pyobject_fastcall-2.patch: More complete changes. Sorry, the patch also contains unrelated refactoring! It's a more advanced implementation which tries to reduce the depth of the C backtrace. For example, _PyObject_FastCall() is now inlined manually in _PyObject_FastCallDict(). PyObject_Call() is also rewritten. If the overall approach is validated, I will rewriten the patch differently to limit changes, or push some changes in multiple commits.

Results of testcapi_stacksize.patch + stack_overflow_28870-sp.py (from issue #28870).

Reference:

haypo@smithers$ ../default-ref/python stack_overflow_28870-sp.py 
test_python_call: 8586 calls before crash, stack: 976 bytes/call
test_python_getitem: 9188 calls before crash, stack: 912 bytes/call
test_python_iterator: 7936 calls before crash, stack: 1056 bytes/call

=> total: 25710 calls, 2944 bytes

Patch:

haypo@smithers$ ./python stack_overflow_28870-sp.py 
test_python_call: 9883 calls before crash, stack: 848 bytes/call (-128 B)
test_python_getitem: 10476 calls before crash, stack: 800 bytes/call (- 112 B)
test_python_iterator: 8878 calls before crash, stack: 944 bytes/call (- 112 B)

=> total: 29237 calls (+3616), 2592 bytes (- 352 B)