Optimal code size reduction for software-pipelined loops on DSP applications

Abstract

Code size expansion of software-pipelined loops is a critical problem for DSP systems with strict code size constraint. Some ad-hoc code size reduction techniques were used to try to reduce the prologue/epilogue produced by software pipelining. We present the fundamental understanding of the relationship between code size expansion and software pipelining. Based on the retiming concept, we present a powerful Code-size REDuction (CRED) technique and its application on various kinds of processors. We also provide CRED algorithms integrated with the software pipelining process. One advantage of our algorithms is that it can explore the trade-off space between "perfect" software pipelining and constrained code size. That is, the software pipelining process can be controlled to generate a schedule concerned with code size requirement. The experiment results show the effectiveness of our algorithms in both reducing the code size for software-pipelined loops and exploring the code size/performance trade-off space.