Rollback and roll-forward of architecturally fixed and non-fixed states

After a tolerable fault has been detected, a fault-tolerant system must handle the fault. Optimally, after handling, the system is fault-free and has consistent states. From real-life systems we know that many transient faults can be recovered if the execution is repeated. A rollback is a simple solution in soft- or hardware to reach a previous (sane) state. A fast detection will result in fewer instructions to re-execute and less fault propagation. In this work, we propose two innovations regarding the rollback and roll-forward of architecturally fixed and non-fixed states within a multithreaded microprocessor, both validated by an FPGA and standard-cell implementation. The rollback of fixed states is done by a register access protocol, reducing the time for a retry in average by 19.92 % under SPECint2006_base workloads.