Make Windows NCE code actually compliable

src/common/host_memory.cpp

@@ -110,6 +110,9 @@ using PFN_MapViewOfFile3 = _Ret_maybenull_ PVOID(WINAPI*)(
 using PFN_UnmapViewOfFile2 = BOOL(WINAPI*)(_In_ HANDLE Process, _In_ PVOID BaseAddress,
                                            _In_ ULONG UnmapFlags);
 
+using PFN_VirtualQuery = SIZE_T(WINAPI*) (
+    _In_opt_ LPCVOID lpAddress, _Out_ PMEMORY_BASIC_INFORMATION lpBuffer, _In_ SIZE_T dwLength);
+
 template <typename T>
 static void GetFuncAddress(Common::DynamicLibrary& dll, const char* name, T& pfn) {
     if (!dll.GetSymbol(name, &pfn)) {
@@ -134,10 +137,11 @@ public:
         }
         GetFuncAddress(kernelbase_dll, "CreateFileMapping2", pfn_CreateFileMapping2);
         GetFuncAddress(kernelbase_dll, "VirtualAlloc2", pfn_VirtualAlloc2);
+        GetFuncAddress(kernelbase_dll, "VirtualQuery", pfn_VirtualQuery);
         GetFuncAddress(kernelbase_dll, "MapViewOfFile3", pfn_MapViewOfFile3);
         GetFuncAddress(kernelbase_dll, "UnmapViewOfFile2", pfn_UnmapViewOfFile2);
 
-        if (!pfn_CreateFileMapping2 || !pfn_VirtualAlloc2 || !pfn_MapViewOfFile3 || !pfn_UnmapViewOfFile2) {
+        if (!pfn_CreateFileMapping2 || !pfn_VirtualAlloc2 || !pfn_VirtualQuery || !pfn_MapViewOfFile3 || !pfn_UnmapViewOfFile2) {
             LOG_CRITICAL(HW_Memory, "Failed to find functions for virtual allocs");
             return false;
         }
@@ -162,8 +166,39 @@ public:
             LOG_CRITICAL(HW_Memory, "Failed to map {} MiB of virtual memory", backing_size >> 20);
             return false;
         }
-        // Allocate virtual address placeholder
-        virtual_base = static_cast<u8*>(pfn_VirtualAlloc2(process, nullptr, virtual_size, MEM_RESERVE | MEM_RESERVE_PLACEHOLDER, PAGE_NOACCESS, nullptr, 0));
+
+        // Allocate virtual address placeholder within a 39-bit address space
+        SIZE_T cursor = 0;
+        while (cursor < (1ULL << 39) - virtual_size) {
+            MEMORY_BASIC_INFORMATION info{};
+
+            // find the next mapped region of memory
+            auto res = pfn_VirtualQuery(reinterpret_cast<LPCVOID>(cursor), &info, sizeof(info));
+
+            if (res == 0) {
+                LOG_WARNING(HW_Memory, "Failed to check memory region: {}", GetLastError());
+                continue;
+            }
+
+            auto start_aligned = AlignUp(reinterpret_cast<SIZE_T>(info.BaseAddress), HugePageSize);
+            // is this region free?
+            if (info.State == MEM_FREE && start_aligned < reinterpret_cast<SIZE_T>(info.BaseAddress) + info.RegionSize) {
+                // is this region big enough for us to use?
+                if (info.RegionSize - (start_aligned - reinterpret_cast<SIZE_T>(info.BaseAddress)) >= virtual_size) {
+                    virtual_base = static_cast<u8*>(pfn_VirtualAlloc2
+                        (process, reinterpret_cast<PVOID>(start_aligned), virtual_size, MEM_RESERVE | MEM_RESERVE_PLACEHOLDER, PAGE_NOACCESS, nullptr, 0));
+                }
+            }
+
+            cursor = reinterpret_cast<SIZE_T>(info.BaseAddress) + info.RegionSize;
+        }
+        // Check if we failed to allocate for direct-mapping, otherwise map normally
+        if (!virtual_base) {
+            LOG_WARNING(HW_Memory, "Failed to allocate within 39-bit address space, direct mapping is not supported");
+            virtual_base = static_cast<u8*>(pfn_VirtualAlloc2
+                        (process, nullptr, virtual_size, MEM_RESERVE | MEM_RESERVE_PLACEHOLDER, PAGE_NOACCESS, nullptr, 0));
+        }
+
         if (!virtual_base) {
             Release();
             LOG_CRITICAL(HW_Memory, "Failed to reserve {} GiB of virtual memory", virtual_size >> 30);
@@ -383,6 +418,7 @@ private:
     DynamicLibrary kernelbase_dll;
     PFN_CreateFileMapping2 pfn_CreateFileMapping2{};
     PFN_VirtualAlloc2 pfn_VirtualAlloc2{};
+    PFN_VirtualQuery pfn_VirtualQuery{};
     PFN_MapViewOfFile3 pfn_MapViewOfFile3{};
     PFN_UnmapViewOfFile2 pfn_UnmapViewOfFile2{};
 

src/core/CMakeLists.txt

@@ -1251,9 +1251,7 @@ if (HAS_NCE)
     if (WIN32)
         target_sources(core PRIVATE
                 arm/nce/win/platform_visitor.h
-                arm/nce/win/platform_visitor.cpp
-                arm/nce/win/exceptions.h
-                arm/nce/win/exceptions.cpp)
+                arm/nce/win/platform_visitor.cpp)
     endif ()
 endif()
 

src/core/arm/nce/arm_nce.cpp

@@ -37,19 +37,17 @@
 
 #include "core/hle/kernel/k_process.h"
 
-#ifndef __WIN32
+#ifndef _WIN32
 #include <unistd.h>
 #include <sys/syscall.h>
 #include <signal.h>
-#else
-#include "core/arm/nce/win/exceptions.h"
 #endif
 
 namespace Core {
 
 namespace {
 
-#ifndef __WIN32
+#ifndef _WIN32
 struct sigaction g_orig_bus_action;
 struct sigaction g_orig_segv_action;
 #endif
@@ -72,7 +70,7 @@ void* ArmNce::GetGuestParameters() {
         : [out] "=&r"(nep)
         : [off] "i"((ContextKey - 1) * 8)
         : "memory");
-#elif defined(__WIN32)
+#elif defined(_WIN32)
     asm volatile(
         "mrs %[out], TPIDR_EL0\n"           // load windows TLS storage
         "ldr %[out], [ %[out], #%[off] ]\n"
@@ -106,7 +104,7 @@ void ArmNce::UnlockThreadParameters(void* tpidr) {
     static_cast<NativeExecutionParameters*>(tpidr)->lock.store(SpinLockUnlocked, std::memory_order_release);
 }
 
-#ifndef __WIN32
+#ifndef _WIN32
 YUZU_NAKED
 YUZU_NO_INLINE
 HaltReason ArmNce::ReturnToRunCodeByExceptionLevelChange(int tid, void *tpidr) {
@@ -143,10 +141,10 @@ HaltReason ArmNce::ReturnToRunCodeByExceptionLevelChange(int tid, void *tpidr) {
 }
 YUZU_NAKED_END
 #else
-HaltReason ArmNce::ReturnToRunCodeByExceptionLevelChange(int tid, void *tpidr) {
-    DEBUG_ASSERT(os::TlsGetValue(ContextKey) == tpidr);
+HaltReason ArmNce::ReturnToRunCodeByExceptionLevelChange(void* tid, void *tpidr) {
+    DEBUG_ASSERT(TlsGetValue(ContextKey) == tpidr);
 
-    os::RaiseException(SIGUSR2, 0, 0, nullptr); // TODO: pass tpidr through arguments?
+    RaiseException(ExceptionLevelChangeSignal, 0, 0, nullptr); // TODO: pass tpidr through arguments?
     __builtin_unreachable();
 }
 #endif
@@ -154,7 +152,7 @@ HaltReason ArmNce::ReturnToRunCodeByExceptionLevelChange(int tid, void *tpidr) {
 void ArmNce::ReturnToRunCodeByExceptionLevelChangeSignalHandler(int sig, void *info, void *raw_context) {
     auto tpidr = static_cast<NativeExecutionParameters*>(RestoreGuestContext(raw_context));
 
-#if !defined(__APPLE__) && !defined(__WIN32)
+#if !defined(__APPLE__) && !defined(_WIN32)
     // Save old value of TPIDR_EL0, load guest one
     u64 tpidr_el0;
     asm volatile("mrs %0, TPIDR_EL0\n"
@@ -186,7 +184,7 @@ HaltReason ArmNce::ReturnToRunCodeByTrampoline(void *tpidr, u64 trampoline_addr)
         "ldr x2, [ x0, #%[ctx_off] ]\n"
         "add x5, x2, #%[host_ctx] \n"
 
-#if !defined(__APPLE__) && !defined(__WIN32)
+#if !defined(__APPLE__) && !defined(_WIN32)
         // Load guest tpidr_el0
         "mrs x4, TPIDR_EL0\n"
         "msr TPIDR_EL0, x0\n"
@@ -223,7 +221,7 @@ HaltReason ArmNce::ReturnToRunCodeByTrampoline(void *tpidr, u64 trampoline_addr)
         :: [ctx_off] "i"(offsetof(NativeExecutionParameters, native_context)),
         [sp_off] "i"(offsetof(GuestContext, sp)),
         [host_ctx] "i"(offsetof(GuestContext, host_ctx))
-#if defined(__APPLE__) || defined(__WIN32)
+#if defined(__APPLE__) || defined(_WIN32)
         ,[is_running_off] "i"(offsetof(NativeExecutionParameters, is_actually_running))
 #endif
         );
@@ -232,7 +230,7 @@ YUZU_NAKED_END
 
 static_assert(offsetof(HostContext, host_sp) == 0xE0); // TODO: don't use magic number
 
-#ifndef __WIN32
+#ifndef _WIN32
 
 void ArmNce::BreakFromRunCodeSignalHandler(int sig, void *info, void *raw_context) {
     NativeExecutionParameters* tpidr = static_cast<NativeExecutionParameters *>(GetGuestParameters());
@@ -255,11 +253,9 @@ void ArmNce::BreakFromRunCodeSignalHandler(int sig, void *info, void *raw_contex
 #endif
 
 void ArmNce::GuestMemoryFaultSignalHandler(int sig, void* raw_info, void* raw_context) {
-    DEBUG_ASSERT(sig == SIGSEGV || sig == SIGBUS);
-
     NativeExecutionParameters* nep = static_cast<NativeExecutionParameters*>(GetGuestParameters());
 
-#if defined(__APPLE__) || defined(__WIN32)
+#if defined(__APPLE__) || defined(_WIN32)
     if (nep->is_actually_running) {
         nep->is_actually_running = false;
 #else
@@ -274,10 +270,14 @@ void ArmNce::GuestMemoryFaultSignalHandler(int sig, void* raw_info, void* raw_co
         auto* guest_ctx = static_cast<GuestContext*>(nep->native_context);
         auto& memory = guest_ctx->parent->m_running_thread->GetOwnerProcess()->GetMemory();
 
+#ifndef _WIN32
         if (sig == SIGSEGV) {
+#else
+        if (sig == static_cast<int>(EXCEPTION_ACCESS_VIOLATION)) {
+#endif
             // Try to handle an invalid access.
             // TODO: handle accesses which split a page?
-#ifndef __WIN32
+#ifndef _WIN32
             const Common::ProcessAddress addr =
                 (reinterpret_cast<u64>(static_cast<siginfo_t*>(raw_info)->si_addr) & ~Memory::YUZU_PAGEMASK);
 #else
@@ -288,7 +288,11 @@ void ArmNce::GuestMemoryFaultSignalHandler(int sig, void* raw_info, void* raw_co
                 // We handled the access successfully and are returning to guest code.
                 goto ret;
             }
+#ifndef _WIN32
         } else if (sig == SIGBUS) {
+#else
+        } else if (sig == static_cast<int>(EXCEPTION_DATATYPE_MISALIGNMENT)) {
+#endif
             // Match and execute an instruction.
             auto ctx = KernelContext(raw_context);
             auto next_pc = MatchAndExecuteOneInstruction(memory, &ctx);
@@ -318,7 +322,7 @@ void ArmNce::GuestMemoryFaultSignalHandler(int sig, void* raw_info, void* raw_co
         nep->is_actually_running = true;
 #endif
     }
-#ifndef __WIN32
+#ifndef _WIN32
     else {
         // Host fault, call original handler
         if (sig == SIGSEGV) {
@@ -329,7 +333,7 @@ void ArmNce::GuestMemoryFaultSignalHandler(int sig, void* raw_info, void* raw_co
             UNREACHABLE_MSG("unexpected signal {}", sig);
         }
     }
-#elif
+#else
     is_host_fault = true;
 #endif
 }
@@ -338,7 +342,7 @@ void* ArmNce::RestoreGuestContext(void* raw_context) {
     // Retrieve the host context.
     auto host_ctx = KernelContext(raw_context);
 
-#ifndef __WIN32
+#ifndef _WIN32
     // Thread-local parameters will be located in x9.
     auto* tpidr = reinterpret_cast<NativeExecutionParameters*>(host_ctx.regs()[9]);
 #else
@@ -401,7 +405,11 @@ bool ArmNce::HandleFailedGuestFault(GuestContext* guest_ctx, void* raw_info, voi
     auto host_ctx = KernelContext(raw_context);
 
     // We can't handle the access, so determine why we crashed.
+#ifndef _WIN32
     const bool is_prefetch_abort = *host_ctx.pc() == reinterpret_cast<u64>(static_cast<siginfo_t*>(raw_info)->si_addr);
+#else
+    const bool is_prefetch_abort = *host_ctx.pc() == *static_cast<u64*>(raw_info);
+#endif
 
     // For data aborts, skip the instruction and return to guest code.
     // This will allow games to continue in many scenarios where they would otherwise crash.
@@ -453,7 +461,7 @@ HaltReason ArmNce::RunThread(Kernel::KThread* thread) {
 
 #if defined(__APPLE__)
     ASSERT(pthread_setspecific(ContextKey, &thread_params) == 0);
-#elif defined(__WIN32)
+#elif defined(_WIN32)
     ASSERT(TlsSetValue(ContextKey, &thread_params) == 0);
 #endif
 
@@ -526,6 +534,27 @@ ArmNce::ArmNce(System& system, bool uses_wall_clock, std::size_t core_index)
 
 ArmNce::~ArmNce() = default;
 
+#ifdef _WIN32
+LONG WINAPI ArmNce::VectoredExceptionHandler(PEXCEPTION_POINTERS info) {
+    DWORD code = info->ExceptionRecord->ExceptionCode;
+
+    if (code == EXCEPTION_ACCESS_VIOLATION || code == EXCEPTION_DATATYPE_MISALIGNMENT) {
+        GuestMemoryFaultSignalHandler(code, reinterpret_cast<void*>(&info->ExceptionRecord->ExceptionAddress), info->ContextRecord);
+        if (is_host_fault) {
+            is_host_fault = false;
+            return EXCEPTION_CONTINUE_SEARCH;
+        }
+        return EXCEPTION_CONTINUE_EXECUTION;
+    } else if (code == ExceptionLevelChangeSignal) {
+        ReturnToRunCodeByExceptionLevelChangeSignalHandler(code, reinterpret_cast<void*>(&info->ExceptionRecord->ExceptionAddress), info->ContextRecord);
+        return EXCEPTION_CONTINUE_EXECUTION;
+    } else {
+        // other exception? let it pass
+        return EXCEPTION_CONTINUE_SEARCH;
+    }
+}
+#endif
+
 #ifdef __APPLE__
 // https://github.com/apple-oss-distributions/libpthread/blob/42d026df5b07825070f60134b980a1ec2552dfee/src/pthread_tsd.c#L418-L435
 extern "C" int pthread_key_init_np(int, void (*)(void *));
@@ -542,13 +571,14 @@ void ArmNce::Initialize() {
     if (m_thread_id == -1) {
         m_thread_id = gettid();
     }
-#elif defined(__WIN32)
+#elif defined(_WIN32)
     if (m_thread_id == nullptr) {
         DuplicateHandle(GetCurrentProcess(), GetCurrentThread(), GetCurrentProcess(),
             &m_thread_id, 0, false, DUPLICATE_SAME_ACCESS);
     }
 #endif
 
+#ifndef _WIN32
     // Configure signal stack.
     if (!m_stack) {
         m_stack = std::make_unique<u8[]>(StackSize);
@@ -562,7 +592,6 @@ void ArmNce::Initialize() {
     // Set up signals.
     static std::once_flag flag;
     std::call_once(flag, [] {
-#ifndef __WIN32
         using HandlerType = decltype(sigaction::sa_sigaction);
 
         sigset_t signal_mask;
@@ -600,10 +629,11 @@ void ArmNce::Initialize() {
             reinterpret_cast<HandlerType>(&ArmNce::GuestMemoryFaultSignalHandler);
         access_fault_action.sa_mask = signal_mask;
         Common::SigAction(SIGSEGV, &access_fault_action, &g_orig_segv_action);
+    });
 #else
-        AddVectoredExceptionHandler(1, VectoredExceptionHandler);
+    static std::once_flag flag;
+    std::call_once(flag, [] { AddVectoredExceptionHandler(1, VectoredExceptionHandler); });
 #endif
-    });
 }
 
 void ArmNce::SetTpidrroEl0(u64 value) {
@@ -663,7 +693,7 @@ void ArmNce::SignalInterrupt(Kernel::KThread* thread) {
             "svc #0x80\n"
             :: "r"(static_cast<u64>(m_thread_id)), "r"(static_cast<u64>(SIGURG))
             : "x0", "x1", "x16", "memory", "cc");
-#elif defined(__WIN32)
+#elif defined(_WIN32)
         // TODO: use SetThreadState to emulate BreakFromRunCodeSignalHandler
         SuspendThread(m_thread_id);
         UnlockThreadParameters(params);

src/core/arm/nce/arm_nce.h

@@ -11,6 +11,10 @@
 #include "core/arm/arm_interface.h"
 #include "core/arm/nce/guest_context.h"
 
+#ifdef _WIN32
+#include <winternl.h>
+#endif
+
 #define SpinLockLocked 0
 #define SpinLockUnlocked 1
 
@@ -28,11 +32,14 @@ class System;
 // so we can manually initialize and use it.
 // https://github.com/apple-oss-distributions/libpthread/blob/42d026df5b07825070f60134b980a1ec2552dfee/private/pthread/tsd_private.h#L241-L245
 constexpr pthread_key_t ContextKey = 210;
-#elif __WIN32
+#elif _WIN32
+
+#define ExceptionLevelChangeSignal 0xE0000001
+thread_local bool is_host_fault = false;
 
-static const u32 ContextKey = os::TlsAlloc();
-static const u32 NCEStorage = os::TlsAlloc();
-static const u64 TlsSlots = offsetof(os::TEB, TlsSlots);
+static const u32 ContextKey = TlsAlloc();
+static const u32 NCEStorage = TlsAlloc();
+static const u64 TlsSlots = offsetof(TEB, TlsSlots);
 #endif
 
 
@@ -77,7 +84,12 @@ private:
     static void* GetGuestParameters();
 
     static HaltReason ReturnToRunCodeByTrampoline(void* tpidr, u64 trampoline_addr);
+#ifndef _WIN32
     static HaltReason ReturnToRunCodeByExceptionLevelChange(int tid, void* tpidr);
+#else
+    static HaltReason ReturnToRunCodeByExceptionLevelChange(void* tid, void* tpidr);
+    static LONG VectoredExceptionHandler(PEXCEPTION_POINTERS info);
+#endif
 
     static void ReturnToRunCodeByExceptionLevelChangeSignalHandler(int sig, void* info,
                                                                    void* raw_context);
@@ -96,7 +108,7 @@ public:
 
     // Members set on initialization.
     std::size_t m_core_index{};
-#ifndef __WIN32
+#ifndef _WIN32
     pid_t m_thread_id{-1};
 #else
     void* m_thread_id{};

src/core/arm/nce/guest_context.h

@@ -14,8 +14,9 @@
 
 #ifdef __linux__
 #include <signal.h>
-#elif __WIN32
-#include "core/arm/nce/win/nt_headers.h"
+#elif _WIN32
+#include <windows.h>
+#undef interface
 #endif
 
 namespace Core {
@@ -116,15 +117,15 @@ public:
     u32* pstate() {
         return &ptr->__ss.__cpsr;
     }
-#elif defined(__WIN32)
-    KernelContext(void* ptr) : ptr(static_cast<os::CONTEXT*>(ptr)) {}
+#elif defined(_WIN32)
+    KernelContext(void* ptr) : ptr(static_cast<CONTEXT*>(ptr)) {}
 
     u64* pc() {
-        return ptr->Pc;
+        return &ptr->Pc;
     }
 
     u64* sp() {
-        return ptr->Sp;
+        return &ptr->Sp;
     }
 
     u64* regs() {
@@ -137,15 +138,18 @@ public:
     }
 
     u32* fpcr() {
-        return &ptr->Fpcr;
+        // unsigned long vs unsigned int
+        return reinterpret_cast<u32*>(&ptr->Fpcr);
     }
 
     u32* fpsr() {
-        return &ptr->Fpsr;
+        // unsigned long vs unsigned int
+        return reinterpret_cast<u32*>(&ptr->Fpsr);
     }
 
     u32* pstate() {
-        return &ptr->Cpsr;
+        // unsigned long vs unsigned int
+        return reinterpret_cast<u32*>(&ptr->Cpsr);
     }
 #endif
 
@@ -163,8 +167,8 @@ private:
         }
         return reinterpret_cast<fpsimd_context*>(header);
     }
-#elif defined(__WIN32)
-    os::CONTEXT* ptr;
+#elif defined(_WIN32)
+    CONTEXT* ptr;
 #endif
 };
 

src/core/arm/nce/patcher.cpp

@@ -18,7 +18,7 @@
 #include "core/memory.h"
 #include "core/hle/kernel/k_thread.h"
 
-#ifdef __WIN32
+#ifdef _WIN32
 #include "win/platform_visitor.h"
 #endif
 
@@ -169,16 +169,16 @@ bool Patcher::PatchText(std::span<const u8> program_image, const Kernel::CodeSet
         if (auto exclusive = Exclusive{inst}; exclusive.Verify()) {
             curr_patch->m_exclusives.push_back(i);
         }
-#ifdef __WIN32
-        // TODO: keep track of exclusives?
+#ifdef _WIN32
+        // TODO: do we still keep track of exclusives?
         if (auto scratch = CheckForPlatformRegister(inst); scratch) {
             bool pre_buffer = false;
             auto ret = AddRelocations(pre_buffer);
 
             if (pre_buffer) {
-                WritePlatformRegHandler(ret, inst, scratch, c_pre);
+                WritePlatformRegHandler(ret, inst, *scratch, c_pre);
             } else {
-                WritePlatformRegHandler(ret, inst, scratch, c);
+                WritePlatformRegHandler(ret, inst, *scratch, c);
             }
         }
 #endif
@@ -391,17 +391,16 @@ void Patcher::LoadTLS(oaknut::VectorCodeGenerator& cg, oaknut::XReg out) {
     // https://github.com/apple-oss-distributions/xnu/blob/f6217f891ac0bb64f3d375211650a4c1ff8ca1ea/libsyscall/os/tsd.h#L156-L189
     cg.MRS(out, oaknut::SystemReg::TPIDRRO_EL0);
     cg.LDR(out, out, ContextKey * 8);
-#elif __WIN32
+#elif _WIN32
     // x18 always points to TEB in Windows, we can just use that for TLS storage
-    ASSERT(out != X18);
     cg.LDR(out, X18, TlsSlots + 8 * ContextKey);
 #else
     cg.MRS(out, oaknut::SystemReg::TPIDR_EL0);
 #endif
 }
 
-#ifdef __WIN32
-void Patcher::WritePlatformRegHandler(ModuleDestLabel module_dest, uint32 instruction, oaknut::XReg scratch, oaknut::VectorCodeGenerator& code) {
+#ifdef _WIN32
+void Patcher::WritePlatformRegHandler(ModuleDestLabel module_dest, u32 instruction, oaknut::XReg scratch, oaknut::VectorCodeGenerator& cg) {
     // Save X18 register and scratch register
     cg.STP(X18, scratch, SP, PRE_INDEXED, -16);
 
@@ -410,7 +409,7 @@ void Patcher::WritePlatformRegHandler(ModuleDestLabel module_dest, uint32 instru
     cg.LDR(X18, scratch, TlsSlots + 8 * NCEStorage);
 
     // Perform operation
-    cg.append(instruction);
+    cg.dw(instruction);
 
     // Store x18 register and restore scratch register
     cg.STR(X18, scratch, TlsSlots + 8 * NCEStorage);
@@ -540,7 +539,7 @@ void Patcher::WriteSvcTrampoline(ModuleDestLabel module_dest, u32 svc_id, oaknut
     // Reload host TPIDR_EL0 and SP.
     cg.LDP(X2, X3, X1, offsetof(HostContext, host_sp));
     cg.MOV(SP, X2);
-#if !defined(__APPLE__) && !defined(__WIN32)
+#if !defined(__APPLE__) && !defined(_WIN32)
     static_assert(offsetof(HostContext, host_sp) + 8 == offsetof(HostContext, host_tpidr_el0));
     cg.MSR(oaknut::SystemReg::TPIDR_EL0, X3);
 #endif
@@ -608,8 +607,8 @@ void Patcher::WriteSvcTrampoline(ModuleDestLabel module_dest, u32 svc_id, oaknut
 // Retrieve emulated TLS register from GuestContext.
 void Patcher::WriteMrsHandler(ModuleDestLabel module_dest, oaknut::XReg dest_reg,
                               oaknut::SystemReg src_reg, oaknut::VectorCodeGenerator& cg) {
-#ifdef __WIN32
-    if (dest_reg != X18) {
+#ifdef _WIN32
+    if (dest_reg.index() == 18) {
 #endif
     LoadTLS(cg, dest_reg);
 
@@ -618,7 +617,7 @@ void Patcher::WriteMrsHandler(ModuleDestLabel module_dest, oaknut::XReg dest_reg
     } else {
         cg.LDR(dest_reg, dest_reg, offsetof(NativeExecutionParameters, tpidr_el0));
     }
-#ifdef __WIN32
+#ifdef _WIN32
     } else {
         const auto scratch = dest_reg.index() == 0 ? X1 : X0;
         cg.STR(scratch, SP, PRE_INDEXED, -16);
@@ -646,8 +645,8 @@ void Patcher::WriteMsrHandler(ModuleDestLabel module_dest, oaknut::XReg src_reg,
 
     // Save guest value to NativeExecutionParameters::tpidr_el0.
     LoadTLS(cg, scratch_reg);
-#ifdef __WIN32
-    if (src_reg == X18) {
+#ifdef _WIN32
+    if (src_reg.index() == 18) {
         // Load real x18 value and use that
         cg.LDR(scratch_reg2, X18, TlsSlots + 8 * NCEStorage);
         src_reg = scratch_reg2;

src/core/arm/nce/patcher.h

@@ -82,8 +82,8 @@ private:
     void WriteMrsHandler(ModuleDestLabel module_dest, oaknut::XReg dest_reg, oaknut::SystemReg src_reg, oaknut::VectorCodeGenerator& code);
     void WriteMsrHandler(ModuleDestLabel module_dest, oaknut::XReg src_reg, oaknut::VectorCodeGenerator& code);
     void WriteCntpctHandler(ModuleDestLabel module_dest, oaknut::XReg dest_reg, oaknut::VectorCodeGenerator& code);
-#ifdef __WIN32
-    void WritePlatformRegHandler(ModuleDestLabel module_dest, uint32 instruction, oaknut::XReg scratch, oaknut::VectorCodeGenerator& code);
+#ifdef _WIN32
+    void WritePlatformRegHandler(ModuleDestLabel module_dest, u32 instruction, oaknut::XReg scratch, oaknut::VectorCodeGenerator& code);
 #endif
 
     // Convenience wrappers using default code generator

src/core/arm/nce/win/exceptions.cpp

@@ -1,29 +0,0 @@
-// SPDX-FileCopyrightText: Copyright 2026 Eden Emulator Project
-// SPDX-License-Identifier: GPL-3.0-or-later
-
-#include "core/arm/nce/win/exceptions.h"
-
-#include "core/arm/nce/arm_nce.h"
-
-namespace Core {
-
-static s32 WINAPI VectoredExecptionHandler(os::EXCEPTION_POINTERS* info) {
-    u32 code = info->ExceptionRecord->ExceptionCode;
-
-    if (code == SIGSEGV || code == SIGBUS) {
-        GuestMemoryFaultSignalHandler(code, &info->ExceptionRecord->ExceptionAddress, info->ContextRecord);
-        if (is_host_fault) {
-            is_host_fault = false;
-            return EXCEPTION_CONTINUE_SEARCH;
-        }
-        return EXCEPTION_CONTINUE_EXECUTION;
-    } else if (code == SIGUSR2) {
-        ReturnToGuestByExceptionLevelChangeSignalHandler(code, &info->ExceptionRecord->ExceptionAddress, info->ContextRecord);
-        return EXCEPTION_CONTINUE_EXECUTION;
-    } else {
-        // other exception?? let it pass
-        return EXCEPTION_CONTINUE_SEARCH;
-    }
-}
-
-} // namespace Core

src/core/arm/nce/win/exceptions.h

@@ -1,18 +0,0 @@
-// SPDX-FileCopyrightText: Copyright 2026 Eden Emulator Project
-// SPDX-License-Identifier: GPL-3.0-or-later
-
-#pragma once
-
-#include "core/arm/nce/arm_nce.h"
-
-#define SIGBUS EXCEPTION_DATATYPE_MISALIGNMENT
-#define SIGSEGV EXCEPTION_ACCESS_VIOLATION
-#define SIGUSR2 0xE0000001
-
-namespace Core {
-
-thread_local bool is_host_fault = false;
-
-static s32 WINAPI VectoredExecptionHandler(os::EXCEPTION_POINTERS* info);
-
-} // namespace Core

src/core/arm/nce/win/nt_headers.h

@@ -1,14 +0,0 @@
-// SPDX-FileCopyrightText: Copyright 2026 Eden Emulator Project
-// SPDX-License-Identifier: GPL-3.0-or-later
-
-#pragma once
-
-// Use a separate namespace to not add Windows headers to the global path in header files
-// We can't directly add Windows.h because of winternl.h
-namespace os {
-extern "C" {
-    #include <winternl.h>
-    #include <processthreadsapi.h>
-    #include <errhandlingapi.h>
-}
-} // namespace os

src/core/arm/nce/win/platform_visitor.h

@@ -14,7 +14,7 @@ struct XReg;
 
 namespace Core {
 
-std::optional<oaknut::Reg> CheckForPlatformRegister(u32 instruction);
+std::optional<oaknut::XReg> CheckForPlatformRegister(u32 instruction);
 
 class PlatformVisitor final : public VisitorBase {
 public:

src/core/hle/kernel/k_thread.h

@@ -668,7 +668,7 @@ public:
 public:
     // TODO: This shouldn't be defined in kernel namespace
     struct NativeExecutionParameters {
-#if (defined(__APPLE__) || defined(__WIN32)) && HAS_NCE
+#if (defined(__APPLE__) || defined(_WIN32)) && HAS_NCE
         // Are we in actual guest code?
         bool is_actually_running{};
 #endif