[Bug testsuite/54622] New: gcc.dg/vect test failures for arm big-endian

[Bug testsuite/54622] New: gcc.dg/vect test failures for arm big-endian

[janis at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54622

             Bug #: 54622
           Summary: gcc.dg/vect test failures for arm big-endian
    Classification: Unclassified
           Product: gcc
           Version: 4.8.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: testsuite
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: janis@gcc.gnu.org
                CC: irar@gcc.gnu.org
            Target: arm


Created attachment 28219
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=28219
list of vect test failures for arm big-endian

71 vectorization tests (plus corresponding checks with -flto) pass for ARM
little endian but fail for ARM big endian with current mainline. I'll attach
the list.

Some of them are scans that are gated on effective targets that always fail for
ARM big endian; perhaps that's a coincidence, or perhaps some of these
effective-target checks should require arm_little_endian:

    vect64 (2)
    vect-widen_shift (8)
    vect_perm_byte (1)
    vect_perm_short (1)
    vect_strided3 (1)

Testing for ARM big-endian is straightforward for any ARM build. Avoid
execution tests with a local patch to gcc/testsuite/gcc.dg/vect/vect.exp to add
'set dg-do-what-default "compile"' after calling
check_vect_support_and_set_flags, and then pass the appropriate flags in
RUNTESTFLAGS.

[Bug testsuite/54622] gcc.dg/vect test failures for arm big-endian

[pinskia at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54622

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

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

--- Comment #1 from Andrew Pinski <pinskia at gcc dot gnu.org> 2012-09-19 01:08:58 UTC ---
I think all the tests that pass for little-endian should also pass for
big-endian.  The only difference comes down to how the memory is located from
memory as IIRC neon stays the same in memory.

[Bug testsuite/54622] gcc.dg/vect test failures for arm big-endian

[janis at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54622

--- Comment #2 from Janis Johnson <janis at gcc dot gnu.org> 2012-09-19 01:16:58 UTC ---
Effective targets vect_unpack and vect_pack_trunc both require little endian
for ARM, which is why I thought that perhaps some of the others do as well.

[Bug testsuite/54622] gcc.dg/vect test failures for arm big-endian

[pinskia at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54622

--- Comment #3 from Andrew Pinski <pinskia at gcc dot gnu.org> 2012-09-19 01:29:09 UTC ---
(In reply to comment #2)
> Effective targets vect_unpack and vect_pack_trunc both require little endian
> for ARM, which is why I thought that perhaps some of the others do as well.

Yes but I think it is wrong to do that in the back-end.

[Bug testsuite/54622] gcc.dg/vect test failures for arm big-endian

[rearnsha at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54622

--- Comment #4 from Richard Earnshaw <rearnsha at gcc dot gnu.org> 2012-09-19 22:26:49 UTC ---
The reasons for the vector problems in big-endian largely fall into two areas:

1) Neon vector elements are numbered from the LSB of the vector register in
both big and little-endian modes.  GCC assumes that vector elements in
big-endian are numbered from the MSB.  This means compensation code is needed
when elements are addressed directly (for example shuffle operations).

2) Quad-words in 256-bit vectors are not pure big-endian, the DWords are
swapped in order on loads.  GCC can't currently cope with this in any sensible
manner.

[Bug testsuite/54622] gcc.dg/vect test failures for arm big-endian

[jules at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54622

jules at gcc dot gnu.org changed:

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

--- Comment #5 from jules at gcc dot gnu.org 2012-09-26 16:25:08 UTC ---
Unfortunately (due to the reasons Richard outlined) lots of things don't work
wrt. vectorization for Neon in big-endian mode, so are deliberately disabled in
the backend. IMO, the right thing to do (at least until that is fixed) is to
make the effective-target checks for vectorization features depend on
little-endian ARM mode.

[Bug testsuite/54622] gcc.dg/vect test failures for arm big-endian

[janis at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54622

--- Comment #6 from Janis Johnson <janis at gcc dot gnu.org> 2012-09-26 16:45:25 UTC ---
It's important to continue to run the vectorization tests with ARM big-endian
to detect regressions in the execution tests, so I don't want to exclude it
from all vect effective targets.  I'll propose a set of effective targets used
in the scans that should exclude ARM big-endian.

[Bug testsuite/54622] gcc.dg/vect test failures for arm big-endian

[janis at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54622

--- Comment #7 from Janis Johnson <janis at gcc dot gnu.org> 2012-09-26 21:51:53 UTC ---
I've tried requiring arm_little_endian for various vect_ effective targets, but
the missing support isn't at all clear-cut.  Lots of vectorization takes place
so it's important to continue to pay attention to the execute tests, but for
now I think we'll just handle big-endian expected failures locally.

[Bug testsuite/54622] gcc.dg/vect test failures for arm big-endian

[janis at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54622

--- Comment #8 from Janis Johnson <janis at gcc dot gnu.org> 2013-01-16 18:50:06 UTC ---
Author: janis
Date: Wed Jan 16 18:49:57 2013
New Revision: 195249

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=195249
Log:
    PR testsuite/54622
    * lib/target-supports.exp (check_effective_target_vect_perm_byte,
    check_effective_target_vect_perm_short,
    check_effective_target_vect_widen_mult_qi_to_hi_pattern,
    check_effective_target_vect64): Return 0 for big-endian ARM.
    (check_effective_target_vect_widen_sum_qi_to_hi): Return 1 for ARM.

Modified:
    trunk/gcc/testsuite/ChangeLog
    trunk/gcc/testsuite/lib/target-supports.exp

[Bug testsuite/54622] gcc.dg/vect test failures for arm big-endian

[pinskia at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54622

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2013-03-20
     Ever Confirmed|0                           |1

--- Comment #9 from Andrew Pinski <pinskia at gcc dot gnu.org> 2013-03-20 00:45:42 UTC ---
Confirmed.