Code to report deprecation warnings is in codegen instead of semantic

[asterite]

The code to report deprecation warnings is in codegen/warnings.cr, which is super strange because warnings is a semantic thing. For example if you compile with --no-codegen you don't get the warnings.

I think this should be refactored. The best place to put this might be in CleanUpVisitor.

I might try to improve this, I just don't want to forget about it.

[bcardiff]

It can be moved around sure, but I would prefer to keep the co-location by feature.

[asterite]

What's "co-location by feature"?

[bcardiff]

That the code that makes a features is located near / in the same file. Instead of laying out the code by namespace. The last refactoring on unions.cr is an example of that.

[RX14]

I've been working on and off on refactoring the warnings code towards adding a crystal tool warnings tool and then minimizing the verbostity of warnings output on crystal build. Part of this is abstracting printing frames from the compile error exception.

Part of this I moved the detection to the semantic phase, not codegen, and since the code is only useful for deprecation warnings, I might rename it all to deprecations instead.

[HertzDevil]

There could definitely be other kinds of warnings, this one from 0.35.1 for example, which were also absent if --no-codegen was given:

abstract class Foo
  abstract def foo : Int32
end

class Bar < Foo
  def foo
  end
end

 6 | def foo
         ^--
Warning: this method overrides Foo#foo() which has an explicit return type of Int32.

Please add an explicit return type (Int32 or a subtype of it) to this method as well.

The above warning will become an error in a future Crystal version.

A total of 1 warnings were found.

Same goes for anything that we might want to disallow in 2.0.