[Oberon] *****SPAM***** Re (2): V4, 64 bit version;

[strohm at airmail.net]

On Fri, 15 Jan 2021 14:37:35 -0800, peter at easthope.ca wrote:

> From: Charles Perkins 
> Date: Fri, 15 Jan 2021 13:17:03 -0800
>> On review of the source, I see that LONGREAL is currently aliased to REAL
>> in ORB, suggesting that REAL may stay 32-bit and LONGREAL would be 64-bit.
>> Would the same be done for SET, introducing LONGSET perhaps?
>
> From the new Project Oberon book I had the impression that the gain of
> one type rather than two outweighed the loss of storing some numbers
> in capacities larger than necessary.
>
> If so, wouldn't any machine have only REAL, according to hardware
> capability, whether 32 bits or 64 bits or 128 bits?
>
> What are the requirements motivating interest in 64 bits? Numerical
> modeling?

32-bit REAL means 24 bits of mantissa and 8 bits of exponent.  This just isn't enough to give acceptable numerical stability or precision.  On simple 32-bit machines, the answer is double precision, or LONGREAL, which gives 48 bits of mantissa and 16 bits of exponent, which is acceptable.

By contrast, the Control Data 6600 (and its little brother, the 6400, where PASCAL started) had a 60-bit word.  60-bit REAL on the 6600 gave 12 bits of exponent and 48 bits of mantissa, which was good enough right out of the box.  I don't recall the details of 6600 double precision.



-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.inf.ethz.ch/pipermail/oberon/attachments/20210115/0edfffe0/attachment.html>