🚀 go-pugleaf

In article <1992Sep17.120636.29070@linus.mitre.org>, jrv@truth.uucp (Vanzandt) writes:
> It seems to me that time_t should be a double.  It would allow representation
> of any time during the life of the universe, with good resolution (microseconds)
> for times near the present.  I haven't seen a copy of the standard.  Would
> such an implementation conform?  Surely it's possible for a conforming
> implementation to represent my birthday?  (hint: before 1970).  (There were
> a few astronomical events of interest before 1970, too.)


That's wrong. Using a floating point representation for precise data
like this is very dangerous.

I compiled this code on a 486 running Dell Unix System V R4.0 :

main()
{
	double	a=1000000000000000.0;
	double	b=10000000000000000.0;
	
	printf("a=%f.\n",a);
	a+=1;
	printf("a=%f.\n",a);

	printf("b=%f.\n",b);
	b+=1;
	printf("b=%f.\n",b);
}

Here are the results :

a=1000000000000000.000000.
a=1000000000000001.000000.
b=10000000000000000.000000.
b=10000000000000000.000000.

As you see, "1" is too "far away" from b value to be taken into
consideration by the add operation : adding it to b doesn't change the
mantissa value. If the floating poing mecanism can manage numbers of a
huge size, it's by loosing accuracy is some situations.

To use floating points numbers without understanding this, is often a
way to produce wrong results without knowing why. (for instance
designing a financial package built over floating point numbers with a
small mantissa).

The solution is to use what's call "big numbers" : numbers that can
allow you to perform your calculations of nearly any types with
accuracy but are not (yet ?) directly known by the micro-processor, as
integers are or as (when using a math co-processor) floating points
numbers are.

When you use "big numbers", you are trading speed (you need to use
librairies and not directly processor's intructions to manipulate the
date) for accuracy.


-- 
--------------------------------------------------------------------------
Bernard Fouche  VOX   : +(33) (1) 48.04.56.30  FAX :  (33) (1) 48.04.98.95
Email : bernard@cpio1.fr.mugnet.org
Postal Mail : CPIO S.A., 13-15 Rue de la Verrerie, 75004 Paris, France