So you are suggesting that if the inliner doesn't happen to expose the
argument statically, then we should fall back to unaligned SSE move
instructions?

I'm not necessarily suggesting anything. :) I'm just saying that there's a way to support both align and unaligned versions using one primop. We might still have to expose two versions to users (wrapping the primop) in case we don't believe the compiler will recover the alignment.