Apologies but, having read the answers to the original question this one occurred to me.

Let's talk C instead of javascript and reference the Quake fast inverse square root algorithm...

What's the speed like if your multiply function takes 2 (or an array of) pointers as its input parameters and ...

Fkr each input pointer (/array of)

• cast as pointer to a float (ie cast the pointers, not the values)
• (bit)shift and add/subtract on the "floats"
• return value from (/array of) pointer to (/array of) float result using pointers of original type

.. or just return the result as a (/array of) pointer(s) to floats with the caller reading it using a (array of) pointer to original type?