> I am curious what applications there are that really need a user-specified > comparison function (rather than just a sort key, like dyadic /: takes).
There must some applications out there with some sort of smart string ordering that couldn't be reduced to array ordering. And of course in a language with only scalar comparison, even string ordering is a custom comparison function. I think the way C and Rust sort developers insist on making a generic function to cover both numeric and arbitrary sorting is pretty silly. Glidesort's author says it's 10-15% slower for numeric sorting because of safety concerns that only apply to a custom comparison. And the reason the sorting function is so hard to duplicate is because of all the optimization for instruction-level parallelism that does nothing if comparison is expensive... https://news.ycombinator.com/item?id=34648947 There are some in-between cases with, say, controlling the behavior of NaNs when sorting floats that sort of justify this setup, but I don't think it'd be terribly hard to put together an API for float sorting if it were actually an issue. Marshall On Thu, Dec 21, 2023 at 06:56:43PM -0800, Elijah Stone wrote: > > The problem it has is that if elements compare equal but don't match > > then it'll never get down to an array where ~.y has length 1. > > Right, of course. Shame on me for being lazy and not bothering to write out > *./2 u/\y. > > > Modern quicksort hybrids ... always use one bit at a time. > > I am curious what applications there are that really need a user-specified > comparison function (rather than just a sort key, like dyadic /: takes). > > > ---------------------------------------------------------------------- > For information about J forums see http://www.jsoftware.com/forums.htm ---------------------------------------------------------------------- For information about J forums see http://www.jsoftware.com/forums.htm
