Applications have no way of finding out if a session manager proxy was
successfully created in gtk_application_startup_session_dbus(), so it's not
appropriate for certain public GtkApplication functions to be asserting the
presence of a session manager proxy as if it were a programmer error.
This affects:
gtk_application_inhibit()
gtk_application_is_inhibited()
If sm_proxy is NULL, the function should just return silently.
In the case of gtk_application_uninhibit(), the application should only be
calling this if it obtained a valid cookie, which implies the presence of a
session manager proxy. I noted that with a comment.
https://bugzilla.gnome.org/show_bug.cgi?id=701365
We toss a g_warning() if we can't find the user's signal handler. A good chunk
of the time this will be because they didn't use -rdynamic. Add a note about
that.
We rename the gtk_widget_class_bind_template_child{_internal}
macros by appending a _private to their name. Otherwise, it
would be too magic to pass the 'public' names as arguments,
but affect a member of the Private struct. At the same time,
Add two new macros with the old names,
gtk_widget_class_bind_template_child{_internal} that operate
on members of the instance struct.
The macros and functions are inconsistently named, and are not tied to
the "template" concept - to the point that it seems plausible to use
them without setting the template.
The new naming scheme is as follows:
gtk_widget_class_bind_template_child_full
gtk_widget_class_bind_template_callback_full
With the convenience macros:
gtk_widget_class_bind_template_child
gtk_widget_class_bind_template_child_internal
gtk_widget_class_bind_template_callback
https://bugzilla.gnome.org/show_bug.cgi?id=700898https://bugzilla.gnome.org/show_bug.cgi?id=700896
Using an offset from the struct means you can have children in
both the public and private (via G_PRIVATE_OFFSET) parts of the
instance. It also matches the new private macros nicer.
https://bugzilla.gnome.org/show_bug.cgi?id=702563
Signed-off-by: Emmanuele Bassi <ebassi@gnome.org>
Plug windows weren't redrawing properly because the embedded
window was expecting to get messages for each frame from the
compositor, but the compositor doesn't know about embedded
windows. Simply disable frame sync for GtkPlug's GdkWindow -
extending XEMBED to handle frame sync isn't interesting
at this point.
A new API gdk_x11_window_set_frame_sync_enabled() is added
to allow this to be done.
https://bugzilla.gnome.org/show_bug.cgi?id=701613
Add introspection friendly version of gtk_tree_model_rows_reordered
with a "_with_length" suffix. This allows language bindings to
safely pass an array with length argument. Use the "Rename to:"
annotation to expose as "gtk_tree_model_rows_reordered".
https://bugzilla.gnome.org/show_bug.cgi?id=684558
We've recently a number of classes wholly. For these cases,
move the headers and sources to gtk/deprecated/ and adjust
Makefiles and includes accordingly.
Affected classes:
GtkAction
GtkActionGroup
GtkActivatable
GtkIconFactory
GtkImageMenuItem
GtkRadioAction
GtkRecentAction
GtkStock
GtkToggleAction
GtkUIManager
Declare variables at the top to the block.
G_GNUC_BEGIN_IGNORE_DEPRECATIONS and G_GNUC_END_IGNORE_DEPRECATIONS with
a ';' at the end is still considered a statement, even if these macros
expand to nothing, which C89 compilers don't really like (oddly speaking)-
but this is life.
A vertical action area causes the info bar to be imposingly large.
This defeats the whole point of the info bar which is to be
unobtrusive. Not to mention it is impossible to make it look
and feel good.
When the icon-release signal is emitted on a GtkSearchEntry, the
contents is now cleared only if it's for the secondary icon. The primary
icon can be used for another purpose.
https://bugzilla.gnome.org/show_bug.cgi?id=704164
This feature offers a number of benefits related to providing
feedback to the user when the password is masked. Some experts have
argued that password masking is harmful. I tend to agree with this
setting providing a better and more moderate solution. Some agree:
https://www.schneier.com/blog/archives/2009/07/the_pros_and_co.html
In order to further lessen the impact I've only enabled the feature
on the primary display since the likelyhood of a non-primary display
being visible by others is higher.
Fixed the documentation to clarify that gtk_recent_info_get_visited
and gtk_recent_info_get_modified are actually about the recent info
meta-data rather than the resource itself.
https://bugzilla.gnome.org/show_bug.cgi?id=703827
In add_preedit_attrs, don't free foreground/background colors already
set in the underlying text attributes (style). They will be free'd by
release_style.
https://bugzilla.gnome.org/show_bug.cgi?id=703533
Other code assumes that the widget has a window if it is realized.
Since we might trigger such code indirectly from gtk_window_realize,
don't mark the window as realized before we've registered its window.
Since we now do all drawing propagation on the cairo_t (rather than
exposing multiple independent times on the GdkWindows) we no longer
need the opacity 0.999 hack.
We render the source into a cairo_surface_t so that we can render it
with cairo directly, rather than having to convert it from a pixbuf
every time. We also specify the target window when creating the cairo
surface so that rendering can be faster.
Using cairo surfaces also allows us to seamlessly support window scales.
We also add a GTK_IMAGE_SURFACE source type.
Support scales when falling back to loading icons from the
icon theme.
In order to actually render scaled icons we add
gtk_icon_set_render_icon_surface which renders to a cairo_surface_t
which includes whatever scaling you need for scaled icons.
This draws an icon from a cairo_surface. We want to use this more rather
than render_icon as this means we can skip the pixbuf to surface
conversion (including allocation and alpha premultiplication) at
render time, plus we can use create_similar_image which may allow
faster rendering.
An optional OutputScale integer key has been added to index.theme
subdirs description, so icon themes may provide icons that are
more suitable to render at a (typically 2x) integer upscaled
resolution. This way it is possible to make eg. a 16x16@2x icon has a
real size of 32x32, but contains a similar level of detail to the
16x16 icon so things don't look any more cluttered on high-dpi
screens.
The pixbuf lookup has changed so it prefers a minimal scale change
that yields the minimal real size difference, so if looking up for
a 16x16 icon at 2x, it would first prefer 16x16@2x, then 32x32, and
then any other icon that's closest to match
There is now *_for_scale() variants for all GtkIconTheme ways
to directly or indirectly fetch a GdkPixbuf.
This is based on code by Carlos Garnacho with changes by Alexander
Larsson
We need to be able to compute different GtkCssImage values
depending on the scale, and we need this at compute time so that
we don't need to read any images other than the scale in used (to
e.g. calculate the image size). GtkStyleProviderPrivate is shared
for all style contexts, so its not right.
This is very useful for hidpi where the dpi is scaled to make
non-dpi aware apps larger. In that case a dpi aware gtk+ using
GDK_SCALE will be getting huge fonts. You can the set GDK_DPI_SCALE
to compensate for this.
Group the buttons on each side in a box, and give them different
style classes, so themes can differentiate (this is partially a
workaround for limitations in our css selectors). The boxes also
let us add a margin below the buttons, without affecting the
allocation of the buttons themselves.
We have to uglify the deprecation messages to make gtk-doc
handle them. The get turned into XML attributes, and there
is no automatic quoting to protect quotes in the text.
If a menu is opened and it doesn't fit entirely below or above
the menu bar, gtk+ will place it on top. The button release will
then activate the popup item that happens to appear under the
cursor. Avoid this by ignoring release events if they originated
in the parent menu bar and the duration of the press was too short.
https://bugzilla.gnome.org/show_bug.cgi?id=703069
Removing object properties is too much of an API break, even for
properties we don't expect to be used outside of GTK itself.
This reverts commit 8b811b623c.
When creating separators we were binding the "label" property on the
tracker to the "label" property on the GtkSeparatorMenuItem.
This was problematic for two reasons.
First, it was pointless. The section header label will never change.
Second, it was causing problems: doing the binding caused the value to
be initially synced up, even if it was NULL. Doing this caused
GtkMenuItem to create a GtkAccelLabel and add it as a child, which
prevented the separator from being shown normally.
Change the code a bit so that we just call gtk_menu_item_set_label()
when creating the item, if we find the label to be non-NULL.
Also, show() the separator item at first. GtkMenu manages visibility of
separators internally, but it seems "more correct" to show it ourselves
at first.
This invariant stopped being guaranteed when we moved the visible area
validation from a high priority idle to a tick callback.
Fixes redrawing bugs like row expanding sometimes not having any visual
effect.
When the search is case sensitive, g_utf8_normalize() is not called, so
the skip_decomp argument of the function forward_chars_with_skipping()
must be FALSE.
To verify that, when searching "éb", the count parameter of
forward_chars_with_skipping() have a different value:
- case sensitive: count = 2
- case insensitive: count = 3 (g_utf8_normalize() has been called)
The commit adds unit tests that didn't pass before, and that now pass
(without known regression, obviously).
https://bugzilla.gnome.org/show_bug.cgi?id=702977
In the non-submenu case we bind the 'visibility' attribute to the
tracker, which takes care of showing the item. In the submenu case, we
don't bind all of the properties, so we miss this one.
Deal with it by just show()ing the submenu item.
https://bugzilla.gnome.org/show_bug.cgi?id=702332
When the file chooser changes directories, it tells the GtkFileChooserEntry to
change its base folder (i.e. the folder from which relative pathnames are resolved).
GtkFileChooserEntry then starts loading a GtkFileSystemModel asynchronously.
In the finished_loading_cb(), however, it would always ask the GtkEntryCompletion
to insert the completion prefix, since that finished_loading_cb() is what is also used
while the user is typing *in the entry*.
But when the entry doesn't have the focus (e.g. the user changed directories by double-clicking
on the file list in the file chooser), there's no reason to insert completions at all.
https://bugzilla.gnome.org/show_bug.cgi?id=672271
Signed-off-by: Federico Mena Quintero <federico@gnome.org>
The placeholder string is truncated when there's no
room for it in entry's width.
So, by ellipsizing it, the user can notice the text
was truncated so that they can workaround that by,
for example, maximizing the window.
https://bugzilla.gnome.org/show_bug.cgi?id=702763
The function must return TRUE only if there is a tag toggle _after_ the
iter, not _at_ the iter. So for the end iter, the function must always
return FALSE.
Add also unit tests for gtk_text_iter_forward_to_tag_toggle().
https://bugzilla.gnome.org/show_bug.cgi?id=691266
It was not possible to know if the @limit was for @match_start or
@match_end. It was documented for backward_search(), but not for
forward_search().
https://bugzilla.gnome.org/show_bug.cgi?id=390048
Currently you can only set the current filename, but not
get it. It's useful to be able to get it in save
dialogs, where the user has typed the desired filename
and you are not in a real directory (recent used, for
example).
https://bugzilla.gnome.org/show_bug.cgi?id=702497
By delaying the emission to an idle we avoid a lot of tricky
reentrancy issues. For instance, a normal gtk_icon_theme_choose_icon()
call could in very rare cases (when a user updated an icon theme) emit
a signal which could affect the icon currently being looked up. This
kind of reentrancy is very hard to test against, especially when it is
so rare, so we're better of avoiding it.
There is no real value to get the change signal directly anyway. All
it can do is affect which icon is rendered the next frame, and we will
handle the queued emission before rendering. Not to mention that icon
theme change detection is polled anyway, so it is already delayed.
https://bugzilla.gnome.org/show_bug.cgi?id=694755
When we're reloading the theme in ensure_valid_themes (due to noticing
that a theme dir has changed) we need to also clear the icon cache
as it will not be valid for the new theme.
We already do this in do_theme_change(), but ensure_valid_themes()
was missing this.
https://bugzilla.gnome.org/show_bug.cgi?id=702598
The icon data in GttkIconInfo->data is currently owned by the
IconThemeDir->icon_data hashtable. However, on e.g. a theme change
blow_themes() destroys the dirs and thus the data, meaning any
outstanding GtkIconInfo points to stale data.
We solve this by adding a refcount to GtkIconData and reffing it
from GtkIconInfo.
https://bugzilla.gnome.org/show_bug.cgi?id=702598
GtkBuilder returns GError for _add_from_{file,resource,string}(),
implying that the user should be able to recover from these errors.
Mention in the docs that it's unreasonable to try to do this.
https://bugzilla.gnome.org/show_bug.cgi?id=679930
There's probably an expectation from the app developers that
Ctrl+F or similar common shortcuts for starting a search will
be handled by the widget, when this is not the case. Explicitely
mention it in the docs to avoid confusion.
https://bugzilla.gnome.org/show_bug.cgi?id=701523
We already take care to not track headers when the listbox is hidden,
because that can be very expensive during creation, so we might as
well skip reseparate() too. The only time we don't want to skip it
is when we're about to show the list.
This was added from https://bugzilla.gnome.org/show_bug.cgi?id=691979
where it was wanted in order to get a "all children filtered" kind
of row. However, the solution is not really generic enough, because
filtering can happen for other reasons than a full refilter (i.e. due
to a row change), and the partial fix you can get by counting children
after refilter can be achieved by manually doing it whenever you manually
called refilter anyway.
Its possible we should have some better solution for the "empty state"
though.
Historically the following states propagated to children:
GTK_STATE_FLAG_ACTIVE
GTK_STATE_FLAG_PRELIGHT
GTK_STATE_FLAG_SELECTED
GTK_STATE_FLAG_INSENSITIVE
GTK_STATE_FLAG_INCONSISTENT
GTK_STATE_FLAG_BACKDROP
However, several of these are problematic on containers like GtkListBox.
For instance, if a row is ACTIVE or SELECTED then all children (like e.g
a button) inside the row will *also* look active/selected. This is almost
never right. The right way to theme this is to catch e.g. SELECTED on the
container itself and set e.g. the color and let the children inherit
the color instead of the flag.
We now propagate only these flags:
GTK_STATE_FLAG_INSENSITIVE
GTK_STATE_FLAG_BACKDROP
Which make sense to be recursive as they really affect every widget
inside the container.
However, this is a CSS theme break, and while most things continue working
as-is some themes may need minor tweaks.
Now that we automatically set up the focus_vadjustment we only
need the adjustment to handle the page-up/down scroll length correctly.
We use parent_set to automatically pick up the adjustment from a
scrollable (i.e. viewport) parent which makes add_to_scrolled useless, so
it can be removed.
When adding with a viewport we automatically set the focus
adjustments on the viewport from the scrolled window, so that
when any child widget gets focused we automatically scroll to it.
This is generally nice, but its particularly important for GtkListBox
where focus changing is how we navigate between rows.
We also ensure that the adjustments are always set before adding the
child to the viewport, which we will need later to pick up the
adjustments on add.
Unhighlight rows if they are removed.
Drop the automatic drag motion based auto-scrolling. This is not really
right, as it e.g. hardcodes auto-scroll regions and assumes we're the only
widget inside the scrolled container. Its also imho in the wrong place,
autoscrolling if any should be done on the scrolled-window or viewport.
All the removed code is implementable outside Gtk+ though.
Use g_string_assign to avoid issues with assigning GString's internal
buffer back to the same string. This can happen when no translations are
available and _gtk_builder_parser_translate returns back the same
pointer that was passed in.
This fixes a regression from commit e9f182e37a
that caused GtkComboBoxText <items> from GtkBuilder to show up empty if
no translations are available.
https://bugzilla.gnome.org/show_bug.cgi?id=700629
-Avoid defining variables in the middle of the block
-Include fallback-c89.c in gtkrevealer.c in place of math.h as
fallback-c89.c includes math.h itself, and is needed for round()
Do not unref mount has we grabbed its pointer without taking
a reference.
Otherwise the following call to g_list_free_full on network_mounts
frees it a second time. Segfault in nautilus ensue.
https://bugzilla.gnome.org/show_bug.cgi?id=700979
gtk_widget_get_allocated_height returns the adjusted
allocation, so we can't use it as measure for how much
room to leave when drawing the window background.
The intention of the expand widget flags was to in fact propagate
legacy expand child properties as well. Due to implementation errors,
this was only happening in some cases, but not in others. To avoid
breaking old code assuming traditional expand flag behavior, this
commit removes all attempts to propagate GtkBox::expand.
This was discussed in
https://bugzilla.gnome.org/show_bug.cgi?id=698656 and
https://bugzilla.gnome.org/show_bug.cgi?id=628902
This patch was written by Tristan Van Berkom.
Focusing the text entry without selecting all the text is needed in
some places (GtkTreeView, and some uses of GtkSearchEntry) so
create a private helper to avoid replicating the hacks.
https://bugzilla.gnome.org/show_bug.cgi?id=700787
Work around this by introspecting gtkclipboard.c and gtkdnd.c instead
of the quartz alternatives.
Note that this is temporary: The implementation of GdkSelection
will make the quartz alternatives unnecessary. See bug 571582.
Mouse events that we do not handle should bubble up to the parent
widget, so they can be handled there, instead of disappearing inside
the button. Also use GDK_EVENT_{STOP,PROPAGATE} to make return
values clearer.
https://bugzilla.gnome.org/show_bug.cgi?id=696640
Instead of storing the rect in the bin window, store the row and column
the editable belongs to and compute the rect lazily. This way, we don't
need to keep the rect up to date.
Fixes /TreeView/scrolling/new-row-mixed/path-500 test.
Emit the "changed" signal after 150 msecs, so that searching
through big lists, or doing online searches feels more responsive.
This is something already done in various applications to make
search-as-you type more responsive (gnome-shell, gnome-documents,
gnome-control-center, etc.). The 150 msecs is the value currently
used by gnome-shell, so keep it (invisibly) consistent.
https://bugzilla.gnome.org/show_bug.cgi?id=700229
Now that we're not drawing the GdkWindows by themeselves we need to
propagate into children in the same order as the windows were painted
otherwise apps can't rely on stacking order to overlay children.
This is still not 100% the same as the old behaviour, because we're
treating all windows that are part of a window at the same time (to allow
e.g. opacity groups), and we're only looking at order for the main
windows of a widget. However, this fixes at least the ordering
of the gnome-boxes fullscreen overlay toolbar.
https://bugzilla.gnome.org/show_bug.cgi?id=699970
Instead of making clients inspect the submenu action and decide what
to do based upon that, always request the submenu open and let the
tracker decide what to do.
Add a new class, GtkMenuTrackerItem that represents a menu item, to be
used with GtkMenuTracker.
GtkMenuTracker's insert callback now works in terms of this new type
(instead of passing reference to the model and an index to the item).
GtkMenuShell now handles all of the binding tasks internally, mostly
through the use of property bindings. Having bindings for the label and
visibility attributes, in partiular, will help with supporting upcoming
extensions to GMenuModel.
GtkModelMenu has been reduced to a helper class that has nothing to do
with GMenuModel. It represents something closer to an "ideal" API for
GtkMenuItem if we didn't have compatibility concerns (eg: not emitting
"activate" when setting toggle state, no separate subclasses per menu
item type, supporting icons, etc.) Improvements to GtkMenuItem could
eventually shrink the size of this class or remove the need for it
entirely.
Some GtkActionHelper functionality has been duplicated in
GtkMenuTracker, which is suboptimal. The duplication exists so that
other codebases (such as Unity and gnome-shell) can reuse the
GtkMenuTracker code, whereas GtkActionHelper is very much tied to
GtkWidget. Supporting binding arbitrary GtkWidgets to actions vs.
supporting the full range of GMenuModel features for menu items turns
out to be two overlapping but not entirely similar problems. Some of
the duplication (such as roles) can be removed from GtkActionHelper once
Gtk's internal Mac OS menubar support is ported to GtkMenuTracker.
The intent to reuse the code outside of Gtk is also the reason for the
unusual treatment of the enum type introduced in this comment.
This adds no new "public" API to the Gtk library, other than types that
we cannot make private due to GType limitations.
Rename our internal GActionMuxer, GActionObserver and GActionObservable
classes and interfaces to have names in our own namespace.
These classes were originally intended for GIO but turned out to be too
special-purpose to be useful there, so we never made them public API but
have just been copying them around (without bothering to properly rename
them). Now that other people will be copying them out of Gtk, it's even
more important to prevent this namespace abuse from spreading further.
We were using gtk_menu_item_get_label() from a testcase to determine the
label associated with a menu item. Future changes to GtkModelMenuItem
will cause this to stop working, so try a bit harder to find a label
inside of the item.
GtkApplicationWindow would only update its list of captured accels
when realizing the window. This meant that keyboard shortcuts added
after the window was realised (for example, added by plugins) would
be non-functional.
Solve this by updating our accels every time the accel map changes,
not only when realizing the window.
https://bugzilla.gnome.org/show_bug.cgi?id=700079
Items that act on the selection should not be sensitive if the
selection is empty. This was already the case for the 'Copy file
location' item, but not for the 'Add bookmark' and 'Visit file'
items.
https://bugzilla.gnome.org/show_bug.cgi?id=699104
Usually, educated GtkContainers' forall() implementation returns children
in an order that's safe for the default draw() implementation in GtkContainer.
So for widgets with some stacking notions (eg. GtkOverlay),
_gtk_widget_find_at_coords() needs to recurse within containers in reverse
order so it finds the topmost widget.
As this function is used in both tooltips and DnD code, this improves behavior
of "floating" widgets wrt those two. This could for example be seen in the
"Transparent" GTK+ demo, where dropping text on the entry results on the text
going to the textview.
https://bugzilla.gnome.org/show_bug.cgi?id=699239
Also removed a bogus FIXME comment which might have been true
a long time ago. Clarification: the "domain" attribute specified
in the builder script overrides any domain previously set with
gtk_builder_set_translation_domain(), so the translation of items
here is perfectly correct and does not need to be fixed.
Bring back need_default_size. We need it to preserve this
documented behavior:
The default size of a window only affects the first time a window is
shown; if a window is hidden and re-shown, it will remember the size
it had prior to hiding, rather than using the default size.
With this patch, all of the window sizing tests in
gtk/tests/window pass again.
gtk_widget_preferred_size() is only useful if you want to quickly port a
widget from GTK2 sizing code to GTK3 but does not properly work with
height-for-width as used in GTK. So we don't want to encourage people to
use it. In particular we want people to convert to height-for-width
before adding baseline support to their widgets.
GtkMenu calls gtk_widget_size_allocate on its GtkWindow during
gtk_menu_popup_for_device if the menu has not been realised. This can cause the
allocation of the GtkWindow and the size of the GdkWindow to become out of sync
because a top level GtkWindow does not attempt to re-size the GdkWindow when
its allocation is set.
https://bugzilla.gnome.org/show_bug.cgi?id=695120
We should only draw the cross-fade on the bin window, not doing this
was causing us to draw it multiple times using ADD which resulted
in weird colors.
This makes iterating over the children a lot faster, as we're
not doing lots of intra-library calls and type checks. We're still
in some sence O(n^2) since we iterate over each child window for each
widget, but the profiles look much better.
We register an invalidate handler on the bin window to get told
of child widget invalidations, although we manually need to discard
invalidates from the scroll operation.
Additionally we invalidate all of the pixel cache whenever
the TreeView itself is queue_draw()n to handle e.g. style (bg)
changes, or changes due to model changes causing queue_draw() in
the tree view.
Since gdk_window_move() no longer uses XCopyArea all scrolling
now re-renders everything in the window. To get performance
back we use a GtkPixelCache to store already drawn children,
and we when we expose the viewport we just blit the
offscreen to the right place.
GtkPixelCache is a helper utility that lets you implement
faster scrolling of a viewport of a canvas by using an
offscreen pixmap cache.
You call _gtk_pixel_cache_draw with a callback function that
does the drawing, and additionally you specify the size and the
position of the viewport in the widget, and the position and size
of the canvas wrt the viewport. The callback will be called to
draw on an offscreen surface, and the surface will be drawn
on the window. The next time you do the same, any already drawn
pieces of the surface are re-used from the offscreen and need
not be rendered again.
If things inside the canvas change you need to call
_gtk_pixel_cache_invalidate to tell the cache about this.
Some other details:
* The offscreen surface is generally a bit larger than
the viewport, so scrolling a small amount can often
be done without redrawing children.
* If the canvas is not larger than the viewport no
offscreen surface is used.
GtkPixelCache: Make sure we always copy using SOURCE
We were using OVER for the first copy (from source to group surface.
GtkPixelCache: Fix x/y typos
GtkPixelCache: Allow NULL for invalidate region
gtkpixelcache: Use CONTENT_COLOR for solid bg windows
Since widgets now cache drawn state we allow them to override
queue_draw_region to detect when some region of the widget
should be redrawn. For instance, if a widget draws the
background color in a pixel cache we will need to invalidate
that when the style context changes which queues a repaint.
We do the save/restore when emitting ::draw rather than in a custom
marshaller, as this saves an extra stack frame, which is helpfull now
that we do painting fully recursive. This is also likely to save a few
cycles.
We now consider non-native windows non-opaque, which means any invalid
area in a subwindow will also be invalid all the way up to the nearest
native windows. We take advantage of this by ignoring all expose events
on non-native windows (which typically means just the toplevel) and instead
propagating down the draw() calls to children directly via
gtk_container_propagate_draw.
This is nice as it means we always draw widgets the same way, and it
will let us do some interesting ways in the future.
We also clean up the GtkWidget opacity handling as we can now always
rely on the draing happening via cairo.
We can't really just draw by walking down the widget hierarchy, as
this doesn't get the clipping right (so e.g. widgets doing cairo_paint
may draw outside the expected gdkwindow subarea) nor does it let
us paint window backgrounds.
So, we now do multiple draws for each widget, once for each GdkWindow,
although we still do it on the same base cairo_t that we get for the
toplevel native window. The difference is only the clipping, the rendering
order, and which other widgets we propagate into.
We also collect all the windows of a widget so we can expose them inside
the same opacity group if needed.
NOTE: This change neuters gtk_widget_set_double_buffered for
widgets without native windows. Its impossible to disable
the double buffering in this model.
