It was problematic to maintain Raleigh going forward, as any
changes in it affected all themes. Also, its more robust if
each theme is a full standalone css rather than relying on
an inherited css base.
So, this changes Raleigh to a standalone theme that we can tweak
without accidentally breaking other themes, and makes the
default theme empty. In fact, we don't even add the default
provider anymore as its always empty.
The freeing the GResource in gtk_css_provider_reset caused problems
because it was called from gtk_css_provider_load_from_path() inside
gtk_css_provider_get_named(). We fix this by delaying the
setting of priv->resource until we successfully loaded the file.
This lets themes install a .gresource for efficient loading
of resources. This resource file will (if it exists) be automatically
loaded (mmaped) when the theme is used and unloaded when not used.
We don't want to include the whitespace after an import statement.
Because people tend to add newlines after @import, we report the wrong
lines for failed imports. And that's bad.
'window-unfocused' is too long and mentions "focus" which is historically
loaded with the meaning "input focus".
'backdrop' isn't generally used in GUI speak and still conveys the state the
widgets in an unfocused or background toplevel window are in.
We can't print shorthands, so don't try.
In particular, I want to get away from shorthands being representable
using GValue, and this function kinda requires that.
Passing the length of the widget path that is relevant is not necessary
anymore, it was only useful for inheritance. Instead, we now pass the
state flags and let the selector do the state matching for us.
We now use the GtkStleProviderPrivate interface, which hopefully is
faster and more conformant to CSS. Long term, it definitely should be
both.
I would have liked to split this up into multiple commits, but couldn't
find a way.
By default, a background image is stretched. Instead, it is worth to
have a tiled background.
This patch allows background surfaces to be repeated or not, and should
be compatible with future extensions and CSS.
https://bugzilla.gnome.org/show_bug.cgi?id=663522
- add gtkmodulesprivate.h and move stuff there from gtkprivate.h
- add gtkprivate.c and move stuff there from gtkmain.c
- add gtkwin32.c and move stuff there from gtkmain.c
- don't redefine GTK_DATADIR and friends in gtkprivate.h
- have _gtk_get_datadir() and friends on all platforms
- remove the horrid hacks where gtkprivate.h can't be included,
or must be included later due to redefinition of the compile-time
directories
Also create a separate section for values to associate the values with.
Finally, use this section information when delayed-parsing a widget
style property.
This struct keeps track of an area of text in a CSS file and uses it
when specifying information. Also, the cssprovider keeps track of
sections when parsing a file.
... and implement the CSS font properties:
- font-size
- font-style
- font-family
- font-weight
- font-variant
This is the second try at this. The first was backed out previously due
to bugginess. Let's hope this one survives a bit longer.
Also makes the font-family CSS test work again.
It's useful to set a slice size != border-width, as backgrounds are
clipped to border-width too.
As slices can be half-transparent and overlap the background,
this would not fill the border box properly if we only use a single
property for specifying the width.
Also, this brings us even closer to CSS3.
The hack in gtk_style_context_get_font() was causing segfaults in
combobox code. This is not acceptable and I'm not awake enough to fix
it, so just reverting until it's fixed sanely is easiest.
This reverts commit cf6bfbdb17.
... and take an optional style property as argument. This way, we can
allow custom parse functions for properties. The style property needs to
be optional so that we can use it for widget style properties, too.
Name it _gtk_style_property_print_value() and actually pass it the style
property. This way, we can later change it to use custom print functions
for different style properties.
The call to gtk_border_free() within unpack_border() felt completely
in the wrong place, as the border actually pertains to the GValue
being unpacked. Plus, the GValue itself was also being leaked.
In finalize(), clear all rulesets.
In parse_declaration(), Free the GValue under unhandled error situations.
In gtk_css_provider_load_internal(), Do not leak the file contents.
GValues stored in GtkCssRulesets are gslice managed, so don't
g_memdup() GValues from shorthand properties. This fixes
memory corruptions when reloading contents in a GtkCssProvider.
Shorthand properties are basically the same a in CSS. For storage in
style properties or the CSS provider, they are unpacked into the real
values, so it is possible to partially override them.
No properties are yet converted to the new world yet, this is just the
code for supporting them.
This provides a huge speedup as we only need to preprocess style
properties when they are indeed inherited. This roughly doubles the
performance of the CSS matcher and brings the time taken by
gtk_css_provider_get_style() from 19% to 7% in my favorite benchmark.
One for the style properties, one for the widget style properties.
This way we can make one hash table by pspec which means we don't have
to repeat the pspec lookup.
Keep rulesets as an on-stack/heap structure instead of allocating all
instances separately.
Also, pass a ruleset to the ruleset parser, so we can make the ruleset
parser do lots of fancy things that might be useful for performance.
Previously we got the list of all matching rules and then iterated it to
find the first one that had the property. Now we look while matching
rules, so we don't lookup rules that we don't need.
Instead of relying on GScanner and its idea of syntax, code up a parser
that obeys the CSS spec.
This also has the great side effect of reporting correct line numbers
and positions.
Also included is a reorganization of the returned error values. Instead
of error values describing what type of syntax error was returned, the
code just returns SYNTAX_ERROR. Other messages exist for when actual
values don't work or when errors shouldn't be fatal due to backwards
compatibility.
This is pretty important, because otherwise recursions cause crashes.
And if you accidentally change your theme to one that crashes on load,
all your gonna SEGV and then on reboot, gdm tries to load the theme...
Call gtk_css_provider_load_from_file() instead of the internal function.
This has two advantages:
1) It simplifies the code a lot
2) It gets rid of GMappedFile usage. GMappedFile does not work
everywhere, so this is finally portable.
This way, we achieve two things:
1) We can unify file loading to one location
2) We can emit the error from file loading using the parsing-error
signal. This is very useful for @import handling in particular.
Emits the error without the need for a scanner. Also simplifies
gtk_css_provider_take_error() because we now can assert an available
scanner at all times.
Instead of having an error member in the CSS provider's private struct,
connect a signal handler when an error is passed in. This has two
advantages:
1) It makes the code clearer as we don't have to keep track of an error
member anywhere.
2) It causes a non-emission of the g_warning() when an error was passed
in, because it only triggers when no signal handlers are connected.
So we get identical behavior to GTK 3.0 where warnings where only
emitted when no error was passed in.
Instead of aborting a parse whenever we encounter an error, parse to the
end. But if a GError was passed in, reset the provider completely as if
nothing had been parsed.
Value parsing only sometimes emitted errors. Sometimes it didn't emit
errors but ignored the value, sometimes it took a default, sometimes it
converted it to something it deemed suitable.
While refactoring, I moved the whole GValue <=> char * conversion
routines to a separate file, to make navigating the core css provider
easier.
Previously, we only checked for errors after parsing the full
declaration. Now we detect errors with the property before even
attempting to parse its value.
The benefit here is that the error reporting reports the correct line
and position numbers.
