Automatic and Correct Register Transfer Level Annotations for Low Power Microprocessor Design

We propose instruction-driven slicing, a new technique for annotating microprocessor descriptions at the Register Transfer Level (RTL) in order to achieve lower power dissipation. Our technique automatically annotates existing RTL code to optimize the circuit for lowering power dissipated by switching activity. Our technique can be applied at the architectural level as well, achieving similar power gains. We first demonstrate our technique on architectural and RTL models of a 32-bit OpenRISC pipelined processor (OR1200), showing power gains for the SPEC2000 benchmarks. These annotations achieve reduction in power dissipation by changing the logic of the design. We give a proof that the annotations on the OR1200 processor preserve the original functionality of the machine using the ACL2 theorem prover. We then further extend our technique to an outof-order superscalar core and demonstrate power gains for the same SPEC2000 benchmarks on architectural and RTL models of PUMA, a fixed point out-of-order PowerPC microprocessor.