Register swapping schemes for low power execution

For embedded systems, the power dissipation on buses has become an essential issue in recent years. Many real-time embedded processors, such as DSP processors, adopt the Harvard architecture in which the data and instruction buses are separated to avoid processing-speed degradation. The power dissipation on an instruction bus can be reduced if the switching activities between consecutive instructions on that bus are reduced. Two efficient algorithms, the greedy method and the dynamic programming based method, are proposed to swap commutative source register fields of adjacent instructions. The switching activities on the instruction bus are therefore reduced, without affecting the execution results. Experimental results show that the proposed schemes result in a reduction of as much as 21.43% in the switching activities of consecutive source register fields between commutative blocks. In addition, the proposed schemes can be conveniently integrated with other encoding schemes to further improve the power dissipation on an instruction bus.