Constant-folding in inference for getfield

[timholy]

If possible, it would be ideal to perform constant folding before running inference on getfield. Here's a demo:

julia> abstract Container{N}

julia> type Container1d <: Container{1}
           center1::Float32
           child::Container1d
       end

julia> type Container2d <: Container{2}
           center1::Float32
           center2::Float32
           child::Container2d
       end

julia> getchild{N}(parent::Container{N}) = getfield(parent, N+1)
getchild (generic function with 1 method)

julia> function childpositive(c::Container)
           child = getchild(c)
           child.center1 > 0
       end
childpositive (generic function with 1 method)

julia> code_typed(childpositive, (Container1d,))
1-element Array{Any,1}:
 :($(Expr(:lambda, Any[:c], Any[Any[:child,:_var0,:_var3,:_var4],Any[Any[:c,Container1d,0],Any[:child,Union(Container1d,Float32),18],Any[:_var0,Float32,18],Any[:_var3,(Int64,),0],Any[:_var4,Float32,18]],Any[]], :(begin  # none, line 2:
        child = getfield(c::Container1d,(top(box))(Int64,(top(add_int))(1,1))::Int64)::Union(Container1d,Float32) # line 3:
        _var4 = (top(getfield))(child::Union(Container1d,Float32),:center1)::Float32
        _var0 = (top(box))(Float32,(top(sitofp))(Float32,0))::Float32
        return (top(box))(Bool,(top(or_int))((top(lt_float))(_var0::Float32,_var4::Float32)::Bool,(top(box))(Bool,(top(and_int))((top(eq_float))(_var0::Float32,_var4::Float32)::Bool,(top(box))(Bool,(top(or_int))((top(eq_float))(_var0::Float32,9.223372f18)::Bool,(top(slt_int))(0,(top(box))(Int64,(top(fptosi))(Int64,_var0::Float32))::Int64)::Bool))::Bool))::Bool))::Bool
    end::Bool))))

Note the Union(Container1d,Float32) for the result of getchild(c), which arises presumably because add_int(1,1) isn't evaluated.

For comparison,

julia> function centerpositive(c::Container)
           c.center1 > 0
       end
centerpositive (generic function with 1 method)

julia> code_typed(centerpositive, (Container1d,))
1-element Array{Any,1}:
 :($(Expr(:lambda, Any[:c], Any[Any[:_var0,:_var3,:_var4],Any[Any[:c,Container1d,0],Any[:_var0,Float32,18],Any[:_var3,(Int64,),0],Any[:_var4,Float32,18]],Any[]], :(begin  # none, line 2:
        _var4 = (top(getfield))(c::Container1d,:center1)::Float32
        _var0 = (top(box))(Float32,(top(sitofp))(Float32,0))::Float32
        return (top(box))(Bool,(top(or_int))((top(lt_float))(_var0::Float32,_var4::Float32)::Bool,(top(box))(Bool,(top(and_int))((top(eq_float))(_var0::Float32,_var4::Float32)::Bool,(top(box))(Bool,(top(or_int))((top(eq_float))(_var0::Float32,9.223372f18)::Bool,(top(slt_int))(0,(top(box))(Int64,(top(fptosi))(Int64,_var0::Float32))::Int64)::Bool))::Bool))::Bool))::Bool
    end::Bool))))

is, of course, inferred beautifully.

Now that we have some parse-time constant folding in cartesian.jl, perhaps we could just use that?

[JeffBezanson]

I think this is a dup of #5560.

[timholy]

Agreed. I did search, but used the term "fold."