Detecting floating-point errors via atomic conditions

• DOI: 10.1145/3371128
• Publication Date: 2019-12-20T19:45:25Z
• Authors: Daming Zou; Muhan Zeng; Yingfei Xiong; Zhoulai Fu; Lu Zhang; Zhendong Su
• Keywords: Code (set theory); Implementation

This paper tackles the important, difficult problem of detecting program inputs that trigger large floating-point errors in numerical code. It introduces a novel, principled dynamic analysis leverages mathematically rigorously analyzed condition numbers for atomic operations, which we call conditions , to effectively guide search errors. Compared with existing approaches, our work based on has several distinctive benefits: (1) it does not rely high-precision implementations act as approximate oracles, are obtain general and computationally costly; (2) provide accurate, modular guidance. These benefits combination lead highly effective approach detects more significant real-world code (e.g., widely-used library functions) achieves orders speedups over state-of-the-art, thus making error significantly practical. We expect methodology principles behind benefit other tasks such debugging, repair synthesis. To facilitate reproduction work, have made implementation, evaluation data results publicly available GitHub at <a>https://github.com/FP-Analysis/atomic-condition</a>.