https://gcc.gnu.org/bugzilla/show_bug.cgi?id=104364
--- Comment #6 from Thomas Schwinge <tschwinge at gcc dot gnu.org> ---
Thanks for having confirmed my findings and doubts -- seems I did correctly
understand a thing or two. ;-)
(In reply to Tom de Vries from comment #5)
> (In reply to Thomas Schwinge from comment #0)
> > ... but only seen regressing for: [...]
> >
> > ... and never seen regressing for: [...]
> >
> > (What is the underlying characteristic here?)
>
> Good question.
>
> I've tested this using (recommended) driver 470.94 on boards: [...]
> while iterating over dimensions { -mptx=3.1 , -mptx=6.3 } x {
> GOMP_NVPTX_JIT=-O0, <default> }.
>
> So I'm slightly surprised that I didn't see any regressions.
If indeed we're now generating some bad 'atom' code, it sure is confusing why
execution anyway PASSes for quite a number of configurations? Are we just
"lucky", or is there some more fundamental issue that we're not even properly
using the concurrency in these configurations (and thus don't notice the 'atom'
issues)?
