Unsafe expressions

text/0000-unsafe-expressions.md

@@ -0,0 +1,100 @@
+- Feature Name: `unsafe_expr`
+- Start Date: 2015-10-29
+- RFC PR: (leave this empty)
+- Rust Issue: (leave this empty)
+
+# Summary
+
+Proposal to extend the syntax of the `unsafe` keyword to be followed by an expression
+rather than a block as it is currently defined.
+
+
+# Motivation
+
+`unsafe` is often used for single method calls, and requires one to wrap the expression
+in curly braces. It looks ugly, makes for strange-looking inline blocks, causes code
+to drift rightward in block indenting, and in general just doesn't feel very ergonomic.
+
+
+# Detailed design
+
+Let's take a look at a trivial example that makes use of unsafe:
+
+```rust
+match senders.len() {
+    1 => if unsafe { senders.get_unchecked(0).send(value).is_err() } {
+        // ...
+    },
+    _ => {},
+}
+```
+
+The inline braces are awkward and feel strange from a formatting perspective.
+Now, suppose `unsafe` accepted an expression rather than a block.
+
+```rust
+match senders.len() {
+    1 => if unsafe senders.get_unchecked(0).send(value).is_err() {
+        // ...
+    },
+    _ => {},
+}
+```
+
+We've removed the strange punctuation, and all works as usual. This change
+would be backwards-compatible, as it will allow all existing Rust code to compile.
+
+## Operator Precedence
+
+The `unsafe` keyword would have similar precedence to any other prefix unary operator.
+As a demonstration, consider the following expressions equivalent...
+
+- `unsafe a.m().n()` => `unsafe { a.m().n() }`
+- `unsafe a[i].foo()` => `unsafe { a[i].foo() }`
+- `unsafe a[i].foo() + 1` => `unsafe { a[i].foo() } + 1`
+- `(unsafe a.()).n()` => `unsafe { a.m() }.n()`
+
+We have another similar (unstable) keyword in the language already: `box`. It would
+make sense to tie these two together with the same precedence in the language.
+
+One modification to the above may be to give it the precedence of the `||` closure
+operator, which would make it more greedy:
+
+`unsafe a[i].foo() + 1` => `unsafe { a[i].foo() + 1 }`
+
+# Drawbacks
+
+Less friction to the ergonomics of using `unsafe` is not necessarily a good thing,
+we certainly don't want to encourage its use. I believe this change would still
+be beneficial in general, however.
+
+This may encourage `unsafe` to be more coarse-grained or greedy, rather than
+making the scope of the `unsafe` explicit, one needs to be aware of its operator
+precedence as discussed above. However, I've found the "minimal unsafe" style to not
+be common in the greater rust ecosystem, which more often means the example
+may be written as such:
+
+```rust
+unsafe {
+    match senders.len() {
+        1 => if senders.get_unchecked(0).send(value).is_err() {
+            // ...
+        },
+        _ => {},
+    }
+}
+```
+
+... where much more of the code than is necessary is covered by `unsafe` while the
+unrelated `_ => {}` block would normally not want it at all. Note that this issue
+exists in other contexts as well, such as `unsafe` functions themselves having their
+body implicitly be `unsafe`.
+
+
+# Unresolved questions
+
+- Is there any reason this proposal will not work from a grammar or syntax perspective?
+- Are there any strange interactions with the language here that haven't been addressed?
+- Does this conflict with or complicate parsing due to the `unsafe fn x()` declaration syntax?
+- See the discussion at [rust-lang/rust#21192](https://github.com/rust-lang/rust/issues/21192)
+  for opinions on the exact precedence of keyword unary operators.