ck_hp_fifo: Migrate to ck_hp_set_fence.

On Pentium M, this yields 30%-40% reduction in latency.

include/ck_hp_fifo.h

@@ -88,8 +88,7 @@ ck_hp_fifo_enqueue_mpmc(ck_hp_record_t *record,
 
 	for (;;) {
 		tail = ck_pr_load_ptr(&fifo->tail);
-		ck_hp_set(record, 0, tail);
-		ck_pr_fence_store_load();
+		ck_hp_set_fence(record, 0, tail);
 		if (tail != ck_pr_load_ptr(&fifo->tail))
 			continue;
 
@@ -119,8 +118,7 @@ ck_hp_fifo_tryenqueue_mpmc(ck_hp_record_t *record,
 	ck_pr_fence_store_atomic();
 
 	tail = ck_pr_load_ptr(&fifo->tail);
-	ck_hp_set(record, 0, tail);
-	ck_pr_fence_store_load();
+	ck_hp_set_fence(record, 0, tail);
 	if (tail != ck_pr_load_ptr(&fifo->tail))
 		return false;
 
@@ -147,14 +145,12 @@ ck_hp_fifo_dequeue_mpmc(ck_hp_record_t *record,
 		head = ck_pr_load_ptr(&fifo->head);
 		ck_pr_fence_load();
 		tail = ck_pr_load_ptr(&fifo->tail);
-		ck_hp_set(record, 0, head);
-		ck_pr_fence_store_load();
+		ck_hp_set_fence(record, 0, head);
 		if (head != ck_pr_load_ptr(&fifo->head))
 			continue;
 
 		next = ck_pr_load_ptr(&head->next);
-		ck_hp_set(record, 1, next);
-		ck_pr_fence_store_load();
+		ck_hp_set_fence(record, 1, next);
 		if (head != ck_pr_load_ptr(&fifo->head))
 			continue;
 
@@ -182,14 +178,12 @@ ck_hp_fifo_trydequeue_mpmc(ck_hp_record_t *record,
 	head = ck_pr_load_ptr(&fifo->head);
 	ck_pr_fence_load();
 	tail = ck_pr_load_ptr(&fifo->tail);
-	ck_hp_set(record, 0, head);
-	ck_pr_fence_store_load();
+	ck_hp_set_fence(record, 0, head);
 	if (head != ck_pr_load_ptr(&fifo->head))
 		return NULL;
 
 	next = ck_pr_load_ptr(&head->next);
-	ck_hp_set(record, 1, next);
-	ck_pr_fence_store_load();
+	ck_hp_set_fence(record, 1, next);
 	if (head != ck_pr_load_ptr(&fifo->head))
 		return NULL;