Weak type inference in a special case

[cdsousa]

Sorry for the not so good title. I'm not sure about the right words to describe this issue.
Hence I'll put it as an auto-explainable (I hope) example. I found this issue due to JuliaImages/Images.jl#356.

Let's begin with the following definitions

type A{N}; end
type B{N}; end

getN{N}(::A{N}) = N

f(a) = B{getN(a)}()

In case we call f with a variable whose type has a Datatype as parameter,

@code_warntype f(A{Int}())

Variables:
  a::A{Int64}

Body:
  begin  # In[2], line 3:
      return $(Expr(:new, B{Int64}))
  end::B{Int64}

the final type is correctly inferred.

However, I we call f with a variable whose type has an integer value as parameter,

@code_warntype f(A{123}())

Variables:
  a::A{123}

Body:
  begin  # In[2], line 3:
      return call((top(apply_type))(Main.B,123)::TYPE{_<:B{N}})::B{N}
  end::B{N}

the return type cannot be correctly inferred (end::B{N}).
Can this be considered a bug, or is there any technical difficulty with this?

Moreover, something stranger occurs. If we wrap the integer value within a parametric type, e.g. a Tuple, the inference works:

@code_warntype f(A{Tuple{123}}())

Variables:
  a::A{Tuple{123}}

Body:
  begin  # In[2], line 3:
      return $(Expr(:new, B{Tuple{123}}))
  end::B{Tuple{123}}

[johnmyleswhite]

What is a in your first codeblock?

[yuyichao]

I think this is just because julia is not doing any local constant propagation except for types right now. Dup of #11159 and covered by #3440

[cdsousa]

Thanks @yuyichao.
(@johnmyleswhite, a was meant as a variable of type A.)

[cdsousa]

So this resumes to

t() = Val{123}
n() = 456
f_t() = Val{t()}
f_n() = Val{n()}
@code_warntype f_t()  # --->   end::Type{Val{Val{123}}}  type inferred
@code_warntype f_n()  # --->   end::TYPE{_<:VAL{T}}      type not inferred

@yuyichao, are you sure this is the same issue of #11159? If so, which line of #3440 covers it?

[yuyichao]

are you sure this is the same issue of #11159?

Yes. At least almost.

If so, which line of #3440 covers it?

See Jameson's comment #11159 (comment) .