PHP RFC: Union Types 2.0

• Date: 2019-09-02
• Author: Nikita Popov nikic@php.net
• Proposed Version: PHP 8.0
• RFC PR: php/php-rfcs#0001
• Mailing list thread: https://externals.io/message/106844
• Implementation: php/php-src#4838

A “union type” accepts values of multiple different types, rather than a single one. PHP already supports two special union types:

• Type or null, using the special ?Type syntax.
• array or Traversable, using the special iterable type.

However, arbitrary union types are currently not supported by the language. Instead, phpdoc annotations have to be used, such as in the following example:

class Number {
    /**
     * @var int|float $number
     */
    private $number;
 
    /**
     * @param int|float $number
     */
    public function setNumber($number) {
        $this->number = $number;
    }
 
    /**
     * @return int|float
     */
    public function getNumber() {
        return $this->number;
    }
}

The statistics section shows that the use of union types is indeed pervasive in the open-source ecosystem, as well as PHP's own standard library.

Supporting union types in the language allows us to move more type information from phpdoc into function signatures, with the usual advantages this brings:

• Types are actually enforced, so mistakes can be caught early.
• Because they are enforced, type information is less likely to become outdated or miss edge-cases.
• Types are checked during inheritance, enforcing the Liskov Substitution Principle.
• Types are available through Reflection.
• The syntax is a lot less boilerplate-y than phpdoc.

After generics, union types are currently the largest “hole” in our type declaration system.

Union types are specified using the syntax T1|T2|... and can be used in all positions where types are currently accepted:

class Number {
    private int|float $number;
 
    public function setNumber(int|float $number): void {
        $this->number = $number;
    }
 
    public function getNumber(): int|float {
        return $this->number;
    }
}

Union types support all types currently supported by PHP, with some caveats outlined in the following.

The void type can never be part of a union. As such, types like T|void are illegal in all positions, including return types.

The void type indicates that the function has no return value, and enforces that argument-less return; is used to return from the function. It is fundamentally incompatible with non-void return types.

What is likely intended instead is ?T, which allows returning either T or null.

The null type is supported as part of unions, such that T1|T2|null can be used to create a nullable union. The existing ?T notation is considered a shorthand for the common case of T|null.

An earlier version of this RFC proposed to use ?(T1|T2) for nullable union types instead, to avoid having two ways of expressing nullability in PHP. However, this notation is both rather awkward syntactically, and differs from the well-established T1|T2|null syntax used by phpdoc comments. The discussion feedback was overwhelmingly in favor of supporting the T1|T2|null notation.

?T remains valid syntax that denotes the same type as T|null. It is neither discouraged nor deprecated, and there are no plans to deprecate it in the future. It is merely a shorthand alias for a particularly common union type.

The null type is only allowed as part of a union, and can not be used as a standalone type. Allowing it as a standalone type would make both function foo(): void and function foo(): null legal function signatures, with similar but not identical semantics. This would negatively impact teachability for an unclear benefit.

Union types and the ?T nullable type notation cannot be mixed. Writing ?T1|T2, T1|?T2 or ?(T1|T2) is not supported and T1|T2|null needs to be used instead. I'm open to permitting the ?(T1|T2) syntax though, if this is considered desirable.

While we nowadays encourage the use of null over false as an error or absence return value, for historical reasons many internal functions continue to use false instead. As shown in the statistics section, the vast majority of union return types for internal functions include false.

A classical example is the strpos() family of functions, which returns int|false.

While it would be possible to model this less accurately as int|bool, this gives the false impression that the function can also return a true value, which makes this type information significantly less useful to humans and static analyzers both.

For this reason, support for the false pseudo-type is included in this proposal. A true pseudo-type is not part of the proposal, because similar historical reasons for its necessity do not exist.

The false pseudo-type cannot be used as a standalone type (including nullable standalone type). As such, all of false, false|null and ?false are not permitted.

To catch some simple bugs in union type declarations, redundant types that can be detected without performing class loading will result in a compile-time error. This includes:

• Each name-resolved type may only occur once. Types like int|string|INT result in an error.
• If bool is used, false cannot be used additionally.
• If object is used, class types cannot be used additionally.
• If iterable is used, array and Traversable cannot be used additionally.

This does not guarantee that the type is “minimal”, because doing so would require loading all used class types.

For example, if A and B are class aliases, then A|B remains a legal union type, even though it could be reduced to either A or B. Similarly, if class B extends A {}, then A|B is also a legal union type, even though it could be reduced to just A.

function foo(): int|INT {} // Disallowed
function foo(): bool|false {} // Disallowed
 
use A as B;
function foo(): A|B {} // Disallowed ("use" is part of name resolution)
 
class_alias('X', 'Y');
function foo(): X|Y {} // Allowed (redundancy is only known at runtime)

Excluding the special void type, PHP's type syntax may now be described by the following grammar:

type: simple_type
    | "?" simple_type
    | union_type
    ;

union_type: simple_type "|" simple_type
          | union_type "|" simple_type
          ;

simple_type: "false"          # only legal in unions
           | "null"           # only legal in unions
           | "bool"
           | "int"
           | "float"
           | "string"
           | "array"
           | "object"
           | "iterable"
           | "callable"       # not legal in property types
           | "self"
           | "parent"
           | namespaced_name
           ;

Union types follow the existing variance rules:

• Return types are covariant (child must be subtype).
• Parameter types are contravariant (child must be supertype).
• Property types are invariant (child must be subtype and supertype).

The only change is in how union types interact with subtyping, with three additional rules:

• A union U_1|...|U_n is a subtype of V_1|...|V_m if for each U_i there exists a V_j such that U_i is a subtype of V_j.
• The iterable type is considered to be the same (i.e. both subtype and supertype) as array|Traversable.
• The false pseudo-type is considered a subtype of bool.

In the following, some examples of what is allowed and what isn't are given.

Property types are invariant, which means that types must stay the same during inheritance. However, the “same” type may be expressed in different ways. Prior to union types, one such possibility was to have two aliased classes A and B, in which case a property type may legally change from A to B or vice versa.

Union types expand the possibilities in this area: For example int|string and string|int represent the same type. The following example shows a more complex case:

class A {}
class B extends A {}
 
class Test {
    public A|B $prop;
}
class Test2 extends Test {
    public A $prop;
}

In this example, the union A|B actually represents the same type as just A, and this inheritance is legal, despite the type not being syntactically the same.

Formally, we arrive at this result as follows: First, A is a subtype of A|B, because it is a subtype of A. Second, A|B is a subtype of A, because A is a subtype of A and B is a subtype of A.

It is legal to remove union types in return position and add union types in parameter position:

class Test {
    public function param1(int $param) {}
    public function param2(int|float $param) {}
 
    public function return1(): int|float {}
    public function return2(): int {}
}
 
class Test2 extends Test {
    public function param1(int|float $param) {} // Allowed: Adding extra param type
    public function param2(int $param) {}       // FORBIDDEN: Removing param type
 
    public function return1(): int {}           // Allowed: Removing return type
    public function return2(): int|float {}     // FORBIDDEN: Adding extra return type
}

Similarly, it is possible to restrict a union member in return position, or widen a union member in parameter position:

class A {}
class B extends A {}
 
class Test {
    public function param1(B|string $param) {}
    public function param2(A|string $param) {}
 
    public function return1(): A|string {}
    public function return2(): B|string {}
}
 
class Test2 extends Test {
    public function param1(A|string $param) {} // Allowed: Widening union member B -> A
    public function param2(B|string $param) {} // FORBIDDEN: Restricting union member A -> B
 
    public function return1(): B|string {}     // Allowed: Restricting union member A -> B
    public function return2(): A|string {}     // FORBIDDEN: Widening union member B -> A
}

Of course, the same can also be done with multiple union members at a time, and be combined with the addition/removal of types mentioned previously.

When strict_types is not enabled, scalar type declarations are subject to limited implicit type coercions. These are problematic in conjunction with union types, because it is not always obvious which type the input should be converted to. For example, when passing a boolean to an int|string argument, both 0 and “” would be viable coercion candidates.

If the exact type of the value is not part of the union, then the target type is chosen in the following order of preference:

1. int
2. float
3. string
4. bool

If the type both exists in the union, and the value can be coerced to the type under PHPs existing type checking semantics, then the type is chosen. Otherwise the next type is tried.

As an exception, if the value is a string and both int and float are part of the union, the preferred type is determined by the existing “numeric string” semantics. For example, for “42” we choose int, while for “42.0” we choose float.

Types that are not part of the above preference list are not eligible targets for implicit coercion. In particular no implicit coercions to the null and false types occur.

The following table shows how the above order of preference plays out for different input types, assuming that the exact type is not part of the union:

// int|string
42    --> 42          // exact type
"42"  --> "42"        // exact type
new ObjectWithToString --> "Result of __toString()"
                      // object never compatible with int, fall back to string
42.0  --> 42          // float compatible with int
42.1  --> 42          // float compatible with int
1e100 --> "1.0E+100"  // float too large for int type, fall back to string
INF   --> "INF"       // float too large for int type, fall back to string
true  --> 1           // bool compatible with int
[]    --> TypeError   // array not compatible with int or string
 
// int|float|bool
"45"    --> 45        // int numeric string
"45.0"  --> 45.0      // float numeric string
"45X"   --> 45 + Notice: Non well formed numeric string
                      // int numeric string
""      --> false     // not numeric string, fall back to bool
"X"     --> true      // not numeric string, fall back to bool
[]      --> TypeError // array not compatible with int, float or bool

There are two main alternatives to the preference-based approach used by this proposal:

The first is to specify that union types always use strict typing, thus avoiding any complicated coercion semantics altogether. Apart from the inconsistency this introduces in the language, this has two main disadvantages: First going from a type like float to int|float would actually reduce the number of valid inputs, which is highly unintuitive. Second, it breaks the variance model for union types, because we can no longer say that float is a subtype of int|float.

The second is to perform the coercions based on the order of types. This would mean that int|string and string|int are distinct types, where the former would favor integers and the latter strings. Depending on whether exact type matches are still prioritized, the string type would always be used for the latter case. Once again, this is unintuitive and has very unclear implications for the subtyping relationship on which variance is based.

References to typed properties with union types follow the semantics outlined in the typed properties RFC:

If typed properties are part of the reference set, then the value is checked against each property type. If a type check fails, a TypeError is generated and the value of the reference remains unchanged.

There is one additional caveat: If a type check requires a coercion of the assigned value, it may happen that all type checks succeed, but result in different coerced values. As a reference can only have a single value, this situation also leads to a TypeError.

The interaction with union types was already considered at the time, because it impacts the detailed reference semantics. Repeating the example given there:

class Test {
    public int|string $x;
    public float|string $y;
}
$test = new Test;
$r = "foobar";
$test->x =& $r;
$test->y =& $r;
 
// Reference set: { $r, $test->x, $test->y }
// Types: { mixed, int|string, float|string }
 
$r = 42; // TypeError

The basic issue is that the final assigned value (after type coercions have been performed) must be compatible with all types that are part of the reference set. However, in this case the coerced value will be int(42) for property Test::$x, while it will be float(42.0) for property Test::$y. Because these values are not the same, this is considered illegal and a TypeError is thrown.

An alternative approach would be to cast the value to the only common type string instead, with the major disadvantage that this matches neither of the values you would get from a direct property assignment.

To support union types, a new class ReflectionUnionType is added:

class ReflectionUnionType extends ReflectionType {
    /** @return ReflectionType[] */
    public function getTypes();
 
    /* Inherited from ReflectionType */
    /** @return bool */
    public function allowsNull();
 
    /* Inherited from ReflectionType */
    /** @return string */
    public function __toString();
}

The getTypes() method returns an array of ReflectionTypes that are part of the union. The types may be returned in an arbitrary order that does not match the original type declaration. The types may also be subject to equivalence transformations.

For example, the type int|string may return types in the order [“string”, “int”] instead. The type iterable|array|string might be canonicalized to iterable|string or Traversable|array|string. The only requirement on the Reflection API is that the ultimately represented type is equivalent.

The allowsNull() method returns whether the union contains the type null.

The __toString() method returns a string representation of the type that constitutes a valid code representation of the type in a non-namespaced context. It is not necessarily the same as what was used in the original code.

For backwards-compatibility reasons, union types that only include null and one other type (written as ?T, T|null, or through implicit parameter nullability), will instead use ReflectionNamedType.

// This is one possible output, getTypes() and __toString() could
// also provide the types in the reverse order instead.
function test(): float|int {}
$rt = (new ReflectionFunction('test'))->getReturnType();
var_dump(get_class($rt));    // "ReflectionUnionType"
var_dump($rt->allowsNull()); // false
var_dump($rt->getTypes());   // [ReflectionType("int"), ReflectionType("float")]
var_dump((string) $rt);      // "int|float"
 
function test2(): float|int|null {}
$rt = (new ReflectionFunction('test2'))->getReturnType();
var_dump(get_class($rt));    // "ReflectionUnionType"
var_dump($rt->allowsNull()); // true
var_dump($rt->getTypes());   // [ReflectionType("int"), ReflectionType("float"),
                             //  ReflectionType("null")]
var_dump((string) $rt); // "int|float|null"
 
function test3(): int|null {}
$rt = (new ReflectionFunction('test3'))->getReturnType();
var_dump(get_class($rt));    // "ReflectionNamedType"
var_dump($rt->allowsNull()); // true
var_dump($rt->getName());    // "int"
var_dump((string) $rt);      // "?int"

This RFC does not contain any backwards incompatible changes. However, existing ReflectionType based code will have to be adjusted in order to support processing of code that uses union types.

The features discussed in the following are not part of this proposal.

Intersection types are logically conjugated with union types. Instead of requiring that (at least) a single type constraints is satisfied, all of them must be.

For example Traversable|Countable requires that the passed value is either Traversable or Countable, while Traversable&Countable requires that it is both.

The mixed type allows to explicitly annotate that any value is acceptable. While specifying no type has the same behavior on the surface, it does not make clear whether the type is simply missing (because nobody bothered adding it yet, or because it can't be added for backwards compatibility reasons), or whether genuinely any value is acceptable.

We've held off on adding a mixed type out of fear that it would be used in cases where a more specific union could have been specified. Once union types are supported, it would probably also make sense to add the mixed type.

The false pseudo-type introduced in this RFC is a special case of a “literal type”, such as supported by TypeScript. They allow specifying enum-like types, which are limited to specific values.

type ArrayFilterFlags = 0|ARRAY_FILTER_USE_KEY|ARRAY_FILTER_USE_BOTH;
array_filter(array $array, callable $callback, ArrayFilterFlags $flag): array;

A benefit of using a union of literal types instead of an enum, is that it works directly with values of the underlying type, rather than an opaque enum value. As such, it is easier to retrofit without breaking backwards-compatibility.

This RFC intentionally supports the false type in a maximally restricted form, which is enough to model internal function return values, but avoids unnecessarily constraining a future proposal for introducing first-class literal types. In particular:

• No values implicitly coerce to false, while it would also be possible to follow bool parameter coercion semantics, restricted to input values that coerce to false. Both approaches have advantages, but we pick the conservative option, which permits future extension, here.
• Only false is supported, but not true. Once both are supported, the subtyping relationship between false|true and bool needs to be defined (which is also tightly related to the question of implicit coercions).

As types become increasingly complex, it may be worthwhile to allow reusing type declarations. There are two general ways in which this could work. One is a local alias, such as:

use int|float as number;
 
function foo(number $x) {}

In this case number is a symbol that is only visible locally and will be resolved to the original int|float type during compilation.

The second possibility is an exported typedef:

namespace Foo;
type number = int|float;
 
// Usable as \Foo\number from elsewhere

Simple yes/no vote.

To illustrate the use of union types in the wild, the use of union types in @param and @return annotations in phpdoc comments has been analyzed.

In the top two thousand composer packages there are:

• 25k parameter union types: Full JSON data
• 14k return union types: Full JSON data

In the PHP stubs for internal functions (these are incomplete right now, so the actual numbers should be at least twice as large) there are:

• 336 union return types
• of which 312 include false as a value

This illustrates that the false pseudo-type in unions is necessary to express the return type of many existing internal functions.