https://gcc.gnu.org/bugzilla/show_bug.cgi?id=121908
vekumar at gcc dot gnu.org changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |vekumar at gcc dot gnu.org
--- Comment #10 from vekumar at gcc dot gnu.org ---
(In reply to Robin Dapp from comment #9)
> > I am seeing a 4-5% exec time slowdown of SPEC 2017 xz_r on Zen4 and Zen5,
> > both with -Ofast -march=native -flto=128:
> >
> > https://lnt.opensuse.org/db_default/v4/SPEC/graph?plot.0=959.417.0
> > https://lnt.opensuse.org/db_default/v4/SPEC/graph?plot.0=1288.417.0
>
> This is somewhat expected, unfortunately. A real speedup should be possible
> once Tamar's early-break changes land. Right now we just do double work
> because of vector overhead and a scalar epilogue.
My understanding is that the loop was not vectorized before the patch. Now we
vectorize but with peeling check for mutual alignment. It mostly of the time
fails at runtime. We are doing extra work by placing these guarding checks for
alignment. Additionally, the code alignment of the scalar loop now gets
shifted.
Speed up is possible if we make the page boundary checks pass.
