Box-enclosed destructors can't safely access other boxes.

[bblum]

Ran into this issue when trying to write a destructor for the doubly-linked-list. The following code segfaults after (i believe) one of the boxes has been annihilated and the other's destructor attempts to run.

The solution, I think, is for the box annihilator to run in two phases - one phase to run destructors, and then another to actually annihilate the boxes after that. (EDIT: this is a point of contention.)

class ticky_tacky {
    let data: uint;
    let mut neighbour: @little_box;
    new(data: uint) { self.data = data; self.neighbour = @none; }
    drop {
        alt *self.neighbour {
            some(other_box) {
                #error["Neighbour's value is %u", other_box.data];
            }
            none { }
        }
    }
}

enum little_box { none, some(ticky_tacky) }

fn main() {
    let box1: @little_box = @some(ticky_tacky(42));
    let box2: @little_box = @some(ticky_tacky(31337));

    alt check *box1 {
        some(x) { x.neighbour = box2; }
    }
    alt check *box2 {
        some(x) { x.neighbour = box1; }
    }
    fail;
}

[bblum]

Related: #2047 #910

[graydon]

No, this should not compile and it's not related to those bugs. Values with dtors should not have @ boxes in them, statically. They're supposed to be acyclic, precisely so they can run top-down. This is why resources were originally their own kind. Kind system bug.

[graydon]

Sorry, that sounds yell-y, I don't mean to be unduly harsh; just want to be completely clear that "running dtors in multiple passes" is how finalizers work, not dtors, and we've made a very explicit effort to avoid that up until now. If it's not working, that's a bug, not a "let's switch to finalizers" opportunity.

[bblum]

No problem. I'm trying to come up with a way that your proposal isn't quite complete, involving allocating new @-boxes inside of destructors, but I think it's solid.

I came up with crashing code that respects your proposed kind restriction, but it just ended up being another demonstration of why TLS deserves to be unsafe. (note that this code follows the "tls is supposedly safe if you don't use polymorphic key functions" rule.)

fn key(+_x: @ticky_tacky) { }

class ticky_tacky {
    let data: uint;
    new(data: uint) { self.data = data; }
    drop unsafe {
        do task::local_data_modify(key) |box| {
            alt box {
                some(x) { #error["My own data: %u", x.data]; }
                none    { fail ~"????" }
            }
            box
        }
    }
}

fn main() {
    unsafe { task::local_data_set(key, @ticky_tacky(42)); }
    fail;
}

[bblum]

However, if we forbid destructors in things that contain @-boxes, then the doubly-linked-list cannot possibly have its own auto-clean-up function. If a dlist goes out of scope with elements still inside, they'll stick around until the cycle collector runs (or until we add true GC to the language...).

There's already a dlist::clear() function for people to use to avoid this, but it'd be nicer if it could clear itself automatically.

[bblum]

OK, I think your proposal is right. I was able to break it with three canonical "funny businesses" - global state (TLS, above), captured closure environments, and existentials, and all three of them can be ruled out with the "no box" kind.

Crashing code with a @-closure: http://pastebin.mozilla.org/1716353
And crashing code with an existential: http://pastebin.mozilla.org/1716358

One question still: Will the "no-box" kind be equivalent to "send", or will it be able to permit region pointers too?

[bblum]

these crashes still happen when the 'fail' statement is removed, which convinces me that the box annihilator is orthogonal to the problem.

[catamorphism]

I'm not totally convinced it's no longer possible to write this program, but
I tried and got stuck because I couldn't match on the @-boxes with a ref mut
pattern. So I'm going to assume this is just not allowed anymore. If you can
port the example to trunk and get the same behavior, by all means reopen!

[bblum]

struct TickyTacky {
    data: uint,
    neighbour: @mut Option<TickyTacky>,
}

impl Drop for TickyTacky {
    fn finalize(&self) {
        if (self.neighbour.is_some()) {
                let other_box = option::swap_unwrap(self.neighbour);
                fail_unless!(other_box.data == 42 || other_box.data == 31337);
        }
    }
}

fn main() {
    let box1 = @mut Some(TickyTacky { data: 42, neighbour: @mut None });
    let box2 = @mut Some(TickyTacky { data: 31337, neighbour: box1 });

    *box1 = Some(TickyTacky { data: 42, neighbour: box2 });
}