[libc++][ranges] `elements_view` determines the tuple-like property with C++23 rules in C++20 mode

[frederick-vs-ja]

The has-tuple-element exposition-only concept (used for std::ranges::elements_view) was modified by P2165R4 in C++23. Previously, the constraints were more relaxed.

As a result, the following program should be well-formed in C++20 but ill-formed in C++23.

#include <cstddef>
#include <ranges>
#include <utility>
#include <type_traits>

template<class T, class U>
struct MyPair : std::pair<T, U> { // std::get works with MyPair specializations.
    using std::pair<T, U>::pair;
};

template<class T, class U>
struct std::tuple_size<MyPair<T, U>> : std::integral_constant<std::size_t, 2> {};

template<class T, class U>
struct std::tuple_element<0, MyPair<T, U>> {
    using type = T;
};

template<class T, class U>
struct std::tuple_element<1, MyPair<T, U>> {
    using type = U;
};

int main() {
    MyPair<int, double> mps[42]{};
    (void)(mps | std::views::elements<0>);
}

Currently, libstdc++ and MSVC STL accept it in C++20 and reject it in C++23 as specified by the standard, but libc++ rejects it in C++20 (Godbolt link).

This might be intended, as the commit adding elements_view (9446182, https://reviews.llvm.org/D136268) used the C++23 rules in the first place. I wonder whether we should use the old rules in C++20 mode.

CC @huixie90

[huixie90]

Yes, when I implemented, I used the “current working draft” c++23 rules straightaway. Personally I think allowing such code in c++20 mode makes the user harder in the future to upgrade their language version.

That being said, I don’t recall this being discussed in SG9.