[Bug tree-optimization/102058] New: 450.soplex regressed on x86_64 with -Ofast -march=generic (by 8-15%)

[Bug tree-optimization/102058] New: 450.soplex regressed on x86_64 with -Ofast -march=generic (by 8-15%)

[jamborm at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102058

            Bug ID: 102058
           Summary: 450.soplex regressed on x86_64 with -Ofast
                    -march=generic (by 8-15%)
           Product: gcc
           Version: 12.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: tree-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: jamborm at gcc dot gnu.org
                CC: rguenth at gcc dot gnu.org
            Blocks: 26163
  Target Milestone: ---

All three LNT x86_64 testers have experienced a regression when
running SPECFP 2006 benchmark 450.soplex compiled with -Ofast
-march=generic (as opposed to -march=native builds which seem not to
be affected).

A znver2 machine regressed by 15%:
https://lnt.opensuse.org/db_default/v4/SPEC/graph?plot.0=414.150.0&plot.1=300.150.0&

A znver1 machine regressed by 8%:
https://lnt.opensuse.org/db_default/v4/SPEC/graph?plot.0=451.150.0&plot.1=27.150.0&

An Intel Kabylake machine regressed by 9%:
https://lnt.opensuse.org/db_default/v4/SPEC/graph?plot.0=429.150.0&plot.1=25.150.0&

I have bisected the regression (on another znver2 machine) to revision
r12-2733-g31855ba6b16:

31855ba6b16cd138d7484076a08cd40d609654b8 is the first bad commit
commit 31855ba6b16cd138d7484076a08cd40d609654b8
Author: Richard Biener <rguenther@suse.de>
Date:   Thu Jul 29 14:14:48 2021 +0200

    Add emulated gather capability to the vectorizer

    This adds a gather vectorization capability to the vectorizer
    without target support by decomposing the offset vector, doing
    sclar loads and then building a vector from the result.  This
    is aimed mainly at cases where vectorizing the rest of the loop
    offsets the cost of vectorizing the gather.
    [...]


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)

[Bug tree-optimization/102058] 450.soplex regressed on x86_64 with -Ofast -march=generic (by 8-15%)

[marxin at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102058

Martin Liška <marxin at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
     Ever confirmed|0                           |1
   Last reconfirmed|                            |2021-08-25
             Status|UNCONFIRMED                 |NEW
                 CC|                            |marxin at gcc dot gnu.org

[Bug tree-optimization/102058] 450.soplex regressed on x86_64 with -Ofast -march=generic (by 8-15%)

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102058

--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
Created attachment 51355
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=51355&action=edit
patch to add --param vect-emulated-gather for debugging

When using attached patch then 0 vs. 1 on zen2 indeed reproduces

                                  Estimated                       Estimated
                Base     Base       Base        Peak     Peak       Peak
Benchmarks      Ref.   Run Time     Ratio       Ref.   Run Time     Ratio
-------------- ------  ---------  ---------    ------  ---------  ---------
450.soplex       8340        120       69.5 *    8340        136       61.4 *  

trying to nail it down now.

[Bug tree-optimization/102058] 450.soplex regressed on x86_64 with -Ofast -march=generic (by 8-15%)

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102058

--- Comment #2 from Richard Biener <rguenth at gcc dot gnu.org> ---
perf isn't particularly helpful, pointing at differences where no differences
in assembly occurs.  But we do now vectorize soplex::SPxSteepPR::entered4, in
particular soplex::Vector::operator* which is

   /// inner product.
   Real operator*(const Vector& w) const
   {
      Real x = 0;
      int n = size();
      Element* e = m_elem;

      while (n--)
      {
         x += e->val * w[e->idx];
         e++;
      }
      return x;
   }

and the e->val * w[e->idx] contains the gather we now handle.

Other parts perf points out are once again not vectorized :/

[Bug tree-optimization/102058] [12 regression] 450.soplex regressed on x86_64 with -Ofast -march=generic (by 8-15%)

[hubicka at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102058

Jan Hubicka <hubicka at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |hubicka at gcc dot gnu.org
            Summary|450.soplex regressed on     |[12 regression] 450.soplex
                   |x86_64 with -Ofast          |regressed on x86_64 with
                   |-march=generic (by 8-15%)   |-Ofast -march=generic (by
                   |                            |8-15%)

--- Comment #3 from Jan Hubicka <hubicka at gcc dot gnu.org> ---
This still shows comparing trunk to gcc11 on lnt, so marking as regression

[Bug tree-optimization/102058] [12 regression] 450.soplex regressed on x86_64 with -Ofast -march=generic (by 8-15%)

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102058

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Target Milestone|---                         |12.0

[Bug tree-optimization/102058] [12 regression] 450.soplex regressed on x86_64 with -Ofast -march=generic (by 8-15%)

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102058

--- Comment #4 from Richard Biener <rguenth at gcc dot gnu.org> ---
It's all with generic arch/tune but -Ofast which is not the most interesting
combination.  But we should see to extract a testcase for the reduction
and see to gather runtime data on the size() distribution.  When
vectorized the loop might also turn from nice small to slightly too big
for efficient cross iteration OOO scheduling.

[Bug tree-optimization/102058] [12 regression] 450.soplex regressed on x86_64 with -Ofast -march=generic (by 8-15%)

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102058

--- Comment #5 from Richard Biener <rguenth at gcc dot gnu.org> ---
It was observed that 450.soplex needs more iterations to converge.  As we now
vectorize a reduction that we didn't before that's definitely a thing that can
impact precision with FP.  But this is expected with -Ofast.

[Bug tree-optimization/102058] [12/13 regression] 450.soplex regressed on x86_64 with -Ofast -march=generic (by 8-15%)

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102058

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Target Milestone|12.0                        |12.2

--- Comment #6 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
GCC 12.1 is being released, retargeting bugs to GCC 12.2.

[Bug tree-optimization/102058] [12/13 regression] 450.soplex regressed on x86_64 with -Ofast -march=generic (by 8-15%)

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102058

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |SUSPENDED

--- Comment #7 from Richard Biener <rguenth at gcc dot gnu.org> ---
See comment#5

[Bug tree-optimization/102058] [12/13/14 regression] 450.soplex regressed on x86_64 with -Ofast -march=generic (by 8-15%)

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102058

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Target Milestone|12.3                        |12.4

--- Comment #9 from Richard Biener <rguenth at gcc dot gnu.org> ---
GCC 12.3 is being released, retargeting bugs to GCC 12.4.