C Programming/Coroutines

A little known fact is that most C implementations have built-in primitives that can be used for cooperative multitasking / coroutines. They are setcontext and setjmp.

setjmp

edit

The function setjmp is used in a pair with longjmp to transfer execution to a different point in the code. It relies on an existing jmp_buf declaration.

#include <setjmp.h>

int main (void)
{
  jmp_buf buf1;
  
  if (setjmp(buf1) == 0)
  {
    /* This code is executed on the first call to setjmp. */

    longjmp(buf1, 1);
  } else {
    /* This code is executed once longjmp is called. */
  }
  return 0; 
}

setjmp() stores the current execution point in memory, which remains valid as long as the containing function doesn't return. It initially returns 0. Control is returned to setjmp once longjmp is called with the original jmp_buf and the replacement return value.

Note that jmp_buf is passed to setjmp without using the address-of operator.

The easiest way to understand setjmp and longjmp, is that setjmp stores the state of the cpu which includes program counter, stack pointer, all the registers, including the bits of the flags register, at the location pointed to by jmp_buf , which is defined some LEN+1, which is enough bytes to store the registers of whatever CPU is involved. The longjmp(buf) , never returns, because it restores the CPU from the contents of struct jmp_buf buf previously set by a previous call the setjmp, so execution begins from after setjmp was called, but the return value of setjmp is not 0, but whatever value was used in the second parameter to longjmp. This is similar to the fork() system call, which returns 0 to the child process, and the PID of the child process to the parent.

The internet suggests co-routines are useful for implementing software as state machines cooperating, such as lexer processing input text and emitting tokens , so that a parser can decide to store the token and ask for the next one, or to act on its current set of token . This is not multithreaded programs synchronising on data, with possibly race conditions if a bug forgets to acquire a lock, but with setjmp and longjmp, seems to be cooperative processes that guarantee only one process will run at a time, with no worries about context switches waking up sleeping processes ( using separate jmp_buf static locations, each process can call setjmp for its own jmp_buf, before calling longjmp later on if zero was returned, or continue a loop to process shared data for non-zero returns).