ck_barrier: Use full barrier semantics for RMO.

src/ck_barrier_centralized.c

@@ -50,7 +50,7 @@ ck_barrier_centralized(struct ck_barrier_centralized *barrier,
 		return;
 	}
 
-	ck_pr_fence_load();
+	ck_pr_fence_atomic_load();
 	while (sense != ck_pr_load_uint(&barrier->sense))
 		ck_pr_stall();
 

src/ck_barrier_dissemination.c

@@ -119,5 +119,6 @@ ck_barrier_dissemination(struct ck_barrier_dissemination *barrier,
 		state->sense = ~state->sense;
 
 	state->parity = 1 - state->parity;
+	ck_pr_fence_memory();
 	return;
 }

src/ck_barrier_mcs.c

@@ -136,5 +136,6 @@ ck_barrier_mcs(struct ck_barrier_mcs *barrier,
 	ck_pr_store_uint(barrier[state->vpid].children[0], state->sense);
 	ck_pr_store_uint(barrier[state->vpid].children[1], state->sense);
 	state->sense = ~state->sense;
+	ck_pr_fence_memory();
 	return;
 }

src/ck_barrier_tournament.c

@@ -178,6 +178,7 @@ wakeup:
 	}
 
 leave:
+	ck_pr_fence_memory();
 	state->sense = ~state->sense;
 	return;
 }