Tailoring Software Pipelining for Effective Exploitation of Zero Overhead Loop Buffer

A Zero Overhead Loop Buffer (ZOLB) is an architectural feature that is commonly found in DSPs (Digital Signal Processors). This buffer can be viewed as a compiler (or program) managed cache that can hold a limited number of instructions, which will be executed a specified number of times without incurring any loop overhead. Preliminary versions of the research, which exploit a ZOLB, report significant improvement in execution time with a minimal code size increase [UH99,UH00]. This paper extends the previous compiler efforts to further exploit a ZOLB by employing a new software pipelining methodology. The proposed techniques choose complex instructions, which capitalize on instruction level parallelism across loop iteration boundaries. Unlike the traditional pipelining techniques, the proposed pipelining strategy is tightly coupled with instruction selection so that it can perform register renaming and/or proactively generate additional instruction(s) on the fly to discover more loop parallelism on the ZOLB. This framework reports additional significant improvements in execution time with modest code size increases for various signal processing applications on the DSP16000.