Part of Rust's ownership system are 'lifetimes.' Without getting into the details, they're an annotation you occasionally add to code in certain situations. Like these two:

    fn get_mut<'a>(&'a mut self) -> &'a mut T;
    fn print<'a>(s: &'a str);

The 'a parts are the lifetime annotations.

We used to have a single inference rule for when you didn't need to write the annotation. With RFC 39[1], we added two more. These two extra inference rules removed 87% of the existing annotations, giving you this instead:

    fn get_mut(&mut self) -> &mut T;
    fn print(s: &str);

This is a pretty huge readability win.

1: https://github.com/rust-lang/rfcs/blob/master/active/0039-li...

oh ok, thanks. i knew what lifetimes were, but was confused by the 'elision.' Apparently I didn't connect 'elision' with 'elide.' So, new vocab word for today :)

What kind of cases do you still need to write explicit lifetimes?

I actually haven't had to since elision was implemented, so I can't give you a good rule of thumb from experience. Here's an overview of how the rules applied to the standard library at the time they were implemented, which is where the percentage came from: https://gist.github.com/aturon/da49a6d00099fdb0e861

The most common case I have run into is storing a reference inside another data structure.

Basically, after working with lifetimes for so long we realized that a great many functions that take lifetime parameters follow a similar and straightforward pattern. As of this release, we now allow you to omit lifetime parameters entirely for functions that follow this pattern, which is likely the vast majority of functions in your code that previously required such annotations.

The detailed release notes have a more, er, detailed explanation: https://github.com/rust-lang/rust/wiki/Doc-detailed-release-...

This is a great change.

But to clarify, functions which take 2 or more reference inputs won't benefit from this, since the compiler can't determine which lifetime parameter to use, right?

Edit: Never mind, I RTFM. Turns out it can't benefit.

In cases where a function's input and output have the same lifetime, the lifetime is now inferred. So instead of this:

    fn f<'a>(x: &'a SomeType) -> &'a OtherType

you can now write this:

    fn f(x: &SomeType) -> &OtherType

Same meaning, just more concise syntax.

Uhm, what if I didn't want `&OtherType` to have the same lifetime? Would I then write:

    fn f<'a>(x: &'a SomeType) -> &'b OtherType

?

That would be illegal, because it only makes sense to return a reference when that reference is somehow linked to the lifetime of an input parameter.

(Or when it has a lifetime of `static`, which means the referent is stored in static memory and thus is alive for the whole program.)

What if you're looking up the input in a temporary cache of some kind, and returning a reference to the cached value? The lifetime of the input and output would not be related in that scenario, but the output would not necessarily be 'static (maybe you build a cache, run some functions, tear it all down, then build a new cache with different values and do it all again).

The lifetime would be matched to the lifetime of the cache, which has to be referred to somehow.

Ah cool, I still have a lot to learn about rust :)

You wouldn't use a reference generally, if you're thinking of creating a new independent object you'd just return that. This is for cases when your return object is a reference into your argument, for example

Yes, though this particular code would not compile because one of the lifetimes is not declared.

ah, got it. thanks!