Created attachment 3327 [details] testcase

From hyatt, this'll be a pain to fix -- part of the reason for webkit doing DHTML faster Will be looking at adding a bit field in the next day or two...

Created attachment 3829 [details] patch keep track of child visibility in renderlayer

Created attachment 3830 [details] better test case

I would put the updates in appendChildNode and insertChildNode, since addChild is not called for all RenderTree manipulations (e.g., when generated content is used or anonymous blocks and continuations cause reshuffling of the tree).

Created attachment 4026 [details] updated patch updated patch based on daves comments

Comment on attachment 4026 [details] updated patch I know this is an entirely trivial comment, but there's no need for a return; at the end of RenderLayer::checkVisibility.

See this link: http://dbaron.org/log/2005-10#e20051013a We should try S5, since it sounds like it's a good test case for this.

Comment on attachment 4026 [details] updated patch + if (style()->visibility() != VISIBLE && newChild->style()->visibility()==VISIBLE && !newChild->layer()) { Style nit with the above, needs spaces before and after the ==. + if (!object->firstChild() && object->style()) { + m_visibilityKnown = true; + m_hasVisibleContent = object->style()->visibility() == VISIBLE; + } This seems like a needlessly aggressive optimization to me. checkVisibility() will quickly determine the same thing, right? Your change to collectLayers results in the entire layer tree being walked now. The old code would not descend into child layers if they weren't visible. You could further optimize by propagating bits up the layer tree (i.e., rather than just a "hasVisibleContent" in this layer bit, you could have a "hasVisibleContent" bit for the layer or its descendant layers.

I think I would have to add another bit for tracking if child layer visibility is known (or something comparable). I'm still not convinced this is worth it. Yes, we loop through child layers of an invisible layer which we did not do before, but only thing we do there is test the visibility bit. We do substantially more work if the test is positive, but that is the case we are fixing here in the first place.

Created attachment 10790 [details] another updated patch - update patch to current webkit - added m_hasVisibleDescendant bit to track visibility of child layers, avoiding the need to crawl through the entire layer tree to find visible layers

Created attachment 10791 [details] updated tests

I'll review this (soon).

Created attachment 10812 [details] cosmetic fixes

Comment on attachment 10812 [details] cosmetic fixes r=me, looks good.

*** Bug 5614 has been marked as a duplicate of this bug. ***

*** Bug 9286 has been marked as a duplicate of this bug. ***

*** Bug 7909 has been marked as a duplicate of this bug. ***

there's still a problem when the parent element has its z-index set to a number. adding z-index:0 to #main in antti's original attachment hides #monkoolia. see #nest on atptennis.com, for example.

(In reply to comment #20) > see #nest on atptennis.com, for example. looks like they changed/fixed this.