[Bug target/102473] [12 Regression] 521.wrf_r 5% slower at -Ofast and generic x86_64 tuning after r12-3426-g8f323c712ea76c

[jamborm at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102473

--- Comment #10 from Martin Jambor <jamborm at gcc dot gnu.org> ---
(In reply to Hongtao.liu from comment #6)
> Does it means cycles? 

Basically yes, AFAIK.  Basically I ran both versions under perf record
and then processed the output (so that is not so wide) of perf report
-n --stdio --percent-limit=2 (where -n is the thing that gives you
"samples").

> Vtune data show __module_mp_wsm5_MOD_nislfv_rain_plm has less instructions
> retired and clocksticks after my commit. And the regression comes from
> libc-2.31.so which shoud be the same.

I tend to think that any glibc from 2.29 is good enough to reproduce this.
For what it's worth, the system I tried this on has glib 2.33

My examination was very preliminary, because wrf takes ages to build,
I hoped I would point people to the important bit.  I am not sure I
succeeded though.

(In reply to Hongtao.liu from comment #8)
> 
> I'm going to revert the patch.

This is your call.  I actually dot not think that compiling wrf_r for
pre-AVX2 targets is a very important use case, the regression was just
so consistent that I thought it was worth investigating (and of course
it would be great if it could be avoided).

So it depends whether the patch has speed benefits in more common
circumstances or not.