Add tail_temporaries_shorter_lifetime RFC.

Author: m-ou-se

text/3606-temporary-lifetimes-in-tail-expressions.md

@@ -0,0 +1,189 @@
+# Temporary lifetimes in tail expressions
+
+- Feature Name: `tail_temporaries_shorter_lifetime`
+- Start Date: 2023-05-04
+- RFC PR: [rust-lang/rfcs#3606](https://github.com/rust-lang/rfcs/pull/3606)
+- Rust Issue: [rust-lang/rust#0000](https://github.com/rust-lang/rust/issues/0000)
+
+# Summary
+
+In the next edition, drop temporaries in tail expressions *before* dropping locals, rather than after.
+
+![A diagram showing a function with one let statement "let x = g();" and a tail expression "temp().h()"
+and a visualisation of how long x and temp live before and after this change.
+Before: x is created first, then temp is created, then x is dropped, then temp is dropped.
+After: x is created first, then temp is created, then temp is dropped, then x is dropped.
+](https://hackmd.io/_uploads/HyVB0FtkA.svg)
+
+# Motivation
+
+Temporaries in the tail expression in a block live longer than the block itself,
+so that e.g. `{expr;}` and `{expr}` can behave very differently.
+
+For example, this fails to compile:
+
+```rust
+// This fails to compile!
+fn f() -> usize {
+    let c = RefCell::new("..");
+    c.borrow().len() // ERROR!!!
+}
+```
+
+The temporary `std::cell::Ref` created in the tail expression will be dropped
+after the local `RefCell` is dropped, resulting in a lifetime error.
+
+This leads to having to add seemingly unnecessary extra `let` statements
+or having to add seemingly unnecessary semicolons:
+
+```rust
+fn main() {
+    let c = std::cell::RefCell::new(123);
+
+    if let Ok(mut b) = c.try_borrow_mut() {
+        *b = 321;
+    }; // <-- Error if you remove the semicolon!
+}
+```
+
+Both of these examples will compile fine after the proposed change.
+
+# Guide-level explanation
+
+Temporaries are normally dropped at the end of the statement.
+
+The tail expression of a block
+(such as a function body, if/else body, match arm, block expression, etc.)
+is not a statement, so has its own rule:
+
+- Starting in Rust 2024,
+  temporaries in tail expressions are dropped after evaluating the tail expression,
+  but before dropping any local variables of the block.
+
+For example:
+
+```rust
+fn f() -> usize {
+    let c = RefCell::new("..");
+    c.borrow().len() // Ok in Rust 2024
+}
+```
+
+The `.borrow()` method returns a (temporary) `Ref` object that borrows `c`.
+Starting in Rust 2024, this will compile fine,
+because the temporary `Ref` is dropped before dropping local variable `c`.
+
+# Reference-level explanation
+
+For blocks/bodies/arms whose `{}` tokens come from Rust 2024 code,
+temporaries in the tail expression will be dropped *before* the locals of the block are dropped.
+
+# Breakage
+
+It is tricky to come up with examples that will stop compiling.
+
+For tail expressions of a function body, such code will involve a tail
+expression that injects a borrow to a temporary
+into an already existing local variable that borrows it on drop.
+
+For example:
+
+```rust
+fn why_would_you_do_this() -> bool {
+    let mut x = None;
+    // Make a temporary `RefCell` and put a `Ref` that borrows it in `x`.
+    x.replace(RefCell::new(123).borrow()).is_some()
+}
+```
+
+We expect such patterns to be very rare in real world code.
+
+For tail expressions of block expressions (and if/else bodies and match arms),
+the block could be a subexpression of a larger expression.
+In that case, dropping the (not lifetime extended) temporaries at the end of
+the block (rather than at the end of the statement) can cause subtle breakage.
+For example:
+
+```rust
+    let zero = { String::new().as_str() }.len();
+```
+
+This example compiles if the temporary `String` is kept alive until the end of
+the statement, which is what happens today without the proposed changes.
+However, it will no longer compile with the proposed changes in the next edition,
+since the temporary `String` will be dropped at the end of the block expression,
+before `.len()` is executed on the `&str` that borrows the `String`.
+
+(In this specific case, possible fixes are: removing the `{}`,
+using `()` instead of `{}`, moving the `.len()` call inside the block, or removing `.as_str()`.)
+
+Such situations are less rare than the first breakage example, but likely still uncommon.
+
+The other kind of breakage to consider is code that will still compile, but behave differently.
+However, we also expect code for which it the current drop order is critical is very rare,
+as it will involve a Drop implementation with side effects.
+
+For example:
+
+```rust
+fn f(m: &Mutex<i32>) -> i32 {
+    let _x = PanicOnDrop;
+    *m.lock().unwrap()
+}
+```
+
+This function will always panic, but will today poison the `Mutex`.
+After the proposed change, this code will still panic, but leave the mutex unpoisoned.
+(Because the mutex is unlocked *before* dropping the `PanicOnDrop`,
+which probably better matches expectations.)
+
+# Edition migration
+
+Since this is a breaking change, this should be an edition change,
+even though we expect the impact to be minimal.
+
+We need to investigate any real world cases where this change results in an observable difference.
+Depending on this investigation, we can either:
+
+- Not have any migration lint at all, or
+- Have a migration lint that warns but does not suggest new code, or
+- Have a migration lint that suggests new code for the most basic common cases (e.g. replacing `{}` by `()`), or
+- Have a migration lint that suggests new code for all cases (e.g. using explicit `let` and `drop()` statements).
+
+We highly doubt the last option is necessary.
+If it turns out to be necessary, that might be a reason to not continue with this change.
+
+# Drawbacks
+
+- It introduces another subtle difference between editions.
+  (That's kind of the point of editions, though.)
+
+- There's a very small chance this breaks existing code in a very subtle way. However, we can detect these cases and issue warnings.
+
+# Prior art
+
+- There has been an earlier attempt at changing temporary lifetimes with [RFC 66](https://rust.tf/rfc66).
+  However, it turned out to be too complicated to resolve types prematurely and
+  it introduced inconsistency when generics are involved.
+
+# Unresolved questions
+
+- How advanced should the edition migration be?
+  How uncommon is the situation where this change could existing code?
+
+# Future possibilities
+
+- Not really "future" but more "recent past":
+  Making temporary lifetime extension consistent between block expressions and
+  if/else blocks and match arms. This has already been implemented and approved:
+  https://github.com/rust-lang/rust/pull/121346
+
+- Dropping temporaries in a match scrutinee *before* the arms are evaluated,
+  rather than after, to prevent deadlocks.
+  This has been explored in depth as part of the
+  [temporary lifetimes effort](https://rust-lang.zulipchat.com/#narrow/stream/403629-t-lang.2Ftemporary-lifetimes-2024),
+  but our initial approaches didn't work out.
+  This requires more research and design.
+
+- An explicit way to make use of temporary lifetime extension. (`super let`)
+  This does not require an edition change and will be part of a separate RFC.