Is there any interest for object constructor shorthand *just for stdClass*

3 years ago by tyson andre — view source

unread

Hi internals,

e.g. $x = object{key: object{'escaped-literal': $v2 }};
(equivalent to $x = (object)['key' => (object)['escaped-literal' => $v2]];)

For example, in JS, non-string key literals are surrounded in [] to unambiguously reference them.
Here, that could be object{[$key]: $value, [MY_CONST]: $other} - but (MY_CONST) would also work

https://externals.io/message/52990 was mentioned 9 years ago but would potentially conflict with future block expression RFCs
object{key: $value} is suggested here for similarity to named arguments

Benefits:

Shorter and more readable code - casting an array to an object is an unintuitive and potentially hard to remember
way to create an object literal
Make it easier for users to use objects for generic associative data instead of using arrays both for lists and associative data,
in situations where it may result in more usable APIs

For example,

$x = new stdClass{};
$x->prop = new stdClass{};
$x->prop->a = $a;
$x->prop->b = $b;
// Could be shortened to
$x = object{prop: object{a: $a, b: $b}}
// $x = (object)['prop' => (object)['a' => $a, 'b' => $b]]

This proposal is only for stdClass - there was not much interest in https://wiki.php.net/rfc/compact-object-property-assignment
and the combination of named properties and constructor property promotion helps with the readability of other classes.

Thanks,

Tyson

3 years ago by Larry Garfield — view source

unread

Hi internals,

e.g. $x = object{key: object{'escaped-literal': $v2 }};
(equivalent to $x = (object)['key' => (object)['escaped-literal' => $v2]];)

For example, in JS, non-string key literals are surrounded in [] to
unambiguously reference them.
Here, that could be object{[$key]: $value, [MY_CONST]: $other} - but
(MY_CONST) would also work

https://externals.io/message/52990 was mentioned 9 years ago but
would potentially conflict with future block expression RFCs

object{key: $value} is suggested here for similarity to named
arguments

Benefits:

Shorter and more readable code - casting an array to an object is an
unintuitive and potentially hard to remember
way to create an object literal

Make it easier for users to use objects for generic associative data
instead of using arrays both for lists and associative data,
in situations where it may result in more usable APIs

For example,
$x = new stdClass{};
$x->prop = new stdClass{};
$x->prop->a = $a;
$x->prop->b = $b;
// Could be shortened to
$x = object{prop: object{a: $a, b: $b}}
// $x = (object)['prop' => (object)['a' => $a, 'b' => $b]]
This proposal is only for stdClass - there was not much interest in
https://wiki.php.net/rfc/compact-object-property-assignment
and the combination of named properties and constructor property
promotion helps with the readability of other classes.

Thanks,

Tyson

The only data structure worse than an associative array -- from the point of view of performance, self-documentation, or usability -- is stdClass. :-) I can't recall using it in the last decade at least, and my code has been better for it.

Especially now with constructor promotion and named properties, I'd prefer to just forget that stdClass exists and encourage others to do so as well.

--Larry Garfield

3 years ago by Andreas Bittner — view source

unread

e.g. $x = object{key: object{'escaped-literal': $v2 }};
(equivalent to $x = (object)['key' => (object)['escaped-literal' => $v2]];)

For example, in JS, non-string key literals are surrounded in [] to unambiguously reference them.
Here, that could be object{[$key]: $value, [MY_CONST]: $other} - but (MY_CONST) would also work

https://externals.io/message/52990 was mentioned 9 years ago but would potentially conflict with future block expression RFCs

object{key: $value} is suggested here for similarity to named arguments

Benefits:

Shorter and more readable code - casting an array to an object is an unintuitive and potentially hard to remember
way to create an object literal

Make it easier for users to use objects for generic associative data instead of using arrays both for lists and associative data,
in situations where it may result in more usable APIs

For example,
$x = new stdClass{};
$x->prop = new stdClass{};
$x->prop->a = $a;
$x->prop->b = $b;
// Could be shortened to
$x = object{prop: object{a: $a, b: $b}}
// $x = (object)['prop' => (object)['a' => $a, 'b' => $b]]

Creating a new syntax, that is only used in context of \stdClass, feels
like overkill. Unfortunately, I cannot come up with a use case for this
syntax, outside of the \stdClass context.

I would also debate the point of the new syntax being more readable
compared to the (object) cast, but that very much depends on what the
reader is used to. I would however like to point out that shorter isn't
always better and sometimes more verbose code is easier to understand.
Imho a language shouldn't look like code-golf by recommendation. But I
don't want to and hope I didn't sidetrack the discussion.

With regards to the matter at hand I can see where you are coming from.
(object)['prop' => 1, ...] feels very indirect, needlessly creating an
array instance just to get an object out of it.

Wouldn't it be opportune to just use named arguments?

$x = new \stdClass(
     prop: new \stdClass(
         a: $a,
         b: $b
     )
);

While this isn't as short as the object{} suggestion, it doesn't add any
additional language constructs. The signature of \stdClass constructor
would change to \stdClass(...$defaultProperties). Maybe this is the
kickoff for variadic property promotion __constructor(public mixed ...$properties).

The following would be true:
new \stdClass(1, 2, 3, prop: 1.2) == (object)[1, 2, 3, 'prop' => 1.2];

On the subject of dynamic/variable property names: I would suggest that
a RFC for the syntax functionA($parameterName: 'value') should be
submitted at some point. This would replace the already working
functionA(...[$parameterName => 'value']). Maybe we even find a way of
using constants as well, which would could be something like your
suggested functionA([SOME_CONST]: 'value').
This would allow everyone to benefit from the change, not only \stdClass.

Closing I would like to make a short point with regards to the
(object) cast:
Both a cast of \stdClass to (array) and a cast of arrays to (object)
are common use cases in (JSON-)API development. It is not unheard of for
API signatures to contain certain properties of the JSON response object
only if some conditions are met. Currently the best way to achieve this
is either by dynamically adding properties to an existing \stdClass
object, or by populating an array and (object) casting it later.

Andreas

3 years ago by Hans Henrik Bergan — view source

unread

@Rowan Tommins

not saying this is a good idea or anything, but i just want to point
out that there's another way to do it in
7.4
$object = new class($foo, $foo * 2, $nextBaz) {
public int $foo;
public int $bar;
public int $baz;
public function __construct($foo, $bar, $baz) {
for($i=0;$i<func_num_args();++$i){
$name = ((new
ReflectionParameter([$this,FUNCTION],$i))->getName());
$this->$name = $$name;
}
}
};

haha finally found a use-case for variable-variables

Den 2020-10-21 kl. 22:45, skrev Rowan Tommins:

I'd recommend that if your data is arbitrary, you can use arrays, if you
it's not, you probably always should write typed value objects

This is pretty much what I was going to say. If you have truly dynamic
keys, then having the full set of array functions available - e.g.
array_map, array_filter, all the variations of sort - is almost
definitely a win. If you know the property names in advance, then
defining them on a class is going to help you catch a lot of mistakes.

We considered arrays once when migrating our PHP 5 codebase to PHP 7, but
the cost benefit analyse didn't show any worth while business aspects. The
only thing we needed to do was to add one line with new stdClass for the
PHP 7 migration to succeed.

One typical usecase we have is populating an object with the result
from a DB query or using an object to write to a DB.

This is certainly common, and even has a built-in mode in PDO.
However, I suspect that if we didn't have stdClass, we would no longer
see the need to add it just for these use cases, because they could
use an anonymous class instead.

True, we looked into anonymous classes as another option but it didn't
improve readability, rather the opposite.

Which brings me back to the feature request: what if rather than
stdClass, we look for a better syntax for creating and populating
anonymous class instances? Then you could start with a literal and add
real object behaviour to it - add methods, implement an interface,
constrain the type of properties, etc.

That's a very good point! We have code today that uses stdClass, so it
would probably only come into play for new code since the existing
code runs perfectly fine :-)

I think there are two problems we might want to solve there:

The only way to initialise the object with variables from the outer
scope is to pass them to a constructor, which is a bit messy (although
constructor property promotion improves things a bit, see example below).

There's no way to create a truly dynamic class, i.e. one with some
or all properties determined at run-time, without extensive string
manipulation and eval(). (Although you can simply write "$object = new
class{}; $object->foo=42;" to add a public, untyped property).

The proposed syntax would help with (1), but not really with (2),
since you need to know at least the number of properties in order to
write out the literal.

Examples, and some probably bad ideas, in the hope they inspire
someone to better ideas:

// 7.4
$object = new class($foo, $foo * 2, $nextBaz) {
public int $foo;
public int $bar;
public int $baz;
public function __construct($foo, $bar, $baz) {
$this->foo=$foo;
$this->bar=$bar;
$this->baz=$baz;
}
};

// 8.0
$object = new class($foo, $foo * 2, $nextBaz) {
public function __construct(public int $foo, public int $bar,
public int $baz) {}
};

// Promote properties into named parameters, implicitly defining a
constructor and passing them to it?
$object = new class(public int foo: $foo, public int bar: $foo * 2,
public int baz: $nextBaz) {};

// Extend this to accept an array spread? Means all properties have to
have the same visibility and type, or default to "public mixed"...
$data = ['foo' => 42, 'bar' => 84, 'baz' => 69]; // dynamic data, e.g.
a DB result set row
$object = new class(public int ...$data) {};

Interesting alternatives!

Regards,

r//Björn L

--

To unsubscribe, visit: https://www.php.net/unsub.php

3 years ago by Rowan Tommins — view source

unread

Closing I would like to make a short point with regards to the
(object) cast:
Both a cast of \stdClass to (array) and a cast of arrays to
(object) are common use cases in (JSON-)API development. It is not
unheard of for API signatures to contain certain properties of the
JSON response object only if some conditions are met. Currently the
best way to achieve this is either by dynamically adding properties to
an existing \stdClass object, or by populating an array and (object)
casting it later.

I don't follow; is the resulting JSON different if you cast to object,
or is there some other reason you prefer an object over using an
associative array directly?

Also, to clarify my earlier comment about stdClass not being necessary
now we have anonymous classes, I meant it very directly: every time you
write "new stdClass" you can write "new class {}" instead.

The fact that "(object)$foo" creates an instance of "stdClass" rather
than an instance of "class{}" is just a historical wart, which can be
easily replaced:

function array_to_object($arr): object {
    $obj = new class {};
    foreach ( $arr as $key => $value ) {
        $obj->{$key} = $value;
    }
    return $obj;
}

The only difference I'm aware of is that you can no longer use
"stdClass" as a type constraint, but I'm not sure what value that would
ever have. Indeed, with anonymous classes you could trivially add a much
more meaningful type constraint using an empty interface:

instead of:

function getFromDB(...): stdClass {
      $result = new stdClass;
      foreach ( $this->blah() as $column => $value ) {
            $result->{ $column } = $value;
      }
      return $result;
}

you can write:

interface DynamicDBItem {}

Regards,

--
Rowan Tommins (né Collins)
[IMSoP]

3 years ago by Mike Schinkel — view source

unread

I don't follow; is the resulting JSON different if you cast to object, or is there some other reason you prefer an object over using an associative array directly?

The by-reference semantics of objects vs arrays in PHP.

Also, to clarify my earlier comment about stdClass not being necessary now we have anonymous classes, I meant it very directly: every time you write "new stdClass" you can write "new class {}" instead.

The fact that "(object)$foo" creates an instance of "stdClass" rather than an instance of "class{}" is just a historical wart, which can be easily replaced:

function array_to_object($arr): object {
$obj = new class {};
foreach ( $arr as $key => $value ) {
$obj->{$key} = $value;
}
return $obj;
}

That pattern can have a non-insignificant performance penalty when dealing with a large number of objects, a use-case that is not infrequent when processing JSON, especially responses returned via an HTTP API.

Wouldn't it be opportune to just use named arguments?

$x = new \stdClass(
prop: new \stdClass(
a: $a,
b: $b
)
);

You illustrate of useful pattern here.

What would be nicer, however, would be an object instantiation shorthand that would omit the "new" by using braces:

$x = \stdClass{
prop: \stdClass{
a: $a,
b: $b,
},
}

Also:

$x = \Foo{
prop: \Bar{
a: $a,
b: $b,
},
}

And for instantiating anonymous classes:

$x = {
prop: {
a: $a,
b: $b,
},
}

-Mike

3 years ago by Rowan Tommins — view source

unread

On Oct 23, 2020, at 2:33 PM, Rowan Tommins rowan.collins@gmail.com
wrote:

I don't follow; is the resulting JSON different if you cast to
object, or is there some other reason you prefer an object over using
an associative array directly?

The by-reference semantics of objects vs arrays in PHP.

shrug If you want things to be mutable between functions, stick an & in front of them; if you want copy-on-write performance, leave it to the engine. That doesn't seem to have anything to do with JSON or API as a use case.

That pattern can have a non-insignificant performance penalty when
dealing with a large number of objects, a use-case that is not
infrequent when processing JSON, especially responses returned via an
HTTP API.

That gives us all the more reason to look for efficient ways of populating anonymous class instances, then. (And, indeed, populating actual named classes, if such a penalty applies to them too.)

My point was that if we didn't already have it, making (object)$array create an instance of an anonymous class would feel natural and not lose anything that we have now.

I see stdClass as a badly named and no longer necessary kludge, and would rather see features that help replace it than features that rely on it.

Regards,

--
Rowan Tommins
[IMSoP]

3 years ago by Andreas Bittner — view source

unread

I don't follow; is the resulting JSON different if you cast to
object, or is there some other reason you prefer an object over using
an associative array directly?

The by-reference semantics of objects vs arrays in PHP.

shrug If you want things to be mutable between functions, stick an & in front of them; if you want copy-on-write performance, leave it to the engine. That doesn't seem to have anything to do with JSON or API as a use case.

Agreed. My original point might have been badly phrased and was only made to reinforce that
\stdClass is currently needed and widely used as no other structure exists that offers dynamic
property names while still being forcefully JSON encoded as a JSON object.
We could replace this with another class. Maybe a real map implementation. But without the
replacement being present \stdClass should be removed.

With regards to the JSON looking different when casting to object:

json_encode([1,2,3]); // result: [1,2,3]
json_encode((object)[1,2,3]); // result: {"0":1,"1":2,"2":3}

json_encode([]); // result: []
json_encode((object)[]); // result: {}

json_encode(['key' => 'value']); // result: {"key":"value"}
json_encode((object)['key' => 'value']); // result: {"key":"value"}

While it only applies to empty and sequential arrays I think it best to hand the desired structure
(object or array) from PHP to json_encode and not leave it to interpretation. I usually provide
json_encode with an object where I want to have an object result, even if I know that the source is
already an array with string keys.

That pattern can have a non-insignificant performance penalty when
dealing with a large number of objects, a use-case that is not
infrequent when processing JSON, especially responses returned via an
HTTP API.

That gives us all the more reason to look for efficient ways of populating anonymous class instances, then. (And, indeed, populating actual named classes, if such a penalty applies to them too.)

I am not familiar with internals, but I can't imagine that omitting new could have any performance
impact at all, as long as the constructor would still be called for non \stdClass objects.

For \stdClass objects it might be possible to find a performant way of handling new \stdClass(...)
with close to (if not the same) performance as a proposed {prop: 'value'}.

I'm not clear on the advantages of the new syntax from a language perspective.

Would \Bar{a: $a, b: $b} still call the constructor?
If yes, it would be a cosmetic change only with negligible (if any) performance impact compared to
new \Bar with named parameters.
If no and the arguments would be written directly to the public properties of the object, I would
consider it a bad pattern, given that internals about the setup of the valid object might be unclear
to the outside user. What if the class has additional private properties? The control over valid
construction of an object must be located in the class defining the object imho.

I see stdClass as a badly named and no longer necessary kludge, and would rather see features that help replace it than features that rely on it.

Do we already have a comprehensive list of features and use cases for \stdClass? If not I would
offer to compile one to the best of my knowledge for a separate discussion on how to replace
\stdClass if possible and needed.

3 years ago by Rowan Tommins — view source

unread

With regards to the JSON looking different when casting to object:

json_encode([1,2,3]); // result: [1,2,3]
json_encode((object)[1,2,3]); // result: {"0":1,"1":2,"2":3}

json_encode([]); // result: []
json_encode((object)[]); // result: {}

Thanks, I hadn't thought of that. For cases as trivial as that, we have
JSON_FORCE_OBJECT, but that's no use for creating structures like [{}, {}]

\stdClass is currently needed and widely used as no other structure
exists that offers dynamic property names while still being forcefully
JSON encoded as a JSON object.
We could replace this with another class. Maybe a real map
implementation. But without the replacement being present \stdClass
should be removed.

Just to be clear, any object (except for a few with magic behaviour)
can have dynamic property names added. As far as I know, defining "class
stdClass {}" would be completely equivalent to what stdClass currently
does (i.e. nothing).

The only thing that makes stdClass at all special is the built-in
functionality that returns it, notably the "object cast" short-hand for
creating an instance and dynamically setting properties on it. The
functionality (if not necessarily the performance) of that
short-hand can definitely be replicated with any normal or anonymous
class: https://3v4l.org/fGRBj

Do we already have a comprehensive list of features and use cases for
\stdClass? If not I would offer to compile one to the best of my
knowledge for a separate discussion on how to replace \stdClass if
possible and needed.

I guess you could search the manual for built-in functions and operators
that return stdClass instances. I doubt there are any that explicitly
require them as input.

For userland code, I'm not aware of any scenario where "new stdClass"
could not be replaced by "new class {}", or why a polyfill would be any
more than "if ( ! class_exists('stdClass') ) { class stdClass{} )".

Regards,

--
Rowan Tommins (né Collins)
[IMSoP]