gh-112536: Set up TSAN CI for free-threading

[corona10]

• Issue: Add support for thread sanitizer (TSAN) via --with-thread-sanitizer #112536

[corona10]

Hmm, I locally passed the thread-sanitizer test with clang, I am taking a look why CI has an issue.

python.exe(48341,0x1e0bcdc40) malloc: nano zone abandoned due to inability to reserve vm space.
Using random seed: 2688388199
Raised RLIMIT_NOFILE: 256 -> 1024
0:00:00 load avg: 7.94 Run 44 tests in parallel using 4 worker processes
0:00:01 load avg: 7.94 [ 1/44] test_binop passed
0:00:02 load avg: 7.94 [ 2/44] test_binascii passed
0:00:04 load avg: 7.94 [ 3/44] test_bisect passed
0:00:07 load avg: 7.94 [ 4/44] test_base64 passed
0:00:09 load avg: 7.94 [ 5/44] test_array passed
0:00:09 load avg: 7.71 [ 6/44] test_cmath passed
0:00:14 load avg: 7.71 [ 7/44] test_bz2 passed
0:00:18 load avg: 7.81 [ 8/44] test_complex passed
0:00:24 load avg: 7.82 [ 9/44] test_dataclasses passed
0:00:24 load avg: 7.60 [10/44] test_codecs passed
0:00:26 load avg: 7.60 [11/44] test_collections passed
0:00:41 load avg: 7.10 [12/44] test_difflib passed
0:00:42 load avg: 7.10 [13/44] test_bytes passed
0:00:46 load avg: 7.33 [14/44] test_float passed

[corona10]

checking for builtin __atomic_load_n and __atomic_store_n functions... no
configure: error: --disable-gil requires mimalloc memory allocator (--with-mimalloc).

Weired..

[corona10]

I need to check on my Linux machine later.

./Include/cpython/pyatomic_gcc.h:531:3: warning: ‘atomic_thread_fence’ is not supported with ‘-fsanitize=thread’ [-Wtsan]
  531 | { __atomic_thread_fence(__ATOMIC_SEQ_CST); }

[colesbury]

The version says "11" but the step name says "GCC-10"

[colesbury]

I think we'll want a configuration specifically for tsan. Maybe a --tsan option? We'll only want to run tests that actually use threading.

Brett had a list of tests that fail with PYTHON_GIL=0 -- those are probably a good starting point:

swtaarrs@1fe9165#diff-590c02490d066378045dd48d62e86dc85d3fb4ad934b6a2b2b0b1112e59eaefb

[colesbury]

I've had a better experience using suppression lists rather than marking functions as _Py_NO_SANITIZE_THREAD:

1. Add a file supressions.txt somwhere (maybe Tools/tsan/supressions.txt?)
2. Set the environment variable TSAN_OPTIONS="suppressions=<path_to_supressions.txt>"

See https://github.com/google/sanitizers/wiki/ThreadSanitizerSuppressions

[colesbury]

Do we want --with-pydebug?

[corona10]

Actual PR for this issue will be: #116872