I guess it wouldn't be a big deal to note next to
#define RX_BUFFER_SIZE 128 something like
// use only 16, 32, 64, 128 or 256 !!!
But I agree that % RX_BUFFER_SIZE is fool-proof.
To me, it seems to me that the compiler optimizes % to & when I use unsigned char instead of int. But writing it explicitly would force it.
Pushing and popping R0 and R1 from the stack is weird, especially when it doesn't do the same with the registers that it's actually using in the interrupt.
Why is 1Mbaud a bad idea? It works great for me. Obviously I'm not sending data to Arduino, but from Arduino to PC. I need a real time info about what it is doing, and that produces 50kB/s of data. That's easy to handle by my PC =) But I needed the Arduino's side to work to be able to decode short commands sent from PC. The little hack from my first post just did the trick. I think that everybody would benefit from a smaller faster core library.