maps: document that DeleteFunc fails in presence of NaN keys

[candlerb]

What version of Go are you using (go version)?

Playground 1.21: https://go.dev/play/p/-NzVmrPm-ln

Does this issue reproduce with the latest release?

Yes: https://go.dev/play/p/-NzVmrPm-ln?v=gotip

What operating system and processor architecture are you using (go env)?

N/A

What did you do?

package main

import (
	"fmt"
	"maps"
	"math"
)

func build(n int) map[float64]int {
	m := map[float64]int{}
	for i := 0; i < n; i++ {
		m[math.NaN()] = i
		//m[float64(i)] = i
	}
	return m
}

func main() {
	m := build(10)
	for k, v := range m {
		fmt.Printf("%f %d\n", k, v)
	}
	fmt.Printf("-----\n")
	maps.DeleteFunc(m, func(k float64, v int) bool {
		return v == 7
	})
	for k, v := range m {
		fmt.Printf("%f %d\n", k, v)
	}
}

(Code based on this from @rsc)

What did you expect to see?

One element deleted from the map: the one where DeleteFunc's predicate returns true (v == 7).

What did you see instead?

No elements are deleted from the map.

However, if I replace m[math.NaN()] = i with m[float64(i)] = i, so that the map has all distinct non-NaN keys, it works.

[seankhliao]

I believe this is unfortunate but working as intended.
clear is the only way NaN keys can be removed.

[candlerb]

clear is the only way NaN keys can be removed.

I didn't see this documented either with maps.DeleteFunc, or with map types or clear. Maybe it should be?

Aside: there's some possibly unfortunate language relating to min/max here: "assuming all NaNs are equal". NaNs are not equal - not even to themselves. But I believe that all NaNs are equivalent or fungible.

[seankhliao]

The inequality of NaNs come from:

Two floating-point values are compared as defined by the IEEE-754 standard.

[candlerb]

Yes I understand that. I was querying the language of the golang spec which says "assuming all NaNs are equal"

[randall77]

We should document this behavior, if nothing else. I'll reopen for that.

It is possible we could let DeleteFunc to dive into the runtime to delete entries which have NaN keys. Since DeleteFunc is already released with the current behavior, changing its behavior is a breaking change. Although I think it's pretty easy to justify this as a bug fix, at the same time I'm not seeing any strong reason to fix it, and compatibility argues for keeping the behavior as is.

[bcmills]

(Compare #20660.)

[gopherbot]

Change https://go.dev/cl/553157 mentions this issue: maps: document that DeleteFunc can't remove NaN keys