[Bug c++/66957] [4.9/5/6 Regression] incorrect "is protected within this context" error

["Casey at Carter dot net" <gcc-bugzilla at gcc dot gnu dot org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66957

Casey Carter <Casey at Carter dot net> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |Casey at Carter dot net

--- Comment #8 from Casey Carter <Casey at Carter dot net> ---
The fix for this bug seems to have broken N4381-style customization points
(http://wg21.link/n4381). To wit, this program used to compile:

template <class T> T&& declval();

namespace X {
  namespace Y {
    int f(auto&&);

    struct fn {
      constexpr auto operator()(auto&& o) const {
        return f(o);
      }
    };
  }
  namespace {
    constexpr auto f = Y::fn{};
  }

  struct A {};

  struct B {
    friend double f(B&);
  };
}

static_assert(__is_same_as(int, decltype(X::f(declval<X::A&>()))));
static_assert(__is_same_as(double, decltype(X::f(declval<X::B&>()))));

but now results in the error "error: âfâ is not a member of âXâ" for both
static_asserts. It's possible to workaround the error be declaring the
anonymous namespace inline:

inline namespace {
  constexpr auto f = Y::fn{};
}

It appears that the non-visible friend function f declared in B is incorrectly
blocking name lookup of X::f before the phase of lookup that considers
namespaces nominated by using directives (such as unnamed namespaces).