Type-Safe Optimisation of Plugin Architectures

Abstract

Programmers increasingly implement plugin architectures in type-safe object-oriented languages such as Java. A virtual machine can dynamically load class files containing plugins, and a JIT compiler can do optimisations such as method inlining. Until now, the best known approach to type-safe method inlining in the presence of dynamic class loading is based on Class Hierarchy Analysis. Flow analyses that are more powerful than Class Hierarchy Analysis lead to more inlining but are more time consuming and not known to be type safe. In this paper we present and justify a new approach to type-safe method inlining in the presence of dynamic class loading. First we present experimental results that show that there are major advantages to analysing all locally available plugins at start-up time. If we analyse the locally available plugins at start-up time, then flow analysis is only needed at start-up time and when downloading plugins from the Internet, that is, when long pauses are expected anyway. Second, inspired by the experimental results, we design a new framework for type-safe method inlining which is based on a new type system and an existing flow analysis. In the new type system, a type is a pair of Java types, one from the original program and one that reflects the flow analysis. We prove that method inlining preserves typability, and the experimental results show that the new approach inlines considerably more call sites than Class Hierarchy Analysis.