The problem is that even if XHR is on a local domain and ensured to work, long-polling is still more intensive than, for example a persistant frame, because you have to re-establish a new connection after each message, even if you bundle messages to a one second window, it's still got the overhead of a new connection (and the latency involved therin) which is even longer if HTTP2/SPDY + SSL.
Of course, that may well be enough for some situations, for others, there's too many server-pushed messages to keep up with that model as well.
The overhead of long-polling isn't quite so bad in practice thanks to persistent HTTP connections, where an existing TCP connection is reused for future HTTP requests. Pretty much all browsers support this, and so a re-poll is usually 1 packet.
Of course, long-polling is still more overhead than a streaming connection.
Of course, that may well be enough for some situations, for others, there's too many server-pushed messages to keep up with that model as well.