Improving named-argument support

Previous Topic Next Topic
 
classic Classic list List threaded Threaded
7 messages Options
Reply | Threaded
Open this post in threaded view
|

Improving named-argument support

paulk_asert

We have a number of open issues around improving named-argument support.
I am currently doing a spike around one AST transformation and one metadata
annotation for the type checker. These relate to various issues and pull requests
that already contain some history:


My current spike borrows code from above with some slight renaming.
It currently has a @NamedVariant AST transform targetted for 2.5 if I get it finished soon.
This can be placed on any method or constructor and produces a variant
with a Map as the first parameter that complies with Groovy's existing
named-argument approach. The method looks for any parameter annotated
with @NamedParam or @NamedDelegate. These are conceptually "pulled out"
of the generated method's signature and passed along via the first argument map.
Any parameter annotated with @NamedParam is assumed to correspond to a single key
for the map. For any parameter annotated with @NamedDelegate, the property names
of the type of that argument are assumed to be the keys.

Examples include:

@NamedVariant
def foo(a, @NamedParam String b, c, @NamedParam(required=true) d) {
  println "$a $b $c $d"
}

which produces a method like this:

def foo(Map _it, a, c) {
  assert _it.containsKey('d')
  this(a, _it.b, c, _it.d)
}

If we have this class:

class Animal {
    String type
    String name
}

Then this definition:

@NamedVariant
def describe(@NamedDelegate Animal animal) {
    "${animal.type.toUpperCase()}:$animal.name"
}

produces an additional method like this:

def describe(Map __namedArgs) {
    this.describe((( __namedArgs ) as Animal))
}

which could be called like this:

assert describe(type: 'Dog', name: 'Rover') == 'DOG:Rover'

Currently if no @NamedXXX annotations are found, @NamedDelegate
is assumed for the first argument (so we could have left it out in the example above).
Currently any number of @NamedParam and @NamedDelegate annotations
can be used but keys can't be duplicated.

A more elaborate example is as follows:

// another domain class
class Color {
    Integer r, g, b
}

@NamedVariant
def describe(@NamedDelegate Animal animal,
             @NamedDelegate Color color,
             @NamedParam('dob') Date born) { ... }

produces (approximately since I am ignoring missing keys):

def describe(Map __namedArgs) {
    Map __colorArgs = [:]
    __colorArgs.r = __namedArgs.r
    __colorArgs.g = __namedArgs.g
    __colorArgs.b = __namedArgs.b
    Map __animalArgs = [:]
    __animalArgs.type = __namedArgs.type
    __animalArgs.name = __namedArgs.name
    this.describe(__animalArgs  as Animal, __colorArgs as Color, __namedArgs.dob)
}

This has little impact on type checking since normal type checking will occur on
the non-Map variant which is still retained. We can do a tiny bit more if we keep
the annotation around and check the required flag but that is a minor discussion
point. We could also generate some additional metadata annotation(s) as per
subsequent discussion if needed but I'll defer that discussion for now.

Over and above the @NamedVariant annotation, I am also suggesting a
way to improve type checking on Map-based methods where the @NamedVariant
might not be possible, e.g. Java classes or existing classes which can't easily
be changed. It would look like the following (again heavily based on above references):

@NamedParams(typeHint = SqlTypeHelper)
Sql newInstance(Map<String, Object> args) { ... }

or

@NamedParams([
  @NamedParam("password"),
  @NamedParam(value = "user", type = "String", required = true)])
Sql newInstance(Map<String, Object> args) { ... }

In the first case the property names and types from the typeHint class are used
for improved type checking. In the second case, the information is embedded in
the nested annotations.

Any thoughts or early comments while I continue on the spike(s)?
More detailed comments can wait for the spike(s) so long as this doesn't
seem way off what people would like to see. It's currently one spike
but I'll probably split into two before making the PRs.

Cheers, Paul.

Reply | Threaded
Open this post in threaded view
|

Re: Improving named-argument support

daniil.ovchinnikov
> @NamedVariant
> def foo(a, @NamedParam String b, c, @NamedParam(required=true) d) {
>   println "$a $b $c $d"
> }
>
> which produces a method like this:
>
> def foo(Map _it, a, c) {
>   assert _it.containsKey('d')
>   this(a, _it.b, c, _it.d)
> }

Do you mean there will be additional method while original method will be left as is?

In case of named parameter with default value, there will be tree methods:

def foo(a, @NamedParam b = 42, @NamedParam c) { a+b+c }  // source method

def foo(Map _it, a) { this(a, _it.b, _it.c) } // generated from @NamedParam
def foo(a, c) { this(a, 42, c) }              // generated from default value
def foo(a, b, c) { a+b+c }                    // actual method

Will default be passed if I invoke code with foo(“value for a”, c: 13) ?

How ‘required’ is supposed to work with default parameters?
Why not leverage default params instead of introducing ‘required’? I.e. all named params are required by default.



Daniil Ovchinnikov
Software Developer
JetBrains
jetbrains.com
“Drive to develop”



Reply | Threaded
Open this post in threaded view
|

Re: Improving named-argument support

Nathan Harvey
In reply to this post by paulk_asert
Paul, I am very much in favor of this idea, but I do not like the execution.
The need for those annotations makes it quite verbose and it seems a bit too
complex. I agree with Daniil that having the "required" flag is also
unnecessary. @NamedParam seems like it should be assumed. I understand the
use case of @NamedDelegate but again, it adds complexity. Ideally you would
only need one annotation to the method for a sensible default case.



--
Sent from: http://groovy.329449.n5.nabble.com/Groovy-Users-f329450.html
MG
Reply | Threaded
Open this post in threaded view
|

Re: Improving named-argument support

MG
In reply to this post by paulk_asert
I have created a Jira issue about Groovy non-map based named parameter support:
https://issues.apache.org/jira/browse/GROOVY-8451

On 16.01.2018 06:54, Paul King wrote:

We have a number of open issues around improving named-argument support.
I am currently doing a spike around one AST transformation and one metadata
annotation for the type checker. These relate to various issues and pull requests
that already contain some history:


My current spike borrows code from above with some slight renaming.
It currently has a @NamedVariant AST transform targetted for 2.5 if I get it finished soon.
This can be placed on any method or constructor and produces a variant
with a Map as the first parameter that complies with Groovy's existing
named-argument approach. The method looks for any parameter annotated
with @NamedParam or @NamedDelegate. These are conceptually "pulled out"
of the generated method's signature and passed along via the first argument map.
Any parameter annotated with @NamedParam is assumed to correspond to a single key
for the map. For any parameter annotated with @NamedDelegate, the property names
of the type of that argument are assumed to be the keys.

Examples include:

@NamedVariant
def foo(a, @NamedParam String b, c, @NamedParam(required=true) d) {
  println "$a $b $c $d"
}

which produces a method like this:

def foo(Map _it, a, c) {
  assert _it.containsKey('d')
  this(a, _it.b, c, _it.d)
}

If we have this class:

class Animal {
    String type
    String name
}

Then this definition:

@NamedVariant
def describe(@NamedDelegate Animal animal) {
    "${animal.type.toUpperCase()}:$animal.name"
}

produces an additional method like this:

def describe(Map __namedArgs) {
    this.describe((( __namedArgs ) as Animal))
}

which could be called like this:

assert describe(type: 'Dog', name: 'Rover') == 'DOG:Rover'

Currently if no @NamedXXX annotations are found, @NamedDelegate
is assumed for the first argument (so we could have left it out in the example above).
Currently any number of @NamedParam and @NamedDelegate annotations
can be used but keys can't be duplicated.

A more elaborate example is as follows:

// another domain class
class Color {
    Integer r, g, b
}

@NamedVariant
def describe(@NamedDelegate Animal animal,
             @NamedDelegate Color color,
             @NamedParam('dob') Date born) { ... }

produces (approximately since I am ignoring missing keys):

def describe(Map __namedArgs) {
    Map __colorArgs = [:]
    __colorArgs.r = __namedArgs.r
    __colorArgs.g = __namedArgs.g
    __colorArgs.b = __namedArgs.b
    Map __animalArgs = [:]
    __animalArgs.type = __namedArgs.type
    __animalArgs.name = __namedArgs.name
    this.describe(__animalArgs  as Animal, __colorArgs as Color, __namedArgs.dob)
}

This has little impact on type checking since normal type checking will occur on
the non-Map variant which is still retained. We can do a tiny bit more if we keep
the annotation around and check the required flag but that is a minor discussion
point. We could also generate some additional metadata annotation(s) as per
subsequent discussion if needed but I'll defer that discussion for now.

Over and above the @NamedVariant annotation, I am also suggesting a
way to improve type checking on Map-based methods where the @NamedVariant
might not be possible, e.g. Java classes or existing classes which can't easily
be changed. It would look like the following (again heavily based on above references):

@NamedParams(typeHint = SqlTypeHelper)
Sql newInstance(Map<String, Object> args) { ... }

or

@NamedParams([
  @NamedParam("password"),
  @NamedParam(value = "user", type = "String", required = true)])
Sql newInstance(Map<String, Object> args) { ... }

In the first case the property names and types from the typeHint class are used
for improved type checking. In the second case, the information is embedded in
the nested annotations.

Any thoughts or early comments while I continue on the spike(s)?
More detailed comments can wait for the spike(s) so long as this doesn't
seem way off what people would like to see. It's currently one spike
but I'll probably split into two before making the PRs.

Cheers, Paul.


Reply | Threaded
Open this post in threaded view
|

Re: Improving named-argument support

Nathan Harvey
MG's approach is the one I favor, although I am not sure I like using = in
the calling syntax; the colon makes more sense to me. Small detail, though.



--
Sent from: http://groovy.329449.n5.nabble.com/Groovy-Users-f329450.html
Reply | Threaded
Open this post in threaded view
|

Re: Improving named-argument support

paulk_asert
In reply to this post by Nathan Harvey

Response below.

On Wed, Jan 17, 2018 at 7:44 AM, Nathan Harvey <[hidden email]> wrote:
Paul, I am very much in favor of this idea, but I do not like the execution.
The need for those annotations makes it quite verbose and it seems a bit too
complex. I agree with Daniil that having the "required" flag is also
unnecessary. @NamedParam seems like it should be assumed. I understand the
use case of @NamedDelegate but again, it adds complexity. Ideally you would
only need one annotation to the method for a sensible default case.

 I imagine a very common use case would be to just have the single @NamedVariant
annotation on the method/constructor and nothing else. The other annotations are
just used to cover the more complex use cases that have been spoken about in
various discussions. If you don't need them, don't use them. There are many examples
even in the Groovy codebase where auto `required` wouldn't be appropriate, e.g. the
Sql.newIstance case. There are examples of using that with various combinations:
(url, properties), (url, driverClassName), (url, username, password, driver), etc.
There are other examples where required would make sense.

Reply | Threaded
Open this post in threaded view
|

Re: Improving named-argument support

paulk_asert
In reply to this post by Nathan Harvey
Yes, we have spoken about providing named arguments as a language feature before. It doesn't rule out the need for other options though. I have put some comments in the issue.

On Thu, Jan 18, 2018 at 7:37 AM, Nathan Harvey <[hidden email]> wrote:
MG's approach is the one I favor, although I am not sure I like using = in
the calling syntax; the colon makes more sense to me. Small detail, though.