SWI-Prolog provides two predicates reset/3 and shift/1 to implement delimited continuations. These two predicates are described in Schrijvers et al., 2013 (preprint PDF), which is a very nice paper. However, the names seems a bit confused. I will explain below.
The reset and shift operators originally come from Danvy; Filinski,1990. “Abstracting Control”. In that paper, the authors claim that
Shift abstracts the current context as an ordinary, composable procedure and reset delimits the scope of such a context.
This means that the continuation captured by shift is like a “total function”, i.e. it always returns normally. But this is not the case in SWI-Prolog.
For example,
?- reset((
format("111~n"),
shift(1),
format("222~n"),
shift(2),
format("333~n")
),
Term1, Cont1),
format("444~n"),
call(Cont1),
format("555~n").
111
444
222
ERROR: reset `2' does not exist (No matching reset/3 call)
The ERROR tells us that Cont1 is not “total”, because there is no reset/3 as a delimiter inside Cont1. A workaround is manually wrapping a reset/3 around the continuation applications:
?- reset((
format("111~n"),
shift(1),
format("222~n"),
shift(2),
format("333~n")
),
Term1, Cont1),
format("444~n"),
reset(call(Cont1), Term2, Cont2), % wrap a reset/3
format("555~n").
111
444
222
555
This works. To obtain the original semantics of reset/shift, we can always do this, but still be careful with the semantics differences. In fact, the reset/3 and shift/1 in SWI-Prolog is more like the prompt and control in Felleisen,1988. The Theory and Practice of First-Class Prompts. The Racket reference demonstrates the differences.
(reset val) => val
(reset E[(shift k expr)]) => (reset ((lambda (k) expr)
(lambda (v) (reset E[v]))))
; where E has no reset
vs.
(prompt val) => val
(prompt E[(control k expr)]) => (prompt ((lambda (k) expr)
(lambda (v) E[v])))
; where E has no prompt
Note that, in the reset/shift case, there is a reset in the captured continuation (but SWI-Prolog no this stuff).
Below is a more complicated example to compare the differences:
-
Prompt and control semantics:
SWI-Prolog version:
P.s. SWI-Prolog’s reset/3 and shift/1 is exactly
promptandcontrolin Racket.:- dynamic save1/1. ?- reset(( format("111~n"), shift(1), format("222~n"), shift(2), format("333~n") ), Term1,Cont1), format("444~n"), assertz(save1(Cont1)). 111 444 ?- save1(Cont1), reset(( format("555~n"), call(Cont1), format("666~n") ), Term2, Cont2), format("777~n"). 555 222 777Note that the
format("666~n")is not called, soCont1is not “total”.Racket version:
> (define save1 null) > (match-let ([(cons v cont) (prompt (begin (printf "111\n") (control k (cons 1 k)) (printf "222\n") (control k (cons 2 k)) (printf "333\n") (cons '_ 0) ))]) (printf "444\n") (set! save1 cont)) 111 444 > (match-let ([(cons v cont) (prompt (begin (printf "555\n") (save1 '_) (printf "666\n") (cons '_ 0) ))]) (printf "777\n")) 555 222 777 -
Reset and shift semantics:
SWI-Prolog version:
P.s. Must wrap a reset/3 around the
call(Cont1).?- reset(( format("111~n"), shift(1), format("222~n"), shift(2), format("333~n") ), Term1,Cont1), format("444~n"), assertz(save1(Cont1)). 111 444 ?- save1(Cont1), reset(( format("555~n"), reset(call(Cont1), _, _), format("666~n") ), Term2, Cont2), format("777~n"). 555 222 666 777Now the
format("666~n")is called.Racket version:
> (define save1 null) > (match-let ([(cons v cont) (reset (begin (printf "111\n") (shift k (cons 1 k)) (printf "222\n") (shift k (cons 2 k)) (printf "333\n") (cons '_ 0) ))]) (printf "444\n") (set! save1 cont)) 111 444 > (match-let ([(cons v cont) (reset (begin (printf "555\n") (save1 '_) (printf "666\n") (cons '_ 0) ))]) (printf "777\n")) 555 222 666 777
Therefore, the conclusion is that the reset/3 and shift/1 in SWI-Prolog is exactly prompt and control of Felleisen,1988.
Some useful quotes from Abstracting Control:
While the effects of these operators are very similar to operators control and prompt of [Felleisen 88], there is a significant semantical difference between shift/reset and control/prompt: the context abstracted by shift is determined statically by the program text, while control captures the context up to the nearest dynamically enclosing prompt. In general, this leads to different behavior.
The operator control is often referred to as Felleisen & Friedman’s Ƒ operator. Its operational definition leaves some room for variation, and one restricted form, also known as Ƒ-, is characterized by wrapping a prompt around all continuation applications. The effect is to close the extent of a context at the point of abstraction, instead of joining it to the context at the point of application. This is very much like building an explicit closure with FUNCTION around a LAMBDA in Lisp 1.5, which experience has shown to be preferable in general for both theoretical and practical reasons, leading to lexically scoped procedures in Scheme. In fact, Ƒ- coincides operationally with shift_1 for the simple case of one delimited context, just like static and dynamic scoping agree when the latter is used in a controlled way.
It seems that prompt and control can always simulate reset and shift, but not vice versa. SWI-Prolog’s delimited continuation is essentially prompt and control, that is nice. Perhaps prompt/3 and control/1 are more suitable names for delimited continuations in SWI-Prolog? Just for discussion.