Bug 161879 – Speed up Function.prototype.bind a bit by making it a builtin

• Component: JavaScriptCore
• Status: RESOLVED FIXED

[Saam Barati]

Speedometer spends about 10% of its time in this function. I have a 10-15% speedup by rewriting it as a builtin.
This isn't going to be a huge win for speedometer, but since 10-15% is significant, we might as well land it.

[Oliver Hunt]

There's an older bug somewhere where i did this with a bunch of the required logic. There were issues i was having making the logic universally faster at that time.

(Part of the reason for .@call and .@apply exist :D )

[Saam Barati]

(In reply to comment #1)
> There's an older bug somewhere where i did this with a bunch of the required
> logic. There were issues i was having making the logic universally faster at
> that time.
> 
> (Part of the reason for .@call and .@apply exist :D )

I don't think this needs to @call or @apply anything. It's basically
just inspecting arguments, and maybe building an array. And doing
some property accesses. So I think property accesses, argument length
introspection, happen to be faster in JS, especially once we get to
the JITs and the call may be inlined. I needed to make a few helper
functions for a few operations that Function.prototype.bind does.

[Saam Barati]

Created attachment 288640 [details]
patch

[Saam Barati]

Created attachment 288641 [details]
patch

[Build Bot]

Comment on attachment 288641 [details]
patch

Attachment 288641 [details] did not pass mac-wk2-ews (mac-wk2):
Output: http://webkit-queues.webkit.org/results/2062262

New failing tests:
js/dom/function-bind.html

[Build Bot]

Created attachment 288649 [details]
Archive of layout-test-results from ews107 for mac-yosemite-wk2

The attached test failures were seen while running run-webkit-tests on the mac-wk2-ews.
Bot: ews107  Port: mac-yosemite-wk2  Platform: Mac OS X 10.10.5

[Oliver Hunt]

Comment on attachment 288641 [details]
patch

Oh, i see what you're doing - the prior work endeavored to avoid a custom object for jsboundfunction.

[Filip Pizlo]

(In reply to comment #7)
> Comment on attachment 288641 [details]
> patch
> 
> Oh, i see what you're doing - the prior work endeavored to avoid a custom
> object for jsboundfunction.

@Oliver: not sure if this helps add context, but JSC already has super fast bound function calls for the common case that you didn't add any additional args.  It's the actual function.bind - the creation of the bound function object - that is slow right now.

[Saam Barati]

Created attachment 288651 [details]
patch

update test results for new error message.

[Saam Barati]

(In reply to comment #7)
> Comment on attachment 288641 [details]
> patch
> 
> Oh, i see what you're doing - the prior work endeavored to avoid a custom
> object for jsboundfunction.

Yeah it might be worth it to make this even faster. This was just low
hanging fruit, so I decided to translate the C function into JS.
However, to make this faster, I think we'd ideally want to be able
to inline the allocation of the actual bound function cell into the
JITs. This would be nice. There are some weird things that constructor does,
so it would probably require more work than just translating the creation into
assembly.

[WebKit Commit Bot]

Comment on attachment 288651 [details]
patch

Clearing flags on attachment: 288651

Committed r205848: <http://trac.webkit.org/changeset/205848>

[WebKit Commit Bot]

All reviewed patches have been landed.  Closing bug.

[Oliver Hunt]

(In reply to comment #10)
> (In reply to comment #7)
> > Comment on attachment 288641 [details]
> > patch
> > 
> > Oh, i see what you're doing - the prior work endeavored to avoid a custom
> > object for jsboundfunction.
> 
> Yeah it might be worth it to make this even faster. This was just low
> hanging fruit, so I decided to translate the C function into JS.
> However, to make this faster, I think we'd ideally want to be able
> to inline the allocation of the actual bound function cell into the
> JITs. This would be nice. There are some weird things that constructor does,
> so it would probably require more work than just translating the creation
> into
> assembly.

The old patch literally had code along the lines of

if (arguments.length == 1) {
   let thisValue = arguments[0];
   let self = this;
   return function() { return self.@apply(thisValue, arguments); }
}

let args = [...arguments]
let thisValue = args.shift(); // i think?
return function() {
   return function() {
       if (arguments.length == 0)
           return self.@apply(thisValue, args);
       return self.@apply(thisValue, [...args, ...arguments]);
   }
}

[Filip Pizlo]

(In reply to comment #13)
> (In reply to comment #10)
> > (In reply to comment #7)
> > > Comment on attachment 288641 [details]
> > > patch
> > > 
> > > Oh, i see what you're doing - the prior work endeavored to avoid a custom
> > > object for jsboundfunction.
> > 
> > Yeah it might be worth it to make this even faster. This was just low
> > hanging fruit, so I decided to translate the C function into JS.
> > However, to make this faster, I think we'd ideally want to be able
> > to inline the allocation of the actual bound function cell into the
> > JITs. This would be nice. There are some weird things that constructor does,
> > so it would probably require more work than just translating the creation
> > into
> > assembly.
> 
> The old patch literally had code along the lines of
> 
> if (arguments.length == 1) {
>    let thisValue = arguments[0];
>    let self = this;
>    return function() { return self.@apply(thisValue, arguments); }
> }
> 
> let args = [...arguments]
> let thisValue = args.shift(); // i think?
> return function() {
>    return function() {
>        if (arguments.length == 0)
>            return self.@apply(thisValue, args);
>        return self.@apply(thisValue, [...args, ...arguments]);
>    }
> }

I don't think this is as good though.  For example:

    return function() { return self.@apply(thisValue, arguments); }

This is worse than the current special bound function with its special JIT thunk.  That thunk means that even before any warmup, bound function calls are fast.

This would mean that bound function calls would only become fast if you did enough of them.