kernel32: For Mac 64-bit, re-enable the inlining of various functions, leveraging the TEB self pointer at %gs:0x30.

On other platforms, gsbase is set to point to Wine's TEB.  So, these functions
can use %gs-relative addressing with the field offsets to access the fields of
the TEB.

On the Mac, gsbase points to internals of the pthread implementation and that
wouldn't work.  However, Wine hijacks %gs:0x30 and stores the TEB address there.
So, we access the TEB fields by first loading the TEB address and then
accessing its fields relative to that.

Signed-off-by: Ken Thomases <ken@codeweavers.com>
Signed-off-by: Alexandre Julliard <julliard@winehq.org>

dlls/kernel32/thread.c

@@ -682,7 +682,41 @@ __ASM_STDCALL_FUNC( GetCurrentThreadId, 0, ".byte 0x64\n\tmovl 0x24,%eax\n\tret"
 /* HANDLE WINAPI GetProcessHeap(void) */
 __ASM_STDCALL_FUNC( GetProcessHeap, 0, ".byte 0x64\n\tmovl 0x30,%eax\n\tmovl 0x18(%eax),%eax\n\tret");
 
-#elif defined(__x86_64__) && !defined(__APPLE__)
+#elif defined(__x86_64__)
+
+#ifdef __APPLE__
+
+/***********************************************************************
+ *		SetLastError (KERNEL32.@)
+ */
+/* void WINAPI SetLastError( DWORD error ); */
+__ASM_STDCALL_FUNC( SetLastError, 8, ".byte 0x65\n\tmovq 0x30,%rax\n\tmovl %ecx,0x68(%rax)\n\tret" );
+
+/***********************************************************************
+ *		GetLastError (KERNEL32.@)
+ */
+/* DWORD WINAPI GetLastError(void); */
+__ASM_STDCALL_FUNC( GetLastError, 0, ".byte 0x65\n\tmovq 0x30,%rax\n\tmovl 0x68(%rax),%eax\n\tret" );
+
+/***********************************************************************
+ *		GetCurrentProcessId (KERNEL32.@)
+ */
+/* DWORD WINAPI GetCurrentProcessId(void) */
+__ASM_STDCALL_FUNC( GetCurrentProcessId, 0, ".byte 0x65\n\tmovq 0x30,%rax\n\tmovl 0x40(%rax),%eax\n\tret" );
+
+/***********************************************************************
+ *		GetCurrentThreadId (KERNEL32.@)
+ */
+/* DWORD WINAPI GetCurrentThreadId(void) */
+__ASM_STDCALL_FUNC( GetCurrentThreadId, 0, ".byte 0x65\n\tmovq 0x30,%rax\n\tmovl 0x48(%rax),%eax\n\tret" );
+
+/***********************************************************************
+ *		GetProcessHeap (KERNEL32.@)
+ */
+/* HANDLE WINAPI GetProcessHeap(void) */
+__ASM_STDCALL_FUNC( GetProcessHeap, 0, ".byte 0x65\n\tmovq 0x30,%rax\n\tmovq 0x60(%rax),%rax\n\tmovq 0x30(%rax),%rax\n\tret");
+
+#else
 
 /***********************************************************************
  *		SetLastError (KERNEL32.@)
@@ -714,6 +748,8 @@ __ASM_STDCALL_FUNC( GetCurrentThreadId, 0, ".byte 0x65\n\tmovl 0x48,%eax\n\tret"
 /* HANDLE WINAPI GetProcessHeap(void) */
 __ASM_STDCALL_FUNC( GetProcessHeap, 0, ".byte 0x65\n\tmovq 0x60,%rax\n\tmovq 0x30(%rax),%rax\n\tret");
 
+#endif /* __APPLE__ */
+
 #else  /* __x86_64__ */
 
 /**********************************************************************

include/winbase.h

@@ -2904,13 +2904,19 @@ static FORCEINLINE LONG WINAPI InterlockedDecrement( LONG volatile *dest )
 
 /* A few optimizations for gcc */
 
-#if defined(__GNUC__) && !defined(__MINGW32__) && (defined(__i386__) || (defined(__x86_64__) && !defined(__APPLE__))) && ((__GNUC__ > 3) || ((__GNUC__ == 3) && (__GNUC_MINOR__ >= 2)))
+#if defined(__GNUC__) && !defined(__MINGW32__) && (defined(__i386__) || defined(__x86_64__)) && ((__GNUC__ > 3) || ((__GNUC__ == 3) && (__GNUC_MINOR__ >= 2)))
 
 static FORCEINLINE DWORD WINAPI GetLastError(void)
 {
     DWORD ret;
 #ifdef __x86_64__
+#ifdef __APPLE__
+    DWORD* teb;
+    __asm__ __volatile__( ".byte 0x65\n\tmovq 0x30,%0" : "=r" (teb) );
+    ret = teb[0x68 / sizeof(DWORD)];
+#else
     __asm__ __volatile__( ".byte 0x65\n\tmovl 0x68,%0" : "=r" (ret) );
+#endif
 #else
     __asm__ __volatile__( ".byte 0x64\n\tmovl 0x34,%0" : "=r" (ret) );
 #endif
@@ -2921,7 +2927,13 @@ static FORCEINLINE DWORD WINAPI GetCurrentProcessId(void)
 {
     DWORD ret;
 #ifdef __x86_64__
+#ifdef __APPLE__
+    DWORD* teb;
+    __asm__ __volatile__( ".byte 0x65\n\tmovq 0x30,%0" : "=r" (teb) );
+    ret = teb[0x40 / sizeof(DWORD)];
+#else
     __asm__ __volatile__( ".byte 0x65\n\tmovl 0x40,%0" : "=r" (ret) );
+#endif
 #else
     __asm__ __volatile__( ".byte 0x64\n\tmovl 0x20,%0" : "=r" (ret) );
 #endif
@@ -2932,7 +2944,13 @@ static FORCEINLINE DWORD WINAPI GetCurrentThreadId(void)
 {
     DWORD ret;
 #ifdef __x86_64__
+#ifdef __APPLE__
+    DWORD* teb;
+    __asm__ __volatile__( ".byte 0x65\n\tmovq 0x30,%0" : "=r" (teb) );
+    ret = teb[0x48 / sizeof(DWORD)];
+#else
     __asm__ __volatile__( ".byte 0x65\n\tmovl 0x48,%0" : "=r" (ret) );
+#endif
 #else
     __asm__ __volatile__( ".byte 0x64\n\tmovl 0x24,%0" : "=r" (ret) );
 #endif
@@ -2942,7 +2960,13 @@ static FORCEINLINE DWORD WINAPI GetCurrentThreadId(void)
 static FORCEINLINE void WINAPI SetLastError( DWORD err )
 {
 #ifdef __x86_64__
+#ifdef __APPLE__
+    DWORD* teb;
+    __asm__ __volatile__( ".byte 0x65\n\tmovq 0x30,%0" : "=r" (teb) );
+    teb[0x68 / sizeof(DWORD)] = err;
+#else
     __asm__ __volatile__( ".byte 0x65\n\tmovl %0,0x68" : : "r" (err) : "memory" );
+#endif
 #else
     __asm__ __volatile__( ".byte 0x64\n\tmovl %0,0x34" : : "r" (err) : "memory" );
 #endif
@@ -2952,7 +2976,13 @@ static FORCEINLINE HANDLE WINAPI GetProcessHeap(void)
 {
     HANDLE *pdb;
 #ifdef __x86_64__
+#ifdef __APPLE__
+    HANDLE** teb;
+    __asm__ __volatile__( ".byte 0x65\n\tmovq 0x30,%0" : "=r" (teb) );
+    pdb = teb[0x60 / sizeof(HANDLE*)];
+#else
     __asm__ __volatile__( ".byte 0x65\n\tmovq 0x60,%0" : "=r" (pdb) );
+#endif
     return pdb[0x30 / sizeof(HANDLE)];  /* get dword at offset 0x30 in pdb */
 #else
     __asm__ __volatile__( ".byte 0x64\n\tmovl 0x30,%0" : "=r" (pdb) );