of instructions.

This requires caching the generation of invokers (which itself should
also give a speed boost), since invoker generation needed to analyze
the instructions in the kernel.