gh-112535: Add test on _Py_ThreadId()

[vstinner]

Add also test.support.Py_GIL_DISABLED constant.

• Issue: Make _Py_ThreadId() work on PowerPC, IBM Z, etc. #112535

[vstinner]

@colesbury@corona10: I wrote a short test to spawn 5 threads, call _Py_ThreadId() in the main thread and in the 5 threads, and then make sure that we get 6 unique numbers. What do you think of this test? Is it relevant?

Maybe we can compare _Py_ThreadId() with _thread.get_ident() or _thread.get_native_id()? I added a test to compare with _thread.get_ident() on Linux x86-64.

[vstinner]

On FreeBSD x86-64, _Py_ThreadId() cannot be compared to Thread.ident or Thread.native_id:

vstinner@freebsd$ ./python Python 3.13.0a2+ (heads/pr/112709:bf6b57b95b, Dec 4 2023, 15:05:07) [Clang 16.0.6 (https://github.com/llvm/llvm-project.git llvmorg-16.0.6-0-g7cbf1a on freebsd14 >>> import _testinternalcapi; hex(_testinternalcapi.py_thread_id()) '0x32c2e5025140' >>> import _thread; hex(_thread.get_ident()) '0x32c2e5412000' >>> _thread.get_native_id() 100125 

[vstinner]

On FreeBSD x86-64, _Py_ThreadId() cannot be compared to Thread.ident or Thread.native_id:

Same on Linux ppc64le.

[colesbury]

I think having a test for _PyThread_Id() is good, but I I don't think we should bother comparing with _thread.get_ident().

I think we want the following:

• _PyThread_Id() should be unique and non-zero
• _PyThread_Id() should not change for a given thread. For example, it should remain the same after a short sleep.
• It should be a multiple of four (i.e., at least 4-byte aligned). This isn't necessary for biased reference counting, but it makes implementing recursive mutexes easier and should be true for all our implementations of it.

[colesbury]

You can just use C11 static_assert() now, which has better error messages.

[vstinner]

Py_BUILD_ASSERT() doesn't require an error message which makes my life easier, I don't have to write down an error message :-)

[vstinner]

I think having a test for _PyThread_Id() is good, but I I don't think we should bother comparing with _thread.get_ident().

Ok, I removed this test.

_PyThread_Id() should be unique and non-zero
_PyThread_Id() should not change for a given thread. For example, it should remain the same after a short sleep.

I added tests for that.

It should be a multiple of four (i.e., at least 4-byte aligned). This isn't necessary for biased reference counting, but it makes implementing recursive mutexes easier and should be true for all our implementations of it.

I will let you add a check for that later (in a following PR). I prefer to add tests step by step, especially if it's not strictly needed right now.

[vstinner]

@colesbury: Would you mind to review the updated PR?

[colesbury]

@vstinner - looks like Windows tests are failing (possibly due to missing import for support:

https://github.com/python/cpython/actions/runs/7090430380/job/19297369118?pr=112709

[vstinner]

Ah right, I saw it before. Thanks for the reminder. It should now be fixed.