Note that we do sibling call optimization per LLVM, and you can use trampolines like Clojure does. Sibling call optimization, most notably, includes all self-calls. This means that calls to the currently executing function in tail position will be tail call optimized, if there are no destructors in scope.
The biggest problem here is ABIs: you have to use Pascal calling conventions instead of C calling conventions. It's also difficult when you have destructors: there are not many tail call positions when destructors are in scope.