This Week in WebKit — April 04–10, 2026

426

전체 commit 수

1건

Security fix 수

Medium

108

기여자 수

Top components WebCore · Platform · WebKit · JSC · WTF

Featured

A one-line CustomEffect callback nulls the animation it's running on — and the very next virtual call dereferences zero

WebAnimation::tick() null-checked m_effect, called animationDidTick(), then reused that same pointer for animationBecameReady(). For CustomEffect, animationDidTick() runs the JS callback — and the callback can do animation.effect = null, clearing m_effect mid-tick. The cached pointer never gets re-checked, so the following virtual dispatch loads a vtable from address zero and faults. The fix grabs a fresh keyframeEffect() after the callback returns, and de-virtualises animationBecameReady() so the dispatch path is gone for good.

The same iterate-then-callback shape lives in IntersectionObserver, ResizeObserver, and MutationObserver delivery loops — worth grepping for cached members touched after an observer callback returns.

→ Web Animations null dereference via CustomEffect callback Medium WebCore Web Animations

JSC inlines `arr.length = N` shrinks by zeroing butterfly tail slots — and the new stub opens a fresh JIT-guard audit surface

Every length store used to bottom out in operationPutByIdStrictGaveUp, with the full JSArray::put path running on every call. This commit caches the shrink case as an ArrayLengthStore IC stub: for Int32Shape/ContiguousShape arrays, emitted code zeros the butterfly tail slots and updates publicLength inline. The reward is 3.5–4.3x DFG/FTL on push-then-reset patterns; the cost is a new JIT write site whose guards must prove the array hasn't transitioned shape, been frozen, gone COW, or been targeted via Reflect.set with a mismatched receiver — every guard slip is a write into memory the stub didn't expect.

Warm-then-freeze is the canonical IC invalidation test — the patch ships one, and any new IC type that omits it is worth eyeballing.

→ JSC ArrayLengthStore inline cache for JSArray length assignment optimization

Wasm GC array ops finally reach B3 as first-class opcodes — and the same change lets CSE drop bounds checks against freshly-allocated arrays

OMG used to lower Wasm array.get/set/new/len straight to raw loads and stores during IR generation, throwing away every fact B3 could optimise on. Four new opcodes — WasmArrayGet/Set/New/Length — keep array semantics alive through CSE and range analysis, with bounds checks expressed as separate WasmArrayLength nodes the optimizer can fold; allocations from WasmArrayNew let the null-trap be removed entirely. Final lowering moves late into B3LowerMacros. The audit angles are the standard JIT ones: alias-info errors, ValueKey collisions across different arrays, and packed-type sign extension can each silently drop a bounds check or return a wrong length.

Two WasmArrayLength nodes on different arrays sharing a ValueKey would mean CSE returns the wrong length — a silent OOB primitive in GC heap memory if it ever lands.

→ JSC B3 Wasm GC array opcodes optimization

document.open() shared the caller's SecurityOriginPolicy with the new document — a later document.domain write crossed the boundary silently

SecurityOriginPolicy wraps the SecurityOrigin that enforces same-origin, and each document is supposed to own its own. Document::open() instead called setSecurityOriginPolicy with the caller's policy object — handing over the very same shared instance rather than constructing a new one. Because document.domain writes mutate that SecurityOrigin in place, a caller could rewrite its own domain and have the change reflect on the opened document with no explicit cross-document operation. The fix is a one-line deletion; the same-origin guard at the top of Document::open() was already the real barrier, and Gecko and Blink had never wired the alias in the first place.

Other call sites of setSecurityOriginPolicy — and any code that copies origin state between documents — are worth tracing for the same shared-mutable-reference shape, since aliasing stays invisible until somebody mutates.

→ document.open() security origin aliasing fix bug