More fixes for scoped_refptr operator T* removal.

mojo/system/channel_endpoint.h

Index: mojo/system/channel_endpoint.h
diff --git a/mojo/system/channel_endpoint.h b/mojo/system/channel_endpoint.h
index 24b4b72d5948ac6bba4b041428c74c80318f9b1c..8bffe48d2943a54b76d6faa40736ba1e5ad47273 100644
--- a/mojo/system/channel_endpoint.h
+++ b/mojo/system/channel_endpoint.h
@@ -103,7 +103,8 @@ class Channel;
 class MOJO_SYSTEM_IMPL_EXPORT ChannelEndpoint
     : public base::RefCountedThreadSafe<ChannelEndpoint> {
  public:
-  ChannelEndpoint(MessagePipe* message_pipe, unsigned port);
+  ChannelEndpoint(const scoped_refptr<MessagePipe>& message_pipe,

[vtl, 2014/09/05 05:41:51]

Why? The original form is fine, and avoids the extra indirection. Moreover, in
the case that the object is passed, this form (as well as passing a
scoped_refptr by value), causes ref count churn, which is expensive.

[dcheng, 2014/09/05 06:01:26]

It's documentation. There's already been some discussion about this pattern in
https://codereview.chromium.org/511323002/,
https://codereview.chromium.org/506153002/, and elsewhere. This is the general
pattern that these conversion CLs have been taking. If this is something you
feel strongly about, I think this would be a good discussion to have on
chromium-dev.
If you care about refcount churn, it appears Channel::AttachMessagePipeEndpoint
already unnecessarily causes refcount churn. And many other places in Chrome.

I've updated this to just pass in the result of .get() when invoking the ctor,
because I'd really like to get the conversion disabled on Linux builds tonight.

+                  unsigned port);
 
  private:
   enum State {