Passing a `Fn` to an `fn` now requires explicit lifetimes for the `Fn` and all references passed to it when invoked

[codyps]

use std::env;
use std::ffi::OsString;

/* Explicit lifetimes are required by rust-2015-05-11 nightly (newer than playpen at the moment)
 * Previously implicit lifetimes worked fine */
fn foo<'a, F>(bar: &'a F) -> Option<OsString>
    where F: Fn(&'a str) -> Option<OsString>
{
    bar("HOME")
}

fn main() {
    println!("{:?}", foo(&env::var_os));
}

Playpen: http://is.gd/LYhXZx

Might be caused by: #25212

I'm unsure if this is an actual issue that can be fixed, but figured I'd open an issue in case it was.

[mbrubeck]

Or this might be an intentional breakage from #24461.

[pnkfelix]

I don't see Box in the example, so I am skeptical that this would be #25212 ...

[bluss]

This is a regression from Rust 1.0

[huonw]

cc @nikomatsakis

[pnkfelix]

There is also this approach to working around the problem: use a for <'a> ... in the trait bound to bind the lifetime. (But doing this apparently requires eta-expanding the fn, wrapping it with a trivial closure; I'm not clear about why that is necessary...)

use std::env;
use std::ffi::OsString;

fn foo<F>(bar: &F) -> Option<OsString>
    where F: for <'r> Fn(&'r str) -> Option<OsString>
{
    bar("HOME")
}

fn main() {
    println!("{:?}", foo(&|x|env::var_os(x)));
}

[codyps]

Playpen of the failing-on-nightly-but-not-stable version, for reference:

http://is.gd/UIC7SX (playpen is now on a nightly new enough to reproduce the failure, despite the inline comment)

[brson]

Nominating because @bluss suggested this was a stable regression.

[codyps]

(Just noticed I hadn't included the error message)

<anon>:12:22: 12:25 error: type mismatch: the type `fn(_) -> core::option::Option<std::ffi::os_str::OsString> {std::env::var_os}` implements the trait `core::ops::Fn<(_,)>`, but the trait `for<'r> core::ops::Fn<(&'r str,)>` is required (expected concrete lifetime, found bound lifetime parameter ) [E0281]
<anon>:12     println!("{:?}", foo(&env::var_os));
                               ^~~

[nikomatsakis]

Hmm. I do think that this should work. If you edit the example to use a crate-local fn, for example, it does work: http://is.gd/vaEi9C

use std::env;
use std::ffi::{OsStr, OsString};

fn test<K:?Sized>(_: &K) -> Option<OsString> where K: AsRef<OsStr> { None }

/* Explicit lifetimes are required by rust-2015-05-11 nightly (newer than playpen) */
fn foo<'a, F>(bar: &'a F) -> Option<OsString>
    where F: Fn(&str) -> Option<OsString>
{
    bar("HOME")
}

fn main() {
    println!("{:?}", foo(&test));
}

[nikomatsakis]

triage: P-high

I'm going to mark this as P-high because I agree it is a regression. My guess would be that perhaps it is related to #24615 or some of the other recent changes to trait selection, but I'll have to dive in and trace it out.