Repeat black list check on CAS fail in stack_push_explicit_aux_release

(a cherry-pick of commit aecb72063 from 'master')

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).

1 files changed, 4 insertions, 6 deletions

diff --git a/src/atomic_ops_stack.c b/src/atomic_ops_stack.c
index 5d7995f..c31c7bf 100644
--- a/src/atomic_ops_stack.c
+++ b/src/atomic_ops_stack.c
@@ -67,6 +67,9 @@ void 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);
+    store_before_cas(x, next);
 # if AO_BL_SIZE == 2
   {
     /* Start all loads as close to concurrently as possible. */
@@ -102,12 +105,7 @@ void 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);
-      store_before_cas(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)));
 }
 
 /*