Better Than Optimal: Fast Identification of Custom Instruction Candidates

Asymptotically optimal algorithms do not always yield the fastest practical algorithm on realistic cases. We examine Gutin etal.'s recently published optimal algorithm for enumerating the set of convex subgraphs under input/output constraints with application to custom instruction identification. We show that (i) suppressing some of the machinery in their algorithm results in a sub-optimal algorithm which is significantly faster in practice on real-world examples and that (ii) the constants of proportionality in the running time for both optimal and sub-optimal versions can be significantly improved by using additional output set filtering constraints.