Found bug in type-inference with `apply(f, args...; kwargs...) = f(args...; kwargs...)`

[schlichtanders]

Unfortunately this is currently blocking me ;-) but at least I could simplify it so that other can find a workaround or actually fix it.

in current julia 1.4 I get the following

julia> apply(f, args...; kwargs...) = f(args...; kwargs...)
apply (generic function with 1 method)

julia> Base.promote_op(apply, typeof(apply), typeof(sin), Int)
Any

julia> apply2(f, args...; kwargs...) = f(args...; kwargs...)
apply2 (generic function with 1 method)

julia> Base.promote_op(apply, typeof(apply2), typeof(sin), Int)
Float64

julia> Base.promote_op(apply2, typeof(apply2), typeof(sin), Int)
Float64

julia> Base.promote_op(apply, typeof(apply), typeof(sin), Int)
Any

here again for quick copy and paste to try it yourself

apply(f, args...; kwargs...) = f(args...; kwargs...)
Base.promote_op(apply, typeof(apply), typeof(sin), Int)
apply2(f, args...; kwargs...) = f(args...; kwargs...)
Base.promote_op(apply, typeof(apply2), typeof(sin), Int)
Base.promote_op(apply2, typeof(apply2), typeof(sin), Int)
Base.promote_op(apply, typeof(apply), typeof(sin), Int)

and here with switched names

apply2(f, args...; kwargs...) = f(args...; kwargs...)
Base.promote_op(apply2, typeof(apply2), typeof(sin), Int)
apply(f, args...; kwargs...) = f(args...; kwargs...)
Base.promote_op(apply2, typeof(apply), typeof(sin), Int)
Base.promote_op(apply, typeof(apply), typeof(sin), Int)
Base.promote_op(apply2, typeof(apply2), typeof(sin), Int)

As you yourself can check, the unexpected outputs do not depend on the specific names, but seems to have todo with being the "first" or the "second" apply. Quite unbelievable.

[schlichtanders]

I am so happy :) I just found a workaround

julia> apply(f, args) = f(args...)
apply (generic function with 1 method)

julia> Base.promote_op(apply, typeof(apply), typeof((sin, (1,))))
Float64

[vtjnash]

The result Any is always correct. Asking inference for the possible result types of calling apply(apply, sin, ::Int) is rather aggressive (calling that function wouldn't even return! sorry, mistyped this in testing), and inference is allowed to do whatever inference feels like doing. (indeed, that makes Any always the most correct answer)