A configurable control flow error detection method based on basic block repartition

Abstract

Radiation-induced soft errors have emerged as a critical issue that must be addressed in fields such as aerospace, where control flow errors they cause could lead to catastrophic consequences. However, existing control flow errors detection methods based on signature analysis typically use a single type of signature to record predecessor–successor relationships, which has limited expressive power and struggles to balance program residual failure rate and time overhead. In response to this, we employ a basic block repartitioning technique to adjust the control flow graph of a program to conform to jump rules we set, and design a novel signature analysis scheme that combines node type signatures with predecessor–successor signatures to detect control flow errors inter basic block. We also improve the signatures mechanism to detect control flow errors intra basic block and inter procedures, thus providing the ability to detect these two types of control flow errors and allowing for flexible configuration based on the reliability requirements of the program. Experimental results show that compared to baselines, the proposed method can ensure a low residual failure rate while maintaining a low time overhead.