Is there any difference between these two Kotlin extension functions?
fun Any?.f(o: Any?) = 100
fun <T> T.g(o: T) = 100
Is it possible to rewrite g in such a way that the type of its argument and receiver are forced to be the same?
That is, 10.g(5) and "x".g("y") are OK, but 10.g("y") does not compile.
Edit:
Given this, I guess the answer to my second question is no, uless one adds additional arguments.
I believe this is not possible officially at the time of writing this answer (Kotlin 1.7.20).
However, internally Kotlin compiler supports such case, it allows to change the default behavior and use exact type parameters. This is controlled by the internal #Exact annotation and it is used in many places across the Kotlin stdlib.
With some hacking we can enable this behavior in our own code:
#Suppress("INVISIBLE_REFERENCE", "INVISIBLE_MEMBER")
fun <T> #kotlin.internal.Exact T.g(o: #kotlin.internal.Exact T) = 100
Of course, this is purely a hack and it may stop working in future versions of Kotlin.
Update
Answering your first question on whether there is a difference between using Any and T. Generic functions make the most sense if the type parameter is not only consumed, but also passed somewhere further. For example, if the function returns T or it receives an object that consumes T:
fun main() {
var result = 5.g(7)
}
fun <T> T.g(o: T): T = if (...) this else o
In this case result is of type Int. If we use Any instead of T, result would have to be Any as well.
Related
I tried to implement some type classes from Haskell but confronted the issue that is probably a bug in the Kotlin compiler.
interface Semigroup<Instance> {
infix fun Instance.assocOp(oother: Instance): Instance
}
inline fun <reified T: Semigroup<T>> Iterable<T>.concat() = this.reduce<T, T> { acc: T, t: T -> acc.assocOp(t) }
The error message is "Expected parameter of type T".
IDEA suggests to "Change type from 'T' to 'T'" (does nothing).
I expect acc to belong to the type T mentioned in generics. But because of some reason compiler tries to find some other type T. I tried to
specify the type explicitly/implicitly
build ignoring IDEA message
change used version of Kotlin compiler (I have tried 1.4.20, 1.4.10, 1.3.72).
Nothing worked.
I suppose that writing the function without reduce (manually) may help to deal with it. Also, writing java code doing the same may help to mitigate the problem. But these solutions are only workarounds for the problem. Is the issue my fault or the compiler bug?
The compiler error clearly is not helpful here. However, it is correct that the code should not compile IMO.
You're defining the method assocOp as a member extension function. The extension applies to any type T, but it's a member of the interface Semigroup<T>.
To call that extension, you need both a receiver or type T and a receiver of type Semigroup<T> (acting as a context).
In your case, the type T both plays the role of the generic type parameter and of the Semigroup<T>, but you still need to have 2 "receivers" for your extension, even if both are the same instance.
Maybe try this:
inline fun <reified T : Semigroup<T>> Iterable<T>.concat(): T =
reduce<T, T> { t1: T, t2: T -> with(t1) { t1.assocOp(t2) } }
The with(t1) { ... } provides a context of type Semigroup<T>, while the t1 used in t1.assocOp(t2) acts as the T receiver.
Let's assume that I have two functions which do the same stuff.
First one:
fun doSomething() = someObject.getSomeData()
Second one:
fun doSomething(): SomeData {
return someObject.getSomeData()
}
Are there any technical differences between expression functions and standard function in Kotlin excluding the way how they look?
Is compiled output the same?
Are there any advantages using one instead another?
As #Sơn Phan says, they both compile to exactly the same bytecode.
So the differences are simply about conciseness. The expression form omits the braces and return; it also lets you omit the return type (using type inference as needed). As the question illustrates, the expression form can be shorter — and when all else is equal, shorter tends to be easier to read and understand.
So whether the expression form is appropriate is usually a matter of style rather than correctness. For example, this function could be on one line:
fun String.toPositiveIntegers() = split(",").mapNotNull{ it.toIntOrNull() }.filter{ it >= 0 }
But it's a bit long, and probably better to split it. You could keep the expression form:
fun String.toPositiveIntegers()
= split(",")
.mapNotNull{ it.toIntOrNull() }
.filter{ it >= 0 }
Or use a traditional function form:
fun String.toPositiveIntegers(): List<Int> {
return split(",")
.mapNotNull{ it.toIntOrNull() }
.filter{ it >= 0 }
}
(I tend to prefer the former, but there are arguments both ways.)
Similarly, I rather like using it when the body is a simple lambda, e.g.:
fun createMyObject() = MyObject.apply {
someConfig(someField)
someOtherConfig()
}
…but I expect some folk wouldn't.
One gotcha when using the expression form is the type inference. Generally speaking, in Kotlin it's good to let the compiler figure out the type when it can; but for function return values, that's not always such a good idea. For example:
fun myFun(): String = someProperty.someFunction()
will give a compilation error if the someFunction() is ever changed to return something other than a String — even a nullable String?. However:
fun myFun() = someProperty.someFunction()
…would NOT give a compilation error; it would silently change the function's return type. That can mask bugs, or make them harder to find. (It's not a very common problem, but I've hit it myself.) So you might consider specifying the return type, even though you don't need to, whenever there's a risk of it changing.
One particular case of this is when calling a Java function which doesn't have an annotation specifying its nullability. Kotlin will treat the result as a ‘platform type’ (which means it can't tell whether it's nullable); returning such a platform type is rarely a good idea, and IntelliJ has a warning suggesting that you specify the return type explicitly.
1. Compiled output
Yes the compiled output will be completely the same
2. Advantage
You usually use expression function when the body of a function is only one line of expression to make it a oneliner function. Its advantage mainly about making the code more concise. Imagine instead of all the brackets and return, you only need a = to make things done.
While working on the answer of How does erasure work in Kotlin? I found out some things I did not yet understand, nor did I find any sources why it is that way.
Why is the following not compilable?
fun bar(foo: List<*>) = ""
fun bar(foo: List<*>) = 2
while the following is?
fun bar(foo: List<String>) = ""
fun bar(foo: List<Int>) = 2
For me it even gets more curious, when adding a generic type that isn't even used, i.e. the following compiles too:
fun bar(foo: List<*>) = ""
fun <T> bar(foo: List<*>) = 2 // T isn't even used
As the last one doesn't even use T and as we know, generics are erased at runtime, why does this one work, while the variant without generic type does not?
Within the byte code methods only differing in return type are allowed (already described in the above linked answer).
Any hints, sources and/or references are welcome.
Added this question now also at discuss.kotlinlang.org.
The reason why these functions compile or don't compile is related to Kotlin's overload resolution rules. Kotlin does not use the expected type for resolving overloads, so when you call this function:
val x = bar(listOf(""))
...there is no way for the Kotlin compiler to determine the type, and it does not allow you to disambiguate the call by specifying the type of x explicitly.
In the second case, there is no overload resolution problem because the functions have distinct parameter types, and there is no JVM name conflict problem because the functions have different return types (and thus different erased signatures). Therefore, the code compiles.
I came from the C# background and I know how to implement this in C#, but I'm struggling with Kotlin.
I've got 2 extension functions:
fun <T> Foo<T>.myFunction(func: () -> Unit): Foo<T>
and
fun <T> Foo<T>.myFunction(func: () -> Foo<T>): Foo<T>
Clearly, the return type of func is different in both functions. The first function executes it and returns this, the second executes func and returns the result of func.
But it gives me an error:
"Platform declaration clash: The following declarations have the same JVM signature".
How to implement this correctly in Kotlin?
Your functions have a conflicting signature within the JVM due to type erasure (the internal Function0<T> class being used to represent the function parameters); and you can fix this by giving each of them a JVM specific name. From Kotlin you would still access them by the original name, but from Java or internally another name is actually used. Simply use the #JvmName annotation on the alternative versions:
fun <T> Foo<T>.myFunction(func: () -> Unit): Foo<T>
#JvmName("myfunctionWithFoo")
fun <T> Foo<T>.myFunction(func: () -> Foo<T>): Foo<T>
On the JVM, we have to contend with type erasure. Meaning essentially that the types (T in this case) are thrown away in the compiled bytecode and that required checks are only done at compile time. Given that, you have to look at your function declaration with that in mind.
Kotlin will define your function argument as a Function0 in both cases. Because the types are erased, () -> Unit and () -> Foo<T> both look the same in the bytecode. We can prove this out by decompiling the code you've provided (I renamed one of these myFunction2 to get this to work):
public final class com/ginsberg/KotlinStuffKt {
public final static myFunction(Lcom/ginsberg/Foo;Lkotlin/jvm/functions/Function0;)Lcom/ginsberg/Foo;
public final static myFunction2(Lcom/ginsberg/Foo;Lkotlin/jvm/functions/Function0;)Lcom/ginsberg/Foo;
}
This is what the Kotlin compiler is generating (it does more, but I have removed the non-essential parts from this example). As you can see, our types are gone thanks to type erasure. And if we undo my change (myFunction2 becomes myFunction), there's no way at all to tell these apart. This is what the compiler is complaining about - if you erase the types the JVM can't tell these functions apart.
I want to write an extension function which will be available on any type and accept parameter of the same type or subtype, but not a completely different type.
I tried naive approach but it didn't work:
fun <T> T.f(x: T) {
}
fun main(args: Array<String>) {
"1".f("1") // ok
"1".f(1) // should be error
}
It seems that compiler just uses Any for T. I want T to be fixed to receiver type.
The only way to do it requires telling the compiler what you want.
fun <T> T.f(x: T) {
}
In order to use it, you have to tell Kotlin what you want the type to be.
"1".f<String>("2") // Okay
"1".f(2) // Okay (see voddan's answer for a good explanation)
"1".f<String>(2) // Fails because 2 isn't a String
"1".f<Int>(2) // Fails because "1" isn't an Int
When you call fun <T> T.f(x: T) {} like "1".f(1), the compiler looks for a common super-type of String and Int, which is Any. Then it decides that T is Any, and issues no error. The only way to influence this process is to specify T explicitly: "1".f<String>(1)
Since all the checks are performed by the compiler, the issue has nothing to do with type erasure.
Your issue is like saying "John is 3 years older than Carl, and Carl is 3 years younger than John" ... you still don't know either of their ages without more information. That's the type of evidence you gave the compiler and then you expected it to guess correctly. The only truth you can get from that information is that John is at least 3 years old and Carl is at least 1 day old.
And this type of assumption is just like the compiler finding the common upper bounds of Any. It had two strong literal types to chose from and no ability to vary either. How would it decide if the Int or String is more important, and at the same time you told it that any T with upper bounds of Any? was valid given your type specification. So the safe answer is to see if both literals could meet the criteria of T: Any? and of course they do, they both have ancestors of Any. The compiler met all of your criteria, even if you didn't want it to.
If you had tie-breaking criteria, this would work out differently. For example, if you had a return type of T and a variable of type String receiving the value, then that would influence the decision of Type inference. This for example produces an error:
fun <T: Any> T.f2(x: T): T = x
val something: String = "1".f2(1) // ERROR
Because now the type T is anchored by the "left side" of the expression expecting String without any doubt.
There is also the possibility that this could also be an type inference issue that is not intended, check issues reported in YouTrack or add your own to get a definite answer from the compiler team. I added a feature request as KT-13138 for anchoring a specific type parameter to see how the team responds.
You can fix T to the receiver type by making f an extension property that returns an invokable object:
val <T> T.f: (T) -> Unit
get() = { x -> }
fun main(vararg args: String) {
"1".f("1") // will be OK once KT-10364 is resolved
"1".f(1) // error: The integer literal does not conform to the expected type String
}
Unfortunately "1".f("1") currently causes an error: "Type mismatch: inferred type is String but T was expected". This is a compiler issue. See KT-10364. See also KT-13139. You can vote on and/or watch the issues for updates. Until this is fixed you can still do the following:
"1".f.invoke("1")
/* or */
("1".f)("1")
/* or */
val f = "1".f
f("1")