With the BoxLayout containers in St and Mx and the ClutterBoxLayout
manager, we now have three more or less diverged implementations of
the same layout policy.
While removing StBoxLayout entirely in favor of ClutterLayoutManager
would be the fashionable thing to do, there are obvious drawbacks:
- it is the only actor we have that implements the scrollable interface
- it conveniently exposes its spacing property in CSS
- last but not least, it is used all over the place
So do the next best thing and make our implementation use the
Clutter layout manager internally - that way, the change is
transparent to users, while we get to refer most of the tricky
bits to Clutter. win-win!
https://bugzilla.gnome.org/show_bug.cgi?id=703810
Replace more direct XFixes usage with a the appropriate abstraction
API from mutter, which is guaranteed to work in wayland too.
It doesn't yet replace pointer position tracking, although probably
it should.
Also, because now we're using Mutter API, we lose the standalone
test case.
https://bugzilla.gnome.org/show_bug.cgi?id=705911
Mutter now includes an object with the same purpose and functionality
as ShellXFixesCursor, so we can replace our XFixes code with it
and work under wayland too.
https://bugzilla.gnome.org/show_bug.cgi?id=705911
In the new application model, there is one ID shared by GApplication,
DBus and .desktop files, so we can use that for the association,
instead of fiddling with badly cased wm classes.
https://bugzilla.gnome.org/show_bug.cgi?id=706252
This is needed to handle applications that are converted to
reverse dns notation, if their application ID includes capital
letters (as it is often the case for DBus names)
https://bugzilla.gnome.org/show_bug.cgi?id=706252
This will replace the indicator painted on the stage right now.
This unfortunately does not work for the recorder triggered by the
keybinding -- we'll simply replace the in-shell code with a keybinding
powered by gnome-settings-daemon.
Commit cfecd063c9 changed the allocation logic to not allocate
scrollbars when the *_visible booleans are false. This breaks the
fade effect as well as the NEVER policy. We do not paint scrollbars
when they are not supposed to be visible, so not allocating them
and thus leaving them in a "needs allocation" state just causes problems.
I am not convinced that it solved any problem to begin with (we don't paint
them anyway).
As the previous condition has basically always been true, just do it
unconditionally.
https://bugzilla.gnome.org/show_bug.cgi?id=705664
Chromium (but not google-chrome) has a StartupWMClass in the desktop
file, so we must match the instance part first to have chrome
web apps working.
Also, we must take care of apps without a wm_class or instance at
all.
https://bugzilla.gnome.org/show_bug.cgi?id=673657
Some applications (such as most Java apps, as well as Chrome Web apps) ship
with desktop files that have the wrong name, but whose StartupWMClass
field contains the right value. Therefore first check that key, against
both the class and instance part of WM_CLASS, and only use the filename
if nothing else works.
https://bugzilla.gnome.org/show_bug.cgi?id=673657
The point of fading the icon is to make the text displayed over the
icon more legible. In RTL layouts, the text is displayed on the left
of the icon, so fading the right-hand-side of the icon doesn't work
well.
https://bugzilla.gnome.org/show_bug.cgi?id=704583
We don't set :visible on the scrollbars, but use booleans to track
if they are visible. Thus check the booleans instead of the actor's
properties when allocating the scrollbars.
https://bugzilla.gnome.org/show_bug.cgi?id=704265
Use ClutterActor.allocate_align_fill() so we don't have to do
this math ourselves. At the same time, clean up the RTL handling
so that it's easier to follow.
https://bugzilla.gnome.org/show_bug.cgi?id=702539
When the St theme is changed, the StThemeContext unrefs all the theme
nodes cached in it's internal hash table, then emits a signal to
notify all theme nodes that the current theme has changed.
The problem is that the first StWidget to catch a theme changed signal
will trigger a "style-changed" signal catched by its children first.
So the theme changed signal can't be processed properly to cleanup
StThemeNodePaintState before recomputing the theme.
This patch adds a weak ref to the StThemeNode in the
StThemeNodePaintState to ensure paint states are properly cleaned up
when the associated StThemeNode is freed.
https://bugzilla.gnome.org/show_bug.cgi?id=703859
Commit 318283fc70 optimized box-shadow rendering by not recreating
shadow materials on every allocation change. Other handles cannot
be reused and are updated regularly, however the patch missed the
cached corner materials - while those can be reused, we still need
to ensure that the currently used paint state references them.
https://bugzilla.gnome.org/show_bug.cgi?id=703909
It is the job of layout containers to arrange their children; having
a hidden feature that *also* allows children to be positioned freely
outside the parent's allocation is just odd.
With the last user of the feature gone, kill it.
https://bugzilla.gnome.org/show_bug.cgi?id=703808
Currently the box-shadow is rendering is done like this :
The first time we want to render a node that requires a box-shadow, St
creates an cogl offscreen surface of the size of the allocation and
renders the box into this offscreen buffer using modulation on the
alpha channel, this buffer is then blurred according to the CSS
parameters.
The problem with this method is that every time an StWidget is
resized, its box-shadow offscreen buffer has to be resized and
therefore rendered and blurred.
This patches propose an optimization for this use case by rendering
the box-shadow only once but at a size that is independent of the
StWidget's size. Then every time we need to paint this box-shadow, we
just render this offscreen buffer using a 9-slices.
This method only works when the allocation of the widget is bigger
than the minimum shadow size on which we can apply a 9-slices, that is
given my the radius of the corners. If the allocation is smaller than
this minimum size, we then fallback to the fully render/blur the
shadow (like before this patch).
https://bugzilla.gnome.org/show_bug.cgi?id=689858
The duality of the Clutter's key focus and mutter's window focus has long been
a problem for us in lots of case, and caused us to create large and complicated
hacks to get around the issue, including GrabHelper's focus grab model.
Instead of doing this, tie basic focus management into the core of gnome-shell,
instead of requiring complex "application-level" management to get it done
right.
Do this by making sure that only one of an actor or window can be focused at
the same time, and apply the appropriate logic to drop one or the other,
reactively.
Modals are considered a special case, as we grab all keyboard events, but at
the X level, the client window still has focus. Make sure to not do any input
synchronization when we have a modal.
At the same time, remove the FOCUSED input mode, as it's no longer necessary.
https://bugzilla.gnome.org/show_bug.cgi?id=700735
This causes a debug SpiderMonkey build to fail when it throws an
exception for the missing symbol, but doesn't properly return FALSE
when executing the script.
https://bugzilla.gnome.org/show_bug.cgi?id=703442
While it is obviously still an error to call get_theme_node() on a
widget that hasn't been added to the stage hierarchy yet, asserting
on it hasn't proven too successful in avoiding those errors - it's
likely the most frequent reason for crash reports. Just accept that
there'll always be code paths where we can hit this case and make
it non-fatal.
https://bugzilla.gnome.org/show_bug.cgi?id=610279
We will allow to use mode-specific overrides; in preparation for that,
move the code so that we only override preferences after initializing
the session mode.
https://bugzilla.gnome.org/show_bug.cgi?id=701717
Otherwise they break the "top level window" detection used by the
unredirect code in mutter, causes game windows not to be unredirected
when tray icons are present.
https://bugzilla.gnome.org/show_bug.cgi?id=701224
Currently we simply set the gsettings key when activating an input
source. This obviously introduces a time window, between the event that
activates the switch and when the switch is complete, under which key
events are being delivered to applications and interpreted according
to the previous input source.
The patches in bug 696996 introduce a DBus API in g-s-d that allows us
to know when an input source if effectively active. Using that and
freezing keyboard events in the X server until we hear back from g-s-d
we can ensure that events won't be misinterpreted after an input
source switch.
https://bugzilla.gnome.org/show_bug.cgi?id=697007
As with the screen recorder, the magnifier already adds its own
copy of the system cursor, so we should not add it again. Just
as in the screen recorder case, we don't address the case where
the cursor should not be included in the screenshot, but the
magnifier adds it anyway.
https://bugzilla.gnome.org/show_bug.cgi?id=700488
The magnifier adds its own copy of the system cursor to apply the
expected transformations, so we don't need to add it again in the
recorder; this avoids two different cursors showing up in recordings,
but doesn't address the case where the cursor should not be recorded
at all, but the magnifier adds it anyway.
https://bugzilla.gnome.org/show_bug.cgi?id=700488
It looks a bit unpolished to overlap our own chrome with the recording
icon, which may happen when an existing adds UI at the bottom edge.
Fix this by using the primary monitor's workarea for the position rather
than the entire monitor.
https://bugzilla.gnome.org/show_bug.cgi?id=700409
Attaching gdb and running with G_DEBUG=fatal-warnings is not a very
fast to debug a specific issue (especially if you have warnings at
startup, since then you need to run the shell from a terminal).
Instead, introduce a new SHELL_DEBUG environment variable that can
be set to backtrace-warning, causing a gjs_dumpstack() after every
warning or critical.
https://bugzilla.gnome.org/show_bug.cgi?id=700262
Wrap new GtkMenuTracker API that adds an easy way to bind to
tracker items, and use it to add back support for submenus.
This also adds support for a submenu feature that we didn't
have support for before, action namespaces.
https://bugzilla.gnome.org/show_bug.cgi?id=700257
This pulls in new upstream API that Ryan will maintain, removing
code on our side.
Currently, our implementation of submenus will be gone, but this
will be fixed in a few commits.
https://bugzilla.gnome.org/show_bug.cgi?id=700257
We'll need some of these pieces to be introspectable when we port to
GtkMenuTrackerItem. Due to technical limitations in introspection, we
can't put Gtk-prefixed items in the shell namespace, so add them to
a new introspection library instead.
https://bugzilla.gnome.org/show_bug.cgi?id=700257
This is a hack we have in our local fork as compared to upstream;
work on a generic "hook" system in here is ongoing, but until then,
this is the easiest way to do it.
https://bugzilla.gnome.org/show_bug.cgi?id=700257
This includes a rename from the G* namespace to the Gtk* one, which
will help us with introspecting this code. Note that this removes
some of the custom code we added to GActionMuxer to relay event times
to the remote action group. We'll add this back soon.
https://bugzilla.gnome.org/show_bug.cgi?id=700257
Currently we will always record the entire screen. It has been requested
to support recording a specified area analogous to the screenshot API as
well, so add a set_area() method which allows this.
https://bugzilla.gnome.org/show_bug.cgi?id=696247
It is currently not always possible to predict the actual output filename
of a recording - the file-template does not necessarily use an absolute
path and may contain %d and %t escape sequences.
This is OK for fire-and-forget uses like the existing keyboard shortcut,
but we will soon expose the functionality on DBus and consumers of that
API might very well need to access the file after the recording. So do
the same as our screenshot API and add an optional (out) parameter to
record().
https://bugzilla.gnome.org/show_bug.cgi?id=696247
Some callers of the keyring prompt keep the dialog up while
processing the prompt. Allow the user to cancel the prompt
while in this state.
This is propagated to the caller, who can cancel the operation
in question when this occurs.
https://bugzilla.gnome.org/show_bug.cgi?id=682830
The shadows are currently rendered by painting the actor we want to
apply shadow on, in an offscreen buffer. The problem is that when this
actor has an allocation padding (ie allocation that isn't at 0x0
relatively to its parent), this padding is added within the offscreen
buffer and as a result the shadow rendering is truncated because the
offscreen buffer size is the size of the allocation box, not the
allocation box + padding.
This patch reposition the actor at 0x0 with rendering it by changing
the initial transformation matrix when rendering the actor offscreen.
https://bugzilla.gnome.org/show_bug.cgi?id=698301
This makes it much easier to implement correct popup-menu behavior
in the case of nested bins.
This fixes the context menu key in application search results when a
result has focus.
https://bugzilla.gnome.org/show_bug.cgi?id=699800
The comment clearly intended that for this to be the case, but a typo
prevented this from actually being done. This fixes the focused state
of the search field not working more than once.
https://bugzilla.gnome.org/show_bug.cgi?id=699799
In most cases, we'll transition between two states on hover / focus.
Instead of recalculating and repainting our resources on state change,
simply cache the last state when we transition.
https://bugzilla.gnome.org/show_bug.cgi?id=697274
The background image, background image shadow and border image are
allocation-indepedent, so we can keep these in the node. Given that these are
are likely cached in the StTextureCache, the slight increase in code complexity
may not be worth caching these textures and materials -- we might be better off
just computing when we need to paint.
https://bugzilla.gnome.org/show_bug.cgi?id=697274
This ensures that two widgets sharing the same theme node won't trample
on each other's prerendered materials if the actors are of different
sizes. This also tries to be very careful to share as much as possible
during a transition.
This has the side effect that if a widget changes state a bunch of times,
we won't cache every state. Since we expect that state changes are
infrequent and that most cases we'll be able to use the texture cache
to do most of the heavy lifting, this cost is much more insignificant
than rendering a number of different actors with the same theme node
and different sizes.
https://bugzilla.gnome.org/show_bug.cgi?id=697274
Since we now share theme nodes between, we shouldn't cache the paint state
across all nodes. As a first step towards putting this in the actor, split
out the state into another structure. Keep it in the theme node for now
so that we don't make too many changes in one commit.
It's possible that some of these pieces of drawing state could be shared
between theme nodes. For the sake of simplicity, assume that none of them
are shared or should be shared. A future commit could identify those that
could be shared and move them back into the theme node.
https://bugzilla.gnome.org/show_bug.cgi?id=697274
We want to put the paint state in the actor rather than in the theme
node, as having two actors with different sizes but the same theme node
is now much less efficient.
https://bugzilla.gnome.org/show_bug.cgi?id=697274
This simplifies the code required to build remote menus and
put all the items in the right place, and makes us share our
implementation with GTK+.
https://bugzilla.gnome.org/show_bug.cgi?id=698427
Similar to the existing generic getter methods, add lookup functions
for URL properties like the standard background-image/border-image
properties.
https://bugzilla.gnome.org/show_bug.cgi?id=693688
The changes from commit b4f5f1e461 and b394d184cc increased the
instructions required for the fade fragment shader. This is over the limit
for some hardware (like intel gen3), which causes the driver to fallback
to software rendering for the shader. The result is that painting a scrollview
that has a fade effect takes around 30 (!!) seconds.
So lets go back to the old effect for 3.8 until we find a solution.
https://bugzilla.gnome.org/show_bug.cgi?id=696404
Previously when a client requests that a window should be docked the
shell would reparent the socket window onto the stage's window and
then use ClutterX11TexturePixmap to get a texture to represent the
window. This will not work if Clutter is no longer using the X11
winsys for example if it becomes its own display server. Instead this
patch leaves the socket window as a child of the root window and lets
mutter create a MetaWindow out of it. If Mutter is acting as a display
server then this mechanism will still work via the headless x server.
The ShellGtkEmbed instance now registers for notification of the
‘window-created’ signal of the display so that it can find the
MetaWindow that gets created to represent the socket window. When this
window is found it is prevented from being displayed on the screen by
setting the actor's opacity to 0. An input shape is then set on the
window to prevent it receiving any input.
Instead of being a subclass of ClutterX11TexturePixmap, ShellGtkEmbed
is now a subclass of ClutterClone. When the MetaWindow is found for
the socket window the clone's source is set to the invisible actor for
the window so it can be displayed in the panel as before.
The ShellEmbeddedWindow no longer needs to know what the stage is
because it no longer reparents the socket window. Therefore the
ShellTrayManager doesn't need to know the stage either so
shell_tray_manager_manage_stage has been replaced with just
shell_tray_manager_manage_screen.
https://bugzilla.gnome.org/show_bug.cgi?id=693438
Pointer tracking is broken when the pointer is over the stage input
area. This is apparently fallout from mutter going to XInput2.
This commit changes the mouse event handling code to also use XInput2.
https://bugzilla.gnome.org/show_bug.cgi?id=695324
Curently it is possible to copy the content of password entries,
and paste it elsewhere in clear text. This is undesirable, so
follow GTK+'s behavior and disable the cut/copy actions for
password entries.
https://bugzilla.gnome.org/show_bug.cgi?id=695104
Doing so causes useless full stage redraws and breaks culling
as clutter cannot know how the signal handler affects painting.
So use clutter_threads_add_repaint_func_full instead.
https://bugzilla.gnome.org/show_bug.cgi?id=694988
Since commit b4f5f1e461, the effect is eased in at the scroll
view's edges so that it does not appear out of nowhere. However,
the linear easing used is not the best option, as now the effect
appears so late that content near the edges ends up just being
cut off rather than faded out.
So adjust the easing function to have the effect appear faster.
https://bugzilla.gnome.org/show_bug.cgi?id=694327
g_str_has_prefix() will assert when passed NULL, so we need to make
sure that we are passed a non-NULL log_domain first.
Spotted by <goughost<at>yahoo.com.cn>
https://bugzilla.gnome.org/show_bug.cgi?id=663601
If enabled, scrollbars take away from the allocation given to the
view's content. This is usually preferrable to painting the bars on
top of the content, but there are exceptions, for instance when the
content needs to be centered with regard to the view as a whole.
Add a :overlay-scrollbars property to account for those cases.
https://bugzilla.gnome.org/show_bug.cgi?id=694261
This commit updates the code to use mutter's new background
api, and changes the shell's startup animation to be closer
to the mockups.
Based on initial work by Giovanni Campagna
https://bugzilla.gnome.org/show_bug.cgi?id=682429
We cannot reset the cursor at the next leave event, as that might
happen on a NULL stage and cause a BadWindow error, so do it on
unmap (which is guaranteed to happen before the stage is cleared).
https://bugzilla.gnome.org/show_bug.cgi?id=694057
As we use g_slist_prepend for efficiency when building the list
of results before ordering it, we need to make sure we traverse
the list of previous results backwards so that it's built in
the same order.
https://bugzilla.gnome.org/show_bug.cgi?id=693936
GdkPixbuf is stored in memory as individual bytes, in strict red-green-blue-alpha
order, while cairo image surfaces are stored as 32bits units whose
order depends on the endianess.
It is probably possible to do something better, taking advantage
of cogl and the GL using the actual component order, but for now
it is easier to use the GDK utility to convert the cairo surface.
https://bugzilla.gnome.org/show_bug.cgi?id=693931
While we were relying on gtk_icon_info_load_icon and friends being
thread-safe, there was no such guarantee, and recent caching that
was added to GTK+ made it non-threadsafe. To replace it, _async()
variants of the icon loading code were added that are thread-safe.
Use those instead of using our own worker threads.
https://bugzilla.gnome.org/show_bug.cgi?id=692845
Use GdkPixbuf rather then cairo_surface_write_to_png_stream when saving
screenshots because this allows us to embedded metadata into the file which
is used by the background control panel to filter out screenshots.
https://bugzilla.gnome.org/show_bug.cgi?id=693737
The notifications spec has two hints for playing a sound, sound-file
and sound-name. We can support them using the existing code that
wraps libcanberra.
https://bugzilla.gnome.org/show_bug.cgi?id=642831
We have to mark the cairo_surface dirty after modifying the data behind
cairo's back.
Also use CAIRO_FORMAT_ARGB32 rather then CAIRO_FORMAT_RGB24 for the
surface to be consistent with the rest of the code.
https://bugzilla.gnome.org/show_bug.cgi?id=691715
Instead of using Clutter to add an event filter for X events it now
uses the GDK API. The Clutter API won't work if Clutter is not using
an X11-based backend such as if Mutter is directly running with the
KMS backend. This is a step towards making Mutter be its own display
server and a step towards being a Wayland compositor. In this case GDK
will still be using the X backend because it will connect to the
headless X server.
https://bugzilla.gnome.org/show_bug.cgi?id=693438
As pressure barriers need a signalling mechanism to provide
information about when and where they are hit, an object which
provides a signal is a more appropriate abstraction for a pointer
barrier than a functional ID-based approach. Mutter has gained
pointer barrier wrappers, so use its objects instead of ours.
https://bugzilla.gnome.org/show_bug.cgi?id=677215
Before version 1.2 of GLSL it would not implicitly convert from int to
float which meant that if you compare a float variable with an integer
constant it will generate a compile error. In particular this means
that on GLES2 (which uses GLSL 1.0) the scroll view shader will not
compile on pedantic compilers, which includes Mesa. This patch just
changes it to use floating point constants.
https://bugzilla.gnome.org/show_bug.cgi?id=693339
Instead of directly using symbols from GLX to check for the swap event
notification, the plugin now first verifies that the Cogl renderer is
actually using the GLX winsys and then indirectly fetches the pointers
for the GLX functions using cogl_get_proc_address. That way it will
continue to work if Cogl is using an EGL winsys.
Nothing in the Gnome Shell plugin now directly uses symbols from libGL
so we don't need to link to it. This helps to avoid problems linking
against two GL APIs when cogl is using a non-GL driver such as GLES2.
https://bugzilla.gnome.org/show_bug.cgi?id=693225
Cogl sets this for us since commit 2701b93f159bf2d3387cedf2d06fe921ad5641f3.
Setting it twice is illegal and causes compile failures:
error C0204: version directive must be first statement and may not be repeated.
This commit removes all the code in charge of playing with the database of
mobile providers, which was originally included in order to perform
MCCMNC->OperatorName and SID->OperatorName conversions.
This logic is now exposed by libnm-gtk.
https://bugzilla.gnome.org/show_bug.cgi?id=688943
Environment variables should be checked for zero length,
instead of non-zero. This prevented the script to run
correctly in a jhbuild shell for example.
https://bugzilla.gnome.org/show_bug.cgi?id=693031
Clutter translates keyboard state internally, and clears the lock bits
from modifier state, so translating again results in the wrong keysym.
Given that Clutter already gives us a fine keysym, we don't need this.
https://bugzilla.gnome.org/show_bug.cgi?id=692586
As theme nodes keep a cache of matched properties, we need to make
sure to update it when the list of stylesheets changes. In particular
this fixes a regression from commit dc2ec0a8f9, which caused
extensions with stylesheets to crash the shell when re-enabled (for
instances when coming back from the lock screen).
https://bugzilla.gnome.org/show_bug.cgi?id=692994
StThemeNodes cache matched properties from stylesheets, so when the
list of custom stylesheets changes, the node may miss better matches
(when a stylesheet was added) or have pointers to invalid memory in
the list (when a stylesheet was removed).
In order to allow theme nodes to listen for stylesheet changes, add
an appropriate signal to StTheme.
https://bugzilla.gnome.org/show_bug.cgi?id=692994
gnome-control-center is planning on removing its own tree in the
future. Since it already installs these applications into
/usr/share/applications, just use this for now.
https://bugzilla.gnome.org/show_bug.cgi?id=692483
Since we also support passing a basename now, clients might be
interested in knowing the path used to save the file.
Add an out argument to the interface for that.
https://bugzilla.gnome.org/show_bug.cgi?id=688004
If a non-absolute path is passed to the screenshot methods, treat it as
a basename for the output image, and automatically try to save it in
$XDG_PICTURES_DIR, falling back to $HOME if it doesn't exist.
https://bugzilla.gnome.org/show_bug.cgi?id=688004
According to css3-transition, transition-duration is expressed
as a time, that is, in seconds or milliseconds. Fix that by
recognizing numbers with units and implicitly converting to
milliseconds after parsing.
https://bugzilla.gnome.org/show_bug.cgi?id=681376
The code here before was added as dummy code to satisfy an error
in the missing switch, and wasn't ever tested due to the lack of XI2
in mutter. Use the same math as GtkRange does to calculate scroll bar
positions from raw XI2 deltas to allow for proper smooth scrolling.
https://bugzilla.gnome.org/show_bug.cgi?id=687573
realpath() does a series of lstat() on each path component to resolve
symbolic links, but we just want to get an absolute path, and we don't
really care if it is physical or not. Going through a GFile does the
canonicalization we need, and is a lot faster.
https://bugzilla.gnome.org/show_bug.cgi?id=687881
StWidget considers "same theme node" as an indication that the style
did not change, and skips emitting style-changed in that case. This
means that icon theme changes are not picked up by StIcon.
https://bugzilla.gnome.org/show_bug.cgi?id=689353
Decorations are fairly uncommon in gnome-shell, so it's
worthwhile to avoid effort creating empty attr lists. This
can also help prevent a relayout.
https://bugzilla.gnome.org/show_bug.cgi?id=689400
Coverity rightly complained about the context < 0 comparison.
In addition, context was shadowing the parameter of the same
name, and the loops over the attributes were unnecessarily
running all the way.
https://bugzilla.gnome.org/show_bug.cgi?id=689325
This was due to incorrect pixel clamping, which bounced the height
of the actor between values. Just remove pixel clamping, as Clutter
will correctly do it for us.
https://bugzilla.gnome.org/show_bug.cgi?id=689243
This doesn't (or shouldn't) change the visual appearance of the fade
effect, but does do all the testing math inside the shader, rather
than on the CPU. This will make fading the offset much easier in
the future.
https://bugzilla.gnome.org/show_bug.cgi?id=689249
GLSL 1.20 is a better language, and we'll rely on it in future updates.
This doesn't have any additional constraints, since GLSL 1.20 was
standardized before GLSL-supporting drivers came out.
https://bugzilla.gnome.org/show_bug.cgi?id=689249
Theme nodes are interned and shared with other widgets, so they cannot
be disposed, otherwise we blow useful resources, and in particular we
break the parent-child chain.
https://bugzilla.gnome.org/show_bug.cgi?id=689029
