https://gcc.gnu.org/bugzilla/show_bug.cgi?id=123207
Victor Do Nascimento <victorldn at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Status|UNCONFIRMED |ASSIGNED
Last reconfirmed| |2026-01-05
Assignee|unassigned at gcc dot gnu.org |victorldn at gcc dot
gnu.org
Ever confirmed|0 |1
--- Comment #16 from Victor Do Nascimento <victorldn at gcc dot gnu.org> ---
I've managed to replicate the 526.blender_r error on an AMD Zen5 machine at my
end.
Interestingly, the error doesn't show up in our default x86_64 benchmark
config, given the other flags we use beyond -Ofast and -march.
The key flag that when present suppresses the miscompare over here is the
-mprefer-vector-width=128.
Therefore, while we see the issue with -Ofast -march=x86-64-v3, it disappears
with -mprefer-vector-width=128 -Ofast -march=x86-64-v3.
I've not yet done any further analysis beyond this, but will dig deeper
tomorrow.
