We just about got away with this on x86 which isn't
alignment-sensitive.  The result of the memory load is compared
against a few different values, but there is a fallback case that
happened to be the right thing when the pointer was tagged.  A good
bug to find, nonetheless.