Structured Dagger Code:
=======================
forall [i] (1:N,1) {
  atomic { 
    sum[i] = 0;
  }
  forall [j] (M:1,-1) {
    when e1[i-1](MSG *m1), e2[j-1](MSG *m2) {
      atomic {
        sum[i] += m1->data * m2->data;
      }
    }
  }
  atomic {
    CPrintf("Sum = %d\n", sum[i]);
  }
}

Translated Code:
================
forall_1_begin()
{
  int start = 1, end = N, incr = 1;
  if (start > end) {
    swap(&start, &end);
    incr = -incr;
  }
  Counter *c1 = new Counter(1,N,1)
  for(int i=start; i != end; i+=incr) {
    _slist_1_begin(i, c1);
  }
  forall_1_end(c1);
}

forall_1_end(Counter *c1)
{
  c1->done();
  if(c1->finished()) {
    delete c1;
    // call next
  }
}

slist_1_begin(int i, Counter *c1)
{
  atomic_1(i, c1);
}

slist_1_end(int i, Counter *c1)
{
  forall_1_end(c1);
}

forall_2_begin(int i, Counter *c1)
{
  int start = M, end = 1, incr = -1;
  if (start > end) {
    swap(&start, &end);
    incr = -incr;
  }
  if(start > end)
  Counter *c2 = new Counter(M,1,-1);
  for(int j=start; j >= end; j+= incr) {
    when_1(i, j, c1, c2);
  }
  forall_2_end(i, c1, c2);
}

forall_2_end(int i, Counter *c1, Counter *c2)
{
  c2->done();
  if(c2->finished()) {
    delete c2;
    atomic_3(i, c1);
  }
}

#define E1	1
#define E2	1

#define WHEN_1	1

int when_1(int i, int j, Counter *c1, Counter *c2)
{
  MSG *m1;
  MSG *m2;

  m1 = e1_A->access(i-1);
  m2 = e2_A->access(j-1);
  if(m1 != 0 && m2 != 0) {
    m1 = e1_A->remove(i-1);
    m2 = e2_A->remove(j-1);
    atomic_2(i,j,m1,m2,c1,c2);
    return 1;
  } else {
    int *tmp_entry = new int[2];
    int *tmp_ref = new int[2];
    size_t *args = new size_t[4];
    tmp_entry[0] = E1;
    tmp_ref[0] = i-1;
    tmp_entry[1] = E2;
    tmp_ref[0] = j-1;
    args[0] = i;
    args[1] = j;
    args[2] = c1;
    args[3] = c2;
    when_1_A->register(2, tmp_entry, tmp_ref, args);
    return 0;
  }
}

atomic_1(int i, Counter *c1)
{
  {
    sum[i] = 0;
  }
  forall_2_begin(i, c1);
}

atomic_2(int i, int j, MSG *m1, MSG *m2, Counter *c1, Counter *c2)
{
  {
    sum[i] += m1->data * m2->data;
  }
  forall_2_end(i, c1, c2);
}

atomic_3(int i, Counter *c1)
{
  {
    CPrintf("Sum = %d\n", sum[i]);
  }
  slist_1_end(i, c1);
}

void e1(MSG *m)
{
  int refnum = GetRefNum(m);
  // buffer the message first
  e1_Buffer->putElement(refnum, m);
  // get a list of when blocks waiting for this msg
  TList<someStruct> *e1_A = e1_Buffer->getList(refnum);
  // no one is waiting, so return
  if(e1_A == 0) {
    return;
  }
  for(WhenStruct *tmp=e1_A.begin();!e1_A.end();tmp=e1_A.next()){
    switch(tmp->id) {
      case WHEN_1:
        if(tmp->allDependenciesSatisfied()) {
          // deregister removes element from all entryLists
          tmp->deregister();
          when_1(tmp->args[0], tmp->args[1], tmp->args[2], tmp->args[3]);
          goto out;
        }
        break;
    }
  }
out:
  return;
}

void e2(MSG *m)
{
  int refnum = GetRefNum(m);
  // buffer the message first
  e2_Buffer->putElement(refnum, m);
  // get a list of when blocks waiting for this msg
  TList<someStruct> *e2_A = e2_Buffer->getList(refnum);
  // no one is waiting, so return
  if(e2_A == 0) {
    return;
  }
  for(WhenStruct *tmp=e2_A.begin();!e2_A.end();tmp=e2_A.next()){
    switch(tmp->id) {
      case WHEN_1:
        if(tmp->allDependenciesSatisfied()) {
          // deregister removes element from all entryLists
          tmp->deregister();
          when_1(tmp->args[0], tmp->args[1], tmp->args[2], tmp->args[3]);
          goto out;
        }
        break;
    }
  }
out:
  return;
}

_sdag_classname_init()
{
  // Initialize all the data structures
}