Static return type of Scala macros

Question

So I've got this macro:

import language.experimental.macros
import scala.reflect.macros.Context

class Foo
class Bar extends Foo { def launchMissiles = "launching" }

object FooExample {
  def foo: Foo = macro foo_impl
  def foo_impl(c: Context): c.Expr[Foo] =
    c.Expr[Foo](c.universe.reify(new Bar).tree)
}

I've said three times that I want foo to return a Foo, and yet I can do the following (in 2.10.0-RC3):

scala> FooExample.foo
res0: Bar = Bar@4118f8dd

scala> res0.launchMissiles
res1: String = launching

The same thing happens if I remove the type parameters on either c.Expr. If I really want to make sure that whoever's calling foo can't see that they're getting a Bar, I have to add a type ascription in the tree itself.

This is actually pretty great—it means for example that I can point a macro at a schema of some sort and create an anonymous subclass of some Vocabulary class with member methods representing terms in the vocabulary, and these will be available on the returned object.

I'd like to understand exactly what I'm doing, though, so I have a couple of questions. First, what is the return type on the foo method actually for? Is it just available for (optional) documentation? It clearly constrains the return type (e.g., I can't change it to Int in this case), and if I remove it entirely I get an error like this:

scala> FooExample.foo
<console>:8: error: type mismatch;
 found   : Bar
 required: Nothing
              FooExample.foo
                         ^

But I can change it to Any and still get a statically typed Bar when I call foo.

Second, is this behavior specified somewhere? This seems like a fairly elementary set of issues, but I haven't been able to search up a clear explanation or discussion.

Answer 1

This behavior is underspecified but intended, though it might appear confusing. We plan to elaborate on the role of return type in macro signatures, but at the moment I feel like the flexibility is a good thing to have.

Also at times the behavior is inconsistent, e.g. when the macro is caught in the middle of type inference, its static signature will be used (i.e. Foo in your example), not the type of the actual expansion. That's because macro expansion is intentionally delayed until type inference is done (so that macro implementations get to see inferred types, not type vars). This is a trade-off and not necessarily the best one, so we're planning to revisit it soon: https://issues.scala-lang.org/browse/SI-6755.

Another problem in this department is with implicit macros. When the return type of an implicit macro is generic and needs to be inferred from the requested type of an implicit value, bad things happen. This makes it currently impossible to use macros to generate type tags: https://issues.scala-lang.org/browse/SI-5923.