Predicting Loop Termination to Boost Speculative Thread-Level Parallelism in Embedded Applications