[hryhoriiK97]

Summary:

In its two per-child loops (slicing a node's children, and walking matched child pairs), the differ calls CullingContext::adjustCullingContextIfNeeded for every child and reads ReactNativeFeatureFlags::enableViewCulling() each time. Each call builds a ShadowViewNodePair with a copy of the child's ShadowView (props, state and event emitter shared_ptrs) and returns a new CullingContext (a Rect and a Transform with its operations vector). Without view culling, the result is always an unchanged copy of the parent's context.

View culling is off by default, so every commit pays for this on every child the differ visits. This change reads the flag once per call and, when culling is off, passes the parent's context to the children instead. With culling on, the same calls run in the same order as before. The other adjustCullingContextIfNeeded calls are in rarer paths (flattening, unflattening, deletions) and are left as they are.

I found this while profiling typing in a TextInput on Android: each keystroke commits a state update, and the differ walks the whole screen. On a low-end phone the copies were about 5% of the differ's time.

Changelog:

[GENERAL] [CHANGED] - The differ no longer copies its view culling context for every child when view culling is off

Test Plan:

• yarn fantom with the tester rebuilt from this change:
  • ScrollView-viewCulling-itest.js (runs with enableViewCulling:true): 44 passed.
  • TextInput suites 39 passed, ScrollView suites 86 passed, Mounting-itest.js 56 passed.
  • Full suite: 3316 passed. The other failures are not related to this change: snapshot printing in my checkout's Jest setup (LogBox and expect snapshots), and the known flaky tester crash in one suite (polyfillArrayES2023Methods-itest.js), which passes on rerun.
• clang-format clean (node scripts/clang-format.js --check).
• On a device. RNTester release build on a Galaxy A16 (Android 16), one APK with both versions behind a local switch, typing into a TextInput on the TextInput example page for 12 s per run, call stacks sampled at 1 kHz, 5 rounds alternating old/new. JS thread samples spent in adjustCullingContextIfNeeded and the flag reads:

	before	this PR
per 12 s run	13 / 9 / 12 / 12 / 12	2 / 1 / 1 / 2 / 0

That is about 1.3–1.5% of the JS thread while typing (the differ itself is about 30% of it). Total differ time moves by more than that between runs, so I am not claiming more.

🤖 Generated with Claude Code

[meta-codesync]

@javache has imported this pull request. If you are a Meta employee, you can view this in D128599926.

[christophpurrer]

Review automatically exported from Phabricator review in Meta.