Sublime Text is an awesome editor, and XCompose is very convenient for quickly typing weird Unicode characters. However, these two don’t combine: Sublime Text has an annoying bug which prevents the xim input method, which handles XCompose files, from working.
What to do? If Sublime Text was open-source, I’d make a patch. But it is not. However, I still made a patch.
If you just want XCompose to work, then add the sublime-imethod-fix PPA
to your APT sources, install the
and restart Sublime Text. (That’s it!) Or, build from source if you’re not
However, if you’re interested in all the gory (and extremely boring) details, with an occasional animated gif, read on.
Hunting the bug
To describe the bug, I will first need to explain its natural environment. In Linux, a desktop graphics stack consists of an X11 server and an application using the Xlib library for drawing the windows and handling user input. When it was conceived, a top-notch UI looked like this:
The X11 protocol and Xlib library are quite high-level: originally, you were expected to send compact, high-level instructions over the wire (such as “fill a rectangle at (x,y,x’,y’)”) in order to support thin clients over slow networks. However, thin clients and mainframes vanished, and in their place came a craving for beautiful user interfaces; and X11 protocol, primitive as it is, draws everything as if it came from 1993. (It is also worth noting that X went from X1 to X11 in three years, and has not changed since then.)
The Compose key and XCompose files are a remnant of that era. Xlib has a notion of input method; that is, you would feed raw keypresses (i.e. the coordinates of keys on the keyboard) to Xlib and it would return you whole characters. This ranged from extremely simple US input method (mapping keys to characters 1:1) to more complex input methods for European languages (using a dedicated key to produce composite characters like é and ç) to very intricate Chinese and Japanese input methods with complex mappings between Latin input and ideographic output.
Modern GUI toolkits like GTK and Qt ignore the X11 protocol almost entirely. The only drawing operation in use is “transfer this image and slap it over a rectangular area” (which isn’t even present in the original X11 protocol). Similarly, they pretty much ignore the X input method, favoring more modern scim and uim.
XCompose is probably the only useful part of the whole X11 stack. Unfortunately, native XCompose support is not present anywhere except the original X input method. Fortunately, both GTK and Qt allow changing their input method to XIM. Unfortunately, Sublime Text somehow ignored the X input method completely even when instructed to use it.
Sublime Text draws its own UI entirely to make it look nice on all the platforms. As such, on Linux it has three layers of indirection: first its own GUI toolkit, then GTK, which it uses to avoid dealing with the horror of X11, then X11 itself.
The Xlib interface for communicating with the input method is pretty simple: it’s just the XmbLookupString function. You would feed it the XPressedKeyEvents containing key codes that you receive from the X11 server, and it would give back a string, possibly empty, with the sequence of characters you need to insert in your text area. Also, in order to start communicating, you need to initialize an X input context corresponding to a particular X window. (An X window is what you’d call a window, but also what you’d call a widget—say, a button has its own X11 window.)
GTK packs the input method communication logic in the gtk_im_context_xim_filter_keypress function it has in its wrapper around the X input method. From there, it’s a pretty deep hole:
- gtk_im_context_xim_filter_keypress uses a helper gtk_im_context_xim_get_ic to get the X input context, and if no context is returned, it resorts to a trivial US keymap;
- gtk_im_context_xim_get_ic pulls the X input method handle and associated GTK
settings from the
- which is initialized by the set_ic_client_window helper;
- which refuses to initialize it if
- which is called (through one more layer of indirection used by GTK to change the input methods on the fly) by Sublime Text itself;
- which passes
1 2 3 4 5 6 7 8
What is a GdkWindow? It’s a structure used by GTK to wrap X11 windows on Linux and
other native structures on the rest of platforms. As such, if the GdkWindow and
its underlying X11 window are not yet created, say, because these windows were yet
never shown, the field would contain
NULL. And since Sublime Text attempts to bind
the IM context to the window immediately after creating the latter, this is exactly
the bug which we observe.
It is worth noting that while no input methods that require the window to be know work,
a simple GTK fallback that queries the system for the key configured as Compose key, but
uses internally defined tables with commonly used sequences, does. This is why if you
launch Sublime Text as
GTK_IM_METHOD=whatever-really subl allows you to enter
<Multi_key> <o> <o>, but not customize it by changing any of the XCompose files.
Cooking the meat
How do we fix this? I started with a simple gdb script:
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23
On a high level, the script does four things:
- Sublime Text forks at startup, so the script has to do a little funny dance to attach gdb to the correct process.
- Then, it stops at the point in the initialization sequence where my Sublime Text
build calls gtk_im_context_set_client_window, and captures the
multicontextvariables, which the compiler happened to leave around in spare registers.
- Then, it waits until GTK surely initializes a GdkWindow for the
- Then, it calls gtk_im_context_set_client_window again, exactly as Sublime Text would, but at the right time.
The script works. However, it is slow at startup and not very convenient in general. In particular, I would have to rewrite it every time Sublime Text updates. So, I opted for a better approach.
LD_PRELOAD (see also tutorial: 1, 2) is a convenient feature of Linux dynamic linker which allows to substitute some functions contained in a shared library with different functions contained in another shared library. This is how, for example, fakeroot performs its magic.
Initially I wanted to intercept gtk_window_new and gtk_im_multicontext_new to get the GtkIMMulticontext and the GtkWindow Sublime Text creates—they’re the first ever created—and then gtk_im_context_filter_keypress to call gtk_im_context_set_client_window before the first keypress is handled. But, somehow these calls were not intercepted by LD_PRELOAD; perhaps a weird way Sublime Text calls dlsym? I never figured it out.
So, eventually I settled on intercepting the initialization of the GTK XIM input method plugin (which is loaded by GTK itself and therefore can be intercepted easily) and replacing its filter_keypress handler with my own. A filter_keypress handler receives a GtkIMContext and a GdkEvent, which contains the pointer to GdkWindow, so that would give me all the information I need.
Celebrating the game
Indeed, the goal was achieved in full. It only took me about ten hours, with practically no prior knowledge of libx11 or libgtk internals, access to Sublime Text source, or experience in reverse engineering.
But what was this for? I don’t think I ever needed to type ಠ_ಠ in Sublime Text.
I think I just like the sense of control over my tools.