PHP RFC: Records

• Version: 0.9
• Date: 2024-07-19
• Author: Robert Landers, landers.robert@gmail.com, rob@bottled.codes
• Status: Draft (or Under Discussion or Accepted or Declined)
• First Published at: http://wiki.php.net/rfc/records

This RFC proposes the introduction of record objects, which are immutable classes with value semantics.

Value objects are immutable objects that represent a value. They’re used to store values with a different semantic by wrapping their technical value, adding additional context. For example, a Point object with x and y properties can represent a point in a 2D space, and an ExpirationDate can represent a date when something expires. This prevents developers from accidentally using the wrong value in the wrong context.

Consider this example where a function accepts an integer as a user ID, and the ID is accidentally set to a nonsensical value:

function updateUserRole(int $userId, string $role): void {
    // ...
}
 
$user = getUser(/*...*/)
$uid = $user->id;
// ...
$uid = 5; // accidentally sets uid to an unrelated integer
// ...
updateUserRole($uid, 'admin'); // accidental passes a nonsensical value for uid

Currently, the only solution to this is to use a class, but this requires significant boilerplate code. Further, readonly classes have many edge cases and are rather unwieldy.

Like arrays, strings, and other values, record objects are strongly equal (===) to each other if they contain the same values.

Let’s take a look at an updated example using a record type for UserId. Thus, if someone were to pass an int to updateUserRole, it would throw an error:

record UserId(int $id);
 
function updateUserRole(UserId $userId, string $role): void {
    // ...
}
 
$user = getUser(/*...*/)
$uid = $user->id; // $uid is a UserId object
// ...
$uid = 5;
// ...
updateUserRole($uid, 'admin'); // This will throw a TypeError

Now, if $uid is accidentally set to an integer, the call to updateUserRole will throw a TypeError because the function expects a UserId object instead of a plain integer.

This RFC proposes the introduction of a record keyword in PHP to define immutable value objects. These objects will allow properties to be initialized concisely and will provide built-in methods for common operations such as modifying properties, performing equality checks, and using a function-like instantiation syntax. Records can implement interfaces and use traits but can’t extend other records or classes; composition is allowed, however.

A record is defined by the keyword record, followed by the name of its type (e.g., UserId), and then must list one or more typed parameters (e.g., int $id) that become properties of the record. A parameter may provide private or public modifiers, but are public when not specified. This is referred to as the “inline constructor.”

A record may optionally implement an interface using the implements keyword, which may optionally be followed by a record body enclosed in curly braces {}.

A record may not extend another record or class.

A record may contain a traditional constructor with zero arguments to perform further initialization.

A record body may contain property hooks, methods, and use traits.

A record body may also declare properties whose values are only mutable during a constructor call. At any other time, the property is immutable.

A record body may also contain static methods and properties, which behave identically to static methods and properties in classes. They may be accessed using the :: operator.

As an example, the following code defines a record named Pigment to represent a color, StockPaint to represent paint colors in stock, and PaintBucket to represent a collection of stock paints mixed together. The actual behavior isn’t important, but illustrates the syntax and semantics of records.

namespace Paint;
 
// Define a record with several primary color properties
record Pigment(int $red, int $yellow, int $blue) {
 
  // property hooks are allowed
  public string $hexValue {
    get => sprintf("#%02x%02x%02x", $this->red, $this->yellow, $this->blue),
  }
 
  // methods are allowed
  public function mix(Pigment $other, float $amount): Pigment {
    return $this->with(
      red: $this->red * (1 - $amount) + $other->red * $amount,
      yellow: $this->yellow * (1 - $amount) + $other->yellow * $amount,
      blue: $this->blue * (1 - $amount) + $other->blue * $amount
    );
  }
 
  // all properties are mutable in constructors
  public function __construct() {
    $this->red = max(0, min(255, $this->red));
    $this->yellow = max(0, min(255, $this->yellow));
    $this->blue = max(0, min(255, $this->blue));
  }
 
  public function with() {
    // prevent the creation of a new Pigment from an existing pigment
    throw new \LogicException("Cannot create a new Pigment from an existing pigment");
  }
}
 
// simple records do not need to define a body
record StockPaint(Pigment $color, float $volume);
 
record PaintBucket(StockPaint ...$constituents) {
  public function mixIn(StockPaint $paint): PaintBucket {
    return $this->with(...[...$this->constituents, $paint]);
  }
 
  public function color(): Pigment {
    return array_reduce($this->constituents, fn($color, $paint) => $color->mix($paint->color, $paint->volume), Pigment(0, 0, 0));
  }
}

A record may be used much like a class, as the behavior of the two is very similar, assisting in migrating from one implementation to another:

$gray = $bucket->mixIn($blackPaint)->mixIn($whitePaint);

Records are instantiated in a function format, with & prepended. This provides visual feedback that a record is being created instead of a function call.

$black = &Pigment(0, 0, 0);
$white = &Pigment(255, 255, 255);
$blackPaint = &StockPaint($black, 1);
$whitePaint = &StockPaint($white, 1);
$bucket = &PaintBucket();
 
$gray = $bucket->mixIn($blackPaint)->mixIn($whitePaint);
$grey = $bucket->mixIn($blackPaint)->mixIn($whitePaint);
 
assert($gray === $grey); // true

A record can also be defined with optional parameters that are set if omitted during instantiation.

One or more properties defined in the inline constructor may have a default value declared using the same syntax and rules as any other default parameter in methods/functions. If a property has a default value, it is optional when instantiating the record, and PHP will assign the default value to the property if omitted.

record Rectangle(int $x, int $y = 10);
var_dump(&Rectangle(10)); // output a record with x: 10 and y: 10

To make records more useful, the RFC proposes generating a with method for each record. This method allows for partial updates to the properties, creating a new instance of the record with the specified properties updated.

Named arguments

The with method accepts only named arguments defined in the inline constructor. Properties not defined in the inline constructor can’t be updated by this method.

Variadic arguments

Variadic arguments from the inline constructor don’t require named arguments in the with method. However, mixing named and variadic arguments in the same with method call is not allowed by PHP syntax.

Using named arguments:

record UserId(int $id) {
  public string $serialNumber;
 
  public function __construct() {
    $this->serialNumber = "U{$this->id}";
  }
}
 
$userId = &UserId(1);
$otherId = $userId->with(2); // Fails: Named arguments must be used
$otherId = $userId->with(serialNumber: "U2"); // Error: serialNumber is not defined in the inline constructor
$otherId = $userId->with(id: 2); // Success: id is updated

Using variadic arguments:

record Vector(int $dimensions, int ...$values);
 
$vector = &Vector(3, 1, 2, 3);
$vector = $vector->with(dimensions: 4); // Success: values are updated
$vector = $vector->with(dimensions: 4, 1, 2, 3, 4); // Error: mixing named arguments with variadic arguments is not allowed by PHP syntax
$vector = $vector->with(dimensions: 4)->with(1, 2, 3, 4); // Success: First update dimensions, then values

A developer may define their own with method if they choose, and reference the generated with method using parent::with(). This allows a developer to define policies or constraints on how data can change from instance to instance.

record Planet(string $name, int $population) {
  // create a with method that only accepts population updates
  public function with(int $population): Planet {
    return parent::with(population: $population);
  }
}
$pluto = Planet("Pluto", 0);
// we made it!
$pluto = $pluto->with(population: 1);
// and then we changed the name
$mickey = $pluto->with(name: "Mickey"); // Error: no named argument for population

A record has two types of constructors: the inline constructor and the traditional constructor.

The inline constructor is always required and must define at least one parameter. The traditional constructor is optional and can be used for further initialization logic, but must not accept any arguments.

When a traditional constructor exists and is called, the properties are already initialized to the values from the inline constructor and are mutable until the end of the method, at which point they become immutable.

// Inline constructor defining two properties
record User(string $name, string $emailAddress) {
  public string $id;
 
  // Traditional constructor
  public function __construct() {
    if (!is_valid_email($this->emailAddress)) {
      throw new InvalidArgumentException("Invalid email address");
    }
 
    $this->id = hash('sha256', $this->emailAddress);
    $this->name = ucwords($this->name);
    // all properties are now immutable
  }
}

A record can implement interfaces, but it cannot extend other records or classes, but may use traits:

interface Vehicle {}
 
interface Car extends Vehicle {
  public function drive(): void;
}
 
interface SpaceShip extends Vehicle {
  public function launch(): void;
}
 
record FancyCar(string $model) implements Car {
  public function drive(): void {
    echo "Driving a Fancy Car {$this->model}";
  }
}
 
record SpaceCar(string $model) implements Car, SpaceShip {
  public function drive(): void {
    echo "Driving a Space Car {$this->model}";
  }
 
  public function launch(): void {
    echo "Launching a Space Car {$this->model}";
  }
}
 
record Submarine(string $model) implements Vehicle {
  use Submersible;
}
 
record TowTruct(string $model, private Car $towing) implements Car {
  use Towable;
}

From the perspective of a developer, declaring a record declares an object with the same name. The developer can consider the record function (the inline constructor) as a factory function that creates a new object or retrieves an existing object from an array.

For example, this would be a valid mental model for a Point record:

record Point(int $x, int $y) {
    public float $magnitude;
 
    public function __construct() {
        $this->magnitude = sqrt($this->x ** 2 + $this->y ** 2);
    }
 
    public function add(Point $point): Point {
        return &Point($this->x + $point->x, $this->y + $point->y);
    }
 
    public function dot(Point $point): int {
        return $this->x * $point->x + $this->y * $point->y;
    }
}
 
// similar to declaring the following function and class
 
// used during construction to allow mutability
class Point_Implementation {
    public int $x;
    public int $y;
    public float $magnitude;
 
    public function __construct() {
        $this->magnitude = sqrt($this->x ** 2 + $this->y ** 2);
    }
 
    public function with(...$parameters) {
        // validity checks omitted for brevity
        $parameters = array_merge([$this->x, $this->y], $parameters);
        return Point(...$parameters);
    }
 
    public function add(Point $point): Point {
        return Point($this->x + $point->x, $this->y + $point->y);
    }
 
    public function dot(Point $point): int {
        return $this->x * $point->x + $this->y * $point->y;
    }
}
 
// used to enforce immutability but has nearly the same implementation
readonly class Point {
    public float $magnitude;
 
    public function __construct(public int $x, public int $y) {}
 
    public function with(...$parameters): self {
        // validity checks omitted for brevity
        $parameters = array_merge([$this->x, $this->y], $parameters);
        return Point(...$parameters);
    }
 
    public function add(Point $point): Point {
        return Point($this->x + $point->x, $this->y + $point->y);
    }
 
    public function dot(Point $point): int {
        return $this->x * $point->x + $this->y * $point->y;
    }
}
 
function Point(int $x, int $y): Point {
    static $points = [];
 
    $key = hash_object($mutablePoint);
    if ($points[$key] ?? null) {
        // return an existing point
        return $points[$key];
    }
 
    // create a new point
    $reflector = new \ReflectionClass(Point_Implementation::class);
    $mutablePoint = $reflector->newInstanceWithoutConstructor();
    $mutablePoint->x = $x;
    $mutablePoint->y = $y;
    $mutablePoint->__construct();
 
    // copy properties to an immutable Point and return it
    $point = new Point($mutablePoint->x, $mutablePoint->y);
    $point->magnitude = $mutablePoint->magnitude;
    return $points[$key] = $point;
}

In reality, this is quite different from how it works in the engine, but this provides a mental model of how behavior should be expected to work.

To ensure that records are both performant and memory-efficient, the RFC proposes leveraging PHP’s copy-on-write (COW) semantics (similar to arrays) and interning values. Unlike interned strings, the garbage collector will be allowed to clean up these interned records when they’re no longer referenced.

$point1 = &Point(3, 4);
$point2 = $point1; // No data duplication, $point2 references the same data as $point1
$point3 = Point(3, 4); // No data duplication, it is pointing to the same memory as $point1
 
$point4 = $point1->with(x: 5); // Data duplication occurs here, creating a new instance
$point5 = &Point(5, 4); // No data duplication, it is pointing to the same memory as $point4

Calling clone on a record results in the same record object being returned. As it is a “value” object, it represents a value and is the same thing as saying clone 3—you expect to get back a 3.

If ->with() is called with no arguments, a warning will be emitted, as this is most likely a mistake.

Records are fully serializable and deserializable, even when nested.

record Single(string $value);
record Multiple(string $value1, string $value2);
 
echo $single = serialize(&Single('value')); // Outputs: "O:6:"Single":1:{s:5:"value";s:5:"value";}"
echo $multiple = serialize(&Multiple('value1', 'value2')); // Outputs: "O:8:"Multiple":1:{s:6:"values";a:2:{i:0;s:6:"value1";i:1;s:6:"value2";}}"
 
echo unserialize($single) === &Single('value'); // Outputs: true
echo unserialize($multiple) === &Multiple('value1', 'value2'); // Outputs: true

If a record contains objects or values that are unserializable, the record will not be serializable.

A record is always strongly equal (===) to another record with the same value in the properties, much like an array is strongly equal to another array containing the same elements. For all intents, $recordA === $recordB is the same as $recordA == $recordB.

Comparison operations will behave exactly like they do for classes, which is currently undefined.

For non-trivial values (e.g., objects, closures, resources, etc.), the === operator will return true if the two operands reference the same instances.

For example, if two different DateTime records reference the exact same date and are stored in a record, the records will not be considered equal:

$date1 = DateTime('2024-07-19');
$date2 = DateTime('2024-07-19');
 
record Date(DateTime $date);
 
$dateRecord1 = Date($date1);
$dateRecord2 = Date($date2);
 
echo $dateRecord1 === $dateRecord2; // Outputs: false

However, this can be worked around by being a bit creative (see: mental model) as only the values passed in the constructor are compared:

record Date(string $date) {
    public DateTime $datetime;
 
    public function __construct() {
        $this->datetime = new DateTime($this->date);
    }
}
 
$date1 = &Date('2024-07-19');
$date2 = &Date('2024-07-19');
 
echo $date1->datetime === $date2->datetime ? 'true' : 'false'; // Outputs: true

A \Record interface will be added to the engine to allow type hinting for records. All records implement this interface.

function doSomething(\Record $record): void {
    // ...
}

The only method on the interface is with, which is a variadic method that accepts named arguments and returns self.

A new reflection class will be added to support records: ReflectionRecord which will inherit from ReflectionClass and add a few additional methods:

• ReflectionRecord::finalizeRecord(object $instance): Record: Finalizes a record under construction, making it immutable.
• ReflectionRecord::isRecord(mixed $object): bool: Returns true if the object is a record, and false otherwise.
• ReflectionRecord::getInlineConstructor(): ReflectionFunction: Returns the inline constructor of the record as ReflectionFunction.
• ReflectionRecord::getTraditionalConstructor(): ReflectionMethod: Returns the traditional constructor of the record as ReflectionMethod.
• ReflectionRecord::makeMutable(Record $instance): object: Returns a new record instance with the properties mutable.
• ReflectionRecord::isMutable(Record $instance): bool: Returns true if the record is mutable, and false otherwise.

Using ReflectionRecord will allow developers to inspect records, their properties, and methods, as well as create new instances for testing or custom deserialization.

Attempting to use ReflectionClass or ReflectionFunction on a record will throw a ReflectionException exception.

The finalizeRecord() method is used to make a record immutable and look up its value in the internal cache, returning an instance that represents the finalized record.

Calling finalizeRecord() on a record that has already been finalized will return the same instance. Attempting to finalize a regular object will throw a ReflectionException.

The isRecord() method is used to determine if an object is a record. It returns true if the object is a finalized record.

The getInlineConstructor() method is used to get the inline constructor of a record as a ReflectionFunction. This can be used to inspect inlined properties and their types.

Invoking the invoke() method on the ReflectionFunction will create a finalized record.

The getTraditionalConstructor() method is used to get the traditional constructor of a record as a ReflectionMethod. This can be useful to inspect the constructor for further initialization.

Invoking the invoke() method on the ReflectionMethod on a finalized record will throw an exception.

The makeMutable() method is used to create a new instance of a record with mutable properties. The returned instance doesn’t provide any value semantics and should only be used for testing purposes or when there is no other option.

A mutable record can be finalized again using finalizeRecord(). A mutable record will not be considered a record by isRecord() or implement the \Record interface. It is a regular object with the same properties and methods as the record. For example, var_dump() will output object instead of record.

The isMutable() method is used to determine if a record has been made mutable via makeMutable() or otherwise not yet finalized.

In cases where custom deserialization is required, a developer can use ReflectionRecord to manually construct a new instance of a record.

record Seconds(int $seconds);
 
$example = &Seconds(5);
 
$reflector = new ReflectionRecord(Seconds::class);
$expiration = $reflector->newInstanceWithoutConstructor(); // this is a mutable object
$expiration->seconds = 5;
assert($example !== $expiration); // true
$expiration = $reflector->finalizeRecord($expiration);
assert($example === $expiration); // true

When passed an instance of a record the var_dump() function will output the same as if an equivalent object were passed — e.g., both having the same properties — except the output generated will replace the prefix text “object” with the text “record.”

record(Point)#1 (2) {
  ["x"]=>
  int(1)
  ["y"]=>
  int(2)
}

A record cannot share its name with an existing record, class, interface, trait, or function, just like a class.

Records will be autoloaded in the same way as classes.

To avoid conflicts with existing code, the record keyword will be handled similarly to enum to prevent backward compatibility issues.

Since & is currently a syntax error when prefixed on a function call, it will be used to denote a record instantiation.

PHP 8.5

N/A

N/A

Unknown.

None

None

To-do

None.

• Records for “record-like” types, such as DateTime, DateInterval, and others.

2/3 majority.

TBD

To be completed during a later phase of discussion.

• Value semantics

TBD