Rollup merge of #68004 - nikomatsakis:negative-impls, r=varkor

permit negative impls for non-auto traits

This is a prototype impl that extends `impl !Trait` beyond auto traits. It is not integrated with coherence or anything else, and hence only serves to prevent downstream impls (but not to allow downstream crates to rely on the absence of such impls for coherence purposes).

Fixes #66544

TODO:

- [x] need a test that you can't rely on negative impls for coherence purposes
- [x] test that negative impls cannot specialize positive ones
- [x] test that positive impls cannot specialize negative ones
- [x] extend negative impl to `Clone` in order to fully fix #66544
- [x] and maybe make `CoerceUnsized` unsafe? -- that problem is now split out into #68015
- [x] introduce feature flag and prepare a write-up
- [x] improve diagnostics?

Author: Centril

src/doc/unstable-book/src/language-features/negative-impls.md

@@ -0,0 +1,57 @@
+# `negative_impls`
+
+The tracking issue for this feature is [#68318].
+
+[#68318]: https://github.com/rust-lang/rust/issues/68318
+
+----
+
+With the feature gate `negative_impls`, you can write negative impls as well as positive ones:
+
+```rust
+#![feature(negative_impls)]
+trait DerefMut { }
+impl<T: ?Sized> !DerefMut for &T { }
+```
+
+Negative impls indicate a semver guarantee that the given trait will not be implemented for the given types. Negative impls play an additional purpose for auto traits, described below.
+
+Negative impls have the following characteristics:
+
+* They do not have any items.
+* They must obey the orphan rules as if they were a positive impl.
+* They cannot "overlap" with any positive impls.
+
+## Semver interaction
+
+It is a breaking change to remove a negative impl. Negative impls are a commitment not to implement the given trait for the named types.
+
+## Orphan and overlap rules
+
+Negative impls must obey the same orphan rules as a positive impl. This implies you cannot add a negative impl for types defined in upstream crates and so forth.
+
+Similarly, negative impls cannot overlap with positive impls, again using the same "overlap" check that we ordinarily use to determine if two impls overlap. (Note that positive impls typically cannot overlap with one another either, except as permitted by specialization.)
+
+## Interaction with auto traits
+
+Declaring a negative impl `impl !SomeAutoTrait for SomeType` for an
+auto-trait serves two purposes:
+
+* as with any trait, it declares that `SomeType` will never implement `SomeAutoTrait`;
+* it disables the automatic `SomeType: SomeAutoTrait` impl that would otherwise have been generated.
+
+Note that, at present, there is no way to indicate that a given type
+does not implement an auto trait *but that it may do so in the
+future*. For ordinary types, this is done by simply not declaring any
+impl at all, but that is not an option for auto traits. A workaround
+is that one could embed a marker type as one of the fields, where the
+marker type is `!AutoTrait`.
+
+## Immediate uses
+
+Negative impls are used to declare that `&T: !DerefMut`  and `&mut T: !Clone`, as required to fix the soundness of `Pin` described in [#66544](https://github.com/rust-lang/rust/issues/66544).
+
+This serves two purposes:
+
+* For proving the correctness of unsafe code, we can use that impl as evidence that no `DerefMut` or `Clone` impl exists.
+* It prevents downstream crates from creating such impls.

src/doc/unstable-book/src/language-features/optin-builtin-traits.md

@@ -10,7 +10,8 @@ The `optin_builtin_traits` feature gate allows you to define auto traits.
 
 Auto traits, like [`Send`] or [`Sync`] in the standard library, are marker traits
 that are automatically implemented for every type, unless the type, or a type it contains, 
-has explicitly opted out via a negative impl. 
+has explicitly opted out via a negative impl. (Negative impls are separately controlled
+by the `negative_impls` feature.)
 
 [`Send`]: https://doc.rust-lang.org/std/marker/trait.Send.html
 [`Sync`]: https://doc.rust-lang.org/std/marker/trait.Sync.html
@@ -22,6 +23,7 @@ impl !Trait for Type
 Example:
 
 ```rust
+#![feature(negative_impls)]
 #![feature(optin_builtin_traits)]
 
 auto trait Valid {}
@@ -43,3 +45,63 @@ fn main() {
     // must_be_valid( MaybeValid(False) );
 }
 ```
+
+## Automatic trait implementations
+
+When a type is declared as an `auto trait`, we will automatically
+create impls for every struct/enum/union, unless an explicit impl is
+provided. These automatic impls contain a where clause for each field
+of the form `T: AutoTrait`, where `T` is the type of the field and
+`AutoTrait` is the auto trait in question. As an example, consider the
+struct `List` and the auto trait `Send`:
+
+```rust
+struct List<T> {
+  data: T,
+  next: Option<Box<List<T>>>,
+}
+```
+
+Presuming that there is no explicit impl of `Send` for `List`, the
+compiler will supply an automatic impl of the form:
+
+```rust
+struct List<T> {
+  data: T,
+  next: Option<Box<List<T>>>,
+}
+
+unsafe impl<T> Send for List<T>
+where
+  T: Send, // from the field `data`
+  Option<Box<List<T>>>: Send, // from the field `next`
+{ }
+```
+
+Explicit impls may be either positive or negative. They take the form:
+
+```rust,ignore
+impl<...> AutoTrait for StructName<..> { }
+impl<...> !AutoTrait for StructName<..> { }
+```
+
+## Coinduction: Auto traits permit cyclic matching
+
+Unlike ordinary trait matching, auto traits are **coinductive**. This
+means, in short, that cycles which occur in trait matching are
+considered ok. As an example, consider the recursive struct `List`
+introduced in the previous section. In attempting to determine whether
+`List: Send`, we would wind up in a cycle: to apply the impl, we must
+show that `Option<Box<List>>: Send`, which will in turn require
+`Box<List>: Send` and then finally `List: Send` again. Under ordinary
+trait matching, this cycle would be an error, but for an auto trait it
+is considered a successful match.
+
+## Items
+
+Auto traits cannot have any trait items, such as methods or associated types. This ensures that we can generate default implementations.
+
+## Supertraits
+
+Auto traits cannot have supertraits. This is for soundness reasons, as the interaction of coinduction with implied bounds is difficult to reconcile.
+