[Bug rtl-optimization/116037] New: [15 Regression] wrong code at -O2 with vector masking and add

[Bug rtl-optimization/116037] New: [15 Regression] wrong code at -O2 with vector masking and add

[zsojka at seznam dot cz]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116037

            Bug ID: 116037
           Summary: [15 Regression] wrong code at -O2 with vector masking
                    and add
           Product: gcc
           Version: 15.0
            Status: UNCONFIRMED
          Keywords: wrong-code
          Severity: normal
          Priority: P3
         Component: rtl-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: zsojka at seznam dot cz
  Target Milestone: ---
              Host: x86_64-pc-linux-gnu
            Target: x86_64-pc-linux-gnu

Created attachment 58722
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=58722&action=edit
reduced testcase

Compiler output:
$ x86_64-pc-linux-gnu-gcc -O2 testcase.c -Wno-psabi
$ ./a.out 
Aborted

The values are missing the bit 32 set.

$ x86_64-pc-linux-gnu-gcc -v
Using built-in specs.
COLLECT_GCC=/repo/gcc-trunk/binary-latest-amd64/bin/x86_64-pc-linux-gnu-gcc
COLLECT_LTO_WRAPPER=/repo/gcc-trunk/binary-trunk-r15-2206-20240722194717-g6f81b7fa799-checking-yes-rtl-df-extra-nobootstrap-amd64/bin/../libexec/gcc/x86_64-pc-linux-gnu/15.0.0/lto-wrapper
Target: x86_64-pc-linux-gnu
Configured with: /repo/gcc-trunk//configure --enable-languages=c,c++
--enable-valgrind-annotations --disable-nls --enable-checking=yes,rtl,df,extra
--disable-bootstrap --with-cloog --with-ppl --with-isl
--build=x86_64-pc-linux-gnu --host=x86_64-pc-linux-gnu
--target=x86_64-pc-linux-gnu --with-ld=/usr/bin/x86_64-pc-linux-gnu-ld
--with-as=/usr/bin/x86_64-pc-linux-gnu-as --enable-libsanitizer
--disable-libstdcxx-pch
--prefix=/repo/gcc-trunk//binary-trunk-r15-2206-20240722194717-g6f81b7fa799-checking-yes-rtl-df-extra-nobootstrap-amd64
Thread model: posix
Supported LTO compression algorithms: zlib zstd
gcc version 15.0.0 20240722 (experimental) (GCC)

[Bug rtl-optimization/116037] [15 Regression] wrong code at -O2 with vector masking and add

[sjames at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116037

Sam James <sjames at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |law at gcc dot gnu.org,
                   |                            |sjames at gcc dot gnu.org

--- Comment #1 from Sam James <sjames at gcc dot gnu.org> ---
It's -fext-dce.

[Bug rtl-optimization/116037] [15 Regression] wrong code at -O2 with vector masking and add

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116037

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
     Ever confirmed|0                           |1
   Last reconfirmed|                            |2024-07-23
             Status|UNCONFIRMED                 |ASSIGNED

--- Comment #2 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
.

[Bug rtl-optimization/116037] [15 Regression] wrong code at -O2 with vector masking and add

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116037

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Target Milestone|---                         |15.0

[Bug rtl-optimization/116037] [15 Regression] wrong code at -O2 with vector masking and add

[law at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116037

--- Comment #3 from Jeffrey A. Law <law at gcc dot gnu.org> ---
So this is fixed by a patch I'm still working on.  Essentially I want to stop
relying on an empty LIVE_TMP to denote that we skipped a destination's set. 
That's not quite ready yet, but I think it's damn close.

[Bug rtl-optimization/116037] [15 Regression] wrong code at -O2 with vector masking and add

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116037

--- Comment #4 from GCC Commits <cvs-commit at gcc dot gnu.org> ---
The master branch has been updated by Jeff Law <law@gcc.gnu.org>:

https://gcc.gnu.org/g:679086172b84be18c55fdbb9cda7e97806e7c083

commit r15-2275-g679086172b84be18c55fdbb9cda7e97806e7c083
Author: Jeff Law <jlaw@ventanamicro.com>
Date:   Wed Jul 24 11:16:26 2024 -0600

    [rtl-optimization/116037] Explicitly track if a destination was skipped in
ext-dce

    So this has been in the hopper since the first bugs were reported against
    ext-dce.  It'd been holding off committing as I was finding other issues in
    terms of correctness of live computations.  There's still problems in that
    space, but I think it's time to push this chunk forward.  I'm marking it as
    116037, but it may impact other bugs.

    This patch starts explicitly tracking if set processing skipped a
destination,
    which can happen for wide modes (TI+), vectors, certain subregs, etc.  This
is
    computed during ext_dce_set_processing.

    During use processing we use that flag to determine reliably if we need to
make
    the inputs fully live and to avoid even trying to eliminate an extension if
we
    skipped output processing.

    While testing this I found that a recent change to fix cases where we had
two
    subreg input operands mucked up the code to make things like a shift/rotate
    count fully live.  So that goof has been fixed.

    Bootstrapped and regression tested on x86.  Most, but not all, of these
changes
    have also been tested on the crosses.  Pushing to the trunk.

    I'm not including it in this patch but I'm poking at converting this code
to
    use note_uses/note_stores to make it more maintainable.  The SUBREG and
    STRICT_LOW_PART handling of note_stores is problematical, but I think it's
    solvable.  I haven't tried a conversion to note_uses yet.

            PR rtl-optimization/116037
    gcc/
            * ext-dce.cc (ext_dce_process_sets): Note if we ever skip a dest
            and return that info explicitly.
            (ext_dce_process_uses): If a set was skipped, then consider all
bits
            in every input as live.  Do not try to optimize away an extension
if
            we skipped processing a destination in the same insn.  Restore code
            to make shift/rotate count fully live.
            (ext_dce_process_bb): Handle API changes for ext_dce_process_sets.

    gcc/testsuite/
            * gcc.dg/torture/pr116037.c: New test

[Bug rtl-optimization/116037] [15 Regression] wrong code at -O2 with vector masking and add

[law at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116037

Jeffrey A. Law <law at gcc dot gnu.org> changed:

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

--- Comment #5 from Jeffrey A. Law <law at gcc dot gnu.org> ---
Fixed on the trunk.

[Bug rtl-optimization/116037] [15 Regression] wrong code at -O2 with vector masking and add

[law at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116037

--- Comment #6 from Jeffrey A. Law <law at gcc dot gnu.org> ---
*** Bug 116067 has been marked as a duplicate of this bug. ***

[Bug rtl-optimization/116037] [15 Regression] wrong code at -O2 with vector masking and add

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116037

--- Comment #7 from GCC 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:df4062c54b0b3c5f5c6a1f1a2454abf965100e3a

commit r15-2753-gdf4062c54b0b3c5f5c6a1f1a2454abf965100e3a
Author: Jakub Jelinek <jakub@redhat.com>
Date:   Tue Aug 6 11:52:35 2024 +0200

    testsuite: Fix up pr116037.c test [PR116245]

    The test FAILs on big endian targets, because VV is a vector of unsigned
__int128
    and VC vector of unsigned char and so ((VC) vv)[0] is 0x01 on little endian
    but 0xff on big endian and PDP endian.
    As I believe it is intentional to test it as it is written on little
endian,
    the following patch just adds another case for big endian and for other
    endians instead of figuring out what exactly to fetch it fetches the whole
    unsigned __int128 and casts it to unsigned char.  Not that pdp11 has
    __int128 support...

    2024-08-06  Jakub Jelinek  <jakub@redhat.com>

            PR rtl-optimization/116037
            PR testsuite/116245
            * gcc.dg/torture/pr116037.c (foo): Fix up for big end middle
endian.