in which the unused-must-use lint adapts to the world of impl-Trait

[zackmdavis]

The message names the concrete type. If we wanted it to say impl TraitName, that would be a little more work—and would also be potentially misleading, because other implementors of the trait might not be must-use. (Perhaps this is an argument for introducing must-use traits, as was floated in the thread? I guess it depends on whether we want to think of impl Trait as mere syntactic sugar (in which case I think this is fine), or if nothing about the concrete type is supposed to leak into user-visibility.)

Resolves #51560.

r? @oli-obk

[scottmcm]

I don't know if anything other than fn-level must_use can really work here, since it could be

fn x() -> impl Iterator<Item = i32> { std::fs::remove_file("/etc/passwd").unwrap(); vec![1, 2, 3].into_iter().skip(1) }

[zackmdavis]

@scottmcm Could you elaborate? I don't see how the fact that some functions could contain malicious lines of code (like the remove_file("/etc/passwd") in your example) has anything to do with this issue.

[scottmcm]

@zackmdavis Sorry, I should have picked a different filename, since the point was just to have a side-effect, not to be malicious. ensure_db_tables_exist() or similar would also make it reasonable to call the function and ignore the returned iterator.

In some ways, the iterator adapters being lazy is actually an advantage for that, since one is always free to return them after doing something useful since creating them is essentially free.

[oli-obk]

@scottmcm you should use let _ = x(); if you want to ignore an iterator.

@zackmdavis I don't like exposing the real type of an impl Trait. Not even if it's just for diagnostics. I'd be fine with treating all impl Trait returning functions as #[must_use] no matter of the real type, but that needs an RFC imo.