A New Critical Variable Analysis in Processor-Based Systems

Determining the dependability of integrated systems with respect to soft errors is necessary for a growing number of applications. The most critical information must be identified to achieve good efficiency/cost trade-offs when selective hardening is necessary. In processor-based systems, the most critical variables and registers must thus be identified for the target application program. An improved algorithm for critical register identification is described and compared to previous work. Fault injection results in a system based on Leon2 are also reported. They demonstrate the impact of micro-architectural characteristics on evaluating the real criticality of the register file. New refinements are suggested for further work based on data dependency analysis.