Scala Macros to the Rescue

Tagged as scala
Written on 2018-12-20 21:38:45+01:00

Did you know Scala has macros? Coming from Common Lisp they serve pretty much the same purpose, doing things that the (plethora of) other language features don't support and to shortcut the resulting boilerplate code. And even the S-expressions can be had when macro debugging is turned on, though the pretty-printed Scala code is arguably much more useful here.

Why would you realistically use them then? Turns out I had to deal with some auto-generated code dealing with Protobuf messages. The generated classes for any message look something like this (in Java syntax since that's what the generated code is):

public interface ExampleResponseOrBuilder

public static final class ExampleResponse
  implements ExampleReponseOrBuilder {

  public static Builder newBuilder();

  public static final class Builder
    implements ExampleResponseOrBuilder;

That is, we have one interface, two classes, one of them conveniently gives you a builder for new objects of the class. That's used like this (back to Scala here):

val builder: ExampleResponse.Builder = Example.newBuilder()
val result: ExampleResponse =

If you try and make a generic builder here, you'll quickly notice that this is rather hard as the generic types don't really express the relationship between ExampleResponse and ExampleResponse.Builder well.

As an aside, you want to have a generic builder parametrised on the return type to be able to write something like this:

val result = build[ExampleResponse](stream)

Without ever having to pass through the type as a value. Better even if you just specify the result type and the type parameter for build is then automatically derived.

These builders look something like this then:

trait ProtobufBuilder[T <: Message] {
  def underlying(): Message.Builder

  def build(string: String)(implicit parser: JsonFormat.Parser): T = {
    val builder = underlying()
    parser.merge(string, builder)[T]

class ExampleResponseBuilder() extends ProtobufBuilder[ExampleResponse] {
  override def underlying(): ExampleResponse.Builder =

This then allows us to use some implicit magic to pass these through to the decoder (sttp's framework in this case) to correctly decode the incoming data.

But, we've to 1. write one class for each type, 2. instantiate it. This is roughly five lines of code per type depending on the formatting.

Macros to the rescue!

Inspired by the circe derivation API I finally got all the pieces together to create such a macro:

def deriveProtobufBuilder[T <: Message]: ProtobufBuilder[T] = macro deriveProtobufBuilder_impl[T]

def deriveProtobufBuilder_impl[T <: Message: c.WeakTypeTag](
    c: blackbox.Context): c.Expr[ProtobufBuilder[T]] = {
  import c.universe._

  val messageType   = weakTypeOf[T]
  val companionType = messageType.typeSymbol.companion

    new ProtobufBuilder[$messageType] {
      override def underlying(): $companionType.Builder = $companionType.newBuilder()

Used then like this:

private implicit val exampleResponseBuilder: ProtobufBuilder[ExampleResponse] = deriveProtobufBuilder

That's one or two lines and the types are only mentioned once (the variable name can be changed). Unfortunately getting rid of the variable name doesn't seem to be possible.

Easy, wasn't it? Unfortunately all of this is hampered by the rather undocumented APIs, you really have to search for existing code or Stackoverflow questions to figure this out.

One thing that helped immensly was the -Ymacro-debug-lite option, which prints the expanded macro when used in sbt via compile.


Unless otherwise credited all material Creative Commons License by Olof-Joachim Frahm