Precise temporal analysis of source code histories at scale

Abstract

Software systems play a central role in contemporary societies, continuously expanding and adjusting to meet emerging requirements and practices. Over the years, through extensive development efforts and numerous code updates, those systems can accumulate millions of lines of code. Moreover, they exhibit complexity, configurability, and multi-language characteristics, relying on large and complex build pipelines to target multiple platforms and hardware. This requires extensive code analyzes to maintain code, monitor quality and catch bugs.

Automated code analysis is a resource-intensive process primarily designed to examine a single specific software version. Presently, existing code analysis methods struggle to efficiently evaluate multiple software versions within a single analysis, taking up an impractical amount of time. This type of code analysis, which scrutinizes software code over several moments of its existence, is referred to as “temporal code analysis.”

Temporal code analyses open new opportunities for improving software quality and reliability. For example, such analyses would permit to fully study how code and their tests co-evolve in a code history. To overcome the different challenges that prevent such analyses to run at large scale, this thesis makes the following contributions. This thesis first demonstrates the feasibility of analyzing source code changes to identify causality relations between changes (ie. co-evolutions).

The second contribution addresses the efficiency of computing fine-grained changes and their impacts from code histories. This required to revisit how source code histories are represented and processed, leveraging the structured nature of code and its stability through time. It led to an approach, called HyperAST, that incrementally computes referential dependencies.

The third contribution is a novel code differencing technique for diffing commits. This last contribution, called HyperDiff, complements HyperAST to compare commits at scale.