Implement ES2021 WeakRef and FinalizationRegistry

[anivar]

Implements ES2021 WeakRef and FinalizationRegistry following the NativeWeakMap pattern.

Implementation

Two self-contained classes, about 400 lines total:

• NativeWeakRef.java - uses Java's WeakReference directly
• NativeFinalizationRegistry.java - uses PhantomReference + ReferenceQueue for GC detection

The finalization approach is simple: poll the ReferenceQueue during register() calls to process any pending cleanups. No separate queue management or Context modifications needed.

Android Compatibility

Uses PhantomReference instead of the Cleaner API since Cleaner isn't available on Android. This keeps the implementation portable across all Java 8+ platforms.

ES2021 Standard Compliance

Correctly implements symbol handling per spec:

• Allows unregistered symbols as weak targets
• Rejects registered symbols created with Symbol.for since they persist globally
• Validates all target/heldValue/token constraints
• Proper Symbol.toStringTag on prototypes

Test Results

All 76 test262 tests passing:

• 29 WeakRef tests
• 47 FinalizationRegistry tests

Also fixes 2 Object.seal tests.

Thread-safe implementation using ConcurrentHashMap for registration tracking.

[gbrail]

Thanks -- this looks like good progress. I have a bunch of questions, please take a look -- thanks!

[aardvark179]

I thought I'd written a review of this, but I must have closed the tab or something.

So, according to the spec the cleanup routines should be run when the weak ref target is no EMPTY, so this should probably be written using PhantomReferences (they are put on the queue when their referent has been collected). I am also unsure what the effect is on the JVM of create a large number of processor queues, and they can be tricky to manage correctly.

Could you try implementing this using a single queue (maybe using a java.lang.ref.Cleaner and putting the tasks from there onto a queue in each Context. We could likely combine processing of those with the micro task loop run as the context is exited.

I think that would help avoid large reference queues hanging around, and would make it easier to understand when things can be run.

[anivar]

@aardvark179 Thank you for the detailed review! Yes, he did leave inline comments earlier and I responded to them and followed the suggestions. However, after refactoring the implementation, the filenames changed and I can't access those same comments anymore.

I understand your architectural suggestions:

1. Using PhantomReferences instead of WeakReferences for more deterministic cleanup behavior
2. Implementing a single shared ReferenceQueue across all FinalizationRegistry instances for efficiency
3. Integrating the cleanup process with the Context's micro-task loop for better timing control

[anivar]

@aardvark179 Implementation completely rewritten per your feedback:

PhantomReferences now replace WeakReferences for correct GC semantics. Added shared ReferenceQueue with Context integration for cleanup processing in the JavaScript execution loop as requested.

Key improvements:

• Thread-safe operations with ConcurrentHashMap
• O(1) token-based unregistration
• Memory leak prevention with proper cleanup
• Significant code simplification
• All deprecation and ErrorProne warnings properly resolved
• 72+ tests passing including cross-realm scenarios

Ready for re-review!

[aardvark179]

Sorry I haven’t got to re-reviewing this one yet. It’s been a busy few weeks, but I promise I will get to it.

It would also be great to have a chat at some point so I can understand your Rhino use case better and maybe help plan out which features are the best ones to tackle first to enable that.

[anivar]

No worries @aardvark179!

I'd be interested in that chat about Rhino use cases and feature prioritization. Feel free to reach out whenever works for your schedule.

Thanks for the architectural guidance on PhantomReference and shared queue - it made the implementation much cleaner.

[anivar]

Force pushed a simplified implementation that follows the NativeWeakMap pattern. All tests passing, Android-compatible, no Context modifications needed. Ready for review.