This guarantees we only create 1 extra style cascade in total for hidpi
and not one per style context.
Style cascades are now nested like this:
GtkSettings root cascade (scale == 1)
|
+-- GtkSettings per scale cascade (for any scale, no custom providers)
|
+-- GtkStyleContext custom cascade (for any scale, custom providers)
This requires a bunch of care when changing cascade-related properties
inside GtkStyleContext, so that it ends up with a properly setup
cascade, but I think I got those cases right.
The only thing we don't do yet is reverting to a GtkSettings cascade
when the last custom provider is removed from a custom cascade.
Instead of asserting, just print a g_warning() and try to work around
the problem.
I hope that g_warning() isn't too spammy for people that are hit with
it.
Also clarify the docs that not restore()ing after a save() is a bad
idea.
https://bugzilla.gnome.org/show_bug.cgi?id=743101
When creating the query path, explicitly specify if it's for a root node
or for a child. Relying on gtk_style_context_is_saved() is unreliable
(for example when updating the cache).
This gets rid of a bunch of awkward transitions. It's not a perfect
solution to the problem of "should we transition from this state" but it
gets rid of the ugliest offenders.
If a gtk_style_context_invalidate() is called on a widget's style
context (which nobody should ever do, sheesh!) and we're animating, stop
the animations.
Fixes crashers in Nautilus.
We need to clear the cache manually on full revalidates because
_gtk_css_change_for_child() will clear the full revalidation flags.
And then gtk_style_context_update_cache() will not do the right thing
(which is to clear itself).
Because we can switch from animating to non-animating pretty much
anywhere, do the check for animations unconditionally instead of trying
to cram it into the correct if path (and failing).
We now cache the results of lookups on the parent GtkCssStyle. This
allows sharing styles between widgets (recursively). However, this
only works if the styles can't potentially depend on siblings -
neither directly via sibling selectors or via position pseudo-classes
like :first-child.
Unfortunately, Adwaita currently uses first-child a lot, and in
particular for labels, which are the most common widgets.
The big benefits of this change are both less CPU - due to not needing
to compute styles again - and less memory usage - due to sharing of
the styles between widgets. Here's some nonscientific numbers I
collected while pondering the usefulness of this patch:
glade glade widget
demo demo factory
runtime styles styles
Adwaita before 19.1s 5800 1150
Adwaita now 18.9s 3800 970
Adwaita hacked now 14.5s 3100 910
simple before 11.3s 5800 1150
simple now 10.8s 1300 590
Adwaita: Adwaita as provided by GTK
Adwaita hacked: Adwaita with the first/last-child for GtkLabel removed
simple: A 250 lines simple GTK theme I use for testing
before: This patch not applied
now: this patch applied
glade demo runtime: Starting glade opening a large file and closing it
glade demo styles: GtkCssStaticStyle objects after opening glade with
the large file as per inspector
widget factory styles: GtkCssStaticStyle objects after startup as per
inspector
Previously we looked at the save/restore state when determining the
parent. This is wrong in the case where we're updating the cache though.
So we now save the parent in the style info.
After the parent changes in commit
3a337156d1 we need to refresh the cached
styles after the current style. After all, they now depend on the base
style.
After 3a337156d1 style lookups still used
the parent context's style as the parent style, even though after a
gtk_style_context_save() the root style of the style context is the
proper parent.
Testcase attached.
Instead of keeping an animated style everywhere, we replace it with the
static style when nothing gets animated.
Apart from making the code cleaner, this gets rid of a bunch of animated
style values that do nothing but wrap a static style.
The only style that is animated is the style of the unsaved primary
node. So there's no need to create animated style objects for the other
ones.
There is a bunch of ugliness in the code currently. Further commits are
expected to fix them.
GtkCssStyle is the base class to be used for all types of styles that do
exist.
GtkCssAnimatedStyle is the only implementation so far, that is exactly a
copy/paste of the old GtkCssStyle code.
After b49c7c3421 we were no longer doing a
full revalidate after GTK_CSS_CHANGE_SOURCE changes.
This fixes spurious failures of widgets not properly updating when
changing the theme.
GtkStyleContext was not properly handling the style cascade when
setting a screen, causing the inspector global CSS to affect the
inspector window, even though the inspector is using a different
screen now.
... and make it the default. This takes over the meaning from "none" for
this property in that it draws the fallback builtin image.
"none" now literally means no image will be drawn.
When a getter function (like get_color()) is called and the passed in
state doesn't match the current state returned via get_state(), we used
to do a trick: We called save()/set_state() on the context before
getting the values.
Unfortunately, since 3a337156d1 this
has the unfortunate side effect that it also creates a child element.
This breaks various old codebases (spinbutton has been fixed in
998feeb2bc, Webkit is fixed in
https://bugs.webkit.org/show_bug.cgi?id=137803 ) unfortunately.
So instead, look up the values manually ensuring that no child element
is created but the correct state is used.
Keeping them is a bad idea now where the widget paths are actually
changed by a save(). And almost all of the time, state or style classes
will be changed anyway.
Looking them up again is just a hash table lookup anyway.
GtkCssNodeDeclaration is a new struct with copy-on-write semantics.
It encapsulated the properties used to define a node in the CSS tree.
The idea is to use it in various places for caching, in particular as
key in hash tables.
This is a change for how CSS is applied.
Previously, subelements (I'll take GtkEntry icons as an example) were
treated as having the same parent as the regular elements. So a selector
such as
.entry
would match an entry inside a window. But it'd also match the icon image
inside the entry. So CSS like
.entry { padding: 10px; }
would add 10px of padding to both the entry itself and to the icon image
inside the entry, so the icon would effective have 20px padding. To get
around that, one would have to unset it again like so:
.entry { padding: 10px; }
.entry.image { padding: unset; }
This is getting more and more of a problem as we make subelements
respect more properties that aren't inherited by default anyway, like
backgrounds and padding/margin/border.
This patch has one caveat though: It makes calling
gtk_style_context_save() the first time have an important side effect.
It's important to audit code to make sure it is only used for
subelements.
And last but not least, this patch is only useful if code unsets
parent's style classes that it doesn't want to apply any longer. Because
style classes are inherited by default (and I don't think we want to
change that), the example will still apply until the subelements no
longer contain the .entry style class.
- gtk_style_context_get_background_color()
- gtk_style_context_get_border_color()
Those functions shouldn't be used anymore, because they don't represent
anything from the CSS styling we support. The background color often
isn't used due to background images and there are actually 4 different
border colors (1 for each side) - if there isn't also a border image in
use.
This was introduced as a hackish way in 3.6 to make font updates
propagate properly. But since then, font handling has been changed and
this flag is no longer necessary.
gtk_style_context_invalidate_internal() will respect only the current
saved state of the style context, which is wrong when updating the scale.
In that case, the whole style context needs updating.
Now that widget paths are allowed to have a state, use that state when
querying style properties. This uses a fast path in gtkcssprovider.c and
that is great.
Don't take a state when constructing the CSS matcher. Instead, rely on
the newly introduced state in the widget path.
This way, the state can be queried not only on the first element, but on
all elements of the widget path.
Set the widget path state flags with the state flags of the style
context.
We do not update the state flags but replace the previous one because we
want to be able to have save()/restore() unset state flags.
When validating the style context, we are copying the animations
from one StyleValues instance to another, and cancel the old ones.
It turns out that sometimes the old and the new StyleValues are
the same, and in this case, we end up cancelling the animations
for good.
One case where breakage from this was observed is that the spinners
in gtk3-widget-factory stop spinning when you open and close a modal
dialog on page 2. This depends a bit on the window manager; the problem
can only be seen if opening the dialog causes a transition to backdrop.
Regions are done in a very non-css way. They don't fit the DOM in that
they don't integrate into the CSS tree and they have very weird matching
behavior in selectors.
So I'm deprecating them now. GtkNotebook and GtkTreeview will continue
to use them and as long as they do, we can't remove the code for it.
But once those are ported it might be safe to remove the code as it will
clean up lots of places in the code by quite a bit.
... from per style data to only existing once per style context. This is
technically an API break because it no longer allows getting different
style properties between save()/restore() pairs, but I don't think this
was ever intended to work that way, as the style property API was to be
used and is used via gtk_widget_get_style().
And it simplifies code a lot.
Add two new icon lookup flags, GTK_ICON_LOOKUP_DIR_LTR and _RTL,
which tell GtkIconTheme to look for icon variants which have a
-ltr or -rtl suffix. GtkIconHelper adds these lookup flags when
looking up icons.
Note that due to the way this overlaps with symbolic icon lookup,
directional variants of symbolic icons must be called -symbolic-rtl, not
-rtl-symbolic.
https://bugzilla.gnome.org/show_bug.cgi?id=729980
disconnect_by_func() is slow, and this becomes particularly evident
when disposing a number of widgets (and their associated style
context) at once, such as when using a language binding which
uses a GC.
https://bugzilla.gnome.org/show_bug.cgi?id=723183
Fixes a tiny typo in commit f51c9d4154
which manifested itself in GtkSpinButton's panels being drawn with an
incorrect, not updated state.
This patch took me more hours than you might think! :P
https://bugzilla.gnome.org/show_bug.cgi?id=709491
Commit 719dd636a9 replaces
margin-left/right with margin-start/end. CSS does not have
margin-start/margin-end properties, the sed script was a bit overeager.
Fwiw, CSS implements RTL margin styling via :dir(rtl) selectors.
Resize modes don't work anymore, both because nobody ever uses them and
because the frame clock changed the way things work quite a bit. So we
don't want to advertise them as a good idea.
https://bugzilla.gnome.org/show_bug.cgi?id=708787
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
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.
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.
The following CSS would infloop:
@define-color self @self
as it would infinitely lookup the color named "self" and try to resolve
it. This patch adds detection of such cycles to the resolve function by
keeping a list of currently resolving colors in the cycle_list variable.
... instead of taking the last one we find. This is necessary as
attached widgets (mostly menus) can be attached to an invisible widget,
but we still want to invalidate styles for them.
https://bugzilla.gnome.org/show_bug.cgi?id=695772
We only FORCE_INVALIDATE when something weird changes that the CSS
machinery can't detect. But now that our style_updated functions skip
recomputations when some properties don't change we want to make sure
these recomputations are still run. So we just claim all properties
changed.
http://bugzilla.gnome.org/show_bug.cgi?id=695482
Add an API to start or stop continually updating the frame clock.
This is a slight convenience for applcations and avoids the problem
of getting one more frame run after an animation stops, but the
primary motivation for this is because it looks like we might have
to use timeBeginPeriod()/timeEndPeriod() on Windows to get reasonably
accurate timing, and for that we'll need to know if there is an
animation running.
https://bugzilla.gnome.org/show_bug.cgi?id=693934
Add a very simple GtkWidget function for an "tick" callback, which
is connected to the ::update signal of GdkFrameClock.
Remove:
- GtkTimeline. The consensus is that it is too complex.
- GdkPaintClockTarget. In the rare cases where tick callbacks
aren't sufficient, it's possible to track the
paint clock with ::realize/::unrealize/::hierarchy-changed.
GtkTimeline is kept using ::update directly to allow using a GtkTimeline
with a paint clock but no widget.
Switch GtkStyleContext to using GdkFrameClock. To do this, add a new
UPDATE phase to GdkFrameClock.
Add a GdkFrameClockTarget interface with a single set_clock() method,
and use this to deal with the fact that GtkWidget only has a frame
clock when realized.
https://bugzilla.gnome.org/show_bug.cgi?id=685460
It's a lot uglier now, but it shouldn't crash anymore.
We must update the font description for animations, but we can't free it
on query, because some paths call gtk_style_context_get_font() twice in
a row without stopping the use of the first call. So us just creating a
new font description all the time and unreffing the old one is not a
good idea. So we just mere the new one into the old one.
https://bugzilla.gnome.org/show_bug.cgi?id=691186
This function is just a sophisitcated optimization.
If we know the GDK window's background will be opaque, we mark it as
opaque. This is so GDK can do all the optimizations it does for opaque
windows and be fast.
This is mainly used when scrolling.
The previous code didn't get this right, in particular it didn't enforce
a transparent background when it knew the background was not opaque.
This is for a very simple reason: The getter is returning a const value
and the font isn't const anymore. So we need to store the font
description somewhere but we can't reuse it as it's changing all the
time (yay animations, yay inherited values). Sucks.
So keep the hack in here but deprecate the function.
This is necessary because values in a GtkCssComputedValues can change
now. So if the font-size is inherited or animated, the cached value will
be outdated.
Fixes the fontchooser preview not updating.
Symbolic colors are an implementation detail of the CSS engine and have
been superceded by GtkCssColorValue. We don't want them clobbering the
public API. In particular because the only use I could find in the
public API is people using it to shade colors.
Make _gtk_style_provider_private_get_color() return a GtkCssValue (a
GtkCssColorValue to be exact) instead of GtkSymbolicColor.
With this, the symbolic color usage inside GTK is minimized.
If lookup->missing is empty we don't need to continue looking.
We short circuit in several places as this can happen
after iteratively makeign lookup->missing smaller.
Old code tried to use the "background-image" proeprty for setting the
default image background. While this used to work in the early days of
GTK3, today it is grossly misleading as the backgronud image may be
resized, repositioned and semi-translucent which causes very weird
artifacts when rendering.
So we use the background-color only instead.
Here's the shortest description of the bug I can come up with:
When computing values, we have 3 kinds of dependencies:
(1) other properties ("currentColor" or em values)
(2) inherited properties ("inherit")
(3) generic things from the theme (@keyframes or @define-color)
Previously, we passed the GtkStyleContext as an argument, because it
provided these 3 things using:
(1) _gtk_style_context_peek_property()
(2) _gtk_style_context_peek_property(gtk_style_context_get_parent())
(3) context->priv->cascade
However, this makes it impossible to lookup values other than the ones
accessible via _gtk_style_context_peek_property(). And this is exactly
what we are doing in gtk_style_context_update_cache(). So when the cache
updates encountered case (1), they were looking up the values from the
wrong style data.
So this large patch essentially does nothing but replace the
context argument in all compute functions with new arguments for the 3
cases above:
(1) values
(2) parent_values
(3) provider
We apparently have a lot of computing code.
While regular animations should always be created, transitions should
not. This patch allows to express this by passing NULL as the values to
transition from.
It also adds a gtk_style_context_should_create_transitions() function
that returns TRUE when transitions should be created.
... that actually was both wrong, a performance failure and has been
there since the original checkin.
Updating the cached style data absolutely does not mean clearing all
cached style data first. There's nothing to update then.
This adds the GtkCssAnimation class and the code needed to hook it into
GtkStyleContext. It takes the values out of the CSS "animation"
properties and does animations. See
http://dev.w3.org/csswg/css3-animations/
for details.
Note that the code for starting and stopping animations with widget
visibility doesn't work yet.
... in the case where no change of the DOM tree actually happened.
We don't do anything yet with that information, this patch just
correctly computes it.
See inline comments for what it does. Its main use is figuring out if
something has been caused by GTK's caching of CSS properties or if it's
a different problem.
Instead of using 1 global queue for both resizes and style validation,
use 2 queues. This makes the code a lot simpler and fixes a bug where we
could accidentally stop restylying for very delayed restyles.
We now animate the core style information (see comment in
gtk_style_context_save()). A lot of widgets save + set custom style
classes/states during drawing and so can't be animated. It does work for
labels, menus and buttons though.
This has two goals:
1) Move invalidation code out of a nested if branch. Invalidation is
actually the most important thing this function does.
2) Have the changes bitmask available. It will needed for invalidate
calls to children later.
The design principles were:
- synchronized
If multiple style contexts are animating, they should all do an
animation step at the same time.
- degrades well
Even when there's thousands of style contexts all animating at the same
time, the animation steps don't starve the CPU. This is achieved by
making sure the timeout is really fast. It just sets a bunch of flags.
- no hidden bottlenecks
Turning animatability on or off on a style context is O(1).
So far it is unused.
This is only a small performance boost by itself, but it's necessary
for animations, so we need it.
Benchmark numbers for my Glade benchmark for interested people:
GTK 3.4.0 last commit this commit
Raleigh
real 0m41.879s 0m10.176s 0m9.900s
user 0m41.394s 0m9.895s 0m9.628s
sys 0m0.111s 0m0.096s 0m0.102s
Adwaita (*)
real 0m51.049s 0m13.432s 0m14.848s 0m12.253s
user 0m50.487s 0m13.034s 0m13.218s 0m11.927s
sys 0m0.117s 0m0.151s 0m0.147s 0m0.107s
Ambiance (patched to not use private GTK APIs)
real 0m52.167s 0m13.115s 0m13.117s 0m12.944s
user 0m51.576s 0m12.739s 0m12.768s 0m12.651s
sys 0m0.119s 0m0.137s 0m0.136s 0m0.118s
(*) Adwaita and unico currently use custom properties, and
_gtk_css_value_compare() for custom properties always returns FALSE,
which makes this optimization never trigger. So I modified
_gtk_css_value_compare() to return TRUE for these properties instead and
reran the benchmark. Those are the numbers.
