Don't access DisplayItemClient::VisualRect() for cached display items.

third_party/WebKit/Source/platform/graphics/ContentLayerDelegate.cpp

Index: third_party/WebKit/Source/platform/graphics/ContentLayerDelegate.cpp
diff --git a/third_party/WebKit/Source/platform/graphics/ContentLayerDelegate.cpp b/third_party/WebKit/Source/platform/graphics/ContentLayerDelegate.cpp
index 65aca21855cf57ef77f0deb8a28a7a81dd7de2a6..05702a7d4cf8766d6537cfb8282c3a5d64153d0a 100644
--- a/third_party/WebKit/Source/platform/graphics/ContentLayerDelegate.cpp
+++ b/third_party/WebKit/Source/platform/graphics/ContentLayerDelegate.cpp
@@ -90,6 +90,7 @@ void ContentLayerDelegate::PaintContents(
     graphics_layer_->Paint(nullptr, disabled_mode);
 
   paint_controller.GetPaintArtifact().AppendToWebDisplayItemList(
+      graphics_layer_->OffsetFromLayoutObjectWithSubpixelAccumulation(),

[wkorman, 2017/05/19 00:55:08]

Does this add some small overhead vs. ToT, as (I think) we used to only do this
once per new item and save thereafter, whereas now we do it each time we append
regardless of old or new?

Seems ok but figured worth checkpoint as part of the review.

[Xianzhu, 2017/05/19 01:25:11]

Good catch.

The overhead happens when the GraphicsLayer has nothing invalidated, and we
reuse the whole cached DisplayItemList. I think we can optimize the case by also
reusing the original WebDisplayItemList instead of reconstructing it from the
cached DisplayItemList.

If anything is invalidated, regardless of this patch, we re-apply the offset for
each cached or new display item. Previously we called
DisplayItemClient::VisualRect() and now we call DisplayItem::VisualRect()
instead. The latter might be slightly faster because the call is inlined instead
of virtual.

Let's see how this affects performance tests.

       web_display_item_list);
 
   paint_controller.SetDisplayItemConstructionIsDisabled(false);