Performance analysis of VLIW compilation techniques

VLIW machines derive their performance advantage from the parallel execution of independent instructions that have been scheduled by the compiler. The paper evaluates the performance impact of a set of important VLIW compilation techniques on non-numerical integer programs. In particular, several key scheduling approaches, including software pipelining versus loop unrolling, DAG-based versus trace-based global scheduling, all-path versus profiled speculation, and restricted versus unrestricted speculative loads, are compared. The evaluation is performed on a uniform VLIW testbed where a relatively fair comparison of these scheduling approaches can be made. The result provides a meaningful insight into the relative benefits of each approach