Incorrect handling of incomplete types in `std::derived_from`

[cor3ntin]

Per https://eel.is/c++draft/meta#lib:is_base_of,

Derived shall be a complete type

This is generally enforced by __is_base_of (the builtin used to implement is_base_of_v)

However, in libc++, derived_from is defined as

llvm-project/libcxx/include/__concepts/derived_from.h

Line 27 in cb1ad98

concept derived_from = is_base_of_v<_Bp, _Dp> && is_convertible_v<const volatile _Dp*, const volatile _Bp*>;

And because the constraints are only enforced in __is_base_of which is not in the immediate context of the is_base_of_v constraint, substituting an incomplete class in is_base_of_v does not result in a substitution failure in the immediate context (which should just cause derived_from to not be satisfied per https://eel.is/c++draft/temp.constr.atomic#3.sentence-2) but instead makes the program ill-formed, which i believe is non-conforming.

Note that GCC diverges in all cases. MSVC and EDG exposes what I believe to be the correct behavior.

I think there are 2 solutions:

• Modify derived_from to use __is_base_of directly (which is what libstdc++ and MS STL do)
• Add a require clause to std::is_base_of_v to check for completeness

https://godbolt.org/z/zGrKEKEao

[cor3ntin]

Note that https://cplusplus.github.io/LWG/lwg-active.html#3929 and https://cplusplus.github.io/LWG/lwg-active.html#2939 are related

The fact that LWG suggest making it a Mandates (rather than a constraint) makes me question whether GCC is correct and this should always be a hard error (in which case it might be a clang bug)

[philnik777]

IMO we should reject such code, since it can easily result in really hard-to-find bugs. Given that the LWG issues you linked seem to agree with me here, I'd rather have Clang reject both implementations of derived_from. I'm actually quite surprised that there is a difference between the two, which makes me question whether I should land #114840, or at least the <concepts> part of it.