Floats are 64-bits wide, and need at least 1 bit to indicate that the
object reference holds a float.  To make a float an immediate object,
one would then need at least 96 bits for each object reference (vs. 32
bits now on a 32-bit platform).

This is doable, but does the performance gain outweigh the cost of the
additional memory required for applications that don't use Floats?  Or
would a solution like an object pool produce a happier medium?

Paul