The new css tree may change the order of selectors (keeping the
same semantics). This affects how the selectors are printed later,
which causes some css parsing tests to not match the references.
Fortunately the order is consistent between runs given the same
css, so we just have to switch around the order in some of the
.ref.css files.
Now we use the selector tree everywhere, so there is no need to
keep around the linear selectors unless we're using them to
verify the tree correctness, so free them.
We add some "artificial" ordering to the otherwise unordered
tree nodes. This means the tree will be the same every time for the
same input. This is good because e.g. tree order affects the
reordering of the simple selectors, which may affect how
css providers are printed, which need to be consistent for
the css tests to work.
When building the tree we generally reorder the selectors inside
the same simple selector in order to pick a good first selector
to balance the tree better. However, some kinds of selectors
can't really be reordered, even thought they are simple.
This is since the matching code for some types handle
the existance of a directly preceeding selector differently:
REGION and ANY selectors look for a DESCENDANT previous
POSITION selector look for a REGION previous
From a set of GtkCssSelectors and the rulesets they match to
we create a large decision tree that lets us efficitently match
against all the rules and return the set of matched rulesets.
The tree is created such that at each level we pick the initial rule[1]
in all the considered selectors for that level and use put the
one that is in most selectors in the node. All selectors matching that
are put in the previous part of the tree.
This returns true if the matcher matches *anything*. We need
to check this later, because such matchers are dangerous in loops
that iterate over all parents/siblings since such loops would not
terminate.
With the previous commit all loads of the same icon will share a single
GtkIconInfo, which typicallty means the pixbuf is shared via Info->pixbuf.
However, atm we don't share symbolic icons, which causes these to be re-read
and re-parsed every time. This is especially bad if the icon is used many times
in some form of list. So, we cache the pixbufs and reuse them.
https://bugzilla.gnome.org/show_bug.cgi?id=689081
In order to avoid loading and keeping around the same icon multiple times
we keep a cache of all outstanding GtkIconInfo objects for a given theme.
Additionally we return to the app not the normal pixbuf from the info,
but rather a proxy copy of it sharing the same data, but no extra
reference. This allows us to track when the app is no longer using
the pixbuf, and we can thus ensure that the GtkIconInfo in the cache
stays around for at least as long as the pixbuf is alive.
When the app unrefs the pixbuf we put the Info on a short LRU list
to keep it alive a bit longer, in case the app needs it in a short
while.
https://bugzilla.gnome.org/show_bug.cgi?id=689081
This was broken since commit b2aaa94 in 2008. Its commit message
clearly states that the intention was to check for GTK_GRAB,
GTK_UNGRAB and STATE_CHANGED. Lets do that, then.
This was found by Coverity.
Both flashing a window and setting the window opacity were using
incorrect declarations for function pointers. They were missing the
WINAPI annotation as defined in windows.h. As a result, the stack
could be corrupted when these functions were invoked.
Fixes https://bugzilla.gnome.org/show_bug.cgi?id=689235
(cherry picked from commit 5637ef1f97)
