You check the tag before doing float operations

And afterwards, because floating-point arithmetic can change the value of the tag. This isn't necessary with NaN-boxing, because it uses NaN bit patterns that the hardware never generates.