A dictionary construction technique for code compression systems with echo instructions

Dictionary compression mechanisms identify redundant sequences of instructions that occur in a program. The sequences are extracted and copied to a dictionary. Each sequence is then replaced with a codeword that acts as an index into the dictionary, thereby enabling decompression of the program at runtime. The problem of optimally organizing a dictionary consisting solely of redundant sequences in order to maximize compression has long been known to be NP-Complete [23]. This paper addresses the problem of dictionary construction when redundant code fragments are represented as Data Flow Graphs (DFGs) rather than linear sequences of instructions. Since there are generally multiple legal schedules for a given DFG G, a compiler must determine a schedule for G so that other DFGs that are subgraphs of G can reference some substring of G's final code sequence. This reduces the size of the dictionary, and in turn, the size of the compressed program. Our experiments with 10 MediaBench [18] applications yielded reductions in dictionary size ranging from 21.14% to 29.76% compared to a naïve approach.