[Bug target/99825] New: ICE in vect_build_combine_node

[Bug target/99825] New: ICE in vect_build_combine_node

[asolokha at gmx dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=99825

            Bug ID: 99825
           Summary: ICE in vect_build_combine_node
           Product: gcc
           Version: 11.0
            Status: UNCONFIRMED
          Keywords: ice-on-valid-code
          Severity: normal
          Priority: P3
         Component: target
          Assignee: unassigned at gcc dot gnu.org
          Reporter: asolokha at gmx dot com
  Target Milestone: ---
            Target: aarch64-linux-gnu

gfortran-11.0.1-alpha20210328 snapshot
(g:499fa254ae8c9752d8c2cf3130b13ffddfd83546) ICEs when compiling the following
testcase, reduced from gcc/testsuite/gfortran.dg/inline_matmul_11.f90, w/
-march=armv8.3-a -O3:

program main
  complex, dimension(3, 2) :: a
  complex, dimension(2, 4) :: b
  complex, dimension(3, 4) :: c, res1

  data a /0, (-5., -7.), (11., -13.), 0, 0, 0/
  data b /0, 0, 0, 0, 0, 0, 0, 0/
  data res1 /0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0/

  c = matmul(a, conjg(b))
  if (any(res1 /= c)) stop 2
end program main

% aarch64-linux-gnu-gfortran-11.0.1 -march=armv8.3-a -O3 -c xfxklbt7.f90
during GIMPLE pass: vect
xfxklbt7.f90:12:16:

   12 | end program main
      |                ^
internal compiler error: in quick_push, at vec.h:1021
0x7c477e vec<std::pair<unsigned int, unsigned int>, va_heap,
vl_embed>::quick_push(std::pair<unsigned int, unsigned int> const&)
       
/var/tmp/portage/cross-aarch64-linux-gnu/gcc-11.0.1_alpha20210328/work/gcc-11-20210328/gcc/vec.h:1021
0x7c489a vec<std::pair<unsigned int, unsigned int>, va_heap,
vl_embed>::quick_push(std::pair<unsigned int, unsigned int> const&)
       
/var/tmp/portage/cross-aarch64-linux-gnu/gcc-11.0.1_alpha20210328/work/gcc-11-20210328/gcc/tree-vect-slp-patterns.c:907
0x7c489a vec<std::pair<unsigned int, unsigned int>, va_heap,
vl_ptr>::quick_push(std::pair<unsigned int, unsigned int> const&)
       
/var/tmp/portage/cross-aarch64-linux-gnu/gcc-11.0.1_alpha20210328/work/gcc-11-20210328/gcc/vec.h:1875
0x7c489a vect_build_combine_node
       
/var/tmp/portage/cross-aarch64-linux-gnu/gcc-11.0.1_alpha20210328/work/gcc-11-20210328/gcc/tree-vect-slp-patterns.c:891
0x123f945 complex_mul_pattern::build(vec_info*)
       
/var/tmp/portage/cross-aarch64-linux-gnu/gcc-11.0.1_alpha20210328/work/gcc-11-20210328/gcc/tree-vect-slp-patterns.c:1068
0x1232418 vect_match_slp_patterns_2
       
/var/tmp/portage/cross-aarch64-linux-gnu/gcc-11.0.1_alpha20210328/work/gcc-11-20210328/gcc/tree-vect-slp.c:2430
0x12323c4 vect_match_slp_patterns_2
       
/var/tmp/portage/cross-aarch64-linux-gnu/gcc-11.0.1_alpha20210328/work/gcc-11-20210328/gcc/tree-vect-slp.c:2422
0x12323c4 vect_match_slp_patterns_2
       
/var/tmp/portage/cross-aarch64-linux-gnu/gcc-11.0.1_alpha20210328/work/gcc-11-20210328/gcc/tree-vect-slp.c:2422
0x12366b7 vect_match_slp_patterns
       
/var/tmp/portage/cross-aarch64-linux-gnu/gcc-11.0.1_alpha20210328/work/gcc-11-20210328/gcc/tree-vect-slp.c:2458
0x12366b7 vect_analyze_slp(vec_info*, unsigned int)
       
/var/tmp/portage/cross-aarch64-linux-gnu/gcc-11.0.1_alpha20210328/work/gcc-11-20210328/gcc/tree-vect-slp.c:2901
0x1213157 vect_analyze_loop_2
       
/var/tmp/portage/cross-aarch64-linux-gnu/gcc-11.0.1_alpha20210328/work/gcc-11-20210328/gcc/tree-vect-loop.c:2316
0x1214eb3 vect_analyze_loop(loop*, vec_info_shared*)
       
/var/tmp/portage/cross-aarch64-linux-gnu/gcc-11.0.1_alpha20210328/work/gcc-11-20210328/gcc/tree-vect-loop.c:2985
0x1245aa7 try_vectorize_loop_1
       
/var/tmp/portage/cross-aarch64-linux-gnu/gcc-11.0.1_alpha20210328/work/gcc-11-20210328/gcc/tree-vectorizer.c:1009
0x1246540 vectorize_loops()
       
/var/tmp/portage/cross-aarch64-linux-gnu/gcc-11.0.1_alpha20210328/work/gcc-11-20210328/gcc/tree-vectorizer.c:1243

[Bug target/99825] [11 Regression] ICE in vect_build_combine_node

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=99825

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
            Summary|ICE in                      |[11 Regression] ICE in
                   |vect_build_combine_node     |vect_build_combine_node
   Target Milestone|---                         |11.0
     Ever confirmed|0                           |1
                 CC|                            |jakub at gcc dot gnu.org
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2021-03-30
           Priority|P3                          |P1

--- Comment #1 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Started with r11-6734-gad2603433853129e847cade5e269c6a5f889a020

[Bug target/99825] [11 Regression] ICE in vect_build_combine_node

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=99825

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

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

--- Comment #2 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
887       perm.create (SLP_TREE_LANES (rep));
888     
889       for (unsigned x = 0; x < SLP_TREE_LANES (rep); x+=2)
890         {
891           perm.quick_push (std::make_pair (0, x));
892           perm.quick_push (std::make_pair (1, x+1));
893         }

that was added in r11-6693-ge09173d84dabd186ad2463f47350c1bcabfeab8f
can't really work, it allocates just SLP_TREE_LANES (resp) elements,
but can quick_push one more elements - when SLP_TREE_LANES is odd (in this
testcase 3).  So, either some caller needs to ensure SLP_TREE_LANES is even or
this function isn't called, or it needs to be changed to handle even odd
numbers.

[Bug target/99825] [11 Regression] ICE in vect_build_combine_node

[tnfchris at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=99825

Tamar Christina <tnfchris at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Assignee|unassigned at gcc dot gnu.org      |tnfchris at gcc dot gnu.org

--- Comment #3 from Tamar Christina <tnfchris at gcc dot gnu.org> ---
(In reply to Jakub Jelinek from comment #2)
> 887	  perm.create (SLP_TREE_LANES (rep));
> 888	
> 889	  for (unsigned x = 0; x < SLP_TREE_LANES (rep); x+=2)
> 890	    {
> 891	      perm.quick_push (std::make_pair (0, x));
> 892	      perm.quick_push (std::make_pair (1, x+1));
> 893	    }
> 
> that was added in r11-6693-ge09173d84dabd186ad2463f47350c1bcabfeab8f
> can't really work, it allocates just SLP_TREE_LANES (resp) elements,
> but can quick_push one more elements - when SLP_TREE_LANES is odd (in this
> testcase 3).  So, either some caller needs to ensure SLP_TREE_LANES is even
> or this function isn't called, or it needs to be changed to handle even odd
> numbers.

It can't work on odd lanes since the operation needs to update both a real and
imaginary component. So this should have been rejected much earlier.

Testing a patch.

[Bug tree-optimization/99825] [11 Regression] ICE in vect_build_combine_node

[tnfchris at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=99825

Tamar Christina <tnfchris at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
          Component|target                      |tree-optimization
             Status|NEW                         |ASSIGNED

[Bug tree-optimization/99825] [11 Regression] ICE in vect_build_combine_node

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=99825

--- Comment #4 from CVS Commits <cvs-commit at gcc dot gnu.org> ---
The master branch has been updated by Tamar Christina <tnfchris@gcc.gnu.org>:

https://gcc.gnu.org/g:9c68e2abe294a48385224cd7617eca0720144b5c

commit r11-7909-g9c68e2abe294a48385224cd7617eca0720144b5c
Author: Tamar Christina <tamar.christina@arm.com>
Date:   Tue Mar 30 14:16:03 2021 +0100

    slp: reject non-multiple of 2 laned SLP trees (PR99825)

    TWO_OPERANDS allows any order or number of combinations of + and -
operations
    but the pattern matcher only supports pairs of operations.

    This patch has the pattern matcher for complex numbers reject SLP trees
where
    the lanes are not a multiple of 2.

    gcc/ChangeLog:

            PR tree-optimization/99825
            * tree-vect-slp-patterns.c (vect_check_evenodd_blend):
            Reject non-mult 2 lanes.

    gcc/testsuite/ChangeLog:

            PR tree-optimization/99825
            * gfortran.dg/vect/pr99825.f90: New test.

[Bug tree-optimization/99825] [11 Regression] ICE in vect_build_combine_node

[tnfchris at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=99825

Tamar Christina <tnfchris at gcc dot gnu.org> changed:

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

--- Comment #5 from Tamar Christina <tnfchris at gcc dot gnu.org> ---
Fixed.