How much of a difference does this make to asyncio_tcp_ssl?

How much of a difference does this make to asyncio_tcp_ssl?

It is ~5% faster on macOS with this change.

I don't understand why this is only done for Py_BUILD_CORE code. Is there a reason to exclude non-stdlib users of critical sections?

Putting the thread state in a temp stack variable for all users seems to work well: https://github.com/python/cpython/pull/146066/files

I wouldn't necessarily expect it to help non-stdlib users. There are basically three cases:

1. Python interpreter core (main executable)
2. Python extensions that don't build with Py_BUILD_CORE
3. Python stdlib extensions that use Py_BUILD_CORE (e.g., _ssl)

In the main executable, everything gets inlined and the access to _Py_tss_tstate is typically very fast. Additionally, in some cases, the compiler can remove redundant loads of _Py_tss_tstate.

In non-stdlib extensions, the calls are not inlined. The access to _Py_tss_tstate is again typically in the main executable, so it's very fast (often a single memory load).

In extensions like _ssl, the critical section calls are inlined, but now the access to _Py_tss_tstate is happening in a shared library. The access to TLS from a shared library usually requires a function call, so it's slower. That's the unique case that this PR was meant to address.