Unbounded Scalable Hardware Verification.

Abstract

Model checking is a formal verification method that has been successfully applied to real-world hardware and software designs. Model checking tools, however, encounter the so-called state-explosion problem, since the size of the state spaces of such designs is exponential in the number of their state elements. In this thesis, we address this problem by exploiting the power of two complementary approaches: (a) counterexample-guided abstraction and refinement (CEGAR) of the design's datapath; and (b) the recently-introduced incremental induction algorithms for approximate reachability. These approaches are well-suited for the verification of control-centric properties in hardware designs consisting of wide datapaths and complex control logic. They also handle most complex design errors in typical hardware designs. Datapath abstraction prunes irrelevant bit-level details of datapath elements, thus greatly reducing the size of the state space that must be analyzed and allowing the verification to be focused on the control logic, where most errors originate. The induction-based approximate reachability algorithms offer the potential of significantly reducing the number of iterations needed to prove/disprove given properties by avoiding the implicit or explicit enumeration of reachable states. Our implementation of this verification framework, which we call the Averroes system, extends the approximate reachability algorithms at the bit level to first-order logic with equality and uninterpreted functions. To facilitate this extension, we formally define the solution space and state space of the abstract transition system produced by datapath abstraction. In addition, we develop an efficient way to represent sets of abstract solutions involving present- and next-states and a systematic way to project such solutions onto the space of just the present-state variables. To further increase the scalability of the Averroes verification system, we introduce the notion of structural abstraction, which extends datapath abstraction with two optimizations for better classification of state variables as either datapath or control, and with efficient memory abstraction techniques. We demonstrate the scalability of this approach by showing that Averroes significantly outperforms bit-level verification on a number of industrial benchmarks.