Improve performance by switching from babel

[o-alexandrov]

Describe the enhancement

Switch from babel to swc.

Motivation

Performance should be improved. It's extremely slow in development.

Possible implementations

https://swc.rs/docs/migrating-from-babel

Related Issues

• Vite plugin: babel transform performance issue

[layershifter]

To be clear, the parser is only a part of the problem.

---

If we will switch to SWC as a parser we will gain almost nothing:

• SWC emits non estree AST (ESTree compatible AST swc-project/swc#2123) ➡️ write a custom walker
• SWC is not the fastest parser, too (https://github.com/oxc-project/bench-javascript-parser-written-in-rust)
• We will still do transforms/serializing AST in JS

---

To get real benefits, we should move the most of processing to Rust, change approaches and preferably use Deno or another runtime JS runtime to eval code.

And it's a complete rewrite 😊

[chin2km]

@layershifter Thanks for the explanations 👍🏻
Would be totally worth it since packages like linaria, pigment-css are gaining market share from styled-components.
Also, Vite with swc is on the rise.

Other than a complete rewrite, can other smaller effort changes improve things?
Like using alternatives like recast | esprima | acorn for estree ast generation. But in combination with @swc/core/transform. So basically we get the benefits of SWC transform. But use other faster alternatives to babel or even babel-parser itself to parse the swc transformed code. In effect, a complete rewrite would not be needed and we can still get benefits of fast SWC transforms?

I'm working on a huge project. Our vite build using @vitejs/plugin-react-swc takes ~2 minutes, but if I disable the linaria plugin and paths alias linaria with styled-components, the build time is just ~1.2 miniutes. Thats a big difference :(

[Anber]

Right now, we are experimenting with Rust and oxc.rc, but I can't promise anything.

[Anber]

A minor update: I managed to build a PoC on Rust, oxc and Deno. However, it is too early to talk about performance measurements since a minimal set of features is supported so far, and there is no caching at all.

[Anber]

The cli version shows almost x10 performance boost.