Match type widens the resulting type

Compiler version

3.2.0

Minimized code

enum A {
  case AA
  case AB
}
enum B {
  case BA
  case BB
  case BC
}
type P[X <: A] = X match {
  case A.AA.type => B.BA.type | B.BB.type
  case A.AB.type => B.BC.type
}
def p[X <: A](x: X): P[X] = x match {
  case x: A.AA.type => B.BA
  case x: A.AB.type => B.BC
}
p[A.AA.type](A.AA) // widened, BA: B
p[A.AB.type](A.AB) // widened, BC: B

sealed trait C
object CA extends C
object CB extends C
object CC extends C
type P2[X <: A] = X match {
  case A.AA.type => CA.type | CB.type
  case A.AB.type => CC.type
}
def p2[X <: A](x: X): P2[X] = x match {
  case x: A.AA.type => CA
  case x: A.AB.type => CC
}
p2[A.AA.type](A.AA) // widened, CA: C
p2[A.AB.type](A.AB) // not widened, CC: CC

Output

Expectation

I expect them not to be widened.

[som-snytt]

There are many tickets about enums, unions, widening, and inference as an intersection of features.

It would be nice if there were a FAQ page summarizing the common questions. That page could also clarify what is “established behavior” and what might be “avenues of research”. In particular, it could list the use cases and the idioms.

The other new bit is that enums are preferentially widened, as mentioned in the docs.

(I started listing things that could go in the FAQ, after reading some tickets, but that is tedious and dreary to write and to read.)

On this ticket, the “further minimization” seems uninteresting, so I can’t tell if the OP is about the speculative tickets on “more precise types”.

But I sympathize with the need for a summary view.

[dwijnand]

Basically singleton types, except for those on actual objects, get widened to their underlying type.

Tried to find where I’d seen this (while thinking about #16107) and it took me too long. So, for future reference:

https://github.com/lampepfl/dotty/blob/e587a813d45a237a46dcee56c94ffd93cd4adf77/compiler/src/dotty/tools/dotc/core/ConstraintHandling.scala#L674-L675