Generic software pipelining at the assembly level

Software used in embedded systems is subject to strict timing and space constraints. The growing software complexity creates an urgent need for fast program execution under the constraint of very limited code size. However, even modern compilers produce code whose quality often is far away from the optimum. The PROPAN system is a postpass optimization framework that enables high-quality machine-dependent postpass optimizers to be generated from a concise hardware specification. The postpass approach allows to enhance the code quality of existing compilers and offers a smooth integration into existing development tool chains. In this article we present an adaptation of the modulo scheduling software pipelining algorithm to the postpass level. The implementation is fully retargetable and has been incorporated in the PROPAN system. The differences of postpass modulo scheduling compared to the standard version of the algorithm are outlined. Experimental results conducted on the Philips TriMedia TM1000 processor demonstrate that modulo scheduling can be applied at the postpass level and allows to achieve a significant code speedup with moderate code size increase.