With -Yexplicit-nulls
import language.strictEquality
val x: String | Null = null
val check = x == null //error
Should the explicit nulls feature introduce CanEqual[T | Null, Null] and CanEqual[T | Null, T]?
Could be relevant for the discussion, how to generally handle strict equality for unions:
This seems required for these two features to interact in a safe way
Is there a way to allow that, but disallow "null" == null (Since "null" is a String | null) ?
I’ve tried trivial fix with
given [A, B](using CanEqual[A, B]): CanEqual[A | Null, B | Null] = CanEqual.derived
And it seems to work in most obvious cases
Except it allows null == ""
That’s actually useful. Even if things are declared with non-null types they can still be null at runtime because they might not be initialized. So, a test like null == s where s is a String is sometimes necessary and it would be annoying if it was not allowed.
In my humble opinion, any access to an uninitialized field is a bug. It is in the same category as accessing an array out of bounds. This isn’t something that one should need to branch on, and I wouldn’t let any code that does this pass code review.
And regarding the original issue, I don’t think it’s a real problem because while x == null doesn’t compile, x eq null works fine, and it makes more sense too because checking for null is a check for reference equality, not for value equality, which is what == is meant to be used for.
There is only one potential issue here: many users don’t know eq because it’s not something that is needed very often. If this is of real concern, it can be fixed with a simple compiler hack. When the compiler sees an == null check and a suitable CanEqual cannot be found, the error message should suggest trying eq.
Maybe we can add methods isNull and isNotNull in Any.
Not even in an assert(s != null)? In code that you don’t control fully? I agree eq/ne would be an alternative, but we want to make it easy and straightforward to write such asserts. Someone who adds such an assert in a desperate debug session would not appreciate the technicalities of CanEqual here.
Well, assert is a special case because it is a tool specifically made to detect bugs at run-time. In pretty much all other cases, I don’t think one should ever branch on a condition that can only ever be true when there’s a bug in the program. And adding a language feature that is always available but whose only legitimate use is inside an assert statement doesn’t seem reasonable to me, especially given that eq and ne work just fine.
Again, I think a compiler hint to use eq or ne is fine.
We had many issues with language.strictEquality before; unfortunately, we couldn’t find anyone who is actively maintaining this feature and discuss the expected behaviour.
Hence, the current behaviour of explicit nulls is based on rules without strictEquality.
After thinking about this again, I have changed my mind. eq and ne are not suitable for testing against null because those are only available on AnyRef. This makes sense at first glance, but unfortunately there are types in Scala that are not subtypes of AnyRef but can still be null, like Int | Null. That means that neither == nor eq can be used to perform a null check on this type, so another solution is needed.
It was proposed to add CanEqual[T | Null, Null] by default when the explicit-nulls feature is used, but due to contravariance that’s just the same as CanEqual[Any, Null], allowing many meaningless checks.
So I’ve had a different idea. == null and != null are already treated specially by the compiler when explicit-nulls are turned on (flow typing). So let’s just make it a little bit more special: == null and != null don’t require a CanEqual instance; instead, the other side’s type must be a supertype of Null. This neatly allows comparisons against Int | Null while disallowing "null" == null.
What’s not to like?
The issue with this is that expression do not really have “a type” (in Scala), instead they have a tower of gradually more general types (or maybe a latice or something, but tower gives a clearer mental image)
And this tower will always include a nullable type: T | Null
And “the most specific type for this expression” is not necessarily well defined, there are cases where an expression can have multiple valid types that are not subtypes of each other:
Array(1, 1, 1) // both Array[1] and Array[Int] (and many other types)
Even variables are weird:
val a: Int = 1
There are two types that a has that could be said to be most specific:
Inta.type (I think this was a confusing choice of syntax)a.type <:< Int)No, that’s not an issue. We have other cases in the compiler where we check for similar things, notably the strictEqualityPatternMatching feature in Scala 3.8. It checks that the scrutinee’s type must be a supertype of the pattern’s type and the new rule won’t kick in otherwise. By your logic, that shouldn’t be possible because every scrutinee has type Any, among others, meaning it’s always a supertype of the pattern’s type.
There are also other cases, like this one:
def foo[A](a: A)(using Null <:< A) = ()
foo(42) // error: Cannot prove that Null <:< Int
foo(42: Int | Null) // OK
So this kind of subtyping check is possible to do.
Oh my bad then !
My intuition was that foo(42) would have worked, but I guess the compiler does not look ahead
I’ve tried implementing this idea and the actual logic is a single line.
This looks like an easy win to me.