Diagnosing Abstraction Failure for Separation Logic–Based Analyses
Abstraction refinement is an effective verification technique for automatically proving safety properties of software. Application of this technique in shape analyses has proved impractical as core components of existing refinement techniques such as backward analysis, general conjunction, and identification of unreachable but doomed states are computationally infeasible in such domains.
We propose a new method to diagnose proof failures to be used in a refinement scheme for Separation Logic–based shape analyses. To check feasibility of abstract error traces, we perform Bounded Model Checking over the traces using a novel encoding into SMT. A subsequent diagnosis finds discontinuities on infeasible traces, and identifies doomed states admitted by the abstraction. To construct doomed states, we give a model-finding algorithm for “symbolic heap” Separation Logic formulas, employing the execution machinery of the feasibility checker to search for concrete counter-examples. The diagnosis has been implemented in SLAyer, and we present a simple scheme for refining the abstraction of hierarchical data structures, and illustrate its effectiveness on benchmarks from the SLAyer test suite.
Unable to display preview. Download preview PDF.