returns Addressable number of VGPRs for given subtarget \p STI.
unsigned getAddressableNumVGPRs(const MCSubtargetInfo &STI,
                                unsigned DynamicVGPRBlockSize);

/// \returns Minimum number of VGPRs that meets given number of waves per
/// execution unit requirement for given subtarget \p STI.
unsigned getMinNumVGPRs(const MCSubtargetInfo &STI, unsigned WavesPerEU,
                        unsigned DynamicVGPRBlockSize);

/// \returns Maximum number of VGPRs that meets given number of waves per
/// execution unit requirement for given subtarget \p STI.
unsigned getMaxNumVGPRs(const MCSubtargetInfo &STI, unsigned WavesPerEU,
                        unsigned DynamicVGPRBlockSize);

/// \returns Number of waves reachable for a given \p NumVGPRs usage for given
/// subtarget \p STI.
unsigned getNumWavesPerEUWithNumVGPRs(const MCSubtargetInfo &STI,
                                      unsigned NumVGPRs,
                                      unsigned DynamicVGPRBlockSize);

/// \returns Number of waves reachable for a given \p NumVGPRs usage, \p Granule
/// size, \p MaxWaves possible, and \p TotalNumVGPRs available.
unsigned getNumWavesPerEUWithNumVGPRs(unsigned NumVGPRs, unsigned Granule,
                                      unsigned MaxWaves,
                                      unsigned TotalNumVGPRs);

/// \returns Whether allocated SGPRs can reduce occupancy on subtarget \p STI
/// (true pre-GFX10). One named capability so callers don't test the version.
bool isSGPROccupancyLimited(const MCSubtargetInfo &STI);

/// \returns SGPR-limited occupancy (waves per EU) for subtarget \p STI: the
/// inverse of getMaxNumSGPRs(). Unlike getMaxNumSGPRs() the budget is not
/// clamped to the addressable count, since the allocated count callers pass in
/// can exceed it.
unsigned getOccupancyWithNumSGPRs(const MCSubtargetInfo &STI, unsigned SGPRs);

/// \returns SGPR-limited occupancy computed from explicit budget parameters
/// (\p MaxWaves, \p TotalNumSGPRs, \p Granule, \p TrapReserve). Subtarget-free
/// core shared by the overload above and the occupancy MCExpr. Callers must
/// check isSGPROccupancyLimited() first.
unsigned getOccupancyWithNumSGPRs(unsigned SGPRs, unsigned MaxWaves,
                                  unsigned TotalNumSGPRs, unsigned Granule,
                                  unsigned TrapReserve);

/// \returns Number of VGPR blocks that need to be allocated for the given
/// subtarget \p STI when \p NumVGPRs are used.
unsigned getAllocatedNumVGPRBlocks(
    const MCSubtargetInfo &STI, unsigned NumVGPRs,
    unsigned DynamicVGPRBlockSize,
    std::optional<bool> EnableWavefrontSize32 = std::nullopt);

} // end namespace IsaInfo

// Represents a field in an encoded value.
template <unsigned HighBit, unsigned LowBit, unsigned D = 0>
struct EncodingField {
  static_assert(HighBit >= LowBit, "Invalid bit range!");
  static constexpr unsigned Offset = LowBit;
  static constexpr unsigned Width = HighBit - LowBit + 1;

  using ValueType = unsigned;
  static_assert(Width <= sizeof(ValueType) * 8);
  static constexpr ValueType Default = D;

  ValueType Value;
  constexpr EncodingField(ValueType Value) : Value(Value) {}

  constexpr uint64_t encode() const { return Value; }
  static ValueType decode(uint64_t Encoded) {
    return static_cast<ValueType>(Encoded);
  }
};

// Represents a single bit in an encoded value.
template <unsigned Bit, unsigned D = 0>
using EncodingBit = EncodingField<Bit, Bit, D>;

// A helper for encoding and decoding multiple fields.
template <typename... Fields> struct EncodingFields {
  static constexpr uint64_t encode(Fields... Values) {
    return ((Values.encode() << Values.Offset) | ...);
  }

  static std::tuple<typename Fields::ValueType...> decode(uint64_t Encoded) {
    return {Fields::decode((Encoded >> Fields::Offset) &
                           maxUIntN(Fields::Width))...};
  }
};

LLVM_READONLY
inline bool hasNamedOperand(uint32_t Opcode, OpName NamedIdx) {
  return getNamedOperandIdx(Opcode, NamedIdx) != -1;
}

LLVM_READONLY
int32_t getSOPPWithRelaxation(uint32_t Opcode);

struct MIMGBaseOpcodeInfo {
  MIMGBaseOpcode BaseOpcode;
  bool Store;
  bool Atomic;
  bool AtomicX2;
  bool Sampler;
  bool Gather4;

  uint8_t NumExtraArgs;
  bool Gradients;
  bool G16;
  bool Coordinates;
  bool LodOrClampOrMip;
  bool HasD16;
  bool MSAA;
  bool BVH;
  bool A16;
  bool NoReturn;
  bool PointSampleAccel;
};

LLVM_READONLY
const MIMGBaseOpcodeInfo *getMIMGBaseOpcode(unsigned Opc);

LLVM_READONLY
const MIMGBaseOpcodeInfo *getMIMGBaseOpcodeInfo(unsigned BaseOpcode);

struct MIMGDimInfo {
  MIMGDim Dim;
  MIMGDim NonArrayDim;
  uint8_t NumCoords;
  uint8_t NumGradients;
  bool MSAA;
  bool DA;
  uint8_t Encoding;
  StringTable::Offset AsmSuffix;
};

LLVM_READONLY
const MIMGDimInfo *getMIMGDimInfo(unsigned DimEnum);

LLVM_READONLY StringRef getMIMGDimInfoStr(StringTable::Offset);

LLVM_READONLY
const MIMGDimInfo *getMIMGDimInfoByEncoding(uint8_t DimEnc);

LLVM_READONLY
const MIMGDimInfo *getMIMGDimInfoByAsmSuffix(StringRef AsmSuffix);

struct MIMGLZMappingInfo {
  MIMGBaseOpcode L;
  MIMGBaseOpcode LZ;
};

struct MIMGMIPMappingInfo {
  MIMGBaseOpcode MIP;
  MIMGBaseOpcode NONMIP;
};

struct MIMGBiasMappingInfo {
  MIMGBaseOpcode Bias;
  MIMGBaseOpcode NoBias;
};

struct MIMGOffsetMappingInfo {
  MIMGBaseOpcode Offset;
  MIMGBaseOpcode NoOffset;
};

struct MIMGG16MappingInfo {
  MIMGBaseOpcode G;
  MIMGBaseOpcode G16;
};

LLVM_READONLY
const MIMGLZMappingInfo *getMIMGLZMappingInfo(unsigned L);

struct WMMAOpcodeMappingInfo {
  unsigned Opcode2Addr;
  unsigned Opcode3Addr;
};

LLVM_READONLY
const MIMGMIPMappingInfo *getMIMGMIPMappingInfo(unsigned MIP);

LLVM_READONLY
const MIMGBiasMappingInfo *getMIMGBiasMappingInfo(unsigned Bias);

LLVM_READONLY
const MIMGOffsetMappingInfo *getMIMGOffsetMappingInfo(unsigned Offset);

LLVM_READONLY
const MIMGG16MappingInfo *getMIMGG16MappingInfo(unsigned G);

LLVM_READONLY
int getMIMGOpcode(unsigned BaseOpcode, unsigned MIMGEncoding,
                  unsigned VDataDwords, unsigned VAddrDwords,
                  bool IndexedRsrc = false, bool IndexedSamp = false);

LLVM_READONLY
int getMaskedMIMGOp(unsigned Opc, unsigned NewChannels);

LLVM_READONLY
unsigned getAddrSizeMIMGOp(const MIMGBaseOpcodeInfo *BaseOpcode,
                           const MIMGDimInfo *Dim, bool IsA16,
                           bool IsG16Supported);

struct MIMGInfo {
  uint32_t Opcode;
  uint32_t BaseOpcode;
  uint8_t MIMGEncoding;
  uint8_t VDataDwords;
  uint8_t VAddrDwords;
  uint8_t VAddrOperands;
  bool IndexedRsrc;
  bool IndexedSamp;
};

LLVM_READONLY
const MIMGInfo *getMIMGInfo(unsigned Opc);

LLVM_READONLY
int getMTBUFBaseOpcode(unsigned Opc);

LLVM_READONLY
int getMTBUFOpcode(unsigned BaseOpc, unsigned Elements);

LLVM_READONLY
int getMTBUFElements(unsigned Opc);

LLVM_READONLY
bool getMTBUFHasVAddr(unsigned Opc);

LLVM_READONLY
bool getMTBUFHasSrsrc(unsigned Opc);

LLVM_READONLY
bool getMTBUFHasSoffset(unsigned Opc);

LLVM_READONLY
int getMUBUFBaseOpcode(unsigned Opc);

LLVM_READONLY
int getMUBUFOpcode(unsigned BaseOpc, unsigned Elements);

LLVM_READONLY
int getMUBUFElements(unsigned Opc);

LLVM_READONLY
bool getMUBUFHasVAddr(unsigned Opc);

LLVM_READONLY
bool getMUBUFHasSrsrc(unsigned Opc);

LLVM_READONLY
bool getMUBUFHasSoffset(unsigned Opc);

LLVM_READONLY
bool getMUBUFIsBufferInv(unsigned Opc);

LLVM_READONLY
bool getMUBUFTfe(unsigned Opc);

LLVM_READONLY
bool getSMEMIsBuffer(unsigned Opc);

LLVM_READONLY
bool getVOP1IsSingle(unsigned Opc);

LLVM_READONLY
bool getVOP2IsSingle(unsigned Opc);

LLVM_READONLY
bool getVOP3IsSingle(unsigned Opc);

LLVM_READONLY
bool isVOPC64DPP(unsigned Opc);

LLVM_READONLY
bool isVOPCAsmOnly(unsigned Opc);

/// Returns true if this VOP3P opcode has no op_sel_hi field, in which case
/// those bits are encoded as zero instead of defaulting to one.
LLVM_READONLY
bool hasNoOpSelHi(unsigned Opc);

/// Returns true if MAI operation is a double precision GEMM.
LLVM_READONLY
bool getMAIIsDGEMM(unsigned Opc);

LLVM_READONLY
bool getMAIIsGFX940XDL(unsigned Opc);

LLVM_READONLY
bool getWMMAIsXDL(unsigned Opc);

LLVM_READONLY
bool getHasMatrixScale(unsigned Opc);

// Get an equivalent BitOp3 for a binary logical \p Opc.
// \returns BitOp3 modifier for the logical operation or zero.
// Used in VOPD3 conversion.
unsigned getBitOp2(unsigned Opc);

struct CanBeVOPD {
  bool X;
  bool Y;
};

/// \returns SIEncodingFamily used for VOPD encoding on a \p ST.
LLVM_READONLY
unsigned getVOPDEncodingFamily(const MCSubtargetInfo &ST);

LLVM_READONLY
CanBeVOPD getCanBeVOPD(unsigned Opc, unsigned EncodingFamily, bool VOPD3);

LLVM_READNONE
uint8_t mfmaScaleF8F6F4FormatToNumRegs(unsigned EncodingVal);

LLVM_READONLY
const MFMA_F8F6F4_Info *getMFMA_F8F6F4_WithFormatArgs(unsigned CBSZ,
                                                      unsigned BLGP,
                                                      unsigned F8F8Opcode);

LLVM_READNONE
uint8_t wmmaScaleF8F6F4FormatToNumRegs(unsigned Fmt);

LLVM_READONLY
const MFMA_F8F6F4_Info *getWMMA_F8F6F4_WithFormatArgs(unsigned FmtA,
                                                      unsigned FmtB,
                                                      unsigned F8F8Opcode);

/// \return true if this combination is listed as valid.
LLVM_READONLY
bool isValidWMMAScaleFmtCombination(unsigned AFmt, unsigned AScale,
                                    unsigned BFmt, unsigned BScale);

/// \returns the matrix_a_fmt and matrix_b_fmt operands of an f8f6f4 WMMA.
/// Works on both MCInst and MachineInstr.
template <typename InstT>
std::pair<unsigned, unsigned> getWMMAMatrixFmts(const InstT &MI) {
  unsigned Opc = MI.getOpcode();
  int AIdx = getNamedOperandIdx(Opc, OpName::matrix_a_fmt);
  int BIdx = getNamedOperandIdx(Opc, OpName::matrix_b_fmt);
  assert(AIdx != -1 && BIdx != -1 && "expected an f8f6f4 WMMA");
  return {static_cast<unsigned>(MI.getOperand(AIdx).getImm()),
          static_cast<unsigned>(MI.getOperand(BIdx).getImm())};
}

/// \returns true if either matrix input of an f8f6f4 WMMA is f8.
template <typename InstT> bool isWMMAAnyF8(const InstT &MI) {
  auto [AFmt, BFmt] = getWMMAMatrixFmts(MI);
  return AFmt <= WMMA::MATRIX_FMT_BF8 || BFmt <= WMMA::MATRIX_FMT_BF8;
}

/// \returns true if both matrix inputs of an f8f6f4 WMMA are f4.
template <typename InstT> bool isWMMABothF4(const InstT &MI) {
  auto [AFmt, BFmt] = getWMMAMatrixFmts(MI);
  return AFmt == WMMA::MATRIX_FMT_FP4 && BFmt == WMMA::MATRIX_FMT_FP4;
}

LLVM_READONLY
const GcnBufferFormatInfo *getGcnBufferFormatInfo(uint8_t BitsPerComp,
                                                  uint8_t NumComponents,
                                                  uint8_t NumFormat,
                                                  const MCSubtargetInfo &STI);
LLVM_READONLY
const GcnBufferFormatInfo *getGcnBufferFormatInfo(uint8_t Format,
                                                  const MCSubtargetInfo &STI);

LLVM_READONLY
int32_t getMCOpcode(uint32_t Opcode, unsigned Gen);

LLVM_READONLY
unsigned getVOPDOpcode(unsigned Opc, bool VOPD3);

LLVM_READONLY
int getVOPDFull(unsigned OpX, unsigned OpY, unsigned EncodingFamily,
                bool VOPD3);

LLVM_READONLY
bool isVOPD(unsigned Opc);

LLVM_READNONE
bool isMAC(unsigned Opc);

LLVM_READNONE
bool isPermlane16(unsigned Opc);

LLVM_READNONE
bool isGenericAtomic(unsigned Opc);

LLVM_READNONE
bool isCvt_F32_Fp8_Bf8_e64(unsigned Opc);

namespace VOPD {

enum Component : unsigned {
  DST = 0,
  SRC0,
  SRC1,
  SRC2,

  DST_NUM = 1,
  MAX_SRC_NUM = 3,
  MAX_OPR_NUM = DST_NUM + MAX_SRC_NUM
};

// LSB mask for VGPR banks per VOPD component operand.
// 4 banks result in a mask 3, setting 2 lower bits.
constexpr unsigned VOPD_VGPR_BANK_MASKS[] = {1, 3, 3, 1};
constexpr unsigned VOPD3_VGPR_BANK_MASKS[] = {1, 3, 3, 3};
// GFX11 VOPD interlock hazard requires SRC0/SRC1 to have
// different parities, not just on different banks. Else,
// non-deterministic forwarding error may occur.
constexpr unsigned VOPD_GFX11_VGPR_BANK_MASKS[] = {1, 1, 1, 1};

enum ComponentIndex : unsigned { X = 0, Y = 1 };
constexpr unsigned COMPONENTS[] = {ComponentIndex::X, ComponentIndex::Y};
constexpr unsigned COMPONENTS_NUM = 2;

// Properties of VOPD components.
class ComponentProps {
private:
  unsigned SrcOperandsNum = 0;
  unsigned MandatoryLiteralIdx = ~0u;
  bool HasSrc2Acc = false;
  unsigned NumVOPD3Mods = 0;
  unsigned Opcode = 0;
  bool IsVOP3 = false;

public:
  ComponentProps() = default;
  ComponentProps(const MCInstrDesc &OpDesc, bool VOP3Layout = false);

  // Return the total number of src operands this component has.
  unsigned getCompSrcOperandsNum() const { return SrcOperandsNum; }

  // Return the number of src operands of this component visible to the parser.
  unsigned getCompParsedSrcOperandsNum() const {
    return SrcOperandsNum - HasSrc2Acc;
  }

  // Return true iif this component has a mandatory literal.
  bool hasMandatoryLiteral() const { return MandatoryLiteralIdx != ~0u; }

  // If this component has a mandatory literal, return component operand
  // index of this literal (i.e. either Component::SRC1 or Component::SRC2).
  unsigned getMandatoryLiteralCompOperandIndex() const {
    assert(hasMandatoryLiteral());
    return MandatoryLiteralIdx;
  }

  // Return true iif this component has operand
  // with component index CompSrcIdx and this operand may be a register.
  bool hasRegSrcOperand(unsigned CompSrcIdx) const {
    assert(CompSrcIdx < Component::MAX_SRC_NUM);
    return SrcOperandsNum > CompSrcIdx && !hasMandatoryLiteralAt(CompSrcIdx);
  }

  // Return true iif this component has tied src2.
  bool hasSrc2Acc() const { return HasSrc2Acc; }

  // Return a number of source modifiers if instruction is used in VOPD3.
  unsigned getCompVOPD3ModsNum() const { return NumVOPD3Mods; }

  // Return opcode of the component.
  unsigned getOpcode() const { return Opcode; }

  // Returns if component opcode is in VOP3 encoding.
  unsigned isVOP3() const { return IsVOP3; }

  // Return index of BitOp3 operand or -1.
  int getBitOp3OperandIdx() const;

private:
  bool hasMandatoryLiteralAt(unsigned CompSrcIdx) const {
    assert(CompSrcIdx < Component::MAX_SRC_NUM);
    return MandatoryLiteralIdx == Component::DST_NUM + CompSrcIdx;
  }
};

enum ComponentKind : unsigned {
  SINGLE = 0,  // A single VOP1 or VOP2 instruction which may be used in VOPD.
  COMPONENT_X, // A VOPD instruction, X component.
  COMPONENT_Y, // A VOPD instruction, Y component.
  MAX = COMPONENT_Y
};

// Interface functions of this class map VOPD component operand indices
// to indices of operands in MachineInstr/MCInst or parsed operands array.
//
// Note that this class operates with 3 kinds of indices:
// - VOPD component operand indices (Component::DST, Component::SRC0, etc.);
// - MC operand indices (they refer operands in a MachineInstr/MCInst);
// - parsed operand indices (they refer operands in parsed operands array).
//
// For SINGLE components mapping between these indices is trivial.
// But things get more complicated for COMPONENT_X and
// COMPONENT_Y because these components share the same
// MachineInstr/MCInst and the same parsed operands array.
// Below is an example of component operand to parsed operand
// mapping for the following instruction:
//
//   v_dual_add_f32 v255, v4, v5 :: v_dual_mov_b32 v6, v1
//
//                          PARSED        COMPONENT         PARSED
// COMPONENT               OPERANDS     OPERAND INDEX    OPERAND INDEX
// -------------------------------------------------------------------
//                     "v_dual_add_f32"                        0
// v_dual_add_f32            v255          0 (DST)    -->      1
//                           v4            1 (SRC0)   -->      2
//                           v5            2 (SRC1)   -->      3
//                          "::"                               4
//                     "v_dual_mov_b32"                        5
// v_dual_mov_b32            v6            0 (DST)    -->      6
//                           v1            1 (SRC0)   -->      7
// -------------------------------------------------------------------
//
class ComponentLayout {
private:
  // Regular MachineInstr/MCInst operands are ordered as follows:
  //   dst, src0 [, other src operands]
  // VOPD MachineInstr/MCInst operands are ordered as follows:
  //   dstX, dstY, src0X [, other OpX operands], src0Y [, other OpY operands]
  // Each ComponentKind has operand indices defined below.
  static constexpr unsigned MC_DST_IDX[] = {0, 0, 1};

  // VOPD3 instructions may have 2 or 3 source modifiers, src2 modifier is not
  // used if there is tied accumulator. Indexing of this array:
  // MC_SRC_IDX[VOPD3ModsNum][SrcNo]. This returns an index for a SINGLE
  // instruction layout, add 1 for COMPONENT_X or COMPONENT_Y. For the second
  // component add OpX.MCSrcNum + OpX.VOPD3ModsNum.
  // For VOPD1/VOPD2 use column with zero modifiers.
  static constexpr unsigned SINGLE_MC_SRC_IDX[4][3] = {
      {1, 2, 3}, {2, 3, 4}, {2, 4, 5}, {2, 4, 6}};

  // Parsed operands of regular instructions are ordered as follows:
  //   Mnemo dst src0 [vsrc1 ...]
  // Parsed VOPD operands are ordered as follows:
  //   OpXMnemo dstX src0X [vsrc1X|imm vsrc1X|vsrc1X imm] '::'
  //   OpYMnemo dstY src0Y [vsrc1Y|imm vsrc1Y|vsrc1Y imm]
  // Each ComponentKind has operand indices defined below.
  static constexpr unsigned PARSED_DST_IDX[] = {1, 1,
                                                4 /* + OpX.ParsedSrcNum */};
  static constexpr unsigned FIRST_PARSED_SRC_IDX[] = {
      2, 2, 5 /* + OpX.ParsedSrcNum */};

private:
  const ComponentKind Kind;
  const ComponentProps PrevComp;
  const unsigned VOPD3ModsNum;
  const int BitOp3Idx; // Index of bitop3 operand or -1

public:
  // Create layout for COMPONENT_X or SINGLE component.
  ComponentLayout(ComponentKind Kind, unsigned VOPD3ModsNum, int BitOp3Idx)
      : Kind(Kind), VOPD3ModsNum(VOPD3ModsNum), BitOp3Idx(BitOp3Idx) {
    assert(Kind == ComponentKind::SINGLE || Kind == ComponentKind::COMPONENT_X);
  }

  // Create layout for COMPONENT_Y which depends on COMPONENT_X layout.
  ComponentLayout(const ComponentProps &OpXProps, unsigned VOPD3ModsNum,
                  int BitOp3Idx)
      : Kind(ComponentKind::COMPONENT_Y), PrevComp(OpXProps),
        VOPD3ModsNum(VOPD3ModsNum), BitOp3Idx(BitOp3Idx) {}

public:
  // Return the index of dst operand in MCInst operands.
  unsigned getIndexOfDstInMCOperands() const { return MC_DST_IDX[Kind]; }

  // Return the index of the specified src operand in MCInst operands.
  unsigned getIndexOfSrcInMCOperands(unsigned CompSrcIdx, bool VOPD3) const {
    assert(CompSrcIdx < Component::MAX_SRC_NUM);

    if (Kind == SINGLE && CompSrcIdx == 2 && BitOp3Idx != -1)
      return BitOp3Idx;

    if (VOPD3) {
      return SINGLE_MC_SRC_IDX[VOPD3ModsNum][CompSrcIdx] + getPrevCompSrcNum() +
             getPrevCompVOPD3ModsNum() + (Kind != SINGLE ? 1 : 0);
    }

    return SINGLE_MC_SRC_IDX[0][CompSrcIdx] + getPrevCompSrcNum() +
           (Kind != SINGLE ? 1 : 0);
  }

  // Return the index of dst operand in the parsed operands array.
  unsigned getIndexOfDstInParsedOperands() const {
    return PARSED_DST_IDX[Kind] + getPrevCompParsedSrcNum();
  }

  // Return the index of the specified src operand in the parsed operands array.
  unsigned getIndexOfSrcInParsedOperands(unsigned CompSrcIdx) const {
    assert(CompSrcIdx < Component::MAX_SRC_NUM);
    return FIRST_PARSED_SRC_IDX[Kind] + getPrevCompParsedSrcNum() + CompSrcIdx;
  }

private:
  unsigned getPrevCompSrcNum() const {
    return PrevComp.getCompSrcOperandsNum();
  }
  unsigned getPrevCompParsedSrcNum() const {
    return PrevComp.getCompParsedSrcOperandsNum();
  }
  unsigned getPrevCompVOPD3ModsNum() const {
    return PrevComp.getCompVOPD3ModsNum();
  }
};

// Layout and properties of VOPD components.
class ComponentInfo : public ComponentProps, public ComponentLayout {
public:
  // Create ComponentInfo for COMPONENT_X or SINGLE component.
  ComponentInfo(const MCInstrDesc &OpDesc,
                ComponentKind Kind = ComponentKind::SINGLE,
                bool VOP3Layout = false)
      : ComponentProps(OpDesc, VOP3Layout),
        ComponentLayout(Kind, getCompVOPD3ModsNum(), getBitOp3OperandIdx()) {}

  // Create ComponentInfo for COMPONENT_Y which depends on COMPONENT_X layout.
  ComponentInfo(const MCInstrDesc &OpDesc, const ComponentProps &OpXProps,
                bool VOP3Layout = false)
      : ComponentProps(OpDesc, VOP3Layout),
        ComponentLayout(OpXProps, getCompVOPD3ModsNum(),
                        getBitOp3OperandIdx()) {}

  // Map component operand index to parsed operand index.
  // Return 0 if the specified operand does not exist.
  unsigned getIndexInParsedOperands(unsigned CompOprIdx) const;
};

// Properties of VOPD instructions.
class InstInfo {
private:
  const ComponentInfo CompInfo[COMPONENTS_NUM];

public:
  using RegIndices = std::array<MCRegister, Component::MAX_OPR_NUM>;

  InstInfo(const MCInstrDesc &OpX, const MCInstrDesc &OpY)
      : CompInfo{OpX, OpY} {}

  InstInfo(const ComponentInfo &OprInfoX, const ComponentInfo &OprInfoY)
      : CompInfo{OprInfoX, OprInfoY} {}

  const ComponentInfo &operator[](size_t ComponentIdx) const {
    assert(ComponentIdx < COMPONENTS_NUM);
    return CompInfo[ComponentIdx];
  }

  // Check VOPD operands constraints.
  // GetRegIdx(Component, MCOperandIdx) must return a VGPR register index
  // for the specified component and MC operand. The callback must return 0
  // if the operand is not a register or not a VGPR.
  // If \p SkipSrc is set to true then constraints for source operands are not
  // checked.
  // If \p AllowSameVGPR is set then same VGPRs are allowed for X and Y sources
  // even though it violates requirement to be from different banks.
  // If \p VOPD3 is set to true both dst registers allowed to be either odd
  // or even and instruction may have real src2 as opposed to tied accumulator.
  // If \p HasGFX11InterlockHazard is set then X/Y SRC0 and SRC1 VGPRs
  // must have different register-number parity.
  bool
  hasInvalidOperand(std::function<MCRegister(unsigned, unsigned)> GetRegIdx,
                    const MCRegisterInfo &MRI, bool SkipSrc = false,
                    bool AllowSameVGPR = false, bool VOPD3 = false,
                    bool HasGFX11InterlockHazard = false) const {
    return getInvalidCompOperandIndex(GetRegIdx, MRI, SkipSrc, AllowSameVGPR,
                                      VOPD3, HasGFX11InterlockHazard)
        .has_value();
  }

  // Check VOPD operands constraints.
  // Return the index of an invalid component operand, if any.
  // If \p SkipSrc is set to true then constraints for source operands are not
  // checked except for being from the same halves of VGPR file on gfx1250.
  // If \p AllowSameVGPR is set then same VGPRs are allowed for X and Y sources
  // even though it violates requirement to be from different banks.
  // If \p VOPD3 is set to true both dst registers allowed to be either odd
  // or even and instruction may have real src2 as opposed to tied accumulator.
  // If \p HasGFX11InterlockHazard is set then X/Y SRC0 and SRC1 VGPRs
  // must have different register-number parity.
  std::optional<unsigned> getInvalidCompOperandIndex(
      std::function<MCRegister(unsigned, unsigned)> GetRegIdx,
      const MCRegisterInfo &MRI, bool SkipSrc = false,
      bool AllowSameVGPR = false, bool VOPD3 = false,
      bool HasGFX11InterlockHazard = false) const;

private:
  RegIndices
  getRegIndices(unsigned ComponentIdx,
                std::function<MCRegister(unsigned, unsigned)> GetRegIdx,
                bool VOPD3) const;
};

} // namespace VOPD

LLVM_READONLY
std::pair<unsigned, unsigned> getVOPDComponents(unsigned VOPDOpcode);

LLVM_READONLY
// Get properties of 2 single VOP1/VOP2 instructions
// used as components to create a VOPD instruction.
VOPD::InstInfo getVOPDInstInfo(const MCInstrDesc &OpX, const MCInstrDesc &OpY);

LLVM_READONLY
// Get properties of VOPD X and Y components.
VOPD::InstInfo getVOPDInstInfo(unsigned VOPDOpcode,
                               const MCInstrInfo *InstrInfo);

LLVM_READONLY
bool isAsyncStore(unsigned Opc);
LLVM_READONLY
bool isTensorStore(unsigned Opc);
LLVM_READONLY
unsigned getTemporalHintType(const MCInstrDesc TID);

LLVM_READONLY
bool isTrue16Inst(unsigned Opc);

LLVM_READONLY
FPType getFPDstSelType(unsigned Opc);

bool isDPMACCInstruction(unsigned Opc);

LLVM_READONLY
unsigned mapWMMA2AddrTo3AddrOpcode(unsigned Opc);

void initDefaultAMDKernelCodeT(AMDGPUMCKernelCodeT &Header,
                               const MCSubtargetInfo &STI);

bool isReadOnlySegment(const GlobalValue *GV);

/// \returns True if constants should be emitted to .text section for given
/// target triple \p TT, false otherwise.
bool shouldEmitConstantsToTextSection(const Triple &TT);

/// Returns a valid charcode or 0 in the first entry if this is a valid physical
/// register name. Followed by the start register number, and the register
/// width. Does not validate the number of registers exists in the class. Unlike
/// parseAsmConstraintPhysReg, this does not expect the name to be wrapped in
/// "{}".
std::tuple<char, unsigned, unsigned> parseAsmPhysRegName(StringRef TupleString);

/// Returns a valid charcode or 0 in the first entry if this is a valid physical
/// register constraint. Followed by the start register number, and the register
/// width. Does not validate the number of registers exists in the class.
std::tuple<char, unsigned, unsigned>
parseAsmConstraintPhysReg(StringRef Constraint);

/// \returns A pair of integer values requested using \p F's \p Name attribute
/// in "first[,second]" format ("second" is optional unless \p OnlyFirstRequired
/// is false).
///
/// \returns \p Default if attribute is not present.
///
/// \returns \p Default and emits error if one of the requested values cannot be
/// converted to integer, or \p OnlyFirstRequired is false and "second" value is
/// not present.
std::pair<unsigned, unsigned>
getIntegerPairAttribute(const Function &F, StringRef Name,
                        std::pair<unsigned, unsigned> Default,
                        bool OnlyFirstRequired = false);

/// \returns A pair of integer values requested using \p F's \p Name attribute
/// in "first[,second]" format ("second" is optional unless \p OnlyFirstRequired
/// is false).
///
/// \returns \p std::nullopt if attribute is not present.
///
/// \returns \p std::nullopt and emits error if one of the requested values
/// cannot be converted to integer, or \p OnlyFirstRequired is false and
/// "second" value is not present.
std::optional<std::pair<unsigned, std::optional<unsigned>>>
getIntegerPairAttribute(const Function &F, StringRef Name,
                        bool OnlyFirstRequired = false);

/// \returns Generate a vector of integer values requested using \p F's \p Name
/// attribute.
/// \returns A vector of size \p Size, with all elements set to \p DefaultVal,
/// if any error occurs. The corresponding error will also be emitted.
SmallVector<unsigned> getIntegerVecAttribute(const Function &F, StringRef Name,
                                             unsigned Size,
                                             unsigned DefaultVal);
/// Similar to the function above, but returns std::nullopt if any error occurs.
std::optional<SmallVector<unsigned>>
getIntegerVecAttribute(const Function &F, StringRef Name, unsigned Size);

/// \returns The maximum number of workgroups for the function.
SmallVector<unsigned> getMaxNumWorkGroups(const Function &F);

inline bool isTgSplitEnabled(const Function &F) {
  return F.hasFnAttribute("amdgpu-tg-split");
}

/// Checks if \p Val is inside \p MD, a !range-like metadata.
bool hasValueInRangeLikeMetadata(const MDNode &MD, int64_t Val);

// The following methods are only meaningful on targets that support
// S_WAITCNT.

/// \returns Vmcnt bit mask for given isa \p Version.
unsigned getVmcntBitMask(const IsaVersion &Version);

/// \returns Expcnt bit mask for given isa \p Version.
unsigned getExpcntBitMask(const IsaVersion &Version);

/// \returns Lgkmcnt bit mask for given isa \p Version.
unsigned getLgkmcntBitMask(const IsaVersion &Version);

/// \returns Waitcnt bit mask for given isa \p Version.
unsigned getWaitcntBitMask(const IsaVersion &Version);

/// \returns Decoded Vmcnt from given \p Waitcnt for given isa \p Version.
unsigned decodeVmcnt(const IsaVersion &Version, unsigned Waitcnt);

/// \returns Decoded Expcnt from given \p Waitcnt for given isa \p Version.
unsigned decodeExpcnt(const IsaVersion &Version, unsigned Waitcnt);

/// \returns Decoded Lgkmcnt from given \p Waitcnt for given isa \p Version.
unsigned decodeLgkmcnt(const IsaVersion &Version, unsigned Waitcnt);

/// \returns Decoded Loadcnt from given \p Waitcnt for given isa \p Version.
unsigned decodeLoadcnt(const IsaVersion &Version, unsigned Waitcnt);

/// \returns Decoded Storecnt from given \p Waitcnt for given isa \p Version.
unsigned decodeStorecnt(const IsaVersion &Version, unsigned Waitcnt);

/// \returns Decoded Dscnt from given \p Waitcnt for given isa \p Version.
unsigned decodeDscnt(const IsaVersion &Version, unsigned Waitcnt);

/// Decodes Vmcnt, Expcnt and Lgkmcnt from given \p Waitcnt for given isa
/// \p Version, and writes decoded values into \p Vmcnt, \p Expcnt and
/// \p Lgkmcnt respectively. Should not be used on gfx12+, the instruction
/// which needs it is deprecated
///
/// \details \p Vmcnt, \p Expcnt and \p Lgkmcnt are decoded as follows:
///     \p Vmcnt = \p Waitcnt[3:0]        (pre-gfx9)
///     \p Vmcnt = \p Waitcnt[15:14,3:0]  (gfx9,10)
///     \p Vmcnt = \p Waitcnt[15:10]      (gfx11)
///     \p Expcnt = \p Waitcnt[6:4]       (pre-gfx11)
///     \p Expcnt = \p Waitcnt[2:0]       (gfx11)
///     \p Lgkmcnt = \p Waitcnt[11:8]     (pre-gfx10)
///     \p Lgkmcnt = \p Waitcnt[13:8]     (gfx10)
///     \p Lgkmcnt = \p Waitcnt[9:4]      (gfx11)
///
void decodeWaitcnt(const IsaVersion &Version, unsigned Waitcnt, unsigned &Vmcnt,
                   unsigned &Expcnt, unsigned &Lgkmcnt);

/// \returns \p Waitcnt with encoded \p Vmcnt for given isa \p Version.
unsigned encodeVmcnt(const IsaVersion &Version, unsigned Waitcnt,
                     unsigned Vmcnt);

/// \returns \p Waitcnt with encoded \p Expcnt for given isa \p Version.
unsigned encodeExpcnt(const IsaVersion &Version, unsigned Waitcnt,
                      unsigned Expcnt);

/// \returns \p Waitcnt with encoded \p Lgkmcnt for given isa \p Version.
unsigned encodeLgkmcnt(const IsaVersion &Version, unsigned Waitcnt,
                       unsigned Lgkmcnt);

/// Encodes \p Vmcnt, \p Expcnt and \p Lgkmcnt into Waitcnt for given isa
/// \p Version. Should not be used on gfx12+, the instruction which needs
/// it is deprecated
///
/// \details \p Vmcnt, \p Expcnt and \p Lgkmcnt are encoded as follows:
///     Waitcnt[2:0]   = \p Expcnt      (gfx11+)
///     Waitcnt[3:0]   = \p Vmcnt       (pre-gfx9)
///     Waitcnt[3:0]   = \p Vmcnt[3:0]  (gfx9,10)
///     Waitcnt[6:4]   = \p Expcnt      (pre-gfx11)
///     Waitcnt[9:4]   = \p Lgkmcnt     (gfx11)
///     Waitcnt[11:8]  = \p Lgkmcnt     (pre-gfx10)
///     Waitcnt[13:8]  = \p Lgkmcnt     (gfx10)
///     Waitcnt[15:10] = \p Vmcnt       (gfx11)
///     Waitcnt[15:14] = \p Vmcnt[5:4]  (gfx9,10)
///
/// \returns Waitcnt with encoded \p Vmcnt, \p Expcnt and \p Lgkmcnt for given
/// isa \p Version.
///
unsigned encodeWaitcnt(const IsaVersion &Version, unsigned Vmcnt,
                       unsigned Expcnt, unsigned Lgkmcnt);

/// \returns Waitcnt with encoded \p Loadcnt and \p Dscnt for given isa \p
/// Version.
unsigned encodeLoadcntDscnt(const IsaVersion &Version, unsigned Loadcnt,
                            unsigned Dscnt);

/// \returns Waitcnt with encoded \p Storecnt and \p Dscnt for given isa \p
/// Version.
unsigned encodeStorecntDscnt(const IsaVersion &Version, unsigned Storecnt,
                             unsigned Dscnt);

// The following methods are only meaningful on targets that support
// S_WAIT_*CNT, introduced with gfx12.

/// \returns Loadcnt bit mask for given isa \p Version.
/// Returns 0 for versions that do not support LOADcnt
unsigned getLoadcntBitMask(const IsaVersion &Version);

/// \returns Samplecnt bit mask for given isa \p Version.
/// Returns 0 for versions that do not support SAMPLEcnt
unsigned getSamplecntBitMask(const IsaVersion &Version);

/// \returns Bvhcnt bit mask for given isa \p Version.
/// Returns 0 for versions that do not support BVHcnt
unsigned getBvhcntBitMask(const IsaVersion &Version);

/// \returns Asynccnt bit mask for given isa \p Version.
/// Returns 0 for versions that do not support Asynccnt
unsigned getAsynccntBitMask(const IsaVersion &Version);

/// \returns Dscnt bit mask for given isa \p Version.
/// Returns 0 for versions that do not support DScnt
unsigned getDscntBitMask(const IsaVersion &Version);

/// \returns Dscnt bit mask for given isa \p Version.
/// Returns 0 for versions that do not support KMcnt
unsigned getKmcntBitMask(const IsaVersion &Version);

/// \returns Xcnt bit mask for given isa \p Version.
/// Returns 0 for versions that do not support Xcnt.
unsigned getXcntBitMask(const IsaVersion &Version);

/// \return STOREcnt or VScnt bit mask for given isa \p Version.
/// returns 0 for versions that do not support STOREcnt or VScnt.
/// STOREcnt and VScnt are the same counter, the name used
/// depends on the ISA version.
unsigned getStorecntBitMask(const IsaVersion &Version);

/// \returns barrier member count bit mask in M0 for given isa \p Version.
unsigned getBarrierMemberCountBitMask(const IsaVersion &Version);

namespace Hwreg {

using HwregId = EncodingField<5, 0>;
using HwregOffset = EncodingField<10, 6>;

struct HwregSize : EncodingField<15, 11, 32> {
  using EncodingField::EncodingField;
  constexpr uint64_t encode() const { return Value - 1; }
  static ValueType decode(uint64_t Encoded) {
    return static_cast<ValueType>(Encoded + 1);
  }
};

using HwregEncoding = EncodingFields<HwregId, HwregOffset, HwregSize>;

} // namespace Hwreg

namespace DepCtr {

int getDefaultDepCtrEncoding(const MCSubtargetInfo &STI);
int encodeDepCtr(const StringRef Name, int64_t Val, unsigned &UsedOprMask,
                 const MCSubtargetInfo &STI);
bool isSymbolicDepCtrEncoding(unsigned Code, bool &HasNonDefaultVal,
                              const MCSubtargetInfo &STI);
bool decodeDepCtr(unsigned Code, int &Id, StringRef &Name, unsigned &Val,
                  bool &IsDefault, const MCSubtargetInfo &STI);

/// \returns Maximum VaVdst value that can be encoded.
unsigned getVaVdstBitMask();

/// \returns Maximum VaSdst value that can be encoded.
unsigned getVaSdstBitMask();

/// \returns Maximum VaSsrc value that can be encoded.
unsigned getVaSsrcBitMask();

/// \returns Maximum HoldCnt value that can be encoded.
unsigned getHoldCntBitMask(const IsaVersion &Version);

/// \returns Maximum VmVsrc value that can be encoded.
unsigned getVmVsrcBitMask();

/// \returns Maximum VaVcc value that can be encoded.
unsigned getVaVccBitMask();

/// \returns Maximum SaSdst value that can be encoded.
unsigned getSaSdstBitMask();

/// \returns Decoded VaVdst from given immediate \p Encoded.
unsigned decodeFieldVaVdst(unsigned Encoded);

/// \returns Decoded VmVsrc from given immediate \p Encoded.
unsigned decodeFieldVmVsrc(unsigned Encoded);

/// \returns Decoded SaSdst from given immediate \p Encoded.
unsigned decodeFieldSaSdst(unsigned Encoded);

/// \returns Decoded VaSdst from given immediate \p Encoded.
unsigned decodeFieldVaSdst(unsigned Encoded);

/// \returns Decoded VaVcc from given immediate \p Encoded.
unsigned decodeFieldVaVcc(unsigned Encoded);

/// \returns Decoded SaSrc from given immediate \p Encoded.
unsigned decodeFieldVaSsrc(unsigned Encoded);

/// \returns Decoded HoldCnt from given immediate \p Encoded.
unsigned decodeFieldHoldCnt(unsigned Encoded, const IsaVersion &Version);

/// \returns \p VmVsrc as an encoded Depctr immediate.
unsigned encodeFieldVmVsrc(unsigned VmVsrc, const MCSubtargetInfo &STI);

/// \returns \p Encoded combined with encoded \p VmVsrc.
unsigned encodeFieldVmVsrc(unsigned Encoded, unsigned VmVsrc);

/// \returns \p VaVdst as an encoded Depctr immediate.
unsigned encodeFieldVaVdst(unsigned VaVdst, const MCSubtargetInfo &STI);

/// \returns \p Encoded combined with encoded \p VaVdst.
unsigned encodeFieldVaVdst(unsigned Encoded, unsigned VaVdst);

/// \returns \p SaSdst as an encoded Depctr immediate.
unsigned encodeFieldSaSdst(unsigned SaSdst, const MCSubtargetInfo &STI);

/// \returns \p Encoded combined with encoded \p SaSdst.
unsigned encodeFieldSaSdst(unsigned Encoded, unsigned SaSdst);

/// \returns \p VaSdst as an encoded Depctr immediate.
unsigned encodeFieldVaSdst(unsigned VaSdst, const MCSubtargetInfo &STI);

/// \returns \p Encoded combined with encoded \p VaSdst.
unsigned encodeFieldVaSdst(unsigned Encoded, unsigned VaSdst);

/// \returns \p VaVcc as an encoded Depctr immediate.
unsigned encodeFieldVaVcc(unsigned VaVcc, const MCSubtargetInfo &STI);

/// \returns \p Encoded combined with encoded \p VaVcc.
unsigned encodeFieldVaVcc(unsigned Encoded, unsigned VaVcc);

/// \returns \p HoldCnt as an encoded Depctr immediate.
unsigned encodeFieldHoldCnt(unsigned HoldCnt, const MCSubtargetInfo &STI);

/// \returns \p Encoded combined with encoded \p HoldCnt.
unsigned encodeFieldHoldCnt(unsigned Encoded, unsigned HoldCnt,
                            const IsaVersion &Version);

/// \returns \p VaSsrc as an encoded Depctr immediate.
unsigned encodeFieldVaSsrc(unsigned VaSsrc, const MCSubtargetInfo &STI);

/// \returns \p Encoded combined with encoded \p VaSsrc.
unsigned encodeFieldVaSsrc(unsigned Encoded, unsigned VaSsrc);

} // namespace DepCtr

namespace Exp {

bool getTgtName(unsigned Id, StringRef &Name, int &Index);

LLVM_READONLY
unsigned getTgtId(const StringRef Name);

LLVM_READNONE
bool isSupportedTgtId(unsigned Id, const MCSubtargetInfo &STI);

} // namespace Exp

namespace MTBUFFormat {

LLVM_READNONE
int64_t encodeDfmtNfmt(unsigned Dfmt, unsigned Nfmt);

void decodeDfmtNfmt(unsigned Format, unsigned &Dfmt, unsigned &Nfmt);

int64_t getDfmt(const StringRef Name);

StringRef getDfmtName(unsigned Id);

int64_t getNfmt(const StringRef Name, const MCSubtargetInfo &STI);

StringRef getNfmtName(unsigned Id, const MCSubtargetInfo &STI);

bool isValidDfmtNfmt(unsigned Val, const MCSubtargetInfo &STI);

bool isValidNfmt(unsigned Val, const MCSubtargetInfo &STI);

int64_t getUnifiedFormat(const StringRef Name, const MCSubtargetInfo &STI);

StringRef getUnifiedFormatName(unsigned Id, const MCSubtargetInfo &STI);

bool isValidUnifiedFormat(unsigned Val, const MCSubtargetInfo &STI);

int64_t convertDfmtNfmt2Ufmt(unsigned Dfmt, unsigned Nfmt,
                             const MCSubtargetInfo &STI);

bool isValidFormatEncoding(unsigned Val, const MCSubtargetInfo &STI);

unsigned getDefaultFormatEncoding(const MCSubtargetInfo &STI);

} // namespace MTBUFFormat

namespace SendMsg {

LLVM_READNONE
bool isValidMsgId(int64_t MsgId, const MCSubtargetInfo &STI);

LLVM_READNONE
bool isValidMsgOp(int64_t MsgId, int64_t OpId, const MCSubtargetInfo &STI,
                  bool Strict = true);

LLVM_READNONE
bool isValidMsgStream(int64_t MsgId, int64_t OpId, int64_t StreamId,
                      const MCSubtargetInfo &STI, bool Strict = true);

LLVM_READNONE
bool msgRequiresOp(int64_t MsgId, const MCSubtargetInfo &STI);

LLVM_READNONE
bool msgSupportsStream(int64_t MsgId, int64_t OpId, const MCSubtargetInfo &STI);

void decodeMsg(unsigned Val, uint16_t &MsgId, uint16_t &OpId,
               uint16_t &StreamId, const MCSubtargetInfo &STI);

LLVM_READNONE
uint64_t encodeMsg(uint64_t MsgId, uint64_t OpId, uint64_t StreamId);

/// Returns true if the message does not use the m0 operand.
bool msgDoesNotUseM0(int64_t MsgId, const MCSubtargetInfo &STI);

} // namespace SendMsg

unsigned getInitialPSInputAddr(const Function &F);

bool getHasColorExport(const Function &F);

bool getHasDepthExport(const Function &F);

// Returns the value of the "amdgpu-dynamic-vgpr-block-size" attribute, or 0 if
// the attribute is missing or its value is invalid.
unsigned getDynamicVGPRBlockSize(const Function &F);

LLVM_READNONE
constexpr bool isShader(CallingConv::ID CC) {
  switch (CC) {
  case CallingConv::AMDGPU_VS:
  case CallingConv::AMDGPU_LS:
  case CallingConv::AMDGPU_HS:
  case CallingConv::AMDGPU_ES:
  case CallingConv::AMDGPU_GS:
  case CallingConv::AMDGPU_PS:
  case CallingConv::AMDGPU_CS_Chain:
  case CallingConv::AMDGPU_CS_ChainPreserve:
  case CallingConv::AMDGPU_CS:
    return true;
  default:
    return false;
  }
}

LLVM_READNONE
constexpr bool isGraphics(CallingConv::ID CC) {
  return isShader(CC) || CC == CallingConv::AMDGPU_Gfx ||
         CC == CallingConv::AMDGPU_Gfx_WholeWave;
}

LLVM_READNONE
constexpr bool isCompute(CallingConv::ID CC) {
  return !isGraphics(CC) || CC == CallingConv::AMDGPU_CS;
}

LLVM_READNONE
constexpr bool isEntryFunctionCC(CallingConv::ID CC) {
  switch (CC) {
  case CallingConv::AMDGPU_KERNEL:
  case CallingConv::SPIR_KERNEL:
  case CallingConv::AMDGPU_VS:
  case CallingConv::AMDGPU_GS:
  case CallingConv::AMDGPU_PS:
  case CallingConv::AMDGPU_CS:
  case CallingConv::AMDGPU_ES:
  case CallingConv::AMDGPU_HS:
  case CallingConv::AMDGPU_LS:
    return true;
  default:
    return false;
  }
}

LLVM_READNONE
constexpr bool isChainCC(CallingConv::ID CC) {
  switch (CC) {
  case CallingConv::AMDGPU_CS_Chain:
  case CallingConv::AMDGPU_CS_ChainPreserve:
    return true;
  default:
    return false;
  }
}

// These functions are considered entrypoints into the current module, i.e. they
// are allowed to be called from outside the current module. This is different
// from isEntryFunctionCC, which is only true for functions that are entered by
// the hardware. Module entry points include all entry functions but also
// include functions that can be called from other functions inside or outside
// the current module. Module entry functions are allowed to allocate LDS.
//
// AMDGPU_CS_Chain is intended for externally callable chain functions, so it is
// treated as a module entrypoint. AMDGPU_CS_ChainPreserve is used for internal
// helper functions (e.g. retry helpers), so it is not a module entrypoint.
LLVM_READNONE
constexpr bool isModuleEntryFunctionCC(CallingConv::ID CC) {
  switch (CC) {
  case CallingConv::AMDGPU_Gfx:
  case CallingConv::AMDGPU_CS_Chain:
    return true;
  default:
    return isEntryFunctionCC(CC);
  }
}

LLVM_READNONE
constexpr inline bool isKernel(CallingConv::ID CC) {
  switch (CC) {
  case CallingConv::AMDGPU_KERNEL:
  case CallingConv::SPIR_KERNEL:
    return true;
  default:
    return false;
  }
}

inline bool isKernel(const Function &F) { return isKernel(F.getCallingConv()); }

LLVM_READNONE
constexpr bool canGuaranteeTCO(CallingConv::ID CC) {
  return CC == CallingConv::Fast;
}

/// Return true if we might ever do TCO for calls with this calling convention.
LLVM_READNONE
constexpr bool mayTailCallThisCC(CallingConv::ID CC) {
  switch (CC) {
  case CallingConv::C:
  case CallingConv::AMDGPU_Gfx:
  case CallingConv::AMDGPU_Gfx_WholeWave:
    return true;
  default:
    return canGuaranteeTCO(CC);
  }
}

bool hasMIMG_R128(const MCSubtargetInfo &STI);
bool hasA16(const MCSubtargetInfo &STI);
bool hasG16(const MCSubtargetInfo &STI);
bool hasPackedD16(const MCSubtargetInfo &STI);
bool hasGDS(const MCSubtargetInfo &STI);
unsigned getNSAMaxSize(const MCSubtargetInfo &STI, bool HasSampler = false);
unsigned getMaxNumUserSGPRs(const MCSubtargetInfo &STI);

bool isSI(const MCSubtargetInfo &STI);
bool isCI(const MCSubtargetInfo &STI);
bool isVI(const MCSubtargetInfo &STI);
bool isGFX9(const MCSubtargetInfo &STI);
bool isGFX9_GFX10(const MCSubtargetInfo &STI);
bool isGFX9_GFX10_GFX11(const MCSubtargetInfo &STI);
bool isGFX8_GFX9_GFX10(const MCSubtargetInfo &STI);
bool isGFX8Plus(const MCSubtargetInfo &STI);
bool isGFX9Plus(const MCSubtargetInfo &STI);
bool isNotGFX9Plus(const MCSubtargetInfo &STI);
bool isGFX10(const MCSubtargetInfo &STI);
bool isGFX10_GFX11(const MCSubtargetInfo &STI);
bool isGFX10Plus(const MCSubtargetInfo &STI);
bool isNotGFX10Plus(const MCSubtargetInfo &STI);
bool isGFX10Before1030(const MCSubtargetInfo &STI);
bool isGFX11(const MCSubtargetInfo &STI);
bool isGFX11Plus(const MCSubtargetInfo &STI);
bool isGFX12(const MCSubtargetInfo &STI);
bool isGFX12Plus(const MCSubtargetInfo &STI);
bool isGFX1250(const MCSubtargetInfo &STI);
bool isGFX1250Plus(const MCSubtargetInfo &STI);
bool isGFX13(const MCSubtargetInfo &STI);
bool isGFX13Plus(const MCSubtargetInfo &STI);

/// \returns true if a work-group's waves run on all four SIMD32s (one
/// contiguous LDS) and not just on two.
bool isFullSIMDMode(const MCSubtargetInfo &STI);

bool supportsWGP(const MCSubtargetInfo &STI);
bool isNotGFX12Plus(const MCSubtargetInfo &STI);
bool isNotGFX11Plus(const MCSubtargetInfo &STI);
bool isGCN3Encoding(const MCSubtargetInfo &STI);
bool isGFX10_BEncoding(const MCSubtargetInfo &STI);
bool hasGFX10_3Insts(const MCSubtargetInfo &STI);
bool isGFX10_3_GFX11(const MCSubtargetInfo &STI);
bool isGFX90A(const MCSubtargetInfo &STI);
bool isGFX940(const MCSubtargetInfo &STI);
bool hasArchitectedFlatScratch(const MCSubtargetInfo &STI);
bool hasMAIInsts(const MCSubtargetInfo &STI);
bool hasPopsExitingWaveID(const MCSubtargetInfo &STI);

/// \returns true if the src_private_base and src_private_limit aperture
/// registers are available on \p STI. Targets with globally addressable
/// scratch have no private aperture and expose src_flat_scratch_base instead.
bool hasPrivateApertureRegs(const MCSubtargetInfo &STI);

bool hasVOPD(const MCSubtargetInfo &STI);
bool hasDPPSrc1SGPR(const MCSubtargetInfo &STI);

int getTotalNumVGPRs(bool has90AInsts, int32_t ArgNumAGPR, int32_t ArgNumVGPR);
unsigned hasKernargPreload(const MCSubtargetInfo &STI);
bool hasSMRDSignedImmOffset(const MCSubtargetInfo &ST);

/// Is Reg - scalar register
bool isSGPR(MCRegister Reg, const MCRegisterInfo *TRI);

/// \returns true if \p Reg is an indexed-resource index register, i.e. either a
/// 32-bit SGPR (uniform-indexed form) or a Lo256 VGPR (per-lane indexed form).
bool isRsrcIndexReg(MCRegister Reg, const MCRegisterInfo &MRI);

/// \returns if \p Reg occupies the high 16-bits of a 32-bit register.
bool isHi16Reg(MCRegister Reg, const MCRegisterInfo &MRI);

/// If \p Reg is a pseudo reg, return the correct hardware register given
/// \p STI otherwise return \p Reg.
MCRegister getMCReg(MCRegister Reg, const MCSubtargetInfo &STI);

/// Convert hardware register \p Reg to a pseudo register
LLVM_READNONE
MCRegister mc2PseudoReg(MCRegister Reg);

LLVM_READNONE
bool isInlineValue(MCRegister Reg);

/// Is this an AMDGPU specific source operand? These include registers,
/// inline constants, literals and mandatory literals (KImm).
constexpr bool isSISrcOperand(const MCOperandInfo &OpInfo) {
  return OpInfo.OperandType >= AMDGPU::OPERAND_SRC_FIRST &&
         OpInfo.OperandType <= AMDGPU::OPERAND_SRC_LAST;
}

inline bool isSISrcOperand(const MCInstrDesc &Desc, unsigned OpNo) {
  return isSISrcOperand(Desc.operands()[OpNo]);
}

/// Is this a scalar (i.e. not packed) bf16 source operand?
constexpr bool isBF16SrcOperand(const MCOperandInfo &OpInfo) {
  return OpInfo.OperandType == AMDGPU::OPERAND_REG_IMM_BF16 ||
         OpInfo.OperandType == AMDGPU::OPERAND_REG_INLINE_C_BF16;
}

/// Is this a KImm operand?
bool isKImmOperand(const MCInstrDesc &Desc, unsigned OpNo);

/// Is this floating-point operand?
bool isSISrcFPOperand(const MCInstrDesc &Desc, unsigned OpNo);

/// Get the size in bits of a register from the register class \p RC.
unsigned getRegBitWidth(unsigned RCID);

/// Get the size in bits of a register from the register class \p RC.
unsigned getRegBitWidth(const MCRegisterClass &RC);

LLVM_READNONE
inline unsigned getOperandSize(const MCOperandInfo &OpInfo) {
  switch (OpInfo.OperandType) {
  case AMDGPU::OPERAND_REG_IMM_INT32:
  case AMDGPU::OPERAND_REG_IMM_FP32:
  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_V2INT32:
  case AMDGPU::OPERAND_REG_IMM_V2FP32:
  case AMDGPU::OPERAND_KIMM32:
  case AMDGPU::OPERAND_KIMM16: // mandatory literal is always size 4
  case AMDGPU::OPERAND_INLINE_SPLIT_BARRIER_INT32:
    return 4;

  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:
  case AMDGPU::OPERAND_REG_IMM_V2FP64:
  case AMDGPU::OPERAND_REG_IMM_V2INT64:
  case AMDGPU::OPERAND_KIMM64:
    return 8;

  case AMDGPU::OPERAND_REG_IMM_INT16:
  case AMDGPU::OPERAND_REG_IMM_BF16:
  case AMDGPU::OPERAND_REG_IMM_FP16:
  case AMDGPU::OPERAND_REG_IMM_NOINLINE_FP16:
  case AMDGPU::OPERAND_REG_INLINE_C_INT16:
  case AMDGPU::OPERAND_REG_INLINE_C_BF16:
  case AMDGPU::OPERAND_REG_INLINE_C_FP16:
  case AMDGPU::OPERAND_REG_INLINE_C_V2INT16:
  case AMDGPU::OPERAND_REG_INLINE_C_V2BF16:
  case AMDGPU::OPERAND_REG_INLINE_C_V2FP16:
  case AMDGPU::OPERAND_REG_IMM_V2INT16:
  case AMDGPU::OPERAND_REG_IMM_V2BF16:
  case AMDGPU::OPERAND_REG_IMM_V2FP16:
  case AMDGPU::OPERAND_REG_IMM_V2FP16_SPLAT:
  case AMDGPU::OPERAND_REG_IMM_NOINLINE_V2FP16:
    return 2;

  default:
    llvm_unreachable("unhandled operand type");
  }
}

LLVM_READNONE
inline unsigned getOperandSize(const MCInstrDesc &Desc, unsigned OpNo) {
  return getOperandSize(Desc.operands()[OpNo]);
}

/// Is this literal inlinable, and not one of the values intended for floating
/// point values.
LLVM_READNONE
inline bool isInlinableIntLiteral(int64_t Literal) {
  return Literal >= -16 && Literal <= 64;
}

/// Is this literal inlinable
LLVM_READNONE
bool isInlinableLiteral64(int64_t Literal, bool HasInv2Pi);

LLVM_READNONE
bool isInlinableLiteral32(int32_t Literal, bool HasInv2Pi);

LLVM_READNONE
bool isInlinableLiteralBF16(int16_t Literal, bool HasInv2Pi);

LLVM_READNONE
bool isInlinableLiteralFP16(int16_t Literal, bool HasInv2Pi);

LLVM_READNONE
bool isInlinableLiteralI16(int32_t Literal, bool HasInv2Pi);

LLVM_READNONE
std::optional<unsigned> getInlineEncodingV2I16(uint32_t Literal);

LLVM_READNONE
std::optional<unsigned> getInlineEncodingV2BF16(uint32_t Literal);

LLVM_READNONE
std::optional<unsigned> getInlineEncodingV2F16(uint32_t Literal);

LLVM_READNONE
std::optional<unsigned> getPKFMACF16InlineEncoding(uint32_t Literal,
                                                   bool IsGFX11Plus);

LLVM_READNONE
bool isInlinableLiteralV216(uint32_t Literal, uint8_t OpType);

LLVM_READNONE
bool isInlinableLiteralV2I16(uint32_t Literal);

LLVM_READNONE
bool isInlinableLiteralV2BF16(uint32_t Literal);

LLVM_READNONE
bool isInlinableLiteralV2F16(uint32_t Literal);

LLVM_READNONE
bool isPKFMACF16InlineConstant(uint32_t Literal, bool IsGFX11Plus);

LLVM_READNONE
bool isValid32BitLiteral(uint64_t Val, bool IsFP64);

LLVM_READNONE
int64_t encode32BitLiteral(int64_t Imm, OperandType Type, bool IsLit);

bool isArgPassedInSGPR(const Argument *Arg);

bool isArgPassedInSGPR(const CallBase *CB, unsigned ArgNo);

/// The opcode is a packed fp32 instruction which only reads low 32 bits of
/// a scalar operand and propagates it to high channel.
LLVM_READONLY bool isPackedSingleSGPRFP32Inst(unsigned Opc);

/// The opcode is a packed 64-bit instruction which only reads low 64 bits of
/// a scalar operand and propagates it to high channel.
LLVM_READONLY bool isPackedSingleSGPR64BitInst(unsigned Opc);

/// Packed instructions that read a single SGPR for SGPR operands, except for
/// 64-bit elements which read two SGPRs.
LLVM_READONLY bool isSingleSGPRReadInst(unsigned Opc);

LLVM_READONLY
bool isLegalSMRDEncodedUnsignedOffset(const MCSubtargetInfo &ST,
                                      int64_t EncodedOffset);

LLVM_READONLY
bool isLegalSMRDEncodedSignedOffset(const MCSubtargetInfo &ST,
                                    int64_t EncodedOffset, bool IsBuffer);

/// Convert \p ByteOffset to dwords if the subtarget uses dword SMRD immediate
/// offsets.
uint64_t convertSMRDOffsetUnits(const MCSubtargetInfo &ST, uint64_t ByteOffset);

/// \returns The encoding that will be used for \p ByteOffset in the
/// SMRD offset field, or std::nullopt if it won't fit. On GFX9 and GFX10
/// S_LOAD instructions have a signed offset, on other subtargets it is
/// unsigned. S_BUFFER has an unsigned offset for all subtargets.
std::optional<int64_t> getSMRDEncodedOffset(const MCSubtargetInfo &ST,
                                            int64_t ByteOffset, bool IsBuffer,
                                            bool HasSOffset = false);

/// \return The encoding that can be used for a 32-bit literal offset in an SMRD
/// instruction. This is only useful on CI.s
std::optional<int64_t> getSMRDEncodedLiteralOffset32(const MCSubtargetInfo &ST,
                                                     int64_t ByteOffset);

/// For pre-GFX12 FLAT instructions the offset must be positive;
/// MSB is ignored and forced to zero.
///
/// \return The number of bits available for the signed offset field in flat
/// instructions. Note that some forms of the instruction disallow negative
/// offsets.
unsigned getNumFlatOffsetBits(const MCSubtargetInfo &ST);

LLVM_READNONE
inline bool isLegalDPALU_DPPControl(const MCSubtargetInfo &ST, unsigned DC) {
  if (isGFX12(ST))
    return DC >= DPP::ROW_SHARE_FIRST && DC <= DPP::ROW_SHARE_LAST;
  if (isGFX90A(ST))
    return DC >= DPP::ROW_NEWBCAST_FIRST && DC <= DPP::ROW_NEWBCAST_LAST;
  return false;
}

/// \returns true if an instruction is a DP ALU DPP without any 64-bit operands.
bool isDPALU_DPP32BitOpc(unsigned Opc);

/// \returns true if an instruction is a DP ALU DPP.
bool isDPALU_DPP(const MCInstrDesc &OpDesc, const MCInstrInfo &MII,
                 const MCSubtargetInfo &ST);

/// \returns true if the intrinsic is divergent
bool isIntrinsicSourceOfDivergence(unsigned IntrID);

/// \returns true if the intrinsic is uniform
bool isIntrinsicAlwaysUniform(unsigned IntrID);

/// \returns true if the intrinsic executes an LDSDMA operation
bool isLDSDMAIntrinsic(unsigned IntrID);

/// \returns a register class for the physical register \p Reg if it is a VGPR
/// or nullptr otherwise.
const MCRegisterClass *getVGPRPhysRegClass(MCRegister Reg,
                                           const MCRegisterInfo &MRI);

/// \returns the MODE bits which have to be set by the S_SET_VGPR_MSB for the
/// physical register \p Reg.
unsigned getVGPREncodingMSBs(MCRegister Reg, const MCRegisterInfo &MRI);

/// If \p Reg is a low VGPR return a corresponding high VGPR with \p MSBs set.
MCRegister getVGPRWithMSBs(MCRegister Reg, unsigned MSBs,
                           const MCRegisterInfo &MRI);

/// \returns VGPR MSBs encoded in a S_SETREG_IMM32_B32 \p MI if it sets
/// it. If \p HasSetregVGPRMSBFixup is true then size of the ID_MODE mask is
/// ignored.
std::optional<unsigned> convertSetRegImmToVgprMSBs(const MachineInstr &MI,
                                                   bool HasSetregVGPRMSBFixup);

/// \returns VGPR MSBs encoded in a S_SETREG_IMM32_B32 \p MI if it sets
/// it. If \p HasSetregVGPRMSBFixup is true then size of the ID_MODE mask is
/// ignored.
std::optional<unsigned> convertSetRegImmToVgprMSBs(const MCInst &MI,
                                                   bool HasSetregVGPRMSBFixup);

// Returns a table for the opcode with a given \p Desc to map the VGPR MSB
// set by the S_SET_VGPR_MSB to one of 4 sources. In case of VOPD returns 2
// maps, one for X and one for Y component.
std::pair<const AMDGPU::OpName *, const AMDGPU::OpName *>
getVGPRLoweringOperandTables(const MCInstrDesc &Desc);

/// \returns true if a memory instruction supports scale_offset modifier.
bool supportsScaleOffset(const MCInstrInfo &MII, unsigned Opcode);

class ClusterDimsAttr {
public:
  enum class Kind { Unknown, NoCluster, VariableDims, FixedDims };

  ClusterDimsAttr() = default;

  Kind getKind() const { return AttrKind; }

  bool isUnknown() const { return getKind() == Kind::Unknown; }

  bool isNoCluster() const { return getKind() == Kind::NoCluster; }

  bool isFixedDims() const { return getKind() == Kind::FixedDims; }

  bool isVariableDims() const { return getKind() == Kind::VariableDims; }

  void setUnknown() { *this = ClusterDimsAttr(Kind::Unknown); }

  void setVariableDims() { *this = ClusterDimsAttr(Kind::VariableDims); }

  /// \returns the dims stored. Note that this function can only be called if
  /// the kind is \p Fixed.
  const std::array<unsigned, 3> &getDims() const;

  bool operator==(const ClusterDimsAttr &RHS) const {
    return AttrKind == RHS.AttrKind && Dims == RHS.Dims;
  }

  std::string to_string() const;

  static ClusterDimsAttr get(const Function &F);

private:
  enum Encoding { EncoNoCluster = 0, EncoVariableDims = 1024 };

  ClusterDimsAttr(Kind AttrKind) : AttrKind(AttrKind) {}

  std::array<unsigned, 3> Dims = {0, 0, 0};

  Kind AttrKind = Kind::Unknown;
};

/// Evaluate the constant-folded result of v_rcp for \p Val, accounting for
/// the hardware's denormal flushing on f32/f64 and its approximate rounding.
/// Returns std::nullopt if the hardware result is not guaranteed to match the
/// exact reciprocal.
std::optional<APFloat> evaluateRcp(const APFloat &Val);

} // namespace AMDGPU

raw_ostream &operator<<(raw_ostream &OS, const AMDGPU::TargetIDSetting S);

} // end namespace llvm

#endif // LLVM_LIB_TARGET_AMDGPU_UTILS_AMDGPUBASEINFO_H