https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96898
--- Comment #4 from Tom de Vries <vries at gcc dot gnu.org> ---
(In reply to Jakub Jelinek from comment #3)
> For OpenMP reductions, we really don't care what kind of mutex protects the
> updates, as long as it is the same for all updates of the same reduction.
> I believe we don't rely on any other synchronization effects.
> So, I think we should change omp-low.c so that it emits __atomic_* calls
> with __ATOMIC_RELAXED rather than __sync_* calls.
That sounds like a good idea.
> And could just use
> libatomic with its own locking if we didn't go the GOMP_atomic_{start,end}
> route (that one is done if there are multiple reductions or the atomics
> aren't available or there are user defined reductions we don't understand
> (or all?), perhaps we should consider also using atomics perhaps even for
> two simple reductions or similar.
> And nvptx certainly could just use libatomic...
If we use libatomic as fallback for openmp, shouldn't we then use the same lock
in both?
