cc: Use a ListContainer for DisplayItemList to reduce allocations.

cc/resources/display_item_list.cc

Index: cc/resources/display_item_list.cc
diff --git a/cc/resources/display_item_list.cc b/cc/resources/display_item_list.cc
index 69bdd03a87474207a764abe851fb3585ab759184..12550b158718db5fbfbac9c6fb699fb87c20d048 100644
--- a/cc/resources/display_item_list.cc
+++ b/cc/resources/display_item_list.cc
@@ -12,6 +12,7 @@
 #include "cc/debug/picture_debug_util.h"
 #include "cc/debug/traced_picture.h"
 #include "cc/debug/traced_value.h"
+#include "cc/resources/largest_display_item.h"
 #include "third_party/skia/include/core/SkCanvas.h"
 #include "third_party/skia/include/core/SkDrawPictureCallback.h"
 #include "third_party/skia/include/core/SkPictureRecorder.h"
@@ -31,18 +32,24 @@ bool PictureTracingEnabled() {
   return tracing_enabled;
 }
 
+const int kDefaultNumDisplayItemsToReserve = 100;
+
 }  // namespace
 
 DisplayItemList::DisplayItemList(gfx::Rect layer_rect,
                                  bool use_cached_picture,
                                  bool retain_individual_display_items)
-    : recorder_(new SkPictureRecorder()),
+    : items_(LargestDisplayItemSize(), kDefaultNumDisplayItemsToReserve),
+      recorder_(new SkPictureRecorder()),
       use_cached_picture_(use_cached_picture),
       retain_individual_display_items_(retain_individual_display_items),
       layer_rect_(layer_rect),
       is_suitable_for_gpu_rasterization_(true),
       approximate_op_count_(0),
       picture_memory_usage_(0) {
+#if DCHECK_IS_ON()
+  needs_process_ = false;
+#endif
   if (use_cached_picture_) {
     SkRTreeFactory factory;
     recorder_.reset(new SkPictureRecorder());
@@ -72,12 +79,12 @@ DisplayItemList::~DisplayItemList() {
 void DisplayItemList::Raster(SkCanvas* canvas,
                              SkDrawPictureCallback* callback,
                              float contents_scale) const {
+  DCHECK(ProcessAppendedItemsCalled());
   if (!use_cached_picture_) {
     canvas->save();
     canvas->scale(contents_scale, contents_scale);
-    for (size_t i = 0; i < items_.size(); ++i) {
-      items_[i]->Raster(canvas, callback);
-    }
+    for (auto* item : items_)
+      item->Raster(canvas, callback);
     canvas->restore();
   } else {
     DCHECK(picture_);
@@ -99,7 +106,44 @@ void DisplayItemList::Raster(SkCanvas* canvas,
   }
 }
 
+void DisplayItemList::ProcessAppendedItemsOnTheFly() {
+  if (retain_individual_display_items_)
+    return;
+  if (items_.size() >= kDefaultNumDisplayItemsToReserve) {
+    ProcessAppendedItems();
+    // This function exists to keep the |items_| from growing indefinitely if
+    // we're not going to store them anyway. So the items better be deleted
+    // after |items_| grows too large and we process it.
+    DCHECK(items_.empty());
+  }
+}
+
+void DisplayItemList::ProcessAppendedItems() {
+#if DCHECK_IS_ON()
+  needs_process_ = false;
+#endif
+  for (DisplayItem* item : items_) {
+    is_suitable_for_gpu_rasterization_ &= item->IsSuitableForGpuRasterization();
+    approximate_op_count_ += item->ApproximateOpCount();
+
+    if (use_cached_picture_) {
+      DCHECK(canvas_);
+      item->Raster(canvas_.get(), NULL);
+    }
+
+    if (retain_individual_display_items_) {
+      // Warning: this double-counts SkPicture data if use_cached_picture_ is
+      // also true.
+      picture_memory_usage_ += item->PictureMemoryUsage();
+    }
+  }
+
+  if (!retain_individual_display_items_)
+    items_.clear();
+}
+
 void DisplayItemList::CreateAndCacheSkPicture() {
+  DCHECK(ProcessAppendedItemsCalled());
   // Convert to an SkPicture for faster rasterization.
   DCHECK(use_cached_picture_);
   DCHECK(!picture_);
@@ -110,24 +154,8 @@ void DisplayItemList::CreateAndCacheSkPicture() {
   canvas_.clear();
 }
 
-void DisplayItemList::AppendItem(scoped_ptr<DisplayItem> item) {
-  is_suitable_for_gpu_rasterization_ &= item->IsSuitableForGpuRasterization();
-  approximate_op_count_ += item->ApproximateOpCount();
-
-  if (use_cached_picture_) {
-    DCHECK(canvas_);
-    item->Raster(canvas_.get(), NULL);
-  }
-
-  if (retain_individual_display_items_) {
-    // Warning: this double-counts SkPicture data if use_cached_picture_ is also
-    // true.
-    picture_memory_usage_ += item->PictureMemoryUsage();
-    items_.push_back(item.Pass());
-  }
-}
-
 bool DisplayItemList::IsSuitableForGpuRasterization() const {
+  DCHECK(ProcessAppendedItemsCalled());
   // This is more permissive than Picture's implementation, since none of the
   // items might individually trigger a veto even though they collectively have
   // enough "bad" operations that a corresponding Picture would get vetoed.
@@ -135,10 +163,12 @@ bool DisplayItemList::IsSuitableForGpuRasterization() const {
 }
 
 int DisplayItemList::ApproximateOpCount() const {
+  DCHECK(ProcessAppendedItemsCalled());
   return approximate_op_count_;
 }
 
 size_t DisplayItemList::PictureMemoryUsage() const {
+  DCHECK(ProcessAppendedItemsCalled());
   // We double-count in this case. Produce zero to avoid being misleading.
   if (use_cached_picture_ && retain_individual_display_items_)
     return 0;
@@ -149,6 +179,7 @@ size_t DisplayItemList::PictureMemoryUsage() const {
 
 scoped_refptr<base::trace_event::ConvertableToTraceFormat>
 DisplayItemList::AsValue() const {
+  DCHECK(ProcessAppendedItemsCalled());
   scoped_refptr<base::trace_event::TracedValue> state =
       new base::trace_event::TracedValue();
 
@@ -178,6 +209,7 @@ DisplayItemList::AsValue() const {
 }
 
 void DisplayItemList::EmitTraceSnapshot() const {
+  DCHECK(ProcessAppendedItemsCalled());
   TRACE_EVENT_OBJECT_SNAPSHOT_WITH_ID(
       TRACE_DISABLED_BY_DEFAULT("cc.debug.picture") ","
       TRACE_DISABLED_BY_DEFAULT("devtools.timeline.picture"),
@@ -185,6 +217,7 @@ void DisplayItemList::EmitTraceSnapshot() const {
 }
 
 void DisplayItemList::GatherPixelRefs(const gfx::Size& grid_cell_size) {
+  DCHECK(ProcessAppendedItemsCalled());
   // This should be only called once, and only after CreateAndCacheSkPicture.
   DCHECK(picture_);
   DCHECK(!pixel_refs_);