Idiom Analysis: a Novel Approach to Dataflow Redundancy Identification

Computer architects have exploited properties of repeating patterns of computation in many applications ranging from instruction set architecture specification to cluster scheduling. Most existing techniques for detecting such patterns involve either identifying loops with high trip counts, or considering only adjacent instructions. Heretofore the general problem of exhaustively detecting patterns of computation in long instruction streams has been considered intractable. This thesis will describe a heuristic technique which performs a nearly exhaustive analysis on selected regions of an application’s instruction stream, and is capable of discovering patterns that have both a large number of dynamic occurrences in the instruction stream, as well as a large number of static occurrences in the binary. The patterns found through this technique, called idioms, have the interesting property that they constitute a connected segment of dataflow. Thus, an idiom’s instructions may not be contiguous either statically or dynamically. The contributions of this thesis are as follows: (1) a description of the algorithms employed in the detection of idioms, (2) a catalog of a sampling of the most common idioms in the SPEC2000 integer benchmarks in the Alpha ISA, (3) an analysis of some of the properties of idioms, and (4) a study on the effects of applying application-specific constraints to idiom analysis.