We need to properly track if a node needs to propagate invalidation
state information to its children. We didn't do this properly before and
that could lead to us forgetting to invalidate nodes in corner cases.
Do not propagate the TIMESTAMP change through the node tree, as that
causes lots of uneeded markings of nodes as invalid.
Instead, walk the node tree and find the nodes that have a non-static
style and only invalidate timestamps on those.
Only invalidate timestamps if the node is marked as invalid. We overload
the meaning of "invalid" as "tracks timestamps".
While I don't like the way this is written, it is an important
optimization because 95+% of nodes don't animate so timestamps don't
matter to them. But timestamps are invalidated 60x per second.
We don't return a NULL style to mean "no changes" anymore, instead
we check new_style == old_style to mean that.
Make sure the code reflects this, otherwise we'll send
GTK_CSS_CHANGE_PARENT_STYLE invalidations everywhere and screw up
performance.
Now that the widget node recomputes styles on update_style() we can just
call it during validate(). That way, we don't need the widget node to
manually compute its style.
If CSS values are queried from a widget, recompute them if necessary. Do
not emit style-updated until the validation phase however.
This way, we don't run into performance traps when style-update causes
invalidations that cause new style-updated to be emitted.
... and pass it to the API that computes new styles.
A special timestamp of 0 means "please don't animate" and is used when
no frame clock is available for a node.
This is mainly an attempt to merge the update_style() and validte()
vfuncs. Code is not there yet, but that's the idea.
Also, gtk_css_node_set_style() should not be public. And this gets
closer to that goal, too.
After measuring it, I realized the optimization never triggers for
Adwaita and rarely ever triggers for simple themes. So it is not
useful to keep it around.
This allows adding more API for it.
It also includes code that tracks modifications and invalidates siblings
and their positions whenever nodes get added or removed.
If we know the parent's get_path_for_child() implementation is safe to
be used with GtkCssNode because it doesn't do anything special, we do
that. Unfortunately that requires whitelisting vfuncs because the vfunc
is public API so anyone can override it.
Instead, use gtk_widget_get_path() which makes GtkWidget cache the path.
This is a temporary solution until we can get rid of widget paths.
This increases memory usage quite noticably.
The node declaration has the same functionality as
gtk_css_node_declaration_add_to_widget_path(). So instead of using that
function on a path, you can use the original path and the declaration in
a matcher.
So far the vfunc is kinda quirky (the path argument is an out argument
for something you have to free when you're done with the matcher), but
I'm about to change that.
We don't want to add the current classes to the widget path - which
might potentially be different after a gtk_style_context_save() - but
the root node's ones. So what better thing to do than actually using the
root node?
