[FastTextAutosizer] Fix the inline overlap issue.

Source/core/rendering/FastTextAutosizer.cpp

Index: Source/core/rendering/FastTextAutosizer.cpp
diff --git a/Source/core/rendering/FastTextAutosizer.cpp b/Source/core/rendering/FastTextAutosizer.cpp
index 3f0f5573fecb952b4728169817acdf3908740cdd..44a9838846fee583dd9180ac82b9b4cb0a9aec6b 100644
--- a/Source/core/rendering/FastTextAutosizer.cpp
+++ b/Source/core/rendering/FastTextAutosizer.cpp
@@ -802,6 +802,10 @@ void FastTextAutosizer::applyMultiplier(RenderObject* renderer, float multiplier
     style->setTextAutosizingMultiplier(multiplier);
     style->setUnique();
     renderer->setStyleInternal(style.release());
+
+    // Notify inline renderers that their width may have changed.
+    if (renderer->parent() && renderer->parent()->isRenderInline())
+        renderer->setPreferredLogicalWidthsDirty(MarkOnlyThis);

[skobes, 2014/03/31 17:19:58]

Do you know why this works?  I would have guessed MarkContainingBlockChain would
be needed here.

[pdr., 2014/04/01 06:06:43]

A good question. In looking into this I found our table layout code is very
broken. Lets look into that tomorrow (crbug.com/358510).

In standard block layout where the block contains inline children, FTA::inflate
gets called after the block's width is computed but before laying out the
children.
In normal layout the FTA flow looks like:
1) At the top of RenderBlock:Layout we do the inflation
2) During inflation we invalidate any pref widths of children
3) Later in RenderBlock::Layout we query the widths of all text nodes
(layoutInlineChildren)
4) Querying the widths with dirty pref widths will recompute them.

Because all children are walked, we can get away with MarkOnlyThis.

You may be wondering... how did my test exercise stale preferred logical widths?
It did, but only by accident. The two-pass layout caused by scrollbars caused
the pref widths to get invalidated. I've hidden the scrollbars so this test now
passes before and after my change, but I think this test is still useful to keep
around.

I've added a new test, table-inline-width, which does properly exercise the new
inline width code. It's not clear how inflateTable should work here so I've
moved the setPreferredLogicalWidthsDirty call into FTA::inflate until the table
code is more sane.

 }
 
 bool FastTextAutosizer::isWiderOrNarrowerDescendant(Cluster* cluster)