Issue with NaN not comparable so piling up in Set[float64]

[ldemailly]

https://go.dev/play/p/2L3NR09zIJu

package main

import (
	"fmt"
	"math"

	"fortio.org/sets"
)

func test(v float64) {
	set := sets.New(v, v, v, v)
	set.Add(v, v, v)
	fmt.Printf("Using %f : %d elems %v and has: %v\n", v, len(set), set, set.Has(v))
}

func main() {
	test(math.Pi)
	test(math.NaN())
}

outputs

# As expected
Using 3.141593 : 1 elems 3.141592653589793 and has: true
# Not great
Using NaN : 7 elems NaN,NaN,NaN,NaN,NaN,NaN,NaN and has: false

[ldemailly]

Just leaving as a doc issue because otherwise would need to pepper the API and implementation with float/NaN checks

[dolmen]

I think that NaN should be forbidden in New, Add, Remove, Has: just panic. Because it is better to fail early.

The cmp package has an isNaN function.

[ldemailly]

x != x is indeed a clever IsNaN() that was mentioned on gophers slack when this came up

panic for NaN is a reasonable suggestion, I wonder if there is a noticeable perf effect of adding that check, probably not when inlined