The code in your example uses volatile. That prevents lots of compiler optimizations. In my experience compilers and CPU do a better optimization job than humans until the human factor interferes with the compiler. Even 40% might not be slower than calling memcpy() for every block or processing the input byte by byte instead of uint64 by uint64...

I can't comment on ARM and Barry's ARM box is dead at the moment. Distributors or users can select different and more ARM-friendly code, too. After all the hash code is easily interchangeable. :)