As they require a draw context and the draw context is already bound to
the surface, it makes much more sense and reduces abiguity by moving
these APIs to the draw context.
As a side effect, we simplify GdkSurface APIs to a point where
GdkSurface now does not concern itself with drawing anymore at all,
apart from being the object that creates draw contexts.
Also, split it into its own file - which was the original reason for
looking at this code, the rewrite was an unintentional side effect.
This changes the context to create surfaces on demand.
So whenever the compositor holds onto a surface while GDK wants to
render, it just creates a new surface. If the compositor releases
surfaces, we will retain one for the next frame to be rendered, but free
all extra ones.
This way, we should get to a stage where we have exactly as many
surfaces as needed and never allocate/free any.
Also, don't implement SurfaceClass.ref_cairo_surface() anymore. This
means calls to it will crash now. But as they only happen in the generic
GdkCairoContext implementation, we shouldn't be affected by that.
Plus, once all backends have been ported, that call is going away
anyway.
And make the GdkCairoContext as abstract.
The idea of this and thje following commits is to get rid of all
Cairo code in gdksurface.c (and $backend/gdksurface-$backend.c)
by moving that code into the Cairo context files.
In particular, the GdkSurfaceClass.begin_frame/end_frame()
functions (which are currently exclusively used by the Cairo code
should end up being moved to GdkDrawContextClass.begin/end_frame().
This has multiple benefits:
1. It unifies code between the different drawing contexts.
GL lives in GLContext, Vulkan in VulkanContext and Cairo in
CairoContext. In turn, this makes it way easier to reason about
what's going on in surface-specific code. Currently pretty much
all backends do things wrong when they want to sync to drawing
or to the frame clock.
2. It makes the API of GdkSurface smaller. No drawing code (apart
from creating the contexts) needs to remain.
3. It confines Cairo to the Drawcontext, thereby making it way
more obvious when backends are still using it in situations
where it may now conflict with OpenGL (like when doing the dnd
failed animation or in the APIs that I'm removing in this
branch).
4. We have 2 very different types of Cairo contexts: The X/win32
model, where we have a natively supported Cairo backend but do
double buffering ourselves and use similar surfaces and the
Wayland/Broadway model where we use image surfaces without any
Cairo backend support and have to submit the buffers manually.
By not sharing code between those 2 versions, we can make the
actual code way smaller. We also get around the need to create
1x1 image surfaces in the Wayland backend where we pretend
there's a native Cairo surface.
This does nothing but disallow passing NULL to gdk_surface_begin_paint()
and instead require this context.
The ultimate goal is to split out Cairo drawing into its own source file
so it doesn't clutter up the generic rendering path.
The shortcuts inhibitors hash table is created when we create a
GdkWaylandWindow implementation for a GdkWindow, and it's destroyed once
we finalize the instance. The fake "root" window we create for the
Wayland display does not have a backing native window, so the shortcuts
inhibitors hash table is set to NULL; this causes a critical error
message when calling g_hash_table_destroy() on it. The finalization of
the root window happens when we close a display connection.
We should use g_clear_pointer(), instead, as it's NULL safe.
Without this change, the displayclose test fails, as all warnings are
considered fatal.
We can't freeze the frame clock on commit, but only after-paint,
otherwise the frameclock will resume in the paint stage.
So freeze the frame clock at the end of the frame if we are waiting for
a frame callback.
Note; The diff is only lage because of indentation changes due to
avoiding early returns in favor of a branch.
Instead of going through an ancillary script to strip away the
`WL_EXPORT` annotation from the generated code, we should bump up the
required version of Wayland, and use the `private-code` argument for
wayland-scanner, which does the right thing for us.
* Previous commit had misleading info. The code was
added to begin_paint() instead of end_paint(). Though
that did not affect its performance in any visible way.
* Company advised to move the code to an "after_paint" signal
handler, so that it works on all renderers, not just Cairo.
This change caused high fluctuation in FPS values in fishbowl
when it is put in a situation where it cannot achieve 60fps
(such as using Cairo renderer at ultra-high resolution).
This seems to be deliberate and not a bug.
There is no easily apparent way of being notified when frame updates
happene exactly, so we just query frame info at the end of each paint.
If we query too often (faster than DWM refresh rate), we just get
the same values twice in a row, but that is, hopefully, highly unlikely.
When asked for a nonexistent (positive) monitor number,
gdk_x11_display_get_monitor would (at best) return an uninitialized pointer,
instead of returning NULL.
That way, we can store the right region there: The actual painted area
instead of the exposed area (which is way too small).
Also, the GL context is the only user of this data, so storing it there
seems way smarter.
This error message is printed into the journal if a GTK app can't
connect to eithre Wayland or X11. Make it at least mention who is not
capable of connecting to a server.
Unrelated, we might want to improve our error reporting when a GTK app
can't start, so that debugging issues with system startup / login get
easier to resolve.
Instead of calling gdk_surface_invalidate_region(), just
gdk_surface_queue_expose() and rely on the renderer computing the diff
from the previous rendering.
This ensures that the frame clock gets updated with correct presentation
times even if nothing was drawn.
This is necessary for benchmarking but would also be relevant for videos
that want to sync to the frame clock but draw frames a lot less.
This commit ensures that each GdkSurface impl remembers the
cursor that GDK sets for it, and that this cursor is set
each time WM_SETCURSOR is called for that sufrace's HWND.
This is needed because W32, unlike X, has no per-window cursors -
the cursor on W32 is a global resource, and we need to keep track
of which cursor should be set when pointer is over which surface
ourselves (WM_SETCURSOR exists exactly for this reason).
This commit also makes GDK remember the surface that has an implicit
grab (since implicit grabs are gone from the upper levels of the toolkit),
and ensures that crossing events are correctly synthesized and the grab
is broken when surface focus changes. This fixes a bug where opening
a new window (by clicking something in some other, pre-existing window)
will make that new window not get any mouse input due to the fact
that the mouse-button-down event from that click caused an implicit
grab on the pre-existing window, and that grab was not released afterward.
Instead of now-unused GdkWin32Cursor class (a subclass of GdkCursor),
add a stand-alone GdkWin32HCursor class that is a wrapper around
HCURSOR handle.
On creation it's given a display instance, a HCURSOR handle and a boolean
that indicates whether the HCURSOR handle can or cannot be destroyed
(this depends on how the handle was obtained).
That information is stored in a hash table inside the GdkWin32Display
singleton, each entry of that table has reference count.
When the GdkWin32HCursor object is finalized, it reduces the reference
count on the table entry in the GdkWin32Display. When it's created,
it either adds such an entry or refs an existing one.
This way two pieces of code (or the same piece of code called
multiple times) that independently obtain the same HCURSOR from the OS
will get to different GdkWin32HCursor instances, but GdkWin32Display
will know that both use the same handle.
Once the reference count reaches 0 on the table entry, it is freed
and the handle (if destroyable) is put on the destruction list,
and an idle destruction function is queued.
If the same handle is once again registered for use before the
idle destructior is invoked (this happens, for example, when
an old cursor is destroyed and then replaced with a new one),
the handle gets removed from the destruction list.
The destructor just calls DestroyCursor() on each handle, calling
SetCursor(NULL) before doing that when the handle is in use.
This ensures that SetCursor(NULL) (which will cause cursor to disappear,
which is bad by itself, and which will also cause flickering if the
cursor is set to a non-NULL again shortly afterward)
is almost never called, unless GTK messes up and keeps using a cursor
beyond its lifetime.
This scheme also ensures that non-destructable cursors are not destroyed.
It's also possible to call _gdk_win32_display_hcursor_ref()
and _gdk_win32_display_hcursor_unref() manually instead of creating
GdkWin32HCursor objects, but that is not recommended.
