This was supposed to be rolled out with the rwh v0.6 update, but it
was left behind for some reason. I've added this type back.
Signed-off-by: John Nunley <dev@notgull.net>
Previously we had a sort of artificial split between these, but both were accessing each other's state, since it's really the same state!
It was especially difficult to follow what happens to the fullscreen state.
So instead, we basically merge the window and the delegate files.
This does unfortunately screw a bit with the git history, apologies to whoever reads this in the future!
* make `EventLoopWindowTarget` independent of UserEvent type
the `EventLoopWindowTarget` is needed for window creation. conceptually,
only `EventLoop` and `EventLoopProxy` need to be parameterized, and all
other parts of the backend should be agnostic about the user event type,
parallel to how `Event<T>` is parameterized, but `WindowEvent` is not.
this change removes the dependency on the type of user events from the
`EventLoopWindowTarget` for the Windows backend, but keep a phantom data
to keep the API intact. to achieve this, I moved the `Receiver` end of
the mpsc channel from `ThreadMsgTargetData` into `EventLoop` itself, so
the `UserEvent` is only passed between `EventLoop` and `EventLoopProxy`,
all other part of the backend just use unit type as a placeholder for
user events.
it's similar to the macos backend where an erased `EventHandler` trait
object is used so all component except `EventLoop` and `EventLoopProxy`
need to be parameterized. however `EventLoop` of the Windows backend
already use an `Box<dyn FnMut>` to wrap the user provided event handler
callback, so no need for an dedicated trait object, I just modified the
wrapper to replace the placeholder user event with real value pulled
from the channel. I find this is the approach which need minimum change
to be made to existing code. but it does the job and could serve as a
starting point to future Windows backend re-works.
* fix CI clippy failure.
* make UserEventPlaceholder a new type instead of alias
* invariance is maintained by top-level EventLoopWindowTarget<T>
this field is transitional and her to keep API compatibility only.
the correct variance and such is already ensured by the top-level
`EventLoopWindowTarget`, just use `PhantomData<T>` here.
xdotool will update modifiers before Xkb will actually send event
updating them, thus the modifiers will be updating even before the
actual update, which is unfortunate.
Links: https://github.com/alacritty/alacritty/issues/7502
The user may change the size during the on-going resize, meaning that
the size will desync with winit's internal loop which breaks viewporter
setup with fractional scaling.
Links: https://github.com/alacritty/alacritty/issues/7474
This appears to be the solution for the elusive #3335 issue. Previously,
in the Xlib backend, we used the "XRRQueryVersion" function to query for
the Xrandr version, and used that to determine whether we should use the
"GetScreenResources" call or the "GetScreenResourcesCurrent" call.
However, we passed the version "0, 0" into "XRRQueryVersion".
Previously with Xlib this wasn't a problem, as Xlib ignores the version
you pass in and substitutes it with the version of RandR it expects.
https://gitlab.freedesktop.org/xorg/lib/libxrandr/-/blob/master/src/Xrandr.c?ref_type=heads#L386-387
The way that "XRRQueryVersion" is implemented on the server end, it
compares the version passed into the request with the version supported
by the server. If the server's version is greater than the client
version, it just returns the client version. If the client's version is
greater, it passes the server's version. Since we were passing in "0, 0"
this means that the server returned "0, 0".
https://gitlab.freedesktop.org/xorg/xserver/-/blob/master/randr/rrdispatch.c?ref_type=heads#L50-59
To determine whether we use "GetScreenResources" or
"GetScreenResourcesCurrent", we compare the version returned by the
server against "1, 3". Since we got "0, 0"- a version of XRandR so old
it doesn't even exist- we use "GetScreenResources".
The problem manifests in that "GetScreenResources" can take several
seconds to query the screen state based on the current hardware
configuration. On the other hand, "GetScreenResourcesCurrent" is fast;
it uses the server's hardware cache if it is available.
This problem is visible in XTrace. On the latest `master`:
```
000:<:00c2: 12: RANDR-Request(140,0): QueryVersion major-version=0 minor-version=0
000:>:00c2:32: Reply to QueryVersion: major-version=0 minor-version=0
000:<:00c3: 8: RANDR-Request(140,8): GetScreenResources window=0x0000076e
000:>:00c3:1600: Reply to GetScreenResources:
```
On the `v0.28.0` tag:
```
000:<:0019: 12: RANDR-Request(140,0): QueryVersion major-version=1 minor-version=6
000:>:0019:32: Reply to QueryVersion: major-version=1 minor-version=6
...later
000:<:002d: 8: RANDR-Request(140,25): GetScreenResourcesCurrent window=0x0000076e
000:>:002d:1600: Reply to GetScreenResourcesCurrent
```
This commit fixes this issue by requesting "1, 3" instead. This returns
the version we expect, where we can now use "GetScreenResourcesCurrent"
properly.
Fixes#3335
Signed-off-by: John Nunley <dev@notgull.net>
