|
| 1 | +- Feature Name: `split_maydangle` |
| 2 | +- Start Date: 2023-02-13 |
| 3 | +- RFC PR: [rust-lang/rfcs#0000](https://github.com/rust-lang/rfcs/pull/0000) |
| 4 | +- Rust Issue: [rust-lang/rust#0000](https://github.com/rust-lang/rust/issues/0000) |
| 5 | + |
| 6 | +# Summary |
| 7 | +[summary]: #summary |
| 8 | + |
| 9 | +Add `#[needs_drop]`, ignore `PhantomData` for outlives requirements. |
| 10 | + |
| 11 | +# Motivation |
| 12 | +[motivation]: #motivation |
| 13 | + |
| 14 | +This fails to compile: |
| 15 | + |
| 16 | +```rust |
| 17 | +use core::marker::PhantomData; |
| 18 | + |
| 19 | +struct PrintOnDrop<'s>(&'s str); |
| 20 | +impl<'s> Drop for PrintOnDrop<'s> { |
| 21 | + fn drop(&mut self) { |
| 22 | + println!("{}", self.0); |
| 23 | + } |
| 24 | +} |
| 25 | + |
| 26 | +fn to_pd<T>(_: T) -> PhantomData<T> { |
| 27 | + PhantomData |
| 28 | +} |
| 29 | + |
| 30 | +pub fn foo() { |
| 31 | + let mut x; |
| 32 | + { |
| 33 | + let s = String::from("temporary"); |
| 34 | + let p = PrintOnDrop(&s); |
| 35 | + x = (to_pd(p), String::new()); |
| 36 | + } |
| 37 | +} |
| 38 | +``` |
| 39 | + |
| 40 | +And yet, this compiles: |
| 41 | + |
| 42 | +```rust |
| 43 | +use core::marker::PhantomData; |
| 44 | + |
| 45 | +struct PrintOnDrop<'s>(&'s str); |
| 46 | +impl<'s> Drop for PrintOnDrop<'s> { |
| 47 | + fn drop(&mut self) { |
| 48 | + println!("{}", self.0); |
| 49 | + } |
| 50 | +} |
| 51 | + |
| 52 | +fn to_pd<T>(_: T) -> PhantomData<T> { |
| 53 | + PhantomData |
| 54 | +} |
| 55 | + |
| 56 | +pub fn foo() { |
| 57 | + let mut x; |
| 58 | + { |
| 59 | + let s = String::from("temporary"); |
| 60 | + let p = PrintOnDrop(&s); |
| 61 | + x = (to_pd(p), ()); |
| 62 | + } |
| 63 | +} |
| 64 | +``` |
| 65 | + |
| 66 | +Since the values in the tuple are unrelated, they should not affect each other. |
| 67 | + |
| 68 | +# Guide-level explanation |
| 69 | +[guide-level-explanation]: #guide-level-explanation |
| 70 | + |
| 71 | +A type marked `#[needs_drop]` gets checked for liveness at drop. This is |
| 72 | +necessary for `Vec`: |
| 73 | + |
| 74 | +```rust |
| 75 | +struct Vec<T> { |
| 76 | + ... |
| 77 | +} |
| 78 | + |
| 79 | +unsafe impl<#[needs_drop] #[may_dangle] T> Drop for Vec<T> { |
| 80 | + fn drop(&mut self) { |
| 81 | + ... |
| 82 | + } |
| 83 | +} |
| 84 | +``` |
| 85 | + |
| 86 | +So that this compiles: |
| 87 | + |
| 88 | +```rust |
| 89 | +fn main() { |
| 90 | + let mut v = vec![]; |
| 91 | + { |
| 92 | + v.push(&String::from("temporary")); |
| 93 | + } |
| 94 | +} |
| 95 | +``` |
| 96 | + |
| 97 | +But this cannot compile, as it would be unsound: |
| 98 | + |
| 99 | +```rust |
| 100 | +struct PrintOnDrop<'s>(&'s str); |
| 101 | +impl<'s> Drop for PrintOnDrop<'s> { |
| 102 | + fn drop(&mut self) { |
| 103 | + println!("{}", self.0); |
| 104 | + } |
| 105 | +} |
| 106 | + |
| 107 | +fn main() { |
| 108 | + let mut v = vec![]; |
| 109 | + { |
| 110 | + v.push(PrintOnDrop(&*String::from("temporary"))); |
| 111 | + } |
| 112 | +} |
| 113 | +``` |
| 114 | + |
| 115 | +# Reference-level explanation |
| 116 | +[reference-level-explanation]: #reference-level-explanation |
| 117 | + |
| 118 | +This RFC removes the dropck/outlives constraints from `PhantomData` and moves |
| 119 | +them into the relevant `Drop` impls instead. |
| 120 | + |
| 121 | +# Drawbacks |
| 122 | +[drawbacks]: #drawbacks |
| 123 | + |
| 124 | +Requires mild churn to update things to the new way. Failing to update wouldn't |
| 125 | +break existing code, but would allow unsound code to compile. |
| 126 | + |
| 127 | +# Rationale and alternatives |
| 128 | +[rationale-and-alternatives]: #rationale-and-alternatives |
| 129 | + |
| 130 | +A type which doesn't need drop should never have dropck/outlives contraints, |
| 131 | +but due to the rushed way in which `may_dangle` was implemented, `PhantomData` |
| 132 | +ended up having this unfortunate behaviour. This RFC removes this behaviour and |
| 133 | +allows strictly more code to compile. |
| 134 | + |
| 135 | +# Prior art |
| 136 | +[prior-art]: #prior-art |
| 137 | + |
| 138 | +- Compiler MCP 563: It is the exact same thing as this RFC, but a full RFC |
| 139 | + seemed appropriate due to observable changes on stable, even if they are |
| 140 | + fairly obscure. |
| 141 | +- Unsound dropck elaboration for `BTreeMap`: <https://github.com/rust-lang/rust/pull/99413> |
| 142 | +- `may_dangle`: RFC 1238, RFC 1327 |
| 143 | +- This is effectively split from RFC PR 3390 and is not intended for |
| 144 | + stabilization. |
| 145 | + |
| 146 | +# Unresolved questions |
| 147 | +[unresolved-questions]: #unresolved-questions |
| 148 | + |
| 149 | +N/A |
| 150 | + |
| 151 | +# Future possibilities |
| 152 | +[future-possibilities]: #future-possibilities |
| 153 | + |
| 154 | +The full RFC 3390, and stabilization. |
0 commit comments