Perf: QuickJS interop hot-path optimizations (audit findings A-F)

[Singtaa]

Tracking issue for the QuickJS interop hot-path performance work, implemented from an audit of the shipped-app runtime hot path (native / QuickJS interpreter). Findings are labeled A-F. Work spans three repos; commit refs are noted per item.

Done

• D - Batched Painter2D API. PainterBridge replays a flat command buffer onto Painter2D in one JS->C# crossing instead of one reflection crossing per op (and per new Vector2/new Color). JS-side Painter recorder + batchedVisualContent / useBatchedVectorContent.
  • OneJS 393d426, onejs-react feeb7fd, Website c20d0e2
• E - Fast value-type ctors. Zero-alloc ctor fast path (resolved by type-name hash) for Vector2/3/4, Color, Quaternion. Same InteropValue packing as the reflection ctor path, so JS sees an identical value.
  • OneJS 393d426
• C - FastPath not seeded for hot UI Toolkit types. Addressed per-type after hitting the concrete-type keying wall (instance-method fast paths key by the concrete element type, so VisualElement.Add can't be seeded for Button/Label/etc.):
  • Element ctors -> fast ctor path (keying-safe; ctors resolve by type name).
  • VisualElement.Add/Insert/RemoveFromHierarchy -> NodeBridge (static, registered by type name, type-agnostic).
  • IStyle members -> StyleBridge typed setters (S1).
  • Vector2/Color -> E; Painter2D -> D.
  • OneJS 393d426, onejs-react feeb7fd
• B - Element-handle resolution under a global lock. FindElementHandle's parent-chain walk now runs under a single lock acquisition (GetHandleForElementOrAncestor) instead of one lock + dict lookup per hop. Behavior-identical; collapses N locks to 1 on deep (composite-control) event targets.
  • OneJS fbc379f

(Prior related work, before this audit pass: a6e6eba delivered the GC delta-gate, per-element pointer + wheel fast paths, and batched task drain.)

Deferred (intentional)

• A - ExecutePendingJobs after every event dispatch. pointermove is already throttled to ~1/frame, and each drain is load-bearing (dispatch returns the suppression-flag bitmask synchronously, so the reconcile must complete per event). Real but low value, with correctness risk. Revisit only if a profiler flags it.
• F - FastPath/ZeroAlloc hand-registration DX. A [OneJSFastPath] attribute / source generator would remove the re-register-in-Awake step. DX-only (built-in registrations self-heal on domain reload), but it is also the AOT-safe path to broad fast member invocation (see the I2 note below). Defer until custom FastPath usage is common or I2-style coverage is wanted.

Next pickup (broader structural tier, not part of the A-F audit)

• JSON->binary buffer marshaling - tracked in Perf: binary typed-array buffer marshaling (replace JSON __csArray for hot buffers) #107. An enabler for GPUBridge (G2) / WebSocket (N1/N2), not standalone perf work.

I2 (general cached-delegate invokers for the reflection slow path) - dropped as a standalone item. It does not survive scrutiny: Expression.Compile is not AOT-safe (IL2CPP, i.e. the no-JIT native targets QuickJS exists for), and Delegate.CreateDelegate can't be invoked generically without DynamicInvoke (slower than reflection). The AOT-safe way to get broad fast invokers is FastPath registration, ideally automated by F (a source generator). So I2 folds into F + FastPath expansion rather than being a separate runtime-delegate layer. The remaining reflection hits after D/E/C/B are the cold long tail anyway.

Notes

Architectural lessons worth keeping: FastPath instance methods key by concrete runtime type; ctors key by type name (keying-safe); Expression.Compile is not AOT-safe, so S1 uses a static typed-setter switch rather than compiled delegates.