As of mid-2020, there is some discussion of adding enumerations (enums) to PHP. There are many good reasons to do so, most around enabling better data modeling and type checking, but that doesn't suggest how to do it. Enumerations in practice refer to a very wide range of functionality depending on the language, from barely above constants to a core part of the type system.
As I am wont to do, I decided the best thing to do would be to survey the existing marketplace and see what other languages did, and what we can steal outright. (As the saying goes, "PHP evolves by beating up other languages in dark alleys and going through their pockets for loose syntax.") I therefore looked at 11 different languages with some kind of native enumeration support. The survey below is intended as a reasonably fair overview and summary of the available languages. My own thoughts and analysis are included at the end. For some languages I have included runnable sample code in the appropriate subdirectory. Whether or not there is sample code depends primarily on whether I had a runtime for the language already installed.
I deliberately excluded languages with no native enum support. Languages such as Javascript, Go, or Ruby do not (as far as I can tell) have any native enumerations, although there are various hacky ways to simulate them in user space. That is not of interest to us at this time.
If you spot any errors in the survey below, please let me know.
In C, enumerations are really just a wrapper for named integer constants. They are defined with the keyword enum, usually in combination with typedef. For example:
#include<stdio.h>
typedef enum { Monday, Tuesday, Wednesday, Thursday, Friday, Saturday, Sunday } Day;
void printer(Day d) {
printf("The day is: %d\n", d);
}
int main() {
Day d = Tuesday;
printer(d);
printer(4);
return 0;
}Even though printer() takes a Day parameter, passing an integer literal works fine. If no integer value for an enum element is specified, the compiler assigns one automatically starting from 0. So Monday is 0, Tuesday is 1, etc. You can specify an equivalent integer for an enum value, including making multiple values refer to the same integer:
typedef enum {
Working,
Failed = 5,
Busted = 5;
} Status;In Java, enums are, unsurprisingly, a shorthand for classes with class constants. They can be defined standalone or within a class, since Java supports inner classes. As a result, enums can support arbitrary methods. The specific values can map to internal integer values, or they can be auto-assigned by the compiler.
The simple case looks like this:
enum Suit {
HEARTS,
DIAMONDS,
CLUBS,
SPADES
}Enum values have a number of methods on them by default to access metadata, including Suit.valueOf("HEARTS") (returns "HEARTS") and Suit.valueOf("HEARTS").ordinal() (returns 0).
The values of an enum can be iterated as a set:
for (Suit s : Suit.values()){
System.out.println(s);
} Because they're built on classes, enums can have methods.
enum Suit {
HEARTS,
DIAMONDS,
CLUBS,
SPADES;
public String color() {
switch (this) {
case SPADES:
return "Swords of a soldier";
case CLUBS:
return "Weapons of war";
case DIAMONDS:
return "Money for this art";
default:
return "Shape of my heart";
}
}
}The switch statement is not exhaustive on enums, however.
Enum constructors are always private, so they can only be called from the definition of enum members. They also support interfaces.
enum Suit {
HEARTS("H"),
DIAMONDS("D"),
CLUBS("C"),
SPADES("S");
private String abbrev;
Suit(String abbrev) {
this.abbrev = abbrev;
}
String shortName() {
return abbrev;
}
}Further reading: https://www.javatpoint.com/enum-in-java
Python builds its enum support on top of classes. An "enum class" is simply a class that extends the enum.Enum parent, which has a lot of methods pre-implemented to provide Enum-ish behavior. All properties of the class are enum members:
import enum
class Suit(enum.Enum):
HEARTS = enum.auto()
DIAMONDS = enum.auto()
CLUBS = 'C'
SPADES = "S"Enum members can be any int or string primitive, or can be auto-generated. The auto-generation logic can also be overridden by defining a _generate_next_value_() method in the class. When an enum value is cast to a string, it always shows as Card.CLUBS or similar, but can be overridden by implementing the __str__ method.
Enum member names must be unique, but values need not be. If two members have the same value then the syntactically first one wins, and all others are alises to it. The aliases will be skipped when iterating an enum or casting it to a list. If needed, you can get the original list with Card.__members__.items().
As a class, an enum can also have methods. However, the methods have no native way to vary depending on which enum value they're on. You can check the value within the method, though:
class Suit(enum.Enum):
HEARTS = enum.auto()
DIAMONDS = enum.auto()
CLUBS = 'C'
SPADES = "S"
def color(self):
if self in [self.CLUBS, self.SPADES]:
return "Black"
else:
return "Red"Because Python lacks any meaningful type declarations on variables, parameters, or return values, there's no way to restrict a value to an enum list. Enum classes also cannot be extended.
The Enum class also has an alternate function-style syntax for simple cases:
Suit = Enum('Suit', 'HEARTS DIAMONDS CLUBS SPADES')Further reading: https://docs.python.org/3/library/enum.html
Typescript supports primitive enumerations, including both constant and runtime-defined values. Depending on the details they may or may not get compiled away to literal constants in code. It has its own dedicated keyword.
enum Suit {
Hearts,
Diamonds,
Clubs,
Spades,
}is equivalent to
enum Suit {
Hearts = 0,
Diamonds = 1,
Clubs = 2,
Spades =3,
}Enums can also have string values if specified explicitly. Values can be set based on some other value, even function definitions:
enum FileAccess {
// constant members
None,
Read = 1 << 1,
Write = 1 << 2,
ReadWrite = Read | Write,
// computed members
UserSetting = userDefaultValue()
}Normally enums exist at runtime, but a fully-constant enum can also be flagged to compile-away to raw constants in the source code:
const enum ShouldWe {
No,
Yes,
}Enum types can be used as type declarations:
function pickCard(desiredSuit: Suit): Card { }Further reading: https://www.typescriptlang.org/docs/handbook/enums.html
Strictly speaking Haskell doesn't have enums, but the way its type system works gives you something close enough that I'm going to include it. In Haskell, you define a new data type with the data keyword, which can be defined in terms of other data types and type constructors.
It's really hard to explain without going into the whole type system, so I'll stick to some examples:
data Suit = Hearts | Diamonds | Clubs | SpadesThe type "Suit" has only four values, one for each suit. They are not backed by a primitive value but literally are those values only. Haskell doesn't have methods as we'd understand them in the OOP world, and I've not been able to wrap my brain around Haskell enough to say if you can attach methods consistently to types of an Enum. The can, however, be used in pattern matching:
data Color = Red | Black
suitColor :: Suit -> Color
suitColor Hearts | Diamonds = Red
suitColor Clubs | Spades = BlackBecause type values are technically not values but "type constructors" they can be parameterized by other values. For instance, the infamous Maybe Monad is defined as:
data Maybe a = Just a | NothingThat is, a "Maybe" can be either the literal Nothing or a Just combined with some other value, which can then be extracted later using pattern matching.
stuff :: Maybe a -> Int
stuff Nothing = 0
stuff Just a = aFurther reading: https://wiki.haskell.org/Type
F#, in what seems to be a very on-brand move, has both union types and enums. They are very similar but not quite the same thing.
Union types in F# look and act an awful lot like Haskell, including the requirement that the unioned types start with a capital.
type SuitUnion = Hearts | Diamonds | Clubs | SpadesThey have no underlying primitive equivalent. F#'s match directive forces you to enumerate all possible values, to help avoid errors:
let color = match x with
| Hearts -> Red
| Diamonds -> Red
| Clubs -> Black
| Spades -> BlackEnums in F#, by contrast, are backed by underlying integer primitives that you specify. Strings are not allowed. They can be all lowercase if you want, but have to be qualified when referencing to them:
type SuitEnum = Hearts = 1 | diamonds = 2 | Clubs = 3 | Spades = 4
let color = match x with
| SuitEnum.Hearts -> "Red"
| SuitEnum.diamonds -> "Red"
| SuitEnum.Clubs -> "Black"
| SuitEnum.Spades -> "Black"
| _ -> "What kind of deck are you using?"Enums can be cast to and from integers. That also, oddly, allow you to define an enum value that is out of range.
// This is, amazingly, legal.
let horseshoe = enum<SuitEnum>(5)For that reason, the _ fallback match arm is required for enums, but not for unions.
Because F# doesn't have function parameter or return types, neither unions nor enums can be type defined in a function signature.
Further reading: https://fsharpforfunandprofit.com/posts/enum-types/
C# enums are explicitly just named integer constants, much like in C. They can be defined within a class like constants, or (I think) stand-alone with a namespace.
enum Suits
{
Hearts = 0,
Diamonds,
Clubs,
Spades
}If a value is not specified, it will be set to the highest existing value + 1. 0 is the default first value but you can set your own. They are referenced scoped, so Suits.Diamonds, Suits.Spaces, etc.
Values can also be defined based on other enum values, bitmask style, such as RedCards = Hearts|Diamonds. However, that only works if the explicit values are defined as bit flags.
Enums need to be cast to an integer explicitly in order to use as an int.
Console.WriteLine((int)WeekDays.Monday);An Enum class contains various static methods for manipulating enumerations further. For instance, to get a list of the names in a given enumeration:
foreach (string str in Enum.GetNames(typeof(WeekDays))) {
Console.WriteLine(str);
}Or this somewhat crazy way to cast an integer up to an enum member:
WeekDays wdEnum;
Enum.TryParse<WeekDays>("1", out wdEnum);
Console.WriteLine(wdEnum);Although they're not a class, you can technically add "extension methods" to enums that end up looking kind of like them. For instance:
public static string Color(this Suit s) {
switch (s)
{
case Hearts: return "Red";
case Diamonds: return "Red";
case Clubs: return "Black";
case Spades: return "Black";
}
}
var theColor = Suit.Clubs.Color();Further reading: https://www.tutorialsteacher.com/csharp/csharp-enum
Swift's enumerations are closer to union types, but still called enumerations. (Go figure.) They form a full fledged type with limited legal values. That means the type has to be capitalized, and the values not.
enum Suit {
case hearts
case diamonds
case clubs
case spaces
}
// or
enum Suit {
case hearts, diamonds, clubs, spaces
}Once defined, values can be defined of that type, and Swift's type inference capability can shorten the syntax somewhat.
var card = Suit.clubs
// since card is now bound to the type Suit, you can now do this:
card = .spadesYou can match on an enum value with switch, and it must either be exhaustive or have a default:
switch card {
case .spades:
print("The swords of a soldier.")
case .clubs:
print("Weapons of war.")
case .diamonds:
print("Money for this art.")
default:
print("That's not the shape of my heart.")
}Enums are not natively iterable, but they can be converted into that easily:
enum Suit: CaseIterable {
case hearts, diamonds, clubs, spaces
}
for s in Suit.allCases {
print(s)
}Swift allows enums to have what it calls "associated values," creating what is variously called a "discriminated union" or "tagged union" depending on whom you ask. Each value can have its own set of associated values that could be the same or different.
case Suit {
case hearts(String)
case diamonds(String)
case clubs(String)
case spades(String)
}
var threeOfDiamonds = Suit.diamond("3")Each instance of an associated value enum is then not equal to another, even if they're of the same enum value. Seemingly the only way to get those values back out, though, is with pattern matching:
switch card {
case .spades(let value):
print("The \(value) of Spades")
case .clubs(let value):
print("The \(value) of Clubs")
case let .diamonds(value):
print("The \(value) of Diamonds")
case let .hearts(value):
print("The \(value) of Hearts")
}For one-off cases, you can use if let.
if case let .clubs(val) = card {
print ("The \(val) of Clubs")
}(Those all do the same thing, but digging into the intricacies of Swift's pattern matching is out of scope for now.)
Enums can also support "raw values," if specified explicitly, but they must be of the same primitive type:
enum Suit: Character {
case hearts = "H"
case diamonds = "D"
case clubs = "C"
case spades = "S"
}If you list only one raw value, Swift will try to generate a raw value for the rest based on the type used. It's also possible to initialize an enum case from a raw value, if one was defined:
let card = Suit(rawValue: "B")This actually creates an "optional" of type Suit?, meaning it may or may not be legal and you have to explicitly check it. (Optionals are essentially a syntactic Maybe Monad, and way off topic.)
You can even define an enumeration in terms of itself, which is just all kinds of weird. From the documentation:
indirect enum ArithmeticExpression {
case number(Int)
case addition(ArithmeticExpression, ArithmeticExpression)
case multiplication(ArithmeticExpression, ArithmeticExpression)
}And they go further by supporting methods on enumerations of all of the above types, the body of which would meaningfully have to be a switch:
case Suit {
case hearts(String)
case diamonds(String)
case clubs(String)
case spades(String)
func color: String {
switch self {
case .hearts: return "Red"
case .diamonds: return "Red"
case .clubs: return "Black"
case .spades: return "Black"
}
}
}
print(Suit.clubs("3").color());
// Prints "Black"Further reading: https://docs.swift.org/swift-book/LanguageGuide/Enumerations.html
As Rust's main syntactic goal seems to have been "Haskell, but with lots of curly braces," the language supports enumerations with and without associated values, either positional or named.
All of the following are legal:
// The values themselves.
enum Suit {
Hearts,
Diamonds,
Clubs,
Spades,
}
// With one or more tuple associated values.
enum Card {
Hearts(i8),
Diamonds(i8),
Clubs(i8),
Spades(i8),
}
// With one or more struct-associated values.
enum Card {
Hearts{val: i8},
Diamonds{val: i8},
Clubs{val: i8},
Spades{val: i8},
}
// With an integer (only) explicit value.
enum Suit {
Hearts = 3,
Diamonds = 4,
Clubs = 5,
Spades = 6,
}Enum values can be referenced scoped from their type, Suit::Heart, or first imported with use Suit::* and then used unqualified. Because they're a full type, they can be used in function signatures.
Enums are almost always used with either match or if let, the latter of which being a sort of inverted way to care about only a single branch of a match. The match version must be exhaustive or have a default.
let msg = match card {
Suit::Spades => "Swords of a soldier".to_string(),
Suit::Clubs => "Weapons of war".to_string(),
Suit::Diamonds => "Money for this art".to_string(),
_ => "Shape of my heart".to_string(),
};
if let Diamonds(val) = card {
println!("{} diamonds are a girl's best friend", val)
}The only way to extract associated values out of the enum is with pattern matching, which in Rust is almost absurdly robust:
use Card::*;
let the_val = match Card {
Clubs(x) | Hearts(x) | Spades(x) | Diamonds(x) => x
};As it's a full type, it can also have methods. Or in Rust-speak, "implementations," including of traits (what most languages would call an interface). They can do pretty much everything a struct can. Their body will in most cases be just a bit match.
impl Suit {
fn color(&self) -> String {
match self {
Self::Hearts => "Red".to_string(),
Self::Diamonds => "Red".to_string(),
Self::Clubs => "Black".to_string(),
Self::Spades => "Black".to_string(),
}
}
}(The capitalized Self in this case is an implicit alias to Suit.)
Further reading: https://doc.rust-lang.org/rust-by-example/custom_types/enum.html
Kotlin also has not one but two enum-esque systems: Enums and Sealed Classes. The difference between them is subtle and confusing.
Enums are a class that inherits from an Enum class implicitly.
enum class Suit {
HEARTS,
DIAMONDS,
CLUBS,
SPADES
}Each enum value is technically a "constant object." By default they're bare, but can also take int or string values constructor-style.
enum class Suit(val abbrev: String) {
HEARTS("H"),
DIAMONDS("D"),
CLUBS(C"),
SPADES("S")
}Enums in Kotlin support methods, and unlike most languages here the methods may be defined separately for each value. Technically they're all implemented as subclasses, with the parent as an abstract base class. An enum can even support interfaces.
interface Colorable {
fun color()
}
enum class Suit(val abbrev: String) {
HEARTS("H") {
override fun color(): String = "Red"
},
DIAMONDS("D") {
override fun color(): String = "Red"
},
CLUBS(C") {
override fun color(): String = "Black"
},
SPADES("S") {
override fun color(): String = "Black"
};
abstract fun color(): String
fun abbreviation(): String {
return this.abbrev
}
}Enums have a number of built-in methods and properties, which make it possible to iterate an enum or get its value. That also makes them trivially serializable, unlike Sealed Classes.
Sealed Classes, meanwhile, are almost like normal classes except that the list of subclasses is fixed at compile time and they must appear in the same source file.
Whereas Enums are singletons, sealed classes may be singleton or instance-based.
sealed class Action
// This is a singleton sealed class
object Quit: Action()
// This is an instance-able sealed class
class Move(val dir: String): Action() Since they're objects/methods in their own right, they can have whatever methods you want, inherited or not. However, they do not have the automatic methods or properties of Enums that make them serializable.
Kotlin supports a when syntax as an alternative to switch that is an expression, and can, in some cases, detect exhaustiveness.
var result = when (card) {
Suit.SPADES -> "The swords of a soldier"
Suit.CLUBS -> "Weapons of war"
Suit.DIAMONDS -> "Money for this art"
else -> "The shape of my heart"
}Further reading: https://blog.kotlin-academy.com/enum-vs-sealed-class-which-one-to-choose-dc92ce7a4df5
Scala enums are also built on objects.
package com.crell.poker {
object Suit extends Enumeration {
type Suit = Value
val HEARTS, DIAMONDS, CLUBS, SPADES = Value
}
}
// ...
object Main extends App {
import com.crell.poker.Suit._
var s = CLUBS
// Iteration
Suit.values foreach println
}They can carry values, including multiple values, which must be pre-set and not vary by instance. They also can support methods that way, although my Scala-fu is not strong enough to know if my syntax here is entirely correct. :-)
object Suit extends Enumeration {
protected case class Val(abbrev: String) extends super.Val {
def color: String = abbrev.match {
case Suit.HEARTS => "Red"
case Suit.DIAMONDS => "Red"
case Suit.CLUBS => "Black"
case Suit.SPADES => "Black"
}
}
type Suit = Value
val HEARTS = Val("H")
val DIAMONDS = Val("D")
val CLUBS = Val("C")
val SPADES = Val("S")
}Further reading: https://www.scala-lang.org/api/current/scala/Enumeration.html
Folded into a convenient table, a feature summary looks like this:
| Language | C/C++ | Java | Python | Typescript | Haskell | F# (Union) | F# (Enum) | C# | Swift | Rust | Kotlin (Enums) | Kotlin (Sealed) | Scala |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Unit values | No | No | No | No | Yes | Yes | No | No | Yes | Yes | Yes | Ish? | Yes |
| Int values | Yes | Yes | Yes | Yes | No | No | Yes | Yes | Yes | Yes | Yes | Ish? | Yes |
| String values | No | Yes | Yes | Yes | No | No | No | No | Yes | No | Yes | Ish? | Yes |
| Associated values | No | No | No | No | Yes | No | No | No | Yes | Yes | No | Yes | No |
| Methods | No | Yes | Yes | No | No? | No | No | Ish | Yes | Yes | Yes | Yes | Yes |
| Type checked | Ish | Yes | No | Yes | Yes | No | Ish | Yes | Yes | Yes | Yes | Yes | Yes? |
| Iterable | No | Yes | Yes | No | No | No | No | Yes | Yes | No | Yes | No | No |
In terms of overall capability, Swift appears to have the edge with Rust a very close second. However, Rust also seems to have more powerful associated values ability (tuples or structs), and the usefulness of iterating enum types is debatable. I'm going to call it a qualified tie between those two in raw expressive power.
Broadly speaking, I would separate the languages into a few categories:
- Fancy Constants: C, Typescript, F#
- Fancy Objects: Python, Java, C#, Scala
- Algebraic Data Types: Haskell, Swift, Rust, Kotlin
While they are superficially similar, and often use the same terminology, they approach the problem from different ways. The Fancy Constants languages are offering a syntactic convenience, but little else. Often they get compiled away at runtime, and their type checking may be incomplete.
The Fancy Objects languages take that a step further and offer methods on enum types, which offers a centralized place to put a switch, match, or whatever branching syntax for RTTI. That is helpful, and helps with data modeling in ways that Fancy Constants do not. If the methods need to vary by enum type more than just a little, though, you run into some contortions and may find yourself better off with normal objects and interfaces.
The main differentiator for ADT languages, as I'm using them here, is that they can be parameterized with different values. That offers another layer again of potential functionality and data modeling. It also becomes a natural and easy way to implement Monads in user space, and Haskell, Swift, and Rust all do exactly that in their core libraries, particularly for Maybe/Optional. That makes them an extremely robust way to handle data modeling in your application, and to "make invalid states unrepresentable," which is an excellent feature if you can get it.
The downside is that once you start parameterizing enum values, you no longer get a guarantee that a Club is a Club is a Club. They may well be two different Clubs. The implementation details here around equality (a tricky subject in the best of circumstances) are the devil's hiding place. The other catch is that, as far as I can tell, no language with parameterized enum values lets you get at them easily without doing pattern matching. Depending on your use case that may be no big deal or may be a deal-breaker. In practice, I think it largely comes down to how easy the syntax is for pattern matching; Of all things I'd say Haskell is the nicest here, followed by Swift, then Rust. (Or possibly Rust then Swift, depending on your tastes. Rust gets very tricky when you have struct-parameterized enums.)
As far as borrowing ideas for a PHP implementation, it seems silly at this point to not go all the way to ADT support if possible. "If possible" being the operative phrase, as PHP also lacks relevant features that some ADT languages use for their enums, such as pattern matching. In practice, I believe the only question is between Fancy Objects and ADTs.
The main implementation questions would be:
In practice, I see little reason to not build on objects. They're already there, and if we want methods and associated data and type checking then you're already 90% to objects. Most of the ADT implementations above build on objects, either implicitly or explicitly.
The majority of ADT/fancy class languages here use a dedicated enum syntax of some variety. The exceptions are Kotlin's sealed classes and Haskell's union types, which are not strictly speaking enums but close enough.
It has been suggested that the addition of a typedef to PHP to allow pre-definition of union types (already added in PHP 8.0, coming soon) would be "close enough" to enums to render any further effort unnecessary. That is true up to a point; the main limitation would be no way to enforce that all of the unioned types are type compatible: that they share an interface that makes it possible to type against them properly. It would also make the degenerate case of "I just want a closed list of options" more verbose.
typedef Suit = Hearts | Diamonds | Clubs | Spades;
// This is just weird
interface SuitInterface {}
class Hearts implements SuitInterface {}
class Diamonds implements SuitInterface {}
class Clubs implements SuitInterface {}
class Spades implements SuitInterface {}Sealed classes would look essentially the same, give or take some syntax. Ideally we want an approach that "scales" cleanly from the basic case to the highly complex case. I think a properly designed dedicated Enum syntax is the best way to achieve that (even if it's just syntactic sugar that decomposes to the same as above).
There are largely 3 types of "single associated value" that enums can have: Unit (just the enum itself), Int only, or any primitive. The latter two get tricky if implemented as objects in PHP, unless it's just still-more sugar. I think Kotlin has the right idea here, though; Rather than giving an enum value its own direct primitive equivalent, make that a constant associated value. That is, no, you cannot define "Diamonds" to be equal to 3, but you can define "Diamonds" to be an object with a single property whose value is always 3.
This one gets really tricky, as equality is troublesome at the best of times. Already in PHP, when two objects are "equal" is somewhat confusing.
$a = new Foo();
$b = new Foo();
print $a == $a . PHP_EOL; // True
print $a === $b . PHP_EOL; // False
print $a == $b . PHP_EOL; // TrueDepending on the implementation this may end up doing all kinds of weird thing that may or may not make sense.
To simplify an implementation for PHP, I would recommend the following:
- Enums are, mostly, sugar for abstract classes and inheritance. Yes, inheritance has its issues, but in this context its behavior is actually desireable.
- Enum values may only be object instances. They may not map to primitives. However, a well-designed
__toString()method can get us 80% of the way there with 10% of the effort, which seems like a good trade-off. Plus it emphasizes enum-as-unit, which I would argue is preferable from a modeling perspective in the majority case. - We allow some conceptual leakage in return for not needing pattern matching. This is another "80% of the benefit for 10% of the effort" case, and is consistent with Kotlin's approach anyway.
For the simple case:
enum Suit {
case Hearts;
case Diamonds;
case Clubs;
case Spades;
}is equivalent to:
class Suit extends Enum implements IteratorAggregate {
public static Hearts $Hearts;
public static Diamonds $Diamonds;
public static Clubs $Clubs;
public static Spades $Spades;
public function getIterator(): array {
return [static::Hearts, static::Diamonds, static::Clubs, static::Spades];
}
}
class Hearts extends Suit {}
class Diamonds extends Suit {}
class Clubs extends Suit {}
class Spades extends Suit {}
Suit::$Hearts = new Hearts();
Suit::$Diamonds = new Diamonds();
Suit::$Clubs = new Clubs();
Suit::$Spades = new Spades();And assumes a base class like:
abstract class Enum {
public static abstract function values(): array
public function __toString(): string {
// Via reflection:
// If this object has no properties, return the name of the class.
// If it has one property, return that property.
// If it has multiple, concatenate them in lexical order and return that.
// This method may of course be overridden.
}
}The automatic IteratorAggregate implementation happens if and only if all of the defined cases is a unit. If any of them may have associated data, then it is not generated and the enum is not iterable. (This is consistent with Swift.)
For enum members that can have one or more constant values applied to them:
enum Suit (public string $abbrev) {
case Hearts("H");
case Diamonds("D");
case Clubs("C");
case Spades("S");
}That is equivalent to:
class Suit extends Enum {
public static Hearts $Hearts;
public static Diamonds $Diamonds;
public static Clubs $Clubs;
public static Spades $Spades;
public function __construct(public string $abbrev) {}
public function getIterator(): array {
return [static::Hearts, static::Diamonds, static::Clubs, static::Spades];
}
}
class Hearts extends Suit {}
class Diamonds extends Suit {}
class Clubs extends Suit {}
class Spades extends Suit {}
Suit::$Hearts = new Hearts("H");
Suit::$Diamonds = new Diamonds("D");
Suit::$Clubs = new Clubs("C");
Suit::$Spades = new Spades("S");Of note, if enum members are going to have constant values, they must all have the same constant values, and they must be defined in the enum itself using compact constructor syntax. Whether the value is public or private is then an explicit decision of the enum author. The enum author may not, however, directly implement a constructor.
If any one of the enum members defines its own parameters, however, that means the following:
- The enum itself may not define any constant values.
- No iteration is defined.
- It follows that any other enum members must be either units (no data at all) or have their own parameters.
enum Optional {
case None;
case Some(public mixed $value);
}Is equivalent to:
class Optional extends Enum {
public static None $None;
public static function Some(public mixed $value);
}
class None extends Optional {}
class Some extends Optional {
public function __construct(public mixed $value);
}
Optional::$None = new None();Again, it's the implementer's choice if the associated values are public or private.
In all three cases, both the enum and its members may have methods, with the following restrictions:
- Any methods defined on an individual member must be defined on the enum itself, either fully or abstract.
- It therefore follows that all members must have an implementation of the same methods.
This logic is lifted directly from Kotlin.
enum Optional {
case None {
public function valueOr(mixed $default): mixed {
return $default;
}
};
case Some(protected mixed $value) {
public function valueOr(mixed $default): mixed {
return $this->value;
}
};
abstract public function valueOr(mixed $default): mixed;
public function bind(callable $c): static {
if ($this instanceof Optional::None) {
return $this;
}
return static::Some($c($this->value));
}
}Is equivalent to:
class Optional extends Enum {
public static None $None;
public static function Some(protected mixed $value);
public function bind(callable $c): static {
if ($this instanceof Optional::None) {
return $this;
}
return static::Some($c($this->value));
}
}
class None extends Optional {
public function valueOr(mixed $default): mixed {
return $default
}
}
class Some extends Optional {
public function __construct(public mixed $value);
public function valueOr(mixed $default): mixed {
return $this->value;
}
}
Optional::$None = new None();As demonstrated above, whether a method is single-instanced and checks for the enum type internally or is split into separate methods is up to the implementer.
The main enum may also implement interfaces if desired, in which case all the usual rules about inheritance and interfaces apply.
- Most existing plumbing for classes and objects applies to enums if needed. They behave in a mostly predictable fashion. That includes being able to autoload the enum itself. If and when named parameters are adopted, creating a new associated-data enum member will automatically support that syntax.
- While value-bound enums are not technically supported, the approach above combind with the
__toString()method offer a "close enough" equivalent. - In the unit case, because member instances are singletons they will always
===each other. In the associated data case, they will not. - Whether associated data should be public or private is punted to the implementer. Given PHP's general approach to visibility that seems the best option. On the plus side, it means any improvements made to PHP's visibility in the future should "just work" on enums as well. (Eg, if asymmetric visibility were ever added in a constructor-promotion-compatible way, they would become available to enums.)
- Enum members do not support arbitrary member variables beyond those defined through constructor promotion. That is a deliberate limitation because if you need that much functionality, you don't want an enum. You want a traditional class.
- Presumably Enums and Members could support attributes. What you would do with them I do not know, but I see no reason to not allow them.
- The existence of a base
Enumclass gives us a place to put future extensions or functionality, such as a default__serialize,__debugInfo, or other such magic method implementations.
- As shown here, enum members end up as public static class properties. That is suboptimal. Ideally they would be be write-once and exposed as though they were constants. This may be something that can be special cased in the engine.
- It's unclear how to handle
instanceofand references to the class. If implemented as pure sugar, that means members do become their own stand-alone objects and you'd do$c instance of Diamonds. That is suboptimal. It would be preferable to always make the name scoped to the enum itself, ie,$c instanceof Suit::Diamonds. However, it's unclear how that would interact with theDiamonds()member itself or with the static factory method in the case of associated data. This requires more investigation and possibly engine trickry. - If implemented as pure sugar, a side effect of this approach is that members become accessible as stand-alone classes to instantiate. That is not desireable. Ideally the actual implementation would be more robust and disallow that somehow.