https://gcc.gnu.org/bugzilla/show_bug.cgi?id=93439
--- Comment #6 from Jakub Jelinek <jakub at gcc dot gnu.org> --- (In reply to Richard Biener from comment #5) > OK, so the reason is that when we actually create the parallel decl > last_clique > is not yet up to its maximum since we later loop_version another loop in the > source function. The solution is to move the last clique setting to > function outlining I guess. > > Testing patch. I guess that doesn't hurt, but do we really want to parallelize inner loops of loops that we've parallelized already? I mean, that is nested parallelism and won't always be very beneficial (especially when in OpenMP 4.5 and earlier dynamic parallelism is disabled by default and needs to be explicitly enabled).
