gtkmm 4 is alive
We (mostly Kjell Ahlstedt and myself) have been quietly working away on gtkmm 4 and an associated ABI-breaking version of glibmm. We’ve been tracking GTK+ 4 from git master, making sure that gtkmm builds against it, and making various API-breaking or ABI-breaking changes that had been left in the code as TODO comments waiting for a chance like this.
This includes simple ABI-breaking (but not API-breaking) stuff such as adding a base classes to widget esor changing the type sof a method parameters. Many other changes are about updating the code to be more modern C++, trying to do things the right way and avoiding duplication with new API in the C++ Standard library.
These changes pleases us as purists but, honestly, as an application developer it isn’t going to give you interesting new behaviour in your applications. If you too are enthusiastic about the C++ renaissance, porting to the new APIs might feel rewarding and correct, and you’ll have to do it someday anyway to get whatever new features arrive later, but I cannot claim there is a more compelling reason to do the work. You might get shiny new features via GTK+ 4, but so far it feels like a similar exercise in internal cleanup and removal of unloved API.
There are still people stuck on GTK+ 2, because most people only port code when they need to. I don’t think the transition to GTK+ 4 or gtkmm 4 will be any faster. Certainly not until the GTK+ 4 porting advice gets a lots better. Porting gtkmm and glom to GTK+ 4 has not been fun, particularly as the trend for API changes without explanation has only increased.
Anyway, here are some of the more significant changes so far in glibmm-2.54 (a horrible ABI name, but there are reasons) and gtkmm-4.0, both still thoroughly unstable and subject to API change:
(Lots of this is currently on in git master but will be in tarball releases soonish, when there are glib and gtk+ releases we can depend on.)
Deprecated API is gone
Anything that was deprecated in gtkmm 3 (or glibmm-2.4) is now completely removed from gtkmm 4 (or glibmm-2.54). You’ll want to build your application with gtkmm 3 with all deprecated API disabled before attempting to build against gtkmm 4.
In some cases this included deprecated base classes that we couldn’t let you optionally disable without breaking ABI, but now they are really really gone.
This is a perfect example of API changes that make us feel better but which are really not of much direct benefit to you as an application developer. It’s for your own good, really.
Glib::RefPtr is now std::shared_ptr
In gtkmm 3, Glib::RefPtr<> is a reference-counting smart pointer, mostly used to hide the manual GObject reference-counting that’s needed in C. It worked well, but C++11 introduced std::shared_ptr so we saw a chance to delete some code and make our API even more “standard”. So, Glib::RefPtr is now just an alias for std::shared_ptr and we will probably change our APIs to use std::shared_ptr instead of Glib::RefPtr.
Glib::RefPtr is an intrusive smart pointer, meaning that it needs, and uses, a reference count in the underlying object rather than in the smartpointer itself. But std::shared_ptr is non-intrusive. So we now just take one reference when we instantiate the std::shared_ptr and release that one reference when the last std::shared_ptr is destroyed, via its Deleter callback. That means we always need to instantiate the std::shared_ptr via Glib::make_refptr_for_instance(), but we hide that inside glibmm/gtkmm code, and application developers would rarely need to do this anyway.
std::shared_ptr<> is a familiar type, so using it in our API should make it easier for the average person to reason about our API. And using it makes us part of the wider ongoing conversation in the C++ community about how to indicate and enforce ownership. For instance, we might start taking Thing* parameters instead of std::shared_ptr<Thing> parameters when we know that the called method will not need to share ownership. I mentioned this idea in an earlier post. However, we often cannot assume much about what the underlying C function really does.
This is a big change. Hopefully it will work.
Now uses libsigc++-3.0 instead of libsigc++-2.0
I rewrote libsigc++ for modern C++, using variadic templates and some other modern C++ techniques. It feels like libsigc++-2.0, but the code is much simpler. Compiler errors might be slightly less cryptic. This requires C++14.
Enums are inside related classes
Enums that were only used with a particular class are now inside that class. For instance, Gio::ApplicationFlags is now Gio::Application::Flags. This makes the API slightly clearer. This didn’t need C++11, but it did need an API break.
Enums are now C++11 enum classes
Enums are now declared as “enum class” (scoped enumerations) instead of “enum” (unscoped enumerations), so they cannot be implicitly converted to other types such as bool or int. That lets the compiler find some subtle programmer errors.
The enum values must now be prefixed by the enum name rather than having a prefix as part of the value name. For instance, we now use Gio::Application::Flags::HANDLES_OPEN instead of Gio::Application::FLAGS_HANDLES_OPEN (actually Gio::APPLICATION_FLAGS_HANDLES_OPEN before we put enums inside classes).
Gtk::TreeView and Gtk::TextView now have real const_iterators
This lets us make the API more const-correct, requiring less arbitrary const_casts<>s in application code.
Removed the old intermediate ListHandle/SListHandle/ArrayHandle/SArrayHandle types
Long ago, the gtkmm API used these intermediate types, wrapping glib types such as GList, GSList, and C arrays, so you didn’t need to choose between using a std::list, std::vector, or other standard container. Since gtkmm 3 we decided that this was more trouble than it was worth, and decided to just uses std::vector everywhere, but it’s only now that we’ve been able to remove them from the glibmm, pangomm, and atkmm APIs.
A possible change: Not using Glib::ustring
We are still considering whether to replace uses of Glib::ustring in our API with std::string, maybe just keeping Glib::ustring for when people really want to manipulate UTF-8 “characters”. I would much prefer standard C++ to have real UTF-8 support, for instance letting us step through and insert characters in a UTF-8 string, but that doesn’t look like it will happen in the foreseeable future.
Glib::ustring still wraps useful UTF-8 APIs in glib, in a std::string-like API, so we wouldn’t want to remove it.
Also, currently it’s useful to know that, for instance, a gtkmm method that returns a Glib::ustring is giving us a UTF-8 string (such as Gtk::FileChooser::get_current_name()), rather than something of unknown encoding (such as Gtk::FileChooser::get_filename()). We allow implicit conversions, for convenience, so we can’t use the compiler to check for awareness of these encoding differences, but having it in the method signature still feels nicer than having to read a method’s documentation.