-O breaks KeyPath resolution
#78948
Comments
lhunath
added
bug
xedin
added
compiler
|
Thanks for the report. Are you able to provide a fully self-contained example? I stubbed out enough stuff from what you've shown here to get it to build, but it's not crashing for me. I'm not sure if I've broken some relevant part or if the bug is hiding, so it would be good to have a known broken |
|
I do apologize, I thought I had purged the code of external references, but not well enough. As I am trying to recreate the situation today, I am struggling to get a reproducing case going. I will update once I have more information. |
|
No worries, I know how hard it can be to extract a standalone example sometimes. |
|
I've reproduced the crash with the following self-contained SwiftUI application, running on the iOS Simulator through Xcode, with a new project modified only to set the I'm not entirely sure why the crash wasn't reproducing earlier today 🤷. If you need the whole project, let me know. I'm unsure at this point how crucial SwiftUI is to the repro, maybe I'll try reducing it to a pure swift cli package. import SwiftUI
import Combine
public struct Customer {
public var tags: [String]
}
public class Container<Value>: ObservableObject {
@Published
public var selection: [Value] = []
@PublishedProxy<Value?, Container>(
get: { $0.selection.first },
set: { $0.selection = $1.flatMap { [$0] } ?? [] }
)
public var selected: Value?
}
@propertyWrapper
public struct PublishedProxy<Value, T: ObservableObject> {
public static subscript(
_enclosingInstance observed: T,
wrapped wrappedKeyPath: ReferenceWritableKeyPath<T, Value>,
storage storageKeyPath: ReferenceWritableKeyPath<T, Self>
)
-> Value {
get {
observed[keyPath: storageKeyPath].get(observed)
}
set {
// swiftlint:disable:next force_cast
(observed.objectWillChange as! ObservableObjectPublisher).send()
observed[keyPath: storageKeyPath].set(observed, newValue)
}
}
public init(get: @escaping (T) -> Value, set: @escaping (T, Value) -> Void) {
self.get = get
self.set = set
}
@available(*, unavailable, message: "@PublishedComputed can only be used in object types.")
public var wrappedValue: Value {
get { fatalError() }
// swiftlint:disable:next unused_setter_value - https://github.com/realm/SwiftLint/issues/3863
set { fatalError() }
}
@available(*, unavailable, message: "@PublishedComputed can only be used in object types.")
public var projectedValue: AnyPublisher<Value, Never> {
get { fatalError() }
// swiftlint:disable:next unused_setter_value - https://github.com/realm/SwiftLint/issues/3863
set { fatalError() }
}
// - Private
private var get: (T) -> Value, set: (T, Value) -> Void
}
struct ContentView: View {
@ObservedObject
var users = Container<Customer>()
var body: some View {
Color.green
.onAppear {
self.users.selected?.tags = []
}
}
}
@main
struct DemoApp: App {
var body: some Scene {
WindowGroup {
ContentView()
}
}
}
#Preview {
ContentView()
} |
|
Nice, that crashes for me when built as a Mac command-line binary with |
Description
It appears that the Swift optimizer interferes with resolution of key paths for
_enclosingInstancein certain situations.Reproduction
Consider an
ObservableObjectwhose property is modified indirectly:(
PublishedComputedis a custom property wrapper which uses_enclosingInstance, see below)This works fine when the optimizer is off (
-Onone) but when optimizing (-O/-Os), crashes with:(Note: the runtime for these backtraces is iOS 18.1.1, 22B91)
Furthermore, when the type of the
Containeris an existential, the problem becomes further obfuscated:This time, we do not observe the
Fatal errorassertion, indicating the source of the issue. We just crash, attempting to resolve the metadata:As mentioned, these examples rely on a property wrapper which observes the enclosing instance:
Expected behavior
The code should have the same runtime effect regardless of whether
-Oor-Ononeare enabled.Environment
Runtime: iOS 18.1.1, 22B91, iPhone 16 Pro
Additional information
No response
The text was updated successfully, but these errors were encountered: