Question on P3074R5: trivial destructor

[huixie90]

The paper presented that the union has to have an empty destructor. However, to solve the triviality problem, we can just have two overloads of the destructor. And that is what we did in std::expected.
https://github.com/llvm/llvm-project/blob/0cb2fe5183c9b25bb96140c27d12b1ad4a80aa92/libcxx/include/__expected/expected.h#L250

(And we are still very keen on seeing this paper solving the issue of starting lifetime of an array)

[frederick-vs-ja]

I think the intent is making union U { std::string arr[N]; }; and its friends trivially destructible, which is not yet possible in C++23.

[huixie90]

@frederick-vs-ja why do you want union U { std::string arr[N]; }; to be trivially destructible? surely inplace_vector<string, N> can never be trivially destructible unless you can make string trivially destructible, which will never happen?

[frederick-vs-ja]

Oh. I must have been confused. Surely, types exposed as standard features only at most need conditionally trivial destructors, which is already available.

LWG3634 and CWG2929 may be related. Implementations want these pointed to or referenced to objects to be immortal (and are already using some tricks for this purpose). Currently, the involved class type don't seem to have constexpr destructors, and are certainly not trivially destructible (since the dtors are virtual). It seems desired to allow a union wrapping such a type to be immortal and trivially destructible.