695592 - http://waltzstep.com/n22d/n22d.html is much slower in FF than Chrome

Status: RESOLVED INCOMPLETE

(Core :: JavaScript Engine, defect)

Description

[Jeff Muizelaar [:jrmuizel]]

I haven't profiled it yet, but my guess is that it's JS

Comment 1

[Jeff Muizelaar [:jrmuizel]]

The source for the slow one is at: https://github.com/adrianbg/n22d/commit/c8d171d3cc964bfb1f270b057532221289f0b61f

The hottest pieces are:
531.0ms    3.4%	 	js::mjit::stubs::FixupArity(js::VMFrame&, unsigned int)
424.0ms    2.7%	 	CallCompiler::generateNativeStub()
396.0ms    2.5%	 	bool js::gc::Arena::finalize<JSObject>(JSContext*, js::gc::AllocKind, unsigned long)

Comment 2

[Boris Zbarsky [:bzbarsky]]

Assuming I did the git magic right to check out the right thing....

Looks like 42% of the time is methodjit generated code, 7% is under array_slice, another 5% doing generateNativeStub stuff around the array_slice.  This is all under the NativeCall ic.

8% under array_extra, half of this calling back into JS.  5% allocating arrays directly from jitcode.  Lots of stubcalls (FixupArity, CreateThis, GetElem, StrictEq, Arguments, FlatLambda).

2% actual mjit compilation via TryCompile.

All those percentages are of total time, not main thread time.  Main thread is about 92% of total; the remainder looks like the GC finalization thread, almost all self time in FinalizeArenas.