Just-in-Time Pptimizations for High-Performance Java Programs

Our previous experience withan off-lineJava optimizer [6] has shown that some traditional algorithms used in compilers are too slow for a JIT compiler. In this paper we propose and implement faster ways of performing analyses needed for our optimizations. For instance we have replaced reaching definition with constant values and loop induction variables with loop-defined variables. As the result Briki, our JIT compiler, is extremely fast, so that its running time is negligible even when the data sets used with our benchmarks result in execution times of a few seconds. The impact for the same benchmarks running on more realistic problem sizes would be even smaller. Currently the speedups resulting from applying our optimizations are between 10% and 20%, but when the JIT compiler will perform standard optimizations which are absent in the current version of the JIT compiler used by us, the speedups should be similar to the ones observed for Fortran programs—about 50%.