The use of class libraries increases programmer productivity by allowing programmers to focus on the functionality unique to their application. However, library classes are generally designed with some typical usage pattern in mind, and performance may be suboptimal if the actual usage differs.
We present an approach for rewriting applications to use customized versions of library classes that are generated using a combination of static analysis and profile information. Type constraints are used to determine where customized classes may be used, and profile information is used to determine where customization is likely to be profitable.
We applied this approach to a number of Java applications by customizing various standard container classes and the omnipresent StringBuffer class, and measured speedups up to 78% and memory footprint reductions up to 46%. The increase in application size due to the added custom classes is limited to 12% for all but the smallest programs.