/src/WasmEdge/lib/llvm/compiler.cpp

Source
// SPDX-License-Identifier: Apache-2.0
// SPDX-FileCopyrightText: Copyright The WasmEdge Authors

#include "llvm/compiler.h"

#include "aot/version.h"
#include "common/defines.h"
#include "common/filesystem.h"
#include "common/spdlog.h"
#include "data.h"
#include "llvm.h"
#include "spdlog/spdlog.h"
#include "system/allocator.h"

#include <algorithm>
#include <array>
#include <cinttypes>
#include <cstdint>
#include <cstdlib>
#include <functional>
#include <limits>
#include <memory>
#include <numeric>
#include <string>
#include <string_view>
#include <system_error>

namespace LLVM = WasmEdge::LLVM;
using namespace std::literals;

namespace {

struct RAIICleanup {
  RAIICleanup(LLVM::Compiler::CompileContext *&ContextRef,
              LLVM::Compiler::CompileContext *NewContext)
      : Context(ContextRef) {
    Context = NewContext;
  }
  ~RAIICleanup() { Context = nullptr; }
  LLVM::Compiler::CompileContext *&Context;
};

static bool
isVoidReturn(WasmEdge::Span<const WasmEdge::ValType> ValTypes) noexcept;
static LLVM::Type toLLVMType(LLVM::Context LLContext,
                             const WasmEdge::ValType &ValType) noexcept;
static std::vector<LLVM::Type>
toLLVMArgsType(LLVM::Context LLContext, LLVM::Type ExecCtxPtrTy,
               WasmEdge::Span<const WasmEdge::ValType> ValTypes) noexcept;
static LLVM::Type
toLLVMRetsType(LLVM::Context LLContext,
               WasmEdge::Span<const WasmEdge::ValType> ValTypes) noexcept;
static LLVM::Type
toLLVMType(LLVM::Context LLContext, LLVM::Type ExecCtxPtrTy,
           const WasmEdge::AST::FunctionType &FuncType) noexcept;
static LLVM::Value
toLLVMConstantZero(LLVM::Context LLContext, const WasmEdge::ValType &ValType,
                   WasmEdge::Span<const WasmEdge::AST::CompositeType *const>
                       CompositeTypes) noexcept;
static std::vector<LLVM::Value> unpackStruct(LLVM::Builder &Builder,
                                             LLVM::Value Struct) noexcept;
class FunctionCompiler;

// XXX: Misalignment handler not implemented yet, forcing unalignment
// force unalignment load/store
static inline constexpr const bool kForceUnalignment = true;

// force checking div/rem on zero
static inline constexpr const bool kForceDivCheck = true;

// Size of a ValVariant
static inline constexpr const uint32_t kValSize = sizeof(WasmEdge::ValVariant);

// Translate Compiler::OptimizationLevel to llvm::PassBuilder version
#if LLVM_VERSION_MAJOR >= 13
static inline const char *
toLLVMLevel(WasmEdge::CompilerConfigure::OptimizationLevel Level) noexcept {
  using OL = WasmEdge::CompilerConfigure::OptimizationLevel;
  switch (Level) {
  case OL::O0:
    return "default<O0>,function(tailcallelim)";
  case OL::O1:
    return "default<O1>,function(tailcallelim)";
  case OL::O2:
    return "default<O2>";
  case OL::O3:
    return "default<O3>";
  case OL::Os:
    return "default<Os>";
  case OL::Oz:
    return "default<Oz>";
  default:
    assumingUnreachable();
  }
}
#else
static inline std::pair<unsigned int, unsigned int>
toLLVMLevel(WasmEdge::CompilerConfigure::OptimizationLevel Level) noexcept {
  using OL = WasmEdge::CompilerConfigure::OptimizationLevel;
  switch (Level) {
  case OL::O0:
    return {0, 0};
  case OL::O1:
    return {1, 0};
  case OL::O2:
    return {2, 0};
  case OL::O3:
    return {3, 0};
  case OL::Os:
    return {2, 1};
  case OL::Oz:
    return {2, 2};
  default:
    assumingUnreachable();
  }
}
#endif

static inline LLVMCodeGenOptLevel toLLVMCodeGenLevel(
    WasmEdge::CompilerConfigure::OptimizationLevel Level) noexcept {
  using OL = WasmEdge::CompilerConfigure::OptimizationLevel;
  switch (Level) {
  case OL::O0:
    return LLVMCodeGenLevelNone;
  case OL::O1:
    return LLVMCodeGenLevelLess;
  case OL::O2:
    return LLVMCodeGenLevelDefault;
  case OL::O3:
    return LLVMCodeGenLevelAggressive;
  case OL::Os:
    return LLVMCodeGenLevelDefault;
  case OL::Oz:
    return LLVMCodeGenLevelDefault;
  default:
    assumingUnreachable();
  }
}
} // namespace

struct LLVM::Compiler::CompileContext {
  LLVM::Context LLContext;
  std::reference_wrapper<LLVM::Module> LLModule;
  LLVM::Attribute Cold;
  LLVM::Attribute NoAlias;
  LLVM::Attribute NoInline;
  LLVM::Attribute NoReturn;
  LLVM::Attribute ReadOnly;
  LLVM::Attribute StrictFP;
  LLVM::Attribute UWTable;
  LLVM::Attribute NoStackArgProbe;
  LLVM::Type VoidTy;
  LLVM::Type Int8Ty;
  LLVM::Type Int16Ty;
  LLVM::Type Int32Ty;
  LLVM::Type Int64Ty;
  LLVM::Type Int128Ty;
  LLVM::Type FloatTy;
  LLVM::Type DoubleTy;
  LLVM::Type Int8x16Ty;
  LLVM::Type Int16x8Ty;
  LLVM::Type Int32x4Ty;
  LLVM::Type Floatx4Ty;
  LLVM::Type Int64x2Ty;
  LLVM::Type Doublex2Ty;
  LLVM::Type Int128x1Ty;
  LLVM::Type Int8PtrTy;
  LLVM::Type Int32PtrTy;
  LLVM::Type Int64PtrTy;
  LLVM::Type Int128PtrTy;
  LLVM::Type Int8PtrPtrTy;
  LLVM::Type ExecCtxTy;
  LLVM::Type ExecCtxPtrTy;
  LLVM::Type IntrinsicsTableTy;
  LLVM::Type IntrinsicsTablePtrTy;
  LLVM::Message SubtargetFeatures;

#if defined(__x86_64__)
#if defined(__XOP__)
  bool SupportXOP = true;
#else
  bool SupportXOP = false;
#endif

#if defined(__SSE4_1__)
  bool SupportSSE4_1 = true;
#else
  bool SupportSSE4_1 = false;
#endif

#if defined(__SSSE3__)
  bool SupportSSSE3 = true;
#else
  bool SupportSSSE3 = false;
#endif

#if defined(__SSE2__)
  bool SupportSSE2 = true;
#else
  bool SupportSSE2 = false;
#endif
#endif

#if defined(__aarch64__)
#if defined(__ARM_NEON__) || defined(__ARM_NEON) || defined(__ARM_NEON_FP)
  bool SupportNEON = true;
#else
  bool SupportNEON = false;
#endif
#endif

  std::vector<const AST::CompositeType *> CompositeTypes;
  std::vector<LLVM::Value> FunctionWrappers;
  std::vector<std::tuple<uint32_t, LLVM::FunctionCallee,
                         const WasmEdge::AST::CodeSegment *>>
      Functions;
  std::vector<LLVM::Value> LazyJITCacheVars;
  uint32_t ImportCount = 0;
  std::vector<LLVM::Type> MemoryAddrTypes;
  std::vector<LLVM::Type> TableAddrTypes;
  std::vector<LLVM::Type> Globals;
  LLVM::Value IntrinsicsTable;
  LLVM::FunctionCallee Trap;
  CompileContext(LLVM::Context C, LLVM::Module &M,
                 bool IsGenericBinary) noexcept
      : LLContext(C), LLModule(M),
        Cold(LLVM::Attribute::createEnum(C, LLVM::Core::Cold, 0)),
        NoAlias(LLVM::Attribute::createEnum(C, LLVM::Core::NoAlias, 0)),
        NoInline(LLVM::Attribute::createEnum(C, LLVM::Core::NoInline, 0)),
        NoReturn(LLVM::Attribute::createEnum(C, LLVM::Core::NoReturn, 0)),
        ReadOnly(LLVM::Attribute::createEnum(C, LLVM::Core::ReadOnly, 0)),
        StrictFP(LLVM::Attribute::createEnum(C, LLVM::Core::StrictFP, 0)),
        UWTable(LLVM::Attribute::createEnum(C, LLVM::Core::UWTable,
                                            LLVM::Core::UWTableDefault)),
        NoStackArgProbe(
            LLVM::Attribute::createString(C, "no-stack-arg-probe"sv, {})),
        VoidTy(LLContext.getVoidTy()), Int8Ty(LLContext.getInt8Ty()),
        Int16Ty(LLContext.getInt16Ty()), Int32Ty(LLContext.getInt32Ty()),
        Int64Ty(LLContext.getInt64Ty()), Int128Ty(LLContext.getInt128Ty()),
        FloatTy(LLContext.getFloatTy()), DoubleTy(LLContext.getDoubleTy()),
        Int8x16Ty(LLVM::Type::getVectorType(Int8Ty, 16)),
        Int16x8Ty(LLVM::Type::getVectorType(Int16Ty, 8)),
        Int32x4Ty(LLVM::Type::getVectorType(Int32Ty, 4)),
        Floatx4Ty(LLVM::Type::getVectorType(FloatTy, 4)),
        Int64x2Ty(LLVM::Type::getVectorType(Int64Ty, 2)),
        Doublex2Ty(LLVM::Type::getVectorType(DoubleTy, 2)),
        Int128x1Ty(LLVM::Type::getVectorType(Int128Ty, 1)),
        Int8PtrTy(Int8Ty.getPointerTo()), Int32PtrTy(Int32Ty.getPointerTo()),
        Int64PtrTy(Int64Ty.getPointerTo()),
        Int128PtrTy(Int128Ty.getPointerTo()),
        Int8PtrPtrTy(Int8PtrTy.getPointerTo()),
        ExecCtxTy(LLVM::Type::getStructType(
            "ExecCtx",
            std::initializer_list<LLVM::Type>{
                // Memory
                Int8PtrTy.getPointerTo(),
                // Globals
                Int128PtrTy.getPointerTo(),
                // InstrCount
                Int64PtrTy,
                // CostTable
                LLVM::Type::getArrayType(Int64Ty, UINT16_MAX + 1)
                    .getPointerTo(),
                // Gas
                Int64PtrTy,
                // GasLimit
                Int64Ty,
                // StopToken
                Int32PtrTy,
            })),
        ExecCtxPtrTy(ExecCtxTy.getPointerTo()),
        IntrinsicsTableTy(LLVM::Type::getArrayType(
            Int8Ty.getPointerTo(),
            static_cast<uint32_t>(Executable::Intrinsics::kIntrinsicMax))),
        IntrinsicsTablePtrTy(IntrinsicsTableTy.getPointerTo()),
        IntrinsicsTable(LLModule.get().addGlobal(IntrinsicsTablePtrTy, true,
                                                 LLVMExternalLinkage,
                                                 LLVM::Value(), "intrinsics")) {
    Trap.Ty = LLVM::Type::getFunctionType(VoidTy, {Int32Ty});
    Trap.Fn = LLModule.get().addFunction(Trap.Ty, LLVMPrivateLinkage, "trap");
    Trap.Fn.setDSOLocal(true);
    Trap.Fn.addFnAttr(NoStackArgProbe);
    Trap.Fn.addFnAttr(StrictFP);
    Trap.Fn.addFnAttr(UWTable);
    Trap.Fn.addFnAttr(NoReturn);
    Trap.Fn.addFnAttr(Cold);
    Trap.Fn.addFnAttr(NoInline);

    if (!IsGenericBinary) {
      SubtargetFeatures = LLVM::getHostCPUFeatures();
      auto Features = SubtargetFeatures.string_view();
      while (!Features.empty()) {
        std::string_view Feature;
        if (auto Pos = Features.find(','); Pos != std::string_view::npos) {
          Feature = Features.substr(0, Pos);
          Features = Features.substr(Pos + 1);
        } else {
          Feature = std::exchange(Features, std::string_view());
        }
        if (Feature[0] != '+') {
          continue;
        }
        Feature = Feature.substr(1);

#if defined(__x86_64__)
        if (!SupportXOP && Feature == "xop"sv) {
          SupportXOP = true;
        }
        if (!SupportSSE4_1 && Feature == "sse4.1"sv) {
          SupportSSE4_1 = true;
        }
        if (!SupportSSSE3 && Feature == "ssse3"sv) {
          SupportSSSE3 = true;
        }
        if (!SupportSSE2 && Feature == "sse2"sv) {
          SupportSSE2 = true;
        }
#elif defined(__aarch64__)
        if (!SupportNEON && Feature == "neon"sv) {
          SupportNEON = true;
        }
#endif
      }
    }

    compileTrap();
  }
  LLVM::Value getMemory(LLVM::Builder &Builder, LLVM::Value ExecCtx,
                        uint32_t Index) noexcept {
    auto Array = Builder.createExtractValue(ExecCtx, 0);
#if WASMEDGE_ALLOCATOR_IS_STABLE
    auto VPtr = Builder.createLoad(
        Int8PtrTy, Builder.createInBoundsGEP1(Int8PtrTy, Array,
                                              LLContext.getInt64(Index)));
    VPtr.setMetadata(LLContext, LLVM::Core::InvariantGroup,
                     LLVM::Metadata(LLContext, {}));
#else
    auto VPtrPtr = Builder.createLoad(
        Int8PtrPtrTy, Builder.createInBoundsGEP1(Int8PtrPtrTy, Array,
                                                 LLContext.getInt64(Index)));
    VPtrPtr.setMetadata(LLContext, LLVM::Core::InvariantGroup,
                        LLVM::Metadata(LLContext, {}));
    auto VPtr = Builder.createLoad(
        Int8PtrTy,
        Builder.createInBoundsGEP1(Int8PtrTy, VPtrPtr, LLContext.getInt64(0)));
#endif
    return Builder.createBitCast(VPtr, Int8PtrTy);
  }
  std::pair<LLVM::Type, LLVM::Value> getGlobal(LLVM::Builder &Builder,
                                               LLVM::Value ExecCtx,
                                               uint32_t Index) noexcept {
    auto Ty = Globals[Index];
    auto Array = Builder.createExtractValue(ExecCtx, 1);
    auto VPtr = Builder.createLoad(
        Int128PtrTy, Builder.createInBoundsGEP1(Int8PtrTy, Array,
                                                LLContext.getInt64(Index)));
    VPtr.setMetadata(LLContext, LLVM::Core::InvariantGroup,
                     LLVM::Metadata(LLContext, {}));
    auto Ptr = Builder.createBitCast(VPtr, Ty.getPointerTo());
    return {Ty, Ptr};
  }
  LLVM::Value getInstrCount(LLVM::Builder &Builder,
                            LLVM::Value ExecCtx) noexcept {
    return Builder.createExtractValue(ExecCtx, 2);
  }
  LLVM::Value getCostTable(LLVM::Builder &Builder,
                           LLVM::Value ExecCtx) noexcept {
    return Builder.createExtractValue(ExecCtx, 3);
  }
  LLVM::Value getGas(LLVM::Builder &Builder, LLVM::Value ExecCtx) noexcept {
    return Builder.createExtractValue(ExecCtx, 4);
  }
  LLVM::Value getGasLimit(LLVM::Builder &Builder,
                          LLVM::Value ExecCtx) noexcept {
    return Builder.createExtractValue(ExecCtx, 5);
  }
  LLVM::Value getStopToken(LLVM::Builder &Builder,
                           LLVM::Value ExecCtx) noexcept {
    return Builder.createExtractValue(ExecCtx, 6);
  }
  LLVM::FunctionCallee getIntrinsic(LLVM::Builder &Builder,
                                    Executable::Intrinsics Index,
                                    LLVM::Type Ty) noexcept {
    const auto Value = static_cast<uint32_t>(Index);
    auto PtrTy = Ty.getPointerTo();
    auto PtrPtrTy = PtrTy.getPointerTo();
    auto IT = Builder.createLoad(IntrinsicsTablePtrTy, IntrinsicsTable);
    IT.setMetadata(LLContext, LLVM::Core::InvariantGroup,
                   LLVM::Metadata(LLContext, {}));
    auto VPtr =
        Builder.createInBoundsGEP2(IntrinsicsTableTy, IT, LLContext.getInt64(0),
                                   LLContext.getInt64(Value));
    auto Ptr = Builder.createBitCast(VPtr, PtrPtrTy);
    return {Ty, Builder.createLoad(PtrTy, Ptr)};
  }
  void compileTrap() noexcept {
    LLVM::Builder Builder(LLContext);
    Builder.positionAtEnd(
        LLVM::BasicBlock::create(LLContext, Trap.Fn, "entry"));
    auto FnTy = LLVM::Type::getFunctionType(VoidTy, {Int32Ty});
    auto CallTrap = Builder.createCall(
        getIntrinsic(Builder, Executable::Intrinsics::kTrap, FnTy),
        {Trap.Fn.getFirstParam()});
    CallTrap.addCallSiteAttribute(NoReturn);
    Builder.createUnreachable();
  }
  void addVersionGlobal() noexcept {
    LLModule.get().addGlobal(
        Int32Ty, true, LLVMExternalLinkage,
        LLVM::Value::getConstInt(Int32Ty, AOT::kBinaryVersion), "version");
  }
  void finalizeIntrinsicsTable() noexcept {
    if (auto Table = LLModule.get().getNamedGlobal("intrinsics")) {
      Table.setInitializer(LLVM::Value::getConstNull(Table.getType()));
      Table.setGlobalConstant(false);
    } else {
      LLModule.get().addGlobal(IntrinsicsTablePtrTy, false, LLVMExternalLinkage,
                               LLVM::Value::getConstNull(IntrinsicsTablePtrTy),
                               "intrinsics");
    }
  }
  std::pair<std::vector<ValType>, std::vector<ValType>>
  resolveBlockType(const BlockType &BType) const noexcept {
    using VecT = std::vector<ValType>;
    using RetT = std::pair<VecT, VecT>;
    if (BType.isEmpty()) {
      return RetT{};
    }
    if (BType.isValType()) {
      return RetT{{}, {BType.getValType()}};
    } else {
      // Type index case. t2* = type[index].returns
      const uint32_t TypeIdx = BType.getTypeIndex();
      const auto &FType = CompositeTypes[TypeIdx]->getFuncType();
      return RetT{
          VecT(FType.getParamTypes().begin(), FType.getParamTypes().end()),
          VecT(FType.getReturnTypes().begin(), FType.getReturnTypes().end())};
    }
  }
};

namespace {

using namespace WasmEdge;

static bool isVoidReturn(Span<const ValType> ValTypes) noexcept {
  return ValTypes.empty();
}

static LLVM::Type toLLVMType(LLVM::Context LLContext,
                             const ValType &ValType) noexcept {
  switch (ValType.getCode()) {
  case TypeCode::I32:
    return LLContext.getInt32Ty();
  case TypeCode::I64:
    return LLContext.getInt64Ty();
  case TypeCode::Ref:
  case TypeCode::RefNull:
  case TypeCode::V128:
    return LLVM::Type::getVectorType(LLContext.getInt64Ty(), 2);
  case TypeCode::F32:
    return LLContext.getFloatTy();
  case TypeCode::F64:
    return LLContext.getDoubleTy();
  default:
    assumingUnreachable();
  }
}

static LLVM::Type toLLVMType(LLVM::Context LLContext,
                             const AddressType AddrType) noexcept {
  switch (AddrType) {
  case AddressType::I32:
    return LLContext.getInt32Ty();
  case AddressType::I64:
    return LLContext.getInt64Ty();
  default:
    assumingUnreachable();
  }
}

static std::vector<LLVM::Type>
toLLVMTypeVector(LLVM::Context LLContext,
                 Span<const ValType> ValTypes) noexcept {
  std::vector<LLVM::Type> Result;
  Result.reserve(ValTypes.size());
  for (const auto &Type : ValTypes) {
    Result.push_back(toLLVMType(LLContext, Type));
  }
  return Result;
}

static std::vector<LLVM::Type>
toLLVMArgsType(LLVM::Context LLContext, LLVM::Type ExecCtxPtrTy,
               Span<const ValType> ValTypes) noexcept {
  auto Result = toLLVMTypeVector(LLContext, ValTypes);
  Result.insert(Result.begin(), ExecCtxPtrTy);
  return Result;
}

static LLVM::Type toLLVMRetsType(LLVM::Context LLContext,
                                 Span<const ValType> ValTypes) noexcept {
  if (isVoidReturn(ValTypes)) {
    return LLContext.getVoidTy();
  }
  if (ValTypes.size() == 1) {
    return toLLVMType(LLContext, ValTypes.front());
  }
  std::vector<LLVM::Type> Result;
  Result.reserve(ValTypes.size());
  for (const auto &Type : ValTypes) {
    Result.push_back(toLLVMType(LLContext, Type));
  }
  return LLVM::Type::getStructType(Result);
}

static LLVM::Type toLLVMType(LLVM::Context LLContext, LLVM::Type ExecCtxPtrTy,
                             const AST::FunctionType &FuncType) noexcept {
  auto ArgsTy =
      toLLVMArgsType(LLContext, ExecCtxPtrTy, FuncType.getParamTypes());
  auto RetTy = toLLVMRetsType(LLContext, FuncType.getReturnTypes());
  return LLVM::Type::getFunctionType(RetTy, ArgsTy);
}

static LLVM::Value toLLVMConstantZero(
    LLVM::Context LLContext, const ValType &ValType,
    Span<const AST::CompositeType *const> CompositeTypes) noexcept {
  switch (ValType.getCode()) {
  case TypeCode::I32:
    return LLVM::Value::getConstNull(LLContext.getInt32Ty());
  case TypeCode::I64:
    return LLVM::Value::getConstNull(LLContext.getInt64Ty());
  case TypeCode::Ref:
  case TypeCode::RefNull: {
    std::array<uint8_t, 16> Data{};
    if (ValType.isAbsHeapType()) {
      // Abstract heap types are already fine for null refs.
      const auto Raw = ValType.getRawData();
      std::copy(Raw.begin(), Raw.end(), Data.begin());
    } else {
      // For non-abstract heap types (concrete type indices), convert to the
      // abstract heap type so that ref.cast/ref.test won't dereference a null
      // pointer when checking the type.
      assuming(ValType.getTypeIndex() < CompositeTypes.size());
      const auto *CompType = CompositeTypes[ValType.getTypeIndex()];
      assuming(CompType != nullptr);
      WasmEdge::ValType VType =
          CompType->isFunc() ? TypeCode::NullFuncRef : TypeCode::NullRef;
      std::copy_n(VType.getRawData().cbegin(), 8, Data.begin());
    }
    return LLVM::Value::getConstVector8(LLContext, Data);
  }
  case TypeCode::V128:
    return LLVM::Value::getConstNull(
        LLVM::Type::getVectorType(LLContext.getInt64Ty(), 2));
  case TypeCode::F32:
    return LLVM::Value::getConstNull(LLContext.getFloatTy());
  case TypeCode::F64:
    return LLVM::Value::getConstNull(LLContext.getDoubleTy());
  default:
    assumingUnreachable();
  }
}

class FunctionCompiler {
  struct Control;

public:
  FunctionCompiler(LLVM::Compiler::CompileContext &Context,
                   LLVM::FunctionCallee F, Span<const ValType> Locals,
                   bool Interruptible, bool InstructionCounting,
                   bool GasMeasuring, bool IsLazyJIT) noexcept
      : Context(Context), LLContext(Context.LLContext),
        Interruptible(Interruptible), IsLazyJIT(IsLazyJIT), F(F),
        Builder(LLContext) {
    if (F.Fn) {
      Builder.positionAtEnd(LLVM::BasicBlock::create(LLContext, F.Fn, "entry"));
      ExecCtx = Builder.createLoad(Context.ExecCtxTy, F.Fn.getFirstParam());

      if (InstructionCounting) {
        LocalInstrCount = Builder.createAlloca(Context.Int64Ty);
        Builder.createStore(LLContext.getInt64(0), LocalInstrCount);
      }

      if (GasMeasuring) {
        LocalGas = Builder.createAlloca(Context.Int64Ty);
        Builder.createStore(LLContext.getInt64(0), LocalGas);
      }

      for (LLVM::Value Arg = F.Fn.getFirstParam().getNextParam(); Arg;
           Arg = Arg.getNextParam()) {
        LLVM::Type Ty = Arg.getType();
        LLVM::Value ArgPtr = Builder.createAlloca(Ty);
        Builder.createStore(Arg, ArgPtr);
        Local.emplace_back(Ty, ArgPtr);
      }

      for (const auto &Type : Locals) {
        LLVM::Type Ty = toLLVMType(LLContext, Type);
        LLVM::Value ArgPtr = Builder.createAlloca(Ty);
        Builder.createStore(
            toLLVMConstantZero(LLContext, Type, Context.CompositeTypes),
            ArgPtr);
        Local.emplace_back(Ty, ArgPtr);
      }
    }
  }

  LLVM::BasicBlock getTrapBB(ErrCode::Value Error) noexcept {
    if (auto Iter = TrapBB.find(Error); Iter != TrapBB.end()) {
      return Iter->second;
    }
    auto BB = LLVM::BasicBlock::create(LLContext, F.Fn, "trap");
    TrapBB.emplace(Error, BB);
    return BB;
  }

  Expect<void>
  compile(const AST::CodeSegment &Code,
          std::pair<std::vector<ValType>, std::vector<ValType>> Type) noexcept {
    auto RetBB = LLVM::BasicBlock::create(LLContext, F.Fn, "ret");
    Type.first.clear();
    enterBlock(RetBB, {}, {}, {}, std::move(Type));
    EXPECTED_TRY(compile(Code.getExpr().getInstrs()));
    assuming(ControlStack.empty());
    compileReturn();

    for (auto &[Error, BB] : TrapBB) {
      Builder.positionAtEnd(BB);
      updateInstrCount();
      updateGasAtTrap();
      auto CallTrap = Builder.createCall(
          Context.Trap, {LLContext.getInt32(static_cast<uint32_t>(Error))});
      CallTrap.addCallSiteAttribute(Context.NoReturn);
      Builder.createUnreachable();
    }
    return {};
  }

  Expect<void> compile(AST::InstrView Instrs) noexcept {
    auto Dispatch = [this](const AST::Instruction &Instr) -> Expect<void> {
      switch (Instr.getOpCode()) {
      // Control instructions (for blocks)
      case OpCode::Block: {
        auto Block = LLVM::BasicBlock::create(LLContext, F.Fn, "block");
        auto EndBlock = LLVM::BasicBlock::create(LLContext, F.Fn, "block.end");
        Builder.createBr(Block);

        Builder.positionAtEnd(Block);
        auto Type = Context.resolveBlockType(Instr.getBlockType());
        const auto Arity = Type.first.size();
        std::vector<LLVM::Value> Args(Arity);
        if (isUnreachable()) {
          for (size_t I = 0; I < Arity; ++I) {
            auto Ty = toLLVMType(LLContext, Type.first[I]);
            Args[I] = LLVM::Value::getUndef(Ty);
          }
        } else {
          for (size_t I = 0; I < Arity; ++I) {
            const size_t J = Arity - 1 - I;
            Args[J] = stackPop();
          }
        }
        enterBlock(EndBlock, {}, {}, std::move(Args), std::move(Type));
        checkStop();
        updateGas();
        return {};
      }
      case OpCode::Loop: {
        auto Curr = Builder.getInsertBlock();
        auto Loop = LLVM::BasicBlock::create(LLContext, F.Fn, "loop");
        auto EndLoop = LLVM::BasicBlock::create(LLContext, F.Fn, "loop.end");
        Builder.createBr(Loop);

        Builder.positionAtEnd(Loop);
        auto Type = Context.resolveBlockType(Instr.getBlockType());
        const auto Arity = Type.first.size();
        std::vector<LLVM::Value> Args(Arity);
        if (isUnreachable()) {
          for (size_t I = 0; I < Arity; ++I) {
            auto Ty = toLLVMType(LLContext, Type.first[I]);
            auto Value = LLVM::Value::getUndef(Ty);
            auto PHINode = Builder.createPHI(Ty);
            PHINode.addIncoming(Value, Curr);
            Args[I] = PHINode;
          }
        } else {
          for (size_t I = 0; I < Arity; ++I) {
            const size_t J = Arity - 1 - I;
            auto Value = stackPop();
            auto PHINode = Builder.createPHI(Value.getType());
            PHINode.addIncoming(Value, Curr);
            Args[J] = PHINode;
          }
        }
        enterBlock(Loop, EndLoop, {}, std::move(Args), std::move(Type));
        checkStop();
        updateGas();
        return {};
      }
      case OpCode::If: {
        auto Then = LLVM::BasicBlock::create(LLContext, F.Fn, "then");
        auto Else = LLVM::BasicBlock::create(LLContext, F.Fn, "else");
        auto EndIf = LLVM::BasicBlock::create(LLContext, F.Fn, "if.end");
        LLVM::Value Cond;
        if (isUnreachable()) {
          Cond = LLVM::Value::getUndef(LLContext.getInt1Ty());
        } else {
          Cond = Builder.createICmpNE(stackPop(), LLContext.getInt32(0));
        }
        Builder.createCondBr(Cond, Then, Else);

        Builder.positionAtEnd(Then);
        auto Type = Context.resolveBlockType(Instr.getBlockType());
        const auto Arity = Type.first.size();
        std::vector<LLVM::Value> Args(Arity);
        if (isUnreachable()) {
          for (size_t I = 0; I < Arity; ++I) {
            auto Ty = toLLVMType(LLContext, Type.first[I]);
            Args[I] = LLVM::Value::getUndef(Ty);
          }
        } else {
          for (size_t I = 0; I < Arity; ++I) {
            const size_t J = Arity - 1 - I;
            Args[J] = stackPop();
          }
        }
        enterBlock(EndIf, {}, Else, std::move(Args), std::move(Type));
        return {};
      }
      case OpCode::Try_table:
        // TODO: EXCEPTION - implement the AOT.
        return Unexpect(ErrCode::Value::AOTNotImpl);
      case OpCode::End: {
        auto Entry = leaveBlock();
        if (Entry.ElseBlock) {
          auto Block = Builder.getInsertBlock();
          Builder.positionAtEnd(Entry.ElseBlock);
          enterBlock(Block, {}, {}, std::move(Entry.Args),
                     std::move(Entry.Type), std::move(Entry.ReturnPHI));
          Entry = leaveBlock();
        }
        buildPHI(Entry.Type.second, Entry.ReturnPHI);
        return {};
      }
      case OpCode::Else: {
        auto Entry = leaveBlock();
        Builder.positionAtEnd(Entry.ElseBlock);
        enterBlock(Entry.JumpBlock, {}, {}, std::move(Entry.Args),
                   std::move(Entry.Type), std::move(Entry.ReturnPHI));
        return {};
      }
      default:
        break;
      }

      if (isUnreachable()) {
        return {};
      }

      switch (Instr.getOpCode()) {
      // Control instructions
      case OpCode::Unreachable:
        Builder.createBr(getTrapBB(ErrCode::Value::Unreachable));
        setUnreachable();
        Builder.positionAtEnd(
            LLVM::BasicBlock::create(LLContext, F.Fn, "unreachable.end"));
        break;
      case OpCode::Nop:
        break;
      case OpCode::Throw:
      case OpCode::Throw_ref:
        // TODO: EXCEPTION - implement the AOT.
        return Unexpect(ErrCode::Value::AOTNotImpl);
      case OpCode::Br: {
        const auto Label = Instr.getJump().TargetIndex;
        setLableJumpPHI(Label);
        Builder.createBr(getLabel(Label));
        setUnreachable();
        Builder.positionAtEnd(
            LLVM::BasicBlock::create(LLContext, F.Fn, "br.end"));
        break;
      }
      case OpCode::Br_if: {
        const auto Label = Instr.getJump().TargetIndex;
        auto Cond = Builder.createICmpNE(stackPop(), LLContext.getInt32(0));
        setLableJumpPHI(Label);
        auto Next = LLVM::BasicBlock::create(LLContext, F.Fn, "br_if.end");
        Builder.createCondBr(Cond, getLabel(Label), Next);
        Builder.positionAtEnd(Next);
        break;
      }
      case OpCode::Br_table: {
        auto LabelTable = Instr.getLabelList();
        assuming(LabelTable.size() <= std::numeric_limits<uint32_t>::max());
        const auto LabelTableSize =
            static_cast<uint32_t>(LabelTable.size() - 1);
        auto Value = stackPop();
        setLableJumpPHI(LabelTable[LabelTableSize].TargetIndex);
        auto Switch = Builder.createSwitch(
            Value, getLabel(LabelTable[LabelTableSize].TargetIndex),
            LabelTableSize);
        for (uint32_t I = 0; I < LabelTableSize; ++I) {
          setLableJumpPHI(LabelTable[I].TargetIndex);
          Switch.addCase(LLContext.getInt32(I),
                         getLabel(LabelTable[I].TargetIndex));
        }
        setUnreachable();
        Builder.positionAtEnd(
            LLVM::BasicBlock::create(LLContext, F.Fn, "br_table.end"));
        break;
      }
      case OpCode::Br_on_null: {
        const auto Label = Instr.getJump().TargetIndex;
        auto Value = Builder.createBitCast(stackPop(), Context.Int64x2Ty);
        auto Cond = Builder.createICmpEQ(
            Builder.createExtractElement(Value, LLContext.getInt64(1)),
            LLContext.getInt64(0));
        setLableJumpPHI(Label);
        auto Next = LLVM::BasicBlock::create(LLContext, F.Fn, "br_on_null.end");
        Builder.createCondBr(Cond, getLabel(Label), Next);
        Builder.positionAtEnd(Next);
        stackPush(Value);
        break;
      }
      case OpCode::Br_on_non_null: {
        const auto Label = Instr.getJump().TargetIndex;
        auto Cond = Builder.createICmpNE(
            Builder.createExtractElement(
                Builder.createBitCast(Stack.back(), Context.Int64x2Ty),
                LLContext.getInt64(1)),
            LLContext.getInt64(0));
        setLableJumpPHI(Label);
        auto Next =
            LLVM::BasicBlock::create(LLContext, F.Fn, "br_on_non_null.end");
        Builder.createCondBr(Cond, getLabel(Label), Next);
        Builder.positionAtEnd(Next);
        stackPop();
        break;
      }
      case OpCode::Br_on_cast:
      case OpCode::Br_on_cast_fail: {
        auto Ref = Builder.createBitCast(Stack.back(), Context.Int64x2Ty);
        const auto Label = Instr.getBrCast().Jump.TargetIndex;
        std::array<uint8_t, 16> Buf = {0};
        std::copy_n(Instr.getBrCast().RType2.getRawData().cbegin(), 8,
                    Buf.begin());
        auto VType = Builder.createExtractElement(
            Builder.createBitCast(LLVM::Value::getConstVector8(LLContext, Buf),
                                  Context.Int64x2Ty),
            LLContext.getInt64(0));
        auto IsRefTest = Builder.createCall(
            Context.getIntrinsic(Builder, Executable::Intrinsics::kRefTest,
                                 LLVM::Type::getFunctionType(
                                     Context.Int32Ty,
                                     {Context.Int64x2Ty, Context.Int64Ty},
                                     false)),
            {Ref, VType});
        auto Cond =
            (Instr.getOpCode() == OpCode::Br_on_cast)
                ? Builder.createICmpNE(IsRefTest, LLContext.getInt32(0))
                : Builder.createICmpEQ(IsRefTest, LLContext.getInt32(0));
        setLableJumpPHI(Label);
        auto Next = LLVM::BasicBlock::create(LLContext, F.Fn, "br_on_cast.end");
        Builder.createCondBr(Cond, getLabel(Label), Next);
        Builder.positionAtEnd(Next);
        break;
      }
      case OpCode::Return:
        compileReturn();
        setUnreachable();
        Builder.positionAtEnd(
            LLVM::BasicBlock::create(LLContext, F.Fn, "ret.end"));
        break;
      case OpCode::Call:
        updateInstrCount();
        updateGas();
        compileCallOp(Instr.getTargetIndex());
        break;
      case OpCode::Call_indirect:
        updateInstrCount();
        updateGas();
        compileIndirectCallOp(Instr.getSourceIndex(), Instr.getTargetIndex());
        break;
      case OpCode::Return_call:
        updateInstrCount();
        updateGas();
        compileReturnCallOp(Instr.getTargetIndex());
        setUnreachable();
        Builder.positionAtEnd(
            LLVM::BasicBlock::create(LLContext, F.Fn, "ret_call.end"));
        break;
      case OpCode::Return_call_indirect:
        updateInstrCount();
        updateGas();
        compileReturnIndirectCallOp(Instr.getSourceIndex(),
                                    Instr.getTargetIndex());
        setUnreachable();
        Builder.positionAtEnd(
            LLVM::BasicBlock::create(LLContext, F.Fn, "ret_call_indir.end"));
        break;
      case OpCode::Call_ref:
        updateInstrCount();
        updateGas();
        compileCallRefOp(Instr.getTargetIndex());
        break;
      case OpCode::Return_call_ref:
        updateInstrCount();
        updateGas();
        compileReturnCallRefOp(Instr.getTargetIndex());
        setUnreachable();
        Builder.positionAtEnd(
            LLVM::BasicBlock::create(LLContext, F.Fn, "ret_call_ref.end"));
        break;
      case OpCode::Try_table:
        // TODO: EXCEPTION - implement the AOT.
        return Unexpect(ErrCode::Value::AOTNotImpl);

      // Reference Instructions
      case OpCode::Ref__null: {
        std::array<uint8_t, 16> Buf = {0};
        // For null references, dynamic type downscaling is needed.
        ValType VType;
        if (Instr.getValType().isAbsHeapType()) {
          switch (Instr.getValType().getHeapTypeCode()) {
          case TypeCode::NullFuncRef:
          case TypeCode::FuncRef:
            VType = TypeCode::NullFuncRef;
            break;
          case TypeCode::NullExternRef:
          case TypeCode::ExternRef:
            VType = TypeCode::NullExternRef;
            break;
          case TypeCode::NullExnRef:
          case TypeCode::ExnRef:
            VType = TypeCode::NullExnRef;
            break;
          case TypeCode::NullRef:
          case TypeCode::AnyRef:
          case TypeCode::EqRef:
          case TypeCode::I31Ref:
          case TypeCode::StructRef:
          case TypeCode::ArrayRef:
            VType = TypeCode::NullRef;
            break;
          default:
            assumingUnreachable();
          }
        } else {
          assuming(Instr.getValType().getTypeIndex() <
                   Context.CompositeTypes.size());
          const auto *CompType =
              Context.CompositeTypes[Instr.getValType().getTypeIndex()];
          assuming(CompType != nullptr);
          if (CompType->isFunc()) {
            VType = TypeCode::NullFuncRef;
          } else {
            VType = TypeCode::NullRef;
          }
        }
        std::copy_n(VType.getRawData().cbegin(), 8, Buf.begin());
        stackPush(Builder.createBitCast(
            LLVM::Value::getConstVector8(LLContext, Buf), Context.Int64x2Ty));
        break;
      }
      case OpCode::Ref__is_null:
        stackPush(Builder.createZExt(
            Builder.createICmpEQ(
                Builder.createExtractElement(
                    Builder.createBitCast(stackPop(), Context.Int64x2Ty),
                    LLContext.getInt64(1)),
                LLContext.getInt64(0)),
            Context.Int32Ty));
        break;
      case OpCode::Ref__func:
        stackPush(Builder.createCall(
            Context.getIntrinsic(Builder, Executable::Intrinsics::kRefFunc,
                                 LLVM::Type::getFunctionType(Context.Int64x2Ty,
                                                             {Context.Int32Ty},
                                                             false)),
            {LLContext.getInt32(Instr.getTargetIndex())}));
        break;
      case OpCode::Ref__eq: {
        LLVM::Value RHS = stackPop();
        LLVM::Value LHS = stackPop();
        stackPush(Builder.createZExt(
            Builder.createICmpEQ(
                Builder.createExtractElement(LHS, LLContext.getInt64(1)),
                Builder.createExtractElement(RHS, LLContext.getInt64(1))),
            Context.Int32Ty));
        break;
      }
      case OpCode::Ref__as_non_null: {
        auto Next =
            LLVM::BasicBlock::create(LLContext, F.Fn, "ref_as_non_null.ok");
        Stack.back() = Builder.createBitCast(Stack.back(), Context.Int64x2Ty);
        auto IsNotNull = Builder.createLikely(Builder.createICmpNE(
            Builder.createExtractElement(Stack.back(), LLContext.getInt64(1)),
            LLContext.getInt64(0)));
        Builder.createCondBr(IsNotNull, Next,
                             getTrapBB(ErrCode::Value::CastNullToNonNull));
        Builder.positionAtEnd(Next);
        break;
      }

      // Reference Instructions (GC proposal)
      case OpCode::Struct__new:
      case OpCode::Struct__new_default: {
        LLVM::Value Args = LLVM::Value::getConstPointerNull(Context.Int8PtrTy);
        assuming(Instr.getTargetIndex() < Context.CompositeTypes.size());
        const auto *CompType = Context.CompositeTypes[Instr.getTargetIndex()];
        assuming(CompType != nullptr && !CompType->isFunc());
        auto ArgSize = CompType->getFieldTypes().size();
        if (Instr.getOpCode() == OpCode::Struct__new) {
          std::vector<LLVM::Value> ArgsVec(ArgSize, nullptr);
          for (size_t I = 0; I < ArgSize; ++I) {
            ArgsVec[ArgSize - I - 1] = stackPop();
          }
          Args = Builder.createArray(ArgSize, kValSize);
          Builder.createArrayPtrStore(ArgsVec, Args, Context.Int8Ty, kValSize);
        } else {
          ArgSize = 0;
        }
        stackPush(Builder.createCall(
            Context.getIntrinsic(
                Builder, Executable::Intrinsics::kStructNew,
                LLVM::Type::getFunctionType(
                    Context.Int64x2Ty,
                    {Context.Int32Ty, Context.Int8PtrTy, Context.Int32Ty},
                    false)),
            {LLContext.getInt32(Instr.getTargetIndex()), Args,
             LLContext.getInt32(static_cast<uint32_t>(ArgSize))}));
        break;
      }
      case OpCode::Struct__get:
      case OpCode::Struct__get_u:
      case OpCode::Struct__get_s: {
        assuming(static_cast<size_t>(Instr.getTargetIndex()) <
                 Context.CompositeTypes.size());
        const auto *CompType = Context.CompositeTypes[Instr.getTargetIndex()];
        assuming(CompType != nullptr && !CompType->isFunc());
        assuming(static_cast<size_t>(Instr.getSourceIndex()) <
                 CompType->getFieldTypes().size());
        const auto &StorageType =
            CompType->getFieldTypes()[Instr.getSourceIndex()].getStorageType();
        auto Ref = stackPop();
        auto IsSigned = (Instr.getOpCode() == OpCode::Struct__get_s)
                            ? LLContext.getInt8(1)
                            : LLContext.getInt8(0);
        LLVM::Value Ret = Builder.createAlloca(Context.Int64x2Ty);
        Builder.createCall(
            Context.getIntrinsic(
                Builder, Executable::Intrinsics::kStructGet,
                LLVM::Type::getFunctionType(Context.VoidTy,
                                            {Context.Int64x2Ty, Context.Int32Ty,
                                             Context.Int32Ty, Context.Int8Ty,
                                             Context.Int8PtrTy},
                                            false)),
            {Ref, LLContext.getInt32(Instr.getTargetIndex()),
             LLContext.getInt32(Instr.getSourceIndex()), IsSigned, Ret});

        switch (StorageType.getCode()) {
        case TypeCode::I8:
        case TypeCode::I16:
        case TypeCode::I32: {
          stackPush(Builder.createValuePtrLoad(Context.Int32Ty, Ret,
                                               Context.Int64x2Ty));
          break;
        }
        case TypeCode::I64: {
          stackPush(Builder.createValuePtrLoad(Context.Int64Ty, Ret,
                                               Context.Int64x2Ty));
          break;
        }
        case TypeCode::F32: {
          stackPush(Builder.createValuePtrLoad(Context.FloatTy, Ret,
                                               Context.Int64x2Ty));
          break;
        }
        case TypeCode::F64: {
          stackPush(Builder.createValuePtrLoad(Context.DoubleTy, Ret,
                                               Context.Int64x2Ty));
          break;
        }
        case TypeCode::V128:
        case TypeCode::Ref:
        case TypeCode::RefNull: {
          stackPush(Builder.createValuePtrLoad(Context.Int64x2Ty, Ret,
                                               Context.Int64x2Ty));
          break;
        }
        default:
          assumingUnreachable();
        }
        break;
      }
      case OpCode::Struct__set: {
        auto Val = stackPop();
        auto Ref = stackPop();
        LLVM::Value Arg = Builder.createAlloca(Context.Int64x2Ty);
        Builder.createValuePtrStore(Val, Arg, Context.Int64x2Ty);
        Builder.createCall(
            Context.getIntrinsic(Builder, Executable::Intrinsics::kStructSet,
                                 LLVM::Type::getFunctionType(
                                     Context.VoidTy,
                                     {Context.Int64x2Ty, Context.Int32Ty,
                                      Context.Int32Ty, Context.Int8PtrTy},
                                     false)),
            {Ref, LLContext.getInt32(Instr.getTargetIndex()),
             LLContext.getInt32(Instr.getSourceIndex()), Arg});
        break;
      }
      case OpCode::Array__new: {
        auto Length = stackPop();
        auto Val = stackPop();
        LLVM::Value Arg = Builder.createAlloca(Context.Int64x2Ty);
        Builder.createValuePtrStore(Val, Arg, Context.Int64x2Ty);
        stackPush(Builder.createCall(
            Context.getIntrinsic(Builder, Executable::Intrinsics::kArrayNew,
                                 LLVM::Type::getFunctionType(
                                     Context.Int64x2Ty,
                                     {Context.Int32Ty, Context.Int32Ty,
                                      Context.Int8PtrTy, Context.Int32Ty},
                                     false)),
            {LLContext.getInt32(Instr.getTargetIndex()), Length, Arg,
             LLContext.getInt32(1)}));
        break;
      }
      case OpCode::Array__new_default: {
        auto Length = stackPop();
        LLVM::Value Arg = LLVM::Value::getConstPointerNull(Context.Int8PtrTy);
        stackPush(Builder.createCall(
            Context.getIntrinsic(Builder, Executable::Intrinsics::kArrayNew,
                                 LLVM::Type::getFunctionType(
                                     Context.Int64x2Ty,
                                     {Context.Int32Ty, Context.Int32Ty,
                                      Context.Int8PtrTy, Context.Int32Ty},
                                     false)),
            {LLContext.getInt32(Instr.getTargetIndex()), Length, Arg,
             LLContext.getInt32(0)}));
        break;
      }
      case OpCode::Array__new_fixed: {
        const auto ArgSize = Instr.getSourceIndex();
        std::vector<LLVM::Value> ArgsVec(ArgSize, nullptr);
        for (size_t I = 0; I < ArgSize; ++I) {
          ArgsVec[ArgSize - I - 1] = stackPop();
        }
        LLVM::Value Args = Builder.createArray(ArgSize, kValSize);
        Builder.createArrayPtrStore(ArgsVec, Args, Context.Int8Ty, kValSize);
        stackPush(Builder.createCall(
            Context.getIntrinsic(Builder, Executable::Intrinsics::kArrayNew,
                                 LLVM::Type::getFunctionType(
                                     Context.Int64x2Ty,
                                     {Context.Int32Ty, Context.Int32Ty,
                                      Context.Int8PtrTy, Context.Int32Ty},
                                     false)),
            {LLContext.getInt32(Instr.getTargetIndex()),
             LLContext.getInt32(ArgSize), Args, LLContext.getInt32(ArgSize)}));
        break;
      }
      case OpCode::Array__new_data:
      case OpCode::Array__new_elem: {
        auto Length = stackPop();
        auto Start = stackPop();
        stackPush(Builder.createCall(
            Context.getIntrinsic(
                Builder,
                ((Instr.getOpCode() == OpCode::Array__new_data)
                     ? Executable::Intrinsics::kArrayNewData
                     : Executable::Intrinsics::kArrayNewElem),
                LLVM::Type::getFunctionType(Context.Int64x2Ty,
                                            {Context.Int32Ty, Context.Int32Ty,
                                             Context.Int32Ty, Context.Int32Ty},
                                            false)),
            {LLContext.getInt32(Instr.getTargetIndex()),
             LLContext.getInt32(Instr.getSourceIndex()), Start, Length}));
        break;
      }
      case OpCode::Array__get:
      case OpCode::Array__get_u:
      case OpCode::Array__get_s: {
        assuming(static_cast<size_t>(Instr.getTargetIndex()) <
                 Context.CompositeTypes.size());
        const auto *CompType = Context.CompositeTypes[Instr.getTargetIndex()];
        assuming(CompType != nullptr && !CompType->isFunc());
        assuming(static_cast<size_t>(1) == CompType->getFieldTypes().size());
        const auto &StorageType = CompType->getFieldTypes()[0].getStorageType();
        auto Idx = stackPop();
        auto Ref = stackPop();
        auto IsSigned = (Instr.getOpCode() == OpCode::Array__get_s)
                            ? LLContext.getInt8(1)
                            : LLContext.getInt8(0);
        LLVM::Value Ret = Builder.createAlloca(Context.Int64x2Ty);
        Builder.createCall(
            Context.getIntrinsic(
                Builder, Executable::Intrinsics::kArrayGet,
                LLVM::Type::getFunctionType(Context.VoidTy,
                                            {Context.Int64x2Ty, Context.Int32Ty,
                                             Context.Int32Ty, Context.Int8Ty,
                                             Context.Int8PtrTy},
                                            false)),
            {Ref, LLContext.getInt32(Instr.getTargetIndex()), Idx, IsSigned,
             Ret});

        switch (StorageType.getCode()) {
        case TypeCode::I8:
        case TypeCode::I16:
        case TypeCode::I32: {
          stackPush(Builder.createValuePtrLoad(Context.Int32Ty, Ret,
                                               Context.Int64x2Ty));
          break;
        }
        case TypeCode::I64: {
          stackPush(Builder.createValuePtrLoad(Context.Int64Ty, Ret,
                                               Context.Int64x2Ty));
          break;
        }
        case TypeCode::F32: {
          stackPush(Builder.createValuePtrLoad(Context.FloatTy, Ret,
                                               Context.Int64x2Ty));
          break;
        }
        case TypeCode::F64: {
          stackPush(Builder.createValuePtrLoad(Context.DoubleTy, Ret,
                                               Context.Int64x2Ty));
          break;
        }
        case TypeCode::V128:
        case TypeCode::Ref:
        case TypeCode::RefNull: {
          stackPush(Builder.createValuePtrLoad(Context.Int64x2Ty, Ret,
                                               Context.Int64x2Ty));
          break;
        }
        default:
          assumingUnreachable();
        }
        break;
      }
      case OpCode::Array__set: {
        auto Val = stackPop();
        auto Idx = stackPop();
        auto Ref = stackPop();
        LLVM::Value Arg = Builder.createAlloca(Context.Int64x2Ty);
        Builder.createValuePtrStore(Val, Arg, Context.Int64x2Ty);
        Builder.createCall(
            Context.getIntrinsic(Builder, Executable::Intrinsics::kArraySet,
                                 LLVM::Type::getFunctionType(
                                     Context.VoidTy,
                                     {Context.Int64x2Ty, Context.Int32Ty,
                                      Context.Int32Ty, Context.Int8PtrTy},
                                     false)),
            {Ref, LLContext.getInt32(Instr.getTargetIndex()), Idx, Arg});
        break;
      }
      case OpCode::Array__len: {
        auto Ref = stackPop();
        stackPush(Builder.createCall(
            Context.getIntrinsic(
                Builder, Executable::Intrinsics::kArrayLen,
                LLVM::Type::getFunctionType(Context.Int32Ty,
                                            {Context.Int64x2Ty}, false)),
            {Ref}));
        break;
      }
      case OpCode::Array__fill: {
        auto Cnt = stackPop();
        auto Val = stackPop();
        auto Off = stackPop();
        auto Ref = stackPop();
        LLVM::Value Arg = Builder.createAlloca(Context.Int64x2Ty);
        Builder.createValuePtrStore(Val, Arg, Context.Int64x2Ty);
        Builder.createCall(
            Context.getIntrinsic(
                Builder, Executable::Intrinsics::kArrayFill,
                LLVM::Type::getFunctionType(Context.VoidTy,
                                            {Context.Int64x2Ty, Context.Int32Ty,
                                             Context.Int32Ty, Context.Int32Ty,
                                             Context.Int8PtrTy},
                                            false)),
            {Ref, LLContext.getInt32(Instr.getTargetIndex()), Off, Cnt, Arg});
        break;
      }
      case OpCode::Array__copy: {
        auto Cnt = stackPop();
        auto SrcOff = stackPop();
        auto SrcRef = stackPop();
        auto DstOff = stackPop();
        auto DstRef = stackPop();
        Builder.createCall(
            Context.getIntrinsic(
                Builder, Executable::Intrinsics::kArrayCopy,
                LLVM::Type::getFunctionType(Context.VoidTy,
                                            {Context.Int64x2Ty, Context.Int32Ty,
                                             Context.Int32Ty, Context.Int64x2Ty,
                                             Context.Int32Ty, Context.Int32Ty,
                                             Context.Int32Ty},
                                            false)),
            {DstRef, LLContext.getInt32(Instr.getTargetIndex()), DstOff, SrcRef,
             LLContext.getInt32(Instr.getSourceIndex()), SrcOff, Cnt});
        break;
      }
      case OpCode::Array__init_data:
      case OpCode::Array__init_elem: {
        auto Cnt = stackPop();
        auto SrcOff = stackPop();
        auto DstOff = stackPop();
        auto Ref = stackPop();
        Builder.createCall(
            Context.getIntrinsic(
                Builder,
                ((Instr.getOpCode() == OpCode::Array__init_data)
                     ? Executable::Intrinsics::kArrayInitData
                     : Executable::Intrinsics::kArrayInitElem),
                LLVM::Type::getFunctionType(Context.VoidTy,
                                            {Context.Int64x2Ty, Context.Int32Ty,
                                             Context.Int32Ty, Context.Int32Ty,
                                             Context.Int32Ty, Context.Int32Ty},
                                            false)),
            {Ref, LLContext.getInt32(Instr.getTargetIndex()),
             LLContext.getInt32(Instr.getSourceIndex()), DstOff, SrcOff, Cnt});
        break;
      }
      case OpCode::Ref__test:
      case OpCode::Ref__test_null: {
        auto Ref = stackPop();
        std::array<uint8_t, 16> Buf = {0};
        std::copy_n(Instr.getValType().getRawData().cbegin(), 8, Buf.begin());
        auto VType = Builder.createExtractElement(
            Builder.createBitCast(LLVM::Value::getConstVector8(LLContext, Buf),
                                  Context.Int64x2Ty),
            LLContext.getInt64(0));
        stackPush(Builder.createCall(
            Context.getIntrinsic(Builder, Executable::Intrinsics::kRefTest,
                                 LLVM::Type::getFunctionType(
                                     Context.Int32Ty,
                                     {Context.Int64x2Ty, Context.Int64Ty},
                                     false)),
            {Ref, VType}));
        break;
      }
      case OpCode::Ref__cast:
      case OpCode::Ref__cast_null: {
        auto Ref = stackPop();
        std::array<uint8_t, 16> Buf = {0};
        std::copy_n(Instr.getValType().getRawData().cbegin(), 8, Buf.begin());
        auto VType = Builder.createExtractElement(
            Builder.createBitCast(LLVM::Value::getConstVector8(LLContext, Buf),
                                  Context.Int64x2Ty),
            LLContext.getInt64(0));
        stackPush(Builder.createCall(
            Context.getIntrinsic(Builder, Executable::Intrinsics::kRefCast,
                                 LLVM::Type::getFunctionType(
                                     Context.Int64x2Ty,
                                     {Context.Int64x2Ty, Context.Int64Ty},
                                     false)),
            {Ref, VType}));
        break;
      }
      case OpCode::Any__convert_extern: {
        std::array<uint8_t, 16> RawRef = {0};
        auto Ref = stackPop();
        auto PtrVal = Builder.createExtractElement(Ref, LLContext.getInt64(1));
        auto IsNullBB =
            LLVM::BasicBlock::create(LLContext, F.Fn, "any_conv_extern.null");
        auto NotNullBB = LLVM::BasicBlock::create(LLContext, F.Fn,
                                                  "any_conv_extern.not_null");
        auto IsExtrefBB = LLVM::BasicBlock::create(LLContext, F.Fn,
                                                   "any_conv_extern.is_extref");
        auto EndBB =
            LLVM::BasicBlock::create(LLContext, F.Fn, "any_conv_extern.end");
        auto CondIsNull = Builder.createICmpEQ(PtrVal, LLContext.getInt64(0));
        Builder.createCondBr(CondIsNull, IsNullBB, NotNullBB);

        Builder.positionAtEnd(IsNullBB);
        auto VT = ValType(TypeCode::RefNull, TypeCode::NullRef);
        std::copy_n(VT.getRawData().cbegin(), 8, RawRef.begin());
        auto Ret1 = Builder.createBitCast(
            LLVM::Value::getConstVector8(LLContext, RawRef), Context.Int64x2Ty);
        Builder.createBr(EndBB);

        Builder.positionAtEnd(NotNullBB);
        auto Ret2 = Builder.createBitCast(
            Builder.createInsertElement(
                Builder.createBitCast(Ref, Context.Int8x16Ty),
                LLContext.getInt8(0), LLContext.getInt64(1)),
            Context.Int64x2Ty);
        auto HType = Builder.createExtractElement(
            Builder.createBitCast(Ret2, Context.Int8x16Ty),
            LLContext.getInt64(3));
        auto CondIsExtref = Builder.createOr(
            Builder.createICmpEQ(HType, LLContext.getInt8(static_cast<uint8_t>(
                                            TypeCode::ExternRef))),
            Builder.createICmpEQ(HType, LLContext.getInt8(static_cast<uint8_t>(
                                            TypeCode::NullExternRef))));
        Builder.createCondBr(CondIsExtref, IsExtrefBB, EndBB);

        Builder.positionAtEnd(IsExtrefBB);
        VT = ValType(TypeCode::Ref, TypeCode::AnyRef);
        std::copy_n(VT.getRawData().cbegin(), 8, RawRef.begin());
        auto Ret3 = Builder.createInsertElement(
            Builder.createBitCast(
                LLVM::Value::getConstVector8(LLContext, RawRef),
                Context.Int64x2Ty),
            PtrVal, LLContext.getInt64(1));
        Builder.createBr(EndBB);

        Builder.positionAtEnd(EndBB);
        auto Ret = Builder.createPHI(Context.Int64x2Ty);
        Ret.addIncoming(Ret1, IsNullBB);
        Ret.addIncoming(Ret2, NotNullBB);
        Ret.addIncoming(Ret3, IsExtrefBB);
        stackPush(Ret);
        break;
      }
      case OpCode::Extern__convert_any: {
        std::array<uint8_t, 16> RawRef = {0};
        auto Ref = stackPop();
        auto IsNullBB =
            LLVM::BasicBlock::create(LLContext, F.Fn, "extern_conv_any.null");
        auto NotNullBB = LLVM::BasicBlock::create(LLContext, F.Fn,
                                                  "extern_conv_any.not_null");
        auto EndBB =
            LLVM::BasicBlock::create(LLContext, F.Fn, "extern_conv_any.end");
        auto CondIsNull = Builder.createICmpEQ(
            Builder.createExtractElement(Ref, LLContext.getInt64(1)),
            LLContext.getInt64(0));
        Builder.createCondBr(CondIsNull, IsNullBB, NotNullBB);

        Builder.positionAtEnd(IsNullBB);
        auto VT = ValType(TypeCode::RefNull, TypeCode::NullExternRef);
        std::copy_n(VT.getRawData().cbegin(), 8, RawRef.begin());
        auto Ret1 = Builder.createBitCast(
            LLVM::Value::getConstVector8(LLContext, RawRef), Context.Int64x2Ty);
        Builder.createBr(EndBB);

        Builder.positionAtEnd(NotNullBB);
        auto Ret2 = Builder.createBitCast(
            Builder.createInsertElement(
                Builder.createBitCast(Ref, Context.Int8x16Ty),
                LLContext.getInt8(1), LLContext.getInt64(1)),
            Context.Int64x2Ty);
        Builder.createBr(EndBB);

        Builder.positionAtEnd(EndBB);
        auto Ret = Builder.createPHI(Context.Int64x2Ty);
        Ret.addIncoming(Ret1, IsNullBB);
        Ret.addIncoming(Ret2, NotNullBB);
        stackPush(Ret);
        break;
      }
      case OpCode::Ref__i31: {
        std::array<uint8_t, 16> RawRef = {0};
        auto VT = ValType(TypeCode::Ref, TypeCode::I31Ref);
        std::copy_n(VT.getRawData().cbegin(), 8, RawRef.begin());
        auto Ref = Builder.createBitCast(
            LLVM::Value::getConstVector8(LLContext, RawRef), Context.Int64x2Ty);
        auto Val = Builder.createZExt(
            Builder.createOr(
                Builder.createAnd(stackPop(), LLContext.getInt32(0x7FFFFFFFU)),
                LLContext.getInt32(0x80000000U)),
            Context.Int64Ty);
        stackPush(Builder.createInsertElement(Ref, Val, LLContext.getInt64(1)));
        break;
      }
      case OpCode::I31__get_s: {
        auto Next = LLVM::BasicBlock::create(LLContext, F.Fn, "i31.get.ok");
        auto Ref = Builder.createBitCast(stackPop(), Context.Int64x2Ty);
        auto Val = Builder.createTrunc(
            Builder.createExtractElement(Ref, LLContext.getInt64(1)),
            Context.Int32Ty);
        auto IsNotNull = Builder.createLikely(Builder.createICmpNE(
            Builder.createAnd(Val, LLContext.getInt32(0x80000000U)),
            LLContext.getInt32(0)));
        Builder.createCondBr(IsNotNull, Next,
                             getTrapBB(ErrCode::Value::AccessNullI31));
        Builder.positionAtEnd(Next);
        Val = Builder.createAnd(Val, LLContext.getInt32(0x7FFFFFFFU));
        stackPush(Builder.createOr(
            Val, Builder.createShl(
                     Builder.createAnd(Val, LLContext.getInt32(0x40000000U)),
                     LLContext.getInt32(1))));
        break;
      }
      case OpCode::I31__get_u: {
        auto Next = LLVM::BasicBlock::create(LLContext, F.Fn, "i31.get.ok");
        auto Ref = Builder.createBitCast(stackPop(), Context.Int64x2Ty);
        auto Val = Builder.createTrunc(
            Builder.createExtractElement(Ref, LLContext.getInt64(1)),
            Context.Int32Ty);
        auto IsNotNull = Builder.createLikely(Builder.createICmpNE(
            Builder.createAnd(Val, LLContext.getInt32(0x80000000U)),
            LLContext.getInt32(0)));
        Builder.createCondBr(IsNotNull, Next,
                             getTrapBB(ErrCode::Value::AccessNullI31));
        Builder.positionAtEnd(Next);
        stackPush(Builder.createAnd(Val, LLContext.getInt32(0x7FFFFFFFU)));
        break;
      }

      // Parametric Instructions
      case OpCode::Drop:
        stackPop();
        break;
      case OpCode::Select:
      case OpCode::Select_t: {
        auto Cond = Builder.createICmpNE(stackPop(), LLContext.getInt32(0));
        auto False = stackPop();
        auto True = stackPop();
        stackPush(Builder.createSelect(Cond, True, False));
        break;
      }

      // Variable Instructions
      case OpCode::Local__get: {
        const auto &L = Local[Instr.getTargetIndex()];
        stackPush(Builder.createLoad(L.first, L.second));
        break;
      }
      case OpCode::Local__set:
        Builder.createStore(stackPop(), Local[Instr.getTargetIndex()].second);
        break;
      case OpCode::Local__tee:
        Builder.createStore(Stack.back(), Local[Instr.getTargetIndex()].second);
        break;
      case OpCode::Global__get: {
        const auto G =
            Context.getGlobal(Builder, ExecCtx, Instr.getTargetIndex());
        stackPush(Builder.createLoad(G.first, G.second));
        break;
      }
      case OpCode::Global__set:
        Builder.createStore(
            stackPop(),
            Context.getGlobal(Builder, ExecCtx, Instr.getTargetIndex()).second);
        break;

      // Table Instructions
      case OpCode::Table__get: {
        auto Off = Builder.createZExt(stackPop(), Context.Int64Ty);
        stackPush(Builder.createCall(
            Context.getIntrinsic(
                Builder, Executable::Intrinsics::kTableGet,
                LLVM::Type::getFunctionType(Context.Int64x2Ty,
                                            {Context.Int32Ty, Context.Int64Ty},
                                            false)),
            {LLContext.getInt32(Instr.getTargetIndex()), Off}));
        break;
      }
      case OpCode::Table__set: {
        auto Ref = stackPop();
        auto Off = Builder.createZExt(stackPop(), Context.Int64Ty);
        Builder.createCall(
            Context.getIntrinsic(
                Builder, Executable::Intrinsics::kTableSet,
                LLVM::Type::getFunctionType(
                    Context.Int64Ty,
                    {Context.Int32Ty, Context.Int64Ty, Context.Int64x2Ty},
                    false)),
            {LLContext.getInt32(Instr.getTargetIndex()), Off, Ref});
        break;
      }
      case OpCode::Table__init: {
        auto Len = stackPop();
        auto Src = stackPop();
        auto Dst = Builder.createZExt(stackPop(), Context.Int64Ty);
        Builder.createCall(
            Context.getIntrinsic(
                Builder, Executable::Intrinsics::kTableInit,
                LLVM::Type::getFunctionType(Context.VoidTy,
                                            {Context.Int32Ty, Context.Int32Ty,
                                             Context.Int64Ty, Context.Int32Ty,
                                             Context.Int32Ty},
                                            false)),
            {LLContext.getInt32(Instr.getTargetIndex()),
             LLContext.getInt32(Instr.getSourceIndex()), Dst, Src, Len});
        break;
      }
      case OpCode::Elem__drop: {
        Builder.createCall(
            Context.getIntrinsic(Builder, Executable::Intrinsics::kElemDrop,
                                 LLVM::Type::getFunctionType(
                                     Context.VoidTy, {Context.Int32Ty}, false)),
            {LLContext.getInt32(Instr.getTargetIndex())});
        break;
      }
      case OpCode::Table__copy: {
        auto Len = Builder.createZExt(stackPop(), Context.Int64Ty);
        auto Src = Builder.createZExt(stackPop(), Context.Int64Ty);
        auto Dst = Builder.createZExt(stackPop(), Context.Int64Ty);
        Builder.createCall(
            Context.getIntrinsic(
                Builder, Executable::Intrinsics::kTableCopy,
                LLVM::Type::getFunctionType(Context.VoidTy,
                                            {Context.Int32Ty, Context.Int32Ty,
                                             Context.Int64Ty, Context.Int64Ty,
                                             Context.Int64Ty},
                                            false)),
            {LLContext.getInt32(Instr.getTargetIndex()),
             LLContext.getInt32(Instr.getSourceIndex()), Dst, Src, Len});
        break;
      }
      case OpCode::Table__grow: {
        auto NewSize = Builder.createZExt(stackPop(), Context.Int64Ty);
        auto Val = stackPop();
        stackPush(Builder.createTrunc(
            Builder.createCall(
                Context.getIntrinsic(
                    Builder, Executable::Intrinsics::kTableGrow,
                    LLVM::Type::getFunctionType(
                        Context.Int64Ty,
                        {Context.Int32Ty, Context.Int64x2Ty, Context.Int64Ty},
                        false)),
                {LLContext.getInt32(Instr.getTargetIndex()), Val, NewSize}),
            Context.TableAddrTypes[Instr.getTargetIndex()]));
        break;
      }
      case OpCode::Table__size: {
        stackPush(Builder.createTrunc(
            Builder.createCall(
                Context.getIntrinsic(
                    Builder, Executable::Intrinsics::kTableSize,
                    LLVM::Type::getFunctionType(Context.Int64Ty,
                                                {Context.Int32Ty}, false)),
                {LLContext.getInt32(Instr.getTargetIndex())}),
            Context.TableAddrTypes[Instr.getTargetIndex()]));
        break;
      }
      case OpCode::Table__fill: {
        auto Len = Builder.createZExt(stackPop(), Context.Int64Ty);
        auto Val = stackPop();
        auto Off = Builder.createZExt(stackPop(), Context.Int64Ty);
        Builder.createCall(
            Context.getIntrinsic(Builder, Executable::Intrinsics::kTableFill,
                                 LLVM::Type::getFunctionType(
                                     Context.Int32Ty,
                                     {Context.Int32Ty, Context.Int64Ty,
                                      Context.Int64x2Ty, Context.Int64Ty},
                                     false)),
            {LLContext.getInt32(Instr.getTargetIndex()), Off, Val, Len});
        break;
      }

      // Memory Instructions
      case OpCode::I32__load:
        compileLoadOp(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                      Instr.getMemoryAlign(), Context.Int32Ty);
        break;
      case OpCode::I64__load:
        compileLoadOp(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                      Instr.getMemoryAlign(), Context.Int64Ty);
        break;
      case OpCode::F32__load:
        compileLoadOp(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                      Instr.getMemoryAlign(), Context.FloatTy);
        break;
      case OpCode::F64__load:
        compileLoadOp(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                      Instr.getMemoryAlign(), Context.DoubleTy);
        break;
      case OpCode::I32__load8_s:
        compileLoadOp(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                      Instr.getMemoryAlign(), Context.Int8Ty, Context.Int32Ty,
                      true);
        break;
      case OpCode::I32__load8_u:
        compileLoadOp(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                      Instr.getMemoryAlign(), Context.Int8Ty, Context.Int32Ty,
                      false);
        break;
      case OpCode::I32__load16_s:
        compileLoadOp(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                      Instr.getMemoryAlign(), Context.Int16Ty, Context.Int32Ty,
                      true);
        break;
      case OpCode::I32__load16_u:
        compileLoadOp(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                      Instr.getMemoryAlign(), Context.Int16Ty, Context.Int32Ty,
                      false);
        break;
      case OpCode::I64__load8_s:
        compileLoadOp(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                      Instr.getMemoryAlign(), Context.Int8Ty, Context.Int64Ty,
                      true);
        break;
      case OpCode::I64__load8_u:
        compileLoadOp(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                      Instr.getMemoryAlign(), Context.Int8Ty, Context.Int64Ty,
                      false);
        break;
      case OpCode::I64__load16_s:
        compileLoadOp(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                      Instr.getMemoryAlign(), Context.Int16Ty, Context.Int64Ty,
                      true);
        break;
      case OpCode::I64__load16_u:
        compileLoadOp(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                      Instr.getMemoryAlign(), Context.Int16Ty, Context.Int64Ty,
                      false);
        break;
      case OpCode::I64__load32_s:
        compileLoadOp(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                      Instr.getMemoryAlign(), Context.Int32Ty, Context.Int64Ty,
                      true);
        break;
      case OpCode::I64__load32_u:
        compileLoadOp(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                      Instr.getMemoryAlign(), Context.Int32Ty, Context.Int64Ty,
                      false);
        break;
      case OpCode::I32__store:
        compileStoreOp(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                       Instr.getMemoryAlign(), Context.Int32Ty);
        break;
      case OpCode::I64__store:
        compileStoreOp(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                       Instr.getMemoryAlign(), Context.Int64Ty);
        break;
      case OpCode::F32__store:
        compileStoreOp(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                       Instr.getMemoryAlign(), Context.FloatTy);
        break;
      case OpCode::F64__store:
        compileStoreOp(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                       Instr.getMemoryAlign(), Context.DoubleTy);
        break;
      case OpCode::I32__store8:
      case OpCode::I64__store8:
        compileStoreOp(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                       Instr.getMemoryAlign(), Context.Int8Ty, true);
        break;
      case OpCode::I32__store16:
      case OpCode::I64__store16:
        compileStoreOp(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                       Instr.getMemoryAlign(), Context.Int16Ty, true);
        break;
      case OpCode::I64__store32:
        compileStoreOp(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                       Instr.getMemoryAlign(), Context.Int32Ty, true);
        break;
      case OpCode::Memory__size:
        stackPush(Builder.createTrunc(
            Builder.createCall(
                Context.getIntrinsic(
                    Builder, Executable::Intrinsics::kMemSize,
                    LLVM::Type::getFunctionType(Context.Int64Ty,
                                                {Context.Int32Ty}, false)),
                {LLContext.getInt32(Instr.getTargetIndex())}),
            Context.MemoryAddrTypes[Instr.getTargetIndex()]));
        break;
      case OpCode::Memory__grow: {
        auto NewPageSize = Builder.createZExt(stackPop(), Context.Int64Ty);
        stackPush(Builder.createTrunc(
            Builder.createCall(
                Context.getIntrinsic(Builder, Executable::Intrinsics::kMemGrow,
                                     LLVM::Type::getFunctionType(
                                         Context.Int64Ty,
                                         {Context.Int32Ty, Context.Int64Ty},
                                         false)),
                {LLContext.getInt32(Instr.getTargetIndex()), NewPageSize}),
            Context.MemoryAddrTypes[Instr.getTargetIndex()]));
        break;
      }
      case OpCode::Memory__init: {
        auto Len = stackPop();
        auto Src = stackPop();
        auto Dst = Builder.createZExt(stackPop(), Context.Int64Ty);
        Builder.createCall(
            Context.getIntrinsic(
                Builder, Executable::Intrinsics::kMemInit,
                LLVM::Type::getFunctionType(Context.VoidTy,
                                            {Context.Int32Ty, Context.Int32Ty,
                                             Context.Int64Ty, Context.Int32Ty,
                                             Context.Int32Ty},
                                            false)),
            {LLContext.getInt32(Instr.getTargetIndex()),
             LLContext.getInt32(Instr.getSourceIndex()), Dst, Src, Len});
        break;
      }
      case OpCode::Data__drop: {
        Builder.createCall(
            Context.getIntrinsic(Builder, Executable::Intrinsics::kDataDrop,
                                 LLVM::Type::getFunctionType(
                                     Context.VoidTy, {Context.Int32Ty}, false)),
            {LLContext.getInt32(Instr.getTargetIndex())});
        break;
      }
      case OpCode::Memory__copy: {
        auto Len = Builder.createZExt(stackPop(), Context.Int64Ty);
        auto Src = Builder.createZExt(stackPop(), Context.Int64Ty);
        auto Dst = Builder.createZExt(stackPop(), Context.Int64Ty);
        Builder.createCall(
            Context.getIntrinsic(
                Builder, Executable::Intrinsics::kMemCopy,
                LLVM::Type::getFunctionType(Context.VoidTy,
                                            {Context.Int32Ty, Context.Int32Ty,
                                             Context.Int64Ty, Context.Int64Ty,
                                             Context.Int64Ty},
                                            false)),
            {LLContext.getInt32(Instr.getTargetIndex()),
             LLContext.getInt32(Instr.getSourceIndex()), Dst, Src, Len});
        break;
      }
      case OpCode::Memory__fill: {
        auto Len = Builder.createZExt(stackPop(), Context.Int64Ty);
        auto Val = Builder.createTrunc(stackPop(), Context.Int8Ty);
        auto Off = Builder.createZExt(stackPop(), Context.Int64Ty);
        Builder.createCall(
            Context.getIntrinsic(
                Builder, Executable::Intrinsics::kMemFill,
                LLVM::Type::getFunctionType(Context.VoidTy,
                                            {Context.Int32Ty, Context.Int64Ty,
                                             Context.Int8Ty, Context.Int64Ty},
                                            false)),
            {LLContext.getInt32(Instr.getTargetIndex()), Off, Val, Len});
        break;
      }

      // Const Numeric Instructions
      case OpCode::I32__const:
        stackPush(LLContext.getInt32(Instr.getNum().get<uint32_t>()));
        break;
      case OpCode::I64__const:
        stackPush(LLContext.getInt64(Instr.getNum().get<uint64_t>()));
        break;
      case OpCode::F32__const:
        stackPush(LLContext.getFloat(Instr.getNum().get<float>()));
        break;
      case OpCode::F64__const:
        stackPush(LLContext.getDouble(Instr.getNum().get<double>()));
        break;

      // Unary Numeric Instructions
      case OpCode::I32__eqz:
        stackPush(Builder.createZExt(
            Builder.createICmpEQ(stackPop(), LLContext.getInt32(0)),
            Context.Int32Ty));
        break;
      case OpCode::I64__eqz:
        stackPush(Builder.createZExt(
            Builder.createICmpEQ(stackPop(), LLContext.getInt64(0)),
            Context.Int32Ty));
        break;
      case OpCode::I32__clz:
        assuming(LLVM::Core::Ctlz != LLVM::Core::NotIntrinsic);
        stackPush(Builder.createIntrinsic(LLVM::Core::Ctlz, {Context.Int32Ty},
                                          {stackPop(), LLContext.getFalse()}));
        break;
      case OpCode::I64__clz:
        assuming(LLVM::Core::Ctlz != LLVM::Core::NotIntrinsic);
        stackPush(Builder.createIntrinsic(LLVM::Core::Ctlz, {Context.Int64Ty},
                                          {stackPop(), LLContext.getFalse()}));
        break;
      case OpCode::I32__ctz:
        assuming(LLVM::Core::Cttz != LLVM::Core::NotIntrinsic);
        stackPush(Builder.createIntrinsic(LLVM::Core::Cttz, {Context.Int32Ty},
                                          {stackPop(), LLContext.getFalse()}));
        break;
      case OpCode::I64__ctz:
        assuming(LLVM::Core::Cttz != LLVM::Core::NotIntrinsic);
        stackPush(Builder.createIntrinsic(LLVM::Core::Cttz, {Context.Int64Ty},
                                          {stackPop(), LLContext.getFalse()}));
        break;
      case OpCode::I32__popcnt:
      case OpCode::I64__popcnt:
        assuming(LLVM::Core::Ctpop != LLVM::Core::NotIntrinsic);
        stackPush(Builder.createUnaryIntrinsic(LLVM::Core::Ctpop, stackPop()));
        break;
      case OpCode::F32__abs:
      case OpCode::F64__abs:
        assuming(LLVM::Core::Fabs != LLVM::Core::NotIntrinsic);
        stackPush(Builder.createUnaryIntrinsic(LLVM::Core::Fabs, stackPop()));
        break;
      case OpCode::F32__neg:
      case OpCode::F64__neg:
        stackPush(Builder.createFNeg(stackPop()));
        break;
      case OpCode::F32__ceil:
      case OpCode::F64__ceil:
        assuming(LLVM::Core::Ceil != LLVM::Core::NotIntrinsic);
        stackPush(Builder.createUnaryIntrinsic(LLVM::Core::Ceil, stackPop()));
        break;
      case OpCode::F32__floor:
      case OpCode::F64__floor:
        assuming(LLVM::Core::Floor != LLVM::Core::NotIntrinsic);
        stackPush(Builder.createUnaryIntrinsic(LLVM::Core::Floor, stackPop()));
        break;
      case OpCode::F32__trunc:
      case OpCode::F64__trunc:
        assuming(LLVM::Core::Trunc != LLVM::Core::NotIntrinsic);
        stackPush(Builder.createUnaryIntrinsic(LLVM::Core::Trunc, stackPop()));
        break;
      case OpCode::F32__nearest:
      case OpCode::F64__nearest: {
        const bool IsFloat = Instr.getOpCode() == OpCode::F32__nearest;
        LLVM::Value Value = stackPop();

#if LLVM_VERSION_MAJOR >= 12 && !defined(__s390x__)
        assuming(LLVM::Core::Roundeven != LLVM::Core::NotIntrinsic);
        if (LLVM::Core::Roundeven != LLVM::Core::NotIntrinsic) {
          stackPush(Builder.createUnaryIntrinsic(LLVM::Core::Roundeven, Value));
          break;
        }
#endif

        // The VectorSize is only used when SSE4_1 or NEON is supported.
        [[maybe_unused]] const uint32_t VectorSize = IsFloat ? 4 : 2;
#if defined(__x86_64__)
        if (Context.SupportSSE4_1) {
          auto Zero = LLContext.getInt64(0);
          auto VectorTy =
              LLVM::Type::getVectorType(Value.getType(), VectorSize);
          LLVM::Value Ret = LLVM::Value::getUndef(VectorTy);
          Ret = Builder.createInsertElement(Ret, Value, Zero);
          auto ID = IsFloat ? LLVM::Core::X86SSE41RoundSs
                            : LLVM::Core::X86SSE41RoundSd;
          assuming(ID != LLVM::Core::NotIntrinsic);
          Ret = Builder.createIntrinsic(ID, {},
                                        {Ret, Ret, LLContext.getInt32(8)});
          Ret = Builder.createExtractElement(Ret, Zero);
          stackPush(Ret);
          break;
        }
#endif

#if defined(__aarch64__)
        if (Context.SupportNEON &&
            LLVM::Core::AArch64NeonFRIntN != LLVM::Core::NotIntrinsic) {
          auto Zero = LLContext.getInt64(0);
          auto VectorTy =
              LLVM::Type::getVectorType(Value.getType(), VectorSize);
          LLVM::Value Ret = LLVM::Value::getUndef(VectorTy);
          Ret = Builder.createInsertElement(Ret, Value, Zero);
          Ret =
              Builder.createUnaryIntrinsic(LLVM::Core::AArch64NeonFRIntN, Ret);
          Ret = Builder.createExtractElement(Ret, Zero);
          stackPush(Ret);
          break;
        }
#endif

        // Fallback case.
        // If the SSE4.1 is not supported on the x86_64 platform or
        // the NEON is not supported on the aarch64 platform,
        // then fallback to this.
        assuming(LLVM::Core::Nearbyint != LLVM::Core::NotIntrinsic);
        stackPush(Builder.createUnaryIntrinsic(LLVM::Core::Nearbyint, Value));
        break;
      }
      case OpCode::F32__sqrt:
      case OpCode::F64__sqrt:
        assuming(LLVM::Core::Sqrt != LLVM::Core::NotIntrinsic);
        stackPush(Builder.createUnaryIntrinsic(LLVM::Core::Sqrt, stackPop()));
        break;
      case OpCode::I32__wrap_i64:
        stackPush(Builder.createTrunc(stackPop(), Context.Int32Ty));
        break;
      case OpCode::I32__trunc_f32_s:
        compileSignedTrunc(Context.Int32Ty);
        break;
      case OpCode::I32__trunc_f64_s:
        compileSignedTrunc(Context.Int32Ty);
        break;
      case OpCode::I32__trunc_f32_u:
        compileUnsignedTrunc(Context.Int32Ty);
        break;
      case OpCode::I32__trunc_f64_u:
        compileUnsignedTrunc(Context.Int32Ty);
        break;
      case OpCode::I64__extend_i32_s:
        stackPush(Builder.createSExt(stackPop(), Context.Int64Ty));
        break;
      case OpCode::I64__extend_i32_u:
        stackPush(Builder.createZExt(stackPop(), Context.Int64Ty));
        break;
      case OpCode::I64__trunc_f32_s:
        compileSignedTrunc(Context.Int64Ty);
        break;
      case OpCode::I64__trunc_f64_s:
        compileSignedTrunc(Context.Int64Ty);
        break;
      case OpCode::I64__trunc_f32_u:
        compileUnsignedTrunc(Context.Int64Ty);
        break;
      case OpCode::I64__trunc_f64_u:
        compileUnsignedTrunc(Context.Int64Ty);
        break;
      case OpCode::F32__convert_i32_s:
      case OpCode::F32__convert_i64_s:
        stackPush(Builder.createSIToFP(stackPop(), Context.FloatTy));
        break;
      case OpCode::F32__convert_i32_u:
      case OpCode::F32__convert_i64_u:
        stackPush(Builder.createUIToFP(stackPop(), Context.FloatTy));
        break;
      case OpCode::F64__convert_i32_s:
      case OpCode::F64__convert_i64_s:
        stackPush(Builder.createSIToFP(stackPop(), Context.DoubleTy));
        break;
      case OpCode::F64__convert_i32_u:
      case OpCode::F64__convert_i64_u:
        stackPush(Builder.createUIToFP(stackPop(), Context.DoubleTy));
        break;
      case OpCode::F32__demote_f64:
        stackPush(Builder.createFPTrunc(stackPop(), Context.FloatTy));
        break;
      case OpCode::F64__promote_f32:
        stackPush(Builder.createFPExt(stackPop(), Context.DoubleTy));
        break;
      case OpCode::I32__reinterpret_f32:
        stackPush(Builder.createBitCast(stackPop(), Context.Int32Ty));
        break;
      case OpCode::I64__reinterpret_f64:
        stackPush(Builder.createBitCast(stackPop(), Context.Int64Ty));
        break;
      case OpCode::F32__reinterpret_i32:
        stackPush(Builder.createBitCast(stackPop(), Context.FloatTy));
        break;
      case OpCode::F64__reinterpret_i64:
        stackPush(Builder.createBitCast(stackPop(), Context.DoubleTy));
        break;
      case OpCode::I32__extend8_s:
        stackPush(Builder.createSExt(
            Builder.createTrunc(stackPop(), Context.Int8Ty), Context.Int32Ty));
        break;
      case OpCode::I32__extend16_s:
        stackPush(Builder.createSExt(
            Builder.createTrunc(stackPop(), Context.Int16Ty), Context.Int32Ty));
        break;
      case OpCode::I64__extend8_s:
        stackPush(Builder.createSExt(
            Builder.createTrunc(stackPop(), Context.Int8Ty), Context.Int64Ty));
        break;
      case OpCode::I64__extend16_s:
        stackPush(Builder.createSExt(
            Builder.createTrunc(stackPop(), Context.Int16Ty), Context.Int64Ty));
        break;
      case OpCode::I64__extend32_s:
        stackPush(Builder.createSExt(
            Builder.createTrunc(stackPop(), Context.Int32Ty), Context.Int64Ty));
        break;

      // Binary Numeric Instructions
      case OpCode::I32__eq:
      case OpCode::I64__eq: {
        LLVM::Value RHS = stackPop();
        LLVM::Value LHS = stackPop();
        stackPush(Builder.createZExt(Builder.createICmpEQ(LHS, RHS),
                                     Context.Int32Ty));
        break;
      }
      case OpCode::I32__ne:
      case OpCode::I64__ne: {
        LLVM::Value RHS = stackPop();
        LLVM::Value LHS = stackPop();
        stackPush(Builder.createZExt(Builder.createICmpNE(LHS, RHS),
                                     Context.Int32Ty));
        break;
      }
      case OpCode::I32__lt_s:
      case OpCode::I64__lt_s: {
        LLVM::Value RHS = stackPop();
        LLVM::Value LHS = stackPop();
        stackPush(Builder.createZExt(Builder.createICmpSLT(LHS, RHS),
                                     Context.Int32Ty));
        break;
      }
      case OpCode::I32__lt_u:
      case OpCode::I64__lt_u: {
        LLVM::Value RHS = stackPop();
        LLVM::Value LHS = stackPop();
        stackPush(Builder.createZExt(Builder.createICmpULT(LHS, RHS),
                                     Context.Int32Ty));
        break;
      }
      case OpCode::I32__gt_s:
      case OpCode::I64__gt_s: {
        LLVM::Value RHS = stackPop();
        LLVM::Value LHS = stackPop();
        stackPush(Builder.createZExt(Builder.createICmpSGT(LHS, RHS),
                                     Context.Int32Ty));
        break;
      }
      case OpCode::I32__gt_u:
      case OpCode::I64__gt_u: {
        LLVM::Value RHS = stackPop();
        LLVM::Value LHS = stackPop();
        stackPush(Builder.createZExt(Builder.createICmpUGT(LHS, RHS),
                                     Context.Int32Ty));
        break;
      }
      case OpCode::I32__le_s:
      case OpCode::I64__le_s: {
        LLVM::Value RHS = stackPop();
        LLVM::Value LHS = stackPop();
        stackPush(Builder.createZExt(Builder.createICmpSLE(LHS, RHS),
                                     Context.Int32Ty));
        break;
      }
      case OpCode::I32__le_u:
      case OpCode::I64__le_u: {
        LLVM::Value RHS = stackPop();
        LLVM::Value LHS = stackPop();
        stackPush(Builder.createZExt(Builder.createICmpULE(LHS, RHS),
                                     Context.Int32Ty));
        break;
      }
      case OpCode::I32__ge_s:
      case OpCode::I64__ge_s: {
        LLVM::Value RHS = stackPop();
        LLVM::Value LHS = stackPop();
        stackPush(Builder.createZExt(Builder.createICmpSGE(LHS, RHS),
                                     Context.Int32Ty));
        break;
      }
      case OpCode::I32__ge_u:
      case OpCode::I64__ge_u: {
        LLVM::Value RHS = stackPop();
        LLVM::Value LHS = stackPop();
        stackPush(Builder.createZExt(Builder.createICmpUGE(LHS, RHS),
                                     Context.Int32Ty));
        break;
      }
      case OpCode::F32__eq:
      case OpCode::F64__eq: {
        LLVM::Value RHS = stackPop();
        LLVM::Value LHS = stackPop();
        stackPush(Builder.createZExt(Builder.createFCmpOEQ(LHS, RHS),
                                     Context.Int32Ty));
        break;
      }
      case OpCode::F32__ne:
      case OpCode::F64__ne: {
        LLVM::Value RHS = stackPop();
        LLVM::Value LHS = stackPop();
        stackPush(Builder.createZExt(Builder.createFCmpUNE(LHS, RHS),
                                     Context.Int32Ty));
        break;
      }
      case OpCode::F32__lt:
      case OpCode::F64__lt: {
        LLVM::Value RHS = stackPop();
        LLVM::Value LHS = stackPop();
        stackPush(Builder.createZExt(Builder.createFCmpOLT(LHS, RHS),
                                     Context.Int32Ty));
        break;
      }
      case OpCode::F32__gt:
      case OpCode::F64__gt: {
        LLVM::Value RHS = stackPop();
        LLVM::Value LHS = stackPop();
        stackPush(Builder.createZExt(Builder.createFCmpOGT(LHS, RHS),
                                     Context.Int32Ty));
        break;
      }
      case OpCode::F32__le:
      case OpCode::F64__le: {
        LLVM::Value RHS = stackPop();
        LLVM::Value LHS = stackPop();
        stackPush(Builder.createZExt(Builder.createFCmpOLE(LHS, RHS),
                                     Context.Int32Ty));
        break;
      }
      case OpCode::F32__ge:
      case OpCode::F64__ge: {
        LLVM::Value RHS = stackPop();
        LLVM::Value LHS = stackPop();
        stackPush(Builder.createZExt(Builder.createFCmpOGE(LHS, RHS),
                                     Context.Int32Ty));
        break;
      }
      case OpCode::I32__add:
      case OpCode::I64__add: {
        LLVM::Value RHS = stackPop();
        LLVM::Value LHS = stackPop();
        stackPush(Builder.createAdd(LHS, RHS));
        break;
      }
      case OpCode::I32__sub:
      case OpCode::I64__sub: {
        LLVM::Value RHS = stackPop();
        LLVM::Value LHS = stackPop();

        stackPush(Builder.createSub(LHS, RHS));
        break;
      }
      case OpCode::I32__mul:
      case OpCode::I64__mul: {
        LLVM::Value RHS = stackPop();
        LLVM::Value LHS = stackPop();
        stackPush(Builder.createMul(LHS, RHS));
        break;
      }
      case OpCode::I32__div_s:
      case OpCode::I64__div_s: {
        LLVM::Value RHS = stackPop();
        LLVM::Value LHS = stackPop();
        if constexpr (kForceDivCheck) {
          const bool Is32 = Instr.getOpCode() == OpCode::I32__div_s;
          LLVM::Value IntZero =
              Is32 ? LLContext.getInt32(0) : LLContext.getInt64(0);
          LLVM::Value IntMinusOne =
              Is32 ? LLContext.getInt32(static_cast<uint32_t>(INT32_C(-1)))
                   : LLContext.getInt64(static_cast<uint64_t>(INT64_C(-1)));
          LLVM::Value IntMin = Is32 ? LLContext.getInt32(static_cast<uint32_t>(
                                          std::numeric_limits<int32_t>::min()))
                                    : LLContext.getInt64(static_cast<uint64_t>(
                                          std::numeric_limits<int64_t>::min()));

          auto NoZeroBB =
              LLVM::BasicBlock::create(LLContext, F.Fn, "div.nozero");
          auto OkBB = LLVM::BasicBlock::create(LLContext, F.Fn, "div.ok");

          auto IsNotZero =
              Builder.createLikely(Builder.createICmpNE(RHS, IntZero));
          Builder.createCondBr(IsNotZero, NoZeroBB,
                               getTrapBB(ErrCode::Value::DivideByZero));

          Builder.positionAtEnd(NoZeroBB);
          auto NotOverflow = Builder.createLikely(
              Builder.createOr(Builder.createICmpNE(LHS, IntMin),
                               Builder.createICmpNE(RHS, IntMinusOne)));
          Builder.createCondBr(NotOverflow, OkBB,
                               getTrapBB(ErrCode::Value::IntegerOverflow));

          Builder.positionAtEnd(OkBB);
        }
        stackPush(Builder.createSDiv(LHS, RHS));
        break;
      }
      case OpCode::I32__div_u:
      case OpCode::I64__div_u: {
        LLVM::Value RHS = stackPop();
        LLVM::Value LHS = stackPop();
        if constexpr (kForceDivCheck) {
          const bool Is32 = Instr.getOpCode() == OpCode::I32__div_u;
          LLVM::Value IntZero =
              Is32 ? LLContext.getInt32(0) : LLContext.getInt64(0);
          auto OkBB = LLVM::BasicBlock::create(LLContext, F.Fn, "div.ok");

          auto IsNotZero =
              Builder.createLikely(Builder.createICmpNE(RHS, IntZero));
          Builder.createCondBr(IsNotZero, OkBB,
                               getTrapBB(ErrCode::Value::DivideByZero));
          Builder.positionAtEnd(OkBB);
        }
        stackPush(Builder.createUDiv(LHS, RHS));
        break;
      }
      case OpCode::I32__rem_s:
      case OpCode::I64__rem_s: {
        LLVM::Value RHS = stackPop();
        LLVM::Value LHS = stackPop();
        // handle INT32_MIN % -1
        const bool Is32 = Instr.getOpCode() == OpCode::I32__rem_s;
        LLVM::Value IntMinusOne =
            Is32 ? LLContext.getInt32(static_cast<uint32_t>(INT32_C(-1)))
                 : LLContext.getInt64(static_cast<uint64_t>(INT64_C(-1)));
        LLVM::Value IntMin = Is32 ? LLContext.getInt32(static_cast<uint32_t>(
                                        std::numeric_limits<int32_t>::min()))
                                  : LLContext.getInt64(static_cast<uint64_t>(
                                        std::numeric_limits<int64_t>::min()));
        LLVM::Value IntZero =
            Is32 ? LLContext.getInt32(0) : LLContext.getInt64(0);

        auto NoOverflowBB =
            LLVM::BasicBlock::create(LLContext, F.Fn, "no.overflow");
        auto EndBB = LLVM::BasicBlock::create(LLContext, F.Fn, "end.overflow");

        if constexpr (kForceDivCheck) {
          auto OkBB = LLVM::BasicBlock::create(LLContext, F.Fn, "rem.ok");

          auto IsNotZero =
              Builder.createLikely(Builder.createICmpNE(RHS, IntZero));
          Builder.createCondBr(IsNotZero, OkBB,
                               getTrapBB(ErrCode::Value::DivideByZero));
          Builder.positionAtEnd(OkBB);
        }

        auto CurrBB = Builder.getInsertBlock();

        auto NotOverflow = Builder.createLikely(
            Builder.createOr(Builder.createICmpNE(LHS, IntMin),
                             Builder.createICmpNE(RHS, IntMinusOne)));
        Builder.createCondBr(NotOverflow, NoOverflowBB, EndBB);

        Builder.positionAtEnd(NoOverflowBB);
        auto Ret1 = Builder.createSRem(LHS, RHS);
        Builder.createBr(EndBB);

        Builder.positionAtEnd(EndBB);
        auto Ret = Builder.createPHI(Ret1.getType());
        Ret.addIncoming(Ret1, NoOverflowBB);
        Ret.addIncoming(IntZero, CurrBB);

        stackPush(Ret);
        break;
      }
      case OpCode::I32__rem_u:
      case OpCode::I64__rem_u: {
        LLVM::Value RHS = stackPop();
        LLVM::Value LHS = stackPop();
        if constexpr (kForceDivCheck) {
          LLVM::Value IntZero = Instr.getOpCode() == OpCode::I32__rem_u
                                    ? LLContext.getInt32(0)
                                    : LLContext.getInt64(0);
          auto OkBB = LLVM::BasicBlock::create(LLContext, F.Fn, "rem.ok");

          auto IsNotZero =
              Builder.createLikely(Builder.createICmpNE(RHS, IntZero));
          Builder.createCondBr(IsNotZero, OkBB,
                               getTrapBB(ErrCode::Value::DivideByZero));
          Builder.positionAtEnd(OkBB);
        }
        stackPush(Builder.createURem(LHS, RHS));
        break;
      }
      case OpCode::I32__and:
      case OpCode::I64__and: {
        LLVM::Value RHS = stackPop();
        LLVM::Value LHS = stackPop();
        stackPush(Builder.createAnd(LHS, RHS));
        break;
      }
      case OpCode::I32__or:
      case OpCode::I64__or: {
        LLVM::Value RHS = stackPop();
        LLVM::Value LHS = stackPop();
        stackPush(Builder.createOr(LHS, RHS));
        break;
      }
      case OpCode::I32__xor:
      case OpCode::I64__xor: {
        LLVM::Value RHS = stackPop();
        LLVM::Value LHS = stackPop();
        stackPush(Builder.createXor(LHS, RHS));
        break;
      }
      case OpCode::I32__shl:
      case OpCode::I64__shl: {
        LLVM::Value Mask = Instr.getOpCode() == OpCode::I32__shl
                               ? LLContext.getInt32(31)
                               : LLContext.getInt64(63);
        LLVM::Value RHS = Builder.createAnd(stackPop(), Mask);
        LLVM::Value LHS = stackPop();
        stackPush(Builder.createShl(LHS, RHS));
        break;
      }
      case OpCode::I32__shr_s:
      case OpCode::I64__shr_s: {
        LLVM::Value Mask = Instr.getOpCode() == OpCode::I32__shr_s
                               ? LLContext.getInt32(31)
                               : LLContext.getInt64(63);
        LLVM::Value RHS = Builder.createAnd(stackPop(), Mask);
        LLVM::Value LHS = stackPop();
        stackPush(Builder.createAShr(LHS, RHS));
        break;
      }
      case OpCode::I32__shr_u:
      case OpCode::I64__shr_u: {
        LLVM::Value Mask = Instr.getOpCode() == OpCode::I32__shr_u
                               ? LLContext.getInt32(31)
                               : LLContext.getInt64(63);
        LLVM::Value RHS = Builder.createAnd(stackPop(), Mask);
        LLVM::Value LHS = stackPop();
        stackPush(Builder.createLShr(LHS, RHS));
        break;
      }
      case OpCode::I32__rotl: {
        LLVM::Value RHS = stackPop();
        LLVM::Value LHS = stackPop();
        assuming(LLVM::Core::FShl != LLVM::Core::NotIntrinsic);
        stackPush(Builder.createIntrinsic(LLVM::Core::FShl, {Context.Int32Ty},
                                          {LHS, LHS, RHS}));
        break;
      }
      case OpCode::I32__rotr: {
        LLVM::Value RHS = stackPop();
        LLVM::Value LHS = stackPop();
        assuming(LLVM::Core::FShr != LLVM::Core::NotIntrinsic);
        stackPush(Builder.createIntrinsic(LLVM::Core::FShr, {Context.Int32Ty},
                                          {LHS, LHS, RHS}));
        break;
      }
      case OpCode::I64__rotl: {
        LLVM::Value RHS = stackPop();
        LLVM::Value LHS = stackPop();
        assuming(LLVM::Core::FShl != LLVM::Core::NotIntrinsic);
        stackPush(Builder.createIntrinsic(LLVM::Core::FShl, {Context.Int64Ty},
                                          {LHS, LHS, RHS}));
        break;
      }
      case OpCode::I64__rotr: {
        LLVM::Value RHS = stackPop();
        LLVM::Value LHS = stackPop();
        assuming(LLVM::Core::FShr != LLVM::Core::NotIntrinsic);
        stackPush(Builder.createIntrinsic(LLVM::Core::FShr, {Context.Int64Ty},
                                          {LHS, LHS, RHS}));
        break;
      }
      case OpCode::F32__add:
      case OpCode::F64__add: {
        LLVM::Value RHS = stackPop();
        LLVM::Value LHS = stackPop();
        stackPush(Builder.createFAdd(LHS, RHS));
        break;
      }
      case OpCode::F32__sub:
      case OpCode::F64__sub: {
        LLVM::Value RHS = stackPop();
        LLVM::Value LHS = stackPop();
        stackPush(Builder.createFSub(LHS, RHS));
        break;
      }
      case OpCode::F32__mul:
      case OpCode::F64__mul: {
        LLVM::Value RHS = stackPop();
        LLVM::Value LHS = stackPop();
        stackPush(Builder.createFMul(LHS, RHS));
        break;
      }
      case OpCode::F32__div:
      case OpCode::F64__div: {
        LLVM::Value RHS = stackPop();
        LLVM::Value LHS = stackPop();
        stackPush(Builder.createFDiv(LHS, RHS));
        break;
      }
      case OpCode::F32__min:
      case OpCode::F64__min: {
        LLVM::Value RHS = stackPop();
        LLVM::Value LHS = stackPop();
        auto FpTy = Instr.getOpCode() == OpCode::F32__min ? Context.FloatTy
                                                          : Context.DoubleTy;
        auto IntTy = Instr.getOpCode() == OpCode::F32__min ? Context.Int32Ty
                                                           : Context.Int64Ty;

        auto UEQ = Builder.createFCmpUEQ(LHS, RHS);
        auto UNO = Builder.createFCmpUNO(LHS, RHS);

        auto LHSInt = Builder.createBitCast(LHS, IntTy);
        auto RHSInt = Builder.createBitCast(RHS, IntTy);
        auto OrInt = Builder.createOr(LHSInt, RHSInt);
        auto OrFp = Builder.createBitCast(OrInt, FpTy);

        auto AddFp = Builder.createFAdd(LHS, RHS);

        assuming(LLVM::Core::MinNum != LLVM::Core::NotIntrinsic);
        auto MinFp = Builder.createIntrinsic(LLVM::Core::MinNum,
                                             {LHS.getType()}, {LHS, RHS});

        auto Ret = Builder.createSelect(
            UEQ, Builder.createSelect(UNO, AddFp, OrFp), MinFp);
        stackPush(Ret);
        break;
      }
      case OpCode::F32__max:
      case OpCode::F64__max: {
        LLVM::Value RHS = stackPop();
        LLVM::Value LHS = stackPop();
        auto FpTy = Instr.getOpCode() == OpCode::F32__max ? Context.FloatTy
                                                          : Context.DoubleTy;
        auto IntTy = Instr.getOpCode() == OpCode::F32__max ? Context.Int32Ty
                                                           : Context.Int64Ty;

        auto UEQ = Builder.createFCmpUEQ(LHS, RHS);
        auto UNO = Builder.createFCmpUNO(LHS, RHS);

        auto LHSInt = Builder.createBitCast(LHS, IntTy);
        auto RHSInt = Builder.createBitCast(RHS, IntTy);
        auto AndInt = Builder.createAnd(LHSInt, RHSInt);
        auto AndFp = Builder.createBitCast(AndInt, FpTy);

        auto AddFp = Builder.createFAdd(LHS, RHS);

        assuming(LLVM::Core::MaxNum != LLVM::Core::NotIntrinsic);
        auto MaxFp = Builder.createIntrinsic(LLVM::Core::MaxNum,
                                             {LHS.getType()}, {LHS, RHS});

        auto Ret = Builder.createSelect(
            UEQ, Builder.createSelect(UNO, AddFp, AndFp), MaxFp);
        stackPush(Ret);
        break;
      }
      case OpCode::F32__copysign:
      case OpCode::F64__copysign: {
        LLVM::Value RHS = stackPop();
        LLVM::Value LHS = stackPop();
        assuming(LLVM::Core::CopySign != LLVM::Core::NotIntrinsic);
        stackPush(Builder.createIntrinsic(LLVM::Core::CopySign, {LHS.getType()},
                                          {LHS, RHS}));
        break;
      }

      // Saturating Truncation Numeric Instructions
      case OpCode::I32__trunc_sat_f32_s:
        compileSignedTruncSat(Context.Int32Ty);
        break;
      case OpCode::I32__trunc_sat_f32_u:
        compileUnsignedTruncSat(Context.Int32Ty);
        break;
      case OpCode::I32__trunc_sat_f64_s:
        compileSignedTruncSat(Context.Int32Ty);
        break;
      case OpCode::I32__trunc_sat_f64_u:
        compileUnsignedTruncSat(Context.Int32Ty);
        break;
      case OpCode::I64__trunc_sat_f32_s:
        compileSignedTruncSat(Context.Int64Ty);
        break;
      case OpCode::I64__trunc_sat_f32_u:
        compileUnsignedTruncSat(Context.Int64Ty);
        break;
      case OpCode::I64__trunc_sat_f64_s:
        compileSignedTruncSat(Context.Int64Ty);
        break;
      case OpCode::I64__trunc_sat_f64_u:
        compileUnsignedTruncSat(Context.Int64Ty);
        break;

      // SIMD Memory Instructions
      case OpCode::V128__load:
        compileVectorLoadOp(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                            Instr.getMemoryAlign(), Context.Int128x1Ty);
        break;
      case OpCode::V128__load8x8_s:
        compileVectorLoadOp(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                            Instr.getMemoryAlign(),
                            LLVM::Type::getVectorType(Context.Int8Ty, 8),
                            Context.Int16x8Ty, true);
        break;
      case OpCode::V128__load8x8_u:
        compileVectorLoadOp(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                            Instr.getMemoryAlign(),
                            LLVM::Type::getVectorType(Context.Int8Ty, 8),
                            Context.Int16x8Ty, false);
        break;
      case OpCode::V128__load16x4_s:
        compileVectorLoadOp(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                            Instr.getMemoryAlign(),
                            LLVM::Type::getVectorType(Context.Int16Ty, 4),
                            Context.Int32x4Ty, true);
        break;
      case OpCode::V128__load16x4_u:
        compileVectorLoadOp(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                            Instr.getMemoryAlign(),
                            LLVM::Type::getVectorType(Context.Int16Ty, 4),
                            Context.Int32x4Ty, false);
        break;
      case OpCode::V128__load32x2_s:
        compileVectorLoadOp(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                            Instr.getMemoryAlign(),
                            LLVM::Type::getVectorType(Context.Int32Ty, 2),
                            Context.Int64x2Ty, true);
        break;
      case OpCode::V128__load32x2_u:
        compileVectorLoadOp(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                            Instr.getMemoryAlign(),
                            LLVM::Type::getVectorType(Context.Int32Ty, 2),
                            Context.Int64x2Ty, false);
        break;
      case OpCode::V128__load8_splat:
        compileSplatLoadOp(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                           Instr.getMemoryAlign(), Context.Int8Ty,
                           Context.Int8x16Ty);
        break;
      case OpCode::V128__load16_splat:
        compileSplatLoadOp(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                           Instr.getMemoryAlign(), Context.Int16Ty,
                           Context.Int16x8Ty);
        break;
      case OpCode::V128__load32_splat:
        compileSplatLoadOp(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                           Instr.getMemoryAlign(), Context.Int32Ty,
                           Context.Int32x4Ty);
        break;
      case OpCode::V128__load64_splat:
        compileSplatLoadOp(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                           Instr.getMemoryAlign(), Context.Int64Ty,
                           Context.Int64x2Ty);
        break;
      case OpCode::V128__load32_zero:
        compileVectorLoadOp(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                            Instr.getMemoryAlign(), Context.Int32Ty,
                            Context.Int128Ty, false);
        break;
      case OpCode::V128__load64_zero:
        compileVectorLoadOp(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                            Instr.getMemoryAlign(), Context.Int64Ty,
                            Context.Int128Ty, false);
        break;
      case OpCode::V128__store:
        compileStoreOp(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                       Instr.getMemoryAlign(), Context.Int128x1Ty, false, true);
        break;
      case OpCode::V128__load8_lane:
        compileLoadLaneOp(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                          Instr.getMemoryAlign(), Instr.getMemoryLane(),
                          Context.Int8Ty, Context.Int8x16Ty);
        break;
      case OpCode::V128__load16_lane:
        compileLoadLaneOp(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                          Instr.getMemoryAlign(), Instr.getMemoryLane(),
                          Context.Int16Ty, Context.Int16x8Ty);
        break;
      case OpCode::V128__load32_lane:
        compileLoadLaneOp(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                          Instr.getMemoryAlign(), Instr.getMemoryLane(),
                          Context.Int32Ty, Context.Int32x4Ty);
        break;
      case OpCode::V128__load64_lane:
        compileLoadLaneOp(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                          Instr.getMemoryAlign(), Instr.getMemoryLane(),
                          Context.Int64Ty, Context.Int64x2Ty);
        break;
      case OpCode::V128__store8_lane:
        compileStoreLaneOp(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                           Instr.getMemoryAlign(), Instr.getMemoryLane(),
                           Context.Int8Ty, Context.Int8x16Ty);
        break;
      case OpCode::V128__store16_lane:
        compileStoreLaneOp(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                           Instr.getMemoryAlign(), Instr.getMemoryLane(),
                           Context.Int16Ty, Context.Int16x8Ty);
        break;
      case OpCode::V128__store32_lane:
        compileStoreLaneOp(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                           Instr.getMemoryAlign(), Instr.getMemoryLane(),
                           Context.Int32Ty, Context.Int32x4Ty);
        break;
      case OpCode::V128__store64_lane:
        compileStoreLaneOp(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                           Instr.getMemoryAlign(), Instr.getMemoryLane(),
                           Context.Int64Ty, Context.Int64x2Ty);
        break;

      // SIMD Const Instructions
      case OpCode::V128__const: {
        const auto Value = Instr.getNum().get<uint64x2_t>();
        auto Vector =
            LLVM::Value::getConstVector64(LLContext, {Value[0], Value[1]});
        stackPush(Builder.createBitCast(Vector, Context.Int64x2Ty));
        break;
      }

      // SIMD Shuffle Instructions
      case OpCode::I8x16__shuffle: {
        auto V2 = Builder.createBitCast(stackPop(), Context.Int8x16Ty);
        auto V1 = Builder.createBitCast(stackPop(), Context.Int8x16Ty);
        const auto V3 = Instr.getNum().get<uint128_t>();
        std::array<uint8_t, 16> Mask;
        for (size_t I = 0; I < 16; ++I) {
          auto Num = static_cast<uint8_t>(V3 >> (I * 8));
          if constexpr (Endian::native == Endian::little) {
            Mask[I] = Num;
          } else {
            Mask[15 - I] = Num < 16 ? 15 - Num : 47 - Num;
          }
        }
        stackPush(Builder.createBitCast(
            Builder.createShuffleVector(
                V1, V2, LLVM::Value::getConstVector8(LLContext, Mask)),
            Context.Int64x2Ty));
        break;
      }

      // SIMD Lane Instructions
      case OpCode::I8x16__extract_lane_s:
        compileExtractLaneOp(Context.Int8x16Ty, Instr.getMemoryLane(),
                             Context.Int32Ty, true);
        break;
      case OpCode::I8x16__extract_lane_u:
        compileExtractLaneOp(Context.Int8x16Ty, Instr.getMemoryLane(),
                             Context.Int32Ty, false);
        break;
      case OpCode::I8x16__replace_lane:
        compileReplaceLaneOp(Context.Int8x16Ty, Instr.getMemoryLane());
        break;
      case OpCode::I16x8__extract_lane_s:
        compileExtractLaneOp(Context.Int16x8Ty, Instr.getMemoryLane(),
                             Context.Int32Ty, true);
        break;
      case OpCode::I16x8__extract_lane_u:
        compileExtractLaneOp(Context.Int16x8Ty, Instr.getMemoryLane(),
                             Context.Int32Ty, false);
        break;
      case OpCode::I16x8__replace_lane:
        compileReplaceLaneOp(Context.Int16x8Ty, Instr.getMemoryLane());
        break;
      case OpCode::I32x4__extract_lane:
        compileExtractLaneOp(Context.Int32x4Ty, Instr.getMemoryLane());
        break;
      case OpCode::I32x4__replace_lane:
        compileReplaceLaneOp(Context.Int32x4Ty, Instr.getMemoryLane());
        break;
      case OpCode::I64x2__extract_lane:
        compileExtractLaneOp(Context.Int64x2Ty, Instr.getMemoryLane());
        break;
      case OpCode::I64x2__replace_lane:
        compileReplaceLaneOp(Context.Int64x2Ty, Instr.getMemoryLane());
        break;
      case OpCode::F32x4__extract_lane:
        compileExtractLaneOp(Context.Floatx4Ty, Instr.getMemoryLane());
        break;
      case OpCode::F32x4__replace_lane:
        compileReplaceLaneOp(Context.Floatx4Ty, Instr.getMemoryLane());
        break;
      case OpCode::F64x2__extract_lane:
        compileExtractLaneOp(Context.Doublex2Ty, Instr.getMemoryLane());
        break;
      case OpCode::F64x2__replace_lane:
        compileReplaceLaneOp(Context.Doublex2Ty, Instr.getMemoryLane());
        break;

      // SIMD Numeric Instructions
      case OpCode::I8x16__swizzle:
        compileVectorSwizzle();
        break;
      case OpCode::I8x16__splat:
        compileSplatOp(Context.Int8x16Ty);
        break;
      case OpCode::I16x8__splat:
        compileSplatOp(Context.Int16x8Ty);
        break;
      case OpCode::I32x4__splat:
        compileSplatOp(Context.Int32x4Ty);
        break;
      case OpCode::I64x2__splat:
        compileSplatOp(Context.Int64x2Ty);
        break;
      case OpCode::F32x4__splat:
        compileSplatOp(Context.Floatx4Ty);
        break;
      case OpCode::F64x2__splat:
        compileSplatOp(Context.Doublex2Ty);
        break;
      case OpCode::I8x16__eq:
        compileVectorCompareOp(Context.Int8x16Ty, LLVMIntEQ);
        break;
      case OpCode::I8x16__ne:
        compileVectorCompareOp(Context.Int8x16Ty, LLVMIntNE);
        break;
      case OpCode::I8x16__lt_s:
        compileVectorCompareOp(Context.Int8x16Ty, LLVMIntSLT);
        break;
      case OpCode::I8x16__lt_u:
        compileVectorCompareOp(Context.Int8x16Ty, LLVMIntULT);
        break;
      case OpCode::I8x16__gt_s:
        compileVectorCompareOp(Context.Int8x16Ty, LLVMIntSGT);
        break;
      case OpCode::I8x16__gt_u:
        compileVectorCompareOp(Context.Int8x16Ty, LLVMIntUGT);
        break;
      case OpCode::I8x16__le_s:
        compileVectorCompareOp(Context.Int8x16Ty, LLVMIntSLE);
        break;
      case OpCode::I8x16__le_u:
        compileVectorCompareOp(Context.Int8x16Ty, LLVMIntULE);
        break;
      case OpCode::I8x16__ge_s:
        compileVectorCompareOp(Context.Int8x16Ty, LLVMIntSGE);
        break;
      case OpCode::I8x16__ge_u:
        compileVectorCompareOp(Context.Int8x16Ty, LLVMIntUGE);
        break;
      case OpCode::I16x8__eq:
        compileVectorCompareOp(Context.Int16x8Ty, LLVMIntEQ);
        break;
      case OpCode::I16x8__ne:
        compileVectorCompareOp(Context.Int16x8Ty, LLVMIntNE);
        break;
      case OpCode::I16x8__lt_s:
        compileVectorCompareOp(Context.Int16x8Ty, LLVMIntSLT);
        break;
      case OpCode::I16x8__lt_u:
        compileVectorCompareOp(Context.Int16x8Ty, LLVMIntULT);
        break;
      case OpCode::I16x8__gt_s:
        compileVectorCompareOp(Context.Int16x8Ty, LLVMIntSGT);
        break;
      case OpCode::I16x8__gt_u:
        compileVectorCompareOp(Context.Int16x8Ty, LLVMIntUGT);
        break;
      case OpCode::I16x8__le_s:
        compileVectorCompareOp(Context.Int16x8Ty, LLVMIntSLE);
        break;
      case OpCode::I16x8__le_u:
        compileVectorCompareOp(Context.Int16x8Ty, LLVMIntULE);
        break;
      case OpCode::I16x8__ge_s:
        compileVectorCompareOp(Context.Int16x8Ty, LLVMIntSGE);
        break;
      case OpCode::I16x8__ge_u:
        compileVectorCompareOp(Context.Int16x8Ty, LLVMIntUGE);
        break;
      case OpCode::I32x4__eq:
        compileVectorCompareOp(Context.Int32x4Ty, LLVMIntEQ);
        break;
      case OpCode::I32x4__ne:
        compileVectorCompareOp(Context.Int32x4Ty, LLVMIntNE);
        break;
      case OpCode::I32x4__lt_s:
        compileVectorCompareOp(Context.Int32x4Ty, LLVMIntSLT);
        break;
      case OpCode::I32x4__lt_u:
        compileVectorCompareOp(Context.Int32x4Ty, LLVMIntULT);
        break;
      case OpCode::I32x4__gt_s:
        compileVectorCompareOp(Context.Int32x4Ty, LLVMIntSGT);
        break;
      case OpCode::I32x4__gt_u:
        compileVectorCompareOp(Context.Int32x4Ty, LLVMIntUGT);
        break;
      case OpCode::I32x4__le_s:
        compileVectorCompareOp(Context.Int32x4Ty, LLVMIntSLE);
        break;
      case OpCode::I32x4__le_u:
        compileVectorCompareOp(Context.Int32x4Ty, LLVMIntULE);
        break;
      case OpCode::I32x4__ge_s:
        compileVectorCompareOp(Context.Int32x4Ty, LLVMIntSGE);
        break;
      case OpCode::I32x4__ge_u:
        compileVectorCompareOp(Context.Int32x4Ty, LLVMIntUGE);
        break;
      case OpCode::I64x2__eq:
        compileVectorCompareOp(Context.Int64x2Ty, LLVMIntEQ);
        break;
      case OpCode::I64x2__ne:
        compileVectorCompareOp(Context.Int64x2Ty, LLVMIntNE);
        break;
      case OpCode::I64x2__lt_s:
        compileVectorCompareOp(Context.Int64x2Ty, LLVMIntSLT);
        break;
      case OpCode::I64x2__gt_s:
        compileVectorCompareOp(Context.Int64x2Ty, LLVMIntSGT);
        break;
      case OpCode::I64x2__le_s:
        compileVectorCompareOp(Context.Int64x2Ty, LLVMIntSLE);
        break;
      case OpCode::I64x2__ge_s:
        compileVectorCompareOp(Context.Int64x2Ty, LLVMIntSGE);
        break;
      case OpCode::F32x4__eq:
        compileVectorCompareOp(Context.Floatx4Ty, LLVMRealOEQ,
                               Context.Int32x4Ty);
        break;
      case OpCode::F32x4__ne:
        compileVectorCompareOp(Context.Floatx4Ty, LLVMRealUNE,
                               Context.Int32x4Ty);
        break;
      case OpCode::F32x4__lt:
        compileVectorCompareOp(Context.Floatx4Ty, LLVMRealOLT,
                               Context.Int32x4Ty);
        break;
      case OpCode::F32x4__gt:
        compileVectorCompareOp(Context.Floatx4Ty, LLVMRealOGT,
                               Context.Int32x4Ty);
        break;
      case OpCode::F32x4__le:
        compileVectorCompareOp(Context.Floatx4Ty, LLVMRealOLE,
                               Context.Int32x4Ty);
        break;
      case OpCode::F32x4__ge:
        compileVectorCompareOp(Context.Floatx4Ty, LLVMRealOGE,
                               Context.Int32x4Ty);
        break;
      case OpCode::F64x2__eq:
        compileVectorCompareOp(Context.Doublex2Ty, LLVMRealOEQ,
                               Context.Int64x2Ty);
        break;
      case OpCode::F64x2__ne:
        compileVectorCompareOp(Context.Doublex2Ty, LLVMRealUNE,
                               Context.Int64x2Ty);
        break;
      case OpCode::F64x2__lt:
        compileVectorCompareOp(Context.Doublex2Ty, LLVMRealOLT,
                               Context.Int64x2Ty);
        break;
      case OpCode::F64x2__gt:
        compileVectorCompareOp(Context.Doublex2Ty, LLVMRealOGT,
                               Context.Int64x2Ty);
        break;
      case OpCode::F64x2__le:
        compileVectorCompareOp(Context.Doublex2Ty, LLVMRealOLE,
                               Context.Int64x2Ty);
        break;
      case OpCode::F64x2__ge:
        compileVectorCompareOp(Context.Doublex2Ty, LLVMRealOGE,
                               Context.Int64x2Ty);
        break;
      case OpCode::V128__not:
        Stack.back() = Builder.createNot(Stack.back());
        break;
      case OpCode::V128__and: {
        auto RHS = stackPop();
        auto LHS = stackPop();
        stackPush(Builder.createAnd(LHS, RHS));
        break;
      }
      case OpCode::V128__andnot: {
        auto RHS = stackPop();
        auto LHS = stackPop();
        stackPush(Builder.createAnd(LHS, Builder.createNot(RHS)));
        break;
      }
      case OpCode::V128__or: {
        auto RHS = stackPop();
        auto LHS = stackPop();
        stackPush(Builder.createOr(LHS, RHS));
        break;
      }
      case OpCode::V128__xor: {
        auto RHS = stackPop();
        auto LHS = stackPop();
        stackPush(Builder.createXor(LHS, RHS));
        break;
      }
      case OpCode::V128__bitselect: {
        auto C = stackPop();
        auto V2 = stackPop();
        auto V1 = stackPop();
        stackPush(Builder.createXor(
            Builder.createAnd(Builder.createXor(V1, V2), C), V2));
        break;
      }
      case OpCode::V128__any_true:
        compileVectorAnyTrue();
        break;
      case OpCode::I8x16__abs:
        compileVectorAbs(Context.Int8x16Ty);
        break;
      case OpCode::I8x16__neg:
        compileVectorNeg(Context.Int8x16Ty);
        break;
      case OpCode::I8x16__popcnt:
        compileVectorPopcnt();
        break;
      case OpCode::I8x16__all_true:
        compileVectorAllTrue(Context.Int8x16Ty);
        break;
      case OpCode::I8x16__bitmask:
        compileVectorBitMask(Context.Int8x16Ty);
        break;
      case OpCode::I8x16__narrow_i16x8_s:
        compileVectorNarrow(Context.Int16x8Ty, true);
        break;
      case OpCode::I8x16__narrow_i16x8_u:
        compileVectorNarrow(Context.Int16x8Ty, false);
        break;
      case OpCode::I8x16__shl:
        compileVectorShl(Context.Int8x16Ty);
        break;
      case OpCode::I8x16__shr_s:
        compileVectorAShr(Context.Int8x16Ty);
        break;
      case OpCode::I8x16__shr_u:
        compileVectorLShr(Context.Int8x16Ty);
        break;
      case OpCode::I8x16__add:
        compileVectorVectorAdd(Context.Int8x16Ty);
        break;
      case OpCode::I8x16__add_sat_s:
        compileVectorVectorAddSat(Context.Int8x16Ty, true);
        break;
      case OpCode::I8x16__add_sat_u:
        compileVectorVectorAddSat(Context.Int8x16Ty, false);
        break;
      case OpCode::I8x16__sub:
        compileVectorVectorSub(Context.Int8x16Ty);
        break;
      case OpCode::I8x16__sub_sat_s:
        compileVectorVectorSubSat(Context.Int8x16Ty, true);
        break;
      case OpCode::I8x16__sub_sat_u:
        compileVectorVectorSubSat(Context.Int8x16Ty, false);
        break;
      case OpCode::I8x16__min_s:
        compileVectorVectorSMin(Context.Int8x16Ty);
        break;
      case OpCode::I8x16__min_u:
        compileVectorVectorUMin(Context.Int8x16Ty);
        break;
      case OpCode::I8x16__max_s:
        compileVectorVectorSMax(Context.Int8x16Ty);
        break;
      case OpCode::I8x16__max_u:
        compileVectorVectorUMax(Context.Int8x16Ty);
        break;
      case OpCode::I8x16__avgr_u:
        compileVectorVectorUAvgr(Context.Int8x16Ty);
        break;
      case OpCode::I16x8__abs:
        compileVectorAbs(Context.Int16x8Ty);
        break;
      case OpCode::I16x8__neg:
        compileVectorNeg(Context.Int16x8Ty);
        break;
      case OpCode::I16x8__all_true:
        compileVectorAllTrue(Context.Int16x8Ty);
        break;
      case OpCode::I16x8__bitmask:
        compileVectorBitMask(Context.Int16x8Ty);
        break;
      case OpCode::I16x8__narrow_i32x4_s:
        compileVectorNarrow(Context.Int32x4Ty, true);
        break;
      case OpCode::I16x8__narrow_i32x4_u:
        compileVectorNarrow(Context.Int32x4Ty, false);
        break;
      case OpCode::I16x8__extend_low_i8x16_s:
        compileVectorExtend(Context.Int8x16Ty, true, true);
        break;
      case OpCode::I16x8__extend_high_i8x16_s:
        compileVectorExtend(Context.Int8x16Ty, true, false);
        break;
      case OpCode::I16x8__extend_low_i8x16_u:
        compileVectorExtend(Context.Int8x16Ty, false, true);
        break;
      case OpCode::I16x8__extend_high_i8x16_u:
        compileVectorExtend(Context.Int8x16Ty, false, false);
        break;
      case OpCode::I16x8__shl:
        compileVectorShl(Context.Int16x8Ty);
        break;
      case OpCode::I16x8__shr_s:
        compileVectorAShr(Context.Int16x8Ty);
        break;
      case OpCode::I16x8__shr_u:
        compileVectorLShr(Context.Int16x8Ty);
        break;
      case OpCode::I16x8__add:
        compileVectorVectorAdd(Context.Int16x8Ty);
        break;
      case OpCode::I16x8__add_sat_s:
        compileVectorVectorAddSat(Context.Int16x8Ty, true);
        break;
      case OpCode::I16x8__add_sat_u:
        compileVectorVectorAddSat(Context.Int16x8Ty, false);
        break;
      case OpCode::I16x8__sub:
        compileVectorVectorSub(Context.Int16x8Ty);
        break;
      case OpCode::I16x8__sub_sat_s:
        compileVectorVectorSubSat(Context.Int16x8Ty, true);
        break;
      case OpCode::I16x8__sub_sat_u:
        compileVectorVectorSubSat(Context.Int16x8Ty, false);
        break;
      case OpCode::I16x8__mul:
        compileVectorVectorMul(Context.Int16x8Ty);
        break;
      case OpCode::I16x8__min_s:
        compileVectorVectorSMin(Context.Int16x8Ty);
        break;
      case OpCode::I16x8__min_u:
        compileVectorVectorUMin(Context.Int16x8Ty);
        break;
      case OpCode::I16x8__max_s:
        compileVectorVectorSMax(Context.Int16x8Ty);
        break;
      case OpCode::I16x8__max_u:
        compileVectorVectorUMax(Context.Int16x8Ty);
        break;
      case OpCode::I16x8__avgr_u:
        compileVectorVectorUAvgr(Context.Int16x8Ty);
        break;
      case OpCode::I16x8__extmul_low_i8x16_s:
        compileVectorExtMul(Context.Int8x16Ty, true, true);
        break;
      case OpCode::I16x8__extmul_high_i8x16_s:
        compileVectorExtMul(Context.Int8x16Ty, true, false);
        break;
      case OpCode::I16x8__extmul_low_i8x16_u:
        compileVectorExtMul(Context.Int8x16Ty, false, true);
        break;
      case OpCode::I16x8__extmul_high_i8x16_u:
        compileVectorExtMul(Context.Int8x16Ty, false, false);
        break;
      case OpCode::I16x8__q15mulr_sat_s:
        compileVectorVectorQ15MulSat();
        break;
      case OpCode::I16x8__extadd_pairwise_i8x16_s:
        compileVectorExtAddPairwise(Context.Int8x16Ty, true);
        break;
      case OpCode::I16x8__extadd_pairwise_i8x16_u:
        compileVectorExtAddPairwise(Context.Int8x16Ty, false);
        break;
      case OpCode::I32x4__abs:
        compileVectorAbs(Context.Int32x4Ty);
        break;
      case OpCode::I32x4__neg:
        compileVectorNeg(Context.Int32x4Ty);
        break;
      case OpCode::I32x4__all_true:
        compileVectorAllTrue(Context.Int32x4Ty);
        break;
      case OpCode::I32x4__bitmask:
        compileVectorBitMask(Context.Int32x4Ty);
        break;
      case OpCode::I32x4__extend_low_i16x8_s:
        compileVectorExtend(Context.Int16x8Ty, true, true);
        break;
      case OpCode::I32x4__extend_high_i16x8_s:
        compileVectorExtend(Context.Int16x8Ty, true, false);
        break;
      case OpCode::I32x4__extend_low_i16x8_u:
        compileVectorExtend(Context.Int16x8Ty, false, true);
        break;
      case OpCode::I32x4__extend_high_i16x8_u:
        compileVectorExtend(Context.Int16x8Ty, false, false);
        break;
      case OpCode::I32x4__shl:
        compileVectorShl(Context.Int32x4Ty);
        break;
      case OpCode::I32x4__shr_s:
        compileVectorAShr(Context.Int32x4Ty);
        break;
      case OpCode::I32x4__shr_u:
        compileVectorLShr(Context.Int32x4Ty);
        break;
      case OpCode::I32x4__add:
        compileVectorVectorAdd(Context.Int32x4Ty);
        break;
      case OpCode::I32x4__sub:
        compileVectorVectorSub(Context.Int32x4Ty);
        break;
      case OpCode::I32x4__mul:
        compileVectorVectorMul(Context.Int32x4Ty);
        break;
      case OpCode::I32x4__min_s:
        compileVectorVectorSMin(Context.Int32x4Ty);
        break;
      case OpCode::I32x4__min_u:
        compileVectorVectorUMin(Context.Int32x4Ty);
        break;
      case OpCode::I32x4__max_s:
        compileVectorVectorSMax(Context.Int32x4Ty);
        break;
      case OpCode::I32x4__max_u:
        compileVectorVectorUMax(Context.Int32x4Ty);
        break;
      case OpCode::I32x4__extmul_low_i16x8_s:
        compileVectorExtMul(Context.Int16x8Ty, true, true);
        break;
      case OpCode::I32x4__extmul_high_i16x8_s:
        compileVectorExtMul(Context.Int16x8Ty, true, false);
        break;
      case OpCode::I32x4__extmul_low_i16x8_u:
        compileVectorExtMul(Context.Int16x8Ty, false, true);
        break;
      case OpCode::I32x4__extmul_high_i16x8_u:
        compileVectorExtMul(Context.Int16x8Ty, false, false);
        break;
      case OpCode::I32x4__extadd_pairwise_i16x8_s:
        compileVectorExtAddPairwise(Context.Int16x8Ty, true);
        break;
      case OpCode::I32x4__extadd_pairwise_i16x8_u:
        compileVectorExtAddPairwise(Context.Int16x8Ty, false);
        break;
      case OpCode::I32x4__dot_i16x8_s: {
        auto ExtendTy = Context.Int16x8Ty.getExtendedElementVectorType();
        auto Undef = LLVM::Value::getUndef(ExtendTy);
        auto LHS = Builder.createSExt(
            Builder.createBitCast(stackPop(), Context.Int16x8Ty), ExtendTy);
        auto RHS = Builder.createSExt(
            Builder.createBitCast(stackPop(), Context.Int16x8Ty), ExtendTy);
        auto M = Builder.createMul(LHS, RHS);
        auto L = Builder.createShuffleVector(
            M, Undef,
            LLVM::Value::getConstVector32(LLContext, {0U, 2U, 4U, 6U}));
        auto R = Builder.createShuffleVector(
            M, Undef,
            LLVM::Value::getConstVector32(LLContext, {1U, 3U, 5U, 7U}));
        auto V = Builder.createAdd(L, R);
        stackPush(Builder.createBitCast(V, Context.Int64x2Ty));
        break;
      }
      case OpCode::I64x2__abs:
        compileVectorAbs(Context.Int64x2Ty);
        break;
      case OpCode::I64x2__neg:
        compileVectorNeg(Context.Int64x2Ty);
        break;
      case OpCode::I64x2__all_true:
        compileVectorAllTrue(Context.Int64x2Ty);
        break;
      case OpCode::I64x2__bitmask:
        compileVectorBitMask(Context.Int64x2Ty);
        break;
      case OpCode::I64x2__extend_low_i32x4_s:
        compileVectorExtend(Context.Int32x4Ty, true, true);
        break;
      case OpCode::I64x2__extend_high_i32x4_s:
        compileVectorExtend(Context.Int32x4Ty, true, false);
        break;
      case OpCode::I64x2__extend_low_i32x4_u:
        compileVectorExtend(Context.Int32x4Ty, false, true);
        break;
      case OpCode::I64x2__extend_high_i32x4_u:
        compileVectorExtend(Context.Int32x4Ty, false, false);
        break;
      case OpCode::I64x2__shl:
        compileVectorShl(Context.Int64x2Ty);
        break;
      case OpCode::I64x2__shr_s:
        compileVectorAShr(Context.Int64x2Ty);
        break;
      case OpCode::I64x2__shr_u:
        compileVectorLShr(Context.Int64x2Ty);
        break;
      case OpCode::I64x2__add:
        compileVectorVectorAdd(Context.Int64x2Ty);
        break;
      case OpCode::I64x2__sub:
        compileVectorVectorSub(Context.Int64x2Ty);
        break;
      case OpCode::I64x2__mul:
        compileVectorVectorMul(Context.Int64x2Ty);
        break;
      case OpCode::I64x2__extmul_low_i32x4_s:
        compileVectorExtMul(Context.Int32x4Ty, true, true);
        break;
      case OpCode::I64x2__extmul_high_i32x4_s:
        compileVectorExtMul(Context.Int32x4Ty, true, false);
        break;
      case OpCode::I64x2__extmul_low_i32x4_u:
        compileVectorExtMul(Context.Int32x4Ty, false, true);
        break;
      case OpCode::I64x2__extmul_high_i32x4_u:
        compileVectorExtMul(Context.Int32x4Ty, false, false);
        break;
      case OpCode::F32x4__abs:
        compileVectorFAbs(Context.Floatx4Ty);
        break;
      case OpCode::F32x4__neg:
        compileVectorFNeg(Context.Floatx4Ty);
        break;
      case OpCode::F32x4__sqrt:
        compileVectorFSqrt(Context.Floatx4Ty);
        break;
      case OpCode::F32x4__add:
        compileVectorVectorFAdd(Context.Floatx4Ty);
        break;
      case OpCode::F32x4__sub:
        compileVectorVectorFSub(Context.Floatx4Ty);
        break;
      case OpCode::F32x4__mul:
        compileVectorVectorFMul(Context.Floatx4Ty);
        break;
      case OpCode::F32x4__div:
        compileVectorVectorFDiv(Context.Floatx4Ty);
        break;
      case OpCode::F32x4__min:
        compileVectorVectorFMin(Context.Floatx4Ty);
        break;
      case OpCode::F32x4__max:
        compileVectorVectorFMax(Context.Floatx4Ty);
        break;
      case OpCode::F32x4__pmin:
        compileVectorVectorFPMin(Context.Floatx4Ty);
        break;
      case OpCode::F32x4__pmax:
        compileVectorVectorFPMax(Context.Floatx4Ty);
        break;
      case OpCode::F32x4__ceil:
        compileVectorFCeil(Context.Floatx4Ty);
        break;
      case OpCode::F32x4__floor:
        compileVectorFFloor(Context.Floatx4Ty);
        break;
      case OpCode::F32x4__trunc:
        compileVectorFTrunc(Context.Floatx4Ty);
        break;
      case OpCode::F32x4__nearest:
        compileVectorFNearest(Context.Floatx4Ty);
        break;
      case OpCode::F64x2__abs:
        compileVectorFAbs(Context.Doublex2Ty);
        break;
      case OpCode::F64x2__neg:
        compileVectorFNeg(Context.Doublex2Ty);
        break;
      case OpCode::F64x2__sqrt:
        compileVectorFSqrt(Context.Doublex2Ty);
        break;
      case OpCode::F64x2__add:
        compileVectorVectorFAdd(Context.Doublex2Ty);
        break;
      case OpCode::F64x2__sub:
        compileVectorVectorFSub(Context.Doublex2Ty);
        break;
      case OpCode::F64x2__mul:
        compileVectorVectorFMul(Context.Doublex2Ty);
        break;
      case OpCode::F64x2__div:
        compileVectorVectorFDiv(Context.Doublex2Ty);
        break;
      case OpCode::F64x2__min:
        compileVectorVectorFMin(Context.Doublex2Ty);
        break;
      case OpCode::F64x2__max:
        compileVectorVectorFMax(Context.Doublex2Ty);
        break;
      case OpCode::F64x2__pmin:
        compileVectorVectorFPMin(Context.Doublex2Ty);
        break;
      case OpCode::F64x2__pmax:
        compileVectorVectorFPMax(Context.Doublex2Ty);
        break;
      case OpCode::F64x2__ceil:
        compileVectorFCeil(Context.Doublex2Ty);
        break;
      case OpCode::F64x2__floor:
        compileVectorFFloor(Context.Doublex2Ty);
        break;
      case OpCode::F64x2__trunc:
        compileVectorFTrunc(Context.Doublex2Ty);
        break;
      case OpCode::F64x2__nearest:
        compileVectorFNearest(Context.Doublex2Ty);
        break;
      case OpCode::I32x4__trunc_sat_f32x4_s:
        compileVectorTruncSatS32(Context.Floatx4Ty, false);
        break;
      case OpCode::I32x4__trunc_sat_f32x4_u:
        compileVectorTruncSatU32(Context.Floatx4Ty, false);
        break;
      case OpCode::F32x4__convert_i32x4_s:
        compileVectorConvertS(Context.Int32x4Ty, Context.Floatx4Ty, false);
        break;
      case OpCode::F32x4__convert_i32x4_u:
        compileVectorConvertU(Context.Int32x4Ty, Context.Floatx4Ty, false);
        break;
      case OpCode::I32x4__trunc_sat_f64x2_s_zero:
        compileVectorTruncSatS32(Context.Doublex2Ty, true);
        break;
      case OpCode::I32x4__trunc_sat_f64x2_u_zero:
        compileVectorTruncSatU32(Context.Doublex2Ty, true);
        break;
      case OpCode::F64x2__convert_low_i32x4_s:
        compileVectorConvertS(Context.Int32x4Ty, Context.Doublex2Ty, true);
        break;
      case OpCode::F64x2__convert_low_i32x4_u:
        compileVectorConvertU(Context.Int32x4Ty, Context.Doublex2Ty, true);
        break;
      case OpCode::F32x4__demote_f64x2_zero:
        compileVectorDemote();
        break;
      case OpCode::F64x2__promote_low_f32x4:
        compileVectorPromote();
        break;

      // Relaxed SIMD Instructions
      case OpCode::I8x16__relaxed_swizzle:
        compileVectorSwizzle();
        break;
      case OpCode::I32x4__relaxed_trunc_f32x4_s:
        compileVectorTruncSatS32(Context.Floatx4Ty, false);
        break;
      case OpCode::I32x4__relaxed_trunc_f32x4_u:
        compileVectorTruncSatU32(Context.Floatx4Ty, false);
        break;
      case OpCode::I32x4__relaxed_trunc_f64x2_s_zero:
        compileVectorTruncSatS32(Context.Doublex2Ty, true);
        break;
      case OpCode::I32x4__relaxed_trunc_f64x2_u_zero:
        compileVectorTruncSatU32(Context.Doublex2Ty, true);
        break;
      case OpCode::F32x4__relaxed_madd:
        compileVectorVectorMAdd(Context.Floatx4Ty);
        break;
      case OpCode::F32x4__relaxed_nmadd:
        compileVectorVectorNMAdd(Context.Floatx4Ty);
        break;
      case OpCode::F64x2__relaxed_madd:
        compileVectorVectorMAdd(Context.Doublex2Ty);
        break;
      case OpCode::F64x2__relaxed_nmadd:
        compileVectorVectorNMAdd(Context.Doublex2Ty);
        break;
      case OpCode::I8x16__relaxed_laneselect:
      case OpCode::I16x8__relaxed_laneselect:
      case OpCode::I32x4__relaxed_laneselect:
      case OpCode::I64x2__relaxed_laneselect: {
        auto C = stackPop();
        auto V2 = stackPop();
        auto V1 = stackPop();
        stackPush(Builder.createXor(
            Builder.createAnd(Builder.createXor(V1, V2), C), V2));
        break;
      }
      case OpCode::F32x4__relaxed_min:
        compileVectorVectorFMin(Context.Floatx4Ty);
        break;
      case OpCode::F32x4__relaxed_max:
        compileVectorVectorFMax(Context.Floatx4Ty);
        break;
      case OpCode::F64x2__relaxed_min:
        compileVectorVectorFMin(Context.Doublex2Ty);
        break;
      case OpCode::F64x2__relaxed_max:
        compileVectorVectorFMax(Context.Doublex2Ty);
        break;
      case OpCode::I16x8__relaxed_q15mulr_s:
        compileVectorVectorQ15MulSat();
        break;
      case OpCode::I16x8__relaxed_dot_i8x16_i7x16_s:
        compileVectorRelaxedIntegerDotProduct();
        break;
      case OpCode::I32x4__relaxed_dot_i8x16_i7x16_add_s:
        compileVectorRelaxedIntegerDotProductAdd();
        break;

      // Atomic Instructions
      case OpCode::Atomic__fence:
        compileMemoryFence();
        break;
      case OpCode::Memory__atomic__notify:
        compileAtomicNotify(Instr.getTargetIndex(), Instr.getMemoryOffset());
        break;
      case OpCode::Memory__atomic__wait32:
        compileAtomicWait(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                          Context.Int32Ty, 32);
        break;
      case OpCode::Memory__atomic__wait64:
        compileAtomicWait(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                          Context.Int64Ty, 64);
        break;
      case OpCode::I32__atomic__load:
        compileAtomicLoad(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                          Instr.getMemoryAlign(), Context.Int32Ty,
                          Context.Int32Ty, true);
        break;
      case OpCode::I64__atomic__load:
        compileAtomicLoad(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                          Instr.getMemoryAlign(), Context.Int64Ty,
                          Context.Int64Ty, true);
        break;
      case OpCode::I32__atomic__load8_u:
        compileAtomicLoad(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                          Instr.getMemoryAlign(), Context.Int32Ty,
                          Context.Int8Ty);
        break;
      case OpCode::I32__atomic__load16_u:
        compileAtomicLoad(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                          Instr.getMemoryAlign(), Context.Int32Ty,
                          Context.Int16Ty);
        break;
      case OpCode::I64__atomic__load8_u:
        compileAtomicLoad(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                          Instr.getMemoryAlign(), Context.Int64Ty,
                          Context.Int8Ty);
        break;
      case OpCode::I64__atomic__load16_u:
        compileAtomicLoad(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                          Instr.getMemoryAlign(), Context.Int64Ty,
                          Context.Int16Ty);
        break;
      case OpCode::I64__atomic__load32_u:
        compileAtomicLoad(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                          Instr.getMemoryAlign(), Context.Int64Ty,
                          Context.Int32Ty);
        break;
      case OpCode::I32__atomic__store:
        compileAtomicStore(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                           Instr.getMemoryAlign(), Context.Int32Ty,
                           Context.Int32Ty, true);
        break;
      case OpCode::I64__atomic__store:
        compileAtomicStore(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                           Instr.getMemoryAlign(), Context.Int64Ty,
                           Context.Int64Ty, true);
        break;
      case OpCode::I32__atomic__store8:
        compileAtomicStore(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                           Instr.getMemoryAlign(), Context.Int32Ty,
                           Context.Int8Ty, true);
        break;
      case OpCode::I32__atomic__store16:
        compileAtomicStore(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                           Instr.getMemoryAlign(), Context.Int32Ty,
                           Context.Int16Ty, true);
        break;
      case OpCode::I64__atomic__store8:
        compileAtomicStore(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                           Instr.getMemoryAlign(), Context.Int64Ty,
                           Context.Int8Ty, true);
        break;
      case OpCode::I64__atomic__store16:
        compileAtomicStore(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                           Instr.getMemoryAlign(), Context.Int64Ty,
                           Context.Int16Ty, true);
        break;
      case OpCode::I64__atomic__store32:
        compileAtomicStore(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                           Instr.getMemoryAlign(), Context.Int64Ty,
                           Context.Int32Ty, true);
        break;
      case OpCode::I32__atomic__rmw__add:
        compileAtomicRMWOp(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                           Instr.getMemoryAlign(), LLVMAtomicRMWBinOpAdd,
                           Context.Int32Ty, Context.Int32Ty, true);
        break;
      case OpCode::I64__atomic__rmw__add:
        compileAtomicRMWOp(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                           Instr.getMemoryAlign(), LLVMAtomicRMWBinOpAdd,
                           Context.Int64Ty, Context.Int64Ty, true);
        break;
      case OpCode::I32__atomic__rmw8__add_u:
        compileAtomicRMWOp(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                           Instr.getMemoryAlign(), LLVMAtomicRMWBinOpAdd,
                           Context.Int32Ty, Context.Int8Ty);
        break;
      case OpCode::I32__atomic__rmw16__add_u:
        compileAtomicRMWOp(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                           Instr.getMemoryAlign(), LLVMAtomicRMWBinOpAdd,
                           Context.Int32Ty, Context.Int16Ty);
        break;
      case OpCode::I64__atomic__rmw8__add_u:
        compileAtomicRMWOp(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                           Instr.getMemoryAlign(), LLVMAtomicRMWBinOpAdd,
                           Context.Int64Ty, Context.Int8Ty);
        break;
      case OpCode::I64__atomic__rmw16__add_u:
        compileAtomicRMWOp(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                           Instr.getMemoryAlign(), LLVMAtomicRMWBinOpAdd,
                           Context.Int64Ty, Context.Int16Ty);
        break;
      case OpCode::I64__atomic__rmw32__add_u:
        compileAtomicRMWOp(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                           Instr.getMemoryAlign(), LLVMAtomicRMWBinOpAdd,
                           Context.Int64Ty, Context.Int32Ty);
        break;
      case OpCode::I32__atomic__rmw__sub:
        compileAtomicRMWOp(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                           Instr.getMemoryAlign(), LLVMAtomicRMWBinOpSub,
                           Context.Int32Ty, Context.Int32Ty, true);
        break;
      case OpCode::I64__atomic__rmw__sub:
        compileAtomicRMWOp(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                           Instr.getMemoryAlign(), LLVMAtomicRMWBinOpSub,
                           Context.Int64Ty, Context.Int64Ty, true);
        break;
      case OpCode::I32__atomic__rmw8__sub_u:
        compileAtomicRMWOp(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                           Instr.getMemoryAlign(), LLVMAtomicRMWBinOpSub,
                           Context.Int32Ty, Context.Int8Ty);
        break;
      case OpCode::I32__atomic__rmw16__sub_u:
        compileAtomicRMWOp(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                           Instr.getMemoryAlign(), LLVMAtomicRMWBinOpSub,
                           Context.Int32Ty, Context.Int16Ty);
        break;
      case OpCode::I64__atomic__rmw8__sub_u:
        compileAtomicRMWOp(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                           Instr.getMemoryAlign(), LLVMAtomicRMWBinOpSub,
                           Context.Int64Ty, Context.Int8Ty);
        break;
      case OpCode::I64__atomic__rmw16__sub_u:
        compileAtomicRMWOp(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                           Instr.getMemoryAlign(), LLVMAtomicRMWBinOpSub,
                           Context.Int64Ty, Context.Int16Ty);
        break;
      case OpCode::I64__atomic__rmw32__sub_u:
        compileAtomicRMWOp(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                           Instr.getMemoryAlign(), LLVMAtomicRMWBinOpSub,
                           Context.Int64Ty, Context.Int32Ty);
        break;
      case OpCode::I32__atomic__rmw__and:
        compileAtomicRMWOp(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                           Instr.getMemoryAlign(), LLVMAtomicRMWBinOpAnd,
                           Context.Int32Ty, Context.Int32Ty, true);
        break;
      case OpCode::I64__atomic__rmw__and:
        compileAtomicRMWOp(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                           Instr.getMemoryAlign(), LLVMAtomicRMWBinOpAnd,
                           Context.Int64Ty, Context.Int64Ty, true);
        break;
      case OpCode::I32__atomic__rmw8__and_u:
        compileAtomicRMWOp(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                           Instr.getMemoryAlign(), LLVMAtomicRMWBinOpAnd,
                           Context.Int32Ty, Context.Int8Ty);
        break;
      case OpCode::I32__atomic__rmw16__and_u:
        compileAtomicRMWOp(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                           Instr.getMemoryAlign(), LLVMAtomicRMWBinOpAnd,
                           Context.Int32Ty, Context.Int16Ty);
        break;
      case OpCode::I64__atomic__rmw8__and_u:
        compileAtomicRMWOp(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                           Instr.getMemoryAlign(), LLVMAtomicRMWBinOpAnd,
                           Context.Int64Ty, Context.Int8Ty);
        break;
      case OpCode::I64__atomic__rmw16__and_u:
        compileAtomicRMWOp(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                           Instr.getMemoryAlign(), LLVMAtomicRMWBinOpAnd,
                           Context.Int64Ty, Context.Int16Ty);
        break;
      case OpCode::I64__atomic__rmw32__and_u:
        compileAtomicRMWOp(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                           Instr.getMemoryAlign(), LLVMAtomicRMWBinOpAnd,
                           Context.Int64Ty, Context.Int32Ty);
        break;
      case OpCode::I32__atomic__rmw__or:
        compileAtomicRMWOp(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                           Instr.getMemoryAlign(), LLVMAtomicRMWBinOpOr,
                           Context.Int32Ty, Context.Int32Ty, true);
        break;
      case OpCode::I64__atomic__rmw__or:
        compileAtomicRMWOp(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                           Instr.getMemoryAlign(), LLVMAtomicRMWBinOpOr,
                           Context.Int64Ty, Context.Int64Ty, true);
        break;
      case OpCode::I32__atomic__rmw8__or_u:
        compileAtomicRMWOp(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                           Instr.getMemoryAlign(), LLVMAtomicRMWBinOpOr,
                           Context.Int32Ty, Context.Int8Ty);
        break;
      case OpCode::I32__atomic__rmw16__or_u:
        compileAtomicRMWOp(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                           Instr.getMemoryAlign(), LLVMAtomicRMWBinOpOr,
                           Context.Int32Ty, Context.Int16Ty);
        break;
      case OpCode::I64__atomic__rmw8__or_u:
        compileAtomicRMWOp(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                           Instr.getMemoryAlign(), LLVMAtomicRMWBinOpOr,
                           Context.Int64Ty, Context.Int8Ty);
        break;
      case OpCode::I64__atomic__rmw16__or_u:
        compileAtomicRMWOp(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                           Instr.getMemoryAlign(), LLVMAtomicRMWBinOpOr,
                           Context.Int64Ty, Context.Int16Ty);
        break;
      case OpCode::I64__atomic__rmw32__or_u:
        compileAtomicRMWOp(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                           Instr.getMemoryAlign(), LLVMAtomicRMWBinOpOr,
                           Context.Int64Ty, Context.Int32Ty);
        break;
      case OpCode::I32__atomic__rmw__xor:
        compileAtomicRMWOp(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                           Instr.getMemoryAlign(), LLVMAtomicRMWBinOpXor,
                           Context.Int32Ty, Context.Int32Ty, true);
        break;
      case OpCode::I64__atomic__rmw__xor:
        compileAtomicRMWOp(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                           Instr.getMemoryAlign(), LLVMAtomicRMWBinOpXor,
                           Context.Int64Ty, Context.Int64Ty, true);
        break;
      case OpCode::I32__atomic__rmw8__xor_u:
        compileAtomicRMWOp(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                           Instr.getMemoryAlign(), LLVMAtomicRMWBinOpXor,
                           Context.Int32Ty, Context.Int8Ty);
        break;
      case OpCode::I32__atomic__rmw16__xor_u:
        compileAtomicRMWOp(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                           Instr.getMemoryAlign(), LLVMAtomicRMWBinOpXor,
                           Context.Int32Ty, Context.Int16Ty);
        break;
      case OpCode::I64__atomic__rmw8__xor_u:
        compileAtomicRMWOp(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                           Instr.getMemoryAlign(), LLVMAtomicRMWBinOpXor,
                           Context.Int64Ty, Context.Int8Ty);
        break;
      case OpCode::I64__atomic__rmw16__xor_u:
        compileAtomicRMWOp(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                           Instr.getMemoryAlign(), LLVMAtomicRMWBinOpXor,
                           Context.Int64Ty, Context.Int16Ty);
        break;
      case OpCode::I64__atomic__rmw32__xor_u:
        compileAtomicRMWOp(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                           Instr.getMemoryAlign(), LLVMAtomicRMWBinOpXor,
                           Context.Int64Ty, Context.Int32Ty);
        break;
      case OpCode::I32__atomic__rmw__xchg:
        compileAtomicRMWOp(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                           Instr.getMemoryAlign(), LLVMAtomicRMWBinOpXchg,
                           Context.Int32Ty, Context.Int32Ty, true);
        break;
      case OpCode::I64__atomic__rmw__xchg:
        compileAtomicRMWOp(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                           Instr.getMemoryAlign(), LLVMAtomicRMWBinOpXchg,
                           Context.Int64Ty, Context.Int64Ty, true);
        break;
      case OpCode::I32__atomic__rmw8__xchg_u:
        compileAtomicRMWOp(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                           Instr.getMemoryAlign(), LLVMAtomicRMWBinOpXchg,
                           Context.Int32Ty, Context.Int8Ty);
        break;
      case OpCode::I32__atomic__rmw16__xchg_u:
        compileAtomicRMWOp(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                           Instr.getMemoryAlign(), LLVMAtomicRMWBinOpXchg,
                           Context.Int32Ty, Context.Int16Ty);
        break;
      case OpCode::I64__atomic__rmw8__xchg_u:
        compileAtomicRMWOp(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                           Instr.getMemoryAlign(), LLVMAtomicRMWBinOpXchg,
                           Context.Int64Ty, Context.Int8Ty);
        break;
      case OpCode::I64__atomic__rmw16__xchg_u:
        compileAtomicRMWOp(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                           Instr.getMemoryAlign(), LLVMAtomicRMWBinOpXchg,
                           Context.Int64Ty, Context.Int16Ty);
        break;
      case OpCode::I64__atomic__rmw32__xchg_u:
        compileAtomicRMWOp(Instr.getTargetIndex(), Instr.getMemoryOffset(),
                           Instr.getMemoryAlign(), LLVMAtomicRMWBinOpXchg,
                           Context.Int64Ty, Context.Int32Ty);
        break;
      case OpCode::I32__atomic__rmw__cmpxchg:
        compileAtomicCompareExchange(
            Instr.getTargetIndex(), Instr.getMemoryOffset(),
            Instr.getMemoryAlign(), Context.Int32Ty, Context.Int32Ty, true);
        break;
      case OpCode::I64__atomic__rmw__cmpxchg:
        compileAtomicCompareExchange(
            Instr.getTargetIndex(), Instr.getMemoryOffset(),
            Instr.getMemoryAlign(), Context.Int64Ty, Context.Int64Ty, true);
        break;
      case OpCode::I32__atomic__rmw8__cmpxchg_u:
        compileAtomicCompareExchange(
            Instr.getTargetIndex(), Instr.getMemoryOffset(),
            Instr.getMemoryAlign(), Context.Int32Ty, Context.Int8Ty);
        break;
      case OpCode::I32__atomic__rmw16__cmpxchg_u:
        compileAtomicCompareExchange(
            Instr.getTargetIndex(), Instr.getMemoryOffset(),
            Instr.getMemoryAlign(), Context.Int32Ty, Context.Int16Ty);
        break;
      case OpCode::I64__atomic__rmw8__cmpxchg_u:
        compileAtomicCompareExchange(
            Instr.getTargetIndex(), Instr.getMemoryOffset(),
            Instr.getMemoryAlign(), Context.Int64Ty, Context.Int8Ty);
        break;
      case OpCode::I64__atomic__rmw16__cmpxchg_u:
        compileAtomicCompareExchange(
            Instr.getTargetIndex(), Instr.getMemoryOffset(),
            Instr.getMemoryAlign(), Context.Int64Ty, Context.Int16Ty);
        break;
      case OpCode::I64__atomic__rmw32__cmpxchg_u:
        compileAtomicCompareExchange(
            Instr.getTargetIndex(), Instr.getMemoryOffset(),
            Instr.getMemoryAlign(), Context.Int64Ty, Context.Int32Ty);
        break;

      default:
        assumingUnreachable();
      }
      return {};
    };

    for (const auto &Instr : Instrs) {
      // Update instruction count
      if (LocalInstrCount) {
        Builder.createStore(
            Builder.createAdd(
                Builder.createLoad(Context.Int64Ty, LocalInstrCount),
                LLContext.getInt64(1)),
            LocalInstrCount);
      }
      if (LocalGas) {
        auto NewGas = Builder.createAdd(
            Builder.createLoad(Context.Int64Ty, LocalGas),
            Builder.createLoad(
                Context.Int64Ty,
                Builder.createConstInBoundsGEP2_64(
                    LLVM::Type::getArrayType(Context.Int64Ty, UINT16_MAX + 1),
                    Context.getCostTable(Builder, ExecCtx), 0,
                    uint16_t(Instr.getOpCode()))));
        Builder.createStore(NewGas, LocalGas);
      }

      // Make the instruction node according to Code.
      EXPECTED_TRY(Dispatch(Instr));
    }
    return {};
  }
  void compileSignedTrunc(LLVM::Type IntType) noexcept {
    auto NormBB = LLVM::BasicBlock::create(LLContext, F.Fn, "strunc.norm");
    auto NotMinBB = LLVM::BasicBlock::create(LLContext, F.Fn, "strunc.notmin");
    auto NotMaxBB = LLVM::BasicBlock::create(LLContext, F.Fn, "strunc.notmax");
    auto Value = stackPop();
    const auto [Precise, MinFp, MaxFp] =
        [IntType, Value]() -> std::tuple<bool, LLVM::Value, LLVM::Value> {
      const auto BitWidth = IntType.getIntegerBitWidth();
      const auto [Min, Max] = [BitWidth]() -> std::tuple<int64_t, int64_t> {
        switch (BitWidth) {
        case 32:
          return {std::numeric_limits<int32_t>::min(),
                  std::numeric_limits<int32_t>::max()};
        case 64:
          return {std::numeric_limits<int64_t>::min(),
                  std::numeric_limits<int64_t>::max()};
        default:
          assumingUnreachable();
        }
      }();
      auto FPType = Value.getType();
      assuming(FPType.isFloatTy() || FPType.isDoubleTy());
      const auto FPWidth = FPType.getFPMantissaWidth();
      return {BitWidth <= FPWidth, LLVM::Value::getConstReal(FPType, Min),
              LLVM::Value::getConstReal(FPType, Max)};
    }();

    auto IsNotNan = Builder.createLikely(Builder.createFCmpORD(Value, Value));
    Builder.createCondBr(IsNotNan, NormBB,
                         getTrapBB(ErrCode::Value::InvalidConvToInt));

    Builder.positionAtEnd(NormBB);
    assuming(LLVM::Core::Trunc != LLVM::Core::NotIntrinsic);
    auto Trunc = Builder.createUnaryIntrinsic(LLVM::Core::Trunc, Value);
    auto IsNotUnderflow =
        Builder.createLikely(Builder.createFCmpOGE(Trunc, MinFp));
    Builder.createCondBr(IsNotUnderflow, NotMinBB,
                         getTrapBB(ErrCode::Value::IntegerOverflow));

    Builder.positionAtEnd(NotMinBB);
    auto IsNotOverflow = Builder.createLikely(
        Builder.createFCmp(Precise ? LLVMRealOLE : LLVMRealOLT, Trunc, MaxFp));
    Builder.createCondBr(IsNotOverflow, NotMaxBB,
                         getTrapBB(ErrCode::Value::IntegerOverflow));

    Builder.positionAtEnd(NotMaxBB);
    stackPush(Builder.createFPToSI(Trunc, IntType));
  }
  void compileSignedTruncSat(LLVM::Type IntType) noexcept {
    auto CurrBB = Builder.getInsertBlock();
    auto NormBB = LLVM::BasicBlock::create(LLContext, F.Fn, "ssat.norm");
    auto NotMinBB = LLVM::BasicBlock::create(LLContext, F.Fn, "ssat.notmin");
    auto NotMaxBB = LLVM::BasicBlock::create(LLContext, F.Fn, "ssat.notmax");
    auto EndBB = LLVM::BasicBlock::create(LLContext, F.Fn, "ssat.end");
    auto Value = stackPop();
    const auto [Precise, MinInt, MaxInt, MinFp, MaxFp] = [IntType, Value]()
        -> std::tuple<bool, uint64_t, uint64_t, LLVM::Value, LLVM::Value> {
      const auto BitWidth = IntType.getIntegerBitWidth();
      const auto [Min, Max] = [BitWidth]() -> std::tuple<int64_t, int64_t> {
        switch (BitWidth) {
        case 32:
          return {std::numeric_limits<int32_t>::min(),
                  std::numeric_limits<int32_t>::max()};
        case 64:
          return {std::numeric_limits<int64_t>::min(),
                  std::numeric_limits<int64_t>::max()};
        default:
          assumingUnreachable();
        }
      }();
      auto FPType = Value.getType();
      assuming(FPType.isFloatTy() || FPType.isDoubleTy());
      const auto FPWidth = FPType.getFPMantissaWidth();
      return {BitWidth <= FPWidth, static_cast<uint64_t>(Min),
              static_cast<uint64_t>(Max),
              LLVM::Value::getConstReal(FPType, Min),
              LLVM::Value::getConstReal(FPType, Max)};
    }();

    auto IsNotNan = Builder.createLikely(Builder.createFCmpORD(Value, Value));
    Builder.createCondBr(IsNotNan, NormBB, EndBB);

    Builder.positionAtEnd(NormBB);
    assuming(LLVM::Core::Trunc != LLVM::Core::NotIntrinsic);
    auto Trunc = Builder.createUnaryIntrinsic(LLVM::Core::Trunc, Value);
    auto IsNotUnderflow =
        Builder.createLikely(Builder.createFCmpOGE(Trunc, MinFp));
    Builder.createCondBr(IsNotUnderflow, NotMinBB, EndBB);

    Builder.positionAtEnd(NotMinBB);
    auto IsNotOverflow = Builder.createLikely(
        Builder.createFCmp(Precise ? LLVMRealOLE : LLVMRealOLT, Trunc, MaxFp));
    Builder.createCondBr(IsNotOverflow, NotMaxBB, EndBB);

    Builder.positionAtEnd(NotMaxBB);
    auto IntValue = Builder.createFPToSI(Trunc, IntType);
    Builder.createBr(EndBB);

    Builder.positionAtEnd(EndBB);
    auto PHIRet = Builder.createPHI(IntType);
    PHIRet.addIncoming(LLVM::Value::getConstInt(IntType, 0, true), CurrBB);
    PHIRet.addIncoming(LLVM::Value::getConstInt(IntType, MinInt, true), NormBB);
    PHIRet.addIncoming(LLVM::Value::getConstInt(IntType, MaxInt, true),
                       NotMinBB);
    PHIRet.addIncoming(IntValue, NotMaxBB);

    stackPush(PHIRet);
  }
  void compileUnsignedTrunc(LLVM::Type IntType) noexcept {
    auto NormBB = LLVM::BasicBlock::create(LLContext, F.Fn, "utrunc.norm");
    auto NotMinBB = LLVM::BasicBlock::create(LLContext, F.Fn, "utrunc.notmin");
    auto NotMaxBB = LLVM::BasicBlock::create(LLContext, F.Fn, "utrunc.notmax");
    auto Value = stackPop();
    const auto [Precise, MinFp, MaxFp] =
        [IntType, Value]() -> std::tuple<bool, LLVM::Value, LLVM::Value> {
      const auto BitWidth = IntType.getIntegerBitWidth();
      const auto [Min, Max] = [BitWidth]() -> std::tuple<uint64_t, uint64_t> {
        switch (BitWidth) {
        case 32:
          return {std::numeric_limits<uint32_t>::min(),
                  std::numeric_limits<uint32_t>::max()};
        case 64:
          return {std::numeric_limits<uint64_t>::min(),
                  std::numeric_limits<uint64_t>::max()};
        default:
          assumingUnreachable();
        }
      }();
      auto FPType = Value.getType();
      assuming(FPType.isFloatTy() || FPType.isDoubleTy());
      const auto FPWidth = FPType.getFPMantissaWidth();
      return {BitWidth <= FPWidth, LLVM::Value::getConstReal(FPType, Min),
              LLVM::Value::getConstReal(FPType, Max)};
    }();

    auto IsNotNan = Builder.createLikely(Builder.createFCmpORD(Value, Value));
    Builder.createCondBr(IsNotNan, NormBB,
                         getTrapBB(ErrCode::Value::InvalidConvToInt));

    Builder.positionAtEnd(NormBB);
    assuming(LLVM::Core::Trunc != LLVM::Core::NotIntrinsic);
    auto Trunc = Builder.createUnaryIntrinsic(LLVM::Core::Trunc, Value);
    auto IsNotUnderflow =
        Builder.createLikely(Builder.createFCmpOGE(Trunc, MinFp));
    Builder.createCondBr(IsNotUnderflow, NotMinBB,
                         getTrapBB(ErrCode::Value::IntegerOverflow));

    Builder.positionAtEnd(NotMinBB);
    auto IsNotOverflow = Builder.createLikely(
        Builder.createFCmp(Precise ? LLVMRealOLE : LLVMRealOLT, Trunc, MaxFp));
    Builder.createCondBr(IsNotOverflow, NotMaxBB,
                         getTrapBB(ErrCode::Value::IntegerOverflow));

    Builder.positionAtEnd(NotMaxBB);
    stackPush(Builder.createFPToUI(Trunc, IntType));
  }
  void compileUnsignedTruncSat(LLVM::Type IntType) noexcept {
    auto CurrBB = Builder.getInsertBlock();
    auto NormBB = LLVM::BasicBlock::create(LLContext, F.Fn, "usat.norm");
    auto NotMaxBB = LLVM::BasicBlock::create(LLContext, F.Fn, "usat.notmax");
    auto EndBB = LLVM::BasicBlock::create(LLContext, F.Fn, "usat.end");
    auto Value = stackPop();
    const auto [Precise, MinInt, MaxInt, MinFp, MaxFp] = [IntType, Value]()
        -> std::tuple<bool, uint64_t, uint64_t, LLVM::Value, LLVM::Value> {
      const auto BitWidth = IntType.getIntegerBitWidth();
      const auto [Min, Max] = [BitWidth]() -> std::tuple<uint64_t, uint64_t> {
        switch (BitWidth) {
        case 32:
          return {std::numeric_limits<uint32_t>::min(),
                  std::numeric_limits<uint32_t>::max()};
        case 64:
          return {std::numeric_limits<uint64_t>::min(),
                  std::numeric_limits<uint64_t>::max()};
        default:
          assumingUnreachable();
        }
      }();
      auto FPType = Value.getType();
      assuming(FPType.isFloatTy() || FPType.isDoubleTy());
      const auto FPWidth = FPType.getFPMantissaWidth();
      return {BitWidth <= FPWidth, Min, Max,
              LLVM::Value::getConstReal(FPType, Min),
              LLVM::Value::getConstReal(FPType, Max)};
    }();

    assuming(LLVM::Core::Trunc != LLVM::Core::NotIntrinsic);
    auto Trunc = Builder.createUnaryIntrinsic(LLVM::Core::Trunc, Value);
    auto IsNotUnderflow =
        Builder.createLikely(Builder.createFCmpOGE(Trunc, MinFp));
    Builder.createCondBr(IsNotUnderflow, NormBB, EndBB);

    Builder.positionAtEnd(NormBB);
    auto IsNotOverflow = Builder.createLikely(
        Builder.createFCmp(Precise ? LLVMRealOLE : LLVMRealOLT, Trunc, MaxFp));
    Builder.createCondBr(IsNotOverflow, NotMaxBB, EndBB);

    Builder.positionAtEnd(NotMaxBB);
    auto IntValue = Builder.createFPToUI(Trunc, IntType);
    Builder.createBr(EndBB);

    Builder.positionAtEnd(EndBB);
    auto PHIRet = Builder.createPHI(IntType);
    PHIRet.addIncoming(LLVM::Value::getConstInt(IntType, MinInt), CurrBB);
    PHIRet.addIncoming(LLVM::Value::getConstInt(IntType, MaxInt), NormBB);
    PHIRet.addIncoming(IntValue, NotMaxBB);

    stackPush(PHIRet);
  }

  void compileAtomicCheckOffsetAlignment(LLVM::Value Offset,
                                         LLVM::Type IntType) noexcept {
    const auto BitWidth = IntType.getIntegerBitWidth();
    auto BWMask = LLContext.getInt64((BitWidth >> 3) - 1);
    auto Value = Builder.createAnd(Offset, BWMask);
    auto OkBB = LLVM::BasicBlock::create(LLContext, F.Fn, "address_align_ok");
    auto IsAddressAligned = Builder.createLikely(
        Builder.createICmpEQ(Value, LLContext.getInt64(0)));
    Builder.createCondBr(IsAddressAligned, OkBB,
                         getTrapBB(ErrCode::Value::UnalignedAtomicAccess));

    Builder.positionAtEnd(OkBB);
  }

  void compileMemoryFence() noexcept {
    Builder.createFence(LLVMAtomicOrderingSequentiallyConsistent);
  }
  void compileAtomicNotify(unsigned MemoryIndex,
                           uint64_t MemoryOffset) noexcept {
    auto Count = Builder.createZExt(stackPop(), Context.Int64Ty);
    auto Offset = Builder.createZExt(stackPop(), Context.Int64Ty);
    if (MemoryOffset != 0) {
      Offset = Builder.createAdd(Offset, LLContext.getInt64(MemoryOffset));
    }
    compileAtomicCheckOffsetAlignment(Offset, Context.Int32Ty);
    stackPush(Builder.createTrunc(
        Builder.createCall(
            Context.getIntrinsic(
                Builder, Executable::Intrinsics::kMemAtomicNotify,
                LLVM::Type::getFunctionType(
                    Context.Int64Ty,
                    {Context.Int32Ty, Context.Int64Ty, Context.Int64Ty},
                    false)),
            {LLContext.getInt32(MemoryIndex), Offset, Count}),
        Context.MemoryAddrTypes[MemoryIndex]));
  }
  void compileAtomicWait(unsigned MemoryIndex, uint64_t MemoryOffset,
                         LLVM::Type TargetType, uint32_t BitWidth) noexcept {
    auto Timeout = stackPop();
    auto ExpectedValue = Builder.createZExtOrTrunc(stackPop(), Context.Int64Ty);
    auto Offset = Builder.createZExt(stackPop(), Context.Int64Ty);
    if (MemoryOffset != 0) {
      Offset = Builder.createAdd(Offset, LLContext.getInt64(MemoryOffset));
    }
    compileAtomicCheckOffsetAlignment(Offset, TargetType);
    stackPush(Builder.createTrunc(
        Builder.createCall(
            Context.getIntrinsic(
                Builder, Executable::Intrinsics::kMemAtomicWait,
                LLVM::Type::getFunctionType(Context.Int64Ty,
                                            {Context.Int32Ty, Context.Int64Ty,
                                             Context.Int64Ty, Context.Int64Ty,
                                             Context.Int32Ty},
                                            false)),
            {LLContext.getInt32(MemoryIndex), Offset, ExpectedValue, Timeout,
             LLContext.getInt32(BitWidth)}),
        Context.MemoryAddrTypes[MemoryIndex]));
  }
  void compileAtomicLoad(unsigned MemoryIndex, uint64_t MemoryOffset,
                         unsigned Alignment, LLVM::Type IntType,
                         LLVM::Type TargetType, bool Signed = false) noexcept {

    auto Offset = Builder.createZExt(Stack.back(), Context.Int64Ty);
    if (MemoryOffset != 0) {
      Offset = Builder.createAdd(Offset, LLContext.getInt64(MemoryOffset));
    }
    compileAtomicCheckOffsetAlignment(Offset, TargetType);
    auto VPtr = Builder.createInBoundsGEP1(
        Context.Int8Ty, Context.getMemory(Builder, ExecCtx, MemoryIndex),
        Offset);

    auto Ptr = Builder.createBitCast(VPtr, TargetType.getPointerTo());
    auto Load = switchEndian(Builder.createLoad(TargetType, Ptr, true));
    Load.setAlignment(1 << Alignment);
    Load.setOrdering(LLVMAtomicOrderingSequentiallyConsistent);

    if (Signed) {
      Stack.back() = Builder.createSExt(Load, IntType);
    } else {
      Stack.back() = Builder.createZExt(Load, IntType);
    }
  }
  void compileAtomicStore(unsigned MemoryIndex, uint64_t MemoryOffset,
                          unsigned Alignment, LLVM::Type, LLVM::Type TargetType,
                          bool Signed = false) noexcept {
    auto V = stackPop();

    if (Signed) {
      V = Builder.createSExtOrTrunc(V, TargetType);
    } else {
      V = Builder.createZExtOrTrunc(V, TargetType);
    }
    V = switchEndian(V);
    auto Offset = Builder.createZExt(Stack.back(), Context.Int64Ty);
    if (MemoryOffset != 0) {
      Offset = Builder.createAdd(Offset, LLContext.getInt64(MemoryOffset));
    }
    compileAtomicCheckOffsetAlignment(Offset, TargetType);
    auto VPtr = Builder.createInBoundsGEP1(
        Context.Int8Ty, Context.getMemory(Builder, ExecCtx, MemoryIndex),
        Offset);
    auto Ptr = Builder.createBitCast(VPtr, TargetType.getPointerTo());
    auto Store = Builder.createStore(V, Ptr, true);
    Store.setAlignment(1 << Alignment);
    Store.setOrdering(LLVMAtomicOrderingSequentiallyConsistent);
  }

  void compileAtomicRMWOp(unsigned MemoryIndex, uint64_t MemoryOffset,
                          [[maybe_unused]] unsigned Alignment,
                          LLVMAtomicRMWBinOp BinOp, LLVM::Type IntType,
                          LLVM::Type TargetType, bool Signed = false) noexcept {
    auto Value = Builder.createSExtOrTrunc(stackPop(), TargetType);
    auto Offset = Builder.createZExt(Stack.back(), Context.Int64Ty);
    if (MemoryOffset != 0) {
      Offset = Builder.createAdd(Offset, LLContext.getInt64(MemoryOffset));
    }
    compileAtomicCheckOffsetAlignment(Offset, TargetType);
    auto VPtr = Builder.createInBoundsGEP1(
        Context.Int8Ty, Context.getMemory(Builder, ExecCtx, MemoryIndex),
        Offset);
    auto Ptr = Builder.createBitCast(VPtr, TargetType.getPointerTo());

    LLVM::Value Ret;
    if constexpr (Endian::native == Endian::big) {
      if (BinOp == LLVMAtomicRMWBinOp::LLVMAtomicRMWBinOpAdd ||
          BinOp == LLVMAtomicRMWBinOp::LLVMAtomicRMWBinOpSub) {
        auto AtomicBB = LLVM::BasicBlock::create(LLContext, F.Fn, "atomic.rmw");
        auto OkBB = LLVM::BasicBlock::create(LLContext, F.Fn, "atomic.rmw.ok");
        Builder.createBr(AtomicBB);
        Builder.positionAtEnd(AtomicBB);

        auto Load = Builder.createLoad(TargetType, Ptr, true);
        Load.setOrdering(LLVMAtomicOrderingMonotonic);
        Load.setAlignment(1 << Alignment);

        LLVM::Value New;
        if (BinOp == LLVMAtomicRMWBinOp::LLVMAtomicRMWBinOpAdd)
          New = Builder.createAdd(switchEndian(Load), Value);
        else if (BinOp == LLVMAtomicRMWBinOp::LLVMAtomicRMWBinOpSub) {
          New = Builder.createSub(switchEndian(Load), Value);
        } else {
          assumingUnreachable();
        }
        New = switchEndian(New);

        auto Exchange = Builder.createAtomicCmpXchg(
            Ptr, Load, New, LLVMAtomicOrderingSequentiallyConsistent,
            LLVMAtomicOrderingSequentiallyConsistent);

        Ret = Builder.createExtractValue(Exchange, 0);
        auto Success = Builder.createExtractValue(Exchange, 1);
        Builder.createCondBr(Success, OkBB, AtomicBB);
        Builder.positionAtEnd(OkBB);
      } else {
        Ret = Builder.createAtomicRMW(BinOp, Ptr, switchEndian(Value),
                                      LLVMAtomicOrderingSequentiallyConsistent);
      }
    } else {
      Ret = Builder.createAtomicRMW(BinOp, Ptr, switchEndian(Value),
                                    LLVMAtomicOrderingSequentiallyConsistent);
    }
    Ret = switchEndian(Ret);
#if LLVM_VERSION_MAJOR >= 13
    Ret.setAlignment(1 << Alignment);
#endif
    if (Signed) {
      Stack.back() = Builder.createSExt(Ret, IntType);
    } else {
      Stack.back() = Builder.createZExt(Ret, IntType);
    }
  }
  void compileAtomicCompareExchange(unsigned MemoryIndex, uint64_t MemoryOffset,
                                    [[maybe_unused]] unsigned Alignment,
                                    LLVM::Type IntType, LLVM::Type TargetType,
                                    bool Signed = false) noexcept {

    auto Replacement = Builder.createSExtOrTrunc(stackPop(), TargetType);
    auto Expected = Builder.createSExtOrTrunc(stackPop(), TargetType);
    auto Offset = Builder.createZExt(Stack.back(), Context.Int64Ty);
    if (MemoryOffset != 0) {
      Offset = Builder.createAdd(Offset, LLContext.getInt64(MemoryOffset));
    }
    compileAtomicCheckOffsetAlignment(Offset, TargetType);
    auto VPtr = Builder.createInBoundsGEP1(
        Context.Int8Ty, Context.getMemory(Builder, ExecCtx, MemoryIndex),
        Offset);
    auto Ptr = Builder.createBitCast(VPtr, TargetType.getPointerTo());

    auto Ret = Builder.createAtomicCmpXchg(
        Ptr, switchEndian(Expected), switchEndian(Replacement),
        LLVMAtomicOrderingSequentiallyConsistent,
        LLVMAtomicOrderingSequentiallyConsistent);
#if LLVM_VERSION_MAJOR >= 13
    Ret.setAlignment(1 << Alignment);
#endif
    auto OldVal = Builder.createExtractValue(Ret, 0);
    OldVal = switchEndian(OldVal);
    if (Signed) {
      Stack.back() = Builder.createSExt(OldVal, IntType);
    } else {
      Stack.back() = Builder.createZExt(OldVal, IntType);
    }
  }

  void compileReturn() noexcept {
    updateInstrCount();
    updateGas();
    auto Ty = F.Ty.getReturnType();
    if (Ty.isVoidTy()) {
      Builder.createRetVoid();
    } else if (Ty.isStructTy()) {
      const auto Count = Ty.getStructNumElements();
      std::vector<LLVM::Value> Ret(Count);
      for (unsigned I = 0; I < Count; ++I) {
        const unsigned J = Count - 1 - I;
        Ret[J] = stackPop();
      }
      Builder.createAggregateRet(Ret);
    } else {
      Builder.createRet(stackPop());
    }
  }

  void updateInstrCount() noexcept {
    if (LocalInstrCount) {
      auto Store [[maybe_unused]] = Builder.createAtomicRMW(
          LLVMAtomicRMWBinOpAdd, Context.getInstrCount(Builder, ExecCtx),
          Builder.createLoad(Context.Int64Ty, LocalInstrCount),
          LLVMAtomicOrderingMonotonic);
#if LLVM_VERSION_MAJOR >= 13
      Store.setAlignment(8);
#endif
      Builder.createStore(LLContext.getInt64(0), LocalInstrCount);
    }
  }

  void updateGas() noexcept {
    if (LocalGas) {
      auto CurrBB = Builder.getInsertBlock();
      auto CheckBB = LLVM::BasicBlock::create(LLContext, F.Fn, "gas_check");
      auto OkBB = LLVM::BasicBlock::create(LLContext, F.Fn, "gas_ok");
      auto EndBB = LLVM::BasicBlock::create(LLContext, F.Fn, "gas_end");

      auto Cost = Builder.createLoad(Context.Int64Ty, LocalGas);
      Cost.setAlignment(64);
      auto GasPtr = Context.getGas(Builder, ExecCtx);
      auto GasLimit = Context.getGasLimit(Builder, ExecCtx);
      auto Gas = Builder.createLoad(Context.Int64Ty, GasPtr);
      Gas.setAlignment(64);
      Gas.setOrdering(LLVMAtomicOrderingMonotonic);
      Builder.createBr(CheckBB);
      Builder.positionAtEnd(CheckBB);

      auto PHIOldGas = Builder.createPHI(Context.Int64Ty);
      auto NewGas = Builder.createAdd(PHIOldGas, Cost);
      auto IsGasRemain =
          Builder.createLikely(Builder.createICmpULE(NewGas, GasLimit));
      Builder.createCondBr(IsGasRemain, OkBB,
                           getTrapBB(ErrCode::Value::CostLimitExceeded));
      Builder.positionAtEnd(OkBB);

      auto RGasAndSucceed = Builder.createAtomicCmpXchg(
          GasPtr, PHIOldGas, NewGas, LLVMAtomicOrderingMonotonic,
          LLVMAtomicOrderingMonotonic);
#if LLVM_VERSION_MAJOR >= 13
      RGasAndSucceed.setAlignment(8);
#endif
      RGasAndSucceed.setWeak(true);
      auto RGas = Builder.createExtractValue(RGasAndSucceed, 0);
      auto Succeed = Builder.createExtractValue(RGasAndSucceed, 1);
      Builder.createCondBr(Builder.createLikely(Succeed), EndBB, CheckBB);
      Builder.positionAtEnd(EndBB);

      Builder.createStore(LLContext.getInt64(0), LocalGas);

      PHIOldGas.addIncoming(Gas, CurrBB);
      PHIOldGas.addIncoming(RGas, OkBB);
    }
  }

  void updateGasAtTrap() noexcept {
    if (LocalGas) {
      auto Update [[maybe_unused]] = Builder.createAtomicRMW(
          LLVMAtomicRMWBinOpAdd, Context.getGas(Builder, ExecCtx),
          Builder.createLoad(Context.Int64Ty, LocalGas),
          LLVMAtomicOrderingMonotonic);
#if LLVM_VERSION_MAJOR >= 13
      Update.setAlignment(8);
#endif
    }
  }

private:
  void compileCallOp(const unsigned int FuncIndex) noexcept {
    const auto &FuncType =
        Context.CompositeTypes[std::get<0>(Context.Functions[FuncIndex])]
            ->getFuncType();
    const auto &Function = std::get<1>(Context.Functions[FuncIndex]);
    const auto &ParamTypes = FuncType.getParamTypes();

    std::vector<LLVM::Value> Args(ParamTypes.size() + 1);
    Args[0] = F.Fn.getFirstParam();
    for (size_t I = 0; I < ParamTypes.size(); ++I) {
      const size_t J = ParamTypes.size() - 1 - I;
      Args[J + 1] = stackPop();
    }

    LLVM::Value Ret;
    if (IsLazyJIT) {
      bool IsImport = std::get<2>(Context.Functions[FuncIndex]) == nullptr;
      if (IsImport) {
        Ret = Builder.createCall(Function, Args);
      } else {
        auto FTy = toLLVMType(LLContext, Context.ExecCtxPtrTy, FuncType);

        if (Context.LazyJITCacheVars.size() <= FuncIndex) {
          Context.LazyJITCacheVars.resize(Context.Functions.size());
        }
        auto &CacheVar = Context.LazyJITCacheVars[FuncIndex];
        if (!CacheVar) {
          CacheVar = Context.LLModule.get().addGlobal(
              FTy.getPointerTo(), false, LLVMPrivateLinkage,
              LLVM::Value::getConstNull(FTy.getPointerTo()), "");
        }

        auto CheckBB = LLVM::BasicBlock::create(LLContext, F.Fn, "ic.check");
        auto ResolveBB =
            LLVM::BasicBlock::create(LLContext, F.Fn, "ic.resolve");
        auto CallBB = LLVM::BasicBlock::create(LLContext, F.Fn, "ic.call");

        Builder.createBr(CheckBB);
        Builder.positionAtEnd(CheckBB);

        auto CachedPtr =
            Builder.createLoad(FTy.getPointerTo(), CacheVar, false);
        CachedPtr.setAlignment(8);
        CachedPtr.setOrdering(LLVMAtomicOrderingAcquire);
        auto IsNull = Builder.createIsNull(CachedPtr);
        auto IsNotNull = Builder.createLikely(Builder.createNot(IsNull));
        Builder.createCondBr(IsNotNull, CallBB, ResolveBB);

        Builder.positionAtEnd(ResolveBB);
        auto FPtr = Builder.createCall(
            Context.getIntrinsic(
                Builder, Executable::Intrinsics::kFuncGetFuncSymbol,
                LLVM::Type::getFunctionType(FTy.getPointerTo(),
                                            {Context.Int32Ty}, false)),
            {LLContext.getInt32(FuncIndex)});
        auto Store = Builder.createStore(FPtr, CacheVar);
        Store.setAlignment(8);
        Store.setOrdering(LLVMAtomicOrderingRelease);
        Builder.createBr(CallBB);

        Builder.positionAtEnd(CallBB);
        auto FinalPtr = Builder.createPHI(FTy.getPointerTo());
        FinalPtr.addIncoming(CachedPtr, CheckBB);
        FinalPtr.addIncoming(FPtr, ResolveBB);

        Ret = Builder.createCall(LLVM::FunctionCallee(FTy, FinalPtr), Args);
      }
    } else {
      Ret = Builder.createCall(Function, Args);
    }

    auto Ty = Ret.getType();
    if (Ty.isVoidTy()) {
      // nothing to do
    } else if (Ty.isStructTy()) {
      for (auto Val : unpackStruct(Builder, Ret)) {
        stackPush(Val);
      }
    } else {
      stackPush(Ret);
    }
  }

  void compileIndirectCallOp(const uint32_t TableIndex,
                             const uint32_t FuncTypeIndex) noexcept {
    auto NotNullBB = LLVM::BasicBlock::create(LLContext, F.Fn, "c_i.not_null");
    auto IsNullBB = LLVM::BasicBlock::create(LLContext, F.Fn, "c_i.is_null");
    auto EndBB = LLVM::BasicBlock::create(LLContext, F.Fn, "c_i.end");

    LLVM::Value FuncIndex = stackPop();
    const auto &FuncType = Context.CompositeTypes[FuncTypeIndex]->getFuncType();
    auto FTy = toLLVMType(Context.LLContext, Context.ExecCtxPtrTy, FuncType);
    auto RTy = FTy.getReturnType();

    const size_t ArgSize = FuncType.getParamTypes().size();
    const size_t RetSize =
        RTy.isVoidTy() ? 0 : FuncType.getReturnTypes().size();
    std::vector<LLVM::Value> ArgsVec(ArgSize + 1, nullptr);
    ArgsVec[0] = F.Fn.getFirstParam();
    for (size_t I = 0; I < ArgSize; ++I) {
      const size_t J = ArgSize - I;
      ArgsVec[J] = stackPop();
    }

    std::vector<LLVM::Value> FPtrRetsVec;
    FPtrRetsVec.reserve(RetSize);
    {
      auto FPtr = Builder.createCall(
          Context.getIntrinsic(
              Builder, Executable::Intrinsics::kTableGetFuncSymbol,
              LLVM::Type::getFunctionType(
                  FTy.getPointerTo(),
                  {Context.Int32Ty, Context.Int32Ty, Context.Int32Ty}, false)),
          {LLContext.getInt32(TableIndex), LLContext.getInt32(FuncTypeIndex),
           FuncIndex});
      Builder.createCondBr(
          Builder.createLikely(Builder.createNot(Builder.createIsNull(FPtr))),
          NotNullBB, IsNullBB);
      Builder.positionAtEnd(NotNullBB);

      auto FPtrRet =
          Builder.createCall(LLVM::FunctionCallee{FTy, FPtr}, ArgsVec);
      if (RetSize == 0) {
        // nothing to do
      } else if (RetSize == 1) {
        FPtrRetsVec.push_back(FPtrRet);
      } else {
        for (auto Val : unpackStruct(Builder, FPtrRet)) {
          FPtrRetsVec.push_back(Val);
        }
      }
    }

    Builder.createBr(EndBB);
    Builder.positionAtEnd(IsNullBB);

    std::vector<LLVM::Value> RetsVec;
    {
      LLVM::Value Args = Builder.createArray(ArgSize, kValSize);
      LLVM::Value Rets = Builder.createArray(RetSize, kValSize);
      Builder.createArrayPtrStore(
          Span<LLVM::Value>(ArgsVec.begin() + 1, ArgSize), Args, Context.Int8Ty,
          kValSize);

      Builder.createCall(
          Context.getIntrinsic(
              Builder, Executable::Intrinsics::kCallIndirect,
              LLVM::Type::getFunctionType(Context.VoidTy,
                                          {Context.Int32Ty, Context.Int32Ty,
                                           Context.Int32Ty, Context.Int8PtrTy,
                                           Context.Int8PtrTy},
                                          false)),
          {LLContext.getInt32(TableIndex), LLContext.getInt32(FuncTypeIndex),
           FuncIndex, Args, Rets});

      if (RetSize == 0) {
        // nothing to do
      } else if (RetSize == 1) {
        RetsVec.push_back(
            Builder.createValuePtrLoad(RTy, Rets, Context.Int8Ty));
      } else {
        RetsVec = Builder.createArrayPtrLoad(RetSize, RTy, Rets, Context.Int8Ty,
                                             kValSize);
      }
      Builder.createBr(EndBB);
      Builder.positionAtEnd(EndBB);
    }

    for (unsigned I = 0; I < RetSize; ++I) {
      auto PHIRet = Builder.createPHI(FPtrRetsVec[I].getType());
      PHIRet.addIncoming(FPtrRetsVec[I], NotNullBB);
      PHIRet.addIncoming(RetsVec[I], IsNullBB);
      stackPush(PHIRet);
    }
  }

  void compileReturnCallOp(const unsigned int FuncIndex) noexcept {
    const auto &FuncType =
        Context.CompositeTypes[std::get<0>(Context.Functions[FuncIndex])]
            ->getFuncType();
    const auto &Function = std::get<1>(Context.Functions[FuncIndex]);
    const auto &ParamTypes = FuncType.getParamTypes();

    std::vector<LLVM::Value> Args(ParamTypes.size() + 1);
    Args[0] = F.Fn.getFirstParam();
    for (size_t I = 0; I < ParamTypes.size(); ++I) {
      const size_t J = ParamTypes.size() - 1 - I;
      Args[J + 1] = stackPop();
    }

    LLVM::Value Ret;
    if (IsLazyJIT) {
      bool IsImport = std::get<2>(Context.Functions[FuncIndex]) == nullptr;
      if (IsImport) {
        Ret = Builder.createCall(Function, Args);
      } else {
        auto FTy = toLLVMType(LLContext, Context.ExecCtxPtrTy, FuncType);

        if (Context.LazyJITCacheVars.size() <= FuncIndex) {
          Context.LazyJITCacheVars.resize(Context.Functions.size());
        }
        auto &CacheVar = Context.LazyJITCacheVars[FuncIndex];
        if (!CacheVar) {
          CacheVar = Context.LLModule.get().addGlobal(
              FTy.getPointerTo(), false, LLVMPrivateLinkage,
              LLVM::Value::getConstNull(FTy.getPointerTo()), "");
        }

        auto CheckBB = LLVM::BasicBlock::create(LLContext, F.Fn, "rc.check");
        auto ResolveBB =
            LLVM::BasicBlock::create(LLContext, F.Fn, "rc.resolve");
        auto CallBB = LLVM::BasicBlock::create(LLContext, F.Fn, "rc.call");

        Builder.createBr(CheckBB);
        Builder.positionAtEnd(CheckBB);

        auto CachedPtr =
            Builder.createLoad(FTy.getPointerTo(), CacheVar, false);
        CachedPtr.setAlignment(8);
        CachedPtr.setOrdering(LLVMAtomicOrderingAcquire);
        auto IsNull = Builder.createIsNull(CachedPtr);
        auto IsNotNull = Builder.createLikely(Builder.createNot(IsNull));
        Builder.createCondBr(IsNotNull, CallBB, ResolveBB);

        Builder.positionAtEnd(ResolveBB);
        auto FPtr = Builder.createCall(
            Context.getIntrinsic(
                Builder, Executable::Intrinsics::kFuncGetFuncSymbol,
                LLVM::Type::getFunctionType(FTy.getPointerTo(),
                                            {Context.Int32Ty}, false)),
            {LLContext.getInt32(FuncIndex)});
        auto Store = Builder.createStore(FPtr, CacheVar);
        Store.setAlignment(8);
        Store.setOrdering(LLVMAtomicOrderingRelease);
        Builder.createBr(CallBB);

        Builder.positionAtEnd(CallBB);
        auto FinalPtr = Builder.createPHI(FTy.getPointerTo());
        FinalPtr.addIncoming(CachedPtr, CheckBB);
        FinalPtr.addIncoming(FPtr, ResolveBB);

        Ret = Builder.createCall(LLVM::FunctionCallee(FTy, FinalPtr), Args);
      }
    } else {
      Ret = Builder.createCall(Function, Args);
    }

    Ret.setMustTailCall();
    auto Ty = Ret.getType();
    if (Ty.isVoidTy()) {
      Builder.createRetVoid();
    } else {
      Builder.createRet(Ret);
    }
  }

  void compileReturnIndirectCallOp(const uint32_t TableIndex,
                                   const uint32_t FuncTypeIndex) noexcept {
    auto NotNullBB = LLVM::BasicBlock::create(LLContext, F.Fn, "c_i.not_null");
    auto IsNullBB = LLVM::BasicBlock::create(LLContext, F.Fn, "c_i.is_null");

    LLVM::Value FuncIndex = stackPop();
    const auto &FuncType = Context.CompositeTypes[FuncTypeIndex]->getFuncType();
    auto FTy = toLLVMType(Context.LLContext, Context.ExecCtxPtrTy, FuncType);
    auto RTy = FTy.getReturnType();

    const size_t ArgSize = FuncType.getParamTypes().size();
    const size_t RetSize =
        RTy.isVoidTy() ? 0 : FuncType.getReturnTypes().size();
    std::vector<LLVM::Value> ArgsVec(ArgSize + 1, nullptr);
    ArgsVec[0] = F.Fn.getFirstParam();
    for (size_t I = 0; I < ArgSize; ++I) {
      const size_t J = ArgSize - I;
      ArgsVec[J] = stackPop();
    }

    {
      auto FPtr = Builder.createCall(
          Context.getIntrinsic(
              Builder, Executable::Intrinsics::kTableGetFuncSymbol,
              LLVM::Type::getFunctionType(
                  FTy.getPointerTo(),
                  {Context.Int32Ty, Context.Int32Ty, Context.Int32Ty}, false)),
          {LLContext.getInt32(TableIndex), LLContext.getInt32(FuncTypeIndex),
           FuncIndex});
      Builder.createCondBr(
          Builder.createLikely(Builder.createNot(Builder.createIsNull(FPtr))),
          NotNullBB, IsNullBB);
      Builder.positionAtEnd(NotNullBB);

      auto FPtrRet =
          Builder.createCall(LLVM::FunctionCallee(FTy, FPtr), ArgsVec);
      FPtrRet.setMustTailCall();
      if (RetSize == 0) {
        Builder.createRetVoid();
      } else {
        Builder.createRet(FPtrRet);
      }
    }

    Builder.positionAtEnd(IsNullBB);

    {
      LLVM::Value Args = Builder.createArray(ArgSize, kValSize);
      LLVM::Value Rets = Builder.createArray(RetSize, kValSize);
      Builder.createArrayPtrStore(
          Span<LLVM::Value>(ArgsVec.begin() + 1, ArgSize), Args, Context.Int8Ty,
          kValSize);

      Builder.createCall(
          Context.getIntrinsic(
              Builder, Executable::Intrinsics::kCallIndirect,
              LLVM::Type::getFunctionType(Context.VoidTy,
                                          {Context.Int32Ty, Context.Int32Ty,
                                           Context.Int32Ty, Context.Int8PtrTy,
                                           Context.Int8PtrTy},
                                          false)),
          {LLContext.getInt32(TableIndex), LLContext.getInt32(FuncTypeIndex),
           FuncIndex, Args, Rets});

      if (RetSize == 0) {
        Builder.createRetVoid();
      } else if (RetSize == 1) {
        Builder.createRet(
            Builder.createValuePtrLoad(RTy, Rets, Context.Int8Ty));
      } else {
        Builder.createAggregateRet(Builder.createArrayPtrLoad(
            RetSize, RTy, Rets, Context.Int8Ty, kValSize));
      }
    }
  }

  void compileCallRefOp(const unsigned int TypeIndex) noexcept {
    auto NotNullBB = LLVM::BasicBlock::create(LLContext, F.Fn, "c_r.not_null");
    auto IsNullBB = LLVM::BasicBlock::create(LLContext, F.Fn, "c_r.is_null");
    auto EndBB = LLVM::BasicBlock::create(LLContext, F.Fn, "c_i.end");

    auto Ref = Builder.createBitCast(stackPop(), Context.Int64x2Ty);
    auto OkBB = LLVM::BasicBlock::create(LLContext, F.Fn, "c_r.ref_not_null");
    auto IsRefNotNull = Builder.createLikely(Builder.createICmpNE(
        Builder.createExtractElement(Ref, LLContext.getInt64(1)),
        LLContext.getInt64(0)));
    Builder.createCondBr(IsRefNotNull, OkBB,
                         getTrapBB(ErrCode::Value::AccessNullFunc));
    Builder.positionAtEnd(OkBB);

    const auto &FuncType = Context.CompositeTypes[TypeIndex]->getFuncType();
    auto FTy = toLLVMType(Context.LLContext, Context.ExecCtxPtrTy, FuncType);
    auto RTy = FTy.getReturnType();

    const size_t ArgSize = FuncType.getParamTypes().size();
    const size_t RetSize =
        RTy.isVoidTy() ? 0 : FuncType.getReturnTypes().size();
    std::vector<LLVM::Value> ArgsVec(ArgSize + 1, nullptr);
    ArgsVec[0] = F.Fn.getFirstParam();
    for (size_t I = 0; I < ArgSize; ++I) {
      const size_t J = ArgSize - I;
      ArgsVec[J] = stackPop();
    }

    std::vector<LLVM::Value> FPtrRetsVec;
    FPtrRetsVec.reserve(RetSize);
    {
      auto FPtr = Builder.createCall(
          Context.getIntrinsic(
              Builder, Executable::Intrinsics::kRefGetFuncSymbol,
              LLVM::Type::getFunctionType(FTy.getPointerTo(),
                                          {Context.Int64x2Ty}, false)),
          {Ref});
      Builder.createCondBr(
          Builder.createLikely(Builder.createNot(Builder.createIsNull(FPtr))),
          NotNullBB, IsNullBB);
      Builder.positionAtEnd(NotNullBB);

      auto FPtrRet =
          Builder.createCall(LLVM::FunctionCallee{FTy, FPtr}, ArgsVec);
      if (RetSize == 0) {
        // nothing to do
      } else if (RetSize == 1) {
        FPtrRetsVec.push_back(FPtrRet);
      } else {
        for (auto Val : unpackStruct(Builder, FPtrRet)) {
          FPtrRetsVec.push_back(Val);
        }
      }
    }

    Builder.createBr(EndBB);
    Builder.positionAtEnd(IsNullBB);

    std::vector<LLVM::Value> RetsVec;
    {
      LLVM::Value Args = Builder.createArray(ArgSize, kValSize);
      LLVM::Value Rets = Builder.createArray(RetSize, kValSize);
      Builder.createArrayPtrStore(
          Span<LLVM::Value>(ArgsVec.begin() + 1, ArgSize), Args, Context.Int8Ty,
          kValSize);

      Builder.createCall(
          Context.getIntrinsic(
              Builder, Executable::Intrinsics::kCallRef,
              LLVM::Type::getFunctionType(
                  Context.VoidTy,
                  {Context.Int64x2Ty, Context.Int8PtrTy, Context.Int8PtrTy},
                  false)),
          {Ref, Args, Rets});

      if (RetSize == 0) {
        // nothing to do
      } else if (RetSize == 1) {
        RetsVec.push_back(
            Builder.createValuePtrLoad(RTy, Rets, Context.Int8Ty));
      } else {
        RetsVec = Builder.createArrayPtrLoad(RetSize, RTy, Rets, Context.Int8Ty,
                                             kValSize);
      }
      Builder.createBr(EndBB);
      Builder.positionAtEnd(EndBB);
    }

    for (unsigned I = 0; I < RetSize; ++I) {
      auto PHIRet = Builder.createPHI(FPtrRetsVec[I].getType());
      PHIRet.addIncoming(FPtrRetsVec[I], NotNullBB);
      PHIRet.addIncoming(RetsVec[I], IsNullBB);
      stackPush(PHIRet);
    }
  }

  void compileReturnCallRefOp(const unsigned int TypeIndex) noexcept {
    auto NotNullBB = LLVM::BasicBlock::create(LLContext, F.Fn, "c_r.not_null");
    auto IsNullBB = LLVM::BasicBlock::create(LLContext, F.Fn, "c_r.is_null");

    auto Ref = Builder.createBitCast(stackPop(), Context.Int64x2Ty);
    auto OkBB = LLVM::BasicBlock::create(LLContext, F.Fn, "c_r.ref_not_null");
    auto IsRefNotNull = Builder.createLikely(Builder.createICmpNE(
        Builder.createExtractElement(Ref, LLContext.getInt64(1)),
        LLContext.getInt64(0)));
    Builder.createCondBr(IsRefNotNull, OkBB,
                         getTrapBB(ErrCode::Value::AccessNullFunc));
    Builder.positionAtEnd(OkBB);

    const auto &FuncType = Context.CompositeTypes[TypeIndex]->getFuncType();
    auto FTy = toLLVMType(Context.LLContext, Context.ExecCtxPtrTy, FuncType);
    auto RTy = FTy.getReturnType();

    const size_t ArgSize = FuncType.getParamTypes().size();
    const size_t RetSize =
        RTy.isVoidTy() ? 0 : FuncType.getReturnTypes().size();
    std::vector<LLVM::Value> ArgsVec(ArgSize + 1, nullptr);
    ArgsVec[0] = F.Fn.getFirstParam();
    for (size_t I = 0; I < ArgSize; ++I) {
      const size_t J = ArgSize - I;
      ArgsVec[J] = stackPop();
    }

    {
      auto FPtr = Builder.createCall(
          Context.getIntrinsic(
              Builder, Executable::Intrinsics::kRefGetFuncSymbol,
              LLVM::Type::getFunctionType(FTy.getPointerTo(),
                                          {Context.Int64x2Ty}, false)),
          {Ref});
      Builder.createCondBr(
          Builder.createLikely(Builder.createNot(Builder.createIsNull(FPtr))),
          NotNullBB, IsNullBB);
      Builder.positionAtEnd(NotNullBB);

      auto FPtrRet =
          Builder.createCall(LLVM::FunctionCallee(FTy, FPtr), ArgsVec);
      FPtrRet.setMustTailCall();
      if (RetSize == 0) {
        Builder.createRetVoid();
      } else {
        Builder.createRet(FPtrRet);
      }
    }

    Builder.positionAtEnd(IsNullBB);

    {
      LLVM::Value Args = Builder.createArray(ArgSize, kValSize);
      LLVM::Value Rets = Builder.createArray(RetSize, kValSize);
      Builder.createArrayPtrStore(
          Span<LLVM::Value>(ArgsVec.begin() + 1, ArgSize), Args, Context.Int8Ty,
          kValSize);

      Builder.createCall(
          Context.getIntrinsic(
              Builder, Executable::Intrinsics::kCallRef,
              LLVM::Type::getFunctionType(
                  Context.VoidTy,
                  {Context.Int64x2Ty, Context.Int8PtrTy, Context.Int8PtrTy},
                  false)),
          {Ref, Args, Rets});

      if (RetSize == 0) {
        Builder.createRetVoid();
      } else if (RetSize == 1) {
        Builder.createRet(
            Builder.createValuePtrLoad(RTy, Rets, Context.Int8Ty));
      } else {
        Builder.createAggregateRet(Builder.createArrayPtrLoad(
            RetSize, RTy, Rets, Context.Int8Ty, kValSize));
      }
    }
  }

  void compileLoadOp(unsigned MemoryIndex, uint64_t Offset, unsigned Alignment,
                     LLVM::Type LoadTy) noexcept {
    if constexpr (kForceUnalignment) {
      Alignment = 0;
    }
    auto Off = Builder.createZExt(stackPop(), Context.Int64Ty);
    if (Offset != 0) {
      Off = Builder.createAdd(Off, LLContext.getInt64(Offset));
    }

    auto VPtr = Builder.createInBoundsGEP1(
        Context.Int8Ty, Context.getMemory(Builder, ExecCtx, MemoryIndex), Off);
    auto Ptr = Builder.createBitCast(VPtr, LoadTy.getPointerTo());
    auto LoadInst = Builder.createLoad(LoadTy, Ptr, true);
    LoadInst.setAlignment(1 << Alignment);
    stackPush(switchEndian(LoadInst));
  }
  void compileLoadOp(unsigned MemoryIndex, uint64_t Offset, unsigned Alignment,
                     LLVM::Type LoadTy, LLVM::Type ExtendTy,
                     bool Signed) noexcept {
    compileLoadOp(MemoryIndex, Offset, Alignment, LoadTy);
    if (Signed) {
      Stack.back() = Builder.createSExt(Stack.back(), ExtendTy);
    } else {
      Stack.back() = Builder.createZExt(Stack.back(), ExtendTy);
    }
  }
  void compileVectorLoadOp(unsigned MemoryIndex, uint64_t Offset,
                           unsigned Alignment, LLVM::Type LoadTy) noexcept {
    compileLoadOp(MemoryIndex, Offset, Alignment, LoadTy);
    Stack.back() = Builder.createBitCast(Stack.back(), Context.Int64x2Ty);
  }
  void compileVectorLoadOp(unsigned MemoryIndex, uint64_t Offset,
                           unsigned Alignment, LLVM::Type LoadTy,
                           LLVM::Type ExtendTy, bool Signed) noexcept {
    compileLoadOp(MemoryIndex, Offset, Alignment, LoadTy, ExtendTy, Signed);
    Stack.back() = Builder.createBitCast(Stack.back(), Context.Int64x2Ty);
  }
  void compileSplatLoadOp(unsigned MemoryIndex, uint64_t Offset,
                          unsigned Alignment, LLVM::Type LoadTy,
                          LLVM::Type VectorTy) noexcept {
    compileLoadOp(MemoryIndex, Offset, Alignment, LoadTy);
    compileSplatOp(VectorTy);
  }
  void compileLoadLaneOp(unsigned MemoryIndex, uint64_t Offset,
                         unsigned Alignment, unsigned Index, LLVM::Type LoadTy,
                         LLVM::Type VectorTy) noexcept {
    auto Vector = stackPop();
    compileLoadOp(MemoryIndex, Offset, Alignment, LoadTy);
    if constexpr (Endian::native == Endian::big) {
      Index = VectorTy.getVectorSize() - 1 - Index;
    }
    auto Value = Stack.back();
    Stack.back() = Builder.createBitCast(
        Builder.createInsertElement(Builder.createBitCast(Vector, VectorTy),
                                    Value, LLContext.getInt64(Index)),
        Context.Int64x2Ty);
  }
  void compileStoreOp(uint32_t MemoryIndex, uint64_t Offset, uint32_t Alignment,
                      LLVM::Type LoadTy, bool Trunc = false,
                      bool BitCast = false) noexcept {
    if constexpr (kForceUnalignment) {
      Alignment = 0;
    }
    auto V = stackPop();
    auto Off = Builder.createZExt(stackPop(), Context.Int64Ty);
    if (Offset != 0) {
      Off = Builder.createAdd(Off, LLContext.getInt64(Offset));
    }

    if (Trunc) {
      V = Builder.createTrunc(V, LoadTy);
    }
    if (BitCast) {
      V = Builder.createBitCast(V, LoadTy);
    }
    V = switchEndian(V);
    auto VPtr = Builder.createInBoundsGEP1(
        Context.Int8Ty, Context.getMemory(Builder, ExecCtx, MemoryIndex), Off);
    auto Ptr = Builder.createBitCast(VPtr, LoadTy.getPointerTo());
    auto StoreInst = Builder.createStore(V, Ptr, true);
    StoreInst.setAlignment(1 << Alignment);
  }
  void compileStoreLaneOp(uint32_t MemoryIndex, uint64_t Offset,
                          uint32_t Alignment, uint8_t Index, LLVM::Type LoadTy,
                          LLVM::Type VectorTy) noexcept {
    auto Vector = Stack.back();
    if constexpr (Endian::native == Endian::big) {
      Index = static_cast<uint8_t>(VectorTy.getVectorSize() - Index - 1);
    }
    Stack.back() = Builder.createExtractElement(
        Builder.createBitCast(Vector, VectorTy), LLContext.getInt64(Index));
    compileStoreOp(MemoryIndex, Offset, Alignment, LoadTy);
  }
  void compileSplatOp(LLVM::Type VectorTy) noexcept {
    auto Undef = LLVM::Value::getUndef(VectorTy);
    auto Zeros = LLVM::Value::getConstNull(
        LLVM::Type::getVectorType(Context.Int32Ty, VectorTy.getVectorSize()));
    auto Value = Builder.createTrunc(Stack.back(), VectorTy.getElementType());
    auto Vector =
        Builder.createInsertElement(Undef, Value, LLContext.getInt64(0));
    Vector = Builder.createShuffleVector(Vector, Undef, Zeros);

    Stack.back() = Builder.createBitCast(Vector, Context.Int64x2Ty);
  }
  void compileExtractLaneOp(LLVM::Type VectorTy, unsigned Index) noexcept {
    auto Vector = Builder.createBitCast(Stack.back(), VectorTy);
    if constexpr (Endian::native == Endian::big) {
      Index = VectorTy.getVectorSize() - Index - 1;
    }
    Stack.back() =
        Builder.createExtractElement(Vector, LLContext.getInt64(Index));
  }
  void compileExtractLaneOp(LLVM::Type VectorTy, unsigned Index,
                            LLVM::Type ExtendTy, bool Signed) noexcept {
    compileExtractLaneOp(VectorTy, Index);
    if (Signed) {
      Stack.back() = Builder.createSExt(Stack.back(), ExtendTy);
    } else {
      Stack.back() = Builder.createZExt(Stack.back(), ExtendTy);
    }
  }
  void compileReplaceLaneOp(LLVM::Type VectorTy, unsigned Index) noexcept {
    auto Value = Builder.createTrunc(stackPop(), VectorTy.getElementType());
    auto Vector = Stack.back();
    if constexpr (Endian::native == Endian::big) {
      Index = VectorTy.getVectorSize() - Index - 1;
    }
    Stack.back() = Builder.createBitCast(
        Builder.createInsertElement(Builder.createBitCast(Vector, VectorTy),
                                    Value, LLContext.getInt64(Index)),
        Context.Int64x2Ty);
  }
  void compileVectorCompareOp(LLVM::Type VectorTy,
                              LLVMIntPredicate Predicate) noexcept {
    auto RHS = stackPop();
    auto LHS = stackPop();
    auto Result = Builder.createSExt(
        Builder.createICmp(Predicate, Builder.createBitCast(LHS, VectorTy),
                           Builder.createBitCast(RHS, VectorTy)),
        VectorTy);
    stackPush(Builder.createBitCast(Result, Context.Int64x2Ty));
  }
  void compileVectorCompareOp(LLVM::Type VectorTy, LLVMRealPredicate Predicate,
                              LLVM::Type ResultTy) noexcept {
    auto RHS = stackPop();
    auto LHS = stackPop();
    auto Result = Builder.createSExt(
        Builder.createFCmp(Predicate, Builder.createBitCast(LHS, VectorTy),
                           Builder.createBitCast(RHS, VectorTy)),
        ResultTy);
    stackPush(Builder.createBitCast(Result, Context.Int64x2Ty));
  }
  template <typename Func>
  void compileVectorOp(LLVM::Type VectorTy, Func &&Op) noexcept {
    auto V = Builder.createBitCast(Stack.back(), VectorTy);
    Stack.back() = Builder.createBitCast(Op(V), Context.Int64x2Ty);
  }
  void compileVectorAbs(LLVM::Type VectorTy) noexcept {
    compileVectorOp(VectorTy, [this](auto V) noexcept {
      return Builder.createIntrinsic(LLVM::Core::Abs, {V.getType()},
                                     {V, LLContext.getFalse()});
    });
  }
  void compileVectorNeg(LLVM::Type VectorTy) noexcept {
    compileVectorOp(VectorTy,
                    [this](auto V) noexcept { return Builder.createNeg(V); });
  }
  void compileVectorPopcnt() noexcept {
    compileVectorOp(Context.Int8x16Ty, [this](auto V) noexcept {
      assuming(LLVM::Core::Ctpop != LLVM::Core::NotIntrinsic);
      return Builder.createUnaryIntrinsic(LLVM::Core::Ctpop, V);
    });
  }
  template <typename Func>
  void compileVectorReduceIOp(LLVM::Type VectorTy, Func &&Op) noexcept {
    auto V = Builder.createBitCast(Stack.back(), VectorTy);
    Stack.back() = Builder.createZExt(Op(V), Context.Int32Ty);
  }
  void compileVectorAnyTrue() noexcept {
    compileVectorReduceIOp(Context.Int128x1Ty, [this](auto V) noexcept {
      auto Zero = LLVM::Value::getConstNull(Context.Int128x1Ty);
      return Builder.createBitCast(Builder.createICmpNE(V, Zero),
                                   LLContext.getInt1Ty());
    });
  }
  void compileVectorAllTrue(LLVM::Type VectorTy) noexcept {
    compileVectorReduceIOp(VectorTy, [this, VectorTy](auto V) noexcept {
      const auto Size = VectorTy.getVectorSize();
      auto IntType = LLContext.getIntNTy(Size);
      auto Zero = LLVM::Value::getConstNull(VectorTy);
      auto Cmp = Builder.createBitCast(Builder.createICmpEQ(V, Zero), IntType);
      auto CmpZero = LLVM::Value::getConstInt(IntType, 0);
      return Builder.createICmpEQ(Cmp, CmpZero);
    });
  }
  void compileVectorBitMask(LLVM::Type VectorTy) noexcept {
    compileVectorReduceIOp(VectorTy, [this, VectorTy](auto V) noexcept {
      const auto Size = VectorTy.getVectorSize();
      auto IntType = LLContext.getIntNTy(Size);
      auto Zero = LLVM::Value::getConstNull(VectorTy);
      return Builder.createBitCast(Builder.createICmpSLT(V, Zero), IntType);
    });
  }
  template <typename Func>
  void compileVectorShiftOp(LLVM::Type VectorTy, Func &&Op) noexcept {
    const bool Trunc = VectorTy.getElementType().getIntegerBitWidth() < 32;
    const uint32_t Mask = VectorTy.getElementType().getIntegerBitWidth() - 1;
    auto N = Builder.createAnd(stackPop(), LLContext.getInt32(Mask));
    auto RHS = Builder.createVectorSplat(
        VectorTy.getVectorSize(),
        Trunc ? Builder.createTrunc(N, VectorTy.getElementType())
              : Builder.createZExtOrTrunc(N, VectorTy.getElementType()));
    auto LHS = Builder.createBitCast(stackPop(), VectorTy);
    stackPush(Builder.createBitCast(Op(LHS, RHS), Context.Int64x2Ty));
  }
  void compileVectorShl(LLVM::Type VectorTy) noexcept {
    compileVectorShiftOp(VectorTy, [this](auto LHS, auto RHS) noexcept {
      return Builder.createShl(LHS, RHS);
    });
  }
  void compileVectorLShr(LLVM::Type VectorTy) noexcept {
    compileVectorShiftOp(VectorTy, [this](auto LHS, auto RHS) noexcept {
      return Builder.createLShr(LHS, RHS);
    });
  }
  void compileVectorAShr(LLVM::Type VectorTy) noexcept {
    compileVectorShiftOp(VectorTy, [this](auto LHS, auto RHS) noexcept {
      return Builder.createAShr(LHS, RHS);
    });
  }
  template <typename Func>
  void compileVectorVectorOp(LLVM::Type VectorTy, Func &&Op) noexcept {
    auto RHS = Builder.createBitCast(stackPop(), VectorTy);
    auto LHS = Builder.createBitCast(stackPop(), VectorTy);
    stackPush(Builder.createBitCast(Op(LHS, RHS), Context.Int64x2Ty));
  }
  void compileVectorVectorAdd(LLVM::Type VectorTy) noexcept {
    compileVectorVectorOp(VectorTy, [this](auto LHS, auto RHS) noexcept {
      return Builder.createAdd(LHS, RHS);
    });
  }
  void compileVectorVectorAddSat(LLVM::Type VectorTy, bool Signed) noexcept {
    auto ID = Signed ? LLVM::Core::SAddSat : LLVM::Core::UAddSat;
    assuming(ID != LLVM::Core::NotIntrinsic);
    compileVectorVectorOp(
        VectorTy, [this, VectorTy, ID](auto LHS, auto RHS) noexcept {
          return Builder.createIntrinsic(ID, {VectorTy}, {LHS, RHS});
        });
  }
  void compileVectorVectorSub(LLVM::Type VectorTy) noexcept {
    compileVectorVectorOp(VectorTy, [this](auto LHS, auto RHS) noexcept {
      return Builder.createSub(LHS, RHS);
    });
  }
  void compileVectorVectorSubSat(LLVM::Type VectorTy, bool Signed) noexcept {
    auto ID = Signed ? LLVM::Core::SSubSat : LLVM::Core::USubSat;
    assuming(ID != LLVM::Core::NotIntrinsic);
    compileVectorVectorOp(
        VectorTy, [this, VectorTy, ID](auto LHS, auto RHS) noexcept {
          return Builder.createIntrinsic(ID, {VectorTy}, {LHS, RHS});
        });
  }
  void compileVectorVectorMul(LLVM::Type VectorTy) noexcept {
    compileVectorVectorOp(VectorTy, [this](auto LHS, auto RHS) noexcept {
      return Builder.createMul(LHS, RHS);
    });
  }
  void compileVectorSwizzle() noexcept {
    auto Index = Builder.createBitCast(stackPop(), Context.Int8x16Ty);
    auto Vector = Builder.createBitCast(stackPop(), Context.Int8x16Ty);

#if defined(__x86_64__)
    if (Context.SupportSSSE3) {
      auto Magic = Builder.createVectorSplat(16, LLContext.getInt8(112));
      auto Added = Builder.createAdd(Index, Magic);
      auto NewIndex = Builder.createSelect(
          Builder.createICmpUGT(Index, Added),
          LLVM::Value::getConstAllOnes(Context.Int8x16Ty), Added);
      assuming(LLVM::Core::X86SSSE3PShufB128 != LLVM::Core::NotIntrinsic);
      stackPush(Builder.createBitCast(
          Builder.createIntrinsic(LLVM::Core::X86SSSE3PShufB128, {},
                                  {Vector, NewIndex}),
          Context.Int64x2Ty));
      return;
    }
#endif

#if defined(__aarch64__)
    if (Context.SupportNEON) {
      assuming(LLVM::Core::AArch64NeonTbl1 != LLVM::Core::NotIntrinsic);
      stackPush(Builder.createBitCast(
          Builder.createIntrinsic(LLVM::Core::AArch64NeonTbl1,
                                  {Context.Int8x16Ty}, {Vector, Index}),
          Context.Int64x2Ty));
      return;
    }
#endif

    auto Mask = Builder.createVectorSplat(16, LLContext.getInt8(15));
    auto Zero = Builder.createVectorSplat(16, LLContext.getInt8(0));

#if defined(__s390x__)
    assuming(LLVM::Core::S390VPerm != LLVM::Core::NotIntrinsic);
    auto Exceed = Builder.createICmpULE(Index, Mask);
    Index = Builder.createSub(Mask, Index);
    auto Result = Builder.createIntrinsic(LLVM::Core::S390VPerm, {},
                                          {Vector, Zero, Index});
    Result = Builder.createSelect(Exceed, Result, Zero);
    stackPush(Builder.createBitCast(Result, Context.Int64x2Ty));
    return;
#endif

    // Fallback case.
    // If the SSSE3 is not supported on the x86_64 platform or
    // the NEON is not supported on the aarch64 platform,
    // then fallback to this.
    auto IsOver = Builder.createICmpUGT(Index, Mask);
    auto InboundIndex = Builder.createAnd(Index, Mask);
    auto Array = Builder.createArray(16, 1);
    for (size_t I = 0; I < 16; ++I) {
      Builder.createStore(
          Builder.createExtractElement(Vector, LLContext.getInt64(I)),
          Builder.createInBoundsGEP1(Context.Int8Ty, Array,
                                     LLContext.getInt64(I)));
    }
    LLVM::Value Ret = LLVM::Value::getUndef(Context.Int8x16Ty);
    for (size_t I = 0; I < 16; ++I) {
      auto Idx =
          Builder.createExtractElement(InboundIndex, LLContext.getInt64(I));
      auto Value = Builder.createLoad(
          Context.Int8Ty,
          Builder.createInBoundsGEP1(Context.Int8Ty, Array, Idx));
      Ret = Builder.createInsertElement(Ret, Value, LLContext.getInt64(I));
    }
    Ret = Builder.createSelect(IsOver, Zero, Ret);
    stackPush(Builder.createBitCast(Ret, Context.Int64x2Ty));
  }

  void compileVectorVectorQ15MulSat() noexcept {
    compileVectorVectorOp(
        Context.Int16x8Ty, [this](auto LHS, auto RHS) noexcept -> LLVM::Value {
#if defined(__x86_64__)
          if (Context.SupportSSSE3) {
            assuming(LLVM::Core::X86SSSE3PMulHrSw128 !=
                     LLVM::Core::NotIntrinsic);
            auto Result = Builder.createIntrinsic(
                LLVM::Core::X86SSSE3PMulHrSw128, {}, {LHS, RHS});
            auto IntMaxV = Builder.createVectorSplat(
                8, LLContext.getInt16(UINT16_C(0x8000)));
            auto NotOver = Builder.createSExt(
                Builder.createICmpEQ(Result, IntMaxV), Context.Int16x8Ty);
            return Builder.createXor(Result, NotOver);
          }
#endif

#if defined(__aarch64__)
          if (Context.SupportNEON) {
            assuming(LLVM::Core::AArch64NeonSQRDMulH !=
                     LLVM::Core::NotIntrinsic);
            return Builder.createBinaryIntrinsic(
                LLVM::Core::AArch64NeonSQRDMulH, LHS, RHS);
          }
#endif

          // Fallback case.
          // If the SSSE3 is not supported on the x86_64 platform or
          // the NEON is not supported on the aarch64 platform,
          // then fallback to this.
          auto ExtTy = Context.Int16x8Ty.getExtendedElementVectorType();
          auto Offset = Builder.createVectorSplat(
              8, LLContext.getInt32(UINT32_C(0x4000)));
          auto Shift =
              Builder.createVectorSplat(8, LLContext.getInt32(UINT32_C(15)));
          auto ExtLHS = Builder.createSExt(LHS, ExtTy);
          auto ExtRHS = Builder.createSExt(RHS, ExtTy);
          auto Result = Builder.createTrunc(
              Builder.createAShr(
                  Builder.createAdd(Builder.createMul(ExtLHS, ExtRHS), Offset),
                  Shift),
              Context.Int16x8Ty);
          auto IntMaxV = Builder.createVectorSplat(
              8, LLContext.getInt16(UINT16_C(0x8000)));
          auto NotOver = Builder.createSExt(
              Builder.createICmpEQ(Result, IntMaxV), Context.Int16x8Ty);
          return Builder.createXor(Result, NotOver);
        });
  }
  void compileVectorVectorSMin(LLVM::Type VectorTy) noexcept {
    compileVectorVectorOp(VectorTy, [this](auto LHS, auto RHS) noexcept {
      return Builder.createIntrinsic(LLVM::Core::SMin, {LHS.getType()},
                                     {LHS, RHS});
    });
  }
  void compileVectorVectorUMin(LLVM::Type VectorTy) noexcept {
    compileVectorVectorOp(VectorTy, [this](auto LHS, auto RHS) noexcept {
      return Builder.createIntrinsic(LLVM::Core::UMin, {LHS.getType()},
                                     {LHS, RHS});
    });
  }
  void compileVectorVectorSMax(LLVM::Type VectorTy) noexcept {
    compileVectorVectorOp(VectorTy, [this](auto LHS, auto RHS) noexcept {
      return Builder.createIntrinsic(LLVM::Core::SMax, {LHS.getType()},
                                     {LHS, RHS});
    });
  }
  void compileVectorVectorUMax(LLVM::Type VectorTy) noexcept {
    compileVectorVectorOp(VectorTy, [this](auto LHS, auto RHS) noexcept {
      return Builder.createIntrinsic(LLVM::Core::UMax, {LHS.getType()},
                                     {LHS, RHS});
    });
  }
  void compileVectorVectorUAvgr(LLVM::Type VectorTy) noexcept {
    auto ExtendTy = VectorTy.getExtendedElementVectorType();
    compileVectorVectorOp(
        VectorTy,
        [this, VectorTy, ExtendTy](auto LHS, auto RHS) noexcept -> LLVM::Value {
#if defined(__x86_64__)
          if (Context.SupportSSE2) {
            const auto ID = [VectorTy]() noexcept {
              switch (VectorTy.getElementType().getIntegerBitWidth()) {
              case 8:
                return LLVM::Core::X86SSE2PAvgB;
              case 16:
                return LLVM::Core::X86SSE2PAvgW;
              default:
                assumingUnreachable();
              }
            }();
            assuming(ID != LLVM::Core::NotIntrinsic);
            return Builder.createIntrinsic(ID, {}, {LHS, RHS});
          }
#endif

#if defined(__aarch64__)
          if (Context.SupportNEON) {
            assuming(LLVM::Core::AArch64NeonURHAdd != LLVM::Core::NotIntrinsic);
            return Builder.createBinaryIntrinsic(LLVM::Core::AArch64NeonURHAdd,
                                                 LHS, RHS);
          }
#endif

          // Fallback case.
          // If the SSE2 is not supported on the x86_64 platform or
          // the NEON is not supported on the aarch64 platform,
          // then fallback to this.
          auto EL = Builder.createZExt(LHS, ExtendTy);
          auto ER = Builder.createZExt(RHS, ExtendTy);
          auto One = Builder.createZExt(
              Builder.createVectorSplat(ExtendTy.getVectorSize(),
                                        LLContext.getTrue()),
              ExtendTy);
          return Builder.createTrunc(
              Builder.createLShr(
                  Builder.createAdd(Builder.createAdd(EL, ER), One), One),
              VectorTy);
        });
  }
  void compileVectorNarrow(LLVM::Type FromTy, bool Signed) noexcept {
    auto [MinInt,
          MaxInt] = [&]() noexcept -> std::tuple<LLVM::Value, LLVM::Value> {
      switch (FromTy.getElementType().getIntegerBitWidth()) {
      case 16: {
        const auto Min =
            static_cast<int16_t>(Signed ? std::numeric_limits<int8_t>::min()
                                        : std::numeric_limits<uint8_t>::min());
        const auto Max =
            static_cast<int16_t>(Signed ? std::numeric_limits<int8_t>::max()
                                        : std::numeric_limits<uint8_t>::max());
        return {LLContext.getInt16(static_cast<uint16_t>(Min)),
                LLContext.getInt16(static_cast<uint16_t>(Max))};
      }
      case 32: {
        const auto Min =
            static_cast<int32_t>(Signed ? std::numeric_limits<int16_t>::min()
                                        : std::numeric_limits<uint16_t>::min());
        const auto Max =
            static_cast<int32_t>(Signed ? std::numeric_limits<int16_t>::max()
                                        : std::numeric_limits<uint16_t>::max());
        return {LLContext.getInt32(static_cast<uint32_t>(Min)),
                LLContext.getInt32(static_cast<uint32_t>(Max))};
      }
      default:
        assumingUnreachable();
      }
    }();
    const auto Count = FromTy.getVectorSize();
    auto VMin = Builder.createVectorSplat(Count, MinInt);
    auto VMax = Builder.createVectorSplat(Count, MaxInt);

    auto TruncTy = FromTy.getTruncatedElementVectorType();

    auto F2 = Builder.createBitCast(stackPop(), FromTy);
    F2 = Builder.createSelect(Builder.createICmpSLT(F2, VMin), VMin, F2);
    F2 = Builder.createSelect(Builder.createICmpSGT(F2, VMax), VMax, F2);
    F2 = Builder.createTrunc(F2, TruncTy);

    auto F1 = Builder.createBitCast(stackPop(), FromTy);
    F1 = Builder.createSelect(Builder.createICmpSLT(F1, VMin), VMin, F1);
    F1 = Builder.createSelect(Builder.createICmpSGT(F1, VMax), VMax, F1);
    F1 = Builder.createTrunc(F1, TruncTy);

    std::vector<uint32_t> Mask(Count * 2);
    std::iota(Mask.begin(), Mask.end(), 0);
    auto V = Endian::native == Endian::little
                 ? Builder.createShuffleVector(
                       F1, F2, LLVM::Value::getConstVector32(LLContext, Mask))
                 : Builder.createShuffleVector(
                       F2, F1, LLVM::Value::getConstVector32(LLContext, Mask));
    stackPush(Builder.createBitCast(V, Context.Int64x2Ty));
  }
  void compileVectorExtend(LLVM::Type FromTy, bool Signed, bool Low) noexcept {
    auto ExtTy = FromTy.getExtendedElementVectorType();
    const auto Count = FromTy.getVectorSize();
    std::vector<uint32_t> Mask(Count / 2);
    if constexpr (Endian::native == Endian::big) {
      Low = !Low;
    }
    std::iota(Mask.begin(), Mask.end(), Low ? 0 : Count / 2);
    auto R = Builder.createBitCast(Stack.back(), FromTy);
    if (Signed) {
      R = Builder.createSExt(R, ExtTy);
    } else {
      R = Builder.createZExt(R, ExtTy);
    }
    R = Builder.createShuffleVector(
        R, LLVM::Value::getUndef(ExtTy),
        LLVM::Value::getConstVector32(LLContext, Mask));
    Stack.back() = Builder.createBitCast(R, Context.Int64x2Ty);
  }
  void compileVectorExtMul(LLVM::Type FromTy, bool Signed, bool Low) noexcept {
    auto ExtTy = FromTy.getExtendedElementVectorType();
    const auto Count = FromTy.getVectorSize();
    std::vector<uint32_t> Mask(Count / 2);
    std::iota(Mask.begin(), Mask.end(), Low ? 0 : Count / 2);
    auto Extend = [this, FromTy, Signed, ExtTy, &Mask](LLVM::Value R) noexcept {
      R = Builder.createBitCast(R, FromTy);
      if (Signed) {
        R = Builder.createSExt(R, ExtTy);
      } else {
        R = Builder.createZExt(R, ExtTy);
      }
      return Builder.createShuffleVector(
          R, LLVM::Value::getUndef(ExtTy),
          LLVM::Value::getConstVector32(LLContext, Mask));
    };
    auto RHS = Extend(stackPop());
    auto LHS = Extend(stackPop());
    stackPush(
        Builder.createBitCast(Builder.createMul(RHS, LHS), Context.Int64x2Ty));
  }
  void compileVectorExtAddPairwise(LLVM::Type VectorTy, bool Signed) noexcept {
    compileVectorOp(
        VectorTy, [this, VectorTy, Signed](auto V) noexcept -> LLVM::Value {
          auto ExtTy = VectorTy.getExtendedElementVectorType()
                           .getHalfElementsVectorType();
#if defined(__x86_64__)
          const auto Count = VectorTy.getVectorSize();
          if (Context.SupportXOP) {
            const auto ID = [Count, Signed]() noexcept {
              switch (Count) {
              case 8:
                return Signed ? LLVM::Core::X86XOpVPHAddWD
                              : LLVM::Core::X86XOpVPHAddUWD;
              case 16:
                return Signed ? LLVM::Core::X86XOpVPHAddBW
                              : LLVM::Core::X86XOpVPHAddUBW;
              default:
                assumingUnreachable();
              }
            }();
            assuming(ID != LLVM::Core::NotIntrinsic);
            return Builder.createUnaryIntrinsic(ID, V);
          }
          if (Context.SupportSSSE3 && Count == 16) {
            assuming(LLVM::Core::X86SSSE3PMAddUbSw128 !=
                     LLVM::Core::NotIntrinsic);
            if (Signed) {
              return Builder.createIntrinsic(
                  LLVM::Core::X86SSSE3PMAddUbSw128, {},
                  {Builder.createVectorSplat(16, LLContext.getInt8(1)), V});
            } else {
              return Builder.createIntrinsic(
                  LLVM::Core::X86SSSE3PMAddUbSw128, {},
                  {V, Builder.createVectorSplat(16, LLContext.getInt8(1))});
            }
          }
          if (Context.SupportSSE2 && Count == 8) {
            assuming(LLVM::Core::X86SSE2PMAddWd != LLVM::Core::NotIntrinsic);
            if (Signed) {
              return Builder.createIntrinsic(
                  LLVM::Core::X86SSE2PMAddWd, {},
                  {V, Builder.createVectorSplat(8, LLContext.getInt16(1))});
            } else {
              V = Builder.createXor(
                  V, Builder.createVectorSplat(8, LLContext.getInt16(0x8000)));
              V = Builder.createIntrinsic(
                  LLVM::Core::X86SSE2PMAddWd, {},
                  {V, Builder.createVectorSplat(8, LLContext.getInt16(1))});
              return Builder.createAdd(
                  V, Builder.createVectorSplat(4, LLContext.getInt32(0x10000)));
            }
          }
#endif

#if defined(__aarch64__)
          if (Context.SupportNEON) {
            const auto ID = Signed ? LLVM::Core::AArch64NeonSAddLP
                                   : LLVM::Core::AArch64NeonUAddLP;
            assuming(ID != LLVM::Core::NotIntrinsic);
            return Builder.createIntrinsic(ID, {ExtTy, VectorTy}, {V});
          }
#endif

          // Fallback case.
          // If the XOP, SSSE3, or SSE2 is not supported on the x86_64 platform
          // or the NEON is not supported on the aarch64 platform,
          // then fallback to this.
          auto Width = LLVM::Value::getConstInt(
              ExtTy.getElementType(),
              VectorTy.getElementType().getIntegerBitWidth());
          Width = Builder.createVectorSplat(ExtTy.getVectorSize(), Width);
          auto EV = Builder.createBitCast(V, ExtTy);
          LLVM::Value L, R;
          if (Signed) {
            L = Builder.createAShr(EV, Width);
            R = Builder.createAShr(Builder.createShl(EV, Width), Width);
          } else {
            L = Builder.createLShr(EV, Width);
            R = Builder.createLShr(Builder.createShl(EV, Width), Width);
          }
          return Builder.createAdd(L, R);
        });
  }
  void compileVectorFAbs(LLVM::Type VectorTy) noexcept {
    compileVectorOp(VectorTy, [this](auto V) noexcept {
      assuming(LLVM::Core::Fabs != LLVM::Core::NotIntrinsic);
      return Builder.createUnaryIntrinsic(LLVM::Core::Fabs, V);
    });
  }
  void compileVectorFNeg(LLVM::Type VectorTy) noexcept {
    compileVectorOp(VectorTy,
                    [this](auto V) noexcept { return Builder.createFNeg(V); });
  }
  void compileVectorFSqrt(LLVM::Type VectorTy) noexcept {
    compileVectorOp(VectorTy, [this](auto V) noexcept {
      assuming(LLVM::Core::Sqrt != LLVM::Core::NotIntrinsic);
      return Builder.createUnaryIntrinsic(LLVM::Core::Sqrt, V);
    });
  }
  void compileVectorFCeil(LLVM::Type VectorTy) noexcept {
    compileVectorOp(VectorTy, [this](auto V) noexcept {
      assuming(LLVM::Core::Ceil != LLVM::Core::NotIntrinsic);
      return Builder.createUnaryIntrinsic(LLVM::Core::Ceil, V);
    });
  }
  void compileVectorFFloor(LLVM::Type VectorTy) noexcept {
    compileVectorOp(VectorTy, [this](auto V) noexcept {
      assuming(LLVM::Core::Floor != LLVM::Core::NotIntrinsic);
      return Builder.createUnaryIntrinsic(LLVM::Core::Floor, V);
    });
  }
  void compileVectorFTrunc(LLVM::Type VectorTy) noexcept {
    compileVectorOp(VectorTy, [this](auto V) noexcept {
      assuming(LLVM::Core::Trunc != LLVM::Core::NotIntrinsic);
      return Builder.createUnaryIntrinsic(LLVM::Core::Trunc, V);
    });
  }
  void compileVectorFNearest(LLVM::Type VectorTy) noexcept {
    compileVectorOp(VectorTy, [&](auto V) noexcept {
#if LLVM_VERSION_MAJOR >= 12 && !defined(__s390x__)
      assuming(LLVM::Core::Roundeven != LLVM::Core::NotIntrinsic);
      if (LLVM::Core::Roundeven != LLVM::Core::NotIntrinsic) {
        return Builder.createUnaryIntrinsic(LLVM::Core::Roundeven, V);
      }
#endif

#if defined(__x86_64__)
      if (Context.SupportSSE4_1) {
        const bool IsFloat = VectorTy.getElementType().isFloatTy();
        auto ID =
            IsFloat ? LLVM::Core::X86SSE41RoundPs : LLVM::Core::X86SSE41RoundPd;
        assuming(ID != LLVM::Core::NotIntrinsic);
        return Builder.createIntrinsic(ID, {}, {V, LLContext.getInt32(8)});
      }
#endif

#if defined(__aarch64__)
      if (Context.SupportNEON &&
          LLVM::Core::AArch64NeonFRIntN != LLVM::Core::NotIntrinsic) {
        return Builder.createUnaryIntrinsic(LLVM::Core::AArch64NeonFRIntN, V);
      }
#endif

      // Fallback case.
      // If the SSE4.1 is not supported on the x86_64 platform or
      // the NEON is not supported on the aarch64 platform,
      // then fallback to this.
      assuming(LLVM::Core::Nearbyint != LLVM::Core::NotIntrinsic);
      return Builder.createUnaryIntrinsic(LLVM::Core::Nearbyint, V);
    });
  }
  void compileVectorVectorFAdd(LLVM::Type VectorTy) noexcept {
    compileVectorVectorOp(VectorTy, [this](auto LHS, auto RHS) noexcept {
      return Builder.createFAdd(LHS, RHS);
    });
  }
  void compileVectorVectorFSub(LLVM::Type VectorTy) noexcept {
    compileVectorVectorOp(VectorTy, [this](auto LHS, auto RHS) noexcept {
      return Builder.createFSub(LHS, RHS);
    });
  }
  void compileVectorVectorFMul(LLVM::Type VectorTy) noexcept {
    compileVectorVectorOp(VectorTy, [this](auto LHS, auto RHS) noexcept {
      return Builder.createFMul(LHS, RHS);
    });
  }
  void compileVectorVectorFDiv(LLVM::Type VectorTy) noexcept {
    compileVectorVectorOp(VectorTy, [this](auto LHS, auto RHS) noexcept {
      return Builder.createFDiv(LHS, RHS);
    });
  }
  void compileVectorVectorFMin(LLVM::Type VectorTy) noexcept {
    compileVectorVectorOp(VectorTy, [this](auto LHS, auto RHS) noexcept {
      auto LNaN = Builder.createFCmpUNO(LHS, LHS);
      auto RNaN = Builder.createFCmpUNO(RHS, RHS);
      auto OLT = Builder.createFCmpOLT(LHS, RHS);
      auto OGT = Builder.createFCmpOGT(LHS, RHS);
      auto Ret = Builder.createBitCast(
          Builder.createOr(Builder.createBitCast(LHS, Context.Int64x2Ty),
                           Builder.createBitCast(RHS, Context.Int64x2Ty)),
          LHS.getType());
      Ret = Builder.createSelect(OGT, RHS, Ret);
      Ret = Builder.createSelect(OLT, LHS, Ret);
      Ret = Builder.createSelect(LNaN, LHS, Ret);
      Ret = Builder.createSelect(RNaN, RHS, Ret);
      return Ret;
    });
  }
  void compileVectorVectorFMax(LLVM::Type VectorTy) noexcept {
    compileVectorVectorOp(VectorTy, [this](auto LHS, auto RHS) noexcept {
      auto LNaN = Builder.createFCmpUNO(LHS, LHS);
      auto RNaN = Builder.createFCmpUNO(RHS, RHS);
      auto OLT = Builder.createFCmpOLT(LHS, RHS);
      auto OGT = Builder.createFCmpOGT(LHS, RHS);
      auto Ret = Builder.createBitCast(
          Builder.createAnd(Builder.createBitCast(LHS, Context.Int64x2Ty),
                            Builder.createBitCast(RHS, Context.Int64x2Ty)),
          LHS.getType());
      Ret = Builder.createSelect(OLT, RHS, Ret);
      Ret = Builder.createSelect(OGT, LHS, Ret);
      Ret = Builder.createSelect(LNaN, LHS, Ret);
      Ret = Builder.createSelect(RNaN, RHS, Ret);
      return Ret;
    });
  }
  void compileVectorVectorFPMin(LLVM::Type VectorTy) noexcept {
    compileVectorVectorOp(VectorTy, [this](auto LHS, auto RHS) noexcept {
      auto Cmp = Builder.createFCmpOLT(RHS, LHS);
      return Builder.createSelect(Cmp, RHS, LHS);
    });
  }
  void compileVectorVectorFPMax(LLVM::Type VectorTy) noexcept {
    compileVectorVectorOp(VectorTy, [this](auto LHS, auto RHS) noexcept {
      auto Cmp = Builder.createFCmpOGT(RHS, LHS);
      return Builder.createSelect(Cmp, RHS, LHS);
    });
  }
  void compileVectorTruncSatS32(LLVM::Type VectorTy, bool PadZero) noexcept {
    compileVectorOp(VectorTy, [this, VectorTy, PadZero](auto V) noexcept {
      const auto Size = VectorTy.getVectorSize();
      auto FPTy = VectorTy.getElementType();
      auto IntMin = LLContext.getInt32(
          static_cast<uint32_t>(std::numeric_limits<int32_t>::min()));
      auto IntMax = LLContext.getInt32(
          static_cast<uint32_t>(std::numeric_limits<int32_t>::max()));
      auto IntMinV = Builder.createVectorSplat(Size, IntMin);
      auto IntMaxV = Builder.createVectorSplat(Size, IntMax);
      auto IntZeroV = LLVM::Value::getConstNull(IntMinV.getType());
      auto FPMin = Builder.createSIToFP(IntMin, FPTy);
      auto FPMax = Builder.createSIToFP(IntMax, FPTy);
      auto FPMinV = Builder.createVectorSplat(Size, FPMin);
      auto FPMaxV = Builder.createVectorSplat(Size, FPMax);

      auto Normal = Builder.createFCmpORD(V, V);
      auto NotUnder = Builder.createFCmpUGE(V, FPMinV);
      auto NotOver = Builder.createFCmpULT(V, FPMaxV);
      V = Builder.createFPToSI(
          V, LLVM::Type::getVectorType(LLContext.getInt32Ty(), Size));
      V = Builder.createSelect(Normal, V, IntZeroV);
      V = Builder.createSelect(NotUnder, V, IntMinV);
      V = Builder.createSelect(NotOver, V, IntMaxV);
      if (PadZero) {
        std::vector<uint32_t> Mask(Size * 2);
        std::iota(Mask.begin(), Mask.end(), 0);
        if constexpr (Endian::native == Endian::little) {
          V = Builder.createShuffleVector(
              V, IntZeroV, LLVM::Value::getConstVector32(LLContext, Mask));
        } else {
          V = Builder.createShuffleVector(
              IntZeroV, V, LLVM::Value::getConstVector32(LLContext, Mask));
        }
      }
      return V;
    });
  }
  void compileVectorTruncSatU32(LLVM::Type VectorTy, bool PadZero) noexcept {
    compileVectorOp(VectorTy, [this, VectorTy, PadZero](auto V) noexcept {
      const auto Size = VectorTy.getVectorSize();
      auto FPTy = VectorTy.getElementType();
      auto IntMin = LLContext.getInt32(std::numeric_limits<uint32_t>::min());
      auto IntMax = LLContext.getInt32(std::numeric_limits<uint32_t>::max());
      auto IntMinV = Builder.createVectorSplat(Size, IntMin);
      auto IntMaxV = Builder.createVectorSplat(Size, IntMax);
      auto FPMin = Builder.createUIToFP(IntMin, FPTy);
      auto FPMax = Builder.createUIToFP(IntMax, FPTy);
      auto FPMinV = Builder.createVectorSplat(Size, FPMin);
      auto FPMaxV = Builder.createVectorSplat(Size, FPMax);

      auto NotUnder = Builder.createFCmpOGE(V, FPMinV);
      auto NotOver = Builder.createFCmpULT(V, FPMaxV);
      V = Builder.createFPToUI(
          V, LLVM::Type::getVectorType(LLContext.getInt32Ty(), Size));
      V = Builder.createSelect(NotUnder, V, IntMinV);
      V = Builder.createSelect(NotOver, V, IntMaxV);
      if (PadZero) {
        auto IntZeroV = LLVM::Value::getConstNull(IntMinV.getType());
        std::vector<uint32_t> Mask(Size * 2);
        std::iota(Mask.begin(), Mask.end(), 0);
        if constexpr (Endian::native == Endian::little) {
          V = Builder.createShuffleVector(
              V, IntZeroV, LLVM::Value::getConstVector32(LLContext, Mask));
        } else {
          V = Builder.createShuffleVector(
              IntZeroV, V, LLVM::Value::getConstVector32(LLContext, Mask));
        }
      }
      return V;
    });
  }
  void compileVectorConvertS(LLVM::Type VectorTy, LLVM::Type FPVectorTy,
                             bool Low) noexcept {
    compileVectorOp(VectorTy,
                    [this, VectorTy, FPVectorTy, Low](auto V) noexcept {
                      if (Low) {
                        const auto Size = VectorTy.getVectorSize() / 2;
                        std::vector<uint32_t> Mask(Size);
                        if constexpr (Endian::native == Endian::little) {
                          std::iota(Mask.begin(), Mask.end(), 0);
                        } else {
                          std::iota(Mask.begin(), Mask.end(), Size);
                        }
                        V = Builder.createShuffleVector(
                            V, LLVM::Value::getUndef(VectorTy),
                            LLVM::Value::getConstVector32(LLContext, Mask));
                      }
                      return Builder.createSIToFP(V, FPVectorTy);
                    });
  }
  void compileVectorConvertU(LLVM::Type VectorTy, LLVM::Type FPVectorTy,
                             bool Low) noexcept {
    compileVectorOp(VectorTy,
                    [this, VectorTy, FPVectorTy, Low](auto V) noexcept {
                      if (Low) {
                        const auto Size = VectorTy.getVectorSize() / 2;
                        std::vector<uint32_t> Mask(Size);
                        if constexpr (Endian::native == Endian::little) {
                          std::iota(Mask.begin(), Mask.end(), 0);
                        } else {
                          std::iota(Mask.begin(), Mask.end(), Size);
                        }
                        V = Builder.createShuffleVector(
                            V, LLVM::Value::getUndef(VectorTy),
                            LLVM::Value::getConstVector32(LLContext, Mask));
                      }
                      return Builder.createUIToFP(V, FPVectorTy);
                    });
  }
  void compileVectorDemote() noexcept {
    compileVectorOp(Context.Doublex2Ty, [this](auto V) noexcept {
      auto Demoted = Builder.createFPTrunc(
          V, LLVM::Type::getVectorType(Context.FloatTy, 2));
      auto ZeroV = LLVM::Value::getConstNull(Demoted.getType());
      if constexpr (Endian::native == Endian::little) {
        return Builder.createShuffleVector(
            Demoted, ZeroV,
            LLVM::Value::getConstVector32(LLContext, {0u, 1u, 2u, 3u}));
      } else {
        return Builder.createShuffleVector(
            Demoted, ZeroV,
            LLVM::Value::getConstVector32(LLContext, {3u, 2u, 1u, 0u}));
      }
    });
  }
  void compileVectorPromote() noexcept {
    compileVectorOp(Context.Floatx4Ty, [this](auto V) noexcept {
      auto UndefV = LLVM::Value::getUndef(V.getType());
      auto Low = Builder.createShuffleVector(
          V, UndefV, LLVM::Value::getConstVector32(LLContext, {0u, 1u}));
      return Builder.createFPExt(
          Low, LLVM::Type::getVectorType(Context.DoubleTy, 2));
    });
  }

  void compileVectorVectorMAdd(LLVM::Type VectorTy) noexcept {
    auto C = Builder.createBitCast(stackPop(), VectorTy);
    auto RHS = Builder.createBitCast(stackPop(), VectorTy);
    auto LHS = Builder.createBitCast(stackPop(), VectorTy);
    stackPush(Builder.createBitCast(
        Builder.createFAdd(Builder.createFMul(LHS, RHS), C),
        Context.Int64x2Ty));
  }

  void compileVectorVectorNMAdd(LLVM::Type VectorTy) noexcept {
    auto C = Builder.createBitCast(stackPop(), VectorTy);
    auto RHS = Builder.createBitCast(stackPop(), VectorTy);
    auto LHS = Builder.createBitCast(stackPop(), VectorTy);
    stackPush(Builder.createBitCast(
        Builder.createFAdd(Builder.createFMul(Builder.createFNeg(LHS), RHS), C),
        Context.Int64x2Ty));
  }

  void compileVectorRelaxedIntegerDotProduct() noexcept {
    auto OriTy = Context.Int8x16Ty;
    auto ExtTy = Context.Int16x8Ty;
    auto RHS = Builder.createBitCast(stackPop(), OriTy);
    auto LHS = Builder.createBitCast(stackPop(), OriTy);
#if defined(__x86_64__)
    if (Context.SupportSSSE3) {
      assuming(LLVM::Core::X86SSSE3PMAddUbSw128 != LLVM::Core::NotIntrinsic);
      // WebAssembly Relaxed SIMD spec: signed(LHS) * unsigned/signed(RHS)
      // But PMAddUbSw128 is unsigned(LHS) * signed(RHS). Therefore swap both
      // side to match the WebAssembly spec
      return stackPush(Builder.createBitCast(
          Builder.createIntrinsic(LLVM::Core::X86SSSE3PMAddUbSw128, {},
                                  {RHS, LHS}),
          Context.Int64x2Ty));
    }
#endif
    auto Width = LLVM::Value::getConstInt(
        ExtTy.getElementType(), OriTy.getElementType().getIntegerBitWidth());
    Width = Builder.createVectorSplat(ExtTy.getVectorSize(), Width);
    auto EA = Builder.createBitCast(LHS, ExtTy);
    auto EB = Builder.createBitCast(RHS, ExtTy);

    LLVM::Value AL, AR, BL, BR;
    AL = Builder.createAShr(EA, Width);
    AR = Builder.createAShr(Builder.createShl(EA, Width), Width);
    BL = Builder.createAShr(EB, Width);
    BR = Builder.createAShr(Builder.createShl(EB, Width), Width);

    return stackPush(Builder.createBitCast(
        Builder.createAdd(Builder.createMul(AL, BL), Builder.createMul(AR, BR)),
        Context.Int64x2Ty));
  }

  void compileVectorRelaxedIntegerDotProductAdd() noexcept {
    auto OriTy = Context.Int8x16Ty;
    auto ExtTy = Context.Int16x8Ty;
    auto FinTy = Context.Int32x4Ty;
    auto VC = Builder.createBitCast(stackPop(), FinTy);
    auto RHS = Builder.createBitCast(stackPop(), OriTy);
    auto LHS = Builder.createBitCast(stackPop(), OriTy);
    LLVM::Value IM;
#if defined(__x86_64__)
    if (Context.SupportSSSE3) {
      assuming(LLVM::Core::X86SSSE3PMAddUbSw128 != LLVM::Core::NotIntrinsic);
      // WebAssembly Relaxed SIMD spec: signed(LHS) * unsigned/signed(RHS)
      // But PMAddUbSw128 is unsigned(LHS) * signed(RHS). Therefore swap both
      // side to match the WebAssembly spec
      IM = Builder.createIntrinsic(LLVM::Core::X86SSSE3PMAddUbSw128, {},
                                   {RHS, LHS});
    } else
#endif
    {
      auto Width = LLVM::Value::getConstInt(
          ExtTy.getElementType(), OriTy.getElementType().getIntegerBitWidth());
      Width = Builder.createVectorSplat(ExtTy.getVectorSize(), Width);
      auto EA = Builder.createBitCast(LHS, ExtTy);
      auto EB = Builder.createBitCast(RHS, ExtTy);

      LLVM::Value AL, AR, BL, BR;
      AL = Builder.createAShr(EA, Width);
      AR = Builder.createAShr(Builder.createShl(EA, Width), Width);
      BL = Builder.createAShr(EB, Width);
      BR = Builder.createAShr(Builder.createShl(EB, Width), Width);
      IM = Builder.createAdd(Builder.createMul(AL, BL),
                             Builder.createMul(AR, BR));
    }

    auto Width = LLVM::Value::getConstInt(
        FinTy.getElementType(), ExtTy.getElementType().getIntegerBitWidth());
    Width = Builder.createVectorSplat(FinTy.getVectorSize(), Width);
    auto IME = Builder.createBitCast(IM, FinTy);
    auto L = Builder.createAShr(IME, Width);
    auto R = Builder.createAShr(Builder.createShl(IME, Width), Width);

    return stackPush(Builder.createBitCast(
        Builder.createAdd(Builder.createAdd(L, R), VC), Context.Int64x2Ty));
  }

  void
  enterBlock(LLVM::BasicBlock JumpBlock, LLVM::BasicBlock NextBlock,
             LLVM::BasicBlock ElseBlock, std::vector<LLVM::Value> Args,
             std::pair<std::vector<ValType>, std::vector<ValType>> Type,
             std::vector<std::tuple<std::vector<LLVM::Value>, LLVM::BasicBlock>>
                 ReturnPHI = {}) noexcept {
    assuming(Type.first.size() == Args.size());
    for (auto &Value : Args) {
      stackPush(Value);
    }
    const auto Unreachable = isUnreachable();
    ControlStack.emplace_back(Stack.size() - Args.size(), Unreachable,
                              JumpBlock, NextBlock, ElseBlock, std::move(Args),
                              std::move(Type), std::move(ReturnPHI));
  }

  Control leaveBlock() noexcept {
    Control Entry = std::move(ControlStack.back());
    ControlStack.pop_back();

    auto NextBlock = Entry.NextBlock ? Entry.NextBlock : Entry.JumpBlock;
    if (!Entry.Unreachable) {
      const auto &ReturnType = Entry.Type.second;
      if (!ReturnType.empty()) {
        std::vector<LLVM::Value> Rets(ReturnType.size());
        for (size_t I = 0; I < Rets.size(); ++I) {
          const size_t J = Rets.size() - 1 - I;
          Rets[J] = stackPop();
        }
        Entry.ReturnPHI.emplace_back(std::move(Rets), Builder.getInsertBlock());
      }
      Builder.createBr(NextBlock);
    } else {
      Builder.createUnreachable();
    }
    Builder.positionAtEnd(NextBlock);
    Stack.erase(Stack.begin() + static_cast<int64_t>(Entry.StackSize),
                Stack.end());
    return Entry;
  }

  void checkStop() noexcept {
    if (!Interruptible) {
      return;
    }
    auto NotStopBB = LLVM::BasicBlock::create(LLContext, F.Fn, "NotStop");
    auto StopToken = Builder.createAtomicRMW(
        LLVMAtomicRMWBinOpXchg, Context.getStopToken(Builder, ExecCtx),
        LLContext.getInt32(0), LLVMAtomicOrderingMonotonic);
#if LLVM_VERSION_MAJOR >= 13
    StopToken.setAlignment(32);
#endif
    auto NotStop = Builder.createLikely(
        Builder.createICmpEQ(StopToken, LLContext.getInt32(0)));
    Builder.createCondBr(NotStop, NotStopBB,
                         getTrapBB(ErrCode::Value::Interrupted));

    Builder.positionAtEnd(NotStopBB);
  }

  void setUnreachable() noexcept {
    if (ControlStack.empty()) {
      IsUnreachable = true;
    } else {
      ControlStack.back().Unreachable = true;
    }
  }

  bool isUnreachable() const noexcept {
    if (ControlStack.empty()) {
      return IsUnreachable;
    } else {
      return ControlStack.back().Unreachable;
    }
  }

  void
  buildPHI(Span<const ValType> RetType,
           Span<const std::tuple<std::vector<LLVM::Value>, LLVM::BasicBlock>>
               Incomings) noexcept {
    if (isVoidReturn(RetType)) {
      return;
    }
    std::vector<LLVM::Value> Nodes;
    if (Incomings.size() == 0) {
      const auto &Types = toLLVMTypeVector(LLContext, RetType);
      Nodes.reserve(Types.size());
      for (LLVM::Type Type : Types) {
        Nodes.push_back(LLVM::Value::getUndef(Type));
      }
    } else if (Incomings.size() == 1) {
      Nodes = std::move(std::get<0>(Incomings.front()));
    } else {
      const auto &Types = toLLVMTypeVector(LLContext, RetType);
      Nodes.reserve(Types.size());
      for (size_t I = 0; I < Types.size(); ++I) {
        auto PHIRet = Builder.createPHI(Types[I]);
        for (auto &[Value, BB] : Incomings) {
          assuming(Value.size() == Types.size());
          PHIRet.addIncoming(Value[I], BB);
        }
        Nodes.push_back(PHIRet);
      }
    }
    for (auto &Val : Nodes) {
      stackPush(Val);
    }
  }

  void setLableJumpPHI(unsigned int Index) noexcept {
    assuming(Index < ControlStack.size());
    auto &Entry = *(ControlStack.rbegin() + Index);
    if (Entry.NextBlock) { // is loop
      std::vector<LLVM::Value> Args(Entry.Type.first.size());
      for (size_t I = 0; I < Args.size(); ++I) {
        const size_t J = Args.size() - 1 - I;
        Args[J] = stackPop();
      }
      for (size_t I = 0; I < Args.size(); ++I) {
        Entry.Args[I].addIncoming(Args[I], Builder.getInsertBlock());
        stackPush(Args[I]);
      }
    } else if (!Entry.Type.second.empty()) { // has return value
      std::vector<LLVM::Value> Rets(Entry.Type.second.size());
      for (size_t I = 0; I < Rets.size(); ++I) {
        const size_t J = Rets.size() - 1 - I;
        Rets[J] = stackPop();
      }
      for (size_t I = 0; I < Rets.size(); ++I) {
        stackPush(Rets[I]);
      }
      Entry.ReturnPHI.emplace_back(std::move(Rets), Builder.getInsertBlock());
    }
  }

  LLVM::BasicBlock getLabel(unsigned int Index) const noexcept {
    return (ControlStack.rbegin() + Index)->JumpBlock;
  }

  void stackPush(LLVM::Value Value) noexcept { Stack.push_back(Value); }
  LLVM::Value stackPop() noexcept {
    assuming(!ControlStack.empty() || !Stack.empty());
    assuming(ControlStack.empty() ||
             Stack.size() > ControlStack.back().StackSize);
    auto Value = Stack.back();
    Stack.pop_back();
    return Value;
  }

  LLVM::Value switchEndian(LLVM::Value Value) {
    if constexpr (Endian::native == Endian::big) {
      auto Type = Value.getType();
      if ((Type.isIntegerTy() && Type.getIntegerBitWidth() > 8) ||
          (Type.isVectorTy() && Type.getVectorSize() == 1)) {
        return Builder.createUnaryIntrinsic(LLVM::Core::Bswap, Value);
      }
      if (Type.isVectorTy()) {
        LLVM::Type VecType = Type.getElementType().getIntegerBitWidth() == 128
                                 ? Context.Int128Ty
                                 : Context.Int64Ty;
        Value = Builder.createBitCast(Value, VecType);
        Value = Builder.createUnaryIntrinsic(LLVM::Core::Bswap, Value);
        return Builder.createBitCast(Value, Type);
      }
      if (Type.isFloatTy() || Type.isDoubleTy()) {
        LLVM::Type IntType =
            Type.isFloatTy() ? Context.Int32Ty : Context.Int64Ty;
        Value = Builder.createBitCast(Value, IntType);
        Value = Builder.createUnaryIntrinsic(LLVM::Core::Bswap, Value);
        return Builder.createBitCast(Value, Type);
      }
    }
    return Value;
  }

  LLVM::Compiler::CompileContext &Context;
  LLVM::Context LLContext;
  std::vector<std::pair<LLVM::Type, LLVM::Value>> Local;
  std::vector<LLVM::Value> Stack;
  LLVM::Value LocalInstrCount = nullptr;
  LLVM::Value LocalGas = nullptr;
  std::unordered_map<ErrCode::Value, LLVM::BasicBlock> TrapBB;
  bool IsUnreachable = false;
  bool Interruptible = false;
  struct Control {
    size_t StackSize;
    bool Unreachable;
    LLVM::BasicBlock JumpBlock;
    LLVM::BasicBlock NextBlock;
    LLVM::BasicBlock ElseBlock;
    std::vector<LLVM::Value> Args;
    std::pair<std::vector<ValType>, std::vector<ValType>> Type;
    std::vector<std::tuple<std::vector<LLVM::Value>, LLVM::BasicBlock>>
        ReturnPHI;
    Control(size_t S, bool U, LLVM::BasicBlock J, LLVM::BasicBlock N,
            LLVM::BasicBlock E, std::vector<LLVM::Value> A,
            std::pair<std::vector<ValType>, std::vector<ValType>> T,
            std::vector<std::tuple<std::vector<LLVM::Value>, LLVM::BasicBlock>>
                R) noexcept
        : StackSize(S), Unreachable(U), JumpBlock(J), NextBlock(N),
          ElseBlock(E), Args(std::move(A)), Type(std::move(T)),
          ReturnPHI(std::move(R)) {}
    Control(const Control &) = default;
    Control(Control &&) = default;
    Control &operator=(const Control &) = default;
    Control &operator=(Control &&) = default;
  };
  bool IsLazyJIT;
  std::vector<Control> ControlStack;
  LLVM::FunctionCallee F;
  LLVM::Value ExecCtx;
  LLVM::Builder Builder;
};

std::vector<LLVM::Value> unpackStruct(LLVM::Builder &Builder,
                                      LLVM::Value Struct) noexcept {
  const auto N = Struct.getType().getStructNumElements();
  std::vector<LLVM::Value> Ret;
  Ret.reserve(N);
  for (unsigned I = 0; I < N; ++I) {
    Ret.push_back(Builder.createExtractValue(Struct, I));
  }
  return Ret;
}

} // namespace

namespace WasmEdge {
namespace LLVM {

Expect<void> Compiler::checkConfigure() noexcept {
  // Note: Although the Exception Handling and Memory64 proposals are not
  // implemented in AOT yet, we should not trap here because the default
  // configuration has become WASM 3.0, which contains these proposals.
  if (Conf.hasProposal(Proposal::ExceptionHandling)) {
    spdlog::warn("Proposal Exception Handling is not yet supported in WasmEdge "
                 "AOT/JIT. The compilation will be trapped when related data "
                 "structure or instructions found in WASM.");
  }
  if (Conf.hasProposal(Proposal::Annotations)) {
    spdlog::error(ErrCode::Value::InvalidAOTConfigure);
    spdlog::error("    Proposal Custom Annotation Syntax is not yet supported "
                  "in WasmEdge AOT/JIT.");
    return Unexpect(ErrCode::Value::InvalidAOTConfigure);
  }
  return {};
}

Expect<void> Compiler::optimize(LLVM::Module &LLModule,
                                LLVM::TargetMachine &TM) noexcept {
  spdlog::info("optimize start"sv);
  auto Triple = LLModule.getTarget();
  auto [TheTarget, ErrorMessage] = LLVM::Target::getFromTriple(Triple);
  if (ErrorMessage) {
    spdlog::error("getFromTriple failed:{}"sv, ErrorMessage.string_view());
    return Unexpect(ErrCode::Value::IllegalPath);
  }

  std::string CPUName;
#if defined(__riscv) && __riscv_xlen == 64
  CPUName = "generic-rv64"s;
#else
  if (!Conf.getCompilerConfigure().isGenericBinary()) {
    CPUName = LLVM::getHostCPUName().string_view();
  } else {
    CPUName = "generic"s;
  }
#endif

  // On RISC-V we use generic-rv64 as the CPU, so also use default
  // features; host features under QEMU can be inconsistent (e.g.
  // zvl*b without v) which LLVM >= 20 rejects.
  TM = LLVM::TargetMachine::create(
      TheTarget, Triple, CPUName.c_str(),
#if defined(__riscv) && __riscv_xlen == 64
      "",
#else
      LLVM::getHostCPUFeatures().unwrap(),
#endif
      toLLVMCodeGenLevel(Conf.getCompilerConfigure().getOptimizationLevel()),
      LLVMRelocPIC, LLVMCodeModelDefault);

#if LLVM_VERSION_MAJOR >= 13
  auto PBO = LLVM::PassBuilderOptions::create();
  if (auto Error = PBO.runPasses(
          LLModule,
          toLLVMLevel(Conf.getCompilerConfigure().getOptimizationLevel()),
          TM)) {
    spdlog::error("{}"sv, Error.message().string_view());
  }
#else
  auto FP = LLVM::PassManager::createForModule(LLModule);
  auto MP = LLVM::PassManager::create();

  TM.addAnalysisPasses(MP);
  TM.addAnalysisPasses(FP);
  {
    auto PMB = LLVM::PassManagerBuilder::create();
    auto [OptLevel, SizeLevel] =
        toLLVMLevel(Conf.getCompilerConfigure().getOptimizationLevel());
    PMB.setOptLevel(OptLevel);
    PMB.setSizeLevel(SizeLevel);
    PMB.populateFunctionPassManager(FP);
    PMB.populateModulePassManager(MP);
  }
  switch (Conf.getCompilerConfigure().getOptimizationLevel()) {
  case CompilerConfigure::OptimizationLevel::O0:
  case CompilerConfigure::OptimizationLevel::O1:
    FP.addTailCallEliminationPass();
    break;
  default:
    break;
  }

  FP.initializeFunctionPassManager();
  for (auto Fn = LLModule.getFirstFunction(); Fn; Fn = Fn.getNextFunction()) {
    FP.runFunctionPassManager(Fn);
  }
  FP.finalizeFunctionPassManager();
  MP.runPassManager(LLModule);
#endif

  spdlog::info("optimize done"sv);
  return {};
}

// Initialize the LLVM module held by the data for compilation: set the
// target triple and the PIC level, and return the LLVM context.
static LLVM::Context initLLVMModule(LLVM::Data &D) noexcept {
  auto LLContext = D.extract().getLLContext();
  LLVM::Core::init(LLContext.unwrap());
  auto &LLModule = D.extract().LLModule;
  LLModule.setTarget(LLVM::getDefaultTargetTriple().unwrap());
  LLModule.addFlag(LLVMModuleFlagBehaviorError, "PIC Level"sv, 2);
  return LLContext;
}

Expect<Data> Compiler::compile(const AST::Module &Module) noexcept {
  // Check that the module is validated.
  if (unlikely(!Module.getIsValidated())) {
    spdlog::error(ErrCode::Value::NotValidated);
    return Unexpect(ErrCode::Value::NotValidated);
  }

  std::unique_lock Lock(Mutex);
  spdlog::info("compile start"sv);

  LLVM::Data D;
  auto LLContext = initLLVMModule(D);
  auto &LLModule = D.extract().LLModule;

  CompileContext NewContext(LLContext, LLModule,
                            Conf.getCompilerConfigure().isGenericBinary());
  RAIICleanup Cleanup(Context, &NewContext);
  Context->addVersionGlobal();

  // Compile all sections and the function declarations.
  compileSections(Module, false);
  // Compile all function bodies.
  const auto DefinedCount = Module.getDefinedFuncCount();
  for (uint32_t I = 0; I < DefinedCount; ++I) {
    EXPECTED_TRY(compileFunctionBody(I));
  }
  // Compile ExportSection.
  compile(Module.getExportSection());
  // StartSection is not required for compilation.

  spdlog::info("verify start"sv);
  LLModule.verify(LLVMPrintMessageAction);

  auto &TM = D.extract().TM;
  EXPECTED_TRY(optimize(LLModule, TM));

  // Set initializer for constant value
  Context->finalizeIntrinsicsTable();
  return Expect<Data>{std::move(D)};
}

void Compiler::compile(const AST::TypeSection &TypeSec,
                       bool DeclarationsOnly) noexcept {
  auto WrapperTy =
      LLVM::Type::getFunctionType(Context->VoidTy,
                                  {Context->ExecCtxPtrTy, Context->Int8PtrTy,
                                   Context->Int8PtrTy, Context->Int8PtrTy},
                                  false);
  auto SubTypes = TypeSec.getContent();
  const auto Size = SubTypes.size();
  if (Size == 0) {
    return;
  }
  Context->CompositeTypes.reserve(Size);
  Context->FunctionWrappers.reserve(Size);

  auto SetFuncAttributes = [&](auto FDecl) {
    FDecl.setVisibility(LLVMProtectedVisibility);
    FDecl.setDSOLocal(true);
    FDecl.setDLLStorageClass(LLVMDLLExportStorageClass);
    FDecl.addFnAttr(Context->NoStackArgProbe);
    FDecl.addFnAttr(Context->StrictFP);
    FDecl.addFnAttr(Context->UWTable);
    FDecl.addParamAttr(0, Context->ReadOnly);
    FDecl.addParamAttr(0, Context->NoAlias);
    FDecl.addParamAttr(1, Context->NoAlias);
    FDecl.addParamAttr(2, Context->NoAlias);
    FDecl.addParamAttr(3, Context->NoAlias);
  };

  // Iterate and compile types.
  for (size_t I = 0; I < Size; ++I) {
    const auto &CompType = SubTypes[I].getCompositeType();
    const auto Name = fmt::format("t{}"sv, Context->CompositeTypes.size());
    if (CompType.isFunc()) {
      // Check that the function type is unique.
      {
        bool Unique = true;
        for (size_t J = 0; J < I; ++J) {
          if (Context->CompositeTypes[J] &&
              Context->CompositeTypes[J]->isFunc()) {
            const auto &OldFuncType = Context->CompositeTypes[J]->getFuncType();
            if (OldFuncType == CompType.getFuncType()) {
              Unique = false;
              Context->CompositeTypes.push_back(Context->CompositeTypes[J]);
              if (DeclarationsOnly) {
                auto FDecl = Context->LLModule.get().addFunction(
                    WrapperTy, LLVMExternalLinkage, Name.c_str());
                SetFuncAttributes(FDecl);
                Context->FunctionWrappers.push_back(FDecl);
              } else {
                auto F = Context->FunctionWrappers[J];
                Context->FunctionWrappers.push_back(F);
                auto A = Context->LLModule.get().addAlias(WrapperTy, F,
                                                          Name.c_str());
                A.setLinkage(LLVMExternalLinkage);
                A.setVisibility(LLVMProtectedVisibility);
                A.setDSOLocal(true);
                A.setDLLStorageClass(LLVMDLLExportStorageClass);
              }
              break;
            }
          }
        }
        if (!Unique) {
          continue;
        }
      }

      // Create Wrapper
      auto F = Context->LLModule.get().addFunction(
          WrapperTy, LLVMExternalLinkage, Name.c_str());
      {
        SetFuncAttributes(F);

        if (!DeclarationsOnly) {
          LLVM::Builder Builder(Context->LLContext);
          Builder.positionAtEnd(
              LLVM::BasicBlock::create(Context->LLContext, F, "entry"));

          auto FTy = toLLVMType(Context->LLContext, Context->ExecCtxPtrTy,
                                CompType.getFuncType());
          auto RTy = FTy.getReturnType();
          std::vector<LLVM::Type> FPTy(FTy.getNumParams());
          FTy.getParamTypes(FPTy);

          const size_t ArgCount = FPTy.size() - 1;
          auto ExecCtxPtr = F.getFirstParam();
          auto RawFunc = LLVM::FunctionCallee{
              FTy, Builder.createBitCast(ExecCtxPtr.getNextParam(),
                                         FTy.getPointerTo())};
          auto RawArgs = ExecCtxPtr.getNextParam().getNextParam();
          auto RawRets = RawArgs.getNextParam();

          std::vector<LLVM::Value> Args;
          Args.reserve(FTy.getNumParams());
          Args.push_back(ExecCtxPtr);
          for (size_t J = 0; J < ArgCount; ++J) {
            Args.push_back(Builder.createValuePtrLoad(
                FPTy[J + 1], RawArgs, Context->Int8Ty, J * kValSize));
          }

          auto Ret = Builder.createCall(RawFunc, Args);
          if (RTy.isVoidTy()) {
            // nothing to do
          } else if (RTy.isStructTy()) {
            auto Rets = unpackStruct(Builder, Ret);
            Builder.createArrayPtrStore(Rets, RawRets, Context->Int8Ty,
                                        kValSize);
          } else {
            Builder.createValuePtrStore(Ret, RawRets, Context->Int8Ty);
          }
          Builder.createRetVoid();
        }
      }
      // Copy wrapper, param and return lists to module instance.
      Context->FunctionWrappers.push_back(F);
    } else {
      // Non function type case. Create empty wrapper.
      auto F = Context->LLModule.get().addFunction(
          WrapperTy, LLVMExternalLinkage, Name.c_str());
      {
        SetFuncAttributes(F);

        if (!DeclarationsOnly) {
          LLVM::Builder Builder(Context->LLContext);
          Builder.positionAtEnd(
              LLVM::BasicBlock::create(Context->LLContext, F, "entry"));
          Builder.createRetVoid();
        }
      }
      Context->FunctionWrappers.push_back(F);
    }
    Context->CompositeTypes.push_back(&CompType);
  }
}

void Compiler::compile(const AST::ImportSection &ImportSec) noexcept {
  // Iterate and compile import descriptions.
  for (const auto &ImpDesc : ImportSec.getContent()) {
    // Get data from import description.
    const auto &ExtType = ImpDesc.getExternalType();

    // Add the imports to the module instance.
    switch (ExtType) {
    case ExternalType::Function: // Function type index
    {
      const auto FuncID = static_cast<uint32_t>(Context->Functions.size());
      // Get the function type index in module.
      uint32_t TypeIdx = ImpDesc.getExternalFuncTypeIdx();
      assuming(TypeIdx < Context->CompositeTypes.size());
      assuming(Context->CompositeTypes[TypeIdx]->isFunc());
      const auto &FuncType = Context->CompositeTypes[TypeIdx]->getFuncType();
      auto FTy =
          toLLVMType(Context->LLContext, Context->ExecCtxPtrTy, FuncType);
      auto RTy = FTy.getReturnType();
      auto F =
          LLVM::FunctionCallee{FTy, Context->LLModule.get().addFunction(
                                        FTy, LLVMInternalLinkage,
                                        fmt::format("f{}"sv, FuncID).c_str())};
      F.Fn.setDSOLocal(true);
      F.Fn.addFnAttr(Context->NoStackArgProbe);
      F.Fn.addFnAttr(Context->StrictFP);
      F.Fn.addFnAttr(Context->UWTable);
      F.Fn.addParamAttr(0, Context->ReadOnly);
      F.Fn.addParamAttr(0, Context->NoAlias);

      LLVM::Builder Builder(Context->LLContext);
      Builder.positionAtEnd(
          LLVM::BasicBlock::create(Context->LLContext, F.Fn, "entry"));

      const auto ArgSize = FuncType.getParamTypes().size();
      const auto RetSize =
          RTy.isVoidTy() ? 0 : FuncType.getReturnTypes().size();

      LLVM::Value Args = Builder.createArray(ArgSize, kValSize);
      LLVM::Value Rets = Builder.createArray(RetSize, kValSize);

      auto Arg = F.Fn.getFirstParam();
      for (unsigned I = 0; I < ArgSize; ++I) {
        Arg = Arg.getNextParam();
        Builder.createValuePtrStore(Arg, Args, Context->Int8Ty, I * kValSize);
      }

      Builder.createCall(
          Context->getIntrinsic(
              Builder, Executable::Intrinsics::kCall,
              LLVM::Type::getFunctionType(
                  Context->VoidTy,
                  {Context->Int32Ty, Context->Int8PtrTy, Context->Int8PtrTy},
                  false)),
          {Context->LLContext.getInt32(FuncID), Args, Rets});

      if (RetSize == 0) {
        Builder.createRetVoid();
      } else if (RetSize == 1) {
        Builder.createRet(
            Builder.createValuePtrLoad(RTy, Rets, Context->Int8Ty));
      } else {
        Builder.createAggregateRet(Builder.createArrayPtrLoad(
            RetSize, RTy, Rets, Context->Int8Ty, kValSize));
      }

      Context->Functions.emplace_back(TypeIdx, F, nullptr);
      Context->ImportCount++;
      break;
    }
    case ExternalType::Table: // Table type
    {
      // Get table address type. External type checked in validation.
      const auto &TabType = ImpDesc.getExternalTableType();
      const auto AddrType = TabType.getLimit().getAddrType();
      auto Type = toLLVMType(Context->LLContext, AddrType);
      Context->TableAddrTypes.push_back(Type);
      break;
    }
    case ExternalType::Memory: // Memory type
    {
      // Get memory address type. External type checked in validation.
      const auto &MemType = ImpDesc.getExternalMemoryType();
      const auto AddrType = MemType.getLimit().getAddrType();
      auto Type = toLLVMType(Context->LLContext, AddrType);
      Context->MemoryAddrTypes.push_back(Type);
      break;
    }
    case ExternalType::Global: // Global type
    {
      // Get global type. External type checked in validation.
      const auto &GlobType = ImpDesc.getExternalGlobalType();
      const auto &ValType = GlobType.getValType();
      auto Type = toLLVMType(Context->LLContext, ValType);
      Context->Globals.push_back(Type);
      break;
    }
    case ExternalType::Tag: // Tag type
    {
      // TODO: EXCEPTION - implement the AOT.
      break;
    }
    default:
      assumingUnreachable();
    }
  }
}

void Compiler::compile(const AST::ExportSection &) noexcept {}

void Compiler::compile(const AST::GlobalSection &GlobalSec) noexcept {
  for (const auto &GlobalSeg : GlobalSec.getContent()) {
    const auto &ValType = GlobalSeg.getGlobalType().getValType();
    auto Type = toLLVMType(Context->LLContext, ValType);
    Context->Globals.push_back(Type);
  }
}

void Compiler::compile(const AST::MemorySection &MemorySec,
                       const AST::DataSection &) noexcept {
  for (const auto &MemType : MemorySec.getContent()) {
    const auto AddrType = MemType.getLimit().getAddrType();
    auto Type = toLLVMType(Context->LLContext, AddrType);
    Context->MemoryAddrTypes.push_back(Type);
  }
}

void Compiler::compile(const AST::TableSection &TableSec,
                       const AST::ElementSection &) noexcept {
  for (const auto &TableSeg : TableSec.getContent()) {
    const auto AddrType = TableSeg.getTableType().getLimit().getAddrType();
    auto Type = toLLVMType(Context->LLContext, AddrType);
    Context->TableAddrTypes.push_back(Type);
  }
}

void Compiler::compileSections(const AST::Module &Module,
                               bool DeclarationsOnly) noexcept {
  // Compile Function Types
  compile(Module.getTypeSection(), DeclarationsOnly);
  // Compile ImportSection
  compile(Module.getImportSection());
  // Compile GlobalSection
  compile(Module.getGlobalSection());
  // Compile MemorySection (MemorySec, DataSec)
  compile(Module.getMemorySection(), Module.getDataSection());
  // Compile TableSection (TableSec, ElemSec)
  compile(Module.getTableSection(), Module.getElementSection());
  // Create function declarations without compiling bodies. (FunctionSec,
  // CodeSec)
  compileFunctionDeclarations(Module.getFunctionSection(),
                              Module.getCodeSection());
}

void Compiler::compileFunctionDeclarations(
    const AST::FunctionSection &FunctionSec,
    const AST::CodeSection &CodeSec) noexcept {
  const auto &TypeIdxs = FunctionSec.getContent();
  const auto &CodeSegs = CodeSec.getContent();
  assuming(TypeIdxs.size() == CodeSegs.size());

  for (size_t I = 0; I < CodeSegs.size(); ++I) {
    const auto &TypeIdx = TypeIdxs[I];
    const auto &Code = CodeSegs[I];
    assuming(TypeIdx < Context->CompositeTypes.size());
    assuming(Context->CompositeTypes[TypeIdx]->isFunc());
    const auto &FuncType = Context->CompositeTypes[TypeIdx]->getFuncType();
    const auto FuncID = Context->Functions.size();
    auto FTy = toLLVMType(Context->LLContext, Context->ExecCtxPtrTy, FuncType);
    LLVM::FunctionCallee F = {FTy, Context->LLModule.get().addFunction(
                                       FTy, LLVMExternalLinkage,
                                       fmt::format("f{}"sv, FuncID).c_str())};
    F.Fn.setVisibility(LLVMProtectedVisibility);
    F.Fn.setDSOLocal(true);
    F.Fn.setDLLStorageClass(LLVMDLLExportStorageClass);
    F.Fn.addFnAttr(Context->NoStackArgProbe);
    F.Fn.addFnAttr(Context->StrictFP);
    F.Fn.addFnAttr(Context->UWTable);
    F.Fn.addParamAttr(0, Context->ReadOnly);
    F.Fn.addParamAttr(0, Context->NoAlias);

    Context->Functions.emplace_back(TypeIdx, F, &Code);
  }
}

Expect<void> Compiler::compileFunctionBody(uint32_t LocalFuncIndex) noexcept {
  // Find the function in the Functions list
  // LocalFuncIndex is relative to the defined functions (not imports)
  uint32_t GlobalFuncIndex = Context->ImportCount + LocalFuncIndex;
  if (GlobalFuncIndex >= Context->Functions.size()) {
    spdlog::error("[lazy-jit]: function index {} out of range"sv,
                  LocalFuncIndex);
    return Unexpect(ErrCode::Value::IllegalPath);
  }

  auto &[T, F, Code] = Context->Functions[GlobalFuncIndex];
  if (!Code) {
    spdlog::error("[lazy-jit]: cannot compile import function {}"sv,
                  LocalFuncIndex);
    return Unexpect(ErrCode::Value::IllegalPath);
  }

  // Check if already compiled (function has basic blocks)
  if (F.Fn.countBasicBlocks() > 0) {
    spdlog::debug("[lazy-jit]: function {} already compiled"sv, LocalFuncIndex);
    return {};
  }

  spdlog::debug("[lazy-jit]: compiling function {}"sv, LocalFuncIndex);

  std::vector<ValType> Locals;
  for (const auto &Local : Code->getLocals()) {
    for (unsigned I = 0; I < Local.first; ++I) {
      Locals.push_back(Local.second);
    }
  }

  FunctionCompiler FC(
      *Context, F, Locals, Conf.getCompilerConfigure().isInterruptible(),
      Conf.getStatisticsConfigure().isInstructionCounting(),
      Conf.getStatisticsConfigure().isCostMeasuring(),
      Conf.getRuntimeConfigure().getRunMode() == RunMode::LazyJIT);
  auto Type = Context->resolveBlockType(T);
  EXPECTED_TRY(FC.compile(*Code, std::move(Type)));
  F.Fn.eliminateUnreachableBlocks();

  return {};
}

Expect<LLVM::Data>
LLVM::Compiler::compileInfrastructure(const AST::Module &Module) noexcept {
  // Check the module is validated.
  if (unlikely(!Module.getIsValidated())) {
    spdlog::error(ErrCode::Value::NotValidated);
    return Unexpect(ErrCode::Value::NotValidated);
  }

  std::unique_lock Lock(Mutex);
  spdlog::info("[lazy-jit]: compile infrastructure start"sv);

  Data D;
  auto LLContext = initLLVMModule(D);
  auto &LLModule = D.extract().LLModule;

  CompileContext NewContext(LLContext, LLModule,
                            Conf.getCompilerConfigure().isGenericBinary());
  RAIICleanup Cleanup(Context, &NewContext);
  Context->addVersionGlobal();

  // Compile all sections and the function declarations without bodies.
  compileSections(Module, false);
  // Compile ExportSection
  compile(Module.getExportSection());

  // Set initializer for constant value
  Context->finalizeIntrinsicsTable();
  LLModule.verify(LLVMPrintMessageAction);

  spdlog::info("[lazy-jit]: infrastructure compilation done"sv);

  return Expect<Data>{std::move(D)};
}

Expect<LLVM::Data>
Compiler::compileFunctions(Data &&LLData, const AST::Module &Module,
                           Span<const uint32_t> LocalFuncIndices) noexcept {
  if (unlikely(!Module.getIsValidated())) {
    spdlog::error(ErrCode::Value::NotValidated);
    return Unexpect(ErrCode::Value::NotValidated);
  }
  if (unlikely(LocalFuncIndices.empty())) {
    spdlog::error("[lazy-jit]: compileFunctions with empty index list"sv);
    return Unexpect(ErrCode::Value::IllegalPath);
  }

  std::unique_lock Lock(Mutex);
  std::vector<uint32_t> Sorted(LocalFuncIndices.begin(),
                               LocalFuncIndices.end());
  std::sort(Sorted.begin(), Sorted.end());
  Sorted.erase(std::unique(Sorted.begin(), Sorted.end()), Sorted.end());

  spdlog::debug("[lazy-jit]: compile functions batch ({}) start"sv,
                Sorted.size());

  // Each batch starts from a fresh module sharing the same thread-safe
  // context: on success the previous batch module was consumed by the JIT,
  // and after a failed batch the leftover module must be discarded so its
  // declarations are not re-added on top of themselves.
  LLData.extract().resetModule();
  auto LLContext = initLLVMModule(LLData);
  auto &LLModule = LLData.extract().LLModule;

  CompileContext NewContext(LLContext, LLModule,
                            Conf.getCompilerConfigure().isGenericBinary());
  RAIICleanup Cleanup(Context, &NewContext);

  // Emit the type wrappers as external declarations resolved against the
  // infrastructure module, then declare the functions and compile the
  // requested bodies.
  compileSections(Module, true);

  for (uint32_t FuncIndex : Sorted) {
    EXPECTED_TRY(compileFunctionBody(FuncIndex));
  }

  spdlog::info("[lazy-jit]: verify batch ({} funcs) start"sv, Sorted.size());
  LLModule.verify(LLVMPrintMessageAction);
  spdlog::info("[lazy-jit]: verify batch ({} funcs) done"sv, Sorted.size());

  auto &TM = LLData.extract().TM;
  EXPECTED_TRY(optimize(LLModule, TM));

  spdlog::debug("[lazy-jit]: compile functions batch ({}) done"sv,
                Sorted.size());
  return Expect<Data>{std::move(LLData)};
}

} // namespace LLVM
} // namespace WasmEdge

Coverage Report

Created: 2026-06-30 06:10