Recently, with increasing system complexity and advanced technology scaling, there is a severe need for accurate fault injection (FI) techniques in the reliability evaluation of safety-critical systems against transient hardware faults, like soft errors. Since compiler-based FI techniques operate at a high intermediate representation (IR) code, their accuracy is insufficient to assess the resilience of safety-critical systems against soft errors. Although binary-level FI techniques can provide high accuracy, error propagation analysis is challenging due to missing program structures. This paper proposes an accurate GCC compiler-based FI technique called (GCFI) to assess the resilience of software against soft errors. GCFI operates at the back-end of the GCC compiler and instruments the very low-level IR code through a compiler extension. GCFI only performs instrumentation once right after the completion of optimization passes, assuring one-to-one correspondence of IR code with assembly code. The effectiveness of GCFI is evaluated by employing it to conduct many FI experiments on different benchmarks compiled for x86 and ARM architectures. We compare the results with high-level and binary-level software FI techniques to demonstrate the accuracy of GCFI. The results show that GCFI can assess the resilience of programs against soft errors with high accuracy similar to binary-level FI.

Keywords