I've always been a little bit uncomfortable with the line of reasoning that TCO is a semantics change. For any list of programs (and inputs) that would return

I've always been a little bit uncomfortable with the line of reasoning that TCO is a semantics change. For any list of programs (and inputs) that would return with TCO but would otherwise fail with a memory error, there is some amount of memory additional memory you could add that would also cause the program to return. I don't normally consider buying RAM to change the semantics of my programming language. It certainly changes what's practical to do with constrained resources in a language, but the same could be said of most things that are considered "optimizations." On the other hand, I do sympathize with the idea that TCO is somehow different that most optimizations. It potentially turns O(\inf) memory usage into O(1). If anything, it feels like an algorithmic optimization more than the implementation optimizations that compilers are generally limited to.