Mistaken "dropped here while still borrowed" with impl trait + match

[alecmocatta]

This manifests identically on stable and nightly without and with nll.

struct X;
impl X {
    fn my_method<'a>(&'a self) -> impl Abc + 'a {
        ()
    }
}

trait Abc {}
impl Abc for () {}

fn main() {
    let my_struct = X;
    match my_struct.my_method() {
        _ => (),
    }
    // let _ = my_struct.my_method(); // uncommenting this makes it work
}

(Playground)

Errors:

error[E0597]: `my_struct` does not live long enough
  --> src/main.rs:13:11
   |
13 |     match my_struct.my_method() {
   |           ^^^^^^^^^ borrowed value does not live long enough
...
17 | }
   | - `my_struct` dropped here while still borrowed
   |
   = note: values in a scope are dropped in the opposite order they are created

[nikomatsakis]

Curious, needs some more investigation.

[pnkfelix]

assigning to self for preliminary investigation.

[pnkfelix]

(removing NLL label since the bug is stated as occurring with and without NLL.)

[eddyb]

You don't need that specific statement, even just adding a semicolon fixes it.

[pnkfelix]

Discussed at meeting. This is similar to a number of other cases I was looking at today (namely #53528 and #53450), in that impl Trait's are assumed to potentially have a destructor, which in turn have implications for how they are treated during borrow-checking.

In our current implementation of the rvalue-lifetime and drop-check rules in particular #21114 and #46413, you can end up in scenarios where the last expression in a block needs to have a semi-colon added to it, or you get borrow-check errors.

It would be good for us to address those cases, potentially as part of longer term NLL work, but its not something we're treating as a high priority item in the short term.

Therefore, I'm closing this bug as a duplicate of #21114