[Bug testsuite/99955] New: gcc.c-torture/execute/pr92618.c violates strict aliasing rules

[Bug testsuite/99955] New: gcc.c-torture/execute/pr92618.c violates strict aliasing rules

[acoplan at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=99955

            Bug ID: 99955
           Summary: gcc.c-torture/execute/pr92618.c violates strict
                    aliasing rules
           Product: gcc
           Version: 11.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: testsuite
          Assignee: unassigned at gcc dot gnu.org
          Reporter: acoplan at gcc dot gnu.org
  Target Milestone: ---

The above execution test trips the strict aliasing warnings:

aarch64-linux-gnu-gcc -O2 -Wall gcc/testsuite/gcc.c-torture/execute/pr92618.c
-static
gcc/testsuite/gcc.c-torture/execute/pr92618.c: In function ‘baz’:
gcc/testsuite/gcc.c-torture/execute/pr92618.c:44:4: warning: dereferencing
type-punned pointer will break strict-aliasing rules [-Wstrict-aliasing]
   44 |   *(__m128i *) &a[0] = c;
      |    ^~~~~~~~~~~~~~~~~
gcc/testsuite/gcc.c-torture/execute/pr92618.c:45:4: warning: dereferencing
type-punned pointer will break strict-aliasing rules [-Wstrict-aliasing]
   45 |   *(__m128i *) &a[2] = d;
      |    ^~~~~~~~~~~~~~~~~

but it doesn't set -fno-strict-aliasing. It looks like the test should set
-fno-strict-aliasing.

[Bug c/99955] gcc.c-torture/execute/pr92618.c violates strict aliasing rules

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=99955

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
           Keywords|                            |diagnostic, wrong-code
     Ever confirmed|0                           |1
                 CC|                            |jsm28 at gcc dot gnu.org
          Component|testsuite                   |c
   Last reconfirmed|                            |2021-04-07

--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
It is

typedef long long __m128i __attribute__((__may_alias__, __vector_size__(2 *
sizeof (long long))));

so should be OK unless the FE somehow magically screws up.  And it looks
like it does!?  The may-alias attribute isn't added to the built vector type
(but vector type building is a bit iffy).

It works when doing

typedef long long __m128i __attribute__((__vector_size__(2 * sizeof (long
long)), __may_alias__));

instead.  Maybe we want to diagnose when the vector_size attribute isn't the
first attribute?  Maybe we want to handle all this gracefully anyway?

Joseph?

Meanwhile Alex, it's pre-approved to adjust the testcase to make __may_alias__
come last ...

[Bug c/99955] gcc.c-torture/execute/pr92618.c violates strict aliasing rules

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=99955

--- Comment #2 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:0d6ad10c170e3670f0b5e3709e0fa6e76b7065b3

commit r11-8026-g0d6ad10c170e3670f0b5e3709e0fa6e76b7065b3
Author: Richard Biener <rguenther@suse.de>
Date:   Wed Apr 7 13:21:23 2021 +0200

    testsuite/99955 - fix may_alias declaration of vector

    This fixes the order of the type attributes to preserve may_alias
    for the vector type.

    2021-04-07  Richard Biener  <rguenther@suse.de>

            PR testsuite/99955
            * gcc.c-torture/execute/pr92618.c: Move may_alias attributes
            last.

[Bug c/99955] gcc.c-torture/execute/pr92618.c violates strict aliasing rules

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=99955

--- Comment #3 from Richard Biener <rguenth at gcc dot gnu.org> ---
I fixed the testcase - we should still diagnose the type declaration or make it
DTRT.

[Bug c/99955] gcc.c-torture/execute/pr92618.c violates strict aliasing rules

[joseph at codesourcery dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=99955

--- Comment #4 from joseph at codesourcery dot com <joseph at codesourcery dot com> ---
I think may_alias ought to work together with vector_size, whatever its 
position in the attribute list, i.e. there is a front-end bug here for 
which making that combination of attributes work properly together in any 
order is the right fix.

[Bug c/99955] gcc.c-torture/execute/pr92618.c violates strict aliasing rules

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=99955

--- Comment #5 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
handle_vector_size_attribute does:

  /* Build back pointers if needed.  */
  tree new_type = build_vector_type (type, nunits);

  /* Build back pointers if needed.  */
  *node = lang_hooks.types.reconstruct_complex_type (*node, new_type);


But the default reconstruct_complex_type never reapplies the attributes back on
it. (I Have not looked into the c++ front-end's implementation to about the
same).