[Bug c++/50864] [4.6/4.7 Regression] ICE with decltype and "declval" from another namespace

[daniel.kruegler at googlemail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=50864

--- Comment #8 from Daniel Krügler <daniel.kruegler at googlemail dot com> 2011-10-25 18:08:01 UTC ---
(In reply to comment #7)
> The arrow operator (vs, eg, +) seems also essential.

That makes sense to me, because the code could never be valid, so I would
suggest that the keyword is changed to ice-on-invalid-code. The proper code

namespace impl {
  template <class T> T create();
}

template < class lhs, class rhs >
struct is_arrow_operable_impl
{
  template <class T, class U,
        class = decltype(impl::create<T>()->*impl::create<U>())>
  void test();
};


works as expected. I guess the reason for the ICE of the original code is that
the compiler runs into a branch where the second impl::create is interpreted as
a member access, which could never be valid, because there is no (base) class
impl. 

But this is probably only part of the whole story, because if I change the
erroneous example to

struct impl {
  template <class T> static T create();
};

template < class lhs, class rhs >
struct is_arrow_operable_impl
{
  template <class T, class U,
        class = decltype(impl::create<T>()->impl::create<U>())
  >
  void test();
};


I'm getting "internal compiler error: Segmentation fault" without further
context information. Note that *this* example *can* lead to valid code when
attempting to instantiate is_arrow_operable_impl with impl* and int,
respectively, for example.