On the extraction and analysis of prevalent dataflow patterns

The complexity-effectiveness of modern wire-dominated architectures is heavily influenced by operand movement patterns within workloads. Unfortunately, the study of these common patterns is burdensome given the NP-completeness of the problem and the size of the dataflow graphs in modern applications. In response we present CPX, a fast and memory-efficient tool for the extraction of common dataflow subgraphs from application binaries. Using this tool and a practical metric of pattern popularity, we analyze Media-Bench and Spec2000int benchmarks and present their most frequent communication patterns. Results confirm the intuition of prior research that dependence chains dominate integer code, but more importantly demonstrate that dataflow communication is restricted to a tractable set of templates. A set of only ten small patterns characterizes over 90% of Spec2000int and over 75% of MediaBench dynamic instructions. These common dataflow idioms are amenable to dynamic optimization, more efficient code representations, and reducing the broadcast nature of micro-architectural resources.