Experimental Evaluation of Numerical Domains for Inferring Ranges

Abstract

Among the numerical abstract domains for detecting linear relationships between program variables, the polyhedra domain is, from a purely theoretical point of view, the most precise one. Other domains, such as intervals, octagons and parallelotopes, are less expressive but generally more efficient. We focus our attention on interval constraints and, using a suite of benchmarks, we experimentally show that, in practice, polyhedra may often compute results less precise than the other domains, due to the use of the widening operator.