[piercifani]

Stack 2/21 · base: jni/01-main-actor-thunks · previous: #11 · next: #3

Problem

In JNI mode:

1. Escaping closures whose parameters are jextracted types (structs, classes, enums) produced Swift that didn't compile. The closure's @JavaInterface wrapper declared the Swift type (not a JavaValue). The statements that wrap the value in its Java class were rendered into a printer that was then discarded, leaving _0Class / _0Pointer undefined.
2. Function-type attributes were matched by substring. Only @escaping was understood; @Sendable and @MainActor were dropped, and without @escaping any attribute made the type unsupported. A stored property of type @MainActor @Sendable (Item) -> Void was skipped.
3. The closure wrapper keeps the JNIEnv of the thread that created it, so a @Sendable closure called from another thread used the wrong environment.
4. A closure-typed property's setter went through the non-escaping lowering, so the stored closure captured the call's JNIEnv and a local reference that are invalid once the call returns.

Changes

• Parse function-type attributes one by one: @escaping, @Sendable, @MainActor, @convention(c|swift). Other attributes are still tolerated next to @escaping, as before.
• In the closure's @JavaInterface, declare jextracted parameters as their Java class (JavaItem?, and JavaOptional<JavaItem>? for optionals), and emit the argument conversions inside the closure body.
• Without Java callbacks (when wrap-java doesn't run), print a minimal @JavaClass wrapper with wrapMemoryAddressUnsafe for each such type, once per module. With callbacks on, wrap-java's classes are used.
• For @Sendable closures, re-wrap the Java object with the calling thread's JNIEnv (JavaVirtualMachine.shared().environment()) on each call.
• Treat a closure property's setter newValue as escaping.
• Docs: escaping closures with jextracted types; @Sendable / @MainActor closure types.

Why Skip needs it

Skip Fuse view models take a Callbacks struct of @MainActor @Sendable closures whose parameters are app structs. In a production Skip Fuse app (16 Skip modules), 849 of 1,351 bridged closures have non-primitive parameter or result types, and 419 are main-actor closures.

Kotlin in this PR

ClosuresTest.kt with Closures.swift: Kotlin lambdas passed as escaping closures taking a struct, as a closure with an optional struct parameter, and as a @Sendable closure that Swift calls on another thread. The test also builds a Callbacks struct of @MainActor @Sendable closures (macOS only).

Testing

• Kotlin sample: ./gradlew :Samples:SwiftJavaExtractJNIKotlinSampleApp:test passes at this point in the stack.
• New JNIEscapingClosureJExtractedTypesTests (7 tests); all 7 fail on the base branch and pass with this change.
• Full swift test --disable-experimental-prebuilts --disable-sandbox passes.
• JNI sample: two new fixtures (deliverStruct, deliverOptionalStruct) with Java tests; ./gradlew :Samples:SwiftJavaExtractJNISampleApp:test passes 275/275. The sample builds with Java callbacks enabled, so that mode is covered.
• A standalone Swift 6 package, without Java callbacks, passes Kotlin tests on the JVM: a struct-taking escaping closure, a closure with an optional struct parameter, a @Sendable closure invoked from a background thread, and a Callbacks struct of @MainActor @Sendable closures.
• swift format lint --strict on changed files; git diff --check.

Not in this PR

• Optional closure properties, non-escaping closures taking structs, closures as return values, and async closures are still dropped (silently).
• With Java callbacks on, wrap-java emits an unescaped init holder type for initializers that take closures (JavaFoo.init.onOpen), which doesn't compile. That's a separate wrap-java fix.

🤖 Generated with Claude Code