gh-126019 Fix inspect.getsource for classes created in PyREPL

[devdanzin]

inspect.getsource would fetch sourcecode from _pyrepl.__main__ for classes defined in the REPL when using PyREPL because sys.modules["__main__"] points to that module.

This PR makes inspect.getsource raise OSError('source code not available') for classes defined in the REPL by detecting that .__module__ is "__main__". Tests pass with this change, but if you know of a corner case that would be broken by it please say so.

• Issue: inspect.getsource returns irrelevant text when used in repl on a class defined in place  #126019

[ZeroIntensity]

Hmm, I'm worried about this being a breaking change. __module__ now has more precedent over __file__--maybe people could have been relying on that?

[pablogsal]

Yeah I share similar concerns. Also why this only affects pyREPL and not any other main module ?

[devdanzin]

Also why this only affects pyREPL and not any other main module ?

I think because for other __main__ modules, __main__.__file__ points to the correct source code.

In basic REPL, sys.modules["__main__"] doesn't have a __file__ attribute, so it raises OSError('source code not available') when inspected.

But with PyREPL, sys.modules["__main__"].__file__ is _pyrepl.__main__, which is where the source code is being pulled from. So we could instead skip loading from __file__ if we detect PyREPL is in use, what do you think?

[devdanzin]

Wait, we could just delete __file__ from __main__, right?

diff --git a/Lib/_pyrepl/main.py b/Lib/_pyrepl/main.py
index a6f824dcc4a..17c55eb9c9f 100644
--- a/Lib/_pyrepl/main.py
+++ b/Lib/_pyrepl/main.py
@@ -35,6 +35,7 @@ def interactive_console(mainmodule=None, quiet=False, pythonstartup=False):
         import __main__
         namespace = __main__.__dict__
         namespace.pop("__pyrepl_interactive_console", None)
+        namespace.pop("__file__", None)

     # sys._baserepl() above does this internally, we do it here
     startup_path = os.getenv("PYTHONSTARTUP")

[pablogsal]

That fixes the problem no?

[ZeroIntensity]

Yeah, I think that would be much less invasive.

[devdanzin]

I've changed the approach to removing __file__ from the namespace created by PyREPL, so that inspect.getsource() doesn't fetch bogus code for classes defined in the REPL.

Some changes had to be made to tests that expected __file__ to be there and to the method of detecting PyREPL in tests.

Wording of the NEWS entry seems poor to me, would gladly apply any suggested improvements.

[Eclips4]

I think that testing whether inspect.getsource works on other objects (functions, methods, traceback, frames) is worthwhile, but I'm not sure if it should be done in this PR. So, I'd be fine with doing it in a follow-up

[python-cla-bot]

All commit authors signed the Contributor License Agreement.