[boost] Re: std_min/std_max reference return value is problematic with boost::interval

From: Michael Stevens (Michael.Stevens_at_[hidden])
Date: 2004-03-09 04:41:22

Eric, Guillaume, (sorry about the previously half send while I was composing)

Thanks for the info, I think the solution is partly simple, partly complex.

> Eric: What strikes me as odd with this overload, though, is that max(a,b) in 
> this case will return something that is not equal to either a or b. That 
> seems counter-intuitive. Is this right?

Firstly it is now clear to me (as stated in the documentation) that 
max(interval_a,interval_b) is not std::max. The difference in semantics is 
however only represented by the difference in return value.

> Me: Why gcc thinks it needs a temporary here is beyond me.
I was confused here. I thought std_max would be choosing the std::max overload 
in this case and not the boost::numeric::interval::max overload. The later is 
however more specialised.

> Guillaume: Consequently, since std_max has been defined so that it
> returns a reference, it can't be used when one argument is an interval
> and GCC is right to complain.
GCC is correctly complain. The new std_max/ std_min should ONLY be used as a 
where we specifically require std::max/min. Therefore the committed changes 
to interval (particular utility.hpp) are incorrect. 

> Eric: The fix is to detect whether the unqualified call to max returns an 
> lvalue or not and define the return type of std_max appropriately. Easy 
> enough.
> 
Given that std::max/min in the standard are defined to return a reference this 
fix would seem unnecessary/wrong. We don't want to use boost::std_max in this 
case.

> However, in the particular case of uBLAS, the problem lies elsewhere
> imho. At lines 620-621, 688-689, 753-754 of ublas/traits.hpp in the
> equals function, uBLAS should not be using the norm_inf function. Indeed
> norm_inf relies on abs(interval); and what is really needed is a
> norm_inf_for_equals function that would rely on norm(interval). The same
> problem will also appear later with complex intervals.
Thanks for following this up and the info. The maths needs checking here. Lets 
see what Joerg thinks. For a Boost 1.31.x release we should change the 
mentioned lines in traits.hpp to use boost::numeric::interval::min/max so we 
at least do not have a regression.

Michael

-- 
___________________________________
Michael Stevens Systems Engineering
34128 Kassel, Germany
Phone/Fax: [phone]
Navigation Systems, Estimation  and
                 Bayesian Filtering
    http://www.sf.net/Bayes++
___________________________________