So you're saying that doesn't work?!

That fixes the alignment crash on sparc64, and my sparc64 machine runs
fine.  It is not stressed normally.

  What assembly got generated as the result of the above?

It's really hard to follow (a maze of twisty inlines), but it looks ok.
My best guess is that under low-memory conditions there's a packet in an
mbuf at the edge of memory or a page and there is not enough data pulled
up.  But I can't figure out what sizes are supposed to be there.
Perhaps we should add a KASSERT that the mbuf has enough before the
IP6_NEQ.

My i386 box has 2 GB of RAM, but it is often very busy (builds, 20G of
backups to tape).