Extracting Source Models from Java Programs: Parse, Disassemble, or Profile?

Source models of software systems are often created during re-engineering to aid in performing tasks such as reachability analysis and software architecture recovery. It is therefore vital to be able to create source models that are both detailed and accurate. However, in practice the creation of these models is difficult and error prone: extraction tools often tell only part of the story. We have been working on the automated extraction of source models for programs written in Java. Our approach considers Java programs from three points of view: parsing, disassembly, and profiling. We have found that these three techniques have advantages that are complementary. Parsing source code provides the most detailed information, but it is the most complex to implement. Disassembling Java byte code gives similar results to parsing, but is less complex technically. Profiling provides the least amount of detail, but does give important feedback on run-time behaviour, such as polymorphic function calls and reflective instantiation of objects by string input. We have applied our tools to several systems, including Sun’s javac Java compiler and the Jigsaw web server. We compare the source models extracted by each of our tools, and describe reasons for differences in the extracted models.