[tomekzaw]

Summary:

FabricMountingManager::synchronouslyUpdateViewOnUIThread creates a ReadableNativeMap for the props and calls .release() on its local reference, so nothing deletes it. The C++ animation backend calls this once per animated view on every frame: for every view when optimizedAnimatedPropUpdates is off, and for views with layout updates otherwise. So each props map stays alive until the frame callback returns to Java. A frame that animates 600 views holds 600 local references. Android 7, the oldest supported version, limits a thread to 512 local references and aborts above that.

This wraps the reference in jni::adopt_local, like dispatchCommand above it, so each map is released after its call.

Changelog:

[ANDROID] [FIXED] - Release the JNI local reference to the props map in synchronouslyUpdateViewOnUIThread

Test Plan:

Release RNTester (./gradlew :packages:rn-tester:android:app:assembleRelease -PreactNativeArchitectures=arm64-v8a) with these temporary, uncommitted changes:

• The experimental release level, so cxxNativeAnimatedEnabled and useSharedAnimatedBackend are on and optimizedAnimatedPropUpdates is off.
• A Playground screen that animates the translateX of 600 views with useNativeDriver: true.
• Logging that records the slot of a fresh local reference at the start of each synchronouslyUpdateViewOnUIThread call.

On a Galaxy A16 (Android 16):

	Calls per frame	Distinct local reference slots
main	600	600
This PR	600	1

On main, every call starts on a new slot because the earlier maps are still referenced. With this change, every call reuses the same slot.

Performance, with the same setup (600 views moved by one native-driven animation), timed per animation frame (Java timer around driveAnimationBackend, which includes freeing the local references when it returns) and per synchronouslyUpdateViewOnUIThread call (C++ timer):

	Galaxy A16	Pixel 9a	OPPO CPH2269	moto g57 power
One animation frame	40.27 ms → 40.40 ms (no change)	10.70 ms → 10.73 ms (no change)	119.21 ms → 118.06 ms (no change)	26.27 ms → 26.45 ms (no change)
One synchronouslyUpdateViewOnUIThread call	54.00 µs → 54.00 µs (no change)	12.21 µs → 12.23 µs (no change)	145.12 µs → 142.96 µs (no change)	33.12 µs → 33.49 µs (no change)

Setup:

• Release RNTester with R8 on and the experimental release level, compiled with cmd package compile -m speed.
• Temporary timing probes, not part of this PR, and a Playground scenario opened by deep link.
• Each run is a fresh process: a 12 s window after the scenario starts, reported as the median of its samples. Per phone, one discarded run of each build, then 8 A/B/B/A cycles of main and this PR (16 processes per build).
• Each figure is the median of the 16 processes. "No change" means the 95% bootstrap confidence interval includes zero. It is within about ±2% on the Galaxy A16, Pixel 9a and moto g57 power. The OPPO CPH2269 is noisier (about 120 ms per frame, so few samples per run).
• The Galaxy A16, Pixel 9a and moto g57 power run Android 16. The OPPO CPH2269 runs Android 11.

Not run: Android 7, where the 512-reference limit would turn this into a crash; I don't have a device with it.

🤖 Generated with Claude Code

[meta-codesync]

@javache has imported this pull request. If you are a Meta employee, you can view this in D128714083.