bogusly casts its contents around causing alignment faults on sparc64 and
most likely also on at least powerpc. Fix this by copying the contents
bytewise instead as partly already done here. Solving this the right way
costs some space, i.e. 148 bytes with GCC and 16 bytes with clang on x86
there are still some bytes left there though, and an acceptable hack which
tricks the compiler into only using a 2-byte alignment instead of the native
one when accessing the contents turned out to even take up more space that.
9acc72f0df