PHP RFC: New in initializers


This RFC proposes to allow use of new expressions inside parameter default values, attribute arguments, static variable initializers and global constant initializers.

Currently, code such as the following is not permitted:

class Test {
    public function __construct(
        private Logger $logger = new NullLogger,
    ) {}

Instead, it is necessary to write code along the following lines:

class Test {
    private Logger $logger;
    public function __construct(
        ?Logger $logger = null,
    ) {
        $this->logger = $logger ?? new NullLogger;

This makes the actual default value less obvious (from an API contract perspective), and requires the use of a nullable argument.

This RFC proposes to relax this restriction and allow the use of new inside certain initializer expressions.


new expressions are allowed as part of certain initializer expressions. It is possible to pass arguments to the constructor, including the use of named arguments:

// All allowed:
function test(
    $foo = new A,
    $bar = new B(1),
    $baz = new C(x: 2),
) {

The use of a dynamic or non-string class name or an anonymous class is not allowed. The use of argument unpacking is not allowed. The use of unsupported expressions as arguments is not allowed.

// All not allowed (compile-time error):
function test(
    $a = new (CLASS_NAME_CONSTANT)(), // dynamic class name
    $b = new class {}, // anonymous class
    $c = new A(...[]), // argument unpacking
    $d = new B($abc), // unsupported constant expression
) {}

New expressions are allowed in parameter default values, attribute arguments, static variable initializers and global class constant initializers. Parameter default values also include defaults for promoted properties:

static $x = new Foo;
const C = new Foo;
function test($param = new Foo) {}
#[AnAttribute(new Foo)]
class Test {
    public function __construct(
        public $prop = new Foo,
    ) {}

Unsupported positions

New expressions continue to not be supported in (static and non-static) property initializers and class constant initializers. The reasons for this are twofold:

For non-static property initializers, the initializer expression needs to be evaluated on each object creation. There are currently two places where this could happen: As part of object creation, and as part of the constructor call. Doing this as part of object creation can create issues for unserialization and any other process that is based on newInstanceWithoutConstructor() and does not want to implicitly execute potential side-effects.

Performing the initialization by injecting code in the constructor avoids the issue, but requires that constructor to actually be called. In particular, this would necessitate generating constructors for classes that do not explicitly declare them, and the disciplined invocation of such constructors from potential child constructors. The third option would be to introduce an additional initialization phase between creation and construction.

For static property initializers and class constant initializers a different evaluation order issue arises. Currently, these initializers are evaluated lazily the first time a class is used in a certain way (e.g. instantiated). Once initializers can contain potentially side-effecting expressions, it would be preferable to have a more well-defined evaluation order. However, the straightforward approach of evaluating initilizers when the class is declared would break certain existing code patterns. In particular, referencing a class that is declared later in the same file would no longer work.

As such support in these contexts is delayed until such a time as a consensus on the preferred behavior can be reached.

Order of evaluation

Initializer expressions could always contain side-effects through autoloaders or error handlers. However, support for new and the accompanying constructor calls make side-effect a more first-class citizen in initializer expressions, so it is worthwhile to specify when and in what order they are evaluated. For the contexts where new is supported under this proposal:

  • Static variable initializers are evaluated when control flow reaches the static variable declaration.
  • Global constant initializers are evaluated when control flow reaches the constant declaration.
  • Attribute arguments are evaluated from left to right on every call of ReflectionAttribute::getArguments() or ReflectionAttribute::newInstance().
  • Parameter default values are evaluated from left to right on every call to the function where the parameter is not explicitly passed.

Additionally, initializers can be accessed through Reflection, in which case the following evaluation semantics apply:

  • ReflectionFunctionAbstract::getStaticVariables(): Returns the current value of the static variables and also forces evaluation of any initializers that haven't been reached yet.
  • ReflectionParameter::getDefaultValue(): Evaluates the default value (on each call).
  • ReflectionParameter::isDefaultValueConstant() and ReflectionParameter::getDefaultValueConstantName(): Do not evaluate the default value.
  • ReflectionAttribute::getArguments() and ReflectionAttribute::newInstance(): Evaluate attribute arguments on each call.

Nested attributes

It is worth mentioning explicitly that this RFC effectively adds support for nested attributes, which were omitted from the original attributes RFC. For example, attributes of the following form are now possible:

#[Assert\All(new Assert\NotNull, new Assert\Length(max: 6))]

Backward Incompatible Changes

This RFC does not introduce any backwards-incompatible changes, and also should not break any major assumptions. The only case where something genuinely new is possible are nested attributes.

Future Scope

This RFC omits support for new in property and class constant initializers. These could be supported in the future, once we have figured out the evaluation order issues.

With the precedent set by new, it would probably also make sense to allow other call expressions. For example, this would allow use of static factor methods.


Voting opened on 2021-06-30 and closes on 2021-07-14.

Support new in initializers as proposed?
Real name Yes No
alcaeus (alcaeus)  
asgrim (asgrim)  
beberlei (beberlei)  
brzuchal (brzuchal)  
bwoebi (bwoebi)  
carusogabriel (carusogabriel)  
colinodell (colinodell)  
crell (crell)  
daverandom (daverandom)  
derick (derick)  
dharman (dharman)  
didou (didou)  
dmitry (dmitry)  
galvao (galvao)  
ilutov (ilutov)  
irker (irker)  
jasny (jasny)  
kalle (kalle)  
kelunik (kelunik)  
kguest (kguest)  
klaussilveira (klaussilveira)  
kocsismate (kocsismate)  
lcobucci (lcobucci)  
levim (levim)  
marandall (marandall)  
mbeccati (mbeccati)  
mcmic (mcmic)  
narf (narf)  
nicolasgrekas (nicolasgrekas)  
nikic (nikic)  
pajoye (pajoye)  
patrickallaert (patrickallaert)  
petk (petk)  
santiagolizardo (santiagolizardo)  
seld (seld)  
sergey (sergey)  
sirsnyder (sirsnyder)  
stas (stas)  
svpernova09 (svpernova09)  
tandre (tandre)  
theodorejb (theodorejb)  
trowski (trowski)  
twosee (twosee)  
wyrihaximus (wyrihaximus)  
zimt (zimt)  
Final result: 43 2
This poll has been closed.
rfc/new_in_initializers.txt · Last modified: 2021/07/14 07:19 by nikic