Possible bug in implementation of P3310

[higher-performance]

I think #124137 may have introduced a bug. The following code no longer compiles: https://godbolt.org/z/G77fMj1nh

struct None {};
template<class T>
struct Node { using Tail = T; };

template <template<class> class Fixture, typename List>
struct Recursive {
  static void foo() {
    return Recursive<Fixture, typename List::Tail>::foo();
  }
};

template <template<class> class Fixture>
struct Recursive<Fixture, None> {
  static void foo() {}
};

template <class>
class Monadic {};

template <class...>
class Variadic {};

int main() {
  Recursive<Monadic, Node<None> >::foo();  // success
  Recursive<Variadic, Node<None> >::foo();  // error
}

It seems that the partial specialization struct Recursive<Fixture, None> is completely missed by the compiler now... but only when the Fixture is variadic. I assume this is not supposed to be the case?

[llvmbot]

@llvm/issue-subscribers-clang-frontend

Author: None (higher-performance)

I think #124137 may have introduced a bug. The following code no longer compiles: https://godbolt.org/z/G77fMj1nh

struct None {};
template<class T>
struct Node { using Tail = T; };

template <template<class> class Fixture, typename List>
struct Recursive {
  static void foo() {
    return Recursive<Fixture, typename List::Tail>::foo();
  }
};

template <template<class> class Fixture>
struct Recursive<Fixture, None> {
  static void foo() {}
};

template <class>
class Monadic {};

template <class...>
class Variadic {};

int main() {
  Recursive<Monadic, Node<None> >::foo();  // success
  Recursive<Variadic, Node<None> >::foo();  // error
}

It seems that the partial specialization struct Recursive<Fixture, None> is completely missed by the compiler now... but only when the Fixture is variadic. I assume this is not supposed to be the case?

[mizvekov]

Yeah, that is not supposed to happen.

There is some plumbing missing here, in this case we instantiate 'Recursive' only when it needs to be complete, but we match the template arguments against the primary template as soon as we see the specialization. The new flag which modifies the overload resolution behavior needs to be plumbed to a different point in time.