A System for Runtime Loop Optimisation in the Jikes RVM

The creation of a new loop analysis phase for the Jikes RVM and three new optimizing compilation phases. Between 0.25 % (overall) and 3 % (best-case) speed-up of the SPECJVM Client 98 benchmark. An improved array-bound and null check elimination phase that eliminates a greater number of runtime checks than existing analyses. Two new loop optimisation phases that achieve loop unrolling and eliminate redundant branches for FOR loops. How does the work advance the state-of-the-art?: The Jikes RVM has a redundant check and loop unrolling optimisation system. The redundant check system looks for known to be true checks and eliminates them. The loop unrolling optimisation replicates loop bodies and their contents. Our new optimisations use loop analysis and explicit checks to remove a greater number of array-bound and null checks through explicit testing, and to unroll loops without replicated test and branches. Our results show the efficacy of runtime loop analysis and runtime checks to improve loop optimisations, and to remove array-bound and null checks. Motivation (problems addressed): Dynamic optimising compilers must trade the time spent optimising code with time taken away from running it. With Java the implicit throwing of exceptions can cause poor runtime performance. Standard loop optimisations can also greatly improve performance. A loop analysis phase is desirable in an optimising compiler to generate a number of optimisations, however, the efficacy of such an approach at runtime isn't clear given the potentially large computational cost. This work demonstrates the efficacy of the loop analysis approach, which will be built upon in the future to deliver a range of loop optimisations