It looks pretty good; it's been a while since I've looked at generated assembly and haven't groaned 
It looks as though there is some code boating when multiple serial ports are used on a MEGA; presumably the templates generate duplicate code with constants rather than more generic (and slower?) functions. That's probably OK given that devices with extra serial ports have extra codespace as well, and it's not TOO awful. A trivial example with all four ports is 4148 bytes in a (slightly fixed) HardwareSerial, and 4598 bytes using the NewSerial templates.