Clang does not diagnose an explicit specialization or partial specialization of initializer_list

[miyuki]

According to the C++ standard (C++23 [initializer.list.syn] paragraph 2):

If an explicit specialization or partial specialization of initializer_list is declared, the program is ill-formed.

It does not say "no diagnostic is required"; therefore, I think Clang must produce an error, but it does not. Example code:

#include <initializer_list>

template<>
class std::initializer_list<int> {};

template<typename T>
class std::initializer_list<T*> {};

Clang (current trunk version) accepts the code with -std=c++23 -pedantic-errors.

[llvmbot]

@llvm/issue-subscribers-clang-frontend

Author: Mikhail Maltsev (miyuki)

According to the C++ standard (C++23 [initializer.list.syn] paragraph 2):

> If an explicit specialization or partial specialization of initializer_list is declared, the program is ill-formed.

It does not say "no diagnostic is required"; therefore, I think Clang must produce an error, but it does not. Example code:

#include <initializer_list>

template<>
class std::initializer_list<int> {};

template<typename T>
class std::initializer_list<T*> {};

Clang (current trunk version) accepts the code with -std=c++23 -pedantic-errors.

[miyuki]

I noticed that Clang now has a no_specializations attribute, so it should probably be easier to implement this diagnostic in libc++ rather than in the frontend.

[shafik]

CC @ldionne @mordante this looks like a libc++ issue, correct?

[mordante]

Thanks for catching this!

This is LWG2129: User specializations of std::initializer_list which has been closed as completed by 3687d3c. Both GCC and MSVC have implemented this issue in the compiler. Using the no_specializations attributes solves it in libc++. This does not fix the issue for clang using libstdc++.

Since libstdc++ is the default standard library on Linux I feel it should be fixed in the compiler.

@ldionne do we want to also fix this in libc++ using the no_specializations attribute? (If so I can create a patch.)

[ldionne]

@ldionne do we want to also fix this in libc++ using the no_specializations attribute? (If so I can create a patch.)

Yes, I think we should do that in the library. libstdc++ can do the same if they want to.