Efficiently intertwining widening and narrowing

TL;DR

This paper introduces a novel operator combining widening and narrowing for fixpoint computations, enabling more precise and efficient analysis of systems with infinite chains and monotonicity constraints.

Contribution

It proposes a new combined operator for widening and narrowing, along with adapted solving algorithms, ensuring soundness and termination in monotonic systems with finitely many unknowns.

Findings

Guaranteed soundness of the combined operator

Termination assured for monotonic systems with finite unknowns

Practical methods for non-monotonic cases

Abstract

Non-trivial analysis problems require posets with infinite ascending and descending chains. In order to compute reasonably precise post-fixpoints of the resulting systems of equations, Cousot and Cousot have suggested accelerated fixpoint iteration by means of widening and narrowing. The strict separation into phases, however, may unnecessarily give up precision that cannot be recovered later, as over-approximated interim results have to be fully propagated through the equation the system. Additionally, classical two-phased approach is not suitable for equation systems with infinitely many unknowns---where demand driven solving must be used. Construction of an intertwined approach must be able to answer when it is safe to apply narrowing---or when widening must be applied. In general, this is a difficult problem. In case the right-hand sides of equations are monotonic, however, we can always apply narrowing whenever we have reached a post-fixpoint for an equation. The assumption of monotonicity, though, is not met in presence of widening. It is also not met by equation systems corresponding to context-sensitive inter-procedural analysis, possibly combining context-sensitive analysis of local information with flow-insensitive analysis of globals. As a remedy, we present a novel operator that combines a given widening operator with a given narrowing operator. We present adapted versions of round-robin as well as of worklist iteration, local and side-effecting solving algorithms for the combined operator and prove that the resulting solvers always return sound results and are guaranteed to terminate for monotonic systems whenever only finitely many unknowns (constraint variables) are encountered. Practical remedies are proposed for termination in the non-monotonic case.