01.04.2019, 13:11, "Michael Ellerman" [off-list ref]:

This looks OK. It's a bit of a pity to replace the 8-byte-at-a-time copy
with a byte-at-a-time copy, but I suspect it's insignificant compared to
the overhead of calling the comparison and swap functions.

And we could always add a generic 8-byte-at-a-time swap function if it's
a bottleneck.

I am sorry, I forgot to quickly comment on your letter.
Now (after George Spelvin's patches) the generic swap is able
to use u64 or u32 if the alignment and size are divisible
by 4 or 8, so we lose nothing here.