Zhao Yingquan
Associate Research Fellow at Tianjin University
Zhao Yingquan is an Associate Research Fellow and Master’s Supervisor at the School of Cyber Science and Technology, Tianjin University. He received his Ph.D. from Tianjin University in 2025 under the supervision of Professors Chen Junjie and Wang Zan. His main research interests include compiler testing and virtual machine testing. He has proposed a series of program feature-driven testing methods for compiler systems and has discovered numerous previously unknown defects in open-source communities and domestic foundational software. Zhao has published 15 papers in top-tier CCF-A conferences and journals such as ASE, ICSE, ISSTA, TSE, TOSEM, and CCS. He has participated in the National Key R&D Program for Young Scientists and several Huawei-sponsored research projects. He has received the First Prize in the CCF Prototype Competition and was selected for the ChinaSoft 2022 Excellent PhD Student Forum.
Topic
Fuzz Testing Techniques for the MLIR Compiler Infrastructure
MLIR is a multi-level intermediate representation (IR) infrastructure introduced by the LLVM community. It has become a foundational component in modern compiler construction, especially in deep learning compilation frameworks and various heterogeneous computing systems. To enhance the correctness and stability of the MLIR compiler infrastructure, this report focuses on its quality assurance challenges and presents a set of specialized fuzzing techniques tailored to MLIR’s program characteristics. These techniques include automatically generating valid MLIR programs, efficiently exploring the vast combination space of transformation and optimization passes, and detecting silent bugs that existing tools often fail to uncover. To date, the MLIR testing framework built upon these techniques has identified more than 160 previously unknown defects in the open-source community, significantly improving the overall security and reliability of the MLIR compilation ecosystem—including its frontends, dialects, optimization passes, and the LLVM backend.
