Master of Science Thesis λSignatures: Function Profiling and Similarity Analysis for Reducing Compiler Design Effort

λSignatures are proposed as a method to express the behavior of code and judge its simi-larity. We present an approach to monitor execution and build effective behavioral summaries. These summaries consist of graphical models of data dependencies and feature vectors indicative of computational logic. Using this hybrid approach, we build a model which is both microarchi-tecturally sensitive and semantically aware. A method of quantizing similarity of λSignatures is also presented. To evaluate the λSignatures model, we used our system to identify common code patterns in the SPEC benchmark suite. This similarity data could allow compiler writers to quickly study larger amounts of code. Design of compiler optimization involves understanding the operation of application code. To achieve maximum speedups, large amounts of code must be targeted, however large studies as typically infeasible as they are manually conducted. Ideally, code analysts could automatically cluster code which behaves similarly and study representative samples in detail. In this paper, we show that λSignatures can be used to improve the productivity of compiler writers involved in the optimization discovery process