This just does the "if the architecture does efficient unaligned
handling, start the memcmp using 'unsigned long' accesses", since
Nikolay Borisov found a load that cares.

This is basically the minimal patch, and limited to architectures that
are known to not have slow unaligned handling.  We've had the stupid
byte-at-a-time version forever, and nobody has ever even noticed before,
so let's keep the fix minimal.

A potential further improvement would be to align one of the sources in
order to at least minimize unaligned cases, but the only real case of
bigger memcmp() users seems to be the FIDEDUPERANGE ioctl().  As David
Sterba says, the dedupe ioctl is typically called on ranges spanning
many pages so the common case will all be page-aligned anyway.

All the relevant architectures select HAVE_EFFICIENT_UNALIGNED_ACCESS,
so I'm not going to worry about the combination of a very rare use-case
and a rare architecture until somebody actually hits it.  Particularly
since Nikolay also tested the more complex patch with extra alignment
handling code, and it only added overhead.

Link: https://lore.kernel.org/lkml/20210721135926.602840-1-nborisov@suse.com/Reported-by: Nikolay Borisov <nborisov@suse.com>
Cc: David Sterba <dsterba@suse.cz>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>