622302 - Figure out a way to call things sort of like traceable natives for IM, using type inference

Status: REOPENED

(Core :: JavaScript Engine, defect)

Description

[Boris Zbarsky [:bzbarsky]]

That would significantly speed up DOM bindings, by reducing the need for type conversions.

Comment 1

[Till Schneidereit [:till]]

JM is gone

Comment 2

[Boris Zbarsky [:bzbarsky]]

We might want something like this for IM, though...

Comment 3

[Jan de Mooij [:jandem]]

(In reply to Boris Zbarsky (:bz) (reading mail, but on paternity leave) from comment #2)
> We might want something like this for IM, though...

Is this still worth it for DOM calls with the jitinfo stuff etc we have now?

Comment 4

[Boris Zbarsky [:bzbarsky]]

That's a good question.

The primary benefit of the traceable natives was that they avoided the need for argument and return value boxing/unboxing.  The jitinfo stuff does nothing to address this (though it helps with this-value unboxing, and helps speed up return value unboxing a bit by communicating the general kind of return value).

It's pretty easy to write testcases in which the boxing and unboxing is a significant fraction of the time; the question is whether these commonly matter in web code...

The main likely beneficiary would presumably be WebGL; roc has had some thoughts about it in the past.

Comment 5

[Luke Wagner [:luke]]

Ah, so IIUC, the idea would be to have, hanging off a DOM builtin's jitinfo, an array of argument/return-type-specialized overloads that could be called directly?  ISTR, that's what we did for traceable natives.

Comment 6

[Boris Zbarsky [:bzbarsky]]

Or something along those lines, yes.

Comment 7

[Robert O'Callahan (:roc) (email my personal email if necessary)]

I thought for each WebIDL method we'd have one function whose signature matches the corresponding C++ method (with 'this' passed as an explicit parameter), which tail-calls the C++ method. Then we could have IM assemble direct calls to this function, under the right conditions (e.g., all parameter types known).

Currently I don't know of any benchmarks that definitely need this. The closest I know of is canvas. So let's leave this bug open but not work on it until we have a benchmark.