[Bug target/94467] New: [10 Regression] wrong code with -mavx and ssse3 builtins

[Bug target/94467] New: [10 Regression] wrong code with -mavx and ssse3 builtins

[zsojka at seznam dot cz]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=94467

            Bug ID: 94467
           Summary: [10 Regression] wrong code with -mavx and ssse3
                    builtins
           Product: gcc
           Version: 10.0
            Status: UNCONFIRMED
          Keywords: wrong-code
          Severity: normal
          Priority: P3
         Component: target
          Assignee: unassigned at gcc dot gnu.org
          Reporter: zsojka at seznam dot cz
  Target Milestone: ---
              Host: x86_64-pc-linux-gnu

Created attachment 48180
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=48180&action=edit
reduced testcase

Output:
$ x86_64-pc-linux-gnu-gcc -O -mssse3 testcase.c && ./a.out
$ x86_64-pc-linux-gnu-gcc -O -mavx testcase.c && ./a.out 
Aborted

__builtin_ia32_pavgb() doesn't seem to be generated with -mavx.

$ 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-r10-7481-20200331151333-gf14b41d2712-checking-yes-rtl-df-extra-amd64/bin/../libexec/gcc/x86_64-pc-linux-gnu/10.0.1/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
--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 --disable-libstdcxx-pch
--prefix=/repo/gcc-trunk//binary-trunk-r10-7481-20200331151333-gf14b41d2712-checking-yes-rtl-df-extra-amd64
Thread model: posix
Supported LTO compression algorithms: zlib zstd
gcc version 10.0.1 20200331 (experimental) (GCC)

[Bug target/94467] [10 Regression] wrong code with -mavx and ssse3 builtins

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=94467

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
               Host|x86_64-pc-linux-gnu         |
     Ever confirmed|0                           |1
   Last reconfirmed|                            |2020-04-03
   Target Milestone|---                         |10.0
             Target|                            |x86_64-*-* i?86-*-*
           Keywords|                            |needs-bisection
             Status|UNCONFIRMED                 |NEW

--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
Confirmed.  Some mmx-with-sse fallout.

[Bug target/94467] [10 Regression] wrong code with -mavx and ssse3 builtins

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=94467

--- Comment #2 from Richard Biener <rguenth at gcc dot gnu.org> ---
Note using __builtin_ia32_* directly is discouraged.  A testcase using official
intrinsics would be better.

[Bug target/94467] [10 Regression] wrong code with -mavx and ssse3 builtins

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=94467

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |jakub at gcc dot gnu.org
           Keywords|needs-bisection             |

--- Comment #3 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Started with r10-396-g9c1d1db27d421ab7c5ac1ca9b23cae002f13380e

[Bug target/94467] [10 Regression] wrong code with -mavx and ssse3 builtins

[zsojka at seznam dot cz]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=94467

--- Comment #4 from Zdenek Sojka <zsojka at seznam dot cz> ---
(In reply to Richard Biener from comment #2)
> Note using __builtin_ia32_* directly is discouraged.  A testcase using
> official intrinsics would be better.

Thanks; I just took the list from
https://gcc.gnu.org/onlinedocs/gcc/x86-Built-in-Functions.html (it has some
mistakes though; I hope to list & report them 'when I have time for that'). I
assume using the intrinsics will be a bit complicated, because creduce will
probably expand/inline them by default to the __builtin_ia32_*() call.

[Bug target/94467] [10 Regression] wrong code with -mavx and ssse3 builtins

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=94467

--- Comment #5 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Either one has to replace the builtins manually afterwards back with
corresponding intrinsics (something I've done e.g. in the PR94460 patch), or
one could e.g. creduce on mostly preprocessed source where content of
preprocessed #include <x86intrin.h> would be replaced back with the include.

[Bug target/94467] [10 Regression] wrong code with -mavx and ssse3 builtins

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=94467

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |ASSIGNED
           Priority|P3                          |P1

[Bug target/94467] [10 Regression] wrong code with -mavx and ssse3 builtins

[hjl.tools at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=94467

--- Comment #6 from H.J. Lu <hjl.tools at gmail dot com> ---
The bug was introduced by r10-393:

commit 16ed2601ad0a4aa82f11e9df86ea92183f94f979
Author: H.J. Lu <hongjiu.lu@intel.com>
Date:   Wed May 15 15:26:19 2019 +0000

    i386: Emulate MMX pshufb with SSE version

    Emulate MMX version of pshufb with SSE version by masking out the bit 3
    of the shuffle control byte.  Only SSE register source operand is allowed.

            PR target/89021
            * config/i386/sse.md (ssse3_pshufbv8qi3): Changed to
            define_insn_and_split.  Also allow TARGET_MMX_WITH_SSE.  Add
            SSE emulation.

+(define_insn_and_split "ssse3_pshufbv8qi3"
+  [(set (match_operand:V8QI 0 "register_operand" "=y,x,Yv")
+  (unspec:V8QI [(match_operand:V8QI 1 "register_operand" "0,0,Yv")
+           (match_operand:V8QI 2 "register_mmxmem_operand" "ym,x,Yv")]
+          UNSPEC_PSHUFB))
+   (clobber (match_scratch:V4SI 3 "=X,x,Yv"))]  <<< This is an earlyclobber
operand.
+  "(TARGET_MMX || TARGET_MMX_WITH_SSE) && TARGET_SSSE3"
+  "@
+   pshufb\t{%2, %0|%0, %2}
+   #
+   #"
+  "TARGET_MMX_WITH_SSE && reload_completed"
+  [(set (match_dup 3) (match_dup 5))
+   (set (match_dup 3)
+  (and:V4SI (match_dup 3) (match_dup 2)))
+   (set (match_dup 0)
+  (unspec:V16QI [(match_dup 1) (match_dup 4)] UNSPEC_PSHUFB))]

[Bug target/94467] [10 Regression] wrong code with -mavx and ssse3 builtins

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=94467

--- Comment #7 from CVS Commits <cvs-commit at gcc dot gnu.org> ---
The master branch has been updated by H.J. Lu <hjl@gcc.gnu.org>:

https://gcc.gnu.org/g:bbcdf9bb3fd04adc59f41e4e1ff6293c84cbecc4

commit r10-7545-gbbcdf9bb3fd04adc59f41e4e1ff6293c84cbecc4
Author: H.J. Lu <hjl.tools@gmail.com>
Date:   Fri Apr 3 11:49:10 2020 -0700

    x86: Mark scratch operand in ssse3_pshufbv8qi3 as earlyclobber

    commit 16ed2601ad0a4aa82f11e9df86ea92183f94f979
    Author: H.J. Lu <hongjiu.lu@intel.com>
    Date:   Wed May 15 15:26:19 2019 +0000

        i386: Emulate MMX pshufb with SSE version

    has

    +(define_insn_and_split "ssse3_pshufbv8qi3"
    +  [(set (match_operand:V8QI 0 "register_operand" "=y,x,Yv")
    +  (unspec:V8QI [(match_operand:V8QI 1 "register_operand" "0,0,Yv")
    +           (match_operand:V8QI 2 "register_mmxmem_operand" "ym,x,Yv")]
    +          UNSPEC_PSHUFB))
    +   (clobber (match_scratch:V4SI 3 "=X,x,Yv"))]
                                           ^^^  There are earlyclobber.
    +  "(TARGET_MMX || TARGET_MMX_WITH_SSE) && TARGET_SSSE3"
    +  "@
    +   pshufb\t{%2, %0|%0, %2}
    +   #
    +   #"
    +  "TARGET_MMX_WITH_SSE && reload_completed"
    +  [(set (match_dup 3) (match_dup 5))
    +   (set (match_dup 3)
    +  (and:V4SI (match_dup 3) (match_dup 2)))
    +   (set (match_dup 0)
    +  (unspec:V16QI [(match_dup 1) (match_dup 4)] UNSPEC_PSHUFB))]

    If input register operand 2 is dead after this insn, RA may choose it
    as scratch operand.  Since it isn't marked as earlyclobber, operand 2
    becomes unused after split and then it gets optimized out.  Mark scratch
    operand as earlyclobber fixes the issue.

    gcc/

            PR target/94467
            * config/i386/sse.md (ssse3_pshufbv8qi3): Mark scratch operand
            as earlyclobber.

    gcc/testsuite/

            PR target/94467
            * gcc.target/i386/pr94467-1.c: New test.
            * gcc.target/i386/pr94467-2.c: Likewise.

[Bug target/94467] [10 Regression] wrong code with -mavx and ssse3 builtins

[hjl.tools at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=94467

H.J. Lu <hjl.tools at gmail dot com> changed:

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

--- Comment #8 from H.J. Lu <hjl.tools at gmail dot com> ---
Fixed for GCC 10.