Serialization and value classes (IR-specific)

This appendix describes how value classes are handled by kotlinx.serialization.

Features described are available on JVM/IR (enabled by default), JS/IR and Native backends.

Table of contents

Serializable value classes

We can mark value class as serializable:

@Serializable
@JvmInline
value class Color(val rgb: Int)

Value class in Kotlin is stored as its underlying type when possible (i.e. no boxing is required). Serialization framework does not impose any additional restrictions and uses the underlying type where possible as well.

@Serializable
data class NamedColor(val color: Color, val name: String)

fun main() {
  println(Json.encodeToString(NamedColor(Color(0), "black")))
}

In this example, NamedColor is serialized as two primitives: color: Int and name: String without an allocation of Color class. When we run the example, encoding data with JSON format, we get the following output:

{"color": 0, "name": "black"}

As we see, Color class is not included during the encoding, only its underlying data. This invariant holds even if the actual value class is allocated — for example, when value class is used as a generic type argument:

@Serializable
class Palette(val colors: List<Color>)

fun main() {
  println(Json.encodeToString(Palette(listOf(Color(0), Color(255), Color(128)))))
}

The snippet produces the following output:

{"colors":[0, 255, 128]}

Unsigned types support (JSON only)

Kotlin standard library provides ready-to-use unsigned arithmetics, leveraging value classes to represent unsigned types: UByte, UShort, UInt and ULong. Json format has built-in support for them: these types are serialized as theirs string representations in unsigned form. These types are handled as regular serializable types by the compiler plugin and can be freely used in serializable classes:

@Serializable
class Counter(val counted: UByte, val description: String)

fun main() {
    val counted = 239.toUByte()
    println(Json.encodeToString(Counter(counted, "tries")))
}

The output is following:

{"counted":239,"description":"tries"}

Unsigned types are currently supported only in JSON format. Other formats such as ProtoBuf and CBOR use built-in serializers that use an underlying signed representation for unsigned types.

Using value classes in your custom serializers

Let's return to our NamedColor example and try to write a custom serializer for it. Normally, as shown in Hand-written composite serializer, we would write the following code in serialize method:

override fun serialize(encoder: Encoder, value: NamedColor) {
  encoder.encodeStructure(descriptor) {
    encodeSerializableElement(descriptor, 0, Color.serializer(), value.color)
    encodeStringElement(descriptor, 1, value.name)
  }
}

However, since Color is used as a type argument in encodeSerializableElement function, value.color will be boxed to Color wrapper before passing it to the function, preventing the value class optimization. To avoid this, we can use special encodeInlineElement function instead. It uses serial descriptor of Color (retrieved from serial descriptor of NamedColor) instead of KSerializer, does not have type parameters and does not accept any values. Instead, it returns Encoder. Using it, we can encode unboxed value:

override fun serialize(encoder: Encoder, value: NamedColor) {
  encoder.encodeStructure(descriptor) {
    encodeInlineElement(descriptor, 0).encodeInt(value.color)
    encodeStringElement(descriptor, 1, value.name)
  }
}

The same principle goes also with CompositeDecoder: it has decodeInlineElement function that returns Decoder.

If your class should be represented as a primitive (as shown in Primitive serializer section), and you cannot use encodeStructure function, there is a complementary function in Encoder called encodeInline. We will use it to show an example how one can represent a class as an unsigned integer.

Let's start with a UID class:

@Serializable(UIDSerializer::class)
class UID(val uid: Int)

uid type is Int, but suppose we want it to be an unsigned integer in JSON. We can start writing the following custom serializer:

object UIDSerializer: KSerializer<UID> {
  override val descriptor = UInt.serializer().descriptor
}

Note that we are using here descriptor from UInt.serializer() — it means that the class' representation looks like a UInt's one.

Then the serialize method:

override fun serialize(encoder: Encoder, value: UID) {
  encoder.encodeInline(descriptor).encodeInt(value.uid)
}

That's where the magic happens — despite we called a regular encodeInt with a uid: Int argument, the output will contain an unsigned int because of the special encoder from encodeInline function. Since JSON format supports unsigned integers, it recognizes theirs descriptors when they're passed into encodeInline and handles consecutive calls as for unsigned integers.

The deserialize method looks symmetrically:

override fun deserialize(decoder: Decoder): UID {
  return UID(decoder.decodeInline(descriptor).decodeInt())
}

Disclaimer: You can also write such a serializer for value class itself (imagine UID being the value class — there's no need to change anything in the serializer). However, do not use anything in custom serializers for value classes besides encodeInline. As we discussed, calls to value class serializer may be optimized and replaced with a encodeInlineElement calls. encodeInline and encodeInlineElement calls with the same descriptor are considered equivalent and can be replaced with each other — formats should return the same Encoder. If you embed custom logic in custom value class serializer, you may get different results depending on whether this serializer was called at all (and this, in turn, depends on whether value class was boxed or not).