Fix restriction of valid type vs free vars

[bew]

Fixes #6446

---

def bar(a : T.class) forall T
  puts "Hello from T"
end
  
def bar(a : Int32.class)
  puts "Hello from Int32"
end

# Before:
bar(Int32) # "Hello from T"

# After
bar(Int32) # "Hello from Int32"

Compiler hacking is fun 🎉

[asterite]

I'll review it tomorrow.

[asterite]

@bew Another great PR! Thank you!

[bew]

Thanks for the review 😃

Just a note, since I currently can't check from the restriction method if the type is really a free var, this happens:

https://carc.in/#/r/6hcr

# No free vars here, and Foo doesn't exist
def bar(a : Foo.class)
  puts "Hello from Foo"
end
  
def bar(a : Int32.class)
  puts "Hello from Int32"
end

# Before:
bar(Int32) # Error: undefined constant Foo

# After
bar(Int32) # "Hello from Int32"

[asterite]

@bew Yes, I'm not sure. I think it's bad that if you don't invoke the method you don't get the "undefined constant Foo" either way. We should probably have a pass to compute those types before analyzing calls, and then in the restriction you would know whether a type is a free var or not. For now maybe this is fine.

[bew]

I think the new behavior is fine.

I'm not sure how you could fix that, since you could have:

def bar(a : Foo.class)
  puts "Hello from Foo"
end
  
def bar(a : Int32.class) # restriction happens here when this method is processed
  puts "Hello from Int32"
end

record Foo # define Foo here!

bar(Foo) # => "Hello from Foo"

But maybe if there's first a pass to find all types, then a pass to process al methods that could do it, but there will be other problems I think, like macros that creates types / create methods / use @type.methods etc..

[asterite]

@bew Macros run on the very first pass. This second pass would check the types of all method arguments to see if they exist (they must either be in scope or declared as free variables). I'm almost sure it's something doable. But it's tricky because methods are also added on the first pass (though they are not type-checked there, and this is why it's tricky to compute which one is stricter than the other). We should probably also have some context of is_restriction? to know whether something is a free var.

But let's merge this for now.

[bew]

We should probably also have some context of is_restriction? to know whether something is a free var.

Yes that's what I was thinking, for example the owner param could be a ctx param instead, with an owner field, but also a def field representing the current def is any.

[bew]

Actually I've found a much simpler solution while playing with #7543:

class Metaclass
  def restriction_of?(other : Metaclass, ctx)
    self.name.restriction_of?(other.name, ctx)
  end
end

---

Done in #7580