Show HN: Python decorator that enables arbitrarily-deep tail/non-tail recursion

by tylerhou • 5 years ago

119 21 5 years ago

tylerhou7• 5 years ago

Hi HN, I wrote this.

Admittedly Fiber is not the best name — in the current implementation we don't yield to the scheduler except for recursive function calls, and our "scheduler" only runs a function to completion. These things could be fixed of course, but I didn't have time.

I called it Fiber because I thought coroutine + scheduler is approximately equal to a fiber implementation, and other schedulers could be added in the future with more clever behavior. Plus, you could change the yielding behavior to match what is necessary: for example, to yield at an arbitrary point one could modify the decorator to emit "pause" returns whenever a special `yield_()` function is called. Or you could create a list of special functions (e.g. functions associated with normal syscalls) that cause the coroutine to yield, and have a scheduler that will resume the function when the syscall completes.

One point behind this project is to demonstrate that if you want to pause and resume functions, you don't actually need runtime (kernel, VM) support; some hacky AST rewrites are also a feasible (if slow) implementation. One thing to note is that our overhead is mostly per pause/resume; hence the "sum a list" benchmark is really a worst-case benchmark. If you have some recursive function that does a lot of non-recursive work that needs to be paused and resumed (hint hint, React rendering and reconciliation) the performance gap will be much narrower.

Tistron• 5 years ago

I don't generally write Python code, so maybe that is it, but should't the first example:

print(trampoline.run(cache, [1010]) % 10 * 5)

reference `fib` somehow, rather than `cache`?

tylerhou7• 5 years ago

You're right, fixed.

KolenCh• 5 years ago

May be compare to tco which is not as general.

a-dub3• 5 years ago

neat idea, but i'm guessing that a rewritten function would be nightmarish to debug... or no?

wonder if it would be possible to do automated rewrites of recursive functions as iterative ones. specifically for cases where stack memory that grows in a recursive implementation can be replaced by a few long lived temporaries... like the good ol' fibonacci function.

tylerhou7• 5 years ago

Yeah, debugging the written function is not easy, but that could be improved by adding the normal tools for rewrites (sourcemap). Also, you can put print/logging statements inside the function; the rewrite preserves those.

In theory the rewrite shouldn't change behavior, so you could also unittest both the decorated function and the undecorated function.

possible to do automated rewrites of recursive functions as iterative ones.

This is basically what we're doing, except the stack is inside the trampoline. If you inlined the trampoline into the function itself then that is exactly a recursion -> iteration transform.

a-dub3• 5 years ago

This is basically what we're doing, except the stack is inside the trampoline. If you inlined the trampoline into the function itself then that is exactly a recursion -> iteration transform.

yeah but aren't you basically creating a stack on the heap, and then running the recursive algorithm there?

i mean more like rewriting the unbounded space recursive version of the fibonacci function as the constant space two temporary iterative version.

vitus• 5 years ago

Fib is a weird case because it's the poster child of dynamic programming. The challenge in this case is that each call spawns multiple recursive calls, and you're ultimately just adding 1's and 0's (so your runtime is comparable to the number you compute). Meanwhile, the DP / iterative approach understands the repeated structure of your recursion and can use that to eliminate redundant computation.

Compare this to tail-recursive functions, where you can basically update the stack variables and jump back to the beginning of the function.

(Note that you can approximately achieve the performance of the iterative solution by memoizing the results if your functions are pure and your input space is small, e.g. via @functools.cache, but you're paying for it with space, whereas the iterative solution understands when you can discard the intermediate results.)

falcor84• 5 years ago

Fib is a weird case because it's the poster child of dynamic programming.

It's actually got a closed-form solution, so finding the nth fib number is O(1) (assuming O(1) arithmetic). https://en.wikipedia.org/wiki/Fibonacci_number#Closed-form_e...

chrisseaton• 5 years ago

i'm guessing that a rewritten function would be nightmarish to debug... or no?

I'm going to guess that it's likely irreducible! So yes a problem for both human and machine understanding!

jhanschoo• 5 years ago

wonder if it would be possible to do automated rewrites of recursive functions as iterative ones. specifically for cases where stack memory that grows in a recursive implementation can be replaced by a few long lived temporaries.

Yes. Some Lisp implementations do this automatically. See https://en.wikipedia.org/wiki/Continuation-passing_style .

Note that this differs from the POC of OP in that in OP's POC, only one function is defined, and it is transformed into a tree with if-statements to determine where in the code execution should be resumed, but in CPS it's an anonymous function in the tail being called with all context still relevant to it. Then the tail-call can be optimized out.

Note that anonymous functions can be seen as just regular top-level functions with a hidden parameter containing the context relevant to it.

thaumasiotes• 5 years ago

wonder if it would be possible to do automated rewrites of recursive functions as iterative ones.

Of course.

specifically for cases where stack memory that grows in a recursive implementation can be replaced by a few long lived temporaries... like the good ol' fibonacci function.

Rewriting a recursive Fibonacci function to use a loop with an explicit stack doesn't do anything about the growth of memory. You'll need just as many stack elements. It's just an explicit stack that you've assigned to a variable rather than being the function call stack.

coldtea• 5 years ago

Rewriting a recursive Fibonacci function to use a loop with an explicit stack doesn't do anything about the growth of memory.

Sort of. It means the amount of memory you can use is much bigger, and not constrained by the stack size.

Show HN: Python decorator that enables arbitrarily-deep tail/non-tail recursion

Related Stories