changes in HEAD need testing

[muppet <scott asofyet org>]

As discussed at length in IRC with Torsten this weekend, i've  
committed changes to the way objects are registered, and these  
changes should get rid of the problems we've been seeing from gtk+'s  
recent addition of GInitiallyUnowned to the ancestry of GtkObject.   
Yes, there were simpler solutions, but this one also closes a few  
very nasty loopholes with the previous scheme, and should prevent  
this sort of problem happening again.

We are considering backporting these changes to the stable-1-10  
series.  So, we ask everyone to please pick up the next unstable  
release tarball (or CVS), install it in a sandbox (instructions in  
README), and try out your favorite gtk2-perl app with it.  If you  
have any problems whatsoever, please let us know immediately, if not  
sooner.

Details below for the curious.


2006/01/15      muppetman

        * Glib.pm, GObject.xs, GType.xs: Change the way GObject type
        mappings are initialized.
        
        Previously, gperl_register_object() would not only map a GType to a
        perl package name, but would set up the @ISA entry for that type by
        introspection; if the parent type was not yet registered, the
        mapping would go into a "pending isa" list, and the code would walk
        that list on each registration.  This left a nice loophole that was
        triggered by the addition of GInitiallyUnowned into libgobject -- if
        a type's parent is unknown to the bindings, that type simply doesn't
        get its ancestry set up correctly.  This ended up breaking all
        nontrivial gtk2-perl programs.  Shame on me for thinking up such a
        fragile scheme.

        Changed the code in various subtle ways.  Added the idea of "lazy
        loading", inspired by Gtk-Perl.  gperl_register_object() still
        registers the association of a GType and a package name, but adds to
        @ISA only "Glib::Object::_LazyLoader".  ISA setup is performed by
        the internal function class_info_finish_loading(), which replaces
        the _LazyLoader name in @ISA with the class's parent and implemented
        interfaces (preserving manually appended and prepended entries).
        This is triggered one of two ways: either by calling
        gperl_object_package_from_type(), which happens when you bless an
        object; or via an AUTOLOAD, isa, or can invocation in _LazyLoader,
        which happens when you try to invoke methods on the class.  When a
        class is set up, all of its parents are checked as well, so this
        happens fairly quickly.  This should be completely transparent to
        all existing code, as the new _LazyLoader stuff in private, and the
        external interfaces have not changed.  The one spot that could break
        is calling UNIVERSAL::(can|isa) on a Glib::Object-derived package
        name before that class has been set up.

        In addition, instead of giving up on unknown classes, we now create
        fake names for them in a "safe" namespace, along the lines of
        Glib::Object::_Unregistered::$c_type_name, and treat them normally.
        For foreign types, this still effectively means they can be treated
        as the first known parent, but the value of ref() will be different
        (but you shouldn't have been relying on that, anyway).


--
elysse (pregnant): are your hands cold?
me: uh, i suppose so.
elysse: will you put them on me?