In this paper, we present a technique to synthesize machine-code instructions from a semantic specification, given as a Quantifier-Free Bit-Vector (QFBV) logic formula. Our technique uses an instantiation of the CounterExample Guided Inductive Synthesis (CEGIS) framework, in combination with search-space pruning heuristics to synthesize instruction-sequences. To counter the exponential cost inherent in enumerative synthesis, our technique uses a divide-and-conquer strategy to break the input QFBV formula into independent sub-formulas, and synthesize instructions for the sub-formulas. Synthesizers created by our technique could be used to create semantics-based binary rewriting tools such as optimizers, partial evaluators, program obfuscators/de-obfuscators, etc. Our experiments for Intel's IA-32 instruction set show that, in comparison to our baseline algorithm, our search-space pruning heuristics reduce the synthesis time by a factor of 473, and our divide-andconquer strategy reduces the synthesis time by a further 3 to 5 orders of magnitude.

This paper presents an algorithm for off-line partial evaluation of machine code. The algorithm follows the classical two-phase approach of binding-time analysis (BTA) followed by specialization. However, machine-code partial evaluation presents a number of new challenges, and it was necessary to devise new techniques for use in each phase.• Our BTA algorithm makes use of an instruction-rewriting method that "decouples" multiple updates performed by a single instruction. This method counters the cascading imprecision that would otherwise occur with a more naïve approach to BTA.• Our specializer specializes an explicit representation of the semantics of an instruction, and emits residual code via machine-code synthesis. Moreover, to create code that allows the stack and heap to be at different positions at run-time than at specialization-time, the specializer represents specialization-time addresses using symbolic constants, and uses a symbolic state for specialization. Our experiments show that our algorithm can be used to specialize binaries with respect to commonly used inputs to produce faster binaries, as well as to extract an executable component from a bloated binary.