Cyclic traits allow arbitrary traits to be synthesized

[nikomatsakis]

This issue is partially fixed through the introduction of the "obligation jungle" but we still want to impose a limitation that auto traits cannot have supertraits (for now, no supertraits at all seems easiest). This would close the remaining soundness hole described here.

This is, I believe, an unintended outcome of some caching I put in. The problem is that this test compiles fine -- I believe the expected outcome was an internal stack overflow. My preferred fix is to refactor the fulfillment context to track trees in more detail, which is something that we've started but the branch has since bitrotted. This makes me want to get back to that more urgently.

[nikomatsakis]

triage: P-high

[nikomatsakis]

(Probably) related: http://is.gd/2qIl3k (you'd want to extend this in a similar fashion to the original code snippet)

Update: weaponized form from eddyb http://is.gd/xks7Td

Update: unrelated, my mistake. This is #29861.

[eddyb]

cc @james-darkfox (who tripped on this and #29861 in his lense crate)

[WildCryptoFox]

Another simple example as follows: Magic treats anything as Copy, but this could 'imply' any traits like Send + Sync to send *mut ClearlyUnSendableThings.

trait Magic: Copy {}
impl<T: Magic> Magic for T {}

fn copy<T: Magic>(x: T) -> (T, T) { (x, x) }

#[derive(Debug)]
struct NoClone;

fn main() {
    let (a, b) = copy(NoClone);
    println!("{:?} {:?}", a, b);
}

[WildCryptoFox]

@nikomatsakis Earlier I decided to create a "safe" hexdump using this concept. :-)

https://play.rust-lang.org/?gist=0832c0210daba58942c0&version=nightly

[jnicholls]

This is a big issue. I think a fix needs to be put out for 1.4.0+ as soon as possible.

[bstrie]

Related, what's our policy on nominating soundness fixes for backporting to beta?

[arielb1]

@nikomatsakis

In my understanding of the system, the impl was supposed to cause a WF error (why do we think that Self: Trait?) but in general this kind of code was intentionally supposed to actually work (I feel like we intentionally discussed it during the associated types RFC). What is going wrong exactly?