CWG2917 [temp.friend] Some variadic template friend declarations should be explicitly ill-formed.

[cor3ntin]

Consider

struct S {

template <typename T>
friend class Foo<T>::Nested; // #1

template <typename ...Ts>
friend class Foo<Ts>::Nested...; //#2

};

Here #1 declares a template friend declaration such that Foo<T>::Nested is a friend of S for any T, per
https://eel.is/c++draft/temp.friend#4 and https://eel.is/c++draft/temp.friend#5

However #2 doesn't really mean anything sensible. Yet there does not seem to be any wording specifying exactly what
variadic friend declarations are or are not allowed.

P2893 incorrectly claimed that #1 was ill-formed and consequently did not add wording to make #2 ill-formed.

[jensmaurer]

Integrated with CWG2917.

[cor3ntin]

Thanks!
But, I am not sure this is sufficient. Consider

template <typename ...Ts>
struct S {
  template <typename>
  friend class Foo<Ts>::Nested...; 
};

This seems fine (and it is a template declaration and a pack expansion)
We need to say that the friend-type-specifier should not contain an unexpanded template parameter that is introduced by the template declaration.

[jensmaurer]

Maybe: If there is a top-level pack expansion, any packs expanded by that pack expansion shall not have been introduced by the template-declaration.

[t3nsor]

Thanks! But, I am not sure this is sufficient. Consider

template <typename ...Ts>
struct S {
  template <typename>
  friend class Foo<Ts>::Nested...; 
};

This seems fine (and it is a template declaration and a pack expansion) We need to say that the friend-type-specifier should not contain an unexpanded template parameter that is introduced by the template declaration.

Why do we want to ban this example? It has a natural interpretation that Foo<Ts...[0]>::Nested, Foo<Ts...[1]>::Nested, etc must each be a class template with one type template parameter, and each such template is declared as a friend of the S specialization.

[jensmaurer]

@t3nsor , we don't want to ban this example, but my suggested amended wording for CWG2917 does ban it, so @cor3ntin is telling me I need to fix my wording. I've made a suggestion that feels a little better to understand than @cor3ntin 's phrasing.

[jensmaurer]

CWG2917 is fixed

[jensmaurer]

This is still in progress. See cplusplus/papers#2032.

[cor3ntin]

Oups, sorry