[Bug target/97872] New: Missed optimization for less-than comparison on vectors

[Bug target/97872] New: Missed optimization for less-than comparison on vectors

[prathamesh3492 at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97872

            Bug ID: 97872
           Summary: Missed optimization for less-than comparison on
                    vectors
           Product: gcc
           Version: 11.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: target
          Assignee: unassigned at gcc dot gnu.org
          Reporter: prathamesh3492 at gcc dot gnu.org
  Target Milestone: ---

Hi,
For the following test-case:

#include <arm_neon.h>

uint8x8_t f1(int8x8_t a, int8x8_t b) {
  return a < b;
}

uint8x8_t f2(int8x8_t a, int8x8_t b) {
  return vclt_s8 (a, b);
}

Code-gen for f2 uses vcgt insn
f2:
        vcgt.s8 d0, d1, d0
        bx      lr

However code-gen for f1 results in:
f1:
        vmov.i32 d16, #0xffffffff  @ v8qi
        vmov.i32 d17, #0  @ v8qi
        vcgt.s8 d0, d1, d0
        vbsl    d0, d16, d17
        bx      lr

which IIUC is redundant, since vcgt will set all-ones, or all-zeros in d0
depending on the comparison.

The reason this happens is because vclt_s8 uses __builtin_neon_vcgtv8qi that
emits vcgt.s8, while f1 is lowered to using VCOND in optimized dump:

f1 (int8x8_t a, int8x8_t b)
{
  vector(8) signed char _2;
  uint8x8_t _5;

  <bb 2> [local count: 1073741824]:
  _2 = .VCOND (a_3(D), b_4(D), { -1, -1, -1, -1, -1, -1, -1, -1 }, { 0, 0, 0,
0, 0, 0, 0, 0 }, 107);
  _5 = VIEW_CONVERT_EXPR<uint8x8_t>(_2);
  return _5;

}

and correspondingly expanded to:
;; _2 = .VCOND (a_3(D), b_4(D), { -1, -1, -1, -1, -1, -1, -1, -1 }, { 0, 0, 0,
0, 0, 0, 0, 0 }, 107);

(insn 7 6 8 (set (reg:V8QI 117)
        (const_vector:V8QI [
                (const_int -1 [0xffffffffffffffff]) repeated x8
            ])) "foo.c":4:12 -1
     (nil))

(insn 8 7 9 (set (reg:V8QI 118)
        (const_vector:V8QI [
                (const_int 0 [0]) repeated x8
            ])) "foo.c":4:12 -1
     (nil))

(insn 9 8 10 (set (reg:V8QI 119)
        (neg:V8QI (gt:V8QI (reg/v:V8QI 116 [ b ])
                (reg/v:V8QI 115 [ a ])))) "foo.c":4:12 -1
     (nil))

(insn 10 9 0 (set (reg:V8QI 113 [ _2 ])
        (unspec:V8QI [
                (reg:V8QI 119)
                (reg:V8QI 117)
                (reg:V8QI 118)
            ] UNSPEC_VBSL)) "foo.c":4:12 -1
     (nil))

Thanks,
Prathamesh

[Bug target/97872] [ARM NEON] Missed optimization for less-than comparison on vectors

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97872

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Last reconfirmed|                            |2020-11-17
     Ever confirmed|0                           |1
             Status|UNCONFIRMED                 |NEW

--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
So arm doesn't have a vec_cmp pattern that applies here?  Before ISEL we have

  vector(8) <signed-boolean:8> _1;
  vector(8) signed char _2;
  uint8x8_t _5;

  <bb 2> [local count: 1073741824]:
  _1 = a_3(D) < b_4(D);
  _2 = VEC_COND_EXPR <_1, { -1, -1, -1, -1, -1, -1, -1, -1 }, { 0, 0, 0, 0, 0,
0, 0, 0 }>;
  _5 = VIEW_CONVERT_EXPR<uint8x8_t>(_2);
  return _5;

and ISEL should be able to directly expand a_3(D) < b_4(D) via .VEC_CMP

Alternatively the backend should optimize vcond expansion when the
operands are all ones/zeros.

[Bug target/97872] [ARM NEON] Missed optimization for less-than comparison on vectors

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97872

--- Comment #2 from CVS Commits <cvs-commit at gcc dot gnu.org> ---
The master branch has been updated by Prathamesh Kulkarni
<prathamesh3492@gcc.gnu.org>:

https://gcc.gnu.org/g:3a6e3ad38a17a03ee0139b49a0946e7b9ded1eb1

commit r11-5839-g3a6e3ad38a17a03ee0139b49a0946e7b9ded1eb1
Author: Prathamesh Kulkarni <prathamesh.kulkarni@linaro.org>
Date:   Tue Dec 8 14:30:04 2020 +0530

    gimple-isel: Fold x CMP y ? -1 : 0 to x CMP y [PR97872]

    gcc/
    2020-12-08  Prathamesh Kulkarni  <prathamesh.kulkarni@linaro.org>

            PR target/97872
            * gimple-isel.cc (gimple_expand_vec_cond_expr): Try to fold
            x CMP y ? -1 : 0 to x CMP y.

    gcc/testsuite/
    2020-12-08  Prathamesh Kulkarni  <prathamesh.kulkarni@linaro.org>

            PR target/97872
            * gcc.target/arm/pr97872.c: New test.

[Bug target/97872] [ARM NEON] Missed optimization for less-than comparison on vectors

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97872

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |ASSIGNED