Two examples that cause this: dataclasses will assign dunder methods after the type object has been created. That triggers this and since the slots are assigned one-by-one, we stop for each assignment. Not great. Another example is unittest MockMock. Again, it assigns dunder methods after the class is defined.

We can avoid stop the world pauses in both cases by checking if the type is uniquely ref by the current thread, the type object is newly created here and not exposed to other threads so we can assign without stopping the world.

We might be able to makeBEGIN_TYPE_LOCK() a regular mutex acquisition instead of a critical section, but we'll have to be careful that nothing betweenBEGIN_TYPE_LOCK() andEND_TYPE_LOCK() leads to reentrancy.

I've convertedTYPE_LOCK from being used as a critical section to being a regular mutex. There was a number of cases of reentrancy and so things got complex. I think I nearly have that sorted out now. Themro_invoke() function needs work yet. At least, unit tests pass without deadlocking. One somewhat ugly part yet is that_PyType_LookupRefAndVersion() got split into two versions with only slight differences between the functions (taking lock if needed vs caller already holding lock or the lock not being needed). Perhaps I can find a way to refactor it to avoid code duplication.

I'm not convinced that the switch of the type lock from a critical section to a plain mutex is profitable here.

Okay, I've reverted back to using a critical section forTYPE_LOCK.

Major rework, reverting back to usingTYPE_LOCK and critical sections to protect tp_mro, tp_bases, tp_base. In a previous version of this PR, I was making updates to those only while the world was stopped. Because we are so limited on what APIs you can call while the world is stopped it is hard to make it work without major overhaul to how types work. Themro() method on the metatype is particularly troublesome but there are other similar problems.

Here is a summary of what's changed vs the main branch:

• Update tp_flags and type slots (other than tp_base, tp_bases, and tp_mro) only while the world is stopped. This allows us to read those type fields without atomics and without locking.
• Add_Py_TYPE_REVEALED_FLAG as a debug build check to help prevent code that unsafely writes to a type (missing lock or stop-the-world).
• Fix a data race with the pname/ptrs "little cache" forresolve_slotdups(). That cache is not safe in the free-threaded build so disable it.
• A few code cleanup things that are not strictly required but leftover from the previous version of this PR. I think they improve the code so I left them in.

Updated pyperformance run.