[Bug c++/106841] New: ice in vect_get_vec_defs_for_operand, at tree-vect-stmts.cc:1509

[Bug c++/106841] New: ice in vect_get_vec_defs_for_operand, at tree-vect-stmts.cc:1509

[dcb314 at hotmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106841

            Bug ID: 106841
           Summary: ice in vect_get_vec_defs_for_operand, at
                    tree-vect-stmts.cc:1509
           Product: gcc
           Version: 12.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: c++
          Assignee: unassigned at gcc dot gnu.org
          Reporter: dcb314 at hotmail dot com
  Target Milestone: ---

Created attachment 53540
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=53540&action=edit
gzipped C++ source code

The attached C++ code does this:

during GIMPLE pass: vect
msh3.cpp: In member function 'virtual AnyType Square_Op::operator()(Stack)
const':
msh3.cpp:8044:9: internal compiler error: in vect_get_vec_defs_for_operand, at
tree-vect-stmts.cc:1509
 8044 | AnyType Square_Op::operator( )(Stack stack) const {
      |         ^~~~~~~~~
0x20af236 vect_get_vec_defs_for_operand(vec_info*, _stmt_vec_info*, unsigned
int, tree_node*, vec<tree_node*, va_heap, vl_ptr>*, tree_node*)
    ../../trunk.git/gcc/tree-vect-stmts.cc:1509
0x20cf508 vect_get_gather_scatter_ops(_loop_vec_info*, loop*, _stmt_vec_info*,
_slp_tree*, gather_scatter_info*, tree_node**, vec<tree_node*, va_heap,
vl_ptr>*)
    ../../trunk.git/gcc/tree-vect-stmts.cc:3066
0x20bab42 vectorizable_load(vec_info*, _stmt_vec_info*, gimple_stmt_iterator*,
gimple**, _slp_tree*, vec<stmt_info_for_cost, va_heap, vl_ptr>*)
    ../../trunk.git/gcc/tree-vect-stmts.cc:9627

when compiled with -c -w -O3 -ffast-math -march=bdver2.

I will have a go at producing a git range.

[Bug c++/106841] ice in vect_get_vec_defs_for_operand, at tree-vect-stmts.cc:1509

[dcb314 at hotmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106841

--- Comment #1 from David Binderman <dcb314 at hotmail dot com> ---
This bug looks to be at least a month old, since it started sometime
before git hash 21c7aab09805d0c8, dated 20220808.

I will have a go at a reduction.

[Bug c++/106841] ice in vect_get_vec_defs_for_operand, at tree-vect-stmts.cc:1509

[dcb314 at hotmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106841

David Binderman <dcb314 at hotmail dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |dcb314 at hotmail dot com

--- Comment #2 from David Binderman <dcb314 at hotmail dot com> ---
Created attachment 53541
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=53541&action=edit
C++ source code

After a couple of hours of reduction, I have the attached code.

[Bug c++/106841] [12/13 Regression] ICE in vect_get_vec_defs_for_operand, at tree-vect-stmts.cc:1509 since r12-2733-g31855ba6b16cd138

[marxin at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106841

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
            Summary|ice in                      |[12/13 Regression] ICE in
                   |vect_get_vec_defs_for_opera |vect_get_vec_defs_for_opera
                   |nd, at                      |nd, at
                   |tree-vect-stmts.cc:1509     |tree-vect-stmts.cc:1509
                   |                            |since
                   |                            |r12-2733-g31855ba6b16cd138
             Status|UNCONFIRMED                 |NEW
   Target Milestone|---                         |12.3
   Last reconfirmed|                            |2022-09-06
                 CC|                            |marxin at gcc dot gnu.org,
                   |                            |rguenth at gcc dot gnu.org
     Ever confirmed|0                           |1

--- Comment #3 from Martin Liška <marxin at gcc dot gnu.org> ---
Thanks for the report!

Reduced to:

$ cat pr106841.C
struct R3 {
  double z;
  R3(R3 A, R3 B) : z(B.z - A.z) {}
  double norme() { return z; }
};
struct TBoundaryEdge {
  int *vertices[2];
  int &operator[](int i) { return *vertices[i]; }
};
struct Mesh {
  int vertices;
  TBoundaryEdge *bedges;
  int operator()(int &vv) { return &vv - &vertices; }
  TBoundaryEdge be(int i) { return bedges[i]; }
};
template <typename Data> struct GenericElement {
  typedef typename Data::V Vertex;
  static const int nv = Data::NbOfVertices;
  Vertex *vertices[nv];
  double mes;
  void set(int *iv) {
    Vertex *v0;
    for (int i = 0; i < nv; ++i)
      vertices[i] = v0 + iv[i];
    mes = Data::mesure(vertices);
  }
};
struct DataSeg3 {
  static const int NbOfVertices = 2;
  typedef R3 V;
  static double mesure(V *pv[]) { return R3(*pv[0], *pv[1]).norme(); }
};
struct MeshS {
  MeshS();
};
template <class> struct Movemesh_Op { int operator()() const; };
template <> int Movemesh_Op<int>::operator()() const {
  Mesh pTh;
  GenericElement<DataSeg3> *bS = new GenericElement<DataSeg3>;
  for (int ibe; ibe; ibe++) {
    TBoundaryEdge K(pTh.be(ibe));
    int iv[2];
    for (int i = 0; i < 2; i++) {
      int &__trans_tmp_2 = K[i];
      iv[i] = pTh(__trans_tmp_2);
    }
    bS[ibe].set(iv);
  }
  MeshS T_Th;
}

$ g++ pr106841.C -c -w -O3 -ffast-math -march=bdver2
during GIMPLE pass: vect
pr106841.C: In member function ‘int Movemesh_Op< <template-parameter-1-1>
>::operator()() const [with <template-parameter-1-1> = int]’:
pr106841.C:37:17: internal compiler error: in vect_get_vec_defs_for_operand, at
tree-vect-stmts.cc:1509
   37 | template <> int Movemesh_Op<int>::operator()() const {
      |                 ^~~~~~~~~~~~~~~~
0x9ea8b8 vect_get_vec_defs_for_operand(vec_info*, _stmt_vec_info*, unsigned
int, tree_node*, vec<tree_node*, va_heap, vl_ptr>*, tree_node*)
        /home/marxin/Programming/gcc/gcc/tree-vect-stmts.cc:1509
0x2236ade vect_get_gather_scatter_ops
        /home/marxin/Programming/gcc/gcc/tree-vect-stmts.cc:3066
0x2247d85 vectorizable_load
        /home/marxin/Programming/gcc/gcc/tree-vect-stmts.cc:9627
0x2250120 vect_transform_stmt(vec_info*, _stmt_vec_info*,
gimple_stmt_iterator*, _slp_tree*, _slp_instance*)
        /home/marxin/Programming/gcc/gcc/tree-vect-stmts.cc:11344
0x15081f1 vect_transform_loop_stmt
        /home/marxin/Programming/gcc/gcc/tree-vect-loop.cc:9606
0x1523202 vect_transform_loop(_loop_vec_info*, gimple*)
        /home/marxin/Programming/gcc/gcc/tree-vect-loop.cc:10042
0x15649de vect_transform_loops
        /home/marxin/Programming/gcc/gcc/tree-vectorizer.cc:1007
0x1564fbf try_vectorize_loop_1
        /home/marxin/Programming/gcc/gcc/tree-vectorizer.cc:1153
0x1564fbf try_vectorize_loop
        /home/marxin/Programming/gcc/gcc/tree-vectorizer.cc:1183
0x1565354 execute
        /home/marxin/Programming/gcc/gcc/tree-vectorizer.cc:1299
Please submit a full bug report, with preprocessed source (by using
-freport-bug).
Please include the complete backtrace with any bug report.
See <https://gcc.gnu.org/bugs/> for instructions.

And it started with r12-2733-g31855ba6b16cd138

[Bug c++/106841] [12/13 Regression] ICE in vect_get_vec_defs_for_operand, at tree-vect-stmts.cc:1509 since r12-2733-g31855ba6b16cd138

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106841

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|NEW                         |ASSIGNED

--- Comment #4 from Richard Biener <rguenth at gcc dot gnu.org> ---
The issue seems to be that we get

t.ii:40:17: note:   vect_recog_widen_mult_pattern: detected: _56 = _55 * 8;
t.ii:40:17: note:   widen_mult pattern recognized: patt_36 = (long unsigned
int) patt_38;
t.ii:40:17: note:   extra pattern stmt: patt_38 = _76 w* 8;
...
t.ii:40:17: note:  ------>vectorizing statement: _55 = (long unsigned int) _76;
t.ii:40:17: note:  ------>vectorizing statement: patt_63 = _76 w* 8;
t.ii:40:17: note:  transform statement.
...
t.ii:40:17: note:  ------>vectorizing statement: A$z_42 = MEM[(const struct R3
&)_57].z;
t.ii:40:17: note:  multiple-types.
t.ii:40:17: note:  transform statement.
t.ii:40:17: note:  vect_is_simple_use: operand (long unsigned int) _76, type of
def: internal
t.ii:40:17: note:  vect_is_simple_use: vectype vector(2) long unsigned int
t.ii:40:17: note:  transform load. ncopies = 2
t.ii:40:17: note:  vect_get_vec_defs_for_operand: _55

and a SLP tree that seems to run into a bug with hybrid discovery:

t.ii:40:17: note:   node 0x654a7d8 (max_nunits=4, refcnt=2) vector(4) int
t.ii:40:17: note:   op template: _76 = (int) _75;
t.ii:40:17: note:       stmt 0 _76 = (int) _75;
t.ii:40:17: note:       stmt 1 _84 = (int) _83;
t.ii:40:17: note:       children 0x654a860

t.ii:40:17: note:   Processing hybrid candidate : _55 = (long unsigned int)
_76;
t.ii:40:17: note:   Marked SLP consumed stmt pure: _55 = (long unsigned int)
_76;

that is because the use of _55 is

 _56 = _55 * 8;

but that is part of a pattern that is PURE_SLP and we fail to see the
direct use in the gather since gather discovery skips this scaling stmt.

We can probably detect this after the fact but the best fix would be to
somehow mark _55 as hybrid, I'll see if I can manage to do that.

[Bug c++/106841] [12/13 Regression] ICE in vect_get_vec_defs_for_operand, at tree-vect-stmts.cc:1509 since r12-2733-g31855ba6b16cd138

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106841

--- Comment #5 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:e33e61d417eb5e981bb7d709f8681a2f55ed518a

commit r13-2492-ge33e61d417eb5e981bb7d709f8681a2f55ed518a
Author: Richard Biener <rguenther@suse.de>
Date:   Tue Sep 6 10:08:44 2022 +0200

    tree-optimization/106841 - gather and hybrid SLP

    Hybrid SLP detection currently fails to consider a not direct
    offset operand of a scatter/gather operation.  The following fixes
    this.

            PR tree-optimization/106841
            * tree-vect-slp.cc (vect_detect_hybrid_slp): Also process
            scatter/gather offset.

            * g++.dg/vect/pr106841.cc: New testcase.

[Bug c++/106841] [12 Regression] ICE in vect_get_vec_defs_for_operand, at tree-vect-stmts.cc:1509 since r12-2733-g31855ba6b16cd138

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106841

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
            Summary|[12/13 Regression] ICE in   |[12 Regression] ICE in
                   |vect_get_vec_defs_for_opera |vect_get_vec_defs_for_opera
                   |nd, at                      |nd, at
                   |tree-vect-stmts.cc:1509     |tree-vect-stmts.cc:1509
                   |since                       |since
                   |r12-2733-g31855ba6b16cd138  |r12-2733-g31855ba6b16cd138
      Known to fail|                            |12.2.0
           Priority|P3                          |P2
      Known to work|                            |13.0

--- Comment #6 from Richard Biener <rguenth at gcc dot gnu.org> ---
Fixed on trunk sofar.

[Bug c++/106841] [12 Regression] ICE in vect_get_vec_defs_for_operand, at tree-vect-stmts.cc:1509 since r12-2733-g31855ba6b16cd138

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106841

--- Comment #7 from CVS Commits <cvs-commit at gcc dot gnu.org> ---
The releases/gcc-12 branch has been updated by Richard Biener
<rguenth@gcc.gnu.org>:

https://gcc.gnu.org/g:41b4faa130a32b37debb1f92e3fa93b2fe8571fc

commit r12-8752-g41b4faa130a32b37debb1f92e3fa93b2fe8571fc
Author: Richard Biener <rguenther@suse.de>
Date:   Tue Sep 6 10:08:44 2022 +0200

    tree-optimization/106841 - gather and hybrid SLP

    Hybrid SLP detection currently fails to consider a not direct
    offset operand of a scatter/gather operation.  The following fixes
    this.

            PR tree-optimization/106841
            * tree-vect-slp.cc (vect_detect_hybrid_slp): Also process
            scatter/gather offset.

            * g++.dg/vect/pr106841.cc: New testcase.

    (cherry picked from commit e33e61d417eb5e981bb7d709f8681a2f55ed518a)

[Bug c++/106841] [12 Regression] ICE in vect_get_vec_defs_for_operand, at tree-vect-stmts.cc:1509 since r12-2733-g31855ba6b16cd138

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106841

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|ASSIGNED                    |RESOLVED
      Known to work|                            |12.2.1
         Resolution|---                         |FIXED

--- Comment #8 from Richard Biener <rguenth at gcc dot gnu.org> ---
Fixed.

[Bug c++/106841] [12 Regression] ICE in vect_get_vec_defs_for_operand, at tree-vect-stmts.cc:1509 since r12-2733-g31855ba6b16cd138

[kenneth.hoste at ugent dot be]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106841

Kenneth Hoste <kenneth.hoste at ugent dot be> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |kenneth.hoste at ugent dot be

--- Comment #9 from Kenneth Hoste <kenneth.hoste at ugent dot be> ---
I seem to be running into the same problem using GCC 11.3 (when compiling
OpenMM 7.7.0 with CUDA support). Is there a reason why the fix wasn't
backported to GCC 11 too?

[Bug c++/106841] [12 Regression] ICE in vect_get_vec_defs_for_operand, at tree-vect-stmts.cc:1509 since r12-2733-g31855ba6b16cd138

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106841

--- Comment #10 from Richard Biener <rguenth at gcc dot gnu.org> ---
(In reply to Kenneth Hoste from comment #9)
> I seem to be running into the same problem using GCC 11.3 (when compiling
> OpenMM 7.7.0 with CUDA support). Is there a reason why the fix wasn't
> backported to GCC 11 too?

Because we don't have a testcase that fails on the GCC 11 branch.  Can you
confirm that cherry-picking the fix fixes the issue you see?