Instruction Scheduling for Clustered VLIW DSPs

Recent digital signal processors (DSPs) show a homogeneous VLIW-like data path architecture, which allows C compilers to generate efficient code. However, still some special restrictions have to be obeyed in code generation for VLIW DSPs. In order to reduce the number of register file ports needed to provide data for multiple functional units working in parallel, the DSP data path may be clustered into several sub-paths, with very limited capabilities of exchanging values between the different clusters. An example is the well-known Texas Instruments C6201 DSP. For such architecture, the tasks of scheduling and partitioning instructions between the clusters are highly interdependent. This paper presents a new instruction scheduling approach, which in contrast to earlier work, integrates partitioning and scheduling into a single technique, to achieve a high code quality. We show experimentally that the proposed technique is capable of generating more efficient code than a commercial code generator for the TI C6201.