While each stage has at most a GdkFrameClock, the same GdkFrameClock
instance may drive multiple stages per frame. This means that the
mapping between a GdkFrameClock and a ClutterStage is a 1:M one, not a
1:1.
We should store a list of stages associated to each frame clock
instance, so that we can iterate over it when we need to update the
stages.
This commit fixes redraws of applications using multiple stages,
especially when using clutter-gtk.
_cally_actor_get_top_level_origin() uses a compile time check
without runtime check, which will obviously fail when another
backend like wayland is used.
https://bugzilla.gnome.org/show_bug.cgi?id=746575
When multiple relative motion events are received and queued, we can't
base the relative => absolute motion conversion off of the stage pointer
position, since that is only updated when the queue is processed at
the beginning of each frame. The effect of trying to use the stage
pointer position was that subsequent motion events were effectively
dropped.
To improve things, switch to keeping track of the pointer position
ourselves in the evdev backend and adding to that.
This has the side effect of making the internal function
_clutter_input_device_set_coords not effect the internal coordinate
state of the evdev stage, but AFAICS there is nothing depending on that
so that should be fine.
https://bugzilla.gnome.org/show_bug.cgi?id=746328
There's no point in trying to go ahead if we don't have a context to
create the CoglOnscreen framebuffer, and Cogl will crash anyway if we
pass NULL to the constructor.
With server-side buffer allocation, buffers may be returned out of order
(e.g. they may be held onto by external references or hardware). As such
we may see older buffers the frame after we discard the history from
seeing a very young buffer. To overcome this we want to keep the history
in a ring so we can keep track of older entries without keeping an
unbounded list. After converting to a ring, the maximum buffer age
observed during testing was 5 (expected value of 4), but before we could
see ages as high as 9 due to the huge latency spikes caused by doing full
buffer redraws (compounded by external listeners doing readback on the
damaged areas, for example vnc, drm/udl, prime). For this reason, a
maximum age of 16 was chosen to be suitably large enough to prevent these
worst cases from taxing the system.
v2: Fix off-by-one in combining the damage histroy into the clipping
rectangle, and apply copious whitespace fixes.
Bugzilla: https://bugzilla.gnome.org/show_bug.cgi?id=745512
References: https://bugzilla.gnome.org/show_bug.cgi?id=724788
References: https://bugzilla.gnome.org/show_bug.cgi?id=669122
cogl provides an interface to pass along damage with the swap buffers
request. This is useful for the display servers and hardware to minimise
the work done in updating the screen and also reducing the work done by
external listeners (such as vnc, drm/udl and PRIME).
Bugzilla: https://bugzilla.gnome.org/show_bug.cgi?id=745512
If the rectangle is allocate a size smaller than the border, drawing the
border will end up with negative coordinates, and will mess up the whole
thing. Since rectangles don't have a minimum preferred size, we cannot
rely on the allocation being big enough to contain the border and the
background color.
If the rectangle is smaller than the border width value, we just paint
the border color as well.
Point to ClutterImage, ClutterContent, and ClutterActor where needed, so
that developers trying to port their code will have a chance at figuring
out how.
Nobody has been compiling Clutter with profiling enabled in a long time.
UProf itself hasn't been updated in 5 years, and it still depends on
deprecated components like dbus-glib, with no port to GDBus in sight.
The profiling code was moderately useful in the past, but these days
it's probably better to profile Cogl than Clutter itself; timing
information can be extracted by the timestamp on each diagnostic message
that is now available by default in the CLUTTER_NOTE macro, and we can
add ad hoc counters where needed.
Add clutter_x11_set_use_stereo_stage() that can be called
before clutter_init() so that the CoglDisplay we create and all
stages created from that CoglDisplay will be created with a
stereo fbconfig.
This is done in clutter-x11 because of the similarity to the
existing clutter_x11_set_use_argb_visual(), and because it's
not clear without other examples whether the need to have
stereo enabled from before clutter_init() is universal or
somethign specific to GLX.
Cogl required version is increased to 1.20, which has the
required API.
https://bugzilla.gnome.org/show_bug.cgi?id=732706
Added support for Mir, now clutter can natively draw on MirSurfaces.
This depends on latest cogl git.
Run your clutter apps using CLUTTER_BACKEND=mir
Signed-off-by: Emmanuele Bassi <ebassi@gnome.org>
Toolkits may need to paint actors internally outside the normal tree
(for example to create a shadow shape), in which case they need to
control the opacity directly.
https://bugzilla.gnome.org/show_bug.cgi?id=677412
Signed-off-by: Emmanuele Bassi <ebassi@gnome.org>
The _clutter_stage_update_state() function is currently putting events
into the Clutter event queue. This leads to problems in the gdk
backend because there are assumptions upon the numbers of queued
events, and how many of them should be moved from the main event queue
to the ClutterStages' event queues.
This change triggers the processing of the state update events on the
stage directly, so the main event queue retains the expected number of
events.
https://bugzilla.gnome.org/show_bug.cgi?id=744604
If gdk event retrieval has been disabled and gdk's backend is wayland
we must also disable cogl's wayland event dispatching otherwise cogl
will try to dispatch wayland events itself which blocks the main
loop.
https://bugzilla.gnome.org/show_bug.cgi?id=744058
Pointer button events will be received from a device where a button has
been pressed, even though an equivalent button has been pressed (same
button code) on a device connected to the same seat. notify_button()
expects to only be called as if there was only one pointer device
associated with the given seat, so to achieve this, ignore every event
where forwarding it would result in multiple 'pressed' or 'released'
notifications.
https://bugzilla.gnome.org/show_bug.cgi?id=743615
Keyboard key events will be received from a device where a key has
been pressed, even though an equivalent key has been pressed (same
key code) on a device connected to the same seat. notify_key()
expects to only be called as if there was only one keyboard device
associated with the given seat, so to achieve this, ignore every event
where forwarding it would result in multiple 'pressed' or 'released'
notifications.
https://bugzilla.gnome.org/show_bug.cgi?id=743615
We are checking for Cairo ≥ 1.12 and then do an additional check for the
existence of the cairo_surface_set_device_scale() function because there
were no stable releases of Cairo with it. Now that Cairo 1.14 is out, we
can simply bump up the dependency.
libinput's API changed from separate scroll events for vert/horiz scrolling to
a single event that contains both axes if they changed.
Updated by Armin K. to use the discrete axis value for wheel events as done
in Weston.
https://bugzilla.gnome.org/show_bug.cgi?id=742829
We tried once with commit 398a7ac7 and ended up reverting because of
regressions in the input layer and on Wayland.
We should try again, now that those regressions have been fixed.
https://bugzilla.gnome.org/show_bug.cgi?id=734587
The easing state is part of the AnimationInfo structure, which is stored
inside the GObject's datalist. Each instance frees the data stored there
during finalization, so there is no point for us to restore an easing
state (which may or may not be the last one) just to have everything
cleared out once we chain up to GObject's own finalize() implementation.
Bind the preferred size of an actor using a BindConstraint to the
preferred size of the source of the constraint, depending on the
coordinate of the constraint.
Constraints can only update an existing allocation, which means they
live only halfway through the layout management system used by Clutter;
this limitation makes it impossible, for instance, to query the
preferred size of an actor, if the actor is only using constraints to
manage its own size.
And move the only private ClutterConstraint method to it.
This commit also sneaks in a change that makes sense for the debugging
of the update_allocation() method, which checks if the allocation was
effectively changed.
The core pointer concept doesn't really exist anymore in an XI2 world,
so the clutter API is a bit of a mismatch with what X provides. Using
XIGetClientPointer doesn't really help, as far as i can tell the
semantics of XIGetClientPointer are essentially: Whatever the X server
picked when it had to reply with device-dependant data to a query
without a device specifier. Not very useful...
To make matters worse, whether XIGetClientPointer returns a valid
pointer depends on whether there has been a query that forced it to pick
one in the first place, making the whole thing pretty non-deterministic.
This patch changes things around such that instead of using
XIGetClientPointer to determine the core pointer, we simply pick the
first master pointer device. In practise this will essentially always
be the X virtual core pointer.
https://bugzilla.gnome.org/show_bug.cgi?id=729462
XIGetClientPointer() may return the device id '0' when called early.
This patch makes pointer cursors work in nested mutter Wayland
sessions again.
https://bugzilla.gnome.org/show_bug.cgi?id=729462
The installed header should not have private API declarations and
macros. Let's move those into the uninstalled clutter-backend-osx.h
header file instead.
When merging multiple definitions it's possible that the ObjectInfo
fields may get overwritten. Instead of trampling over the fields, we
should reset them only when they actually change — especially the
"is_actor" one, which controls the destruction of the objects when
unmerging happens.
https://bugzilla.gnome.org/show_bug.cgi?id=669743
We want to recompute the content box when changing the content instance,
in case the preferred size is different and the content gravity uses the
preferred size; the change of content with different preferred size and
same gravity should also trigger an implicit transition.
https://bugzilla.gnome.org/show_bug.cgi?id=711182
Some actors want to have a preferred size driven by their content, not
by their children or by their fixed size.
In order to achieve that, we can extend the ClutterRequestMode
enumeration so that clutter_actor_get_preferred_size() defers to the
ClutterContent's own preferred size.
https://bugzilla.gnome.org/show_bug.cgi?id=676326
An easing mode can be set on a frame of a KeyframeTransition.
However, the progress value of the current frame is computed using using
a linear function.
This patch adds a call to clutter_easing_for_mode() to compute
the actual progress value.
Note that parametrized easing modes (bezier and 'step') are not taken
into account.
https://bugzilla.gnome.org/show_bug.cgi?id=740997
If a touchpad is not multitouch, or does not report MT axes (eg. through
the libinput driver), resort to name matching before falling back to
CLUTTER_POINTER_DEVICE.
https://bugzilla.gnome.org/show_bug.cgi?id=741350
In keyboard/mouse wireless combos, it is rather common for the mouse to
claim it contains the multimedia keys, this makes libinput enable both
the pointer and keyboard capabilities on this device, and Clutter thus
to create a keyboard ClutterInputDevice for it.
Ideally clutter devices should be able to reflect their full capabilities,
or maybe account for the fact that certain events can be sent from
seemingly unexpected device types. But this will bring a somewhat better
behavior on such devices.
https://bugzilla.gnome.org/show_bug.cgi?id=740518
Atlasing is fine for smaller textures, but once they get too large its
downsides outweight the benefits. At worst, the larger texture will end
up inside its own atlas, but at worst it will require copying and/or
resizing of an existing atlas.
The cut-off at 512x512 pixels is a bit arbitrary, and we can change it
at any point; it would be nice if we could get the texture limit from
Cogl, and then use a fraction of that size as the cut-off limit. Sadly,
that's not portable, and it's not guaranteed to work either.
For a variety of complicated reasons, ClutterText currently sets fields
on the PangoContext when creating a layout. This causes ClutterText to
behave somewhat erratically in certain cases, since the PangoContext is
currently shared between all actors.
GTK+ creates a PangoContext for every single GtkWidget, so it seems like
we should do the same here.
Move the private code that was previously in clutter-main.c into
clutter-actor.c and clean it up a bit. This gives every actor its own
PangoContext it can mutilate whenever it wants, at its heart's content.
https://bugzilla.gnome.org/show_bug.cgi?id=739050
clutter_set_font_flags already calls clutter_backend_set_font_options,
which emits a signal which our PangoContext listens to, so this is just
duplicate and unneeded code.
https://bugzilla.gnome.org/show_bug.cgi?id=739050
libinput_suspend() will trigger the removal of input devices, but also
the emission of button/key releases pairing everything that is pressed
at that moment. These events are queued, but the ClutterInputDevice
pointers in these will point to invalid memory at the time these are
processed.
Fix this by flushing the event queue, in order to ensure there are no
unprocessed input events after libinput_suspend().
https://bugzilla.gnome.org/show_bug.cgi?id=738520
Just like unmapped children.
Apparently, layers above Clutter allow mapped children without an
allocation, instead of unmapping them. This means we need to ignore
them when computing the paint volume.
Patch originally by: Adel Gadllah <adel.gadllah@gmail.com>
Signed-off by: Emmanuele Bassi <ebassi@gnome.org>
https://bugzilla.gnome.org/show_bug.cgi?id=736682
To support sub-pixel motion events coming from relative events, the
fraction part needs to be stored in the input device state as well. To
do this, simply change the current type from gint to gfloat.
https://bugzilla.gnome.org/show_bug.cgi?id=736413
If gtk_init is called after clutter_init, it can override clutter
AtkUtil implementation. In that situation, we can't say that
the accessibility is enabled, as the root object would be wrong.
In order to provide a way to prevent this:
* clutter_get_accessibility_enabled returns true of false
depending on the current AtkUtil implemented
* cally_accessibility_init always override AtkUtil implementation.
The get_threshold_tigger_egde() method was renamed to fix the typo, but
it obviously broke the ABI. To be fair, nobody in the whole of Debian
was using the symbol, apparently, so it's not like we broke existing
code. Still, it's not nice to break ABI without bumping soname, so let's
put the old symbol back in — obviously, deprecated — as a wrapper to the
newly added one.
If the device manager is created and queried for the client pointer at
a very early stage in application lifetime, the device_id returned would
be 0 as the server hasn't apparently decided yet about the client pointer.
For these situations, doing XSync prior to fetching the client pointer
gets the server to device about the client pointer before we query it.
https://bugzilla.gnome.org/show_bug.cgi?id=735388
When an application sets the scaling factor manually we should mark it as fixed
and not override it when the xsettings change. This matches GDKs behaviour.
In order for this to work we cannot use the same path when setting the value
internally so introduce a _clutter_settings_set_property_internal and use it
for that.
https://bugzilla.gnome.org/show_bug.cgi?id=735244
We need to release the temporary reference we acquired in order for the
signal emission to work.
https://bugzilla.gnome.org/show_bug.cgi?id=734761
Signed-off-by: Emmanuele Bassi <ebassi@gnome.org>
This reverts commit 398a7ac713.
We cannot really use the GDK backend without massive regressions inside
the input layer, like touch events and gestures. The GDK backend is not
entirely up to scratch, and it's late in the cycle.
Let's land this early in 3.15, and get it up to par with X11.
Quite a few people at Guadec complained of pinpoint being broken in
speaker+fullscreen mode, with slides being half displayed. It turns
out that the X11 backend of clutter was being used and this backend
assumes the size of the current monitor is the size of the X screen
(that's not the case with multiple monitors).
To work around the shortcomings of the X11 backend we should probably
position the GDK one before. GDK implements most of the logic the
ClutterStage needs and is probably more tested.
https://bugzilla.gnome.org/show_bug.cgi?id=734587
To get correct premultiplied opacity on a canvas content, white needs
to be assigned to the color that is passed to the texture node. The
content will be very dark for lower opacity values otherwise.
https://bugzilla.gnome.org/show_bug.cgi?id=733385
xkb_state creation has been refactored out of clutter_evdev_set_keyboard_map(),
and used too in clutter_evdev_reclaim_devices(), so the xkb_state is fresh
clean after input is paused/resumed (and keyboard state possibly changed in
between)
https://bugzilla.gnome.org/show_bug.cgi?id=733562
There may be odd situations where full gesture cancellation may be wanted at once
when the first touch is lifted and ::gesture-end is emitted on a gesture action.
Although calling clutter_gesture_action_cancel() within the ::gesture-end handler
causes 2 critical warnings that are otherwise harmless.
https://bugzilla.gnome.org/show_bug.cgi?id=732907
clutter_stage_set_paint_callback() has the disadvantage that it only
works for a single caller, and subsequent callers will overwrite and
break previous callers. Replace it with an ::after-paint signal that is
emitted at the same point - after all painting for the stage is
completed but before the drawing is presented to the screen.
https://bugzilla.gnome.org/show_bug.cgi?id=732342
CLUTTER_ENTER/LEAVE might be processed too, leading to accounting of the
NULL sequence (ie. pointer) in the gesture, and fooling the gesture with
a static extra point that wouldn't go away.
https://bugzilla.gnome.org/show_bug.cgi?id=732235
Until now, touch events sort of rely on XI_Enter/XI_Leave events accompanying
the pointer emulating touch in order to have a stage set on the device, These
events won't happen though if it's not a pointer emulating touch which happens
on the stage, causing touch events to be ignored.
Fix this by ensuring that the input device has a stage on XI_TouchBegin itself,
but only if it's not already set, so we don't possibly steal touch events to
an already interacting stage.
https://bugzilla.gnome.org/show_bug.cgi?id=732234
On X11 the pointer will follow a "pointer emulating" touch sequence, so the
pointer will be effectively left inside the stage after that touch is lifted,
even though the master device stage is unset. This makes pointer events get
ignored until the pointer leaves and enters again the stage.
https://bugzilla.gnome.org/show_bug.cgi?id=732234
The 'state' field should be used for pointer events without button
information. Pointer events that have button information should use
the 'button' field.
https://bugzilla.gnome.org/show_bug.cgi?id=732143
The coordinates translated by the XI2 device manager were being clamped using
the X window size kept by StageX11. However, when the stage is fullscreen,
that size is not updated to the screen size, but kept the same in order to
allow going back to it when the stage goes out of fullscreen.
https://bugzilla.gnome.org/show_bug.cgi?id=731268
The vector of libinput and Wayland pointer axis events are in pointer
motion coordinate space. To convert to clutter's internal representation
the vectors need to be scaled to Xi2 scroll steps.
https://bugzilla.gnome.org/show_bug.cgi?id=723560
And get CLUTTER_EVENT_LEAVE out of the touch event compression logic, as
touches are always implicitly grabbed. If no sequence check is done, only
the last touch update would be emitted, even if multiple sequences got
updated.
https://bugzilla.gnome.org/show_bug.cgi?id=730577
Let's cross fingers and hope nobody notices. If this went unnoticed so far, likely
means this function has never been used. If any complain is raised about this, a
stub function should be added (and marked deprecated).
Those are translated into CLUTTER_TOUCH_* ClutterEvents. As the
"NULL" ClutterEventSequence is special cased, the slot=0 value is
avoided.
Frame events are ignored, as there is no Clutter equivalence, and
Cancel events are sent to all current individual touches.
https://bugzilla.gnome.org/show_bug.cgi?id=728968
And ensure the core pointer shares the same stage than the slave
device when those events are set. This fixes problems on the evdev
backend where the last touch unsets the stage on the device, but
nothing sets it back afterwards.
https://bugzilla.gnome.org/show_bug.cgi?id=728968
When an actor carrying canvas content is repainted, it will currently reupload
the data from the buffer to a texture. While this is not a performance problem
on a desktop, some mobile environments take a big performance hit. This
change tracks data changes and only recreates the texture if necessary.
https://bugzilla.gnome.org/show_bug.cgi?id=729144
By creating and starting the timer on clutter_main() an assumption is made
that that is how the main loop will be run for all clutter applications.
With more and more applications moving to GApplication, this assumption no
longer holds true.
Moving to clutter_init() means we are starting the timer earlier than we
should, and by not stopping it when the main loop quits we are taking a
measure that is later than we should. I believe it is safe to consider
those are close enough to the actual beginning and quitting of the main
loop in practice.
https://bugzilla.gnome.org/show_bug.cgi?id=728521
Instead of just bailing out when initializing the test suite, we can do
a much better job and skip all the tests. This means that the TAP driver
will work correctly instead of dying a horrible death, and we get a nice
report with a proper cause of the test skipping.
Without a paint() implementation in clutter-stage, the function
from clutter-group is used. That class has its own child list,
but attempts to use sort_depth_order, which is empty in this case.
This provides a partial fix by replacing a minimal paint(), see:
https://bugzilla.gnome.org/show_bug.cgi?id=711645
Commit e70a0109 simplified the dispatching of events by passing the event's
owernership to ClutterStage, but it may be so that any.stage is NULL at
some point on Windows, which will either cause _clutter_stage_queue_event()
to crash or issue a critical warning. Avoid this problem by checking
whether event->any.stage is not NULL before trying to call
_clutter_stage_queue_event().
https://bugzilla.gnome.org/show_bug.cgi?id=726765
...so that its entries will reflect the entries that are checked by the
autotools builds on config.h.in. Also take into consideration for MinGW
builds and for newer Visual Studio versions, such as the availability for
inttypes.h. Update the layout of the file cosmetic-wise as well.
Clutter, like GTK+ and GLib, has recently switched to a visibility-based
method of exporting symbols, so update the Visual Studio build files to
do likewise, by using __declspec (dllexport). This eliminats the need to
use a .def file to export the symbols. The pre-configured
config.h.win32.in is also updated accordingly for this purpose. The
clutter.symbols file can be dropped if it is not being used otherwise.
Instead of listing every public symbol inside an ancillary file, we can
use compiler annotations. This scheme is also used by GLib and GTK+.
The symbols file is left in tree until the Visual Studio rules are
fixed, but it's not used any more during distcheck.
I double-checked that the exposed ABI is the same before and after this
change, except for symbols that were never meant to be public in the
first place, and that escaped our attention when we generated the first
version of the symbols file.
All backends follow the same pattern of queueing events first in
ClutterMainContext, then copying them to a ClutterStage queue and
immediately free them. Instead, we can just pass ownership of events
directly to ClutterStage thus avoiding the allocation and copy in
between.
https://bugzilla.gnome.org/show_bug.cgi?id=711857
Keep track of the button modifier mask state in
ClutterInputDeviceWayland and push its state to new button events going
out.
Signed-off-by: Jonas Ådahl <jadahl@gmail.com>
https://bugzilla.gnome.org/show_bug.cgi?id=708781
We do not strictly require the 'swap-region' Cogl feature in order to use
clipped redraws: they work equally well with just the 'buffer-age' Cogl
feature.
https://bugzilla.gnome.org/show_bug.cgi?id=726313
_clutter_stage_window_can_clip_redraws is used to check for clipped redraws
support but can_clip_redraws is not implemented by clutter-stage-wayland so
it always returns FALSE causing full screen redraws.
Fix that by implementing can_clip_redraws in clutter-stage-wayland.
https://bugzilla.gnome.org/show_bug.cgi?id=726315
There could be times when we may not necessarily see a device appear
at initialization time, like when we're VT switched away when we
initialize, and thus we can't ever rely on a main seat appearing.
Always create a main seat with logical pointer/keyboard devices, and
tie the first physical seat that comes in to the main seat.
https://bugzilla.gnome.org/show_bug.cgi?id=726199
We're going to create the main seat at an earlier time, when
we don't have the physical libinput_seat yet, so we need to
do the association later.
https://bugzilla.gnome.org/show_bug.cgi?id=726199
The GetSystemMetrics() function returns wrong values for SM_CXSIZEFRAME,
SM_CYSIZEFRAME, SM_CXFIXEDFRAME and SM_CYFIXEDFRAME when built with Visual
Studio 2012 and 2013 (unless the XP compatibility setting for the
PlatformToolset entry is turned on), causing the window of Clutter programs
to automatically shrink to a point where they become unusable.
This patch uses AdjustWindowRectEx() for builds using Visual Studio 2012
and later, which deduces the required height and width of the Window
properly. Unfortunately we can't use this for the VS 2008/2010 builds as
they cause the Window to continually expand as the program is run.
https://bugzilla.gnome.org/show_bug.cgi?id=725873
We should use the Xkb API to query the direction of the key map,
depending on the group. To get a valid result we need to go over
the Unicode equivalents of the key symbols for each group, so we
should cache the result.
The code used to query and cache the key map direction is taken
from GDK.
https://bugzilla.gnome.org/show_bug.cgi?id=705779
We should set the direction on the PangoContext when creating a
PangoLayout based on a best effort between the contents of the text
itself and the text direction of the widget, in case that fails.
https://bugzilla.gnome.org/show_bug.cgi?id=705779
Currently clutter_device_manager_xi2_get_core_device always
does a round trip to query the client.
So avoid that by caching the client pointer and only update it when the
xi devices change.
https://bugzilla.gnome.org/show_bug.cgi?id=725561
ClutterInputDevice's default initial coordinates is (-1, -1) and since
they're updated from events in a relative way it means that the
pointer can go outside the stage right from the first event.
We usually let this up to higher layers to fix through the pointer
constraint callback but that doesn't work if the first event doesn't
put the pointer immediately inside the stage.
https://bugzilla.gnome.org/show_bug.cgi?id=725103
Doing so is unlikely to work reliably. Instead, switching the keymap
should be done at a time when no key is currently pressed down, but
let's leave that task to higher level code.
This allows us to remove key state tracking at yet another level in
the stack since higher level code likely already tracks this for other
purposes.
https://bugzilla.gnome.org/show_bug.cgi?id=725102