KPHP type system

Here we will learn:

  • what types exist in KPHP
  • how do they occur in source code
  • and how they are expressed in PHPDoc

You are supposed to be aware, that KPHP infers types unless you declare them manually.

KPHP types diagram


int is a whole 64-bit number, same as PHP integer, declared as int64_t in C++. Every whole number is int:

$a = 0;                    // int
const ADMIN_ID = 2180312;  // int

There is no unsigned int, as far as there are no int16 or int32 types. Primarily because they can't be auto inferred.

Addition, multiplication, substraction remain int. Bitwise operations infer ints. This works identically with PHP except for 64-bit overflow:

$a = A::ADMIN_ID * 2 + 1;  // inferred as int
// overflow works differently: PHP converts to double (=9.22E+18), KPHP is int (=INT_MIN)
$a = PHP_INT_MAX + 1;      // inferred as int

Cast operator (int) exists and is a valid PHP syntax:

$id = (int)$user['id'];   // int
$num = (int)"-42";        // int
$ugly = (int)[1,2,3];     // inferred int, but casting array to int is strange, what does it mean?

In PHPDoc, you just write ‘int'. PHP 7 type hint int is the same.

/** @param int $arg */


float is a 64-bit floating-point number, same as PHP real, declared as double in C++.

$a = -0.4;             // float
const EPSILON = 1e-10; // float

Math operations with floats infer float. Int is convertible to float. Division of ints is float:

$a = 3 + 5.5;   // float
$a = 1 / 2;     // float
$a = 2 / 2;     // float (1.0), not int (1)!

Cast operator (float) exists and is a valid PHP syntax:

$coeff = (float)$_GET['coeff'];   // float
$num = (float)"4.42";             // float

In PHPDoc, you just write ‘float'. PHP 7 type hint float is the same.

/** @param float $arg */


bool is a logical true or false value, same as PHP bool, declared as bool in C++.

$a = true;       // bool
$b = false;      // bool unless it is not mixed with other types

Logical operations sometimes infer bool.

$a = 5 > 2;      // bool
$a = !!f();      // bool

Cast operator (bool) exists and is a valid PHP syntax. Its action depends on an operand type:

$str_not_empty = (bool)$some_string;   // converts to bool like PHP, handling "" and "0" strings
$arr_not_empty = (bool)$some_array;    // for example, converting an array to bool is true if len > 0
if ($some_array) {}                    // implicit (bool) conversion 

In PHPDoc, you just write ‘bool' or ‘boolean'. PHP 7 type hint bool is the same.

/** @param bool $arg */


string is a binary-safe sequence of single-chars, same as PHP string, declared as string in C++ (custom implementation of PHP strings).

$a = 'asdf';           // string
$b = "asdf{$a}vasdf";  // string
__FUNCTION__;          // string

Concatenation is string. Taking an index of a string is string. Heredoc/nowdoc are strings.

$s = 5 . "asdf";       // string
$c = $s[0];            // string

Cast operator (string) exists and is a valid PHP syntax:

$id = (string)$user['name'];  // string 
$num = (string)42;            // string

In PHPDoc, you just write ‘string'. PHP 7 type hint string is the same.

/** @param string $arg */


array is a pseudotype meaning “an array of any elements”, behaves like PHP array, declared as array<T> in C++ (custom implementation of PHP arrays).
array<T> and T[] are synonyms.

Like PHP array, it can store both hashmaps and vectors. Storing vectors is memory-optimized.

All elements' type is the same — T. KPHP infers T from all assignments and modifications of an array:

$a = [1,2,3];         // int[]
$u = ['o'=>1,'t'=>2]; // int[]
$b = [1,2,3.5];       // float[]
$c = [new User];      // User[]
$d = [];              // unknown[] unless modified

$e = [1,2,3];
$e[] = 4.12;
// $e infers as float[], as float is inserted (important!)

Getting an element returns an element of type T. Getting an unexisting element returns T, not null.
This is probably the most important difference between PHP and KPHP. It is discussed in a separate section.

Assigning by any index infers array. Modifying a foreach-by-ref re-infers T. Reading at any index returns T:

$options['html'] = 'div';  // string[] (though not initialized earlier)
$s = $options[$any_idx];   // string

// $a will be inferred float[], as mixing ints and floats
$a = [1,2,3];
foreach ($a as &$v)
  $v = (float)$v;

If an array is empty, T can't be inferred, so accessing its elements leads to an error:

$a = [];
$a[1];         // compilation error: Unknown type

Cast operator (array) exists and is a valid PHP syntax:

$arr_of_s = (array)"s";             // creates array ["s"], still a strange intention 
$friends = (array)$user['friends']; // let KPHP know it's array of something when getting from untyped

In PHPDoc, ‘array' means “array of what would be automatically inferred”, same as PHP 7 array type hint.
Better use T[] syntax to manually specify and control types. Nested arrays are also supported.

/** @param int[] $arg */
/** @param bool[][] $options */

You can’t mix unmixable types in an array. For instance, numbers and objects are not mixable.
That’s why array(1, new User) would be a compilation error, as KPHP can’t infer T: such T doesn’t exist.

null, T|null

null is a keyword that adds nullability to a type T used, making the type T|null unless T can store null itself, declared as Optional<T> in C++ (custom implementation for false / null additions).
T|null and ?T are synonyms. ?T is a preferred form.

$id = something ? 1 : null;      // ?int
[1, 2, null, 4];                 // (?int)[]
$arr = something ? ["1"] : null; // ?string[]

Assigning null anywhere adds nullability:

$eq = true;           // will be ?bool
if (something) 
  $eq = null;

Objects and mixed can store null, that's why mixing null with them don't change type:

$user = something ? new User : null;    // User

Cast operator not_null() drops nullability from a type. Use it when you are 100% sure, that var is not null here:

class A {
  /** @var int[]|null */
  static $cache = null;

// suppose you are sure $cache is filled
$c = not_null(A::$cache);   // int[]

In PHPDoc, ‘T|null' and ?T are the same. ‘?' has the least priority: ‘?int[]' is ‘int[]|null', not ‘(int|null)[]'. ‘?' and ‘|' can't mix in one type. For primitives, you can use PHP 7 nullable type hints.

/** @param ?int $arg */
/** @var string|null */
/** @return ?float[][] */

false, T|false

false is a keyword that adds falseability to a type T used, making the type T|false unless T can store false itself, declared as Optional<T> in C++ (custom implementation for false / null additions).

$id = something ? 1 : false;      // int|false
[1, 2, false, 4];                 // (int|false)[]
$arr = something ? ["1"] : false; // string[]|false

Objects can't be mixed with false, only with null:

$user = something ? new User : false;    // error, use null

Lots of standard PHP functions return “something, or false or error”: strpos() returns int|false, substr() returns string|false, etc. That's the main reason why false is handled in such a way. T|false exists primarily for optimization, as it is more efficient than mixed.

$pos = strpos("hello", "h");        // int|false
$sub = substr($str, 0, 2);          // string|false

False can be mixed with null, giving T|false|null:

$pos = something ? strpos(...) : null;   // int|false|null

Cast operator not_false() drops falseability from a type. Use it when you guarantee that a value isn't false actually:

$pos = not_false(strpos("hello", "h"));    // int
not_false(something ? strpos(...) : null); // ?int

In PHPDoc, you write ‘T|false'. Important! Not ‘T|bool', but ‘T|false':

/** @param int|false $arg */
/** @return (int|false)[]|false|null */


mixed == (int OR float OR bool OR null OR string OR array of mixed)

mixed is a type that can store any primitive, declared as var in C++ (implementation of “variant” storage).

If a variable is mixed — then it is bool/int/etc. at runtime. While writing code, you keep in mind, what runtime type it actually holds. Lots of PHP operators work in different ways for different runtime types:

$a + $b;        // mixed + mixed = mixed
// if they are ints/floats, they are just summed
// if they are strings, they are converted to numbers and summed
// if they are arrays (hashtables), it is logical summation with dropping duplicate keys
// if they are arrays (vectors), it's the same, but raises a warning, because you probably meant merge

KPHP runtime implementation behaves exactly like PHP, maybe with the exception in some strange cases (substracting array from a number, for example).

The word mixed is often met in the PHP documentation, meaning anything. In KPHP, ‘mixed' is not anything, it is any primitive. For instance, objects and tuples are not compatible with mixed.

mixed is automatically inferred if you mix various primitive types:

$a = something ? 1 : "1";              // mixed
return $single ? $result : [$result];  // mixed

Getting a value from superglobals is mixed:

$_GET['user_id'];  // mixed

Lots of standard PHP functions return mixed. For example, unserialize() returns mixed, as the serialized string can represent array/string/int/etc. For example, microtime() returns mixed, as if passed true argument it returns float, if false or none it returns an array.

json_decode($str);              // mixed
parse_url($url, PHP_URL_PASS);  // mixed, as without second parameter parse_url() returns array

Array keys are mixed, as they can be integers or strings. If you mix various primitives in an array, it would be inferred as mixed[]. Getting element from mixed[] or mixed is mixed:

$user = ['id' => 3, 'name' => "Alice"];    // mixed[]
$user[$any_key];                           // mixed
$name = $user['name'];                     // mixed
$first_char = $name[0];                    // mixed

To suppress mixed inferring and get cleaner types, you can cast a mixed to a primitive:

$name = (string)$user['name'];  // string
(int)$user['name'];             // int; if ['name'] would be a numeric string — ok, otherwise you get 0 + warning

For mixed arguments, you can pass anything convertible to mixed. T|false and ?T are convertible if T is.

function f($arg) { /* assume $arg inferred/declared as mixed */ }

f(true);         // bool -> mixed
f($_REQUEST);    // mixed[] -> mixed
f(strpos(...));  // int|false -> mixed
f(new User);     // error! objects are not convertible to mixed

But remember: when you pass a primitive, it will be wrapped to mixed at runtime. It is especially noticeable when passing T[] to mixed, as all inner element must be converted:

// suppose $arg was inferred as mixed
function f($arg) { /* ... */ }

f([1,2,3]);  // this leads to int[] -> mixed[] -> mixed runtime conversion; for big arrays takes time

In PHPDoc, you can write mixed or you can pipe any types that lead to a primitive mixture:

/** @param mixed $arg */
/** @var int|string $key */

objects (instances)

object is not a type, it's a synonym to a word instance. There is no “any object” type in KPHP, there is “an instance of exact class”. Objects of class T are declared as class_instance<T> in C++ (custom implementation of refcounting above C++ structures behaving like PHP).
Like in PHP, you declare classes with fields and methods. Static fields are like global variables, but instance fields are the real structure of an object.
KPHP supports almost all PHP features: extends, implements, traits, late static binding, etc. It has some limitations due to compilability and some restrictions on class hierarchy, that you'll never face in practice probably.

class User { /* ... */ }

$u = new User;          // User
$a = [new User, null];  // User[]

objects are not compatible with mixed. You can't pass a string and an object to the same function argument. You can't mix primitives and objects in an array.

return ['exists' => true, 'user' => $user];   // error: invalid types of array elements

Objects are much more effective than associative arrays. All fields of objects are typed, and it’s the reason why they work times faster than in PHP.
The next chapter is fully dedicated to objects.


tuple(T1,T2,…) is an unnamed container to carry different types, declared as std::tuple<T1,T2,…> in C++.

Tuples are especially effective to return multiple values from a function with proper type handling.

Without tuples, types are spoilt (lacking performance) or impossible:

// without tuples: 
return [$html, $friend_ids];     // mixed[]: type spoiling
return [$exists, $user, 4.4];    // error: can't mix primitives with objects

To have cleaner types, you can create an instance with typed fields. But tuples here are much more convenient:

return tuple($html, $friend_ids);   // tuple(string, int[])
return tuple($exists, $user, 4.4);  // tuple(bool, User, float)

You can assign tuple to list(), and types will be tracked:

list($html, $ids) = buildFriendsList();  // $html : string, $ids : int[]

In plain PHP, tuples are just arrays, but KPHP tracks keys independently:

// PHP polyfill
function tuple(...$args) { return $args; }

Tuples are read-only. You can access tuples only by a constant key. foreach and other array-like behavior triggers compilation errors.

$t = tuple(1, "s");
$t[$some_key];        // error! access allowed only by constant keys
$t[0] = 2;            // error! tuples are read-only

Tuples are not convertible to mixed, but tuple(…)|false and ?tuple(…) are allowed. Not only you can return tuples, but also create arrays of them, store in class fields, etc.

class A { 
  public $a = false;   // will be inferred tuple(string, int)|false

  function __construct(int $i) { $this->a = tuple("s", $i); }

In PHPDoc, you write tuple(T1, T2, …) with arbitrary nesting:

/** @param tuple(int, string|false) $arg */
/** @return tuple(User, ?string, tuple(int, Message[])) */

By default, IDE can’t parse tuples in PHPDoc, as it is KPHP-specific. Also, IDE doesn’t suggest arrows:
list($exists, $user) = f(); $user→... // IDE doesn't understand, that $user is User
To fix this, install KPHPStorm plugin, which brings extended PHPDoc types, tuples support, and more.


shape(k1:T1, k2:T2, …) is a named tuple — associative array with fixed structure, declared as shape<std::index_sequence<hash1, hash2, …>, T1, T2, …> in C++ (custom implementation).

Shapes primarily are suitable for passing lots of optional parameters:

// without shapes: $options is just mixed[]
function renderUI($rows, $options) {
  if ($options['surround_with_div']) { /* ... */ }
  $html = $options['boundary_html'];
  // ... what keys are available inside $options? how to get autocomplete and compile-time checks? 

Shapes are like tuples, but with named indexes. Like instances, but without classes and reference behavior:

 * @param shape(surround_with_div: bool, boundary_html: string, cur_user: User) $options
function renderUI($rows, $options) { 
  // you still access shape elements like an array
  // every element has a separate type, primitives and instances as separate keys
  $options['cur_user']->name;   // ok, compiler knows it's User

// while invoking, you MUST wrap the array with shape():
renderUI($rows, shape(['surround_with_div' => false, 'boundary_html' => '', 'cur_user' => new User]));

When calling shape(), you can pass only arrays with constant-string keys. You can't pass a variable or use variables as key names. Accessing shapes is also allowed only by constant keys.

Shapes can have not obligatory elements, declared as nullable types:

 * @param shape(check_rights:bool, server_name:?string, a:?A) $options 'check_rights' is obligatory, others are not
function f($options) { /* ... */ }

f(shape(['check_rights' => false]));                         // ok
f(shape(['server_name' => 's', 'check_rights' => true]));    // ok
f(shape(['server_name' => 's']));                            // error, 'check_rights' missing

Shapes can be vararg: shape(properties, ...). It means: a part is known, a part can be any other keys:

 * @param shape(i:int, ...) $sh
function demo($sh) { /* ... */ }
demo(shape(['i' => 1]));
demo(shape(['i' => 3, 'k1' => 1, 'k2' => new B]));

It is useful when a function accepts $options with a lot of parameters and then passes $options to other functions, which operate concrete fields. To add a new field to the public interface of $options, it's not necessary to add them to private implementations. This is very close to subtyping:

/** @param shape(upload_server_name: string, ...) $options */ 
function connectToUploadServer($options) { /* ... */ }

/** @param shape(min_left_space: int, ...) $options */
function checkSpaceOnServer($server, $options) { /* ... */ }

/** @param shape(upload_server_name: string, min_left_space: int, check_space: ?bool) $options */
function uploadPhoto($options) { 
  $server = connectToUploadServer($options); 
  if ($options['check_space']) 
    checkSpaceOnServer($server, $options);

Shapes are not hashtables, they don't produce overhead. Indexes are constant, that's why everything is pre-calculated during compile-time, there are no names at runtime.

Shapes are essential to transfer old untyped code to a typed variant without involving instances, leaving array access and allowing type mixture.

Don't forget to surround arrays with shape() call while passing. KPHPStorm highlights invalid shapes usage and provides keys autocompletion.


future<T> and future_queue<T> are related to async programming, declared as int in C++.

Read about async programming and coroutines (forks) in the dedicated article.


void is “type absence”, used only at function return when it doesn't return anything, declared as void in C++.

function f() {
  echo "hello\n";
  // function doesn't return anything => it returns void

You can manually declare, that function returns void — to get a compiler error if you later break this rule.

In PHPDoc, you can write @return void, this works identically to PHP 7 type hint void:

/** @return void */

When compiler option KPHP_REQUIRE_FUNCTIONS_TYPING = 1, you need to write @return to all functions.
If missing, @return void is assumed. More info here


callable type indicates that this argument/variable can hold any anonymous function (lambda), almost identical to PHP callable or Closure, declared as an instance of a special lambda-codegenerated class in C++.

function invocator(callable $f, $arg) { $f($arg); }

invocator(function($arg) {});
invocator(function($arg) use($extra) {});

To have compile-type checks, you can use typed callables. In PHPDoc, they are expressed with callable(T1, T2, …): ReturnT:

 * @param int[] $ints
 * @param callable(int): bool $cb
 * @return int[]
function filter($ints, $cb) { /* ... */ }
filter([1,2,3], function($v) { return $v > 2; });  // ok
filter([1,2,3], function($v) { return $v; });      // error: expected bool, got int

If ReturnT is missing, void assumed. Examples:

callable(int, string) : mixed
callable(int, tuple(bool, A), ?User): void
callable(): bool

Lambdas passed to typed callables can capture variables with use: capturing doesn't affect typing. Like any other types, callables can be a field of a class, there may be an array of equally-typed callables, etc:

class OnFinishCallbacks {
  /** @var (callable(): void)[] */
  static private $callbacks = [];
  /** @param callable(): void $cb */
  static function add($cb) { self::$callbacks[] = $cb; }

OnFinishCallbacks::add(function() use($x) { /* ... */ });

Internally, callable is an instance of a class with __invoke() method, classes are codegenerated implicitly.


any is a pseudotype meaning “something that would be auto inferred”. It has no C++ analog, as it is pseudotype.

Using any in @var/@param/@return is identical to absence, as it just relies on type inferring:

 * @param any $arg   "automatically inferred", like no @param at all
 * @return any       "automatically inferred", like no @return at all

any is a valid notation if you can't express a type with PHPDoc properly because it depends on external circumstances. For instance, arguments of template functions are likely to be annotated with any.

any is also useful to deal with covariance/contravariance in inheritance when you want to specify PHPDoc for a base method, but children methods have different types:

class Section {
   * @param int|string $k
   * @param any $v
  static function add($k, $v) { static::$cache[$k] = $v; }

class IntSection extends Section {
  /** @var int[] */
  static $cache = [];

class StringSection extends Section {
  /** @var string[] */
  static $cache = [];

Here if we pass non-int to IntSection::add(), type restrictions will fail; same for StringSection. But we can't write a concrete type to base Section::add(), because neither string nor int is correct, it depends on children, that's why we use any — don't lock type at all.

array pseudotype is in fact synonym to any[]: an array of something, that would be auto inferred.

The type system is rather huge but quite simple.
In two words: 1) primitives, can be combined to mixed; 2) objects/tuples/etc, they can not.
Next, read more about objects, which carry structure and are much more effective than associative arrays.