SHF: Medium: Collaborative Research: Semi and Fully Automated Program Repair and Synthesis via Semantic Code Search

我们经济的许多方面在很大程度上依赖于软件的正常运行。然而，软件错误很常见，经常导致安全漏洞，每年给我们的经济造成数十亿美元的损失。尽管众所周知，软件错误的成本很高，但软件行业仍在努力克服这一挑战，因为报告新错误的速度快于修复它们的速度。最近的研究已经证明了自动程序修复技术解决这一挑战的潜力。在这项研究中，我们开发了新的技术来修复软件错误并自动实现新功能。挑战是在不破坏其他功能的情况下修复代码，并致力于修复日益复杂的代码。这种方法利用了开放源代码的高可用性，这些代码已经实现了新软件项目所需的许多功能。这种方法是在开源项目中搜索相关代码，使用自动软件修复和生成技术使代码适应新的上下文，然后验证更改后的软件。该方法的一个关键组件是语义代码搜索，它查询大型代码数据库以找到满足行为规范的代码片段。该项目开发了新的技术，这些技术(1)将大的、可搜索的代码体编码为行为配置文件，(2)定位错误和特征以编码块、模块和组件，(3)提取这些块、模块和组件所需的行为配置文件，(4)使用提取的配置文件来搜索数据库以寻找潜在的补丁，(5)调整潜在的补丁以适应代码上下文，以及(6)验证潜在的补丁。该项目专注于生成高质量的代码，验证注入的代码不会破坏现有功能。更广泛的影响主要来自通过重用和适配现有代码从根本上提高软件生产率的目标。

Many aspects of our economy rely heavily on software working correctly. However, software errors are common, routinely cause security breaches, and cost our economy billions of dollars annually. Despite the well-known high costs of software errors, the software industry struggles to overcome this challenge, as new errors are reported faster than they can be fixed. Recent research has demonstrated the potential of automated program repair techniques to address this challenge. In this research, we develop new techniques to fix software errors and implement new features automatically. The challenge is to fix code while not breaking other functionality, and to work toward repairing code of increasing complexity.The approach takes advantage of the high availability of open-source code that already implements many functions required for a new software project. The approach is to search for relevant code in open-source projects, adapt that code to its new context using automated software repair and generation techniques, and then validate the changed software. A key component of the approach is semantic code search, which queries large databases of code to find code snippets that satisfy a behavioral specification. The project develops novel techniques that (1) encode large, searchable bodies of code as behavioral profiles, (2) localize bugs and features to code blocks, modules, and components, (3) extract the desired behavioral profiles of those blocks, modules, and components, (4) use the extracted profiles to search the database for potential patches, (5) adapt the potential patches to fit into the code context, and (6) validate the potential patches. The project focuses on producing high-quality code, verifying that the injected code does not break existing functionality. The broader impacts come mainly from goal of radically improving software productivity through reuse and adaptation of existing code.