proposal: spec: enums as an extension to types

[deanveloper]

Yet another enum proposal

Related: #19814, #28438

First of all, what is the issue with const? Why can't we use that instead?

Well first of all, iota of course only works with anything that works with an untyped integer. Also, the namespace for the constants are at the package level, meaning that if your package provides multiple utilities, there is no distinction between them other than their type, which may not be immediately obvious.

For instance if I had my own mat (material) package, I'd want to define mat.Metal, mat.Plastic, and mat.Wood. Then maybe classify my materials as mat.Soft, mat.Neutral, and mat.Hard. Currently, all of these would be in the same namespace. What would be good is to have something along the lines of mat.Material.Metal, mat.Material.Plastic, mat.Material.Wood, and then mat.Hardness.Soft, mat.Hardness.Neutral, and mat.Hardness.Hard.

Another issue with using constants is that they may have a lot of runtime issues. Consider the
following:

var ErrInvalidWeekday = errors.New("invalid weekday")

type Weekday byte

const (
	Sunday Weekday = iota
	Monday
	Tuesday
	// ...
)
func (f Weekday) Valid() bool {
	return f <= Saturday
}

func (d Weekday) Tomorrow() Weekday {
	if !d.Valid() {
		panic(ErrInvalidWeekday)
	}
	
	if d == Sunday {
		return Saturday
	}
	return d + 1
}

Not only is there a lot of boilerplate code where we define the "enum", but there is also a lot of boilerplate whenever we use the "enum", not to mention that it means that we need to do runtime error checking, as there are bitflags that are not valid.

I thought to myself. What even are enums? Let's take a look at some other languages:

C

typedef enum week{Sun,Mon,Tue,Wed,Thu,Fri,Sat} Weekday;

Weekday day = Sun;

This ends up being similar to Go's iota. But it suffers the same pitfalls that we have with iota, of course. But since it has a dedicated type, there is some compile-time checking to make sure that you don't mess up too easily. I had assumed there was compile-time checking to make sure that things like Weekday day = 20 were at least compile-time warnings, but at least with gcc -Wextra -Wall there are no warnings for it.

C++

This section was added in an edit, originally C and C++ were grouped together, but C++11 has added enum class and enum struct which are very similar to Java's (next section). They do have compile-time checking to make sure that you don't compare two different types, or do something like Weekday day = 20. Weeday day = static_cast<Weekday>(20) still works, however. We should not allow something like this. #28987 (comment)

Syntax:

enum class Weekday { sun, mon, tues, ... };

Weekday day = Weekday::sun;
Weekday day2 = static_cast<Weekday>(2); // tuesday

Java

An enum is a kind of class. This class has several static members, named after the enum values you define. The type of the enum value is of the class itself, so each enum value is an object.

enum Weekday {
	SUNDAY(), // parentheses optional, if we define a constructor, we can add arguments here
	MONDAY,
	TUESDAY,
	// ...
	SATURDAY;
	
	// define methods here
	
	public String toString() {
		// ...
	}
}

I personally like this implementation, although I would appreciate if the objects were immutable.

The good thing about this implementation is that you are able to define methods on your enum types, which can be extremely useful. We can do this in Go today, but with Go you need to validate the value at runtime which adds quite a bit of boilerplate and a small efficiency cost. This is not a problem in Java because there are no possible enum values other than the ones you define.

Kotlin

Kotlin, being heavily inspired by Java, has the same implementation. They are even more clearly
objects, as they are called enum class instead of simply enum.

Swift

Proposal #28438 was inspired by these. I personally don't think they're a good fit for Go, but it's a different one, so let's take a look:

enum Weekday {
	case Sunday
	case Monday
	case Tuesday
	// ...
}

The idea becomes more powerful, as you can define "case functions" (syntax is case SomeCase(args...), which allow something like EnumType.number(5) being separate from EnumType.number(6). I personally think it is more fitting to just use a function instead, although it does seem like a powerful idea.

I barely have any Swift experience though, so I don't know the advantages of a lot of the features that come with Swift's implementation.

JavaScript

const Weekday = Object.freeze({
	Sunday:  Symbol("Sunday"),
	Monday:  Symbol("Monday"),
	Tuesday: Symbol("Tuesday"),
	// ...
});

This is probably the best you can do in JavaScript without a static type system. I find this to be a good implementation for JavaScript, though. It also allows the values to actually have behavior.

Okay, so enough with other languages. What about Go?

We need to ask ourselves, what would we want out of enums?

1. Named, Immutable values.
2. Compile-time validation. (We don't want to have to manually check at runtime to see if enum values are valid)
3. A consise way to define the values (the only thing that iota really provides)

And what is an enum? The way that I have always seen it, enums are an exhaustive list of immutable values for a given type.

Proposal

Enums limit what values a type can hold. So really, enums are just an extension on what a type can do. "Extension" perhaps isn't the right word, but the syntax should hopefully make my point.

The enum syntax should reflect this. The proposed syntax would be type TypeName <base type> enum { <values> }

package mat // import "github.com/user/mat"

// iota can be used in enums
type Hardness int enum {
	Soft = iota
	Neutral
	Hard
}

// Enums should be able to be objects similar to Java, but
// they should be required to be immutable. A readonly types
// proposal may help this out. Until then, it may be good just to either
// have it as a special case that enum values' fields cannot be edited,
// or have a `go vet` warning if you try to assign to an enum value's field.
type Material struct {
	Name string
	Strength Hardness
} enum {
	Metal = Material{Name: "Metal", Strength: values(Hardness).Hard } // these would greatly benefit from issue #12854
	Plastic = Material{Name: "Plastic", Strength: values(Hardness).Neutral }
	Foam = Material{Name: "Foam", Strength: values(Hardness).Soft }
}

// We can define functions on `Material` like we can on any type.

// Strong returns true if this is a strong material
func (m Material) Strong() bool {
	return m.Strength >= Hardness.Neutral
}

The following would be true with enums:

• int enum { ... } would be the type that Hardness is based on. int enum { ... } has the underlying type int, so Hardness also has the underlying type int.
• Assigning an untyped constant to a variable with an enum type is allowed, but results in a compile error if the enum does not support the constant expression's value (That's a long winded way of saying var h Hardness = 1 is allowed, but var h Hardness = 100 is not. This is similar how it is a compile error to do var u uint = -5)
• As with normal types, assigning a typed expression to a variable (var h Hardness = int(5)) of a different type is not allowed
• There is a runtime validation check sometimes, although this can be ommited in most cases. The runtime check occurs when converting to the new type. For instance var h Hardness = Hardness(x) where x is an integer variable.
• Using arithmetic operators on enums with underlying arithmetic types should probably either not be allowed, or be a runtime panic with a go vet flag. This is because h + 1 may not be a valid Hardness.

Syntax ideas for reading syntax values:

1. Type.Name
   • It's a common syntax people are familiar with, but it makes Type look like a value.
2. Type#Name, Type@Name, etc
   • Something like these would make the distinction that Type is not a value, but it doesn't feel familiar or intuitive.
3. Type().Name
   • This one doesn't make too much sense to me but it popped in my head.
4. values(Type).Name, enum(Type).Name, etc
   • values would be a builtin function that takes a type, and returns its enumeration values as a struct value. Passing a type that has no enum part would of trivially return struct{}{}. It seems extremely verbose though. It would also clash as values is a pretty common name. Many go vet errors may result from this name. A different name such as enum may be good.

I personally believe values(Type).Name (or something similar) is the best option, although I can see Type.Name being used because of it's familiarity.

I would like more critique on the enum definitions rather than reading the values, as that is mainly what the proposal mainly focuses on. Reading values from an enum is trivial once you have a syntax, so it doesn't really shouldn't need to be critiqued too much. What needs to be critiqued is what the goal of an enum is, how well this solution accomplishes that goal, and if the solution is feasible.

Points of discussion

There has been some discussion in the comments about how we can improve the design, mainly the syntax. I'll take the highlights and put them here. If new things come up and I forget to add them, please remind me.

Value list for the enum should use parentheses instead of curly braces, to match var/const declaration syntax.

• Advantage: More consistent with the rest of the language
• Disadvantage: Doesn't quite look as nice when declaring enums of structs

Perhaps changing the type syntax from <underlying type> enum ( values ) to enum <underlying type> ( values ).

• Advantage: int enum ( ... ) -> enum int ( ... ) and similar become more readable and consistent with other languages.
• Advantage: Ambiguities such as []byte enum ( ... ) get resolved to either enum []byte ( ... ) or []enum byte ( ... ).
• Disadvantage: struct { ... } enum ( ... ) -> enum struct { ... } ( ... ) becomes less readable.
• Disadvantage: (In my eyes) it doesn't illustrate how this enum implementation works quite as well.

Add type inference to enum declarations

• Advantage: Definitions become more concise, especially when declaring inline types with enums.
• Disadvantage: The concise-ness comes with a price to readability, in that the original type of the enum is not in a consistent location.
• My Comment: Type inference in Go is typically done in places which would benefit from it often, like declaring a variable. There really should be very few enum declarations "per capita" of code, so I (personally) think the verbosity of requiring the type is justified.

Use the Type.Value syntax for reading enum values

• I've already talked about advantages and disadvantages to this above, but it was mentioned that we already use Type.Method to reference methods, so it wouldn't be quite as bad to reference enum values as Type.Value.

Ranging over enum values is not discussed

• I forgot about it when writing the original text, but luckily it doesn't undermine the proposal. This is an easy thing to fit in though. We can use Type.Slice which returns a []Type

Regarding zero values

• We have two choices - either the first enum value, or the zero value of the underlying type.
• First enum value: Makes more intuitive sense when you first look at it
• Zero value of type: More consistent with the rest of Go, but may cause a compile error if the zero value of the type is not in the enum
• My Comment: I think the zero value of the type should be used. The zero value of a type is always represented as all-zeros in binary, and this shouldn't change that. On top of that, the only thing the enum "attachment" to a type does is limit what values variables of the type can hold. So under this rationale, I think it makes intuitive sense that if the enum for a type doesn't include the zero-value, then declaring a variable with the zero-value should fail to compile. This may seem strange at first, but as long as the compile error message is something intuitive (ie illegal assignment to fooVar: FooType's enum does not contain value <value>) it shouldn't be much of a problem.

[ianlancetaylor]

Is there a section missing after "Let's take a look at some other languages"?

[ianlancetaylor]

A commonly requested feature for enums is a way to iterate through the valid enum values. That doesn't seem to be supported here.

You discuss converting an integer value to Hardness, but is it also permitted to convert a value to Material? If not, what is the essential difference?

I'm not sure but I suspect that this syntax is going to introduce parsing ambiguities. There are many places where a type can appear. You also have to consider cases like

for _, x := range []struct { f int } enum { A = /* what do I write here? */

[deanveloper]

Is there a section missing after "Let's take a look at some other languages"?

Yes, that was my bad. I've updated to include other languages.

A commonly requested feature for enums is a way to iterate through the valid enum values. That doesn't seem to be supported here.

Oops... That's my bad. Either way, once there is a syntax for it, it should be easily supported. In your upcoming example I will use values(Type).slice, which evaluates to a []Type.

I'm not sure but I suspect that this syntax is going to introduce parsing ambiguities. There are many places where a type can appear. You also have to consider cases like...

That's true. It's confusing if []struct { f int } enum { ... } is an enum of []struct { f int } or a slice of struct { f int } enum { ... }, making the contents of the enum very confusing. This also isn't even really that contrived of a case either, as type TSlice []T isn't uncommon. I'd personally assume that the enum is the last thing that is "applied" to the type, since the syntax is type Type <underlying type> enum <values>, but this isn't immediately obvious.

Under that assumption, the ambiguous code would work as:

for i, x := range values([]struct { f int } enum { A = []struct{f int}{struct{f int}{5},struct{f int}{2} }).slice {
    fmt.Println(i, x)
}

which would have the same output as:

for i, x := range [][]struct{f int}{{struct{f int}{5}, struct{f int}{2}}} {
	fmt.Println(i, x)
}

https://play.golang.org/p/d3HtDbyFZVp

[networkimprov]

Non-primitive enums (i.e. struct, array) is an interesting concept!

We should also consider
a) the var/const (...) declaration syntax
b) defining a namespace from a parent type name

enum Name ( global = "Atlas" ) // defines namespace & typename; underlying type inferred
type Name enum ( global = "Atlas" ) // alternatively

var word enum ( stop = "red"; go = "green" ) // inline enum applies to one variable
                                             // no namespace or typename defined

type Person struct {
   name Name
   status enum ( follows byte = iota; leads ) // namespace is Person
}

func f() {
   v := Person{ status: follows, name: Name.global } // Person namespace is implicit
   v.status = Person.leads

   word = stop // global namespace
   word = Name.global // compile error

   for n := range Name { ... } // iterate over enum with typename
}

type Material struct {
   kind enum ( Metal = 1 ... )
   hardness enum ( Malleable = 1 ... )
}
enum Composite ( gold = Material{Metal, Malleable} )

(My up-thumb is a qualified one :-)

[deanveloper]

a) the var/const (...) declaration syntax

The reason behind picking { ... } over ( ... ) was that it just looked more visually appealing when defining enums of struct types.

example:

type Example struct {
    i int
} enum (
    A = Example{1}
    B = Example{2}
)

versus

type Example struct {
    i int
} enum {
    A = Example{1}
    B = Example{2}
}

The symmetry of } enum { looks much nicer. I would say using parentheses does make more sense, since we are declaring a list of variables. I was pretty tied on which one to use.

b) defining a namespace from a parent type name

I addressed this, I didn't like it because Type.Value makes Type look like a value, even though it isn't. It does feel much more familiar to other languages however.

Something that bothers me about the example code is that I don't like the type inference. I think that since we don't typically need to declare enums too often, the extra verbosity makes the code much more readable. For instance:

type SomeNumbers enum (
    A = 15
    B = 92
    C = 29993
    D = 29.3
    E = 1939
)

What is the underlying type of SomeNumbers? Well you'd look at the first few numbers and think it's an int type, but because the D constant is 29.3, it would instead have to be float64. This is not immediately obvious and would be an especially large problem for long enums. Just using type SomeNumbers float64 enum ( ... ) would mitigate this issue

[networkimprov]

Type.Value makes Type look like a value, even though it isn't

Type.Method is how you reference a method.

I don't like the type inference

Leveraging the const/var (...) pattern, let the first item clarify the type:

type SomeNumbers enum ( A float64 = 15; B = 92 ... )

Which makes type T struct { ... } enum { ... } unnecessary. (It isn't that readable to my eye.)

[deanveloper]

Type.Method is how you reference a method.

Fair point, completely forgot about that. One of those features that don't get used much, haha

Leveraging the const/var (...) pattern, let the first item clarify the type:

I personally think that

type Status int enum (
    Success = iota
    TimedOut
    Interrupted
    // ...
)

is more readable than

type Status enum (
    Success int = iota
    TimedOut
    Interrupted
    // ...
)

Although that's just be a matter of opinion. I think that the extra verbosity helps the readability in this case, and since people shouldn't be declaring enums all the time, it comes at a relatively low cost.

Which makes type T struct { ... } enum { ... } unnecessary. (It isn't that readable to my eye.)

I actually thought about forcing the underlying struct and enum definition to be separate (which this would effectively do). It was actually the initial design, but as I made examples, it raised a few issues.

Now, you have defined two types (an internal one for the structure, and the type with the enum to actually be used). So now you have this useless internal type floating around, which is only used to define an enum. Not a huge deal per se, but it's pretty inconvenient and seems like a waste of typing. It would also clutter up autocompleters for those who use them.

Another issue is documentation. Presumably you wouldn't want that struct to be exported, because it's only use is to be used by the enum. The issue with not exporting the struct is that now it doesn't appear on godoc.org, so people don't know what fields your enum has. So your two choices are to either export your struct, which is bad design since it's not supposed to be used outside of the enum, or to keep the struct unexported, which makes it invisible to godoc. The type T struct { ... } enum { ... } fixes this issue, since the underlying type is defined along with the enum rather than separate.

Also, defining the enum all at once illustrates that enums are an extension on types, and not types on their own. Doing type SomeNumbers enum ( ... ) makes it look like that enum ( ... ) is the underlying type, even though it's actually an int where enum just limits which int values that SomeNumbers can hold. The proposed syntax is a bit more verbose, but I think that a syntax where the underlying type is defined along with the type illustrates how it works a bit better.

Also, if you want to define the struct and enum separately, you still can:

type internalT struct { ... }
type T internalT enum ( ... )

type SomeNumbers enum ( A float64 = 15; B = 92 ... )

Even in the const/var pattern, that doesn't work. B would still be an int since you assign an untyped integer to it: https://play.golang.org/p/NnYCrYIsENm

Either way, this is all just syntax. I'm glad that there haven't been any problems with the actual concept yet!

[jonas-schulze]

I think you are using the wrong associativity: type T enum int {} should resolve the ambiguity.

I like the overall concept, but how would you cover "multi-value" enums like this one: https://play.golang.org/p/V7DAZ1HWkN4? From the top of my head one could use

• enum (which for base type int would be numbered +=1 *=2) and exclusive enum (which for base type int would be numbered *=2 +=1) or
• enum and bitflag (like https://github.com/mvpninjas/go-bitflag does),

but this would add yet another keyword to the language. How would one extract the bit information? Using a & b != 0 feels clumsy, but a ~= b is even more syntax to consider.

Update: I already mixed up the base values. 🙈

[deanveloper]

I think you are using the wrong associativity: type T enum int {} should resolve the ambiguity.

That's true that it would solve ambiguity with existing data structures since "extensions" on types are usually prefixes (ie slices/arrays, channels, etc).

Apply this to structs and it gets a bit more messy in my eyes, especially if we adopt the (...) syntax rather than {...}

type Example enum struct {
    i int
} (
    A = Example{1}
    B = Example{2}
)

versus

type Example struct {
    i int
} enum (
    A = Example{1}
    B = Example{2}
)

I think the second example illustrates that Example is just a struct { i int } and that the enum is simply an extension to that type, rather than enum struct { ... } being some entirely new concept.

[jonas-schulze]

But Example isn't just a struct { int }, is it? I think it really is a struct { int } wrapped as an enum. Putting enum first would also be compatible with future type additions to the language. However, that's all syntax. What do you think about the "bit flag" use case for enums?

[networkimprov]

You only need to specify the value type if its literal form isn't unique, i.e. numeric types other than int & float64.

enum int8 (a=iota; b) is indeed a good way to achieve that, and
enum (a int8 = iota; b) should also work for consistency with var/const.

Anonymous value types have problems...

type E struct { // looks like a struct type
  ...
} enum (        // surprise, an enum type
  a = E{...}    // surprise, it works like a struct type here
)
var x = E{...}  // surprise, error

Anonymous enum types are valuable (as I described above)...

type S struct {
   v enum (a = 1)
}