https://gcc.gnu.org/bugzilla/show_bug.cgi?id=126055

--- Comment #3 from Filip Kastl <pheeck at gcc dot gnu.org> ---
The benchmark indeed sped up again.  I'm okay with closing this bug then (I'll
do that unless someone objects).

Some kind of code alignment issue seems probable to me.  I had trouble
reproducing the slowdown.  In particular, when I ran the benchmark under 'git
bisect run', I couldn't reproduce the slowdown.  But when I ran it manually, I
could.  I originally attributed it to some mistake on my part but maybe the git
bisect run influenced the code alignment somehow (I'm not sure how exactly that
would work, it's just a vague idea)?  And now, a few GCC commits later, GCC
generates slightly different code for the benchmark which also influences
alignment and also gets rid of the slowdown?

Reply via email to