795778 - Significant performance regression calling Array.prototype.slice on a NodeList due to handle API stuff

Status: RESOLVED WORKSFORME

(Core :: JavaScript Engine, defect)

Description

[Boris Zbarsky [:bzbarsky]]

The patch in bug 750733 significantly (30% or more) regressed the testcase in bug 697343, as far as I can tell.

Still trying to figure out _why_ exactly...

Comment 1

[Boris Zbarsky [:bzbarsky]]

Well, the initial regression was because it broke the DoGetElement overloads.  But that seems to have gotten mostly fixed in bug 776579, and the numbers didn't go back down to where they should be...

Comment 2

[Boris Zbarsky [:bzbarsky]]

And in particular, if I actually compare a current build to one from back in May in a profiler, it looks like various stuff is not getting inlined that used to be, and more time is generally being spent on every method call.  Or something.  It's a bit hard to make sense of the data.  :(

Comment 3

[Boris Zbarsky [:bzbarsky]]

I _think_ this might be due to JSObject::getElementIfPresent not being inlined.  Maybe.  I'm not sure whether the old thing got inlined either...

Comment 4

[Boris Zbarsky [:bzbarsky]]

But I think it did.

Comment 5

[Boris Zbarsky [:bzbarsky]]

We killed all this code in bug 697343.

Comment 6

[Terrence Cole [:terrence]]

Great! Eric and I also walked through the optimized bindings instruction-by-instruction at the GC work week and found that everything is working as expected.