Improve atomic ops implementation

When using gcc builtins for atomic operations, provide them
as macros, so gcc can see the builtins and do optimizations.
This change gives considerable speedups in bitlocks, which
use atomic operations heavily, see bug 650458.

Also, don't define G_ATOMIC_OP_MEMORY_BARRIER_NEEDED unconditionally
when using gcc builtins.

https://bugzilla.gnome.org/show_bug.cgi?id=617491

configure.ac

@@ -2468,9 +2468,6 @@ if test x"$GCC" = xyes; then
 	      [glib_cv_gcc_has_builtin_atomic_operations=no])
 
   AC_MSG_RESULT($glib_cv_gcc_has_builtin_atomic_operations)
-  if test $glib_cv_gcc_has_builtin_atomic_operations = yes; then
-    glib_memory_barrier_needed=yes
-  else
     case $host_cpu in
       i386)
         AC_MSG_RESULT([none])
@@ -2567,7 +2564,6 @@ if test x"$GCC" = xyes; then
         glib_memory_barrier_needed=yes
         ;;
     esac
-  fi
 else
   if test $glib_native_win32 = yes; then
     # For Windows but not using gcc. No barriers needed then either.
@@ -2590,7 +2586,7 @@ AC_COMPILE_IFELSE([AC_LANG_PROGRAM([
 int
 main (void)
 {
-  /* not like this actually runs or anything... */
+  /* it is not like this actually runs or anything... */
   syscall (SYS_futex, NULL, FUTEX_WAKE, FUTEX_WAIT);
   return 0;
 }
@@ -3257,7 +3253,10 @@ _______EOF
 	  echo >>$outfile
 	  echo "#define G_ATOMIC_OP_MEMORY_BARRIER_NEEDED 1" >>$outfile
 	fi
-
+	if test x"$g_gcc_atomic_ops" != xno; then
+          echo >>$outfile
+          echo "#define G_ATOMIC_OP_USE_GCC_BUILTINS 1" >>$outfile
+        fi
 	echo >>$outfile
 	g_bit_sizes="16 32 64"
 	for bits in $g_bit_sizes; do
@@ -3634,6 +3633,7 @@ g_system_thread_sizeof="$glib_cv_sizeof_system_thread"
 g_mutex_contents="$glib_cv_byte_contents_gmutex"
 
 g_memory_barrier_needed="$glib_memory_barrier_needed"
+g_gcc_atomic_ops="$glib_cv_gcc_has_builtin_atomic_operations"
 
 g_module_suffix="$glib_gmodule_suffix"
 

glib/gatomic-gcc.c

@@ -22,6 +22,22 @@
 
 #include "gatomic.h"
 
+/* All atomic operations are available as macros evaluating
+ * to gcc builtins (when using gcc builtins for atomic operations).
+ * For ABI stability, we provide functions for them too.
+ *
+ * To avoid interference, undefine the macros first.
+ */
+
+#undef g_atomic_int_exchange_and_add
+#undef g_atomic_int_compare_and_exchange
+#undef g_atomic_int_add
+#undef g_atomic_int_get
+#undef g_atomic_int_set
+#undef g_atomic_pointer_compare_and_exchange
+#undef g_atomic_pointer_get
+#undef g_atomic_pointer_set
+
 gint
 g_atomic_int_exchange_and_add (volatile gint G_GNUC_MAY_ALIAS *atomic,
                                gint                            val)
@@ -58,14 +74,14 @@ _g_atomic_thread_init (void)
 }
 
 gint
-(g_atomic_int_get) (volatile gint G_GNUC_MAY_ALIAS *atomic)
+g_atomic_int_get (volatile gint G_GNUC_MAY_ALIAS *atomic)
 {
   __sync_synchronize ();
   return *atomic;
 }
 
 void
-(g_atomic_int_set) (volatile gint G_GNUC_MAY_ALIAS *atomic,
+g_atomic_int_set (volatile gint G_GNUC_MAY_ALIAS *atomic,
                   gint                            newval)
 {
   *atomic = newval;
@@ -73,14 +89,14 @@ void
 }
 
 gpointer
-(g_atomic_pointer_get) (volatile gpointer G_GNUC_MAY_ALIAS *atomic)
+g_atomic_pointer_get (volatile gpointer G_GNUC_MAY_ALIAS *atomic)
 {
   __sync_synchronize ();
   return *atomic;
 }
 
 void
-(g_atomic_pointer_set) (volatile gpointer G_GNUC_MAY_ALIAS *atomic,
+g_atomic_pointer_set (volatile gpointer G_GNUC_MAY_ALIAS *atomic,
                       gpointer                            newval)
 {
   *atomic = newval;

glib/gatomic.h

@@ -56,12 +56,41 @@ gpointer g_atomic_pointer_get                  (volatile gpointer G_GNUC_MAY_ALI
 void     g_atomic_pointer_set                  (volatile gpointer G_GNUC_MAY_ALIAS *atomic,
 						gpointer           newval);
 
-#ifndef G_ATOMIC_OP_MEMORY_BARRIER_NEEDED
+#if defined(__GNUC__) && defined(G_ATOMIC_OP_USE_GCC_BUILTINS)
+
+#define g_atomic_int_exchange_and_add(atomic,val) \
+  __sync_fetch_and_add((atomic),(val))
+
+#define g_atomic_int_add(atomic,val) \
+  __sync_fetch_and_add((atomic),(val))
+
+#define g_atomic_int_compare_and_exchange(atomic,oldval,newval) \
+  __sync_bool_compare_and_swap((atomic),(oldval),(newval))
+
+#define g_atomic_int_get(atomic) \
+  __extension__ ({ __sync_synchronize(); *(atomic); })
+
+#define g_atomic_int_set(atomic,newval) \
+  __extension__ ({ *(atomic) = (newval); __sync_synchronize(); })
+
+#define g_atomic_pointer_compare_and_exchange(atomic,oldval,newval) \
+  __sync_bool_compare_and_swap((atomic),(oldval),(newval))
+
+#define g_atomic_pointer_get(atomic) \
+  __extension__ ({ __sync_synchronize(); *(atomic); })
+
+#define g_atomic_pointer_set(atomic,newval) \
+  __extension__ ({ *(atomic) = (newval); __sync_synchronize(); })
+
+#elif !defined(G_ATOMIC_OP_MEMORY_BARRIER_NEEDED)
+
 # define g_atomic_int_get(atomic) 		((gint)*(atomic))
 # define g_atomic_int_set(atomic, newval) 	((void) (*(atomic) = (newval)))
 # define g_atomic_pointer_get(atomic) 		((gpointer)*(atomic))
 # define g_atomic_pointer_set(atomic, newval)	((void) (*(atomic) = (newval)))
+
 #else
+
 # define g_atomic_int_get(atomic) \
  ((void) sizeof (gchar [sizeof (*(atomic)) == sizeof (gint) ? 1 : -1]), \
   (g_atomic_int_get) ((volatile gint G_GNUC_MAY_ALIAS *) (volatile void *) (atomic)))
@@ -74,6 +103,7 @@ void     g_atomic_pointer_set                  (volatile gpointer G_GNUC_MAY_ALI
 # define g_atomic_pointer_set(atomic, newval) \
  ((void) sizeof (gchar [sizeof (*(atomic)) == sizeof (gpointer) ? 1 : -1]), \
   (g_atomic_pointer_set) ((volatile gpointer G_GNUC_MAY_ALIAS *) (volatile void *) (atomic), (newval)))
+
 #endif /* G_ATOMIC_OP_MEMORY_BARRIER_NEEDED */
 
 /**