https://gcc.gnu.org/bugzilla/show_bug.cgi?id=120941
--- Comment #20 from H.J. Lu <hjl.tools at gmail dot com> --- (In reply to Filip Kastl from comment #19) > Well, if you want to reproduce the lbm slowdown, you need a Zen2 or Zen5 > machine. I'm not sure how I would produce a testcase that would also > uncover the slowdown on other microarchitectures, sorry. If I understand > correctly, you have a Zen3 machine and don't have a Zen2 or Zen5 machine, > right? > > Then there is also the bwaves slowdown on Zen3. It also bisects to your > commit so it might be caused by the same problem. To reproduce that, > compile 2017 bwaves_r with -Ofast -march=native with > r16-1644-gaba3b9d3a48a07 and with the previous commit and you should get > ~10% slowdown. Or use 2006 bwaves also with -Ofast -march=native and you > should get ~5% slowdown. I don't need to see the slowdown. I just need to reproduce the bad transformation from the source.
