macOS will set the cursor to arrow after milliseconds after the switch
desktop animation ends. So the cursor that was updated immediately after
the focus will be changed again. This also affects toggling fullscreen.
Bypass cocoa's flagsChanged handling and call handleevent ourselves,
this allows us to know exactly what the IME is doing. Hopefully fixes
issues with IMEs that change state on modifier key presses.
Fixes#4541
Apple go on making it harder and harder to support multiple
versions of their pathetic operating system in one code base.
kIOMainPortDefault is the "non-deprecated" replacement for
kIOMasterPortDefault. However, it is now marked as being available only
on macOS 12 and later so it refuses to build in a codebase designed to
run on earlier versions of macOS. Le bubbling sigh.
According to https://developer.apple.com/documentation/iokit/kiomasterportdefault?language=objc
passing NULL has the same effect and avoids the stupid warning.
Hopefully this will work till Apple decides to spend their trillion
dollars doing more stupid shit.
Try to ensure we have a functioning displaylink always.
GLFW skips over sleeping monitors during a poll and also
had a bug where the display link was not re-created for a monitor
that already had a glfw monitor entry.
Also add a bunch more debug reporting
This is caused, as far as I can tell, by CVDisplayLink getting stuck.
Apple apparently are incapable of writing a simple timer robustly.
So if it remains stuck after a second delete and recreate it to force it
to restart.
Fixes#2016
GLFW creates the menu bar in the applicationWillFinishLaunching method, while kitty creates it in `create_os_window()`. This patch changes the behaviour to match GLFW.
In practice, without this change, there can be a short time where the menu bar is not fully populated.
The time is stored in a signed 64 bit integer with nanosecond accuracy. This eliminates the possibility of floating-point inaccuracies.
`monotonic_t` can currently hold values large enough to work correctly for more than 200 years into the future.
Using a typedef instead of directly using `int64_t` everywhere will also allow easily changing the datatype in the future should the need arise for more precise or bigger time values.
Also reduce input latency by ignoring repaint_delay when
there is actual pending input.
Gets rid of request_tick_callback(). Now empty events
result in the tick callback being called so there is only a
single mechanism for waking up the main loop and getting
the tick callback called.
I had added an optimization to not pass messages to
main thread every time the CVDisplayLink timer fired, unless
a render frame for that monitor was actually requested.
However, this optimization is impossible to implement wihtout a deadlock
since CVDisplayLink has its own internal lock that it does not expose.
So I guess macOS users with multiple monitors will simply have to take
the performance hit of useless wakeups sixty times a second for every
extra monitor.
Fixes#1779
performSelectorOnMainThread runs in more loop run modes which means that
the tick callback will behave more like it does on other platforms,
during window resizes and other modal event loops.
