🚀 go-pugleaf

I am quite new to OCaml and interested in the simple mandelbrot computation presented here:
http://metamatix.org/~ocaml/price-of-abstraction.html

What I would like to optimize is this function:
let iters1 max_iter xc yc = 
  let rec aux count x y = 
    if count = max_iter then max_iter else begin
      if x *. x +. y *. y >= 4.0 then count else
      aux (count+1) (x *. x -. y *. y +. xc) (2.0 *. x *. y +. yc)
    end in
  aux 0 xc yc

The main problem with this is that arguments x and y in function aux are always boxed so that contributes to a noticeable slowdown.

I tried several approaches to make this faster, one of them is this:
type complex = {
  x : float;
  y : float;
}
let iters1 max_iter xc yc = 
  let rec aux count c = 
    if count = max_iter then max_iter else begin
      if c.x *. c.x  +. c.y *. c.y >= 4.0 then count else begin
        aux (count+1) { x = c.x *. c.x -. c.y *. c.y +. xc;
                        y = 2.0 *. c.x *. c.y +. yc  }
      end
    end in
  aux 0 { x=xc; y=yc }

This time fields x and y are unboxed, but the record itself is boxed so I gained no speed up. The running time is the same as before.
Using references didn't help either.

I would like to understand the workings of ocamlopt better so I'm curious whether it is possible to avoid boxing in recursive functions like "aux"?

Thanks!