membars for lock-free ref-count decrements...

[Chris Thomasson]

This is the way I am currently doing reference count decrements:

typedef int32_t rc_t;

/* returns zero only if count drops to zero */
int rc_dec(rc_t *_this) {
rc_t l;
int rel_mb = 0;

do {
l = (*_this);

/* release barrier is count > 1 */
if (l > 1 && ! rel_mb) {
membar #LoadStore|#StoreStore;

/* exec the barrier only one time! */
rel_mb = 1;
}

} while (! CAS(_this, l, l - 1));

/* acquire barrier if count dropped to zero */
if (l == 1) {
membar #StoreLoad|#StoreStore;
}

return (l == 1) ? 0 : 1;
}

IIRC, this is the correct way to do things wrt ref-count decrements?

[Peter Dimov]

Chris Thomasson wrote:
> This is the way I am currently doing reference count decrements:
>
>
> typedef int32_t rc_t;
>
>
> /* returns zero only if count drops to zero */
> int rc_dec(rc_t *_this) {
> rc_t l;
> int rel_mb = 0;
>
> do {
> l = (*_this);
>
> /* release barrier is count > 1 */
> if (l > 1 && ! rel_mb) {
> membar #LoadStore|#StoreStore;
>
> /* exec the barrier only one time! */
> rel_mb = 1;
> }
>
> } while (! CAS(_this, l, l - 1));
>
> /* acquire barrier if count dropped to zero */
> if (l == 1) {
> membar #StoreLoad|#StoreStore;
> }

Since CAS is both a load and a store for membar purposes, you can use
#LoadLoad instead of the more expensive #StoreLoad.

Most hardware doesn't allow stores to cross an earlier conditional
branch, so you can probably omit the #StoreStore part if there's a
performance advantage.