NonZero lang item.

text/0000-non-zero-lang-item.md

@@ -0,0 +1,37 @@
+- Start Date: 2014-12-05
+- RFC PR: (leave this empty)
+- Rust Issue: (leave this empty)
+
+# Summary
+
+Add a `NonZero` lang item to be used to hint to the compiler that the wrapped pointer or integral value is never 0/NULL.
+
+# Motivation
+
+We currently take advantage of the fact that we know certain kinds of types either are never NULL themselves (`Box`, `&`, function pointers) or contain some pointer which we know to never be NULL (the pointer in a slice, trait object, closure). Specifically, for certain kinds of enums the contain these types, we don't add in an extra field for the discriminant but rather non-nullness of the type to encode which of two possible variants the enum may be.
+
+This is what enables `size_of::<Box<T>>() == size_of::<Option<Box<T>>>()`.
+
+Unfortunately we can't take advantage of this in library code leading to extra overhead for `Option<Vec<T>>`, `Rc<T>` and the like.
+
+# Detailed design
+
+As part of this we'd introduce a new lang item:
+```
+#[lang = "non_zero"]
+struct NonZero<T>(T);
+```
+
+Then, `trans::adt` can treat `NonZero` fields that wrap a pointer or integral type in the same manner as it currently does for `Box<T>` and the like. Thus, we'd represent such enums as `RawNullablePointer/StructWrappedNullablePointer`.
+
+In using it, creating an instance of `NonZero` would be an unsafe operation (private field, mark `::new` as unsafe). We could also implement `Deref` for convenience (i.e. for `foo: NonZero<*mut i8>`, `*foo` would return `*mut i8` since `*mut i8` is `Copy`).
+
+# Drawbacks
+
+
+
+# Alternatives
+
+We could possibly do this with an attribute instead but it's a bit more annoying to mark it as unsafe beyond just the name or hardcoding that specific attribute in rustc.
+
+# Unresolved questions