Good question! I've been working on it to see what the real issue is. I don't have an answer for you; I'm just documenting what I've found for everyone's benefit.

First, I presume that "inherently recursive" means "cannot be rewritten to be tail-recursive". Fibonacci is an example of an inherently recursive algorithm.

sub fibonacci {
   my ($n) = @_;
   return 0 if $n == 0;
   return 1 if $n == 1;
   return sum fibonacci($n-2), fibonacci($n-1);
}
[download]

At face value, it's quite easy to make it threaded:

sub fibonacci {
   my ($n) = @_;
   return 0 if $n == 0;
   return 1 if $n == 1;
   return sum map $_->join(),
      async { fibonacci($n-2) },
      async { fibonacci($n-1) };
}
[download]

But what if you wanted to limit the number of threads (perhaps to limit overhead)? One solution would be to have a pool of threads and to use them if available.

sub fibonacci {
   my ($n) = @_;
   return 0 if $n == 0;
   return 1 if $n == 1;
   return sum map $_->get(),
      async_maybe { fibonacci($n-2) },
      async_maybe { fibonacci($n-1) };
}
[download]

(The above assumes closures can be shared, but it can be rewritten to not use closures.)

When implementing async_maybe (and the get of the object it returns), one must be extra careful to avoid the situation where a thread is waiting to have it's result collected.

But what if you want a worker model (perhaps to distribute the work to other machines)? Now, that's hard. One would need some kind of a callback system.

sub fibonacci {
   my ($n) = @_;
   my $result;
   process_and_wait(sub {
      fibonacci_task(sub { $result = $_[0]; }, $n );
   });
   return $result;
}

sub fibonacci_task {
   my ($on_complete, $n) = @_;
   return $on_complete->(0) if $n == 0;
   return $on_complete->(1) if $n == 1;
   my ($x,$y);
   process(sub { fibonacci_task(sub { $x = $_[0] }, $n-2) });
   process(sub { fibonacci_task(sub { $y = $_[0] }, $n-1) });

   #TODO: The last of the two tasks to complete
   #      must call $on_complete->($x+$y).

}
[download]

The TODO item is hard to do cleanly. And of course, I'm still using closures even though I think that's not an option for threads threads. Rewriting to avoid closures will likely make the code even longer.