Characterization of Repeating Dynamic Code Fragments

For this study, we analyze the dynamic instruction streams of the SPEC2000 integer benchmarks to find frequently occurring units of computation, or idioms. An idiom, in the broadest sense, is an interdependent piece of a computation dataflow. For example, a load-add-store idiom performs an increment operation through a set of three interdependent instructions. Using a heuristic technique that performs an exhaustive analysis on selected regions of an application’s instruction stream, we are able to derive a small set of idioms each consisting of between three and eight Al­ pha instructions, where the set covers a non-trivial fraction of the overall stream. On the average benchmark, a set consisting of ten idioms (50 total instructions) spans over 26% of the instruction stream. We provide a catalog of the top idiom occurring in each of the benchmarks. This catalog provides interesting insights into the type of small-scale computations that are frequent in general code. For each idiom, we identify the locations in the source code from which it originates. Many idioms occur in multiple static locations. We outline some potential applications for such idioms, including techniques for cache compression, more effective instruction dispersal in a clustered architecture, and specialized instructions for a customiz­ able instruction set. We more deeply investigate an application that can reduce some redundancy in trace caches and potentially boost fetch bandwidth by a careful and systematic encoding of frequent idioms into smaller instruction words. We demonstrate that a simple decoder suffices to reconstitute the instruction stream.