[Bug tree-optimization/60502] New: 4.9 ICE reassociation and vector types.

[Bug tree-optimization/60502] New: 4.9 ICE reassociation and vector types.

[graham.stott at btinternet dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=60502

            Bug ID: 60502
           Summary: 4.9 ICE reassociation and vector types.
           Product: gcc
           Version: 4.9.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: tree-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: graham.stott at btinternet dot com

gcc_unreacble abort in reassociation and vector types

If reassocation manages to simplify an expression which involves any VECTOR
type
it will trigger a gcc_unreachable abort.

On any target which has support for vector types.

------------------------------------------------------------
typedef signed char v16i8 __attribute__ ((vector_size(16)));

v16i8 bug1(v16i8 op0)
{
  v16i8 m1 = {-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1};
  return (op0 ^ m1) | (op0 & m1);
}
------------------------------------------------------------

It hits a gcc_unreachable() in build_int_cst_wide.

bug.c:3:7: internal compiler error: in build_int_cst_wide, at tree.c:1200
 v16i8 bug1(v16i8 op0)
       ^
0xb75583 build_int_cst_wide(tree_node*, unsigned long, long)
    /gcc/gcc/gcc/tree.c:1200
0xaade63 eliminate_not_pairs
    /gcc/gcc/gcc/tree-ssa-reassoc.c:810
0xaade63 optimize_ops_list
    /gcc/gcc/gcc/tree-ssa-reassoc.c:1744
0xab0a99 reassociate_bb
    /gcc/gcc/gcc/tree-ssa-reassoc.c:4436
0xab08a7 reassociate_bb
    /gcc/gcc/gcc/tree-ssa-reassoc.c:4519
0xab311b do_reassoc
    /gcc/gcc/gcc/tree-ssa-reassoc.c:4552
0xab311b execute_reassoc
    /gcc/gcc/gcc/tree-ssa-reassoc.c:4634
0xab311b execute
    /gcc/gcc/gcc/tree-ssa-reassoc.c:4676

[Bug tree-optimization/60502] [4.8/4.9 Regression] ICE reassociation and vector types.

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=60502

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2014-03-11
                 CC|                            |jakub at gcc dot gnu.org
   Target Milestone|---                         |4.8.3
            Summary|4.9 ICE reassociation and   |[4.8/4.9 Regression] ICE
                   |vector types.               |reassociation and vector
                   |                            |types.
     Ever confirmed|0                           |1

--- Comment #1 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Started with r184923.

[Bug tree-optimization/60502] [4.8/4.9 Regression] ICE reassociation and vector types.

[glisse at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=60502

--- Comment #2 from Marc Glisse <glisse at gcc dot gnu.org> ---
Indeed. Probably should call build_all_ones_cst instead of build_low_bits_mask
in eliminate_not_pairs.

[Bug tree-optimization/60502] [4.8/4.9 Regression] ICE reassociation and vector types.

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=60502

--- Comment #3 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
(In reply to Marc Glisse from comment #2)
> Indeed. Probably should call build_all_ones_cst instead of
> build_low_bits_mask in eliminate_not_pairs.

Agreed.  Are you going to prepare a patch, or should I?  Guess we don't want to
use v16i8 in fn return type or argument, it needs to be
typedef signed char v16i8 __attribute__ ((vector_size (16)));
typedef unsigned char v16u8 __attribute__ ((vector_size (16)));

void
foo (v16i8 *x)
{
  v16i8 op0 = *x;
  v16i8 m1 = { -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1
};
  *x = (op0 ^ m1) | (op0 & m1);
}

void
bar (v16u8 *x)
{
  v16u8 op0 = *x;
  v16u8 m1 = { -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1
};
  *x = (op0 ^ m1) | (op0 & m1);
}

instead.

[Bug tree-optimization/60502] [4.8/4.9 Regression] ICE reassociation and vector types.

[glisse at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=60502

Marc Glisse <glisse at gcc dot gnu.org> changed:

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

--- Comment #4 from Marc Glisse <glisse at gcc dot gnu.org> ---
(In reply to Jakub Jelinek from comment #3)
> Are you going to prepare a patch, or should I?

Please go ahead (or I'll do it in a few days).

Note that the &m1 part is not necessary to reproduce: *x |= (*x ^ m1);

[Bug tree-optimization/60502] [4.8/4.9 Regression] ICE reassociation and vector types.

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=60502

--- Comment #5 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Author: jakub
Date: Wed Mar 12 09:21:59 2014
New Revision: 208507

URL: http://gcc.gnu.org/viewcvs?rev=208507&root=gcc&view=rev
Log:
    PR tree-optimization/60502
    * tree-ssa-reassoc.c (eliminate_not_pairs): Use build_all_ones_cst
    instead of build_low_bits_mask.

    * gcc.c-torture/compile/pr60502.c: New test.

Added:
    trunk/gcc/testsuite/gcc.c-torture/compile/pr60502.c
Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/testsuite/ChangeLog
    trunk/gcc/tree-ssa-reassoc.c

[Bug tree-optimization/60502] [4.8 Regression] ICE reassociation and vector types.

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=60502

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |ASSIGNED
      Known to work|                            |4.9.0
           Assignee|unassigned at gcc dot gnu.org      |jakub at gcc dot gnu.org
            Summary|[4.8/4.9 Regression] ICE    |[4.8 Regression] ICE
                   |reassociation and vector    |reassociation and vector
                   |types.                      |types.
      Known to fail|                            |4.8.2

--- Comment #6 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Fixed on the trunk so far.

[Bug tree-optimization/60502] [4.8 Regression] ICE reassociation and vector types.

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=60502

--- Comment #7 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Author: jakub
Date: Thu Apr 10 09:35:39 2014
New Revision: 209274

URL: http://gcc.gnu.org/viewcvs?rev=209274&root=gcc&view=rev
Log:
    Backport from mainline
    2014-03-12  Jakub Jelinek  <jakub@redhat.com>
            Marc Glisse  <marc.glisse@inria.fr>

    PR tree-optimization/60502
    * tree-ssa-reassoc.c (eliminate_not_pairs): Use build_all_ones_cst
    instead of build_low_bits_mask.

    * gcc.c-torture/compile/pr60502.c: New test.

    2013-06-13  Marc Glisse  <marc.glisse@inria.fr>

    * tree.c (build_all_ones_cst): New function.
    * tree.h (build_all_ones_cst): Declare it.

    2013-05-10  Marc Glisse  <marc.glisse@inria.fr>

    * tree.c (build_minus_one_cst): New function.
    * tree.h (build_minus_one_cst): Declare new function.

Modified:
    branches/gcc-4_8-branch/gcc/ChangeLog
    branches/gcc-4_8-branch/gcc/testsuite/ChangeLog
    branches/gcc-4_8-branch/gcc/tree-ssa-reassoc.c
    branches/gcc-4_8-branch/gcc/tree.c
    branches/gcc-4_8-branch/gcc/tree.h

[Bug tree-optimization/60502] [4.8 Regression] ICE reassociation and vector types.

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=60502

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|ASSIGNED                    |RESOLVED
      Known to work|                            |4.8.3
         Resolution|---                         |FIXED

--- Comment #8 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Fixed for 4.8.3+ too.

[Bug tree-optimization/60502] [4.8 Regression] ICE reassociation and vector types.

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=60502

--- Comment #9 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Author: jakub
Date: Mon May  5 08:07:35 2014
New Revision: 210062

URL: http://gcc.gnu.org/viewcvs?rev=210062&root=gcc&view=rev
Log:
    Backport from mainline
    2014-03-12  Jakub Jelinek  <jakub@redhat.com>
            Marc Glisse  <marc.glisse@inria.fr>

    PR tree-optimization/60502
    * tree-ssa-reassoc.c (eliminate_not_pairs): Use build_all_ones_cst
    instead of build_low_bits_mask.

    * gcc.c-torture/compile/pr60502.c: New test.

    2013-06-13  Marc Glisse  <marc.glisse@inria.fr>

    * tree.c (build_all_ones_cst): New function.
    * tree.h (build_all_ones_cst): Declare it.

    2013-05-10  Marc Glisse  <marc.glisse@inria.fr>

    * tree.c (build_minus_one_cst): New function.
    * tree.h (build_minus_one_cst): Declare new function.

Added:
    branches/gcc-4_8-branch/gcc/testsuite/gcc.c-torture/compile/pr60502.c