🚀 go-pugleaf

I don't have any references on distribution of floating point numbers
(and hope you'll be able to point us to them).

However, a few thoughts on "smaller-than-word" floating point:

- My first reaction was that you should use one of the 20 or 24 bit
floating point formats that are used by a number of DSPs - say the
AT&T chip.

- Second, it occurred to me that you could sacrifice the low order bit(s)
of the mantissa, in much the same way that many LISP systems use the
least significant bit of the integer format as a tag. If you used LSB=0
as the tag for your <32 bit tagged FP value, you wouldn't need any masking
operations on the operands before dispatching the FP operation.  However,
you would need a cleanup afterwards, probably implementing correct rounding
in software.  Trouble is, the LSB=0 tag would conflict with the most convenient
tag value for integers.  A little bit of masking for <32 bit FP operations
might be acceptable if it permitted the dynamically more frequent integer
operations to have no tag manipulation overhead. 
    You could still use the native FP hardware.  A machine like the 88K,
where FP and integers share the same register file, might be preferrable for
you.  A seperate FP register set will usually require a bit more footwork
in order to perform the "integer" tag masking operations on the FP value.

--
Andy Glew, aglew@uiuc.edu