Setting four properties is more expensive than calling two C functions,
so move to set_origin()/set_size() calls for our ClutterActorBox
handling.
This gets us down to an average time of 2.1 ms spent in vfunc_allocate()
with 20 windows
Part-of: <https://gitlab.gnome.org/GNOME/gnome-shell/-/merge_requests/1743>
We can save a little bit of time in this loop by iterating directly
over the windowSlots array instead of iterating through children and
then performing a search for the windowSlot. This saves more time and we
only spend 2.2 ms instead of 2.3 ms in vfunc_allocate() now.
Part-of: <https://gitlab.gnome.org/GNOME/gnome-shell/-/merge_requests/1743>
This layout manager is used quite often and the time we spend calling
it's allocate and get_preferred_width/heigth functions increases with
every open window.
We can save a lot of precious time during the layout cycle by moving
this layout manager from JS to C, thus avoiding the overhead of
trampolining between C and JS land.
In a measurement where the average time spent in vfunc_allocate() of the
Workspace actor was measured while opening an overview with 20 windows,
the average time spent went down from 3.1 ms to 2.3 ms.
Part-of: <https://gitlab.gnome.org/GNOME/gnome-shell/-/merge_requests/1743>
With the introduction of the transparent panel in the overview, we
started making the panel text/icon color slightly brighter in the
overview and on the lockscreen to ensure best contrast. Now
unfortunately, setting the text color incurred a relayout of the
underlying ClutterText actor (fixed with
https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/1750), and
setting the icon color causes the icon texture to get regenerated
(fixed with
https://gitlab.gnome.org/GNOME/gnome-shell/-/merge_requests/932).
Regenerating the icon texture will replace the icon actor, which also
causes a relayout.
This relayout of the panel has been measured to add at least 1
millisecond (the numbers fluctuated a lot) to about 5 ms it takes to
layout the first frame when showing the overview.
Since https://gitlab.gnome.org/GNOME/gnome-shell/-/merge_requests/932 is
unlikely to land this cycle, this commit proposes a different solution:
Simply don't use a different color for text in the overview. To avoid
issues with contrast in the overview, make the default color slightly
brighter and change it from #ccc to #ddd.
Part-of: <https://gitlab.gnome.org/GNOME/gnome-shell/-/merge_requests/1733>
Right now we always recreate the icon of the appMenu when calling
_sync(). This will relayout the panel everytime we open the overview,
because we call _sync() in that case.
We can easily avoid that by only recreating the icon actor in case the
app that's opened actually changes. This gets us close to doing no more
relayouts of the panel when opening the overview.
Part-of: <https://gitlab.gnome.org/GNOME/gnome-shell/-/merge_requests/1733>
We already set the AppMenuButton to be non-reactive and transparent when
we hide it. Hiding it completely using clutter_actor_hide() will
additionally make it no longer affect layout and thus queue a relayout.
Since we hide the appMenu in the overview and we want to avoid
relayouting the panel when entering and leaving the overview, don't
completely hide the AppMenuButton to avoid queueing this relayout.
Part-of: <https://gitlab.gnome.org/GNOME/gnome-shell/-/merge_requests/1733>
This updates the use of bind_cairo_surface_property for the changes
from d7cb2eeebc. Now this function returns a StImageContent and not
an Actor like the following code expects, so wrap it in a StIcon.
Also the function lost its size argument.
Part-of: <https://gitlab.gnome.org/GNOME/gnome-shell/-/merge_requests/1718>
This is how the code works and how it is already used from the C side.
There would also be no point in keeping this transfer none, because
textures generated from this function will not be shared and are not
kept in the cache.
Part-of: <https://gitlab.gnome.org/GNOME/gnome-shell/-/merge_requests/1718>
This reverts commit 87558efbf1.
The commit did not fix the bug it was supposed to fix, it just
complicated the code. The hopefully correct fix is in the previous
commit.
The point of this commit was to ensure everything gets removed when
bind->source gets removed. This however was already the case since the
signal handler was already connected to bind->source and has a
destructor registered that takes care of everything. And since gobject
destroys its signal handlers before it clears the weak refs, this new
weak ref was effectively never being used.
Part-of: <https://gitlab.gnome.org/GNOME/gnome-shell/-/merge_requests/1718>
If the icon surface of a window got updated and its type or format no
longer match what we expect, st_texture_cache_reset_texture() might
overwrite the existing image in the bind with a new image, while still
keeping the weak ref on the old image. Due to this the old image might
trigger a st_texture_cache_on_pixbuf_notify() after the bind has already
been freed by g_signal_handlers_destroy() in the bind source. While this
usually would remove the weak ref, the weak ref it tries to remove is
on the new image, not the old one. The call to
st_texture_cache_on_pixbuf_notify() then tries to read the already
free'd memory from the bind which causes the cast to G_OBJECT to fail,
resulting in the crash.
Fixes https://gitlab.gnome.org/GNOME/gnome-shell/-/issues/3785
Part-of: <https://gitlab.gnome.org/GNOME/gnome-shell/-/merge_requests/1718>
Changing the workspace of a window causes the window tracker to remove
and add it to the app again. If this happens from within
_shell_app_add_window() before the window has been added to the windows
list, this will cause the check that is supposed to prevent adding the
same window multiple times to fail and the window to be added twice.
The app will then be considered still running after the last window has
been closed. Then when clicking on the corresponding app icon, the shell
would attempt to switch to a NULL workspace for the closed window
instead of starting a new instance, resulting in a crash.
Changing the workspace also needs to happen after increasing the
interesting window count, because otherwise removal of the window by
the window tracker would trigger a uint underflow leading the app to be
considered running with UINT_MAX interesting windows, despite having no
windows, leading to crashes right after launching the app.
Fixes https://gitlab.gnome.org/GNOME/gnome-shell/-/issues/3833
Part-of: <https://gitlab.gnome.org/GNOME/gnome-shell/-/merge_requests/1745>
If the window wasn't in _windowSlots, the index is -1, so the last
element of the array is removed, leading to
https://gitlab.gnome.org/GNOME/gnome-shell/-/issues/3547
I logged the stack trace from removeWindow() and it seems that when you
move the window from the second monitor to another workspace like shown
in https://gitlab.gnome.org/GNOME/gnome-shell/-/issues/3547,
removeWindow() is called twice for it, presumably once for the second
monitor workspace, and then once for the first monitor workspace for
some reason. It is during that call that _windowSlots doesn't contain
the window, so instead the last element (index -1) is removed, leading
to the animation bug.
Part-of: <https://gitlab.gnome.org/GNOME/gnome-shell/-/merge_requests/1727>
We block state updates while the indicator for the active workspace
is animating. To track that, we check whether the scroll-adjustment's
value matches the active workspace index. That works as long as the
adjustment's value changes after the active workspace, but not when
switching workspaces via SwipeTracker which only changes the active
workspace after transitioning to the new scroll value.
To fix that, update the indicator on workspace changes as well.
Part-of: <https://gitlab.gnome.org/GNOME/gnome-shell/-/merge_requests/1716>
Animating items of the iconGrid involves calling a few more C functions,
which is quite slow. Especially calling ClutterActor.set_easing_delay()
is slow because we override that function in JS to adjust for the
animation slow-down factor. So add a small class variable to make sure
we only animate the icons of the grid when we actually need it.
This makes the average time spent in vfunc_allocate() of the iconGrid go
down to about 0.7 ms.
Part-of: <https://gitlab.gnome.org/GNOME/gnome-shell/-/merge_requests/1713>
It's quite slow to access class variables in JS, especially when they're
backed by GObject properties. To avoid accessing them in every iteration
when we're looping through the children of iconGrid, store those values
to another variable and reuse that inside the loop.
This shaves off another 0.2 ms from iconGrids vfunc_allocate(), getting
the average time spent in that function down from 1.3 ms to 1.1 ms.
Part-of: <https://gitlab.gnome.org/GNOME/gnome-shell/-/merge_requests/1713>
Using a preexisting array to iterate over is much faster than iterating
over the actors children using a "for ... of" loop, that's because the
latter calls into C functions to get the next actor all the time.
Also, stop using array deconstruction here, it turned out that this is
extremely expensive. When profiling this part of the code, it turned out
that only 9% of the time spent in _getChildrenMaxSize() is spent calling
the get_preferred_height/width() methods. When not using array
deconstruction, this time increased to 22%, still not great, but a lot
better.
Part-of: <https://gitlab.gnome.org/GNOME/gnome-shell/-/merge_requests/1713>
We need to access the visible children of a page in inside
vfunc_allocate(), and since getting those children is quite slow (it
involves iterating over all the children of the actor) let's avoid that
and cache the array instead.
This reduces average time spent in vfunc_allocate() of the iconGrid from
1.6 ms down to 1.4 ms.
Part-of: <https://gitlab.gnome.org/GNOME/gnome-shell/-/merge_requests/1713>
We call this._getChildrenMaxSize() from the allocate() vfunc of
IconGridLayout. Since the function is quite expensive, it slows the
layout process down a lot, so instead of re-calculating it on every
relayout, cache the max size of children.
This makes the average time spent in vfunc_allocate() of the iconGrid go
down from 2.3 ms to 1.6 ms.
Part-of: <https://gitlab.gnome.org/GNOME/gnome-shell/-/merge_requests/1713>
Instead of always aligning window previews vertically at the bottom of
their row, only do that if we have multiple rows. If there's only a
single row of windows, align every window vertically centered.
This is a very small step towards the new layout for window previews in
the overview, but since the release of 40 is getting nearer and nearer,
changing more is not feasible anymore.
https://gitlab.gnome.org/GNOME/gnome-shell/-/issues/3634
Part-of: <https://gitlab.gnome.org/GNOME/gnome-shell/-/merge_requests/1737>
To ensure the workspace thumbnails are vertically closer to the window
picker than to the search, scale down the wallpapers by a fixed number
of pixels. Using 24 px for this means we'll take of 12 px at the top and
12 px at the bottom of the wallpaper, that's a better strategy than
always scaling it by a fixed factor since it doesn't change with the
monitor size.
Part-of: <https://gitlab.gnome.org/GNOME/gnome-shell/-/merge_requests/1702>
Use a bit less margin at the top and add some margin at the bottom of
the search entry. This should ensure the search entry is closer to the
panel than to the workspaces and avoid associating the workspace
thumbnails with the search instead of the window picker.
Part-of: <https://gitlab.gnome.org/GNOME/gnome-shell/-/merge_requests/1702>
The Dash currently has a horizontal margin and a top margin applied. The
horizontal margin isn't necessary since there already is a horizontal
padding applied, and the top margin is a bit large, since the
ControlsLayout enforces a spacing between the different containers
already, so slightly decrease the margin-top here.
Part-of: <https://gitlab.gnome.org/GNOME/gnome-shell/-/merge_requests/1702>
The back → transparent transition gives it a very bad look when
booting and running the startup animation.
Use the same transition duration hack to ensure that the panel
starts completely transparent.
Part-of: <https://gitlab.gnome.org/GNOME/gnome-shell/-/merge_requests/1678>
There's this little hack that we do to match the panel transition from
transparent to black with leaving the overview via swipe down. The only
problem is that, while the duration of the panel transition itself is
matches, the corners don't, and they get out of sync.
This isn't very noticeable with the swipe gesture, but it'll be much
more prominent when booting straight into the overview.
Bind the 'style' property of the panel to the corners', so that the
transition duration hack applies to all of them.
Part-of: <https://gitlab.gnome.org/GNOME/gnome-shell/-/merge_requests/1678>
The new startup animation consists of rising the Dash from the bottom,
falling the search entry from the ceiling, and going from HIDDEN to
WINDOW_PICKER with an opacity applied.
One little trick from IconGridLayout was added to ControlsManagerLayout,
which is a promises-based wait for allocation. This is required to make
sure that the transformed position of the search entry is valid, which
is only the case right after an allocation.
This animation also ensures that the overview is shown right on startup.
For session modes that do not have an overview, continue using the same
fade + scale animation.
Part-of: <https://gitlab.gnome.org/GNOME/gnome-shell/-/merge_requests/1678>
We want to avoid updating the input region on startup, since it incurs
in roundtrips to the X server, but not workspaces struts, since they
affect the visible clip of wallpapers in the workspace. Since next
commits will make the overview be the after-boot screen, we really
don't want the wallpaper to be clipped wrongly.
Allow updating regions while starting up, but only workspace struts.
Make sure input is not updated by accounting for 'this._startingUp'
on 'wantsInputRegion'.
Part-of: <https://gitlab.gnome.org/GNOME/gnome-shell/-/merge_requests/1678>
This is an X11-specific routine, and building the list of input region rects
on Wayland is a waste, since it incurs in many trampolines only to throw them
in the trash.
Don't build input region rects on Wayland. By modifying the 'wantsInputRegion'
variable, it also skips actors that only update input, which is another small
optimization for Wayland.
Part-of: <https://gitlab.gnome.org/GNOME/gnome-shell/-/merge_requests/1678>
LayoutManager doesn't update struts when there's any modal running. Turns
out, the Overview itself is a modal. That, and the fact that the Overview
will be the startup state, prevents the workarea to be updated.
Allow updating struts when there's no other modal than the Overview.
Part-of: <https://gitlab.gnome.org/GNOME/gnome-shell/-/merge_requests/1678>
Launching applications on a particular workspace works through
launch contexts and startup notifications. While this is no
longer required by a launcher/WM split, in theory this allows
us to reliably identify the correct window to apply startup
properties to.
However in practice we fail more often than not: Missing support in
toolkits, differences between display protocols, D-Bus activation
and single-instance applications all provide their own pitfalls.
So instead, take advantage of the fact that launcher and WM live in
the same process, and go with the unsophisticated approach: Just
remember the last workspace that was requested when launching an
app, then move the next window that is associated with the app to
that workspace.
This will break X11 applications that set an initial workspace, but
that's legacy functionality anyway (given that there's no wayland
protocol for that functionality), and seems a price worth paying
for making launching apps on workspaces more reliable.
Part-of: <https://gitlab.gnome.org/GNOME/gnome-shell/-/merge_requests/1316>
Let the goToPage call afterwards to take precedence, instead
of resetting the adjustment (thus the view) on the side.
This resulted in strange state when the last page contains
a single icon, and it is dragged. The last page being emptied
triggers a pages-changed signal, which half resets the view
to the first page while DnD is ongoing.
Letting goToPage do its business means we neatly clamp to the
closest page to currentPage, the last page in that case.
Part-of: <https://gitlab.gnome.org/GNOME/gnome-shell/-/merge_requests/1630>
