[Bug middle-end/106449] New: ICE in #pragma omp parallel for simd

[Bug middle-end/106449] New: ICE in #pragma omp parallel for simd

[sandra at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106449

            Bug ID: 106449
           Summary: ICE in #pragma omp parallel for simd
           Product: gcc
           Version: 12.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: middle-end
          Assignee: unassigned at gcc dot gnu.org
          Reporter: sandra at gcc dot gnu.org
  Target Milestone: ---

Test case is derived from c-c++-common/gomp/loop-8.c, with "simd" added:

void
foo (void)
{
  int a[1024];
  int *p, *q;
  #pragma omp parallel for simd collapse(2)
  for (p = &a[0]; p < &a[512]; p++)
    for (q = p + 64; q < p + 128; q++)
      ;
}

$ x86_64-none-linux-gnu-gcc -fopenmp loop-8.c
during GIMPLE pass: ompexp
loop-8.c: In function 'foo':
loop-8.c:6:11: internal compiler error: in build2, at tree.cc:5023
    6 |   #pragma omp parallel for simd collapse(2)
      |           ^~~
0x19e4116 build2(tree_code, tree_node*, tree_node*, tree_node*)
        /path/to/gcc/tree.cc:5022
0xf81b3f build2_loc
        /path/to/gcc/tree.h:4577
0xfc0d03 fold_build2_loc(unsigned int, tree_code, tree_node*, tree_node*,
tree_node*)
        /path/to/gcc/fold-const.cc:13860
0xfb26f3 fold_binary_loc(unsigned int, tree_code, tree_node*, tree_node*,
tree_node*)
        /path/to/gcc/fold-const.cc:10894
0xfc0cdc fold_build2_loc(unsigned int, tree_code, tree_node*, tree_node*,
tree_node*)
        /path/to/gcc/fold-const.cc:13858
0x380f1ca expand_omp_simd
        /path/to/gcc/omp-expand.cc:6882
0x3815c41 expand_omp_for
        /path/to/gcc/omp-expand.cc:8090
0x381e79d expand_omp
        /path/to/gcc/omp-expand.cc:10357
0x381e717 expand_omp
        /path/to/gcc/omp-expand.cc:10343
0x381e717 expand_omp
        /path/to/gcc/omp-expand.cc:10343
0x381edbb execute_expand_omp
        /path/to/gcc/omp-expand.cc:10592
0x381ee95 execute
        /path/to/gcc/omp-expand.cc:10639

[Bug middle-end/106449] ICE in #pragma omp parallel for simd

[burnus at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106449

Tobias Burnus <burnus at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|                            |ice-on-valid-code
     Ever confirmed|0                           |1
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2022-07-26

--- Comment #1 from Tobias Burnus <burnus at gcc dot gnu.org> ---
My impression is that the ICE is due to having
  fold_build2_loc (0, PLUS_EXPR, type, op0, op1)
with
(gdb) p debug_tree(type)
 <pointer_type 0x7ffff6feda80 [...]
(gdb) p debug_tree(op0)
 <integer_cst 0x7ffff7128b70 type <pointer_type 0x7ffff6feda80> constant 256>
(gdb) p debug_tree(op1)
 <var_decl 0x7ffff7148480 p.0
    type <pointer_type 0x7ffff6feda80  [...]

Namely: variable PLUS_EXPR integer_cst  - instead of a POINTER_PLUS_EXPR.


Patch would be something like the following – except that there are several
additional PLUS_EXPR in that function and that fold_build_pointer_plus assumes
that the first argument is the pointer and the second the offset, which is
converted via convert_to_ptrofftype_loc ...

--- a/gcc/omp-expand.cc
+++ b/gcc/omp-expand.cc
@@ -6882 +6882,4 @@ expand_omp_simd (struct omp_region *region, struct
omp_for_data *fd)
-             t = fold_build2 (PLUS_EXPR, TREE_TYPE (t), t, t2);
+             if (POINTER_TYPE_P (TREE_TYPE (t)))
+               t = fold_build_pointer_plus (t, t2);
+             else
+               t = fold_build2 (PLUS_EXPR, TREE_TYPE (t), t, t2);

[Bug middle-end/106449] ICE in #pragma omp parallel for simd

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106449

--- Comment #2 from Richard Biener <rguenth at gcc dot gnu.org> ---
(In reply to Tobias Burnus from comment #1)
> My impression is that the ICE is due to having
>   fold_build2_loc (0, PLUS_EXPR, type, op0, op1)
> with
> (gdb) p debug_tree(type)
>  <pointer_type 0x7ffff6feda80 [...]
> (gdb) p debug_tree(op0)
>  <integer_cst 0x7ffff7128b70 type <pointer_type 0x7ffff6feda80> constant 256>
> (gdb) p debug_tree(op1)
>  <var_decl 0x7ffff7148480 p.0
>     type <pointer_type 0x7ffff6feda80  [...]
> 
> Namely: variable PLUS_EXPR integer_cst  - instead of a POINTER_PLUS_EXPR.
> 
> 
> Patch would be something like the following – except that there are several
> additional PLUS_EXPR in that function and that fold_build_pointer_plus
> assumes that the first argument is the pointer and the second the offset,
> which is converted via convert_to_ptrofftype_loc ...

I think the above also shows that op0 is the offset and op1 the base
and thus the arguments should be reversed here.

> --- a/gcc/omp-expand.cc
> +++ b/gcc/omp-expand.cc
> @@ -6882 +6882,4 @@ expand_omp_simd (struct omp_region *region, struct
> omp_for_data *fd)
> -             t = fold_build2 (PLUS_EXPR, TREE_TYPE (t), t, t2);
> +             if (POINTER_TYPE_P (TREE_TYPE (t)))
> +               t = fold_build_pointer_plus (t, t2);
> +             else
> +               t = fold_build2 (PLUS_EXPR, TREE_TYPE (t), t, t2);

[Bug middle-end/106449] ICE in #pragma omp parallel for simd since r6-4544-ge01d41e553aae245

[marxin at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106449

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |marxin at gcc dot gnu.org
            Summary|ICE in #pragma omp parallel |ICE in #pragma omp parallel
                   |for simd                    |for simd since
                   |                            |r6-4544-ge01d41e553aae245

--- Comment #3 from Martin Liška <marxin at gcc dot gnu.org> ---
Btw. started with r6-4544-ge01d41e553aae245.

[Bug middle-end/106449] ICE in #pragma omp parallel for simd since r6-4544-ge01d41e553aae245

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106449

--- Comment #4 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
void
foo (void)
{
  int a[1024];
  int *p, *q;
  #pragma omp simd collapse(2)
  for (p = &a[0]; p < &a[512]; p++)
    for (q = p + 64; q < p + 128; q++)
      ;
}

is enough to reproduce it.

[Bug middle-end/106449] ICE in #pragma omp parallel for simd since r6-4544-ge01d41e553aae245

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106449

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |ASSIGNED
           Assignee|unassigned at gcc dot gnu.org      |jakub at gcc dot gnu.org

[Bug middle-end/106449] ICE in #pragma omp parallel for simd since r6-4544-ge01d41e553aae245

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106449

--- Comment #5 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Created attachment 53362
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=53362&action=edit
gcc13-pr106449.patch

Untested fix.

[Bug middle-end/106449] ICE in #pragma omp parallel for simd since r6-4544-ge01d41e553aae245

[burnus at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106449

--- Comment #6 from Tobias Burnus <burnus at gcc dot gnu.org> ---
Comment on attachment 53362
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=53362
gcc13-pr106449.patch

> +		  t = fold_convert (sizetype, fd->loops[i + 1].n1);
> +		  t = fold_build_pointer_plus (t2, t);

I think the first line is not needed, given that fold_build_pointer_plus_loc
has:

  return fold_build2_loc (loc, POINTER_PLUS_EXPR, TREE_TYPE (ptr),
                          ptr, convert_to_ptrofftype_loc (loc, off));

and the convert_to_ptrofftype_loc does:

  if (ptrofftype_p (TREE_TYPE (off)))
    return off;
  return fold_convert_loc (loc, sizetype, off);

where the last line matches your first line. Hence, just calling

                  t = fold_build_pointer_plus (t2, fd->loops[i + 1].n1);

should gives the same result and is more compact (avoids a line and a
2-character indent).

[Bug middle-end/106449] ICE in #pragma omp parallel for simd since r6-4544-ge01d41e553aae245

[burnus at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106449

--- Comment #7 from Tobias Burnus <burnus at gcc dot gnu.org> ---
Comment on attachment 53362
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=53362
gcc13-pr106449.patch

The code does not seem to work. If I add:

void foobar(int *a, int*b) {
  __builtin_printf("%lu\n", b-a);
}

and use it in place for the ';' in the loops, when then calling
  foo();
  bar(64, 128);

While foo() with and without OpenMP and - with -fno-openmp - also bar(64,128)
give all identical result. But -fopenmp, 'bar' does not call 'foobar' at all
(no printf output).

[Bug middle-end/106449] ICE in #pragma omp parallel for simd since r6-4544-ge01d41e553aae245

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106449

--- Comment #8 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Created attachment 53369
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=53369&action=edit
gcc13-pr106449-1.patch

Preparation patch to remove unnecessary fold_converts to sizetype.

[Bug middle-end/106449] ICE in #pragma omp parallel for simd since r6-4544-ge01d41e553aae245

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106449

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
  Attachment #53362|0                           |1
        is obsolete|                            |

--- Comment #9 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Created attachment 53370
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=53370&action=edit
gcc13-pr106449-2.patch

Updated fix for this PR which fixes even the runtime case (it was caused by
missing unsharing and regimplification clobbering a shared tree).

[Bug middle-end/106449] ICE in #pragma omp parallel for simd since r6-4544-ge01d41e553aae245

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106449

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

https://gcc.gnu.org/g:97d32048c04e9787fccadc4bae1c042754503e34

commit r13-1887-g97d32048c04e9787fccadc4bae1c042754503e34
Author: Jakub Jelinek <jakub@redhat.com>
Date:   Fri Jul 29 09:49:11 2022 +0200

    openmp: Fix up handling of non-rectangular simd loops with pointer type
iterators [PR106449]

    There were 2 issues visible on this new testcase, one that we didn't have
    special POINTER_TYPE_P handling in a few spots of expand_omp_simd - for
    pointers we need to use POINTER_PLUS_EXPR and need to have the non-pointer
    part in sizetype, for non-rectangular loop on the other side we can rely on
    multiplication factor 1, pointers can't be multiplied, without those
changes
    we'd ICE.  The other issue was that we put n2 expression directly into a
    comparison in a condition and regimplified that, for the &a[512] case that
    and with gimplification being destructed that unfortunately meant
modification
    of original fd->loops[?].n2.  Fixed by unsharing the expression.  This was
    causing a runtime failure on the testcase.

    2022-07-29  Jakub Jelinek  <jakub@redhat.com>

            PR middle-end/106449
            * omp-expand.cc (expand_omp_simd): Fix up handling of pointer
            iterators in non-rectangular simd loops.  Unshare fd->loops[i].n2
            or n2 before regimplifying it inside of a condition.

            * testsuite/libgomp.c-c++-common/pr106449.c: New test.

[Bug middle-end/106449] ICE in #pragma omp parallel for simd since r6-4544-ge01d41e553aae245

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106449

--- Comment #11 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Fixed on the trunk so far.  Will backport later.

[Bug middle-end/106449] ICE in #pragma omp parallel for simd since r6-4544-ge01d41e553aae245

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106449

--- Comment #12 from CVS Commits <cvs-commit at gcc dot gnu.org> ---
The releases/gcc-12 branch has been updated by Jakub Jelinek
<jakub@gcc.gnu.org>:

https://gcc.gnu.org/g:681c73db9bd156f9b65a73ccc6c4a0a697fe70d6

commit r12-8645-g681c73db9bd156f9b65a73ccc6c4a0a697fe70d6
Author: Jakub Jelinek <jakub@redhat.com>
Date:   Fri Jul 29 09:49:11 2022 +0200

    openmp: Fix up handling of non-rectangular simd loops with pointer type
iterators [PR106449]

    There were 2 issues visible on this new testcase, one that we didn't have
    special POINTER_TYPE_P handling in a few spots of expand_omp_simd - for
    pointers we need to use POINTER_PLUS_EXPR and need to have the non-pointer
    part in sizetype, for non-rectangular loop on the other side we can rely on
    multiplication factor 1, pointers can't be multiplied, without those
changes
    we'd ICE.  The other issue was that we put n2 expression directly into a
    comparison in a condition and regimplified that, for the &a[512] case that
    and with gimplification being destructed that unfortunately meant
modification
    of original fd->loops[?].n2.  Fixed by unsharing the expression.  This was
    causing a runtime failure on the testcase.

    2022-07-29  Jakub Jelinek  <jakub@redhat.com>

            PR middle-end/106449
            * omp-expand.cc (expand_omp_simd): Fix up handling of pointer
            iterators in non-rectangular simd loops.  Unshare fd->loops[i].n2
            or n2 before regimplifying it inside of a condition.

            * testsuite/libgomp.c-c++-common/pr106449.c: New test.

    (cherry picked from commit 97d32048c04e9787fccadc4bae1c042754503e34)

[Bug middle-end/106449] ICE in #pragma omp parallel for simd since r6-4544-ge01d41e553aae245

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106449

--- Comment #13 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Fixed for 12.2+ too.

[Bug middle-end/106449] ICE in #pragma omp parallel for simd since r6-4544-ge01d41e553aae245

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106449

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

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

--- Comment #14 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
This was rejected in 11.x and earlier, so I don't understand the
"Btw. started with r6-4544-ge01d41e553aae245."
comment.  So, fixed now.

[Bug middle-end/106449] ICE in #pragma omp parallel for simd since r6-4544-ge01d41e553aae245

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106449

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
      Known to fail|                            |12.1.0
      Known to work|                            |12.2.0, 13.0
   Target Milestone|---                         |12.2