execution time

Best to work in clocks.

The only real variability is the avalon MM transfer, both for code fetch and any data cycles. For fpga internal memory this is probably 2 clock (assuming there isn't contention for the memory).

I've analised and measured the 'f' processor (with branch prediction disabled) executing from tightly coupled memory and the two match. Signaltap did show an undocumented stall if a memory write follows a memory read to the same tightly coupled memory area.