Static Analysis of Mega-Programs

Abstract

As imperative programs get larger and larger, many inter-procedural analyses produce progressively poorer results, even as the (usually) polynomial cost hurts more and more. The problem is not just program size, but that larger programs contain analysis defeating idioms (e.g., manual persistence methods and dispatch vectors) and deeper levels of procedural abstraction. In this talk, I will discuss what my group has learned about analyzing mega-programs (applications exceeding one million lines of non-blank, non-comment source lines). I will report on our approaches that have and haven’t worked, our attempts to get meaningful information from analyses with near-linear complexities, engineering issues that must be resolved when dealing with mega-programs, and our best guesses regarding the eventual feasibility of routinely performing static analysis on mega-programs.