Interpreter: fix 12351

[asterite]

Fixes #12351

[asterite]

It would be great to merge #12355 so that we can easily see if this caused any regressions :-)

[straight-shoota]

Just wondering, could we skip this cast if arg type and target type are identical?

[asterite]

That's true!

I'm now thinking that we only need the cast in the last if, where there's no restriction on the argument. Because otherwise the arg's type is always checked. Let me try it...

[asterite]

Done!

[asterite]

To expand a bit more on this: a multidispatch will generate code like this:

def multidispatch(arg)
  if arg.is_a?(String)
    foo(arg)
  elsif arg.is_a?(Int32)
    foo(arg)
  else
    foo(arg)
  end
end

In this case if arg's type is String | Int32 | Nil the last branch will have arg's type to be Nil. But, the method corresponding to foo might either take Nil or maybe String | Nil, or even String | Int32 | Nil depending on how the overloads are defined.

To make the cast conditionally on the last branch we'd have to know what the type of arg is there: removing any types that were already checked in .is_a?. Given that this is an optimization, I'd like to leave it for later or maybe not even implement it later (at most we just have one Cast node that does nothing)

[asterite]

Is it okay to merge this even if the Windows build is failing?

[straight-shoota]

Yeah, that's probably just a hiccup.

[cyangle]

It seems like this PR introduced a bug. The interpreter failed to run the std specs.