Regression in inference of implicit arguments

[WojciechMazur]

Community build regression for https://github.com/rssh/cps-async-connect (maintainer @rssh)
Based on Open Community Build #2434
Part of #15647 regressions tracker

This issue seems not to be related to avoidance changes introduced in 3.1.3 and reverts of that change as it compiles in both 3.1.1 and 3.1.3. It seems that the problem might be caused by the local Context type

Compiler version

Works with 3.1.1
Works with 3.1.3
Fails in 3.2.0-RC1
Fails in 3.2.1-RC1-bin-20220711-6efd92d-NIGHTLY

Minimized code

trait CpsMonad[F[_]]:
  type Context <: CpsMonadContext[F]
type Aux[F[_], C <: CpsMonadContext[F]] = CpsMonad[F] { type Context = C }
trait CpsMonadContext[F[_]]
trait CpsMonadInstanceContext[F[_]] extends CpsMonad[F]:
  type Context = CpsMonadInstanceContextBody[F]
class CpsMonadInstanceContextBody[F[_]](m: CpsMonadInstanceContext[F]) extends CpsMonadContext[F]
class InferAsyncArg[F[_], C <: CpsMonadContext[F]](using val am: Aux[F, C])

sealed abstract class ZManaged[-R, +E, +A]
type RManaged[-R, +A] = ZManaged[R, Throwable, A]

type ForZManaged[R, E] = [X] =>> ZManaged[R, E, X]
given zManagedCpsMonad[R, E]: CpsMonadInstanceContext[ForZManaged[R, E]] = ???

// Usage
def failing[R, E](using
    CpsMonad[ForZManaged[R, E]]
): InferAsyncArg[ForZManaged[R, E], CpsMonadInstanceContextBody[ForZManaged[R, E]]] =
  new InferAsyncArg

def compiling[R, E](using
    CpsMonad[ForZManaged[R, E]]
) =
  new InferAsyncArg[ForZManaged[R, E], CpsMonadInstanceContextBody[ForZManaged[R, E]]]

Output

[error] ./test.scala:23:3: Found:    InferAsyncArg[[X] =>> ZManaged[R, E, X], x$1.Context]
[error] Required: InferAsyncArg[ForZManaged[R, E], CpsMonadInstanceContextBody[ForZManaged[R, E]]]
[error]   new InferAsyncArg
[error]   ^^^^^^^^^^^^^^^^^

Expectation

It should compile

[odersky]

I think the failure to infer implicit issues need a bisect to pinpoint where they stopped working. Otherwise it's hopeless.

[griggt]

First bad commit is fe40c59 from #14840

[odersky]

I think what we observe here is the fuzz test boundary established by the huge number of projects we test. Many changes will trip something up, and often it is not immediately clear what it is. That's why we need minimizations and bisections. I am really thankful for everyone who makes this work.