[Bug target/55184] New: Invalid codegen with vectors and casts

[Bug target/55184] New: Invalid codegen with vectors and casts

[mathias at gaunard dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55184

             Bug #: 55184
           Summary: Invalid codegen with vectors and casts
    Classification: Unclassified
           Product: gcc
           Version: 4.6.3
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: target
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: mathias@gaunard.com


Created attachment 28600
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=28600
Testcase

On x86-64, when 128-bit vectors of long or unsigned long are aliased, an
invalid optimization seems to occur and causes only the first element of the
vector to be modified.

Output is correct when compiled without optimization, but fails with -O1 or
greater. Usage of -fno-strict-aliasing doesn't change anything.

The problem doesn't seem to appear in the 4.4, 4.5 and 4.7 series.
Nevertheless I'd like to know all the cases it affects and whether there is a
workaround.

[Bug target/55184] [4.6 Regression] Invalid codegen with vectors and casts

[pinskia at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55184

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
      Known to work|                            |4.4.0, 4.5.0, 4.7.0
   Target Milestone|---                         |4.6.4
            Summary|Invalid codegen with        |[4.6 Regression] Invalid
                   |vectors and casts           |codegen with vectors and
                   |                            |casts
      Known to fail|                            |4.6.3

[Bug target/55184] [4.6 Regression] Invalid codegen with vectors and casts

[mathias at gaunard dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55184

--- Comment #1 from Mathias Gaunard <mathias at gaunard dot com> 2012-11-03 00:29:47 UTC ---
Problem seems to occur with i686 as well but only if SSE2 is enabled (you'll
need to replace 'long' by 'long long' for this in the testcase)

[Bug target/55184] [4.6 Regression] Invalid codegen with vectors and casts

[mikpe at it dot uu.se]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55184

--- Comment #2 from Mikael Pettersson <mikpe at it dot uu.se> 2012-11-04 14:13:58 UTC ---
I can't reproduce the error with vanilla gcc-4.6.3 on x86_64-linux.

[Bug target/55184] [4.6 Regression] Invalid codegen with vectors and casts

[mathias at gaunard dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55184

Mathias Gaunard <mathias at gaunard dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
  Attachment #28600|0                           |1
        is obsolete|                            |

--- Comment #3 from Mathias Gaunard <mathias at gaunard dot com> 2012-11-04 17:58:30 UTC ---
Created attachment 28614
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=28614
Fixed testcase

[Bug target/55184] [4.6 Regression] Invalid codegen with vectors and casts

[mathias at gaunard dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55184

--- Comment #4 from Mathias Gaunard <mathias at gaunard dot com> 2012-11-04 18:01:27 UTC ---
Sorry, I edited the file in between and ended up uploading the wrong test case.

Below is the result on my machine with the fixed testcase.

$ gcc --version
gcc (Ubuntu/Linaro 4.6.3-1ubuntu5) 4.6.3
Copyright (C) 2011 Free Software Foundation, Inc.
This is free software; see the source for copying conditions.  There is NO
warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.

$ gcc -dumpmachine
x86_64-linux-gnu
$ gcc test.c -o test && ./test
1 (should be 1)
0 (should be 0)
0 (should be 0)
0 (should be 0)
0 (should be 0)
0 (should be 0)
0 (should be 0)
0 (should be 0)
1 (should be 1)
0 (should be 0)
0 (should be 0)
0 (should be 0)
0 (should be 0)
0 (should be 0)
0 (should be 0)
0 (should be 0)
$ gcc -O1 test.c -o test && ./test
1 (should be 1)
0 (should be 0)
0 (should be 0)
0 (should be 0)
0 (should be 0)
0 (should be 0)
0 (should be 0)
0 (should be 0)
0 (should be 1)
0 (should be 0)
0 (should be 0)
0 (should be 0)
0 (should be 0)
0 (should be 0)
0 (should be 0)
0 (should be 0)

[Bug target/55184] [4.6 Regression] Invalid codegen with vectors and casts

[mikpe at it dot uu.se]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55184

--- Comment #5 from Mikael Pettersson <mikpe at it dot uu.se> 2012-11-04 21:24:44 UTC ---
I can confirm the bug with gcc-4.6.3 and the fixed test case.  However, the bug
has since been fixed on 4.6 branch in r187763, the fix for PR52407.  The test
cases for these bugs are very similar, so I'd consider this bug a duplicate of
PR52407.

[Bug target/55184] [4.6 Regression] Invalid codegen with vectors and casts

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55184

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |RESOLVED
         Resolution|                            |FIXED
   Target Milestone|4.6.4                       |4.7.0

--- Comment #6 from Jakub Jelinek <jakub at gcc dot gnu.org> 2013-04-12 16:17:37 UTC ---
The 4.6 branch has been closed, fixed in GCC 4.7.0.