Using thread-level speculation to simplify manual parallelization