c++/9443: [3.4 regression] ICE when accessing a nonstatic member as A::i

[Volker Reichelt]

For the record:

* The ICE in the case with "A::i" is fixed (mainline as of gcc 20030406).
  The error message, however, could still use some improvement:

  PR9443A.cc: In member function `int B<<anonymous> >::foo() [with int 
     <anonymous> = 0]':
  PR9443A.cc:15:   instantiated from here
  PR9443A.cc:12: error: 'struct B<0>' has no member named 'i'

  We tried to access A::i and not B<0>::i as the error message suggests.
  This is as bad as the message on the 3.3 branch:

  PR9443A.cc: In member function `int B<<anonymous> >::foo() [with int 
   <anonymous> = 0]':
  PR9443A.cc:15:   instantiated from here
  PR9443A.cc:12: error: type `A' is not a base type for type `B<0>'

  In fact you can make the code legal if you derive B from A, so this
  might actually help in some situations.

* The accepts-illegal bug in the case "A::j" still persists on mainline.
  I changed the synopsis to reflect the situation.

* I probably reduced the testcase a little too much. The ICE also happened
  with the following testcase which is indeed legal-code (and reflects the
  problem in the original example more accurately). As a side-effect
  the synopsis was a bit misleading, but I changed that anyway.
  
  ------------------------snip here----------------------
  struct A
  {
      int i;
  };

  template <int> struct B : A
  {
      int foo() { return A::i; }
  };

  void bar() { B<0>().foo(); }
  ------------------------snip here----------------------

  This should answer your question, Giovanni, at least partially: Boost
  doesn't contain illegal code (at least not in that spot).

  The boost problems regarding the 3.3 branch are probably a different
  issue (an ICE in c_expand_expr, at c-common.c happens quite often -
  the reason for the failure is usually buried elsewhere).
  Alas there are no more recent results available on

    http://cci.lbl.gov/boost/results/

  The tests were performed with gcc 3.3-20030306. If the problems with the
  3.3 branch persist, we should probably have a look at it. Could you test
  it with a more recent compiler, Giovanni?

Regards,
Volker