std::move_only_function::operator bool doesn't work for DLLs

[andreeL]

The current implementation:

_NODISCARD bool _Is_null() const noexcept {
    return _Data._Impl == &_Null_move_only_function;
}

compares with a static object (&_Null_move_only_function), which is different for each DLL

[CaseyCarter]

Thanks for the report!

It would be better if the _Impl itself "knew" whether or not it is null, i.e. via a bool member or bool-returning function.

This might not be fixable without breaking ABI, so we'll need to do something here before freezing C++23.

[AlexGuteniev]

It would be better if the _Impl itself "knew" whether or not it is null, i.e. via a bool member or bool-returning function.

I agree that this would be the way to fix the issue, and also don't see good non-ABI-breaking ways.

[StephanTLavavej]

I think we also need to worry about the case where a move_only_function is constructed in DLL A, then used as the source to move-construct a move_only_function in DLL B, then DLL A is unloaded. DLL B's move_only_function should continue to be usable (as long as it didn't refer to a function in DLL A), including when it is empty.

[AlexGuteniev]

I think we also need to worry about the case where a move_only_function is constructed in DLL A, then used as the source to move-construct a move_only_function in DLL B, then DLL A is unloaded. DLL B's move_only_function should continue to be usable (as long as it didn't refer to a function in DLL A), including when it is empty.

Thanks, I hate it.

I see couple of options:

1. Keep nullptr as empty "vtable" indicator. This will solve both issues, but will add a check for nullptr on every "vtable" use. Hopefully the check will cause extra cmov, not real branch.
2. Embed the content of "vtable" to the object itself, keep nullptr as default _Move / _Destroy, make sure nullptr is also an invalid _Call. This will avoid any dangerous indirections, and potentially improve some operations, but will make the used part of the function object larger, and leave less room for the buffer

I'm leaning towards the first option.