Fast evolution of custom machine representations

Described are new approaches for evaluating computer program representations for use in automated search methodologies such as the evolutionary design of software. Previously, program representations have been either evaluated directly on raw hardware, providing high speed but little control and flexibility; or, programs were interpreted by a software interpreter which can incorporate much flexibility into a program´s evaluation, but does so at a large cost in time due to interpretation overheads. In this paper we bridge this gap by providing intermediate compilation techniques for machine representations that approach the speed of running raw bits directly on hardware, but that have all the flexibility and control of custom instruction sets. In particular, we describe two compilation techniques: the first uses just-in-time compilation to convert a custom instruction sequence to machine code; the second compiles an instruction set specification into a specialized interpreter which incurs only small overheads for instruction decoding. We show that both techniques can provide manyfold speedups over direct interpretation while retaining the expressiveness of custom representations