https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116896
--- Comment #26 from Uroš Bizjak <ubizjak at gmail dot com> --- (In reply to Jakub Jelinek from comment #24) > (In reply to Uroš Bizjak from comment #22) > > (In reply to Uroš Bizjak from comment #21) > > > (In reply to Jakub Jelinek from comment #19) > > > > Created attachment 59273 [details] > > > > gcc15-pr116896-inc2.patch > > > > > > > > I've tried to improve the signed int <=> case (the only one which > > > > doesn't > > > > use sbb and so needs two xors first), but unfortunately that doesn't > > > > seem to > > > > work at all, the IL is how I'd like to see it at the end, but then > > > > combine > > > > happily turns those *setcc_qi_slp back into *setcc_si_1_movzbl and > > > > removes > > > > the clear instructions. > > > > So, I think all we can do is some extra peephole2 to deal with it > > > > (though in > > > > the jp case it will have to use to movl $0, %reg). > > > > Huh, another idea is to add a variant of *setcc_qi_slp, perhaps implemented > > with unspec or (better?) add some unspec tag to it. If we already emit > > optimal code at the expand time, then there is not a lot of work left for > > optimization passes anyway. > > That might be best for the float non-fast-math case, and split it after > combine into the strict_low_part. You beat me by the second ;) Yes, for the cases where non-QI mode result is needed. Also, I think this pattern would be universally useful, ix86_expand_clear that is used in combination with SLP SETcc works only after reload.
