gh-126366: Fix crash if __iter__ raises an exception during yield from

[ZeroIntensity]

Apparently, gh-126366 has two issues: one is a thread safety problem inside list.__init__, and then the other is that the GET_YIELD_FROM_ITER instruction crashes if a non-native-generator object raises an exception in its __iter__. Since the latter only applies to 3.14, I'm splitting the fix into two PRs.

• Issue: Abort in free-threaded build due to mutation of ChainMap of a Counter in threads #126366

[ZeroIntensity]

I don't understand how the JIT failures could be related.

[ZeroIntensity]

@savannahostrowski @diegorusso My JIT experts, are these known failures, or is there something wrong with my fix?

[savannahostrowski]

@ZeroIntensity The aarch64 failures are known/expected. This is what we were discussing over Discord a couple of days back. Right now, we are waiting for this to be fixed in the aarch64 macos runners.

As for the emulated Linux failures, sometimes we do see failures here, and that's why we maintain this list of tests to skip. I will note that this test was seen recently passing in https://github.com/python/cpython/actions/runs/11647994504/job/32434030552?pr=126339).

If this PR has nothing to do with test.test_subprocess.POSIXProcessTestCase.test_vfork_used_when_expected (and I suspect it doesn't...), it may need to be added to the skip list.

[ZeroIntensity]

Thank you, Savannah! I'll update the skip list and see if that fixes it.

[colesbury]

LGTM

[vstinner]

Would it be possible to backport the test to 3.12 and 3.13?

[ZeroIntensity]

Maybe, but would it be all that useful? Anything that gets backported to those versions will (probably) have to go through the test on main anyway.

[vstinner]

It's useful to check for non-regression: make sure that we don't break this code path in the future.