LExecution: Learning to Guide and Analyze Program Executions

神经软件分析已成为补充和改进传统的基于逻辑的程序分析的有效途径。今天几乎所有的神经软件分析都集中在源代码和其他与软件相关的静态工件上。相比之下，在利用大规模机器学习与动态程序分析相结合的能力方面，很少有人做过工作。该建议将使用基于学习的技术来实现动态分析，对程序执行进行推理，并最终改进程序的源代码。为此，我们计划探索三个研究方向：(1)学习引导执行，它使用机器学习模型在常规执行会被卡住的情况下进行动态分析。(2)对执行进行预测，例如，根据运行时可用的信息识别错误行为和可能的错误。(3)以执行为导向的代码编辑，它不仅根据代码，而且根据其执行的痕迹来预测如何改进程序的源代码。总的来说，该项目将缩小动态程序分析和神经软件分析之间的差距，如果成功，将产生优于最先进的新型分析技术。

Neural software analysis has become an effective way of complementing and improving traditional, logic-based program analysis. Almost all of today’s neural software analyses focus on source code and other static artifacts associated with software. In contrast, little work has been done toward exploiting the power of large-scale machine learning in combination with dynamic program analysis. This proposal will use learning-based techniques to enable dynamic analysis, reason about program executions, and eventually improve the source code of the program. To this end, we plan to explore three research directions: (1) Learning-guided execution, which uses machine learning models to enable dynamic analysis in situations where a regular execution would get stuck. (2) Making predictions about executions, e.g., by identifying misbehavior and likely bugs based on information available at runtime. (3) Execution-guided code editing, which predicts how to improve the source code of a program based not only on the code, but also on traces of its execution. Overall, this project will close the gap between dynamic program analysis and neural software analysis, and if successful, yield novel analysis techniques that outperform the state-of-the-art.