Support double splatting on NamedTuple union.

[zw963]

Bug Report

Both of following code give same error message

def foo(a, b, c = false)
  p [a, b, c]
end

args = {
  a: 1,
  b: 2,
}

# Both of following code not work

if rand(5) > 3
  args = args.merge(g: true)
end

# ----------------------

if rand(5) > 3
  args = {
    a: 1, b: 2,
  }
else
  args = {
    a: 1, b: 2, c: true,
  }
end

foo(**args)

The very detailed discussion, check here

I didn't see an opened issue for this, considering this issue has been reported by the compiler itself for a long time, and I'm not the only one who meet this issue, I created a new issue.

[devnote-dev]

I'm curious as to when you would ever need to do this; there are so many alternative ways to handle this that it makes me question if this should be implemented at all.

[ysbaddaden]

Work arounds for those that need a solution:

1. the obvious one: avoid creating an union; either by always merging, or with a distinct double splat in each branch of a condition;

   foo(**args.merge(checked: rand(0..2) == 0))

   if rand(0..2) == 0
     foo(**args.merge(checked: true))
   else
     foo(**args)
   end

2. add a method overload that takes a NamedTuple:

   def foo(**args)
     foo(args)
   end
   
   def foo(args : NamedTuple)
     p args
   end
   
   if rand(0..2) == 0
     args = args.merge(checked: true))
   end
   
   foo(args)

I believe libraries should always provide an overload to accept a NamedTuple directly. The point is to spare an expensive Hash when args are known at compile-time. Double splats are mostly sugar coat: they hide a merge (for example foo(**args, checked: true) reads much better than foo(args.merge(checked: true)) and allow intermediary methods to anonymously pass args around.