Bug 72284 – ASSERT in RenderLayerCompositor on /fast/multicol/pagination-v-horizontal-bt.html

• Component: New Bugs
• Status: NEW

[Alexis Menard (darktears)]

ASSERT in RenderLayerCompositor with Qt port.

[Alexis Menard (darktears)]

Created attachment 114972 [details]
Patch

[Alexis Menard (darktears)]

./WebKitTestRunner --verbose ../../../webkit/LayoutTests/fast/multicol/pagination-v-horizontal-bt.html

assert with the Qt port on WK2.

We use AC by default now.

The assert is 

0x00007f35eb3ba376 in WebCore::RenderLayerCompositor::computeCompositingRequirements (this=0x7f359403f7f0, layer=0x7f359403f6b8, overlapMap=0x7fff43cb51d0, compositingState=..., layersChanged=@0x7fff43cb523f)
    at ../../../../webkit/Source/WebCore/rendering/RenderLayerCompositor.cpp:758
758         ASSERT(willBeComposited == needsToBeComposited(layer));

The test triggers a style change in the root layer moving from a normal flow only to a different one.

I'm not very familiar with the code here but it seems that if a layer is not normal flow only it can't be composited but computeCompositingRequirements assume that if we are in composition mode (the state we are before) and AC is turned on (which is always true) then the root layer will be composited but according to canBeComposited the layer can be composited only if layer->isSelfPaintingLayer(); is true which is not true anymore as the latter check !isNormalFlowOnly().

Any suggestion on how I can improve the patch is welcome, or if I spot the right problem in the first place.

Thanks.

[Simon Fraser (smfr)]

canBeComposited() can return false for the root layer? That would break all kinds of stuff.

[Alexis Menard (darktears)]

(In reply to comment #3)
> canBeComposited() can return false for the root layer? That would break all kinds of stuff.

It doesn't seem that there is specific check for that. The root layer should be always composited not matter the style?

[Simon Fraser (smfr)]

(In reply to comment #4)
> (In reply to comment #3)
> > canBeComposited() can return false for the root layer? That would break all kinds of stuff.
> 
> It doesn't seem that there is specific check for that. The root layer should be always composited not matter the style?

If there are any other compositing layers, the root layer is composited, yes. If that's no longer true for multicol, then we'll have to fix it somehow.

[Simon Fraser (smfr)]

*** Bug 74999 has been marked as a duplicate of this bug. ***

[Balazs Kelemen]

(In reply to comment #5)
> (In reply to comment #4)
> > (In reply to comment #3)
> > > canBeComposited() can return false for the root layer? That would break all kinds of stuff.
> > 
> > It doesn't seem that there is specific check for that. The root layer should be always composited not matter the style?
> 
> If there are any other compositing layers, the root layer is composited, yes. If that's no longer true for multicol, then we'll have to fix it somehow.

But we are dealing with the root layer here, rigth? So, there is no other compositing layers yet, and we are deciding whether to composite the root layer. But ts is normalFlowOnly() => !isSelfPainting(), and canBeComposited say that such layers cannot be composited. But I guess that rather means such layers _should_ not be composited, because they are too primitive to be worth to composite them. Note that canBeComposited is called from needToBeComposited. Furthermore, there was no check for whether it is a root layer before http://trac.webkit.org/changeset/63149. I think the preassumption that the root layer must be composited if we have entered into compositing mode is wrong. I don't know exactly what it means "entering into compositing mode" but it seems to me that it doesn't mean to force compositing to all kinds of content. Well, I'm not familiar with the AC machinery, I just try to interpret what I see :) What do you think?

[Balazs Kelemen]

Test results (locally, debug) with the patch: no unexpected regression, no crash on pagination-* tests.