https://gcc.gnu.org/bugzilla/show_bug.cgi?id=120752
Bug ID: 120752
Summary: 5% slowdown of 525.x264_r since
r16-1346-gb0d50cbb42ab2c
Product: gcc
Version: 16.0
Status: UNCONFIRMED
Keywords: missed-optimization
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: pheeck at gcc dot gnu.org
CC: hubicka at gcc dot gnu.org
Blocks: 26163
Target Milestone: ---
Host: x86_64-pc-linux-gnu, aarch64-gnu-linux
Target: x86_64-pc-linux-gnu, aarch64-gnu-linux
As seen here
https://lnt.opensuse.org/db_default/v4/SPEC/graph?plot.0=477.377.0
https://lnt.opensuse.org/db_default/v4/SPEC/graph?plot.0=958.377.0
https://lnt.opensuse.org/db_default/v4/SPEC/graph?plot.0=286.377.0
https://lnt.opensuse.org/db_default/v4/SPEC/graph?plot.0=589.377.0
https://lnt.opensuse.org/db_default/v4/SPEC/graph?plot.0=792.377.0
there was a 5% exec time slowdown of the 525.x264_r SPEC 2017
benchmark when run with -Ofast -march=native -flto -fprofile-use. I've seen it
on x86_64 AMD, x86_64 Intel and Aarch64 machines.
I bisected it to r16-1346-gb0d50cbb42ab2c.
b0d50cbb42ab2ce5fab8a832cb82fc54b371c914 is the first bad commit
commit b0d50cbb42ab2ce5fab8a832cb82fc54b371c914
Author: Jan Hubicka <hubi...@ucw.cz>
Date: Fri Jun 6 17:57:00 2025 +0200
Fix profile updating in ipa-cp
Bootstrapping with autoprofiledbootstrap, LTO and checking enables ICEs in
WPA
because we end up mixing local and IPA count in
ipa-cp.cc:update_specialized_profile. This is because of missing call to
profile_count::adjust_for_ipa_scaling. While looking into that I however
noticed that the function forgets to update indirect call edges. This made
me
to commonize same logic which currently exists in clone_inlined_nodes,
update_specialized_profile, update_profiling_info and
update_counts_for_self_gen_clones.
While testing it I noticed that we also ICE when linking with
-fdump-ipa-all-details-blocks since IPA and local counts are temporarily
mixed
during IPA transformation stage, so I also added check to
profile_count::dump
to not crash and added verifier to gimple_verify_flow_info.
Other problem I also noticed is that while profile updates done by inliner
(via
cgraph_node::clone) are correctly using global0 profiles instead of erasing
profile completely when IPA counts drops to 0, the scaling in ipa-cp is not
doing that, so we lose info and possibly some code quality. I will fix
that
incrementally. Similarly ipa-split, when offlining region with 0 entry
count
may re-do frequency propagation to get something useful.
This is not a regression against GCC 15. See the comparison
here:
https://lnt.opensuse.org/db_default/v4/SPEC/graph?plot.0=1016.377.0&plot.1=1179.377.0&plot.2=958.377.0&;
Btw, if you look at the graphs, at some point we got better execution times
than GCC 15 and this slowdown is just a rebound (sometimes not even entirely).
Referenced Bugs:
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=26163
[Bug 26163] [meta-bug] missed optimization in SPEC (2k17, 2k and 2k6 and 95)
