Manage registering composited elements in PaintArtifactCompositor.

third_party/WebKit/Source/platform/graphics/compositing/PaintArtifactCompositor.cpp

Index: third_party/WebKit/Source/platform/graphics/compositing/PaintArtifactCompositor.cpp
diff --git a/third_party/WebKit/Source/platform/graphics/compositing/PaintArtifactCompositor.cpp b/third_party/WebKit/Source/platform/graphics/compositing/PaintArtifactCompositor.cpp
index 612385a0f360279353a80562b430d51db769a035..55f8ae884d159f53ff5a80b247b6943a4dcf3061 100644
--- a/third_party/WebKit/Source/platform/graphics/compositing/PaintArtifactCompositor.cpp
+++ b/third_party/WebKit/Source/platform/graphics/compositing/PaintArtifactCompositor.cpp
@@ -9,6 +9,7 @@
 #include "cc/layers/picture_layer.h"
 #include "cc/paint/display_item_list.h"
 #include "cc/trees/layer_tree_host.h"
+#include "cc/trees/mutator_host_client.h"
 #include "platform/RuntimeEnabledFeatures.h"
 #include "platform/graphics/GraphicsContext.h"
 #include "platform/graphics/compositing/PaintChunksToCcLayer.h"
@@ -656,6 +657,19 @@ void PaintArtifactCompositor::Update(
     }
 
     root_layer_->AddChild(layer);
+    auto previous_host = layer->layer_tree_host();

[pdr., 2017/05/16 02:46:35]

What does it mean to change layer tree hosts? In my mental model the host is
sort of the controller... would we change it in SPV2 (or even SPV1?)

[wkorman, 2017/05/16 21:42:42]

It appears that the LayerTreeHost of a layer is updated as layers move between
trees at least in SPv1 according to comment here:

https://cs.chromium.org/chromium/src/cc/layers/layer.h?gsn=Create&l=600

SetLayerTreeHost is critical path for maintaining a LayerTreeHost's knowledge of
the layers it's controlling. It's called by Layer::{Add,Insert}Child when they
call SetParent on the Layer.

In SPv2 we create a single root layer and reuse, and that layer appears to have
a stable LayerTreeHost for the duration of a single page.

We remove all layers from the root layer (see RemoveAllChildren() call above),
and re-add, each time PaintArtifactCompositor::Update() is called.
RemoveAllChildren ends up calling SetParent(nullptr) on the layer, which will
un-register it from its LayerTreeHost.

So, we're un-registering and then re-registering all child layers every time
with the same LayerTreeHost which does seem a little silly. I need to understand
LayerTreeHost and LayerTree lifecycle better to consider optimization
opportunities. Maybe let's look at in person whenever works for you?

+    if (previous_host) {
+      previous_host->UnregisterElement(element_id, cc::ElementListType::ACTIVE,
+                                       layer.get());
+    }
+    auto host = root_layer_->layer_tree_host();
+    if (element_id) {
+      host->RegisterElement(element_id, cc::ElementListType::ACTIVE,
+                            layer.get());
+      if (host->mutator_host()->HasAnyAnimation(element_id))

[pdr., 2017/05/16 02:46:35]

Do we set animations on the animation host yet?

[wkorman, 2017/05/16 21:42:42]

If I understand your question correctly, given that most animation test cases
are operational, this must be the case. Active animations end up being added as
"ticking players":

https://cs.chromium.org/chromium/src/cc/animation/animation_host.h?q=animatio...

And are ticked as part of LayerTreeHost{Impl}::Animate().

Adding of compositor animations happens through this code path from Blink into
cc:

https://cs.chromium.org/chromium/src/third_party/WebKit/Source/platform/anima...

+        host->SetNeedsCommit();
+    }
+
     layer->set_property_tree_sequence_number(g_s_property_tree_sequence_number);
     layer->SetTransformTreeIndex(transform_id);
     layer->SetClipTreeIndex(clip_id);