Taking an argument hides lifetime error

[camelid]

This compiles without error:

fn foo(s: &i32) -> &i32 {
    &0
}

(Playground)

Errors:

   Compiling playground v0.0.1 (/playground)
warning: unused variable: `s`
 --> src/lib.rs:1:8
  |
1 | fn foo(s: &i32) -> &i32 {
  |        ^ help: if this is intentional, prefix it with an underscore: `_s`
  |
  = note: `#[warn(unused_variables)]` on by default

warning: function is never used: `foo`
 --> src/lib.rs:1:4
  |
1 | fn foo(s: &i32) -> &i32 {
  |    ^^^
  |
  = note: `#[warn(dead_code)]` on by default

warning: 2 warnings emitted

    Finished dev [unoptimized + debuginfo] target(s) in 0.73s

---

But this errors, even though there shouldn't be a difference as s is not used in the previous example:

fn foo() -> &i32 {
    &0
}

(Playground)

Errors:

   Compiling playground v0.0.1 (/playground)
error[E0106]: missing lifetime specifier
 --> src/lib.rs:1:13
  |
1 | fn foo() -> &i32 {
  |             ^ expected named lifetime parameter
  |
  = help: this function's return type contains a borrowed value, but there is no value for it to be borrowed from
help: consider using the `'static` lifetime
  |
1 | fn foo() -> &'static i32 {
  |             ^^^^^^^^

error: aborting due to previous error

For more information about this error, try `rustc --explain E0106`.
error: could not compile `playground`.

To learn more, run the command again with --verbose.

---

Potentially related to #77910.

[camelid]

To be clear, I believe the bug occurs in the first example; I think the error in the second is correct.

[jonas-schievink]

When did this regress?

[jonas-schievink]

Both cases look correct to me. The first one uses lifetime elision.

[camelid]

Just finished bisecting :)

searched nightlies: from nightly-2019-01-01 to nightly-2020-10-13
regressed nightly: nightly-2020-05-03
searched commits: from 7f65393 to f05a524
regressed commit: dae90c1

bisected with cargo-bisect-rustc v0.5.2

Host triple: x86_64-apple-darwin
Reproduce with:

cargo bisect-rustc --preserve --regress=success --start=2019-01-01 

[camelid]

Both cases look correct to me. The first one uses lifetime elision.

How so? They should be no different in the errors they produce. This is the diff between them:

-fn foo(s: &i32) -> &i32 {
+fn foo() -> &i32 {
    &0
}

[camelid]

Also the bitcode commit :/

Cc @alexcrichton

[jonas-schievink]

What was the behavior before that PR?

This documents the behavior: https://doc.rust-lang.org/reference/lifetime-elision.html?highlight=lifetime,elision#lifetime-elision-in-functions

[camelid]

I'm not sure; Godbolt is showing the same behavior throughout but bisect-rustc found a spot where it changed. Not sure what's going on...

[camelid]

Oops, this is why:

$ cargo +bisector-nightly-2020-05-02-x86_64-apple-darwin rustc
   Compiling rust-issue-77912-a v0.1.0 (/Users/user/rust-issue-77912-a)
error: unknown codegen option: `embed-bitcode`

error: could not compile `rust-issue-77912-a`

To learn more, run the command again with --verbose.

[lcnr]

I also think this behavior is expected, so I don't feel like there is a bug here

[camelid]

Sorry @alexcrichton I think that was a false positive.

[camelid]

Why would this be expected?

[camelid]

Example A takes one argument (but does not use it) and does not error, while Example B takes no arguments and does error.

[jonas-schievink]

Yeah, this is clearly working as specified in the reference (linked above), as well as here: https://rust-lang.github.io/rfcs/0141-lifetime-elision.html#examples

[lcnr]

example a is desugared to

fn foo<'a>(s: &'a i32) -> &'a i32 {
    &0
}

while we currently do not apply any desugarings in the second example. We may want to change this behavior in the future, but as of now the compiler agrees with the language specification here

[camelid]

Huh, yes I find that behavior very confusing. I guess this can be closed then. Thank you for walking me through this! It was really hurting my brain :)