M. Gupta
IBM T.J. Watson Research Centre, USA

A framework for deep analysis of Java programs


We describe a new framework for extensive interprocedural analysis of Java
programs. Our framework is able to handle dynamic features of Java like
dynamic class loading and runtime binding of methods, and yet, it avoids the
high cost associated with runtime compilation. The first part of our talk
presents quasi-static compilation, which reduces the cost of dynamic
compilation, while maintaining compliance with the dynamic features of
Java. The quasi-static compiler generates persistent code images ahead of
time, and during a production run, performs validation checks and adapts
those images to the new execution context while preserving global
optimizations and maintaining binary compatibility. A preliminary
implementation of this approach in the Jalapeno VM has led to performance
improvements ranging from 9% to 91% for the SPECjvm98 benchmarks with size
100, and 54% to 356% for the (shorter running) SPECjvm98 benchmarks with size
10. The second part of the talk describes an extension of the extant analysis
framework (presented recently by Sreedhar et al.) in the context of the
quasi-static compiler, which enables global analyses and optimizations in the
presence of dynamic class loading. Our approach enables transformations like
unguarded devirtualization and/or inlining of virtual methods in cases that
are beyond the scope of previously known techniques.