Get rid of inefficient uses of childNodeCount()

Source/core/dom/ContainerNode.cpp

Index: Source/core/dom/ContainerNode.cpp
diff --git a/Source/core/dom/ContainerNode.cpp b/Source/core/dom/ContainerNode.cpp
index d63b98e908235bd54bd1ea1d68fd517d04e89bab..f1b91faa149f45d0ac1ecf76ff5c3163f33ce5ca 100644
--- a/Source/core/dom/ContainerNode.cpp
+++ b/Source/core/dom/ContainerNode.cpp
@@ -537,7 +537,6 @@ void ContainerNode::removeChildren()
         RenderWidget::UpdateSuspendScope suspendWidgetHierarchyUpdates;
         {
             NoEventDispatchAssertion assertNoEventDispatch;
-            removedChildren.reserveInitialCapacity(childNodeCount());

[Inactive, 2014/02/19 04:13:33]

I don't think this one is worth it as we end up traversing the children twice
instead of once for a likely small win on Vector::append().

[eseidel, 2014/02/19 05:39:54]

I expect walking the child list twice may be cheaper than a possible extra
malloc.  But I could be wrong.  I would expect this change to be visible on
microbenchmarks (one way or the other).  This function has historically ben
quite hot.

[Inactive, 2014/02/19 14:28:15]

Ok then. I prefer to keep this change out of the CL. We can try this one out
separately if we want to.

             while (m_firstChild) {
                 removedChildren.append(m_firstChild);
                 removeBetween(0, m_firstChild->nextSibling(), *m_firstChild);