ck_barrier: Serialize centralized and combining barriers.

Fixes regressions in PPC.

src/ck_barrier_centralized.c

@@ -45,10 +45,12 @@ ck_barrier_centralized(struct ck_barrier_centralized *barrier,
 	value = ck_pr_faa_uint(&barrier->value, 1);
 	if (value == n_threads - 1) {
 		ck_pr_store_uint(&barrier->value, 0);
+		ck_pr_fence_store();
 		ck_pr_store_uint(&barrier->sense, sense);
 		return;
 	}
 
+	ck_pr_fence_memory();
 	while (sense != ck_pr_load_uint(&barrier->sense))
 		ck_pr_stall();
 

src/ck_barrier_combining.c

@@ -182,8 +182,10 @@ ck_barrier_combining_aux(struct ck_barrier_combining *barrier,
 		 * are at the barrier.
 		 */
 		ck_pr_store_uint(&tnode->count, 0);
+		ck_pr_fence_store();
 		ck_pr_store_uint(&tnode->sense, ~tnode->sense);
 	} else {
+		ck_pr_fence_memory();
 		while (sense != ck_pr_load_uint(&tnode->sense))
 			ck_pr_stall();
 	}