Hi all.

The following testcase currently triggers an ICE in 4.1, 4.2 and 4.3:

=== cut here ===
class QGList; // L1
unsigned count() {
  class QGListIterator {
    friend class QGList; // L2
    QGListIterator( const QGList & );
  };
}
=== cut here ===

If I read the standard properly, the friend declaration in L2 introduces a 
local class QGList, which must not be taken into account during name lookup, 
it is therefore marked as "hidden" (front-end term).

When QGList is looked up in QGListIterator's constructor, the local QGList 
is found, and an assertion fails in 'name_lookup_real', hence the ICE. This 
assertion ensures that only namespace-scope bindings are hidden; this 
constraint is not verified in our case: we have a hidden and non 
namespace-scope binding (the local QGList).

The attached patch fixes this by ensuring that if the binding is hidden, we 
are processing a local class. Moreover, the lookup will fail because there is 
no suitable declaration for QGList: the local one coming from the friend 
declaration is not a candidate, nor is the one at L1 (only a declaration in 
the innermost enclosing nonclass scope - that is 'count' - is to be found). In 
other words, this testcase is invalid, and the compiler will output an error 
(however, please note that icc accepts this snippet, so my interpretation 
might be wrong...).

I have successfully regtested this on i686-pc-linux-gnu. Is it OK for 
mainline? For 4.2?

Best regards,
Simon

:ADDPATCH c++: