Regression in summoning valueOf with intersection type

[eejbyfeldt]

Compiler version

3.8.0-RC1, 3.8.1-RC1-bin-20251117-5ccea40-NIGHTLY

Works in 3.7.4

Minimized code

//> using scala 3.8.0-RC1
import scala.compiletime.{erasedValue, summonFrom}

object test {
  private inline def singletons[T, Elem <: Tuple]: Seq[T] =
    inline erasedValue[Elem] match {
      case _: EmptyTuple => Seq.empty
      case _: (h *: t) =>
        summonFrom {
          case _: ValueOf[`h`] => valueOf[`h` & T] +: singletons[T, t]
        }
    }

  enum A { case A1, A2, A3 }

  val _ = singletons[A, (A.A1.type, A.A2.type, A.A3.type)]
}

Output

[error] ./value_of_inline.scala:15:11
[error] cannot reduce summonFrom with
[error]  patterns :  case given ev @ _:ValueOf[(test.A.A1 : test.A) & test.A]
[error]   val _ = singletons[A, (A.A1.type, A.A2.type, A.A3.type)]
[error]           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

Expectation

Compiles like 3.7.4.

Slightly related to #24237

[hamzaremmal]

I'm actually surprised that valueOf[`h` & T] reduced in 3.7.4.
Even in 3.7.4, val x = valueOf[A.A1.type & A]) will fail.

My opinion is: this was a bug and now it is fixed. What should be done to migrate such a snippet of code is the following:

//> using scala 3.8.0-RC1
import scala.compiletime.{erasedValue, summonFrom}

object test {
  private inline def singletons[T, Elem <: Tuple]: Seq[T] =
    inline erasedValue[Elem] match {
      case _: EmptyTuple => Seq.empty
      case _: (h *: t) => valueOf[`h`].asInstanceOf[T] +: singletons[T, t]
    }

  enum A { case A1, A2, A3 }

  val _ = singletons[A, (A.A1.type, A.A2.type, A.A3.type)]
}

[eejbyfeldt]

If one replace the asInstanceOf with something like

inline def staticCast[T](inline value: Any): T = inline value match { case v: T => v }
...
      case _: (h *: t) => staticCast[T](valueOf[`h`]) +: singletons[T, t]
...

so we do not move any failures from compiletime to runtime then it sounds reasonable to me.

It was never clear to me if it was something that as intended to work or just happened to work under some circumstances. And it seemed kind of finicky under to when that was. So if it never intended to work here that sounds good.

[hamzaremmal]

If one replace the asInstanceOf with something like

inline def staticCast[T](inline value: Any): T = inline value match { case v: T => v }
...
      case _: (h *: t) => staticCast[T](valueOf[`h`]) +: singletons[T, t]
...

so we do not move any failures from compiletime to runtime then it sounds reasonable to me.

It was never clear to me if it was something that as intended to work or just happened to work under some circumstances. And it seemed kind of finicky under to when that was. So if it never intended to work here that sounds good.

I'm sure it definitely was not meant to work but that doesn't mean it couldn't (maybe the specification can change?). Nevertheless, I find it fine as it is.
Also, I like the staticCast idea 💡😄. Something worth adding in the compiletime package.