https://gcc.gnu.org/bugzilla/show_bug.cgi?id=125855

--- Comment #9 from Filip Kastl <pheeck at gcc dot gnu.org> ---
So I verified some of the slowdowns and speedups.  Here is what I found:


What r17-1577 certainly slowed down:

https://lnt.opensuse.org/db_default/v4/SPEC/graph?plot.0=956.70.0
5% milc SPEC 2006 with -O2 -g -flto=128 on an Zen4

slowdown lbm 6% Zen2 -Ofast -march=native -g with PGO
https://lnt.opensuse.org/db_default/v4/SPEC/graph?plot.0=299.240.0

6% slowdown on Zen3 with the same flags: milc O2 lto
https://lnt.opensuse.org/db_default/v4/SPEC/graph?plot.0=469.70.0


What r17-1577 certainly sped up:

8% lntzen5.spec2006.gcc-trunk.O2_native:1242 lbm
https://lnt.opensuse.org/db_default/v4/SPEC/graph?plot.0=1242.240.0

11%     benzen.spec2006.gcc-trunk.O2_generic_lto:292 cactusADM
https://lnt.opensuse.org/db_default/v4/SPEC/graph?plot.0=292.100.0

5% speedup: benzen.spec2006.gcc-trunk.O2_native_lto:290 lbm
https://lnt.opensuse.org/db_default/v4/SPEC/graph?plot.0=290.240.0


These speedups certainly weren't caused by r17-1577:

5% speedup      lntzen4.spec2006.gcc-trunk.O2_x86_64v3_lto:1101 zeusmp
https://lnt.opensuse.org/db_default/v4/SPEC/graph?plot.0=1101.80.0

7% speedup lntzen5.spec2006.gcc-trunk.Ofast_native_lto:1284 cactusADM
https://lnt.opensuse.org/db_default/v4/SPEC/graph?plot.0=1284.100.0


What remains unsure:

- This slowdown (but it likely is caused by r17-1577 IMO):
https://lnt.opensuse.org/db_default/v4/SPEC/graph?plot.0=1100.70.0
9-12% Zen4 -O2 -g -flto=128 -U_FORTIFY_SOURCE -D_FORTIFY_SOURCE=3
-fstack-protector-strong -funwind-tables -fasynchronous-unwind-tables
-fstack-clash-protection

- The small intel speedups


So all in all I'd say there are both things that sped up and things that slowed
down.  Without running a lot of benchmarks on a lot of machines with a lot of
cmdline flags combinations (what Josef talked about), I don't think we can
really determine for sure if the patch is net gain or net loss for SPEC CPU. 
If it were up to me, I'd keep the patch.

Reply via email to