Conclusion

We have shown that it is possible to port a JVM implementation from a 64-bit RISC architecture to a 32-bit CISC architecture spending minimal effort and without losing much performance. The achieved performance is competitive with state-of-the-art Java Just-in-time compilers. Nevertheless, there were some pitfalls to get around, including floating-point precision mode, register allocation, and calling convention. On the other hand we had opportunities for further improvement in instruction selection. After all, from our experience it is generally simpler to generate efficient RISC code because of the complexity in addressing modes and instructions of CISC. Disregarding the architectural issues, more architecture-neutral compiler optimization techniques such as [19] and [14] can be implemented to further improve performance.