If anyone wants to make a fancy deepcopy traverse slot they can still do that in the future.

BTW, there's some discussion about long-term plans to make PyObject an opaque struct, which would make this sort of directmemcpy business no longer work. It's already a little questionable in the free-threaded build, since you're not supposed to move aPyMutex.

Yes, agreed we don't have to solve the deepcopy of a custom dtype here. Thanks.

which would make this sort of direct memcpy business no longer work

Interesting, but we do need the size, no? In any-case, there are very few places we do that, and this isn't one of them. This just usesmemcpy on the pointer in case the original array is unaligned.
(Seems even for the output, which is unnecessary, but harmless.)