Efficient SMT Solving for Hardware Model Checking

Kim, Hyondeuk

Graduate Thesis Or Dissertation

Efficient SMT Solving for Hardware Model Checking Public Deposited

Analytics

Download PDF

Citations

Citeable URL: https://scholar.colorado.edu/concern/graduate_thesis_or_dissertations/cj82k756t

Abstract

The Satisfiability Modulo Theories (SMT) problem is a decision problem for the satisfiability of first-order formulas with background theories. In the last few years, decision procedures for SMT have been studied intensively, and they are applied successfully to hardware and software verification, compiler optimization, scheduling, and other design automation areas. In particular, during our study, we have found that they are also applicable to constrained random simulation. SMT solvers have been effectively applied to software verification with predicate abstraction and bounded model checking. Only to a lesser extent, they have been applied to hardware verification. In today's hardware designs, bit-level and word-level operations are often tightly intermingled. On some designs, a bit-level model checker may perform better than a word-level model checker or vice versa. In my dissertation, we study several efficient SMT solving techniques that can be applied to hardware model checking and constrained random simulation. In particular, we present a hybrid approach for integer difference logic that combines finite instantiation method with Bellman-Ford algorithm. In addition, we present an efficient term-ITE conversion method that improves SMT solving by word-level simplifications. Efficiency of these techniques have been shown in our SMT solver SatEEn that won the 1st places in Integer Difference Logic (IDL) and Linear Integer Arithmetic Logic (LIA) divisions of SMT Competition 2009. In SMT-based model checking, an efficient encoding plays an important role along with the efficient SMT solving. For hardware model checking, we propose an SMT-based model checking system that consists of modeling and constraint solving components. The modeling component selectively decides the encoding method by analyzing the model, and the constraint solving component uses either Linear Integer Arithmetic Logic (LIA) or Bit-Vector (BV) solver for the encoding. On the other hand, hardware modeling is nontrivial since the behavior of hardware is described with the detailed event semantics of Standard Verilog; hence we define a subset of Verilog with restrictions that guarantee behavioral equivalence between verification condition and simulation of synchronous hardware. The restrictions lead to a concise verification condition and allow controlled nondeterminism that can be easily eliminated for synthesis. In addition, we propose an encoding method that improves SMT solving by maximizing the use of word-level information. For constrained random simulation, we propose to use word-level simplification that reduces the bit-width of each variable in the design.

Creator