Google reference: b/9004062

Quite similar, but not identical to PR 56641.

Test case:

// --- cut ---
# 1 "t.h"
template < typename T > struct integral_c {
  static const T value = 0;
};
struct is_reference:integral_c < bool > { };
template < class > struct is_function_ptr_helper { };
template < bool > struct is_function_chooser;

template <> struct is_function_chooser <0 >
{
  template < typename T > struct result_:is_function_ptr_helper < T * > { };
};

template < typename T > struct is_function_impl:is_function_chooser <
  is_reference::value >::result_ < T > { };

#5 "t.cc"
namespace
{
  class NonCloneable;
  void fn1 ()
  {
    is_function_impl < NonCloneable > i;
  }
}
// --- cut ---


Using current (r199023) trunk:

g++  -c t.ii 
t.h: In instantiation of 'struct is_function_chooser<false>::result_<{anonymous}::NonCloneable>':
t.h:13:32:   required from 'struct is_function_impl<{anonymous}::NonCloneable>'
t.cc:10:39:   required from here
t.h:10:34: warning: 'is_function_chooser<false>::result_<{anonymous}::NonCloneable>' has a base 'is_function_ptr_helper<{anonymous}::NonCloneable*>' whose type uses the anonymous namespace [enabled by default]



Richard Smiths says:

  I think GCC is upset that a type defined outside of the main source file
  has a template argument that has internal linkage.

  In principle, that's a valuable diagnostic: it's trying to tell you that
  every time you include that header, you're going to get a different type,
  and you're likely to have an ODR diagnostic.

  However, in this case, it's a false positive, because this happened
  while instantiating a template from the main source file (with a type
  defined in the main source file, no less).

In addition, this warning isn't suppressable via -Wno-... parameter.