Experimental 64-bit host support

CMakeLists.txt

@@ -1,10 +1,5 @@
 cmake_minimum_required(VERSION 3.13)
 
-set(CMAKE_C_FLAGS_INIT "-m32")
-set(CMAKE_CXX_FLAGS_INIT "-m32 -fno-exceptions -fno-rtti")
-set(CMAKE_EXE_LINKER_FLAGS_INIT "-m32")
-set(CMAKE_SHARED_LINKER_FLAGS_INIT "-m32")
-
 project(wibo LANGUAGES ASM C CXX)
 
 set(WIBO_VERSION "" CACHE STRING "Version string for the wibo binary; if empty, attempts to use git describe")
@@ -46,11 +41,21 @@ configure_file(
     @ONLY
 )
 
+option(WIBO_64 "Build wibo for 64-bit host" OFF)
 option(WIBO_ENABLE_FIXTURE_TESTS "Enable Win32 fixture tests (requires i686-w64-mingw32)" ON)
 option(WIBO_ENABLE_LIBURING "Enable liburing for asynchronous I/O" OFF)
 set(WIBO_ENABLE_LTO "AUTO" CACHE STRING "Enable link-time optimization (LTO)")
 set_property(CACHE WIBO_ENABLE_LTO PROPERTY STRINGS "AUTO" "ON" "OFF")
 
+add_compile_options(
+    $<IF:$<BOOL:${WIBO_64}>,-m64,-m32>
+    $<$<COMPILE_LANGUAGE:CXX>:-fno-exceptions>
+    $<$<COMPILE_LANGUAGE:CXX>:-fno-rtti>
+)
+add_link_options(
+    $<IF:$<BOOL:${WIBO_64}>,-m64,-m32>
+)
+
 if (WIBO_ENABLE_LTO STREQUAL "AUTO")
     if (CMAKE_BUILD_TYPE STREQUAL "Release" OR CMAKE_BUILD_TYPE STREQUAL "MinSizeRel")
         include(CheckIPOSupported)
@@ -88,6 +93,10 @@ FetchContent_MakeAvailable(mimalloc)
 # Disable `note: the alignment of '_Atomic long long int' fields changed in GCC 11.1`
 target_compile_options(mimalloc-obj PRIVATE -Wno-psabi)
 target_compile_definitions(mimalloc-obj PRIVATE MI_USE_BUILTIN_THREAD_POINTER=1)
+if (WIBO_64)
+    # Disable a noisy warning on startup
+    target_compile_definitions(mimalloc-obj PRIVATE MI_DEBUG=0)
+endif()
 
 if (WIBO_ENABLE_LIBURING)
     FetchContent_Declare(
@@ -126,12 +135,12 @@ endif()
 
 add_executable(wibo
     dll/advapi32.cpp
+    dll/advapi32/md5.c
     dll/advapi32/processthreadsapi.cpp
     dll/advapi32/securitybaseapi.cpp
     dll/advapi32/winbase.cpp
     dll/advapi32/wincrypt.cpp
     dll/advapi32/winreg.cpp
-    dll/advapi32/md5.c
     dll/bcrypt.cpp
     dll/crt.cpp
     dll/kernel32.cpp
@@ -144,8 +153,8 @@ add_executable(wibo
     dll/kernel32/interlockedapi.cpp
     dll/kernel32/ioapiset.cpp
     dll/kernel32/libloaderapi.cpp
-    dll/kernel32/namedpipeapi.cpp
     dll/kernel32/memoryapi.cpp
+    dll/kernel32/namedpipeapi.cpp
     dll/kernel32/processenv.cpp
     dll/kernel32/processthreadsapi.cpp
     dll/kernel32/profileapi.cpp
@@ -162,8 +171,8 @@ add_executable(wibo
     dll/mscoree.cpp
     dll/msvcrt.cpp
     dll/ntdll.cpp
-    dll/rpcrt4.cpp
     dll/ole32.cpp
+    dll/rpcrt4.cpp
     dll/user32.cpp
     dll/vcruntime.cpp
     dll/version.cpp
@@ -173,21 +182,22 @@ add_executable(wibo
     src/errors.cpp
     src/files.cpp
     src/handles.cpp
-    src/loader.cpp
     src/heap.cpp
+    src/loader.cpp
     src/main.cpp
     src/modules.cpp
     src/processes.cpp
     src/resources.cpp
+    src/setup.S
     src/strutil.cpp
     src/tls.cpp
 )
 target_compile_definitions(wibo PRIVATE _GNU_SOURCE _FILE_OFFSET_BITS=64 _TIME_BITS=64)
 target_compile_features(wibo PRIVATE cxx_std_20)
 target_compile_options(wibo PRIVATE -Wall -Wextra)
-target_link_options(wibo PRIVATE -no-pie -Wl,--image-base=0x90000000)
+target_link_options(wibo PRIVATE -no-pie -Wl,--image-base=0x70000000)
 if (CMAKE_CXX_COMPILER_ID STREQUAL "GNU")
-    target_compile_options(wibo PRIVATE -fno-pie -maccumulate-outgoing-args)
+    target_compile_options(wibo PRIVATE -maccumulate-outgoing-args)
     target_link_options(wibo PRIVATE -maccumulate-outgoing-args)
 endif()
 target_include_directories(wibo PRIVATE dll src ${WIBO_GENERATED_HEADER_DIR})
@@ -242,15 +252,17 @@ function(wibo_codegen_module)
                 --dll ${module_NAME}
                 --headers ${module_HEADERS}
                 --namespace ${module_NAME}
-                --arch x86
+                --arch $<IF:$<BOOL:${WIBO_64}>,x86_64,x86>
                 --out-asm ${out_asm}
                 --out-hdr ${out_hdr}
                 -I ${CMAKE_CURRENT_SOURCE_DIR}
                 -I ${CMAKE_CURRENT_SOURCE_DIR}/dll
                 -I ${CMAKE_CURRENT_SOURCE_DIR}/src
-        DEPENDS ${CMAKE_CURRENT_SOURCE_DIR}/tools/gen_trampolines.py ${module_HEADERS})
+        DEPENDS ${CMAKE_CURRENT_SOURCE_DIR}/tools/gen_trampolines.py
+                ${module_HEADERS}
+    )
     target_sources(wibo PRIVATE ${out_asm})
-    set_source_files_properties(${out_asm} PROPERTIES COMPILE_FLAGS "-m32")
+    set_source_files_properties(${out_asm} PROPERTIES COMPILE_FLAGS $<IF:$<BOOL:${WIBO_64}>,-m64,-m32>)
 endfunction()
 
 wibo_codegen_module(NAME advapi32 HEADERS

CMakePresets.json

@@ -53,6 +53,46 @@
       "cacheVariables": {
         "CMAKE_BUILD_TYPE": "Release"
       }
+    },
+    {
+      "name": "debug64",
+      "displayName": "Debug (64-bit)",
+      "inherits": ["ninja-base"],
+      "binaryDir": "${sourceDir}/build/debug64",
+      "cacheVariables": {
+        "CMAKE_BUILD_TYPE": "Debug",
+        "WIBO_64": "ON"
+      }
+    },
+    {
+      "name": "release64",
+      "displayName": "Release (64-bit)",
+      "inherits": ["ninja-base"],
+      "binaryDir": "${sourceDir}/build/release64",
+      "cacheVariables": {
+        "CMAKE_BUILD_TYPE": "Release",
+        "WIBO_64": "ON"
+      }
+    },
+    {
+      "name": "debug64-clang",
+      "displayName": "Debug (64-bit, Clang)",
+      "inherits": ["ninja-base", "clang-base"],
+      "binaryDir": "${sourceDir}/build/debug64-clang",
+      "cacheVariables": {
+        "CMAKE_BUILD_TYPE": "Debug",
+        "WIBO_64": "ON"
+      }
+    },
+    {
+      "name": "release64-clang",
+      "displayName": "Release (64-bit, Clang)",
+      "inherits": ["ninja-base", "clang-base"],
+      "binaryDir": "${sourceDir}/build/release64-clang",
+      "cacheVariables": {
+        "CMAKE_BUILD_TYPE": "Release",
+        "WIBO_64": "ON"
+      }
     }
   ],
   "buildPresets": [
@@ -83,6 +123,34 @@
       "configurePreset": "release-clang",
       "targets": ["wibo", "wibo_test_fixtures"],
       "configuration": "Release"
+    },
+    {
+      "name": "debug64",
+      "displayName": "Build (Debug, 64-bit)",
+      "configurePreset": "debug64",
+      "targets": ["wibo", "wibo_test_fixtures"],
+      "configuration": "Debug"
+    },
+    {
+      "name": "release64",
+      "displayName": "Build (Release, 64-bit)",
+      "configurePreset": "release64",
+      "targets": ["wibo", "wibo_test_fixtures"],
+      "configuration": "Release"
+    },
+    {
+      "name": "debug64-clang",
+      "displayName": "Build (Debug, 64-bit, Clang)",
+      "configurePreset": "debug64-clang",
+      "targets": ["wibo", "wibo_test_fixtures"],
+      "configuration": "Debug"
+    },
+    {
+      "name": "release64-clang",
+      "displayName": "Build (Release, 64-bit, Clang)",
+      "configurePreset": "release64-clang",
+      "targets": ["wibo", "wibo_test_fixtures"],
+      "configuration": "Release"
     }
   ],
   "testPresets": [
@@ -125,6 +193,46 @@
         "outputOnFailure": true,
         "shortProgress": true
       }
+    },
+    {
+      "name": "debug64",
+      "displayName": "Run tests (Debug, 64-bit)",
+      "configurePreset": "debug64",
+      "configuration": "Debug",
+      "output": {
+        "outputOnFailure": true,
+        "shortProgress": true
+      }
+    },
+    {
+      "name": "release64",
+      "displayName": "Run tests (Release, 64-bit)",
+      "configurePreset": "release64",
+      "configuration": "Release",
+      "output": {
+        "outputOnFailure": true,
+        "shortProgress": true
+      }
+    },
+    {
+      "name": "debug64-clang",
+      "displayName": "Run tests (Debug, 64-bit, Clang)",
+      "configurePreset": "debug64-clang",
+      "configuration": "Debug",
+      "output": {
+        "outputOnFailure": true,
+        "shortProgress": true
+      }
+    },
+    {
+      "name": "release64-clang",
+      "displayName": "Run tests (Release, 64-bit, Clang)",
+      "configurePreset": "release64-clang",
+      "configuration": "Release",
+      "output": {
+        "outputOnFailure": true,
+        "shortProgress": true
+      }
     }
   ]
 }

dll/advapi32/internal.h

@@ -6,8 +6,6 @@
 
 constexpr DWORD SECURITY_LOCAL_SYSTEM_RID = 18;
 
-constexpr BYTE kNtAuthority[6] = {0, 0, 0, 0, 0, 5};
-
 struct TokenObject : ObjectBase {
 	static constexpr ObjectType kType = ObjectType::Token;
 

dll/advapi32/securitybaseapi.cpp

@@ -6,6 +6,7 @@
 #include "handles.h"
 #include "internal.h"
 #include "kernel32/internal.h"
+#include "types.h"
 
 #include <algorithm>
 #include <cstring>
@@ -13,6 +14,8 @@
 
 namespace {
 
+constexpr BYTE kNtAuthority[6] = {0, 0, 0, 0, 0, 5};
+
 constexpr size_t kAceAlignment = 4;
 constexpr DWORD ERROR_REVISION_MISMATCH = 1306;
 constexpr DWORD ERROR_INVALID_ACL = 1336;
@@ -202,14 +205,14 @@ BOOL WINAPI AddAccessAllowedAce(PACL pAcl, DWORD dwAceRevision, DWORD AccessMask
 	return TRUE;
 }
 
-BOOL WINAPI FindFirstFreeAce(PACL pAcl, LPVOID *pAce) {
+BOOL WINAPI FindFirstFreeAce(PACL pAcl, GUEST_PTR *pAce) {
 	HOST_CONTEXT_GUARD();
 	DEBUG_LOG("FindFirstFreeAce(%p, %p)\n", pAcl, pAce);
 	if (!pAce) {
 		kernel32::setLastError(ERROR_INVALID_PARAMETER);
 		return FALSE;
 	}
-	*pAce = nullptr;
+	*pAce = GUEST_NULL;
 	if (!pAcl) {
 		kernel32::setLastError(ERROR_INVALID_PARAMETER);
 		return FALSE;
@@ -220,12 +223,12 @@ BOOL WINAPI FindFirstFreeAce(PACL pAcl, LPVOID *pAce) {
 		kernel32::setLastError(ERROR_INVALID_ACL);
 		return FALSE;
 	}
-	*pAce = reinterpret_cast<BYTE *>(pAcl) + used;
+	*pAce = toGuestPtr(reinterpret_cast<BYTE *>(pAcl) + used);
 	pAcl->AclSize = static_cast<WORD>(std::max<size_t>(pAcl->AclSize, used));
 	return TRUE;
 }
 
-BOOL WINAPI GetSecurityDescriptorDacl(PSECURITY_DESCRIPTOR pSecurityDescriptor, LPBOOL lpbDaclPresent, PACL *pDacl,
+BOOL WINAPI GetSecurityDescriptorDacl(PSECURITY_DESCRIPTOR pSecurityDescriptor, LPBOOL lpbDaclPresent, GUEST_PTR *pDacl,
 									  LPBOOL lpbDaclDefaulted) {
 	HOST_CONTEXT_GUARD();
 	DEBUG_LOG("GetSecurityDescriptorDacl(%p, %p, %p, %p)\n", pSecurityDescriptor, lpbDaclPresent, pDacl,
@@ -246,7 +249,7 @@ BOOL WINAPI GetSecurityDescriptorDacl(PSECURITY_DESCRIPTOR pSecurityDescriptor,
 	*lpbDaclPresent = hasDacl;
 	if (!hasDacl) {
 		if (pDacl) {
-			*pDacl = nullptr;
+			*pDacl = GUEST_NULL;
 		}
 		if (lpbDaclDefaulted) {
 			*lpbDaclDefaulted = FALSE;
@@ -425,10 +428,10 @@ BOOL WINAPI InitializeSecurityDescriptor(PSECURITY_DESCRIPTOR pSecurityDescripto
 	pSecurityDescriptor->Revision = static_cast<BYTE>(dwRevision);
 	pSecurityDescriptor->Sbz1 = 0;
 	pSecurityDescriptor->Control = 0;
-	pSecurityDescriptor->Owner = nullptr;
-	pSecurityDescriptor->Group = nullptr;
-	pSecurityDescriptor->Sacl = nullptr;
-	pSecurityDescriptor->Dacl = nullptr;
+	pSecurityDescriptor->Owner = GUEST_NULL;
+	pSecurityDescriptor->Group = GUEST_NULL;
+	pSecurityDescriptor->Sacl = GUEST_NULL;
+	pSecurityDescriptor->Dacl = GUEST_NULL;
 	return TRUE;
 }
 
@@ -446,9 +449,9 @@ BOOL WINAPI SetSecurityDescriptorDacl(PSECURITY_DESCRIPTOR pSecurityDescriptor,
 		if (bDaclDefaulted) {
 			control = static_cast<WORD>(control | SE_DACL_DEFAULTED);
 		}
-		pSecurityDescriptor->Dacl = pDacl;
+		pSecurityDescriptor->Dacl = toGuestPtr(pDacl);
 	} else {
-		pSecurityDescriptor->Dacl = nullptr;
+		pSecurityDescriptor->Dacl = GUEST_NULL;
 	}
 	pSecurityDescriptor->Control = control;
 	return TRUE;

dll/advapi32/securitybaseapi.h

@@ -30,10 +30,10 @@ struct SECURITY_DESCRIPTOR {
 	BYTE Revision;
 	BYTE Sbz1;
 	WORD Control;
-	void *Owner;
-	void *Group;
-	ACL *Sacl;
-	ACL *Dacl;
+	GUEST_PTR Owner;
+	GUEST_PTR Group;
+	GUEST_PTR Sacl;
+	GUEST_PTR Dacl;
 };
 
 using PSECURITY_DESCRIPTOR = SECURITY_DESCRIPTOR *;
@@ -85,7 +85,7 @@ namespace advapi32 {
 
 BOOL WINAPI InitializeAcl(PACL pAcl, DWORD nAclLength, DWORD dwAclRevision);
 BOOL WINAPI AddAccessAllowedAce(PACL pAcl, DWORD dwAceRevision, DWORD AccessMask, PSID pSid);
-BOOL WINAPI FindFirstFreeAce(PACL pAcl, LPVOID *pAce);
+BOOL WINAPI FindFirstFreeAce(PACL pAcl, GUEST_PTR *pAce);
 PSID_IDENTIFIER_AUTHORITY WINAPI GetSidIdentifierAuthority(PSID pSid);
 PUCHAR WINAPI GetSidSubAuthorityCount(PSID pSid);
 PDWORD WINAPI GetSidSubAuthority(PSID pSid, DWORD nSubAuthority);
@@ -95,7 +95,7 @@ BOOL WINAPI DuplicateTokenEx(HANDLE hExistingToken, DWORD dwDesiredAccess, void
 BOOL WINAPI CopySid(DWORD nDestinationSidLength, PSID pDestinationSid, PSID pSourceSid);
 BOOL WINAPI InitializeSid(PSID sid, PSID_IDENTIFIER_AUTHORITY pIdentifierAuthority, BYTE nSubAuthorityCount);
 BOOL WINAPI EqualSid(PSID pSid1, PSID pSid2);
-BOOL WINAPI GetSecurityDescriptorDacl(PSECURITY_DESCRIPTOR pSecurityDescriptor, LPBOOL lpbDaclPresent, PACL *pDacl,
+BOOL WINAPI GetSecurityDescriptorDacl(PSECURITY_DESCRIPTOR pSecurityDescriptor, LPBOOL lpbDaclPresent, GUEST_PTR *pDacl,
 									  LPBOOL lpbDaclDefaulted);
 BOOL WINAPI SetKernelObjectSecurity(HANDLE Handle, SECURITY_INFORMATION SecurityInformation,
 									PSECURITY_DESCRIPTOR SecurityDescriptor);

dll/advapi32/winbase.cpp

@@ -15,6 +15,7 @@ namespace {
 
 constexpr WCHAR kAccountSystem[] = {u'S', u'Y', u'S', u'T', u'E', u'M', u'\0'};
 constexpr WCHAR kDomainNtAuthority[] = {u'N', u'T', u' ', u'A', u'U', u'T', u'H', u'O', u'R', u'I', u'T', u'Y', u'\0'};
+constexpr BYTE kNtAuthority[6] = {0, 0, 0, 0, 0, 5};
 
 std::mutex g_privilegeMapMutex;
 std::unordered_map<std::string, LUID> g_privilegeLuidCache;

dll/advapi32/winbase.h

@@ -17,8 +17,8 @@ enum SID_NAME_USE : DWORD {
 
 namespace advapi32 {
 
-BOOL WINAPI LookupAccountSidW(LPCWSTR lpSystemName, PSID Sid, LPWSTR Name, LPDWORD cchName,
-								LPWSTR ReferencedDomainName, LPDWORD cchReferencedDomainName, SID_NAME_USE *peUse);
+BOOL WINAPI LookupAccountSidW(LPCWSTR lpSystemName, PSID Sid, LPWSTR Name, LPDWORD cchName, LPWSTR ReferencedDomainName,
+							  LPDWORD cchReferencedDomainName, SID_NAME_USE *peUse);
 BOOL WINAPI LookupPrivilegeValueA(LPCSTR lpSystemName, LPCSTR lpName, PLUID lpLuid);
 BOOL WINAPI LookupPrivilegeValueW(LPCWSTR lpSystemName, LPCWSTR lpName, PLUID lpLuid);
 BOOL WINAPI GetUserNameA(LPSTR lpBuffer, LPDWORD pcbBuffer);

dll/advapi32/winreg.cpp

@@ -29,17 +29,17 @@ struct RegistryKeyObject : ObjectBase {
 };
 
 struct PredefinedKeyInfo {
-	uintptr_t value;
+	HKEY value;
 	const char16_t *name;
 };
 
 constexpr PredefinedKeyInfo kPredefinedKeyInfos[] = {
-	{static_cast<uintptr_t>(0x80000000u), u"HKEY_CLASSES_ROOT"},
-	{static_cast<uintptr_t>(0x80000001u), u"HKEY_CURRENT_USER"},
-	{static_cast<uintptr_t>(0x80000002u), u"HKEY_LOCAL_MACHINE"},
-	{static_cast<uintptr_t>(0x80000003u), u"HKEY_USERS"},
-	{static_cast<uintptr_t>(0x80000004u), u"HKEY_PERFORMANCE_DATA"},
-	{static_cast<uintptr_t>(0x80000005u), u"HKEY_CURRENT_CONFIG"},
+	{static_cast<HKEY>(0x80000000u), u"HKEY_CLASSES_ROOT"},
+	{static_cast<HKEY>(0x80000001u), u"HKEY_CURRENT_USER"},
+	{static_cast<HKEY>(0x80000002u), u"HKEY_LOCAL_MACHINE"},
+	{static_cast<HKEY>(0x80000003u), u"HKEY_USERS"},
+	{static_cast<HKEY>(0x80000004u), u"HKEY_PERFORMANCE_DATA"},
+	{static_cast<HKEY>(0x80000005u), u"HKEY_CURRENT_CONFIG"},
 };
 
 constexpr size_t kPredefinedKeyCount = std::size(kPredefinedKeyInfos);
@@ -82,7 +82,7 @@ std::u16string canonicalizeKeySegment(LPCWSTR input) {
 	return canonicalizeKeySegment(wide);
 }
 
-Pin<RegistryKeyObject> predefinedHandleForValue(uintptr_t value) {
+Pin<RegistryKeyObject> predefinedHandleForValue(HKEY value) {
 	static std::array<Pin<RegistryKeyObject>, kPredefinedKeyCount> g_predefinedHandles = [] {
 		std::array<Pin<RegistryKeyObject>, kPredefinedKeyCount> arr;
 		for (size_t i = 0; i < kPredefinedKeyCount; ++i) {
@@ -101,11 +101,10 @@ Pin<RegistryKeyObject> predefinedHandleForValue(uintptr_t value) {
 }
 
 Pin<RegistryKeyObject> handleDataFromHKeyLocked(HKEY hKey) {
-	uintptr_t raw = reinterpret_cast<uintptr_t>(hKey);
-	if (raw == 0) {
+	if (hKey == NO_HANDLE) {
 		return {};
 	}
-	if (auto predefined = predefinedHandleForValue(raw)) {
+	if (auto predefined = predefinedHandleForValue(hKey)) {
 		return predefined;
 	}
 	auto obj = wibo::handles().getAs<RegistryKeyObject>(hKey);
@@ -116,9 +115,8 @@ Pin<RegistryKeyObject> handleDataFromHKeyLocked(HKEY hKey) {
 }
 
 bool isPredefinedKeyHandle(HKEY hKey) {
-	uintptr_t raw = reinterpret_cast<uintptr_t>(hKey);
 	return std::any_of(std::begin(kPredefinedKeyInfos), std::end(kPredefinedKeyInfos),
-					   [raw](const PredefinedKeyInfo &info) { return info.value == raw; });
+					   [hKey](const PredefinedKeyInfo &info) { return info.value == hKey; });
 }
 
 } // namespace
@@ -139,7 +137,7 @@ LSTATUS WINAPI RegCreateKeyExW(HKEY hKey, LPCWSTR lpSubKey, DWORD Reserved, LPWS
 		kernel32::setLastError(ERROR_INVALID_PARAMETER);
 		return ERROR_INVALID_PARAMETER;
 	}
-	*phkResult = nullptr;
+	*phkResult = NO_HANDLE;
 	if (Reserved != 0) {
 		kernel32::setLastError(ERROR_INVALID_PARAMETER);
 		return ERROR_INVALID_PARAMETER;
@@ -220,7 +218,7 @@ LSTATUS WINAPI RegOpenKeyExW(HKEY hKey, LPCWSTR lpSubKey, DWORD ulOptions, REGSA
 		kernel32::setLastError(ERROR_INVALID_PARAMETER);
 		return ERROR_INVALID_PARAMETER;
 	}
-	*phkResult = nullptr;
+	*phkResult = NO_HANDLE;
 	if ((ulOptions & ~REG_OPTION_OPEN_LINK) != 0) {
 		kernel32::setLastError(ERROR_INVALID_PARAMETER);
 		return ERROR_INVALID_PARAMETER;

dll/advapi32/winreg.h

@@ -34,11 +34,9 @@ constexpr REGSAM KEY_WOW64_32KEY = 0x00000200;
 namespace advapi32 {
 
 LSTATUS WINAPI RegCreateKeyExA(HKEY hKey, LPCSTR lpSubKey, DWORD Reserved, LPSTR lpClass, DWORD dwOptions,
-								 REGSAM samDesired, void *lpSecurityAttributes, PHKEY phkResult,
-								 LPDWORD lpdwDisposition);
+							   REGSAM samDesired, void *lpSecurityAttributes, PHKEY phkResult, LPDWORD lpdwDisposition);
 LSTATUS WINAPI RegCreateKeyExW(HKEY hKey, LPCWSTR lpSubKey, DWORD Reserved, LPWSTR lpClass, DWORD dwOptions,
-								 REGSAM samDesired, void *lpSecurityAttributes, PHKEY phkResult,
-								 LPDWORD lpdwDisposition);
+							   REGSAM samDesired, void *lpSecurityAttributes, PHKEY phkResult, LPDWORD lpdwDisposition);
 LSTATUS WINAPI RegOpenKeyExA(HKEY hKey, LPCSTR lpSubKey, DWORD ulOptions, REGSAM samDesired, PHKEY phkResult);
 LSTATUS WINAPI RegOpenKeyExW(HKEY hKey, LPCWSTR lpSubKey, DWORD ulOptions, REGSAM samDesired, PHKEY phkResult);
 LSTATUS WINAPI RegQueryValueExA(HKEY hKey, LPCSTR lpValueName, LPDWORD lpReserved, LPDWORD lpType, BYTE *lpData,

dll/crt.cpp

@@ -6,6 +6,7 @@
 #include "heap.h"
 #include "kernel32/internal.h"
 #include "modules.h"
+#include "types.h"
 
 #include <csignal>
 #include <cstdarg>
@@ -42,24 +43,29 @@ int _fmode = 0;
 std::vector<_PVFV> atexitFuncs;
 _invalid_parameter_handler invalidParameterHandler = nullptr;
 
-void CDECL _initterm(const _PVFV *ppfn, const _PVFV *end) {
+GUEST_PTR guestArgv = GUEST_NULL;
+GUEST_PTR guestEnviron = GUEST_NULL;
+
+void CDECL _initterm(GUEST_PTR *ppfn, GUEST_PTR *end) {
 	HOST_CONTEXT_GUARD();
 	DEBUG_LOG("_initterm(%p, %p)\n", ppfn, end);
 	do {
-		if (_PVFV pfn = *++ppfn) {
+		if (GUEST_PTR pfn = *++ppfn) {
 			DEBUG_LOG("-> calling %p\n", pfn);
-			call__PVFV(pfn);
+			auto fn = reinterpret_cast<_PVFV>(fromGuestPtr(pfn));
+			call__PVFV(fn);
 		}
 	} while (ppfn < end);
 }
 
-int CDECL _initterm_e(const _PIFV *ppfn, const _PIFV *end) {
+int CDECL _initterm_e(GUEST_PTR *ppfn, GUEST_PTR *end) {
 	HOST_CONTEXT_GUARD();
 	DEBUG_LOG("_initterm_e(%p, %p)\n", ppfn, end);
 	do {
-		if (_PIFV pfn = *++ppfn) {
+		if (GUEST_PTR pfn = *++ppfn) {
 			DEBUG_LOG("-> calling %p\n", pfn);
-			int err = call__PIFV(pfn);
+			auto fn = reinterpret_cast<_PIFV>(fromGuestPtr(pfn));
+			int err = call__PIFV(fn);
 			if (err)
 				return err;
 		}
@@ -80,16 +86,16 @@ int CDECL _set_fmode(int mode) {
 	return 0;
 }
 
-int *CDECL __p__commode() {
+GUEST_PTR CDECL __p__commode() {
 	HOST_CONTEXT_GUARD();
 	DEBUG_LOG("__p__commode()\n");
-	return &_commode;
+	return toGuestPtr(&_commode);
 }
 
-int *CDECL __p__fmode() {
+GUEST_PTR CDECL __p__fmode() {
 	HOST_CONTEXT_GUARD();
 	DEBUG_LOG("__p__fmode()\n");
-	return &_fmode;
+	return toGuestPtr(&_fmode);
 }
 
 int CDECL _crt_atexit(void (*func)()) {
@@ -137,28 +143,79 @@ int CDECL _set_new_mode(int newhandlermode) {
 	return 0;
 }
 
-char **CDECL _get_initial_narrow_environment() {
+GUEST_PTR CDECL _get_initial_narrow_environment() {
 	HOST_CONTEXT_GUARD();
 	DEBUG_LOG("_get_initial_narrow_environment()\n");
-	return environ;
+	if (guestEnviron == GUEST_NULL) {
+		int count = 0;
+		while (environ[count]) {
+			count++;
+		}
+		GUEST_PTR *buf = reinterpret_cast<GUEST_PTR *>(wibo::heap::guestMalloc(count * sizeof(GUEST_PTR)));
+		if (!buf) {
+			return GUEST_NULL;
+		}
+		for (int i = 0; i < count; i++) {
+			size_t len = ::strlen(environ[i]);
+			char *str = reinterpret_cast<char *>(wibo::heap::guestMalloc(len + 1));
+			::memcpy(str, environ[i], len + 1);
+			buf[i] = toGuestPtr(str);
+		}
+		guestEnviron = toGuestPtr(buf);
+	}
+	return guestEnviron;
 }
 
-char ***CDECL __p__environ() {
+GUEST_PTR CDECL __p__environ() {
 	HOST_CONTEXT_GUARD();
 	DEBUG_LOG("__p__environ()\n");
-	return &environ;
+	if (guestEnviron == GUEST_NULL) {
+		int count = 0;
+		while (environ[count]) {
+			count++;
+		}
+		GUEST_PTR *buf = reinterpret_cast<GUEST_PTR *>(wibo::heap::guestMalloc(count * sizeof(GUEST_PTR)));
+		if (!buf) {
+			return GUEST_NULL;
+		}
+		for (int i = 0; i < count; i++) {
+			size_t len = ::strlen(environ[i]);
+			char *str = reinterpret_cast<char *>(wibo::heap::guestMalloc(len + 1));
+			::memcpy(str, environ[i], len + 1);
+			buf[i] = toGuestPtr(str);
+		}
+		guestEnviron = toGuestPtr(buf);
+	}
+	return toGuestPtr(&environ);
 }
 
-char ***CDECL __p___argv() {
+GUEST_PTR CDECL __p___argv() {
 	HOST_CONTEXT_GUARD();
 	DEBUG_LOG("__p___argv()\n");
-	return &wibo::argv;
+	if (guestArgv == GUEST_NULL) {
+		int count = 0;
+		while (wibo::argv[count]) {
+			count++;
+		}
+		GUEST_PTR *buf = reinterpret_cast<GUEST_PTR *>(wibo::heap::guestMalloc(count * sizeof(GUEST_PTR)));
+		if (!buf) {
+			return GUEST_NULL;
+		}
+		for (int i = 0; i < count; i++) {
+			size_t len = ::strlen(wibo::argv[i]);
+			char *str = reinterpret_cast<char *>(wibo::heap::guestMalloc(len + 1));
+			::memcpy(str, wibo::argv[i], len + 1);
+			buf[i] = toGuestPtr(str);
+		}
+		guestArgv = toGuestPtr(buf);
+	}
+	return toGuestPtr(&guestArgv);
 }
 
-int *CDECL __p___argc() {
+GUEST_PTR CDECL __p___argc() {
 	HOST_CONTEXT_GUARD();
 	DEBUG_LOG("__p___argc()\n");
-	return &wibo::argc;
+	return toGuestPtr(&wibo::argc);
 }
 
 SIZE_T CDECL strlen(const char *str) {
@@ -258,9 +315,9 @@ int CDECL _initialize_onexit_table(_onexit_table_t *table) {
 		return -1;
 	if (table->first != table->last)
 		return 0;
-	table->first = nullptr;
-	table->last = nullptr;
-	table->end = nullptr;
+	table->first = GUEST_NULL;
+	table->last = GUEST_NULL;
+	table->end = GUEST_NULL;
 	return 0;
 }
 
@@ -269,20 +326,23 @@ int CDECL _register_onexit_function(_onexit_table_t *table, _onexit_t func) {
 	DEBUG_LOG("_register_onexit_function(%p, %p)\n", table, func);
 	if (!table || !func)
 		return -1;
-	if (table->last == table->end) {
-		size_t count = table->end - table->first;
+	GUEST_PTR *first = reinterpret_cast<GUEST_PTR *>(fromGuestPtr(table->first));
+	GUEST_PTR *last = reinterpret_cast<GUEST_PTR *>(fromGuestPtr(table->last));
+	GUEST_PTR *end = reinterpret_cast<GUEST_PTR *>(fromGuestPtr(table->end));
+	if (last == end) {
+		size_t count = end - first;
 		size_t newCount = count + 1;
 		if (newCount <= 0)
 			return -1;
-		_onexit_t *newTable =
-			static_cast<_onexit_t *>(wibo::heap::guestRealloc(table->first, newCount * sizeof(_onexit_t)));
+		GUEST_PTR *newTable = static_cast<GUEST_PTR *>(wibo::heap::guestRealloc(first, newCount * sizeof(GUEST_PTR)));
 		if (!newTable)
 			return -1;
-		table->first = newTable;
-		table->last = newTable + count;
-		table->end = newTable + newCount;
+		table->first = toGuestPtr(newTable);
+		last = newTable + count;
+		table->end = toGuestPtr(newTable + newCount);
 	}
-	*table->last++ = func;
+	*last = toGuestPtr(reinterpret_cast<void *>(func));
+	table->last = toGuestPtr(last + 1);
 	return 0;
 }
 
@@ -291,9 +351,12 @@ int CDECL _execute_onexit_table(_onexit_table_t *table) {
 	DEBUG_LOG("_execute_onexit_table(%p)\n", table);
 	if (!table)
 		return -1;
-	for (auto it = table->first; it != table->last; ++it) {
-		DEBUG_LOG("Calling onexit_table function %p\n", *it);
-		call__onexit_t(*it);
+	GUEST_PTR *first = reinterpret_cast<GUEST_PTR *>(fromGuestPtr(table->first));
+	GUEST_PTR *last = reinterpret_cast<GUEST_PTR *>(fromGuestPtr(table->last));
+	for (auto it = first; it != last; ++it) {
+		_onexit_t fn = reinterpret_cast<_onexit_t>(fromGuestPtr(*it));
+		DEBUG_LOG("Calling onexit_table function %p\n", fn);
+		call__onexit_t(fn);
 	}
 	return 0;
 }
@@ -345,7 +408,7 @@ void *CDECL __acrt_iob_func(unsigned int index) {
 	return nullptr;
 }
 
-int CDECL_NO_CONV __stdio_common_vfprintf(unsigned long long options, _FILE *stream, const char *format, void *locale,
+int CDECL_NO_CONV __stdio_common_vfprintf(ULONGLONG options, _FILE *stream, const char *format, void *locale,
 										  va_list args) {
 	HOST_CONTEXT_GUARD();
 	DEBUG_LOG("__stdio_common_vfprintf(%llu, %p, %s, %p, %p)\n", options, stream, format, locale, args);
@@ -355,8 +418,8 @@ int CDECL_NO_CONV __stdio_common_vfprintf(unsigned long long options, _FILE *str
 	return vfprintf(hostFile, format, args);
 }
 
-int CDECL_NO_CONV __stdio_common_vsprintf(unsigned long long options, char *buffer, SIZE_T len, const char *format,
-										  void *locale, va_list args) {
+int CDECL_NO_CONV __stdio_common_vsprintf(ULONGLONG options, char *buffer, SIZE_T len, const char *format, void *locale,
+										  va_list args) {
 	HOST_CONTEXT_GUARD();
 	DEBUG_LOG("__stdio_common_vsprintf(%llu, %p, %zu, %s, %p, ...)\n", options, buffer, len, format, locale);
 	if (!buffer || !format)

dll/crt.h

@@ -23,9 +23,9 @@ typedef int(_CC_CDECL *sort_compare)(const void *, const void *);
 typedef int(_CC_CDECL *_onexit_t)();
 
 struct _onexit_table_t {
-	_onexit_t *first;
-	_onexit_t *last;
-	_onexit_t *end;
+	GUEST_PTR first;
+	GUEST_PTR last;
+	GUEST_PTR end;
 };
 
 namespace crt {
@@ -33,23 +33,23 @@ namespace crt {
 extern int _commode;
 extern int _fmode;
 
-void CDECL _initterm(const _PVFV *ppfn, const _PVFV *end);
-int CDECL _initterm_e(const _PIFV *ppfn, const _PIFV *end);
+void CDECL _initterm(GUEST_PTR *ppfn, GUEST_PTR *end);
+int CDECL _initterm_e(GUEST_PTR *ppfn, GUEST_PTR *end);
 void CDECL _set_app_type(_crt_app_type type);
 int CDECL _set_fmode(int mode);
-int *CDECL __p__commode();
-int *CDECL __p__fmode();
-int CDECL _crt_atexit(void (*func)());
+GUEST_PTR CDECL __p__commode();
+GUEST_PTR CDECL __p__fmode();
+int CDECL _crt_atexit(_PVFV func);
 int CDECL _configure_narrow_argv(_crt_argv_mode mode);
 _invalid_parameter_handler CDECL _set_invalid_parameter_handler(_invalid_parameter_handler newHandler);
 int CDECL _controlfp_s(unsigned int *currentControl, unsigned int newControl, unsigned int mask);
 int CDECL _configthreadlocale(int per_thread_locale_type);
 int CDECL _initialize_narrow_environment();
 int CDECL _set_new_mode(int newhandlermode);
-char **CDECL _get_initial_narrow_environment();
-char ***CDECL __p__environ();
-char ***CDECL __p___argv();
-int *CDECL __p___argc();
+GUEST_PTR CDECL _get_initial_narrow_environment();
+GUEST_PTR CDECL __p__environ();
+GUEST_PTR CDECL __p___argv();
+GUEST_PTR CDECL __p___argc();
 SIZE_T CDECL strlen(const char *str);
 int CDECL strcmp(const char *lhs, const char *rhs);
 int CDECL strncmp(const char *lhs, const char *rhs, SIZE_T count);
@@ -74,10 +74,12 @@ void CDECL _exit(int status);
 void CDECL abort();
 signal_handler CDECL signal(int signum, signal_handler handler);
 void *CDECL __acrt_iob_func(unsigned int index);
-int CDECL_NO_CONV __stdio_common_vfprintf(unsigned long long options, _FILE *stream, const char *format, void *locale,
+#ifndef __x86_64__ // TODO
+int CDECL_NO_CONV __stdio_common_vfprintf(ULONGLONG options, _FILE *stream, const char *format, void *locale,
 										  va_list args);
-int CDECL_NO_CONV __stdio_common_vsprintf(unsigned long long options, char *buffer, SIZE_T len, const char *format,
+int CDECL_NO_CONV __stdio_common_vsprintf(ULONGLONG options, char *buffer, SIZE_T len, const char *format,
 										  void *locale, va_list args);
+#endif
 void CDECL qsort(void *base, SIZE_T num, SIZE_T size, sort_compare compare);
 int CDECL puts(const char *str);
 

dll/kernel32/errhandlingapi.h

@@ -7,14 +7,14 @@ constexpr DWORD EXCEPTION_MAXIMUM_PARAMETERS = 15;
 struct EXCEPTION_RECORD {
 	DWORD ExceptionCode;
 	DWORD ExceptionFlags;
-	EXCEPTION_RECORD *ExceptionRecord;
-	PVOID ExceptionAddress;
+	GUEST_PTR ExceptionRecord;
+	GUEST_PTR ExceptionAddress;
 	DWORD NumberParameters;
 	ULONG_PTR ExceptionInformation[EXCEPTION_MAXIMUM_PARAMETERS];
 };
 
-using PEXCEPTION_RECORD = EXCEPTION_RECORD *;
-using PCONTEXT = void *;
+using PEXCEPTION_RECORD = GUEST_PTR;
+using PCONTEXT = GUEST_PTR;
 
 struct EXCEPTION_POINTERS {
 	PEXCEPTION_RECORD ExceptionRecord;

dll/kernel32/fileapi.cpp

@@ -12,6 +12,7 @@
 #include "overlapped_util.h"
 #include "strutil.h"
 #include "timeutil.h"
+#include "types.h"
 
 #include <algorithm>
 #include <cctype>
@@ -146,24 +147,24 @@ struct FindSearchHandle {
 };
 
 std::mutex g_findHandleMutex;
-std::unordered_map<FindSearchHandle *, std::unique_ptr<FindSearchHandle>> g_findHandles;
+HANDLE g_nextFindHandle = 1;
+std::unordered_map<HANDLE, std::unique_ptr<FindSearchHandle>> g_findHandles;
 
 HANDLE registerFindHandle(std::unique_ptr<FindSearchHandle> handle) {
 	if (!handle) {
 		return INVALID_HANDLE_VALUE;
 	}
-	FindSearchHandle *raw = handle.get();
 	std::lock_guard lk(g_findHandleMutex);
+	HANDLE raw = g_nextFindHandle++;
 	g_findHandles.emplace(raw, std::move(handle));
-	return reinterpret_cast<HANDLE>(raw);
+	return raw;
 }
 
 FindSearchHandle *lookupFindHandleLocked(HANDLE handle) {
-	if (handle == nullptr) {
+	if (handle == NO_HANDLE) {
 		return nullptr;
 	}
-	auto *raw = reinterpret_cast<FindSearchHandle *>(handle);
-	auto it = g_findHandles.find(raw);
+	auto it = g_findHandles.find(handle);
 	if (it == g_findHandles.end()) {
 		return nullptr;
 	}
@@ -172,8 +173,7 @@ FindSearchHandle *lookupFindHandleLocked(HANDLE handle) {
 
 std::unique_ptr<FindSearchHandle> detachFindHandle(HANDLE handle) {
 	std::lock_guard lk(g_findHandleMutex);
-	auto *raw = reinterpret_cast<FindSearchHandle *>(handle);
-	auto it = g_findHandles.find(raw);
+	auto it = g_findHandles.find(handle);
 	if (it == g_findHandles.end()) {
 		return nullptr;
 	}
@@ -1185,7 +1185,7 @@ DWORD WINAPI SetFilePointer(HANDLE hFile, LONG lDistanceToMove, PLONG lpDistance
 	HOST_CONTEXT_GUARD();
 	DEBUG_LOG("SetFilePointer(%p, %ld, %p, %u)\n", hFile, static_cast<long>(lDistanceToMove), lpDistanceToMoveHigh,
 			  dwMoveMethod);
-	if (hFile == nullptr) {
+	if (hFile == NO_HANDLE) {
 		setLastError(ERROR_INVALID_HANDLE);
 		return INVALID_SET_FILE_POINTER;
 	}
@@ -1224,7 +1224,8 @@ DWORD WINAPI SetFilePointer(HANDLE hFile, LONG lDistanceToMove, PLONG lpDistance
 BOOL WINAPI SetFilePointerEx(HANDLE hFile, LARGE_INTEGER liDistanceToMove, PLARGE_INTEGER lpNewFilePointer,
 							 DWORD dwMoveMethod) {
 	HOST_CONTEXT_GUARD();
-	if (hFile == nullptr) {
+	DEBUG_LOG("SetFilePointerEx(%p, %lld, %p, %u)\n", hFile, liDistanceToMove.QuadPart, lpNewFilePointer, dwMoveMethod);
+	if (hFile == NO_HANDLE) {
 		setLastError(ERROR_INVALID_HANDLE);
 		return FALSE;
 	}
@@ -1537,12 +1538,12 @@ DWORD WINAPI GetFileType(HANDLE hFile) {
 	return type;
 }
 
-DWORD WINAPI GetFullPathNameA(LPCSTR lpFileName, DWORD nBufferLength, LPSTR lpBuffer, LPSTR *lpFilePart) {
+DWORD WINAPI GetFullPathNameA(LPCSTR lpFileName, DWORD nBufferLength, LPSTR lpBuffer, GUEST_PTR *lpFilePart) {
 	HOST_CONTEXT_GUARD();
 	DEBUG_LOG("GetFullPathNameA(%s, %u)\n", lpFileName ? lpFileName : "(null)", nBufferLength);
 
 	if (lpFilePart) {
-		*lpFilePart = nullptr;
+		*lpFilePart = GUEST_NULL;
 	}
 
 	if (!lpFileName) {
@@ -1580,21 +1581,21 @@ DWORD WINAPI GetFullPathNameA(LPCSTR lpFileName, DWORD nBufferLength, LPSTR lpBu
 
 	if (lpFilePart) {
 		if (info.filePartOffset != std::string::npos && info.filePartOffset < pathLen) {
-			*lpFilePart = lpBuffer + info.filePartOffset;
+			*lpFilePart = toGuestPtr(lpBuffer + info.filePartOffset);
 		} else {
-			*lpFilePart = nullptr;
+			*lpFilePart = GUEST_NULL;
 		}
 	}
 
 	return static_cast<DWORD>(pathLen);
 }
 
-DWORD WINAPI GetFullPathNameW(LPCWSTR lpFileName, DWORD nBufferLength, LPWSTR lpBuffer, LPWSTR *lpFilePart) {
+DWORD WINAPI GetFullPathNameW(LPCWSTR lpFileName, DWORD nBufferLength, LPWSTR lpBuffer, GUEST_PTR *lpFilePart) {
 	HOST_CONTEXT_GUARD();
 	DEBUG_LOG("GetFullPathNameW(%p, %u)\n", lpFileName, nBufferLength);
 
 	if (lpFilePart) {
-		*lpFilePart = nullptr;
+		*lpFilePart = GUEST_NULL;
 	}
 
 	if (!lpFileName) {
@@ -1633,9 +1634,9 @@ DWORD WINAPI GetFullPathNameW(LPCWSTR lpFileName, DWORD nBufferLength, LPWSTR lp
 
 	if (lpFilePart) {
 		if (info.filePartOffset != std::string::npos && info.filePartOffset < info.path.size()) {
-			*lpFilePart = lpBuffer + info.filePartOffset;
+			*lpFilePart = toGuestPtr(lpBuffer + info.filePartOffset);
 		} else {
-			*lpFilePart = nullptr;
+			*lpFilePart = GUEST_NULL;
 		}
 	}
 
@@ -1883,7 +1884,7 @@ BOOL WINAPI FindNextFileW(HANDLE hFindFile, LPWIN32_FIND_DATAW lpFindFileData) {
 BOOL WINAPI FindClose(HANDLE hFindFile) {
 	HOST_CONTEXT_GUARD();
 	DEBUG_LOG("FindClose(%p)\n", hFindFile);
-	if (hFindFile == nullptr) {
+	if (hFindFile == NO_HANDLE) {
 		DEBUG_LOG(" -> ERROR_INVALID_HANDLE\n");
 		setLastError(ERROR_INVALID_HANDLE);
 		return FALSE;

dll/kernel32/fileapi.h

@@ -56,8 +56,8 @@ constexpr DWORD INVALID_FILE_SIZE = 0xFFFFFFFF;
 
 namespace kernel32 {
 
-DWORD WINAPI GetFullPathNameA(LPCSTR lpFileName, DWORD nBufferLength, LPSTR lpBuffer, LPSTR *lpFilePart);
-DWORD WINAPI GetFullPathNameW(LPCWSTR lpFileName, DWORD nBufferLength, LPWSTR lpBuffer, LPWSTR *lpFilePart);
+DWORD WINAPI GetFullPathNameA(LPCSTR lpFileName, DWORD nBufferLength, LPSTR lpBuffer, GUEST_PTR *lpFilePart);
+DWORD WINAPI GetFullPathNameW(LPCWSTR lpFileName, DWORD nBufferLength, LPWSTR lpBuffer, GUEST_PTR *lpFilePart);
 DWORD WINAPI GetShortPathNameA(LPCSTR lpszLongPath, LPSTR lpszShortPath, DWORD cchBuffer);
 DWORD WINAPI GetShortPathNameW(LPCWSTR lpszLongPath, LPWSTR lpszShortPath, DWORD cchBuffer);
 UINT WINAPI GetTempFileNameA(LPCSTR lpPathName, LPCSTR lpPrefixString, UINT uUnique, LPSTR lpTempFileName);

dll/kernel32/heapapi.cpp

@@ -18,7 +18,7 @@ using kernel32::HeapObject;
 namespace {
 
 std::once_flag g_processHeapInitFlag;
-HANDLE g_processHeapHandle = nullptr;
+HANDLE g_processHeapHandle = NO_HANDLE;
 HeapObject *g_processHeapRecord = nullptr;
 
 void ensureProcessHeapInitialized() {
@@ -71,7 +71,7 @@ HeapObject::~HeapObject() {
 		heap = nullptr;
 	}
 	if (isProcessHeap) {
-		g_processHeapHandle = nullptr;
+		g_processHeapHandle = NO_HANDLE;
 		g_processHeapRecord = nullptr;
 	}
 }
@@ -83,13 +83,13 @@ HANDLE WINAPI HeapCreate(DWORD flOptions, SIZE_T dwInitialSize, SIZE_T dwMaximum
 	DEBUG_LOG("HeapCreate(%u, %zu, %zu)\n", flOptions, dwInitialSize, dwMaximumSize);
 	if (dwMaximumSize != 0 && dwInitialSize > dwMaximumSize) {
 		setLastError(ERROR_INVALID_PARAMETER);
-		return nullptr;
+		return NO_HANDLE;
 	}
 
 	mi_heap_t *heap = wibo::heap::createGuestHeap();
 	if (!heap) {
 		setLastError(ERROR_NOT_ENOUGH_MEMORY);
-		return nullptr;
+		return NO_HANDLE;
 	}
 
 	auto record = make_pin<HeapObject>(heap);

dll/kernel32/interlockedapi.h

@@ -5,13 +5,13 @@
 namespace kernel32 {
 
 struct SLIST_ENTRY {
-	SLIST_ENTRY *Next;
+	GUEST_PTR Next;
 };
 
 using PSLIST_ENTRY = SLIST_ENTRY *;
 
 struct SLIST_HEADER {
-	SLIST_ENTRY *Head;
+	GUEST_PTR Head;
 	unsigned short Depth;
 	unsigned short Sequence;
 };

dll/kernel32/internal.h

@@ -160,17 +160,15 @@ struct HeapObject : public ObjectBase {
 	[[nodiscard]] inline bool canAccess() const { return isProcessHeap || (isOwner() && heap != nullptr); }
 };
 
-inline constexpr uintptr_t kPseudoCurrentProcessHandleValue = static_cast<uintptr_t>(-1);
-inline constexpr uintptr_t kPseudoCurrentThreadHandleValue = static_cast<uintptr_t>(-2);
+inline constexpr HANDLE kPseudoCurrentProcessHandleValue = static_cast<HANDLE>(-1);
+inline constexpr HANDLE kPseudoCurrentThreadHandleValue = static_cast<HANDLE>(-2);
 
 inline bool isPseudoCurrentProcessHandle(HANDLE h) {
-	uintptr_t rawHandle = reinterpret_cast<uintptr_t>(h);
-	return rawHandle == kPseudoCurrentProcessHandleValue;
+	return h == kPseudoCurrentProcessHandleValue;
 }
 
 inline bool isPseudoCurrentThreadHandle(HANDLE h) {
-	uintptr_t rawHandle = reinterpret_cast<uintptr_t>(h);
-	return rawHandle == kPseudoCurrentThreadHandleValue;
+	return h == kPseudoCurrentThreadHandleValue;
 }
 
 void tryMarkExecutable(void *mem);

dll/kernel32/libloaderapi.cpp

@@ -6,6 +6,7 @@
 #include "modules.h"
 #include "resources.h"
 #include "strutil.h"
+#include "types.h"
 
 #include <algorithm>
 #include <cassert>
@@ -20,13 +21,13 @@ HRSRC findResourceInternal(HMODULE hModule, const wibo::ResourceIdentifier &type
 	auto *exe = wibo::executableFromModule(hModule);
 	if (!exe) {
 		kernel32::setLastError(ERROR_RESOURCE_DATA_NOT_FOUND);
-		return nullptr;
+		return NO_HANDLE;
 	}
 	wibo::ResourceLocation loc;
 	if (!exe->findResource(type, name, language, loc)) {
-		return nullptr;
+		return NO_HANDLE;
 	}
-	return reinterpret_cast<HRSRC>(const_cast<void *>(loc.dataEntry));
+	return toGuestPtr(loc.dataEntry);
 }
 
 } // namespace
@@ -58,7 +59,7 @@ HMODULE WINAPI GetModuleHandleA(LPCSTR lpModuleName) {
 	const auto *module = wibo::findLoadedModule(lpModuleName);
 	if (!module) {
 		setLastError(ERROR_MOD_NOT_FOUND);
-		return nullptr;
+		return NO_HANDLE;
 	}
 	return module->handle;
 }
@@ -111,7 +112,7 @@ DWORD WINAPI GetModuleFileNameA(HMODULE hModule, LPSTR lpFilename, DWORD nSize)
 
 DWORD WINAPI GetModuleFileNameW(HMODULE hModule, LPWSTR lpFilename, DWORD nSize) {
 	HOST_CONTEXT_GUARD();
-	DEBUG_LOG("GetModuleFileNameW(%p, %s, %u)\n", hModule, wideStringToString(lpFilename).c_str(), nSize);
+	DEBUG_LOG("GetModuleFileNameW(%p, %p, %u)\n", hModule, lpFilename, nSize);
 	if (!lpFilename) {
 		setLastError(ERROR_INVALID_PARAMETER);
 		return 0;
@@ -187,25 +188,25 @@ HGLOBAL WINAPI LoadResource(HMODULE hModule, HRSRC hResInfo) {
 	DEBUG_LOG("LoadResource %p %p\n", hModule, hResInfo);
 	if (!hResInfo) {
 		setLastError(ERROR_RESOURCE_DATA_NOT_FOUND);
-		return nullptr;
+		return GUEST_NULL;
 	}
 	auto *exe = wibo::executableFromModule(hModule);
 	if (!exe || !exe->rsrcBase) {
 		setLastError(ERROR_RESOURCE_DATA_NOT_FOUND);
-		return nullptr;
+		return GUEST_NULL;
 	}
 	const auto *entry = reinterpret_cast<const wibo::ImageResourceDataEntry *>(hResInfo);
 	if (!wibo::resourceEntryBelongsToExecutable(*exe, entry)) {
 		setLastError(ERROR_INVALID_PARAMETER);
-		return nullptr;
+		return GUEST_NULL;
 	}
-	return const_cast<void *>(exe->fromRVA<const void>(entry->offsetToData));
+	return toGuestPtr(exe->fromRVA<const void>(entry->offsetToData));
 }
 
 LPVOID WINAPI LockResource(HGLOBAL hResData) {
 	HOST_CONTEXT_GUARD();
 	DEBUG_LOG("LockResource(%p)\n", hResData);
-	return hResData;
+	return (LPVOID)hResData;
 }
 
 DWORD WINAPI SizeofResource(HMODULE hModule, HRSRC hResInfo) {
@@ -234,7 +235,7 @@ HMODULE WINAPI LoadLibraryA(LPCSTR lpLibFileName) {
 	const auto *info = wibo::loadModule(lpLibFileName);
 	if (!info) {
 		// lastError is set by loadModule
-		return nullptr;
+		return NO_HANDLE;
 	}
 	return info->handle;
 }
@@ -242,7 +243,7 @@ HMODULE WINAPI LoadLibraryA(LPCSTR lpLibFileName) {
 HMODULE WINAPI LoadLibraryW(LPCWSTR lpLibFileName) {
 	HOST_CONTEXT_GUARD();
 	if (!lpLibFileName) {
-		return nullptr;
+		return NO_HANDLE;
 	}
 	auto filename = wideStringToString(lpLibFileName);
 	DEBUG_LOG("LoadLibraryW(%s)\n", filename.c_str());

dll/kernel32/memoryapi.cpp

@@ -7,6 +7,7 @@
 #include "heap.h"
 #include "internal.h"
 #include "strutil.h"
+#include "types.h"
 
 #include <cerrno>
 #include <cstring>
@@ -161,8 +162,8 @@ bool mappedViewRegionForAddress(uintptr_t request, uintptr_t pageBase, MEMORY_BA
 		}
 		uintptr_t blockStart = viewStart;
 		uintptr_t blockEnd = alignUp(viewEnd, pageSize);
-		info.BaseAddress = reinterpret_cast<void *>(blockStart);
-		info.AllocationBase = reinterpret_cast<void *>(view.viewBase);
+		info.BaseAddress = toGuestPtr(reinterpret_cast<void *>(blockStart));
+		info.AllocationBase = toGuestPtr(reinterpret_cast<void *>(view.viewBase));
 		info.AllocationProtect = view.allocationProtect;
 		info.RegionSize = blockEnd > blockStart ? blockEnd - blockStart : 0;
 		info.State = MEM_COMMIT;
@@ -189,7 +190,7 @@ HANDLE WINAPI CreateFileMappingA(HANDLE hFile, LPSECURITY_ATTRIBUTES lpFileMappi
 	if (flProtect != PAGE_READONLY && flProtect != PAGE_READWRITE && flProtect != PAGE_WRITECOPY) {
 		DEBUG_LOG("CreateFileMappingA: unsupported protection 0x%x\n", flProtect);
 		setLastError(ERROR_INVALID_PARAMETER);
-		return nullptr;
+		return NO_HANDLE;
 	}
 
 	auto mapping = make_pin<MappingObject>();
@@ -200,25 +201,25 @@ HANDLE WINAPI CreateFileMappingA(HANDLE hFile, LPSECURITY_ATTRIBUTES lpFileMappi
 		mapping->fd = -1;
 		if (size == 0) {
 			setLastError(ERROR_INVALID_PARAMETER);
-			return nullptr;
+			return NO_HANDLE;
 		}
 		mapping->maxSize = size;
 	} else {
 		auto file = wibo::handles().getAs<FileObject>(hFile);
 		if (!file || !file->valid()) {
 			setLastError(ERROR_INVALID_HANDLE);
-			return nullptr;
+			return NO_HANDLE;
 		}
 		int dupFd = fcntl(file->fd, F_DUPFD_CLOEXEC, 0);
 		if (dupFd == -1) {
 			setLastErrorFromErrno();
-			return nullptr;
+			return NO_HANDLE;
 		}
 		mapping->fd = dupFd;
 		if (size == 0) {
 			off_t fileSize = lseek(dupFd, 0, SEEK_END);
 			if (fileSize < 0) {
-				return nullptr;
+				return NO_HANDLE;
 			}
 			size = static_cast<uint64_t>(fileSize);
 		}

dll/kernel32/minwinbase.h

@@ -4,7 +4,7 @@
 
 struct SECURITY_ATTRIBUTES {
 	DWORD nLength;
-	LPVOID lpSecurityDescriptor;
+	GUEST_PTR lpSecurityDescriptor;
 	BOOL bInheritHandle;
 };
 
@@ -86,7 +86,7 @@ typedef struct _OVERLAPPED {
 			DWORD Offset;
 			DWORD OffsetHigh;
 		};
-		PVOID Pointer;
+		GUEST_PTR Pointer;
 	};
 	HANDLE hEvent;
 } OVERLAPPED, *LPOVERLAPPED;

dll/kernel32/namedpipeapi.cpp

@@ -355,8 +355,8 @@ BOOL WINAPI CreatePipe(PHANDLE hReadPipe, PHANDLE hWritePipe, LPSECURITY_ATTRIBU
 		setLastError(ERROR_INVALID_PARAMETER);
 		return FALSE;
 	}
-	*hReadPipe = nullptr;
-	*hWritePipe = nullptr;
+	*hReadPipe = NO_HANDLE;
+	*hWritePipe = NO_HANDLE;
 
 	int pipeFds[2];
 	if (pipe(pipeFds) != 0) {

dll/kernel32/overlapped_util.h

@@ -1,17 +1,14 @@
 #pragma once
 
-#include "errors.h"
 #include "handles.h"
 #include "internal.h"
 #include "minwinbase.h"
 
-#include <cstdint>
-
 namespace kernel32::detail {
 
 inline HANDLE normalizedOverlappedEventHandle(const OVERLAPPED *ov) {
-	if (!ov || (reinterpret_cast<uintptr_t>(ov->hEvent) & 1U) != 0) {
-		return nullptr;
+	if (!ov || (ov->hEvent & 1U) != 0) {
+		return NO_HANDLE;
 	}
 	return ov->hEvent;
 }

dll/kernel32/processenv.cpp

@@ -6,6 +6,7 @@
 #include "heap.h"
 #include "internal.h"
 #include "strutil.h"
+#include "types.h"
 
 #include <cstdlib>
 #include <cstring>
@@ -16,6 +17,9 @@
 
 namespace {
 
+GUEST_PTR g_commandLineA = GUEST_NULL;
+GUEST_PTR g_commandLineW = GUEST_NULL;
+
 std::string convertEnvValueForWindows(const std::string &name, const char *rawValue) {
 	if (!rawValue) {
 		return {};
@@ -42,16 +46,26 @@ std::string convertEnvValueToHost(const std::string &name, const char *rawValue)
 
 namespace kernel32 {
 
-LPSTR WINAPI GetCommandLineA() {
+GUEST_PTR WINAPI GetCommandLineA() {
 	HOST_CONTEXT_GUARD();
 	DEBUG_LOG("GetCommandLineA() -> %s\n", wibo::commandLine.c_str());
-	return const_cast<LPSTR>(wibo::commandLine.c_str());
+	if (g_commandLineA == GUEST_NULL) {
+		void *tmp = wibo::heap::guestCalloc(1, wibo::commandLine.size() + 1);
+		memcpy(tmp, wibo::commandLine.c_str(), wibo::commandLine.size());
+		g_commandLineA = toGuestPtr(tmp);
+	}
+	return g_commandLineA;
 }
 
-LPWSTR WINAPI GetCommandLineW() {
+GUEST_PTR WINAPI GetCommandLineW() {
 	HOST_CONTEXT_GUARD();
 	DEBUG_LOG("GetCommandLineW() -> %s\n", wideStringToString(wibo::commandLineW.data()).c_str());
-	return wibo::commandLineW.data();
+	if (g_commandLineW == GUEST_NULL) {
+		void *tmp = wibo::heap::guestCalloc(1, wibo::commandLineW.size() * sizeof(WCHAR) + sizeof(WCHAR));
+		memcpy(tmp, wibo::commandLineW.data(), wibo::commandLineW.size() * sizeof(WCHAR));
+		g_commandLineW = toGuestPtr(tmp);
+	}
+	return g_commandLineW;
 }
 
 HANDLE WINAPI GetStdHandle(DWORD nStdHandle) {
@@ -66,7 +80,11 @@ BOOL WINAPI SetStdHandle(DWORD nStdHandle, HANDLE hHandle) {
 	return files::setStdHandle(nStdHandle, hHandle);
 }
 
-LPCH WINAPI GetEnvironmentStrings() {
+GUEST_PTR WINAPI GetEnvironmentStrings() {
+	return GetEnvironmentStringsA();
+}
+
+GUEST_PTR WINAPI GetEnvironmentStringsA() {
 	HOST_CONTEXT_GUARD();
 	DEBUG_LOG("GetEnvironmentStrings()\n");
 
@@ -82,7 +100,7 @@ LPCH WINAPI GetEnvironmentStrings() {
 	char *buffer = static_cast<char *>(wibo::heap::guestMalloc(bufSize));
 	if (!buffer) {
 		setLastError(ERROR_NOT_ENOUGH_MEMORY);
-		return nullptr;
+		return GUEST_NULL;
 	}
 	char *ptr = buffer;
 	work = environ;
@@ -96,10 +114,10 @@ LPCH WINAPI GetEnvironmentStrings() {
 	}
 	*ptr = 0;
 
-	return buffer;
+	return toGuestPtr(buffer);
 }
 
-LPWCH WINAPI GetEnvironmentStringsW() {
+GUEST_PTR WINAPI GetEnvironmentStringsW() {
 	HOST_CONTEXT_GUARD();
 	DEBUG_LOG("GetEnvironmentStringsW()\n");
 
@@ -115,7 +133,7 @@ LPWCH WINAPI GetEnvironmentStringsW() {
 	uint16_t *buffer = static_cast<uint16_t *>(wibo::heap::guestMalloc(bufSizeW * sizeof(uint16_t)));
 	if (!buffer) {
 		setLastError(ERROR_NOT_ENOUGH_MEMORY);
-		return nullptr;
+		return GUEST_NULL;
 	}
 	uint16_t *ptr = buffer;
 	work = environ;
@@ -131,7 +149,7 @@ LPWCH WINAPI GetEnvironmentStringsW() {
 	}
 	*ptr = 0;
 
-	return buffer;
+	return toGuestPtr(buffer);
 }
 
 BOOL WINAPI FreeEnvironmentStringsA(LPCH penv) {

dll/kernel32/processenv.h

@@ -4,12 +4,13 @@
 
 namespace kernel32 {
 
-LPSTR WINAPI GetCommandLineA();
-LPWSTR WINAPI GetCommandLineW();
+GUEST_PTR WINAPI GetCommandLineA();
+GUEST_PTR WINAPI GetCommandLineW();
 HANDLE WINAPI GetStdHandle(DWORD nStdHandle);
 BOOL WINAPI SetStdHandle(DWORD nStdHandle, HANDLE hHandle);
-LPCH WINAPI GetEnvironmentStrings();
-LPWCH WINAPI GetEnvironmentStringsW();
+GUEST_PTR WINAPI GetEnvironmentStrings();
+GUEST_PTR WINAPI GetEnvironmentStringsA();
+GUEST_PTR WINAPI GetEnvironmentStringsW();
 BOOL WINAPI FreeEnvironmentStringsA(LPCH penv);
 BOOL WINAPI FreeEnvironmentStringsW(LPWCH penv);
 DWORD WINAPI GetEnvironmentVariableA(LPCSTR lpName, LPSTR lpBuffer, DWORD nSize);

dll/kernel32/processthreadsapi.cpp

@@ -13,6 +13,7 @@
 #include "strutil.h"
 #include "timeutil.h"
 #include "tls.h"
+#include "types.h"
 
 #include <cerrno>
 #include <csignal>
@@ -83,7 +84,7 @@ template <typename StartupInfo> void populateStartupInfo(StartupInfo *info) {
 	info->dwFlags = STARTF_USESHOWWINDOW | STARTF_USESTDHANDLES;
 	info->wShowWindow = SW_SHOWNORMAL;
 	info->cbReserved2 = 0;
-	info->lpReserved2 = nullptr;
+	info->lpReserved2 = GUEST_NULL;
 	info->hStdInput = files::getStdHandle(STD_INPUT_HANDLE);
 	info->hStdOutput = files::getStdHandle(STD_OUTPUT_HANDLE);
 	info->hStdError = files::getStdHandle(STD_ERROR_HANDLE);
@@ -182,7 +183,7 @@ BOOL WINAPI IsProcessorFeaturePresent(DWORD ProcessorFeature) {
 HANDLE WINAPI GetCurrentProcess() {
 	HOST_CONTEXT_GUARD();
 	DEBUG_LOG("GetCurrentProcess() -> %p\n", reinterpret_cast<void *>(static_cast<uintptr_t>(-1)));
-	return reinterpret_cast<HANDLE>(static_cast<uintptr_t>(-1));
+	return kPseudoCurrentProcessHandleValue;
 }
 
 DWORD WINAPI GetCurrentProcessId() {
@@ -395,9 +396,9 @@ LPVOID WINAPI TlsGetValue(DWORD dwTlsIndex) {
 		setLastError(ERROR_INVALID_PARAMETER);
 		return nullptr;
 	}
-	void *result = wibo::tls::getValue(dwTlsIndex);
+	GUEST_PTR result = wibo::tls::getValue(dwTlsIndex);
 	setLastError(ERROR_SUCCESS);
-	return result;
+	return reinterpret_cast<LPVOID>(result);
 }
 
 BOOL WINAPI TlsSetValue(DWORD dwTlsIndex, LPVOID lpTlsValue) {
@@ -407,7 +408,7 @@ BOOL WINAPI TlsSetValue(DWORD dwTlsIndex, LPVOID lpTlsValue) {
 		setLastError(ERROR_INVALID_PARAMETER);
 		return FALSE;
 	}
-	if (!wibo::tls::setValue(dwTlsIndex, lpTlsValue)) {
+	if (!wibo::tls::setValue(dwTlsIndex, toGuestPtr(lpTlsValue))) {
 		setLastError(ERROR_INVALID_PARAMETER);
 		return FALSE;
 	}
@@ -463,7 +464,7 @@ HANDLE WINAPI CreateThread(LPSECURITY_ATTRIBUTES lpThreadAttributes, SIZE_T dwSt
 	if ((dwCreationFlags & ~SUPPORTED_FLAGS) != 0) {
 		DEBUG_LOG("CreateThread: unsupported creation flags 0x%x\n", dwCreationFlags);
 		setLastError(ERROR_NOT_SUPPORTED);
-		return nullptr;
+		return NO_HANDLE;
 	}
 
 	Pin<ThreadObject> obj = make_pin<ThreadObject>(0); // tid set during pthread_create
@@ -554,7 +555,7 @@ int WINAPI GetThreadPriority(HANDLE hThread) {
 
 DWORD WINAPI GetPriorityClass(HANDLE hProcess) {
 	HOST_CONTEXT_GUARD();
-	DEBUG_LOG("GetPriorityClass(%p)\n", hProcess);
+	DEBUG_LOG("STUB: GetPriorityClass(%p)\n", hProcess);
 	(void)hProcess;
 	return NORMAL_PRIORITY_CLASS;
 }

dll/kernel32/processthreadsapi.h

@@ -15,9 +15,9 @@ using LPPROCESS_INFORMATION = PROCESS_INFORMATION *;
 
 struct STARTUPINFOA {
 	DWORD cb;
-	LPSTR lpReserved;
-	LPSTR lpDesktop;
-	LPSTR lpTitle;
+	GUEST_PTR lpReserved;
+	GUEST_PTR lpDesktop;
+	GUEST_PTR lpTitle;
 	DWORD dwX;
 	DWORD dwY;
 	DWORD dwXSize;
@@ -28,7 +28,7 @@ struct STARTUPINFOA {
 	DWORD dwFlags;
 	WORD wShowWindow;
 	WORD cbReserved2;
-	LPBYTE lpReserved2;
+	GUEST_PTR lpReserved2;
 	HANDLE hStdInput;
 	HANDLE hStdOutput;
 	HANDLE hStdError;
@@ -38,9 +38,9 @@ using LPSTARTUPINFOA = STARTUPINFOA *;
 
 struct STARTUPINFOW {
 	DWORD cb;
-	LPWSTR lpReserved;
-	LPWSTR lpDesktop;
-	LPWSTR lpTitle;
+	GUEST_PTR lpReserved;
+	GUEST_PTR lpDesktop;
+	GUEST_PTR lpTitle;
 	DWORD dwX;
 	DWORD dwY;
 	DWORD dwXSize;
@@ -51,7 +51,7 @@ struct STARTUPINFOW {
 	DWORD dwFlags;
 	WORD wShowWindow;
 	WORD cbReserved2;
-	LPBYTE lpReserved2;
+	GUEST_PTR lpReserved2;
 	HANDLE hStdInput;
 	HANDLE hStdOutput;
 	HANDLE hStdError;

dll/kernel32/synchapi.cpp

@@ -4,8 +4,10 @@
 #include "context.h"
 #include "errors.h"
 #include "handles.h"
+#include "heap.h"
 #include "internal.h"
 #include "strutil.h"
+#include "types.h"
 
 #include <algorithm>
 #include <chrono>
@@ -158,7 +160,7 @@ HANDLE WINAPI CreateMutexW(LPSECURITY_ATTRIBUTES lpMutexAttributes, BOOL bInitia
 	if (!mu) {
 		// Name exists but isn't a mutex
 		setLastError(ERROR_INVALID_HANDLE);
-		return nullptr;
+		return NO_HANDLE;
 	}
 	HANDLE h = wibo::handles().alloc(std::move(mu), grantedAccess, handleFlags);
 	setLastError(created ? ERROR_SUCCESS : ERROR_ALREADY_EXISTS);
@@ -222,7 +224,7 @@ HANDLE WINAPI CreateEventW(LPSECURITY_ATTRIBUTES lpEventAttributes, BOOL bManual
 	if (!ev) {
 		// Name exists but isn't an event
 		setLastError(ERROR_INVALID_HANDLE);
-		return nullptr;
+		return NO_HANDLE;
 	}
 	HANDLE h = wibo::handles().alloc(std::move(ev), grantedAccess, handleFlags);
 	DEBUG_LOG("-> %p (created=%d)\n", h, created ? 1 : 0);
@@ -260,7 +262,7 @@ HANDLE WINAPI CreateSemaphoreW(LPSECURITY_ATTRIBUTES lpSemaphoreAttributes, LONG
 	if (!sem) {
 		// Name exists but isn't an event
 		setLastError(ERROR_INVALID_HANDLE);
-		return nullptr;
+		return NO_HANDLE;
 	}
 	HANDLE h = wibo::handles().alloc(std::move(sem), granted, hflags);
 	setLastError(created ? ERROR_SUCCESS : ERROR_ALREADY_EXISTS);
@@ -545,38 +547,37 @@ DWORD WINAPI WaitForMultipleObjects(DWORD nCount, const HANDLE *lpHandles, BOOL
 
 void WINAPI InitializeCriticalSection(LPCRITICAL_SECTION lpCriticalSection) {
 	HOST_CONTEXT_GUARD();
-	VERBOSE_LOG("STUB: InitializeCriticalSection(%p)\n", lpCriticalSection);
-	if (!lpCriticalSection) {
-		return;
-	}
-	std::memset(lpCriticalSection, 0, sizeof(*lpCriticalSection));
+	VERBOSE_LOG("InitializeCriticalSection(%p)\n", lpCriticalSection);
+	InitializeCriticalSectionEx(lpCriticalSection, 0, 0);
 }
 
 BOOL WINAPI InitializeCriticalSectionEx(LPCRITICAL_SECTION lpCriticalSection, DWORD dwSpinCount, DWORD Flags) {
 	HOST_CONTEXT_GUARD();
-	DEBUG_LOG("STUB: InitializeCriticalSectionEx(%p, %u, 0x%x)\n", lpCriticalSection, dwSpinCount, Flags);
+	DEBUG_LOG("InitializeCriticalSectionEx(%p, %u, 0x%x)\n", lpCriticalSection, dwSpinCount, Flags);
 	if (!lpCriticalSection) {
 		setLastError(ERROR_INVALID_PARAMETER);
 		return FALSE;
 	}
-	if (Flags & ~CRITICAL_SECTION_NO_DEBUG_INFO) {
-		setLastError(ERROR_INVALID_PARAMETER);
-		return FALSE;
-	}
 	std::memset(lpCriticalSection, 0, sizeof(*lpCriticalSection));
+	if (Flags & RTL_CRITICAL_SECTION_FLAG_NO_DEBUG_INFO) {
+		lpCriticalSection->DebugInfo = static_cast<GUEST_PTR>(-1);
+	} else {
+		auto *debugInfo = reinterpret_cast<RTL_CRITICAL_SECTION_DEBUG *>(
+			wibo::heap::guestCalloc(1, sizeof(RTL_CRITICAL_SECTION_DEBUG)));
+		debugInfo->CriticalSection = toGuestPtr(lpCriticalSection);
+		debugInfo->ProcessLocksList.Blink = toGuestPtr(&debugInfo->ProcessLocksList);
+		debugInfo->ProcessLocksList.Flink = toGuestPtr(&debugInfo->ProcessLocksList);
+		lpCriticalSection->DebugInfo = toGuestPtr(debugInfo);
+	}
+	lpCriticalSection->LockCount = -1;
 	lpCriticalSection->SpinCount = dwSpinCount;
 	return TRUE;
 }
 
 BOOL WINAPI InitializeCriticalSectionAndSpinCount(LPCRITICAL_SECTION lpCriticalSection, DWORD dwSpinCount) {
 	HOST_CONTEXT_GUARD();
-	DEBUG_LOG("STUB: InitializeCriticalSectionAndSpinCount(%p, %u)\n", lpCriticalSection, dwSpinCount);
-	if (!lpCriticalSection) {
-		setLastError(ERROR_INVALID_PARAMETER);
-		return FALSE;
-	}
-	std::memset(lpCriticalSection, 0, sizeof(*lpCriticalSection));
-	lpCriticalSection->SpinCount = dwSpinCount;
+	DEBUG_LOG("InitializeCriticalSectionAndSpinCount(%p, %u)\n", lpCriticalSection, dwSpinCount);
+	InitializeCriticalSectionEx(lpCriticalSection, dwSpinCount, 0);
 	return TRUE;
 }
 
@@ -598,7 +599,7 @@ void WINAPI LeaveCriticalSection(LPCRITICAL_SECTION lpCriticalSection) {
 	(void)lpCriticalSection;
 }
 
-BOOL WINAPI InitOnceBeginInitialize(LPINIT_ONCE lpInitOnce, DWORD dwFlags, PBOOL fPending, LPVOID *lpContext) {
+BOOL WINAPI InitOnceBeginInitialize(LPINIT_ONCE lpInitOnce, DWORD dwFlags, PBOOL fPending, GUEST_PTR *lpContext) {
 	HOST_CONTEXT_GUARD();
 	DEBUG_LOG("STUB: InitOnceBeginInitialize(%p, %u, %p, %p)\n", lpInitOnce, dwFlags, fPending, lpContext);
 	if (!lpInitOnce) {
@@ -613,7 +614,7 @@ BOOL WINAPI InitOnceBeginInitialize(LPINIT_ONCE lpInitOnce, DWORD dwFlags, PBOOL
 		*fPending = TRUE;
 	}
 	if (lpContext) {
-		*lpContext = nullptr;
+		*lpContext = GUEST_NULL;
 	}
 	return TRUE;
 }

dll/kernel32/synchapi.h

@@ -15,14 +15,20 @@ constexpr DWORD INIT_ONCE_ASYNC = 0x00000002UL;
 constexpr DWORD INIT_ONCE_INIT_FAILED = 0x00000004UL;
 constexpr DWORD INIT_ONCE_CTX_RESERVED_BITS = 2;
 
-constexpr DWORD CRITICAL_SECTION_NO_DEBUG_INFO = 0x01000000UL;
+constexpr DWORD RTL_CRITICAL_SECTION_FLAG_NO_DEBUG_INFO = 0x01000000UL;
+constexpr DWORD RTL_CRITICAL_SECTION_FLAG_DYNAMIC_SPIN = 0x02000000UL;
+constexpr DWORD RTL_CRITICAL_SECTION_FLAG_STATIC_INIT = 0x04000000UL;
+constexpr DWORD RTL_CRITICAL_SECTION_FLAG_RESOURCE_TYPE = 0x08000000UL;
+constexpr DWORD RTL_CRITICAL_SECTION_FLAG_FORCE_DEBUG_INFO = 0x10000000UL;
+constexpr DWORD RTL_CRITICAL_SECTION_ALL_FLAG_BITS = 0xff000000UL;
+constexpr DWORD RTL_CRITICAL_SECTION_FLAG_RESERVED = 0xe0000000UL;
 
 struct RTL_CRITICAL_SECTION;
 
 struct RTL_CRITICAL_SECTION_DEBUG {
 	WORD Type;
 	WORD CreatorBackTraceIndex;
-	RTL_CRITICAL_SECTION *CriticalSection;
+	GUEST_PTR CriticalSection;
 	LIST_ENTRY ProcessLocksList;
 	DWORD EntryCount;
 	DWORD ContentionCount;
@@ -33,7 +39,7 @@ struct RTL_CRITICAL_SECTION_DEBUG {
 };
 
 struct RTL_CRITICAL_SECTION {
-	RTL_CRITICAL_SECTION_DEBUG *DebugInfo;
+	GUEST_PTR DebugInfo;
 	LONG LockCount;
 	LONG RecursionCount;
 	HANDLE OwningThread;
@@ -47,7 +53,7 @@ using PCRITICAL_SECTION = RTL_CRITICAL_SECTION *;
 using PRTL_CRITICAL_SECTION_DEBUG = RTL_CRITICAL_SECTION_DEBUG *;
 
 union RTL_RUN_ONCE {
-	PVOID Ptr;
+	GUEST_PTR Ptr;
 };
 
 using PRTL_RUN_ONCE = RTL_RUN_ONCE *;
@@ -55,17 +61,17 @@ using INIT_ONCE = RTL_RUN_ONCE;
 using PINIT_ONCE = INIT_ONCE *;
 using LPINIT_ONCE = INIT_ONCE *;
 
-constexpr INIT_ONCE INIT_ONCE_STATIC_INIT{nullptr};
+constexpr INIT_ONCE INIT_ONCE_STATIC_INIT{GUEST_NULL};
 
 union RTL_SRWLOCK {
-	PVOID Ptr;
+	GUEST_PTR Ptr;
 };
 
 using SRWLOCK = RTL_SRWLOCK;
 using PSRWLOCK = SRWLOCK *;
 using PRTL_SRWLOCK = SRWLOCK *;
 
-constexpr SRWLOCK SRWLOCK_INIT{nullptr};
+constexpr SRWLOCK SRWLOCK_INIT{GUEST_NULL};
 
 namespace kernel32 {
 
@@ -92,7 +98,7 @@ BOOL WINAPI InitializeCriticalSectionAndSpinCount(LPCRITICAL_SECTION lpCriticalS
 void WINAPI DeleteCriticalSection(LPCRITICAL_SECTION lpCriticalSection);
 void WINAPI EnterCriticalSection(LPCRITICAL_SECTION lpCriticalSection);
 void WINAPI LeaveCriticalSection(LPCRITICAL_SECTION lpCriticalSection);
-BOOL WINAPI InitOnceBeginInitialize(LPINIT_ONCE lpInitOnce, DWORD dwFlags, PBOOL fPending, LPVOID *lpContext);
+BOOL WINAPI InitOnceBeginInitialize(LPINIT_ONCE lpInitOnce, DWORD dwFlags, PBOOL fPending, GUEST_PTR *lpContext);
 BOOL WINAPI InitOnceComplete(LPINIT_ONCE lpInitOnce, DWORD dwFlags, LPVOID lpContext);
 void WINAPI AcquireSRWLockShared(PSRWLOCK SRWLock);
 void WINAPI ReleaseSRWLockShared(PSRWLOCK SRWLock);

dll/kernel32/sysinfoapi.cpp

@@ -4,6 +4,7 @@
 #include "context.h"
 #include "errors.h"
 #include "internal.h"
+#include "ntdll.h"
 #include "timeutil.h"
 
 #include <cstring>
@@ -52,12 +53,12 @@ void WINAPI GetSystemInfo(LPSYSTEM_INFO lpSystemInfo) {
 	}
 	lpSystemInfo->dwPageSize = static_cast<DWORD>(pageSize);
 
-	lpSystemInfo->lpMinimumApplicationAddress = reinterpret_cast<LPVOID>(0x00010000);
-	if (sizeof(void *) == 4) {
-		lpSystemInfo->lpMaximumApplicationAddress = reinterpret_cast<LPVOID>(0x7FFEFFFF);
-	} else {
-		lpSystemInfo->lpMaximumApplicationAddress = reinterpret_cast<LPVOID>(0x00007FFFFFFEFFFFull);
-	}
+	lpSystemInfo->lpMinimumApplicationAddress = toGuestPtr(reinterpret_cast<void *>(0x00010000));
+#ifdef _WIN64
+	lpSystemInfo->lpMaximumApplicationAddress = toGuestPtr(reinterpret_cast<void *>(0x00007FFFFFFEFFFFull));
+#else
+	lpSystemInfo->lpMaximumApplicationAddress = toGuestPtr(reinterpret_cast<void *>(0x7FFEFFFF));
+#endif
 
 	unsigned int cpuCount = 1;
 	long reported = sysconf(_SC_NPROCESSORS_ONLN);
@@ -198,11 +199,62 @@ BOOL WINAPI GetVersionExA(LPOSVERSIONINFOA lpVersionInformation) {
 		setLastError(ERROR_INVALID_PARAMETER);
 		return FALSE;
 	}
-	std::memset(lpVersionInformation, 0, lpVersionInformation->dwOSVersionInfoSize);
-	lpVersionInformation->dwMajorVersion = kMajorVersion;
-	lpVersionInformation->dwMinorVersion = kMinorVersion;
-	lpVersionInformation->dwBuildNumber = kBuildNumber;
-	lpVersionInformation->dwPlatformId = 2;
+
+	DWORD size = lpVersionInformation->dwOSVersionInfoSize;
+	if (size < sizeof(OSVERSIONINFOA)) {
+		setLastError(ERROR_INVALID_PARAMETER);
+		return FALSE;
+	}
+
+	DWORD requestSize = (size >= sizeof(OSVERSIONINFOEXA)) ? sizeof(OSVERSIONINFOEXW) : sizeof(OSVERSIONINFOW);
+	OSVERSIONINFOEXW wideInfo{};
+	wideInfo.dwOSVersionInfoSize = requestSize;
+	NTSTATUS status = ntdll::RtlGetVersion(reinterpret_cast<PRTL_OSVERSIONINFOW>(&wideInfo));
+	if (status != STATUS_SUCCESS) {
+		setLastError(wibo::winErrorFromNtStatus(status));
+		return FALSE;
+	}
+
+	std::memset(lpVersionInformation, 0, size);
+	lpVersionInformation->dwOSVersionInfoSize = size;
+	lpVersionInformation->dwMajorVersion = wideInfo.dwMajorVersion;
+	lpVersionInformation->dwMinorVersion = wideInfo.dwMinorVersion;
+	lpVersionInformation->dwBuildNumber = wideInfo.dwBuildNumber;
+	lpVersionInformation->dwPlatformId = wideInfo.dwPlatformId;
+
+	if (size >= sizeof(OSVERSIONINFOEXA)) {
+		auto extended = reinterpret_cast<OSVERSIONINFOEXA *>(lpVersionInformation);
+		extended->wServicePackMajor = wideInfo.wServicePackMajor;
+		extended->wServicePackMinor = wideInfo.wServicePackMinor;
+		extended->wSuiteMask = wideInfo.wSuiteMask;
+		extended->wProductType = wideInfo.wProductType;
+		extended->wReserved = wideInfo.wReserved;
+	}
+
+	return TRUE;
+}
+
+BOOL WINAPI GetVersionExW(LPOSVERSIONINFOW lpVersionInformation) {
+	HOST_CONTEXT_GUARD();
+	DEBUG_LOG("GetVersionExW(%p)\n", lpVersionInformation);
+	if (!lpVersionInformation) {
+		setLastError(ERROR_INVALID_PARAMETER);
+		return FALSE;
+	}
+
+	DWORD size = lpVersionInformation->dwOSVersionInfoSize;
+	if (size < sizeof(OSVERSIONINFOW)) {
+		setLastError(ERROR_INVALID_PARAMETER);
+		return FALSE;
+	}
+
+	NTSTATUS status = ntdll::RtlGetVersion(reinterpret_cast<PRTL_OSVERSIONINFOW>(lpVersionInformation));
+	if (status != STATUS_SUCCESS) {
+		setLastError(wibo::winErrorFromNtStatus(status));
+		return FALSE;
+	}
+
+	lpVersionInformation->dwOSVersionInfoSize = size;
 	return TRUE;
 }
 

dll/kernel32/sysinfoapi.h

@@ -12,8 +12,8 @@ struct SYSTEM_INFO {
 		};
 	};
 	DWORD dwPageSize;
-	LPVOID lpMinimumApplicationAddress;
-	LPVOID lpMaximumApplicationAddress;
+	GUEST_PTR lpMinimumApplicationAddress;
+	GUEST_PTR lpMaximumApplicationAddress;
 	DWORD_PTR dwActiveProcessorMask;
 	DWORD dwNumberOfProcessors;
 	DWORD dwProcessorType;
@@ -35,6 +35,37 @@ struct OSVERSIONINFOA {
 
 using LPOSVERSIONINFOA = OSVERSIONINFOA *;
 
+struct OSVERSIONINFOW {
+	DWORD dwOSVersionInfoSize;
+	DWORD dwMajorVersion;
+	DWORD dwMinorVersion;
+	DWORD dwBuildNumber;
+	DWORD dwPlatformId;
+	WCHAR szCSDVersion[128];
+};
+
+using LPOSVERSIONINFOW = OSVERSIONINFOW *;
+
+struct OSVERSIONINFOEXA : OSVERSIONINFOA {
+	WORD wServicePackMajor;
+	WORD wServicePackMinor;
+	WORD wSuiteMask;
+	BYTE wProductType;
+	BYTE wReserved;
+};
+
+using LPOSVERSIONINFOEXA = OSVERSIONINFOEXA *;
+
+struct OSVERSIONINFOEXW : OSVERSIONINFOW {
+	WORD wServicePackMajor;
+	WORD wServicePackMinor;
+	WORD wSuiteMask;
+	BYTE wProductType;
+	BYTE wReserved;
+};
+
+using LPOSVERSIONINFOEXW = OSVERSIONINFOEXW *;
+
 namespace kernel32 {
 
 void WINAPI GetSystemInfo(LPSYSTEM_INFO lpSystemInfo);
@@ -44,5 +75,6 @@ void WINAPI GetSystemTimeAsFileTime(LPFILETIME lpSystemTimeAsFileTime);
 DWORD WINAPI GetTickCount();
 DWORD WINAPI GetVersion();
 BOOL WINAPI GetVersionExA(LPOSVERSIONINFOA lpVersionInformation);
+BOOL WINAPI GetVersionExW(LPOSVERSIONINFOW lpVersionInformation);
 
 } // namespace kernel32

dll/kernel32/winbase.cpp

@@ -235,7 +235,7 @@ bool doFree(void *mem) {
 		DEBUG_LOG("doFree(%p) -> virtualFree\n", mem);
 		MEMORY_BASIC_INFORMATION info;
 		auto result = wibo::heap::virtualQuery(mem, &info);
-		if (result != wibo::heap::VmStatus::Success || info.BaseAddress != mem) {
+		if (result != wibo::heap::VmStatus::Success || fromGuestPtr(info.BaseAddress) != mem) {
 			return false;
 		}
 		wibo::heap::virtualFree(mem, info.RegionSize, MEM_RELEASE);
@@ -697,7 +697,7 @@ BOOL WINAPI FindActCtxSectionStringW(DWORD dwFlags, const GUID *lpExtensionGuid,
 	ReturnedData->ulDataFormatVersion = 1;
 	ReturnedData->ulFlags = ACTCTX_SECTION_KEYED_DATA_FLAG_FOUND_IN_ACTCTX;
 	if (dwFlags & FIND_ACTCTX_SECTION_KEY_RETURN_HACTCTX) {
-		ReturnedData->hActCtx = reinterpret_cast<HANDLE>(&g_builtinActCtx);
+		ReturnedData->hActCtx = toGuestPtr(&g_builtinActCtx);
 	}
 
 	if (!matchedEntry) {
@@ -705,9 +705,9 @@ BOOL WINAPI FindActCtxSectionStringW(DWORD dwFlags, const GUID *lpExtensionGuid,
 		return FALSE;
 	}
 
-	ReturnedData->lpData = const_cast<ACTIVATION_CONTEXT_DATA_DLL_REDIRECTION *>(&matchedEntry->dllData);
+	ReturnedData->lpData = toGuestPtr(&matchedEntry->dllData);
 	ReturnedData->ulLength = matchedEntry->dllData.Size;
-	ReturnedData->lpSectionBase = const_cast<ACTIVATION_CONTEXT_DATA_DLL_REDIRECTION *>(&matchedEntry->dllData);
+	ReturnedData->lpSectionBase = toGuestPtr(&matchedEntry->dllData);
 	ReturnedData->ulSectionTotalLength = matchedEntry->dllData.Size;
 	ReturnedData->ulAssemblyRosterIndex = 1;
 	ReturnedData->AssemblyMetadata = {};
@@ -805,20 +805,20 @@ HGLOBAL WINAPI GlobalAlloc(UINT uFlags, SIZE_T dwBytes) {
 	if (uFlags & GMEM_MOVEABLE) {
 		// not implemented rn
 		assert(0);
-		return nullptr;
+		return NO_HANDLE;
 	}
 	bool zero = (uFlags & GMEM_ZEROINIT) != 0;
 	void *ret = doAlloc(static_cast<UINT>(dwBytes), zero);
 	DEBUG_LOG("-> %p\n", ret);
-	return ret;
+	return toGuestPtr(ret);
 }
 
 HGLOBAL WINAPI GlobalFree(HGLOBAL hMem) {
 	HOST_CONTEXT_GUARD();
 	VERBOSE_LOG("GlobalFree(%p)\n", hMem);
-	if (doFree(hMem)) {
+	if (doFree(reinterpret_cast<void *>(hMem))) {
 		DEBUG_LOG("-> success\n");
-		return nullptr;
+		return NO_HANDLE;
 	} else {
 		DEBUG_LOG("-> failure\n");
 		return hMem;
@@ -830,12 +830,12 @@ HGLOBAL WINAPI GlobalReAlloc(HGLOBAL hMem, SIZE_T dwBytes, UINT uFlags) {
 	VERBOSE_LOG("GlobalReAlloc(%p, %zu, %x)\n", hMem, static_cast<size_t>(dwBytes), uFlags);
 	if (uFlags & GMEM_MODIFY) {
 		assert(0);
-		return nullptr;
+		return NO_HANDLE;
 	}
 	bool zero = (uFlags & GMEM_ZEROINIT) != 0;
-	void *ret = doRealloc(hMem, static_cast<UINT>(dwBytes), zero);
+	void *ret = doRealloc(reinterpret_cast<void *>(hMem), static_cast<UINT>(dwBytes), zero);
 	DEBUG_LOG("-> %p\n", ret);
-	return ret;
+	return toGuestPtr(ret);
 }
 
 UINT WINAPI GlobalFlags(HGLOBAL hMem) {
@@ -855,20 +855,24 @@ HLOCAL WINAPI LocalAlloc(UINT uFlags, SIZE_T uBytes) {
 	void *result = doAlloc(static_cast<UINT>(uBytes), zero);
 	if (!result) {
 		setLastError(ERROR_NOT_SUPPORTED);
-		return nullptr;
+		return NO_HANDLE;
 	}
 	// Legacy Windows applications (pre-NX and DEP) may expect executable memory from LocalAlloc.
 	tryMarkExecutable(result);
 	DEBUG_LOG("  -> %p\n", result);
-	return result;
+	return toGuestPtr(result);
 }
 
 HLOCAL WINAPI LocalFree(HLOCAL hMem) {
 	HOST_CONTEXT_GUARD();
 	VERBOSE_LOG("LocalFree(%p)\n", hMem);
-	// Windows returns NULL on success.
-	std::free(hMem);
-	return nullptr;
+	if (doFree(reinterpret_cast<void *>(hMem))) {
+		DEBUG_LOG("-> success\n");
+		return NO_HANDLE;
+	} else {
+		DEBUG_LOG("-> failure\n");
+		return hMem;
+	}
 }
 
 HLOCAL WINAPI LocalReAlloc(HLOCAL hMem, SIZE_T uBytes, UINT uFlags) {
@@ -878,27 +882,27 @@ HLOCAL WINAPI LocalReAlloc(HLOCAL hMem, SIZE_T uBytes, UINT uFlags) {
 	if ((uFlags & LMEM_MOVEABLE) != 0) {
 		DEBUG_LOG("  ignoring LMEM_MOVEABLE\n");
 	}
-	void *result = doRealloc(hMem, static_cast<UINT>(uBytes), zero);
+	void *result = doRealloc(reinterpret_cast<void *>(hMem), static_cast<UINT>(uBytes), zero);
 	if (!result && uBytes != 0) {
 		setLastError(ERROR_NOT_SUPPORTED);
-		return nullptr;
+		return NO_HANDLE;
 	}
 	// Legacy Windows applications (pre-NX and DEP) may expect executable memory from LocalReAlloc.
 	tryMarkExecutable(result);
 	DEBUG_LOG("  -> %p\n", result);
-	return result;
+	return toGuestPtr(result);
 }
 
 HLOCAL WINAPI LocalHandle(LPCVOID pMem) {
 	HOST_CONTEXT_GUARD();
 	VERBOSE_LOG("LocalHandle(%p)\n", pMem);
-	return const_cast<LPVOID>(pMem);
+	return toGuestPtr(pMem);
 }
 
 LPVOID WINAPI LocalLock(HLOCAL hMem) {
 	HOST_CONTEXT_GUARD();
 	VERBOSE_LOG("LocalLock(%p)\n", hMem);
-	return hMem;
+	return reinterpret_cast<void *>(hMem);
 }
 
 BOOL WINAPI LocalUnlock(HLOCAL hMem) {
@@ -911,7 +915,7 @@ BOOL WINAPI LocalUnlock(HLOCAL hMem) {
 SIZE_T WINAPI LocalSize(HLOCAL hMem) {
 	HOST_CONTEXT_GUARD();
 	VERBOSE_LOG("LocalSize(%p)\n", hMem);
-	return hMem ? mi_usable_size(hMem) : 0;
+	return hMem ? mi_usable_size(reinterpret_cast<void *>(hMem)) : 0;
 }
 
 UINT WINAPI LocalFlags(HLOCAL hMem) {

dll/kernel32/winbase.h

@@ -5,21 +5,21 @@
 struct GUID;
 
 struct ACTCTX_SECTION_KEYED_DATA_ASSEMBLY_METADATA {
-	PVOID lpInformation;
-	PVOID lpSectionBase;
+	GUEST_PTR lpInformation;
+	GUEST_PTR lpSectionBase;
 	ULONG ulSectionLength;
-	PVOID lpSectionGlobalData;
+	GUEST_PTR lpSectionGlobalData;
 	ULONG ulSectionGlobalDataLength;
 };
 
 struct ACTCTX_SECTION_KEYED_DATA {
 	ULONG cbSize;
 	ULONG ulDataFormatVersion;
-	PVOID lpData;
+	GUEST_PTR lpData;
 	ULONG ulLength;
-	PVOID lpSectionGlobalData;
+	GUEST_PTR lpSectionGlobalData;
 	ULONG ulSectionGlobalDataLength;
-	PVOID lpSectionBase;
+	GUEST_PTR lpSectionBase;
 	ULONG ulSectionTotalLength;
 	HANDLE hActCtx;
 	ULONG ulAssemblyRosterIndex;
@@ -68,8 +68,8 @@ ATOM WINAPI AddAtomW(LPCWSTR lpString);
 UINT WINAPI GetAtomNameA(ATOM nAtom, LPSTR lpBuffer, int nSize);
 UINT WINAPI GetAtomNameW(ATOM nAtom, LPWSTR lpBuffer, int nSize);
 UINT WINAPI SetHandleCount(UINT uNumber);
-DWORD WINAPI FormatMessageA(DWORD dwFlags, LPCVOID lpSource, DWORD dwMessageId, DWORD dwLanguageId, LPSTR lpBuffer,
-							DWORD nSize, va_list *Arguments);
+// DWORD WINAPI FormatMessageA(DWORD dwFlags, LPCVOID lpSource, DWORD dwMessageId, DWORD dwLanguageId, LPSTR lpBuffer,
+// 							DWORD nSize, va_list *Arguments);
 PVOID WINAPI EncodePointer(PVOID Ptr);
 PVOID WINAPI DecodePointer(PVOID Ptr);
 BOOL WINAPI SetDllDirectoryA(LPCSTR lpPathName);

dll/kernel32/winnls.cpp

@@ -19,6 +19,9 @@ constexpr DWORD kNormIgnoreCase = 0x00000001;
 constexpr DWORD LCID_INSTALLED = 0x00000001;
 constexpr DWORD LCID_SUPPORTED = 0x00000002;
 constexpr DWORD LCID_ALTERNATE_SORTS = 0x00000004;
+constexpr LCID kEnUsLcid = 0x0409;
+constexpr LCID kInvariantLcid = 0x007f;
+constexpr DWORD LOCALE_ALLOW_NEUTRAL_NAMES = 0x08000000;
 
 int compareStrings(const std::string &a, const std::string &b, DWORD dwCmpFlags) {
 	for (size_t i = 0;; ++i) {
@@ -102,6 +105,50 @@ LANGID WINAPI GetUserDefaultUILanguage() {
 	return 0;
 }
 
+int WINAPI GetUserDefaultLocaleName(LPWSTR lpLocaleName, int cchLocaleName) {
+	HOST_CONTEXT_GUARD();
+	DEBUG_LOG("GetUserDefaultLocaleName(%p, %d)\n", lpLocaleName, cchLocaleName);
+	if (!lpLocaleName || cchLocaleName < 0) {
+		setLastError(ERROR_INVALID_PARAMETER);
+		return 0;
+	}
+
+	constexpr char16_t localeName[] = u"en-US";
+	constexpr int requiredChars = static_cast<int>(sizeof(localeName) / sizeof(char16_t));
+	if (cchLocaleName < requiredChars) {
+		setLastError(ERROR_INSUFFICIENT_BUFFER);
+		return 0;
+	}
+
+	std::memcpy(lpLocaleName, localeName, sizeof(localeName));
+	return requiredChars;
+}
+
+LCID WINAPI LocaleNameToLCID(LPCWSTR lpName, DWORD dwFlags) {
+	HOST_CONTEXT_GUARD();
+	DEBUG_LOG("LocaleNameToLCID(%p, 0x%x)\n", lpName, dwFlags);
+	if (dwFlags & ~LOCALE_ALLOW_NEUTRAL_NAMES) {
+		setLastError(ERROR_INVALID_PARAMETER);
+		return 0;
+	}
+	if (!lpName) {
+		return kEnUsLcid;
+	}
+
+	std::string localeName = wideStringToString(lpName);
+	if (localeName.empty()) {
+		return kInvariantLcid;
+	}
+
+	std::string normalized = stringToLower(localeName);
+	if (normalized == "en-us" || normalized == "en_us" || normalized == "!x-sys-default-locale") {
+		return kEnUsLcid;
+	}
+
+	setLastError(ERROR_INVALID_PARAMETER);
+	return 0;
+}
+
 BOOL WINAPI GetCPInfo(UINT CodePage, LPCPINFO lpCPInfo) {
 	HOST_CONTEXT_GUARD();
 	DEBUG_LOG("GetCPInfo(%u, %p)\n", CodePage, lpCPInfo);

dll/kernel32/winnls.h

@@ -19,6 +19,8 @@ namespace kernel32 {
 UINT WINAPI GetACP();
 LANGID WINAPI GetSystemDefaultLangID();
 LANGID WINAPI GetUserDefaultUILanguage();
+int WINAPI GetUserDefaultLocaleName(LPWSTR lpLocaleName, int cchLocaleName);
+LCID WINAPI LocaleNameToLCID(LPCWSTR lpName, DWORD dwFlags);
 BOOL WINAPI GetCPInfo(UINT CodePage, LPCPINFO lpCPInfo);
 int WINAPI CompareStringA(LCID Locale, DWORD dwCmpFlags, LPCSTR lpString1, int cchCount1, LPCSTR lpString2,
 							int cchCount2);

dll/kernel32/wow64apiset.cpp

@@ -5,14 +5,15 @@
 #include "errors.h"
 #include "handles.h"
 #include "internal.h"
+#include "types.h"
 
 namespace kernel32 {
 
-BOOL WINAPI Wow64DisableWow64FsRedirection(PVOID *OldValue) {
+BOOL WINAPI Wow64DisableWow64FsRedirection(GUEST_PTR *OldValue) {
 	HOST_CONTEXT_GUARD();
 	DEBUG_LOG("STUB: Wow64DisableWow64FsRedirection(%p)\n", OldValue);
 	if (OldValue) {
-		*OldValue = nullptr;
+		*OldValue = GUEST_NULL;
 	}
 	return TRUE;
 }

dll/kernel32/wow64apiset.h

@@ -4,7 +4,7 @@
 
 namespace kernel32 {
 
-BOOL WINAPI Wow64DisableWow64FsRedirection(PVOID *OldValue);
+BOOL WINAPI Wow64DisableWow64FsRedirection(GUEST_PTR *OldValue);
 BOOL WINAPI Wow64RevertWow64FsRedirection(PVOID OldValue);
 BOOL WINAPI IsWow64Process(HANDLE hProcess, PBOOL Wow64Process);
 

dll/msvcrt.cpp

@@ -9,6 +9,7 @@
 #include "msvcrt_trampolines.h"
 #include "processes.h"
 #include "strutil.h"
+#include "types.h"
 
 #include <algorithm>
 #include <array>
@@ -97,10 +98,10 @@ int closeGuestFile(_FILE *file) {
 namespace msvcrt {
 	int _commode;
 	int _fmode;
-	char** __initenv;
-	uint16_t** __winitenv;
-	uint16_t* _wpgmptr = nullptr;
-	char* _pgmptr = nullptr;
+	GUEST_PTR __initenv = GUEST_NULL;
+	GUEST_PTR __winitenv = GUEST_NULL;
+	GUEST_PTR _wpgmptr = GUEST_NULL;
+	GUEST_PTR _pgmptr = GUEST_NULL;
 	int __mb_cur_max = 1;
 	_FILE _iob[_IOB_ENTRIES] = {_FILE{STDOUT_FILENO}, _FILE{STDERR_FILENO}, _FILE{STDIN_FILENO}};
 
@@ -381,11 +382,11 @@ namespace msvcrt {
 		return mbCodePageSetting;
 	}
 
-	int* CDECL __p___mb_cur_max() {
+	GUEST_PTR CDECL __p___mb_cur_max() {
 		HOST_CONTEXT_GUARD();
 		ensureMbctypeInitialized();
 		DEBUG_LOG("__p___mb_cur_max() -> %u\n", __mb_cur_max);
-		return &__mb_cur_max;
+		return toGuestPtr(&__mb_cur_max);
 	}
 
 	int CDECL _setmbcp(int codepage) {
@@ -414,19 +415,19 @@ namespace msvcrt {
 		return 0;
 	}
 
-	unsigned char *CDECL __p__mbctype() {
+	GUEST_PTR CDECL __p__mbctype() {
 		HOST_CONTEXT_GUARD();
 		ensureMbctypeInitialized();
 		DEBUG_LOG("__p__mbctype() -> %p\n", mbctypeTable.data());
-		return mbctypeTable.data();
+		return toGuestPtr(mbctypeTable.data());
 	}
 
-	unsigned short **CDECL __p__pctype() {
+	GUEST_PTR CDECL __p__pctype() {
 		HOST_CONTEXT_GUARD();
 		DEBUG_LOG("__p__pctype()\n");
 		static unsigned short *pointer = nullptr;
 		pointer = pctypeTable().data() + 1;
-		return &pointer;
+		return toGuestPtr(&pointer);
 	}
 
 	int CDECL _isctype(int ch, int mask) {
@@ -502,11 +503,11 @@ namespace msvcrt {
 
 		template <typename CharT>
 		struct StringListStorage {
-			std::vector<std::unique_ptr<CharT[]>> strings;
-			std::unique_ptr<CharT*[]> pointers;
+			std::vector<wibo::heap::guest_ptr<CharT[]>> strings;
+			wibo::heap::guest_ptr<GUEST_PTR[]> pointers;
 
 			template <typename Converter>
-			CharT **assign(char **source, Converter convert) {
+			GUEST_PTR *assign(char **source, Converter convert) {
 				if (!source) {
 					strings.clear();
 					pointers.reset();
@@ -520,18 +521,18 @@ namespace msvcrt {
 
 				strings.clear();
 				strings.reserve(count);
-				pointers = std::make_unique<CharT *[]>(count + 1);
+				pointers = wibo::heap::make_guest_unique<GUEST_PTR[]>(count + 1);
 
 				for (SIZE_T i = 0; i < count; ++i) {
 					auto data = convert(source[i]);
-					auto buffer = std::make_unique<CharT[]>(data.size());
+					auto buffer = wibo::heap::make_guest_unique<CharT[]>(data.size());
 					std::copy(data.begin(), data.end(), buffer.get());
 					CharT *raw = buffer.get();
 					strings.emplace_back(std::move(buffer));
-					pointers[i] = raw;
+					pointers[i] = toGuestPtr(raw);
 				}
 
-				pointers[count] = nullptr;
+				pointers[count] = GUEST_NULL;
 				return pointers.get();
 			}
 		};
@@ -554,7 +555,7 @@ namespace msvcrt {
 
 		template <typename CharT, typename Converter>
 		// NOLINTNEXTLINE(readability-non-const-parameter)
-		int getMainArgsCommon(int *argcOut, CharT ***argvOut, CharT ***envOut, Converter convert) {
+		int getMainArgsCommon(int *argcOut, GUEST_PTR *argvOut, GUEST_PTR *envOut, Converter convert) {
 			if (argcOut) {
 				*argcOut = wibo::argc;
 			}
@@ -563,18 +564,18 @@ namespace msvcrt {
 			static StringListStorage<CharT> envStorage;
 
 			if (argvOut) {
-				*argvOut = argvStorage.assign(wibo::argv, convert);
+				*argvOut = toGuestPtr(argvStorage.assign(wibo::argv, convert));
 			}
 
-			CharT **envData = envStorage.assign(environ, convert);
+			GUEST_PTR *envData = envStorage.assign(environ, convert);
 			if (envOut) {
-				*envOut = envData;
+				*envOut = toGuestPtr(envData);
 			}
 
 			if constexpr (std::is_same_v<CharT, uint16_t>) {
-				__winitenv = envData;
+				__winitenv = toGuestPtr(envData);
 			} else if constexpr (std::is_same_v<CharT, char>) {
-				__initenv = envData;
+				__initenv = toGuestPtr(envData);
 			}
 
 			return 0;
@@ -633,7 +634,7 @@ namespace msvcrt {
 			if (!__winitenv) {
 				getMainArgsCommon<uint16_t>(nullptr, nullptr, nullptr, copyWideString);
 			}
-			return __winitenv;
+			return reinterpret_cast<uint16_t**>(__winitenv);
 		}
 	} // namespace
 
@@ -644,42 +645,44 @@ namespace msvcrt {
 		(void)at;
 	}
 
-	int* CDECL __p__fmode() {
+	GUEST_PTR CDECL __p__fmode() {
 		HOST_CONTEXT_GUARD();
 		DEBUG_LOG("__p__fmode() -> %p\n", &_fmode);
-		return &_fmode;
+		return toGuestPtr(&_fmode);
 	}
 
-	int* CDECL __p__commode() {
+	GUEST_PTR CDECL __p__commode() {
 		HOST_CONTEXT_GUARD();
 		DEBUG_LOG("__p__commode() -> %p\n", &_commode);
-		return &_commode;
+		return toGuestPtr(&_commode);
 	}
 
-	void CDECL _initterm(const _PVFV *ppfn, const _PVFV* end) {
+
+	void CDECL _initterm(GUEST_PTR *ppfn, GUEST_PTR *end) {
 		HOST_CONTEXT_GUARD();
 		DEBUG_LOG("_initterm(%p, %p)\n", ppfn, end);
-		for (; ppfn < end; ppfn++) {
-			_PVFV func = *ppfn;
-			if (func) {
-				DEBUG_LOG("_initterm: calling %p\n", func);
-				call__PVFV(func);
-			}
+		do {
+			if (GUEST_PTR pfn = *++ppfn) {
+				DEBUG_LOG("-> calling %p\n", pfn);
+				auto fn = reinterpret_cast<_PVFV>(fromGuestPtr(pfn));
+				call__PVFV(fn);
 			}
+		} while (ppfn < end);
 	}
 
-	int CDECL _initterm_e(const _PIFV *ppfn, const _PIFV *end) {
+	int CDECL _initterm_e(GUEST_PTR *ppfn, GUEST_PTR *end) {
 		HOST_CONTEXT_GUARD();
 		DEBUG_LOG("_initterm_e(%p, %p)\n", ppfn, end);
-		for (; ppfn < end; ppfn++) {
-			_PIFV func = *ppfn;
-			if (func) {
-				int err = call__PIFV(func);
-				DEBUG_LOG("_initterm_e: calling %p -> %d\n", func, err);
-				if (err != 0)
+		do {
+			if (GUEST_PTR pfn = *++ppfn) {
+				DEBUG_LOG("-> calling %p\n", pfn);
+				auto fn = reinterpret_cast<_PIFV>(fromGuestPtr(pfn));
+				int err = call__PIFV(fn);
+				if (err)
 					return err;
 			}
-		}
+		} while (ppfn < end);
+
 		return 0;
 	}
 
@@ -720,7 +723,7 @@ namespace msvcrt {
 	}
 
 	// NOLINTNEXTLINE(readability-non-const-parameter)
-	int CDECL __wgetmainargs(int *wargc, uint16_t ***wargv, uint16_t ***wenv, int doWildcard, int *startInfo) {
+	int CDECL __wgetmainargs(int *wargc, GUEST_PTR *wargv, GUEST_PTR *wenv, int doWildcard, int *startInfo) {
 		HOST_CONTEXT_GUARD();
 		DEBUG_LOG("__wgetmainargs(doWildcard=%d)\n", doWildcard);
 		(void)startInfo;
@@ -733,7 +736,7 @@ namespace msvcrt {
 	}
 
 	// NOLINTNEXTLINE(readability-non-const-parameter)
-	int CDECL __getmainargs(int *argc, char ***argv, char ***env, int doWildcard, int *startInfo) {
+	int CDECL __getmainargs(int *argc, GUEST_PTR *argv, GUEST_PTR *env, int doWildcard, int *startInfo) {
 		HOST_CONTEXT_GUARD();
 		DEBUG_LOG("__getmainargs(doWildcard=%d)\n", doWildcard);
 		(void)startInfo;
@@ -749,10 +752,10 @@ namespace msvcrt {
 		return std::getenv(varname);
 	}
 
-	char*** CDECL __p___initenv() {
+	GUEST_PTR CDECL __p___initenv() {
 		HOST_CONTEXT_GUARD();
 		DEBUG_LOG("__p___initenv() -> %p\n", &__initenv);
-		return &__initenv;
+		return toGuestPtr(&__initenv);
 	}
 
 	char* CDECL strcat(char *dest, const char *src) {
@@ -881,7 +884,7 @@ namespace msvcrt {
 		}
 
 		if (!prev) {
-			long diff = static_cast<long>(current - start);
+			LONG diff = static_cast<LONG>(current - start);
 			DEBUG_LOG("_mbsdec fallback start=%p current=%p diff=%ld first-bytes=%02x %02x %02x %02x\n",
 				start, current, diff,
 				start ? start[0] : 0, start ? start[1] : 0,
@@ -1166,11 +1169,11 @@ namespace msvcrt {
 		return ::close(fd);
 	}
 
-	long CDECL _lseek(int fd, long offset, int origin) {
+	LONG CDECL _lseek(int fd, LONG offset, int origin) {
 		HOST_CONTEXT_GUARD();
 		DEBUG_LOG("_lseek(%d, %ld, %d)\n", fd, offset, origin);
 		off_t result = ::lseek(fd, static_cast<off_t>(offset), origin);
-		return static_cast<long>(result);
+		return static_cast<LONG>(result);
 	}
 
 	int CDECL _unlink(const char *path) {
@@ -1199,7 +1202,7 @@ namespace msvcrt {
 		return ::utime(hostPath.c_str(), &native);
 	}
 
-	int CDECL _chsize(int fd, long size) {
+	int CDECL _chsize(int fd, LONG size) {
 		HOST_CONTEXT_GUARD();
 		DEBUG_LOG("_chsize(%d, %ld)\n", fd, size);
 		return ::ftruncate(fd, static_cast<off_t>(size));
@@ -1238,13 +1241,13 @@ namespace msvcrt {
 		return std::strtok(str, delim);
 	}
 
-	long CDECL _adj_fdiv_r(long value) {
+	LONG CDECL _adj_fdiv_r(LONG value) {
 		HOST_CONTEXT_GUARD();
 		DEBUG_LOG("STUB: _adj_fdiv_r(%ld)\n", value);
 		return value;
 	}
 
-	void CDECL _adjust_fdiv(long n) {
+	void CDECL _adjust_fdiv(LONG n) {
 		HOST_CONTEXT_GUARD();
 		DEBUG_LOG("STUB: _adjust_fdiv(%ld)\n", n);
 		(void)n;
@@ -1261,14 +1264,14 @@ namespace msvcrt {
 		if (gettimeofday(&tv, nullptr) != 0) {
 			return -1;
 		}
-		timeptr->time = tv.tv_sec;
+		timeptr->time = static_cast<LONG>(tv.tv_sec);
 		timeptr->millitm = static_cast<unsigned short>(tv.tv_usec / 1000);
 		timeptr->timezone = 0;
 		timeptr->dstflag = 0;
 		return 0;
 	}
 
-	unsigned long CDECL _ultoa(unsigned long value, char *str, int radix) {
+	ULONG CDECL _ultoa(ULONG value, char *str, int radix) {
 		HOST_CONTEXT_GUARD();
 		DEBUG_LOG("_ultoa(%lu, %p, %d)\n", value, str, radix);
 		if (!str || radix < 2 || radix > 36) {
@@ -1282,15 +1285,15 @@ namespace msvcrt {
 			*--cursor = '0';
 		}
 		while (value > 0) {
-			unsigned long digit = value % static_cast<unsigned long>(radix);
-			value /= static_cast<unsigned long>(radix);
+			ULONG digit = value % static_cast<ULONG>(radix);
+			value /= static_cast<ULONG>(radix);
 			*--cursor = static_cast<char>(digit < 10 ? '0' + digit : 'A' + (digit - 10));
 		}
 		std::strcpy(str, cursor);
-		return static_cast<unsigned long>(std::strlen(str));
+		return static_cast<ULONG>(std::strlen(str));
 	}
 
-	char* CDECL _ltoa(long value, char *str, int radix) {
+	char* CDECL _ltoa(LONG value, char *str, int radix) {
 		HOST_CONTEXT_GUARD();
 		DEBUG_LOG("_ltoa(%ld, %p, %d)\n", value, str, radix);
 		if (!str || radix < 2 || radix > 36) {
@@ -1298,9 +1301,9 @@ namespace msvcrt {
 			return nullptr;
 		}
 		bool negative = value < 0;
-		unsigned long absValue = negative ? static_cast<unsigned long>(-value) : static_cast<unsigned long>(value);
+		ULONG absValue = negative ? static_cast<ULONG>(-value) : static_cast<ULONG>(value);
 		char buffer[65];
-		unsigned long length = _ultoa(absValue, buffer, radix);
+		ULONG length = _ultoa(absValue, buffer, radix);
 		std::string result;
 		if (negative) {
 			result.push_back('-');
@@ -1437,7 +1440,7 @@ namespace msvcrt {
 	TIME_T CDECL time(TIME_T *t) {
 		HOST_CONTEXT_GUARD();
 		DEBUG_LOG("time(%p)\n", t);
-		TIME_T result = std::time(nullptr);
+		TIME_T result = static_cast<TIME_T>(std::time(nullptr));
 		if (t) {
 			*t = result;
 		}
@@ -1468,11 +1471,11 @@ namespace msvcrt {
 		return result;
 	}
 
-	int CDECL _wdupenv_s(uint16_t **buffer, SIZE_T *numberOfElements, const uint16_t *varname){
+	int CDECL _wdupenv_s(GUEST_PTR *buffer, SIZE_T *numberOfElements, const uint16_t *varname){
 		HOST_CONTEXT_GUARD();
 		DEBUG_LOG("_wdupenv_s(%p, %p, %s)\n", buffer, numberOfElements, wideStringToString(varname).c_str());
 		if (buffer) {
-			*buffer = nullptr;
+			*buffer = GUEST_NULL;
 		}
 		if (numberOfElements) {
 			*numberOfElements = 0;
@@ -1504,7 +1507,7 @@ namespace msvcrt {
 
 		wstrncpy(copy, match->value, value_len);
 		copy[value_len] = 0;
-		*buffer = copy;
+		*buffer = toGuestPtr(copy);
 		if (numberOfElements) {
 			*numberOfElements = value_len + 1;
 		}
@@ -1678,10 +1681,18 @@ namespace msvcrt {
 		return copy;
 	}
 
-	unsigned long CDECL strtoul(const char *str, char **endptr, int base) {
+	ULONG CDECL strtoul(const char *str, GUEST_PTR *endptr, int base) {
 		HOST_CONTEXT_GUARD();
 		VERBOSE_LOG("strtoul(%s, %p, %d)\n", str, endptr, base);
-		return ::strtoul(str, endptr, base);
+		int ret;
+		if (endptr != nullptr) {
+			char *endptr_host = reinterpret_cast<char*>(fromGuestPtr(*endptr));
+			ret = ::strtoul(str, &endptr_host, base);
+			*endptr = toGuestPtr(endptr_host);
+		} else {
+			ret = ::strtoul(str, nullptr, base);
+		}
+		return ret;
 	}
 
 	void* CDECL malloc(SIZE_T size){
@@ -1728,7 +1739,7 @@ namespace msvcrt {
 		lockTable()[static_cast<SIZE_T>(locknum)].unlock();
 	}
 
-	_onexit_t CDECL __dllonexit(_onexit_t func, _onexit_t **pbegin, _onexit_t **pend) {
+	_onexit_t CDECL __dllonexit(_onexit_t func, GUEST_PTR *pbegin, GUEST_PTR *pend) {
 		HOST_CONTEXT_GUARD();
 		DEBUG_LOG("__dllonexit(%p, %p, %p)\n", func, pbegin, pend);
 		if (!pbegin || !pend) {
@@ -1740,13 +1751,13 @@ namespace msvcrt {
 			return nullptr;
 		}
 
-		_onexit_t *table = static_cast<_onexit_t *>(wibo::heap::guestRealloc(pbegin, len * sizeof(_onexit_t)));
+		GUEST_PTR *table = static_cast<GUEST_PTR *>(wibo::heap::guestRealloc(pbegin, len * sizeof(GUEST_PTR)));
 		if (!table) {
 			return nullptr;
 		}
-		*pbegin = table;
-		*pend = table + len;
-		table[len - 1] = func;
+		*pbegin = toGuestPtr(table);
+		*pend = toGuestPtr(table + len);
+		table[len - 1] = toGuestPtr(reinterpret_cast<void *>(func));
 		return func;
 	}
 
@@ -1827,14 +1838,14 @@ namespace msvcrt {
 		return isatty(fd);
 	}
 
-	int CDECL fseek(_FILE *stream, long offset, int origin) {
+	int CDECL fseek(_FILE *stream, LONG offset, int origin) {
 		HOST_CONTEXT_GUARD();
 		VERBOSE_LOG("fseek(%p, %ld, %d)\n", stream, offset, origin);
 		FILE* host = mapToHostFile(stream);
 		return std::fseek(host, offset, origin);
 	}
 
-	long CDECL ftell(_FILE *stream) {
+	LONG CDECL ftell(_FILE *stream) {
 		HOST_CONTEXT_GUARD();
 		VERBOSE_LOG("ftell(%p)\n", stream);
 		FILE* host = mapToHostFile(stream);
@@ -1895,7 +1906,7 @@ namespace msvcrt {
 		return std::fgetwc(host);
 	}
 
-	int CDECL _wfopen_s(_FILE **stream, const uint16_t *filename, const uint16_t *mode) {
+	int CDECL _wfopen_s(GUEST_PTR *stream, const uint16_t *filename, const uint16_t *mode) {
 		HOST_CONTEXT_GUARD();
 		DEBUG_LOG("_wfopen_s(%p, %s, %s)\n", stream, wideStringToString(filename).c_str(),
 				  wideStringToString(mode).c_str());
@@ -1907,10 +1918,10 @@ namespace msvcrt {
 		std::string narrowMode = wideStringToString(mode);
 		FILE *handle = std::fopen(narrowName.c_str(), narrowMode.c_str());
 		if (!handle) {
-			*stream = nullptr;
+			*stream = GUEST_NULL;
 			return errno ? errno : EINVAL;
 		}
-		*stream = mapToGuestFile(handle);
+		*stream = toGuestPtr(mapToGuestFile(handle));
 		return 0;
 	}
 
@@ -2061,13 +2072,13 @@ namespace msvcrt {
 		return 0;
 	}
 
-	unsigned long CDECL wcsspn(const uint16_t *str1, const uint16_t *str2) {
+	ULONG CDECL wcsspn(const uint16_t *str1, const uint16_t *str2) {
 		HOST_CONTEXT_GUARD();
 		VERBOSE_LOG("wcsspn(%p, %p)\n", str1, str2);
 		if (!str1 || !str2) {
 			return 0;
 		}
-		unsigned long count = 0;
+		ULONG count = 0;
 		for (const uint16_t *p = str1; *p; ++p) {
 			bool match = false;
 			for (const uint16_t *q = str2; *q; ++q) {
@@ -2084,7 +2095,7 @@ namespace msvcrt {
 		return count;
 	}
 
-	long CDECL _wtol(const uint16_t *str) {
+	LONG CDECL _wtol(const uint16_t *str) {
 		HOST_CONTEXT_GUARD();
 		VERBOSE_LOG("_wtol(%p)\n", str);
 		return wstrtol(str, nullptr, 10);
@@ -2341,13 +2352,13 @@ namespace msvcrt {
 		return 0;
 	}
 
-	long CDECL _XcptFilter(unsigned long code, void *) {
+	LONG CDECL _XcptFilter(ULONG code, void *) {
 		HOST_CONTEXT_GUARD();
 		DEBUG_LOG("STUB: _XcptFilter(%lu)\n", code);
 		return 0;
 	}
 
-	int CDECL _get_wpgmptr(uint16_t **pValue) {
+	int CDECL _get_wpgmptr(GUEST_PTR *pValue) {
 		HOST_CONTEXT_GUARD();
 		DEBUG_LOG("_get_wpgmptr(%p)\n", pValue);
 		if (!pValue) {
@@ -2358,22 +2369,25 @@ namespace msvcrt {
 			return 0;
 		}
 
-		const auto wStr = stringToWideString(wibo::guestExecutablePath.c_str());
-		delete[] _wpgmptr;
+		// TODO
+		// const auto wStr = stringToWideString(wibo::guestExecutablePath.c_str());
+		// delete[] _wpgmptr;
 
-		_wpgmptr = new uint16_t[wStr.size() + 1];
-		std::copy(wStr.begin(), wStr.end(), _wpgmptr);
-		_wpgmptr[wStr.size()] = 0;
+		// _wpgmptr = new uint16_t[wStr.size() + 1];
+		// std::copy(wStr.begin(), wStr.end(), _wpgmptr);
+		// _wpgmptr[wStr.size()] = 0;
 
-		*pValue = _wpgmptr;
+		// *pValue = _wpgmptr;
 		return 0;
 	}
 
-	char** CDECL __p__pgmptr() {
+	GUEST_PTR CDECL __p__pgmptr() {
 		HOST_CONTEXT_GUARD();
 		DEBUG_LOG("__p__pgmptr()\n");
-		_pgmptr = const_cast<char *>(wibo::guestExecutablePath.c_str());
-		return &_pgmptr;
+		// TODO
+		// _pgmptr = const_cast<char *>(wibo::guestExecutablePath.c_str());
+		// return &_pgmptr;
+		return GUEST_NULL;
 	}
 
 	int CDECL _wsplitpath_s(const uint16_t * path, uint16_t * drive, SIZE_T driveNumberOfElements, uint16_t *dir, SIZE_T dirNumberOfElements,
@@ -2564,7 +2578,7 @@ namespace msvcrt {
 		return wstrtol(str, nullptr, 10);
 	}
 
-	int CDECL _ltoa_s(long value, char *buffer, SIZE_T sizeInChars, int radix) {
+	int CDECL _ltoa_s(LONG value, char *buffer, SIZE_T sizeInChars, int radix) {
 		HOST_CONTEXT_GUARD();
 		VERBOSE_LOG("_ltoa_s(%ld, %p, %zu, %d)\n", value, buffer, sizeInChars, radix);
 		if (!buffer || sizeInChars == 0) {
@@ -2782,10 +2796,18 @@ namespace msvcrt {
 		return wstrrchr(str, c);
 	}
 
-	unsigned long CDECL wcstoul(const uint16_t *strSource, uint16_t **endptr, int base){
+	ULONG CDECL wcstoul(const uint16_t *strSource, GUEST_PTR *endptr, int base){
 		HOST_CONTEXT_GUARD();
 		VERBOSE_LOG("wcstoul(%p, %p, %d)\n", strSource, endptr, base);
-		return wstrtoul(strSource, endptr, base);
+		int ret;
+		if (endptr != nullptr) {
+			uint16_t *endptr_host = reinterpret_cast<uint16_t*>(fromGuestPtr(*endptr));
+			ret = ::wstrtoul(strSource, &endptr_host, base);
+			*endptr = toGuestPtr(endptr_host);
+		} else {
+			ret = ::wstrtoul(strSource, nullptr, base);
+		}
+		return ret;
 	}
 
 	_FILE* CDECL _wfsopen(const uint16_t* filename, const uint16_t* mode, int shflag){
@@ -2842,7 +2864,7 @@ namespace msvcrt {
 		return &errno;
 	}
 
-	LONG_PTR CDECL _wspawnvp(int mode, const uint16_t* cmdname, const uint16_t* const * argv) {
+	LONG_PTR CDECL _wspawnvp(int mode, const uint16_t* cmdname, GUEST_PTR *argv) {
 		HOST_CONTEXT_GUARD();
 		if (!cmdname || !argv) {
 			errno = EINVAL;
@@ -2853,8 +2875,9 @@ namespace msvcrt {
 		DEBUG_LOG("_wspawnvp(%d, %s)\n", mode, command.c_str());
 
 		std::vector<std::string> argStorage;
-		for (const uint16_t *const *cursor = argv; *cursor; ++cursor) {
-			argStorage.emplace_back(wideStringToString(*cursor));
+		for (GUEST_PTR *cursor = argv; *cursor; ++cursor) {
+			const WCHAR *arg = reinterpret_cast<const WCHAR*>(fromGuestPtr(*cursor));
+			argStorage.emplace_back(wideStringToString(arg));
 		}
 
 		auto resolved = wibo::resolveExecutable(command, false);
@@ -2891,7 +2914,7 @@ namespace msvcrt {
 		return static_cast<intptr_t>(po->pid);
 	}
 
-	LONG_PTR CDECL _spawnvp(int mode, const char *cmdname, const char * const *argv) {
+	LONG_PTR CDECL _spawnvp(int mode, const char *cmdname, GUEST_PTR *argv) {
 		HOST_CONTEXT_GUARD();
 		if (!cmdname || !argv) {
 			errno = EINVAL;
@@ -2902,8 +2925,9 @@ namespace msvcrt {
 		DEBUG_LOG("_spawnvp(%d, %s)\n", mode, command.c_str());
 
 		std::vector<std::string> argStorage;
-		for (const char * const *cursor = argv; *cursor; ++cursor) {
-			argStorage.emplace_back(*cursor);
+		for (GUEST_PTR *cursor = argv; *cursor; ++cursor) {
+			const char *arg = reinterpret_cast<const char*>(fromGuestPtr(*cursor));
+			argStorage.emplace_back(arg);
 		}
 
 		auto resolved = wibo::resolveExecutable(command, false);

dll/msvcrt.h

@@ -12,8 +12,8 @@ typedef void(_CC_CDECL *signal_handler)(int);
 typedef int(_CC_CDECL *sort_compare)(const void *, const void *);
 
 struct _utimbuf {
-	long actime;
-	long modtime;
+	LONG actime;
+	LONG modtime;
 };
 
 struct _timeb {
@@ -29,10 +29,10 @@ namespace msvcrt {
 
 extern int _commode;
 extern int _fmode;
-extern char **__initenv;
-extern WCHAR **__winitenv;
-extern WCHAR *_wpgmptr;
-extern char *_pgmptr;
+extern GUEST_PTR __initenv;
+extern GUEST_PTR __winitenv;
+extern GUEST_PTR _wpgmptr;
+extern GUEST_PTR _pgmptr;
 extern int __mb_cur_max;
 extern _FILE _iob[_IOB_ENTRIES];
 
@@ -42,23 +42,23 @@ void CDECL setbuf(_FILE *stream, char *buffer);
 void CDECL _splitpath(const char *path, char *drive, char *dir, char *fname, char *ext);
 int CDECL _fileno(_FILE *stream);
 int CDECL _getmbcp();
-int *CDECL __p___mb_cur_max();
+GUEST_PTR CDECL __p___mb_cur_max();
 int CDECL _setmbcp(int codepage);
-unsigned char *CDECL __p__mbctype();
-unsigned short **CDECL __p__pctype();
+GUEST_PTR CDECL __p__mbctype();
+GUEST_PTR CDECL __p__pctype();
 int CDECL _isctype(int ch, int mask);
 void CDECL __set_app_type(int at);
-int *CDECL __p__fmode();
-int *CDECL __p__commode();
-void CDECL _initterm(const _PVFV *ppfn, const _PVFV *end);
-int CDECL _initterm_e(const _PIFV *ppfn, const _PIFV *end);
+GUEST_PTR CDECL __p__fmode();
+GUEST_PTR CDECL __p__commode();
+void CDECL _initterm(GUEST_PTR *ppfn, GUEST_PTR *end);
+int CDECL _initterm_e(GUEST_PTR *ppfn, GUEST_PTR *end);
 unsigned int CDECL _controlfp(unsigned int newControl, unsigned int mask);
 int CDECL _controlfp_s(unsigned int *currentControl, unsigned int newControl, unsigned int mask);
 _onexit_t CDECL _onexit(_onexit_t func);
-int CDECL __wgetmainargs(int *wargc, WCHAR ***wargv, WCHAR ***wenv, int doWildcard, int *startInfo);
-int CDECL __getmainargs(int *argc, char ***argv, char ***env, int doWildcard, int *startInfo);
+int CDECL __wgetmainargs(int *wargc, GUEST_PTR *wargv, GUEST_PTR *wenv, int doWildcard, int *startInfo);
+int CDECL __getmainargs(int *argc, GUEST_PTR *argv, GUEST_PTR *env, int doWildcard, int *startInfo);
 char *CDECL getenv(const char *varname);
-char ***CDECL __p___initenv();
+GUEST_PTR CDECL __p___initenv();
 char *CDECL strcat(char *dest, const char *src);
 char *CDECL strcpy(char *dest, const char *src);
 int CDECL _access(const char *path, int mode);
@@ -84,31 +84,33 @@ int CDECL _ismbcdigit(unsigned int ch);
 int CDECL _stricmp(const char *lhs, const char *rhs);
 int CDECL _strnicmp(const char *lhs, const char *rhs, SIZE_T count);
 int CDECL _memicmp(const void *lhs, const void *rhs, SIZE_T count);
+#ifndef __x86_64__ // TODO
 int CDECL_NO_CONV _vsnprintf(char *buffer, SIZE_T count, const char *format, va_list args);
 int CDECL_NO_CONV _snprintf(char *buffer, SIZE_T count, const char *format, ...);
 int CDECL_NO_CONV sprintf(char *buffer, const char *format, ...);
 int CDECL_NO_CONV printf(const char *format, ...);
 int CDECL_NO_CONV sscanf(const char *buffer, const char *format, ...);
+#endif
 char *CDECL fgets(char *str, int count, _FILE *stream);
 SIZE_T CDECL fread(void *buffer, SIZE_T size, SIZE_T count, _FILE *stream);
 _FILE *CDECL _fsopen(const char *filename, const char *mode, int shflag);
 int CDECL _sopen(const char *path, int oflag, int shflag, int pmode);
 int CDECL _read(int fd, void *buffer, unsigned int count);
 int CDECL _close(int fd);
-long CDECL _lseek(int fd, long offset, int origin);
+LONG CDECL _lseek(int fd, LONG offset, int origin);
 int CDECL _unlink(const char *path);
 int CDECL _utime(const char *path, const _utimbuf *times);
-int CDECL _chsize(int fd, long size);
+int CDECL _chsize(int fd, LONG size);
 char *CDECL strncpy(char *dest, const char *src, SIZE_T count);
 char *CDECL strpbrk(const char *str, const char *accept);
 char *CDECL strstr(const char *haystack, const char *needle);
 char *CDECL strrchr(const char *str, int ch);
 char *CDECL strtok(char *str, const char *delim);
-long CDECL _adj_fdiv_r(long value);
-void CDECL _adjust_fdiv(long n);
+LONG CDECL _adj_fdiv_r(LONG value);
+void CDECL _adjust_fdiv(LONG n);
 int CDECL _ftime(struct _timeb *timeptr);
-unsigned long CDECL _ultoa(unsigned long value, char *str, int radix);
-char *CDECL _ltoa(long value, char *str, int radix);
+ULONG CDECL _ultoa(ULONG value, char *str, int radix);
+char *CDECL _ltoa(LONG value, char *str, int radix);
 char *CDECL _makepath(char *path, const char *drive, const char *dir, const char *fname, const char *ext);
 char *CDECL _fullpath(char *absPath, const char *relPath, SIZE_T maxLength);
 int CDECL _putenv(const char *envString);
@@ -119,7 +121,7 @@ TIME_T CDECL time(TIME_T *t);
 char *CDECL __unDName(char *outputString, const char *mangledName, int maxStringLength, void *(*allocFunc)(SIZE_T),
 					  void (*freeFunc)(void *), unsigned short);
 char *CDECL setlocale(int category, const char *locale);
-int CDECL _wdupenv_s(WCHAR **buffer, SIZE_T *numberOfElements, const WCHAR *varname);
+int CDECL _wdupenv_s(GUEST_PTR *buffer, SIZE_T *numberOfElements, const WCHAR *varname);
 int CDECL _wgetenv_s(SIZE_T *pReturnValue, WCHAR *buffer, SIZE_T numberOfElements, const WCHAR *varname);
 SIZE_T CDECL strlen(const char *str);
 int CDECL strcmp(const char *lhs, const char *rhs);
@@ -129,38 +131,40 @@ int CDECL strcpy_s(char *dest, SIZE_T dest_size, const char *src);
 int CDECL strcat_s(char *dest, SIZE_T numberOfElements, const char *src);
 int CDECL strncpy_s(char *dest, SIZE_T dest_size, const char *src, SIZE_T count);
 char *CDECL _strdup(const char *strSource);
-unsigned long CDECL strtoul(const char *str, char **endptr, int base);
+ULONG CDECL strtoul(const char *str, GUEST_PTR *endptr, int base);
 void *CDECL malloc(SIZE_T size);
 void *CDECL calloc(SIZE_T count, SIZE_T size);
 void *CDECL realloc(void *ptr, SIZE_T size);
 void *CDECL _malloc_crt(SIZE_T size);
 void CDECL _lock(int locknum);
 void CDECL _unlock(int locknum);
-_onexit_t CDECL __dllonexit(_onexit_t func, _onexit_t **pbegin, _onexit_t **pend);
+_onexit_t CDECL __dllonexit(_onexit_t func, GUEST_PTR *pbegin, GUEST_PTR *pend);
 void CDECL free(void *ptr);
 void *CDECL memcpy(void *dest, const void *src, SIZE_T count);
 void *CDECL memmove(void *dest, const void *src, SIZE_T count);
 int CDECL memcmp(const void *lhs, const void *rhs, SIZE_T count);
 void CDECL qsort(void *base, SIZE_T num, SIZE_T size, sort_compare compare);
 int CDECL fflush(_FILE *stream);
+#ifndef __x86_64__ // TODO
 int CDECL_NO_CONV vfwprintf(_FILE *stream, const WCHAR *format, va_list args);
+#endif
 _FILE *CDECL fopen(const char *filename, const char *mode);
 int CDECL _dup2(int fd1, int fd2);
 int CDECL _isatty(int fd);
-int CDECL fseek(_FILE *stream, long offset, int origin);
-long CDECL ftell(_FILE *stream);
+int CDECL fseek(_FILE *stream, LONG offset, int origin);
+LONG CDECL ftell(_FILE *stream);
 int CDECL feof(_FILE *stream);
 int CDECL fputws(const WCHAR *str, _FILE *stream);
 int CDECL _cputws(const WCHAR *string);
 WCHAR *CDECL fgetws(WCHAR *buffer, int size, _FILE *stream);
 WINT_T CDECL fgetwc(_FILE *stream);
-int CDECL _wfopen_s(_FILE **stream, const WCHAR *filename, const WCHAR *mode);
+int CDECL _wfopen_s(GUEST_PTR *stream, const WCHAR *filename, const WCHAR *mode);
 int CDECL _wcsicmp(const WCHAR *lhs, const WCHAR *rhs);
 int CDECL _wmakepath_s(WCHAR *path, SIZE_T sizeInWords, const WCHAR *drive, const WCHAR *dir, const WCHAR *fname,
 					   const WCHAR *ext);
 int CDECL _wputenv_s(const WCHAR *varname, const WCHAR *value);
-unsigned long CDECL wcsspn(const WCHAR *str1, const WCHAR *str2);
-long CDECL _wtol(const WCHAR *str);
+ULONG CDECL wcsspn(const WCHAR *str1, const WCHAR *str2);
+LONG CDECL _wtol(const WCHAR *str);
 int CDECL _wcsupr_s(WCHAR *str, SIZE_T size);
 int CDECL _wcslwr_s(WCHAR *str, SIZE_T size);
 WINT_T CDECL towlower(WINT_T ch);
@@ -172,8 +176,10 @@ int CDECL _crt_debugger_hook(int value);
 int CDECL _configthreadlocale(int mode);
 void CDECL __setusermatherr(void *handler);
 void CDECL _cexit();
+#ifndef __x86_64__ // TODO
 int CDECL_NO_CONV vfprintf(_FILE *stream, const char *format, va_list args);
 int CDECL_NO_CONV fprintf(_FILE *stream, const char *format, ...);
+#endif
 int CDECL fputc(int ch, _FILE *stream);
 SIZE_T CDECL fwrite(const void *buffer, SIZE_T size, SIZE_T count, _FILE *stream);
 char *CDECL strerror(int errnum);
@@ -188,9 +194,9 @@ void CDECL _invoke_watson(const WCHAR *, const WCHAR *, const WCHAR *, UINT, UIN
 void CDECL terminateShim();
 int CDECL _purecall();
 int CDECL _except_handler4_common(void *, void *, void *, void *);
-long CDECL _XcptFilter(unsigned long code, void *);
-int CDECL _get_wpgmptr(WCHAR **pValue);
-char **CDECL __p__pgmptr();
+LONG CDECL _XcptFilter(ULONG code, void *);
+int CDECL _get_wpgmptr(GUEST_PTR *pValue);
+GUEST_PTR CDECL __p__pgmptr();
 int CDECL _wsplitpath_s(const WCHAR *path, WCHAR *drive, SIZE_T driveNumberOfElements, WCHAR *dir,
 						SIZE_T dirNumberOfElements, WCHAR *fname, SIZE_T nameNumberOfElements, WCHAR *ext,
 						SIZE_T extNumberOfElements);
@@ -202,28 +208,32 @@ int CDECL wcsncpy_s(WCHAR *strDest, SIZE_T numberOfElements, const WCHAR *strSou
 int CDECL wcsncat_s(WCHAR *strDest, SIZE_T numberOfElements, const WCHAR *strSource, SIZE_T count);
 int CDECL _itow_s(int value, WCHAR *buffer, SIZE_T size, int radix);
 int CDECL _wtoi(const WCHAR *str);
-int CDECL _ltoa_s(long value, char *buffer, SIZE_T sizeInChars, int radix);
+int CDECL _ltoa_s(LONG value, char *buffer, SIZE_T sizeInChars, int radix);
 int CDECL wcscpy_s(WCHAR *dest, SIZE_T dest_size, const WCHAR *src);
+#ifndef __x86_64__ // TODO
 int CDECL_NO_CONV swprintf_s(WCHAR *buffer, SIZE_T sizeOfBuffer, const WCHAR *format, ...);
 int CDECL_NO_CONV swscanf_s(const WCHAR *buffer, const WCHAR *format, ...);
+#endif
 int *CDECL _get_osfhandle(int fd);
 int CDECL _write(int fd, const void *buffer, unsigned int count);
 void CDECL exit(int status);
 int CDECL wcsncmp(const WCHAR *string1, const WCHAR *string2, SIZE_T count);
+#ifndef __x86_64__ // TODO
 int CDECL_NO_CONV _vswprintf_c_l(WCHAR *buffer, SIZE_T size, const WCHAR *format, ...);
+#endif
 const WCHAR *CDECL wcsstr(const WCHAR *dest, const WCHAR *src);
 int CDECL iswspace(WINT_T w);
 int CDECL iswdigit(WINT_T w);
 const WCHAR *CDECL wcschr(const WCHAR *str, WCHAR c);
 const WCHAR *CDECL wcsrchr(const WCHAR *str, WCHAR c);
-unsigned long CDECL wcstoul(const WCHAR *strSource, WCHAR **endptr, int base);
+ULONG CDECL wcstoul(const WCHAR *strSource, GUEST_PTR *endptr, int base);
 _FILE *CDECL _wfsopen(const WCHAR *filename, const WCHAR *mode, int shflag);
 int CDECL puts(const char *str);
 int CDECL fclose(_FILE *stream);
 int CDECL _flushall();
 int *CDECL _errno();
-LONG_PTR CDECL _wspawnvp(int mode, const WCHAR *cmdname, const WCHAR *const *argv);
-LONG_PTR CDECL _spawnvp(int mode, const char *cmdname, const char *const *argv);
+LONG_PTR CDECL _wspawnvp(int mode, const WCHAR *cmdname, GUEST_PTR *argv);
+LONG_PTR CDECL _spawnvp(int mode, const char *cmdname, GUEST_PTR *argv);
 int CDECL _wunlink(const WCHAR *filename);
 WCHAR *CDECL _wfullpath(WCHAR *absPath, const WCHAR *relPath, SIZE_T maxLength);
 

dll/ntdll.cpp

@@ -24,7 +24,7 @@ namespace {
 
 struct PROCESS_BASIC_INFORMATION {
 	NTSTATUS ExitStatus;
-	PEB *PebBaseAddress;
+	GUEST_PTR PebBaseAddress;
 	ULONG_PTR AffinityMask;
 	LONG BasePriority;
 	ULONG_PTR UniqueProcessId;
@@ -94,8 +94,8 @@ std::string windowsImagePathFor(const ProcessHandleDetails &details) {
 } // namespace
 
 namespace kernel32 {
-BOOL WIN_FUNC SetEvent(HANDLE hEvent);
-BOOL WIN_FUNC ResetEvent(HANDLE hEvent);
+BOOL WINAPI SetEvent(HANDLE hEvent);
+BOOL WINAPI ResetEvent(HANDLE hEvent);
 } // namespace kernel32
 
 namespace ntdll {
@@ -262,7 +262,7 @@ NTSTATUS WINAPI NtWriteFile(HANDLE FileHandle, HANDLE Event, PIO_APC_ROUTINE Apc
 	return status;
 }
 
-NTSTATUS WINAPI NtAllocateVirtualMemory(HANDLE ProcessHandle, PVOID *BaseAddress, ULONG_PTR ZeroBits,
+NTSTATUS WINAPI NtAllocateVirtualMemory(HANDLE ProcessHandle, GUEST_PTR *BaseAddress, ULONG_PTR ZeroBits,
 										PSIZE_T RegionSize, ULONG AllocationType, ULONG Protect) {
 	HOST_CONTEXT_GUARD();
 	DEBUG_LOG("NtAllocateVirtualMemory(%p, %p, %lu, %p, %lu, %lu) ", ProcessHandle, BaseAddress, ZeroBits, RegionSize,
@@ -276,20 +276,24 @@ NTSTATUS WINAPI NtAllocateVirtualMemory(HANDLE ProcessHandle, PVOID *BaseAddress
 		return STATUS_INVALID_PARAMETER;
 	}
 
-	wibo::heap::VmStatus vmStatus =
-		wibo::heap::virtualAlloc(reinterpret_cast<void **>(BaseAddress), reinterpret_cast<std::size_t *>(RegionSize),
-								 static_cast<DWORD>(AllocationType), static_cast<DWORD>(Protect));
+	void *baseAddress = fromGuestPtr(*BaseAddress);
+	size_t regionSize = static_cast<size_t>(*RegionSize);
+	wibo::heap::VmStatus vmStatus = wibo::heap::virtualAlloc(
+		&baseAddress, &regionSize, static_cast<DWORD>(AllocationType), static_cast<DWORD>(Protect));
 	if (vmStatus != wibo::heap::VmStatus::Success) {
 		NTSTATUS status = wibo::heap::ntStatusFromVmStatus(vmStatus);
 		DEBUG_LOG("-> 0x%x\n", status);
 		return status;
 	}
 
+	*BaseAddress = toGuestPtr(baseAddress);
+	*RegionSize = static_cast<SIZE_T>(regionSize);
+
 	DEBUG_LOG("-> 0x%x\n", STATUS_SUCCESS);
 	return STATUS_SUCCESS;
 }
 
-NTSTATUS WINAPI NtProtectVirtualMemory(HANDLE ProcessHandle, PVOID *BaseAddress, PSIZE_T NumberOfBytesToProtect,
+NTSTATUS WINAPI NtProtectVirtualMemory(HANDLE ProcessHandle, GUEST_PTR *BaseAddress, PSIZE_T NumberOfBytesToProtect,
 									   ULONG NewAccessProtection, PULONG OldAccessProtection) {
 	HOST_CONTEXT_GUARD();
 	DEBUG_LOG("NtProtectVirtualMemory(%p, %p, %p, %lu, %p) ", ProcessHandle, BaseAddress, NumberOfBytesToProtect,
@@ -303,8 +307,8 @@ NTSTATUS WINAPI NtProtectVirtualMemory(HANDLE ProcessHandle, PVOID *BaseAddress,
 		return STATUS_INVALID_PARAMETER;
 	}
 
-	void *base = *BaseAddress;
-	std::size_t length = static_cast<std::size_t>(*NumberOfBytesToProtect);
+	void *base = fromGuestPtr(*BaseAddress);
+	size_t length = static_cast<size_t>(*NumberOfBytesToProtect);
 	wibo::heap::VmStatus vmStatus =
 		wibo::heap::virtualProtect(base, length, static_cast<DWORD>(NewAccessProtection), OldAccessProtection);
 	if (vmStatus != wibo::heap::VmStatus::Success) {
@@ -351,7 +355,7 @@ NTSTATUS WINAPI NtQueryInformationProcess(HANDLE ProcessHandle, PROCESSINFOCLASS
 										  PVOID ProcessInformation, ULONG ProcessInformationLength,
 										  PULONG ReturnLength) {
 	HOST_CONTEXT_GUARD();
-	DEBUG_LOG("NtQueryInformationProcess(%p, %u, %p, %u, %p) ", ProcessHandle, ProcessInformationClass,
+	DEBUG_LOG("NtQueryInformationProcess(%d, %u, %p, %u, %p) ", ProcessHandle, ProcessInformationClass,
 			  ProcessInformation, ProcessInformationLength, ReturnLength);
 	if (!ProcessInformation) {
 		DEBUG_LOG("-> 0x%x\n", STATUS_INVALID_PARAMETER);
@@ -378,7 +382,7 @@ NTSTATUS WINAPI NtQueryInformationProcess(HANDLE ProcessHandle, PROCESSINFOCLASS
 		auto *info = reinterpret_cast<PROCESS_BASIC_INFORMATION *>(ProcessInformation);
 		std::memset(info, 0, sizeof(*info));
 		info->ExitStatus = static_cast<NTSTATUS>(details.exitCode);
-		info->PebBaseAddress = details.peb;
+		info->PebBaseAddress = toGuestPtr(details.peb);
 		DWORD_PTR processMask = 0;
 		DWORD_PTR systemMask = 0;
 		if (kernel32::GetProcessAffinityMask(ProcessHandle, &processMask, &systemMask)) {
@@ -440,7 +444,7 @@ NTSTATUS WINAPI NtQueryInformationProcess(HANDLE ProcessHandle, PROCESSINFOCLASS
 		size_t characterCount = widePath.empty() ? 0 : widePath.size() - 1;
 		unicode->Length = static_cast<unsigned short>(characterCount * sizeof(uint16_t));
 		unicode->MaximumLength = static_cast<unsigned short>(widePath.size() * sizeof(uint16_t));
-		unicode->Buffer = buffer;
+		unicode->Buffer = toGuestPtr(buffer);
 		DEBUG_LOG("-> 0x%x\n", STATUS_SUCCESS);
 		return STATUS_SUCCESS;
 	}

dll/ntdll.h

@@ -7,7 +7,7 @@ using PIO_APC_ROUTINE = PVOID;
 typedef struct _IO_STATUS_BLOCK {
 	union {
 		NTSTATUS Status;
-		PVOID Pointer;
+		GUEST_PTR Pointer;
 	};
 	ULONG_PTR Information;
 } IO_STATUS_BLOCK, *PIO_STATUS_BLOCK;
@@ -38,12 +38,13 @@ NTSTATUS WINAPI NtReadFile(HANDLE FileHandle, HANDLE Event, PIO_APC_ROUTINE ApcR
 NTSTATUS WINAPI NtWriteFile(HANDLE FileHandle, HANDLE Event, PIO_APC_ROUTINE ApcRoutine, PVOID ApcContext,
 							PIO_STATUS_BLOCK IoStatusBlock, PVOID Buffer, ULONG Length, PLARGE_INTEGER ByteOffset,
 							PULONG Key);
-NTSTATUS WINAPI NtAllocateVirtualMemory(HANDLE ProcessHandle, PVOID *BaseAddress, ULONG_PTR ZeroBits, PSIZE_T RegionSize,
-										ULONG AllocationType, ULONG Protect);
-NTSTATUS WINAPI NtProtectVirtualMemory(HANDLE ProcessHandle, PVOID *BaseAddress, PSIZE_T NumberOfBytesToProtect,
+NTSTATUS WINAPI NtAllocateVirtualMemory(HANDLE ProcessHandle, GUEST_PTR *BaseAddress, ULONG_PTR ZeroBits,
+										PSIZE_T RegionSize, ULONG AllocationType, ULONG Protect);
+NTSTATUS WINAPI NtProtectVirtualMemory(HANDLE ProcessHandle, GUEST_PTR *BaseAddress, PSIZE_T NumberOfBytesToProtect,
 									   ULONG NewAccessProtection, PULONG OldAccessProtection);
 NTSTATUS WINAPI RtlGetVersion(PRTL_OSVERSIONINFOW lpVersionInformation);
 NTSTATUS WINAPI NtQueryInformationProcess(HANDLE ProcessHandle, PROCESSINFOCLASS ProcessInformationClass,
-										  PVOID ProcessInformation, ULONG ProcessInformationLength, PULONG ReturnLength);
+										  PVOID ProcessInformation, ULONG ProcessInformationLength,
+										  PULONG ReturnLength);
 
 } // namespace ntdll

dll/ole32.cpp

@@ -119,7 +119,7 @@ HRESULT WINAPI CoInitialize(LPVOID pvReserved) {
 }
 
 HRESULT WINAPI CoCreateInstance(const GUID *rclsid, LPVOID pUnkOuter, DWORD dwClsContext, const GUID *riid,
-								LPVOID *ppv) {
+								GUEST_PTR *ppv) {
 	HOST_CONTEXT_GUARD();
 	DEBUG_LOG("STUB: CoCreateInstance(0x%x, %p, %d, 0x%x, %p)\n", rclsid->Data1, pUnkOuter, dwClsContext, riid->Data1,
 			  *ppv);

dll/ole32.h

@@ -5,7 +5,7 @@
 namespace ole32 {
 
 HRESULT WINAPI CoInitialize(LPVOID pvReserved);
-HRESULT WINAPI CoCreateInstance(const GUID *rclsid, LPVOID pUnkOuter, DWORD dwClsContext, const GUID *riid, LPVOID *ppv);
+HRESULT WINAPI CoCreateInstance(const GUID *rclsid, LPVOID pUnkOuter, DWORD dwClsContext, const GUID *riid, GUEST_PTR *ppv);
 HRESULT WINAPI CLSIDFromString(LPCWSTR lpsz, GUID *pclsid);
 
 } // namespace ole32

dll/rpcrt4.cpp

@@ -4,8 +4,8 @@
 #include "context.h"
 #include "heap.h"
 #include "modules.h"
+#include "types.h"
 
-#include <cstdarg>
 #include <cstdlib>
 #include <cstring>
 #include <memory>
@@ -44,7 +44,7 @@ struct BindingHandleData {
 };
 
 std::unordered_map<RPC_WSTR, BindingComponents> g_stringBindings;
-std::unordered_map<RPC_BINDING_HANDLE, std::unique_ptr<BindingHandleData>> g_bindingHandles;
+std::unordered_map<RPC_BINDING_HANDLE, wibo::heap::guest_ptr<BindingHandleData>> g_bindingHandles;
 
 std::u16string toU16(RPC_WSTR str) {
 	if (!str) {
@@ -112,7 +112,7 @@ BindingHandleData *getBinding(RPC_BINDING_HANDLE handle) {
 namespace rpcrt4 {
 
 RPC_STATUS WINAPI RpcStringBindingComposeW(RPC_WSTR objUuid, RPC_WSTR protSeq, RPC_WSTR networkAddr, RPC_WSTR endpoint,
-										   RPC_WSTR options, RPC_WSTR *stringBinding) {
+										   RPC_WSTR options, GUEST_PTR *stringBinding) {
 	HOST_CONTEXT_GUARD();
 	BindingComponents components;
 	components.objectUuid = toU16(objUuid);
@@ -136,18 +136,18 @@ RPC_STATUS WINAPI RpcStringBindingComposeW(RPC_WSTR objUuid, RPC_WSTR protSeq, R
 		buffer[length] = 0;
 		RPC_WSTR result = reinterpret_cast<RPC_WSTR>(buffer);
 		g_stringBindings[result] = components;
-		*stringBinding = result;
+		*stringBinding = toGuestPtr(result);
 	}
 
 	return RPC_S_OK;
 }
 
-RPC_STATUS WINAPI RpcBindingFromStringBindingW(RPC_WSTR stringBinding, RPC_BINDING_HANDLE *binding) {
+RPC_STATUS WINAPI RpcBindingFromStringBindingW(RPC_WSTR stringBinding, GUEST_PTR *binding) {
 	HOST_CONTEXT_GUARD();
 	if (!binding) {
 		return RPC_S_INVALID_ARG;
 	}
-	*binding = nullptr;
+	*binding = GUEST_NULL;
 	if (!stringBinding) {
 		return RPC_S_INVALID_STRING_BINDING;
 	}
@@ -155,13 +155,13 @@ RPC_STATUS WINAPI RpcBindingFromStringBindingW(RPC_WSTR stringBinding, RPC_BINDI
 	if (it == g_stringBindings.end()) {
 		return RPC_S_INVALID_STRING_BINDING;
 	}
-	auto handleData = std::make_unique<BindingHandleData>();
+	auto handleData = wibo::heap::make_guest_unique<BindingHandleData>();
 	handleData->components = it->second;
 	handleData->bindingString = composeString(handleData->components);
 	handleData->serverReachable = false;
 	RPC_BINDING_HANDLE handle = reinterpret_cast<RPC_BINDING_HANDLE>(handleData.get());
 	g_bindingHandles.emplace(handle, std::move(handleData));
-	*binding = handle;
+	*binding = toGuestPtr(handle);
 	DEBUG_LOG("RpcBindingFromStringBindingW(handle=%p)\n", handle);
 	return RPC_S_OK;
 }
@@ -190,12 +190,12 @@ RPC_STATUS WINAPI RpcBindingSetAuthInfoExW(RPC_BINDING_HANDLE binding, RPC_WSTR
 	return RPC_S_OK;
 }
 
-RPC_STATUS WINAPI RpcBindingFree(RPC_BINDING_HANDLE *binding) {
+RPC_STATUS WINAPI RpcBindingFree(GUEST_PTR *binding) {
 	HOST_CONTEXT_GUARD();
 	if (!binding) {
 		return RPC_S_INVALID_ARG;
 	}
-	RPC_BINDING_HANDLE handle = *binding;
+	RPC_BINDING_HANDLE handle = reinterpret_cast<RPC_BINDING_HANDLE>(fromGuestPtr(*binding));
 	if (!handle) {
 		return RPC_S_INVALID_BINDING;
 	}
@@ -204,17 +204,17 @@ RPC_STATUS WINAPI RpcBindingFree(RPC_BINDING_HANDLE *binding) {
 		return RPC_S_INVALID_BINDING;
 	}
 	g_bindingHandles.erase(it);
-	*binding = nullptr;
+	*binding = GUEST_NULL;
 	DEBUG_LOG("RpcBindingFree\n");
 	return RPC_S_OK;
 }
 
-RPC_STATUS WINAPI RpcStringFreeW(RPC_WSTR *string) {
+RPC_STATUS WINAPI RpcStringFreeW(GUEST_PTR *string) {
 	HOST_CONTEXT_GUARD();
 	if (!string) {
 		return RPC_S_INVALID_ARG;
 	}
-	RPC_WSTR value = *string;
+	RPC_WSTR value = reinterpret_cast<RPC_WSTR>(fromGuestPtr(*string));
 	if (!value) {
 		return RPC_S_OK;
 	}
@@ -223,7 +223,7 @@ RPC_STATUS WINAPI RpcStringFreeW(RPC_WSTR *string) {
 		g_stringBindings.erase(it);
 	}
 	std::free(reinterpret_cast<void *>(value));
-	*string = nullptr;
+	*string = GUEST_NULL;
 	return RPC_S_OK;
 }
 

dll/rpcrt4.h

@@ -15,25 +15,27 @@ struct RPC_SECURITY_QOS {
 	ULONG Capabilities;
 	ULONG IdentityTracking;
 	ULONG ImpersonationType;
-	PVOID AdditionalSecurityInfo;
+	GUEST_PTR AdditionalSecurityInfo;
 };
 
 union CLIENT_CALL_RETURN {
-	PVOID Pointer;
+	GUEST_PTR Pointer;
 	LONG_PTR Simple;
 };
 
 namespace rpcrt4 {
 
 RPC_STATUS WINAPI RpcStringBindingComposeW(RPC_WSTR objUuid, RPC_WSTR protSeq, RPC_WSTR networkAddr, RPC_WSTR endpoint,
-										   RPC_WSTR options, RPC_WSTR *stringBinding);
-RPC_STATUS WINAPI RpcBindingFromStringBindingW(RPC_WSTR stringBinding, RPC_BINDING_HANDLE *binding);
+										   RPC_WSTR options, GUEST_PTR *stringBinding);
+RPC_STATUS WINAPI RpcBindingFromStringBindingW(RPC_WSTR stringBinding, GUEST_PTR *binding);
 RPC_STATUS WINAPI RpcBindingSetAuthInfoExW(RPC_BINDING_HANDLE binding, RPC_WSTR serverPrincName, ULONG authnLevel,
 										   ULONG authnSvc, RPC_AUTH_IDENTITY_HANDLE authIdentity, ULONG authzSvc,
 										   RPC_SECURITY_QOS *securityQos);
-RPC_STATUS WINAPI RpcBindingFree(RPC_BINDING_HANDLE *binding);
-RPC_STATUS WINAPI RpcStringFreeW(RPC_WSTR *string);
+RPC_STATUS WINAPI RpcBindingFree(GUEST_PTR *binding);
+RPC_STATUS WINAPI RpcStringFreeW(GUEST_PTR *string);
+#ifndef __x86_64__ // TODO
 CLIENT_CALL_RETURN CDECL_NO_CONV NdrClientCall2(PMIDL_STUB_DESC stubDescriptor, PFORMAT_STRING format, ...);
+#endif
 VOID WINAPI NdrServerCall2(PRPC_MESSAGE message);
 
 } // namespace rpcrt4

dll/user32.cpp

@@ -12,7 +12,7 @@
 namespace user32 {
 
 constexpr uint32_t RT_STRING_ID = 6;
-constexpr uintptr_t kDefaultKeyboardLayout = 0x04090409;
+constexpr HKL kDefaultKeyboardLayout = 0x04090409;
 constexpr int UOI_FLAGS = 1;
 constexpr DWORD WSF_VISIBLE = 0x0001;
 
@@ -135,15 +135,14 @@ int WINAPI MessageBoxA(HWND hwnd, LPCSTR lpText, LPCSTR lpCaption, UINT uType) {
 
 HKL WINAPI GetKeyboardLayout(DWORD idThread) {
 	HOST_CONTEXT_GUARD();
-	DEBUG_LOG("GetKeyboardLayout(%u)\n", idThread);
+	DEBUG_LOG("STUB: GetKeyboardLayout(%u)\n", idThread);
 	(void)idThread;
-	return reinterpret_cast<HKL>(kDefaultKeyboardLayout);
+	return kDefaultKeyboardLayout;
 }
 
 HWINSTA WINAPI GetProcessWindowStation() {
-	DEBUG_LOG("GetProcessWindowStation()\n");
-	static int kWindowStationStub;
-	return reinterpret_cast<HWINSTA>(&kWindowStationStub);
+	DEBUG_LOG("STUB: GetProcessWindowStation()\n");
+	return NO_HANDLE;
 }
 
 BOOL WINAPI GetUserObjectInformationA(HANDLE hObj, int nIndex, PVOID pvInfo, DWORD nLength, LPDWORD lpnLengthNeeded) {
@@ -173,7 +172,7 @@ BOOL WINAPI GetUserObjectInformationA(HANDLE hObj, int nIndex, PVOID pvInfo, DWO
 
 HWND WINAPI GetActiveWindow() {
 	DEBUG_LOG("GetActiveWindow()\n");
-	return nullptr;
+	return NO_HANDLE;
 }
 
 } // namespace user32

dll/version.cpp

@@ -8,6 +8,7 @@
 #include "modules.h"
 #include "resources.h"
 #include "strutil.h"
+#include "types.h"
 
 #include <cstdio>
 #include <cstring>
@@ -239,7 +240,7 @@ UINT WINAPI GetFileVersionInfoA(LPCSTR lptstrFilename, DWORD dwHandle, DWORD dwL
 	return 1;
 }
 
-static unsigned int VerQueryValueImpl(const void *pBlock, const std::string &subBlock, void **lplpBuffer,
+static unsigned int VerQueryValueImpl(const void *pBlock, const std::string &subBlock, GUEST_PTR *lplpBuffer,
 									  unsigned int *puLen) {
 	if (!pBlock)
 		return 0;
@@ -264,20 +265,20 @@ static unsigned int VerQueryValueImpl(const void *pBlock, const std::string &sub
 		std::string narrow = wideStringToString(reinterpret_cast<const uint16_t *>(outPtr), static_cast<int>(outLen));
 		std::memcpy(dest, narrow.c_str(), narrow.size() + 1);
 		if (lplpBuffer)
-			*lplpBuffer = dest;
+			*lplpBuffer = toGuestPtr(dest);
 		if (puLen)
 			*puLen = static_cast<unsigned int>(narrow.size());
 		return 1;
 	}
 
 	if (lplpBuffer)
-		*lplpBuffer = const_cast<uint8_t *>(outPtr);
+		*lplpBuffer = toGuestPtr(outPtr);
 	if (puLen)
 		*puLen = outLen;
 	return 1;
 }
 
-UINT WINAPI VerQueryValueA(LPCVOID pBlock, LPCSTR lpSubBlock, LPVOID *lplpBuffer, PUINT puLen) {
+UINT WINAPI VerQueryValueA(LPCVOID pBlock, LPCSTR lpSubBlock, GUEST_PTR *lplpBuffer, PUINT puLen) {
 	HOST_CONTEXT_GUARD();
 	DEBUG_LOG("VerQueryValueA(%p, %s, %p, %p)\n", pBlock, lpSubBlock ? lpSubBlock : "(null)", lplpBuffer, puLen);
 	if (!lpSubBlock)
@@ -299,7 +300,7 @@ UINT WINAPI GetFileVersionInfoW(LPCWSTR lptstrFilename, DWORD dwHandle, DWORD dw
 	return GetFileVersionInfoA(narrow.c_str(), dwHandle, dwLen, lpData);
 }
 
-UINT WINAPI VerQueryValueW(LPCVOID pBlock, LPCWSTR lpSubBlock, LPVOID *lplpBuffer, PUINT puLen) {
+UINT WINAPI VerQueryValueW(LPCVOID pBlock, LPCWSTR lpSubBlock, GUEST_PTR *lplpBuffer, PUINT puLen) {
 	HOST_CONTEXT_GUARD();
 	if (!lpSubBlock)
 		return 0;

dll/version.h

@@ -6,9 +6,9 @@ namespace version {
 
 UINT WINAPI GetFileVersionInfoSizeA(LPCSTR lptstrFilename, LPDWORD lpdwHandle);
 UINT WINAPI GetFileVersionInfoA(LPCSTR lptstrFilename, DWORD dwHandle, DWORD dwLen, LPVOID lpData);
-UINT WINAPI VerQueryValueA(LPCVOID pBlock, LPCSTR lpSubBlock, LPVOID *lplpBuffer, PUINT puLen);
+UINT WINAPI VerQueryValueA(LPCVOID pBlock, LPCSTR lpSubBlock, GUEST_PTR *lplpBuffer, PUINT puLen);
 UINT WINAPI GetFileVersionInfoSizeW(LPCWSTR lptstrFilename, LPDWORD lpdwHandle);
 UINT WINAPI GetFileVersionInfoW(LPCWSTR lptstrFilename, DWORD dwHandle, DWORD dwLen, LPVOID lpData);
-UINT WINAPI VerQueryValueW(LPCVOID pBlock, LPCWSTR lpSubBlock, LPVOID *lplpBuffer, PUINT puLen);
+UINT WINAPI VerQueryValueW(LPCVOID pBlock, LPCWSTR lpSubBlock, GUEST_PTR *lplpBuffer, PUINT puLen);
 
 } // namespace version

src/files.cpp

@@ -1,11 +1,11 @@
 #include "files.h"
 #include "common.h"
+#include "errors.h"
 #include "handles.h"
 #include "strutil.h"
 
 #include <algorithm>
 #include <cerrno>
-#include <climits>
 #include <cstddef>
 #include <cstdio>
 #include <mutex>
@@ -312,7 +312,7 @@ HANDLE getStdHandle(DWORD nStdHandle) {
 	case STD_ERROR_HANDLE:
 		return stderrHandle;
 	default:
-		return (void *)0xFFFFFFFF;
+		return INVALID_HANDLE_VALUE;
 	}
 }
 

src/handles.cpp

@@ -1,4 +1,5 @@
 #include "handles.h"
+#include "types.h"
 #include <atomic>
 #include <cassert>
 #include <cstdint>
@@ -12,7 +13,7 @@ constexpr uint32_t kCompatMaxIndex = (0xFFFFu >> kHandleAlignShift) - 1;
 constexpr uint32_t kQuarantineLen = 64;
 
 inline uint32_t indexOf(HANDLE h) noexcept {
-	uint32_t v = static_cast<uint32_t>(reinterpret_cast<uintptr_t>(h));
+	uint32_t v = static_cast<uint32_t>(h);
 	if (v == 0 || (v & ((1U << kHandleAlignShift) - 1)) != 0) {
 		return UINT32_MAX;
 	}
@@ -21,7 +22,7 @@ inline uint32_t indexOf(HANDLE h) noexcept {
 
 inline HANDLE makeHandle(uint32_t index) noexcept {
 	uint32_t v = (index + 1) << kHandleAlignShift;
-	return reinterpret_cast<HANDLE>(static_cast<uintptr_t>(v));
+	return static_cast<HANDLE>(v);
 }
 
 inline bool isPseudo(HANDLE h) noexcept { return reinterpret_cast<int32_t>(h) < 0; }
@@ -83,7 +84,7 @@ HANDLE Handles::alloc(Pin<> obj, uint32_t grantedAccess, uint32_t flags) {
 }
 
 Pin<> Handles::get(HANDLE h, HandleMeta *metaOut) {
-	if (h == nullptr || isPseudo(h)) {
+	if (h == NO_HANDLE || isPseudo(h)) {
 		return {}; // pseudo-handles have no entries
 	}
 

src/heap.cpp

@@ -35,6 +35,7 @@
 namespace {
 
 constexpr uintptr_t kLowMemoryStart = 0x00110000UL; // 1 MiB + 64 KiB
+constexpr uintptr_t kHeapMax = 0x60000000UL;		// 1 GiB
 constexpr uintptr_t kTopDownStart = 0x7F000000UL;	// Just below 2GB
 constexpr uintptr_t kTwoGB = 0x80000000UL;
 constexpr std::size_t kGuestArenaSize = 512ULL * 1024ULL * 1024ULL; // 512 MiB
@@ -195,7 +196,7 @@ bool overlapsExistingMapping(uintptr_t base, std::size_t length) {
 		if (info.RegionSize == 0) {
 			continue;
 		}
-		uintptr_t mapStart = reinterpret_cast<uintptr_t>(info.BaseAddress);
+		uintptr_t mapStart = reinterpret_cast<uintptr_t>(fromGuestPtr(info.BaseAddress));
 		uintptr_t mapEnd = mapStart + static_cast<uintptr_t>(info.RegionSize);
 		if (mapEnd <= base) {
 			continue;
@@ -214,8 +215,8 @@ void recordGuestMapping(uintptr_t base, std::size_t size, DWORD allocationProtec
 		return;
 	}
 	MEMORY_BASIC_INFORMATION info{};
-	info.BaseAddress = reinterpret_cast<void *>(base);
-	info.AllocationBase = reinterpret_cast<void *>(base);
+	info.BaseAddress = toGuestPtr(reinterpret_cast<void *>(base));
+	info.AllocationBase = toGuestPtr(reinterpret_cast<void *>(base));
 	info.AllocationProtect = allocationProtect;
 	info.RegionSize = size;
 	info.State = state;
@@ -435,7 +436,7 @@ void initializeImpl() {
 
 	// Map and register guest arena (below 2GB, exclusive)
 	ArenaRange guest;
-	if (mapArena(kGuestArenaSize, kLowMemoryStart, kTopDownStart, true, "wibo guest arena", guest)) {
+	if (mapArena(kGuestArenaSize, kLowMemoryStart, kHeapMax, true, "wibo guest arena", guest)) {
 		bool ok = mi_manage_os_memory_ex(guest.start, guest.size,
 										 /*is_committed*/ false,
 										 /*is_pinned*/ false,
@@ -910,8 +911,8 @@ VmStatus virtualQuery(const void *address, MEMORY_BASIC_INFORMATION *outInfo) {
 			regionEnd = regionStart + pageSize;
 		}
 		allocLock.unlock();
-		outInfo->BaseAddress = reinterpret_cast<void *>(regionStart);
-		outInfo->AllocationBase = nullptr;
+		outInfo->BaseAddress = toGuestPtr(reinterpret_cast<void *>(regionStart));
+		outInfo->AllocationBase = GUEST_NULL;
 		outInfo->AllocationProtect = 0;
 		outInfo->RegionSize = regionEnd - regionStart;
 		outInfo->State = MEM_FREE;
@@ -961,8 +962,8 @@ VmStatus virtualQuery(const void *address, MEMORY_BASIC_INFORMATION *outInfo) {
 	DWORD finalProtect = committed ? pageProtect : PAGE_NOACCESS;
 	allocLock.unlock();
 
-	outInfo->BaseAddress = reinterpret_cast<void *>(blockStart);
-	outInfo->AllocationBase = reinterpret_cast<void *>(region->base);
+	outInfo->BaseAddress = toGuestPtr(reinterpret_cast<void *>(blockStart));
+	outInfo->AllocationBase = toGuestPtr(reinterpret_cast<void *>(region->base));
 	outInfo->AllocationProtect = allocationProtect;
 	outInfo->RegionSize = blockEnd - blockStart;
 	outInfo->State = committed ? MEM_COMMIT : MEM_RESERVE;
@@ -1153,7 +1154,7 @@ static size_t blockLower2GB(MEMORY_BASIC_INFORMATION mappings[MAX_NUM_MAPPINGS])
 		if (numMappings < MAX_NUM_MAPPINGS) {
 			if (numMappings > 0) {
 				auto &prevMapping = mappings[numMappings - 1];
-				uintptr_t prevMapStart = reinterpret_cast<uintptr_t>(prevMapping.BaseAddress);
+				uintptr_t prevMapStart = reinterpret_cast<uintptr_t>(fromGuestPtr(prevMapping.BaseAddress));
 				uintptr_t prevMapEnd = prevMapStart + prevMapping.RegionSize;
 				if (mapStart <= prevMapEnd) {
 					// Extend the previous mapping
@@ -1164,10 +1165,10 @@ static size_t blockLower2GB(MEMORY_BASIC_INFORMATION mappings[MAX_NUM_MAPPINGS])
 				}
 			}
 			mappings[numMappings++] = (MEMORY_BASIC_INFORMATION){
-				.BaseAddress = reinterpret_cast<void *>(mapStart),
-				.AllocationBase = reinterpret_cast<void *>(mapStart),
+				.BaseAddress = toGuestPtr(reinterpret_cast<void *>(mapStart)),
+				.AllocationBase = toGuestPtr(reinterpret_cast<void *>(mapStart)),
 				.AllocationProtect = PAGE_NOACCESS,
-				.RegionSize = mapEnd - mapStart,
+				.RegionSize = static_cast<SIZE_T>(mapEnd - mapStart),
 				.State = MEM_RESERVE,
 				.Protect = PAGE_NOACCESS,
 				.Type = 0, // external
@@ -1225,9 +1226,9 @@ __attribute__((used)) static void wibo_heap_constructor() {
 	g_mappings = new std::map<uintptr_t, MEMORY_BASIC_INFORMATION>;
 	for (size_t i = 0; i < numMappings; ++i) {
 		if (debug) {
-			fprintf(stderr, "Existing %zu: BaseAddress=%p, RegionSize=%lu\n", i, mappings[i].BaseAddress,
+			fprintf(stderr, "Existing %zu: BaseAddress=%x, RegionSize=%u\n", i, mappings[i].BaseAddress,
 					mappings[i].RegionSize);
 		}
-		g_mappings->emplace(reinterpret_cast<uintptr_t>(mappings[i].BaseAddress), mappings[i]);
+		g_mappings->emplace(reinterpret_cast<uintptr_t>(fromGuestPtr(mappings[i].BaseAddress)), mappings[i]);
 	}
 }

src/heap.h

@@ -5,6 +5,8 @@
 #include <cstddef>
 #include <cstdint>
 #include <cstdio>
+#include <limits>
+#include <memory>
 
 struct mi_heap_s;
 typedef struct mi_heap_s mi_heap_t;
@@ -50,4 +52,100 @@ NTSTATUS ntStatusFromVmStatus(VmStatus status);
 
 bool reserveGuestStack(std::size_t stackSizeBytes, void **outStackLimit, void **outStackBase);
 
+//-------------------- deleters --------------------
+template <class U> struct single_deleter {
+	void operator()(U *p) const noexcept {
+		if (!p)
+			return;
+		p->~U();
+		wibo::heap::guestFree(static_cast<void *>(p));
+	}
+};
+
+template <class U> struct array_deleter {
+	std::size_t n{}; // number of elements
+	void operator()(U *p) const noexcept {
+		if (!p)
+			return;
+		for (std::size_t i = n; i > 0; --i)
+			(p + (i - 1))->~U();
+		wibo::heap::guestFree(static_cast<void *>(p));
+	}
+};
+
+//-------------------- pointer alias picking T or T[] --------------------
+template <class T> struct unique_type {
+	using type = std::unique_ptr<T, single_deleter<T>>;
+};
+template <class T> struct unique_type<T[]> {
+	using type = std::unique_ptr<T[], array_deleter<T>>;
+};
+// template <class T, std::size_t> struct unique_type<T[]>; // no bounded arrays (int[N])
+
+template <class T> using guest_ptr = typename unique_type<T>::type;
+
+//-------------------- helpers --------------------
+inline bool mul_overflows(std::size_t a, std::size_t b) {
+	return b != 0 && a > (std::numeric_limits<std::size_t>::max)() / b;
+}
+
+//-------------------- single object --------------------
+template <class T, class... Args>
+	requires(!std::is_array_v<T>)
+guest_ptr<T> make_guest_unique(Args &&...args) noexcept {
+	// Optional: insist on nothrow construction in a no-exception build
+	static_assert(std::is_nothrow_constructible_v<T, Args...>,
+				  "T must be nothrow-constructible when exceptions are disabled");
+
+	void *raw = wibo::heap::guestMalloc(sizeof(T));
+	if (!raw)
+		return {};
+
+	// placement-new without exceptions
+	T *p = ::new (raw) T(std::forward<Args>(args)...);
+	return guest_ptr<T>(p, single_deleter<T>{});
+}
+
+//-------------------- unbounded array: default-construct --------------------
+template <class T>
+	requires std::is_unbounded_array_v<T>
+guest_ptr<T> make_guest_unique(std::size_t n) noexcept {
+	using U = std::remove_extent_t<T>;
+	static_assert(std::is_nothrow_default_constructible_v<U>,
+				  "U must be nothrow default-constructible when exceptions are disabled");
+
+	if (mul_overflows(n, sizeof(U)))
+		return {};
+	void *raw = wibo::heap::guestMalloc(n * sizeof(U));
+	if (!raw)
+		return {};
+	U *p = static_cast<U *>(raw);
+
+	for (std::size_t i = 0; i < n; ++i)
+		::new (p + i) U();
+
+	return guest_ptr<T>(p, array_deleter<U>{n});
+}
+
+//-------------------- unbounded array: per-element args --------------------
+template <class T, class... Args>
+	requires std::is_unbounded_array_v<T>
+guest_ptr<T> make_guest_unique(std::size_t n, Args &&...args) noexcept {
+	using U = std::remove_extent_t<T>;
+	static_assert(std::is_nothrow_constructible_v<U, Args...>,
+				  "U(args...) must be nothrow-constructible when exceptions are disabled");
+
+	if (mul_overflows(n, sizeof(U)))
+		return {};
+	void *raw = wibo::heap::guestMalloc(n * sizeof(U));
+	if (!raw)
+		return {};
+	U *p = static_cast<U *>(raw);
+
+	for (std::size_t i = 0; i < n; ++i)
+		::new (p + i) U(std::forward<Args>(args)...);
+
+	return guest_ptr<T>(p, array_deleter<U>{n});
+}
+
 } // namespace wibo::heap

src/loader.cpp

@@ -86,11 +86,11 @@ struct PESectionHeader {
 	uint32_t characteristics;
 };
 struct PEImportDirectoryEntry {
-	uint32_t *importLookupTable;
+	uint32_t importLookupTable;
 	uint32_t timeDateStamp;
 	uint32_t forwarderChain;
-	char *name;
-	uint32_t *importAddressTable;
+	uint32_t name;
+	uint32_t importAddressTable;
 };
 struct PEHintNameTableEntry {
 	uint16_t hint;
@@ -254,7 +254,9 @@ bool wibo::Executable::loadPE(FILE *file, bool exec) {
 		allocStatus = wibo::heap::virtualAlloc(
 			&allocatedBase, &allocationSize, MEM_RESERVE | MEM_COMMIT, initialProtect, MEM_IMAGE);
 	}
-	if (allocStatus != wibo::heap::VmStatus::Success) {
+	if (allocStatus == wibo::heap::VmStatus::Success) {
+		DEBUG_LOG("loadPE: mapped image at %p\n", allocatedBase);
+	} else {
 		DEBUG_LOG("Image mapping failed (status=%u)\n", static_cast<unsigned>(allocStatus));
 		imageBase = nullptr;
 		return false;
@@ -425,10 +427,10 @@ bool wibo::Executable::resolveImports() {
 	}
 
 	while (dir->name) {
-		char *dllName = fromRVA(dir->name);
+		char *dllName = fromRVA<char>(dir->name);
 		DEBUG_LOG("DLL Name: %s\n", dllName);
-		uint32_t *lookupTable = fromRVA(dir->importLookupTable);
-		uint32_t *addressTable = fromRVA(dir->importAddressTable);
+		uint32_t *lookupTable = fromRVA<uint32_t>(dir->importLookupTable);
+		uint32_t *addressTable = fromRVA<uint32_t>(dir->importAddressTable);
 
 		ModuleInfo *module = loadModule(dllName);
 		if (!module && kernel32::getLastError() != ERROR_MOD_NOT_FOUND) {

src/macros.S

@@ -0,0 +1,40 @@
+.intel_syntax noprefix
+
+.equ TEB_SELF, 0x18     # Self
+.equ TEB_FS_SEL, 0xf98  # CurrentFsSelector
+.equ TEB_GS_SEL, 0xf9a  # CurrentGsSelector
+
+#ifdef __i386__
+
+.equ TEB_SP, 0xf9c      # CurrentStackPointer
+
+#endif // __i386__
+
+#ifdef __x86_64__
+
+.equ TEB_SP, 0xfa0      # CurrentStackPointer
+.equ TEB_FSBASE, 0xfa8  # CurrentFsBase
+
+.equ CS_32, 0x23        # 32-bit code segment (Linux)
+.equ CS_64, 0x33        # 64-bit code segment (Linux)
+.equ DS_32, 0x2b        # 32-bit data segment (Linux)
+
+.macro LJMP32
+	jmp fword ptr [rip] # far jump into 32-bit code
+    .long 1f            # 32-bit code offset
+    .word CS_32         # 32-bit code segment (Linux)
+.code32
+1:
+    endbr32
+.endm
+
+.macro LJMP64
+    push CS_64          # 64-bit code segment (Linux)
+    push offset 1f      # 64-bit code offset
+    retf                # far jump into 64-bit code
+.code64
+1:
+    endbr64
+.endm
+
+#endif // __x86_64__

src/main.cpp

@@ -1,5 +1,4 @@
 #include "common.h"
-#include "context.h"
 #include "entry.h"
 #include "entry_trampolines.h"
 #include "files.h"
@@ -8,9 +7,15 @@
 #include "processes.h"
 #include "strutil.h"
 #include "tls.h"
+#include "types.h"
 #include "version_info.h"
 
+#ifdef __x86_64__
+#include "setup.h"
+#endif
+
 #include <asm/ldt.h>
+#include <asm/prctl.h>
 #include <cstdarg>
 #include <cstdio>
 #include <cstdlib>
@@ -53,13 +58,13 @@ void wibo::debug_log(const char *fmt, ...) {
 }
 
 TEB *wibo::allocateTib() {
-	auto *newTib = static_cast<TEB *>(std::calloc(1, sizeof(TEB)));
+	auto *newTib = static_cast<TEB *>(wibo::heap::guestCalloc(1, sizeof(TEB)));
 	if (!newTib) {
 		return nullptr;
 	}
 	tls::initializeTib(newTib);
-	newTib->Tib.Self = &newTib->Tib;
-	newTib->Peb = processPeb;
+	newTib->Tib.Self = toGuestPtr(newTib);
+	newTib->Peb = toGuestPtr(processPeb);
 	return newTib;
 }
 
@@ -82,8 +87,8 @@ void wibo::initializeTibStackInfo(TEB *tibPtr) {
 		fprintf(stderr, "Failed to reserve guest stack\n");
 		std::abort();
 	}
-	tibPtr->Tib.StackLimit = guestLimit;
-	tibPtr->Tib.StackBase = guestBase;
+	tibPtr->Tib.StackLimit = toGuestPtr(guestLimit);
+	tibPtr->Tib.StackBase = toGuestPtr(guestBase);
 	tibPtr->CurrentStackPointer = guestBase;
 	DEBUG_LOG("initializeTibStackInfo: using guest stack base=%p limit=%p\n", tibPtr->Tib.StackBase,
 			  tibPtr->Tib.StackLimit);
@@ -94,10 +99,18 @@ bool wibo::installTibForCurrentThread(TEB *tibPtr) {
 		return false;
 	}
 
+	currentThreadTeb = tibPtr;
+#ifdef __x86_64__
+	tibEntryNumber = x86_64_thread_setup(tibEntryNumber, tibPtr);
+	if (tibEntryNumber == -1 || tibPtr->CurrentFsSelector == 0) {
+		perror("x86_64_thread_setup failed");
+		return false;
+	}
+#else
 	struct user_desc desc;
 	std::memset(&desc, 0, sizeof(desc));
 	desc.entry_number = tibEntryNumber;
-	desc.base_addr = reinterpret_cast<unsigned int>(tibPtr);
+	desc.base_addr = reinterpret_cast<uintptr_t>(tibPtr);
 	desc.limit = static_cast<unsigned int>(sizeof(TEB) - 1);
 	desc.seg_32bit = 1;
 	desc.contents = 0;
@@ -118,13 +131,10 @@ bool wibo::installTibForCurrentThread(TEB *tibPtr) {
 
 	tibPtr->CurrentFsSelector = static_cast<uint16_t>((desc.entry_number << 3) | 3);
 	tibPtr->CurrentGsSelector = 0;
-	currentThreadTeb = tibPtr;
+#endif
 	return true;
 }
 
-// Make this global to ease debugging
-TEB tib;
-
 static std::string getExeName(const char *argv0) {
 	std::filesystem::path exePath(argv0 ? argv0 : "wibo");
 	return exePath.filename().string();
@@ -347,16 +357,18 @@ int main(int argc, char **argv) {
 
 	files::init();
 
+	// Create PEB
+	PEB *peb = reinterpret_cast<PEB *>(wibo::heap::guestCalloc(1, sizeof(PEB)));
+	peb->ProcessParameters = toGuestPtr(wibo::heap::guestCalloc(1, sizeof(RTL_USER_PROCESS_PARAMETERS)));
+
 	// Create TIB
-	memset(&tib, 0, sizeof(tib));
-	wibo::tls::initializeTib(&tib);
-	tib.Tib.Self = &tib.Tib;
-	tib.Peb = static_cast<PEB *>(calloc(1, sizeof(PEB)));
-	tib.Peb->ProcessParameters =
-		static_cast<RTL_USER_PROCESS_PARAMETERS *>(calloc(1, sizeof(RTL_USER_PROCESS_PARAMETERS)));
-	wibo::processPeb = tib.Peb;
-	wibo::initializeTibStackInfo(&tib);
-	if (!wibo::installTibForCurrentThread(&tib)) {
+	TEB *tib = reinterpret_cast<TEB *>(wibo::heap::guestCalloc(1, sizeof(TEB)));
+	wibo::tls::initializeTib(tib);
+	tib->Tib.Self = toGuestPtr(tib);
+	tib->Peb = toGuestPtr(peb);
+	wibo::processPeb = peb;
+	wibo::initializeTibStackInfo(tib);
+	if (!wibo::installTibForCurrentThread(tib)) {
 		fprintf(stderr, "Failed to install TIB for main thread\n");
 		return 1;
 	}
@@ -506,7 +518,7 @@ int main(int argc, char **argv) {
 	call_EntryProc(entryPoint);
 	DEBUG_LOG("We came back\n");
 	wibo::shutdownModuleRegistry();
-	wibo::tls::cleanupTib(&tib);
+	wibo::tls::cleanupTib(tib);
 
 	return 1;
 }

src/modules.cpp

@@ -6,9 +6,11 @@
 #include "errors.h"
 #include "files.h"
 #include "heap.h"
+#include "kernel32/errhandlingapi.h"
 #include "kernel32/internal.h"
 #include "strutil.h"
 #include "tls.h"
+#include "types.h"
 
 #include <algorithm>
 #include <array>
@@ -38,8 +40,65 @@ extern const wibo::ModuleStub lib_user32;
 extern const wibo::ModuleStub lib_vcruntime;
 extern const wibo::ModuleStub lib_version;
 
+// setup.S
+template <size_t Index> void stubThunk();
+
+/*
+ apiset api-ms-win-core-crt-l1-1-0 = msvcrt.dll
+ apiset api-ms-win-core-crt-l2-1-0 = msvcrt.dll
+ apiset api-ms-win-crt-conio-l1-1-0 = msvcrt.dll
+ apiset api-ms-win-crt-convert-l1-1-0 = msvcrt.dll
+ apiset api-ms-win-crt-environment-l1-1-0 = ucrtbase.dll
+ apiset api-ms-win-crt-filesystem-l1-1-0 = ucrtbase.dll
+ apiset api-ms-win-crt-heap-l1-1-0 = ucrtbase.dll
+ apiset api-ms-win-crt-locale-l1-1-0 = ucrtbase.dll
+ apiset api-ms-win-crt-math-l1-1-0 = ucrtbase.dll
+ apiset api-ms-win-crt-multibyte-l1-1-0 = ucrtbase.dll
+ apiset api-ms-win-crt-private-l1-1-0 = ucrtbase.dll
+ apiset api-ms-win-crt-process-l1-1-0 = ucrtbase.dll
+ apiset api-ms-win-crt-runtime-l1-1-0 = ucrtbase.dll
+ apiset api-ms-win-crt-stdio-l1-1-0 = ucrtbase.dll
+ apiset api-ms-win-crt-string-l1-1-0 = ucrtbase.dll
+ apiset api-ms-win-crt-time-l1-1-0 = ucrtbase.dll
+ apiset api-ms-win-crt-utility-l1-1-0 = ucrtbase.dll
+*/
+
 namespace {
 
+const std::array<std::pair<std::string_view, std::string_view>, 17> kApiSet = {
+	std::pair{"api-ms-win-core-crt-l1-1-0.dll", "msvcrt.dll"},
+	std::pair{"api-ms-win-core-crt-l2-1-0.dll", "msvcrt.dll"},
+	std::pair{"api-ms-win-crt-conio-l1-1-0.dll", "msvcrt.dll"},
+	std::pair{"api-ms-win-crt-convert-l1-1-0.dll", "msvcrt.dll"},
+	std::pair{"api-ms-win-crt-environment-l1-1-0.dll", "ucrtbase.dll"},
+	std::pair{"api-ms-win-crt-filesystem-l1-1-0.dll", "ucrtbase.dll"},
+	std::pair{"api-ms-win-crt-heap-l1-1-0.dll", "ucrtbase.dll"},
+	std::pair{"api-ms-win-crt-locale-l1-1-0.dll", "ucrtbase.dll"},
+	std::pair{"api-ms-win-crt-math-l1-1-0.dll", "ucrtbase.dll"},
+	std::pair{"api-ms-win-crt-multibyte-l1-1-0.dll", "ucrtbase.dll"},
+	std::pair{"api-ms-win-crt-private-l1-1-0.dll", "ucrtbase.dll"},
+	std::pair{"api-ms-win-crt-process-l1-1-0.dll", "ucrtbase.dll"},
+	std::pair{"api-ms-win-crt-runtime-l1-1-0.dll", "ucrtbase.dll"},
+	std::pair{"api-ms-win-crt-stdio-l1-1-0.dll", "ucrtbase.dll"},
+	std::pair{"api-ms-win-crt-string-l1-1-0.dll", "ucrtbase.dll"},
+	std::pair{"api-ms-win-crt-time-l1-1-0.dll", "ucrtbase.dll"},
+	std::pair{"api-ms-win-crt-utility-l1-1-0.dll", "ucrtbase.dll"},
+};
+
+const std::array<std::pair<std::string_view, std::string_view>, 11> kFallbacks = {
+	std::pair{"ucrtbase.dll", "msvcrt.dll"},
+	std::pair{"msvcrt40.dll", "msvcrt.dll"},
+	std::pair{"msvcr70.dll", "msvcrt.dll"},
+	std::pair{"msvcr71.dll", "msvcrt.dll"},
+	std::pair{"msvcr80.dll", "msvcrt.dll"},
+	std::pair{"msvcr90.dll", "msvcrt.dll"},
+	std::pair{"msvcr100.dll", "msvcrt.dll"},
+	std::pair{"msvcr110.dll", "msvcrt.dll"},
+	std::pair{"msvcr120.dll", "msvcrt.dll"},
+	std::pair{"msvcr130.dll", "msvcrt.dll"},
+	std::pair{"msvcr140.dll", "msvcrt.dll"},
+};
+
 constexpr DWORD DLL_PROCESS_DETACH = 0;
 constexpr DWORD DLL_PROCESS_ATTACH = 1;
 constexpr DWORD DLL_THREAD_ATTACH = 2;
@@ -69,6 +128,8 @@ size_t stubIndex = 0;
 std::array<std::string, MAX_STUBS> stubDlls;
 std::array<std::string, MAX_STUBS> stubFuncNames;
 std::unordered_map<std::string, StubFuncType> stubCache;
+std::unordered_map<HANDLE, std::shared_ptr<wibo::ModuleInfo>> g_modules;
+HANDLE g_nextStubHandle = 1;
 
 std::string makeStubKey(const char *dllName, const char *funcName) {
 	std::string key;
@@ -85,11 +146,6 @@ std::string makeStubKey(const char *dllName, const char *funcName) {
 	return key;
 }
 
-template <size_t Index> void stubThunk() {
-	// Call the appropriate entry trampoline
-	thunk_entry_stubBase(Index);
-}
-
 template <size_t... Indices>
 constexpr std::array<void (*)(void), sizeof...(Indices)> makeStubTable(std::index_sequence<Indices...>) {
 	return {{stubThunk<Indices>...}};
@@ -162,8 +218,11 @@ LockedRegistry registry() {
 	if (!reg.initialized) {
 		reg.initialized = true;
 		const wibo::ModuleStub *builtins[] = {
-			&lib_advapi32, &lib_bcrypt, &lib_crt,	 &lib_kernel32, &lib_lmgr,		&lib_mscoree, &lib_msvcrt,
-			&lib_ntdll,	   &lib_ole32,	&lib_rpcrt4, &lib_user32,	&lib_vcruntime, &lib_version, nullptr,
+			&lib_advapi32, &lib_bcrypt,	   /*&lib_crt,*/ &lib_kernel32,
+			&lib_lmgr,	   &lib_mscoree, /*&lib_msvcrt,*/
+			&lib_ntdll,	   &lib_ole32,	   &lib_rpcrt4,
+			&lib_user32,   &lib_vcruntime, &lib_version,
+			nullptr,
 		};
 		for (const wibo::ModuleStub **module = builtins; *module; ++module) {
 			registerBuiltinModule(reg, *module);
@@ -204,11 +263,11 @@ struct ParsedModuleName {
 	bool endsWithDot = false;
 };
 
-ParsedModuleName parseModuleName(const std::string &name) {
+ParsedModuleName parseModuleName(std::string_view name) {
 	ParsedModuleName parsed;
 	parsed.original = name;
 	parsed.base = name;
-	std::string sanitized = name;
+	std::string sanitized{name};
 	std::replace(sanitized.begin(), sanitized.end(), '/', '\\');
 	auto sep = sanitized.find_last_of('\\');
 	if (sep != std::string::npos) {
@@ -282,19 +341,20 @@ bool allocateModuleTlsForThread(wibo::ModuleInfo &module, TEB *tib) {
 	if (info.threadAllocations.find(tib) != info.threadAllocations.end()) {
 		return true;
 	}
-	void *block = nullptr;
+	GUEST_PTR block = GUEST_NULL;
 	const size_t allocationSize = info.allocationSize;
 	if (allocationSize > 0) {
-		block = wibo::heap::guestMalloc(allocationSize);
-		if (!block) {
+		void *ptr = wibo::heap::guestMalloc(allocationSize);
+		if (!ptr) {
 			DEBUG_LOG("  allocateModuleTlsForThread: failed to allocate %zu bytes for %s\n", allocationSize,
 					  module.originalName.c_str());
 			return false;
 		}
-		std::memset(block, 0, allocationSize);
+		std::memset(ptr, 0, allocationSize);
 		if (info.templateData && info.templateSize > 0) {
-			std::memcpy(block, info.templateData, info.templateSize);
+			std::memcpy(ptr, info.templateData, info.templateSize);
 		}
+		block = toGuestPtr(ptr);
 	}
 	info.threadAllocations.emplace(tib, block);
 	if (!wibo::tls::setValue(tib, info.index, block)) {
@@ -322,10 +382,10 @@ void freeModuleTlsForThread(wibo::ModuleInfo &module, TEB *tib) {
 	if (it == info.threadAllocations.end()) {
 		return;
 	}
-	void *block = it->second;
+	GUEST_PTR block = it->second;
 	info.threadAllocations.erase(it);
 	if (wibo::tls::getValue(tib, info.index) == block) {
-		if (!wibo::tls::setValue(tib, info.index, nullptr)) {
+		if (!wibo::tls::setValue(tib, info.index, GUEST_NULL)) {
 			DEBUG_LOG("  freeModuleTlsForThread: failed to clear TLS pointer for %s (index %u)\n",
 					  module.originalName.c_str(), info.index);
 		}
@@ -334,7 +394,7 @@ void freeModuleTlsForThread(wibo::ModuleInfo &module, TEB *tib) {
 		wibo::tls::clearModulePointer(tib, info.loaderIndex);
 	}
 	if (block) {
-		std::free(block);
+		wibo::heap::guestFree(fromGuestPtr(block));
 	}
 }
 
@@ -346,7 +406,7 @@ void runModuleTlsCallbacks(wibo::ModuleInfo &module, DWORD reason) {
 		if (!callback) {
 			continue;
 		}
-		call_PIMAGE_TLS_CALLBACK(callback, module.handle, reason, nullptr);
+		call_PIMAGE_TLS_CALLBACK(callback, reinterpret_cast<PVOID>(module.handle), reason, nullptr);
 	}
 }
 
@@ -502,7 +562,9 @@ void registerBuiltinModule(ModuleRegistry &reg, const wibo::ModuleStub *module)
 		return;
 	}
 	wibo::ModulePtr entry = std::make_shared<wibo::ModuleInfo>();
-	entry->handle = entry.get();
+	HANDLE handle = g_nextStubHandle++;
+	g_modules[handle] = entry;
+	entry->handle = handle;
 	entry->moduleStub = module;
 	entry->refCount = UINT_MAX;
 	entry->originalName = module->names[0] ? module->names[0] : "";
@@ -570,12 +632,10 @@ BOOL callDllMain(wibo::ModuleInfo &info, DWORD reason, LPVOID reserved) {
 			}
 		}
 
-		DEBUG_LOG("  callDllMain: invoking DllMain(%p, %u, %p) for %s\n",
-				  reinterpret_cast<HMODULE>(info.executable->imageBase), callReason, callReserved,
-				  info.normalizedName.c_str());
+		DEBUG_LOG("  callDllMain: invoking DllMain(%p, %u, %p) for %s\n", toGuestPtr(info.executable->imageBase),
+				  callReason, callReserved, info.normalizedName.c_str());
 
-		BOOL result =
-			call_DllEntryProc(dllMain, reinterpret_cast<HMODULE>(info.executable->imageBase), callReason, callReserved);
+		BOOL result = call_DllEntryProc(dllMain, toGuestPtr(info.executable->imageBase), callReason, callReserved);
 		DEBUG_LOG("  callDllMain: %s DllMain returned %d\n", info.normalizedName.c_str(), result);
 		return result;
 	};
@@ -739,7 +799,7 @@ ModuleInfo *registerProcessModule(std::unique_ptr<Executable> executable, std::f
 	std::string normalizedName = normalizedBaseKey(parsed);
 
 	ModulePtr info = std::make_unique<ModuleInfo>();
-	info->handle = executable->imageBase; // Use image base as handle for main module
+	info->handle = toGuestPtr(executable->imageBase); // Use image base as handle for main module
 	info->moduleStub = nullptr;
 	info->originalName = std::move(originalName);
 	info->normalizedName = std::move(normalizedName);
@@ -825,7 +885,7 @@ ModuleInfo *moduleInfoFromHandle(HMODULE module) {
 		if (info->handle == module) {
 			return info;
 		}
-		if (info->executable && info->executable->imageBase == module) {
+		if (info->executable && info->executable->imageBase == reinterpret_cast<void *>(module)) {
 			return info;
 		}
 	}
@@ -883,7 +943,7 @@ bool initializeModuleTls(ModuleInfo &module) {
 	info.callbacks.clear();
 	uintptr_t callbacksArray = resolveModuleAddress(exec, tlsDirectory->AddressOfCallBacks);
 	if (callbacksArray) {
-		auto callbackPtr = reinterpret_cast<uintptr_t *>(callbacksArray);
+		auto callbackPtr = reinterpret_cast<GUEST_PTR *>(callbacksArray);
 		while (callbackPtr && *callbackPtr) {
 			info.callbacks.push_back(reinterpret_cast<PIMAGE_TLS_CALLBACK>(resolveModuleAddress(exec, *callbackPtr)));
 			++callbackPtr;
@@ -937,13 +997,13 @@ bool initializeModuleTls(ModuleInfo &module) {
 	if (!ctx.success) {
 		for (auto &[thread, block] : info.threadAllocations) {
 			if (block) {
-				std::free(block);
+				wibo::heap::guestFree(fromGuestPtr(block));
 			}
 			if (info.loaderIndex != tls::kInvalidTlsIndex) {
 				wibo::tls::clearModulePointer(thread, info.loaderIndex);
 			}
 			if (wibo::tls::getValue(thread, info.index) == block) {
-				wibo::tls::setValue(thread, info.index, nullptr);
+				wibo::tls::setValue(thread, info.index, GUEST_NULL);
 			}
 		}
 		info.threadAllocations.clear();
@@ -971,14 +1031,14 @@ void releaseModuleTls(ModuleInfo &module) {
 	for (auto &[tib, block] : info.threadAllocations) {
 		if (tib) {
 			if (wibo::tls::getValue(tib, info.index) == block) {
-				wibo::tls::setValue(tib, info.index, nullptr);
+				wibo::tls::setValue(tib, info.index, GUEST_NULL);
 			}
 			if (info.loaderIndex != tls::kInvalidTlsIndex) {
 				wibo::tls::clearModulePointer(tib, info.loaderIndex);
 			}
 		}
 		if (block) {
-			std::free(block);
+			wibo::heap::guestFree(fromGuestPtr(block));
 		}
 	}
 	info.threadAllocations.clear();
@@ -1068,18 +1128,9 @@ ModuleInfo *findLoadedModule(const char *name) {
 	return info;
 }
 
-ModuleInfo *loadModule(const char *dllName) {
-	if (!dllName || *dllName == '\0') {
-		kernel32::setLastError(ERROR_INVALID_PARAMETER);
-		return nullptr;
-	}
-	std::string requested(dllName);
-	DEBUG_LOG("loadModule(%s)\n", requested.c_str());
-
+static ModuleInfo *loadModuleInternal(const std::string &dllName) {
 	auto reg = registry();
-
-	ParsedModuleName parsed = parseModuleName(requested);
-
+	ParsedModuleName parsed = parseModuleName(dllName);
 	DWORD diskError = ERROR_SUCCESS;
 
 	auto tryLoadExternal = [&](const std::filesystem::path &path) -> ModuleInfo * {
@@ -1090,7 +1141,7 @@ ModuleInfo *loadModule(const char *dllName) {
 			if (info->refCount != UINT_MAX) {
 				info->refCount++;
 			}
-			registerExternalModuleAliases(*reg, requested, files::canonicalPath(path), info);
+			registerExternalModuleAliases(*reg, dllName, files::canonicalPath(path), info);
 			return info;
 		}
 		reg.lock.unlock();
@@ -1115,9 +1166,11 @@ ModuleInfo *loadModule(const char *dllName) {
 		fclose(file);
 
 		ModulePtr info = std::make_unique<ModuleInfo>();
-		info->handle = info.get();
+		HANDLE handle = g_nextStubHandle++;
+		g_modules[handle] = info;
+		info->handle = handle;
 		info->moduleStub = nullptr;
-		info->originalName = requested;
+		info->originalName = dllName;
 		info->normalizedName = normalizedBaseKey(parsed);
 		info->resolvedPath = files::canonicalPath(path);
 		info->executable = std::move(executable);
@@ -1126,7 +1179,7 @@ ModuleInfo *loadModule(const char *dllName) {
 		reg.lock.lock();
 		ModuleInfo *raw = info.get();
 		reg->modulesByKey[key] = std::move(info);
-		registerExternalModuleAliases(*reg, requested, raw->resolvedPath, raw);
+		registerExternalModuleAliases(*reg, dllName, raw->resolvedPath, raw);
 		reg.lock.unlock();
 		ensureExportsInitialized(*raw);
 		if (!raw->executable->resolveImports()) {
@@ -1165,7 +1218,7 @@ ModuleInfo *loadModule(const char *dllName) {
 	};
 
 	auto resolveAndLoadExternal = [&]() -> ModuleInfo * {
-		auto resolvedPath = resolveModuleOnDisk(*reg, requested, false);
+		auto resolvedPath = resolveModuleOnDisk(*reg, dllName, false);
 		if (!resolvedPath) {
 			DEBUG_LOG("  module not found on disk\n");
 			diskError = ERROR_MOD_NOT_FOUND;
@@ -1177,7 +1230,7 @@ ModuleInfo *loadModule(const char *dllName) {
 	std::string alias = normalizedBaseKey(parsed);
 	ModuleInfo *existing = findByAlias(*reg, alias);
 	if (!existing) {
-		existing = findByAlias(*reg, normalizeAlias(requested));
+		existing = findByAlias(*reg, normalizeAlias(dllName));
 	}
 	if (existing) {
 		DEBUG_LOG("  found existing module alias %s (builtin=%d)\n", alias.c_str(), existing->moduleStub != nullptr);
@@ -1188,10 +1241,10 @@ ModuleInfo *loadModule(const char *dllName) {
 			DEBUG_LOG("  returning existing external module %s\n", existing->originalName.c_str());
 			return existing;
 		}
-		bool pinned = reg->pinnedModules.count(existing) != 0;
+		bool pinned = reg->pinnedModules.contains(existing);
 		if (!pinned) {
 			if (ModuleInfo *external = resolveAndLoadExternal()) {
-				DEBUG_LOG("  replaced builtin module %s with external copy\n", requested.c_str());
+				DEBUG_LOG("  replaced builtin module %s with external copy\n", dllName.c_str());
 				return external;
 			} else if (diskError != ERROR_MOD_NOT_FOUND) {
 				kernel32::setLastError(diskError);
@@ -1203,7 +1256,7 @@ ModuleInfo *loadModule(const char *dllName) {
 	}
 
 	if (ModuleInfo *external = resolveAndLoadExternal()) {
-		DEBUG_LOG("  loaded external module %s\n", requested.c_str());
+		DEBUG_LOG("  loaded external module %s\n", dllName.c_str());
 		return external;
 	} else if (diskError != ERROR_MOD_NOT_FOUND) {
 		kernel32::setLastError(diskError);
@@ -1217,7 +1270,7 @@ ModuleInfo *loadModule(const char *dllName) {
 		builtin = builtinIt->second;
 	}
 	if (!builtin) {
-		builtinIt = reg->builtinAliasMap.find(normalizeAlias(requested));
+		builtinIt = reg->builtinAliasMap.find(normalizeAlias(dllName));
 		if (builtinIt != reg->builtinAliasMap.end()) {
 			builtin = builtinIt->second;
 		}
@@ -1231,6 +1284,46 @@ ModuleInfo *loadModule(const char *dllName) {
 	return nullptr;
 }
 
+ModuleInfo *loadModule(const char* dllName) {
+	if (!dllName || *dllName == '\0') {
+		kernel32::setLastError(ERROR_INVALID_PARAMETER);
+		return nullptr;
+	}
+	DEBUG_LOG("loadModule(%s)\n", dllName);
+	std::string_view requested{dllName};
+
+	const auto parsed = parseModuleName(requested);
+	std::string normalized = normalizedBaseKey(parsed);
+
+	for (auto& [alias, module] : kApiSet) {
+		if (alias == normalized) {
+			DEBUG_LOG("  resolved api set %s -> %s\n", alias.data(), module.data());
+			requested = module;
+			normalized = module;
+			break;
+		}
+	}
+
+	DWORD lastError = kernel32::getLastError();
+	if (auto* info = loadModuleInternal(std::string{requested})) {
+		return info;
+	}
+	if (kernel32::getLastError() != ERROR_MOD_NOT_FOUND) {
+		return nullptr;
+	}
+
+	kernel32::setLastError(lastError);
+	for (auto& [module, fallback] : kFallbacks) {
+		if (module == normalized) {
+			DEBUG_LOG("  trying fallback %s -> %s\n", module.data(), fallback.data());
+			return loadModuleInternal(std::string{fallback});
+		}
+	}
+
+	kernel32::setLastError(ERROR_MOD_NOT_FOUND);
+	return nullptr;
+}
+
 void freeModule(ModuleInfo *info) {
 	auto reg = registry();
 	if (!info || info->refCount == UINT_MAX) {

src/modules.h

@@ -4,6 +4,7 @@
 #include "entry.h"
 #include "msvcrt.h"
 #include "tls.h"
+#include "types.h"
 
 #include <optional>
 #include <unordered_map>
@@ -39,8 +40,7 @@ class Executable {
 	bool findResource(const ResourceIdentifier &type, const ResourceIdentifier &name, std::optional<uint16_t> language,
 					  ResourceLocation &out) const;
 
-	template <typename T> T *fromRVA(uintptr_t rva) const { return (T *)(rva + (uint8_t *)imageBase); }
-	template <typename T> T *fromRVA(T *rva) const { return fromRVA<T>((uintptr_t)rva); }
+	template <typename T> T *fromRVA(uint32_t rva) const { return (T *)(rva + (uint8_t *)imageBase); }
 
 	void *imageBase = nullptr;
 	size_t imageSize = 0;
@@ -77,7 +77,7 @@ struct ModuleTlsInfo {
 	uint32_t characteristics = 0;
 	size_t allocationSize = 0;
 	std::vector<PIMAGE_TLS_CALLBACK> callbacks;
-	std::unordered_map<TEB *, void *> threadAllocations;
+	std::unordered_map<TEB *, GUEST_PTR> threadAllocations;
 };
 
 struct ModuleInfo {
@@ -140,8 +140,8 @@ ModuleInfo *moduleInfoFromAddress(void *addr);
  * otherwise it will be a pointer to a `wibo::ModuleInfo`.
  */
 inline bool isMainModule(HMODULE hModule) {
-	return hModule == nullptr || hModule == reinterpret_cast<HMODULE>(mainModule) ||
-		   (mainModule && mainModule->executable && hModule == mainModule->executable->imageBase);
+	return hModule == NO_HANDLE || (mainModule && mainModule->executable &&
+									reinterpret_cast<void *>(hModule) == mainModule->executable->imageBase);
 }
 
 } // namespace wibo

src/setup.S

@@ -0,0 +1,74 @@
+#include "macros.S"
+.section .note.GNU-stack, "", @progbits
+.text
+
+#ifdef __x86_64__
+
+# int x86_64_thread_setup(int entry_number, void *teb)
+.globl x86_64_thread_setup
+.type x86_64_thread_setup, @function
+x86_64_thread_setup:
+    push rbx                             # save rbx
+	mov r8, rsp                          # save host stack
+	rdfsbase r9                          # read host FS base
+	mov rdx, qword ptr [rsi+TEB_SP]      # fetch guest stack
+	LJMP32                               # far jump into 32-bit code
+	mov ax, 0x2b                         # user data segment (Linux)
+	mov ds, ax                           # setup data segment
+	mov es, ax                           # setup extra segment
+	mov esp, edx                         # switch to guest stack
+	sub esp, 0x10                        # sizeof(user_desc)
+	mov dword ptr [esp], edi             # entry_number (arg 0)
+	mov dword ptr [esp+4], esi           # base_addr (arg 1)
+	mov dword ptr [esp+8], 0xffff        # limit
+	mov dword ptr [esp+12], 0x41         # seg_32bit | usable
+	mov ebx, esp                         # &user_desc
+	mov eax, 0xf3                        # SYS_set_thread_area
+	int 0x80                             # syscall
+	test eax, eax                        # check for error
+	jnz 1f                               # skip selector setup
+	mov eax, dword ptr [esp]             # entry_number
+	cmp eax, -1                          # check for invalid entry_number
+	jz 2f                                # skip selector setup
+	lea ebx, [eax*8+3]                   # create selector
+	mov fs, bx                           # setup fs segment
+	mov word ptr [esi+TEB_FS_SEL], bx    # save selector
+	jmp 2f                               # skip error handling
+1:
+    mov eax, -1                          # return -1
+2:
+	add esp, 0x10                        # cleanup stack
+	LJMP64                               # far jump into 64-bit code
+    cdqe                                 # sign-extend eax to rax
+	mov rsp, r8                          # switch to host stack
+	wrfsbase r9                          # restore host FS base
+	pop rbx                              # restore rbx
+	ret
+.size x86_64_thread_setup, .-x86_64_thread_setup
+
+#endif // __x86_64__
+
+.altmacro
+.code32
+
+.macro stubThunkX number
+#ifdef __x86_64__
+.globl _Z9stubThunkILm\()\number\()EEvv
+.type  _Z9stubThunkILm\()\number\()EEvv, @function
+_Z9stubThunkILm\()\number\()EEvv:
+#else
+.globl _Z9stubThunkILj\()\number\()EEvv
+.type  _Z9stubThunkILj\()\number\()EEvv, @function
+_Z9stubThunkILj\()\number\()EEvv:
+#endif
+    pop eax
+    push \number
+    push eax
+    jmp thunk_entry_stubBase
+.endm
+
+.set i, 0
+.rept 256
+    stubThunkX %i
+    .set i, i+1
+.endr

src/setup.h

@@ -0,0 +1,13 @@
+#pragma once
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#ifdef __x86_64__
+int x86_64_thread_setup(int entry_number, void *teb);
+#endif
+
+#ifdef __cplusplus
+}
+#endif

src/tls.cpp

@@ -1,5 +1,7 @@
 #include "tls.h"
 #include "common.h"
+#include "heap.h"
+#include "types.h"
 
 #include <algorithm>
 #include <array>
@@ -20,7 +22,7 @@ size_t g_expansionCapacity = 0;
 
 struct TlsArray {
 	size_t capacity;
-	void *slots[];
+	GUEST_PTR slots[];
 };
 
 std::unordered_map<TEB *, TlsArray *> g_moduleArrays;
@@ -31,8 +33,8 @@ TlsArray *allocateTlsArray(size_t capacity) {
 	if (capacity == 0 || capacity > kMaxExpansionSlots) {
 		return nullptr;
 	}
-	const size_t bytes = sizeof(TlsArray) + capacity * sizeof(void *);
-	auto *arr = static_cast<TlsArray *>(std::calloc(1, bytes));
+	const size_t bytes = sizeof(TlsArray) + capacity * sizeof(GUEST_PTR);
+	auto *arr = static_cast<TlsArray *>(wibo::heap::guestCalloc(1, bytes));
 	if (!arr) {
 		return nullptr;
 	}
@@ -40,7 +42,7 @@ TlsArray *allocateTlsArray(size_t capacity) {
 	return arr;
 }
 
-inline TlsArray *arrayFromSlots(void *slots) {
+inline TlsArray *arrayFromSlots(GUEST_PTR slots) {
 	return slots ? reinterpret_cast<TlsArray *>(reinterpret_cast<uint8_t *>(slots) - offsetof(TlsArray, slots))
 				 : nullptr;
 }
@@ -56,7 +58,7 @@ void setExpansionArray(TEB *tib, TlsArray *arr) {
 	if (!tib) {
 		return;
 	}
-	tib->TlsExpansionSlots = arr ? arr->slots : nullptr;
+	tib->TlsExpansionSlots = arr ? toGuestPtr(arr->slots) : GUEST_NULL;
 }
 
 size_t chooseCapacity(size_t current, size_t required) {
@@ -202,7 +204,7 @@ bool ensureModuleArrayCapacityLocked(size_t required) {
 	}
 	for (auto &entry : pending) {
 		g_moduleArrays[entry.tib] = entry.newArray;
-		entry.tib->ThreadLocalStoragePointer = entry.newArray->slots;
+		entry.tib->ThreadLocalStoragePointer = toGuestPtr(entry.newArray->slots);
 		if (entry.oldArray) {
 			queueOldModuleArray(entry.tib, entry.oldArray);
 		}
@@ -214,7 +216,7 @@ bool ensureModuleArrayCapacityLocked(size_t required) {
 void zeroSlotForAllTibs(size_t index) {
 	if (index < kTlsSlotCount) {
 		for (TEB *tib : g_activeTibs) {
-			tib->TlsSlots[index] = nullptr;
+			tib->TlsSlots[index] = GUEST_NULL;
 		}
 		return;
 	}
@@ -224,7 +226,7 @@ void zeroSlotForAllTibs(size_t index) {
 		if (!arr || expansionIndex >= arr->capacity) {
 			continue;
 		}
-		arr->slots[expansionIndex] = nullptr;
+		arr->slots[expansionIndex] = GUEST_NULL;
 	}
 }
 
@@ -272,7 +274,7 @@ void cleanupTib(TEB *tib) {
 		}
 		g_moduleGarbage.erase(garbageIt);
 	}
-	tib->ThreadLocalStoragePointer = nullptr;
+	tib->ThreadLocalStoragePointer = GUEST_NULL;
 	auto it = std::find(g_activeTibs.begin(), g_activeTibs.end(), tib);
 	if (it != g_activeTibs.end()) {
 		g_activeTibs.erase(it);
@@ -316,9 +318,9 @@ bool isSlotAllocated(DWORD index) {
 	return index < wibo::tls::kTlsMaxSlotCount && g_slotUsed[index];
 }
 
-void *getValue(TEB *tib, DWORD index) {
+GUEST_PTR getValue(TEB *tib, DWORD index) {
 	if (!tib || index >= static_cast<DWORD>(wibo::tls::kTlsMaxSlotCount)) {
-		return nullptr;
+		return GUEST_NULL;
 	}
 	if (index < static_cast<DWORD>(kTlsSlotCount)) {
 		return tib->TlsSlots[index];
@@ -326,16 +328,16 @@ void *getValue(TEB *tib, DWORD index) {
 	std::lock_guard lock(g_tlsMutex);
 	auto *arr = getExpansionArray(tib);
 	if (!arr) {
-		return nullptr;
+		return GUEST_NULL;
 	}
 	size_t expansionIndex = static_cast<size_t>(index) - kTlsSlotCount;
 	if (expansionIndex >= arr->capacity) {
-		return nullptr;
+		return GUEST_NULL;
 	}
 	return arr->slots[expansionIndex];
 }
 
-bool setValue(TEB *tib, DWORD index, void *value) {
+bool setValue(TEB *tib, DWORD index, GUEST_PTR value) {
 	if (!tib || index >= static_cast<DWORD>(wibo::tls::kTlsMaxSlotCount)) {
 		return false;
 	}
@@ -357,9 +359,9 @@ bool setValue(TEB *tib, DWORD index, void *value) {
 	return true;
 }
 
-void *getValue(DWORD index) { return getValue(currentThreadTeb, index); }
+GUEST_PTR getValue(DWORD index) { return getValue(currentThreadTeb, index); }
 
-bool setValue(DWORD index, void *value) { return setValue(currentThreadTeb, index, value); }
+bool setValue(DWORD index, GUEST_PTR value) { return setValue(currentThreadTeb, index, value); }
 
 void forEachTib(void (*callback)(TEB *, void *), void *context) {
 	if (!callback) {
@@ -380,7 +382,7 @@ bool ensureModulePointerCapacity(size_t capacity) {
 	return ensureModuleArrayCapacityLocked(capacity);
 }
 
-bool setModulePointer(TEB *tib, size_t index, void *value) {
+bool setModulePointer(TEB *tib, size_t index, GUEST_PTR value) {
 	if (!tib) {
 		return false;
 	}
@@ -393,7 +395,7 @@ bool setModulePointer(TEB *tib, size_t index, void *value) {
 		return false;
 	}
 	array->slots[index] = value;
-	tib->ThreadLocalStoragePointer = array->slots;
+	tib->ThreadLocalStoragePointer = toGuestPtr(array->slots);
 	return true;
 }
 
@@ -406,7 +408,7 @@ void clearModulePointer(TEB *tib, size_t index) {
 	if (!array || index >= array->capacity) {
 		return;
 	}
-	array->slots[index] = nullptr;
+	array->slots[index] = GUEST_NULL;
 }
 
 } // namespace wibo::tls

src/tls.h

@@ -13,17 +13,17 @@ void cleanupTib(TEB *tib);
 void forEachTib(void (*callback)(TEB *, void *), void *context);
 
 bool ensureModulePointerCapacity(size_t capacity);
-bool setModulePointer(TEB *tib, size_t index, void *value);
+bool setModulePointer(TEB *tib, size_t index, GUEST_PTR value);
 void clearModulePointer(TEB *tib, size_t index);
 
 DWORD reserveSlot();
 bool releaseSlot(DWORD index);
 bool isSlotAllocated(DWORD index);
 
-void *getValue(TEB *tib, DWORD index);
-bool setValue(TEB *tib, DWORD index, void *value);
+GUEST_PTR getValue(TEB *tib, DWORD index);
+bool setValue(TEB *tib, DWORD index, GUEST_PTR value);
 
-void *getValue(DWORD index);
-bool setValue(DWORD index, void *value);
+GUEST_PTR getValue(DWORD index);
+bool setValue(DWORD index, GUEST_PTR value);
 
 } // namespace wibo::tls

src/types.h

@@ -4,11 +4,6 @@
 #define va_list __builtin_va_list
 #endif
 
-// On Windows, the incoming stack is aligned to a 4 byte boundary.
-// force_align_arg_pointer will realign the stack to match GCC's 16 byte alignment.
-#define WIN_ENTRY __attribute__((cdecl, force_align_arg_pointer))
-#define WIN_FUNC __attribute__((stdcall, force_align_arg_pointer))
-
 // Annotation macros for code generation
 #ifdef WIBO_CODEGEN
 #define WIBO_ANNOTATE(x) __attribute__((annotate(x)))
@@ -32,19 +27,41 @@
 #define _CC_STDCALL WIBO_ANNOTATE("CC:stdcall")
 
 // Instructs codegen to convert between calling conventions
+#ifdef __x86_64__
+#define WINAPI _CC_STDCALL
+#define CDECL _CC_CDECL
+#define CDECL_NO_CONV _CC_CDECL __attribute__((force_align_arg_pointer))
+#else
 #define WINAPI _CC_STDCALL __attribute__((fastcall))
 #define CDECL _CC_CDECL __attribute__((fastcall))
 #define CDECL_NO_CONV _CC_CDECL __attribute__((cdecl, force_align_arg_pointer))
+#endif
 
 // Used for host-to-guest calls
 #define GUEST_STDCALL __attribute__((stdcall))
 
+typedef unsigned int GUEST_PTR;
+constexpr GUEST_PTR GUEST_NULL = 0;
+
+#ifdef __x86_64__
+inline GUEST_PTR toGuestPtr(const void *addr) {
+	unsigned long long addr64 = reinterpret_cast<unsigned long long>(addr);
+	if (addr64 > 0xFFFFFFFF)
+		__builtin_unreachable();
+	return static_cast<GUEST_PTR>(addr64);
+}
+inline void *fromGuestPtr(GUEST_PTR addr) { return reinterpret_cast<void *>(addr); }
+#else
+inline GUEST_PTR toGuestPtr(const void *addr) { return static_cast<GUEST_PTR>(reinterpret_cast<unsigned long>(addr)); }
+inline void *fromGuestPtr(GUEST_PTR addr) { return reinterpret_cast<void *>(addr); }
+#endif
+
 using VOID = void;
-using HANDLE = VOID *;
-using HMODULE = VOID *;
-using HGLOBAL = HANDLE;
-using HLOCAL = HANDLE;
-using HRSRC = HANDLE;
+using HANDLE = int;
+using HMODULE = HANDLE;
+using HGLOBAL = GUEST_PTR;
+using HLOCAL = GUEST_PTR;
+using HRSRC = GUEST_PTR;
 using HINSTANCE = HANDLE;
 using LPHANDLE = HANDLE *;
 using PHANDLE = HANDLE *;
@@ -66,12 +83,9 @@ using ULONG = unsigned int;
 using PULONG = ULONG *;
 using LONGLONG = long long;
 using ULONGLONG = unsigned long long;
-using LONG_PTR = long;
-static_assert(sizeof(LONG_PTR) == sizeof(void *), "LONG_PTR must be pointer-sized");
-using ULONG_PTR = unsigned long;
-static_assert(sizeof(ULONG_PTR) == sizeof(void *), "ULONG_PTR must be pointer-sized");
-using UINT_PTR = unsigned long;
-static_assert(sizeof(UINT_PTR) == sizeof(void *), "UINT_PTR must be pointer-sized");
+using LONG_PTR = int;
+using ULONG_PTR = unsigned int;
+using UINT_PTR = unsigned int;
 using DWORD_PTR = ULONG_PTR;
 using PDWORD_PTR = DWORD_PTR *;
 using SHORT = short;
@@ -97,7 +111,7 @@ using BYTE = unsigned char;
 using BOOLEAN = unsigned char;
 using UINT = unsigned int;
 using PUINT = UINT *;
-using HKEY = VOID *;
+using HKEY = HANDLE;
 using PHKEY = HKEY *;
 using PSID = VOID *;
 using REGSAM = DWORD;
@@ -110,6 +124,8 @@ using PBYTE = BYTE *;
 using LPBYTE = BYTE *;
 using PWSTR = WCHAR *;
 
+constexpr HANDLE NO_HANDLE = 0;
+
 using NTSTATUS = LONG;
 using HRESULT = LONG;
 
@@ -271,20 +287,23 @@ constexpr SIZE_T kTlsSlotCount = 64;
 typedef struct _UNICODE_STRING {
 	USHORT Length;
 	USHORT MaximumLength;
-	PWSTR Buffer;
-} UNICODE_STRING, *PUNICODE_STRING;
+	GUEST_PTR Buffer;
+} UNICODE_STRING;
+typedef GUEST_PTR PUNICODE_STRING;
 
 typedef struct _RTL_USER_PROCESS_PARAMETERS {
 	BYTE Reserved1[16];
-	PVOID Reserved2[10];
+	GUEST_PTR Reserved2[10];
 	UNICODE_STRING ImagePathName;
 	UNICODE_STRING CommandLine;
 } RTL_USER_PROCESS_PARAMETERS, *PRTL_USER_PROCESS_PARAMETERS;
 
 typedef struct _LIST_ENTRY {
-	struct _LIST_ENTRY *Flink;
-	struct _LIST_ENTRY *Blink;
-} LIST_ENTRY, *PLIST_ENTRY, PRLIST_ENTRY;
+	GUEST_PTR Flink;
+	GUEST_PTR Blink;
+} LIST_ENTRY;
+typedef GUEST_PTR PLIST_ENTRY;
+typedef GUEST_PTR PRLIST_ENTRY;
 
 typedef struct _PEB_LDR_DATA {
 	BYTE Reserved1[8];
@@ -299,23 +318,24 @@ typedef struct _PEB {
 	BYTE Reserved1[2];
 	BYTE BeingDebugged;
 	BYTE Reserved2[1];
-	PVOID Reserved3[2];
-	PPEB_LDR_DATA Ldr;
-	PRTL_USER_PROCESS_PARAMETERS ProcessParameters;
-	PVOID Reserved4[3];
-	PVOID AtlThunkSListPtr;
-	PVOID Reserved5;
+	GUEST_PTR Reserved3[2];
+	GUEST_PTR Ldr;
+	GUEST_PTR ProcessParameters;
+	GUEST_PTR Reserved4[3];
+	GUEST_PTR AtlThunkSListPtr;
+	GUEST_PTR Reserved5;
 	ULONG Reserved6;
-	PVOID Reserved7;
+	GUEST_PTR Reserved7;
 	ULONG Reserved8;
 	ULONG AtlThunkSListPtr32;
-	PVOID Reserved9[45];
+	GUEST_PTR Reserved9[45];
 	BYTE Reserved10[96];
-	PPS_POST_PROCESS_INIT_ROUTINE PostProcessInitRoutine;
+	GUEST_PTR PostProcessInitRoutine;
 	BYTE Reserved11[128];
-	PVOID Reserved12[1];
+	GUEST_PTR Reserved12[1];
 	ULONG SessionId;
-} PEB, *PPEB;
+} PEB;
+typedef GUEST_PTR PPEB;
 
 struct CLIENT_ID {
 	HANDLE UniqueProcess;
@@ -325,7 +345,7 @@ struct CLIENT_ID {
 struct _ACTIVATION_CONTEXT;
 
 typedef struct _RTL_ACTIVATION_CONTEXT_STACK_FRAME {
-	struct _RTL_ACTIVATION_CONTEXT_STACK_FRAME *Previous;
+	GUEST_PTR Previous;
 	_ACTIVATION_CONTEXT *ActivationContext;
 	ULONG Flags;
 } RTL_ACTIVATION_CONTEXT_STACK_FRAME, *PRTL_ACTIVATION_CONTEXT_STACK_FRAME;
@@ -339,6 +359,7 @@ typedef struct _ACTIVATION_CONTEXT_STACK {
 } ACTIVATION_CONTEXT_STACK, *PACTIVATION_CONTEXT_STACK;
 
 #define GDI_BATCH_BUFFER_SIZE 0x136
+
 typedef struct _GDI_TEB_BATCH {
 	ULONG Offset;
 	HANDLE HDC;
@@ -346,83 +367,86 @@ typedef struct _GDI_TEB_BATCH {
 } GDI_TEB_BATCH, *PGDI_TEB_BATCH;
 
 typedef struct _NT_TIB {
-	struct _EXCEPTION_REGISTRATION_RECORD *ExceptionList;
-	PVOID StackBase;
-	PVOID StackLimit;
-	PVOID SubSystemTib;
+	GUEST_PTR ExceptionList;
+	GUEST_PTR StackBase;
+	GUEST_PTR StackLimit;
+	GUEST_PTR SubSystemTib;
 	union {
-		PVOID FiberData;
+		GUEST_PTR FiberData;
 		DWORD Version;
 	} DUMMYUNIONNAME;
-	PVOID ArbitraryUserPointer;
-	struct _NT_TIB *Self;
+	GUEST_PTR ArbitraryUserPointer;
+	GUEST_PTR Self;
 } NT_TIB, *PNT_TIB;
 
 typedef struct _TEB {
 	NT_TIB Tib;
-	PVOID EnvironmentPointer;
+	GUEST_PTR EnvironmentPointer;
 	CLIENT_ID ClientId;
-	PVOID ActiveRpcHandle;
-	PVOID ThreadLocalStoragePointer;
+	GUEST_PTR ActiveRpcHandle;
+	GUEST_PTR ThreadLocalStoragePointer;
 	PPEB Peb;
 	ULONG LastErrorValue;
 	ULONG CountOfOwnedCriticalSections;
-	PVOID CsrClientThread;
-	PVOID Win32ThreadInfo;
+	GUEST_PTR CsrClientThread;
+	GUEST_PTR Win32ThreadInfo;
 	ULONG Win32ClientInfo[31]; /* used for user32 private data in Wine */
-	PVOID WOW32Reserved;
+	GUEST_PTR WOW32Reserved;
 	ULONG CurrentLocale;
 	ULONG FpSoftwareStatusRegister;
-	PVOID SystemReserved1[54]; /* used for kernel32 private data in Wine */
-	PVOID Spare1;
+	GUEST_PTR SystemReserved1[54]; /* used for kernel32 private data in Wine */
+	GUEST_PTR Spare1;
 	LONG ExceptionCode;
-	PVOID ActivationContextStackPointer;
+	GUEST_PTR ActivationContextStackPointer;
 	BYTE SpareBytes1[36];
-	PVOID SystemReserved2[10]; /* used for ntdll private data in Wine */
+	GUEST_PTR SystemReserved2[10]; /* used for ntdll private data in Wine */
 	GDI_TEB_BATCH GdiTebBatch;
 	ULONG gdiRgn;
 	ULONG gdiPen;
 	ULONG gdiBrush;
 	CLIENT_ID RealClientId;
-	HANDLE GdiCachedProcessHandle;
+	GUEST_PTR GdiCachedProcessHandle;
 	ULONG GdiClientPID;
 	ULONG GdiClientTID;
-	PVOID GdiThreadLocaleInfo;
-	PVOID UserReserved[5];
-	PVOID glDispatchTable[280];
+	GUEST_PTR GdiThreadLocaleInfo;
+	GUEST_PTR UserReserved[5];
+	GUEST_PTR glDispatchTable[280];
 	ULONG glReserved1[26];
-	PVOID glReserved2;
-	PVOID glSectionInfo;
-	PVOID glSection;
-	PVOID glTable;
-	PVOID glCurrentRC;
-	PVOID glContext;
+	GUEST_PTR glReserved2;
+	GUEST_PTR glSectionInfo;
+	GUEST_PTR glSection;
+	GUEST_PTR glTable;
+	GUEST_PTR glCurrentRC;
+	GUEST_PTR glContext;
 	ULONG LastStatusValue;
 	UNICODE_STRING StaticUnicodeString;
 	WCHAR StaticUnicodeBuffer[261];
-	PVOID DeallocationStack;
-	PVOID TlsSlots[64];
+	GUEST_PTR DeallocationStack;
+	GUEST_PTR TlsSlots[64];
 	LIST_ENTRY TlsLinks;
-	PVOID Vdm;
-	PVOID ReservedForNtRpc;
-	PVOID DbgSsReserved[2];
+	GUEST_PTR Vdm;
+	GUEST_PTR ReservedForNtRpc;
+	GUEST_PTR DbgSsReserved[2];
 	ULONG HardErrorDisabled;
-	PVOID Instrumentation[16];
-	PVOID WinSockData;
+	GUEST_PTR Instrumentation[16];
+	GUEST_PTR WinSockData;
 	ULONG GdiBatchCount;
 	ULONG Spare2;
 	ULONG Spare3;
 	ULONG Spare4;
-	PVOID ReservedForOle;
+	GUEST_PTR ReservedForOle;
 	ULONG WaitingOnLoaderLock;
-	PVOID Reserved5[3];
-	PVOID *TlsExpansionSlots;
+	GUEST_PTR Reserved5[3];
+	GUEST_PTR TlsExpansionSlots;
 	// wibo
 	WORD CurrentFsSelector;
 	WORD CurrentGsSelector;
-	PVOID CurrentStackPointer;
-} TEB, *PTEB;
-
+	void *CurrentStackPointer;
+#ifdef __x86_64__
+	void *CurrentFsBase;
+#endif
+} TEB;
+typedef GUEST_PTR PTEB;
 #ifndef offsetof
 #define offsetof(type, member) __builtin_offsetof(type, member)
 #endif
@@ -436,8 +460,8 @@ static_assert(offsetof(TEB, DeallocationStack) == 0xE0C, "DeallocationStack offs
 static_assert(offsetof(TEB, TlsSlots) == 0xE10, "TLS slots offset mismatch");
 
 typedef struct _MEMORY_BASIC_INFORMATION {
-	PVOID BaseAddress;
-	PVOID AllocationBase;
+	GUEST_PTR BaseAddress;
+	GUEST_PTR AllocationBase;
 	DWORD AllocationProtect;
 	SIZE_T RegionSize;
 	DWORD State;
@@ -451,12 +475,12 @@ typedef struct _iobuf {
 	_iobuf() : _file(-1) {}
 	explicit _iobuf(int file) : _file(file) {}
 
-	char *_ptr = nullptr;
+	GUEST_PTR _ptr = GUEST_NULL;
 	int _cnt = 0;
-	char *_base = nullptr;
+	GUEST_PTR _base = GUEST_NULL;
 	int _flag = 0;
 	int _file;
 	int _charbuf = 0;
 	int _bufsiz = 0;
-	char *_tmpfname = nullptr;
+	GUEST_PTR _tmpfname = GUEST_NULL;
 } _FILE;

tools/gen_trampolines.py

@@ -20,7 +20,7 @@ import os
 import sys
 import tempfile
 from dataclasses import dataclass, field
-from enum import IntEnum
+from enum import Enum, IntEnum
 from pathlib import Path
 from typing import Iterable, List, Optional
 
@@ -50,6 +50,11 @@ if "LIBCLANG_PATH" in os.environ:
         )
 
 
+class Arch(str, Enum):
+    X86 = "x86"
+    X86_64 = "x86_64"
+
+
 class CallingConv(IntEnum):
     """CXCallingConv enum values from clang-c/Index.h"""
 
@@ -91,13 +96,30 @@ def _get_function_calling_conv(func_type: CXType) -> CallingConv:
     return CallingConv(_get_calling_conv(func_type))
 
 
+class ArgClass(str, Enum):
+    INT = "int"
+    MEMORY = "memory"
+
+
 @dataclass
 class ArgInfo:
+    type: Type
+    arg_class: ArgClass
+    sign_extended: bool
+
+
+@dataclass
+class ArgPlacement:
     size: int
     slot_size: int
-    primitive: bool
-    sign_extended: bool
-    type: Type
+    stack_offset: Optional[int] = None
+    register: Optional[str] = None
+
+    def __init__(self, arg: ArgInfo, arch: Arch):
+        self.size = arg.type.get_canonical().get_size()
+        self.slot_size = _slot_size_for_arch(arg, arch)
+        self.register = None
+        self.stack_offset = None
 
 
 @dataclass
@@ -108,7 +130,7 @@ class FuncInfo:
     source_cc: CallingConv
     target_cc: CallingConv
     variadic: bool
-    return_type: Type
+    return_type: ArgInfo
     args: List[ArgInfo] = field(default_factory=list)
 
 
@@ -118,7 +140,7 @@ class TypedefInfo:
     source_cc: CallingConv
     target_cc: CallingConv
     variadic: bool
-    return_type: Type
+    return_type: ArgInfo
     args: List[ArgInfo] = field(default_factory=list)
 
 
@@ -210,43 +232,146 @@ SIGNED_KINDS = [
 ]
 
 
+def _calculate_arg_info(t: Type) -> ArgInfo:
+    canonical = t.get_canonical()
+
+    # if canonical.kind == TypeKind.RECORD:
+    #     arg_class = ArgClass.MEMORY
+    # else:
+    arg_class = ArgClass.INT
+
+    if canonical.kind == TypeKind.POINTER:
+        pointee = canonical.get_pointee()
+        if pointee.kind == TypeKind.POINTER:
+            print(f"Bugprone: Pointer to pointer ({_type_to_string(t)})")
+
+    # Sign-extend signed integers and HANDLE-like typedefs
+    is_sign_extended = canonical.kind in SIGNED_KINDS or _is_handle_typedef(t)
+
+    return ArgInfo(
+        arg_class=arg_class,
+        sign_extended=is_sign_extended,
+        type=t,
+    )
+
+
 def _collect_args(func_type: CXType) -> List[ArgInfo]:
     """Collect argument information for a function."""
     args: List[ArgInfo] = []
     for t in func_type.argument_types():
-        size = t.get_size()
-        canonical = t.get_canonical()
-
-        # Determine if primitive (not struct/union)
-        is_primitive = canonical.kind != TypeKind.RECORD
-
-        # Determine if sign-extended
-        # Sign-extend signed integers and HANDLE-like typedefs
-        is_sign_extended = canonical in SIGNED_KINDS or _is_handle_typedef(t)
-
-        # Calculate stack slot size
-        if size <= 4:
-            slot_size = 4
-        elif size <= 8:
-            slot_size = 8
-        else:
-            raise NotImplementedError(
-                f"Argument size {size} not supported for function {func_type.spelling}"
-            )
-
-        args.append(
-            ArgInfo(
-                size=size,
-                slot_size=slot_size,
-                primitive=is_primitive,
-                sign_extended=is_sign_extended,
-                type=t,
-            )
-        )
+        args.append(_calculate_arg_info(t))
     return args
 
 
-def collect_functions(tu: TranslationUnit, ns_filter: Optional[str]) -> List[FuncInfo]:
+def _slot_size_for_arch(arg: ArgInfo, arch: Arch) -> int:
+    """Return the slot size (in bytes) used to pass an argument on the given architecture."""
+    canonical = arg.type.get_canonical()
+    if canonical.kind == TypeKind.POINTER:
+        return 8 if arch == Arch.X86_64 else 4
+    size = canonical.get_size()
+    if arch == Arch.X86:
+        if size <= 4:
+            return 4
+        if size <= 8:
+            return 8
+    elif arch == Arch.X86_64:
+        if size <= 8:
+            return 8
+    raise NotImplementedError(
+        f"Argument size {size} not supported for architecture {arch.value}"
+    )
+
+
+@dataclass
+class ArgLayout:
+    args: List[ArgPlacement]
+    stack_size: int
+
+
+def compute_arg_layout(
+    args: List[ArgInfo],
+    cc: CallingConv,
+    arch: Arch,
+    stack_offset: int = 0,
+    skip_args: int = 0,
+) -> ArgLayout:
+    """Compute how each argument is passed for the given calling convention and arch."""
+
+    placements: List[ArgPlacement] = []
+    stack_size = 0
+    gpr_order: List[str] = []
+    gpr_index = skip_args
+
+    if arch == Arch.X86 and cc == CallingConv.X86_FASTCALL:
+        gpr_order = ["ecx", "edx"]
+    elif arch == Arch.X86_64 and cc == CallingConv.C:
+        gpr_order = ["rdi", "rsi", "rdx", "rcx", "r8", "r9"]
+
+    # Offset our stack based on number of extra arguments
+    # We assume that every arg represented by skip_args fits in a register
+    register_size = 8 if arch == Arch.X86_64 else 4
+    consumed_stack = max(0, skip_args - len(gpr_order)) * register_size
+    stack_offset += consumed_stack
+    stack_size += consumed_stack
+
+    def _push_stack(arg: ArgInfo) -> None:
+        nonlocal stack_offset
+        nonlocal stack_size
+        placement = ArgPlacement(arg, arch)
+        placement.stack_offset = stack_offset
+        placements.append(placement)
+        stack_offset += placement.slot_size
+        stack_size += placement.slot_size
+
+    def _push_register(arg: ArgInfo) -> None:
+        nonlocal gpr_index
+        placement = ArgPlacement(arg, arch)
+        placement.register = gpr_order[gpr_index]
+        placements.append(placement)
+        gpr_index += 1
+
+    # Special case for x86 fastcall: stop using registers if any spill onto the stack
+    if arch == Arch.X86 and cc == CallingConv.X86_FASTCALL:
+        stack_args_start = 0
+        for i in range(min(len(gpr_order), len(args))):
+            if gpr_index >= len(gpr_order):
+                break
+            arg = args[i]
+            slot_size = _slot_size_for_arch(arg, arch)
+            if arg.arg_class == ArgClass.INT and slot_size == 4:
+                _push_register(arg)
+                stack_args_start += 1
+            else:
+                break
+
+        for i in range(stack_args_start, len(args)):
+            _push_stack(args[i])
+    else:
+        for arg in args:
+            slot_size = _slot_size_for_arch(arg, arch)
+            if (
+                arg.arg_class == ArgClass.INT
+                and slot_size <= register_size
+                and gpr_index < len(gpr_order)
+            ):
+                _push_register(arg)
+            else:
+                _push_stack(arg)
+
+    return ArgLayout(args=placements, stack_size=stack_size)
+
+
+def describe_arg_placement(placement: ArgPlacement) -> str:
+    if placement.register is not None:
+        return f"{placement.register}[{placement.slot_size}]"
+    if placement.stack_offset is not None:
+        return f"stack+{placement.stack_offset}[{placement.slot_size}]"
+    raise ValueError(f"Unassigned placement {placement}")
+
+
+def collect_functions(
+    tu: TranslationUnit, ns_filter: Optional[str], arch: Arch
+) -> List[FuncInfo]:
     want_ns = ns_filter.split("::") if ns_filter else None
     out: dict[str, FuncInfo] = {}
 
@@ -268,7 +393,7 @@ def collect_functions(tu: TranslationUnit, ns_filter: Optional[str]) -> List[Fun
                 source_cc=source_cc,
                 target_cc=_get_function_calling_conv(node.type),
                 variadic=node.type.is_function_variadic(),
-                return_type=node.type.get_result(),
+                return_type=_calculate_arg_info(node.type.get_result()),
                 args=_collect_args(node.type),
             )
 
@@ -292,7 +417,7 @@ def _type_to_string(t: CXType) -> str:
     return spelling
 
 
-def collect_typedefs(tu: TranslationUnit) -> List[TypedefInfo]:
+def collect_typedefs(tu: TranslationUnit, arch: Arch) -> List[TypedefInfo]:
     """Collect function pointer typedefs and type aliases from the translation unit."""
     out: dict[str, TypedefInfo] = {}
 
@@ -309,17 +434,13 @@ def collect_typedefs(tu: TranslationUnit) -> List[TypedefInfo]:
         if target_cc == CallingConv.DEFAULT:
             return  # No CC annotation; skip
 
-        variadic = func_type.is_function_variadic()
-        args = _collect_args(func_type)
-        return_type = func_type.get_result()
-
         out[name] = TypedefInfo(
             name=name,
             source_cc=source_cc,
             target_cc=target_cc,
-            variadic=variadic,
-            return_type=return_type,
-            args=args,
+            variadic=func_type.is_function_variadic(),
+            return_type=_calculate_arg_info(func_type.get_result()),
+            args=_collect_args(func_type),
         )
 
     def visit(node: Cursor) -> None:
@@ -372,17 +493,15 @@ def collect_variables(tu: TranslationUnit, ns_filter: Optional[str]) -> List[Var
     return sorted(out.values(), key=lambda v: v.name)
 
 
-def emit_cc_thunk(f: FuncInfo | TypedefInfo, lines: List[str]):
+def emit_cc_thunk32(f: FuncInfo | TypedefInfo, lines: List[str]):
     if isinstance(f, TypedefInfo):
         # Host-to-guest
-        target = "[eax+4]"
-        arg_off = 8
+        call_target = "[eax+4]"
         align = 0
         host_to_guest = True
     elif isinstance(f, FuncInfo):
         # Guest-to-host
-        target = f.mangled
-        arg_off = 4
+        call_target = f.mangled
         align = 16
         host_to_guest = False
 
@@ -391,49 +510,32 @@ def emit_cc_thunk(f: FuncInfo | TypedefInfo, lines: List[str]):
         assert f.source_cc == CallingConv.C and f.target_cc == CallingConv.C, (
             "Variadic functions must be cdecl"
         )
-        lines.append(f"\tjmp {target}")
+        lines.append(f"\tjmp {call_target}")
         return
 
-    # Compute argument stack offsets
-    offsets: List[int] = []
-    for arg in f.args:
-        offsets.append(arg_off)
-        arg_off += arg.slot_size
-
-    reg_indices: List[int] = []
-    if f.target_cc == CallingConv.X86_FASTCALL:
-        # Store the first two non-record 4-byte args in ECX/EDX for GCC/Clang x86 fastcall
-        if len(f.args) >= 1 and f.args[0].primitive and f.args[0].slot_size == 4:
-            reg_indices.append(0)  # ECX
-            if len(f.args) >= 2 and f.args[1].primitive and f.args[1].slot_size == 4:
-                reg_indices.append(1)  # EDX
-    elif f.target_cc == CallingConv.C or f.target_cc == CallingConv.X86_STDCALL:
-        # No register args for cdecl or stdcall
-        pass
-    else:
-        raise NotImplementedError(
-            f"Unsupported target calling convention {f.target_cc.name} for function {f.name}"
+    source_layout = compute_arg_layout(
+        f.args,
+        f.source_cc,
+        Arch.X86,
+        stack_offset=4,
+        skip_args=1 if host_to_guest else 0,
     )
+    target_layout = compute_arg_layout(f.args, f.target_cc, Arch.X86)
 
-    # Bytes we will push for the call (exclude args passed in registers)
-    stack_bytes = sum(
-        arg.slot_size for i, arg in enumerate(f.args) if i not in reg_indices
-    )
-
-    # Get current TIB
+    # Get current TEB
     if host_to_guest:
         lines.append("\tmov ecx, gs:[currentThreadTeb@ntpoff]")
     else:
-        lines.append("\tmov ecx, fs:[0x18]")
+        lines.append("\tmov ecx, fs:[TEB_SELF]")
 
     # Swap fs and gs
     lines.append("\tmov ax, fs")
-    lines.append("\tmov dx, word ptr [ecx+0xf98]")
-    lines.append("\tmov word ptr [ecx+0xf98], ax")
+    lines.append("\tmov dx, word ptr [ecx+TEB_FS_SEL]")
+    lines.append("\tmov word ptr [ecx+TEB_FS_SEL], ax")
     lines.append("\tmov fs, dx")
     lines.append("\tmov ax, gs")
-    lines.append("\tmov dx, word ptr [ecx+0xf9a]")
-    lines.append("\tmov word ptr [ecx+0xf9a], ax")
+    lines.append("\tmov dx, word ptr [ecx+TEB_GS_SEL]")
+    lines.append("\tmov word ptr [ecx+TEB_GS_SEL], ax")
     lines.append("\tmov gs, dx")
 
     # Store guest stack pointer in eax for arg access
@@ -442,48 +544,60 @@ def emit_cc_thunk(f: FuncInfo | TypedefInfo, lines: List[str]):
 
     # Swap stack pointer
     lines.append("\tpush ebp")
-    lines.append("\tmov ebp, dword ptr [ecx+0xf9c]")
-    lines.append("\tmov dword ptr [ecx+0xf9c], esp")
+    lines.append("\tmov ebp, dword ptr [ecx+TEB_SP]")
+    lines.append("\tmov dword ptr [ecx+TEB_SP], esp")
     lines.append("\tmov esp, ebp")
 
     # Allocate stack space for arguments
-    if stack_bytes > 0:
-        lines.append(f"\tsub esp, {stack_bytes}")
+    if target_layout.stack_size > 0:
+        lines.append(f"\tsub esp, {target_layout.stack_size}")
 
     # Align stack if needed (must be done after allocating args)
     if align > 0:
         lines.append(f"\tand esp, ~{align - 1}")
 
-    # Copy args onto stack
-    cur_off = 0
-    for i, arg in enumerate(f.args):
-        if i in reg_indices:
+    # Copy args onto stack for the callee
+    for idx, target in enumerate(target_layout.args):
+        if target.stack_offset is None:
             continue
-        base = offsets[i]
-        for part_off in range(0, arg.slot_size, 4):
-            lines.append(f"\tmov ecx, [eax+{base + part_off}]")
-            lines.append(f"\tmov [esp+{cur_off + part_off}], ecx")
-        cur_off += arg.slot_size
+
+        source = source_layout.args[idx]
+        if source.stack_offset is None:
+            raise NotImplementedError(
+                f"Source calling convention {f.source_cc.name} requires register argument {idx}; not implemented"
+            )
+
+        if source.slot_size != target.slot_size:
+            raise NotImplementedError(
+                f"Argument {idx} requires size conversion {source.slot_size}->{target.slot_size}; not implemented"
+            )
+
+        for off in range(0, target.slot_size, 4):
+            lines.append(f"\tmov ecx, [eax+{source.stack_offset + off}]")
+            lines.append(f"\tmov [esp+{target.stack_offset + off}], ecx")
 
     # Load args into registers as needed
-    if len(reg_indices) > 0:
-        i = reg_indices[0]
-        offset = offsets[i]
-        lines.append(f"\tmov ecx, [eax+{offset}]")
-    if len(reg_indices) > 1:
-        i = reg_indices[1]
-        offset = offsets[i]
-        lines.append(f"\tmov edx, [eax+{offset}]")
+    for idx, target in enumerate(target_layout.args):
+        if target.register is None:
+            continue
+
+        source = source_layout.args[idx]
+        if source.stack_offset is None:
+            raise NotImplementedError(
+                f"Source calling convention {f.source_cc.name} requires register argument {idx}; not implemented"
+            )
+
+        lines.append(f"\tmov {target.register}, [eax+{source.stack_offset}]")
 
     # Call into target
-    lines.append(f"\tcall {target}")
+    lines.append(f"\tcall {call_target}")
 
     # Determine if we can clobber eax/edx
-    if f.return_type.kind == TypeKind.RECORD:
+    if f.return_type.arg_class != ArgClass.INT:
         raise NotImplementedError(
-            f"Struct return type not supported for function {f.name}"
+            f"Unsupported return type class {f.return_type.arg_class.value} for function {f.name}"
         )
-    return_size = f.return_type.get_size()
+    return_size = f.return_type.type.get_size()
     save_eax = return_size > 0
     save_edx = return_size > 4
     if return_size > 8:
@@ -497,16 +611,16 @@ def emit_cc_thunk(f: FuncInfo | TypedefInfo, lines: List[str]):
     if save_edx:
         lines.append("\tpush edx")
     if host_to_guest:
-        lines.append("\tmov ecx, fs:[0x18]")
+        lines.append("\tmov ecx, fs:[TEB_SELF]")
     else:
         lines.append("\tmov ecx, gs:[currentThreadTeb@ntpoff]")
     lines.append("\tmov ax, fs")
-    lines.append("\tmov dx, word ptr [ecx+0xf98]")
-    lines.append("\tmov word ptr [ecx+0xf98], ax")
+    lines.append("\tmov dx, word ptr [ecx+TEB_FS_SEL]")
+    lines.append("\tmov word ptr [ecx+TEB_FS_SEL], ax")
     lines.append("\tmov fs, dx")
     lines.append("\tmov ax, gs")
-    lines.append("\tmov dx, word ptr [ecx+0xf9a]")
-    lines.append("\tmov word ptr [ecx+0xf9a], ax")
+    lines.append("\tmov dx, word ptr [ecx+TEB_GS_SEL]")
+    lines.append("\tmov word ptr [ecx+TEB_GS_SEL], ax")
     lines.append("\tmov gs, dx")
     if save_edx:
         lines.append("\tpop edx")
@@ -515,29 +629,293 @@ def emit_cc_thunk(f: FuncInfo | TypedefInfo, lines: List[str]):
 
     # Swap stack pointer
     lines.append("\tmov esp, ebp")  # Clean up arg space
-    lines.append("\tmov ebp, dword ptr [ecx+0xf9c]")
-    lines.append("\tmov dword ptr [ecx+0xf9c], esp")
+    lines.append("\tmov ebp, dword ptr [ecx+TEB_SP]")
+    lines.append("\tmov dword ptr [ecx+TEB_SP], esp")
 
     # Restore stack and frame pointer
     lines.append("\tleave")
 
     # Return to guest
-    if f.source_cc == CallingConv.X86_STDCALL:
-        ret_bytes = sum(arg.slot_size for arg in f.args)
-    elif f.source_cc == CallingConv.C:
-        ret_bytes = 0
-    else:
-        raise NotImplementedError(
-            f"Unsupported source calling convention {f.source_cc.name} for function {f.name}"
-        )
-    if ret_bytes > 0:
-        lines.append(f"\tret {ret_bytes}")
+    if f.source_cc == CallingConv.X86_STDCALL and source_layout.stack_size > 0:
+        lines.append(f"\tret {source_layout.stack_size}")
     else:
         lines.append("\tret")
 
 
+def _x64_register_by_slot_size(reg: str, slot_size: int) -> str:
+    if slot_size == 8:
+        return reg
+    if reg in ["rax", "rbx", "rcx", "rdx", "rsi", "rdi", "rbp", "rsp"]:
+        if slot_size == 4:
+            return f"e{reg[1:]}"
+        elif slot_size == 2:
+            return reg[1:]
+        elif slot_size == 1:
+            if reg in ["rax", "rbx", "rcx", "rdx"]:
+                return f"{reg[1]}l"
+            elif reg in ["rsi", "rdi"]:
+                return f"{reg[1]}il"
+            else:
+                return f"{reg[1]}pl"
+    if slot_size == 4:
+        return f"{reg}d"
+    if slot_size == 2:
+        return f"{reg}w"
+    if slot_size == 1:
+        return f"{reg}b"
+    raise NotImplementedError(f"Unsupported register {reg} for slot size {slot_size}")
+
+
+def _x64_ptr_type_by_slot_size(slot_size) -> str:
+    if slot_size == 4:
+        return "dword ptr"
+    elif slot_size == 8:
+        return "qword ptr"
+    else:
+        raise ValueError(f"Unsupported slot size {slot_size}")
+
+
+def emit_cc_thunk64(f: FuncInfo | TypedefInfo, lines: List[str]):
+    if isinstance(f, TypedefInfo):
+        # Host-to-guest
+        call_target = "edi"
+        align = 0
+        host_to_guest = True
+    elif isinstance(f, FuncInfo):
+        # Guest-to-host
+        call_target = f.mangled
+        align = 16
+        host_to_guest = False
+
+    if f.variadic:
+        # Variadic functions are not yet supported for calling convention conversion.
+        assert f.source_cc == CallingConv.C and f.target_cc == CallingConv.C, (
+            "Variadic functions must be cdecl"
+        )
+        lines.append(f"\tjmp {call_target}")
+        return
+
+    source_layout = compute_arg_layout(
+        f.args,
+        f.source_cc,
+        Arch.X86_64 if host_to_guest else Arch.X86,
+        stack_offset=24 if host_to_guest else 16,
+        skip_args=1 if host_to_guest else 0,
+    )
+    target_layout = compute_arg_layout(
+        f.args, f.target_cc, Arch.X86 if host_to_guest else Arch.X86_64
+    )
+
+    if host_to_guest:
+        lines.append(".code64")
+
+        # Save rbx and rbp
+        lines.append("\tpush rbx")
+        lines.append("\tpush rbp")
+
+        # Stash host stack in r10
+        lines.append("\tmov r10, rsp")
+
+        # Get current TEB
+        lines.append("\tmov rcx, fs:[currentThreadTeb@tpoff]")
+
+        # Save FS base
+        lines.append("\trdfsbase r9")
+        lines.append("\tmov qword ptr [rcx+TEB_FSBASE], r9")
+
+        # Save RSP and load guest stack
+        lines.append("\tmov rbp, qword ptr [rcx+TEB_SP]")
+        lines.append("\tmov qword ptr [rcx+TEB_SP], rsp")
+        lines.append("\tmov rsp, rbp")
+
+        # Allocate stack space for arguments
+        if target_layout.stack_size > 0:
+            lines.append(f"\tsub rsp, {target_layout.stack_size}")
+
+        # Align stack if needed (must be done after allocating args)
+        if align > 0:
+            lines.append(f"\tand rsp, ~{align - 1}")
+
+        # Transfer arguments
+        for i, target in enumerate(target_layout.args):
+            if target.stack_offset is None:
+                raise NotImplementedError(f"Unexpected register argument {target}")
+
+            source = source_layout.args[i]
+            if source.stack_offset is not None:
+                ptr_type = _x64_ptr_type_by_slot_size(source.slot_size)
+                register = _x64_register_by_slot_size("rax", target.slot_size)
+                lines.append(
+                    f"\tmov {register}, {ptr_type} [r10+{source.stack_offset}]"
+                )
+                ptr_type = _x64_ptr_type_by_slot_size(target.slot_size)
+                register = _x64_register_by_slot_size("rax", target.slot_size)
+            elif source.register is not None:
+                ptr_type = _x64_ptr_type_by_slot_size(target.slot_size)
+                register = _x64_register_by_slot_size(source.register, target.slot_size)
+            else:
+                raise ValueError(f"Argument {i} is not a register or stack offset")
+            lines.append(f"\tmov {ptr_type} [rsp+{target.stack_offset}], {register}")
+
+        # Jump to 32-bit mode
+        lines.append("\tLJMP32")
+
+        # Setup FS selector
+        lines.append("\tmov ax, word ptr [ecx+TEB_FS_SEL]")
+        lines.append("\tmov fs, ax")
+
+        # Call into target
+        lines.append(f"\tcall {call_target}")
+
+        # Get current TEB
+        lines.append("\tmov ecx, fs:[TEB_SELF]")
+
+        # Jump back to 64-bit
+        lines.append("\tLJMP64")
+
+        # Sign extend return value if necessary
+        if f.return_type.sign_extended:
+            lines.append("\tcdqe")
+
+        # Restore FS base
+        lines.append("\tmov r9, qword ptr [rcx+TEB_FSBASE]")
+        lines.append("\twrfsbase r9")
+
+        # Restore host stack
+        lines.append("\tmov rsp, qword ptr [rcx+TEB_SP]")
+        lines.append("\tmov qword ptr [rcx+TEB_SP], rbp")
+
+        # Restore rbp, rbx and return
+        lines.append("\tpop rbp")
+        lines.append("\tpop rbx")
+        lines.append("\tret")
+    else:
+        lines.append(".code32")
+
+        # Save registers
+        lines.append("\tpush ebp")
+        lines.append("\tpush esi")
+        lines.append("\tpush edi")
+
+        # Get current TEB
+        lines.append("\tmov ecx, fs:[TEB_SELF]")
+
+        # Save fs segment
+        lines.append("\tmov di, fs")
+        lines.append("\tmov word ptr [ecx+TEB_FS_SEL], di")
+
+        # Jump back to 64-bit
+        lines.append("\tLJMP64")
+
+        # Restore FS base
+        lines.append("\tmov r9, qword ptr [rcx+TEB_FSBASE]")
+        lines.append("\twrfsbase r9")
+
+        # Stash guest stack in r10
+        lines.append("\tmov r10, rsp")
+
+        # Restore host stack
+        lines.append("\tmov rbp, qword ptr [rcx+TEB_SP]")
+        lines.append("\tmov qword ptr [rcx+TEB_SP], rsp")
+        lines.append("\tmov rsp, rbp")
+
+        # Allocate stack space for arguments
+        if target_layout.stack_size > 0:
+            lines.append(f"\tsub rsp, {target_layout.stack_size}")
+
+        # Align stack if needed (must be done after allocating args)
+        if align > 0:
+            lines.append(f"\tand rsp, ~{align - 1}")
+
+        # Transfer args
+        for i, target in enumerate(target_layout.args):
+            arg = f.args[i]
+            source = source_layout.args[i]
+
+            if target.stack_offset is not None:
+                if source.stack_offset is not None:
+                    ptr_type = _x64_ptr_type_by_slot_size(source.slot_size)
+                    register = _x64_register_by_slot_size("rax", source.slot_size)
+                    lines.append(
+                        f"\tmov {register}, {ptr_type} [r10+{source.stack_offset}]"
+                    )
+                    ptr_type = _x64_ptr_type_by_slot_size(target.slot_size)
+                    register = _x64_register_by_slot_size("rax", target.slot_size)
+                elif source.register is not None:
+                    ptr_type = _x64_ptr_type_by_slot_size(target.slot_size)
+                    register = _x64_register_by_slot_size(
+                        source.register, target.slot_size
+                    )
+                else:
+                    raise ValueError(f"Argument {i} is not a register or stack offset")
+                lines.append(
+                    f"\tmov {ptr_type} [rsp+{target.stack_offset}], {register}"
+                )
+            elif target.register is not None:
+                ptr_type = _x64_ptr_type_by_slot_size(source.slot_size)
+                if source.slot_size == 4 and target.slot_size == 8:
+                    if arg.sign_extended:
+                        register = _x64_register_by_slot_size(
+                            target.register, source.slot_size
+                        )
+                        lines.append(
+                            f"\tmov {register}, {ptr_type} [r10+{source.stack_offset}]"
+                        )
+                        lines.append(f"\tmovsxd {target.register}, {register}")
+                    else:
+                        register = _x64_register_by_slot_size(
+                            target.register, source.slot_size
+                        )
+                        lines.append(
+                            f"\tmov {register}, {ptr_type} [r10+{source.stack_offset}]"
+                        )
+                elif source.slot_size == 8 and target.slot_size == 8:
+                    lines.append(
+                        f"\tmov {target.register}, {ptr_type} [r10+{source.stack_offset}]"
+                    )
+                else:
+                    raise NotImplementedError(
+                        f"Unsupported conversion from {source.slot_size} to {target.slot_size}"
+                    )
+
+        # Call into target
+        lines.append(f"\tcall {call_target}")
+
+        # Get current TEB
+        lines.append("\tmov rcx, fs:[currentThreadTeb@tpoff]")
+
+        # Restore host stack
+        lines.append("\tmov rsp, qword ptr [rcx+TEB_SP]")
+        lines.append("\tmov qword ptr [rcx+TEB_SP], rbp")
+
+        # Jump to 32-bit mode
+        lines.append("\tLJMP32")
+
+        # Setup FS selector
+        lines.append("\tmov di, word ptr [ecx+TEB_FS_SEL]")
+        lines.append("\tmov fs, di")
+
+        # Restore registers
+        lines.append("\tpop edi")
+        lines.append("\tpop esi")
+        lines.append("\tpop ebp")
+
+        # Return to guest
+        if f.source_cc == CallingConv.X86_STDCALL and source_layout.stack_size > 0:
+            lines.append(f"\tret {source_layout.stack_size}")
+        else:
+            lines.append("\tret")
+
+
+def emit_cc_thunk(f: FuncInfo | TypedefInfo, lines: List[str], arch: Arch):
+    if arch == Arch.X86_64:
+        return emit_cc_thunk64(f, lines)
+    elif arch == Arch.X86:
+        return emit_cc_thunk32(f, lines)
+
+
 def emit_guest_to_host_thunks(
-    lines: List[str], dll: str, funcs: Iterable[FuncInfo]
+    lines: List[str], dll: str, funcs: Iterable[FuncInfo], arch: Arch
 ) -> None:
     for f in funcs:
         thunk = f"thunk_{dll}_{f.name}"
@@ -545,19 +923,24 @@ def emit_guest_to_host_thunks(
         lines.append(
             f"# {f.qualified_ns}::{f.name} (source_cc={f.source_cc.name}, target_cc={f.target_cc.name}, variadic={f.variadic})"
         )
+        source_layout = compute_arg_layout(f.args, f.source_cc, Arch.X86)
+        target_layout = compute_arg_layout(f.args, f.target_cc, arch)
         for i, arg in enumerate(f.args):
-            lines.append(
-                f"\t# Arg {i} (slot_size={arg.slot_size}, primitive={arg.primitive}, sign_extended={arg.sign_extended})"
-            )
+            details: List[str] = []
+            details.append(f"src={describe_arg_placement(source_layout.args[i])}")
+            details.append(f"dst={describe_arg_placement(target_layout.args[i])}")
+            details.append(f"class={arg.arg_class.value}")
+            details.append(f"sign_extended={arg.sign_extended}")
+            lines.append(f"\t# Arg {i} ({', '.join(details)})")
         lines.append(f".globl {thunk}")
         lines.append(f".type {thunk}, @function")
         lines.append(f"{thunk}:")
-        emit_cc_thunk(f, lines)
+        emit_cc_thunk(f, lines, arch)
         lines.append(f".size {thunk}, .-{thunk}")
 
 
 def emit_host_to_guest_thunks(
-    lines: List[str], typedefs: Iterable[TypedefInfo]
+    lines: List[str], typedefs: Iterable[TypedefInfo], arch: Arch
 ) -> None:
     for f in typedefs:
         thunk = f"call_{f.name}"
@@ -565,14 +948,23 @@ def emit_host_to_guest_thunks(
         lines.append(
             f"# {f.name} (target_cc={f.target_cc.name}, variadic={f.variadic})"
         )
+        source_layout = compute_arg_layout(f.args, f.source_cc, arch, skip_args=1)
+        target_layout = compute_arg_layout(f.args, f.target_cc, Arch.X86)
         for i, arg in enumerate(f.args):
-            lines.append(
-                f"\t# Arg {i} (slot_size={arg.slot_size}, primitive={arg.primitive}, sign_extended={arg.sign_extended})"
-            )
+            details: List[str] = []
+            details.append(f"src={describe_arg_placement(source_layout.args[i])}")
+            details.append(f"dst={describe_arg_placement(target_layout.args[i])}")
+            details.append(f"class={arg.arg_class.value}")
+            details.append(f"sign_extended={arg.sign_extended}")
+            lines.append(f"\t# Arg {i} ({', '.join(details)})")
+        # details = []
+        # details.append(f"class={f.return_type.arg_class.value}")
+        # details.append(f"sign_extended={f.return_type.sign_extended}")
+        # lines.append(f"\t# Ret ({', '.join(details)})")
         lines.append(f".weak {thunk}")
         lines.append(f".type {thunk}, @function")
         lines.append(f"{thunk}:")
-        emit_cc_thunk(f, lines)
+        emit_cc_thunk(f, lines, arch)
         lines.append(f".size {thunk}, .-{thunk}")
 
 
@@ -581,6 +973,7 @@ def emit_header_mapping(
     funcs: Iterable[FuncInfo],
     typedefs: Iterable[TypedefInfo],
     variables: Iterable[VarInfo],
+    arch: Arch,
 ) -> str:
     guard = f"WIBO_GEN_{dll.upper()}_THUNKS_H"
     lines: List[str] = []
@@ -619,8 +1012,10 @@ def emit_header_mapping(
             type_str = _canonical_type_str(arg.type)
             args.append(f"{type_str} arg{i}")
         param_list = ", ".join(args)
-        return_type = _canonical_type_str(f.return_type)
-        if f.source_cc == CallingConv.X86_STDCALL:
+        return_type = _canonical_type_str(f.return_type.type)
+        if arch == Arch.X86_64:
+            cc_attr = ""
+        elif f.source_cc == CallingConv.X86_STDCALL:
             cc_attr = "__attribute__((stdcall)) "
         elif f.source_cc == CallingConv.C:
             cc_attr = "__attribute__((cdecl)) "
@@ -642,7 +1037,7 @@ def emit_header_mapping(
             params.append(f"{type_str} arg{i}")
 
         param_list = ", ".join(params)
-        return_type = _type_to_string(td.return_type)
+        return_type = _type_to_string(td.return_type.type)
         lines.append(f"{return_type} {thunk}({param_list});")
 
     lines.append("#ifdef __cplusplus\n}\n#endif")
@@ -672,7 +1067,7 @@ def main() -> int:
     ap.add_argument(
         "--namespace", dest="ns", default=None, help="Namespace filter, e.g. kernel32"
     )
-    ap.add_argument("--arch", choices=["x86"], default="x86")
+    ap.add_argument("--arch", choices=["x86", "x86_64"], default="x86")
     ap.add_argument(
         "--out-asm", type=Path, required=True, help="Output assembly file (.S)"
     )
@@ -682,10 +1077,17 @@ def main() -> int:
     ap.add_argument("-I", dest="incs", action="append", default=[])
     args = ap.parse_args()
 
+    if args.arch == "x86":
+        arch = Arch.X86
+    elif args.arch == "x86_64":
+        arch = Arch.X86_64
+    else:
+        raise ValueError(f"Unsupported architecture: {args.arch}")
+
     target = "i686-pc-linux-gnu" if args.arch == "x86" else "x86_64-pc-linux-gnu"
     tu = parse_tu(args.headers, args.incs, target)
-    funcs = collect_functions(tu, args.ns)
-    typedefs = collect_typedefs(tu)
+    funcs = collect_functions(tu, args.ns, arch)
+    typedefs = collect_typedefs(tu, arch)
     variables = collect_variables(tu, args.ns)
 
     if not funcs and not typedefs and not variables:
@@ -694,15 +1096,15 @@ def main() -> int:
 
     lines: List[str] = []
     lines.append("# Auto-generated thunks; DO NOT EDIT.")
-    lines.append(".intel_syntax noprefix")
+    lines.append('#include "macros.S"')
     lines.append('.section .note.GNU-stack, "", @progbits')
     lines.append(".text")
 
-    emit_guest_to_host_thunks(lines, args.dll, funcs)
-    emit_host_to_guest_thunks(lines, typedefs)
+    emit_guest_to_host_thunks(lines, args.dll, funcs, arch)
+    emit_host_to_guest_thunks(lines, typedefs, arch)
 
     asm = "\n".join(lines) + "\n"
-    hdr = emit_header_mapping(args.dll, funcs, typedefs, variables)
+    hdr = emit_header_mapping(args.dll, funcs, typedefs, variables, arch)
 
     args.out_asm.parent.mkdir(parents=True, exist_ok=True)
     args.out_hdr.parent.mkdir(parents=True, exist_ok=True)