[Bug target/98648] New: Failure to optimize out no-op vector operation using andnot

[Bug target/98648] New: Failure to optimize out no-op vector operation using andnot

[gabravier at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98648

            Bug ID: 98648
           Summary: Failure to optimize out no-op vector operation using
                    andnot
           Product: gcc
           Version: 11.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: target
          Assignee: unassigned at gcc dot gnu.org
          Reporter: gabravier at gmail dot com
  Target Milestone: ---

__m128 f(__m128 val) 
{
    return _mm_andnot_ps(_mm_set_ps1(0.0f), val);
}

This can be optimized to `return val;`. This optimization is done by LLVM, but
not by GCC.

[Bug target/98648] Failure to optimize out no-op vector operation using andnot

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98648

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|                            |ABI

--- Comment #1 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
I think this is an ABI issue.  What does the ABI say about the other bits of
the SSE register for float passing?
This might be a clang/LLVM issue if the ABI says the other bits are not defined
to be 0.

[Bug target/98648] Failure to optimize out no-op vector operation using andnot

[crazylht at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98648

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

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

--- Comment #2 from Hongtao.liu <crazylht at gmail dot com> ---
It looks like not an ABI issue, but gcc didn't simplify for vector all one bits
for floating point.

.277r.combine

Trying 6 -> 7:
    6: r85:V4SF=const_vector
    7: r84:V4SF=~r85:V4SF&r87:V4SF
      REG_DEAD r87:V4SF
      REG_DEAD r85:V4SF
Failed to match this instruction:
(set (reg:V4SF 84)
    (and:V4SF (reg:V4SF 87)
        (const_vector:V4SF [
                (const_double:SF -QNaN [-QNaN]) repeated x4
            ])))

When NAN is all ones(which is situation here), it should be simplified to (set
(reg:V4SF 84) (reg: V4SF 87))

[Bug target/98648] Failure to optimize out no-op vector operation using andnot

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98648

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

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

--- Comment #3 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
I guess the reason why simplify-rtx.c doesn't optimize it is that these
patterns really don't have valid RTL, bitwise and/or/xor/andnot are expected to
have integral or integral vector arguments, not floating point vector.
Don't know if we couldn't e.g. represent these as
(set (match_operand:V4SF 0)
     (subreg:V4SF
       (and:V4SI (subreg:V4SI (match_operand:V4SF 1) 0)
                 (subreg:V4SI (match_operand:V4SF 2) 0)) 0))
etc.

[Bug target/98648] Failure to optimize out no-op vector operation using andnot

[gabravier at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98648

--- Comment #4 from Gabriel Ravier <gabravier at gmail dot com> ---
This code :

typedef int64_t v2di __attribute__((vector_size(16)));

v2di f(__m128 val) 
{
    return (~(v2di)_mm_set_ps1(0.0f) & (v2di)val);
}

is optimized better (and is equivalent, if I understand the semantics of andnps
right). Maybe the builtin for andnot should be thrown out as soon as possible
(i.e. transformed into ~a & b`) ? From what I can see, `~a & b` for vectors in
general is optimized to an andnot operation too.

[Bug target/98648] Failure to optimize out no-op vector operation using andnot

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98648

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
     Ever confirmed|0                           |1
   Last reconfirmed|                            |2021-12-15
             Status|UNCONFIRMED                 |NEW

--- Comment #5 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
(set (reg:V4SF 84)
    (and:V4SF (reg:V4SF 87)
        (const_vector:V4SF [
                (const_double:SF -QNaN [-QNaN]) repeated x4
            ])))

[Bug target/98648] Failure to optimize out no-op vector operation using andnot

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98648

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

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

[Bug target/98648] Failure to optimize out no-op vector operation using andnot

[crazylht at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98648

--- Comment #6 from Hongtao.liu <crazylht at gmail dot com> ---
Fixed by r12-6071-g19dcecd963295b02b96c8cac57933657dbe3234a

[Bug target/98648] Failure to optimize out no-op vector operation using andnot

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98648

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |RESOLVED
         Resolution|---                         |FIXED
   Target Milestone|---                         |12.0

--- Comment #7 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
.