endless compilation when deferred inline + match type are used (LTS only)

[road21]

Compiler version

3.3.5 LTS (3.6.4 works as expected)

Minimized code

//> using scala 3.3.5

import scala.compiletime.*

trait InlineFoldR[Acc[_ <: Tuple]]:
  inline def apply[Elem, T <: Tuple](acc: Acc[T]): Acc[Elem *: T]

object InlineFoldR:
  inline def fold[Acc[_ <: Tuple], T <: Tuple](
      acc: Acc[EmptyTuple],
      f: InlineFoldR[Acc]
  ): Acc[T] =
    inline erasedValue[T] match
      case _: EmptyTuple => acc.asInstanceOf[Acc[T]]
      case _: (h1 *: h2 *: h3 *: h4 *: ts) =>
        f[h1, h2 *: h3 *: h4 *: ts](
          f[h2, h3 *: h4 *: ts](
            f[h3, h4 *: ts](f[h4, ts](fold[Acc, ts](acc, f)))
          )
        ).asInstanceOf[Acc[T]]
      case _: (h *: ts) =>
        f[h, ts](fold[Acc, ts](acc, f)).asInstanceOf[Acc[T]]

type DropUnits[A <: Tuple] <: Tuple = A match
  case hd *: tl =>
    hd match
      case Unit => DropUnits[tl]
      case Any  => hd *: DropUnits[tl]
  case EmptyTuple => EmptyTuple

object DropUnits:
  object DropInlined extends InlineFoldR[[T <: Tuple] =>> T => DropUnits[T]]:
    inline def apply[Elem, T <: Tuple](
        acc: T => DropUnits[T]
    ): (Elem *: T) => DropUnits[Elem *: T] =
      (elem: Elem *: T) =>
        inline erasedValue[Elem & Matchable] match
          case _: Unit => acc(elem.tail)
          case _       => elem.head *: acc(elem.tail)

  inline def drop[A <: Tuple](a: A): DropUnits[A] =
    InlineFoldR
      .fold[[T <: Tuple] =>> T => DropUnits[T], A](
        (_: EmptyTuple) => EmptyTuple,
        DropInlined
      )(a)

val a: (
    Int,
    Int,
    Int,
    Int,
    Int,
    Int,
    Int,
    Int,
    Int,
    Int,
    Int,
    Int,
    Int,
    Int,
    Int,
    Int,
    Int,
    Int,
    Int,
    Int,
    Int,
    Int,
    Int,
    Int,
    Int,
    Int,
    Int,
    Int,
    Unit,
    Int,
    Unit
) =
  (1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
    1, 1, (), 1, ())

println(DropUnits.drop(a))

Output

Endless compilation

Expectation

Compilation succeeded.

Sorry, I don't know how to minimize it. Deferred inline InlineFoldR and match type DropUnits are required to show the bug.

[tgodzik]

Looks like 3.3.7 nightlies also have the bug, it was probably fixed by something that we haven't been able to port. Anyone knows what that could be?

[WojciechMazur]

Bisect points that it was fixed by 9726e1a
Last nightly with timeout: 3.3.2-RC1-bin-20230630-c629090-NIGHTLY
First nightly that compiles 3.3.2-RC1-bin-20230703-0a21ecf-NIGHTLY

I'm not sure what happen so that it still timeouts in 3.3.2 but compiles with last 3.3.2 nightly: 3.3.2-RC1-bin-20230724-ce1ce99-NIGHTLY
We had no nightlies for 3.3.3
I'll try to bisect between last 3.3.2-NIGHTLY (working) and first 3.3.4-NIGHTLY (failing)

[tgodzik]

We could also just backport #18073 ? I don't see why it wasn't backported.

[WojciechMazur]

We could also just backport #18073 ? I don't see why it wasn't backported.

Be aware that this PR has also caused other regression that remains unfixed to this day. So we're trading here infinite compilations vs unfixed typer regressions. In my opinion the infinite compilation is worse, but still it can affect other projects

[tgodzik]

Looks like we can't port the fix as it causes a bunch of regressions that are not fixed yet in Next and we need to be more cautious in LTS. Once they are fixed we could try to port the fix, but it's a larger issue.