Comprehension fuzzer
Generate random code samples including list comprehensions and try running them on a Python build.
To use, create a Python 3.11 venv and pip install -r frozen.txt.
Then start the evalserver with the build of Python you want to test for compatibility with 3.11:
/home/carljm/cpython-builds/dbg/python evalserver.py
Then run pytest.
Two other useful options: --hypothesis-show-statistics and
--logfile="/some/path"; the latter will dump all valid generated samples to
the log file.
Current approach is to generate a bunch of random code samples and filter down to the ones containing list comprehensions, then aim for samples that maximize number of listcomps, lambdas, and classes. The problem is that too many samples are generated without listcomps and then rejected. We probably need a more constrained format for generating the samples initially, but without over-constraining and possibly missing counter-examples.
Note that this will exec() the generated code samples on multiple
Python binaries. As the code samples cannot contain imports or calls
to builtin functions, this is probably safe, but definitely do not
run the evalserver on a publicly accessible port.