Backport #18503 (on `-Wconf` ordering) to the `3.3.x` LTS series

[BalmungSan]

I am aware of #20282 (comment)

However, IMHO, the already "wrong" order that Scala 3 had until #18503 was merged should be considered like a bug and fixed.
While I understand the current order doesn't need to be considered "wrong" per se, note that at least IME it is common to have general rules defined at the top of your build definitions and then provide more specific rules on specific modules. Such an approach, at least in sbt, leads to more specific rules being added later than the general ones; meaning they appear at their right. And sincerely, I don't want to mess with my build just to change the order of flags; especially when in future versions the order will be changed again.
I also, want to point out that since the behavior is different from Scala 2, making a cross-build for 2.13.15 & 3.3.4 would be extremely painful because now you need to maintain two different orders.

I do understand the rationale of not changing semantics in a patch release.
But I do wonder if anyone is finding the current semantics useful and preferable over the other one. I wonder, how many projects would break with the change, and how hard would be fixing them. I also wonder how many projects are not using the Scala 3 linting to its full potential because they need this.

---

For the record, we found this problem while trying to use the new linting options in Scala 3: neotypes/neotypes#745
We wanted to make all warnings verbose but silence UnusedNonUnitValue for ScalaTest Assertions, and at the end we simply decided to opt out of the verbosity: neotypes/neotypes@2253fb3

[SethTisue]

note that at the time this was previously considered (on 20282), Scala 2 hadn't yet aligned with the Scala 3.4+ behavior (in 2.13.15), so circumstances are different now, as Luis points out by writing:

making a cross-build for 2.13.15 & 3.3.4 would be extremely painful because now you need to maintain two different orders

I'm rather torn now about what's best. changing behavior in a patch release still feels wrong, but Luis's arguments are strong, also.

[Gedochao]

We will discuss this on the next core meeting.

[SethTisue]

@WojciechMazur has pointed out that a workaround exists for cross-building: never use the comma form -Wconf:x,y, always use separate options -Wconf:x -Wconf:y. if you never use the comma, then all the Scala versions agree on the behavior

[SethTisue]

At core meeting, opinion was initially about evenly divided. Everyone acknowledged that there are strong arguments on both sides.

But once we realized the workaround exists for cross-building, that tipped sentiment against backporting.

[SethTisue]

Based on this, I added an "Advice on cross-building" section to the PR description of scala/scala#10708

[BalmungSan]

has pointed out that a workaround exists for cross-building: never use the comma form -Wconf:x,y, always use separate options -Wconf:x -Wconf:y

@SethTisue that is just plain wrong. We don't use the comma form, we only use the separate options form, that is the one that is broken, the one I reported in my example.

The problem is that in 3.3.x doing -Wconf:x -Wconf:y means that x wins over y, whereas in Scala 2 the opposite is true. That is what #18503 fixed.

[SethTisue]

The problem is that in 3.3.x doing -Wconf:x -Wconf:y means that x wins over y

are you sure? @prolativ did some experiments and I don't think that's what his results showed

Michał, maybe you could attach your code and the table of the results?

[BalmungSan]

@SethTisue maybe sbt is needed to trigger the wrong behavior? If so, then could this be a sbt bug instead? But why would it only affect Scala 3?

I know a full project is far from a minimal reproducer; I can work in one if you all want / need one.
But, take a quick look at this PR on neotypes: neotypes/neotypes#745 the upgraded sbt-tpolecat version enabled UnusedNonUnitValue warning in Scala 3 (as expected and wanted). However, we need to silence it only for ScalaTest Assertions. For Scala 2 we used this: -Wconf:cat=other-pure-statement&msg=Assertion:s, and for Scala 3 we did this: -Wconf:name=UnusedNonUnitValue&msg=Assertion:s. But, that didn't work.

After some investigation, I found that we also made all warnings verbose using: -Wconf:any:verbose, this is part of a lazy val commonSettings = Def.settings(...) at the top of the build. And then, in the tests module, we do .settings(commonSettings) BEFORE the .settings(...) block where we add the silence config.
I found that if I commented / removed that config then the silence would actually work: neotypes/neotypes@2253fb3#diff-5634c415cd8c8504fdb973a3ed092300b43c4b8fc1e184f7249eb29a55511f91R66

AFAIK, all sbt is doing here is concatenating the scalacOptions and thus, -Wconf:name=UnusedNonUnitValue&msg=Assertion:s should appear AFTER -Wconf:any:verbose.