[Bug tree-optimization/105793] New: Missed vectorisation with conditional-select inside loop

[Bug tree-optimization/105793] New: Missed vectorisation with conditional-select inside loop

[ktkachov at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105793

            Bug ID: 105793
           Summary: Missed vectorisation with conditional-select inside
                    loop
           Product: gcc
           Version: unknown
            Status: UNCONFIRMED
          Keywords: missed-optimization
          Severity: normal
          Priority: P3
         Component: tree-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: ktkachov at gcc dot gnu.org
  Target Milestone: ---

The code:
#define N 1024

float f(const float in[N], unsigned int n) {
    float a = 0.0f;

    for (unsigned  i = 0; i < N; ++i) {
        float b = in[i];
        if (b < 10.f)
            a += b;
        else
            a -= b;
    }

    return a;
}

with -Ofast does not vectorise (on aarch64, for example):
f:
        movi    v0.2s, #0
        add     x1, x0, 4096
        fmov    s3, 1.0e+1
.L5:
        ldr     s1, [x0], 4
        fsub    s2, s0, s1
        fcmpe   s1, s3
        fadd    s0, s0, s1
        fcsel   s0, s0, s2, mi
        cmp     x1, x0
        bne     .L5
        ret

whereas clang can and does. Commenting out the "else a -=b;" line allows GCC to
vectorise it:
f:
        movi    v0.4s, 0
        add     x1, x0, 4096
        fmov    v3.4s, 1.0e+1
.L2:
        ldr     q2, [x0], 16
        fcmgt   v1.4s, v3.4s, v2.4s
        and     v1.16b, v1.16b, v2.16b
        fadd    v0.4s, v0.4s, v1.4s
        cmp     x1, x0
        bne     .L2
        faddp   v0.4s, v0.4s, v0.4s
        faddp   v0.4s, v0.4s, v0.4s
        ret

Examples at https://gcc.godbolt.org/z/qbn6T73qE

[Bug tree-optimization/105793] Missed vectorisation with conditional-select inside loop

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105793

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

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

--- Comment #1 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
I think there is another bug about this. Basically it comes down to recognizing
conditional negative.

[Bug tree-optimization/105793] Missed vectorisation with conditional-select inside loop

[crazylht at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105793

--- Comment #2 from Hongtao.liu <crazylht at gmail dot com> ---
Guess vectorizer expects something like

tmp1 = cond ? b : -b;
a_5 = a_4 + tmp1;

from ifcvt instead of current

  a_13 = b_10 + a_16;
  # DEBUG a => NULL
  _4 = b_10 < 1.0e+1;
  # DEBUG BEGIN_STMT
  a_12 = a_16 - b_10;
  # DEBUG a => NULL
  a_5 = _4 ? a_13 : a_12;

[Bug tree-optimization/105793] Missed vectorisation with conditional-select inside loop

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105793

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
            Version|unknown                     |13.0
             Target|                            |aarch64
     Ever confirmed|0                           |1
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2022-06-01

--- Comment #3 from Richard Biener <rguenth at gcc dot gnu.org> ---
There's already some code in if-conversion to deal with the vectorizers
restrictions with respect to how reductions have to appear.  Basically the
vectorizer currently does not accept

  for (..)
    a = b < 10. ? a + b : a - b;

because there are two uses of 'a' here.  Re-writing this to

  for (..)
    a = a + (b < 10. ? b : -b)

would indeed work.  See is_cond_scalar_reduction for the existing special
casing.