5.5. Nickle Continuations

Arbitrary flow control is accomplished in Nickle with first-class continuations and the functions setjmp and longjmp. These are similar to those in C, but without restrictions on the target.

poly setjmp ( continuation *c, poly retval )
void lomgjmp ( continuation c, poly retval )

Setjmp saves the state of the program, including program counter and names in scope, in c and returns retval.

Longjmp returns retval from the setjmp that set c. There can be two distinctions from this jump and the initial call to setjmp: the return value may differ, and variables that have changed retain their new values.

Continuations are often used to implement control structures that do not exist in the language, interpreters, and escaping from recursive algorithms. For example, the following is a simple binary tree search that uses continuations to jump directly to the top instead of returning up each branch once the result is found.

typedef tree;

typedef struct {
	int key;
	poly data;
	&poly left, right;
} tree;

void function search ( tree t, int i, &continuation c ) {
	if ( i < t.key && ! is_void ( t.left ) )
		search ( t.left, i, &c );
	else if ( i > t.key && ! is_void ( t.right ) )
		search ( t.right, i, &c );
	else if ( i == t.key )
		longjmp ( c, t.data );

tree t = { key = 2, data = "blah", left = reference ( <> ), right = reference ( <> ) };

continuation c;
int i = 0;
	poly p = setjmp ( &c, <> );
	printf ( "I have been here %d times.\n", i );

	if ( is_void ( p ) )
		search ( t, 2, &c );
		printf ( "value = %g\n", p );

This is a pretty normal binary tree search, but notice how it is run: a continuation is set; if setjmp returns <> (which it will the first time), a value is searched for (this is a pretty degenerate example with only one node). If an actual value is returned, it must be from the longjmp in search, and the value is printed; a message is printed to emphasize that setjmp returns twice. This optimizes the return from what can be a very deeply nested search.

Notice that the last part of the code is inside curly braces. This is legal, of course, but ordinarily not very useful. It is used in this case to get around a slight flaw in Nickle: currently, each top-level command is executed in its own thread. Thus when longjmp tries to return from the setjmp, that thread has already finished and the program exits. By placing those statements in curly braces, they will all be executed in the same thread and setjmp will still exist for longjmp to find.

This sort of escape from a nested search is also commonly done with exceptions, raising one when the value is found and catching it at the top, passing the value as an argument to the exception. Actually, that method is more common because of its simplicity, but this example was done using continuations to demonstrate them.