Accelerated filters should filter unpadded primitives.

cc/layers/render_surface_impl.cc

Index: cc/layers/render_surface_impl.cc
diff --git a/cc/layers/render_surface_impl.cc b/cc/layers/render_surface_impl.cc
index 600206faa081522f815ead7ae6d41d0e912123d6..493adf9f70b539cf4a3d6d11e283219220f4b8e4 100644
--- a/cc/layers/render_surface_impl.cc
+++ b/cc/layers/render_surface_impl.cc
@@ -50,6 +50,11 @@ gfx::RectF RenderSurfaceImpl::DrawableContentRect() const {
     drawable_content_rect.Union(MathUtil::MapClippedRect(
         replica_draw_transform_, gfx::RectF(content_rect_)));
   }
+  if (!owning_layer_->filters().IsEmpty()) {
+    int left, top, right, bottom;
+    owning_layer_->filters().GetOutsets(&top, &right, &bottom, &left);
+    drawable_content_rect.Inset(-left, -top, -right, -bottom);

[enne (OOO), 2016/01/19 23:21:19]

Should you do this before the replica unioning? What happens if you have a
replica and a blur? Does the replica also get blurred? Or no?

Can you also write a unit test that tests drawable content rect values directly?
I think the pixel test is good, to make sure that gl renderer is correct. 
However, I'd like something that tests this value more directly if you could.  I
think you could take what you do in EnlargedTextureWithCropOffsetFilter and
create a unit test with similar layer tree construction and put it in
layer_tree_host_common_unittest.cc. 
DrawableAndVisibleContentRectsForSurfaceHierarchy is an example to crib from.

Such a test should include a blur of a surface that contains multiple layers (so
that no layer is the entire size of the surface and the whole thing needs to be
expanded), a scale (like you have in the existing test), and maybe with/without
replica.  I think with that, I'd feel confident that nobody would break this in
the future.

[Stephen White, 2016/01/19 23:59:10]

Good question. Will investigate.
Done.

[Stephen White, 2016/01/20 19:59:01]

-webkit-box-reflect is non-standard, and doesn't seem to have a spec, so I
couldn't verify the correct behaviour (unless there's some other way to create a
replica other than -webkit-box-reflect?). I did some tests with drop-shadow and
reflection followed by blur, and it looks better in this patch than in M47 (no
clipping on the edges). Incidentally, it looks like Safari has bugs in this
case: drop-shadow and reflect is fine, but add a blur and the shadow on the
reflection disappears.

Given the general brokenness of reflections in their current state, I'd rather
get this patch in now and address any other issues as they arise.

+  }
 
   // If the rect has a NaN coordinate, we return empty rect to avoid crashes in
   // functions (for example, gfx::ToEnclosedRect) that are called on this rect.