Anonymous lifetime inference regression

[sfackler]

The following function compiled up to today:

    fn write_messages(&mut self, messages: &[&FrontendMessage]) {
         for &message in messages.iter() {
             self.stream.write_message(message);
         }
         self.stream.flush();
     }

After updating to the from yesterday's Rust nightly to today's, it now fails to compile with the following error:

/home/sfackler/rust/src/rust-postgres/lib.rs:540:13: 540:20 error: in type `&lib::message::FrontendMessage<>`, pointer has a longer lifetime than the data it references
/home/sfackler/rust/src/rust-postgres/lib.rs:540         for &message in messages.iter() {
                                                              ^~~~~~~
/home/sfackler/rust/src/rust-postgres/lib.rs:539:64: 544:5 note: the pointer is valid for the anonymous lifetime #3 defined on the block at 539:64
/home/sfackler/rust/src/rust-postgres/lib.rs:539     fn write_messages(&mut self, messages: &[&FrontendMessage]) {
/home/sfackler/rust/src/rust-postgres/lib.rs:540         for &message in messages.iter() {
/home/sfackler/rust/src/rust-postgres/lib.rs:541             self.stream.write_message(message);
/home/sfackler/rust/src/rust-postgres/lib.rs:542         }
/home/sfackler/rust/src/rust-postgres/lib.rs:543         self.stream.flush();
/home/sfackler/rust/src/rust-postgres/lib.rs:544     }
/home/sfackler/rust/src/rust-postgres/lib.rs:539:64: 544:5 note: but the referenced data is only valid for the anonymous lifetime #4 defined on the block at 539:64
/home/sfackler/rust/src/rust-postgres/lib.rs:539     fn write_messages(&mut self, messages: &[&FrontendMessage]) {
/home/sfackler/rust/src/rust-postgres/lib.rs:540         for &message in messages.iter() {
/home/sfackler/rust/src/rust-postgres/lib.rs:541             self.stream.write_message(message);
/home/sfackler/rust/src/rust-postgres/lib.rs:542         }
/home/sfackler/rust/src/rust-postgres/lib.rs:543         self.stream.flush();
/home/sfackler/rust/src/rust-postgres/lib.rs:544     }
error: aborting due to previous error
task '<unnamed>' failed at 'explicit failure', /build/rust-git/src/rust/src/libsyntax/diagnostic.rs:101
task '<unnamed>' failed at 'receiving on closed channel', /build/rust-git/src/rust/src/libstd/rt/comm.rs:198

Possibly related to #10153?

[jdm]

cc @nikomatsakis

[nikomatsakis]

Interesting. I'm probably going to need more context to diagnose what's going on. @sfackler is it possible to create something self-contained or, failing that, point me at the repo / revision where this occurs?

[sfackler]

Here's a minimal reproduction:

enum Foo<'self> {
    V(&'self str)
}

fn f(arr: &[&Foo]) {
    for &f in arr.iter() {
        println!("{:?}", f);
    }
}

It looks like the key part is that Foo contains borrowed pointers. If I change the &'self str to ~str, it compiles.

[flaper87]

@nikomatsakis updated example:

enum Foo<'s> {
    V(&'s str)
}

fn f(arr: &[&Foo]) {
    for &f in arr.iter() {
        println!("{:?}", f);
    }
}

fn main() {}

[eddyb]

This small testcase I've found (while working on #12809) seems to indicate that we don't infer 'a <= 'b from the existence of &'a &'b T (which may be implied in the original title/description):

struct Foo<'a, 'b> {
    foo: &'a &'b int
}
fn foo<'a, 'b>(x: Foo<'a, 'b>) {
    let _y = x.foo;
}

[pnkfelix]

I spoke with @eddyb and pointed him at regions-variance-covariant-use-contravariant.rs to illustrate that we do infer 'a <= 'b from the existence of &'a &'b T.

The fact that the code above does not compile as-is is interesting. I cannot tell if that is actually related to the original bug filed here or not. I have now filed #13703 for it.

(Moved rest of comment to the issue description for #13703.)

[sfackler]

This now compiles.