Identifying and evaluating a generic set of superinstructions for embedded Java programs

In this paper we present an approach to the optimisation of interpreted Java programs using superinstructions. Unlike existing techniques, we examine the feasibility of identifying a generic set of superinstructions across a suite of programs, and implementing them statically on a JVM. We formally present the sequence analysis algorithm and we describe the resulting sets of superinstructions for programs from the embedded CaffeineMark benchmark suite. We have implemented the approach on the Jam VM, a lightweight JVM, and we present results showing the level of speedup possible from this approach