Regression: rvalue temporary "does not live long enough"

[DanielKeep]

The following code compiles on stable and beta, but not nightly:

/*!
The following is a reduced test case extracted from `cargo-script`, depending on clap 1.4.5.

Fails to compile using 1.5.0-nightly (d3f497861 2015-10-18).

Compiles with rustc 1.4.0-nightly (7bf626a68 2015-09-07).

It compiles without fault on playpen under both stable and beta as of 2015-10-19.
*/

fn main() {
    // This fails to compile:
    drop(vec![&()].into_iter())

    // This *does* compile:
    // drop(vec![&()])

    // So does this:
    // drop([&()].into_iter())
}

For reference, the error produced by the above is:

   Compiling clapsplode v0.1.0 (file:///F:/Programming/Rust/sandbox/cargo-test/clapsplode)
main.rs:13:16: 13:18 error: borrowed value does not live long enough
main.rs:13     drop(vec![&()].into_iter())
                          ^~
main.rs:13:10: 13:19 note: in this expansion of vec! (defined in <std macros>)
main.rs:11:11: 20:2 note: reference must be valid for the destruction scope surrounding block at 11:10...
main.rs:11 fn main() {
main.rs:12     // This fails to compile:
main.rs:13     drop(vec![&()].into_iter())
main.rs:14
main.rs:15     // This *does* compile:
main.rs:16     // drop(vec![&()])
           ...
main.rs:11:11: 20:2 note: ...but borrowed value is only valid for the block at 11:10
main.rs:11 fn main() {
main.rs:12     // This fails to compile:
main.rs:13     drop(vec![&()].into_iter())
main.rs:14
main.rs:15     // This *does* compile:
main.rs:16     // drop(vec![&()])
           ...

The original test case was reduced down to the above thanks to eddyb and bluss. The original problematic line was m.values_of("args").unwrap_or(vec![]).into_iter().map(Into::into).collect() (clap ArgMatches → Option<Vec<&str>> → Vec<&str> → boom). Specifically, the issue was with the .into_iter() method, though the error pointed to an earlier temporary as not living long enough.

If desired, I can provide a complete, in-context example of this going awry.

bluss suggested that this was related to #29006.

[DanielKeep]

cc @brson

[pnkfelix]

Hmm I wonder if the iterators involved need to have the escape hatch for dropck added to them.

[pnkfelix]

Here is a variant (that has also regressed) that does not use the vec! macro:

fn main() {
    drop((match (Vec::new(), &()) { (mut v, b) => { v.push(b); v } })
         .into_iter())
}

[pnkfelix]

(It looks like vec::Drain may need similar treatment too, but I am having trouble actually making a concrete test illustrating the necessity there -- it seems when I call .drain instead of .into_iter in the above test case, even stable and beta complain about lifetime issues. Maybe the old dropck is catching something I'm missing in my manual analysis of vec::Drain, so I'm not going to lump that in until I see evidence that it is warranted.)

Update: Oh, duh, Drain has a lifetime bound (and a Drop impl), so of course the old dropck was going to reject it.