An empirical study of method inlining for a Java just-in-time compiler

Abstract

Method inlining is one of the optimizations that have a significant impact on both system performance and total compilation overhead in a dynamic compilation system. This paper describes an empirical study of online-profile-directed method inlining for obtaining both performance benefits and compilation time reductions in our dynamic compilation system. We rely solely on the profile information in deciding which methods should be inlined along which call paths, without employing the existing static heuristics, except that methods with extremely small bodies are always inlined. The call site distribution and invocation frequency are collected using dynamically generated instrumentation code, and are given to the recompilation controller for analyzing and organizing the inlining requests. The experimental results show that the strategy of relying on the profile information in method inlining decision is quite effective for reducing compilation overhead and/or for improving the performance. It can also provide flexibility for inlining decisions based on the type of profile, such as spiky or flat profiles. The results show that the strategy can be the basis of automatic optimization selection for a future efficient dynamic compilation framework.