C++ PATCH for c++/90537 - Implement P1286R2, Contra CWG DR1778

[Jason Merrill]

On Thu, Sep 5, 2019 at 2:14 PM Marek Polacek <polacek@redhat.com> wrote:
>
> This patch implements C++20 P1286R2, Contra CWG DR1778 as described here
> <http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2019/p1286r2.html>.
> Essentially, it removes the restriction that the exception specification of
> a defaulted special member matches the implicit exception specification.
> "Implementing" it means just removing some code (woo!) that was marking
> a function DECL_DELETED_FN if the exception-specifications didn't match...
>
> ...except there's a snag.  We no longer reject
>
>   struct A {
>     A() noexcept(false) = default;
>   };
>
> because A::A() is no longer deleted.  But this assert fails:
>
>   static_assert(!noexcept(A()), "");
>
> and it doesn't seem to me it should, but it's sort of Catch 22.  The default
> constructor *is* trivial as per [class.default.ctor]/3, so trivial_fn_p is
> true.  When build_over_call sees a default_ctor_p that is trivial_fn_p, it
> elides the call and creates a TARGET_EXPR <D, void_cst>.  When evaluating
> noexcept(A()) expr_noexcept_p gets TARGET_EXPR <D, {}> and since there is
> no CALL_EXPR/AGGR_INIT_EXPR that throws, it says "ok, this can't throw".
>
> [except.spec]/6: An expression e is potentially-throwing if
> (6.1) e is a function call [...] with a potentially-throwing exception specification, or
> (6.2) e implicitly invokes a function [...] that is potentially-throwing, or
> ...
> (6.6) any of the immediate subexpressions of e is potentially-throwing.

> which of these apply here, 6.1?  But such a TARGET_EXPR can't throw, can it?

Ah, interesting.  It seems to me that a trivial function should not be
considered potentially-throwing.  I'm not sure whether it's best to
address this by making it non-trivial or overriding the exception
specification.  I think the definition of potentially-throwing is not
the right place to fix it.  Do you want to email CWG or shall I?

Jason