Class template partial specialization with concept-auto non-type parameter is not supported

[Fedr]

In this program

#include <concepts>

template <auto>
struct A;

template <std::same_as<int> auto p>
struct A<p> {};

A<0> a;

A<auto> is a class template with non-type parameter, and A<std::same_as<int> auto> is its specialization for int parameters. The program is accepted by GCC and MSVC, but Clang complains:

error: class template partial specialization is not more specialized than the primary template [-Winvalid-partial-specialization]

Online demo: https://godbolt.org/z/ee4TcPY5K

[llvmbot]

@llvm/issue-subscribers-clang-frontend

Author: None (Fedr)

In this program ``` #include <concepts>

template <auto>
struct A;

template <std::same_as<int> auto p>
struct A<p> {};

A<0> a;

`A<auto>` is a class template with non-type parameter, and `A<std::same_as<int> auto>` is its specialization for `int` parameters. The program is accepted by GCC and MSVC, but Clang complains:

error: class template partial specialization is not more specialized than the primary template [-Winvalid-partial-specialization]

Online demo: https://godbolt.org/z/ee4TcPY5K
</details>

[cor3ntin]

Interestingly, it does work with a requires clause
https://godbolt.org/z/bYM43r4Yr

@erichkeane

[zyn0217]

We have to compare the template parameter equivalence before going through the constraint check, per [temp.func.order]p6.2.2:

Otherwise, if the corresponding template-parameters of the template-parameter-lists are not equivalent [temp.over.link]... neither template is more specialized than the other.

And the definition for 'equivalent' is at [temp.over.link]p6.3:

Two template-parameters are equivalent under the following conditions:

if they declare non-type template parameters, they have equivalent types ignoring the use of type-constraints for placeholder types

And this is implemented by bc62fb9.

Interestingly, it does work with a requires clause
https://godbolt.org/z/bYM43r4Yr

I think this could also explain why using a requires expression does work -- the parameters are considered equivalent that way.

(I'm not 100% sure the standard is intended to reject such cases. CC @zygoloid for some opinions.)

[zygoloid]

And the definition for 'equivalent' is at [temp.over.link]p6.3:

Two template-parameters are equivalent under the following conditions:
if they declare non-type template parameters, they have equivalent types ignoring the use of type-constraints for placeholder types

And this is implemented by bc62fb9.

In this case, the template parameters are equivalent, so we should continue to [temp.func.order]/6.4 which orders the function templates based on their constraints.