Improving the performance of WAT parser

[thramp]

We (the rust-analyzer team) have been aware of the slowness in Rowan for a while, but other things always took priority. Beyond allocation, Rowan is structured internally as a doubly-linked list to support mutating trees, but: 1. Mutation isn’t really worth it; the API isn’t user-friendly. 2. In most cases, it’s straight up faster to create a new parse tree and replace the existing one. Cache effects of a linked list vs. an arena!

In fairness, I don’t think we predicted just how large L1/L2 caches would get over the coming years.

[testdelacc1]

Is a rewrite of Rowan on the horizon?

[vjerancrnjak]

It’s funny how there is continuous reinvention of parsing approaches.

Why isn’t there already some parser generator with vector instructions, pgo, low stack usage. Just endless rewrites of recursive descent with caching optimizations sprinkled when needed.

[embedding-shape]

Hardware also changes across time, so while something that was initially fast, people with new hardware tries it, finds it now so fast for them, then create their own "fast X". Fast forward 10 more years, someone with new hardware finds that, "huh why isn't it using extension Y" and now we have three libraries all called "Fast X".

[zahlman]

Because you have to learn how to use any given parser generator, naive code is easy to write, and there are tons of applications for parsing that aren't really performance critical.

[high_na_euv]

I'd say because parsing is very specific kind of work heavily dependent on the grammar you're dealing with

[munificent]

A parser spends time:

1. Consuming tokens.

2. Recognizing the grammar.

3. Producing AST nodes.

Steps 1 and 3 are heavily dependent on the data types that make the most sense for the previous (lexing) and next (semantic analysis) phases of the compiler. There is no one Token type that works for every language, nor one AST type.

The recognizing the grammar part is relatively easy, but since so much of the code is consuming and producing datatypes that are unique to a given implementation, it's hard to have very high performance reusable libraries.

[mgaunard]

There are good parser generators, but potentially not as Rust libraries.

[writebetterc]

So it went from parsing at 25MiB/s to 115MiB/s. I feel like 115MiB/s is very slow for a Rust program, I wonder what it's up to that makes it so slow now. No diss to the author, good speedup, and it might be good enough for them.

[mananaysiempre]

115 MiB/s is something like 20 to 30 cycles per byte on a laptop, 50 on a desktop. That’s definitely quite slow as far as a CPU’s capacity to ingest bytes, but unfortunately about as fast as it gets for scalar (machine) code that does meaningful work per byte. There may be another factor of 2 or 3 to be had somewhere, or there may not be. If you want to go meaningfully faster, as in at least at the speed of your disk[1], you need to stop doing work per byte and start vectorizing. For parsers, that is possible but hard.

[1] https://www.youtube.com/watch?v=p6X8BGSrR9w

[high_na_euv]

"for Rust program"?

Isnt it more about the grammar than the prog lang?

[dfajgljsldkjag]

The performance gain from using a single shared vector for the nodes is pretty crazy. It just goes to show how much allocation overhead can slow things down if you are not careful.

[epage]

> Use hand-written parser > > The old parser was written with winnow which is a parser combinator library. While it’s easy to create a parser with parser combinators, it’s generally slower than a hand-written parser, so the first step is to write the parser by hands. Hand-written parser is not only faster but also allows to do more optimizations in the future.

Maintainer of Winnow here. I wish there were more details on this. I switched `toml` / `toml_edit` to being hand written and got some performance boost but I feel like the other things listed would have dwarfed the gains that I got. I wonder if there were sub optimal patterns they employeed that we could find ways to help better guide people.

For anyone going on about "hand written is always better", I disagree. Parser combinators offer a great way to map things back to grammar definitions which makes them much easier to maintainer. Only in extreme circumstances of features and/or performance does it seem worth going hand-written to me.

[gplane]

> "hand written is always better", I disagree. - Yep. As far as I know, winnow provides SIMD in some cases, while for hand written parsers, writing SIMD can be very hard.

[taylorallred]

It seems to me like parser combinators are always more trouble than they're worth. People often have the impression that parsing is difficult and should be outsourced to another library, but often it's pretty simple to hand-roll and usually it makes faster code.

[shevy-java]

Anyone using WebAssembly yet? HTML, CSS, JavaScript - all there.

Just about nobody uses WebAssembly. It first appeared almost ten years ago. This is snail-speed evolution at best.

[anonymous908213]

People use wasm for things that need wasm. My use case is my cross-platform game engine, because running both natively and in the browser was a priority for me. It is a wonderful tool and it is a truly magical feeling to see my native games running in the browser. But 99% of web developers are developing ordinary websites, so they don't need it. That's not an indictment of wasm.

[circuit10]

This is like saying "HTML, CSS and JavaScript are all widely used, but the webcam capture API is used way less, so obviously it's a failure"

In its current scope, WASM is a way to port existing code or accelerate certain computations, which only some applications need. Most websites don't need it, like how most sites don't need to use webcam capture; that doesn't mean it's not useful for those that do

[miki_oomiri]

You have the wrong understanding about wasm. It's absolutely not supposed to be replacing HTML, CSS or JS.

And yes wasm is used wildly. On the web for expensive computation (Google earth, figma, autocad, unity games) or server side for portability and sandboxing (Cloudflare workers, fastly, …)

[embedding-shape]

> Anyone using WebAssembly yet?

Yes, tons. Obviously not all, but large parts of these are WASM: https://itch.io/games/platform-web

Tools like Figma are only performant because of WASM.

[mickael-kerjean]

I use it extensively in my OSS work (https://github.com/mickael-kerjean/filestash) mostly around:

1. creating plugins that get executed in the browser to render files like Parquet, PSD, TIFF, SQLite, EPS, ZIP, TGZ, GIS related files and many more, where C libraries are almost always the reference implementations. There are almost a hundred supported file formats, most of which are supported through WASM

2. creating plugins that get executed in the server to generate your own endpoint or middleware while being sure you can't start exfiltrating data (which can be other people's files, and other sensitive stuff)

3. in the workflow engine to enable people to run their own sandboxed scripts without giving those a blank check to go crazy

[adzm]

I use a wasm xxhash implementation that is 40x faster than the fastest JavaScript version I can find. Drop in replacement. Call overhead is minimal, could be better with stringref if that ever gets available. Also some other audio analysis stuff in wasm I've been using is 400x faster than the JavaScript implementation but admittedly I just went straight to wasm rather than try to optimize the js in that case.

[onion2k]

I'm writing a point and click adventure game, and for that I've built a dialogue editor that uses a local text-to-speech model to turn speech into audio that runs in WASM (or WebGPU if it's available).

From what I can tell WASM is mostly being used to run big libraries from other languages in web apps. That's not a particularly common thing to need, so it's not commonly used. That doesn't mean it's moving too slowly.

[TkTech]

Sure, here's a Rust/WASM procedural skybox generator I threw together the other day, and is much, much faster at 16k renders then Javascript. https://tkte.ch/night-sky/

[demaga]

I saw a few web apps that use Rust crates for physics. I guess they must be using wasm?

[taminka]

wasm isn't meant to supersede html/css/js (unfortunately) and it's regularly used for high performance applications in the browser, web-based cad software, figma, youtube (i think they use wasm for codec fallback when support is spotty) etc

there is also games, stuff to do with video (ffmpeg built for wasm), ml applications (mlc), in fact it's currently impossible to use wasm w/o js to load the wasm binary

as a result, the web stack is a bit upside down now, w/o the seemingly "low level" and "high performance" parts over the slow bits (javascript)

[flohofwoe]

WebAssembly is a virtual ISA, not a replacement for HTML and CSS. It was also never meant to kill Javascript (which is actually a pretty nice language if you stick to the 'good parts' via Typescript and linting), but at most as an alternative or complement to JS, and as that WASM works really well.

[dagi3d]

Figma