Bridging the Gap Between Binary and Source Analysis

Dynamic analyses for software engineering typically operate either at the source code level or at the binary level (possibly postpro-cessing results to source code terms for output). We propose a mixed-level approach that combines the source-level and binary-level approaches throughout the duration of the analysis. Com-pared to a one-level approach, the mixed-level approach simplifies implementation, improves robustness, and enables analyses that are impossible or impractical to perform purely at the source or binary level. We have implemented a dynamic instrumentation toolkit, named Fjalar, that embodies the mixed-level approach, and we present two distinct analyses that are built upon the toolkit. The first tool performs value profiling — outputting a rich set of run-time values for further analysis. The other tool performs value partitioning — determining abstract types for concrete values. Compared to sim-ilar tools that use a source-based approach, the mixed-level tools built upon Fjalar were both easier to implement and more scalable, handling C and C++ programs of hundreds of thousands of lines. 1