Debugging Debug Information With Neural Networks

The correctness of debug information included in optimized binaries has been the subject recent attention by research community. Indeed, it represents a practically important problem, as most software running production is produced an optimizing compiler. Current solutions rely on invariants, human-defined rules that embed desired behavior, whose violation may indicate presence bug. Although this approach proved to be effective discovering several bugs, unable identify bugs do not trigger invariants. In paper, we investigate feasibility using Deep Neural Networks (DNNs) discover incorrect information. We trained set different models borrowed from NLP community unsupervised way large dataset traces and tested their performance two novel datasets propose. Our results are positive show DNNs capable both synthetic real datasets. More interestingly, performed live analysis our them bug detectors fuzzing system. they were able report 12 unknown latest version widely used LLVM toolchain, 2 which have confirmed.