TabContext.onGC has bad scalability on FF

[the8472]

I've obtained the following profile on a FF session with µMatrix and many tabs.

As you can see the onGC function spends many seconds in indexFromBrowser. I think this might be because tabbrowser.browsers does some XUL and proxy object hocuspocus which then leads to a very inefficient indexOf

I think simply keeping a WeakMap of => ID and a Map of ID => weakreference() should reduce the need for cleanup.

All that would be left would be scrubbing of IDs that map to nulled weak references.

[gorhill]

Thanks for the investigation/insight, I didn't know about Components.utils.getWeakReference() -- that would solve the issue I was facing not being able to iterate through a WeakMap. However there is this part in the doc:

just because get() succeeds doesn't mean your object is valid. It might simply not have been garbage collected yet

So the way I understand this, I would have to double-check whether the weak browser reference is really valid by looking up => ID?

[the8472]

This probably refers to weak references simply not being aware of any internal lifecycle that an object might have. I.e. they don't get nulled magically just because you close/invalidate/shutdown some object that supports such things.

Whether that's actually a problem depends on downstream consumers of the object returned from the weak reference. It's no different from fishing some old dom node that has long been detached or a channel that has been closed out of some javascript variable.

[the8472]

Ah, I looked at the code. In addition to the weakmap/ref stuff this check could be replaced with something else that doesn't need the index operation.

Checking whether the <browser> is still attached to the DOM might be the most efficient way.

Alternatively there is browser.ownerDocument.defaultView.top.gBrowser.getTabForBrowser(browser), but I don't know whether that's any more efficient, I think it might also iterate internally. Edit: it uses a weakmap internally, so it should be a fast way to access it.

[gorhill]

If this works well, I will import the changes into uBO as well.

[the8472]

Looks good. Haven't seen any more hangs on 0.9.3.5rc1

[sheepdestroyer]

Running the RC with this fix since release, no pb. Is is still planned to port that fix to uBO?