[Bug target/113079] New: [x86] Fails to generate dot_prod instructions for 64-bit vector.

[Bug target/113079] New: [x86] Fails to generate dot_prod instructions for 64-bit vector.

[liuhongt at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=113079

            Bug ID: 113079
           Summary: [x86] Fails to generate dot_prod instructions for
                    64-bit vector.
           Product: gcc
           Version: 14.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: target
          Assignee: unassigned at gcc dot gnu.org
          Reporter: liuhongt at gcc dot gnu.org
  Target Milestone: ---

int
foo (int n, unsigned char* p, char* pi)
{
    int sum = 0;
    for (int i = 0; i != 8; i++)
    {
        sum += p[i] * pi[i];
    }
    return sum;
}

We can use 128-bit dot_prod instruction + clean upper 64 bits. Currently, gcc
generates a long instruction sequence.

        vmovq   xmm0, QWORD PTR [rsi]
        vmovq   xmm2, QWORD PTR [rdx]
        vpmovzxbw       xmm1, xmm0
        vpsrlq  xmm0, xmm0, 32
        vpmovsxbw       xmm3, xmm2
        vpmullw xmm1, xmm1, xmm3
        vpsrlq  xmm2, xmm2, 32
        vpmovzxbw       xmm0, xmm0
        vpmovsxbw       xmm2, xmm2
        vpmullw xmm0, xmm0, xmm2
        vpmovsxwd       xmm2, xmm1
        vpsrlq  xmm1, xmm1, 32
        vpmovsxwd       xmm1, xmm1
        vpaddd  xmm2, xmm2, xmm1
        vpmovsxwd       xmm1, xmm0
        vpsrlq  xmm0, xmm0, 32
        vpmovsxwd       xmm0, xmm0
        vpaddd  xmm1, xmm1, xmm2
        vpxor   xmm2, xmm2, xmm2
        vpshufb xmm2, xmm2, XMMWORD PTR .LC1[rip]
        vpaddd  xmm0, xmm0, xmm1
        vpshufb xmm1, xmm0, XMMWORD PTR .LC0[rip]
        vpor    xmm1, xmm1, xmm2
        vpaddd  xmm0, xmm0, xmm1
        vmovd   eax, xmm0

[Bug target/113079] [x86] Fails to generate dot_prod instructions for 64-bit vector.

[liuhongt at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=113079

--- Comment #1 from Hongtao Liu <liuhongt at gcc dot gnu.org> ---
(In reply to Hongtao Liu from comment #0)
> int
> foo (int n, unsigned char* p, char* pi)
> {
>     int sum = 0;
>     for (int i = 0; i != 8; i++)
>     {
>         sum += p[i] * pi[i];
>     }
>     return sum;
> }
> 
> We can use 128-bit dot_prod instruction + clean upper 64 bits. Currently,
clean upper is not needed since it's integral operations, no side effect from
upper 64-bits operations.

[Bug target/113079] [x86] Fails to generate dot_prod instructions for 64-bit vector.

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=113079

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
     Ever confirmed|0                           |1
   Last reconfirmed|                            |2023-12-19

--- Comment #2 from Richard Biener <rguenth at gcc dot gnu.org> ---
It's includes a lane reduction so we need to have those zero at least on
GIMPLE (if we'd do it there) because the tree codes do not specify which
lanes are reduced.  The actual x86 instruction is probably fine so if
you make V8QI operation available in the backend that should work without
zeroing.

[Bug target/113079] [x86] Fails to generate dot_prod instructions for 64-bit vector.

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=113079

--- Comment #3 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:fa911365490a7ca308878517a4af6189ffba7ed6

commit r15-235-gfa911365490a7ca308878517a4af6189ffba7ed6
Author: liuhongt <hongtao.liu@intel.com>
Date:   Wed Dec 20 11:43:25 2023 +0800

    Support dot_prod optabs for 64-bit vector.

    gcc/ChangeLog:

            PR target/113079
            * config/i386/mmx.md (usdot_prodv8qi): New expander.
            (sdot_prodv8qi): Ditto.
            (udot_prodv8qi): Ditto.
            (usdot_prodv4hi): Ditto.
            (udot_prodv4hi): Ditto.
            (sdot_prodv4hi): Ditto.

    gcc/testsuite/ChangeLog:

            * gcc.target/i386/pr113079.c: New test.
            * gcc.target/i386/pr113079-2.c: New test.
            * gcc.target/i386/sse4-pr113079-2.c: New test.

[Bug target/113079] [x86] Fails to generate dot_prod instructions for 64-bit vector.

[liuhongt at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=113079

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

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

--- Comment #4 from Hongtao Liu <liuhongt at gcc dot gnu.org> ---
Fixed in GCC15.