[Bug target/104401] New: [x86] Failure to recognize min/max pattern using pcmp+pblendv

[Bug target/104401] New: [x86] Failure to recognize min/max pattern using pcmp+pblendv

[gabravier at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=104401

            Bug ID: 104401
           Summary: [x86] Failure to recognize min/max pattern using
                    pcmp+pblendv
           Product: gcc
           Version: 12.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: target
          Assignee: unassigned at gcc dot gnu.org
          Reporter: gabravier at gmail dot com
  Target Milestone: ---

#include <smmintrin.h>

__m128i min32(__m128i value, __m128i input)
{
  return _mm_blendv_epi8(input, value, _mm_cmplt_epi32(value, input));
}

With -O3 -msse4.1, GCC outputs this:

min32(long long __vector(2), long long __vector(2)):
  movdqa xmm2, xmm0
  movdqa xmm0, xmm1
  movdqa xmm3, xmm1
  pcmpgtd xmm0, xmm2
  pblendvb xmm3, xmm2, xmm0
  movdqa xmm0, xmm3
  ret

LLVM instead outputs this:

min32(long long __vector(2), long long __vector(2)):
  pminsd xmm0, xmm1
  ret

The equivalent code with cmpgt used instead of cmplt can be optimized to
pmaxsd.

[Bug target/104401] [x86] Failure to recognize min/max pattern using pcmp+pblendv

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=104401

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Severity|normal                      |enhancement

[Bug target/104401] [x86] Failure to recognize min/max pattern using pcmp+pblendv

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=104401

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2022-02-05
     Ever confirmed|0                           |1

--- Comment #1 from Andrew Pinski <pinskia at gcc dot gnu.org> ---

(set (reg:V16QI 93)
    (unspec:V16QI [
            (subreg:V16QI (reg/v:V2DI 92 [ input ]) 0)
            (subreg:V16QI (reg/v:V2DI 91 [ value ]) 0)
            (subreg:V16QI (gt:V4SI (subreg:V4SI (reg/v:V2DI 92 [ input ]) 0)
                    (subreg:V4SI (reg/v:V2DI 91 [ value ]) 0)) 0)
        ] UNSPEC_BLENDV))

blend uses unspec which makes this a little harder

[Bug target/104401] [x86] Failure to recognize min/max pattern using pcmp+pblendv

[crazylht at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=104401

--- Comment #2 from Hongtao.liu <crazylht at gmail dot com> ---
(In reply to Andrew Pinski from comment #1)
> (set (reg:V16QI 93)
>     (unspec:V16QI [
>             (subreg:V16QI (reg/v:V2DI 92 [ input ]) 0)
>             (subreg:V16QI (reg/v:V2DI 91 [ value ]) 0)
>             (subreg:V16QI (gt:V4SI (subreg:V4SI (reg/v:V2DI 92 [ input ]) 0)
>                     (subreg:V4SI (reg/v:V2DI 91 [ value ]) 0)) 0)
>         ] UNSPEC_BLENDV))
> 
> blend uses unspec which makes this a little harder

I think it could be done in gimple level.

 18  _10 = VIEW_CONVERT_EXPR<vector(4) int>(value_2(D));
 19  _11 = VIEW_CONVERT_EXPR<vector(4) int>(input_3(D));
 20  _12 = _10 < _11;
 21  _13 = VIEW_CONVERT_EXPR<vector(4) int>(_12);
 22  _5 = VIEW_CONVERT_EXPR<vector(16) char>(_13);
 23  _4 = VIEW_CONVERT_EXPR<vector(16) char>(value_2(D));
 24  _6 = VIEW_CONVERT_EXPR<vector(16) char>(input_3(D));
 25  _8 = .VCOND (_5, { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 }, _4,
_6, 108);
 26  _9 = VIEW_CONVERT_EXPR<__m128i>(_8);

[Bug target/104401] [x86] Failure to recognize min/max pattern using pcmp+pblendv

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=104401

--- Comment #3 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
On a slightly different subject, I wish VCOND would print out the tree code
symbolically rather than just a number. It would be easier to understand what
the condition exec is really.

[Bug target/104401] [x86] Failure to recognize min/max pattern using pcmp+pblendv

[crazylht at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=104401

--- Comment #4 from Hongtao.liu <crazylht at gmail dot com> ---
(In reply to Andrew Pinski from comment #3)
> On a slightly different subject, I wish VCOND would print out the tree code
> symbolically rather than just a number. It would be easier to understand
> what the condition exec is really.

Yes, not quite straightforward.

Here it means
 _8 = _5 < { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 } ? _4 : _6;
(blendv uses sign bit as mask bit).

[Bug target/104401] [x86] Failure to recognize min/max pattern using pcmp+pblendv

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=104401

--- Comment #5 from Richard Biener <rguenth at gcc dot gnu.org> ---
(In reply to Andrew Pinski from comment #3)
> On a slightly different subject, I wish VCOND would print out the tree code
> symbolically rather than just a number. It would be easier to understand
> what the condition exec is really.

Feel free to add a // comment after it ;)

Btw, forwprop1 sees

  <bb 2> :
  _12 = VIEW_CONVERT_EXPR<vector(4) int>(value_3(D));
  _13 = VIEW_CONVERT_EXPR<vector(4) int>(input_4(D));
  _14 = _12 < _13;
  _15 = VEC_COND_EXPR <_14, { -1, -1, -1, -1 }, { 0, 0, 0, 0 }>;
  _16 = VIEW_CONVERT_EXPR<__m128i>(_15);
  _6 = VIEW_CONVERT_EXPR<vector(16) char>(_15);
  _5 = VIEW_CONVERT_EXPR<vector(16) char>(value_3(D));
  _8 = VIEW_CONVERT_EXPR<vector(16) char>(input_4(D));
  _9 = _6 < { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 };
  _10 = VEC_COND_EXPR <_9, _5, _8>;
  _11 = VIEW_CONVERT_EXPR<__m128i>(_10);
  return _11;

the difficulty #1 is the v16qi mode blend vs. the v4si mode compare so
we'd have to pattern match the blend to be done on v4si based on the
_15 value range.

[Bug target/104401] [x86] Failure to recognize min/max pattern using pcmp+pblendv

[liuhongt at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=104401

Hongtao Liu <liuhongt at gcc dot gnu.org> changed:

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

--- Comment #6 from Hongtao Liu <liuhongt at gcc dot gnu.org> ---
On latest trunk, we have

19  <bb 2> [local count: 1073741824]:
20  _6 = VIEW_CONVERT_EXPR<vector(4) int>(value_2(D));
21  _7 = VIEW_CONVERT_EXPR<vector(4) int>(input_3(D));
22  _8 = _6 < _7;
23  _4 = VEC_COND_EXPR <_8, _6, _7>;
24  _5 = VIEW_CONVERT_EXPR<vector(2) long long int>(_4);
25  return _5;
26
27}

And we can match a < b ? a : b for vector modes.

[Bug target/104401] [x86] Failure to recognize min/max pattern using pcmp+pblendv

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=104401

--- Comment #7 from GCC Commits <cvs-commit at gcc dot gnu.org> ---
The master branch has been updated by hongtao Liu <liuhongt@gcc.gnu.org>:

https://gcc.gnu.org/g:6686e16fda419067b91614835dd743ebb82717ea

commit r14-7124-g6686e16fda419067b91614835dd743ebb82717ea
Author: liuhongt <hongtao.liu@intel.com>
Date:   Tue Dec 19 10:39:06 2023 +0800

    Optimize A < B ? A : B to MIN_EXPR.

    Similar for A < B ? B : A to MAX_EXPR.
    There're codes in the frontend to optimize such pattern but failed to
    handle testcase in the PR since it's exposed at gimple level when
    folding backend builtins.

    pr95906 now can be optimized to MAX_EXPR as it's commented in the
    testcase.

    // FIXME: this should further optimize to a MAX_EXPR
     typedef signed char v16i8 __attribute__((vector_size(16)));
     v16i8 f(v16i8 a, v16i8 b)

    gcc/ChangeLog:

            PR target/104401
            * match.pd (VEC_COND_EXPR: A < B ? A : B -> MIN_EXPR): New patten
match.

    gcc/testsuite/ChangeLog:

            * gcc.target/i386/pr104401.c: New test.
            * gcc.dg/tree-ssa/pr95906.c: Adjust testcase.

[Bug target/104401] [x86] Failure to recognize min/max pattern using pcmp+pblendv

[liuhongt at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=104401

Hongtao Liu <liuhongt at gcc dot gnu.org> changed:

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

--- Comment #8 from Hongtao Liu <liuhongt at gcc dot gnu.org> ---
Fixed in GCC14.

[Bug target/104401] [x86] Failure to recognize min/max pattern using pcmp+pblendv

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=104401

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Target Milestone|---                         |14.0