We use raw-window-handle extensive in the public API as well as we
force the users to use it to get some essential data for interop, thus
reexport it.
Fixes: #2913.
Inner panics could make it hard to trouble shoot the issues and for some
users it's not desirable.
The inner panics were left only when they are used to `assert!` during
development.
This reverts commit 9f91bc413fe20618bd7090829832bb074aab15c3 which
reverted the original patch which was merged without a proper review.
Fixes: #500.
Up until now the Android backend has been directly mapping key codes
which essentially just represent the "physical" cap of the key (quoted
since this also related to virtual keyboards).
Since we didn't account for any meta keys either it meant the backend
only supported a 1:1 mapping from key codes, which only covers a tiny
subset of characters. For example you couldn't type a colon since
there's no keycode for that and we didn't try and map Shift+Semicolon
into a colon character.
This has been tricky to support because the `NativeActivity` class doesn't
have direct access to the Java `KeyEvent` object which exposes a more
convenient `getUnicodeChar` API.
It is now possible to query a `KeyCharcterMap` for the device associated
with a `KeyEvent` via the `AndroidApp::device_key_character_map` API
which provides a binding to the SDK `KeyCharacterMap` API in Java:
https://developer.android.com/reference/android/view/KeyCharacterMap
This is effectively what `getUnicodeChar` is implemented based on and is
a bit more general purpose.
`KeyCharacterMap` lets us map a key_code + meta_state from a `KeyEvent`
into either a unicode character or dead key accent that can be combined
with the following key. This mapping is done based on the user's chosen
layout for the keyboard.
To enable support for key character maps the
`AndroidApp::input_events()` API was replaced by
`AndroidApp::input_events_iter()` which returns a (lending) iterator for
events. This was changed because the previous design made it difficult
to allow other AndroidApp APIs to be used while iterating events (mainly
because AndroidApp held a lock over the backend during iteration)
Inner panics could make it hard to trouble shoot the issues and for some
users ints not desirable.
The inner panics were left only when they are used to `assert!` during
development.
At the moment, the with_x11_visual function takes a pointer and
immediately dereferences it to get the visual info inside. As it is safe
to pass a null pointer to this function, it is unsound. This commit
replaces the pointer parameter with a visual ID, and then uses that ID
to look up the actual visual under
the X11 setup. As this is what was already practically happening before,
this change shouldn't cause any performance downgrades.
This is a breaking change, but it's done in the name of soundness so it
should be okay. It should be trivial for end users to accommodate it,
as it's just a matter of getting the visual ID from the pointer to the
visual before passing it in.
Signed-off-by: John Nunley <dev@notgull.net>
There's no need to force the static on the users, given that internally
some backends were not using static in the first place.
Co-authored-by: daxpedda <daxpedda@gmail.com>
Mark it as breaking, since some clients relied on that behavior, simply
because dispatching clients queue always woke up a winit, meaning that
they won't be able to use user events for this sake.
Nothing changed from the user point of view, other than they should
use the `raw-window-handle`, which is objectively better, given that
it reduces the amount of `cfg` guards in downstream code.
#2662 renamed `VirtualKeyCode` to `Key` yet references to the former
type still exist in `src/platform_impl/linux/x11/events.rs`. As it
turns out the `mod events;` in `x11/mod.rs` was removed in the same PR,
but the file accidentally stuck around without being referenced anywhere
else.
Lifetimes don't work nicely when dealing with multithreaded environments
in the current design of the existing winit's event handling model, so
remove it in favor of `InnerSizeWriter` fences passed to client, so they
could try to update the size.
Fixes#1387.
The idea that redraw events are dispatched with a specific ordering
that makes it possible to specifically report when we have finished
dispatching redraw events isn't portable and the way in which we
dispatched RedrawEventsCleared was inconsistent across backends.
More generally speaking, there is no inherent relationship between
redrawing and event loop iterations. An event loop may wake up at any
frequency depending on what sources of input events are being listened
to but redrawing is generally throttled and in some way synchronized
with the display frequency.
Similarly there's no inherent relationship between a single event loop
iteration and the dispatching of any specific kind of "main" event.
An event loop wakes up when there are events to read (e.g. input
events or responses from a display server / compositor) and goes back
to waiting when there's nothing else to read.
There isn't really a special kind of "main" event that is dispatched
in order with respect to other events.
What we can do more portably is emit an event when the event loop
is about to block and wait for new events.
In practice this is very similar to how MainEventsCleared was
implemented except it wasn't the very last event previously since
redraw events could be dispatched afterwards.
The main backend where we don't strictly know when we're going to
wait for events is Web (since the real event loop is internal to
the browser). For now we emulate AboutToWait on Web similar to how
MainEventsCleared was dispatched.
In practice most applications almost certainly shouldn't care about
AboutToWait because the frequency of event loop iterations is
essentially arbitrary and usually irrelevant.
Considering the possibility of re-running an event loop via run_ondemand
then it's more correct to say that the loop is about to exit without
assuming it's going to be destroyed.
This renames all internal implementations of pump_events_with_timeout
to pump_events and makes them public.
Since all platforms that support pump_events support timeouts there's
no need to have a separate API.
This layers pump_events on a pump_events_with_timeout API, like we have
for Linux and Android.
This is just an internal implementation detail for now but we could
consider making pump_events_with_timeout public, or just making it so
that pump_events() takes the timeout argument.
Considering the strict requirement that applications can't keep windows
across run_ondemand calls, this tries to make the window_ondemand example
explicitly wait for its Window to be destroyed before exiting each
run_ondemand iteration.
This updates the example to only `.set_exit()` after it gets a
`Destroyed` event after the Window has been dropped.
On Windows this works to ensure the Window is destroyed before the
example waits for 5 seconds.
Unfortunately though:
1. The Wayland backend doesn't emit `Destroyed` events for windows
2. The macOS backend emits `Destroyed` events before the window is
really destroyed.
and so the example isn't currently portable.
Although we document that applications can't keep windows between
separate run_ondemand calls it's possible that the application has only
just dropped their windows and we need to flush these requests to the
server/compositor.
This fixes the window_ondemand example - by ensuring the window from
the first loop really is destroyed before waiting for 5 seconds
and starting the second loop.
A minimal example that shows an application running the event loop more
than once via `run_ondemand`
There is a 5 second delay between each run to help highlight problems
with destroying the window from the first loop.
This re-works the portable `run()` API that consumes the `EventLoop` and
runs the loop on the calling thread until the app exits.
This can be supported across _all_ platforms and compared to the
previous `run() -> !` API is now able to return a `Result` status on all
platforms except iOS and Web. Fixes: #2709
By moving away from `run() -> !` we stop calling `std::process::exit()`
internally as a means to kill the process without returning which means
it's possible to return an exit status and applications can return from
their `main()` function normally.
This also fixes Android support where an Activity runs in a thread but
we can't assume to have full ownership of the process (other services
could be running in separate threads).
Additionally all examples have generally been updated so that `main()`
returns a `Result` from `run()`
Fixes: #2709
Wayland:
I found the calloop abstraction a little awkward to work with while I was
trying to understand why there was surprising workaround code in the wayland
backend for manually dispatching pending events.
Investigating this further it looks like there may currently be several issues
with the calloop WaylandSource (with how prepare_read is used and with (not)
flushing writes before polling)
Considering the current minimal needs for polling in all winit backends I do
personally tend to think it would be simpler to just own the responsibility for
polling more directly, so the logic for wayland-client `prepare_read` wouldn't
be in a separate crate (and in this current situation would also be easier to fix)
I've tried to maintain the status quo with calloop + workarounds.
X11:
I found that the recent changes (4ac2006cbc) to port the X11 backend
from mio to calloop lost the ability to check for pending events before
needing to poll/dispatch. (The `has_pending` state being queried
before dispatching() was based on state that was filled in during
dispatching)
As part of the rebase this re-introduces the PeekableReceiver and
WakeSender which are small utilities on top of
`std::sync::mpsc::channel()`. This adds a calloop `PingSource`
so we can use a `Ping` as a generic event loop waker.
For taking into account false positive wake ups the X11 source now
tracks when the file descriptor is readable so after we poll via
calloop we can then specifically check if there are new X11 events
or pending redraw/user events when deciding whether to skip the
event loop iteration.
The implementation of `pump_events` essentially works by hooking into the
`RunLoopObserver` and requesting that the app should be stopped the next time
that the `RunLoop` prepares to wait for new events.
Originally I had thought I would poke the `CFRunLoop` for the app directly and
I was originally going to implement `pump_events` based on a timeout which I'd
seen SDL doing.
I found that `[NSApp run]` wasn't actually being stopped by asking the RunLoop
to stop directly and inferred that `NSApp run` will actually catch this and
re-start the loop.
Hooking into the observer and calling `[NSApp stop]` actually seems like a
better solution that doesn't need a hacky constant timeout.
The end result is quite similar to what happens with existing apps that
call `run_return` inside an external loop and cause the loop to exit for
each iteration (that also results in the `NSApp` stopping each
iteration).
A surprising amount of work was required to enable these extensions
on Windows.
I had originally assumed that pump_events was going to be very similar
to run except would use PeekMessageW instead of GetMessageW to avoid
blocking the external loop but I found the Windows backend broke
several assumptions I had.
Overall I think these changes can hopefully be considered a quite a
significant simplification (I think it's a net deletion of a fair amount
of code) and I think it also helps bring it into slightly closer alignment
with other backends too
Key changes:
- I have removed the `wait_thread` that was a fairly fiddly way of handling
`ControlFlow::WaitUntil` timeouts in favor of using `SetTimer` which works
with the same messages picked up by `GetMessage` and `PeekMessage`.
- I have removed the ordering guarantees between `MainEventsCleared`,
`RedrawRequested` and `RedrawEventsCleared` events due to the complexity in
maintaining this artificial ordering, which is already not supported
consistently across backends anyway (in particular this ordering already
isn't compatible with how MacOS / iOS work).
- `RedrawRequested` events are now directly dispatched via `WM_PAINT` messages
- comparable to how `RedrawRequested` is dispatched via `drawRect` in the
MacOS backend.
- I have re-worked how `NewEvents`, `MainEventsCleared`, and `RedrawEventsCleared`
get dispatched to be more in line with the MacOS backend and also more in line
with how we have recently discussed defining them for all platforms.
`NewEvents` is conceptually delivered when the event loop "wakes up" and
`MainEventsCleared` gets dispatched when the event loop is about to ask the
OS to wait for new events.
This is a more portable model, and is already how these events work in the
MacOS backend.
`RedrawEventsCleared` are just delivered after `MainEventsCleared` but this
event no longer has a useful meaning.
Probably the most controversial thing here is that this "breaks" the ordering
rules for redraw event handling, but since my changes interacted with how the
order is maintained I was very reluctant to figure out how to continue
maintaining something that we have recently been discussing changing:
https://github.com/rust-windowing/winit/issues/2640.
Additionally, since the MacOS backend already doesn't strictly maintain this
order it's somewhat academic to see this as a breakage if Winit applications
can't really rely on it already.
This updates the documentation for `request_redraw()` to reflect that we
no longer guarantee that `RedrawRequested` events must be dispatched
after `MainEventsCleared`.
This adds two new extensions for running a Winit event loop which will
replace `EventLoopExtRunReturn`
The `run_return` API is trying to solve multiple problems and address
multiple, unrelated, use cases but in doing so it is not succeeding
at addressing any of them fully.
The notable use cases we have are:
1. Applications want to be able to implement their own external
event loop and call some Winit API to poll / pump events, once
per iteration of their own loop, without blocking the outer,
external loop. Addressing #2706
2. Applications want to be able to re-run separate instantiations
of some Winit-based GUI and want to allow the event loop to exit with
a status, and then later be able to run the loop again for a new
instantiation of their GUI. Addressing #2431
It's very notable that these use cases can't be supported across
all platforms and so they are extensions, similar to
`EventLoopExtRunReturn`
The intention is to support these extensions on:
- Windows
- Linux (X11 + Wayland)
- macOS
- Android
These extensions aren't compatible with Web or iOS though.
Each method of running the loop will behave consistently in terms of how
`NewEvents(Init)`, `Resumed` and `LoopDestroyed` events are dispatched
(so portable application code wouldn't necessarily need to have any awareness
of which method of running the loop was being used)
Once all backends have support for these extensions then we can
remove `EventLoopExtRunReturn`
For simplicity, the extensions are documented with the assumption that
the above platforms will be supported.
This patch makes no functional change, it only introduces these new
extensions so we can then handle adding platform-specific backends
in separate pull requests, so the work can be landed in stages.
Unfortunately this isn't a total removal, for two reasons:
- We still need "libc" for the Xlib XIM implementation, for locales.
- BSD requires libc to check for main-threadedness.
First one we can likely resolve in the near future, not so sure about
the second one without using some weird pthreads trick.
The utils in this module should help the users to activate the windows
they create, as well as manage activation tokens environment variables.
The API is essential for Wayland in the first place, since some
compositors may decide initial focus of the window based on whether
the activation token was during the window creation.
Fixes#2279.
Co-authored-by: John Nunley <jtnunley01@gmail.com>
