https://gcc.gnu.org/bugzilla/show_bug.cgi?id=123606
--- Comment #4 from Andrew Pinski <pinskia at gcc dot gnu.org> --- (In reply to Richard Biener from comment #3) > It's likely that having a latch vs. not might trigger some of the many weird > conditions we have on threading around loops. I was thinking that too which is why I had suspected r13-5103-g7c9f20fcfdc2d8 was the original worse performance back in 2023. Plus there was a better performance happening around r13-2087-gbac07a1da24fbc too. The increase in performance in June 2024 looks to be the implementation of the ccmp hooks to the x86 backend; the decrease in peformance after that looks to be r15-1293-g83a765768510d1 which turns off ccmp code path unless APX is turned on. Which definite means conditional jumps changes branch behavior. Note the decrease/fix in August 2025 to January looks to be due to the upgrade of the system both times (as the other branches have the same effect). This is all based on https://lnt.opensuse.org/db_default/v4/SPEC/graph?plot.0=802.60.0&plot.1=838.60.0&plot.2=821.60.0&plot.3=903.60.0&plot.4=1084.60.0&plot.5=1175.60.0&plot.6=799.60.0 . Note aarch64 (https://lnt.opensuse.org/db_default/v4/SPEC/graph?highlight_run=80219&plot.60=574.60.0): There was a slight increase in October 2024, which seems like it was r15-4759-g3d8cd34a450e9f (even with the noise due to what looks of branch predition or maybe THPs). I am not sure it is jump thread exactly but maybe rather branch prediction related.
