NLL breaks functions with intermingled lifetimes in return type when returned value is a (non-borrowed) associated type

[daboross]

Code which returns this kind of complicated expression involving two different lifetimes compiles under normal compilation conditions, but fails with #![feature(nll)].

Minified example (courtesy of @matthewjasper).

#![feature(nll)]
trait Visitor<'de> {
    type Value;
}

impl<'a, 'de: 'a> Visitor<'de> for &'a () {
    type Value = ();
}

//error: free region `'a` does not outlive free region `'de`
fn visit_seq<'de: 'a, 'a>() -> <&'a () as Visitor<'de>>::Value {}
//                                                             ^^

fn main() {}

working playground: https://play.rust-lang.org/?gist=5f08095fda551daa89fadf0b8566989e&version=nightly
broken playground: https://play.rust-lang.org/?gist=a5d0d7619f610d94fe316c80b0f1dfbb&version=nightly

Original post / serde_derive generated code example

The source code is anything like the following:

#![feature(nll)]
extern crate sede;
#[macro_use]
extern crate serde_derive;

#[derive(Deserialize)]
struct X<'a> {
    #[serde(borrow)]
    field: &'a [u8]
}

Pared down generated code:

#![feature(nll)]
extern crate serde;
pub struct X<'a> {
    pub field: &'a [u8],
}
impl<'de: 'a, 'a> serde::Deserialize<'de> for X<'a> {
    fn deserialize<__D>(__deserializer: __D) -> serde::export::Result<Self, __D::Error>
    where
        __D: serde::Deserializer<'de>,
    {
        struct __Visitor<'de: 'a, 'a> {
            marker: serde::export::PhantomData<X<'a>>,
            lifetime: serde::export::PhantomData<&'de ()>,
        }
        impl<'de: 'a, 'a> serde::de::Visitor<'de> for __Visitor<'de, 'a> {
            type Value = X<'a>;
            fn expecting(
                &self,
                formatter: &mut serde::export::Formatter,
            ) -> serde::export::fmt::Result {
                serde::export::Formatter::write_str(formatter, "struct X")
            }
            #[inline]
            fn visit_seq<__A>(
                self,
                mut __seq: __A,
            ) -> serde::export::Result<Self::Value, __A::Error>
            where
                __A: serde::de::SeqAccess<'de>,
            {
                let __field0 = match serde::de::SeqAccess::next_element::<&'a [u8]>(&mut __seq)? {
                    Some(__value) => __value,
                    None => {
                        return Err(serde::de::Error::invalid_length(
                            0usize,
                            &"tuple of 1 elements",
                        ));
                    }
                };
                Ok(X { field: __field0 })
            }
        }
        unimplemented!()
    }
}
fn main() {}

This code compiles without NLL enabled, but with NLL, fails with:

error: free region `'a` does not outlive free region `'de`
  --> src/main.rs:31:21
   |
31 |                 let __field0 = match serde::de::SeqAccess::next_element::<&'a [u8]>(&mut __seq)? {
   |                     ^^^^^^^^

Playground working: https://play.rust-lang.org/?gist=cb213e15287e06e85fb7a55412967b23&version=nightly

Playground failing: https://play.rust-lang.org/?gist=c4745b7f9e40f16bd7fe62f27d9c534b&version=nightly

[matthewjasper]

Minified.

trait Visitor<'de> {
    type Value;
}

impl<'a, 'de: 'a> Visitor<'de> for &'a () {
    type Value = ();
}

//error: free region `'a` does not outlive free region `'de`
fn visit_seq<'de: 'a, 'a>() -> <&'a () as Visitor<'de>>::Value {}
//                                                             ^^

fn main() {}

[daboross]

This was broken sometime in between nightly-2018-03-07 and nightly-2018-03-15.

There don't seem to be any nightly builds available between those two dates, I don't think I can narrow it down further on my computer.

[davidtwco]

I'll take a go at this issue.

[davidtwco]

An update after a little bit of digging, I've bisected it down to 8c4ff22 (ie. the PR #48411).

[davidtwco]

An update on my progress with this: I've got the logging for the some recent commit near HEAD and a commit before the regression was introduced.

#48411 was a pretty large change and there's no clear place (to my untrained eyes) where it might have introduced the regression. I'm struggling to make heads of tails of what exactly is happening. I've got the logging in a gist.

[daboross]

Thank you for taking this on!

If there's anything I can do to help, I can try. I'm not well-versed in compiler-internals, though, so I probably won't be able to help much.

[nikomatsakis]

@davidtwco ah indeed that is a big PR =) I guess it's not surprising that it has a bit of fallout. I imagine I'll have to dig a bit deeper. I wonder if we can find a more minimal example.

[nikomatsakis]

I wonder if #49824 is related.

[davidtwco]

I'm going to unassign myself from this issue. Not got a lot of time over the next month or so and don't want this priority issue to be held up.