LLVM needs to know statically whether or not an SSE move is aligned---it
can't be computed at runtime. I don't think passing an extra Int#
argument (or whatever) to a primop is going to work.

Ah, but sometimes optimization exposes that information statically (i.e. by inlining). In my case I know Array# is word-aligned and if someone uses memcpy with a 0 offset the compiler might spot this and use the right alignment (this is what happens today). Now, if e.g. ByteArray# can never allocated 16-byte aligned, this doesn't help us much.