The X86 backend shuffle combining is saving us from some poor vectorised IR
extract oneuse(op.with.overflow),1