Fix #14319: Use correct stat context when transforming Block

[noti0na1]

Fix #14319

When transforming a Block, the last expression should have import information from statements above.

[noti0na1]

The akka build fails.

Accoding to tests/neg/3559.scala, it seems they are valid errors? I'm not sure about this.

[som-snytt]

IDK but there are default args in self invocation at the akka failures.

case class WeirdCtor(i: Int, s: Option[String] = None, b: Boolean = true) {
  def this(n: Int) = this(n, b = n > 0)
}

is

    def <init>(n: Int): Unit =
      {
        {
          val b$1: Boolean = n.>(0)
          val s$1: Option[String] @uncheckedVariance = WeirdCtor.$lessinit$greater$default$2
          this(n, s$1, b = b$1)
        }
        ()
      }

[odersky]

I am not sure what this is supposed to achieve. The previous fixes should already do the right thing. Maybe it's just ExtractAPI that needs to be fixed here.

[olhotak]

I am not sure what this is supposed to achieve. The previous fixes should already do the right thing. Maybe it's just ExtractAPI that needs to be fixed here.

A Block consists of stats followed by an expr. Before this PR, MegaPhase was fixed to thread contexts through the stats, but not the final expr. The expr was transformed using the original context from before the stats. The purpose of this PR is to use the context that comes out of the last of the stats as the context for the expr.

[odersky]

Thanks for explaining! Yes, I agree that makes sense. But it needs a different fix that does not append a list of statements on the right. And, one would need to troubleshoot why tests are failing.

[noti0na1]

@odersky I will delete the changes to MegaPhase in this PR. The test should be fixed only by the changes in TreeMapWithPreciseStatContexts.

I will think about how to handle the last expr in MegaPhase at a seperate PR.

[olhotak]

Now LGTM.

[odersky]

This will lead to great tree node churn. Every block will get new statements on every transform, so everything containing a block will be copied, which means you get a complete copy of the AST for a macro transform. I don't think this is acceptable.

[odersky]

I think it is better to have an optional expr argument in transformStats. It's more convoluted but I don't see an alternative.

[noti0na1]

A better fix at #14523