Treat associated types the same as type parameters when it comes to region bounding. Fixes#20303.
Strictly speaking, this is a [breaking-change] (if you are using
associated types). You are no longer free to wantonly violate the type
system rules by closing associated types into objects without any form
of region bound. Instead you should add region bounds like `T::X :
'a`, just as you would with a normal type parameter.
r? @aturon
Fixes#17904. All the cases that I believe we should support are detailed in the test case, let me know if there is there is any more desired behavior. cc @japaric.
r? @nikomatsakis or whoever is appropriate.
While talking on IRC, someone wanted to post a link to the Rust source code, but while the lines of the rendered source code do have anchors (`<span id="[line number]">`), there is no convenient way to make links as they are not clickable. This PR makes them clickable.
Also, a minor fix of the FAQ is included.
This PR removes boxed closures from the language, the closure type syntax (`let f: |int| -> bool = /* ... */`) has been obsoleted. Move all your uses of closures to the new unboxed closure system (i.e. `Fn*` traits).
[breaking-change] patterns
- `lef f = || {}`
This binding used to type check to a boxed closure. Now that boxed closures are gone, you need to annotate the "kind" of the unboxed closure, i.e. you need pick one of these: `|&:| {}`, `|&mut:| {}` or `|:| {}`.
In the (near) future we'll have closure "kind" inference, so the compiler will infer which `Fn*` trait to use based on how the closure is used. Once this inference machinery is in place, we'll be able to remove the kind annotation from most closures.
- `type Alias<'a> = |int|:'a -> bool`
Use a trait object: `type Alias<'a> = Box<FnMut(int) -> bool + 'a>`. Use the `Fn*` trait that makes sense for your use case.
- `fn foo(&self, f: |uint| -> bool)`
In this case you can use either a trait object or an unboxed closure:
``` rust
fn foo(&self, f: F) where F: FnMut(uint) -> bool;
// or
fn foo(&self, f: Box<FnMut(uint) -> bool>);
```
- `struct Struct<'a> { f: |uint|:'a -> bool }`
Again, you can use either a trait object or an unboxed closure:
``` rust
struct Struct<F> where F: FnMut(uint) -> bool { f: F }
// or
struct Struct<'a> { f: Box<FnMut(uint) -> bool + 'a> }
```
- Using `|x, y| f(x, y)` for closure "borrows"
This comes up in recursive functions, consider the following (contrived) example:
``` rust
fn foo(x: uint, f: |uint| -> bool) -> bool {
//foo(x / 2, f) && f(x) // can't use this because `f` gets moved away in the `foo` call
foo(x / 2, |x| f(x)) && f(x) // instead "borrow" `f` in the `foo` call
}
```
If you attempt to do the same with unboxed closures you'll hit ""error: reached the recursion limit during monomorphization" (see #19596):
``` rust
fn foo<F>(x: uint, mut f: F) -> bool where F: FnMut(uint) -> bool {
foo(x / 2, |x| f(x)) && f(x)
//~^ error: reached the recursion limit during monomorphization
}
```
Instead you *should* be able to write this:
``` rust
fn foo<F>(x: uint, mut f: F) -> bool where F: FnMut(uint) -> bool {
foo(x / 2, &mut f) && f(x)
//~^ error: the trait `FnMut` is not implemented for the type `&mut F`
}
```
But as you see above `&mut F` doesn't implement the `FnMut` trait. `&mut F` *should* implement the `FnMut` and the above code *should* work, but due to a bug (see #18835) it doesn't (for now).
You can work around the issue by rewriting the function to take `&mut F` instead of `F`:
``` rust
fn foo<F>(x: uint, f: &mut F) -> bool where F: FnMut(uint) -> bool {
foo(x / 2, f) && (*f)(x)
}
```
This finally works! However writing `foo(0, &mut |x| x == 0)` is unergonomic. So you can use a private helper function to avoid this:
``` rust
// public API function
pub fn foo<F>(x: uint, mut f: F) -> bool where F: FnMut(uint) -> bool {
foo_(x, &mut f)
}
// private helper function
fn foo_<F>(x: uint, f: &mut F) -> bool where F: FnMut(uint) -> bool {
foo_(x / 2, f) && (*f)(x)
}
```
Closes#14798
---
There is more cleanup to do: like renaming functions/types from `unboxed_closure` to just `closure`, removing more dead code, simplify functions which now have unused arguments, update the documentation, etc. But that can be done in another PR.
r? @nikomatsakis @aturon (You probably want to focus on the deleted/modified tests.)
cc @eddyb