Bug 18064 and covariant returns

[Mark Mitchell]

Nathan Sidwell wrote:
> We apparently have a gnu extension that allows a virtual function
> returning void * to be overridden by one returning T *, provided
> T is not a class type
> 
> search.c:
>       else if (!pedantic
>            && can_convert (TREE_TYPE (base_type), TREE_TYPE (over_type)))
>     /* GNU extension, allow trivial pointer conversions such as
>        converting to void *, or qualification conversion.  */
>     {
>       /* can_convert will permit user defined conversion from a
>          (reference to) class type. We must reject them.  */
>       over_return = non_reference (TREE_TYPE (over_type));
>       if (CLASS_TYPE_P (over_return))
>         fail = 2;
> 
> This is not documented in extend.texi.
> 
> Do we want to keep the extension?
> 
> My preference would be to deprecate and remove it (it's easier to
> remove than document), but I'm not passionate about that.

I would be in favor of removing this extension.  However, as extensions 
go, this is relatively non-evil, so I would not be violently opposed. 
(If we're going to keep the extension, it seems like it would actually 
make more sense to allow any pointer type convertible to "void *".  In 
general, one could extend C++ to allow any covariant type, so that, for 
example, you could override a function returning "int" with one 
returning "short".  Not that I think that's a great idea.)

-- 
Mark Mitchell
CodeSourcery, LLC
mark@codesourcery.com