Open Bug 1109816 Opened 5 years ago Updated 5 years ago

Devirtualize JS traversal in the cycle collector

Categories

(Core :: XPCOM, defect)

defect
Not set

Tracking

()

People

(Reporter: mccr8, Unassigned)

References

(Blocks 1 open bug)

Details

Attachments

(3 files)

In bug 1105069, Terrence is doing some epic work making GC things strongly typed, pushing it all the way into the CC.  I'm not sure how the details would quite work out, but I think we might be able to allocate GCThings in the CC graph into a separate set of nodes that would be JS::GCCellPtrs instead of a void* plus participant pair.  Then we could make a direct call the GC tracing code, rather than the indirect calls we have now, which could possibly be much faster.  We could even explicitly separate out scripts and objects, though that may be over kill.  Zones are rare enough I'd just leave them as regular CC participant based things.
Depends on: 1109928
Though a much simpler way to implement that would be to have a constant, non-null magic value for the participant for GC things, and then just test for that, and do a direct call.  You'd pay an extra branch against a constant compared to the other approach, but it would be much much simpler.
No longer depends on: 1109928
I hacked up a prototype patch that does what I said in comment 1, and it does seem to help a bit.  On a TechCrunch page close, with about 9000 refcounted objects and about 40000 JS objects, it reduced the total time spent marking from 48ms to 40ms, which cut out a slice.  Of course, that's just with one example of each, so maybe that's just noise.  And that shouldn't really affect max pause time.
Summary: Strongly type GC things in the CC graph → Stop using a participant for JS things in the CC graph
This probably doesn't matter given that both methods are defined in the same .cpp file but whatever.
We can just remove the participant now.  Split into a separate patch for simplicity.
Summary: Stop using a participant for JS things in the CC graph → Devirtualize JS traversal in the cycle collector
You need to log in before you can comment on or make changes to this bug.