Hi Eugene! I was asking users on Scala discord about whether this issue bothers Scala users in practice.

The feedback I got is that basically everyone already has setup to only use optimizer when producing release binary. Hence this issue do not actually impact anyone.

Should we instead not proceed with the PR and instead close the two linked issues sbt/sbt#4125, #537 as won't fix?

Since the change itself is relatively small, I see no reason we should reject a good change. While it's true, that once you start using optimizer many things would break so people have turned from past injuries to turn them off during normal development, but perhaps changes like this would gradually improve the situation.

This PR looks right and is an improvement. However in the vast majority of cases methods are not inlined because they carry the @inline annotation, but due to other inliner heuristics. The official documentation even recommends to avoid using @inline in general (https://docs.scala-lang.org/overviews/compiler-options/optimizer.html#inliner-heuristics-and-inline). So I don't think this PR will help much in practise, but it can be considered a first step?

However in the vast majority of cases methods are not inlined because they carry the @inline annotation, but due to other inliner heuristics.

I had never worked with optimizer, but I think it would be possible for optimizer to provide information for each inlining that took place (caller & callee, along with the class name caller and callee belongs to) and store them in CompilationUnit, and zinc Analyzer phase can then use the information to handle inlining dependency. (e.g. via introducing a new dependency called DependencyByInlining and handle them similarly as DependencyByInheritance.). This would catch inlining dependency due to heuristics.

If that solution is indeed reasonable, we can then decide whether we should implement the solution in a follow up PR.

It's a little confusing that isMacro doesn't mean isMacro any more :-/

The inliner can already produce a log, so adding the right calls to record dependencies should be doable.

➜ sandbox cat A.scala
object A {
  def f = 1
}

object B {
  def g = A.f
}

object C {
  def h = B.g
}
➜ sandbox qsc -opt:inline:'**' A.scala -Vinline _
Inlining into B$.g: inlined A$.f (the callee is a small trivial method). Before: 7 ins, after: 22 ins.
Inlining into C$.h: inlined B$.g (the callee is a forwarder or alias method). Before: 7 ins, after: 40 ins.

But I'm not convinced it's worth the effort. Enabling the optimizer in development is discouraged as it increases compilation times. How about issuing a warning when zinc performs an incremental compilation (not all sources) with the optimizer enabled?