On 05/28/2018 04:15 PM, Gavin Hu wrote:
do {
+const uint32_t cons_tail = r->cons.tail;
+
/* Reset n to the initial burst count */
n = max;
*old_head = __atomic_load_n(&r->prod.head,
__ATOMIC_ACQUIRE);
-const uint32_t cons_tail = r->cons.tail;
+
[Gavin Hu] The ACQUIRE and RELEASE pair protects anything that between the two
must be visible to other threads when they perform an acquire operation on the
same memory address. Your changes broke this semantics. I advise to move the
declaration before and keep the assignment in the old place.
I see, thanks for the tip.
How about just get rid of this temp altogether if access to it is locked
during this sequence anyway? It's not like we had to sample it after
the lock then, or it's bringing anything else to the party.
So instead of cons_tail / prod_tail at all, replace directly with
r->cons.tail / r->prod.tail at the single usage for each.
-Andy
