18 Oct
2011
18 Oct
'11
10:57 a.m.
On 18 Oct 2011 at 16:25, vintagecoder at aol.com wrote:
That suggests to me there is a guarantee against loss
of precision.
Why do you feel otherwise?
When using floating-point is there any language that automatically
tracks the precision of the result? There probably is, but I've
never run into it.
Floating-point/exponential notation with a fixed number of places in
the mantissa is like a loaded shotgun. Arithmetic asociative and
distributive rules go out the window (something, sadly, a lot of
compiler writers don't understand). Every printed digit is accepted
as being part of the precise answer.
So you get 20 digits of garbage.
Some hardware floating point is worse than others. In particular,
S/360 single-precision floating point was particularly awful.
--Chuck