Abstract
This paper presents a tool Altair that automatically generates API function cross-references, which emphasizes reliable structural measures and does not depend on specific client code. Altair ranks related API functions for a given query according to pair-wise overlap, i.e., how they share state, and clusters tightly related ones into meaningful modules. Experiments against several popular C software packages show that Altair recommends related API functions for a given query with remarkably more precise and complete results than previous tools, that it can extract modules from moderate-sized software (e.g., Apache with 1000+ functions) at high precision and recall rates (e.g., both exceeding 70% for two modules in Apache), and that the computation can finish within a few seconds.
