https://gcc.gnu.org/bugzilla/show_bug.cgi?id=122017
--- Comment #1 from Martin Jambor <jamborm at gcc dot gnu.org> --- (In reply to Jan Hubicka from comment #0) > On Znver5 I measured 6% regression of parest and 4% regression of deepsjeng > with ipa-cp. What do you mean by "with IPA-CP?" Because of IPA-CP? And is this a regression using the same compiler but with and without PGO? (As opposed to a regression against a previous version?) With parest_r that indeed looks like it is the case but it seems to be because the non-PGO case improved while the PGO one did not, see: https://lnt.opensuse.org/db_default/v4/SPEC/graph?plot.2=1288.457.0&plot.4=1285.457.0&; https://lnt.opensuse.org/db_default/v4/SPEC/graph?plot.2=959.457.0&plot.3=958.457.0&; A recent deepsjeng regression happening around the time this was reported seems to have been fixed: https://lnt.opensuse.org/db_default/v4/SPEC/graph?plot.0=958.387.0 However, it seems that PGO is hurting long-term, see for example https://lnt.opensuse.org/db_default/v4/SPEC/graph?plot.0=1288.387.0&plot.1=1285.387.0&; https://lnt.opensuse.org/db_default/v4/SPEC/graph?plot.0=959.387.0&plot.1=958.387.0&; > Inreasing ipa-cp parameters using --param > ipa-cp-unit-growth=1000 --param ipa-cp-eval-threshold=10 further regresses > x264 by 11%. I believe this is mostly related to the heuristics choosing > one cloning over other. Did you really mean x264 or one of the other two benchmarks described before? I think this will need to get re-examined because of the new speculative edges, though I would not be surprised if there were still issues.
