Call Py_Finalize at exit using libc::atexit.

[m-ou-se]

This makes sure buffers are flushed, threads are joined, etc. when exiting the process.

[m-ou-se]

Maybe this also removes the need for this hack:

pyo3/src/lib.rs

Lines 300 to 305 in 072be6c

	e.print($py);
	// So when this c api function the last line called printed the error to stderr,
	// the output is only written into a buffer which is never flushed because we
	// panic before flushing. This is where this hack comes into place
	$py.run("import sys; sys.stderr.flush()", None, None)
	.unwrap();

[davidhewitt]

Seems like a good idea to me. I wonder if this will also resolve #660

[davidhewitt]

Maybe this also removes the need for this hack:

Quite possibly. Does libc::atexit handler also get called on panic?

[m-ou-se]

Does libc::atexit handler also get called on panic?

Yes.

[m-ou-se]

Does libc::atexit handler also get called on panic?

Yes.

Well, not if the panic is caught and the program keeps running. But it runs if the panic causes the program to exit, yes.

[davidhewitt]

Great, if you could please add a CHANGELOG and then I'm happy to merge this. I suggest just your commit message minus the implementation detail:

### Changed
- Call `Py_Finalize` at exit. [#943](https://github.com/PyO3/pyo3/pull/943)

But will take any other CHANGELOG message you think is good.

[m-ou-se]

Oh right, the changelog. Keep forgetting that file. Updated.

[davidhewitt]

Great, thanks 😄

I think regarding the hack - you're right that it probably isn't needed any more. But I think order of output still will change. Currently it's flush buffers then panic handler runs, if we remove the hack it'll be panic handler then flush buffers during finalize. So let's leave it there for now, but we might clean up later.

[m-ou-se]

Yeah. It's a shame that the CPython API doesn't have a Py_FlushStreams or something.

[davidhewitt]

Thank you!

[kngwyu]

Thank you!
I think it's OK to remove the import sys; sys.stderr.flush() hack, though we can test that it works.