proposal: add package 'operators'

[mrwonko]

When using the github.com/google/go-cmp/cmp/cmpopts package, I sometimes find myself writing code like

cmpopts.SortSlices(func(a, b string) bool { return a < b })

Now that we have generics, there could be an operators package like in Python that provides the operators as functions, like

func Lt[T constraints.Ordered](lhs, rhs T) bool {
	return lhs < rhs
}

so that I can write

cmpopts.SortSlices(operators.Lt[string])

instead. Or maybe it should be LT, given Go capitalises initialisms.

[seankhliao]

see also #21498

[DeedleFake]

While a package like this could certainly be useful, I think a more general solution, like lightweight functions, would be better overall.

[golightlyb]

One generic implementation here:
https://pkg.go.dev/github.com/tawesoft/golib/v2/operator

e.g. cmpopts.SortSlices(operator.LT[string])

[mrwonko]

A quick way to turn a generic operator into a concrete function would address this better, agreed. But that requires a language-level change, while this is a library-only solution. But I'd be happy with either.

[earthboundkid]

I dismissed this proposal at first because I was distracted by the usecase which doesn't strike me as very compelling, but on further reflection, this is a great idea for what to do about math/v2. I think the proposal should be to add a new package called math/ops which has Min, Max, Abs, Compare(T) int and the operators: Eq, Gt, Gte, etc. Instead of using constraints.Ordered, it should define its own constraint, ops.Orderable and the constraints package should be quietly abandoned, see #52427.

[earthboundkid]

Can the title of the issue be changed to "proposal: add package math/ops" or to use mrwonko's original package name "proposal: add package operators"? The current title is a bit misleading.

[golightlyb]

I dismissed this proposal at first because I was distracted by the usecase which doesn't strike me as very compelling, but on further reflection, this is a great idea for what to do about math/v2. I think the proposal should be to add a new package called math/ops which has Min, Max, Abs, Compare(T) int and the operators: Eq, Gt, Gte, etc. Instead of using constraints.Ordered, it should define its own constraint, ops.Orderable and the constraints package should be quietly abandoned, see #52427.

Strings can be compared GT, LTE etc. too but i wouldn't call that maths.

[earthboundkid]

How about package sort/ops?

[golightlyb]

How about package sort/ops?

Makes sense. Or even just the sort package itself.

[earthboundkid]

Python's operators package has add, mul, mod, etc. I don't think those really make sense for Go. How about this:

package cmp // import "sort/cmp"

func Eq[T comparable](a, b T) bool
func Ne[T comparable](a, b T) bool

type Orderable interface { /* ... */ }

func Gt[T Orderable](a, b T) bool
func Gte[T Orderable](a, b T) bool
func Lt[T Orderable](a, b T) bool
func Lte[T Orderable](a, b T) bool

func Compare[T Orderable](a, b T) int

func Min[T Orderable](a, b T) T
func Max[T Orderable](a, b T) T
func Abs[T Orderable](t T) T

Any objections? Additions? Note that Compare doesn't really work for floats because of NaN, so you should use Go 1.21's math.Compare instead.

Edited.

[DeedleFake]

@carlmjohnson What do the comparison functions compare to? You've got, for example, func Eq[T comparable](other T) bool. What does other get compared against? Edit: 'twas fixed. Never mind.

Which makes me wonder: As Go doesn't have partial application, would a second set of functions that return another function make sense? Something like func EqTo[T comparable](a T) func(b T) bool such that Eq(1)(1) is true?

[earthboundkid]

D'oh, yes, that's a thinko. I somehow got thinking about it as a method. Will edit the list.

[mrwonko]

As Go doesn't have partial application, would a second set of functions that return another function make sense?

That feels inelegant, a special case solution to a general problem. Maybe lightweight anonymous functions (#21498) could be a better solution for that? Or some kind of func Partial2[T any, U any, V any](func(T, U) V, T) func(U) V helper in a separate package, although without variadic generics, you end up needing one of those per arity.

[DeedleFake]

Note that there is a proposal for variadic generics: #56462. Partial() feels kind of weird to me, though. Maybe it's just the name. Hmmm... slices.ContainsFunc(s, Partial(cmp.Eq, 3)). Yeah, it feels weird to me compared to slices.ContainsFunc(s, cmp.EqTo(3)).

[earthboundkid]

How about this then:

package cmp // import "sort/cmp"

func Eq[T comparable](a, b T) bool
func Ne[T comparable](a, b T) bool

func EqTo[T comparable](T) func(T) bool
func NeTo[T comparable](T) func(T) bool

type Orderable interface { /* ... */ }

func Gt[T Orderable](a, b T) bool
func Gte[T Orderable](a, b T) bool
func Lt[T Orderable](a, b T) bool
func Lte[T Orderable](a, b T) bool

func GtOf[T Orderable](T) func(T) bool  // GtTo isn't grammatical, so Of?
func GteTo[T Orderable](T) func(T) bool
func LtOf[T Orderable](T) func(T) bool
func LteTo[T Orderable](T) func(T) bool

func Compare[T Orderable](a, b T) int

func Min[T Orderable](val, vals ...T) T
func Max[T Orderable](va, vals ...T) T
func Clamp[T Orderable](min, val, max T) T  // Should add this too
func Abs[T Orderable](t T) T

[DeedleFake]

func Min(val, vals ...T) won't compile. I also feel like func Min(val T, vals ...T) T might tie it to variadic use too much. You would be able to do Min(1, 2, 3), but if you already had a slice you'd have to do Min(s[0], s[1:]...), which is weird, though I suppose that it would avoid the case of figuring out what to return if given no arguments or an empty slice.

Edit: greater than of doesn't work any more than greater than to does, unfortunately. I don't think there's any short thing that could be added that would really work there, unfortunately. a and b are equal vs. a is equal to... works for Eq() and EqTo(), I think, but a is greater than b is already directional, so...

[mrwonko]

Partial() feels kind of weird to me, though. Maybe it's just the name.

Another popular name for that kind of function is Bind(), as seen e.g. in C++.

[golightlyb]

These to/of functions look reqlly ugly to me and I'm not sure there's a much need compared to the plain operators commonly used for sorting. There are other uses for operators as functions but sorting is the major one.

We already have Contains in /x/slices, so no need for EqTo in a ContainsFunc. What would you want these partially applied functions for that its worth them being added to the stdlib?

[DeedleFake]

Bind() sounds a lot better to me. JavaScript uses bind(), too, though it's as a method.

[earthboundkid]

These to/of functions look reqlly ugly to me and I'm not sure there's a much need compared to the plain operators commonly used for sorting. There are other uses for operators as functions but sorting is the major one.

Yeah. I think the main use is to just pass op.Gt to slices.SortFunc. My guess is that, not counting min/max/clamp, Gt will get the most use out of anything in the package, then maybe Compare, followed by everything else in the long tail.

[earthboundkid]

New try:

package cmp // import "sort/cmp"

func Eq[T comparable](a, b T) bool
func Ne[T comparable](a, b T) bool

type Orderable interface { /* ... */ }

func Gt[T Orderable](a, b T) bool
func Gte[T Orderable](a, b T) bool
func Lt[T Orderable](a, b T) bool
func Lte[T Orderable](a, b T) bool

type Type[T Orderable] struct{ Val T }

// For type inference
func Op[T Orderable](t T) Type[T] { return Type[T]{t} }

func (Type[T]) Eq(other T) bool
func (Type[T]) Ne(other T) bool
func (Type[T]) Gt(other T) bool
func (Type[T]) Gte(other T) bool
func (Type[T]) Lt(other T) bool
func (Type[T]) Lte(other T) bool
func (Type[T]) Compare(other T) int

func Compare[T Orderable](a, b T) int

func Min[T Orderable](val T, vals ...T) T // fixed typo
func Max[T Orderable](val T, vals ...T) T
func Clamp[T Orderable](min, val, max T) T
func Abs[T Orderable](t T) T

In another package slices.BinarySearchFunc(s, cmp.Op(1).Gte).

[DeedleFake]

Introducing a type in between is an interesting option. If it wasn't for the fact that it's a wrapper, method expressions and method values would allow that single method to function in either way:

partial := Op(a).Lt // func(T) bool
full := Op[T].Lt // func(Type[T], T) bool

[gophun]

This proposal seems to be motivated by a desire for saving a couple of keystrokes, because seeing the actual operator literally on the page is clearer than hiding it behind a cryptic name.

cmpopts.SortSlices(func(a, b string) bool { return a < b })

cmpopts.SortSlices(operators.Lt[string])

[jarrodhroberson]

This reeks of worst kitchen sink Java/Python bloat. Support for defining proper definitions for operators on structs as user definable functions would be much better and help solve the comparable problem with struct we have now.

[earthboundkid]

#59488 has basically obsoleted this proposal. You could try to get more functions added to cmp though.

[rsc]

This proposal has been added to the active column of the proposals project
and will now be reviewed at the weekly proposal review meetings.
— rsc for the proposal review group

[willfaught]

Support for defining proper definitions for operators on structs as user definable functions would be much better and help solve the comparable problem with struct we have now.

Agreed. Built-in types could have methods for these things, that operations are syntactic sugar for, at least conceptually. int16 could have an Equals(int16) bool method. Generic constraints could describe these methods where appropriate.

[ianlancetaylor]

We have several times considered and rejected defining methods for built-in types. That is not this proposal. If you want to discuss that, please do it on a forum or in a different issue. Thanks.

[ianlancetaylor]

The stated reason for this proposal is to have comparison functions for sorting purposes. As @carlmjohnson points out, this has been done in the upcoming 1.21 release with the new cmp.Less and cmp.Compare functions. Is there any other strong reason to adopt this proposal? Thanks.

[rsc]

To repeat what Ian said two weeks ago, we've standardized on cmp.Compare for comparison, not less-than functions. It sounds like cmpopts should add a new function that takes a 3-way compare function.

[rsc]

Based on the discussion above, this proposal seems like a likely decline.
— rsc for the proposal review group

[rsc]

No change in consensus, so declined.
— rsc for the proposal review group