[Bug target/106562] New: PRU: Inefficient code for zero check of 64-bit AND result

[Bug target/106562] New: PRU: Inefficient code for zero check of 64-bit AND result

[dimitar at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106562

            Bug ID: 106562
           Summary: PRU: Inefficient code for zero check of 64-bit AND
                    result
           Product: gcc
           Version: 13.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: target
          Assignee: unassigned at gcc dot gnu.org
          Reporter: dimitar at gcc dot gnu.org
  Target Milestone: ---

GCC generates inefficient code for the following C snippet:

char test(uint64_t a, uint64_t b)
{
        return a && b;
}

test:
        or      r0.b0, r14.b0, r14.b1
        or      r0.b0, r0.b0, r14.b2
        or      r0.b0, r0.b0, r14.b3
        or      r0.b0, r0.b0, r15.b0
        or      r0.b0, r0.b0, r15.b1
        or      r0.b0, r0.b0, r15.b2
        or      r0.b0, r0.b0, r15.b3
        qbeq    .L4, r0.b0, 0
        mov     r14, r16
        mov     r15, r17
        sub     r2, r2, 2
        rsb     r0, r16, 0
        rsc     r1, r17, 0
        mov     r17, r14
        mov     r18, r15
        sbbo    r3.b2, r2, 0, 2
        ldi     r16.b0, (63) & 0xffff
        or      r14, r0, r17
        or      r15, r1, r18
        call    %label(__pruabi_lsrll)
        lbbo    r3.b2, r2, 0, 2
        add     r2, r2, 2
        ret

[Bug target/106562] PRU: Inefficient code for zero check of 64-bit AND result

[dimitar at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106562

Dimitar Dimitrov <dimitar at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
     Ever confirmed|0                           |1
             Status|UNCONFIRMED                 |ASSIGNED
   Last reconfirmed|                            |2022-08-08
           Assignee|unassigned at gcc dot gnu.org      |dimitar at gcc dot gnu.org

[Bug target/106562] PRU: Inefficient code for zero check of 64-bit (boolean) AND result

[dimitar at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106562

--- Comment #1 from Dimitar Dimitrov <dimitar at gcc dot gnu.org> ---
I explored setting REGMODE_NATURAL_SIZE=4 for PRU.  This required adjustments
in many places in middle end to use REGMODE_NATURAL_SIZE instead of word_mode.
That however proved too intrusive. And I don't see how other targets would
benefit.

I'll instead go with defining an expansion of cbranchdi4 for PRU. That would be
contained in the PRU backend only, and thus safer.

[Bug target/106562] PRU: Inefficient code for zero check of 64-bit (boolean) AND result

[dimitar at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106562

--- Comment #2 from Dimitar Dimitrov <dimitar at gcc dot gnu.org> ---
With cbranchdi4 defined, the generated code is now 10 instructions:

test:
        qbne    .L5, r15, 0
        qbeq    .L4, r14, 0
.L5:
        rsb     r0, r16, 0
        rsc     r1, r17, 0
        or      r0, r0, r16
        or      r1, r1, r17
        lsr     r14, r1, 31
        ret
.L4:
        ldi     r14, 0
        ret

Defining BRANCH_COST=0, as avr backend does, shrinks the above to only 7
instructions.

[Bug target/106562] PRU: Inefficient code for zero check of 64-bit (boolean) AND result

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106562

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

https://gcc.gnu.org/g:e95e91eccd022a4a3a86da2749809fbad9afd20e

commit r13-3180-ge95e91eccd022a4a3a86da2749809fbad9afd20e
Author: Dimitar Dimitrov <dimitar@dinux.eu>
Date:   Sun Sep 18 16:27:18 2022 +0300

    pru: Add cbranchdi4 pattern

    Manually expanding into 32-bit comparisons is much more efficient than
    the default expansion into word-size comparisons.  Note that word for PRU
    is 8-bit.

            PR target/106562

    gcc/ChangeLog:

            * config/pru/pru-protos.h (pru_noteq_condition): New
            function declaration.
            * config/pru/pru.cc (pru_noteq_condition): New function.
            * config/pru/pru.md (cbranchdi4): Define new pattern.

    gcc/testsuite/ChangeLog:

            * gcc.target/pru/pr106562-1.c: New test.
            * gcc.target/pru/pr106562-2.c: New test.
            * gcc.target/pru/pr106562-3.c: New test.
            * gcc.target/pru/pr106562-4.c: New test.

    Signed-off-by: Dimitar Dimitrov <dimitar@dinux.eu>

[Bug target/106562] PRU: Inefficient code for zero check of 64-bit (boolean) AND result

[dimitar at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106562

--- Comment #4 from Dimitar Dimitrov <dimitar at gcc dot gnu.org> ---
The ideal PRU code sequence for the snippet would be:

char test(uint64_t a, uint64_t b)
{
        return a && b;
}
        or      r14, r14, r15
        or      r16, r16, r17
        umin    r14, r14, 1
        umin    r14, r14, r16
        ret

Thus I'm trying to implementing the following conversion in
emit_store_flag_int():

   "X != 0" -> "UMIN (X, 1)

[Bug target/106562] PRU: Inefficient code for zero check of 64-bit (boolean) AND result

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106562

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
           See Also|                            |https://gcc.gnu.org/bugzill
                   |                            |a/show_bug.cgi?id=104296

--- Comment #5 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
(In reply to Dimitar Dimitrov from comment #4)
> Thus I'm trying to implementing the following conversion in
> emit_store_flag_int():
> 
>    "X != 0" -> "UMIN (X, 1)

That is basically what I mention in PR 104296.

[Bug target/106562] PRU: Inefficient code for zero check of 64-bit (boolean) AND result

[dimitar at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106562

--- Comment #6 from Dimitar Dimitrov <dimitar at gcc dot gnu.org> ---
https://gcc.gnu.org/pipermail/gcc-patches/2022-August/599276.html gives a good
analysis why deferring expansion decisions to the backend is preferred.

Most backends already define cstore patterns, so it would not be valuable to
add a generic code in emit_store_flag_int() as a fallback if cstore expansion
fails. Such fallback would simply not be utilized on most architectures.

Hence I intend do add a cstore pattern for PRU as a non-intrusive fix for this
PR.

[Bug target/106562] PRU: Inefficient code for zero check of 64-bit (boolean) AND result

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106562

--- Comment #7 from CVS Commits <cvs-commit at gcc dot gnu.org> ---
The master branch has been updated by Dimitar Dimitrov <dimitar@gcc.gnu.org>:

https://gcc.gnu.org/g:ee077d0c5793e1d4ad8d3b033ef2f0225ba6bd59

commit r14-3578-gee077d0c5793e1d4ad8d3b033ef2f0225ba6bd59
Author: Dimitar Dimitrov <dimitar@dinux.eu>
Date:   Tue Jun 13 22:20:13 2023 +0300

    pru: Add cstore expansion patterns

    Add cstore patterns for the two specific operations which can be
    efficiently expanded using the UMIN instruction:
      X != 0
      X == 0
    The rest of the operations are rejected, and left to be expanded
    by the common expansion code.

            PR target/106562

    gcc/ChangeLog:

            * config/pru/predicates.md (const_0_operand): New predicate.
            (pru_cstore_comparison_operator): Ditto.
            * config/pru/pru.md (cstore<mode>4): New pattern.
            (cstoredi4): Ditto.

    gcc/testsuite/ChangeLog:

            * gcc.target/pru/pr106562-10.c: New test.
            * gcc.target/pru/pr106562-11.c: New test.
            * gcc.target/pru/pr106562-5.c: New test.
            * gcc.target/pru/pr106562-6.c: New test.
            * gcc.target/pru/pr106562-7.c: New test.
            * gcc.target/pru/pr106562-8.c: New test.
            * gcc.target/pru/pr106562-9.c: New test.

    Signed-off-by: Dimitar Dimitrov <dimitar@dinux.eu>

[Bug target/106562] PRU: Inefficient code for zero check of 64-bit (boolean) AND result

[dimitar at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106562

Dimitar Dimitrov <dimitar at gcc dot gnu.org> changed:

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

--- Comment #8 from Dimitar Dimitrov <dimitar at gcc dot gnu.org> ---
Should be fixed now in trunk.