[Bug c++/99176] [8/9/10 Regression] GCC rejects const_cast of null pointer in constant expressions

["cvs-commit at gcc dot gnu.org" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=99176

--- Comment #6 from CVS Commits <cvs-commit at gcc dot gnu.org> ---
The releases/gcc-10 branch has been updated by Marek Polacek
<mpolacek@gcc.gnu.org>:

https://gcc.gnu.org/g:581e87b67233556d566df3a0ea33be9c19fbcb2f

commit r10-9687-g581e87b67233556d566df3a0ea33be9c19fbcb2f
Author: Marek Polacek <polacek@redhat.com>
Date:   Fri Apr 9 18:41:21 2021 -0400

    c++: const_cast of null pointer in constant expr [PR99176]

    Here we reject

      constexpr const int *p = nullptr;
      constexpr int *q = const_cast<int*>(p);

    with "conversion of 'const int*' null pointer to 'int*' is not a
    constant expression", which seems bogus.  This code has been rejected
    since r238909 which added the can_convert check when converting a null
    pointer.  I'm not finding any standard rule that this check was supposed
    to enforce.  The original discussion was here
    <https://gcc.gnu.org/legacy-ml/gcc-patches/2016-06/msg01447.html>
    and here
    <https://gcc.gnu.org/legacy-ml/gcc-patches/2016-07/msg00280.html>.

    Since can_convert never assumes a C-style cast, it rejects casting
    away constness as in the test above and in:

      constexpr int *q = (int *)(const int *) nullptr;

    Removing the check only breaks constexpr-nullptr-2.C by not giving any
    diagnostic for line 229:

      constexpr B *pb2 = static_cast<B*>(pa0);  // { dg-error "not a constant
expression" }

    but the cast seems to be valid: we do [expr.static.cast]/7, and
    [expr.const] only says that a reinterpreter_cast and converting from
    void* is invalid in constexpr.  The can_convert check rejected convering
    from void *, but only when converting from a null pointer, so it's not
    good enough.  So I've added a check to catch conversions from cv void*.
    I realize it's not a great time to be adding additional checking, but
    removing the can_convert check would then technically be a regression.

    Let's limit the new check to only trigger for integer_zerop and then remove
    it in GCC 12.

    gcc/cp/ChangeLog:

            DR 1312
            PR c++/99176
            * constexpr.c (is_std_construct_at): New overload.
            (is_std_allocator_allocate): New overload.
            (cxx_eval_call_expression): Use the new overloads.
            (cxx_eval_constant_expression): Reject casting
            from void * as per DR 1312.  Don't check can_convert.

    gcc/testsuite/ChangeLog:

            DR 1312
            PR c++/99176
            * g++.dg/cpp0x/constexpr-nullptr-2.C: Adjust dg-error.
            * g++.dg/cpp0x/constexpr-cast2.C: New test.
            * g++.dg/cpp0x/constexpr-cast3.C: New test.