Automated Repair of Compilation Errors in Embedded Continuous Integration
Time: Fri 2026-11-27 13.00
Location: Kollegiesalen, Brinellvägen 8, Stockholm
Language: English
Subject area: Computer Science
Doctoral student: Han Fu , Teoretisk datalogi
Opponent: Senior lecturer Neil Walkinshaw, University of Sheffield, Sheffield, England, UK
Supervisor: Professor Cyrille Artho, Teoretisk datalogi
QC 20261106
Abstract
My thesis contributes empirical studies, diagnostic and repair techniques, as well as infrastructure for reproducible analysis to support the understanding and automated repair of compilation failures in industrial and open-source embedded continuous integration (CI).
CI provides rapid feedback on software changes, but embedded systems introduce challenges from hardware–software dependencies, cross-compilation, board-specific configurations, heterogeneous toolchains, and complex build environments. Compilation failures are particularly disruptive because they prevent subsequent testing and validation, yet their underlying causes and potential for automated repair remain insufficiently understood.
The work first investigates CI and compilation failures in an industrial embeddedsetting, characterizing their types, causes, and repair difficulty. Building on these findings, Shadow Job is developed to diagnose and repair industrial compilation failures using large language models (LLMs). The investigation is then extended toopen-source embedded systems through PhantomRun, which reproduces and repairs compilation failures across heterogeneous projects and build environments. Finally, infrastructure for reconstructing historical CI failures is developed to support reproducible analysis and evaluation.
The results establish compilation as a prominent source of failures in embedded CI and identify dependency-related errors as a major contributor. LLM-based repair achieves CI pass rates of up to 63 % in the industrial setting and repairs up to 45 % of the studied open-source failures. Moreover, 91.8 % of 4628 historical failing CI runsare successfully reconstructed, with 98 % of evaluated builds preserving their originaloutcomes.
Together, these contributions connect empirical understanding, automated repair, and reproducible CI reconstruction to reduce manual debugging effort and improve embedded software development.