CAREER: Adaptive Large-Scale Program Analysis

在过去三十年中开发的自动化程序分析已经证明了证明真实世界程序的非平凡性质的能力。反过来，这种能力也适用于安全、软件定义的网络、网络物理系统等领域的新兴软件挑战。这类应用程序的多样性要求底层程序分析在可伸缩性、适用性和准确性方面适应客户需求。然而，今天的程序分析并没有提供有用的调谐旋钮。这项研究的目标是一种通用的计算机辅助方法，以有效地使程序分析适应不同的客户。它弥合了数十年的程序分析研究和构建在它们之上的各种构件之间的鸿沟，以应对新出现的软件挑战。通过这样做，它扩大并增强了程序分析对其用户以及其质量受程序分析影响的软件用户的好处。该研究有三个关键组成部分。首先，它提出了优化问题，暴露了一大组选择，以适应程序分析的各个方面，如程序分析的成本、结果的准确性以及它对丢失信息所做的假设。其次，它通过新的搜索算法解决了这些优化问题，这些算法可以高效地导航大搜索空间，在噪音存在的情况下进行推理，与用户交互，并跨程序学习。第三，构建了一个程序分析平台，方便用户指定和编写分析，使搜索算法能够对分析进行推理，并允许使用大规模计算资源来并行化分析。这种方法是在分析移动应用程序的背景下演示的，移动应用程序是在智能手机和平板电脑等高级移动设备上运行的程序。移动应用程序代表着越来越多的非专家程序员的使用，它们可能会在各种不同和要求苛刻的条件下用于广泛的用户，这些用户可以从程序分析提供的假设分析中受益。

Automated program analyses developed over the last three decades have demonstrated the ability to prove non-trivial properties of real-world programs. This ability, in turn, has applications to emerging software challenges in security, software-defined networking, cyber-physical systems, and beyond. The diversity of such applications necessitates adapting the underlying program analyses to client needs, in aspects of scalability, applicability, and accuracy. Today's program analyses, however, do not provide useful tuning knobs. The goal of this research is a general computer-assisted approach to effectively adapt program analyses to diverse clients. It bridges the gap between decades of program analysis research on one hand and diverse artifacts built atop them to address emerging software challenges on the other. In doing so, it broadens and enhances the benefits of program analysis to its users, as well as users of software whose quality is impacted by program analysis.The research has three key ingredients. First, it poses optimization problems that expose a large set of choices to adapt various aspects of a program analysis, such as its cost, the accuracy of its result, and the assumptions it makes about missing information. Second, it solves those optimization problems by new search algorithms that efficiently navigate large search spaces, reason in the presence of noise, interact with users, and learn across programs. Third, it builds a program analysis platform that facilitates users to specify and compose analyses, enables search algorithms to reason about analyses, and allows using large-scale computing resources to parallelize analyses. The approach is demonstrated in the context of analyzing mobile apps -- programs that run on advanced mobile devices such as smartphones and tablets. Mobile apps represent an increasing use of non-expert programmers and they are likely to be used across a wide range of users in heterogeneous and demanding conditions that can benefit from what-if analyses that program analysis can offer.