(This is a sub-task of bpo-36876, "Global C variables are a problem.".)

Currently Argument Clinic generates "_PyArg_Parser _parser" as a static variable.  Dropping "static" solves the problem of thread safety (e.g. for subinterpreters not sharing the GIL).

Dropping "static" will hit performance. _PyArg_Parser is used at first place because it can cache some data between calls. If drop static, you should drop also _PyArg_Parser and return to old slow arguments parsing.

Would adding a mutex for thread-safe initialization of _Py_Identifier and _PyArg_Parser solve the problem with subinterpreters?

That might work. :)

There are two key problems under subinterpreters that do not share the GIL:

* races on refcount operations
* cache thrashing due to refcount operations (under multi-core threads)

A lock would certainly mitigate the first problem.  I'm not sure how much the second would actually be a problem.

shared objects x threads = contention for notification of invalidated cache lines

If you're not running multiple threads, there's no problem.  If it's only a few shared objects, it probably wouldn't be a big deal.  As they say in medicine: "the dose makes the poison."