Coroutines in one page of C

I've written a single page implementation of coroutines in C using setjmp/longjmp and just a little bit of inline assembly. The article is at Here's an example C program using coroutines – equivalent to the Python code:

def iterate(max_x, max_y):
  for x in range(max_x):
    for y in range(max_y):
      yield x,y

for x,y in iterate(2,2):
  print x,y

In C:

#include <stdio.h>
#include "coroutine.h"

typedef struct {
  coroutine* c;
  int max_x, max_y;
  int x, y;
} iter;

void iterate(void* p) {
  iter* it = (iter*)p;
  int x,y;
  for(x=0; x<it->max_x; x++) {
    for(y=0; y<it->max_y; y++) {
      it->x = x;
      it->y = y;

#define N 1024

int main() {
  coroutine c;
  int stack[N];
  iter it = {&c, 3, 2};
  start(&c, &iterate, &it, stack+N);
  while(next(&c)) {
    printf("x=%d y=%dn", it.x, it.y);

Yeah it's a bit longer, or perhaps more than a bit longer. But it's still very useful at times. Check it out!


#1 Tanim Islam on 08.20.13 at 7:51 am

In your python routine, should you not replace the "range" with "xrange," at least to preserve the spirit of using a pure python iterable?

#2 John Regehr on 08.20.13 at 9:49 am

This line looks like undefined behavior to me, but probably it is reliably compiled in practice:

start_params* p = ((start_params*)sp)-1;

#3 Yossi Kreinin on 08.20.13 at 11:49 am

@Tanim Islam: yeah, I just figured "range" is more mnemonic for someone not knowing Python and so reading it as pseudocode.

@John Regehr: very possibly it's undefined :) I'm not very good at C's semantics, really; I know I can carve out a struct out of a byte array in some cases but not others, and I was never completely sure which are which. With this type of code, I know I'll be able to maintain it as compilers/platforms change over the years but it's not because I'm trusting my formal understanding of C as much as because I trust that I can make sense of the assembly coming out of the compiler and how it's wrong and why. I wonder if start() in general is at all possible to defend in terms of C's semantics – for instance the fiddling with $sp and $fp, not to mention the FRAME_SZ constant.

BTW I discovered that on my platform for which I originally started writing this stuff, setjmp is broken in that it assumes 32b floating point registers on 32b processor variants; that's because the newlib version is from 1993 when there were just two variants of the machine – all registers 32b or all registers 64b, ints as well as floats. So the part which is supposedly correct according to C's semantics will need to be rewritten…

#4 David Janke on 08.20.13 at 11:55 am

also, xrange is not part of python 3… and range is now a Sequence

#5 Jon on 09.06.13 at 5:57 pm

Hi – this is cool. I'm going to play with it over the weekend. :)

I'm wondering if you've seen this other implementation of C coroutines? It probably has an uglier API, but I think it works by expanding macros into switch statements, which is possibly more portable, I'm not sure. It'd be interesting to see a comparison between the two.

#6 Yossi Kreinin on 09.06.13 at 11:27 pm

Indeed it's more portable; the downside is, you can only yield at the top level of your coroutine – that is, the coroutine can't itself call functions and have one of them yield control to the producer. That's because calling functions that yield requires multiple call stacks.

#7 Michael Moser on 09.24.13 at 8:05 pm

Thanks for the article,

Now there is another trick for solving lack of getcontext: if jmp_buf is an array of values then one can infer the index that stores the stack pointer and change jmp_buf directly.

Also: a very useful feature for coroutines is to have a guard page at the end of the stack; otherwise one will never know if the program ran out of the limited stack space (if one wants to service lots of coroutines, then that's what it gets)

#8 Yossi Kreinin on 09.24.13 at 10:36 pm

You could change the sp (and fp) directly by knowing which libc you're targeting; you'd need to deal with the fact that your local variables could have been rendered invalid by that when comping back from setjmp through longjmp though.

As to guard pages – if you have paging, you probably have getcontext… Non-Unix operating systems might have paging or they could have other memory protection features; I'm on an OS which doesn't use paging, for instance.

#9 Sagi on 04.20.14 at 4:09 am

For c++ programmers Boost appears to include Coroutines too

#10 kevin on 10.27.14 at 5:54 pm

Hi Yosef, Nice writeup! So, I have tried this and it works on MACOSX 64bit and Windows 32bit very nicely, but Win64/x64 inline asm is not allowed in visual c++.

Any ideas how to write the set/get_sp/fp functions on _WIN64 target? I'm trying MASM but I think the function wind/unwind stuff is getting in the way, I'm no ASM guru.

I tried the TonyFinch/Picoro stuff from his github and it's awesomely portable and simple to read, but limited by stack space. so I am preferring your method which allows me to create my own stack space on the heap, and patch it in to each coroutine…

Wishing I could use ucontext but it seems deprecated by POSIX group and broken on MACOSX (my code which works on linux hangs on MACOSX). And finally, windows Fibers seems limited to stack space as well, I get around 2028 coroutines… (and it isn't portable). I even found a ucontext implementation for windows.

Some work I did on top of ucontext, for the curious:

#11 Yossi Kreinin on 10.27.14 at 6:25 pm

No inline asm at all? How come?

My solution for Windows was actually to cross-compile with mingw, so I don't know how to do it in VS.

#12 kevin on 10.27.14 at 6:41 pm

Right, visual C++ doesn't support inline ASM for x64 targets, I guess they didn't want to have it interfering with their optimizer

#13 kevin on 10.30.14 at 5:33 pm

This page descries the rationale for deprecating ucontext:

They think that making people port to threads (i.e. pthreads) is easier/better than writing a new interface for ucontext which conforms to C standards. I guess they leave us with no POSIX way to implement coroutines, which seems shortsighted to me.

#14 Yossi Kreinin on 10.30.14 at 5:59 pm

To be fair (?), C has the problem of handling stack overflow in coroutines that is kinda solvable for the OS, especially if you have plentiful pointer bits (by trapping stack overflow and mapping more memory pages on demand), but not solvable in userspace on today's systems. Solved in Go by having a calling convention handling stack overflow, AFAIK… So deprecating ucontext is like saying, "we take away a portable method to implement a construct but not really", which can seem reasonable; "if you're willing to risk stack overflow then you should also be willing to fiddle with assembly." On the other hand, C always lets you shoot yourself in the foot… so why not make this shot easier is not really clear…

#15 Jesse Lactin on 03.30.17 at 6:08 am

Hi, Yossi. Cool write up. Coroutines have really opened my eyes to how flexible and clean iteration can be (C++ iterators are put to shame for their verbosity and implementation concerns). I think your one page API could be made to look more like the Python example if you're a glutton for punishment and want to mix variable argument lists and setjmp/longjmp/makecontext. But I'm at a loss of how to go about it, as my experience with setjmp/longjmp is limited, and I can't reason about the resulting control flow. Do you have any ideas as to where I could improve my setjmp mastery?

#16 Yossi Kreinin on 03.30.17 at 11:02 am

I think setjmp/longjmp sink in once you fiddle with them for a few hours, in the same way recursion does. C varargs are different though, they're always ugly at the callee code and never safe. I think you need C++ templates to make this look pretty (as in, any coroutine prototype works and you call it naturally.)

#17 Jesse Lactin on 03.31.17 at 12:28 am

Thanks for getting back to me so quickly. I was thinking of an API that adds varargs to next() and yield() like this.

void yield(coroutine *c, void *arg0, …);
int next(corotuine *, void *arg0, …);

start() would be left unchanged. If arg0 is NULL, the code would terminate vararg handling. This API would require a simple change to the coroutine struct (a fixed-size buffer of void *s, and void ** pointing the next slot). It's not the prettiest solution, but it seems to work. I have no PC right now, so I can't run GDB to figure out why it crashes. I'm stuck with online C compilers ATM.

Have you heard of libffi? It allows you to call any function when you only know the API at runtime. It's the perfect solution if for calling an arbitrary coroutine. Here's an example Simon Tatham wrote:

#18 Yossi Kreinin on 04.01.17 at 11:42 am

The problem with varargs is you need some sort of a convention for figuring out which arguments were actually passed (printf has one hairy one which isn't suitable for anything except printf.) With these complications, I'm not sure the code is gonna be prettier or safer with varargs than with a hand-coded params struct, and maybe the reverse. A C++11 variadic template instantiating a tuple struct from the parameters and then calling the entry point functions with these arguments might be as clean as what you're looking for.

#19 Jesse Lactin on 04.03.17 at 1:28 am

My convention for this particular function is that the list of args is terminated with a NULL pointer. You were right about setjmp; I get it now. One last note, I got the API changes working, but it required volatile-qualifying x and y in iterate(). It printed garbage without it, and it seems that it's UB otherwise.

#20 Jesse Lactin on 04.10.17 at 7:28 pm

I got it working. Here's the gist of it:

I like this API a little more because it doesn't require me to make a struct. For mixing varargs and setjmp, I think it's a pretty clean interface. Once I got my head around setjmp (next() longjmps back into start(), and calls the entry point function), it all seemed to fall into place for me.

#21 Yossi Kreinin on 04.18.17 at 10:27 pm

I guess I prefer almost anything to varargs except when I'm ultimately forwarding arguments for printf, because of how unsafe and unreadable accessing varargs is, but to each his own :-) I don't think my code makes a good library anyway, it's more of an example one would mold to suit their use case, which I guess is what you did there.

#22 Jesse Lactin on 04.21.17 at 6:17 am

I think your right. Varargs are ugly and error prone. My modification asserts when not enough arguments are missing. I haven't worked out all the corner cases yet. Maybe explicitly passing the expected number of arguments is safer. I saw a simple Lisp implementation use varargs for constructing cons-cells. I thought it was an alright solution instead of having multiple constructor functions. Would you be interested in seeing the code on Gist? I could format, comment, and post it.

Leave a Comment