Thanks to the courage you all gave me, I found the strength to go on. Template instantiation seems to be the easiest path, so I decided to follow it. I think I solved the problem about primitives and FFI I though it had. Namely, how the hell do I compile primitive calls with an arbitrary number of arguments. Starting from the Tutorial of Peyton Jones and Lester, I added a new node type to help guide the order of evaluation. Basically, this "Seq" node takes the address of two other nodes. It pushes them both on the stack, like an application node would. (That is because I don't want a dump, but I could as well create a new stack and and push the previous one on the dump.) Note the nodes which are evaluated this way must not have a functional head normal form (primitive or supercombinator). That would leak past the beginning of the stack. When the node is finally evaluated (e.g. points to a HNF, like a boxed integer), the topmost and second topmost pointers on the stack are swapped. (or the topmost pointer of the stack and the topmost pointer of the topmost stack on the dump are swapped). With a suitable compilation scheme, that should guarantee that primitives have their arguments already evaluated. (And avoid infinite loops). Now, I just have to implement it. :-) Cheers, Loup.