I believe this is meant to be used in combination with a low-bits pointer tagging scheme. One of the low-bits tags represents a float value, and you insert this

I believe this is meant to be used in combination with a low-bits pointer tagging scheme. One of the low-bits tags represents a float value, and you insert this tag into the NaN-space of a IEEE-754 single-precision float in the full lower 32-bits. The upper 32-bits contains the single-precision float you are interested in. The optimization is only for single-precision floats. Other types just follow the low-bits tagging scheme.