Improving yt-dlp-ejs with Rust: Smaller and Faster

#rust #ytdlp #swc #quickjs

Motivation

In the mpv-easy project, there's a web tool for customizing mpv players and scripts, including yt-dlp. To leverage GitHub CDN (which requires files under 25MB), file size is a constraint. Node/Bun/Deno runtimes are nearly impossible to compress below 25MB, and QuickJS performance in yt-dlp-ejs is suboptimal.

The Problem

The yt-dlp/ejs implementation uses meriyah to parse JavaScript, modifies the AST, regenerates code, and executes it via a JS engine.

Profiling reveals that for lightweight engines like QuickJS and Boa, the bottleneck is code generation. When generating code from a large AST, extensive string concatenation occurs—something these engines don't optimize well. (Note: QuickJS recently added optimizations for this scenario.)

In contrast, SWC's code generation is highly efficient. This explains why SWC + QuickJS outperforms Deno/Node/Bun—their CLI startup time is comparable to JS execution time.

[TRACE] → readFile
[TRACE] ← readFile (2.14ms)
[TRACE] → main
  [TRACE] → preprocessPlayer
    [TRACE] → parse
    [TRACE] ← parse (3187.57ms)
    [TRACE] → filterExpressions
    [TRACE] ← filterExpressions (251.01ms)
    [TRACE] → generate

Solution

Using Rust, SWC replaces meriyah for AST manipulation. For Deno/Node/Bun, code is saved to a temp file and executed via CLI. For QuickJS/Boa, code is executed directly, eliminating CLI startup overhead.