https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102178
Bug ID: 102178
Summary: SPECFP 2006 470.lbm regressions on AMD Zen CPUs after
r12-897-gde56f95afaaa22
Product: gcc
Version: 11.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: jamborm at gcc dot gnu.org
Blocks: 26163
Target Milestone: ---
Host: x86_64-linix
Target: x86_64-linux
LNT has detected an 18% regression of SPECFP 2006 benchmark 470.lbm
when it is compiled with -Ofast -march=native on a Zen2 machine:
https://lnt.opensuse.org/db_default/v4/SPEC/graph?plot.0=421.240.0&plot.1=301.240.0&;
...and similarly a 6% regression when it is run on the same machine
with -Ofast:
https://lnt.opensuse.org/db_default/v4/SPEC/graph?plot.0=450.240.0&plot.1=24.240.0&;
I have bisected both on another zen2 machine to commit
r12-897-gde56f95afaaa22 (Run pass_sink_code once more before
store_merging).
Zen1 machine has also seen a similar -march=native regression in the
same time frame:
https://lnt.opensuse.org/db_default/v4/SPEC/graph?plot.0=450.240.0&plot.1=24.240.0&;
Zen1 -march=generic seems to be unaffected, which is also the case for
the Intel machines we track.
Although lbm has been known to have weird regressions caused entirely
by code layout where the compiler was not really at fault, the fact
that both generic code-gen and Zen1 are affected seems to indicate this
is not the case.
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)
