Add PyPy to the build matrix and adapt for pypy pybind11 issue

[mattip]

This is a result of the pypy failure in the conda feedstock. The problem with pybind11 is tracked here pybind/pybind11#3408 (comment)

[bobfang1992]

Umm... let's do this, let me merge this, and I can fix the clang-format issue on my local machine, and then I can publish the new version. Happy for me to do that? Or do you prefer fixing the clang format on your side?

[mattip]

I fixed the formatting and removed the debug cruft.

[mattip]

This now builds (and tests) pytomlpp-1.0.7-pp37-pypy37_pp73-*.whl files for pypy3.7 on linux, windows, macOS.

[mattip]

Thanks!

[henryiii]

FYI, unrelated, you can use pre-commit.ci and the pip installable clang-format to automatically fix format issues on PRs (and easily run them anywhere with full version pinning). https://pypi.org/project/clang-format/ & https://scikit-hep.org/developer/style#clang-format-c-only

[henryiii]

You can remove the skipping of cp27 & cp35 since cibuildwheel 2.0 (in fact, it should be printing a warning stating these are not needed).

[bobfang1992]

yeah that's for the suggestion, but I think we intentionally skipped cp27, because the underlying C++ library won't compile and python2.7 is reaching end of life. But I cannot remember clearly why we skipped cp35 though. But I am pretty sure certain it was because an issue, though I am not sure what the issue was. Do you have a need for these two to work? If so I can take a look later probably.

[henryiii]

Cibuildwheel 2.0 dropped Python 2.7 and Python 3.5, so there's no need to skip them anymore, is what I meant. :)

[bobfang1992]

Ah make sense! Thanks for letting me knnow.

[henryiii]

Just to leave a record of what I'm doing, I stuck in the following noxfile:

import nox


PYTHON_VERISONS = ["3.6", "3.7", "3.8", "3.9", "3.10", "pypy3.7", "pypy3.8"]


@nox.session(python=PYTHON_VERISONS)
def tests(session: nox.Session) -> None:
    """
    Run the tests (requires a compiler).
    """
    session.install("-r", "tests/requirements.txt")
    session.install(".", env={"MACOSX_DEPLOYMENT_TARGET": "10.9"})
    session.run("pytest", "tests", *session.posargs)

And then nox -s tests-pypy3.7 runs the build and tests.

(MACOSX_DEPLOYMENT_TARGET is necessary for PyPy to work around the fact it tries to use 10.7, and that's been invalid since around macOS 10.14 or so due to usage of the removed stdlibc++ that occurs when this is set below 10.9)

[henryiii]

I don't understand why this is happening yet, but the Python way to do this also happens to work correctly on PyPy:

diff --git a/src/encoding_decoding.cpp b/src/encoding_decoding.cpp
index 0c78830..23d0057 100644
--- a/src/encoding_decoding.cpp
+++ b/src/encoding_decoding.cpp
@@ -74,16 +74,9 @@ toml::array py_list_to_toml_array(const py::list &list) {
       arr.push_back(time_value);
     } else {
       std::stringstream ss;
-#ifdef PYPY_VERSION
-      // see
-      // https://github.com/conda-forge/pytomlpp-feedstock/pull/1#issuecomment-972738986
-      // and
-      // https://github.com/pybind/pybind11/issues/3408#issuecomment-972752210
-      ss << "not a valid type for conversion " << std::endl;
-#else
-      ss << "not a valid type for conversion " << it << std::endl;
-#endif
-      throw py::type_error(ss.str());
+      throw py::type_error(py::str("not a valid type for conversion {}").format(it));
     }
   }
   return arr;

It seems to be tripping up at the compiler level, which makes me think there might be something disabled to workaround a PyPy bug that messes up the overload here.

[bobfang1992]

@henryiii sorry I might be missing something here, what's the actual bug? I am failing to grasp the issue you are facing 😭

[henryiii]

The bug is the one avoided by this PR by the addition of a define to skip printing the object in PyPy. std::cout << it is fine on CPython, but doesn't work on PyPy for some weird reason (haven't been able to make a MWE that reproduces this). But if you just use py::str().format(), as shown above, it's fine, and that way you still get the object in the error.

[bobfang1992]

@henryiii ah I see what you mean. Care to submit a PR? I think it would be great to get this in. No worries if you find it trouble-some, I can do it later. Thanks!