[Bug tree-optimization/125750] Excessive vector unrolling leads to excessive spilling

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=125750

--- Comment #5 from GCC Commits <cvs-commit at gcc dot gnu.org> ---
The master branch has been updated by Kyrylo Tkachov <ktkachov@gcc.gnu.org>:

https://gcc.gnu.org/g:6bbe931ea0e8f412c81fed8a3df4a07df97821f4

commit r17-1528-g6bbe931ea0e8f412c81fed8a3df4a07df97821f4
Author: Kyrylo Tkachov <ktkachov@nvidia.com>
Date:   Fri Jun 12 00:00:00 2026 +0000

    match.pd: fold (0/1) * -(0/1) into -((0/1) & (0/1))

    For operands known to be in the range [0, 1], multiplying a 0/1 value by a
    negated 0/1 value is the negation of their bitwise AND:

      x * -y == -(x & y)   when x, y are in { 0, 1 }.

    This complements the existing "{ 0, 1 } * { 0, 1 } -> { 0, 1 } & { 0, 1 }"
    simplification, which does not handle a negated operand.  For the
    comparison-derived 0/1 masks produced by if-conversion this exposes a plain
    bitwise AND of the original conditions to later passes (replacing a
    COND_EXPR).

    This triggers a few times in astcenc in SPEC2026 where it simplifies the
    codegen of one of the hot kernels and gives a ~2.4% improvement on
    aarch64, though the real winners for that kernel are described in
    PR125750. This is just a small cleanup.

    Signed-off-by: Kyrylo Tkachov <ktkachov@nvidia.com>

    gcc/ChangeLog:

            PR tree-optimization/125750
            * match.pd (mult of a 0/1 value by a negated 0/1 value): New
            simplification.

    gcc/testsuite/ChangeLog:

            PR tree-optimization/125750
            * g++.dg/tree-ssa/mult-negate-zeroone-1.C: New test.
            * g++.dg/tree-ssa/mult-negate-zeroone-2.C: New test.