https://gcc.gnu.org/bugzilla/show_bug.cgi?id=127432
Drea Pinski <pinskia at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
See Also| |https://gcc.gnu.org/bugzill
| |a/show_bug.cgi?id=126052,
| |https://gcc.gnu.org/bugzill
| |a/show_bug.cgi?id=125542
--- Comment #9 from Drea Pinski <pinskia at gcc dot gnu.org> ---
(In reply to Josef Melcr from comment #8)
> The graph came down. I reproduced the speedup on the machine which
> originated this report, bisects to r17-4456-gb1dcde565577e8:
>
> commit b1dcde565577e8ae1b6297e4a230f040e6aa6547
> Author: Andrea Pinski <[email protected]>
> Date: Wed Sep 16 00:01:13 2026 -0700
>
> phiopt: Better pre-vect heurstics for factoring out loads [PR125945]
>
> Not quite sure why that is though, I don't see how these two are linked.
Hmm, The only thing I can think of what this might be doing (and how it
relates to the previous patch) is there are some bouncing up and down due to
profiling data/count differences. My patch would have caused earlier
commonalization of loads which in theory could cause differences in the count
on the branches happening. The other thing that my patch could allow for
commonalization of the load (that is the same on both sides) before
vectorization which might allowed for more vectorization or other
optimizations; but that would be more likely seen in the non-pgo case too which
we don't. This is why I am thinking the count on the bb changes which is
indirectly causing improvements.
