Increasing instruction-level parallelism through multi-way branching

Abstract

Sequential execution of conditional branches in nonnumerical code limits the exploitation of instruction-level parallelism (ILP). In order to cope with this limitation, exploitation of parallelism must be ertended to concurrent execution of data and branch instructions, as well as parallel execution of multiple branches in a single cycle. This paper introduces a new representation of concurrency, and describes an empirical study aimed to evaluate the performance improvement that can be expected from it. The results indicate that, as more data instructions are executed in parallel, the performance benefits arising from the proposed representation become quite notable.