Re: [boost] [floating point utilities]

From: Johan Råde (rade_at_[hidden])
Date: 2007-03-19 14:23:17

Matthias Schabel wrote:
> While porting the proposed Boost.Units (currently mcs::units) library  
> to some (mainly older)
> platforms, we have run into problems with <cmath> functions being  
> unavailable. For example,
> gcc 3.4.4 on Cygwin does not implement nexttoward, llrint, llround  
> (possibly others). Given
> that the proposed fp utilities library already gives portable  
> implementations of some TR1
> functions in <cmath>, is it possible to extend support to provide a  
> TR1 support for all functions
> aside from the special functions that John and Paul are working on  
> providing?
> 
> Matthias
> _______________________________________________
> Unsubscribe & other changes: http://lists.boost.org/mailman/listinfo.cgi/boost
> 

llrint:

I don't see how to do a platform independent implementation.
The function rounds to the nearest integer,
using the current rounding direction.
You get the current rounding direction using fegetround.
It would be very hard to do a platform independent
implementation of fegetround.

I think it will be very hard to do platform independent
implementations of any of the fe... functions in <cmath>.
You probably need assembler to implement them at all.

Does the Boost.Units library really need these functions?

nextafter, nextbefore, nexttoward:

Doable.

To get the next floating point number after a given floating point 
number you copy the bits into an integer, add 1, and copy them back.
You also need to do some special handling of

1: 0
2: std::numeric_limis<T>::min (if there are no denormals)
3: std::numeric_limits<max>::max
4: infinity
5: NaN

If there is a need for this, I might do it, but not during the next two 
months.

-- 
Johan Råde