Hacker News Reader

Comment by badminton1 | Hacker News Reader

Sad that Scala.js will not use WebAssembly, since WebAssembly doesn't provide GC.

WebAssembly will most likely eventually provide a GC, at which point we'll do what we can to support compiling to wasm.

They don't seem be on Chrome roadmap. We'll all grow full beards by the time DOM and GC become available. https://wasmdash.appspot.com

GC support can hardly be on Chrome's roadmap when the design of the spec has hardly anything more than "we want to have it, and here are some vague goals".

So yes, we'll grow beards, but it will happen. Remember though that wasm as whole happened in only a few years!

My two cents: to me WebAssembly seems like something Scala Native might address, more than Scala.js. Or even better, both project can sort of blend for this purpose as on one hand (SN) addresses LLVM compilation and the other (SJS) is perfect JS interop (not so sure how WebAssembly does that).

> not so sure how WebAssembly does that

Basically: badly. That's the number one reason we're not trying to compile Scala.js to wasm at the moment. For wasm to get the level of JS interop that Scala.js has, it needs to integrate with JS objects, which means with the JS GC, which means basically have a GC of its own. Hence we're waiting for the wasm GC spec to take life :)

I, sort of, know the limitations WebAssembly has right now when it comes to JS interop (such as only being able to use functions with numbers as parameters!) but my guess is this is temporary. It'd be curious to see experiments doing something Scala Native -> Emscripten -> JS-land.
Regardless, thanks so much for the work you're doing with Scala.ks (Scala is still my main language and I'm working on removing JS from my toolbox soon).

https://youtu.be/HktWin_LPf4?t=22m31s

Scala.js could write their own GC to hook into WASM as a library, and ship it with Scala.js until WASM has its own GC. Indeed, such competition could spur on the the WASM team(s) to accelerate work on GC.

That would not solve that essential part:

> it needs to integrate with JS objects, which means with the JS GC

Otherwise we would have done it already.

Could you not have a radical separation between data to be collected by Scala.js GC, and data GCed by Javascript? Data that crosses the boundary is deep-copied so as to avoid interference between both collectors?

Not good for performance, but possible in principle, and a stop-gap measure until the WASM GC lands.

I'm not sure what you mean by "deep-copied" in this case. But I suspect it would involve losing the identity of the objects, which is unacceptable for correctness.

Moreover, typical Scala.js code constructs cyclic dependencies between Scala objects and JS objects routinely. If Scala.js has its own GC, you'll have to manage those cyclic heap references, which will involve weak maps and other trickery so complicated that we would most likely lose all the performance that wasm would otherwise bring us.

It's just not worth it.

What do you exactly expect from targetting WASM instead of JS? Higher performance? Smaller bundles and faster load times? Better interoperability?

Better interoperability, definitely not. Scala.js already has the best interoperability out there, excluding languages with the same run-time semantics as JS (such as TypeScript). In fact, it is the lack of good interoperability in wasm that currently prevents us from compiling Scala.js to that target.

We could hope for smaller bundles and faster load times, yes. But mostly we should be able to get faster "cruising" execution. Scala.js being statically typed, there is a lot of static guarantees that we lose when compiling to JS. JS engines are very good and can "recover" a lot of it through JIT'ing, but it's still 3x slower than Scala/JVM on average (after warmup of both), so clearly we lose something non-negligible. By compiling to wasm, we would lose the JIT, but keep the static types. It is unclear how fast we would manage to be, but I would hope we can get within 2x of the JVM instead of 3x-ish.

There is a lot of misconception on the Internet that "wasm makes everything faster". For example, a lot of people want to "compile JS down to wasm" to make it faster. That's bullshit, because your language does not have static types to begin with, so you lose the JIT and gain nothing in exchange, which means you'll run much slower. But if you compile a statically typed language down to wasm, you should be able to gain some significant speed.

Yeah, but Scala started out on a JIT runtime. I'm skeptical a non-JIT would really be that much better.

As I said in the parent, in the case of Scala.js, I believe we lose more performance because of the loss of static types than we would if we lost the JIT.

Performance measurements in Scala Native, which has no JIT and is still quite young (so not very optimized yet), suggest than we can easily be within 2x of the JVM for Scala code, even without JIT.

> a lot of people want to "compile JS down to wasm" to make it faster. That's bullshit, because your language does not have static types to begin with, so you lose the JIT and gain nothing in exchange, which means you'll run much slower.

According to Mozilla's recent blog[1], WASM will in many cases be faster than JIT. In your experience is this not true?

[1] https://hacks.mozilla.org/2017/02/what-makes-webassembly-fas...

I read that article. They do not say anything wrong. But if you read between the lines, the article assumes your source language has types and does not have managed memory. Under those conditions, wasm will beat the JIT every time.

But if your language has dynamic types and/or managed memory (but mostly dynamic types), then you need a JIT to get anywhere close to decent performance.

The silent assumptions in that quite famous article is one reason many people believe that "compiling JS down to wasm" will make it faster. It is not the only article that has completely failed to point out the limitations, though. It's very sad to see, from where I stand.

Note that I am not the only one thinking this. Core wasm developers agree. [1]

[1] https://github.com/WebAssembly/design/issues/219

thanks, TIL. That Mozilla blog is misleading.

Reply on news.ycombinator.com