[Bug target/105965] New: x86: single-element vectors don't have scalar FMA insns used anymore

[Bug target/105965] New: x86: single-element vectors don't have scalar FMA insns used anymore

[jbeulich at suse dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105965

            Bug ID: 105965
           Summary: x86: single-element vectors don't have scalar FMA
                    insns used anymore
           Product: gcc
           Version: 12.1.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: target
          Assignee: unassigned at gcc dot gnu.org
          Reporter: jbeulich at suse dot com
  Target Milestone: ---

While this used to work fine up to gcc8, gcc9 and newer use
vmuls[sdh]+vadds[sdh] instead. No similar issue exists when operating on
scalars, or when operating on multi-element vectors not matching any available
register size (so my guess of "target" as the component may not be correct).

This has regressed the test harness of the Xen Project's insn emulator [1],
which no longer exercises any scalar FMA insns because of the compiler not
emitting any. (Note that using intrinsics is not really an option, as the
primary goal is to test insns with memory operands. Yet the intrinsics don't
lend themselves to such because of using 128-bit parameter types.)

The issue is uniform for FMA, FMA4, AVX512F, and AVX512-FP16. It can be easily
seen by compiling

T test(T x, T y, T z) {
        return x * y + z;
}

#define TEST(n) \
typedef T __attribute__((vector_size(n * sizeof(T)))) v##n##_t; \
v##n##_t test##n(v##n##_t x, v##n##_t y, v##n##_t z) { \
        return x * y + z; \
}

TEST(1)
TEST(2)
TEST(4)
TEST(8)
TEST(16)
TEST(32)
TEST(64)

with e.g. "-mfpmath=sse -O3 -c -mfma -DT=float", but obvious other option
combinations similarly demonstrate the issue.

[1] https://xenbits.xen.org/gitweb/?p=xen.git;a=tree;f=tools/tests/x86_emulator

[Bug target/105965] x86: single-element vectors don't have scalar FMA insns used anymore

[crazylht at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105965

Hongtao.liu <crazylht at gmail dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |crazylht at gmail dot com

--- Comment #1 from Hongtao.liu <crazylht at gmail dot com> ---
It looks like a regression since GCC9

typedef float v1sf __attribute__((vector_size(4)));

v1sf
foo43 (v1sf a, v1sf b, v1sf c)
{
    return a * b + c;
}

gcc9 also don't generate vfmaddXXXss.

        pushq   %rbp
        movq    %rdi, %rax
        movq    %rsp, %rbp
        andq    $-32, %rsp
        vmovss  24(%rbp), %xmm0
        vmulss  16(%rbp), %xmm0, %xmm0
        vaddss  32(%rbp), %xmm0, %xmm0
        vmovss  %xmm0, -64(%rsp)
        movl    -64(%rsp), %edx
        movl    %edx, (%rdi)
        leave
        ret

[Bug target/105965] [10/11/12/13 Regression] x86: single-element vectors don't have scalar FMA insns used anymore

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105965

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Assignee|unassigned at gcc dot gnu.org      |rguenth at gcc dot gnu.org
   Target Milestone|---                         |10.4
           Priority|P3                          |P2
             Status|UNCONFIRMED                 |ASSIGNED
            Summary|x86: single-element vectors |[10/11/12/13 Regression]
                   |don't have scalar FMA insns |x86: single-element vectors
                   |used anymore                |don't have scalar FMA insns
                   |                            |used anymore
   Last reconfirmed|                            |2022-06-14
     Ever confirmed|0                           |1
           Keywords|                            |missed-optimization

--- Comment #2 from Richard Biener <rguenth at gcc dot gnu.org> ---
The widen-mul pass now sees

  <bb 2> [local count: 1073741824]:
  _8 = VIEW_CONVERT_EXPR<float>(a_3(D));
  _9 = VIEW_CONVERT_EXPR<float>(b_4(D));
  _10 = _8 * _9;
  _1 = {_10};
  _11 = VIEW_CONVERT_EXPR<float>(_1);
  _12 = VIEW_CONVERT_EXPR<float>(c_5(D));
  _13 = _11 + _12;
  BIT_FIELD_REF <<retval>, 32, 0> = _13;
  return <retval>;

which confuses it.  The above is the result from vector lowering which
presumably
sees that V1SFmode isn't supported.  In GCC 8 the above is instead

  <bb 2> [local count: 1073741825]:
  _8 = BIT_FIELD_REF <a_3(D), 32, 0>;
  _9 = BIT_FIELD_REF <b_4(D), 32, 0>;
  _10 = _8 * _9;
  _11 = BIT_FIELD_REF <c_5(D), 32, 0>;
  _12 = _10 + _11;
  _2 = {_12};
  <retval> = _2;

that means we are at least missing a match.pd pattern to simplify

  _1 = {_10};
  _11 = VIEW_CONVERT_EXPR<float>(_1);

[Bug target/105965] [10/11/12/13 Regression] x86: single-element vectors don't have scalar FMA insns used anymore

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105965

--- Comment #3 from CVS Commits <cvs-commit at gcc dot gnu.org> ---
The master branch has been updated by Richard Biener <rguenth@gcc.gnu.org>:

https://gcc.gnu.org/g:90467f0ad649d0817f9e034596a0fb85605b55af

commit r13-1085-g90467f0ad649d0817f9e034596a0fb85605b55af
Author: Richard Biener <rguenther@suse.de>
Date:   Tue Jun 14 10:59:49 2022 +0200

    middle-end/105965 - add missing v_c_e <{ el }> simplification

    When we got the simplification of bit-field-ref to view-convert
    we lost the ability to detect FMAs since we cannot look through

      _1 = {_10};
      _11 = VIEW_CONVERT_EXPR<float>(_1);

    the following amends the (view_convert CONSTRUCTOR) pattern
    to handle this case.

    2022-06-14  Richard Biener  <rguenther@suse.de>

            PR middle-end/105965
            * match.pd (view_convert CONSTRUCTOR): Handle single-element
            CTOR case.

            * gcc.target/i386/pr105965.c: New testcase.

[Bug target/105965] [10/11/12 Regression] x86: single-element vectors don't have scalar FMA insns used anymore

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105965

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
            Summary|[10/11/12/13 Regression]    |[10/11/12 Regression] x86:
                   |x86: single-element vectors |single-element vectors
                   |don't have scalar FMA insns |don't have scalar FMA insns
                   |used anymore                |used anymore
      Known to work|                            |13.0

--- Comment #4 from Richard Biener <rguenth at gcc dot gnu.org> ---
Fixed on trunk sofar.

[Bug tree-optimization/105965] [10/11/12 Regression] x86: single-element vectors don't have scalar FMA insns used anymore

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105965

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
          Component|target                      |tree-optimization
           See Also|                            |https://gcc.gnu.org/bugzill
                   |                            |a/show_bug.cgi?id=101540

--- Comment #5 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
(In reply to Richard Biener from comment #2)
> that means we are at least missing a match.pd pattern to simplify
> 
>   _1 = {_10};
>   _11 = VIEW_CONVERT_EXPR<float>(_1);

I had an alternative patch which I had submitted in November:
https://gcc.gnu.org/pipermail/gcc-patches/2021-November/585597.html

I never followed up and such. Sorry about that.

[Bug tree-optimization/105965] [10/11/12 Regression] x86: single-element vectors don't have scalar FMA insns used anymore

[rguenther at suse dot de]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105965

--- Comment #6 from rguenther at suse dot de <rguenther at suse dot de> ---
On Tue, 21 Jun 2022, pinskia at gcc dot gnu.org wrote:

> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105965
> 
> Andrew Pinski <pinskia at gcc dot gnu.org> changed:
> 
>            What    |Removed                     |Added
> ----------------------------------------------------------------------------
>           Component|target                      |tree-optimization
>            See Also|                            |https://gcc.gnu.org/bugzill
>                    |                            |a/show_bug.cgi?id=101540
> 
> --- Comment #5 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
> (In reply to Richard Biener from comment #2)
> > that means we are at least missing a match.pd pattern to simplify
> > 
> >   _1 = {_10};
> >   _11 = VIEW_CONVERT_EXPR<float>(_1);
> 
> I had an alternative patch which I had submitted in November:
> https://gcc.gnu.org/pipermail/gcc-patches/2021-November/585597.html
> 
> I never followed up and such. Sorry about that.

Hmm, that's actually the opposite case where we want a vector type.
But sure, when we replace _1 = {_10} with _1 = V_C_E <..> (_10); then
the V_C_E series should fold away (hopefully ;))

[Bug tree-optimization/105965] [10/11/12 Regression] x86: single-element vectors don't have scalar FMA insns used anymore

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105965

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Target Milestone|10.4                        |10.5

--- Comment #7 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
GCC 10.4 is being released, retargeting bugs to GCC 10.5.

[Bug tree-optimization/105965] [10/11/12 Regression] x86: single-element vectors don't have scalar FMA insns used anymore

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105965

--- Comment #8 from CVS Commits <cvs-commit at gcc dot gnu.org> ---
The releases/gcc-12 branch has been updated by Richard Biener
<rguenth@gcc.gnu.org>:

https://gcc.gnu.org/g:1fe7321a6ce0dcb05763c8f1850a066824516342

commit r12-8578-g1fe7321a6ce0dcb05763c8f1850a066824516342
Author: Richard Biener <rguenther@suse.de>
Date:   Tue Jun 14 10:59:49 2022 +0200

    middle-end/105965 - add missing v_c_e <{ el }> simplification

    When we got the simplification of bit-field-ref to view-convert
    we lost the ability to detect FMAs since we cannot look through

      _1 = {_10};
      _11 = VIEW_CONVERT_EXPR<float>(_1);

    the following amends the (view_convert CONSTRUCTOR) pattern
    to handle this case.

    2022-06-14  Richard Biener  <rguenther@suse.de>

            PR middle-end/105965
            * match.pd (view_convert CONSTRUCTOR): Handle single-element
            CTOR case.

            * gcc.target/i386/pr105965.c: New testcase.

    (cherry picked from commit 90467f0ad649d0817f9e034596a0fb85605b55af)

[Bug tree-optimization/105965] [10/11 Regression] x86: single-element vectors don't have scalar FMA insns used anymore

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105965

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
            Summary|[10/11/12 Regression] x86:  |[10/11 Regression] x86:
                   |single-element vectors      |single-element vectors
                   |don't have scalar FMA insns |don't have scalar FMA insns
                   |used anymore                |used anymore
      Known to work|                            |12.1.1

--- Comment #9 from Richard Biener <rguenth at gcc dot gnu.org> ---
Fixed on the GCC 12 branch as well.  I'm not considering to backport further to
the more mature branches.

[Bug tree-optimization/105965] [11 Regression] x86: single-element vectors don't have scalar FMA insns used anymore

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105965

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Target Milestone|10.5                        |11.5

--- Comment #10 from Richard Biener <rguenth at gcc dot gnu.org> ---
GCC 10 branch is being closed.