Bug 116076 - 4.5% slowdown of 433.milc on AMD Zen4 since r15-2054-g1e3aa9c9278db6 shows faster operation with non-sensical target tuning
Summary: 4.5% slowdown of 433.milc on AMD Zen4 since r15-2054-g1e3aa9c9278db6 shows fa...
Status: UNCONFIRMED
Alias: None
Product: gcc
Classification: Unclassified
Component: target (show other bugs)
Version: 15.0
: P3 normal
Target Milestone: ---
Assignee: Not yet assigned to anyone
URL:
Keywords: missed-optimization
Depends on:
Blocks: spec
  Show dependency treegraph
 
Reported: 2024-07-24 17:17 UTC by Filip Kastl
Modified: 2024-12-02 09:03 UTC (History)
1 user (show)

See Also:
Host: x86_64-pc-linux-gnu
Target: x86_64-pc-linux-gnu
Build:
Known to work:
Known to fail:
Last reconfirmed:


Attachments

Note You need to log in before you can comment on or make changes to this bug.
Description Filip Kastl 2024-07-24 17:17:13 UTC
As seen here

https://lnt.opensuse.org/db_default/v4/SPEC/graph?plot.0=992.70.0

there was a 4.5% exec time slowdown of the 433.milc SPEC 2006 benchmark between commits

r15-2045-g61c21a719e205f
r15-2061-gacd4ab8d6e8938

when run with -Ofast -march=native -flto on an AMD Zen4 machine.
Comment 2 Filip Kastl 2024-09-23 15:09:04 UTC
(In reply to Andrew Pinski from comment #1)
> Might be:
> https://gcc.gnu.org/git/?p=gcc.git;a=commit;
> h=1e3aa9c9278db69d4bdb661a750a7268789188d6

I've just checked and it really is r15-2054-g1e3aa9c9278db6.  Thanks for the tip, Andrew.  Cc-ing richi.

Though I must add that on the machine where I just ran the benchmarks (Zen 4) the slowdown is only ~3%.
Comment 3 Richard Biener 2024-09-24 06:51:36 UTC
While it might be interesting to investigate why making unaligned loads and stores costed faster than aligned ones the fix will be not restoring
behavior to before the change.

Given it's with LTO and only 3% tracking it down will be difficult.

Note the change was backported to all affected branches.
Comment 4 Filip Kastl 2024-09-24 07:37:25 UTC
Should I leave this bug open or close it, then?  What do you think, Richard?
Comment 5 Richard Biener 2024-09-24 07:46:02 UTC
Let's keep it open as non-regression with amended Summary.