Helmut Eller asked what this example should return?
(let ((events '())) (guard (c (#t #f)) (guard (c ((dynamic-wind (lambda () (set! events (cons 'c events))) (lambda () #f) (lambda () (set! events (cons 'd events)))) #f)) (dynamic-wind (lambda () (set! events (cons 'a events))) (lambda () (raise 'error)) (lambda () (set! events (cons 'b events)))))) (reverse events))Is it (a b c d a b) or (a c d b) or (a b c d) or unspecified?
Aaron Hsu replied:
The important parts here are the dynamic extent in which the cond-clauses are evaluated, and the dynamic extent of the implicit raise that occurs if none of the clauses fire. The extent/continuation of the cond evaluation is that of the whole guard, whereas the re-raise is that of the original raise.
This means that the first raise will trigger the A and B setters, and then the C and D setters will trigger. At this point, since the result is #f, the implementation should re-rais the object from the original calling extent, thus triggering A and B setters again, before finally returning without re-entering again.
This gives (a b c d a b) as the only valid result.
My tests:
We do indeed get (a b c d a b) from Chez, Ikarus, Vicare, Larceny, Ypsilon, Mosh, Chibi.
However, IronScheme, Racket (in #lang r6rs mode), STklos all return (a b c d). In the case of IronScheme at least, that is because it supports escape continuations only, and so the outer continuation in guard cannot re-enter the dynamic extent of the body.
SigScheme returns (a c d b), apparently evaluating the test of the cond-clause in the dynamic environment of raise and unwinding the stack only when the test returns true. That's arguably "better" that (a b c d) as this will call other handlers in the correct environment if the test returns #f.
The other Schemes all report errors, typically about guard or raise being undefined, or that (#t #f) is not a valid procedure call.