https://gcc.gnu.org/bugzilla/show_bug.cgi?id=104364
--- Comment #1 from Tom de Vries <vries at gcc dot gnu.org> ---
(In reply to Thomas Schwinge from comment #0)
> 'nvptx-none/mgomp/libatomic/cas_1_.o' (complete diff):
>
> @@ -113,7 +113,7 @@
> .loc 3 80 9
> or.b64 %r61,%r60,%r39;
> .loc 3 82 11
> -@ %r71 atom.cas.b64 %r62,[%r35],%r29,%r61;
> +atom.cas.b64 %r62,[%r35],%r29,%r61;
> mov.b64 {%r69,%r70},%r62;
> shfl.idx.b32 %r69,%r69,%r68,31;
> shfl.idx.b32 %r70,%r70,%r68,31;
>
Interesting. This could be fallout from nvptx_reorg_uniform_simt.
Basically, the function assumes that atom.cas is predicable, which it's not
(anymore).
What is odd is that the resulting insn is still validated, I would have
expected that to fail.