Implement InterlockedCompareExchange using CAS

This implementation is only usable on 486 or newer CPUs, so there is
also code that detects at runtime whether the AC bit of the EFLAGS
register can be set. On 386, this register always returns the same
value.

The CAS implementation uses the "lock cmpxchg" instruction, which is
only executed if the above runtime check is satisfied. The overhead of
calling this implementation is a "cmp" followed by a "jnz" instruction,
which should have a negligible cost when execution already had to cross
a DLL boundary to get here.

Using the command

find -name "*.dll" -o -name "*.vxd" | perl -nE "print if grep { $_ =~ /InterlockedCompareExchange/ } `dumpbin /imports $_`;"

Here is a list of .dll and .vxd files in the installer that import the
"InterlockedCompareExchange" symbol and thus benefit from this:

$WINDIR/Microsoft.NET/Framework/sbscmp10.dll
$WINDIR/Microsoft.NET/Framework/sbscmp20_mscorwks.dll
$WINDIR/Microsoft.NET/Framework/sbscmp20_perfcounter.dll
$WINDIR/Microsoft.NET/Framework/SharedReg12.dll
$WINDIR/Microsoft.NET/Framework/v1.0.3705/mscormmc.dll
$WINDIR/Microsoft.NET/Framework/v2.0.50727/AdoNetDiag.dll
$WINDIR/Microsoft.NET/Framework/v2.0.50727/alink.dll
$WINDIR/Microsoft.NET/Framework/v2.0.50727/aspnet_filter.dll
$WINDIR/Microsoft.NET/Framework/v2.0.50727/aspnet_isapi.dll
$WINDIR/Microsoft.NET/Framework/v2.0.50727/Aspnet_perf.dll
$WINDIR/Microsoft.NET/Framework/v2.0.50727/CORPerfMonExt.dll
$WINDIR/Microsoft.NET/Framework/v2.0.50727/cscomp.dll
$WINDIR/Microsoft.NET/Framework/v2.0.50727/Culture.dll
$WINDIR/Microsoft.NET/Framework/v2.0.50727/dfdll.dll
$WINDIR/Microsoft.NET/Framework/v2.0.50727/diasymreader.dll
$WINDIR/Microsoft.NET/Framework/v2.0.50727/fusion.dll
$WINDIR/Microsoft.NET/Framework/v2.0.50727/MmcAspExt.dll
$WINDIR/Microsoft.NET/Framework/v2.0.50727/mscordacwks.dll
$WINDIR/Microsoft.NET/Framework/v2.0.50727/mscordbc.dll
$WINDIR/Microsoft.NET/Framework/v2.0.50727/mscordbi.dll
$WINDIR/Microsoft.NET/Framework/v2.0.50727/mscorie.dll
$WINDIR/Microsoft.NET/Framework/v2.0.50727/mscorjit.dll
$WINDIR/Microsoft.NET/Framework/v2.0.50727/mscorld.dll
$WINDIR/Microsoft.NET/Framework/v2.0.50727/mscorpe.dll
$WINDIR/Microsoft.NET/Framework/v2.0.50727/mscorsec.dll
$WINDIR/Microsoft.NET/Framework/v2.0.50727/mscorsn.dll
$WINDIR/Microsoft.NET/Framework/v2.0.50727/mscorsvc.dll
$WINDIR/Microsoft.NET/Framework/v2.0.50727/mscortim.dll
$WINDIR/Microsoft.NET/Framework/v2.0.50727/mscorwks.dll
$WINDIR/Microsoft.NET/Framework/v2.0.50727/normalization.dll
$WINDIR/Microsoft.NET/Framework/v2.0.50727/PerfCounter.dll
$WINDIR/Microsoft.NET/Framework/v2.0.50727/peverify.dll
$WINDIR/Microsoft.NET/Framework/v2.0.50727/sbscmp20_mscorlib.dll
$WINDIR/Microsoft.NET/Framework/v2.0.50727/shfusion.dll
$WINDIR/Microsoft.NET/Framework/v2.0.50727/ShFusRes.dll
$WINDIR/Microsoft.NET/Framework/v2.0.50727/System.Data.dll
$WINDIR/Microsoft.NET/Framework/v2.0.50727/System.Data.OracleClient.dll
$WINDIR/Microsoft.NET/Framework/v2.0.50727/System.EnterpriseServices.Wrapper.dll
$WINDIR/Microsoft.NET/Framework/v2.0.50727/System.Transactions.dll
$WINDIR/Microsoft.NET/Framework/v2.0.50727/VsaVb7rt.dll
$WINDIR/Microsoft.NET/Framework/v2.0.50727/webengine.dll
$WINDIR/Microsoft.NET/Framework/v2.0.50727/WMINet_Utils.dll
$WINDIR/RegisteredPackages/{D5D40355-5FB0-48fb-A231-CDC637FA16E0}/NETFXMigration.dll
$WINDIR/System/dfshim.dll
$WINDIR/System/mscoree.dll
$WINDIR/System/mscories.dll
$WINDIR/System/msvcm80.dll
$WINDIR/System/msvcp80.dll
$WINDIR/System/WBEM/Wmidcad.dll
$WINDIR/winsxs/x86_microsoft.vc80.crt_1fc8b3b9a1e18e3b_8.0.50727.42_none_db5f52fb98cb24ad/msvcm80.dll
$WINDIR/winsxs/x86_microsoft.vc80.crt_1fc8b3b9a1e18e3b_8.0.50727.42_none_db5f52fb98cb24ad/msvcp80.dll
$WINDIR/winsxs/x86_Microsoft.VC80.CRT_1fc8b3b9a1e18e3b_8.0.50727.42_x-ww_0de06acd/msvcm80.dll
$WINDIR/winsxs/x86_Microsoft.VC80.CRT_1fc8b3b9a1e18e3b_8.0.50727.42_x-ww_0de06acd/msvcp80.dll

wrappers/CMakeLists.txt

@@ -1,30 +1,54 @@
 cmake_minimum_required(VERSION 3.13)
 
-project(corkel32 C)
+project(corkel32 C ASM_MASM)
 
-set(CMAKE_BUILD_TYPE Release)
+add_compile_definitions(_WIN32_WINNT=0x0400)
 
-add_library(corkdebug OBJECT
+add_library(
+  corkdebug OBJECT
   debug.c
 )
-link_libraries(corkdebug)
+target_compile_definitions(corkdebug PRIVATE _CRT_SECURE_NO_WARNINGS=1)
+
+add_library(
+  detect486 OBJECT
+  detect486.asm
+)
+
+add_library(
+  cas OBJECT
+  cas.asm
+)
+
+include(CheckSymbolExists)
+check_symbol_exists(RtlUnwind Windows.h HAS_RTL_UNWIND)
 
 # KERNEL32 => CORKEL32
-add_library(corkel32 SHARED
+add_library(
+  corkel32 MODULE
   kernel32.c
   advapi32.c
   comdlg32.c
   corkel32.def
 )
+target_link_libraries(corkel32 PRIVATE corkdebug detect486 cas)
+target_compile_definitions(corkel32 PRIVATE _CRT_SECURE_NO_WARNINGS=1)
 
 # NTDLL => CORNT
-add_library(cornt SHARED
+add_library(
+  cornt MODULE
   ntdll.c
   cornt.def
 )
+target_link_libraries(cornt PRIVATE corkdebug)
+if(HAS_RTL_UNWIND)
+  target_compile_definitions(cornt PRIVATE HAS_RTL_UNWIND=1)
+endif()
 
 # USER32 => CORUSR
-add_library(corusr SHARED
+add_library(
+  corusr MODULE
   user32.c
   corusr.def
 )
+target_link_libraries(corusr PRIVATE corkdebug)

wrappers/advapi32.c

@@ -1,4 +1,3 @@
-#define _WIN32_WINNT 0x0400
 #include <windows.h>
 
 #include "debug.h"

wrappers/cas.asm

@@ -0,0 +1,17 @@
+.486
+.MODEL FLAT
+
+.CODE
+
+OPTION PROLOGUE:NONE
+OPTION EPILOGUE:NONE
+
+_InterlockedCompareExchange_486@12 PROC
+  mov ecx, DWORD PTR [esp + 4] ; dest
+  mov edx, DWORD PTR [esp + 8] ; exchange
+  mov eax, DWORD PTR [esp + 12] ; compare
+  lock cmpxchg DWORD PTR [ecx], edx
+  ret 12
+_InterlockedCompareExchange_486@12 ENDP
+
+END

wrappers/cas.h

@@ -0,0 +1,10 @@
+#ifndef CAS
+#define CAS
+
+#ifndef STDCALL
+#define STDCALL __stdcall
+#endif
+
+long STDCALL InterlockedCompareExchange_486(long* dest, long exchange, long compare);
+
+#endif // CAS

wrappers/comdlg32.c

@@ -1,4 +1,3 @@
-#define _WIN32_WINNT 0x0400
 #include <windows.h>
 
 #include "debug.h"

wrappers/detect486.asm

@@ -0,0 +1,28 @@
+.386
+.MODEL FLAT, STDCALL
+
+.CODE
+
+is_cpu_486_or_newer PROC
+  pushfd
+  pop eax
+  mov ebx, eax
+  xor eax, 40000h ; toggle the AC bit in EFLAGS (only available in 486 or newer)
+  push eax
+  popfd
+  pushfd
+  pop eax
+  cmp eax, ebx
+  jz is_386
+  push ebx
+  popfd
+  xor eax, eax
+  inc eax
+  ret
+
+is_386:
+  xor eax, eax
+  ret
+is_cpu_486_or_newer ENDP
+
+END

wrappers/detect486.h

@@ -0,0 +1,10 @@
+#ifndef DETECT_486
+#define DETECT_486
+
+#ifndef STDCALL
+#define STDCALL __stdcall
+#endif
+
+int STDCALL is_cpu_486_or_newer(void);
+
+#endif // DETECT_486

wrappers/kernel32.c

@@ -1,9 +1,11 @@
-#define _WIN32_WINNT 0x0400
-
 #include <stdio.h>
 #include <windows.h>
 
+#include "cas.h"
 #include "debug.h"
+#include "detect486.h"
+
+static int has_cmpxchg = 0;
 
 DWORD WINAPI CORKEL32_GetLastError()
 {
@@ -154,10 +156,14 @@ BOOL WINAPI CORKEL32_DeviceIoControl(HANDLE hDevice, DWORD dwIoControlCode, LPVO
 }
 
 // Reimplemented
-LONG WINAPI CORKEL32_InterlockedCompareExchange(LONG *dest,  LONG xchg,  LONG compare)
+LONG WINAPI CORKEL32_InterlockedCompareExchange(LONG* dest, LONG xchg, LONG compare)
 {
-  LONG temp = *dest;
+  LONG temp;
+  if (has_cmpxchg) {
+    return InterlockedCompareExchange_486(dest, xchg, compare);
+  }
 
+  temp = *dest;
   Trace(TRACE_FORCE_DONT_PRINT, "InterlockedCompareExchange");
 
   if (compare == *dest) {
@@ -303,3 +309,12 @@ BOOL WINAPI CORKEL32_InitializeCriticalSectionAndSpinCount(LPCRITICAL_SECTION cr
 
   return TRUE;
 }
+
+BOOL WINAPI DllMain(HINSTANCE hinstDLL, DWORD fdwReason, LPVOID lpvReserved)
+{
+  if (fdwReason == DLL_PROCESS_ATTACH) {
+    has_cmpxchg = is_cpu_486_or_newer();
+  }
+
+  return TRUE;
+}

wrappers/ntdll.c

@@ -1,16 +1,18 @@
-#define _WIN32_WINNT 0x0400
 #include <windows.h>
 
 #include "debug.h"
 
-ULONG WINAPI CORNT_RtlNtStatusToDosError(NTSTATUS)
+ULONG WINAPI CORNT_RtlNtStatusToDosError(NTSTATUS param_0)
 {
   Trace(TRACE_UNIMPLEMENTED, "RtlNtStatusToDosError");
   // TODO: Stub
   return 0;
 }
 
+#ifndef HAS_RTL_UNWIND
 extern NTAPI RtlUnwind(void* param_0, void* param_1, struct _EXCEPTION_RECORD* param_2, void* param_3);
+#endif
+
 VOID NTAPI CORNT_RtlUnwind(void *p0,void *p1,struct _EXCEPTION_RECORD *p2, void *p3)
 {
   Trace(TRACE_PASSTHROUGH, "RtlUnwind");

wrappers/user32.c

@@ -1,4 +1,3 @@
-#define _WIN32_WINNT 0x0400
 #include <windows.h>
 
 #include "debug.h"