Practical Floating-Point Tests with Integer Code

Testing integer software with symbolic execution is wellestablished but floating-point remains a specialty feature. Modern symbolic floating-point tactics include concretization, lexical analysis, floating-point solvers, and intricate theories, but mostly ignore the default integer-only capabilities. If a symbolic executor is already highperformance, then software-emulation, common to integer-only machines, becomes a compelling choice for symbolic floating-point.

We propose a software floating-point emulation extension for symbolic execution of binary programs. First, supporting a soft floating-point library requires little effort, so multiple models are cheap; our executor has five distinct open source soft floating-point code bases. For integrity, test cases from symbolic execution of library code itself are hardware validated; mismatches with hardware appear in every tested library, a justin- time compiler, a machine decoder, and several floating-point solvers. In practice, the executor finds program faults involving floating-point in hundreds of Linux binaries.