660959 - Figure out a better setup for figuring out link state

Status: RESOLVED FIXED

(Core :: DOM: Core & HTML, defect)

Description

[Boris Zbarsky [:bzbarsky]]

It's expensive, so we do it lazily.... the problem is when.  Right now we have hacks in style resolution to trigger it, basically, but that's not so hot if we want to parallelize.  Can we do better?

Comment 1

[David Zbarsky (:dzbarsky)]

Attachment: Update link state before resolving style

Comment 2

[Boris Zbarsky [:bzbarsky]]

Comment on attachment 541461 [details] [diff] [review]
Update link state before resolving style

Does this not break querySelector?

What about resolutions coming via ReResolveStyleContext?

I'm pretty sure this change is wrong....

Comment 3

[David Zbarsky (:dzbarsky)]

Attachment: Update link state less lazily

Comment 4

[Boris Zbarsky [:bzbarsky]]

Comment on attachment 542144 [details] [diff] [review]
Update link state less lazily

Hmm.  I think this should work, but can you add some of the tests your initial patch would have failed, please?

And fix the commit message!

Comment 5

[David Zbarsky (:dzbarsky)]

This patch does have tests.  Do you mean some other tests?

Comment 6

[David Zbarsky (:dzbarsky)]

Attachment: Patch to check in r=bz

Comment 7

[Boris Zbarsky [:bzbarsky]]

Comment on attachment 542339 [details] [diff] [review]
Patch to check in r=bz

Er, I somehow completely missed the tests, sorry.

The first test is wrong.  It needs to do $("content").querySelector, and ideally put an id on the <a> and make sure querySelector returned the right thing.  So something like:

  <div id="content" style="display: none">
    <a href="#" id="testa"></a>
  </div>

  <script>
    is($("content").querySelector(":link, :visited"), $("testa"),
       "Should find a link even in a display:none subtree");
  </script>

The second test is fine, though you can replace ".getPropertyValue('color').toString()" with ".color" there.

Comment 8

[David Zbarsky (:dzbarsky)]

Attachment: Patch

Comment 9

[Boris Zbarsky [:bzbarsky]]

Comment on attachment 542592 [details] [diff] [review]
Patch

r=me

Comment 10

[David Zbarsky (:dzbarsky)]

Attachment: Patch

This one doesn't have a slowdown on the selector test

Comment 11

[Boris Zbarsky [:bzbarsky]]

Comment on attachment 543006 [details] [diff] [review]
Patch

I just realized that the nsGlobalWindow change is wrong.  You should QI to nsIContent, then check that it's non-null and IsElement() and then call AsElement().... r=me with that.

Comment 12

[David Zbarsky (:dzbarsky)]

Attachment: Patch to check in r=bz

Comment 13

[Boris Zbarsky [:bzbarsky]]

Actually, I think there's still a problem here...  Consider a display:none subtree containing this markup:

  <a href="http://www.example.com"></a>
  <div id="foo">
    <span></span>
  </div>

And a call to document.getElementById("foo").querySelector(":link + * span, :visited + * span").

This should return the span.  I bet with your changes it doesn't....

Comment 14

[David Zbarsky (:dzbarsky)]

Attachment: Patch

Comment 15

[Boris Zbarsky [:bzbarsky]]

Two comments offhand:

1)  Why not just use the hashtable as a member, instead of pointer-to-hashtable?
    This hashtable will be used on basically every page....
2)  You probably need to either have a scriptblocker in FlushPendingLinkUpdates
    or stick all the element pointers into an array before making the
    RequestLinkStateUpdate calls, because I don't quite trust those calls to not
    add/remove things in this hashtable.  Scriptblocker is probably simpler.

I'll look through the rest carefully tomorrow.

Comment 16

[David Zbarsky (:dzbarsky)]

Attachment: Patch

Dromaeo DOM results

Before patch:
    appendChild: 653.94runs/s ±0.89%
    insertBefore: 442.44runs/s ±1.46%

After patch:
    appendChild: 643.07runs/s ±0.99%
    insertBefore: 437.16runs/s ±2.30%

Comment 17

[Boris Zbarsky [:bzbarsky]]

Comment on attachment 545959 [details] [diff] [review]
Patch

Please document your new methods and member.  

I don't understand the block at the beginning of UnregisterPendingLinkUpdate.  Why is it there?

In UnbindFromTree, you can set doc to GetCurrentDoc so you only unregister when you might possibly be registered.

To address the Element precondition thinkg you added, do you want to just switch the document hashtable to using Link instead of Element?  That seems like it would be cleaner...

r=me with those changes.  Thank you!

Comment 18

[David Zbarsky (:dzbarsky)]

Attachment: Patch r=bz

> I don't understand the block at the beginning of UnregisterPendingLinkUpdate.  > Why is it there?

I removed the Get call. There's no point in performing the hash lookup if there are no elements in the table.

Comment 19

[Boris Zbarsky [:bzbarsky]]

That last patch still needs comments.  It's also missing the changes to Link.h and Link.cpp, and the nsDocument code is still trying to call RequestLinkStateUpdate on Element.  Did you forget to qref somewhere?

Comment 20

[David Zbarsky (:dzbarsky)]

Attachment: Patch r=bz

This time with comments

Comment 21

[Boris Zbarsky [:bzbarsky]]

It's too bad that you can't call the method Element().  At least document that it can't return null?

Comment 22

[David Zbarsky (:dzbarsky)]

Attachment: Patch r=bz

Comment 23

[David Zbarsky (:dzbarsky)]

Attachment: Patch r=bz

Comment 24

[Boris Zbarsky [:bzbarsky]]

Attachment: Same thing, updated to tip so it can actually land

Comment 25

[Boris Zbarsky [:bzbarsky]]

https://hg.mozilla.org/integration/mozilla-inbound/rev/61094ae6da18

Comment 26

[Ed Morley [:emorley]]

https://hg.mozilla.org/mozilla-central/rev/61094ae6da18