When the easing state has a duration of zero milliseconds we can skip
the entire create_transition() call inside set_width() and set_height(),
to avoid what may be a costly call to get_preferred_*.
If we update a transition that is currently playing, we need to check
the current easing state, and look at the eventual duration, in case
the user wants to cancel the transition.
Instead of checking the duration of the current easing state we should
check if there's a transition in progress, and update it
unconditionally.
If there is no easing state, or the easing state has a duration of zero
milliseconds, then create_transition() should bail out early and set the
requested final state.
This allows us to write:
clutter_actor_save_easing_state (actor);
clutter_actor_set_x (actor, 200);
clutter_actor_restore_easing_state (actor);
[...]
clutter_actor_set_x (actor, 100);
and have the second set_x() update the easing in progress, instead of
being ignored.
https://bugzilla.gnome.org/show_bug.cgi?id=672945
Commit 80626e7584 removed an
IN_DESTRUCTION check from within the add_child_internal() method,
outlining an option for bringing it back. It was too late for the 1.10
cycle to do it, and eventually pick up the pieces, but now that we're
at the beginning of the 1.11 cycle we can restore it, and add checks
elsewhere to balance it.
Should not have been there in the first place: the animatable will be
set either using ClutterTransition API, or when adding the transition
to a ClutterActor.
When adding a transition to a ClutterActor, the actor should hold a
reference on it, and release it only when we remove it. This makes
transitions just like other objects held by ClutterActor.
While you can get a per-transition notification of completion, it can be
convenient to also have a way to notify that all the transitions
involving an actor are complete. A simple signal triggered by the
removal of the last transition fits the bill pretty neatly.
If restore_easing_state() is called on the last easing state on the
stack, clean up the stack, so that we don't leave stale pointers
around to later segfault on.
When setting the easing mode, duration, or delay without having ever
called clutter_actor_save_easing_state(). It's confusing, and not
really nice.
In the future, we'll have a default easing state implicitly created by
the actor itself, but for the time being explicitly opting in is
preferrable.
Yes, it's not really the proper GL name for a linear-on-every-axis of a
texture plus linear-between-mipmap-levels minification filter, but it
has three redeeming qualities as a name:
- LINEAR_MIPMAP_LINEAR sucks, as it introduces GL concepts like
mipmaps in the API naming, while we're trying to avoid that;
- people using GL already know what 'trilinear' means in this context
without going all Khronos on their asses;
- we're using 2D textures anyway, so 'linear on two axes and linear
between mipmap levels' can be effectively approximated to
'trilinear'.
I mean, if even the OpenGL official wiki says:
Unfortunately, what most people think of as "trilinear" is not linear
filtering of a 3D texture, but what in OpenGL terms is GL_LINEAR mag
filter and GL_LINEAR_MIPMAP_LINEAR in the min filter in a 2D texture.
That is, it is bilinear filtering of each appropriate mipmap level,
and doing a third linear filter between the adjacent mipmap levels.
Hence the term "trilinear".
-- http://www.opengl.org/wiki/Texture
then the horse has already been flogged to death, and I don't intend to
be accused of necrophilia and sadism by flogging it some more.
Prior art: every single GL tutorial in the history of ever;
CoreAnimation's scaling filter enumerations.
If people want to start using 1D or 3D textures they they are probably
going to be using Cogl API directly, and that has the GL naming scheme
for minification and magnification filters anyway.
It's a bit late in the game for changing the emission of the paint
signal with actors that use paint nodes - mostly because we have both
implicit paint nodes (background color, content) and explicit paint
nodes (the paint_node virtual).
When we branch for 1.12 we can revert this change.
The ::paint signal is the old way to paint an actor; the paint_node()
virtual function is the new way. It's still not possible to traverse the
whole scene graph and build a render tree of PaintNode instances, but
with this change we simultaneously cut out the ::paint signal emission
from the critical path for actors that are using the new PaintNode-based
API, and we retain backward compatibility in the interim period between
1.10 and 2.0.
ClutterContent is an interface for creating delegate objects that handle
what an actor is going to paint.
Since they are a newly added type, they only hook into the new PaintNode
based API.
The position and size of the content is controlled in part by the
content's own preferred size, and by the ClutterContentGravity
enumeration.
The ::paint-node virtual inside ClutterActor is what we want people to
use when painting their actors.
Right now, it's a new code path, that gets called while painting; the
paint_node() implementation should only paint the actor itself, and not
its children — they will get their own paint_node() called when needed.
Internally, ClutterActor will automatically create a dummy PaintNode and
paint the background color; then control will be handed out to the
implementation on the class. This is required to maintain compatibility
with the old ::paint signal emission.
Once we are able to get rid of the paint (and pick) sequences, we'll
switch to a fully retained render tree.
As it turns out, we do end up recursing inside the ::paint signal
emission - especially inside the conformance test suite.
This thoroughly sucks - and we'll only be able to fix it properly
when we bump API for 2.0.
ClutterActor should be able to hold all transitions, even the ones that
have been explicitly created.
This will allow to add new transitions types in the future, like the
keyframe-based one, or the transition group.
It should be possible to set up the delay of a transition, but since
we start the Transition instance before returning control to the caller,
we cannot use clutter_actor_get_transition() to do it without something
extra-awkward, like:
transition = clutter_actor_get_transition (actor, "width");
clutter_timeline_stop (transition);
clutter_timeline_set_delay (transition, 1000);
clutter_timeline_start (transition);
for each property involved. It's much easier to add a delay to the
easing state of an actor.
Clutter is meant to be, and I quote from the README, a toolkit:
for creating fast, compelling, portable, and dynamic graphical
user interfaces
and yet the default mode of operation for setting an actor's state on
the scene graph (position, size, opacity, rotation, scaling, depth,
etc.) is *not* dynamic. We assume a static UI, and then animate it.
This is the wrong way to design an API for a toolkit meant to be used to
create animated user interfaces. The default mode of operation should be
to implicitly animate every state transition, and only allow skipping
the animation if the user consciously decides to do so — i.e. the design
tenet of the API should be to make The Right Thing™ by default, and make
it really hard (or even impossible) to do The Wrong Thing™.
So we should identify "animatable" properties, i.e. those properties
that should be implicitly animated by ClutterActor, and use the
animation framework we provide to tween the transitions between the
current state and the desired state; the implicit animation should
happen when setting these properties using the public accessors, and not
through some added functionality. For instance, the following:
clutter_actor_set_position (actor, newX, newY);
should not make the actor jump to the (newX, newY) point; it should
tween the actor's position between the current point and the desired
point.
Since we have to maintain backward compatibility with existing
applications, we still need to mark the transitions explicitly, but we
can be smart about it, and treat transition states as a stack that can
be pushed and popped, e.g.:
clutter_actor_save_easing_state (actor);
clutter_actor_set_easing_duration (actor, 500);
clutter_actor_set_position (actor, newX, newY);
clutter_actor_set_opacity (actor, newOpacity);
clutter_actor_restore_easing_state (actor);
And we can even start stacking animations, e.g.:
clutter_actor_save_easing_state (actor);
clutter_actor_set_easing_duration (actor, 500);
clutter_actor_set_position (actor, newX, newY);
clutter_actor_save_easing_state (actor);
clutter_actor_set_easing_duration (actor, 500);
clutter_actor_set_easing_mode (actor, CLUTTER_LINEAR);
clutter_actor_set_opacity (actor, newOpacity);
clutter_actor_set_depth (actor, newDepth);
clutter_actor_restore_easing_state (actor);
clutter_actor_restore_easing_state (actor);
And so on, and so forth.
The implementation takes advantage of the newly added Transition API,
which uses only ClutterTimeline sub-classes and ClutterInterval, to cut
down the amount of signal emissions and memory management of object
instances; as well of using the ClutterAnimatable interface for custom
properties and interpolation of values.
The ::paint, ::queue-redraw, and ::queue-relayout signals should be
marked as no-recurse and no-hooks; these signals are emitted *a lot*
during each frame, and since GLib has a bunch of optimizations for
signals with no closures, we should try and squeeze every single CPU
cycle we can.
In theory, handlers connected to the ::allocation-changed signal may be
able to modify the actor's real allocation and allocation flags,
especially now that we use STATIC_SCOPE; let's avoid this, so that we
don't regret it later.
The ActorBox passed to the ::allocation-changed signal should be
annotated as STATIC_SCOPE, given that it's a pointer to a structure
inside ClutterActorPrivate - hence there's no risk of it actually being
freed from a signal handler. This allows the GSignal machinery to avoid
a costly copy/free for each signal emission.
If the redraw entry is not cleared, queueing a redraw from a signal
handler could reinsert the same object in the stage redraw list,
causing the segfault later (as the object is immediately freed)
https://bugzilla.gnome.org/show_bug.cgi?id=671173
We currently check for the IN_DESTRUCTION flag inside the
add_child_internal() function.
This check disallows calling methods that change the stacking order
within the destruction sequence, by triggering a critical warning first,
and leaving the actor in an undefined state, which then ends up being
caught by an assertion.
The reproducible sequence is:
- actor gets destroyed;
- another actor, linked to the first, will try to change the
stacking order of the first actor;
- changing the stacking order is a composite operation composed
by the following steps:
1. ref() the child;
2. remove_child_internal(), which removes the reference;
3. add_child_internal(), which adds a reference;
- the state of the actor is not changed between (2) and (3), as
it could be an expensive recomputation;
- if (3) bails out, then the actor is in an undefined state, but
still alive;
- the destruction sequence terminates, but the actor is unparented
while its state indicates being parented instead.
- assertion failure.
The obvious fix would be to decompose each set_child_*_sibling() method
into proper remove_child()/add_child(), with state validation; this may
cause excessive work, though, and trigger a cascade of other bugs in
code that assumes that a change in the stacking order is an atomic
operation.
Another potential fix is to just remove this check here, and let code
doing stacking order changes inside the destruction sequence of an actor
continue doing the work.
The third fix is to silently bail out early from every
set_child_*_sibling() and set_child_at_index() method, and avoid doing
work.
I have a preference for the second solution, since it involves the least
amount of work, and the least amount of code duplication.
See bug: https://bugzilla.gnome.org/show_bug.cgi?id=670647
