Suppress layer creation for descendants of GPU-rasterized layers

Source/core/rendering/RenderObject.cpp

 /*
  * Copyright (C) 1999 Lars Knoll (knoll@kde.org)
  *           (C) 1999 Antti Koivisto (koivisto@kde.org)
  *           (C) 2000 Dirk Mueller (mueller@kde.org)
  *           (C) 2004 Allan Sandfeld Jensen (kde@carewolf.com)
  * Copyright (C) 2004, 2005, 2006, 2007, 2008, 2011 Apple Inc. All rights reserved.
  * Copyright (C) 2009 Google Inc. All rights reserved.
  * Copyright (C) 2009 Torch Mobile Inc. All rights reserved. (http://www.torchmobile.com/)
  *
  * This library is free software; you can redistribute it and/or
@@ skipping 1830 matching lines @@
             RenderBlock* block = toRenderBlock(parent())->createAnonymousBlock();
             RenderObjectChildList* childlist = parent()->virtualChildren();
             childlist->insertChildNode(parent(), block, this);
             block->children()->appendChildNode(block, childlist->removeChildNode(parent(), this));
         }
     }
 }
 
 StyleDifference RenderObject::adjustStyleDifference(StyleDifference diff, unsigned contextSensitiveProperties) const
 {
+    // FIXME: The calls to hasDirectReasonsForCompositing are using state that may not be up to date.
+    DisableCompositingQueryAsserts disabler;

[abarth-chromium, 2014/04/17 17:29:51]

This should be asking whether we have any style-determined reasons for
compositing.

Maybe this will cause a bug with your CL because we'll think we have compositing
reasons but then we won't actually create a graphics layer because we'll
suppress the creation of the layer?

[ajuma, 2014/04/17 19:24:44]

Yes, switching to styleDeterminedReasons() does cause a bug, where we fail to
repaint because we think we have a composited layer when we don't have one (test
case: http://jsfiddle.net/c59gC/). I'll switch this to styleDeterminedReasons()
(and fix the resulting bug) in a follow-up.

     // If transform changed, and the layer does not paint into its own separate backing, then we need to do a layout.
     // FIXME: The comment above is what the code does, but it is technically not following spec. This means we will
     // not to layout for 3d transforms, but we should be invoking a simplified relayout. Is it possible we are avoiding
     // doing this for some performance reason at this time?
     if (contextSensitiveProperties & ContextSensitivePropertyTransform) {
         // Text nodes share style with their parents but transforms don't apply to them,
         // hence the !isText() check.
         // FIXME: when transforms are taken into account for overflow, we will need to do a layout.
         if (!isText() && (!hasLayer() || !toRenderLayerModelObject(this)->layer()->hasDirectReasonsForCompositing())) {
             // We need to set at least SimplifiedLayout, but if PositionedMovementOnly is already set
@@ skipping 1512 matching lines @@
 {
     if (object1) {
         const WebCore::RenderObject* root = object1;
         while (root->parent())
             root = root->parent();
         root->showRenderTreeAndMark(object1, "*", object2, "-", 0);
     }
 }
 
 #endif