I think this is an excellent idea, and I have had similar in mind for possible code generation for a data binding tool. The reasoning is along the lines that resources used for generating specific code to replace data-driven (and reflection-using) must be worth it: new class will be created, and speed benefit is likely just incremental (less than 2x throughput). So I was thinking of similar counter-based approach; and perhaps just JITing one class-handler at a time (top match). I suspect that a simple solution is plenty good at first. I'd be interested in how your solution works, if and when you get a chance to implement it.