Regression: "a match type could not be fully reduced" when flattening a list of tuples

[hegetim]

Compiler version

Affected: 3.5.0 and newer (tested on Scastie up to 3.6.1)
Latest working version: 3.4.3

Minimized code

case class MyNum(i: Int)

val l = List((MyNum(1), MyNum(1)))

l.flatMap(_.toList)

https://scastie.scala-lang.org/ZXMSKeMdT9SdHzvHdxVaQw

Output

Does not compile on 3.5.0 and newer:

    Found:    List[Tuple.Union[(_$1 : (Playground.MyNum, Playground.MyNum))]]
    Required: IterableOnce[Playground.MyNum]

    Note: a match type could not be fully reduced:

      trying to reduce  Tuple.Union[(_$1 : (Playground.MyNum, Playground.MyNum))]
      trying to reduce  Tuple.Fold[(_$1 : (Playground.MyNum, Playground.MyNum)), Nothing,
      [x, y] =>> x | y]
      failed since selector (_$1 : (Playground.MyNum, Playground.MyNum))
      does not uniquely determine parameters h, t in
        case h *: t => h | Tuple.Fold[t, Nothing, [x, y] =>> x | y]
      The computed bounds for the parameters are:
        h <: Playground.MyNum
        t <: Playground.MyNum *: EmptyTuple.type

Expectation

Should compile (as it does on 3.4.3) and give a flat list of MyNums.

[Gedochao]

Last good release: 3.5.0-RC1-bin-20240403-ece87c3-NIGHTLY
First bad release: 3.5.0-RC1-bin-20240404-a7f00e2-NIGHTLY
Bisect points at 0bf43b2

cc @sjrd

[sjrd]

Given the bisect commit, I suspect that the new error is legitimate. It fixed an unsoundness issue (#19746), so it may very well prevent compilation of code that worked more-or-less by chance before.

[hegetim]

Not sure if I get this right.

Is my usage of Tuple's toList function unsound? Or does the current implementation of toList just incorrectly rely on unsound behavior that has been fixed in 3.5.0?

The signature of toList is

inline def toList: List[Union[this.type]]

In its current state toList is quite limited. Any chance to get that improved?

[Gedochao]

@hegetim I think @sjrd meant that #19746 was an unsoundness issue (check the soundness label), and the fix for that caused this regression (that's where the bisect script points). We need to fix it separately.