On Friday, 25 July 2014 at 20:13:57 UTC, H. S. Teoh via
Digitalmars-d wrote:
Nobody said anything
about making it *impossible* to define non-linear orderings.
But opCmp() already make it impossible to define some binary
relations that are order-like for <, <= etc. You cannot return
both -1 and 1 at the same time. And >= is defined a the
complement of <. What does 0 signify, does it signify equality or
that two values cannot be ordered? Take for instance intervals:
Defining [0,1] < [2,3] makes sense, but should [1,2] vs [0,3]
give 0, or is that only for [1,2] vs [1,2]?
It makes a lot of sense to provide "sortability traits" for the
type, such as total-order etc, but then one shouldn't limit the
implementation of operators like this without tying it to a trait
of some kind. Either the presence of opCmp is a trait of the
type, or the trait has to be provided by some other means.
Related to some other comments in the thread, some sort
algorithms designed for scenarios where you have many similar
values need equality for proper or efficient sorting. Like
quicksort implementations where you partition into 3 sets (left,
pivot-equality, right). On most CPUs you get both less-than and
equality information from a single compare so how you implement
opCmp and sorting is crucial for performance…
