Skip to content



Repository files navigation

Graves-CPA - Graph-based Verifier Selector

Graves is an algorithm selection tool which uses graph representations of programs to decide the order in which to run a variety of verification algorithms. Using state of the art graph neural network techniques, Graves predicts the optimal ordering of several CPAChecker configurations. This is a fork of the PeSCo repositpory, which itself is a fork of the original CPAChecker repository.


Beyond the normal dependencies of CPAChecker, Graves requires both PyTorch>=1.9.1 and PyTorch-Geometric>=2.0.1 (Other versions may work, but these are the officially supported version). Both of these libraries have CPU only installs, which is all Graves requires. These will be installed by the setup script, as long as conda or pip are installed.

Graves also requires llvm be built from source for the portion of the tool which constructs program graphs, which requires CMake and a build tool, like make or ninja. Luckily, this process can be expedited by only building what is needed for the graph builder. The setup script supports both ninja and make, but it is generally faster if ninja is installed.

Graves is built on CPAChecker, so ant and a Java SDK of version 11 or greater are needed to install it. These can be installed via package manager. For example, on ubuntu sudo apt install ant and sudo apt-get install openjdk-11-jdk


To setup Graves perform the following steps:

  1. git clone --recursive
  2. cd cpachecker
  3. ./

Running Graves

In the scripts directory run:

./ -svcomp22-graves [file.c]

where file.c is the C program you wish to verify.

Getting Started with CPAchecker

Installation Instructions: Develop and Contribute: doc/

More documentation can be found in the doc folder.

License and Copyright

CPAchecker is licensed under the Apache 2.0 License with copyright by Dirk Beyer and others (cf. for full list of all contributors). Third-party libraries are under various other licenses and copyrights, cf. lib/java-runtime-licenses.txt for an overview and the files in the directory LICENSES for the full license texts. In particular, MathSAT is available for research and evaluation purposes only (cf. LICENSES/LicenseRef-MathSAT-CPAchecker.txt), so make sure to use a different SMT solver if necessary. Note that although a GPL program is distributed together with CPAchecker, CPAchecker is separate from that program and thus not under the terms of the GPL.

Prepare Programs for Verification by CPAchecker

All programs need to pre-processed with the C pre-processor, i.e., they may not contain #define and #include directives. You can enable pre-processing inside CPAchecker by specifying -preprocess on the command line. Multiple C files can be given and will be linked together and verified as a single program (experimental feature).

CPAchecker is able to parse and analyze a large subset of (GNU)C. If parsing fails for your program, please send a report to

Verifying a Program with CPAchecker

  1. Choose a source code file that you want to be checked. If you use your own program, remember to pre-process it as mentioned above. Example: doc/examples/example.c or doc/examples/example_bug.c A good source for more example programs is the benchmark set of the International Competition on Software Verification, which can be checked out from

  2. If you want to enable certain analyses like predicate analysis, choose a configuration file. This file defines for example which CPAs are used. Standard configuration files can be found in the directory config/. If you do not want a specific analysis, we recommend config/ However, note that if you are on MacOS you need to provide specifically-compiled MathSAT binaries for this configuration to work (or use Docker in order to run the Linux version of CPAchecker). The configuration of CPAchecker is explained in doc/

  3. Choose a specification file (you may not need this for some CPAs). The standard configuration files use config/specification/default.spc as the default specification. With this one, CPAchecker will look for labels named ERROR (case insensitive) and assertions in the source code file. Other examples for specifications can be found in config/specification/

  4. Execute scripts/ [ -config <CONFIG_FILE> ] [ -spec <SPEC_FILE> ] <SOURCE_FILE> Either a configuration file or a specification file needs to be given. The current directory should be the CPAchecker project directory. Additional command line switches are described in doc/ Example: scripts/ -config config/ doc/examples/example.c This example can also be abbreviated to: scripts/ -default doc/examples/example.c Java 11 or later is necessary. If it is not in your PATH, you need to specify it in the environment variable JAVA. Example: export JAVA=/usr/lib/jvm/java-11-openjdk-amd64/bin/java for 64bit OpenJDK 11 on Ubuntu. On Windows (without Cygwin or WSL), you need to use cpa.bat instead of

    Please note that not all analysis configurations are available for MacOS because we do not ship binaries for SMT solvers for this platform. You either need to build the appropriate binaries yourself or use less powerful analyses that work with Java-based solvers, for example this one instead of -default: -predicateAnalysis-linear -setprop solver.solver=SMTInterpol Of course you can also use solutions like Docker for executing the Linux version of CPAchecker.

    If you installed CPAchecker using Docker, the above example command line would look like this: docker run -v $(pwd):/workdir -u $UID:$GID -default /cpachecker/doc/examples/example.c This command makes the current directory available in the container, so to verify a program in the current directory just provide its file name instead of the example that is bundled with CPAchecker. Output files of CPAchecker will be placed in ./output/.

  5. Additionally to the console output, an interactive HTML report is generated in the directory output/, either named Report.html (for result TRUE) or Counterexample.*.html (for result FALSE). Open these files in a browser to view the CPAchecker analysis result (cf. doc/

There are also additional output files in the directory output/:

  • Visualization of abstract reachability tree (Graphviz format)
  • cfa*.dot: Visualization of control flow automaton (Graphviz format)
  • Visualization of control flow automaton with the abstract states visualized on top (Graphviz format)
  • Coverage information (similar to those of testing tools) in Gcov format Use the following command line to generate an HTML report as output/index.html: genhtml output/ --output-directory output --legend
  • Counterexample.*.txt: A path through the program that leads to an error
  • Counterexample.*.assignment.txt: Assignments for all variables on the error path.
  • predmap.txt: Predicates used by predicate analysis to prove program safety
  • reached.txt: Dump of all reached abstract states
  • Statistics.txt: Time statistics (can also be printed to console with -stats)

Note that not all of these files will be available for all configurations. Also some of these files are only produced if an error is found (or vice-versa). CPAchecker will overwrite files in this directory!

Validating a Program with CPA-witness2test

You can validate violation witnesses with CPA-witness2test, which is part of CPAchecker.

  1. To do so, you need a violation witness, a specification file that fits the violation witness, and the source code file that fits the violation witness.

  2. To validate the witness, execute the following command:

    scripts/ -witness <WITNESS_FILE> -spec <SPEC_FILE> <SOURCE_FILE>`

    Addtional command line switches are viewed with scripts/ -h.

  3. When finished, and if the violation witness is successfully validated, the console output shows Verification result: FALSE. Additionally to the console output, CPA-witness2test also creates a file output/*.harness.c. This file can be compiled against the source file to create an executable test that reflects the violation witness.

Note that if the violation witness does not contain enough information to create an executable test, the validation result will be ERROR and the console output will contain the following line: Could not export a test harness, some test-vector values are missing.


Graves, Graph-based Verifier Selector







No releases published


No packages published


  • Java 35.4%
  • SWIG 27.4%
  • Assembly 26.0%
  • C 10.0%
  • Python 0.5%
  • JavaScript 0.3%
  • Other 0.4%