Using indexed data structures for program specialization

Abstract

Given a program and values of static (fixed) inputs, program specialization generates an optimized version of the program that only requires dynamic (run-time) inputs. It has been an useful tool for such areas as operating systems, multimedia applications, and scientific applications. However, the size of specialized code may grow up exponentially which makes program specialization impractical for many applications. In this paper, we present a mechanism to address this problem by using indexed data structures. Unlike traditional program specialization, which encodes the result of specialization only into run-time code, our method encodes the values of multi-valued static expressions into indexed data structures and single-valued static expressions into run-time code. Because the sizes of the indexed data structures are much smaller than that of program code, we can overcome the size problem of program specialization. With a preliminary implementation for Java, we achieved improvement in performance up to a factor of 3 with very low memory and space requirements and overheads.