Considering there will be support of fluent PHP configuration (#23834), I'd rather put effort to create closure-based factories, e.g.

<?phpdi\services()    ->instanceof(Bar::class)    ->factory(function () {returnnewBar(newFoo());     })

I tried to bring this feature in the past (#12115), but it was rejected because "It adds a lot of complexity for very rare use cases that you can do another way". However, with fluent PHP configuration in the core it wouldn't be "rare case" anymore I guess.

Even if one prefers YAML/XML formats, he still can create fluent PHP config file with the factories instead of "service factory" which you propose.

@unkind nice.. didnt even thought of such an approach :)

Yet we cannot dump such a closure factory. So im not sure it works out; the current approach mostly leverages existing features.

For the config part is was planning to have someServiceFactoryInterface::configure($id, Definition $def). Keeping everything at the same place.

I like the fact we can create powerful services with a minimal config.

@unkind Closure-based factories are incompatible with the dumping of an optimized container, as we cannot dump the closure (due to a closure being able to reference variables from its defining scope). And dumping the final container to a PHP class is necessary (at least if you care about performance).

Btw, I see a big drawback with this approach: any time your factory references another of its methods to get a dependency, it will create its own instance of the services, which will not be the same than the shared service used by the container. This can lead to hard-to-debug behaviors (overwriting thefoo service will not overwrite the dependency used bybar, as it does not use thefoo service but its own instance with a similar configuration than the originalfoo service).

Btw, I see a big drawback with this approach: any time your factory references another of its methods to get a dependency, it will create its own instance of the services, which will not be the same than the shared service used by the container.

Fair point.

We could set shared to false by default, but it really depends what the user does. I.e. if it mimics sharing/memoization in PHP. We dont know that...

I tend not to care about using non-shared$this->someFoo() internally, while it maybe shared externally ($container->get('some_foo')). IMHO you are responsible to set this right, as needed. But if we can do something here that be nice for sure.

Closure-based factories are incompatible with the dumping of an optimized container, as we cannot dump the closure (due to a closure being able to reference variables from its defining scope).

Don't reference such variables. There is no need in it.

@unkind even then, getting the source code of a closure to be able to replicate it in the dumped source code cannot be done in a clean way (and is impossible if you have multiple closures defined on the same line, as closures only have a start and end line in reflection, like other functions, and they don't have a name to resolve ambiguities when having multiple closures on the line).

even then, getting the source code of a closure to be able to replicate it in the dumped source code cannot be done in a clean way

There are third party tools which do it quite well.

and is impossible if you have multiple closures defined on the same line

I know these restrictions, but they are too artificial and can be detected at container compilation time. Likewise, I can say that Symfony doesn't have full support of YAML, but it seems like it isn't a disaster.

@unkind please open a dedicated RFC if you want to push this idea. On my side, I agree with@stof. We don't have proxy-manager in core because it's already a too particular piece of code. We won't add superclosure or similar for the same reason to me.

Agree with@nicolas-grekas, this is not about proposedsetFactory(function() {}) vs. currentsetFactory(array(Some::class, 'foo')). But this PR basically does that by using class methods as a factory, with an automated config approach as much as possible.

Parsing@return would so help this PR :) or adding more annotations. I must agree it's convenient for config... yet im also fine with a static PHP callback to further process if we go this way.

In fact, this PR made me realize we have a "bug" with autowiring (that we should fix as a feature in 3.4):
if a service is defined as autowired and has a factory, we should autowire the factory method instead of the constructor of the service!
@ro0NL would you like to submit a PR fixing this? It would be a nice prerequisite for this PR also (yet, I don't if this PR will make it - but it's nice to be able to think about the idea with an implementation!)

@nicolas-grekas dont we already do that.. looking atAbstractRecursivePass::getConstructor :S

oh indeed!

Closing after some discussion with@nicolas-grekas

This PR can go many ways, and i want to play with#23834 first actually :)