On 9/10/26 4:21 AM, Tobias Burnus wrote:
Hi Jerry!
To make it easier for others reading this email:
This is abouthttps://gcc.gnu.org/PR126964:
[17 Regression] 4-5% slowdown of tonto of Zen{2,3,4,5} since
r17-3342-gf3943597388db4
and
Jerry D wrote:
In the subject PR Comment #10, there is a patch attached that is recommended
by Claude to give further performance improvement.
And while Jerry linked to the attachment of comment 10 - the latest
update is in Comment 17,
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=126964#c17
Both patches affect gcc/gimple-loop-versioning.cc
* * *
It is not within the bounds of commits I can do and I can only test with made-
up tests for the spec tonto case. (I do not have access to SPEC)
I am going to leave it alone. However, someone might be interested in looking
at it and maybe even using it.
It is also out of the realm of my expertise, which is Fortran and
OpenMP (+ OpenACC / offload) related.
[However, we have a recent SPEC CPU, if there is something that makes
sense to test. But others closer related to those topics like some
SUSE developers have as well. (We also have SPEC Accel and SPEC HPC,
if I should test something with those.)]
Cheers,
Tobias
Thanks Tobias,
In the original PR report there were links to charts vs time for the tonto
benchmark. I provided in the PR all the information I have about the claude
suggested loop versioning change. (I do not know what that even means)
It might be useful to see before and after results of the patch attached to the
PR. I updated the patch to include the ChangeLog commit message.
Best regards
Jerry
