Migrate neon to napi-rs

[Brooooooklyn]

Started at napi-rs/node-rs#34

Benefits

• Minimal overhead
• Stable native API, easy to release Releases

First benchmark in my local machine

swc (es3) x 695 ops/sec ±1.92% (82 runs sampled)
swc (es5) x 699 ops/sec ±1.14% (83 runs sampled)
swc (es2015) x 667 ops/sec ±1.70% (85 runs sampled)
swc (es2016) x 1,654 ops/sec ±1.62% (84 runs sampled)
swc (es2017) x 1,630 ops/sec ±2.06% (85 runs sampled)
swc (es2018) x 2,177 ops/sec ±1.63% (88 runs sampled)
@node-rs/swc (es2018) x 5,538 ops/sec ±1.04% (86 runs sampled)
swc-optimize (es3) x 1,555 ops/sec ±1.31% (85 runs sampled)
babel (es5) x 32.78 ops/sec ±4.99% (52 runs sampled)

[kdy1]

@Brooooooklyn Does it support invoking javascript function from rust (in a thread created by rayon) and awaiting promise from rust (again, within any thread)?

I was considering switching to node-bindegen, but I'm open with other options, if the features above are implemented in easy-to-use manner.

[Brooooooklyn]

Does it support invoking javascript function from rust

ThreadSafeFunction has not been implemented yet, I'm planning to finish it in this month.

But you can spawn task to libuv thread via napi_create_async_work, see https://github.com/napi-rs/node-rs/pull/34/files#diff-ec133f03d8eaec6d1f6e2e0bffee7e50R64 . So you can perform spawn(task) -> promise easily

node-bindegen is a great library, but the overhead is still too large if you are considering about performance: https://github.com/jdsaund/rust-node-perf

[kdy1]

@Brooooooklyn I think the benchmark does not fit for swc. The benchmark is about adding two numbers, but swc does lots of things in the background task.

I didn't update the README.md because I'm too lazy to do it, swc is 20+ faster than babel even with neon. Clearly, ffi overhead is a bottleneck in swc, but I'm not sure if it worths migration.

[Brooooooklyn]

I didn't update the README.md because I'm too lazy to do it

The benchmark I mentioned above is produced in my machine right now... and it's over two times faster @node-rs/swc (es2018) x 5,538 ops/sec ±1.04% (86 runs sampled) vs swc (es2018) x 2,177 ops/sec ±1.63% (88 runs sampled) if my implement was not wrong. I think this gain of performance is really worth to consider about. Especially napi-rs won't bring more effort in native side.

The benchmark is about adding two numbers, but swc does lots of things in the background task

I agree with this point, and I'm planning to setup more benchmark suites in napi-rs project.

[kdy1]

@Brooooooklyn Wow, it's cool. I'll see how much performance differs in the case of almost all work being done on native.

[Brooooooklyn]

Does it support invoking javascript function from rust

napi@0.4 support it now

[kdy1]

@Brooooooklyn Is there any documentation?
I want to improve swc and spack.

[Brooooooklyn]

I will write a simple guide in this week, and you can also find the api use cases in this project: https://github.com/napi-rs/node-rs

[kdy1]

@Brooooooklyn Can you help migration?
How can I expose a class from rust code?

[kdy1]

@Brooooooklyn Also, do you have a plan to provide an alternative neon_serde? swc requires passing object from each side to the other side