https://gcc.gnu.org/bugzilla/show_bug.cgi?id=121778

--- Comment #8 from GCC Commits <cvs-commit at gcc dot gnu.org> ---
The master branch has been updated by Jeff Law <[email protected]>:

https://gcc.gnu.org/g:4fbc0bbc03162f3962ea79bac29d36952867c90f

commit r16-6592-g4fbc0bbc03162f3962ea79bac29d36952867c90f
Author: Shreya Munnangi <[email protected]>
Date:   Thu Jan 8 21:29:38 2026 -0700

    [PR target/121778] Improving rotation detection

    In this PR we're getting code like this out of the gimple optimizers:

    >   _1 = a_4(D) << 63;
    >   _2 = a_4(D) >> 1;
    >   _3 = _2 ^ 1;
    >   _5 = _1 | _3;

    Note the XOR in that sequence.  It spoils our ability to recognize the
    rotation.  As a result we get code like this for rv64gcb:

    >         srli    a5,a0,1
    >         xori    a5,a5,1
    >         slli    a0,a0,63
    >         or      a0,a5,a0

    We can reassociate the operations when the XOR only flips bits resulting
from
    the right or left shift, but not both.  So after reassociation in gimple we
    get:

    >   _1 = a_2(D) r>> 1;
    >   _3 = _1 ^ 1;

    Which results in:

    >         rori    a0,a0,1
    >         xori    a0,a0,1

    We don't bother with the transformation when the XOR is flipping a bit
known to
    be zero (ie, a high bit of the result of the right shift or a low bit on
the
    result of the left shift).  For those cases we already figure out that the
XOR
    is just an IOR and the right things already "just happen".

    This triggered some code generation changes on the SH (not surprising
because
    this BZ was derived from an older SH BZ).  It doesn't seem to significantly
    improve the SH code, though it does turn a cmp/pz + rotate through carry
with a
    rotate + xor with immediate.    That may be a latency win on the SH, I
really
    don't know.

    Shreya did the bulk of the work here.  My contribution was the sister
pattern
    which has the XOR on the other operand and testcase development.

    Bootstrapped and regression tested on x86 & riscv.  Also tested across the
    various embedded targets without any regressions.

            PR target/121778
    gcc/
            * match.pd: Add pattern to recognize rotate with one or more
            bits flipped via xor.
            * config/sh/sh.md (*rotcl); New variant which handles the output
            we get after the match.pd change above.

    gcc/testsuite/
            * gcc.target/riscv/pr121778.c: New test.

    Co-Authored-By: Jeff Law <[email protected]>
  • [Bug target/121778] Improve rot... cvs-commit at gcc dot gnu.org via Gcc-bugs

Reply via email to