[Bug target/98060] New: Failure to optimize cmp+setnb+add to cmp+sbb

[Bug target/98060] New: Failure to optimize cmp+setnb+add to cmp+sbb

[gabravier at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98060

            Bug ID: 98060
           Summary: Failure to optimize cmp+setnb+add to cmp+sbb
           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: ---

int r(unsigned v0, unsigned v1, int v2)
{
    return (v0 >= v1) + v2;
}

This code, on x86, can be implemented with `cmp` followed by `sbb`. This
optimization is done by LLVM, but not by GCC.

-O3 x86 output on LLVM :

r:
  mov eax, edx
  cmp edi, esi
  sbb eax, -1
  ret

On GCC :

r:
  xor eax, eax
  cmp edi, esi
  setnb al
  add eax, edx
  ret

[Bug target/98060] Failure to optimize cmp+setnb+add to cmp+sbb

[ubizjak at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98060

Uroš Bizjak <ubizjak at gmail dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Severity|normal                      |enhancement
   Last reconfirmed|                            |2020-11-30
             Status|UNCONFIRMED                 |NEW
     Ever confirmed|0                           |1

--- Comment #1 from Uroš Bizjak <ubizjak at gmail dot com> ---
There are several other cases where sbb/adc can be used:

--cut here--
int r1 (unsigned v0, unsigned v1, int v2)
{
    return (v0 >= v1) + v2;
}

int r2 (unsigned v0, unsigned v1, int v2)
{
    return (v1 > v0) + v2;
}

int r3 (unsigned v0, unsigned v1, int v2)
{
    return (v0 <= v1) + v2;
}

int r4 (unsigned v0, unsigned v1, int v2)
{
    return (v1 < v0) + v2;
}
--cut here--

r1 and r3 can be implemented with sbb $-1, reg, r2 and r4 with adc $0, reg.

gcc currently converts only r4.

[Bug target/98060] Failure to optimize cmp+setnb+add to cmp+sbb

[ubizjak at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98060

--- Comment #2 from Uroš Bizjak <ubizjak at gmail dot com> ---
Created attachment 49662
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=49662&action=edit
Testcases

[Bug target/98060] Failure to optimize cmp+setnb+add to cmp+sbb

[ubizjak at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98060

Uroš Bizjak <ubizjak at gmail dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Assignee|unassigned at gcc dot gnu.org      |ubizjak at gmail dot com
             Status|NEW                         |ASSIGNED

--- Comment #3 from Uroš Bizjak <ubizjak at gmail dot com> ---
Created attachment 49663
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=49663&action=edit
Proposed patch

Proposed patch does a couple of things:

a) introduces reversed adc and sbb insn patterns
b) introduces carry_flag_unset_operator
c) SUBSTANTIALLY cleans carry flag checking predicates
d) rearranges integer compares to allow more combines with carry flag
e) enhances ix86_unary_operator_ok to allow input memory operand for adc and
sbb

I don't think this patch is appropriate for stage-3+, let's wait for stage-1 to
reopen.

[Bug target/98060] Failure to optimize cmp+setnb+add to cmp+sbb

[ubizjak at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98060

Uroš Bizjak <ubizjak at gmail dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
  Attachment #49663|0                           |1
        is obsolete|                            |

--- Comment #4 from Uroš Bizjak <ubizjak at gmail dot com> ---
Created attachment 49666
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=49666&action=edit
Updated patch

[Bug target/98060] Failure to optimize cmp+setnb+add to cmp+sbb

[ubizjak at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98060

Uroš Bizjak <ubizjak at gmail dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Target Milestone|---                         |12.0

[Bug target/98060] Failure to optimize cmp+setnb+add to cmp+sbb

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98060

--- Comment #5 from CVS Commits <cvs-commit at gcc dot gnu.org> ---
The master branch has been updated by Uros Bizjak <uros@gcc.gnu.org>:

https://gcc.gnu.org/g:c111f6066043d3b7bc4141ca0411eae9294aa6c5

commit r12-311-gc111f6066043d3b7bc4141ca0411eae9294aa6c5
Author: Uros Bizjak <ubizjak@gmail.com>
Date:   Fri Apr 30 10:15:26 2021 +0200

    i386: Introduce reversed ADC and SBB patterns [PR98060]

    The compiler is able to merge LTU comparisons with PLUS or MINUS pattern to
    form addition with carry (ADC) and subtraction with borrow (SBB)
instructions:

            op = op + carry         [ADC $0, op]
            op = op - carry         [SBB $0, op]

    The patch introduces reversed ADC and SBB insn patterns:

            op = op + !carry        [SBB $-1, op]
            op = op - !carry        [ADC $-1, op]

    allowing the compiler to also merge GEU comparisons.

    2021-04-30  Uroš Bizjak  <ubizjak@gmail.com>

    gcc/
            PR target/98060
            * config/i386/i386.md (*add<mode>3_carry_0r): New insn pattern.
            (*addsi3_carry_zext_0r): Ditto.
            (*sub<mode>3_carry_0): Ditto.
            (*subsi3_carry_zext_0r): Ditto.
            * config/i386/predicates.md (ix86_carry_flag_unset_operator):
            New predicate.
            * config/i386/i386.c (ix86_rtx_costs) <case PLUS, case MINUS>:
            Also consider ix86_carry_flag_unset_operator to calculate
            the cost of adc/sbb insn.

    gcc/testsuite/

            PR target/98060
            * gcc.target/i386/pr98060.c: New test.

[Bug target/98060] Failure to optimize cmp+setnb+add to cmp+sbb

[ubizjak at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98060

Uroš Bizjak <ubizjak at gmail dot com> changed:

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

--- Comment #6 from Uroš Bizjak <ubizjak at gmail dot com> ---
Implemented in gcc-12.