[Bug middle-end/101910] New: tsvc regressions for -O2 -ftree-loop-vectorize at zen hardware

[Bug middle-end/101910] New: tsvc regressions for -O2 -ftree-loop-vectorize at zen hardware

[hubicka at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101910

            Bug ID: 101910
           Summary: tsvc regressions for -O2 -ftree-loop-vectorize at zen
                    hardware
           Product: gcc
           Version: 12.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: middle-end
          Assignee: unassigned at gcc dot gnu.org
          Reporter: hubicka at gcc dot gnu.org
  Target Milestone: ---

https://lnt.opensuse.org/db_default/v4/CPP/latest_runs_report?younger_in_days=14&older_in_days=0&all_changes=on&min_percentage_change=0.001&revisions=e54acea9e5a821448af97c66e94a1e4c4f977d5d%2Ce87209a1269622017bf3d98bf71502dcb0f893aa%2C73474527aaa24d9236aca074c5494a07f40ce058&include_user_branches=on

shows large regressions on tsvc for benchmarks s253, s271, s318, s3110, s281,
s292, s443

Maybe generic model needs retuning here.

[Bug tree-optimization/101910] tsvc regressions for -O2 -ftree-loop-vectorize at zen hardware

[crazylht at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101910

--- Comment #1 from Hongtao.liu <crazylht at gmail dot com> ---
We can only reproduce 5% regression for s3110, no regressions for others on
CLX.

And 5% regression seems to be related to code alignment which would impact
decoding efficiency, it could be considered as noise.

[Bug tree-optimization/101910] [12 Regression] tsvc regressions for -O2 -ftree-loop-vectorize at zen hardware

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101910

Andrew Pinski <pinskia at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
            Summary|tsvc regressions for -O2    |[12 Regression] tsvc
                   |-ftree-loop-vectorize at    |regressions for -O2
                   |zen hardware                |-ftree-loop-vectorize at
                   |                            |zen hardware
   Target Milestone|---                         |12.0

[Bug tree-optimization/101910] [12 Regression] tsvc regressions for -O2 -ftree-loop-vectorize at zen hardware

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101910

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |marxin at gcc dot gnu.org,
                   |                            |rguenth at gcc dot gnu.org

--- Comment #2 from Richard Biener <rguenth at gcc dot gnu.org> ---
https://lnt.opensuse.org/db_default/v4/CPP/latest_runs_report shows regressions
for s422, in the GCC testsuite copy we do not vectorize this with -O2 since
it requires versioning for aliasing.

/home/rguenther/src/gcc3/gcc/testsuite/gcc.dg/vect/tsvc/vect-tsvc-s422.c:21:34:
missed:   versioning for alias required: can't determine dependence between
flat_2d_array[_1] and *_7
consider run-time aliasing test between flat_2d_array[_1] and *_7.  That's
because the code does

    xx = flat_2d_array + 4;

    for (int nl = 0; nl < 8*iterations; nl++) {
        for (int i = 0; i < LEN_1D; i++) {
            xx[i] = flat_2d_array[i + 8] + a[i];
        }
        dummy(a, b, c, d, e, aa, bb, cc, 0.);
    }

but 'dummy' is not known to not clobber 'xx'.  While xx is declared __restrict
that dummy call still gets in the way.

So I wonder what variant LNT exactly benchmarks.

[Bug tree-optimization/101910] [12 Regression] tsvc regressions for -O2 -ftree-loop-vectorize at zen hardware

[marxin at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101910

--- Comment #3 from Martin Liška <marxin at gcc dot gnu.org> ---
> So I wonder what variant LNT exactly benchmarks.

It uses:
https://gitlab.suse.de/marxin/cpp-benchmarks

Note the benchmark tends to be extremely sensitive to a layout that
dramatically influences speed. Similarly, one can see the rapid difference
between ld.bfd and ld.gold.

[Bug tree-optimization/101910] [12 Regression] tsvc regressions for -O2 -ftree-loop-vectorize at zen hardware

[marxin at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101910

--- Comment #4 from Martin Liška <marxin at gcc dot gnu.org> ---
E.g this one is really caused by system update:
https://lnt.opensuse.org/db_default/v4/CPP/graph?plot.0=176.922.0

So I would close this unless Honza bisects that and provides a perf numbers.

[Bug tree-optimization/101910] [12 Regression] tsvc regressions for -O2 -ftree-loop-vectorize at zen hardware

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101910

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
         Resolution|---                         |INVALID
             Status|UNCONFIRMED                 |RESOLVED

--- Comment #5 from Richard Biener <rguenth at gcc dot gnu.org> ---
I see.