https://gcc.gnu.org/bugzilla/show_bug.cgi?id=128453

--- Comment #1 from GCC Commits <cvs-commit at gcc dot gnu.org> ---
The master branch has been updated by Richard Biener <[email protected]>:

https://gcc.gnu.org/g:a2d7bbb3494b789291d90d89e865b5ff063493e4

commit r17-6314-ga2d7bbb3494b789291d90d89e865b5ff063493e4
Author: Richard Biener <[email protected]>
Date:   Thu Nov 5 11:53:55 2026 +0100

    tree-optimization/128453 - avoid truncating pow exponents in reassoc

    acceptable_pow_call accepts any integral pow exponent that fits a
    HOST_WIDE_INT, but the exponent is recorded in operand_entry::count,
    which is an unsigned int, so for example pow (x, 4294967296.0) gets a
    count of zero.  attempt_builtin_powi then removes that entry from the
    operand list for free, dropping the pow call from the product
    entirely.  Likewise the accumulated occurrence count of a factor is
    used as the int exponent of __builtin_powi without checking it fits.

    Widen operand_entry::count to HOST_WIDE_INT, reject exponents in
    acceptable_pow_call that do not fit the exponent argument type of
    __builtin_powi, avoid overflowing the accumulated occurrence counts
    and split powers exceeding the range of the __builtin_powi exponent
    argument over multiple calls.

            PR tree-optimization/128453
            * tree-ssa-reassoc.h (operand_entry::count): Change type to
            HOST_WIDE_INT.
            * tree-ssa-reassoc.cc (acceptable_pow_call): Reject exponents
            not fitting integer_type_node.
            (attempt_builtin_powi): Record a factor in a separate element
            when accumulating its occurrence count would overflow.  Limit
            the power of __builtin_powi calls to the maximum value of
            integer_type_node.

            * gcc.dg/pr128453-1.c: New testcase.
            * gcc.dg/pr128453-2.c: Likewise.

Reply via email to