Wouldn't this kind of development target be better served by optimizing small C/++ programs instead of trying to optimize to some abstract virtual machine implemented on top of the hardware? I mean if speed really is your goal, why not do it correctly instead of hitting yourself in the face with an extra tree before starting?