Compilers and interpreters can (more easily) optimize tail-recursive algorithms to avoid the normal cost of creating a new stack frame and making a normal function call, which means the recursive algorithm has performance more like that of an iterative implementation.
https://en.wikipedia.org/wiki/Tail_call
https://llvm.org/docs/CodeGenerator.html#tail-call-optimizat...
Now Python doesn't do tail call optimization out of the box (surprise). but there is a module that is adding some magical decorator that fixes that: https://pypi.org/project/tail-recursive/
(actually need to look how they implement this decorator, it must be some serious hack.)
If you can only optimise calls to yourself, you have a very limited form of tail call optimisation. It's an important point in the context of Scheme, where they made a big deal out of experimenting with continuation passing style. In CPS, all function calls are tail calls, and you want to optimise them all, even though most are not recursive.
Eliminating recursion might be difficult to detect if multiple functions form a mutually-recursive set, or if data structures have to be introduced to compensate for no stack (e.g., converting a depth-first search from recursive to iterative form requires a stack).