827534 - Mark nsXBLDocumentInfo held by nsXULPrototypeCache black during GC

Status: RESOLVED INCOMPLETE

(Core :: XBL, defect)

Description

[Andrew McCreight [:mccr8]]

The times are kind of all over the place, but the CleanupJS times are similar between these two, so they are probably comparable. The time added to the GC was 129us.

Without patch:
  cc: MarkWindowList mediator took 410us
  cc: MarkWindowList watcher took 58us
  cc: AppShellService took 34us
  cc: xulCache took 1136us
  cc: CleanupJS took 1171us
  cc: TOTAL took 2855us

With patch:
  cc: MarkWindowList mediator took 268us
  cc: MarkWindowList watcher took 76us
  cc: AppShellService took 45us
  cc: xulCache took 35us
  cc: CleanupJS took 1209us
  cc: TOTAL took 1679us

As you can see, this knocks about 1ms off the time for nsCCUncollectableMarker::Observe.

Comment 1

[Olli Pettay [:smaug][[e-mail]]]

Which phase of GC does this make slower? Can we ensure we're not making a slow GC phase even
slower?

Comment 2

[Olli Pettay [:smaug][[e-mail]]]

But yes, this is great :)

Comment 3

[Andrew McCreight [:mccr8]]

It is making gray marking slower, which does tend to be one of the slower phases.

I'm not sure if a 0.129ms slower GC is worth the tradeoff for a 1ms fast CC cleanup. I could probably tweak it a bit (only looking at the globals instead of the protos too) and maybe that would make it faster.

Comment 4

[Andrew McCreight [:mccr8]]

Now that I think about it, mozilla::dom::TraceBlackJS is interesting because it isn't needed for correctness. Potentially we could shove it into its own slice, depending on how often this causes us to miss things.

Comment 5

[Andrew McCreight [:mccr8]]

Attachment: WIP, with timing cruft

Comment 6

[Andrew McCreight [:mccr8]]

I took some fresh measurements here, and the XUL cache stuff was taking less than 100us so this seems less important now.  I'm not sure why things are so much faster.

Comment 7

[Andrew McCreight [:mccr8]]

Attachment: time various phases of forget skippable

Comment 8

[Olli Pettay [:smaug][[e-mail]]]

When getting the times it would be good to check few different cases, like the one
having just couple of tabs or having lots of tabs.
And then what happens when closing a "heavy" tab etc.


(I'm still in process to get access to some games so we could check also how forget skippable behaves there)

Comment 9

[Olli Pettay [:smaug][[e-mail]]]

Messagemanager unmarking is kind of guaranteed to take more time the more we have tabs.
And looks like it is also a case to optimize for e10s/child-process.

Comment 10

[Andrew McCreight [:mccr8]]

This is probably okay for now...