mingw32: use same strategy for thread-locals as MSVC

The gcc implementation of `_thread' seems to be much
slower, so we use the inline-assembly technique from
the MSVC build plus a hack in "main.c".

src/racket/include/schthread.h

@@ -27,7 +27,7 @@ extern "C" {
 
 #if defined(MZ_USE_PLACES) || defined(MZ_USE_FUTURES)
 # define USE_THREAD_LOCAL
-# if _MSC_VER
+# ifdef _WIN32
 #  ifdef _WIN64
 #   define THREAD_LOCAL __declspec(thread)
 #   define MZ_THREAD_EXTERN extern
@@ -36,10 +36,6 @@ extern "C" {
 #   define THREAD_LOCAL /* empty */
 #   define IMPLEMENT_THREAD_LOCAL_VIA_WIN_TLS
 #  endif
-# elif __MINGW32__
-#  define THREAD_LOCAL __thread
-#  define MZ_THREAD_EXTERN extern
-#  define IMPLEMENT_THREAD_LOCAL_EXTERNALLY_VIA_PROC
 # elif (defined(__APPLE__) && defined(__MACH__)) || defined(GC2_PLACES_TESTING)
 #  define IMPLEMENT_THREAD_LOCAL_VIA_PTHREADS
 #  if defined(__x86_64__) || defined(__i386__)
@@ -422,6 +418,19 @@ MZ_EXTERN uintptr_t scheme_tls_delta;
 MZ_EXTERN int scheme_tls_index;
 #  endif
 static __inline Thread_Local_Variables **scheme_get_thread_local_variables_ptr(void) {
+# ifdef __MINGW32__
+  Thread_Local_Variables **x;
+  asm (
+       "mov %%fs:(0x2C), %%eax;"
+       "mov (%%eax), %%eax;"
+       "add %1, %%eax;"
+       "mov %%eax, %0;"
+       :"=r"(x)        /* output */
+       :"r"(scheme_tls_delta)
+       :"%eax"  /* clobbered register */
+       );
+  return x;
+# else
   __asm { mov eax, FS:[0x2C]
 #  ifdef MZ_USE_WIN_TLS_VIA_DLL
           add eax, scheme_tls_index
@@ -429,6 +438,7 @@ static __inline Thread_Local_Variables **scheme_get_thread_local_variables_ptr(v
           mov eax, [eax]
           add eax, scheme_tls_delta }
   /* result is in eax */
+# endif
 }
 static __inline Thread_Local_Variables *scheme_get_thread_local_variables(void) {
   return *scheme_get_thread_local_variables_ptr();

src/racket/main.c

@@ -292,11 +292,11 @@ START_XFORM_SKIP;
 
 #ifdef IMPLEMENT_THREAD_LOCAL_VIA_WIN_TLS
 extern intptr_t _tls_index;
-static __declspec(thread) void *tls_space;
+static __thread void *tls_space;
 #endif
 
 #ifdef DOS_FILE_SYSTEM
-static void load_delayed()
+void load_delayed()
 {
   /* Order matters: load dependencies first */
 # ifndef MZ_PRECISE_GC
@@ -305,7 +305,19 @@ static void load_delayed()
   load_delayed_dll(NULL, "libracket" DLL_3M_SUFFIX "xxxxxxx.dll");
   record_dll_path();
 # ifdef IMPLEMENT_THREAD_LOCAL_VIA_WIN_TLS
+#  ifdef __MINGW32__
+  {
+    /* gcc declares space for the thread-local variable in a way that
+       the OS can set up, but its doesn't actually map variables
+       through the OS-supplied mechanism. Just assume that the first
+       thread-local variable is ours. */
+    void **base;
+    asm("mov %%fs:(0x2C), %0;" :"=r"(base));
+    scheme_register_tls_space(*base, _tls_index);
+  }
+#  else
   scheme_register_tls_space(&tls_space, _tls_index);
+#  endif
 # endif
 }
 #endif

src/racket/src/salloc.c

@@ -281,8 +281,13 @@ void scheme_setup_thread_local_key_if_needed() XFORM_SKIP_PROC
   {
     void **base;
 
+# ifdef __MINGW32__
+    asm("mov %%fs:(0x2C), %0;"
+	:"=r"(base));        /* output */
+# else
     __asm { mov ecx, FS:[0x2C]
             mov base, ecx }
+# endif
     scheme_tls_delta -= (uintptr_t)base[scheme_tls_index];
     scheme_tls_index *= sizeof(void*);
   }