Sorting floating-point values, and NaN

[tn]

On Tuesday, 12 November 2013 at 19:39:19 UTC, Vladimir Panteleev 
wrote:
> On Tuesday, 12 November 2013 at 19:26:12 UTC, tn wrote:
>> Thus, the correct solution would be to modify the functions to 
>> take "lessOrEqual" as a template parameter instead of "less".
>
> Too late for that now... just like it's too late to change IEEE 
> comparison logic.

Yes, I was expecting this. :)

>> If "less" is kept, then we need another template argument to 
>> separate equality and incomparability (something like 
>> "equals(a,b)", or "incomparable(a,b)").
>
> Is this to allow generic partial ordering (rather than to fix 
> sorting NaNs specifically)? I remember seeing a thread about 
> partial ordering and opCmp:
> http://forum.dlang.org/post/dpfbtu$1ocf$1@digitaldaemon.com

I had generic partial orderings in my mind too, but I think it is 
also needed for user defined floating point types such as 
bigfloat. (Unless "isnan" function is defined for the bigfloat 
type too and detected automatically.)