Disable fatal warnings on 2.12 if cross-building?

[armanbilge]

At least a few Typelevel projects are affected by the changes in scala/scala#9890 (particularly those containing JS facades / Native bindings). Actually that PR improves warnings, but the problem is it is a significant divergence from 2.12. This makes cross-compiling across 2.12/2.13 with fatals enabled on both no longer possible, at least with the default sbt-typelevel-settings.

Some random ideas:

1. Disable fatals on 2.12 and move on.
2. Disable just unused warnings on 2.12.
3. Don't warn for ununused nowarns.
4. Create a @nowarn212, @nowarn213, and @nowarn3 annotations to help navigate this mess.

/cc @satorg @DavidGregory084 my resident scalac option experts :)

[som-snytt]

I think -Wconf is a more likely solution.

It would be nice to have a plugin to automate silencing by version, but you're silencing "parameter ... is never used". Or there are "categories" for each flavor. -Wconf:cat="something".

[satorg]

@armanbilge just to clarify: you mean sbt-typelevel 0.5.x line (which is not released yet), but not the 0.4.x one, don't you?
Because in 0.4 all unused warnings are still disabled for 2.12.

[armanbilge]

Oh they are! I completely forgot about that 😂 now I feel silly.

But ok, we still need to think about this for 0.5.x then.

[som-snytt]

all unused warnings are still disabled for 2.12.

I don't use the dis- word anymore, but it would be nice to activate ALL the warnings (insert meme) all the time.

Probably that will be in the form of a sbt-tpolecat that enforces a "profile" per version. An ideal was that -Xlint enforces a baseline.

But maybe it must have the form of -Wconf per version to dial back false positives.

[satorg]

But ok, we still need to think about this for 0.5.x then.

Oh yeah, I do think about it. But the thing is that if we'd like to switch to scalac-option while maintaining the same level of option preciseness, we get to extend the latter with several improvements like allowing a particular option to be adjusted in several ways before being rendered into an option string.

But maybe it must have the form of -Wconf per version to dial back false positives.

Right... There's a great draft proposal in scalac-option to extend it with -Wconf DSL.
Still may require some refinements to be finally incorporated into the lib.

Then, if we manage to add a possibility to tune ScalacOption instance for a particular version range (regardless of whether it is supported or not), then your suggestion will become feasible.

[som-snytt]

thanks, I'll look up scalac-option as I'm out of the loop. Option precision FTW!

[satorg]

@som-snytt it would be great if you could take a look at typelevel/scalac-options#9
and share your thoughts/suggestions in there. thanks!

[satorg]

@armanbilge is this thread related to your initial concern?

[som-snytt]

I commented on the linked ticket that it should be trivial to use a less-trivial but trivially efficient syntax.

[DavidGregory084]

Hmm this one is interesting - so just to state the actual problem clearly, the issue is that:

• The precision of the unused warning was improved in 2.13.9 so that it ignores trivial RHS
• This is causing @nowarn annotations that were added to silence previous warnings to trigger their own "does not silence any warnings" warning
• We still get unused warnings for the same code on 2.12 unless we keep the @nowarn annotations, which poses problems for cross-building

I guess the only solution is to add -Wunused:-nowarn or -Wconf:cat=unused-nowarn:silent for 2.13 only and keep the @nowarn annotations? Does that sound right @som-snytt?

[som-snytt]

The solution on the mouse ticket was to make the RHS less trivial by adding rhs: Unit.

I started looking at why the check for unused @nowarn is difficult or impossible to suppress selectively. If I figure that out, maybe I can get a solution into 2.13.10.

It seems that having unused @nowarn check in -Xlint:unused is the hassle for cross-building. In the absence of selective -Wconf, I agree that -Wunused:-nowarn is simplest way to turn it off.