color: Fix color not being reset between layout tests

cc/trees/layer_tree_host_impl.cc

Index: cc/trees/layer_tree_host_impl.cc
diff --git a/cc/trees/layer_tree_host_impl.cc b/cc/trees/layer_tree_host_impl.cc
index 54dc7bbd336bc73f606cdad27254392eed83b148..c2368225e123ef67c53bcbe136c718332e44ccf6 100644
--- a/cc/trees/layer_tree_host_impl.cc
+++ b/cc/trees/layer_tree_host_impl.cc
@@ -1352,11 +1352,14 @@ void LayerTreeHostImpl::SetIsLikelyToRequireADraw(
 }
 
 gfx::ColorSpace LayerTreeHostImpl::GetRasterColorSpace() const {
+  gfx::ColorSpace result;
   if (!settings_.enable_color_correct_rasterization)
-    return gfx::ColorSpace();
-  if (!sync_tree())
-    return gfx::ColorSpace::CreateSRGB();
-  return sync_tree()->raster_color_space();
+    return result;
+  if (sync_tree())
+    result = sync_tree()->raster_color_space();

[enne (OOO), 2017/05/20 03:32:21]

I realize you're not changing this here, but this is a little weird to me that
this only asks the pending tree and not the active tree? For threaded
compositors, sync_tree only exists during commit.  Would that be why you're
getting invalid color spaces in these cases?

[ccameron, 2017/05/22 23:27:08]

Oh, LTHI::sync_tree() returns
  return pending_tree_ ? pending_tree_.get() : active_tree_.get();
That's not defined outside of commit?

In this particular case, we were getting an invalid color space because
sync_tree() existed, but had had its color space explicitly set to the
default-ctor color space (which is invalid).

[enne (OOO), 2017/05/23 00:13:43]

Ok, you're correct that as written, sync tree will return you a tree outside of
commit.  However, it really just means "the tree you're committing to" and
probably it should DCHECK that you're in the middle of commit when you call it.

The reason this is problematic is you can get a weird state here where both of
the trees exist and the pending tree has a different color space, and suddenly
the active tree starts using that color space to do raster in.

This is outside of this patch at this point but I think there should be at the
very least a comment here explaining what the process is when the color space is
different on both trees and why this code is ok.  It'd also be better to not
call sync tree and just include the logic to say "use the pending tree if it
exists, otherwise use the active tree, and this is ok because X". 
 
Ok, in that case this patch lgtm.

[ccameron, 2017/05/23 06:28:39]

Yeah, this doesn't have guarantees about when it will be called (it doesn't need
to be during commit).
Added a comment (prefer pending tree because that color space is more recently
sent from the browser).

+  if (!result.IsValid())
+    result = gfx::ColorSpace::CreateSRGB();
+  return result;
 }
 
 void LayerTreeHostImpl::RequestImplSideInvalidation() {