Regression in type inference in closures

[davidanthoff]

This is the repo code:

function get_f()
    a = 0.2

    function f2()
        b = 2/a

        return b
    end

    return f2
end

f = get_f()

@code_warntype f()

On julia 0.4 I get:

Variables:
  b::Float64

Body:
  begin  # C:\Users\anthoff\source\4sD.jl\test.jl, line 5:
      b = (Base.box)(Base.Float64,(Base.div_float)((Base.box)(Float64,(Base.sitofp)(Float64,2)),a::Float64)) # C:\Users\anthoff\source\4sD.jl\test.jl, line 7:
      return b::Float64
  end::Float64

On julia 0.5 I get:

Variables:
  #self#::#f2#1
  b::Any

Body:
  begin  # C:\Users\anthoff\source\4sD.jl\test.jl, line 5:
      b::Any = (2 / (top(getfield))((top(getfield))(#self#::#f2#1,:a)::Any,:contents)::Any)::Any # C:\Users\anthoff\source\4sD.jl\test.jl, line 7:
      return b::Any
  end::Any

[yuyichao]

Dup of #15276

[davidanthoff]

Is this a dup? My problem has no arrays or comprehensions in it at all. Even if this has the same root cause, isn't this a much simpler repo than #15276?

[yuyichao]

It is a dup. Both of them are caused by closure lowering. It is very easy to construct simple repros and there's no reason to keep multiple issue around for exactly the same issue.

julia> function f()
           a = 2
           (()->a)()
       end
f (generic function with 1 method)

julia> @code_warntype f()
Variables:
  #self#::#f
  a::Core.Box
  #1::##1#2

Body:
  begin  # REPL[1], line 2:
      a::Core.Box = $(Expr(:new, :((top(getfield))(Core,:Box)::Type{Core.Box})))
      (top(setfield!))(a::Core.Box,:contents,2)::Int64 # REPL[1], line 3:
      GenSym(0) = a::Core.Box
      return (top(getfield))(GenSym(0),:contents)::Any
  end::Any

[davidanthoff]

Alright, thanks for looking into it!