[Bug target/122777] [16 Regression] 5-8% slowdown of 454.calculix (PGO) and 548.exchange2_r (non-pgo) on AMD Zen{4,5} since r16-5258-g1d8e2d51e5c5cb

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=122777

--- Comment #12 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
(In reply to Filip Kastl from comment #9)
> (In reply to Andrew Pinski from comment #8)
> > (In reply to Andrew Pinski from comment #6)
> > > calculix for zen5 is fixed by r16-5975-gfae0c626250563 (and now better than
> > > ever before).
> > 
> > Looks like it is broken again :(.
> 
> The new slowdown bisects to r16-6069-g9f27ac9b83af5e

Which might mean some RTL optimization is messing up the profile data. Or RA
depends on the bad profile data after all.


exchange2_r is most likely similar to PR 123070 which is causing extra spilling
now because the RTL level didn't resolve overlapping live ranges. Once I have
PR 123070 resolved I will check exchange2_r code generation.