[Bug libgcj/14296] New: 3.4 regression: Method.Invoke

["aph at gcc dot gnu dot org" <gcc-bugzilla at gcc dot gnu dot org>]

gcj 3.4 is broken because of a change to natMethod.cc in
http://gcc.gnu.org/ml/java-patches/2003-q4/txt00009.txt.  A test case
is attached.

It seems like _Jv_isVirtualMethod is wrong, and has always been wrong.
Here is the logic in the runtime:

// Determine if METH gets an entry in a VTable.
static inline jboolean _Jv_isVirtualMethod (_Jv_Method *meth)
{
  using namespace java::lang::reflect;
  return (((meth->accflags & (Modifier::STATIC | Modifier::PRIVATE)) == 0)
          && meth->name->data[0] != '<');
}

and here is the corresponding logic in the compiler:

     if (super_method != NULL_TREE && ! METHOD_PRIVATE (super_method))
	{
	  DECL_VINDEX (method_decl) = DECL_VINDEX (super_method);
	  if (DECL_VINDEX (method_decl) == NULL_TREE 
	      && !CLASS_FROM_SOURCE_P (this_class))
	    error ("%Jnon-static method '%D' overrides static method",
                   method_decl, method_decl);
	}
      else if (! METHOD_FINAL (method_decl)
	       && ! METHOD_PRIVATE (method_decl)
	       && ! CLASS_FINAL (TYPE_NAME (this_class))
	       && dtable_count)
	{
	  DECL_VINDEX (method_decl) = dtable_count;
	  dtable_count = fold (build (PLUS_EXPR, integer_type_node,
				      dtable_count, integer_one_node));
	}

The important thing to note here is that the compiler doesn't allocate
a new vtable entry for a method if the whole class is final.  This
seems reasonable enough, but the runtime doesn't know that and it
tries to use the vtable index.  Which is -1.  Bang.

-- 
           Summary: 3.4 regression: Method.Invoke
           Product: gcc
           Version: 3.4.0
            Status: UNCONFIRMED
          Severity: critical
          Priority: P1
         Component: libgcj
        AssignedTo: aph at redhat dot com
        ReportedBy: aph at gcc dot gnu dot org
                CC: gcc-bugs at gcc dot gnu dot org


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=14296