[piercifani]

Stack 12/21 · base: jni/11-async-throwing-stream · previous: #15 · next: #17 (jni/13-java-enums)

Problem

JNI arrays are converted eagerly, and only primitive, String, jextracted-class and nested-array elements were handled. Any other element type made jextract skip the declaration, including:

• [Tag?], an array of optionals, which failed with unsupportedSwiftType(Tag?);
• [[String: Int64]], an array of dictionaries, which failed with unsupportedSwiftType([String: Int64]).

Dictionaries and sets didn't have this problem, because they convert their keys and values through bridges. [String: [Tag]] and [Tag]? already worked.

Changes

Arrays whose elements are optionals, dictionaries or sets now go through the existing ArrayBridge, which converts element by element with the same bridges dictionaries use.

• Swift thunk: ArrayBridge<…>.toJavaObject(result, in: environment) for results and ArrayBridge<…>.fromJavaObject(array, in: environment) for parameters.
• Java: the public method is typed: Optional<Tag>[], SwiftDictionaryMap<String, Long>[], SwiftSet<String>[]. The native method takes and returns Object[], the array's static type on the Swift side. The wrapper casts the result with a new .cast conversion step, which renders (T[]) $f().
• Element types: OptionalBridge always creates a java.util.Optional, so an optional primitive element is Optional<Long>, not OptionalLong. bridgedArrayElementJavaType computes that type.

Kotlin in this PR

CollectionsTest.kt with Collections.swift:

Watchlist.quotes().map { it.getOrNull()?.id }   // [Item?] → listOf("a", null, "c")
Watchlist.pages()[1]["c"]                        // [[String: Int64]] → 3
Watchlist.total(arrayOf(Optional.of(1L), Optional.empty(), Optional.of(3L)))  // [Int64?] parameter

Why Skip needs it

A production Skip Fuse app (16 Skip modules) bridges 2 signatures with [T?] and 19 with nested collections. This PR fixes those whose outer collection is an array; dictionaries and sets of collections, and [T]? (85 signatures), already work on this branch. With strict skipped-declaration checks, each unsupported one fails the build. Without the checks, Kotlin silently loses those members.

Testing

• Kotlin sample: ./gradlew :Samples:SwiftJavaExtractJNIKotlinSampleApp:test passes (33 tests).
• New JNIBridgedArrayElementsTests (4 tests): the Java signature, cast and native Object[] signature; the Swift ArrayBridge calls for results and parameters; dictionary elements; jextracted optional elements.
• JNI sample ArraysTest: [Int64?] and [MySwiftClass?] round trips, [[String: Int64]] and [Set<String>] results.
• Full swift test --disable-experimental-prebuilts --disable-sandbox (658 tests) and ./gradlew :Samples:SwiftJavaExtractJNISampleApp:test (292 tests) pass.
• JNISkippedDeclarationsTests (from JExtract: report every skipped declaration; add failOnSkippedDeclarations #3) used [Tag?] and [[String: Int64]] as its unsupported examples; it now uses arrays of tuples and of closures.
• swift format lint --strict on changed files; git diff --check.

Not in this PR

• Arrays of other element types without a direct mapping (tuples, closures, Date, UUID).
• Kotlin-friendly nullability (Item? instead of Optional<Item>) is a separate gap.
• Arrays returned from Java-implemented protocol requirements still use the earlier array rules.

🤖 Generated with Claude Code