Dedicated Search Strategies For Finding Critical Counterexamples In Programs With Floating Point Computations

Constraint Programming (CP) is an efficient technique for searching counter-examples that violate a property of the program to verify. However, the search process can become very long and costly when the program to check contains Floating Point (FP) computations. In this paper we discuss the capabilities of a set of variable choice strategies and subdomain selection strategies that take advantages of the very specific properties of the floats. We also outline the capabilities of a CP techniques for computing ranges of error values as well as input values that exercise a given error; a critical issue to take advantage of cancellation phenomena when searching counterexamples.