//===-- NeonEmitter.cpp - Generate arm_neon.h for use with clang ----------===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//
//
// This tablegen backend is responsible for emitting arm_neon.h, which includes
// a declaration and definition of each function specified by the ARM NEON
// compiler interface.  See ARM document DUI0348B.
//
// Each NEON instruction is implemented in terms of 1 or more functions which
// are suffixed with the element type of the input vectors.  Functions may be
// implemented in terms of generic vector operations such as +, *, -, etc. or
// by calling a __builtin_-prefixed function which will be handled by clang's
// CodeGen library.
//
// Additional validation code can be generated by this file when runHeader() is
// called, rather than the normal run() entry point.
//
// See also the documentation in include/clang/Basic/arm_neon.td.
//
//===----------------------------------------------------------------------===//

#include "TableGenBackends.h"
#include "llvm/ADT/ArrayRef.h"
#include "llvm/ADT/DenseMap.h"
#include "llvm/ADT/STLExtras.h"
#include "llvm/ADT/SmallVector.h"
#include "llvm/ADT/StringExtras.h"
#include "llvm/ADT/StringRef.h"
#include "llvm/Support/Casting.h"
#include "llvm/Support/ErrorHandling.h"
#include "llvm/Support/raw_ostream.h"
#include "llvm/TableGen/AArch64ImmCheck.h"
#include "llvm/TableGen/Error.h"
#include "llvm/TableGen/Record.h"
#include "llvm/TableGen/SetTheory.h"
#include <algorithm>
#include <cassert>
#include <cctype>
#include <cstddef>
#include <cstdint>
#include <deque>
#include <map>
#include <optional>
#include <set>
#include <sstream>
#include <string>
#include <unordered_map>
#include <utility>
#include <vector>

using namespace llvm;

namespace {

// While globals are generally bad, this one allows us to perform assertions
// liberally and somehow still trace them back to the def they indirectly
// came from.
static const Record *CurrentRecord = nullptr;
static void assert_with_loc(bool Assertion, const std::string &Str) {
  if (!Assertion) {
    if (CurrentRecord)
      PrintFatalError(CurrentRecord->getLoc(), Str);
    else
      PrintFatalError(Str);
  }
}

enum ClassKind {
  ClassNone,
  ClassI,     // generic integer instruction, e.g., "i8" suffix
  ClassS,     // signed/unsigned/poly, e.g., "s8", "u8" or "p8" suffix
  ClassW,     // width-specific instruction, e.g., "8" suffix
  ClassB,     // bitcast arguments with enum argument to specify type
  ClassL,     // Logical instructions which are op instructions
              // but we need to not emit any suffix for in our
              // tests.
  ClassNoTest // Instructions which we do not test since they are
              // not TRUE instructions.
};

/// NeonTypeFlags - Flags to identify the types for overloaded Neon
/// builtins.  These must be kept in sync with the flags in
/// include/clang/Basic/TargetBuiltins.h.
namespace NeonTypeFlags {

enum { EltTypeMask = 0xf, UnsignedFlag = 0x10, QuadFlag = 0x20 };

enum EltType {
  Int8,
  Int16,
  Int32,
  Int64,
  Poly8,
  Poly16,
  Poly64,
  Poly128,
  Float16,
  Float32,
  Float64,
  BFloat16,
  MFloat8 // Not used by Sema or CodeGen in Clang
};

} // end namespace NeonTypeFlags

class NeonEmitter;

//===----------------------------------------------------------------------===//
// TypeSpec
//===----------------------------------------------------------------------===//

/// A TypeSpec is just a simple wrapper around a string, but gets its own type
/// for strong typing purposes.
///
/// A TypeSpec can be used to create a type.
class TypeSpec : public std::string {
public:
  static std::vector<TypeSpec> fromTypeSpecs(StringRef Str) {
    std::vector<TypeSpec> Ret;
    TypeSpec Acc;
    for (char I : Str.str()) {
      if (islower(I)) {
        Acc.push_back(I);
        Ret.push_back(TypeSpec(Acc));
        Acc.clear();
      } else {
        Acc.push_back(I);
      }
    }
    return Ret;
  }
};

//===----------------------------------------------------------------------===//
// Type
//===----------------------------------------------------------------------===//

/// A Type. Not much more to say here.
class Type {
private:
  TypeSpec TS;

  enum TypeKind { Void, Float, SInt, UInt, Poly, BFloat16, MFloat8 };
  TypeKind Kind;
  bool Immediate, Constant, Pointer;
  // ScalarForMangling and NoManglingQ are really not suited to live here as
  // they are not related to the type. But they live in the TypeSpec (not the
  // prototype), so this is really the only place to store them.
  bool ScalarForMangling, NoManglingQ;
  unsigned Bitwidth, ElementBitwidth, NumVectors;

public:
  Type()
      : Kind(Void), Immediate(false), Constant(false),
        Pointer(false), ScalarForMangling(false), NoManglingQ(false),
        Bitwidth(0), ElementBitwidth(0), NumVectors(0) {}

  Type(TypeSpec TS, StringRef CharMods)
      : TS(std::move(TS)), Kind(Void), Immediate(false),
        Constant(false), Pointer(false), ScalarForMangling(false),
        NoManglingQ(false), Bitwidth(0), ElementBitwidth(0), NumVectors(0) {
    applyModifiers(CharMods);
  }

  /// Returns a type representing "void".
  static Type getVoid() { return Type(); }

  bool operator==(const Type &Other) const { return str() == Other.str(); }
  bool operator!=(const Type &Other) const { return !operator==(Other); }

  //
  // Query functions
  //
  bool isScalarForMangling() const { return ScalarForMangling; }
  bool noManglingQ() const { return NoManglingQ; }

  bool isPointer() const { return Pointer; }
  bool isValue() const { return !isVoid() && !isPointer(); }
  bool isScalar() const { return isValue() && NumVectors == 0; }
  bool isVector() const { return isValue() && NumVectors > 0; }
  bool isConstPointer() const { return Constant; }
  bool isFloating() const { return Kind == Float; }
  bool isInteger() const { return Kind == SInt || Kind == UInt; }
  bool isPoly() const { return Kind == Poly; }
  bool isSigned() const { return Kind == SInt; }
  bool isImmediate() const { return Immediate; }
  bool isFloat() const { return isFloating() && ElementBitwidth == 32; }
  bool isDouble() const { return isFloating() && ElementBitwidth == 64; }
  bool isHalf() const { return isFloating() && ElementBitwidth == 16; }
  bool isChar() const { return ElementBitwidth == 8; }
  bool isShort() const { return isInteger() && ElementBitwidth == 16; }
  bool isInt() const { return isInteger() && ElementBitwidth == 32; }
  bool isLong() const { return isInteger() && ElementBitwidth == 64; }
  bool isVoid() const { return Kind == Void; }
  bool isBFloat16() const { return Kind == BFloat16; }
  bool isMFloat8() const { return Kind == MFloat8; }
  unsigned getNumElements() const { return Bitwidth / ElementBitwidth; }
  unsigned getSizeInBits() const { return Bitwidth; }
  unsigned getElementSizeInBits() const { return ElementBitwidth; }
  unsigned getNumVectors() const { return NumVectors; }

  //
  // Mutator functions
  //
  void makeUnsigned() {
    assert(!isVoid() && "not a potentially signed type");
    Kind = UInt;
  }
  void makeSigned() {
    assert(!isVoid() && "not a potentially signed type");
    Kind = SInt;
  }

  void makeInteger(unsigned ElemWidth, bool Sign) {
    assert(!isVoid() && "converting void to int probably not useful");
    Kind = Sign ? SInt : UInt;
    Immediate = false;
    ElementBitwidth = ElemWidth;
  }

  void makeImmediate(unsigned ElemWidth) {
    Kind = SInt;
    Immediate = true;
    ElementBitwidth = ElemWidth;
  }

  void makeScalar() {
    Bitwidth = ElementBitwidth;
    NumVectors = 0;
  }

  void makeOneVector() {
    assert(isVector());
    NumVectors = 1;
  }

  void make32BitElement() {
    assert_with_loc(Bitwidth > 32, "Not enough bits to make it 32!");
    ElementBitwidth = 32;
  }

  void doubleLanes() {
    assert_with_loc(Bitwidth != 128, "Can't get bigger than 128!");
    Bitwidth = 128;
  }

  void halveLanes() {
    assert_with_loc(Bitwidth != 64, "Can't get smaller than 64!");
    Bitwidth = 64;
  }

  /// Return the C string representation of a type, which is the typename
  /// defined in stdint.h or arm_neon.h.
  std::string str() const;

  /// Return the string representation of a type, which is an encoded
  /// string for passing to the BUILTIN() macro in Builtins.def.
  std::string builtin_str() const;

  /// Return the value in NeonTypeFlags for this type.
  unsigned getNeonEnum() const;

  /// Parse a type from a stdint.h or arm_neon.h typedef name,
  /// for example uint32x2_t or int64_t.
  static Type fromTypedefName(StringRef Name);

private:
  /// Creates the type based on the typespec string in TS.
  /// Sets "Quad" to true if the "Q" or "H" modifiers were
  /// seen. This is needed by applyModifier as some modifiers
  /// only take effect if the type size was changed by "Q" or "H".
  void applyTypespec(bool &Quad);
  /// Applies prototype modifiers to the type.
  void applyModifiers(StringRef Mods);
};

//===----------------------------------------------------------------------===//
// Variable
//===----------------------------------------------------------------------===//

/// A variable is a simple class that just has a type and a name.
class Variable {
  Type T;
  std::string N;

public:
  Variable() : T(Type::getVoid()) {}
  Variable(Type T, std::string N) : T(std::move(T)), N(std::move(N)) {}

  Type getType() const { return T; }
  std::string getName() const { return "__" + N; }
};

//===----------------------------------------------------------------------===//
// Intrinsic
//===----------------------------------------------------------------------===//

/// The main grunt class. This represents an instantiation of an intrinsic with
/// a particular typespec and prototype.
class Intrinsic {
  /// The Record this intrinsic was created from.
  const Record *R;
  /// The unmangled name.
  std::string Name;
  /// The input and output typespecs. InTS == OutTS except when
  /// CartesianProductWith is non-empty - this is the case for vreinterpret.
  TypeSpec OutTS, InTS;
  /// The base class kind. Most intrinsics use ClassS, which has full type
  /// info for integers (s32/u32). Some use ClassI, which doesn't care about
  /// signedness (i32), while some (ClassB) have no type at all, only a width
  /// (32).
  ClassKind CK;
  /// The list of DAGs for the body. May be empty, in which case we should
  /// emit a builtin call.
  const ListInit *Body;
  /// The architectural ifdef guard.
  std::string ArchGuard;
  /// The architectural target() guard.
  std::string TargetGuard;
  /// Set if the Unavailable bit is 1. This means we don't generate a body,
  /// just an "unavailable" attribute on a declaration.
  bool IsUnavailable;
  /// Is this intrinsic safe for big-endian? or does it need its arguments
  /// reversing?
  bool BigEndianSafe;

  /// The types of return value [0] and parameters [1..].
  std::vector<Type> Types;

  SmallVector<ImmCheck, 2> ImmChecks;
  /// The index of the key type passed to CGBuiltin.cpp for polymorphic calls.
  int PolymorphicKeyType;
  /// The local variables defined.
  std::map<std::string, Variable, std::less<>> Variables;
  /// NeededEarly - set if any other intrinsic depends on this intrinsic.
  bool NeededEarly;
  /// UseMacro - set if we should implement using a macro or unset for a
  ///            function.
  bool UseMacro;
  /// The set of intrinsics that this intrinsic uses/requires.
  std::set<Intrinsic *> Dependencies;
  /// The "base type", which is Type('d', OutTS). InBaseType is only
  /// different if CartesianProductWith is non-empty (for vreinterpret).
  Type BaseType, InBaseType;
  /// The return variable.
  Variable RetVar;
  /// A postfix to apply to every variable. Defaults to "".
  std::string VariablePostfix;

  NeonEmitter &Emitter;
  std::stringstream OS;

  bool isBigEndianSafe() const {
    if (BigEndianSafe)
      return true;

    for (const auto &T : Types){
      if (T.isVector() && T.getNumElements() > 1)
        return false;
    }
    return true;
  }

public:
  Intrinsic(const Record *R, StringRef Name, StringRef Proto, TypeSpec OutTS,
            TypeSpec InTS, ClassKind CK, const ListInit *Body,
            NeonEmitter &Emitter, StringRef ArchGuard, StringRef TargetGuard,
            bool IsUnavailable, bool BigEndianSafe)
      : R(R), Name(Name.str()), OutTS(OutTS), InTS(InTS), CK(CK), Body(Body),
        ArchGuard(ArchGuard.str()), TargetGuard(TargetGuard.str()),
        IsUnavailable(IsUnavailable), BigEndianSafe(BigEndianSafe),
        PolymorphicKeyType(0), NeededEarly(false), UseMacro(false),
        BaseType(OutTS, "."), InBaseType(InTS, "."), Emitter(Emitter) {
    // Modify the TypeSpec per-argument to get a concrete Type, and create
    // known variables for each.
    // Types[0] is the return value.
    unsigned Pos = 0;
    Types.emplace_back(OutTS, getNextModifiers(Proto, Pos));
    StringRef Mods = getNextModifiers(Proto, Pos);
    while (!Mods.empty()) {
      Types.emplace_back(InTS, Mods);
      if (Mods.contains('!'))
        PolymorphicKeyType = Types.size() - 1;

      Mods = getNextModifiers(Proto, Pos);
    }

    for (const auto &Type : Types) {
      // If this builtin takes an immediate argument, we need to #define it rather
      // than use a standard declaration, so that SemaChecking can range check
      // the immediate passed by the user.

      // Pointer arguments need to use macros to avoid hiding aligned attributes
      // from the pointer type.

      // It is not permitted to pass or return an __fp16 by value, so intrinsics
      // taking a scalar float16_t must be implemented as macros.
      if (Type.isImmediate() || Type.isPointer() ||
          (Type.isScalar() && Type.isHalf()))
        UseMacro = true;
    }

    int ArgIdx, Kind, TypeArgIdx;
    for (const Record *I : R->getValueAsListOfDefs("ImmChecks")) {
      unsigned EltSizeInBits = 0, VecSizeInBits = 0;

      ArgIdx = I->getValueAsInt("ImmArgIdx");
      TypeArgIdx = I->getValueAsInt("TypeContextArgIdx");
      Kind = I->getValueAsDef("Kind")->getValueAsInt("Value");

      assert((ArgIdx >= 0 && Kind >= 0) &&
             "ImmArgIdx and Kind must be nonnegative");

      if (TypeArgIdx >= 0) {
        Type ContextType = getParamType(TypeArgIdx);

        // Element size cannot be set for intrinscs that map to polymorphic
        // builtins.
        if (CK != ClassB)
          EltSizeInBits = ContextType.getElementSizeInBits();

        VecSizeInBits = ContextType.getSizeInBits();
      }

      ImmChecks.emplace_back(ArgIdx, Kind, EltSizeInBits, VecSizeInBits);
    }
    sort(ImmChecks.begin(), ImmChecks.end(),
         [](const ImmCheck &a, const ImmCheck &b) {
           return a.getImmArgIdx() < b.getImmArgIdx();
         }); // Sort for comparison with other intrinsics which map to the
             // same builtin
  }

  /// Get the Record that this intrinsic is based off.
  const Record *getRecord() const { return R; }
  /// Get the set of Intrinsics that this intrinsic calls.
  /// this is the set of immediate dependencies, NOT the
  /// transitive closure.
  const std::set<Intrinsic *> &getDependencies() const { return Dependencies; }
  /// Get the architectural guard string (#ifdef).
  std::string getArchGuard() const { return ArchGuard; }
  std::string getTargetGuard() const { return TargetGuard; }
  ArrayRef<ImmCheck> getImmChecks() const { return ImmChecks; }
  /// Get the non-mangled name.
  std::string getName() const { return Name; }

  /// Return true if the intrinsic takes an immediate operand.
  bool hasImmediate() const {
    return any_of(Types, [](const Type &T) { return T.isImmediate(); });
  }

  // Return if the supplied argument is an immediate
  bool isArgImmediate(unsigned idx) const {
    return Types[idx + 1].isImmediate();
  }

  unsigned getNumParams() const { return Types.size() - 1; }
  Type getReturnType() const { return Types[0]; }
  Type getParamType(unsigned I) const { return Types[I + 1]; }
  Type getBaseType() const { return BaseType; }
  Type getPolymorphicKeyType() const { return Types[PolymorphicKeyType]; }

  /// Return true if the prototype has a scalar argument.
  bool protoHasScalar() const;

  /// Return the index that parameter PIndex will sit at
  /// in a generated function call. This is often just PIndex,
  /// but may not be as things such as multiple-vector operands
  /// and sret parameters need to be taken into account.
  unsigned getGeneratedParamIdx(unsigned PIndex) {
    unsigned Idx = 0;
    if (getReturnType().getNumVectors() > 1)
      // Multiple vectors are passed as sret.
      ++Idx;

    for (unsigned I = 0; I < PIndex; ++I)
      Idx += std::max(1U, getParamType(I).getNumVectors());

    return Idx;
  }

  bool hasBody() const { return Body && !Body->getValues().empty(); }

  void setNeededEarly() { NeededEarly = true; }

  bool operator<(const Intrinsic &Other) const {
    // Sort lexicographically on a three-tuple (ArchGuard, TargetGuard, Name)
    if (ArchGuard != Other.ArchGuard)
      return ArchGuard < Other.ArchGuard;
    if (TargetGuard != Other.TargetGuard)
      return TargetGuard < Other.TargetGuard;
    return Name < Other.Name;
  }

  ClassKind getClassKind(bool UseClassBIfScalar = false) {
    if (UseClassBIfScalar && !protoHasScalar())
      return ClassB;
    return CK;
  }

  /// Return the name, mangled with type information.
  /// If ForceClassS is true, use ClassS (u32/s32) instead
  /// of the intrinsic's own type class.
  std::string getMangledName(bool ForceClassS = false) const;
  /// Return the type code for a builtin function call.
  std::string getInstTypeCode(Type T, ClassKind CK) const;
  /// Return the type string for a BUILTIN() macro in Builtins.def.
  std::string getBuiltinTypeStr();

  /// Generate the intrinsic, returning code.
  std::string generate();
  /// Perform type checking and populate the dependency graph, but
  /// don't generate code yet.
  void indexBody();

private:
  StringRef getNextModifiers(StringRef Proto, unsigned &Pos) const;

  std::string mangleName(std::string Name, ClassKind CK) const;

  void initVariables();
  std::string replaceParamsIn(std::string S);

  void emitBodyAsBuiltinCall();

  void generateImpl(bool ReverseArguments,
                    StringRef NamePrefix, StringRef CallPrefix);
  void emitReturn();
  void emitBody(StringRef CallPrefix);
  void emitShadowedArgs();
  void emitArgumentReversal();
  void emitReturnVarDecl();
  void emitReturnReversal();
  void emitReverseVariable(Variable &Dest, Variable &Src);
  void emitNewLine();
  void emitClosingBrace();
  void emitOpeningBrace();
  void emitPrototype(StringRef NamePrefix);

  class DagEmitter {
    Intrinsic &Intr;
    StringRef CallPrefix;

  public:
    DagEmitter(Intrinsic &Intr, StringRef CallPrefix) :
      Intr(Intr), CallPrefix(CallPrefix) {
    }
    std::pair<Type, std::string> emitDagArg(const Init *Arg,
                                            std::string ArgName);
    std::pair<Type, std::string> emitDagSaveTemp(const DagInit *DI);
    std::pair<Type, std::string> emitDagSplat(const DagInit *DI);
    std::pair<Type, std::string> emitDagDup(const DagInit *DI);
    std::pair<Type, std::string> emitDagDupTyped(const DagInit *DI);
    std::pair<Type, std::string> emitDagShuffle(const DagInit *DI);
    std::pair<Type, std::string> emitDagCast(const DagInit *DI, bool IsBitCast);
    std::pair<Type, std::string> emitDagCall(const DagInit *DI,
                                             bool MatchMangledName);
    std::pair<Type, std::string> emitDagNameReplace(const DagInit *DI);
    std::pair<Type, std::string> emitDagLiteral(const DagInit *DI);
    std::pair<Type, std::string> emitDagOp(const DagInit *DI);
    std::pair<Type, std::string> emitDag(const DagInit *DI);
  };
};

//===----------------------------------------------------------------------===//
// NeonEmitter
//===----------------------------------------------------------------------===//

class NeonEmitter {
  const RecordKeeper &Records;
  DenseMap<const Record *, ClassKind> ClassMap;
  std::map<std::string, std::deque<Intrinsic>, std::less<>> IntrinsicMap;
  unsigned UniqueNumber;

  void createIntrinsic(const Record *R, SmallVectorImpl<Intrinsic *> &Out);
  void genBuiltinsDef(raw_ostream &OS, SmallVectorImpl<Intrinsic *> &Defs);
  void genStreamingSVECompatibleList(raw_ostream &OS,
                                     SmallVectorImpl<Intrinsic *> &Defs);
  void genOverloadTypeCheckCode(raw_ostream &OS,
                                SmallVectorImpl<Intrinsic *> &Defs);
  bool areRangeChecksCompatible(const ArrayRef<ImmCheck> ChecksA,
                                const ArrayRef<ImmCheck> ChecksB);
  void genIntrinsicRangeCheckCode(raw_ostream &OS,
                                  SmallVectorImpl<Intrinsic *> &Defs);

public:
  /// Called by Intrinsic - this attempts to get an intrinsic that takes
  /// the given types as arguments.
  Intrinsic &getIntrinsic(StringRef Name, ArrayRef<Type> Types,
                          std::optional<std::string> MangledName);

  /// Called by Intrinsic - returns a globally-unique number.
  unsigned getUniqueNumber() { return UniqueNumber++; }

  NeonEmitter(const RecordKeeper &R) : Records(R), UniqueNumber(0) {
    const Record *SI = R.getClass("SInst");
    const Record *II = R.getClass("IInst");
    const Record *WI = R.getClass("WInst");
    const Record *SOpI = R.getClass("SOpInst");
    const Record *IOpI = R.getClass("IOpInst");
    const Record *WOpI = R.getClass("WOpInst");
    const Record *LOpI = R.getClass("LOpInst");
    const Record *NoTestOpI = R.getClass("NoTestOpInst");

    ClassMap[SI] = ClassS;
    ClassMap[II] = ClassI;
    ClassMap[WI] = ClassW;
    ClassMap[SOpI] = ClassS;
    ClassMap[IOpI] = ClassI;
    ClassMap[WOpI] = ClassW;
    ClassMap[LOpI] = ClassL;
    ClassMap[NoTestOpI] = ClassNoTest;
  }

  // Emit arm_neon.h.inc
  void run(raw_ostream &o);

  // Emit arm_fp16.h.inc
  void runFP16(raw_ostream &o);

  // Emit arm_bf16.h.inc
  void runBF16(raw_ostream &o);

  void runVectorTypes(raw_ostream &o);

  // Emit all the __builtin prototypes used in arm_neon.h, arm_fp16.h and
  // arm_bf16.h
  void runHeader(raw_ostream &o);
};

} // end anonymous namespace

//===----------------------------------------------------------------------===//
// Type implementation
//===----------------------------------------------------------------------===//

std::string Type::str() const {
  if (isVoid())
    return "void";
  std::string S;

  if (isInteger() && !isSigned())
    S += "u";

  if (isPoly())
    S += "poly";
  else if (isFloating())
    S += "float";
  else if (isBFloat16())
    S += "bfloat";
  else if (isMFloat8())
    S += "mfloat";
  else
    S += "int";

  S += utostr(ElementBitwidth);
  if (isVector())
    S += "x" + utostr(getNumElements());
  if (NumVectors > 1)
    S += "x" + utostr(NumVectors);
  S += "_t";

  if (Constant)
    S += " const";
  if (Pointer)
    S += " *";

  return S;
}

std::string Type::builtin_str() const {
  std::string S;
  if (isVoid())
    return "v";

  if (isPointer()) {
    // All pointers are void pointers.
    S = "v";
    if (isConstPointer())
      S += "C";
    S += "*";
    return S;
  } else if (isInteger())
    switch (ElementBitwidth) {
    case 8: S += "c"; break;
    case 16: S += "s"; break;
    case 32: S += "i"; break;
    case 64: S += "Wi"; break;
    case 128: S += "LLLi"; break;
    default: llvm_unreachable("Unhandled case!");
    }
  else if (isBFloat16()) {
    assert(ElementBitwidth == 16 && "BFloat16 can only be 16 bits");
    S += "y";
  } else if (isMFloat8()) {
    assert(ElementBitwidth == 8 && "MFloat8 can only be 8 bits");
    S += "m";
  } else
    switch (ElementBitwidth) {
    case 16: S += "h"; break;
    case 32: S += "f"; break;
    case 64: S += "d"; break;
    default: llvm_unreachable("Unhandled case!");
    }

  // FIXME: NECESSARY???????????????????????????????????????????????????????????????????????
  if (isChar() && !isPointer() && isSigned())
    // Make chars explicitly signed.
    S = "S" + S;
  else if (isInteger() && !isSigned())
    S = "U" + S;

  // Constant indices are "int", but have the "constant expression" modifier.
  if (isImmediate()) {
    assert(isInteger() && isSigned());
    S = "I" + S;
  }

  if (isScalar())
    return S;

  std::string Ret;
  for (unsigned I = 0; I < NumVectors; ++I)
    Ret += "V" + utostr(getNumElements()) + S;

  return Ret;
}

unsigned Type::getNeonEnum() const {
  unsigned Addend;
  switch (ElementBitwidth) {
  case 8: Addend = 0; break;
  case 16: Addend = 1; break;
  case 32: Addend = 2; break;
  case 64: Addend = 3; break;
  case 128: Addend = 4; break;
  default: llvm_unreachable("Unhandled element bitwidth!");
  }

  unsigned Base = (unsigned)NeonTypeFlags::Int8 + Addend;
  if (isPoly()) {
    // Adjustment needed because Poly32 doesn't exist.
    if (Addend >= 2)
      --Addend;
    Base = (unsigned)NeonTypeFlags::Poly8 + Addend;
  }
  if (isFloating()) {
    assert(Addend != 0 && "Float8 doesn't exist!");
    Base = (unsigned)NeonTypeFlags::Float16 + (Addend - 1);
  }

  if (isBFloat16()) {
    assert(Addend == 1 && "BFloat16 is only 16 bit");
    Base = (unsigned)NeonTypeFlags::BFloat16;
  }

  if (isMFloat8()) {
    Base = (unsigned)NeonTypeFlags::MFloat8;
  }

  if (Bitwidth == 128)
    Base |= (unsigned)NeonTypeFlags::QuadFlag;
  if (isInteger() && !isSigned())
    Base |= (unsigned)NeonTypeFlags::UnsignedFlag;

  return Base;
}

Type Type::fromTypedefName(StringRef Name) {
  Type T;
  T.Kind = SInt;

  if (Name.consume_front("u"))
    T.Kind = UInt;

  if (Name.consume_front("float")) {
    T.Kind = Float;
  } else if (Name.consume_front("poly")) {
    T.Kind = Poly;
  } else if (Name.consume_front("bfloat")) {
    T.Kind = BFloat16;
  } else if (Name.consume_front("mfloat")) {
    T.Kind = MFloat8;
  } else {
    assert(Name.starts_with("int"));
    Name = Name.drop_front(3);
  }

  unsigned I = 0;
  for (I = 0; I < Name.size(); ++I) {
    if (!isdigit(Name[I]))
      break;
  }
  Name.substr(0, I).getAsInteger(10, T.ElementBitwidth);
  Name = Name.drop_front(I);

  T.Bitwidth = T.ElementBitwidth;
  T.NumVectors = 1;

  if (Name.consume_front("x")) {
    unsigned I = 0;
    for (I = 0; I < Name.size(); ++I) {
      if (!isdigit(Name[I]))
        break;
    }
    unsigned NumLanes;
    Name.substr(0, I).getAsInteger(10, NumLanes);
    Name = Name.drop_front(I);
    T.Bitwidth = T.ElementBitwidth * NumLanes;
  } else {
    // Was scalar.
    T.NumVectors = 0;
  }
  if (Name.consume_front("x")) {
    unsigned I = 0;
    for (I = 0; I < Name.size(); ++I) {
      if (!isdigit(Name[I]))
        break;
    }
    Name.substr(0, I).getAsInteger(10, T.NumVectors);
    Name = Name.drop_front(I);
  }

  assert(Name.starts_with("_t") && "Malformed typedef!");
  return T;
}

void Type::applyTypespec(bool &Quad) {
  std::string S = TS;
  ScalarForMangling = false;
  Kind = SInt;
  ElementBitwidth = ~0U;
  NumVectors = 1;

  for (char I : S) {
    switch (I) {
    case 'S':
      ScalarForMangling = true;
      break;
    case 'H':
      NoManglingQ = true;
      Quad = true;
      break;
    case 'Q':
      Quad = true;
      break;
    case 'P':
      Kind = Poly;
      break;
    case 'U':
      Kind = UInt;
      break;
    case 'c':
      ElementBitwidth = 8;
      break;
    case 'h':
      Kind = Float;
      [[fallthrough]];
    case 's':
      ElementBitwidth = 16;
      break;
    case 'f':
      Kind = Float;
      [[fallthrough]];
    case 'i':
      ElementBitwidth = 32;
      break;
    case 'd':
      Kind = Float;
      [[fallthrough]];
    case 'l':
      ElementBitwidth = 64;
      break;
    case 'k':
      ElementBitwidth = 128;
      // Poly doesn't have a 128x1 type.
      if (isPoly())
        NumVectors = 0;
      break;
    case 'b':
      Kind = BFloat16;
      ElementBitwidth = 16;
      break;
    case 'm':
      Kind = MFloat8;
      ElementBitwidth = 8;
      break;
    default:
      llvm_unreachable("Unhandled type code!");
    }
  }
  assert(ElementBitwidth != ~0U && "Bad element bitwidth!");

  Bitwidth = Quad ? 128 : 64;
}

void Type::applyModifiers(StringRef Mods) {
  bool AppliedQuad = false;
  applyTypespec(AppliedQuad);

  for (char Mod : Mods) {
    switch (Mod) {
    case '.':
      break;
    case 'v':
      Kind = Void;
      break;
    case 'S':
      Kind = SInt;
      break;
    case 'U':
      Kind = UInt;
      break;
    case 'B':
      Kind = BFloat16;
      ElementBitwidth = 16;
      break;
    case 'F':
      Kind = Float;
      break;
    case 'P':
      Kind = Poly;
      break;
    case '>':
      assert(ElementBitwidth < 128);
      ElementBitwidth *= 2;
      break;
    case '<':
      assert(ElementBitwidth > 8);
      ElementBitwidth /= 2;
      break;
    case '1':
      NumVectors = 0;
      break;
    case '2':
      NumVectors = 2;
      break;
    case '3':
      NumVectors = 3;
      break;
    case '4':
      NumVectors = 4;
      break;
    case '*':
      Pointer = true;
      break;
    case 'c':
      Constant = true;
      break;
    case 'Q':
      Bitwidth = 128;
      break;
    case 'q':
      Bitwidth = 64;
      break;
    case 'I':
      Kind = SInt;
      ElementBitwidth = Bitwidth = 32;
      NumVectors = 0;
      Immediate = true;
      break;
    case 'p':
      if (isPoly())
        Kind = UInt;
      break;
    case '!':
      // Key type, handled elsewhere.
      break;
    default:
      llvm_unreachable("Unhandled character!");
    }
  }
}

//===----------------------------------------------------------------------===//
// Intrinsic implementation
//===----------------------------------------------------------------------===//

StringRef Intrinsic::getNextModifiers(StringRef Proto, unsigned &Pos) const {
  if (Proto.size() == Pos)
    return StringRef();
  else if (Proto[Pos] != '(')
    return Proto.substr(Pos++, 1);

  size_t Start = Pos + 1;
  size_t End = Proto.find(')', Start);
  assert_with_loc(End != StringRef::npos, "unmatched modifier group paren");
  Pos = End + 1;
  return Proto.slice(Start, End);
}

std::string Intrinsic::getInstTypeCode(Type T, ClassKind CK) const {
  char typeCode = '\0';
  bool printNumber = true;

  if (CK == ClassB && TargetGuard == "neon")
    return "";

  if (T.isBFloat16())
    return "bf16";

  if (T.isMFloat8())
    return "mfp8";

  if (T.isPoly())
    typeCode = 'p';
  else if (T.isInteger())
    typeCode = T.isSigned() ? 's' : 'u';
  else
    typeCode = 'f';

  if (CK == ClassI) {
    switch (typeCode) {
    default:
      break;
    case 's':
    case 'u':
    case 'p':
      typeCode = 'i';
      break;
    }
  }
  if (CK == ClassB && TargetGuard == "neon") {
    typeCode = '\0';
  }

  std::string S;
  if (typeCode != '\0')
    S.push_back(typeCode);
  if (printNumber)
    S += utostr(T.getElementSizeInBits());

  return S;
}

std::string Intrinsic::getBuiltinTypeStr() {
  ClassKind LocalCK = getClassKind(true);
  std::string S;

  Type RetT = getReturnType();
  if ((LocalCK == ClassI || LocalCK == ClassW) && RetT.isScalar() &&
      !RetT.isFloating() && !RetT.isBFloat16() && !RetT.isMFloat8())
    RetT.makeInteger(RetT.getElementSizeInBits(), false);

  // Since the return value must be one type, return a vector type of the
  // appropriate width which we will bitcast.  An exception is made for
  // returning structs of 2, 3, or 4 vectors which are returned in a sret-like
  // fashion, storing them to a pointer arg.
  if (RetT.getNumVectors() > 1) {
    S += "vv*"; // void result with void* first argument
  } else {
    if (RetT.isPoly())
      RetT.makeInteger(RetT.getElementSizeInBits(), false);
    if (!RetT.isScalar() && RetT.isInteger() && !RetT.isSigned())
      RetT.makeSigned();

    if (LocalCK == ClassB && RetT.isValue() && !RetT.isScalar())
      // Cast to vector of 8-bit elements.
      RetT.makeInteger(8, true);

    S += RetT.builtin_str();
  }

  for (unsigned I = 0; I < getNumParams(); ++I) {
    Type T = getParamType(I);
    if (T.isPoly())
      T.makeInteger(T.getElementSizeInBits(), false);

    if (LocalCK == ClassB && !T.isScalar())
      T.makeInteger(8, true);
    // Halves always get converted to 8-bit elements.
    if (T.isHalf() && T.isVector() && !T.isScalarForMangling())
      T.makeInteger(8, true);

    if (LocalCK == ClassI && T.isInteger())
      T.makeSigned();

    if (isArgImmediate(I))
      T.makeImmediate(32);

    S += T.builtin_str();
  }

  // Extra constant integer to hold type class enum for this function, e.g. s8
  if (LocalCK == ClassB)
    S += "i";

  return S;
}

std::string Intrinsic::getMangledName(bool ForceClassS) const {
  // Check if the prototype has a scalar operand with the type of the vector
  // elements.  If not, bitcasting the args will take care of arg checking.
  // The actual signedness etc. will be taken care of with special enums.
  ClassKind LocalCK = CK;
  if (!protoHasScalar())
    LocalCK = ClassB;

  return mangleName(Name, ForceClassS ? ClassS : LocalCK);
}

std::string Intrinsic::mangleName(std::string Name, ClassKind LocalCK) const {
  std::string typeCode = getInstTypeCode(BaseType, LocalCK);
  std::string S = Name;

  if (Name == "vcvt_f16_f32" || Name == "vcvt_f32_f16" ||
      Name == "vcvt_f32_f64" || Name == "vcvt_f64_f32" ||
      Name == "vcvt_f32_bf16")
    return Name;

  if (!typeCode.empty()) {
    // If the name ends with _xN (N = 2,3,4), insert the typeCode before _xN.
    if (Name.size() >= 3 && isdigit(Name.back()) &&
        Name[Name.length() - 2] == 'x' && Name[Name.length() - 3] == '_')
      S.insert(S.length() - 3, "_" + typeCode);
    else
      S += "_" + typeCode;
  }

  if (BaseType != InBaseType) {
    // A reinterpret - out the input base type at the end.
    S += "_" + getInstTypeCode(InBaseType, LocalCK);
  }

  if (LocalCK == ClassB && TargetGuard == "neon")
    S += "_v";

  // Insert a 'q' before the first '_' character so that it ends up before
  // _lane or _n on vector-scalar operations.
  if (BaseType.getSizeInBits() == 128 && !BaseType.noManglingQ()) {
    size_t Pos = S.find('_');
    S.insert(Pos, "q");
  }

  char Suffix = '\0';
  if (BaseType.isScalarForMangling()) {
    switch (BaseType.getElementSizeInBits()) {
    case 8: Suffix = 'b'; break;
    case 16: Suffix = 'h'; break;
    case 32: Suffix = 's'; break;
    case 64: Suffix = 'd'; break;
    default: llvm_unreachable("Bad suffix!");
    }
  }
  if (Suffix != '\0') {
    size_t Pos = S.find('_');
    S.insert(Pos, &Suffix, 1);
  }

  return S;
}

std::string Intrinsic::replaceParamsIn(std::string S) {
  while (S.find('$') != std::string::npos) {
    size_t Pos = S.find('$');
    size_t End = Pos + 1;
    while (isalpha(S[End]))
      ++End;

    std::string VarName = S.substr(Pos + 1, End - Pos - 1);
    assert_with_loc(Variables.find(VarName) != Variables.end(),
                    "Variable not defined!");
    S.replace(Pos, End - Pos, Variables.find(VarName)->second.getName());
  }

  return S;
}

void Intrinsic::initVariables() {
  Variables.clear();

  // Modify the TypeSpec per-argument to get a concrete Type, and create
  // known variables for each.
  for (unsigned I = 1; I < Types.size(); ++I) {
    char NameC = '0' + (I - 1);
    std::string Name = "p";
    Name.push_back(NameC);

    Variables[Name] = Variable(Types[I], Name + VariablePostfix);
  }
  RetVar = Variable(Types[0], "ret" + VariablePostfix);
}

void Intrinsic::emitPrototype(StringRef NamePrefix) {
  if (UseMacro) {
    OS << "#define ";
  } else {
    OS << "__ai ";
    if (TargetGuard != "")
      OS << "__attribute__((target(\"" << TargetGuard << "\"))) ";
    OS << Types[0].str() << " ";
  }

  OS << NamePrefix.str() << mangleName(Name, ClassS) << "(";

  for (unsigned I = 0; I < getNumParams(); ++I) {
    if (I != 0)
      OS << ", ";

    char NameC = '0' + I;
    std::string Name = "p";
    Name.push_back(NameC);
    assert(Variables.find(Name) != Variables.end());
    Variable &V = Variables[Name];

    if (!UseMacro)
      OS << V.getType().str() << " ";
    OS << V.getName();
  }

  OS << ")";
}

void Intrinsic::emitOpeningBrace() {
  if (UseMacro)
    OS << " __extension__ ({";
  else
    OS << " {";
  emitNewLine();
}

void Intrinsic::emitClosingBrace() {
  if (UseMacro)
    OS << "})";
  else
    OS << "}";
}

void Intrinsic::emitNewLine() {
  if (UseMacro)
    OS << " \\\n";
  else
    OS << "\n";
}

void Intrinsic::emitReverseVariable(Variable &Dest, Variable &Src) {
  if (Dest.getType().getNumVectors() > 1) {
    emitNewLine();

    for (unsigned K = 0; K < Dest.getType().getNumVectors(); ++K) {
      OS << "  " << Dest.getName() << ".val[" << K << "] = "
         << "__builtin_shufflevector("
         << Src.getName() << ".val[" << K << "], "
         << Src.getName() << ".val[" << K << "]";
      for (int J = Dest.getType().getNumElements() - 1; J >= 0; --J)
        OS << ", " << J;
      OS << ");";
      emitNewLine();
    }
  } else {
    OS << "  " << Dest.getName()
       << " = __builtin_shufflevector(" << Src.getName() << ", " << Src.getName();
    for (int J = Dest.getType().getNumElements() - 1; J >= 0; --J)
      OS << ", " << J;
    OS << ");";
    emitNewLine();
  }
}

void Intrinsic::emitArgumentReversal() {
  if (isBigEndianSafe())
    return;

  // Reverse all vector arguments.
  for (unsigned I = 0; I < getNumParams(); ++I) {
    std::string Name = "p" + utostr(I);
    std::string NewName = "rev" + utostr(I);

    Variable &V = Variables[Name];
    Variable NewV(V.getType(), NewName + VariablePostfix);

    if (!NewV.getType().isVector() || NewV.getType().getNumElements() == 1)
      continue;

    OS << "  " << NewV.getType().str() << " " << NewV.getName() << ";";
    emitReverseVariable(NewV, V);
    V = NewV;
  }
}

void Intrinsic::emitReturnVarDecl() {
  assert(RetVar.getType() == Types[0]);
  // Create a return variable, if we're not void.
  if (!RetVar.getType().isVoid()) {
    OS << "  " << RetVar.getType().str() << " " << RetVar.getName() << ";";
    emitNewLine();
  }
}

void Intrinsic::emitReturnReversal() {
  if (isBigEndianSafe())
    return;
  if (!getReturnType().isVector() || getReturnType().isVoid() ||
      getReturnType().getNumElements() == 1)
    return;
  emitReverseVariable(RetVar, RetVar);
}

void Intrinsic::emitShadowedArgs() {
  // Macro arguments are not type-checked like inline function arguments,
  // so assign them to local temporaries to get the right type checking.
  if (!UseMacro)
    return;

  for (unsigned I = 0; I < getNumParams(); ++I) {
    // Do not create a temporary for an immediate argument.
    // That would defeat the whole point of using a macro!
    if (getParamType(I).isImmediate())
      continue;
    // Do not create a temporary for pointer arguments. The input
    // pointer may have an alignment hint.
    if (getParamType(I).isPointer())
      continue;

    std::string Name = "p" + utostr(I);

    assert(Variables.find(Name) != Variables.end());
    Variable &V = Variables[Name];

    std::string NewName = "s" + utostr(I);
    Variable V2(V.getType(), NewName + VariablePostfix);

    OS << "  " << V2.getType().str() << " " << V2.getName() << " = "
       << V.getName() << ";";
    emitNewLine();

    V = V2;
  }
}

bool Intrinsic::protoHasScalar() const {
  return any_of(Types,
                [](const Type &T) { return T.isScalar() && !T.isImmediate(); });
}

void Intrinsic::emitBodyAsBuiltinCall() {
  std::string S;

  // If this builtin returns a struct 2, 3, or 4 vectors, pass it as an implicit
  // sret-like argument.
  bool SRet = getReturnType().getNumVectors() >= 2;

  StringRef N = Name;
  ClassKind LocalCK = CK;
  if (!protoHasScalar())
    LocalCK = ClassB;

  if (!getReturnType().isVoid() && !SRet)
    S += "(" + RetVar.getType().str() + ") ";

  S += "__builtin_neon_" + mangleName(std::string(N), LocalCK) + "(";

  if (SRet)
    S += "&" + RetVar.getName() + ", ";

  for (unsigned I = 0; I < getNumParams(); ++I) {
    Variable &V = Variables["p" + utostr(I)];
    Type T = V.getType();

    // Handle multiple-vector values specially, emitting each subvector as an
    // argument to the builtin.
    if (T.getNumVectors() > 1) {
      // Check if an explicit cast is needed.
      std::string Cast;
      if (LocalCK == ClassB) {
        Type T2 = T;
        T2.makeOneVector();
        T2.makeInteger(8, /*Sign=*/true);
        Cast = "(" + T2.str() + ")";
      }

      for (unsigned J = 0; J < T.getNumVectors(); ++J)
        S += Cast + V.getName() + ".val[" + utostr(J) + "], ";
      continue;
    }

    std::string Arg = V.getName();
    Type CastToType = T;

    // Check if an explicit cast is needed.
    if (CastToType.isVector() &&
        (LocalCK == ClassB || (T.isHalf() && !T.isScalarForMangling()))) {
      CastToType.makeInteger(8, true);
      Arg = "(" + CastToType.str() + ")" + Arg;
    } else if (CastToType.isVector() && LocalCK == ClassI) {
      if (CastToType.isInteger())
        CastToType.makeSigned();
      Arg = "(" + CastToType.str() + ")" + Arg;
    }

    S += Arg + ", ";
  }

  // Extra constant integer to hold type class enum for this function, e.g. s8
  if (getClassKind(true) == ClassB) {
    S += utostr(getPolymorphicKeyType().getNeonEnum());
  } else {
    // Remove extraneous ", ".
    S.pop_back();
    S.pop_back();
  }
  S += ");";

  std::string RetExpr;
  if (!SRet && !RetVar.getType().isVoid())
    RetExpr = RetVar.getName() + " = ";

  OS << "  " << RetExpr << S;
  emitNewLine();
}

void Intrinsic::emitBody(StringRef CallPrefix) {
  std::vector<std::string> Lines;

  if (!Body || Body->getValues().empty()) {
    // Nothing specific to output - must output a builtin.
    emitBodyAsBuiltinCall();
    return;
  }

  // We have a list of "things to output". The last should be returned.
  for (auto *I : Body->getValues()) {
    if (const auto *SI = dyn_cast<StringInit>(I)) {
      Lines.push_back(replaceParamsIn(SI->getAsString()));
    } else if (const auto *DI = dyn_cast<DagInit>(I)) {
      DagEmitter DE(*this, CallPrefix);
      Lines.push_back(DE.emitDag(DI).second + ";");
    }
  }

  assert(!Lines.empty() && "Empty def?");
  if (!RetVar.getType().isVoid())
    Lines.back().insert(0, RetVar.getName() + " = ");

  for (auto &L : Lines) {
    OS << "  " << L;
    emitNewLine();
  }
}

void Intrinsic::emitReturn() {
  if (RetVar.getType().isVoid())
    return;
  if (UseMacro)
    OS << "  " << RetVar.getName() << ";";
  else
    OS << "  return " << RetVar.getName() << ";";
  emitNewLine();
}

std::pair<Type, std::string> Intrinsic::DagEmitter::emitDag(const DagInit *DI) {
  // At this point we should only be seeing a def.
  const DefInit *DefI = cast<DefInit>(DI->getOperator());
  std::string Op = DefI->getAsString();

  if (Op == "cast" || Op == "bitcast")
    return emitDagCast(DI, Op == "bitcast");
  if (Op == "shuffle")
    return emitDagShuffle(DI);
  if (Op == "dup")
    return emitDagDup(DI);
  if (Op == "dup_typed")
    return emitDagDupTyped(DI);
  if (Op == "splat")
    return emitDagSplat(DI);
  if (Op == "save_temp")
    return emitDagSaveTemp(DI);
  if (Op == "op")
    return emitDagOp(DI);
  if (Op == "call" || Op == "call_mangled")
    return emitDagCall(DI, Op == "call_mangled");
  if (Op == "name_replace")
    return emitDagNameReplace(DI);
  if (Op == "literal")
    return emitDagLiteral(DI);
  assert_with_loc(false, "Unknown operation!");
  return std::make_pair(Type::getVoid(), "");
}

std::pair<Type, std::string>
Intrinsic::DagEmitter::emitDagOp(const DagInit *DI) {
  std::string Op = cast<StringInit>(DI->getArg(0))->getAsUnquotedString();
  if (DI->getNumArgs() == 2) {
    // Unary op.
    std::pair<Type, std::string> R =
        emitDagArg(DI->getArg(1), std::string(DI->getArgNameStr(1)));
    return std::make_pair(R.first, Op + R.second);
  } else {
    assert(DI->getNumArgs() == 3 && "Can only handle unary and binary ops!");
    std::pair<Type, std::string> R1 =
        emitDagArg(DI->getArg(1), std::string(DI->getArgNameStr(1)));
    std::pair<Type, std::string> R2 =
        emitDagArg(DI->getArg(2), std::string(DI->getArgNameStr(2)));
    assert_with_loc(R1.first == R2.first, "Argument type mismatch!");
    return std::make_pair(R1.first, R1.second + " " + Op + " " + R2.second);
  }
}

std::pair<Type, std::string>
Intrinsic::DagEmitter::emitDagCall(const DagInit *DI, bool MatchMangledName) {
  std::vector<Type> Types;
  std::vector<std::string> Values;
  for (unsigned I = 0; I < DI->getNumArgs() - 1; ++I) {
    std::pair<Type, std::string> R =
        emitDagArg(DI->getArg(I + 1), std::string(DI->getArgNameStr(I + 1)));
    Types.push_back(R.first);
    Values.push_back(R.second);
  }

  // Look up the called intrinsic.
  std::string N;
  if (const auto *SI = dyn_cast<StringInit>(DI->getArg(0)))
    N = SI->getAsUnquotedString();
  else
    N = emitDagArg(DI->getArg(0), "").second;
  std::optional<std::string> MangledName;
  if (MatchMangledName) {
    if (Intr.getRecord()->getValueAsString("Name").contains("laneq"))
      N += "q";
    MangledName = Intr.mangleName(N, ClassS);
  }
  Intrinsic &Callee = Intr.Emitter.getIntrinsic(N, Types, MangledName);

  // Make sure the callee is known as an early def.
  Callee.setNeededEarly();
  Intr.Dependencies.insert(&Callee);

  // Now create the call itself.
  std::string S;
  if (!Callee.isBigEndianSafe())
    S += CallPrefix.str();
  S += Callee.getMangledName(true) + "(";
  for (unsigned I = 0; I < DI->getNumArgs() - 1; ++I) {
    if (I != 0)
      S += ", ";
    S += Values[I];
  }
  S += ")";

  return std::make_pair(Callee.getReturnType(), S);
}

std::pair<Type, std::string>
Intrinsic::DagEmitter::emitDagCast(const DagInit *DI, bool IsBitCast) {
  // (cast MOD* VAL) -> cast VAL to type given by MOD.
  std::pair<Type, std::string> R =
      emitDagArg(DI->getArg(DI->getNumArgs() - 1),
                 std::string(DI->getArgNameStr(DI->getNumArgs() - 1)));
  Type castToType = R.first;
  for (unsigned ArgIdx = 0; ArgIdx < DI->getNumArgs() - 1; ++ArgIdx) {

    // MOD can take several forms:
    //   1. $X - take the type of parameter / variable X.
    //   2. The value "R" - take the type of the return type.
    //   3. a type string
    //   4. The value "U" or "S" to switch the signedness.
    //   5. The value "H" or "D" to half or double the bitwidth.
    //   6. The value "8" to convert to 8-bit (signed) integer lanes.
    if (!DI->getArgNameStr(ArgIdx).empty()) {
      assert_with_loc(Intr.Variables.find(DI->getArgNameStr(ArgIdx)) !=
                          Intr.Variables.end(),
                      "Variable not found");
      castToType =
          Intr.Variables[std::string(DI->getArgNameStr(ArgIdx))].getType();
    } else {
      const auto *SI = dyn_cast<StringInit>(DI->getArg(ArgIdx));
      assert_with_loc(SI, "Expected string type or $Name for cast type");

      if (SI->getAsUnquotedString() == "R") {
        castToType = Intr.getReturnType();
      } else if (SI->getAsUnquotedString() == "U") {
        castToType.makeUnsigned();
      } else if (SI->getAsUnquotedString() == "S") {
        castToType.makeSigned();
      } else if (SI->getAsUnquotedString() == "H") {
        castToType.halveLanes();
      } else if (SI->getAsUnquotedString() == "D") {
        castToType.doubleLanes();
      } else if (SI->getAsUnquotedString() == "8") {
        castToType.makeInteger(8, true);
      } else if (SI->getAsUnquotedString() == "32") {
        castToType.make32BitElement();
      } else {
        castToType = Type::fromTypedefName(SI->getAsUnquotedString());
        assert_with_loc(!castToType.isVoid(), "Unknown typedef");
      }
    }
  }

  std::string S;
  if (IsBitCast) {
    // Emit a reinterpret cast. The second operand must be an lvalue, so create
    // a temporary.
    std::string N = "reint";
    unsigned I = 0;
    while (Intr.Variables.find(N) != Intr.Variables.end())
      N = "reint" + utostr(++I);
    Intr.Variables[N] = Variable(R.first, N + Intr.VariablePostfix);

    Intr.OS << R.first.str() << " " << Intr.Variables[N].getName() << " = "
            << R.second << ";";
    Intr.emitNewLine();

    S = "*(" + castToType.str() + " *) &" + Intr.Variables[N].getName() + "";
  } else {
    // Emit a normal (static) cast.
    S = "(" + castToType.str() + ")(" + R.second + ")";
  }

  return std::make_pair(castToType, S);
}

std::pair<Type, std::string>
Intrinsic::DagEmitter::emitDagShuffle(const DagInit *DI) {
  // See the documentation in arm_neon.td for a description of these operators.
  class LowHalf : public SetTheory::Operator {
  public:
    void apply(SetTheory &ST, const DagInit *Expr, SetTheory::RecSet &Elts,
               ArrayRef<SMLoc> Loc) override {
      SetTheory::RecSet Elts2;
      ST.evaluate(Expr->arg_begin(), Expr->arg_end(), Elts2, Loc);
      Elts.insert(Elts2.begin(), Elts2.begin() + (Elts2.size() / 2));
    }
  };

  class HighHalf : public SetTheory::Operator {
  public:
    void apply(SetTheory &ST, const DagInit *Expr, SetTheory::RecSet &Elts,
               ArrayRef<SMLoc> Loc) override {
      SetTheory::RecSet Elts2;
      ST.evaluate(Expr->arg_begin(), Expr->arg_end(), Elts2, Loc);
      Elts.insert(Elts2.begin() + (Elts2.size() / 2), Elts2.end());
    }
  };

  class Rev : public SetTheory::Operator {
    unsigned ElementSize;

  public:
    Rev(unsigned ElementSize) : ElementSize(ElementSize) {}

    void apply(SetTheory &ST, const DagInit *Expr, SetTheory::RecSet &Elts,
               ArrayRef<SMLoc> Loc) override {
      SetTheory::RecSet Elts2;
      ST.evaluate(Expr->arg_begin() + 1, Expr->arg_end(), Elts2, Loc);

      int64_t VectorSize = cast<IntInit>(Expr->getArg(0))->getValue();
      VectorSize /= ElementSize;

      std::vector<const Record *> Revved;
      for (unsigned VI = 0; VI < Elts2.size(); VI += VectorSize) {
        for (int LI = VectorSize - 1; LI >= 0; --LI) {
          Revved.push_back(Elts2[VI + LI]);
        }
      }

      Elts.insert(Revved.begin(), Revved.end());
    }
  };

  class MaskExpander : public SetTheory::Expander {
    unsigned N;

  public:
    MaskExpander(unsigned N) : N(N) {}

    void expand(SetTheory &ST, const Record *R,
                SetTheory::RecSet &Elts) override {
      unsigned Addend = 0;
      if (R->getName() == "mask0")
        Addend = 0;
      else if (R->getName() == "mask1")
        Addend = N;
      else
        return;
      for (unsigned I = 0; I < N; ++I)
        Elts.insert(R->getRecords().getDef("sv" + utostr(I + Addend)));
    }
  };

  // (shuffle arg1, arg2, sequence)
  std::pair<Type, std::string> Arg1 =
      emitDagArg(DI->getArg(0), std::string(DI->getArgNameStr(0)));
  std::pair<Type, std::string> Arg2 =
      emitDagArg(DI->getArg(1), std::string(DI->getArgNameStr(1)));
  assert_with_loc(Arg1.first == Arg2.first,
                  "Different types in arguments to shuffle!");

  SetTheory ST;
  SetTheory::RecSet Elts;
  ST.addOperator("lowhalf", std::make_unique<LowHalf>());
  ST.addOperator("highhalf", std::make_unique<HighHalf>());
  ST.addOperator("rev",
                 std::make_unique<Rev>(Arg1.first.getElementSizeInBits()));
  ST.addExpander("MaskExpand",
                 std::make_unique<MaskExpander>(Arg1.first.getNumElements()));
  ST.evaluate(DI->getArg(2), Elts, {});

  std::string S = "__builtin_shufflevector(" + Arg1.second + ", " + Arg2.second;
  for (auto &E : Elts) {
    StringRef Name = E->getName();
    assert_with_loc(Name.starts_with("sv"),
                    "Incorrect element kind in shuffle mask!");
    S += ", " + Name.drop_front(2).str();
  }
  S += ")";

  // Recalculate the return type - the shuffle may have halved or doubled it.
  Type T(Arg1.first);
  if (Elts.size() > T.getNumElements()) {
    assert_with_loc(
        Elts.size() == T.getNumElements() * 2,
        "Can only double or half the number of elements in a shuffle!");
    T.doubleLanes();
  } else if (Elts.size() < T.getNumElements()) {
    assert_with_loc(
        Elts.size() == T.getNumElements() / 2,
        "Can only double or half the number of elements in a shuffle!");
    T.halveLanes();
  }

  return std::make_pair(T, S);
}

std::pair<Type, std::string>
Intrinsic::DagEmitter::emitDagDup(const DagInit *DI) {
  assert_with_loc(DI->getNumArgs() == 1, "dup() expects one argument");
  std::pair<Type, std::string> A =
      emitDagArg(DI->getArg(0), std::string(DI->getArgNameStr(0)));
  assert_with_loc(A.first.isScalar(), "dup() expects a scalar argument");

  Type T = Intr.getBaseType();
  assert_with_loc(T.isVector(), "dup() used but default type is scalar!");
  std::string S = "(" + T.str() + ") {";
  for (unsigned I = 0; I < T.getNumElements(); ++I) {
    if (I != 0)
      S += ", ";
    S += A.second;
  }
  S += "}";

  return std::make_pair(T, S);
}

std::pair<Type, std::string>
Intrinsic::DagEmitter::emitDagDupTyped(const DagInit *DI) {
  assert_with_loc(DI->getNumArgs() == 2, "dup_typed() expects two arguments");
  std::pair<Type, std::string> B =
      emitDagArg(DI->getArg(1), std::string(DI->getArgNameStr(1)));
  assert_with_loc(B.first.isScalar(),
                  "dup_typed() requires a scalar as the second argument");
  Type T;
  // If the type argument is a constant string, construct the type directly.
  if (const auto *SI = dyn_cast<StringInit>(DI->getArg(0))) {
    T = Type::fromTypedefName(SI->getAsUnquotedString());
    assert_with_loc(!T.isVoid(), "Unknown typedef");
  } else
    T = emitDagArg(DI->getArg(0), std::string(DI->getArgNameStr(0))).first;

  assert_with_loc(T.isVector(), "dup_typed() used but target type is scalar!");
  std::string S = "(" + T.str() + ") {";
  for (unsigned I = 0; I < T.getNumElements(); ++I) {
    if (I != 0)
      S += ", ";
    S += B.second;
  }
  S += "}";

  return std::make_pair(T, S);
}

std::pair<Type, std::string>
Intrinsic::DagEmitter::emitDagSplat(const DagInit *DI) {
  assert_with_loc(DI->getNumArgs() == 2, "splat() expects two arguments");
  std::pair<Type, std::string> A =
      emitDagArg(DI->getArg(0), std::string(DI->getArgNameStr(0)));
  std::pair<Type, std::string> B =
      emitDagArg(DI->getArg(1), std::string(DI->getArgNameStr(1)));

  assert_with_loc(B.first.isScalar(),
                  "splat() requires a scalar int as the second argument");

  std::string S = "__builtin_shufflevector(" + A.second + ", " + A.second;
  for (unsigned I = 0; I < Intr.getBaseType().getNumElements(); ++I) {
    S += ", " + B.second;
  }
  S += ")";

  return std::make_pair(Intr.getBaseType(), S);
}

std::pair<Type, std::string>
Intrinsic::DagEmitter::emitDagSaveTemp(const DagInit *DI) {
  assert_with_loc(DI->getNumArgs() == 2, "save_temp() expects two arguments");
  std::pair<Type, std::string> A =
      emitDagArg(DI->getArg(1), std::string(DI->getArgNameStr(1)));

  assert_with_loc(!A.first.isVoid(),
                  "Argument to save_temp() must have non-void type!");

  std::string N = std::string(DI->getArgNameStr(0));
  assert_with_loc(!N.empty(),
                  "save_temp() expects a name as the first argument");

  assert_with_loc(Intr.Variables.find(N) == Intr.Variables.end(),
                  "Variable already defined!");
  Intr.Variables[N] = Variable(A.first, N + Intr.VariablePostfix);

  std::string S =
      A.first.str() + " " + Intr.Variables[N].getName() + " = " + A.second;

  return std::make_pair(Type::getVoid(), S);
}

std::pair<Type, std::string>
Intrinsic::DagEmitter::emitDagNameReplace(const DagInit *DI) {
  std::string S = Intr.Name;

  assert_with_loc(DI->getNumArgs() == 2, "name_replace requires 2 arguments!");
  std::string ToReplace = cast<StringInit>(DI->getArg(0))->getAsUnquotedString();
  std::string ReplaceWith = cast<StringInit>(DI->getArg(1))->getAsUnquotedString();

  size_t Idx = S.find(ToReplace);

  assert_with_loc(Idx != std::string::npos, "name should contain '" + ToReplace + "'!");
  S.replace(Idx, ToReplace.size(), ReplaceWith);

  return std::make_pair(Type::getVoid(), S);
}

std::pair<Type, std::string>
Intrinsic::DagEmitter::emitDagLiteral(const DagInit *DI) {
  std::string Ty = cast<StringInit>(DI->getArg(0))->getAsUnquotedString();
  std::string Value = cast<StringInit>(DI->getArg(1))->getAsUnquotedString();
  return std::make_pair(Type::fromTypedefName(Ty), Value);
}

std::pair<Type, std::string>
Intrinsic::DagEmitter::emitDagArg(const Init *Arg, std::string ArgName) {
  if (!ArgName.empty()) {
    assert_with_loc(!Arg->isComplete(),
                    "Arguments must either be DAGs or names, not both!");
    assert_with_loc(Intr.Variables.find(ArgName) != Intr.Variables.end(),
                    "Variable not defined!");
    Variable &V = Intr.Variables[ArgName];
    return std::make_pair(V.getType(), V.getName());
  }

  assert(Arg && "Neither ArgName nor Arg?!");
  const auto *DI = dyn_cast<DagInit>(Arg);
  assert_with_loc(DI, "Arguments must either be DAGs or names!");

  return emitDag(DI);
}

std::string Intrinsic::generate() {
  // Avoid duplicated code for big and little endian
  if (isBigEndianSafe()) {
    generateImpl(false, "", "");
    return OS.str();
  }
  // Little endian intrinsics are simple and don't require any argument
  // swapping.
  OS << "#ifdef __LITTLE_ENDIAN__\n";

  generateImpl(false, "", "");

  OS << "#else\n";

  // Big endian intrinsics are more complex. The user intended these
  // intrinsics to operate on a vector "as-if" loaded by (V)LDR,
  // but we load as-if (V)LD1. So we should swap all arguments and
  // swap the return value too.
  //
  // If we call sub-intrinsics, we should call a version that does
  // not re-swap the arguments!
  generateImpl(true, "", "__noswap_");

  // If we're needed early, create a non-swapping variant for
  // big-endian.
  if (NeededEarly) {
    generateImpl(false, "__noswap_", "__noswap_");
  }
  OS << "#endif\n\n";

  return OS.str();
}

void Intrinsic::generateImpl(bool ReverseArguments,
                             StringRef NamePrefix, StringRef CallPrefix) {
  CurrentRecord = R;

  // If we call a macro, our local variables may be corrupted due to
  // lack of proper lexical scoping. So, add a globally unique postfix
  // to every variable.
  //
  // indexBody() should have set up the Dependencies set by now.
  for (auto *I : Dependencies)
    if (I->UseMacro) {
      VariablePostfix = "_" + utostr(Emitter.getUniqueNumber());
      break;
    }

  initVariables();

  emitPrototype(NamePrefix);

  if (IsUnavailable) {
    OS << " __attribute__((unavailable));";
  } else {
    emitOpeningBrace();
    // Emit return variable declaration first as to not trigger
    // -Wdeclaration-after-statement.
    emitReturnVarDecl();
    emitShadowedArgs();
    if (ReverseArguments)
      emitArgumentReversal();
    emitBody(CallPrefix);
    if (ReverseArguments)
      emitReturnReversal();
    emitReturn();
    emitClosingBrace();
  }
  OS << "\n";

  CurrentRecord = nullptr;
}

void Intrinsic::indexBody() {
  CurrentRecord = R;

  initVariables();
  // Emit return variable declaration first as to not trigger
  // -Wdeclaration-after-statement.
  emitReturnVarDecl();
  emitBody("");
  OS.str("");

  CurrentRecord = nullptr;
}

//===----------------------------------------------------------------------===//
// NeonEmitter implementation
//===----------------------------------------------------------------------===//

Intrinsic &NeonEmitter::getIntrinsic(StringRef Name, ArrayRef<Type> Types,
                                     std::optional<std::string> MangledName) {
  // First, look up the name in the intrinsic map.
  assert_with_loc(IntrinsicMap.find(Name) != IntrinsicMap.end(),
                  ("Intrinsic '" + Name + "' not found!").str());
  auto &V = IntrinsicMap.find(Name)->second;
  std::vector<Intrinsic *> GoodVec;

  // Create a string to print if we end up failing.
  std::string ErrMsg = "looking up intrinsic '" + Name.str() + "(";
  for (unsigned I = 0; I < Types.size(); ++I) {
    if (I != 0)
      ErrMsg += ", ";
    ErrMsg += Types[I].str();
  }
  ErrMsg += ")'\n";
  ErrMsg += "Available overloads:\n";

  // Now, look through each intrinsic implementation and see if the types are
  // compatible.
  for (auto &I : V) {
    ErrMsg += "  - " + I.getReturnType().str() + " " + I.getMangledName();
    ErrMsg += "(";
    for (unsigned A = 0; A < I.getNumParams(); ++A) {
      if (A != 0)
        ErrMsg += ", ";
      ErrMsg += I.getParamType(A).str();
    }
    ErrMsg += ")\n";

    if (MangledName && MangledName != I.getMangledName(true))
      continue;

    if (I.getNumParams() != Types.size())
      continue;

    unsigned ArgNum = 0;
    bool MatchingArgumentTypes = all_of(Types, [&](const auto &Type) {
      return Type == I.getParamType(ArgNum++);
    });

    if (MatchingArgumentTypes)
      GoodVec.push_back(&I);
  }

  assert_with_loc(!GoodVec.empty(),
                  "No compatible intrinsic found - " + ErrMsg);
  assert_with_loc(GoodVec.size() == 1, "Multiple overloads found - " + ErrMsg);

  return *GoodVec.front();
}

void NeonEmitter::createIntrinsic(const Record *R,
                                  SmallVectorImpl<Intrinsic *> &Out) {
  std::string Name = std::string(R->getValueAsString("Name"));
  std::string Proto = std::string(R->getValueAsString("Prototype"));
  std::string Types = std::string(R->getValueAsString("Types"));
  const Record *OperationRec = R->getValueAsDef("Operation");
  bool BigEndianSafe  = R->getValueAsBit("BigEndianSafe");
  std::string ArchGuard = std::string(R->getValueAsString("ArchGuard"));
  std::string TargetGuard = std::string(R->getValueAsString("TargetGuard"));
  bool IsUnavailable = OperationRec->getValueAsBit("Unavailable");
  std::string CartesianProductWith = std::string(R->getValueAsString("CartesianProductWith"));

  // Set the global current record. This allows assert_with_loc to produce
  // decent location information even when highly nested.
  CurrentRecord = R;

  const ListInit *Body = OperationRec->getValueAsListInit("Ops");

  std::vector<TypeSpec> TypeSpecs = TypeSpec::fromTypeSpecs(Types);

  ClassKind CK = ClassNone;
  if (R->getSuperClasses().size() >= 2)
    CK = ClassMap[R->getSuperClasses()[1].first];

  std::vector<std::pair<TypeSpec, TypeSpec>> NewTypeSpecs;
  if (!CartesianProductWith.empty()) {
    std::vector<TypeSpec> ProductTypeSpecs = TypeSpec::fromTypeSpecs(CartesianProductWith);
    for (auto TS : TypeSpecs) {
      Type DefaultT(TS, ".");
      for (auto SrcTS : ProductTypeSpecs) {
        Type DefaultSrcT(SrcTS, ".");
        if (TS == SrcTS ||
            DefaultSrcT.getSizeInBits() != DefaultT.getSizeInBits())
          continue;
        NewTypeSpecs.push_back(std::make_pair(TS, SrcTS));
      }
    }
  } else {
    for (auto TS : TypeSpecs) {
      NewTypeSpecs.push_back(std::make_pair(TS, TS));
    }
  }

  sort(NewTypeSpecs);
  NewTypeSpecs.erase(std::unique(NewTypeSpecs.begin(), NewTypeSpecs.end()),
		     NewTypeSpecs.end());
  auto &Entry = IntrinsicMap[Name];

  for (auto &I : NewTypeSpecs) {
    Entry.emplace_back(R, Name, Proto, I.first, I.second, CK, Body, *this,
                       ArchGuard, TargetGuard, IsUnavailable, BigEndianSafe);
    Out.push_back(&Entry.back());
  }

  CurrentRecord = nullptr;
}

/// genBuiltinsDef: Generate the BuiltinsARM.def and  BuiltinsAArch64.def
/// declaration of builtins, checking for unique builtin declarations.
void NeonEmitter::genBuiltinsDef(raw_ostream &OS,
                                 SmallVectorImpl<Intrinsic *> &Defs) {
  OS << "#ifdef GET_NEON_BUILTINS\n";

  // We only want to emit a builtin once, and we want to emit them in
  // alphabetical order, so use a std::set.
  std::set<std::pair<std::string, std::string>> Builtins;

  for (auto *Def : Defs) {
    if (Def->hasBody())
      continue;

    std::string S = "__builtin_neon_" + Def->getMangledName() + ", \"";
    S += Def->getBuiltinTypeStr();
    S += "\", \"n\"";

    Builtins.emplace(S, Def->getTargetGuard());
  }

  for (auto &S : Builtins) {
    if (S.second == "")
      OS << "BUILTIN(";
    else
      OS << "TARGET_BUILTIN(";
    OS << S.first;
    if (S.second == "")
      OS << ")\n";
    else
      OS << ", \"" << S.second << "\")\n";
  }

  OS << "#endif\n\n";
}

void NeonEmitter::genStreamingSVECompatibleList(
    raw_ostream &OS, SmallVectorImpl<Intrinsic *> &Defs) {
  OS << "#ifdef GET_NEON_STREAMING_COMPAT_FLAG\n";

  std::set<std::string> Emitted;
  for (auto *Def : Defs) {
    // If the def has a body (that is, it has Operation DAGs), it won't call
    // __builtin_neon_* so we don't need to generate a definition for it.
    if (Def->hasBody())
      continue;

    std::string Name = Def->getMangledName();
    if (Emitted.find(Name) != Emitted.end())
      continue;

    // FIXME: We should make exceptions here for some NEON builtins that are
    // permitted in streaming mode.
    OS << "case NEON::BI__builtin_neon_" << Name
       << ": BuiltinType = ArmNonStreaming; break;\n";
    Emitted.insert(Name);
  }
  OS << "#endif\n\n";
}

/// Generate the ARM and AArch64 overloaded type checking code for
/// SemaChecking.cpp, checking for unique builtin declarations.
void NeonEmitter::genOverloadTypeCheckCode(raw_ostream &OS,
                                           SmallVectorImpl<Intrinsic *> &Defs) {
  OS << "#ifdef GET_NEON_OVERLOAD_CHECK\n";

  // We record each overload check line before emitting because subsequent Inst
  // definitions may extend the number of permitted types (i.e. augment the
  // Mask). Use std::map to avoid sorting the table by hash number.
  struct OverloadInfo {
    uint64_t Mask = 0ULL;
    int PtrArgNum = 0;
    bool HasConstPtr = false;
    OverloadInfo() = default;
  };
  std::map<std::string, OverloadInfo> OverloadMap;

  for (auto *Def : Defs) {
    // If the def has a body (that is, it has Operation DAGs), it won't call
    // __builtin_neon_* so we don't need to generate a definition for it.
    if (Def->hasBody())
      continue;
    // Functions which have a scalar argument cannot be overloaded, no need to
    // check them if we are emitting the type checking code.
    if (Def->protoHasScalar())
      continue;

    uint64_t Mask = 0ULL;
    Mask |= 1ULL << Def->getPolymorphicKeyType().getNeonEnum();

    // Check if the function has a pointer or const pointer argument.
    int PtrArgNum = -1;
    bool HasConstPtr = false;
    for (unsigned I = 0; I < Def->getNumParams(); ++I) {
      const auto &Type = Def->getParamType(I);
      if (Type.isPointer()) {
        PtrArgNum = I;
        HasConstPtr = Type.isConstPointer();
      }
    }

    // For sret builtins, adjust the pointer argument index.
    if (PtrArgNum >= 0 && Def->getReturnType().getNumVectors() > 1)
      PtrArgNum += 1;

    std::string Name = Def->getName();
    // Omit type checking for the pointer arguments of vld1_lane, vld1_dup,
    // vst1_lane, vldap1_lane, and vstl1_lane intrinsics.  Using a pointer to
    // the vector element type with one of those operations causes codegen to
    // select an aligned load/store instruction.  If you want an unaligned
    // operation, the pointer argument needs to have less alignment than element
    // type, so just accept any pointer type.
    if (Name == "vld1_lane" || Name == "vld1_dup" || Name == "vst1_lane" ||
        Name == "vldap1_lane" || Name == "vstl1_lane") {
      PtrArgNum = -1;
      HasConstPtr = false;
    }

    if (Mask) {
      OverloadInfo &OI = OverloadMap[Def->getMangledName()];
      OI.Mask |= Mask;
      OI.PtrArgNum |= PtrArgNum;
      OI.HasConstPtr = HasConstPtr;
    }
  }

  for (auto &I : OverloadMap) {
    OverloadInfo &OI = I.second;

    OS << "case NEON::BI__builtin_neon_" << I.first << ": ";
    OS << "mask = 0x" << Twine::utohexstr(OI.Mask) << "ULL";
    if (OI.PtrArgNum >= 0)
      OS << "; PtrArgNum = " << OI.PtrArgNum;
    if (OI.HasConstPtr)
      OS << "; HasConstPtr = true";
    OS << "; break;\n";
  }
  OS << "#endif\n\n";
}

inline bool
NeonEmitter::areRangeChecksCompatible(const ArrayRef<ImmCheck> ChecksA,
                                      const ArrayRef<ImmCheck> ChecksB) {
  // If multiple intrinsics map to the same builtin, we must ensure that the
  // intended range checks performed in SemaArm.cpp do not contradict each
  // other, as these are emitted once per-buitlin.
  //
  // The arguments to be checked and type of each check to be performed must be
  // the same. The element types may differ as they will be resolved
  // per-intrinsic as overloaded types by SemaArm.cpp, though the vector sizes
  // are not and so must be the same.
  bool compat =
      std::equal(ChecksA.begin(), ChecksA.end(), ChecksB.begin(), ChecksB.end(),
                 [](const auto &A, const auto &B) {
                   return A.getImmArgIdx() == B.getImmArgIdx() &&
                          A.getKind() == B.getKind() &&
                          A.getVecSizeInBits() == B.getVecSizeInBits();
                 });

  return compat;
}

void NeonEmitter::genIntrinsicRangeCheckCode(
    raw_ostream &OS, SmallVectorImpl<Intrinsic *> &Defs) {
  std::unordered_map<std::string, ArrayRef<ImmCheck>> Emitted;

  OS << "#ifdef GET_NEON_IMMEDIATE_CHECK\n";
  for (auto &Def : Defs) {
    // If the Def has a body (operation DAGs), it is not a __builtin_neon_
    if (Def->hasBody() || !Def->hasImmediate())
      continue;

    // Sorted by immediate argument index
    ArrayRef<ImmCheck> Checks = Def->getImmChecks();

    const auto it = Emitted.find(Def->getMangledName());
    if (it != Emitted.end()) {
      assert(areRangeChecksCompatible(Checks, it->second) &&
             "Neon intrinsics with incompatible immediate range checks cannot "
             "share a builtin.");
      continue; // Ensure this is emitted only once
    }

    // Emit builtin's range checks
    OS << "case NEON::BI__builtin_neon_" << Def->getMangledName() << ":\n";
    for (const auto &Check : Checks) {
      OS << " ImmChecks.emplace_back(" << Check.getImmArgIdx() << ", "
         << Check.getKind() << ", " << Check.getElementSizeInBits() << ", "
         << Check.getVecSizeInBits() << ");\n"
         << " break;\n";
    }
    Emitted[Def->getMangledName()] = Checks;
  }

  OS << "#endif\n\n";
}

/// runHeader - Emit a file with sections defining:
/// 1. the NEON section of BuiltinsARM.def and BuiltinsAArch64.def.
/// 2. the SemaChecking code for the type overload checking.
/// 3. the SemaChecking code for validation of intrinsic immediate arguments.
void NeonEmitter::runHeader(raw_ostream &OS) {
  SmallVector<Intrinsic *, 128> Defs;
  for (const Record *R : Records.getAllDerivedDefinitions("Inst"))
    createIntrinsic(R, Defs);

  // Generate shared BuiltinsXXX.def
  genBuiltinsDef(OS, Defs);

  // Generate ARM overloaded type checking code for SemaChecking.cpp
  genOverloadTypeCheckCode(OS, Defs);

  genStreamingSVECompatibleList(OS, Defs);

  // Generate ARM range checking code for shift/lane immediates.
  genIntrinsicRangeCheckCode(OS, Defs);
}

static void emitNeonTypeDefs(const std::string& types, raw_ostream &OS) {
  std::string TypedefTypes(types);
  std::vector<TypeSpec> TDTypeVec = TypeSpec::fromTypeSpecs(TypedefTypes);

  // Emit vector typedefs.
  bool InIfdef = false;
  for (auto &TS : TDTypeVec) {
    bool IsA64 = false;
    Type T(TS, ".");
    if (T.isDouble() || T.isMFloat8())
      IsA64 = true;

    if (InIfdef && !IsA64) {
      OS << "#endif\n";
      InIfdef = false;
    }
    if (!InIfdef && IsA64) {
      OS << "#if defined(__aarch64__) || defined(__arm64ec__)\n";
      InIfdef = true;
    }

    if (T.isMFloat8())
      OS << "typedef __MFloat8x";
    else if (T.isPoly())
      OS << "typedef __attribute__((neon_polyvector_type(";
    else
      OS << "typedef __attribute__((neon_vector_type(";

    Type T2 = T;
    T2.makeScalar();
    OS << T.getNumElements();
    if (T.isMFloat8())
      OS << "_t ";
    else
      OS << "))) " << T2.str();
    OS << " " << T.str() << ";\n";
  }
  if (InIfdef)
    OS << "#endif\n";
  OS << "\n";

  // Emit struct typedefs.
  InIfdef = false;
  for (unsigned NumMembers = 2; NumMembers <= 4; ++NumMembers) {
    for (auto &TS : TDTypeVec) {
      bool IsA64 = false;
      Type T(TS, ".");
      if (T.isDouble() || T.isMFloat8())
        IsA64 = true;

      if (InIfdef && !IsA64) {
        OS << "#endif\n";
        InIfdef = false;
      }
      if (!InIfdef && IsA64) {
        OS << "#if defined(__aarch64__) || defined(__arm64ec__)\n";
        InIfdef = true;
      }

      const char Mods[] = { static_cast<char>('2' + (NumMembers - 2)), 0};
      Type VT(TS, Mods);
      OS << "typedef struct " << VT.str() << " {\n";
      OS << "  " << T.str() << " val";
      OS << "[" << NumMembers << "]";
      OS << ";\n} ";
      OS << VT.str() << ";\n";
      OS << "\n";
    }
  }
  if (InIfdef)
    OS << "#endif\n";
}

/// run - Read the records in arm_neon.td and output arm_neon.h.  arm_neon.h
/// is comprised of type definitions and function declarations.
void NeonEmitter::run(raw_ostream &OS) {
  OS << "/*===---- arm_neon.h - ARM Neon intrinsics "
        "------------------------------"
        "---===\n"
        " *\n"
        " * Permission is hereby granted, free of charge, to any person "
        "obtaining "
        "a copy\n"
        " * of this software and associated documentation files (the "
        "\"Software\"),"
        " to deal\n"
        " * in the Software without restriction, including without limitation "
        "the "
        "rights\n"
        " * to use, copy, modify, merge, publish, distribute, sublicense, "
        "and/or sell\n"
        " * copies of the Software, and to permit persons to whom the Software "
        "is\n"
        " * furnished to do so, subject to the following conditions:\n"
        " *\n"
        " * The above copyright notice and this permission notice shall be "
        "included in\n"
        " * all copies or substantial portions of the Software.\n"
        " *\n"
        " * THE SOFTWARE IS PROVIDED \"AS IS\", WITHOUT WARRANTY OF ANY KIND, "
        "EXPRESS OR\n"
        " * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF "
        "MERCHANTABILITY,\n"
        " * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT "
        "SHALL THE\n"
        " * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR "
        "OTHER\n"
        " * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, "
        "ARISING FROM,\n"
        " * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER "
        "DEALINGS IN\n"
        " * THE SOFTWARE.\n"
        " *\n"
        " *===-----------------------------------------------------------------"
        "---"
        "---===\n"
        " */\n\n";

  OS << "#ifndef __ARM_NEON_H\n";
  OS << "#define __ARM_NEON_H\n\n";

  OS << "#ifndef __ARM_FP\n";
  OS << "#error \"NEON intrinsics not available with the soft-float ABI. "
        "Please use -mfloat-abi=softfp or -mfloat-abi=hard\"\n";
  OS << "#else\n\n";

  OS << "#include <stdint.h>\n\n";

  OS << "#include <arm_bf16.h>\n";

  OS << "#include <arm_vector_types.h>\n";

  // For now, signedness of polynomial types depends on target
  OS << "#if defined(__aarch64__) || defined(__arm64ec__)\n";
  OS << "typedef uint8_t poly8_t;\n";
  OS << "typedef uint16_t poly16_t;\n";
  OS << "typedef uint64_t poly64_t;\n";
  OS << "typedef __uint128_t poly128_t;\n";
  OS << "#else\n";
  OS << "typedef int8_t poly8_t;\n";
  OS << "typedef int16_t poly16_t;\n";
  OS << "typedef int64_t poly64_t;\n";
  OS << "#endif\n";
  emitNeonTypeDefs("PcQPcPsQPsPlQPl", OS);

  OS << "#define __ai static __inline__ __attribute__((__always_inline__, "
        "__nodebug__))\n\n";

  SmallVector<Intrinsic *, 128> Defs;
  for (const Record *R : Records.getAllDerivedDefinitions("Inst"))
    createIntrinsic(R, Defs);

  for (auto *I : Defs)
    I->indexBody();

  stable_sort(Defs, deref<std::less<>>());

  // Only emit a def when its requirements have been met.
  // FIXME: This loop could be made faster, but it's fast enough for now.
  bool MadeProgress = true;
  std::string InGuard;
  while (!Defs.empty() && MadeProgress) {
    MadeProgress = false;

    for (SmallVector<Intrinsic *, 128>::iterator I = Defs.begin();
         I != Defs.end(); /*No step*/) {
      bool DependenciesSatisfied = true;
      for (auto *II : (*I)->getDependencies()) {
        if (is_contained(Defs, II))
          DependenciesSatisfied = false;
      }
      if (!DependenciesSatisfied) {
        // Try the next one.
        ++I;
        continue;
      }

      // Emit #endif/#if pair if needed.
      if ((*I)->getArchGuard() != InGuard) {
        if (!InGuard.empty())
          OS << "#endif\n";
        InGuard = (*I)->getArchGuard();
        if (!InGuard.empty())
          OS << "#if " << InGuard << "\n";
      }

      // Actually generate the intrinsic code.
      OS << (*I)->generate();

      MadeProgress = true;
      I = Defs.erase(I);
    }
  }
  assert(Defs.empty() && "Some requirements were not satisfied!");
  if (!InGuard.empty())
    OS << "#endif\n";

  OS << "\n";
  OS << "#undef __ai\n\n";
  OS << "#endif /* if !defined(__ARM_NEON) */\n";
  OS << "#endif /* ifndef __ARM_FP */\n";
}

/// run - Read the records in arm_fp16.td and output arm_fp16.h.  arm_fp16.h
/// is comprised of type definitions and function declarations.
void NeonEmitter::runFP16(raw_ostream &OS) {
  OS << "/*===---- arm_fp16.h - ARM FP16 intrinsics "
        "------------------------------"
        "---===\n"
        " *\n"
        " * Permission is hereby granted, free of charge, to any person "
        "obtaining a copy\n"
        " * of this software and associated documentation files (the "
				"\"Software\"), to deal\n"
        " * in the Software without restriction, including without limitation "
				"the rights\n"
        " * to use, copy, modify, merge, publish, distribute, sublicense, "
				"and/or sell\n"
        " * copies of the Software, and to permit persons to whom the Software "
				"is\n"
        " * furnished to do so, subject to the following conditions:\n"
        " *\n"
        " * The above copyright notice and this permission notice shall be "
        "included in\n"
        " * all copies or substantial portions of the Software.\n"
        " *\n"
        " * THE SOFTWARE IS PROVIDED \"AS IS\", WITHOUT WARRANTY OF ANY KIND, "
        "EXPRESS OR\n"
        " * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF "
        "MERCHANTABILITY,\n"
        " * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT "
        "SHALL THE\n"
        " * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR "
        "OTHER\n"
        " * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, "
        "ARISING FROM,\n"
        " * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER "
        "DEALINGS IN\n"
        " * THE SOFTWARE.\n"
        " *\n"
        " *===-----------------------------------------------------------------"
        "---"
        "---===\n"
        " */\n\n";

  OS << "#ifndef __ARM_FP16_H\n";
  OS << "#define __ARM_FP16_H\n\n";

  OS << "#include <stdint.h>\n\n";

  OS << "typedef __fp16 float16_t;\n";

  OS << "#define __ai static __inline__ __attribute__((__always_inline__, "
        "__nodebug__))\n\n";

  SmallVector<Intrinsic *, 128> Defs;
  for (const Record *R : Records.getAllDerivedDefinitions("Inst"))
    createIntrinsic(R, Defs);

  for (auto *I : Defs)
    I->indexBody();

  stable_sort(Defs, deref<std::less<>>());

  // Only emit a def when its requirements have been met.
  // FIXME: This loop could be made faster, but it's fast enough for now.
  bool MadeProgress = true;
  std::string InGuard;
  while (!Defs.empty() && MadeProgress) {
    MadeProgress = false;

    for (SmallVector<Intrinsic *, 128>::iterator I = Defs.begin();
         I != Defs.end(); /*No step*/) {
      bool DependenciesSatisfied = true;
      for (auto *II : (*I)->getDependencies()) {
        if (is_contained(Defs, II))
          DependenciesSatisfied = false;
      }
      if (!DependenciesSatisfied) {
        // Try the next one.
        ++I;
        continue;
      }

      // Emit #endif/#if pair if needed.
      if ((*I)->getArchGuard() != InGuard) {
        if (!InGuard.empty())
          OS << "#endif\n";
        InGuard = (*I)->getArchGuard();
        if (!InGuard.empty())
          OS << "#if " << InGuard << "\n";
      }

      // Actually generate the intrinsic code.
      OS << (*I)->generate();

      MadeProgress = true;
      I = Defs.erase(I);
    }
  }
  assert(Defs.empty() && "Some requirements were not satisfied!");
  if (!InGuard.empty())
    OS << "#endif\n";

  OS << "\n";
  OS << "#undef __ai\n\n";
  OS << "#endif /* __ARM_FP16_H */\n";
}

void NeonEmitter::runVectorTypes(raw_ostream &OS) {
  OS << "/*===---- arm_vector_types - ARM vector type "
        "------===\n"
        " *\n"
        " *\n"
        " * Part of the LLVM Project, under the Apache License v2.0 with LLVM "
        "Exceptions.\n"
        " * See https://llvm.org/LICENSE.txt for license information.\n"
        " * SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception\n"
        " *\n"
        " *===-----------------------------------------------------------------"
        "------===\n"
        " */\n\n";
  OS << "#if !defined(__ARM_NEON_H) && !defined(__ARM_SVE_H)\n";
  OS << "#error \"This file should not be used standalone. Please include"
        " arm_neon.h or arm_sve.h instead\"\n\n";
  OS << "#endif\n";
  OS << "#ifndef __ARM_NEON_TYPES_H\n";
  OS << "#define __ARM_NEON_TYPES_H\n";
  OS << "typedef float float32_t;\n";
  OS << "typedef __fp16 float16_t;\n";

  OS << "#if defined(__aarch64__) || defined(__arm64ec__)\n";
  OS << "typedef __mfp8 mfloat8_t;\n";
  OS << "typedef double float64_t;\n";
  OS << "#endif\n\n";

  OS << R"(
typedef uint64_t fpm_t;

enum __ARM_FPM_FORMAT { __ARM_FPM_E5M2, __ARM_FPM_E4M3 };

enum __ARM_FPM_OVERFLOW { __ARM_FPM_INFNAN, __ARM_FPM_SATURATE };

static __inline__ fpm_t __attribute__((__always_inline__, __nodebug__))
__arm_fpm_init(void) {
  return 0;
}

static __inline__ fpm_t __attribute__((__always_inline__, __nodebug__))
__arm_set_fpm_src1_format(fpm_t __fpm, enum __ARM_FPM_FORMAT __format) {
  return (__fpm & ~7ull) | (fpm_t)__format;
}

static __inline__ fpm_t __attribute__((__always_inline__, __nodebug__))
__arm_set_fpm_src2_format(fpm_t __fpm, enum __ARM_FPM_FORMAT __format) {
  return (__fpm & ~0x38ull) | ((fpm_t)__format << 3u);
}

static __inline__ fpm_t __attribute__((__always_inline__, __nodebug__))
__arm_set_fpm_dst_format(fpm_t __fpm, enum __ARM_FPM_FORMAT __format) {
  return (__fpm & ~0x1c0ull) | ((fpm_t)__format << 6u);
}

static __inline__ fpm_t __attribute__((__always_inline__, __nodebug__))
__arm_set_fpm_overflow_mul(fpm_t __fpm, enum __ARM_FPM_OVERFLOW __behaviour) {
  return (__fpm & ~0x4000ull) | ((fpm_t)__behaviour << 14u);
}

static __inline__ fpm_t __attribute__((__always_inline__, __nodebug__))
__arm_set_fpm_overflow_cvt(fpm_t __fpm, enum __ARM_FPM_OVERFLOW __behaviour) {
  return (__fpm & ~0x8000ull) | ((fpm_t)__behaviour << 15u);
}

static __inline__ fpm_t __attribute__((__always_inline__, __nodebug__))
__arm_set_fpm_lscale(fpm_t __fpm, uint64_t __scale) {
  return (__fpm & ~0x7f0000ull) | (__scale << 16u);
}

static __inline__ fpm_t __attribute__((__always_inline__, __nodebug__))
__arm_set_fpm_nscale(fpm_t __fpm, int64_t __scale) {
  return (__fpm & ~0xff000000ull) | (((fpm_t)__scale & 0xffu) << 24u);
}

static __inline__ fpm_t __attribute__((__always_inline__, __nodebug__))
__arm_set_fpm_lscale2(fpm_t __fpm, uint64_t __scale) {
  return (uint32_t)__fpm | (__scale << 32u);
}

)";

  emitNeonTypeDefs("cQcsQsiQilQlUcQUcUsQUsUiQUiUlQUlmQmhQhfQfdQd", OS);

  emitNeonTypeDefs("bQb", OS);
  OS << "#endif // __ARM_NEON_TYPES_H\n";
}

void NeonEmitter::runBF16(raw_ostream &OS) {
  OS << "/*===---- arm_bf16.h - ARM BF16 intrinsics "
        "-----------------------------------===\n"
        " *\n"
        " *\n"
        " * Part of the LLVM Project, under the Apache License v2.0 with LLVM "
        "Exceptions.\n"
        " * See https://llvm.org/LICENSE.txt for license information.\n"
        " * SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception\n"
        " *\n"
        " *===-----------------------------------------------------------------"
        "------===\n"
        " */\n\n";

  OS << "#ifndef __ARM_BF16_H\n";
  OS << "#define __ARM_BF16_H\n\n";

  OS << "typedef __bf16 bfloat16_t;\n";

  OS << "#define __ai static __inline__ __attribute__((__always_inline__, "
        "__nodebug__))\n\n";

  SmallVector<Intrinsic *, 128> Defs;
  for (const Record *R : Records.getAllDerivedDefinitions("Inst"))
    createIntrinsic(R, Defs);

  for (auto *I : Defs)
    I->indexBody();

  stable_sort(Defs, deref<std::less<>>());

  // Only emit a def when its requirements have been met.
  // FIXME: This loop could be made faster, but it's fast enough for now.
  bool MadeProgress = true;
  std::string InGuard;
  while (!Defs.empty() && MadeProgress) {
    MadeProgress = false;

    for (SmallVector<Intrinsic *, 128>::iterator I = Defs.begin();
         I != Defs.end(); /*No step*/) {
      bool DependenciesSatisfied = true;
      for (auto *II : (*I)->getDependencies()) {
        if (is_contained(Defs, II))
          DependenciesSatisfied = false;
      }
      if (!DependenciesSatisfied) {
        // Try the next one.
        ++I;
        continue;
      }

      // Emit #endif/#if pair if needed.
      if ((*I)->getArchGuard() != InGuard) {
        if (!InGuard.empty())
          OS << "#endif\n";
        InGuard = (*I)->getArchGuard();
        if (!InGuard.empty())
          OS << "#if " << InGuard << "\n";
      }

      // Actually generate the intrinsic code.
      OS << (*I)->generate();

      MadeProgress = true;
      I = Defs.erase(I);
    }
  }
  assert(Defs.empty() && "Some requirements were not satisfied!");
  if (!InGuard.empty())
    OS << "#endif\n";

  OS << "\n";
  OS << "#undef __ai\n\n";

  OS << "#endif\n";
}

void clang::EmitNeon(const RecordKeeper &Records, raw_ostream &OS) {
  NeonEmitter(Records).run(OS);
}

void clang::EmitFP16(const RecordKeeper &Records, raw_ostream &OS) {
  NeonEmitter(Records).runFP16(OS);
}

void clang::EmitBF16(const RecordKeeper &Records, raw_ostream &OS) {
  NeonEmitter(Records).runBF16(OS);
}

void clang::EmitNeonSema(const RecordKeeper &Records, raw_ostream &OS) {
  NeonEmitter(Records).runHeader(OS);
}

void clang::EmitVectorTypes(const RecordKeeper &Records, raw_ostream &OS) {
  NeonEmitter(Records).runVectorTypes(OS);
}

void clang::EmitNeonTest(const RecordKeeper &Records, raw_ostream &OS) {
  llvm_unreachable("Neon test generation no longer implemented!");
}