https://gcc.gnu.org/bugzilla/show_bug.cgi?id=119495
Bug ID: 119495
Summary: 8% slowdown of 436.cactusADM on AMD Zen2 since
r15-7895-gb191e8bdecf881
Product: gcc
Version: 15.0
Status: UNCONFIRMED
Keywords: missed-optimization, ra
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: pheeck at gcc dot gnu.org
CC: rsandifo at gcc dot gnu.org
Blocks: 26163
Target Milestone: ---
Host: x86_64-linux
Target: x86_64-linux
As seen here (the first spike of March):
https://lnt.opensuse.org/db_default/v4/SPEC/graph?plot.0=290.100.0
there was an 8% slowdown of 436.cactusADM SPEC 2006 benchmark on AMD Zen2 when
run with options -O2 -march=native -flto
I've bisected this to
r15-7895-gb191e8bdecf881
Author: Richard Sandiford <richard.sandif...@arm.com>
Date: Thu Mar 6 11:06:25 2025 +0000
ira: Add new hooks for callee-save vs spills [PR117477]
Following on from the discussion in:
https://gcc.gnu.org/pipermail/gcc-patches/2025-February/675256.html
this patch removes TARGET_IRA_CALLEE_SAVED_REGISTER_COST_SCALE and
replaces it with two hooks: one that controls the cost of using an
extra callee-saved register and one that controls the cost of allocating
a frame for the first spill.
(The patch does not attempt to address the shrink-wrapping part of
the thread above.)
On AArch64, this is enough to fix PR117477, as verified by the new tests.
The patch does not change the SPEC2017 scores significantly. (I saw a
slight improvement in fotonik3d and roms, but I'm not convinced that
the improvements are real.)
The patch makes IRA use caller saves for gcc.target/aarch64/pr103350-1.c,
which is a scan-dump correctness test that relies on not using
caller saves. The decision to use caller saves looks appropriate,
and saves an instruction, so I've just added -fno-caller-saves
to the test options.
The x86 parts were written by Honza. ix86_callee_save_cost is updated
by H.J. to replace gcc_checking_assert with returning 1 if mem_cost <= 2.
However:
1. This isn't a regression against GCC 14. Comparison here:
https://lnt.opensuse.org/db_default/v4/SPEC/graph?plot.8=1033.100.0&plot.9=290.100.0&;
2. cactusADM performance is prone to change a lot with minor changes to how
registers are allocated (I source this information from Richard Biener's
comment in pr119044 :)).
So my understanding is that this slowdown isn't really that important.
However, it seemed reasonable to at least notify Richard Sandiford about this
in case he wants to investigate it. Otherwise, I would be fine with closing
this as WONTFIX or something like that.
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)
