Fixed complex gcc assembler issues with 64-bit operations on 32-bit platforms...

Fixed complex gcc assembler issues with 64-bit operations on 32-bit platforms using PIC codes, commented x86-gcc.h a lot more

include/atomic/x86-gcc.h

@@ -59,39 +59,68 @@
   asm volatile (LOCK_prefix "; cmpxchg %3, %0; setz %2;"	\
                : "+m" (*a), "+a" (*cmp), "=q" (ret): "r" (set))
 
-#define make_atomic_cas_bodyptr make_atomic_cas_body32
-
-#ifndef __x86_64__
+#ifdef __x86_64__
 #define make_atomic_add_body64 make_atomic_add_body32
 #define make_atomic_cas_body64 make_atomic_cas_body32
+
+#define make_atomic_fas_body(S)                                        \
+  asm volatile ("xchg %0, %1;" : "+r" (v) , "+m" (*a))
+
+/*
+  Actually 32-bit reads/writes are always atomic on x86
+  But we add LOCK_prefix here anyway to force memory barriers
+*/
+#define make_atomic_load_body(S)                                \
+  ret=0;                                                        \
+  asm volatile (LOCK_prefix "; cmpxchg %2, %0"                  \
+                : "+m" (*a), "+a" (ret): "r" (ret))
+#define make_atomic_store_body(S)                               \
+  asm volatile ("; xchg %0, %1;" : "+m" (*a), "+r" (v))
+
 #else
+/*
+  Use default implementations of 64-bit operations since we solved
+  the 64-bit problem on 32-bit platforms for CAS, no need to solve it
+  once more for ADD, LOAD, STORE and FAS as well.
+  Since we already added add32 support, we need to define add64
+  here, but we haven't defined fas, load and store at all, so
+  we can fallback on default implementations.
+*/
 #define make_atomic_add_body64                                  \
   int64 tmp=*a;                                                 \
   while (!my_atomic_cas64(a, &tmp, tmp+v));                     \
   v=tmp;
+
+/*
+  On some platforms (e.g. Mac OS X and Solaris) the ebx register
+  is held as a pointer to the global offset table. Thus we're not
+  allowed to use the b-register on those platforms when compiling
+  PIC code, to avoid this we push ebx and pop ebx and add a movl
+  instruction to avoid having ebx in the interface of the assembler
+  instruction.
+
+  cmpxchg8b works on both 32-bit platforms and 64-bit platforms but
+  the code here is only used on 32-bit platforms, on 64-bit
+  platforms the much simpler make_atomic_cas_body32 will work
+  fine.
+*/
 #define make_atomic_cas_body64                                  \
   int32 ebx=(set & 0xFFFFFFFF), ecx=(set >> 32);                \
-  asm volatile (LOCK_prefix "; cmpxchg8b %0; setz %2;"         \
+  asm volatile ("push %%ebx; movl %3, %%ebx;"                   \
+                LOCK_prefix "; cmpxchg8b %0; setz %2; pop %%ebx"\
                : "+m" (*a), "+A" (*cmp), "=q" (ret)             \
-               :"b" (ebx), "c" (ecx))
+               :"m" (ebx), "c" (ecx))
 #endif
 
-#define make_atomic_fas_body(S)                                        \
-  asm volatile ("xchg %0, %1;" : "+r" (v) , "+m" (*a))
+/*
+  The implementation of make_atomic_cas_body32 is adaptable to
+  the OS word size, so on 64-bit platforms it will automatically
+  adapt to 64-bits and so it will work also on 64-bit platforms
+*/
+#define make_atomic_cas_bodyptr make_atomic_cas_body32
 
 #ifdef MY_ATOMIC_MODE_DUMMY
 #define make_atomic_load_body(S)   ret=*a
 #define make_atomic_store_body(S)  *a=v
-#else
-/*
-  Actually 32-bit reads/writes are always atomic on x86
-  But we add LOCK_prefix here anyway to force memory barriers
-*/
-#define make_atomic_load_body(S)				\
-  ret=0;							\
-  asm volatile (LOCK_prefix "; cmpxchg %2, %0"			\
-               : "+m" (*a), "+a" (ret): "r" (ret))
-#define make_atomic_store_body(S)				\
-  asm volatile ("; xchg %0, %1;" : "+m" (*a), "+r" (v))
 #endif
 #endif /* ATOMIC_X86_GCC_INCLUDED */