Sorting floating-point values, and NaN

[Andrei Alexandrescu]

On 11/12/13 10:31 AM, tn wrote:
> On Tuesday, 12 November 2013 at 15:56:25 UTC, Andrei Alexandrescu wrote:
>> On 11/12/13 12:54 AM, tn wrote:
>>> ...
>>> An alternative would be to check for nans explicitly.
>>
>> I think NaNs are singular unordered values that make invalid inputs
>> for either sort or isSorted. We should simply add an assert to isSorted.
>
> But sort and isSorted both allow user to provide a custom "less"
> function. What if the user needs NaNs and, for example, wants to sort
> them before all the other values? Then he calls "arr.sort!((a,b) => a <
> b || (isnan(a) && !isnan(b)))" and the code asserts?

I think this is a misunderstanding. I never advocated inserting 
specialized asserts for floating-point numbers. http://goo.gl/3dGGkf 
takes care of that the right way - by making sure that the less 
predicate is sensible for the kind of work isSorted and sort do. That is 
regardless of the nature of the objects being sorted.

Andrei