Metal hal seems to accumulate CAMetalLayers in view on iOS

[jimblandy]

On iOS, this code seems to be appending a new sublayer to the UIView's main layer each time the function is called:

wgpu/wgpu-hal/src/metal/surface.rs

Lines 125 to 134 in a377ae2

	#[cfg(target_os = "ios")]
	{
	// Unlike NSView, UIView does not allow to replace main layer.
	let () = msg_send![main_layer, addSublayer: new_layer];
	// On iOS, "from_view" may be called before the application initialization is complete,
	// `msg_send![view, window]` and `msg_send![window, screen]` will get null.
	let screen: *mut Object = msg_send![class!(UIScreen), mainScreen];
	let scale_factor: CGFloat = msg_send![screen, nativeScale];
	let () = msg_send![view, setContentScaleFactor: scale_factor];
	};

If the view's layer is not a CAMetalLayer, we create one, and add to the end of view's layer's list of sublayers. But the view's layer itself doesn't change, so if we call this function again, we're just going to create and add a new CAMetalLayer, not return the one we created last time, which seems to be what the function is trying to do.

I'm not sure this is actually a bug; I just noticed it while looking at the code. Putting it here just in case.

[jimblandy]

Cc @jinleili

[jinleili]

get_metal_layer is actually only called when create_surface(), so is it enough to change visibility to pub(crate)?

[jimblandy]

If that's the case, then is the code that checks for an existing CAMetalLayer needed? Will people ever pass in views that already have a CAMetalLayer?

[jimblandy]

Yes, making it only callable from the crate would make sense.

It confused me because it looked like it was designed to be idempotent, but then it wasn't.

[jinleili]

If that's the case, then is the code that checks for an existing CAMetalLayer needed? Will people ever pass in views that already have a CAMetalLayer?

It is indeed possible for a user to pass a UIView with CAMetalLayer, and I need to think about how to better check CAMetalLayer.