Difference between revisions of "GtkMMification"

From Inkscape Wiki
Jump to navigation Jump to search
(Add a (probably annoying) reply.)
Line 1: Line 1:
Gtk+ and [[GtkMM]] interact cleanly.  This task is to replace individual widgets and dialogs with their [[GtkMM]] ports so as to reduce the code size and make code reading easier.
Gtk+ and GtkMM interact cleanly.  This task is to replace individual widgets and dialogs with their [[GtkMM]] ports so as to reduce the code size and make code reading easier.

We won't be using libglade [1], but we may consider other [[GtkMM]]-based derived widgetsets, such as the ones used for the new [[The_Gimp]] interface.
We won't be using libglade [1], but we may consider other [[GtkMM]]-based derived widgetsets, such as the ones used for the new [[The_Gimp]] interface.

Revision as of 19:33, 29 February 2012

Gtk+ and GtkMM interact cleanly. This task is to replace individual widgets and dialogs with their GtkMM ports so as to reduce the code size and make code reading easier.

We won't be using libglade [1], but we may consider other GtkMM-based derived widgetsets, such as the ones used for the new The_Gimp interface. There are several reasons for not using libglade. First, it imposes an extra dependency that has proven problematic in the past. Second, while it allows for layout of custom widgets, it isn't really designed for that purpose; Inkscape will have its fair share of custom widgets. Third, it is not felt to be suitable to dynamically laid-out dialogs; we wish to make the new dialogs be responsive to changes in the Action system, so will wish to make the dialogs very dynamic. Fourth, we will need to be able to specify UI changes triggered by user actions, which glade does not provide.

That said, use of Glade is encouraged for prototyping and for module/extension interfaces. But for core UI, any Glade-generated code will need to be coded up properly using clean Gtkmm. You'll probably find, though, that Gtkmm's packing toolkit makes visual form layout editing unnecessary.

[1] Comment from Murray Cumming: I can not emphasis enough the great benefits of using libglade(mm). I use it even for layouts that use custom widgets, using get_widget_derived(). Even if you aren't sure about that, I'm sure that you have lots of completely static dialog boxes for which you could obviously use libglade. Using libglade leads to better UI. Murray.

Thanks, yes it does seem quite nice, yet the rationale above still seems to hold. Glade seems helpful for laying out simple static dialogs, however those are also fairly straightforward to code by hand. The extra dependencies, plus the complexity of code partly written by hand and partly generated, plus the level of dynamicism needed still suggest that we'll get less out of glade than the cost of inclusion.
Comment from Murray Cumming: I think you are overstating the additional "dependencies". It's one library, which is as widely available as GTK+. And the libglade replacement, GtkBuilder, is now part of GTK+ itself. The need to keep IDs in sync between the code and the XML is a bit annoying sometimes, but otherwise I have no problem using Glade with custom widgets and with code written by hand (it's all written by hand), so I find it well worth it. Inkscape has a lot of small UI problems that could be easily fixed if it was using Glade.


As part of doing the migration to gtkmm you should try and get icon theming working properly. While not directly related to the port I think fixing it when redoing the GUI anyway is a good time. There are a couple of things to keep in mind. For icon shared by Inkscape with all GTK+ and GNOME apps (file new, file open) etc. you should just use the standard icons. Look at gedit for sample code of how it is done. For icons that Inkscape shares with other drawing apps like The_Gimp you should adopt the stock icons defined by The_Gimp (floodfill is a good example here).

For the remaining icons you would need to define them yourself, good thing here would be to make sure Sodipodi could use the same stock icons definitions so that when someone icon themes Sodipodi or Inkscape both apps picks up the correct icons. -- Uraeus


Do not include the generic <gtkmm.h> header, as that includes all gtkmm headers (about a megabyte or so). Instead include the specific headers needed.

[ n.b. if pch is used, including all the gtkmm headers would not impact things so badly. ]

An easy first-step in converting existing code to Gtkmm would be to replace existing character strings with Glib::ustring, which has a similar interface to std::string but supports UTF-8. Watch out for pointer arithmetic or functions like strlen().

We will need to create a menubar from Gtk::MenuBar. Ideally, this should be dynamically built based on registered verbs (Actions) in the system, as modified by appropriate user preferences.

The XML Editor will require refactoring to make use of the TreeView widget. A few other proposed dialogs have been conceived to "be like the xml editor, except..." so it would probably be wise to create a general purpose TreeEditor class that these dialogs could inherit from, and that provides the general functionality of handling trees and editing their contents.

Change the text properties editor to use TextView. Can we get more powerful text editing than is normally available? E.g., search-and-replace, spellcheck, etc.

The right-click Popup Menu needs to be redeveloped to also tie into the Action system, such that the contents of this menu change dynamically based on what the mouse is hovering over - for text, it should provide text-modification actions, whereas over a node in a path it should show node-editing capabilities.

Toolbars will need to be built from the Action tree, as modified by the user preference settings. There are four principle toolbars: The master, auxillary, extension, and custom toolbars. By default these are positioned on the left, top, bottom, and right sides of the canvas, respectively. The master toolbar is a set of radio buttons corresponding to different aux toolbars; changing the selected radio button changes which aux toolbar is shown. The extension toolbar is used to show third-party tools. The custom toolbar is left to be defined by the user.

Inkscape will require an easy way to get to a variety of dialog pages for setting properties, selecting options, etc. etc. In release 0.35 this was handled via a plethora of popup dialogs, however this is not felt to be an adequate solution. In the new The_Gimp 2.0 look and feel, access to properties are through a combined dialog, where each page of the dialog is accessed via tabbed pages. This is what is being discussed and planned on PreferencesDialog. Other approaches should be researched and evaluated.

The standard Gtk font and color dialogs won't be suited to our needs; we'll need to fold in similar capabilities into the combined dialog mentioned above.

Drag and Drop will need further investigation. Widgets can be specified as sources or destinations for drag and drop actions. It should be possible to drag text from the canvas, other widgets, or external sources into any text entry areas. Dragging of SVG objects off of the canvas into a style editing widget or panel should cause it to adopt the style settings of the dragged object.

Every widget should have a unique, patternized name and take signals in order to operate it. This is to allow scripting such as for automated testing purposes. It should be possible to programmatically drive Inkscape through invoking signals externally, which should result in testable transforms of the document. We can keep a library of these invocation scripts and use them for regression testing.

Dialog windows that aren't shown at startup should be built-up after program initialization during idle times or timeout events. Action button images that are not shown initially should also be rendered for display during idle times, after the program has initiated. The goal is to get the app to the point that it can accept editing commands as rapidly as possible by putting off program initialization work.

We'll need to create a set of our own signals. This can replace situations where pointers are being passed around from object to object. Normal SigC::Signals objects can be used for this.


This will muchly work in the same manner as it has in the codebase, using the english versions of text in gettext() function calls via the _() macro. The xgettext script extracts the strings into the inkscape.pot file. Translators copy this file to languagename.po and edit to suit. The msmerge script updates the .po files from the regenerated .pot file.

Info on getting the gnome-i18n project to help with translations can be found here: http://www.gtkmm.org/gtkmm2/docs/tutorial/html/ch20s04.html

GObject vs. NRObject vs. C++

With GObject, classes are represented by integer ids which are allocated when classes are registered. The usual pattern is to provide some accessor function (normally class_name_here_get_type()) to get a class id for a particular class, which checks if the class has been registered yet, and if not it does so. The registration function itself allocates sufficient memory for the class structure, then copies a populated class structure from the superclass over it (thereby filling in the fields that are common with the superclass). At that point it calls the class init function which takes care of overriding any superclass functions, and populating any new ones that aren't provided by the superclass.

Now, in terms of creating a new object instance, that generally happens via a call to g_object_new(), which takes a type identifier that allocates enough memory for the given type, then walks down the inheritance chain calling init functions in succession. The init functions are responsible for initializing any members.

Now, there are a couple important points to note:

  1. GObject zero-fills object memory before starting the initialization process
  2. As part of initialization, it fills in a couple fields in the GObject header; mainly a pointer to an instance of the appropriate class object (which acts as a vtable, among other things)
  3. the presence of virtual functions in derived classes won't alter the layout (since there's already a vtable pointer in NRObject), but multiple inheritance will still break the expectations of the GObject-style casting macros.

Generally it's safest to stick with single inheritance below NRObject.

The first item in the SPView structure is a 'GObject object'; this is really intended to be an 'is-a' relationship. I.e., object is really the parent object that SPView is "derived" from. For clarity, this should be rewritten to be:

class SPView : public GObject {

Next, SPView should be converted to use NRObject in place of GObject, as a transitional step towards full C++/Gtkmm. In doing this, all the SPView-related GObject macros and functions will need to be converted to their NRObject equivalents. But before you can do that, all of the SPView signals need to be converted to sigc++ signals. These signals have constructors that need to be called, so those constructors will need to be called manually from the init function, using placement new notation.

In the most generic sense, "Placement new" refers to the notation used to pass additional arguments to the new operator, but in its classic usage it's used to pass a single void*, which is used as the address of memory to initialize for the given object type, rather than allocating new memory itself. In essense, `new (ptr) Blah();` means to run Blah's constructor on the memory pointed to by ptr.

Also, the destructor will need to be manually called. This is done via the notation `ptr->~Blah();` In the case of a template class (like signals), you leave off the signal parameters. This needs to be done (as members) for each signal in the init()/dispose() functions for a GObject class with signals. Once you convert to NRObject, you'll need to get rid of those, as the initialization will be done automatically at that point.

Before converting to NRObject, it's imperative to start explicitly initializing every member. GObject allows skipping initialization since it automatically zero-fills the structure, but pure C++ will not. NRObject deliberately initializes with garbage to make sure you don't forget.  ;-)

Finally, SPView should be converted fully into C++. This can be done in several steps. The class should be renamed from SPView to just View, put into the Inkscape::UI::View namespace, and moved into the src/ui/view/ directory. This will require changing all code referring to SPView, obviously. Referrers should be changed to use the member function style instead of macros or procedural style calls.