Wrong Vararg expansion in type intersection

[martinholters]

julia> T = Tuple{Val{N}, Any, Any, Vararg{Any,N}} where N
Tuple{Val{N},Any,Any,Vararg{Any,N}} where N

julia> typeintersect(T, Tuple{Val{1}, Vararg{Any}})
Tuple{Val{1},Any,Any,Any,Any,Any}

julia> typeintersect(T, Tuple{Val{1}, Any, Vararg{Any}})
Tuple{Val{1},Any,Any,Any,Any}

julia> typeintersect(T, Tuple{Val{1}, Any, Any, Vararg{Any}})
Tuple{Val{1},Any,Any,Any}

Clearly, the result should be Tuple{Val{1},Any,Any,Any} in all three cases.

Of course, this can have an impact on inference, e.g.

julia> foo(::Val{N}, unused, args::Vararg{Any,N}) where {N} = (Val(N), args)
foo (generic function with 1 method)

julia> code_warntype(foo, Tuple{Val{1}, Any, Vararg}) # good
Variables
  #self#::Core.Const(foo, false)
  #unused#::Core.Const(Val{1}(), false)
  unused::Any
  args::Tuple{Any}

Body::Tuple{Val{1},Tuple{Any}}
1 ─ %1 = Main.Val($(Expr(:static_parameter, 1)))::Core.Const(Val{1}(), false)
│   %2 = Core.tuple(%1, args)::Core.PartialStruct(Tuple{Val{1},Tuple{Any}}, Any[Core.Const(Val{1}(), false), Tuple{Any}])
└──      return %2

julia> code_warntype(foo, Tuple{Val{1}, Vararg}) # bad
Variables
  #self#::Core.Const(foo, false)
  #unused#::Core.Const(Val{1}(), false)
  unused::Any
  args::Tuple{Any,Any}

Body::Tuple{Val{_A} where _A,Tuple{Any,Any}}
1 ─ %1 = Main.Val($(Expr(:static_parameter, 1)))::Val{_A} where _A
│   %2 = Core.tuple(%1, args)::Tuple{Val{_A} where _A,Tuple{Any,Any}}
└──      return %2

Note that in the bad case, not only is inference of the first return value needlessly imprecise (because N wasn't fixed), but the type of the second return value is plain wrong.

I've stumbled upon this when looking at suboptimal inference of this constructor:

julia/base/bitarray.jl

Line 28 in 4b42f14

function BitArray{N}(::UndefInitializer, dims::Vararg{Int,N}) where N

So this is not a purely theoretical issue.