[Bug tree-optimization/97494] New: [11 regression] several vector test case failures starting with r11-3966

[Bug tree-optimization/97494] New: [11 regression] several vector test case failures starting with r11-3966

[seurer at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97494

            Bug ID: 97494
           Summary: [11 regression] several vector test case failures
                    starting with r11-3966
           Product: gcc
           Version: 11.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: tree-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: seurer at gcc dot gnu.org
  Target Milestone: ---

g:429ad0bb0d3dc77e44f95620341da4938d49168e, r11-3966: 109 failures

On all powerpc64 both BE and LE:

FAIL: gcc.dg/vect/slp-11b.c -flto -ffat-lto-objects  scan-tree-dump-times vect
"vectorizing stmts using SLP" 0
FAIL: gcc.dg/vect/slp-11b.c scan-tree-dump-times vect "vectorizing stmts using
SLP" 0
FAIL: gcc.dg/vect/slp-21.c -flto -ffat-lto-objects  scan-tree-dump-times vect
"vectorizing stmts using SLP" 2
FAIL: gcc.dg/vect/slp-21.c scan-tree-dump-times vect "vectorizing stmts using
SLP" 2

Some extras on power7 BE:

FAIL: gcc.dg/vect/pr97428.c -flto -ffat-lto-objects  scan-tree-dump-times vect
"vectorizing stmts using SLP" 2
FAIL: gcc.dg/vect/pr97428.c scan-tree-dump-times vect "vectorizing stmts using
SLP" 2
FAIL: gcc.dg/vect/vect-complex-5.c -flto -ffat-lto-objects 
scan-tree-dump-times vect "vectorizing stmts using SLP" 2
FAIL: gcc.dg/vect/vect-complex-5.c scan-tree-dump-times vect "vectorizing stmts
using SLP" 2


commit 429ad0bb0d3dc77e44f95620341da4938d49168e (HEAD)
Author: Richard Biener <rguenther@suse.de>
Date:   Thu Oct 15 11:55:53 2020 +0200

    tree-optimization/97428 - split SLP groups for loop vectorization

    This enables SLP store group splitting also for loop vectorization.
    For the existing testcase gcc.dg/vect/vect-complex-5.c this then
    generates much better code, likewise for the PR97428 testcase.

    Both of those have a splitting opportunity splitting the group
    into two equal (vector-sized) halves, still the patch enables
    quite arbitrary splitting since generally the interleaving scheme
    results in quite awkward code for even small groups.  If any
    problems surface with this it's easy to restrict the splitting
    to known-good cases.

    2020-10-15  Richard Biener  <rguenther@suse.de>

            PR tree-optimization/97428
            * tree-vect-slp.c (vect_analyze_slp_instance): Split store
            groups also for loop vectorization.

            * gcc.dg/vect/vect-complex-5.c: Expect to SLP.
            * gcc.dg/vect/pr97428.c: Likewise.

[Bug tree-optimization/97494] [11 regression] several vector test case failures starting with r11-3966

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97494

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Assignee|unassigned at gcc dot gnu.org      |rguenth at gcc dot gnu.org
             Status|UNCONFIRMED                 |ASSIGNED
   Target Milestone|---                         |11.0
   Last reconfirmed|                            |2020-10-19
     Ever confirmed|0                           |1

--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
Hmm, I'll have a look how/if to adjust testcases.  Help appreciated.

[Bug tree-optimization/97494] [11 regression] several vector test case failures starting with r11-3966

[clyon at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97494

Christophe Lyon <clyon at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |clyon at gcc dot gnu.org

--- Comment #2 from Christophe Lyon <clyon at gcc dot gnu.org> ---
Similar problems seen on arm too.

[Bug tree-optimization/97494] [11 regression] several vector test case failures starting with r11-3966

[ro at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97494

Rainer Orth <ro at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
              Build|powerpc64*-linux-gnu        |
             Target|powerpc64*-linux-gnu        |powerpc64*-linux-gnu,
                   |                            |sparc-sun-solaris2.11
                 CC|                            |ro at gcc dot gnu.org
               Host|powerpc64*-linux-gnu        |

--- Comment #3 from Rainer Orth <ro at gcc dot gnu.org> ---
Also seen on 32 and 64-bit sparc.

[Bug tree-optimization/97494] [11 regression] several vector test case failures starting with r11-3966

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97494

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Priority|P3                          |P1

[Bug tree-optimization/97494] [11 regression] several vector test case failures starting with r11-3966

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97494

--- Comment #4 from Richard Biener <rguenth at gcc dot gnu.org> ---
OK, so for gcc.dg/vect/slp-11b.c we're now doing hybrid vectorization if a { 0
2 1 } load permute is supported.  The reason this is needed is that

      out[i*4] = (in[i*4] + 2) * 3;
      out[i*4 + 1] = (in[i*4 + 2] + 2) * 7;
      out[i*4 + 2] = (in[i*4 + 1] + 7) * 3;
      out[i*4 + 3] = (in[i*4 + 3] + 3) * 4;

is "inconsistently" folded like

      out[i*4] = (in[i*4] * 3 + 6);
      out[i*4 + 1] = (in[i*4 + 2] * 7 + 14);
      out[i*4 + 2] = (in[i*4 + 1] * 3 + 21);
      out[i*4 + 3] = (in[i*4 + 3] + 3) * 4;

so for (x + 3) * 4 we're _not_ associating.  That breaks SLP discovery
but with splitting we're now using SLP for the first three lane and
interleaving for the last.

[Bug tree-optimization/97494] [11 regression] several vector test case failures starting with r11-3966

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97494

--- Comment #5 from Richard Biener <rguenth at gcc dot gnu.org> ---
I guess this folding detail wasn't what the original testcase was supposed to
test.  Doing

      out[i*4] = (in[i*4] + 2) * 3;
      out[i*4 + 1] = (in[i*4 + 2] + 2) * 7;
      out[i*4 + 2] = (in[i*4 + 1] + 7) * 3;
      out[i*4 + 3] = (in[i*4 + 3] + 3) * 7;

will fix that but then we still need some good target selector given likely
not all targets can do this permute but we do not have "good" target-supports
for specific permutes.  Maybe

/* { dg-final { scan-tree-dump-times "vectorized 1 loops" 1 "vect" { target { {
vect_strided4 || vect_perm} && vect_int_mult } } } } */
/* { dg-final { scan-tree-dump-times "vectorizing stmts using SLP" 1 "vect" {
target { vect_perm && vect_int_mult } } } } */

will do though.  I will try that.

[it would be nice to have separate PRs for different test fails]

[Bug tree-optimization/97494] [11 regression] several vector test case failures starting with r11-3966

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97494

--- Comment #6 from CVS Commits <cvs-commit at gcc dot gnu.org> ---
The master branch has been updated by Richard Biener <rguenth@gcc.gnu.org>:

https://gcc.gnu.org/g:e393f03b1a73d75901d1bc49c99123bdf534e120

commit r11-6770-ge393f03b1a73d75901d1bc49c99123bdf534e120
Author: Richard Biener <rguenther@suse.de>
Date:   Mon Jan 18 15:18:15 2021 +0100

    testsuite/97494 - adjust gcc.dg/vect/slp-11b.c

    Support for loop SLP splitting exposed that slp-11b.c has
    folding that breaks SLP discovery which isn't what was intended
    when the testcase was written.  The following makes it SLP-able
    and "only" run into the issue that a load permutation is required.

    And tries to adjust the target selectors accordingly.

    2021-01-18  Richard Biener  <rguenther@suse.de>

            PR testsuite/97494
            * gcc.dg/vect/slp-11b.c: Adjust.

[Bug tree-optimization/97494] [11 regression] several vector test case failures starting with r11-3966

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97494

--- Comment #7 from CVS Commits <cvs-commit at gcc dot gnu.org> ---
The master branch has been updated by Richard Biener <rguenth@gcc.gnu.org>:

https://gcc.gnu.org/g:9d4ad8570c3bab2e145aa7b39ce8c66bafbf1945

commit r11-7620-g9d4ad8570c3bab2e145aa7b39ce8c66bafbf1945
Author: Richard Biener <rguenther@suse.de>
Date:   Thu Mar 11 13:30:56 2021 +0100

    testsuite/97494 - amend gcc.dg/vect/slp-21.c

    As reported in the PR all powerpc64 targets fail

    FAIL: gcc.dg/vect/slp-21.c scan-tree-dump-times vect "vectorizing stmts
using SLP" 2

    because like on arm we now vectorize 4 opportunities.  This adjusts
    the testcase to follow the arm example.

    2021-03-11  Richard Biener  <rguenther@suse.de>

            PR testsuite/97494
            * gcc.dg/vect/slp-21.c: Adjust for powerpc64*-*-*.

[Bug tree-optimization/97494] [11 regression] several vector test case failures starting with r11-3966

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97494

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |rsandifo at gcc dot gnu.org

--- Comment #8 from Richard Biener <rguenth at gcc dot gnu.org> ---
The power7 failure with

FAIL: gcc.dg/vect/vect-complex-5.c scan-tree-dump-times vect "vectorizing stmts
using SLP" 2

Shows that alignment checking as done in get_group_load_store_type does not
match what we later code-generate.  The first group, storing into

      c[i].f1 = a1[i] + b1[i];

is said to be aligned while the second

      c[i].f2 = a2[i] + b2[i];

is said to be unaligned and thus unsupported.  In reality both are
VMAT_STRIDED_SLP which can end up with a variety of alignment requirements.
In this case we emit a series of SImode stores to SImode aligned memory
in both cases.

Thus,

static bool
get_group_load_store_type (vec_info *vinfo, stmt_vec_info stmt_info,
...
  if (*memory_access_type == VMAT_GATHER_SCATTER
      || *memory_access_type == VMAT_ELEMENTWISE)
    *alignment_support_scheme = dr_unaligned_supported;
  else
    *alignment_support_scheme
      = vect_supportable_dr_alignment (vinfo, first_dr_info, false);

is incomplete and misses special-casing of VMAT_STRIDED_SLP.  For
VMAT_ELEMENTWISE we seem to simply assume that element (scalar)
misaligned accesses are OK (ISTR we reject non-element aligned vectorization),
but with VMAT_STRIDED_SLP we eventually use SImode to access two HImode
components which are not necessarily aligned according to SImode.

There's the related bug where the alignment we compute for a vector DR
is biased when the stride is negative and which also results in bogus
alignment to be used for checks.

ISTR I wrote somewhere that vectorizable_load/store analysis & costing
should possibly simply run through a common code path for analysis
and code generation (fending off actual stmt generation for analysis)
to avoid this kind of disconnect between code gen assumptions and
analysis parts.  Thus the classical if (!vec_stmt) return early; approach
isn't good for maintainance.

That said, gcc.dg/vect/vect-complex-5.c is a missed optimization on
!hw-misaligned targets, thus I'm going to XFAIL it for ! vect_hw_misalign.

[Bug tree-optimization/97494] [11 regression] several vector test case failures starting with r11-3966

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97494

--- Comment #9 from CVS Commits <cvs-commit at gcc dot gnu.org> ---
The master branch has been updated by Richard Biener <rguenth@gcc.gnu.org>:

https://gcc.gnu.org/g:f6c1d026c3a9109de58d1148e3f54b5273ebb86a

commit r11-7621-gf6c1d026c3a9109de58d1148e3f54b5273ebb86a
Author: Richard Biener <rguenther@suse.de>
Date:   Thu Mar 11 14:03:55 2021 +0100

    testsuite/97494 - XFAIL gcc.dg/vect/vect-complex-5.c on !vect_hw_misalign

    This is a missed optimization due to bogus alignment analysis.

    2021-03-11  Richard Biener  <rguenther@suse.de>

            PR testsuite/97494
            * gcc.dg/vect/vect-complex-5.c: XFAIL on !vect_hw_misalign.

[Bug tree-optimization/97494] [11 regression] several vector test case failures starting with r11-3966

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97494

--- Comment #10 from CVS Commits <cvs-commit at gcc dot gnu.org> ---
The master branch has been updated by Richard Biener <rguenth@gcc.gnu.org>:

https://gcc.gnu.org/g:aa0d8a3e28cd3365a2910884ce3e7cb67b07f866

commit r11-7622-gaa0d8a3e28cd3365a2910884ce3e7cb67b07f866
Author: Richard Biener <rguenther@suse.de>
Date:   Thu Mar 11 14:11:08 2021 +0100

    testsuite/97494 - XFAIL gcc.dg/vect/pr97428.c on !vect_hw_misalign

    While we could at least vectorize it on targets which support
    re-alignment tokens we fail to do this because of imperfections in
    alignment analysis.  XFAIL when the HW cannot deal with misaligned
    vector accesses for now.

    2021-03-11  Richard Biener  <rguenther@suse.de>

            PR testsuite/97494
            * gcc.dg/vect/pr97428.c: XFAIL on !vect_hw_misalign.

[Bug tree-optimization/97494] [11 regression] several vector test case failures starting with r11-3966

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97494

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|ASSIGNED                    |RESOLVED
         Resolution|---                         |FIXED

--- Comment #11 from Richard Biener <rguenth at gcc dot gnu.org> ---
All cases in the original report have been addressed.  If other cases
persist/pop up, please open a new PR.  If the originally reported FAILs persist
fell free to re-open (I only have limited ppc testing capabilities).

[Bug tree-optimization/97494] [11 regression] several vector test case failures starting with r11-3966

[seurer at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97494

--- Comment #12 from seurer at gcc dot gnu.org ---
I looked through the test results on all our testers and it looks like
everything is cleaned up now.  Thanks!