I just realised I'm slightly wrong about identity checks -- there is a very very small chance where if the type object occupies the same address and the attribute is in a dynamically allocated __dict__ (and not some static slot), we can trick the checks and a segfault may occur for LOAD_ATTR using the old opcache in 3.10.

I've tried running tests for 30 repetitions, but I cannot get the segfault to appear. @Pablo I'll leave it up to you if you feel it's backport worthy. I don't think people use sys._clear_type_cache or PyType_ClearCache normally so the chances of this causing any crash is astronomically small.