[Nickle]Numeric type heirarchy

Keith Packard keithp@keithp.com
Fri, 23 Feb 2001 09:58:24 -0800


Nickle currently incorporates several internal numeric representations and 
exposes several numeric datatypes.  While these use the same names, the 
language groups several representations together as a single datatype.

In the implementation, there are:

	type_int		32 bit integer
	type_integer		arbitrary precision natural + sign
	type_rational		two naturals + sign
	type_double		C 'double'.

In the language, there are:

	int		type_int
	integer		type_int or type_integer
	rational	type_int, type_integer or type_rational
	double		type_int, type_integer, type_rational or type_double

I think I should remove 32-bit integers from the language; they're not 
useful.

I'm about to add a 'type_float' which uses a natural mantissa, natural
exponent two signs and a 32-bit integer indicating the number of bits in 
the mantissa.  This will eliminate the inability to convert large or small
values into an imprecise representation.

Once this new representation is available, I'd like to change the
internal names to:

	type_int		natural + sign
	type_ratio		two naturals + sign
	type_float		two naturals, two signs and a 32-bit precision

The language datatypes become:

	integer			type_int
	rational		type_ratio or type_int
	real			type_float, type_ratio or type_int

As for datatype conversions:

real imprecise (real value, integer bits)

	Convert 'value' to an imprecise repsentation using no more than 
	'bits' bits.

integer floor (real value)
integer ceil (real value)
integer round (real value)

	Convert 'value' to the nearest integer in the indicated direction.

Is there any conventional function name for converting an imprecise number 
into a nearby rational approximation?  Certainly it's easy to do:

	rational function nearby (real value, integer denom) {
		return round (value * denom) / denom;
	}

-keith