Revisit passing tstate to function calls

[rhettinger]

In https://bugs.python.org/issue38644 many internal functions added overhead by calling _PyThreadState_GET().

The effect was pervasive (over 200 occurrences throughout the core) and hit many time critical code paths. For example in commit 1726909094, type_call() became slower for every single object instantiation. The cost varies depending on the cost of memory fencing and the compiler's implementation of atomic loads. It doesn't look like these changes had any discussion, review, tests, or demonstrable user benefits.

[vstinner]

The effect was pervasive (over 200 occurrences throughout the core) and hit many time critical code paths. For example in commit 1726909094, type_call() became slower for every single object instantiation. The cost varies depending on the cost of memory fencing and the compiler's implementation of atomic loads. It doesn't look like these changes had any discussion, review, tests, or demonstrable user benefits.

Can you explain how you ran your benchmark (what is the benchmark)? What is your OS and your versions (OS, Python, compiler, etc.)?

I tried to only pass tstate when tstate was already used in the "regular" code path.

One of the motivations of this change was to prepare Python for https://bugs.python.org/issue40522 : get the Python thread state from thread local variable (thread_local, TLS, pthread_getspecific, etc.). The problem is that this change may make the internal _PyThreadState_GET() function slower on C compilers which doesn't support C11 thread_local (use slower but portable pthread_getspecific()), and it may make the public PyThreadState_Get() function slower as well.

[vstinner]

It doesn't look like these changes had any discussion

Please read again the issue. See for example the python-dev thread: https://mail.python.org/archives/list/python-dev@python.org/thread/PIXJAJPWKDGHSQD65VOO2B7FDLU2QLHH/

See also this very old sub-intepreter bug which should be fixed: https://bugs.python.org/issue15751

[vstinner]

The impact on adding an additional argument to function on performance was discussed in python/cpython#17052