On Wed, Nov 25, 2015 at 03:52:10PM +0100, Richard Biener wrote:
> On Wed, Nov 25, 2015 at 2:18 PM, Michael Matz <m...@suse.de> wrote:
> > Hi,
> >
> > On Wed, 25 Nov 2015, Richard Biener wrote:
> >
> >> I don't think so. Btw, if you want to add this please add a new gimple
> >> predicate to identify "memory barrier" (any call or asm with a VDEF).
> >
> > if (is_gimple_call (stmt) && !nonfreeing_call_p (stmt))
> > nt_call_phase++;
>
> That looks bogus to me. It misses asm()s and at least today
> nonfreeing_call_p too much checks what it sounds like it checks.
> In practice it might work though. At least all the __sync_* and
> __atomic_* calls are _not_ barriers this way. A pthread_mutex_lock is
> though as we don't have a builtin for it.
>
> I'd change the above to a conservative
>
> if ((is_gimple_call (stmt) || is_gimple_asm (stmt)) && gimple_vuse (stmt))
nonfreeing_call_p is one necessary condition (if that is true, it means
the call could mean that the first access does not trap while the second one
does).
But I agree that we need a predicate for nonbarrier_call_p or similar.
Some atomic builtins are not barriers though and IMHO should not be treated
as such, at least relaxed atomic loads/stores and pehraps even other relaxed
atomics. And it would be nice to have IPA discovery of nonbarrier_call_p
calls, like we have for nonfreeing_fn.
Jakub
