Never used Turbo Pascal overlays but we had a product on OS/2 using a Realia compiler that when ported to DOS didn't fit. The overlays destroyed performance. I ended up making a segmented that spilt the call graph into mostly subtree by selecting the subroots. A small set of the shared pets got to live in non overlay memory. There was still a little thrashing because the splits weren't exclusive or of similar sizes but it worked remarkably well.