How Alibaba’s JS AOT Supercharges Web Performance: Design, Implementation, and Results

Background

JavaScript’s dynamic nature makes it easy for developers but forces recompilation before each execution, hurting performance compared with native code. Alibaba’s U4 engine previously used code‑cache techniques, yet still lagged behind native in first‑launch scenarios.

Technical Options

Three feasible AOT approaches were examined:

Native Machine Code – Directly compiling JS to assembly was rejected because JavaScript’s weak typing leads to massive code size, slow compilation, and CPU‑architecture‑specific binaries.

Full Code Cache (Bytecode) – Caching complete bytecode avoids runtime compilation but suffers low function coverage, large size (≈2.6× source), and high deserialization cost; performance gains are modest and inconsistent.

Partial Code Cache (Selective Bytecode) – Generates bytecode only for hot functions, but faces fragmentation across engine versions, CPU‑architecture‑specific constants, and device‑specific compatibility issues.

Design of the AOT Solution

The final design combines two strategies to ensure effectiveness and compatibility:

Effectiveness

PGO (Profile‑Guided Optimization) : Collect runtime function usage data from a pre‑release page visit, process it on the server, and pre‑compile only the hot functions into AOT code.

Prior‑knowledge Rules : Predict hot functions based on code depth and size, generating bytecode for shallow, frequently executed code.

Compatibility

Online Generation : Pre‑warm AOT during idle time or generate on‑demand when a page is opened, storing the result on disk for fast loading.

Offline Generation : Use a build‑time tool to compile AOT from source and bundle it with the release, suitable for cold‑start scenarios and rarely‑changed scripts.

Small JS files skip AOT, medium‑size files use full bytecode cache, and large files compile only the top three function layers, with incremental updates as pages run.

Optimization Effects

In production, pages using the prior‑rule AOT reduced JavaScript execution time by roughly 35%. PGO‑based AOT delivered average performance improvements of over 49%, while rule‑based AOT achieved about 34% gains. Specific case studies (e.g., Quark exam app) showed first‑screen speedups of 17.6%.

Future Outlook

U4 5.0 will introduce a Sparkplug baseline compiler that translates bytecode directly to unoptimized assembly, improving branch prediction and overall speed. Early estimates suggest an additional 20% performance boost when combined with AOT.