dangerous type inference failure

[vtjnash]

Start by defining a function for which it is impossible to generate exact information:

julia> function g()
        g(i,2)
       end
# methods for generic function g
g() at none:2

julia> g(::Any,::Int) = 2
# methods for generic function g
g() at none:2
g(::Any,::Int64) at none:1

Looking at the finfer result, we see julia has assumed too much, even though it will call jl_apply_generic and does not know the run-time return type:

julia> finfer(g,())
:($(Expr(:lambda, {}, {{},{},{}}, quote  # none, line 2:
        return g(i,2)::Int64
    end)))

And in fact, we can break it quite easily:

julia> g(::Int,::Int) = 3.0
# methods for generic function g
g() at none:2
g(::Int64,::Int64) at none:1
g(::Any,::Int64) at none:1

julia> finfer(g,())
:($(Expr(:lambda, {}, {{},{},{}}, quote  # none, line 2:
        return g(i,2)::Int64
    end)))

julia> i=1
1

julia> g()
4613937818241073152

julia> reinterpret(Float64, ans)
3.0

[JeffBezanson]

Same issue as #265. Seems bad, but in practice this doesn't come up much.

[vtjnash]

It seems there should at least be a typeassert in there.

[vtjnash]

closing, since this is a duplicate