Nightly regression: lifetime error with autoderef

[eefriedman]

Testcase:

use std::collections::HashMap;
pub fn named_lints<'a>(names: &[&str],
                       transforms: &'a HashMap<&'static str, u32>)
                       -> Option<&'a u32> {
    transforms.get(&names[0])
}
fn main(){}

<anon>:5:20: 5:29 error: cannot infer an appropriate lifetime due to conflicting requirements [E0495]
<anon>:5     transforms.get(&names[0])
                            ^~~~~~~~~
<anon>:2:1: 6:2 help: consider using an explicit lifetime parameter as shown: fn named_lints<'a>(names: &[&'a str],
                   transforms: &'a HashMap<&'static str, u32>)
 -> Option<&'a u32>
<anon>:2 pub fn named_lints<'a>(names: &[&str],
<anon>:3                        transforms: &'a HashMap<&'static str, u32>)
<anon>:4                        -> Option<&'a u32> {
<anon>:5     transforms.get(&names[0])
<anon>:6 }
error: aborting due to previous error
playpen: application terminated with error code 101

It looks rather similar to #28839, but I'm not sure whether it's the same thing. Originally reported by @jonas-schievink (#28853 (comment)) .

[alexcrichton]

triage: I-nominated

Tagging as T-compiler, but may also be T-lang (not sure). This compiles ok on stable/beta and fails to compile on nightly.

[arielb1]

This and #28853 are the same issue.

The extraneous address-of operation (to names[0] here, caps.at() in #28854) causes the Q type parameter of HashMap::get to be inferred to &'α str, with the argument type being &'β &'α str, and no deref-coercion taking place, and the The T: Borrow<T> impl is selected.

As one of the T-s is &'static str, this code compiling is suspect. The only reason this code compiles is because the type of transforms is a supertype of &'a HashMap<&'c str, u32> (where 'c is the lifetime of the strings in names), which matches with 'α = 'c. This is, obviously, unsound, as get is allowed to assume that 'c: 'a - I am not sure how bad this can be with HashMap::get because of parametricity, but it would certainly be a hole with specialization.

[arielb1]

This problem is easily fixed by removing the extraneous &.

@aturon: I think it is cases like this that led to the removal of find_equiv etc.

[eefriedman]

Oh... I see, this is kind of awkward: the signature is asserting that &'static str implements Borrow<&'a str>... which it doesn't, really. It implements Borrow<str> and Borrow<&'static str>.

This particular case is actually safe because &'static str could implement Borrow<&'a str>... the only problem is that it overlaps with impl<T> Borrow<T> for T.

[nikomatsakis]

@arielb1

This is, obviously, unsound, as get is allowed to assume that 'c: 'a - I am not sure how bad this can be with HashMap::get because of parametricity, but it would certainly be a hole with specialization.

this is presumably #25860

[arielb1]

@nikomatsakis

Borrow is covariant in practice, so this doesn't actually form a soundness hole. #25860 involves contravariance which does not occur here.

[nikomatsakis]

triage: P-high