FEX-Emu--FEX/Source/Windows/ARM64EC/Module.cpp

// SPDX-License-Identifier: MIT
/*
$info$
tags: Bin|ARM64EC
desc: Implements the ARM64EC BT module API using FEXCore
$end_info$
*/

#include <FEXCore/fextl/fmt.h>
#include <FEXCore/Core/X86Enums.h>
#include <FEXCore/Core/SignalDelegator.h>
#include <FEXCore/Core/Context.h>
#include <FEXCore/Core/CoreState.h>
#include <FEXCore/Debug/InternalThreadState.h>
#include <FEXCore/HLE/SyscallHandler.h>
#include <FEXCore/Config/Config.h>
#include <FEXCore/Utils/Allocator.h>
#include <FEXCore/Utils/LogManager.h>
#include <FEXCore/Utils/Threads.h>
#include <FEXCore/Utils/EnumOperators.h>
#include <FEXCore/Utils/EnumUtils.h>
#include <FEXCore/Utils/FPState.h>
#include <FEXCore/Utils/ArchHelpers/Arm64.h>
#include <FEXCore/Utils/MathUtils.h>
#include <FEXCore/Utils/TypeDefines.h>

#include "Common/Config.h"
#include "Common/InvalidationTracker.h"
#include "Common/TSOHandlerConfig.h"
#include "Common/CPUFeatures.h"
#include "Common/Logging.h"
#include "DummyHandlers.h"
#include "BTInterface.h"

#include <cstdint>
#include <cstdio>
#include <type_traits>
#include <mutex>
#include <optional>
#include <unordered_map>
#include <utility>
#include <ntstatus.h>
#include <windef.h>
#include <winternl.h>
#include <winnt.h>
#include <wine/debug.h>

class ECSyscallHandler;

extern "C" {
void* X64ReturnInstr; // See Module.S
extern void* ExitFunctionEC;

// Wine doesn't support issuing direct system calls with SVC, and unlike Windows it doesn't have a 'stable' syscall number for NtContinue
void* WineSyscallDispatcher;
// TODO: this really shouldn't be hardcoded, once wine gains proper syscall thunks this can be dropped.
uint64_t WineNtContinueSyscallId = 0x1a;
}

struct ThreadCPUArea {
  static constexpr size_t TEBCPUAreaOffset = 0x1788;
  CHPE_V2_CPU_AREA_INFO* Area;

  explicit ThreadCPUArea(_TEB* TEB)
    : Area(*reinterpret_cast<CHPE_V2_CPU_AREA_INFO**>(reinterpret_cast<uintptr_t>(TEB) + TEBCPUAreaOffset)) {}

  uint64_t& EmulatorStackLimit() const {
    return Area->EmulatorStackLimit;
  }

  uint64_t& EmulatorStackBase() const {
    return Area->EmulatorStackBase;
  }

  ARM64EC_NT_CONTEXT& ContextAmd64() const {
    return *Area->ContextAmd64;
  }

  FEXCore::Core::CpuStateFrame*& StateFrame() const {
    return reinterpret_cast<FEXCore::Core::CpuStateFrame*&>(Area->EmulatorData[0]);
  }

  FEXCore::Core::InternalThreadState*& ThreadState() const {
    return reinterpret_cast<FEXCore::Core::InternalThreadState*&>(Area->EmulatorData[1]);
  }

  uint64_t& DispatcherLoopTopEnterEC() const {
    return reinterpret_cast<uint64_t&>(Area->EmulatorData[2]);
  }

  uint64_t& DispatcherLoopTopEnterECFillSRA() const {
    return reinterpret_cast<uint64_t&>(Area->EmulatorData[3]);
  }
};


extern "C" NTSTATUS NtContinueNative(ARM64_NT_CONTEXT* NativeContext, BOOLEAN Alert);

namespace {
fextl::unique_ptr<FEXCore::Context::Context> CTX;
fextl::unique_ptr<FEX::DummyHandlers::DummySignalDelegator> SignalDelegator;
fextl::unique_ptr<ECSyscallHandler> SyscallHandler;
std::optional<FEX::Windows::InvalidationTracker> InvalidationTracker;
std::optional<FEX::Windows::CPUFeatures> CPUFeatures;

std::recursive_mutex ThreadCreationMutex;
// Map of TIDs to their FEX thread state, `ThreadCreationMutex` must be locked when accessing
std::unordered_map<DWORD, FEXCore::Core::InternalThreadState*> Threads;


std::pair<NTSTATUS, ThreadCPUArea> GetThreadCPUArea(HANDLE Thread) {
  THREAD_BASIC_INFORMATION Info;
  const NTSTATUS Err = NtQueryInformationThread(Thread, ThreadBasicInformation, &Info, sizeof(Info), nullptr);
  return {Err, ThreadCPUArea(reinterpret_cast<_TEB*>(Info.TebBaseAddress))};
}

ThreadCPUArea GetCPUArea() {
  return ThreadCPUArea(NtCurrentTeb());
}

bool IsEmulatorStackAddress(uint64_t Address) {
  return Address <= GetCPUArea().EmulatorStackBase() && Address >= GetCPUArea().EmulatorStackLimit();
}

bool IsDispatcherAddress(uint64_t Address) {
  const auto& Config = SignalDelegator->GetConfig();
  return Address >= Config.DispatcherBegin && Address < Config.DispatcherEnd;
}

// GetProcAddress on ARM64EC returns a pointer to an x64 fast forward sequence to allow for redirecting to the JIT if functions are
// hotpatched. This looks up the procedure address of the native code even if the fast forward sequence has been patched.
uintptr_t GetRedirectedProcAddress(HMODULE Module, const char* ProcName) {
  const uintptr_t Proc = reinterpret_cast<uintptr_t>(GetProcAddress(Module, ProcName));
  if (!Proc) {
    return 0;
  }

  ULONG Size;
  const auto* LoadConfig =
    reinterpret_cast<_IMAGE_LOAD_CONFIG_DIRECTORY64*>(RtlImageDirectoryEntryToData(Module, true, IMAGE_DIRECTORY_ENTRY_LOAD_CONFIG, &Size));
  const auto* CHPEMetadata = reinterpret_cast<IMAGE_ARM64EC_METADATA*>(LoadConfig->CHPEMetadataPointer);
  const uintptr_t ModuleBase = reinterpret_cast<uintptr_t>(Module);
  const uintptr_t ProcRVA = Proc - ModuleBase;
  const auto* RedirectionTableBegin = reinterpret_cast<IMAGE_ARM64EC_REDIRECTION_ENTRY*>(ModuleBase + CHPEMetadata->RedirectionMetadata);
  const auto* RedirectionTableEnd = RedirectionTableBegin + CHPEMetadata->RedirectionMetadataCount;
  const auto* It =
    std::lower_bound(RedirectionTableBegin, RedirectionTableEnd, ProcRVA, [](const auto& Entry, uintptr_t RVA) { return Entry.Source < RVA; });
  if (It->Source != ProcRVA) {
    return 0;
  }
  return ModuleBase + It->Destination;
}
} // namespace

namespace Exception {
static std::optional<FEX::Windows::TSOHandlerConfig> HandlerConfig;
static uintptr_t KiUserExceptionDispatcher;

static EXCEPTION_RECORD HandleGuestException(const EXCEPTION_RECORD& Src, ARM64_NT_CONTEXT& Context) {
  auto* Thread = GetCPUArea().ThreadState();
  auto& Fault = Thread->CurrentFrame->SynchronousFaultData;
  EXCEPTION_RECORD Dst = Src;
  Dst.ExceptionAddress = reinterpret_cast<void*>(Context.Pc);

  // Windows always clears TF, DF and AF when handling an exception, restoring after.
  // TODO: Check windows behaviour for the restoring after, quite awkward to achieve with the BT API. Would need to fixup flags after a
  // rethrow and keep track of context pointers on the stack so if a SEH handler changes flags they can be restored in BeginContext after
  // the NtContinue syscall (which will convert to an ARM64 context and back, losing these flags).
  uint32_t EFlags = CTX->ReconstructCompactedEFLAGS(Thread, true, Context.X, Context.Cpsr);
  EFlags &= ~((1 << FEXCore::X86State::RFLAG_DF_RAW_LOC) | (1 << FEXCore::X86State::RFLAG_TF_LOC) | (1 << FEXCore::X86State::RFLAG_AF_RAW_LOC));
  CTX->SetFlagsFromCompactedEFLAGS(Thread, EFlags);

  if (!Fault.FaultToTopAndGeneratedException) {
    return Dst;
  }
  Fault.FaultToTopAndGeneratedException = false;

  Dst.ExceptionFlags = 0;
  Dst.NumberParameters = 0;

  switch (Fault.Signal) {
  case FEXCore::Core::FAULT_SIGILL: Dst.ExceptionCode = EXCEPTION_ILLEGAL_INSTRUCTION; return Dst;
  case FEXCore::Core::FAULT_SIGTRAP:
    switch (Fault.TrapNo) {
    case FEXCore::X86State::X86_TRAPNO_DB: Dst.ExceptionCode = EXCEPTION_SINGLE_STEP; return Dst;
    case FEXCore::X86State::X86_TRAPNO_BP:
      Context.Pc -= 1;
      Dst.ExceptionAddress = reinterpret_cast<void*>(Context.Pc);
      Dst.ExceptionCode = EXCEPTION_BREAKPOINT;
      Dst.NumberParameters = 1;
      Dst.ExceptionInformation[0] = 0;
      return Dst;
    default: LogMan::Msg::EFmt("Unknown SIGTRAP trap: {}", Fault.TrapNo); break;
    }
    break;
  case FEXCore::Core::FAULT_SIGSEGV:
    switch (Fault.TrapNo) {
    case FEXCore::X86State::X86_TRAPNO_GP:
      if ((Fault.err_code & 0b111) == 0b010) {
        switch (Fault.err_code >> 3) {
        case 0x2d:
          Context.Pc += 2;
          Dst.ExceptionCode = EXCEPTION_BREAKPOINT;
          Dst.ExceptionAddress = reinterpret_cast<void*>(Context.Pc + 1);
          Dst.NumberParameters = 1;
          Dst.ExceptionInformation[0] = Context.X8; // RAX
          // Note that ExceptionAddress doesn't equal the reported context RIP here, this discrepancy expected and not having it can trigger anti-debug logic.
          return Dst;
        default: LogMan::Msg::EFmt("Unknown interrupt: 0x{:X}", Fault.err_code >> 3); break;
        }
      } else {
        Dst.ExceptionCode = EXCEPTION_PRIV_INSTRUCTION;
        return Dst;
      }
      break;
    case FEXCore::X86State::X86_TRAPNO_OF: Dst.ExceptionCode = EXCEPTION_INT_OVERFLOW; return Dst;
    default: LogMan::Msg::EFmt("Unknown SIGSEGV trap: {}", Fault.TrapNo); break;
    }
    break;
  default: LogMan::Msg::EFmt("Unknown signal type: {}", Fault.Signal); break;
  }

  // Default to SIGILL
  Dst.ExceptionCode = EXCEPTION_ILLEGAL_INSTRUCTION;
  return Dst;
}

static bool HandleUnalignedAccess(ARM64_NT_CONTEXT& Context) {
  if (!CTX->IsAddressInCodeBuffer(GetCPUArea().ThreadState(), Context.Pc)) {
    return false;
  }

  const auto Result = FEXCore::ArchHelpers::Arm64::HandleUnalignedAccess(GetCPUArea().ThreadState(),
                                                                         HandlerConfig->GetUnalignedHandlerType(), Context.Pc, &Context.X0);
  if (!Result.first) {
    return false;
  }

  Context.Pc += Result.second;
  return true;
}

static void LoadStateFromECContext(FEXCore::Core::InternalThreadState* Thread, CONTEXT& Context) {
  auto& State = Thread->CurrentFrame->State;

  if (Context.ContextFlags & CONTEXT_INTEGER) {
    // General register state
    State.gregs[FEXCore::X86State::REG_RAX] = Context.Rax;
    State.gregs[FEXCore::X86State::REG_RCX] = Context.Rcx;
    State.gregs[FEXCore::X86State::REG_RDX] = Context.Rdx;
    State.gregs[FEXCore::X86State::REG_RBX] = Context.Rbx;

    State.gregs[FEXCore::X86State::REG_RSI] = Context.Rsi;
    State.gregs[FEXCore::X86State::REG_RDI] = Context.Rdi;
    State.gregs[FEXCore::X86State::REG_R8] = Context.R8;
    State.gregs[FEXCore::X86State::REG_R9] = Context.R9;
    State.gregs[FEXCore::X86State::REG_R10] = Context.R10;
    State.gregs[FEXCore::X86State::REG_R11] = Context.R11;
    State.gregs[FEXCore::X86State::REG_R12] = Context.R12;
    State.gregs[FEXCore::X86State::REG_R13] = Context.R13;
    State.gregs[FEXCore::X86State::REG_R14] = Context.R14;
    State.gregs[FEXCore::X86State::REG_R15] = Context.R15;
  }

  if (Context.ContextFlags & CONTEXT_CONTROL) {
    State.rip = Context.Rip;
    State.gregs[FEXCore::X86State::REG_RSP] = Context.Rsp;
    State.gregs[FEXCore::X86State::REG_RBP] = Context.Rbp;
    CTX->SetFlagsFromCompactedEFLAGS(Thread, Context.EFlags);
  }

  if (Context.ContextFlags & CONTEXT_SEGMENTS) {
    State.es_idx = Context.SegEs & 0xffff;
    State.cs_idx = Context.SegCs & 0xffff;
    State.ss_idx = Context.SegSs & 0xffff;
    State.ds_idx = Context.SegDs & 0xffff;
    State.fs_idx = Context.SegFs & 0xffff;
    State.gs_idx = Context.SegGs & 0xffff;

    // The TEB is the only populated GDT entry by default
    const auto TEB = reinterpret_cast<uint64_t>(NtCurrentTeb());
    State.gdt[(Context.SegGs & 0xffff) >> 3].base = TEB;
    State.gs_cached = TEB;
    State.fs_cached = 0;
    State.es_cached = 0;
    State.cs_cached = 0;
    State.ss_cached = 0;
    State.ds_cached = 0;
  }

  if (Context.ContextFlags & CONTEXT_FLOATING_POINT) {
    // Floating-point register state
    CTX->SetXMMRegistersFromState(Thread, reinterpret_cast<const __uint128_t*>(Context.FltSave.XmmRegisters), nullptr);
    memcpy(State.mm, Context.FltSave.FloatRegisters, sizeof(State.mm));

    State.FCW = Context.FltSave.ControlWord;
    State.flags[FEXCore::X86State::X87FLAG_C0_LOC] = (Context.FltSave.StatusWord >> 8) & 1;
    State.flags[FEXCore::X86State::X87FLAG_C1_LOC] = (Context.FltSave.StatusWord >> 9) & 1;
    State.flags[FEXCore::X86State::X87FLAG_C2_LOC] = (Context.FltSave.StatusWord >> 10) & 1;
    State.flags[FEXCore::X86State::X87FLAG_C3_LOC] = (Context.FltSave.StatusWord >> 14) & 1;
    State.flags[FEXCore::X86State::X87FLAG_TOP_LOC] = (Context.FltSave.StatusWord >> 11) & 0b111;
    State.AbridgedFTW = Context.FltSave.TagWord;
  }
}

static void ReconstructThreadState(ARM64_NT_CONTEXT& Context) {
  const auto& Config = SignalDelegator->GetConfig();
  auto* Thread = GetCPUArea().ThreadState();
  auto& State = Thread->CurrentFrame->State;

  State.rip = CTX->RestoreRIPFromHostPC(Thread, Context.Pc);

  // Spill all SRA GPRs
  for (size_t i = 0; i < Config.SRAGPRCount; i++) {
    State.gregs[i] = Context.X[Config.SRAGPRMapping[i]];
  }

  // Spill all SRA FPRs
  for (size_t i = 0; i < Config.SRAFPRCount; i++) {
    memcpy(State.xmm.sse.data[i], &Context.V[Config.SRAFPRMapping[i]], sizeof(__uint128_t));
  }
}

// Reconstructs an x64 context from the input context within the JIT, packed into a regular ARM64 context following the ARM64EC register mapping
static ARM64_NT_CONTEXT ReconstructPackedECContext(ARM64_NT_CONTEXT& Context) {
  ReconstructThreadState(Context);
  ARM64_NT_CONTEXT ECContext {};

  ECContext.ContextFlags = CONTEXT_ARM64_FULL;

  auto* Thread = GetCPUArea().ThreadState();
  auto& State = Thread->CurrentFrame->State;

  ECContext.X8 = State.gregs[FEXCore::X86State::REG_RAX];
  ECContext.X0 = State.gregs[FEXCore::X86State::REG_RCX];
  ECContext.X1 = State.gregs[FEXCore::X86State::REG_RDX];
  ECContext.X27 = State.gregs[FEXCore::X86State::REG_RBX];
  ECContext.Sp = State.gregs[FEXCore::X86State::REG_RSP];
  ECContext.Fp = State.gregs[FEXCore::X86State::REG_RBP];
  ECContext.X25 = State.gregs[FEXCore::X86State::REG_RSI];
  ECContext.X26 = State.gregs[FEXCore::X86State::REG_RDI];
  ECContext.X2 = State.gregs[FEXCore::X86State::REG_R8];
  ECContext.X3 = State.gregs[FEXCore::X86State::REG_R9];
  ECContext.X4 = State.gregs[FEXCore::X86State::REG_R10];
  ECContext.X5 = State.gregs[FEXCore::X86State::REG_R11];
  ECContext.X19 = State.gregs[FEXCore::X86State::REG_R12];
  ECContext.X20 = State.gregs[FEXCore::X86State::REG_R13];
  ECContext.X21 = State.gregs[FEXCore::X86State::REG_R14];
  ECContext.X22 = State.gregs[FEXCore::X86State::REG_R15];

  ECContext.Pc = State.rip;

  CTX->ReconstructXMMRegisters(Thread, reinterpret_cast<__uint128_t*>(&ECContext.V[0]), nullptr);

  ECContext.Lr = State.mm[0][0];
  ECContext.X6 = State.mm[1][0];
  ECContext.X7 = State.mm[2][0];
  ECContext.X9 = State.mm[3][0];
  ECContext.X16 = (State.mm[3][1] & 0xffff) << 48 | (State.mm[2][1] & 0xffff) << 32 | (State.mm[1][1] & 0xffff) << 16 | (State.mm[0][1] & 0xffff);
  ECContext.X10 = State.mm[4][0];
  ECContext.X11 = State.mm[5][0];
  ECContext.X12 = State.mm[6][0];
  ECContext.X15 = State.mm[7][0];
  ECContext.X17 = (State.mm[7][1] & 0xffff) << 48 | (State.mm[6][1] & 0xffff) << 32 | (State.mm[5][1] & 0xffff) << 16 | (State.mm[4][1] & 0xffff);

  // Zero all disallowed registers
  ECContext.X13 = 0;
  ECContext.X14 = 0;
  ECContext.X18 = 0;
  ECContext.X23 = 0;
  ECContext.X24 = 0;
  ECContext.X28 = 0;

  // NZCV will be converted into EFlags by ntdll, the rest are lost during exception handling.
  // See HandleGuestException
  ECContext.Cpsr = Context.Cpsr;
  ECContext.Fpcr = Context.Fpcr;
  ECContext.Fpsr = Context.Fpsr;

  return ECContext;
}

static void RethrowGuestException(const EXCEPTION_RECORD& Rec, ARM64_NT_CONTEXT& Context) {
  const auto& Config = SignalDelegator->GetConfig();
  uint64_t GuestSp = Context.X[Config.SRAGPRMapping[static_cast<size_t>(FEXCore::X86State::REG_RSP)]];
  struct DispatchArgs {
    ARM64_NT_CONTEXT Context;
    EXCEPTION_RECORD Rec;
    uint64_t Align;
    uint64_t Redzone[2];
  }* Args = reinterpret_cast<DispatchArgs*>(FEXCore::AlignDown(GuestSp, 64)) - 1;

  LogMan::Msg::DFmt("Reconstructing context");
  Args->Context = ReconstructPackedECContext(Context);
  LogMan::Msg::DFmt("pc: {:X} rip: {:X}", Context.Pc, Args->Context.Pc);
  Args->Rec = HandleGuestException(Rec, Args->Context);
  Context.Sp = reinterpret_cast<uint64_t>(Args);
  Context.Pc = KiUserExceptionDispatcher;
}
} // namespace Exception

class ECSyscallHandler : public FEXCore::HLE::SyscallHandler, public FEXCore::Allocator::FEXAllocOperators {
public:
  ECSyscallHandler() {
    OSABI = FEXCore::HLE::SyscallOSABI::OS_GENERIC;
  }

  uint64_t HandleSyscall(FEXCore::Core::CpuStateFrame* Frame, FEXCore::HLE::SyscallArguments* Args) override {
    return 0;
  }

  FEXCore::HLE::SyscallABI GetSyscallABI(uint64_t Syscall) override {
    return {.NumArgs = 0, .HasReturn = false, .HostSyscallNumber = -1};
  }

  FEXCore::HLE::AOTIRCacheEntryLookupResult LookupAOTIRCacheEntry(FEXCore::Core::InternalThreadState* Thread, uint64_t GuestAddr) override {
    return {0, 0};
  }

  void MarkGuestExecutableRange(FEXCore::Core::InternalThreadState* Thread, uint64_t Start, uint64_t Length) override {
    InvalidationTracker->ReprotectRWXIntervals(Start, Length);
  }
};

extern "C" void SyncThreadContext(CONTEXT* Context) {
  auto* Thread = GetCPUArea().ThreadState();
  Exception::LoadStateFromECContext(Thread, *Context);
}

void ProcessInit() {
  FEX::Config::InitializeConfigs();
  FEXCore::Config::Initialize();
  FEXCore::Config::AddLayer(FEX::Config::CreateGlobalMainLayer());
  FEXCore::Config::AddLayer(FEX::Config::CreateMainLayer());
  FEXCore::Config::Load();
  FEXCore::Config::ReloadMetaLayer();
  FEX::Windows::Logging::Init();

  FEXCore::Config::EraseSet(FEXCore::Config::CONFIG_IS64BIT_MODE, "1");

  // Not applicable to Windows
  FEXCore::Config::EraseSet(FEXCore::Config::ConfigOption::CONFIG_TSOAUTOMIGRATION, "0");

  FEXCore::Context::InitializeStaticTables(FEXCore::Context::MODE_64BIT);

  SignalDelegator = fextl::make_unique<FEX::DummyHandlers::DummySignalDelegator>();
  SyscallHandler = fextl::make_unique<ECSyscallHandler>();
  Exception::HandlerConfig.emplace();

  CTX = FEXCore::Context::Context::CreateNewContext();
  CTX->SetSignalDelegator(SignalDelegator.get());
  CTX->SetSyscallHandler(SyscallHandler.get());
  CTX->InitCore();
  InvalidationTracker.emplace(*CTX, Threads);
  CPUFeatures.emplace(*CTX);

  X64ReturnInstr = ::VirtualAlloc(nullptr, FEXCore::Utils::FEX_PAGE_SIZE, MEM_COMMIT, PAGE_EXECUTE_READWRITE);
  *reinterpret_cast<uint8_t*>(X64ReturnInstr) = 0xc3;

  const auto NtDll = GetModuleHandle("ntdll.dll");
  Exception::KiUserExceptionDispatcher = GetRedirectedProcAddress(NtDll, "KiUserExceptionDispatcher");
  const auto WineSyscallDispatcherPtr = reinterpret_cast<void**>(GetProcAddress(NtDll, "__wine_syscall_dispatcher"));
  if (WineSyscallDispatcherPtr) {
    WineSyscallDispatcher = *WineSyscallDispatcherPtr;
  }
}

void ProcessTerm() {}

class ScopedCallbackDisable {
private:
  bool Prev;

public:
  ScopedCallbackDisable() {
    Prev = GetCPUArea().Area->InSyscallCallback;
    GetCPUArea().Area->InSyscallCallback = true;
  }

  ~ScopedCallbackDisable() {
    GetCPUArea().Area->InSyscallCallback = Prev;
  }
};

NTSTATUS ResetToConsistentState(EXCEPTION_POINTERS* Ptrs, ARM64_NT_CONTEXT* Context, BOOLEAN* Continue) {
  ScopedCallbackDisable Guard;
  const auto* Exception = Ptrs->ExceptionRecord;
  if (Exception->ExceptionCode == EXCEPTION_DATATYPE_MISALIGNMENT && Exception::HandleUnalignedAccess(*Context)) {
    LogMan::Msg::DFmt("Handled unaligned atomic: new pc: {:X}", Context->Pc);
    *Continue = true;
    return STATUS_SUCCESS;
  }

  if (Exception->ExceptionCode == EXCEPTION_ACCESS_VIOLATION) {
    const auto FaultAddress = static_cast<uint64_t>(Exception->ExceptionInformation[1]);

    bool HandledRWX = false;
    if (InvalidationTracker && GetCPUArea().ThreadState()) {
      std::scoped_lock Lock(ThreadCreationMutex);
      HandledRWX = InvalidationTracker->HandleRWXAccessViolation(FaultAddress);
    }

    if (HandledRWX) {
      LogMan::Msg::DFmt("Handled self-modifying code: pc: {:X} fault: {:X}", Context->Pc, FaultAddress);
      *Continue = true;
      return STATUS_SUCCESS;
    }
  }

  if (!CTX->IsAddressInCodeBuffer(GetCPUArea().ThreadState(), Context->Pc) && !IsDispatcherAddress(Context->Pc)) {
    return STATUS_SUCCESS;
  }


  if (IsEmulatorStackAddress(reinterpret_cast<uint64_t>(__builtin_frame_address(0)))) {
    Exception::RethrowGuestException(*Exception, *Context);
    LogMan::Msg::DFmt("Rethrowing onto guest stack: {:X}", Context->Sp);
    *Continue = true;
    return STATUS_SUCCESS;
  } else {
    LogMan::Msg::EFmt("Unexpected exception in JIT code on guest stack");
    return STATUS_SUCCESS;
  }
}

void NotifyMemoryAlloc(void* Address, SIZE_T Size, ULONG Type, ULONG Prot) {
  if (!InvalidationTracker || !GetCPUArea().ThreadState()) {
    return;
  }

  std::scoped_lock Lock(ThreadCreationMutex);
  InvalidationTracker->HandleMemoryProtectionNotification(reinterpret_cast<uint64_t>(Address), static_cast<uint64_t>(Size), Prot);
}

void NotifyMemoryFree(void* Address, SIZE_T Size, ULONG FreeType) {
  if (!InvalidationTracker || !GetCPUArea().ThreadState()) {
    return;
  }

  std::scoped_lock Lock(ThreadCreationMutex);
  if (!Size) {
    InvalidationTracker->InvalidateContainingSection(reinterpret_cast<uint64_t>(Address), true);
  } else if (FreeType & MEM_DECOMMIT) {
    InvalidationTracker->InvalidateAlignedInterval(reinterpret_cast<uint64_t>(Address), static_cast<uint64_t>(Size), true);
  }
}

void NotifyMemoryProtect(void* Address, SIZE_T Size, ULONG NewProt) {
  if (!InvalidationTracker || !GetCPUArea().ThreadState()) {
    return;
  }

  std::scoped_lock Lock(ThreadCreationMutex);
  InvalidationTracker->HandleMemoryProtectionNotification(reinterpret_cast<uint64_t>(Address), static_cast<uint64_t>(Size), NewProt);
}

void NotifyUnmapViewOfSection(void* Address) {
  if (!InvalidationTracker || !GetCPUArea().ThreadState()) {
    return;
  }

  std::scoped_lock Lock(ThreadCreationMutex);
  InvalidationTracker->InvalidateContainingSection(reinterpret_cast<uint64_t>(Address), true);
}

void BTCpu64FlushInstructionCache(const void* Address, SIZE_T Size) {
  if (!InvalidationTracker || !GetCPUArea().ThreadState()) {
    return;
  }

  std::scoped_lock Lock(ThreadCreationMutex);
  InvalidationTracker->InvalidateAlignedInterval(reinterpret_cast<uint64_t>(Address), static_cast<uint64_t>(Size), false);
}

NTSTATUS ThreadInit() {
  static constexpr size_t EmulatorStackSize = 0x40000;
  const uint64_t EmulatorStack = reinterpret_cast<uint64_t>(::VirtualAlloc(nullptr, EmulatorStackSize, MEM_COMMIT | MEM_RESERVE, PAGE_READWRITE));
  GetCPUArea().EmulatorStackLimit() = EmulatorStack;
  GetCPUArea().EmulatorStackBase() = EmulatorStack + EmulatorStackSize;

  const auto CPUArea = GetCPUArea();

  auto* Thread = CTX->CreateThread(0, 0);
  Thread->CurrentFrame->Pointers.Common.ExitFunctionEC = reinterpret_cast<uintptr_t>(&ExitFunctionEC);
  CPUArea.StateFrame() = Thread->CurrentFrame;

  uint64_t EnterEC = Thread->CurrentFrame->Pointers.Common.DispatcherLoopTopEnterEC;
  CPUArea.DispatcherLoopTopEnterEC() = EnterEC;

  uint64_t EnterECFillSRA = Thread->CurrentFrame->Pointers.Common.DispatcherLoopTopEnterECFillSRA;
  CPUArea.DispatcherLoopTopEnterECFillSRA() = EnterECFillSRA;

  CPUArea.ContextAmd64() = {.ContextFlags = CONTEXT_CONTROL | CONTEXT_SEGMENTS | CONTEXT_INTEGER | CONTEXT_FLOATING_POINT,
                            .AMD64_SegCs = 0x33,
                            .AMD64_SegDs = 0x2b,
                            .AMD64_SegEs = 0x2b,
                            .AMD64_SegFs = 0x53,
                            .AMD64_SegGs = 0x2b,
                            .AMD64_SegSs = 0x2b,
                            .AMD64_EFlags = 0x202,
                            .AMD64_MxCsr = 0x1f80,
                            .AMD64_MxCsr_copy = 0x1f80,
                            .AMD64_ControlWord = 0x27f};
  Exception::LoadStateFromECContext(Thread, CPUArea.ContextAmd64().AMD64_Context);

  {
    std::scoped_lock Lock(ThreadCreationMutex);
    Threads.emplace(GetCurrentThreadId(), Thread);
  }

  CPUArea.ThreadState() = Thread;
  return STATUS_SUCCESS;
}

NTSTATUS ThreadTerm(HANDLE Thread) {
  const auto [Err, CPUArea] = GetThreadCPUArea(Thread);
  if (Err) {
    return Err;
  }
  auto* OldThreadState = CPUArea.ThreadState();
  CPUArea.ThreadState() = nullptr;

  {
    THREAD_BASIC_INFORMATION Info;
    if (NTSTATUS Err = NtQueryInformationThread(Thread, ThreadBasicInformation, &Info, sizeof(Info), nullptr); Err) {
      return Err;
    }

    const auto ThreadTID = reinterpret_cast<uint64_t>(Info.ClientId.UniqueThread);
    std::scoped_lock Lock(ThreadCreationMutex);
    Threads.erase(ThreadTID);
  }

  CTX->DestroyThread(OldThreadState);
  ::VirtualFree(reinterpret_cast<void*>(GetCPUArea().EmulatorStackLimit()), 0, MEM_RELEASE);
  return STATUS_SUCCESS;
}

BOOLEAN BTCpu64IsProcessorFeaturePresent(UINT Feature) {
  return CPUFeatures->IsFeaturePresent(Feature) ? TRUE : FALSE;
}

void UpdateProcessorInformation(SYSTEM_CPU_INFORMATION* Info) {
  CPUFeatures->UpdateInformation(Info);
}