I'm having a hard time believing that a benchmark that produces a 300% speedup just from eliminating CALL and RET could be realistic. The cost of accessing cached memory on the program stack is close to the cost of accessing a register.