Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Compile-time Extension Interfaces #87

Closed
wants to merge 85 commits into from
Closed
Changes from 25 commits
Commits
Show all changes
85 commits
Select commit Hold shift + click to select a range
13df9dd
Type Classes via natural extensions in Kotlin
raulraja Oct 2, 2017
7494c84
Adapted code examples to new proposed syntax https://github.com/Kotli…
raulraja Oct 2, 2017
a5f9659
Fixed typo
raulraja Oct 3, 2017
de8032b
Included section on overcoming `inline` `reified` limitations as show…
raulraja Oct 4, 2017
47e1348
Adapt proposal examples to new style using `given`
raulraja Oct 9, 2017
eb4f687
replace `extension class` for `extension object` where possible addre…
raulraja Oct 10, 2017
3777fa6
added imports to code examples
raulraja Oct 10, 2017
374e848
Add language changes and instance resolution rules order
raulraja Oct 12, 2017
2f2f27e
fixed wrong type param reference
raulraja Oct 24, 2017
3b568bf
code review comments on `most` vs `immediately`
raulraja Nov 7, 2017
7d10901
Reverted to use `typeclass` and `instance`
raulraja Nov 7, 2017
f8dce43
Added sentence clarifying the Reified example.
raulraja Nov 12, 2017
323a550
`Monoid.empty` is a value
fvasco Feb 19, 2018
a105496
Use IntMonoind in the example
fvasco Feb 27, 2018
505619c
Specify package in 'with'
fvasco Feb 27, 2018
115a483
Remove extra import in example 4
fvasco Feb 27, 2018
8a7dd15
Merge pull request #4 from fvasco/import4
raulraja Apr 9, 2018
8a74e37
Merge pull request #1 from fvasco/empty-value
raulraja Apr 9, 2018
20976f6
Merge branch 'master' into IntMonoind_empty
raulraja Apr 9, 2018
73b5d33
Merge pull request #2 from fvasco/IntMonoind_empty
raulraja Apr 9, 2018
7027bde
Merge pull request #3 from fvasco/intext_package
raulraja Apr 9, 2018
4a98c86
Using extension keyword
fvasco Apr 10, 2018
0496c73
Fix lambda declaration
fvasco Apr 11, 2018
da913fe
Compile resolution rules #4 clarification
fvasco Apr 11, 2018
6a10e93
Merge pull request #5 from fvasco/extension-interface
raulraja Apr 11, 2018
00d6aca
Update type class KEEP
Jun 25, 2018
d47eeef
Merge pull request #6 from ClaireNeveu/master
raulraja Jun 28, 2018
f495563
Update type-classes.md
raulraja Jun 28, 2018
d70f40c
Remove out-of-date section.
Jun 28, 2018
23b5276
Merge pull request #7 from ClaireNeveu/master
raulraja Jun 28, 2018
4671389
anonymous parameter clarification (#8)
fvasco Jun 29, 2018
84d16ce
Update proposal based on the initial implementation (#10)
truizlop Nov 14, 2018
b632229
Fix misuse of encoding (#9)
pakoito Nov 14, 2018
2a8b95a
Linguistic improvements within type-classes.md (#11)
TAGC Nov 15, 2018
3b74240
Fix typo: The intro… allow -> The intro… allows (#12)
LouisCAD Apr 10, 2019
5f7b61c
Updates contributors list.
JorgeCastilloPrz Apr 11, 2019
ce47c2e
Renames keep file.
JorgeCastilloPrz Apr 11, 2019
2417af2
First pass on Summary and Motivation.
JorgeCastilloPrz Apr 11, 2019
34d89c8
Description first pass.
JorgeCastilloPrz Apr 11, 2019
97f772f
Pass over inline reified section.
JorgeCastilloPrz Apr 11, 2019
c6835ea
Pass over composition and chain evidences section.
JorgeCastilloPrz Apr 11, 2019
9b6c6fd
Pass over Language changes section.
JorgeCastilloPrz Apr 11, 2019
f4ba65a
Polish Language changes a bit.
JorgeCastilloPrz Apr 11, 2019
ae5c0a7
Review resolution order.
JorgeCastilloPrz Apr 11, 2019
8c0737a
Remove type constructors section.
JorgeCastilloPrz Apr 11, 2019
eff56ad
Remove reified and inline generics section.
JorgeCastilloPrz Apr 11, 2019
33f55fc
Improve resolution order and add error reporting.
JorgeCastilloPrz Apr 11, 2019
988c8a9
Adds how to try sections.
JorgeCastilloPrz Apr 11, 2019
94a1a99
Polish how to test sections.
JorgeCastilloPrz Apr 11, 2019
bcbd746
Ploishes Orphan instances description a bit.
JorgeCastilloPrz Apr 11, 2019
dfc0b64
Completes first pass.
JorgeCastilloPrz Apr 11, 2019
a4fd7f4
Adds more details to some sections.
JorgeCastilloPrz Apr 11, 2019
a32eaa5
Adds some details to description.
JorgeCastilloPrz Apr 11, 2019
c811a73
Moar polishments.
JorgeCastilloPrz Apr 11, 2019
e6ba5e6
Fixes a typo.
JorgeCastilloPrz Apr 11, 2019
5485db1
Fixes a typo.
JorgeCastilloPrz Apr 11, 2019
ed8561e
Fixes a typo on a snippet.
JorgeCastilloPrz Apr 11, 2019
535c453
Includes horizontal composition in Summary.
JorgeCastilloPrz Apr 12, 2019
fba1db3
Updates repo to use Map property and adds A.save() extension.
JorgeCastilloPrz Apr 12, 2019
c563312
Update proposals/compile-time-dependency-resolution.md
raulraja Apr 12, 2019
21aeaed
swap How to try approaches.
JorgeCastilloPrz Apr 12, 2019
3dd43f5
Merge branch 'jc-keep-rewording' of github.com:47deg/KEEP into jc-kee…
JorgeCastilloPrz Apr 12, 2019
c8bf869
Fixes a typo in one of the snippets where a monoid was wrongly mentio…
JorgeCastilloPrz Apr 12, 2019
7792dd7
Adds proper package to User definition snippet.
JorgeCastilloPrz Apr 12, 2019
67c8f3e
Polishes wording over resolution order.
JorgeCastilloPrz Apr 12, 2019
6ba2577
Drops implicits from comparison.
JorgeCastilloPrz Apr 12, 2019
87d1c43
Fix wrong typing in loadById call sites.
JorgeCastilloPrz Apr 12, 2019
531711d
Rethink wording about named extensions.
JorgeCastilloPrz Apr 12, 2019
bd1416d
Rethink wording about named extensions.
JorgeCastilloPrz Apr 12, 2019
7bed642
fix conflicts.
JorgeCastilloPrz Apr 12, 2019
87eafe6
Remove FP constructs mentions from Error reporting screenshots.
JorgeCastilloPrz Apr 12, 2019
3753301
Refactor keep to reflect proper wording on first resolution scope.
JorgeCastilloPrz Apr 12, 2019
951ea3d
Reflect how syntax becomes available in methods and functions bodies.
JorgeCastilloPrz Apr 12, 2019
297e65f
Adds fun and val extension providers to the TODO list.
JorgeCastilloPrz Apr 15, 2019
071d6b7
Adds both big still open issues to the KEEP.
JorgeCastilloPrz Apr 15, 2019
6ac0d48
Uploads and links fixed Keep87Sample zip project using proper interna…
JorgeCastilloPrz Apr 15, 2019
e564b61
Adds clarification for internal modifier requirement in some scopes.
JorgeCastilloPrz Apr 15, 2019
21003c6
Switches chained to nested wording.
JorgeCastilloPrz Apr 15, 2019
83a1a94
Adds internal flag to required snippet extensions.
JorgeCastilloPrz Apr 15, 2019
abcd547
small lang changes
MaureenElsberry Apr 16, 2019
46e6a34
Merge pull request #14 from 47deg/jc-keep-rewording
JorgeCastilloPrz Apr 17, 2019
e863b25
Move Group data class to GroupRepository (#15)
bloderxd Apr 18, 2019
d488d79
Update compile-time-dependency-resolution.md
bassjacob May 5, 2019
78f5459
Merge pull request #16 from bassjacob/patch-1
JorgeCastilloPrz May 5, 2019
da6a274
Update proposals/compile-time-dependency-resolution.md
raulraja May 31, 2019
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Jump to
Jump to file
Failed to load files.
Diff view
Diff view
296 changes: 296 additions & 0 deletions proposals/type-classes.md
@@ -0,0 +1,296 @@
# Type classes

* **Type**: Design proposal
* **Author**: Raul Raja
* **Status**: New
* **Prototype**: -

## Summary

The goal of this proposal is to enable `type classes` and lightweight `Higher Kinded Types` in Kotlin to enable ad-hoc polymorphism and better extension syntax.
Type classes is the most important feature that Kotlin lacks in order to support a broader range of FP idioms.
Kotlin already has an excellent extension mechanism where this proposal fits nicely. As a side effect `Type classes as extensions` also allows for compile time
dependency injection which will improve the current landscape where trivial applications rely on heavy frameworks based on runtime Dependency Injection.
Furthermore introduction of type classes improves usages for `reified` generic functions with a more robust approach that does not require those to be `inline` or `reified`.

## Motivation

* Support Type class evidence compile time verification.
* Support a broader range of Typed FP patterns.
* Enable multiple extension functions groups for type declarations.
* Enable compile time DI through the use of the Type class pattern.
* Enable better compile reified generics without the need for explicit inlining.
* Enable definition of polymorphic functions whose constrains can be verified at compile time in call sites.

## Description

We propose to use the existing `interface` semantics allowing for generic definition of type classes and their instances with the same style interfaces are defined

```kotlin
extension interface Monoid<A> {
fun A.combine(b: A): A
val empty: A
}
```

The above declaration can serve as target for implementations for any arbitrary data type.
In the implementation below we provide evidence that there is a `Monoid<Int>` instance that enables `combine` and `empty` on `Int`

```kotlin
package intext

extension object IntMonoid : Monoid<Int> {
fun Int.combine(b: Int): Int = this + b
val empty: Int = 0
}
```

```
import intext.IntMonoid

1.combine(2) // 3
IntMonoid.empty() // 0
```

Because of this constrain where we are stating that there is a `Monoid` constrain for a given type `A` we can also encode polymorphic definitions based on those constrains:

```kotlin
import intext.IntMonoid

fun <A> add(a: A, b: A, with Monoid<A>): A = a.combine(b)
add(1, 1) // compiles
add("a", "b") // does not compile: No `String: Monoid` instance defined in scope
```

## Compile Time Dependency Injection

On top of the value this brings to typed FP in Kotlin it also helps in OOP contexts where dependencies can be provided at compile time:

```kotlin
extension interface Context<A> {
fun A.config(): Config
}
```

```kotlin
package prod

extension object ProdContext: Context<Service> {
fun Service.config(): Config = ProdConfig
}
```

```kotlin
package test

extension object TestContext: Context<Service> {
fun Service.config(): Config = TestConfig
}
```

```kotlin
package prod

service.config() // ProdConfig
```

```kotlin
package test

service.config() // TestConfig
```

## Overcoming `inline` + `reified` limitations

Type classes allow us to workaround `inline` `reified` generics and their limitations and express those as type classes instead:

```kotlin
extension interface Reified<A> {
val A.selfClass: KClass<A>
}
```

Now a function that was doing something like:

```kotlin
inline fun <reified A> foo() { .... A::class ... }
```

can be replaced with:

```kotlin
fun <A> fooTC(with Reified<A>): Klass<A> { .... A.selfClass ... }
```

This allows us to obtain generics info without the need to declare the functions `inline` or `reified` overcoming the current limitations of inline reified functions that can't be invoked unless made concrete from non reified contexts.

Not this does not remove the need to use `inline reified` where one tries to instrospect generic type information at runtime with reflection. This particular case is only relevant for those cases where you know the types you want `Reified` ahead of time and you need to access to their class value.

```kotlin
extension class Foo<A> {
val someKlazz = foo<A>() //won't compile because class disallow reified type args.
}
```

```kotlin
extension class Foo<A> {
val someKlazz = fooTC<A>() //works and does not requires to be inside an `inline reified` context.
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Why does this work? I don't see any evidence that Reified<A> is satisfied here.

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Wouldn't this require putting reified on the classes generic to work properly?

}
```

## Composition and chain of evidences

Type class instances and declarations can encode further constrains in their generic args so they can be composed nicely:

```kotlin
package optionext

extension class OptionMonoid<A>(with Monoid<A>): Monoid<Option<A>> {

val empty: Option<A> = None

fun Option.combine(ob: Option<A>): Option<A> =
when (this) {
is Some<A> -> when (ob) {
is Some<A> -> Some(this.value.combine(b.value)) //works because there is evidence of a Monoid<A>
is None -> ob
}
is None -> this
}

}
```

The above instance declares a `Monoid<Option<A>>` as long as there is a `Monoid<A>` in scope.

```kotlin
import optionext.OptionMonoid
import intext.IntMonoid

Option(1).combine(Option(1)) // Option(2)
Option("a").combine(Option("b")) // does not compile. Found `Monoid<Option<A>>` instance providing `combine` but no `Monoid<String>` instance was in scope
```

We believe the above proposed encoding fits nicely with Kotlin's philosophy of extensions and will reduce the boilerplate compared to other langs that also support typeclasses such as Scala where this is done via implicits.

## Typeclasses over type constructors

We recommend if this proposal is accepted that a lightweight version of higher kinds support is included to unveil the true power of typeclasses through the extensions mechanisms

A syntax that would allow for higher kinds in these definitions may look like this:

```kotlin
extension interface FunctionK<F<_>, G<_>> {
fun <A> invoke(fa: F<A>): G<A>
}

extension object Option2List : FunctionK<Option, List> {
fun <A> invoke(fa: Option<A>): List<A> =
fa.fold({ emptyList() }, { listOf(it) })
}
```

Here `F<_>` refers to a type constructor meaning a type that has a hole on it such as `Option`, `List`, etc.
Copy link

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Should F<_> be F<*>?

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I made a comment below, on the main thread, as to why IMHO we should avoid using Scala-like syntax for Higher-Kind Type Parameters.


A use of this declaration in a polymorphic function would look like:

```kotlin
fun <F<_>, A, B> transform(fa: F<A>, f: (A) -> B, with Functor<F>): F<B> = F.map(fa, f)

transform(Option(1), { it + 1 }) // Option(2)
transform("", { it + "b" }) // Does not compile: `String` is not type constructor with shape F<_>
transform(listOf(1), { it + 1 }) // does not compile: No `Functor<List>` instance defined in scope.
```

## Language Changes

- Add `with` to require instances evidences in both function and class/object declarations as demonstrated by previous and below examples:
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

What about properties?

Copy link
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Hi @ilya-g . If you are referring to flag properties as extension providers where in this example would be something like:

//somehere else in code
fun <A> evidence(with A): A = this

//inside a class or package
extension val optionMonoid: Monoid<Option<A>>
    get() = evidence<Monoid<Option<A>>()

//or with special sytnax

extension val optionMonoid: Monoid<Option<A>> with Monoid<Option<A>>
    get() = this

That could work but the issue with properties is that since they don't accept arguments we would need to provide a special way to encode the with constrains in the property itself and not sure how that would translate in java. If we add the constrain to an argument in the getter then it no longer has the getter shape expected by frameworks like java beans etc.

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This confuses me somehow. I can understand the first example, although my first impression is, that this is not how the extension mechanism should be applied. Is it a bad idea to use a local function here instead, and just limit the extension mechanism to functions?

The second example doesn't make sense to me^^ Could you give another example please? Although things are brought into scope, there should be a method call somewhere, shouldn't it?

Copy link
Author

@raulraja raulraja May 23, 2018

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@hannespernpeintner there has been no discussion to supporting extension for properties. I'll be happy to limit to functions for now since object is already supported and can already provide a singleton value of the extension instance.

I realized after your comment the example was wrong, it should have looked like this:

extension val optionMonoid: Monoid<Option<A>> with Monoid<A>
    get() = evidence<Monoid<Option<A>>() //

That should be the correct form since to construct an Monoid<Option<A>> we need first evidence of Monoid<A>

Copy link

@Takhion Takhion May 31, 2018

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@raulraja: ...the issue with properties is that since they don't accept arguments we would need to provide a special way to encode the with constrains in the property itself and not sure how that would translate in java

If you think about it, extension properties with a receiver do accept an implicit parameter:

val SomeClass.someProperty get() = "<$this>"
// equivalent to:
fun someProperty(`this`: SomeClass) = "<$`this`>"

Or even two:

class ParentClass {
    val SomeClass.someProperty get() = "<${this@ParentClass}>|<$this>"
}
// equivalent to:
fun someProperty(thisParenClass: ParentClass, `this`: SomeClass) = "<$thisParentClass>|<$`this`>"

You can also see it "reflected" (haha metajoke) in the difference between the possible types of kotlin.reflect.KProperty:

  • KProperty0<out R>
  • KProperty1<T, out R>
  • KProperty2<D, E, out R>

Copy link
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Thanks @Takhion, then we can probably support that use case too. Thanks for clarifying this.

```kotlin
extension class OptionMonoid<A>(with Monoid<A>) : Monoid<Option<A>> //class position using argument "Monoid"
extension class OptionMonoid<A>(with M: Monoid<A>) : Monoid<Option<A>> //class position using argument "M"

fun <A> add(a: A, b: A, with Monoid<A>): A = a.combine(b) //function position using argument "Monoid"
fun <A> add(a: A, b: A, with M: Monoid<A>): A = a.combine(b) //function position argument "M"
```
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

What is the difference between named and unnamed variants? Are these two options to choose between?

Copy link

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Hi @ilya-g
the initial idea is to name "Monoid" the Monoid argument, now I consider this pretty messy.
"Monoid" is the argument name, the parameter class name or the Monoid typeclass.

I consider better treat is as anonymous declaration, so

fun <A> add(a: A, b: A, with Monoid<A>): A

is same as

fun <A> add(a: A, b: A, with _: Monoid<A>): A

In such case it is possible to use a type class easy, however if you want to reference to it then you have to declare explicilty its name.


## Compiler Changes

- The type checker will declare the below definition as valid since the `with` clause provides evidence that call sites won't be able to compile calls to this function unless a `Monoid<A>` is in scope.
```kotlin
fun <A> add(a: A, b: A, with Monoid<A>): A = a.combine(b) //compiles
```
- The type checker will declare the below definition as invalid since there is no `Monoid<Int>` in scope.
```kotlin
add(1, 2)
```
- The type checker will declare the below definition as valid since there is a `Monoid<Int>` in scope.
```kotlin
import intext.IntMonoid
add(1, 2)
```
- The type checker will declare the below definition as valid since there is a `Monoid<Int>` in scope.
```kotlin
fun addInts(a: Int, b: Int, with Monoid<Int>): Int = add(a, b)
```
- The type checker will declare the below definition as valid since there is a `with` block around the concrete `IntMonoid` in scope.
```kotlin
fun addInts(a: Int, b: Int): Int = with(intext.IntMonoid) { add(a, b) }
```

## Compile resolution rules

When the compiler finds a call site invoking a function that has type class instances constrains declared with `with` as in the example below:

Declaration:
```kotlin
fun <A> add(a: A, b: A, with Monoid<A>): A = a.combine(b)
```
Call site:
```kotlin
extension class AddingInts {
fun addInts(): Int = add(1, 2)
}
```
The compiler may choose the following order for resolving the evidence that a `Monoid<Int>` exists in scope.

1. Look in the most immediate scope for declarations of `with Monoid<Int>` in this case the function `addInts`

This will compile because the responsibility of providing `Monoid<Int>` is passed into the callers of `addInts()`:
```kotlin
extension object AddingInts {
fun addInts(with Monoid<Int>): Int = add(1, 2)
}
```

2. Look in the immediately outher class/interface scope for declarations of `with Monoid<Int>` in this case the class `AddingInts`:
```kotlin
extension class AddingInts(with Monoid<Int>) {
fun addInts(): Int = add(1, 2)
}
```
This will compile because the responsibility of providing `Monoid<Int>` is passed unto the callers of `AddingInts()`

3. Look in the import declarations for an explicitly imported instance that satisfies the constrain `Monoid<Int>`:
```kotlin
import intext.IntMonoid
extension object AddingInts {
fun addInts(): Int = add(1, 2)
}
```
This will compile because the responsibility of providing `Monoid<Int>` is satisfied by `import intext.IntMonoid`

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

So, no *-imports? People may not be happy with it.

I'd suspect that we'll need to think in advance of a strategy for code completion for such functions: where does the IDE look for instances and what does it import if needed?

Copy link
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I would like * imports. The only issue with * imports is that if the compiler found ambiguous instances it should bail with a proper message. It was suggested at some point that those so be explicit so that newbies were not confused as to where the instances where getting applied from but if we want to support import kategory.* for example to bring all of our instance into scope that would be awesome. The IDE is doing something similar in the Scala plugin in the case of implicits. It looks in all symbols imported in a given scope trying to find candidates for resolution and it if it finds one it activates the syntax.

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

*-imports will likely pose a performance challenge for the compiler

Copy link
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Something worth mentioning is that instances are only resolved at call sites where the invocation is concrete so the compiler does not need to look into all * where they are declared just where functions are invoked. I believe is the same way resolution now works to bring other symbols into scope without fully qualifying them with the full package name.

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

just where functions are invoked.

It's very-very many places in the code :)

I believe is the same way resolution now works to bring other symbols into scope without fully qualifying them with the full package name.

Not quite. Other symbols are bound by name, here we are binding by type, and it's a lot more work for the compiler

Copy link

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

What about cases where you want to import two implementations of the Monoid<Int> into the scope of one file?

Do you have to resolve this conflict by calling each scoped by a with block??

import intext.IntMonoid1
import intext.IntMonoid2

fun addInts1(a: Int, b: Int): Int = with(IntMonoid1) { add(a, b) }
fun addInts2(a: Int, b: Int): Int = with(IntMonoid2) { add(a, b) }

Edit:
NVM: This is defined below

4. Fail to compile if neither outer scopes nor explicit imports fail to provide evidence that there is a constrain satisfied by an instance in scope.
```kotlin
extension object AddingInts {
fun addInts(): Int = add(1, 2)
}
```
Fails to compile lacking evidence that you can invoke `add(1,2)` since `add` is a polymorphic function that requires a `Monoid<Int>` inferred by `1` and `2` being of type `Int`.
In such case the extension instance have to be explicitly defined.
```kotlin
extension object AddingInts {
fun addInts(): Int = with(intext.IntMonoid) { add(1, 2) }
}
```

Some of these examples where originally proposed by Roman Elizarov and the Kategory contributors where these features where originally discussed https://github.com/Kotlin/KEEP/pull/87