We explore the effectiveness of using traces in optimization. We build a trace collection system for the Jikes Research Virtual Machine and create traces based on the execution of the SPECjvm98 and Java Grande benchmarks. We evaluate the use of traces for inlining in Jikes, and find that the use of traces leads to a decrease in execution time of 10%, when compared to providing similar information from Jikes’s adaptive system from a previous execution. This increase in performance is achieved at the expense of a code expansion of 47%. Further, this performance is slightly better than that achieved when using a greedy algorithm. We conclude that traces may be used effectively to perform inlining, although code expansion and trace collection overhead must be addressed.
Borys J. Bradel, Tarek S. Abdelrahman