[piercifani]

Stack 1/8 · base: main · next: #2 (jni/02-closures-jextracted-types)

This is upstream swiftlang/swift-java#938 by Himeshi (teamhimeh on GitHub), carried unchanged with its authorship, so the rest of this stack can build on it. It stays here until swiftlang#938 is merged upstream, then this branch is dropped and the stack is rebased onto main.

What it does (from swiftlang#938)

JNI extraction kept synchronous Java signatures for @MainActor APIs but emitted nonisolated Swift thunks, so the generated bindings didn't compile when they touched isolated initializers, methods, properties or subscripts.

• Synchronous JNI entry points are annotated @MainActor when the declaration, its enclosing type or its extension is explicitly main-actor isolated.
• The thunk calls MainActor.preconditionIsolated() before converting arguments. Java callers must already be on the main actor's executor; there's no hop.
• Explicitly nonisolated members and async downcalls are unchanged.

Why Skip needs it

In Skip Fuse apps every view model is a @MainActor class read from Compose on the UI thread. In a production Skip Fuse app (16 Skip modules), 4,180 bridged entrypoints are main-actor isolated, and all 190 view models write @MainActor explicitly, so swiftlang#938's explicit-attribute detection covers them.

Kotlin in this PR

This PR's first commit adds Samples/SwiftJavaExtractJNIKotlinSampleApp. Its Swift modules build in Swift 6 language mode, and its bindings are used from Kotlin: one test file per feature, kotlin.test plus kotlinx-coroutines for await() on the generated futures. It's in the CI sample matrices like the other samples. Each later PR in the stack adds its own Swift fixture and Kotlin test, so every PR shows the Kotlin it enables. (It was briefly a separate PR, #9, but GitHub doesn't allow changing the base of the first PR in a stack.)

• BaselineTest.kt: what works on main today. A struct, a throwing function, an async function awaited as a coroutine, an enum with a payload, and a type from the other module.
• MainActorTest.kt with MainActorViewModel.swift: a @MainActor view model called synchronously from Kotlin. It runs on macOS only, where the JVM test thread can be the process's main thread (-XstartOnFirstThread), as Android's UI thread is; on Linux the JVM never runs tests on the process's main thread.

Extra validation done here

• Kotlin sample: ./gradlew :Samples:SwiftJavaExtractJNIKotlinSampleApp:test passes (the baseline tests also pass on main without this PR).
• JNI: preserve MainActor isolation in synchronous thunks swiftlang/swift-java#938's 10 generator tests pass.
• A standalone Swift 6 package with a @MainActor view model, called from Kotlin on the JVM, passes when the test JVM runs on the process's first thread (-XstartOnFirstThread on macOS, which matches Android's UI thread being the process main thread). Off the main thread the JVM stops with SIGTRAP from the precondition, as intended.

Note for Skip: today Skip's bridge uses assumeMainActorUnchecked, which lets off-main calls through silently. With this change they trap, which will surface latent threading bugs in apps.

🤖 Generated with Claude Code

[piercifani]

Superseded by a replacement PR (linked below once opened). The Kotlin sample commits now sit on upstream/pr-938, a branch at the exact commit of swiftlang#938 by Himeshi. Upstream code stays out of this fork's PR diffs, the way upstream syncs to main do, so the CLA check only covers Skip's own commits. The CLA bot comments above were automatic and need no action from anyone.

[piercifani]

Replaced by #11.