Common JS shell infrastructure

[lars-t-hansen]

trafficstars

cc @jfbastien @saambarati @kmiller68 @binji @MikeHolman

JF brought up the issue that we'll want a shared system in the JS shells to be able to test wasm threaded code. (And JS code for that matter - everyone will love us.)

To kick us off, I propose the following strawman:

The main thread can create new workers:

var w = new Worker(text)  // where text is defined below
w.postMessage(obj)           // what you think
w.onmessage = function (ev) { ... } // where ev.data is something sent by postMessage
dispatch()                           // enter the event loop

The worker can then:

onmessage = function (ev) { ... } where ev.data is something sent by postMessage
postMessage(obj)       // what you think

Note,

• There is an implicit event loop in the worker.
• There is an explicit event loop in the main thread by means of dispatch(); when it returns the main thread continues and may exit. (Debatable how good this explicit event loop is, not sure what d8 does here. Discuss.)
• The worker cannot create new workers (keeps things simple).
• The object to be posted is anything structured-cloneable
• I haven't bothered to worry about transfering objects here

The "text" for a worker is a string that starts with 'text/plain;', this is for web compat so that tests can just carry over (ideally).

[binji]

This is pretty close to what is currently implemented in d8. The primary difference is that there is currently no event loop on the main thread. Instead, messages are queued:

// Main thread.
var w = new Worker(text);
w.postMessage(obj);
w.getMessage();  // Get object from worker, or undefined if none available.
w.terminate();

// Worker.
postMessage(obj);
onmessage = function(m) {
  ...
};

• There is an implicit event loop in the worker.
• There is no event loop on the main thread.
• The worker can't create new workers
• Objects posted must be structured-cloneable or transferable
• Text is just raw text.

[lars-t-hansen]

@binji, so you just end up busy-waiting on the main thread? Or do you also have a sleep() like the SpiderMonkey shell has? Busy-waiting is a little nasty but I suppose we can live with it if we must.

It looks like for d8 one could perhaps create a dispatch function that just gets messages from the workers (round-robin with some kind of relaxation?) and dispatches them to installed onmessage handlers, if we wanted to find common ground and be closer to the web platform.

[binji]

Yeah, the tests just spin wait. But @mtrofin pointed out that we've recently added support for using an event loop on the main thread, so perhaps we can just rework to use that.

[mtrofin]

We have a testRunner.waitUntilDone() and testRunner.notifyWait(), akin what's used in LayoutTests. The former starts an event loop on the main thread. The latter ends it when all pending work is done.

We don't busy-wait, instead, we wait on a semaphore to get more work.

[saambarati]

@lars-t-hansen This proposal looks good to me. I like it being similar to web spec.

What dictates when the event loop can return back inside dispatch()?

[lars-t-hansen]

@saambarati, I think we need something like the v8 infrastructure has, eg, a returnFromDispatch() function that can be invoked by one of the registered event handlers on the main thread when it discovers that the execution can end.

I think that with an event loop on the main thread we won't need the w.getMessage() that the v8 shell has; that seems to be an artifact of not having an event loop previously.

[lars-t-hansen]

I've been stalled on this issue, no work since July. Hope to get to it soon.

[lars-t-hansen]

There's a concrete API description (based closely on the discussion above) and a sample implementation for the SpiderMonkey shell here: https://github.com/lars-t-hansen/Worker. See the file API.md in that repo. Note the "Notes" and "Questions" sections in that doc.

@saambarati, @binji, @kmiller68, @mtrofin, @MikeHolman - is this something you can support in your shells? Obviously we can bicker about names, notably enterEventLoop() and exitEventLoop() are new.

@binji, do we really need a w.getMessage() if we have event handling on both sides?

As far as the discussion is concerned, probably best to keep it in this ticket, so that we have it all in one place.

[jfbastien]

We'll track our implementation here: https://bugs.webkit.org/show_bug.cgi?id=179319

[binji]

@lars-t-hansen No, getMessage isn't necessary. I added it because we didn't have event loop support in d8 at the time, but it would be better to have the API more closely match the browser, IMO.