[Bug tree-optimization/126637] [17 Regression] wrong code at -O{s,2,3} with -fwhole-program on x86_64-linux-gnu

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=126637

--- Comment #4 from GCC Commits <cvs-commit at gcc dot gnu.org> ---
The master branch has been updated by Aldy Hernandez <[email protected]>:

https://gcc.gnu.org/g:b9c9f9350ca8cb8d0a59cb99b5e48f6f7328adc6

commit r17-3038-gb9c9f9350ca8cb8d0a59cb99b5e48f6f7328adc6
Author: Aldy Hernandez <[email protected]>
Date:   Thu Aug 6 07:01:38 2026 +0000

    Handle sub-ranges in frange == and != operators [PR126637]

    When the operands do not intersect, == and != still allow equality if
    each range holds a zero of opposite sign, hence -0.0 == 0.0.  The
    check for this looked at the hull endpoints, ignoring the inner
    sub-range boundaries.  With sub-ranges the zero can be inner,
    e.g. [-1.0, -0.0][1.0, 1.0] with hull [-1.0, 1.0], so it was missed
    and e.g. "e != 0.0" folded to true.

    Tested on ppc64le Linux: regstrap and LAPACK.

            PR tree-optimization/126637

    gcc/ChangeLog:

            * range-op-float.cc (operator_equal::fold_range): Use the
            contains_zero_p method to detect a contained zero instead of
            inspecting the hull endpoints.
            (operator_not_equal::fold_range): Likewise.

    gcc/testsuite/ChangeLog:

            * gcc.dg/tree-ssa/pr126637.c: New test.