Exploring Efficient Architecture Design For Thread-Level Speculation---Power And Performance Perspectives.