Well you could instrument the source code and measure the number of calls, and compare that to the rough number of non-call ops per function. Not a perfect metric but it's something. His claims do seem far fetched though.