Regression trouble inferring type of java wildcard in 2.12.10

[frosforever]

reproduction steps

Using the dependency: "io.grpc" % "grpc-core" % "1.18.0", the following code fails to compile in 2.12.10 and 2.13.1 but compiles just fine in 2.11.12

import io.grpc.ManagedChannelBuilder

trait Test {
  val grpcChannel = ManagedChannelBuilder
    .forAddress(
      "host",
      1234
    )
    .maxInboundMessageSize(20)
    .usePlaintext()
    .build()
}

problem

Fails with:

 value usePlaintext is not a member of ?0
[error] possible cause: maybe a semicolon is missing before `value usePlaintext'?
[error]     .usePlaintext()
[error]      ^
[error] one error found

expectation

This compiles fine in 2.11.12. It also compiles if Test is an object instead of a trait. From what I can tell, this was also an issue in 2.10 that got fixed in 2.11 but has now come back in 2.12 and 2.13.

See also grpc/grpc-java#2813 (comment) and https://stackoverflow.com/questions/44899041/unable-to-build-grpc-managedchannel-with-scala for historical context.

[nogurenn]

I don't know what ?0, but when I tried separating the last two method calls, the compiler complained:

val grpcChannel = ManagedChannelBuilder.forAddress("host",1234).maxInboundMessageSize(20) 
val x = grpcChannel.usePlaintext().build()

cmd2.sc:3: value usePlaintext is not a member of Object
val x = grpcChannel.usePlaintext().build()
                    ^
Compilation Failed

// Another thing I found is that the returned value of `.forAddress()`
// is properly seen(?) by the compiler
@ val grpcChannel = ManagedChannelBuilder.forAddress("host",1234)
// ...
// calling the involved types by-name to avoid errors
res: () => ManagedChannelBuilder[?0] forSome { type ?0 <: ManagedChannelBuilder[?0] }

// but scala only sees the returned value
// of `.maxInboundMessageSize()` as `[java.lang.]Object` / ?0

So, in summary,

// given this call chain
ManagedChannelBuilder.forAddress("host",1234).maxInboundMessageSize(20).usePlaintext().build()

// scala can see this returned value
.forAddress(): static ManagedChannelBuilder<?>

// but it cannot see the next returned value in the chain
.maxInboundMessageSize(): T

// so for some reason, scala complains with seeing
.maxInboundMessageSize(): Object

// since, by this stage, the compiler thinks it's a java.lang.Object,
// it says usePlaintext() is not a member, which is technically correct

ManagedChannelBuilder javadoc here

[SethTisue]

for convenient testing in the REPL, https://github.com/dwijnand/scala-runners is helpful here, so you can scala --scala-version 2.13.2 -Cio.grpc:grpc-core:1.18.0 (Ammonite also lets you add dependencies on the fly, but when investigating scala/bug tickets, scala-runners' ability to easily run any Scala version, including old nightlies and such, is damn near essential; see com-lihaoyi/Ammonite#1093)

?0 is a sure sign an existential type is involved

with -Xlint (which my scala shell alias enables, since without it you're flying half-blind), .maxInboundMessageSize(20) produces the warning warning: inferred existential type ?0 forSome { type ?0 <: io.grpc.ManagedChannelBuilder[?0] }, which cannot be expressed by wildcards, should be enabled by making the implicit value scala.language.existentials visible, so that confirms that an existential type is involved

and ?0 forSome { type ?0 <: io.grpc.ManagedChannelBuilder[?0] } shows that F-bounded polymorphism is in play here

it's odd (and probably significant) that in the REPL, it compiles without the trait wrapper

it also compiles if you use def instead of val

[SethTisue]

I've added the f-bounds label, which is then helpful for finding related tickets. Perhaps this is a duplicate of #4745?

[SethTisue]

(and btw thx @nogurenn for the triage help!)

[frosforever]

@SethTisue that sounds right. Not sure what version #4745 was against but it's interesting this did work for scala 2.11.12 and has now regressed.

[SethTisue]

it's fixed in Dotty. tested via cs launch io.grpc:grpc-core:1.18.0 dotty-repl

[SethTisue]

Not sure what version #4745 was against but it's interesting this did work for scala 2.11.12 and has now regressed.

Scala 2.11.0 was released in 2014, so #4745 (2011) must be a report against some older version. The example here in this ticket fails on 2.10.7, so it could still be a duplicate.