[3.3] Fix for 10428 breaks template instantiation

[Michael Matz <matz at suse dot de>]

Hi Mark,

I noticed that the current 3.3 branch does not compile the below code
anymore:

-------- snip -------
template<typename T>
struct A {
  ~A() {p->unref();};
  T* p;
};

struct B;

struct C {
  C() {}
  A<B> b;
};

struct B{
  void unref(void);
};
void B::unref(void) {}

C c;
-------- snap -------

It now breaks with this error:
opensp-bug.ii:10:   instantiated from here
opensp-bug.ii:3: error: `unref' undeclared (first use this function)
opensp-bug.ii:3: error: (Each undeclared identifier is reported only once
for   each function it appears in.)

It was compiling on the branch from 2003-04-22 but fails on 2003-04-23.
The ChangeLog for cp/ shows this:

+2003-04-22  Mark Mitchell  <mark@codesourcery.com>
+
+       PR c++/10428
+       * decl.c (check_elaborated_type_specifier): New function, split
+       out from ...
+       (xref_tag): ... here.  Use the new function in more places.
+
+2003-04-21  Mark Mitchell  <mark@codesourcery.com>
+
+       * call.c (build_over_call): Use build_cxx_call.
+       (build_cxx_call): New method, split out of build_over_call.
+       * cp-tree.h (language_function): Add can_throw.
+       (build_cxx_call): Declare it.
+       * decl.c (finish_function): If a function does not contain any
+       calls to functions that can throw an exception, indicate that
+       fact.
+       * decl2.c (mark_used): Do not defer the instantiation of
+       functions, if the current function does not throw.
+       * optimize.c (maybe_clone_body): Copy TREE_NOTHROW to the clones.
+       * pt.c (instantiate_decl): Make sure import_export_decl is called
+       before emitting things.
+       * rtti.c (throw_bad_cast): Use build_cxx_call.
+       (build_dynamic_cast_1): Likewise.
+       * typeck.c (build_function_call): Likewise.

If I revert back just the cp/ directory before those patches went in, it
works again.  I'm not totally sure if this is even valid C++, although it
worked before.  After all, the ~A() calls a function on a still
incomplete type, although if one would defer the instantiation of A<B> to
the place where actually needed (the "C c;") B would be complete.  I just
want to be sure, that it's really invalid C++, and not just an oversight
in your patches.  (The testcase was constructed out of OpenSP, i.e.
there's at least one real world case using this possibly broken code).


Ciao,
Michael.