scottrogowski / code2flow

Pretty good call graphs for dynamic languages
MIT License
4.01k stars 296 forks source link

code2flow logo

Version 2.5.1 Build passing Coverage 100% License MIT

Code2flow generates call graphs for dynamic programming language. Code2flow supports Python, JavaScript, Ruby, and PHP.

The basic algorithm is simple:

  1. Translate your source files into ASTs.
  2. Find all function definitions.
  3. Determine where those functions are called.
  4. Connect the dots.

Code2flow is useful for:

Code2flow provides a pretty good estimate of your project's structure. No algorithm can generate a perfect call graph for a dynamic language – even less so if that language is duck-typed. See the known limitations in the section below.

(Below: Code2flow running against a subset of itself code2flow code2flow/engine.py code2flow/python.py --target-function=code2flow --downstream-depth=3)

code2flow running against a subset of itself

Installation

pip3 install code2flow

If you don't have it already, you will also need to install graphviz. Installation instructions can be found here.

Additionally, depending on the language you want to parse, you may need to install additional dependencies:

Usage

To generate a DOT file, run something like:

code2flow mypythonfile.py

Or, for Javascript:

code2flow myjavascriptfile.js

You can specify multiple files or import directories:

code2flow project/directory/source_a.js project/directory/source_b.js
code2flow project/directory/*.js
code2flow project/directory --language js

To pull out a subset of the graph, try something like:

code2flow mypythonfile.py --target-function my_func --upstream-depth=1 --downstream-depth=1

There are a ton of command line options, to see them all, run:

code2flow --help

How code2flow works

Code2flow approximates the structure of projects in dynamic languages. It is not possible to generate a perfect callgraph for a dynamic language.

Detailed algorithm:

  1. Generate an AST of the source code
  2. Recursively separate groups and nodes. Groups are files, modules, or classes. More precisely, groups are namespaces where functions live. Nodes are the functions themselves.
  3. For all nodes, identify function calls in those nodes.
  4. For all nodes, identify in-scope variables. Attempt to connect those variables to specific nodes and groups. This is where there is some ambiguity in the algorithm because it is impossible to know the types of variables in dynamic languages. So, instead, heuristics must be used.
  5. For all calls in all nodes, attempt to find a match from the in-scope variables. This will be an edge.
  6. If a definitive match from in-scope variables cannot be found, attempt to find a single match from all other groups and nodes.
  7. Trim orphaned nodes and groups.
  8. Output results.

Why is it impossible to generate a perfect call graph?

Consider this toy example in Python

def func_factory(param):
    if param < .5:
        return func_a
    else:
        return func_b

func = func_factory(important_variable)
func()

We have no way of knowing whether func will point to func_a or func_b until runtime. In practice, ambiguity like this is common and is present in most non-trivial applications.

Known limitations

Code2flow is internally powered by ASTs. Most limitations stem from a token not being named what code2flow expects it to be named.

As an imported library

You can work with code2flow as an imported Python library in much the same way as you work with it from the CLI.

import code2flow
code2flow.code2flow(['path/to/filea', 'path/to/fileb'], 'path/to/outputfile')

The keyword arguments to code2flow.code2flow are roughly the same as the CLI parameters. To see all available parameters, refer to the code2flow function in engine.py.

How to contribute

  1. Open an issue: Code2flow is not perfect and there is a lot that can be improved. If you find a problem parsing your source that you can identify with a simplified example, please open an issue.
  2. Create a PR: Even better, if you have a fix for the issue you identified that passes unit tests, please open a PR.
  3. Add a language: While dense, each language implementation is between 250-400 lines of code including comments. If you want to implement another language, the existing implementations can be your guide.

Unit tests

Test coverage is 100%. To run:

    pip install -r requirements_dev.txt
    make test

License

Code2flow is licensed under the MIT license. Prior to the rewrite in April 2021, code2flow was licensed under LGPL. The last commit under that license was 24b2cb854c6a872ba6e17409fbddb6659bf64d4c. The April 2021 rewrite was substantial, so it's probably reasonable to treat code2flow as completely MIT-licensed.

Acknowledgements

Unrelated projects

The name, "code2flow", has been used for several unrelated projects. Specifically, the domain, code2flow.com, has no association with this project. I've never heard anything from them and it doesn't appear like they use anything from here.

Feedback / Issues / Contact

If you have an issue using code2flow or a feature request, please post it in the issues tab. In general, I don't provide help over email. Answering a question publicly helps way more people. For everything else, please do email! scottmrogowski@gmail.com

Feature Requests

Email me. Usually, I'm spread thin across a lot of projects, so I will, unfortunately, turn down most requests. However, I am open to paid development for compelling features.