Rollup merge of #76099 - camelid:patch-8, r=jyn514

Add info about `!` and `impl Trait`

Fixes #76094.

@rustbot modify labels: T-doc C-enhancement

Author: tmandry

library/std/src/primitive_docs.rs

@@ -1,7 +1,6 @@
 #[doc(primitive = "bool")]
 #[doc(alias = "true")]
 #[doc(alias = "false")]
-//
 /// The boolean type.
 ///
 /// The `bool` represents a value, which could only be either `true` or `false`. If you cast
@@ -12,8 +11,8 @@
 /// `bool` implements various traits, such as [`BitAnd`], [`BitOr`], [`Not`], etc.,
 /// which allow us to perform boolean operations using `&`, `|` and `!`.
 ///
-/// `if` always demands a `bool` value. [`assert!`], being an important macro in testing,
-/// checks whether an expression returns `true`.
+/// `if` always demands a `bool` value. [`assert!`], which is an important macro in testing,
+/// checks whether an expression returns `true` and panics if it isn't.
 ///
 /// ```
 /// let bool_val = true & false | false;
@@ -194,14 +193,48 @@ mod prim_bool {}
 /// # `!` and traits
 ///
 /// When writing your own traits, `!` should have an `impl` whenever there is an obvious `impl`
-/// which doesn't `panic!`. As it turns out, most traits can have an `impl` for `!`. Take [`Debug`]
+/// which doesn't `panic!`. The reason is that functions returning an `impl Trait` where `!`
+/// does not have an `impl` of `Trait` cannot diverge as their only possible code path. In other
+/// words, they can't return `!` from every code path. As an example, this code doesn't compile:
+///
+/// ```compile_fail
+/// use core::ops::Add;
+///
+/// fn foo() -> impl Add<u32> {
+///     unimplemented!()
+/// }
+/// ```
+///
+/// But this code does:
+///
+/// ```
+/// use core::ops::Add;
+///
+/// fn foo() -> impl Add<u32> {
+///     if true {
+///         unimplemented!()
+///     } else {
+///         0
+///     }
+/// }
+/// ```
+///
+/// The reason is that, in the first example, there are many possible types that `!` could coerce
+/// to, because many types implement `Add<u32>`. However, in the second example,
+/// the `else` branch returns a `0`, which the compiler infers from the return type to be of type
+/// `u32`. Since `u32` is a concrete type, `!` can and will be coerced to it. See issue [#36375]
+/// for more information on this quirk of `!`.
+///
+/// [#36375]: https://github.com/rust-lang/rust/issues/36375
+///
+/// As it turns out, though, most traits can have an `impl` for `!`. Take [`Debug`]
 /// for example:
 ///
 /// ```
 /// #![feature(never_type)]
 /// # use std::fmt;
 /// # trait Debug {
-/// # fn fmt(&self, formatter: &mut fmt::Formatter<'_>) -> fmt::Result;
+/// #     fn fmt(&self, formatter: &mut fmt::Formatter<'_>) -> fmt::Result;
 /// # }
 /// impl Debug for ! {
 ///     fn fmt(&self, formatter: &mut fmt::Formatter<'_>) -> fmt::Result {