3.11: Performance regression in ctypes

[mdboom]

Bug report

There is an approximately 8% performance regression in function call overhead in ctypes in py3.11.0a7 vs. py3.10.4. On my machine, 280 ns/call vs. 235 ns/call.

Gist to reproduce.

Additional context in faster-cpython/ideas#370.

Your environment

Both Python versions built from source with pgo on Fedora Linux 36, x86_64.