Regression on path dependent types from 3.2

[AnnelineD]

Compiler version

Works on 3.2.2 and fails on 3.3.2 and 3.5.0.

Minimized code

object PatchG:
   inline def apply[EL, MG <: DNIELMG[EL]](inline mg: MG)(inline ptps: List[(mg.Node | C.type, mg.Node | C.type)]): PatchG[EL, MG]

https://github.com/Adam-Vandervorst/LLGraph/blob/10f010e536beb1c8fb5f5a8bf80f7c308dc72768/src/main/scala/be.adamv.llgraph/graphs/PatchG.scala#L112

Output

[error] -- [E083] Type Error: /home/anneline/IdeaProjects/LLGraph/src/main/scala/be.adamv.llgraph/graphs/PatchG.scala:112:76 
[error] 112 |  inline def apply[EL, MG <: DNIELMG[EL]](inline mg: MG)(inline ptps: List[(mg.Node | C.type, mg.Node | C.type)]): PatchG[EL, MG] =
[error]     |                                                                            ^^
[error]     |(mg : MG) is not a valid type prefix, since it is not an immutable path
[error]     |---------------------------------------------------------------------------
[error]     | Explanation (enabled by `-explain`)
[error]     |- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
[error]     | An immutable path is
[error]     |  - a reference to an immutable value, or
[error]     |  - a reference to `this`, or
[error]     |  - a selection of an immutable path with an immutable value.
[error]      ---------------------------------------------------------------------------

Expectation

To compile, since the inlined version does type-check.
Example usage:

  val rhs = DNIELMG[Char]()
  val r = rhs.newNodes(2)
  rhs.connect(r(0), r(0), 'b')

  val rpg = PatchG[Char, rhs.type](rhs)(List(C -> r(0), r(1) -> C))

https://github.com/Adam-Vandervorst/LLGraph/blob/10f010e536beb1c8fb5f5a8bf80f7c308dc72768/src/test/scala/PatchGTest.scala#L24

[AnnelineD]

I guess this relates to #16804 but the answer there is not satisfactory, since dropping the path-dependent-types from this (and its down the line) project seriously hampers its type safety.

What is the workaround? Can this be expressed via implicits somehow since they are unique (the cause identified in @soronpo his issue)?

[Gedochao]

cc @jchyb

[jchyb]

Hi @AnnelineD. Looking at the linked project, it looks to me like it should be enough to just drop the inline modifier from the mg parameter, since its "inlineness" does not seem to be required anywhere:

object PatchG:
   inline def apply[EL, MG <: DNIELMG[EL]](mg: MG)(inline ptps: List[(mg.Node | C.type, mg.Node | C.type)]): PatchG[EL, MG]

This will make the mg into an immutable path, and inline into roughly the same form as in the Expectation field, with the val mg = DNIELMG[Char]() being a separate val (with an immutable path), instead of being directly inlined.

[soronpo]

@Gedochao I'm looking again on the original issue, and it seems that people run into it several more times and spent time to minimize it. I think we should change the problem to a reporting issue. The compiler needs to suggest to remove the inline to solve the problem.

[Gedochao]

@soronpo Agreed. Can you re-raise this as a reporting issue? linking back to #21536 and #16804.
The scope/requirements are different enough to warrant a separate ticket IMO, if we're to address it.
The behaviour is as intended, we just need to improve the error.
This issue is a won'tfix as it's currently defined.

[Gedochao]

Closing this, the reporting improvement should be tracked under #21538