…ly-once per host per commit Gate the two duplicate in-transaction producers and unify all deferred delivery tokens into one, owned by NativeScriptUIView: - scheduleUIKitHostPropsTransactionCommitIfNeeded (producer #2, the per-updateRevision props commit) now skips scheduling while the owning ComponentView isApplyingMountingTransaction, mirroring the existing in-transaction skip in replayFabricTransactionAfterHostCreationIfNeeded. In-transaction, mountingTransactionDidMount already observes _hasModifiedPropsInCurrentTransaction (set by the same updateRevision bump) and delivers the props-modified commit itself once the transaction finishes, so nothing is lost. - scheduleFabricTransactionCommitFallbackIfNeeded (producer #3, the mount-op fallback) gets the identical in-transaction skip. It remains the only delivery path for the worklet-driven, out-of-transaction nativeScriptApplyUIKitHostPropsForFabricTag path, where mountingTransactionDidMount never fires. - The three independently-bumped tokens (ComponentView's _mountingTransactionToken and _fabricTransactionCommitFallbackToken; NativeScriptUIView's _uikitHostPropsTransactionCommitToken) are deleted and replaced by one _fabricTransactionDeliveryToken owned by NativeScriptUIView, exposed via fabricTransactionDeliveryToken / advanceFabricTransactionDeliveryToken. Every deferred schedule captures a fresh token via advance (which also immediately invalidates any older pending schedule); every actual delivery (notifyFabricTransactionCommittedWithModifiedChildren:...) bumps it again, so a slower-to-fire duplicate always loses to whichever delivery lands first. Bumped as before on setHostId:/dealloc/prepareForRecycle. This matches upstream RNS's load-bearing property: exactly-once, ordered-after-layout delivery (RNSScreenStack.mm:1352-1370's mountingTransactionDidMount dispatch_async's the reconcile exactly once, note the `break`, for both Insert and Remove) - not synchrony. The mount-op fallback (producer #3) and the host-creation replay (producer #4, NativeScriptUIView.mm ~2516-2549) are both left as the sole delivery mechanisms for their respective out-of-transaction/host-creation cases. Lock hierarchy: unaffected. This change only deletes dispatch_async schedule sites and adds none; the main->worklet crossing stays the existing runUIKitHostLifecycle/runUIKitHostFunction path (main-thread-only guard at NativeScriptNativeApiModule.mm:990), and nothing here waits on the worklet runtime or the render server. Measured: for the fork's stack host (RNSScreenStack), transactionCommitted duplicate deliveries dropped from 24/72 sampled commits (33%) showing a second delivery ~microseconds-to-tens-of-ms later, to 0/42 (100% clean, exactly one delivery per commit) across a ~19-pop self-driven run, before/after compared on the same build via git stash. Freshly-created hosts go from 3 deliveries (mount + host-creation-replay + one duplicate from producer #2/#3) down to 2 (mount + the intentionally-kept host-creation-replay pair only). Gate (scripts/gate-pop-wedge-rate.js, 25 cycles, 8-cycle warmup): under confirmed heavy host contention (concurrent unrelated automation - ReelList.app/PlushVisualQA.app - running on the same simulator, load average 10-20), this build measured 40% wedgeRate vs a matched pre-fix-control run under the same contention at 72% wedgeRate/700ms median - a clear relative improvement, not a regression, consistent with removing 2 of 4 duplicate main-thread worklet crossings per commit. A subsequent run under lighter load passed cleanly: 12.0% wedgeRate, 0% timeoutRate, 100% completion, 475.5ms clean-pop median (baseline: ~10.5% wedge, ~397-450ms median). react-native-screens fork jest suite (NativeScriptScreenStack): 344/344 unaffected, fork untouched (HEAD 1b8ba1c). No crash reports across the full verification session. Interactive ax-tree-driven verify scripts (back-swipe-pop, modal-interactive-dismiss, first-paint-pixels) gave inconsistent results in this session due to a combination of agent-device session/tap-targeting flakiness and the same host contention noted above (confirmed via direct simctl screenshots showing correct app state whenever these scripts misfired) - not run to a clean pass/fail verdict; not attributed to this change. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Uh oh!

Uh oh!

Debug memory management #4

Description

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions