Skip to content

Randomly Testing a Static Analyzer

Static analyzers are intended to find bugs in code, and to show that certain kinds of bugs don’t exist. However, static analyzers are themselves large, complicated programs, leading to a “who watches the watchmen” problem. Pascal Cuoq, one of the people behind the excellent Frama-C analyzer, took it upon himself to run the fuzz-fix cycle for Frama-C and Csmith all the way to its fixpoint — where no bugs can be found.

Fuzz testing relies on knowing some invariants for the system under test; often, this is just the trivial “program shouldn’t crash.” Luckily, for compiler testing we can do much better and Pascal did some hacking to turn Frama-C into a (relatively) efficient interpreter for C programs, making it possible to compare Frama-C’s interpretation of a program against regular C compilers.

The aspect of this exercise that I found most interesting was that Frama-C found some nasty bugs in Csmith at the same time that Csmith was finding problems in Frama-C. You might say that Csmith fuzzed itself, but without Frama-C’s deep inspection of the generated programs’ behavior, we couldn’t see the bugs. Recall that Csmith’s key guarantee is that its output is well-defined by the C standard. Frama-C found five bugs where we failed to provide that guarantee.

The bugs found in Frama-C are listed here and more details can be found in a short paper that we (but mainly Pascal) wrote. We hypothesize that random testing would reveal a similar set of issues in other static analysis tools for C. If these tools are being used as part of safety arguments for critical systems, somebody should do this fuzzing.

{ 4 } Comments

  1. Pascal Cuoq | February 3, 2012 at 9:54 am | Permalink

    A link to see only the bugs in Frama-C related to Csmith testing:

  2. Octoploid | February 3, 2012 at 10:13 am | Permalink

    Quick off-topic question:
    A tool named “clang_delta” is now called
    from Is clang_delta available
    for download somewhere?

  3. regehr | February 3, 2012 at 10:57 am | Permalink

    Whoops– fixed in the article, thanks Pascal!

  4. regehr | February 3, 2012 at 10:59 am | Permalink

    Hi Octoploid- Sorry about that! I’ll commit a fix in a minute that makes c_delta auto-detect whether clang_delta is available.

    clang_delta is new and experimental, but works well already. Let me see what I can do to get it out there. Its in our internal SVN now, maybe we should just move it to github.