Statistical Analysis of Inlining Heuristics in Jikes RVM

While inlining is one of the heavily used optimizations for compiling Java programs, its overly aggressive usage can actually hurt the program performance. Therefore, the inlining heuristics used in various Java virtual machines need to be appropriately tuned in order to maximize inlining benefits without introducing its potential cost. However, the current tuning process is usually performed manually, which can not achieve the optimal solution. Several algorithms (e.g., the genetic algorithm) have been developed to automate the tuning process, but most of them are very time-consuming due to the large number of parameters used in an inlining heuristic. In this paper, we use multivariate statistical methods to find the parameters which actually impact the performance of an inlining heuristic. By focusing on these ”effective” parameters, we can greatly shorten the tuning process and further simplify the original inlining heuristic while maintaining the similar performance.