Repeat black list check on CAS fail in stack_push_explicit_aux_release

(a cherry-pick of commit ae0f8e7 from 'release-7_4')

Also, execute the first read in a loop with an acquire barrier, and
place black list checking as close to CAS as possible.

* src/atomic_ops_stack.c [AO_USE_ALMOST_LOCK_FREE]
(AO_stack_push_explicit_aux_release): Use acquire barrier to read list
value (stored to next local variable); read list value and store it to
x element before iterating over AO_stack_bl (and, thus, retry iterating
over AO_stack_bl if CAS failed).

src/atomic_ops_stack.c

@@ -56,6 +56,9 @@ AO_stack_push_explicit_aux_release(volatile AO_t *list, AO_t *x,
   /* No deletions of x can start here, since x is not currently in the  */
   /* list.                                                              */
  retry:
+  do {
+    next = AO_load_acquire(list);
+    *x = next;
 # if AO_BL_SIZE == 2
   {
     /* Start all loads as close to concurrently as possible. */
@@ -91,12 +94,7 @@ AO_stack_push_explicit_aux_release(volatile AO_t *list, AO_t *x,
   }
 # endif
   /* x_bits is not currently being deleted */
-  do
-    {
-      next = AO_load(list);
-      *x = next;
-    }
-  while (AO_EXPECT_FALSE(!AO_compare_and_swap_release(list, next, x_bits)));
+  } while (AO_EXPECT_FALSE(!AO_compare_and_swap_release(list, next, x_bits)));
 }
 
 /*