This paper advances the state of the art in abstract interpretation of machine code. It tackles two of the biggest challenges in machine-code analysis: (1) holding onto invariants about values in memory, and (2)identifying affine-inequality invariants while handling overflow in arithmetic operations over bit-vector data-types.

Most current approaches either capture relations only among registers (and ignore memory entirely), or make potentially unsound assumptions when handling memory. Furthermore, existing bit-vector domains are able to represent either relational affine equalities or non-relational inequalities (e.g., intervals).

The key insight to tackling both challenges is to define a new domain combinator (denoted by V), called the view-product combinator. V constructs a reduced product of two domains in which shared view-variables are used to communicate information among the domains. V applied to a non-relational memory domain and a relational bit-vector affine-equality domain constructs the Bit-Vector Memory-Equality Domain (BVME), a domain of bit-vector affine-equalities over memory and registers. V applied to the BVME domain and a bit-vector interval domain constructs the Bit-Vector Memory-Inequality Domain, a domain of relational bit-vector affine-inequalities over memory and registers.

(Click here to access the paper: PDF.)