Convert arrays to Java with one existential open and one ArrayList call

[piercifani]

Summary

Array.toJavaObject called AnyBridging.toJavaObject(_: Any?) for every element, which boxes the element into Any and runs a dynamic as? JConvertible cast, then made one ArrayList.add JNI call per element. A profile of a Kotlin loop reading [Int] × 200 put about a quarter of the getter in swift_dynamicCast and another quarter in the ArrayList.add calls.

Now:

• Element.self as? JConvertible.Type is checked once per array, and the conversion runs through the statically-typed conformance (withMemoryRebound over the array buffer, which is valid because the cast proved Element == E).
• Elements go into a Java Object[] via SetObjectArrayElement (no Java method invocation per element), then one ArrayList(Arrays.asList(...)) call builds the list. The result is still a mutable java.util.ArrayList, which skip.lib.Array(nocopy:) relies on.
• Element types that are not JConvertible (Any, existentials, Optional<Any>, errors) keep the per-element AnyBridging path.

Set and Dictionary are unchanged: a per-element generic helper measured slower than the existing cast, so I left them alone.

Numbers

	Pixel stock	Pixel patched		Mac stock	Mac patched
[Int] × 200 getter	113 µs	76 µs	1.49×	65 µs	33 µs	1.99×
[Struct] × 200 getter	180 µs	120 µs	1.50×	94 µs	52 µs	1.79×
[Struct] × 200 getter + read every field	297 µs	220 µs	1.35×	127 µs	88 µs	1.43×
[String] × 200 getter (with skiptools/swift-jni#11)	123 µs	62 µs	1.97×	64 µs	26 µs	2.45×

App-shaped types (Pixel only; the rows above use a deliberately small 4-field struct). ChatMessage mirrors a production chat message: 13 fields including a nested 10-field user struct, Dates, an enum with associated values and arrays of nested structs. FeedRowID is a String-raw enum plus a String key.

	Pixel stock	Pixel patched
[ChatMessage] × 50 getter	70 µs	49 µs	1.43×
[ChatMessage] × 50 getter + read 4 fields per row	349 µs	331 µs	1.05×
[FeedRowID] × 200 getter	147 µs	112 µs	1.31×

Once a row's fields are read, per-field getters dominate and the gain from this change shrinks accordingly.

What is left per element is mostly the Kotlin peer allocation (NewObjectA on a class with finalize()) and the SwiftValueTypeBox.

How this was measured

A benchmark test in SkipBridgeToKotlinSamplesTests (not part of these PRs) reads a bridged BenchModel from the transpiled Kotlin test: scalar, String, struct and 200-element array getters, plus == on two 200-element [Struct] lists (what Compose's changed() does). Each row is the best of 15 trials, two interleaved stock/patched rounds.

• Device: Pixel 10 Pro Fold, Android 17, connectedAndroidTest with testBuildType = "release". The debug test APK is debuggable, so ART turns on CheckJNI ("Late-enabling -Xcheck:jni"), which roughly doubled every JNI cost and would have overstated these changes.
• Mac: the same test under Robolectric, Swift built with -O.

Rows these changes don't touch moved by up to ±12% between rounds on the device; treat that as the noise floor.

Tests

• skip test: 153/153 (macOS + Robolectric).
• Polymarket's Android app builds against it (assembleDebug).

Related: skiptools/swift-jni#11, skiptools/skipstone#292 (independent; each can land on its own)

🤖 Generated with Claude Code