From 7c776c83fe66c8c74929585891931b3918ccfdd8 Mon Sep 17 00:00:00 2001 From: Roxane Date: Tue, 6 Jul 2021 12:41:44 -0400 Subject: [PATCH 1/9] Update closure types documentation so it includes information about RFC2229 --- src/types/closure.md | 301 ++++++++++++++++++++++++++++++++++++------- 1 file changed, 254 insertions(+), 47 deletions(-) diff --git a/src/types/closure.md b/src/types/closure.md index eecdb038f..54993e056 100644 --- a/src/types/closure.md +++ b/src/types/closure.md @@ -2,38 +2,46 @@ A [closure expression] produces a closure value with a unique, anonymous type that cannot be written out. A closure type is approximately equivalent to a -struct which contains the captured variables. For instance, the following +struct which contains the captured values. For instance, the following closure: ```rust +#[derive(Debug)] +struct Point { x:i32, y:i32 } +struct Rectangle { left_top: Point, right_bottom: Point } + fn f String> (g: F) { println!("{}", g()); } -let mut s = String::from("foo"); -let t = String::from("bar"); - -f(|| { - s += &t; - s -}); -// Prints "foobar". +let mut rect = Rectangle { + left_top: Point { x: 1, y: 1 }, + right_bottom: Point { x: 0, y: 0 } +}; + +let c = || { + rect.left_top.x += 1; + rect.right_bottom.x += 1; + format!("{:?}", rect.left_top) +}; +// Prints "Point { x: 2, y: 1 }". ``` generates a closure type roughly like the following: - + ```rust,ignore struct Closure<'a> { - s : String, - t : &'a String, + left_top : &'a mut Point, + right_bottom_x : &'a mut i32, } impl<'a> FnOnce<()> for Closure<'a> { type Output = String; fn call_once(self) -> String { - self.s += &*self.t; - self.s + self.left_top.x += 1; + self.right_bottom_x += 1; + format!("{:?}", self.left_top) } } ``` @@ -42,48 +50,150 @@ so that the call to `f` works as if it were: ```rust,ignore -f(Closure{s: s, t: &t}); +f(Closure{ left_top: rect.left_top, right_bottom_x: rect.left_top.x }); ``` ## Capture modes -The compiler prefers to capture a closed-over variable by immutable borrow, +The compiler prefers to capture a value by immutable borrow, followed by unique immutable borrow (see below), by mutable borrow, and finally -by move. It will pick the first choice of these that is compatible with how the -captured variable is used inside the closure body. The compiler does not take -surrounding code into account, such as the lifetimes of involved variables, or -of the closure itself. +by move. It will pick the first choice of these that allows the closure to +compile. The choice is made only with regards to the contents of the closure +expression; the compiler does not take into account surrounding code, such as +the lifetimes of involved variables or fields. +>>>>>>> 881f305... Update closure types documentation so it includes information about RFC2229 -If the `move` keyword is used, then all captures are by move or, for `Copy` -types, by copy, regardless of whether a borrow would work. The `move` keyword is -usually used to allow the closure to outlive the captured values, such as if the -closure is being returned or used to spawn a new thread. +## Capture Precision + +The precise path that gets captured is typically the full path that is used in the closure, but there are cases where we will only capture a prefix of the path. + + +### Shared prefix -Composite types such as structs, tuples, and enums are always captured entirely, -not by individual fields. It may be necessary to borrow into a local variable in -order to capture a single field: +In the case where a path and one of the ancestor’s of that path are both captured by a closure, the ancestor path is captured with the highest capture mode among the two captures,`CaptureMode = max(AncestorCaptureMode, DescendantCaptureMode)`, using the strict weak ordering + +`ImmBorrow < UniqueImmBorrow < MutBorrow < ByValue`. + +Note that this might need to be applied recursively. + +```rust= +let s = String::new("S"); +let t = (s, String::new("T")); +let mut u = (t, String::new("U")); + +let c = || { + println!("{:?}", u); // u captured by ImmBorrow + u.0.truncate(0); // u.0 captured by MutBorrow + move_value(u.0.0); // u.0.0 captured by ByValue +}; +``` + +Overall the closure will capture `u` by `ByValue`. + +### Wild Card Patterns +Closures only capture data that needs to be read, which means the following closures will not capture `x` ```rust -# use std::collections::HashSet; -# -struct SetVec { - set: HashSet, - vec: Vec -} +let x = 10; +let c = || { + let _ = x; +}; + +let c = || match x { + _ => println!("Hello World!") +}; +``` -impl SetVec { - fn populate(&mut self) { - let vec = &mut self.vec; - self.set.iter().for_each(|&n| { - vec.push(n); - }) - } -} +### Capturing references in move contexts + +Rust doesn't allow moving fields out of references. As a result, in the case of move closures, when values accessed through a shared references are moved into the closure body, the compiler, instead of moving the values out of the reference, would reborrow the data. + +```rust +struct T(String, String); + +let mut t = T(String::from("foo"), String::from("bar")); +let t = &mut t; +let c = move || t.0.truncate(0); // closure captures (&mut t.0) ``` -If, instead, the closure were to use `self.vec` directly, then it would attempt -to capture `self` by mutable reference. But since `self.set` is already -borrowed to iterate over, the code would not compile. +### Raw pointer dereference +In Rust, it's `unsafe` to dereference a raw pointer. Therefore, closures will only capture the prefix of a path that runs up to, but not including, the first dereference of a raw pointer. + +```rust, +struct T(String, String); + +let t = T(String::from("foo"), String::from("bar")); +let t = &t as *const T; + +let c = || unsafe { + println!("{}", (*t).0); // closure captures t +}; +``` + +### Reference into unaligned `struct`s + +In Rust, it's `unsafe` to hold references to unaligned fields in a structure, and therefore, closures will only capture the prefix of the path that runs up to, but not including, the first field access into an unaligned structure. + +```rust +#[repr(packed)] +struct T(String, String); + +let t = T(String::from("foo"), String::from("bar")); +let c = || unsafe { + println!("{}", t.0); // closure captures t +}; +``` + + +### `Box` vs other `Deref` implementations + +The compiler treats the implementation of the Deref trait for `Box` differently, as it is considered a special entity. + +For example, let us look at examples involving `Rc` and `Box`. The `*rc` is desugared to a call to the trait method `deref` defined on `Rc`, but since `*box` is treated differently by the compiler, the compiler is able to do precise capture on contents of the `Box`. + +#### Non `move` closure + +In a non `move` closure, if the contents of the `Box` are not moved into the closure body, the contents of the `Box` are precisely captured. + +```rust +# use std::rc::Rc; + +struct S(i32); + +let b = Box::new(S(10)); +let c_box = || { + println!("{}", (*b).0); // closure captures `(*b).0` +}; + +let r = Rc::new(S(10)); +let c_rc = || { + println!("{}", (*r).0); // closure caprures `r` +}; +``` + +However, if the contents of the `Box` are moved into the closure, then the box is entirely captured. This is done so the amount of data that needs to be moved into the closure is minimized. + +```rust +struct S(i32); + +let b = Box::new(S(10)); +let c_box = || { + let x = (*b).0; // closure captures `b` +}; +``` + +#### `move` closure + +Similarly to moving contents of a `Box` in a non-`move` closure, reading the contents of a `Box` in a `move` closure will capture the `Box` entirely. + +```rust +struct S(i32); + +let b = Box::new(S(10)); +let c_box = || { + println!("{}", (*b).0); // closure captures `b` +}; +``` ## Unique immutable borrows in captures @@ -113,6 +223,7 @@ the declaration of `y` will produce an error because it would violate the uniqueness of the closure's borrow of `x`; the declaration of z is valid because the closure's lifetime has expired at the end of the block, releasing the borrow. + ## Call traits and coercions Closure types all implement [`FnOnce`], indicating that they can be called once @@ -156,12 +267,13 @@ following traits if allowed to do so by the types of the captures it stores: The rules for [`Send`] and [`Sync`] match those for normal struct types, while [`Clone`] and [`Copy`] behave as if [derived]. For [`Clone`], the order of -cloning of the captured variables is left unspecified. +cloning of the captured values is left unspecified. + Because captures are often by reference, the following general rules arise: -* A closure is [`Sync`] if all captured variables are [`Sync`]. -* A closure is [`Send`] if all variables captured by non-unique immutable +* A closure is [`Sync`] if all captured values are [`Sync`]. +* A closure is [`Send`] if all values captured by non-unique immutable reference are [`Sync`], and all values captured by unique immutable or mutable reference, copy, or move are [`Send`]. * A closure is [`Clone`] or [`Copy`] if it does not capture any values by @@ -178,3 +290,98 @@ Because captures are often by reference, the following general rules arise: [`Sync`]: ../special-types-and-traits.md#sync [closure expression]: ../expressions/closure-expr.md [derived]: ../attributes/derive.md + +## Drop Order + +If a closure captures a field of a composite types such as structs, tuples, and enums by value, the field's lifetime would now be tied to the closure. As a result, it is possible for disjoint fields of a composite types to be dropped at different times. + +```rust +{ + let tuple = + (String::from("foo"), String::from("bar")); // --+ + { // | + let c = || { // ----------------------------+ | + // tuple.0 is captured into the closure | | + drop(tuple.0); // | | + }; // | | + } // 'c' and 'tuple.0' dropped here ------------+ | +} // tuple.1 dropped here -----------------------------+ +``` + +# Edition 2018 and before + +## Closure types difference + +In Edition 2018 and before, a closure would capture variables in its entirety. This means that for the example used in the [Closure types](#closure-types) section, the generated closure type would instead look something like this: + + +```rust,ignore +struct Closure<'a> { + rect : &'a mut Rectangle, +} + +impl<'a> FnOnce<()> for Closure<'a> { + type Output = String; + fn call_once(self) -> String { + self.rect.left_top.x += 1; + self.rect.right_bottom.x += 1; + format!("{:?}", self.rect.left_top) + } +} +``` +and the call to `f` would work as follows: + +```rust,ignore +f(Closure { rect: rect }); +``` + +## Capture precision difference + +Composite types such as structs, tuples, and enums are always captured in its intirety, +not by individual fields. As a result, it may be necessary to borrow into a local variable in order to capture a single field: + +```rust +# use std::collections::HashSet; +# +struct SetVec { + set: HashSet, + vec: Vec +} + +impl SetVec { + fn populate(&mut self) { + let vec = &mut self.vec; + self.set.iter().for_each(|&n| { + vec.push(n); + }) + } +} +``` + +If, instead, the closure were to use `self.vec` directly, then it would attempt +to capture `self` by mutable reference. But since `self.set` is already +borrowed to iterate over, the code would not compile. + +If the `move` keyword is used, then all captures are by move or, for `Copy` +types, by copy, regardless of whether a borrow would work. The `move` keyword is +usually used to allow the closure to outlive the captured values, such as if the +closure is being returned or used to spawn a new thread. + +Regardless of if the data will be read by the closure, i.e. in case of wild card patterns, if a variable defined outside the closure is mentioned within the closure the variable will be captured in its entirety. + +## Drop order difference + +As composite types are captured in their entirety, a closure which captures one of those composite types by value would drop the entire captured variable at the same time as the closure gets dropped. + +```rust +{ + let tuple = + (String::from("foo"), String::from("bar")); + { + let c = || { // --------------------------+ + // tuple is captured into the closure | + drop(tuple.0); // | + }; // | + } // 'c' and 'tuple' dropped here ------------+ +} +``` From 124ece696a7a1538ede41db0047ebb8776c0dd40 Mon Sep 17 00:00:00 2001 From: Roxane Date: Thu, 8 Jul 2021 20:10:43 -0400 Subject: [PATCH 2/9] Address comments --- src/types/closure.md | 24 ++++++++++++++---------- 1 file changed, 14 insertions(+), 10 deletions(-) diff --git a/src/types/closure.md b/src/types/closure.md index 54993e056..1158b242d 100644 --- a/src/types/closure.md +++ b/src/types/closure.md @@ -76,14 +76,15 @@ In the case where a path and one of the ancestor’s of that path are both captu Note that this might need to be applied recursively. -```rust= -let s = String::new("S"); -let t = (s, String::new("T")); -let mut u = (t, String::new("U")); +```rust +# fn move_value(_: T){} +let s = String::from("S"); +let t = (s, String::from("T")); +let mut u = (t, String::from("U")); let c = || { println!("{:?}", u); // u captured by ImmBorrow - u.0.truncate(0); // u.0 captured by MutBorrow + u.1.truncate(0); // u.0 captured by MutBorrow move_value(u.0.0); // u.0.0 captured by ByValue }; ``` @@ -106,7 +107,7 @@ let c = || match x { ### Capturing references in move contexts -Rust doesn't allow moving fields out of references. As a result, in the case of move closures, when values accessed through a shared references are moved into the closure body, the compiler, instead of moving the values out of the reference, would reborrow the data. +Moving fields out of references is not allowed. As a result, in the case of move closures, when values accessed through a shared references are moved into the closure body, the compiler, instead of moving the values out of the reference, would reborrow the data. ```rust struct T(String, String); @@ -117,7 +118,7 @@ let c = move || t.0.truncate(0); // closure captures (&mut t.0) ``` ### Raw pointer dereference -In Rust, it's `unsafe` to dereference a raw pointer. Therefore, closures will only capture the prefix of a path that runs up to, but not including, the first dereference of a raw pointer. +Because it is `unsafe` to dereference a raw pointer, closures will only capture the prefix of a path that runs up to, but not including, the first dereference of a raw pointer. ```rust, struct T(String, String); @@ -132,7 +133,7 @@ let c = || unsafe { ### Reference into unaligned `struct`s -In Rust, it's `unsafe` to hold references to unaligned fields in a structure, and therefore, closures will only capture the prefix of the path that runs up to, but not including, the first field access into an unaligned structure. +Because it is `unsafe` to hold references to unaligned fields in a structure, closures will only capture the prefix of the path that runs up to, but not including, the first field access into an unaligned structure. ```rust #[repr(packed)] @@ -147,10 +148,13 @@ let c = || unsafe { ### `Box` vs other `Deref` implementations -The compiler treats the implementation of the Deref trait for `Box` differently, as it is considered a special entity. +The implementation of the [`Deref`] trait for [`Box`] is treated differently from other `Deref` implementations, as it is considered a special entity. For example, let us look at examples involving `Rc` and `Box`. The `*rc` is desugared to a call to the trait method `deref` defined on `Rc`, but since `*box` is treated differently by the compiler, the compiler is able to do precise capture on contents of the `Box`. +[`Box`]: ../special-types-and-traits.md#boxt +[`Deref`]: ../special-types-and-traits.md#deref-and-derefmut + #### Non `move` closure In a non `move` closure, if the contents of the `Box` are not moved into the closure body, the contents of the `Box` are precisely captured. @@ -337,7 +341,7 @@ f(Closure { rect: rect }); ## Capture precision difference -Composite types such as structs, tuples, and enums are always captured in its intirety, +Composite types such as structs, tuples, and enums are always captured in its entirety, not by individual fields. As a result, it may be necessary to borrow into a local variable in order to capture a single field: ```rust From fa56c013aa7b5e642b2c2f734eedd5580062cbc5 Mon Sep 17 00:00:00 2001 From: Aman Arora Date: Tue, 20 Jul 2021 04:10:29 -0400 Subject: [PATCH 3/9] Copy capture analyis alogrithm from hackmd --- src/types/closure.md | 133 +++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 133 insertions(+) diff --git a/src/types/closure.md b/src/types/closure.md index 1158b242d..915970254 100644 --- a/src/types/closure.md +++ b/src/types/closure.md @@ -312,6 +312,139 @@ If a closure captures a field of a composite types such as structs, tuples, and } // tuple.1 dropped here -----------------------------+ ``` + +## Overall Capture analysis algorithm + +* Input: + * Analyzing the closure C yields a set of `(Mode, Place)` pairs that are accessed + * Access mode is `ref`, `ref uniq`, `ref mut`, or `by-value` (ordered least to max) + * Closure mode is `ref` or `move` +* Output: + * Minimal `(Mode, Place)` pairs that are actually captured +* Cleanup and truncation + * Generate C' by mapping each (Mode, Place) in C: + * `(Mode1, Place1) = ref_opt(unsafe_check(copy_type(Mode, Place)))` + * if this is a ref closure: + * Add `ref_xform(Mode1, Place1)` to C' + * else: + * Add `move_xform(Mode1, Place1)` to C' +* Minimization + * Until no rules apply: + * For each two places (M1, P1), (M2, P2) where P1 is a prefix of P2: + * Remove both places from the set + * Add (max(M1, M2), P1) into the set +* Helper functions: + * `copy_type(Mode, Place) -> (Mode, Place)` + * "By-value use of a copy type is a ref" + * If Mode = "by-value" and type(Place) is `Copy`: + * Return (ref, Place) + * Else + * Return (Mode, Place) + * `unsafe_check(Mode, Place) -> (Mode, Place)` + * "Ensure unsafe accesses occur within the closure" + * If Place contains a deref of a raw pointer: + * Let Place1 = Place truncated just before the deref + * Return (Mode, Place1) + * If Mode is `ref *` and the place contains a field of a packed struct: + * Let Place1 = Place truncated just before the field + * Return (Mode, Place1) + * Else + * Return (Mode, Place1) + * `move_xform(Mode, Place) -> (Mode, Place)` (For move closures) + * "Take ownership if data being accessed is owned by the variable used to access it (or if closure attempts to move data that it doesn't own)." + * "When taking ownership, only capture data found on the stack." + * "Otherwise, reborrow the reference." + * If Mode is `ref mut` and the place contains a deref of an `&mut`: + * Return (Mode, Place) + * Else if Mode is `ref *` and the place contains a deref of an `&`: + * Return (Mode, Place) + * Else if place contains a deref: + * Let Place1 = Place truncated just before the deref + * Return (ByValue, Place1) + * Else: + * Return (ByValue, Place) + * `ref_xform(Mode, Place) -> (Mode, Place)` (for ref closures) + * "If taking ownership of data, only move data from enclosing stack frame." + * Generate C' by mapping each (Mode, Place) in C + * If Mode is ByValue and place contains a deref: + * Let Place1 = Place truncated just before the deref + * Return (ByValue, Place1) + * Else: + * Return (Mode, Place) + * `ref_opt(Mode, Place) -> (Mode, Place)` (for ref closures) + * "Optimization: borrow the ref, not data owned by ref." + * If Place contains a deref of an `&`... + * ...or something + +## Key examples + +### box-mut + +```rust +fn box_mut() { + let mut s = Foo { x: 0 } ; + + let px = &mut s; + let bx = Box::new(px); + + + let c = #[rustc_capture_analysis] move || bx.x += 10; + // Mutable reference to this place: + // (*(*bx)).x + // ^ ^ + // | a Box + // a &mut +} +``` + +``` +Closure mode = move +C = { + (ref mut, (*(*bx)).x) +} +C' = C +``` + +Output is the same: `C' = C` + +### Packed-field-ref-and-move + +When you have a closure that both references a packed field (which is unsafe) and moves from it (which is safe) we capture the entire struct, rather than just moving the field. This is to aid in predictability, so that removing the move doesn't make the closure become unsafe: + +```rust +print(&packed.x); +move_value(packed.x); +``` + + +```rust +struct Point { x: i32, y: i32 } +fn f(p: &Point) -> impl Fn() { + let c = move || { + let x = p.x; + }; + + // x.x -> ByValue + // after rules x -> ByValue + + c +} + +struct Point { x: i32, y: i32 } +fn g(p: &mut Point) -> impl Fn() { + let c = move || { + let x = p.x; // ought to: (ref, (*p).x) + }; + + move || { + p.y += 1; + } + + + // x.x -> ByValue + +``` + # Edition 2018 and before ## Closure types difference From 52a9af968e4b23b365c6af252365925e0968b7f2 Mon Sep 17 00:00:00 2001 From: Aman Arora Date: Tue, 20 Jul 2021 04:46:00 -0400 Subject: [PATCH 4/9] Update copy type, optimization, make algorithm (place, mode) --- src/types/closure.md | 137 ++++++++++++++++++++++++++----------------- 1 file changed, 82 insertions(+), 55 deletions(-) diff --git a/src/types/closure.md b/src/types/closure.md index 915970254..91007ec6f 100644 --- a/src/types/closure.md +++ b/src/types/closure.md @@ -316,71 +316,77 @@ If a closure captures a field of a composite types such as structs, tuples, and ## Overall Capture analysis algorithm * Input: - * Analyzing the closure C yields a set of `(Mode, Place)` pairs that are accessed + * Analyzing the closure C yields a mapping of `Place -> Mode` that are accessed * Access mode is `ref`, `ref uniq`, `ref mut`, or `by-value` (ordered least to max) + * For a `Place` that is used in two different acess modes within the same closure, the mode reported from closure analysis is the maximum access mode. + * Note: `ByValue` use of a `Copy` type is seen as a `ref` access mode. * Closure mode is `ref` or `move` * Output: - * Minimal `(Mode, Place)` pairs that are actually captured + * Minimal `(Place, Mode)` pairs that are actually captured * Cleanup and truncation * Generate C' by mapping each (Mode, Place) in C: - * `(Mode1, Place1) = ref_opt(unsafe_check(copy_type(Mode, Place)))` + * `(Place1, Mode1) = ref_opt(unsafe_check(Place, Mode))` * if this is a ref closure: - * Add `ref_xform(Mode1, Place1)` to C' + * Add `ref_xform(Place1, Mode1)` to C' * else: - * Add `move_xform(Mode1, Place1)` to C' + * Add `move_xform(Place1, Mode1)` to C' * Minimization * Until no rules apply: - * For each two places (M1, P1), (M2, P2) where P1 is a prefix of P2: + * For each two places (P1, M1), (P2, M2) where P1 is a prefix of P2: * Remove both places from the set - * Add (max(M1, M2), P1) into the set + * Add (P1, max(M1, M2)) into the set * Helper functions: - * `copy_type(Mode, Place) -> (Mode, Place)` - * "By-value use of a copy type is a ref" - * If Mode = "by-value" and type(Place) is `Copy`: - * Return (ref, Place) - * Else - * Return (Mode, Place) - * `unsafe_check(Mode, Place) -> (Mode, Place)` + * `unsafe_check(Place, Mode) -> (Place, Mode)` * "Ensure unsafe accesses occur within the closure" * If Place contains a deref of a raw pointer: * Let Place1 = Place truncated just before the deref - * Return (Mode, Place1) + * Return (Place1, Mode) * If Mode is `ref *` and the place contains a field of a packed struct: * Let Place1 = Place truncated just before the field - * Return (Mode, Place1) + * Return (Place1, Mode) * Else - * Return (Mode, Place1) - * `move_xform(Mode, Place) -> (Mode, Place)` (For move closures) + * Return (Place, Mode) + * `move_xform(Place, Mode) -> (Place, Mode)` (For move closures) * "Take ownership if data being accessed is owned by the variable used to access it (or if closure attempts to move data that it doesn't own)." * "When taking ownership, only capture data found on the stack." * "Otherwise, reborrow the reference." * If Mode is `ref mut` and the place contains a deref of an `&mut`: - * Return (Mode, Place) + * Return (Place, Mode) * Else if Mode is `ref *` and the place contains a deref of an `&`: - * Return (Mode, Place) + * Return (Place, Mode) * Else if place contains a deref: * Let Place1 = Place truncated just before the deref - * Return (ByValue, Place1) + * Return (Place1, ByValue) * Else: - * Return (ByValue, Place) - * `ref_xform(Mode, Place) -> (Mode, Place)` (for ref closures) + * Return (Place, ByValue) + * `ref_xform(Place, Mode) -> (Place, Mode)` (for ref closures) * "If taking ownership of data, only move data from enclosing stack frame." * Generate C' by mapping each (Mode, Place) in C * If Mode is ByValue and place contains a deref: * Let Place1 = Place truncated just before the deref - * Return (ByValue, Place1) + * Return (Place1, ByValue) * Else: - * Return (Mode, Place) - * `ref_opt(Mode, Place) -> (Mode, Place)` (for ref closures) + * Return (Place, Mode) + * `ref_opt(Place, Mode) -> (Place, Mode)` * "Optimization: borrow the ref, not data owned by ref." - * If Place contains a deref of an `&`... - * ...or something + * Disjoint capture over immutable reference doesn't add too much value because the fields can either be borrowed immutably or copied. + * Edge case: Field that is accessed via the referece lives longer than the reference. + * Resolution: Only consider the last Deref + * If Place is (Base, Projections), where Projections is a list of size N. + * For all `i, 0 <= i < N`, Projections[i] != Deref + * Return (Place, Mode) + * If `l, 0 <= l < N` is the last/rightmost Deref Projection i.e. for any `i, l < i < N` Projection[i] != Deref, + and `Place.type_before_projection(l) = ty::Ref(.., Mutability::Not)` + * Let Place1 = (Base, Projections[0..=l]) + * Return (Place1, Ref) ## Key examples ### box-mut ```rust +struct Foo { x: i32 } + fn box_mut() { let mut s = Foo { x: 0 } ; @@ -388,7 +394,7 @@ fn box_mut() { let bx = Box::new(px); - let c = #[rustc_capture_analysis] move || bx.x += 10; + let c = move || bx.x += 10; // Mutable reference to this place: // (*(*bx)).x // ^ ^ @@ -397,7 +403,8 @@ fn box_mut() { } ``` -``` + +```ignore Closure mode = move C = { (ref mut, (*(*bx)).x) @@ -412,37 +419,57 @@ Output is the same: `C' = C` When you have a closure that both references a packed field (which is unsafe) and moves from it (which is safe) we capture the entire struct, rather than just moving the field. This is to aid in predictability, so that removing the move doesn't make the closure become unsafe: ```rust -print(&packed.x); -move_value(packed.x); +#[repr(packed)] +struct Packed { x: String } + +# fn use_ref(_: &T) {} +# fn move_value(_: T) {} + +fn main() { + let packed = Packed { x: String::new() }; + + let c = || { + use_ref(&packed.x); + move_value(packed.x); + }; + + c(); +} +``` + + +```ignore +Closure mode = ref +C = { + (ref mut, packed) +} +C' = C ``` +### Optimization-Edge-Case +```edition2021 +struct Int(i32); +struct B<'a>(&'a i32); -```rust -struct Point { x: i32, y: i32 } -fn f(p: &Point) -> impl Fn() { - let c = move || { - let x = p.x; - }; - - // x.x -> ByValue - // after rules x -> ByValue +struct MyStruct<'a> { + a: &'static Int, + b: B<'a>, +} +fn foo<'a, 'b>(m: &'a MyStruct<'b>) -> impl FnMut() + 'static { + let c = || drop(&m.a.0); c -} +} -struct Point { x: i32, y: i32 } -fn g(p: &mut Point) -> impl Fn() { - let c = move || { - let x = p.x; // ought to: (ref, (*p).x) - }; - - move || { - p.y += 1; - } - - - // x.x -> ByValue - +``` + + +```ignore +Closure mode = ref +C = { + (ref mut, *m.a) +} +C' = C ``` # Edition 2018 and before From 7944512783684dc48cff00d7b2a8d898f67c4705 Mon Sep 17 00:00:00 2001 From: Roxane Date: Tue, 20 Jul 2021 14:41:15 -0400 Subject: [PATCH 5/9] fix stray merge conflict --- src/types/closure.md | 9 ++++----- 1 file changed, 4 insertions(+), 5 deletions(-) diff --git a/src/types/closure.md b/src/types/closure.md index 91007ec6f..f8accef83 100644 --- a/src/types/closure.md +++ b/src/types/closure.md @@ -57,11 +57,10 @@ f(Closure{ left_top: rect.left_top, right_bottom_x: rect.left_top.x }); The compiler prefers to capture a value by immutable borrow, followed by unique immutable borrow (see below), by mutable borrow, and finally -by move. It will pick the first choice of these that allows the closure to -compile. The choice is made only with regards to the contents of the closure -expression; the compiler does not take into account surrounding code, such as -the lifetimes of involved variables or fields. ->>>>>>> 881f305... Update closure types documentation so it includes information about RFC2229 +by move. It will pick the first choice of these that is compatible with how the +captured value is used inside the closure body. The compiler does not take +surrounding code into account, such as the lifetimes of involved variables or fields, or +of the closure itself. ## Capture Precision From 55c0829f2cb1d70b272562c239a4be5c85629bf4 Mon Sep 17 00:00:00 2001 From: Aman Arora Date: Mon, 2 Aug 2021 03:11:26 -0400 Subject: [PATCH 6/9] Minor edits from Feedback --- src/types/closure.md | 19 ++++++++++++------- 1 file changed, 12 insertions(+), 7 deletions(-) diff --git a/src/types/closure.md b/src/types/closure.md index f8accef83..440756735 100644 --- a/src/types/closure.md +++ b/src/types/closure.md @@ -317,7 +317,7 @@ If a closure captures a field of a composite types such as structs, tuples, and * Input: * Analyzing the closure C yields a mapping of `Place -> Mode` that are accessed * Access mode is `ref`, `ref uniq`, `ref mut`, or `by-value` (ordered least to max) - * For a `Place` that is used in two different acess modes within the same closure, the mode reported from closure analysis is the maximum access mode. + * For a `Place` that is used in two different access modes within the same closure, the mode reported from closure analysis is the maximum access mode. * Note: `ByValue` use of a `Copy` type is seen as a `ref` access mode. * Closure mode is `ref` or `move` * Output: @@ -383,6 +383,8 @@ If a closure captures a field of a composite types such as structs, tuples, and ### box-mut +This test shows how a `move` closure can sometimes capture values by mutable reference, if they are reached via a `&mut` reference. + ```rust struct Foo { x: i32 } @@ -405,10 +407,10 @@ fn box_mut() { ```ignore Closure mode = move -C = { +C_in = { (ref mut, (*(*bx)).x) } -C' = C +C_out = C_in ``` Output is the same: `C' = C` @@ -439,13 +441,16 @@ fn main() { ```ignore Closure mode = ref -C = { +C_in = { (ref mut, packed) } -C' = C +C_out = C_in ``` ### Optimization-Edge-Case + +This test shows an interesting edge case. Normally, when we see a borrow of something behind a shared reference (`&T`), we truncate to capture the entire reference, because that is more efficient (and we can always use that reference to reach all the data it refers to). However, in the case where we are dereferencing two shared references, we have to be sure to preserve the full path, since otherwise the resulting closure could have a shorter lifetime than is necessary. + ```edition2021 struct Int(i32); struct B<'a>(&'a i32); @@ -465,10 +470,10 @@ fn foo<'a, 'b>(m: &'a MyStruct<'b>) -> impl FnMut() + 'static { ```ignore Closure mode = ref -C = { +C_in = { (ref mut, *m.a) } -C' = C +C_out = C_in ``` # Edition 2018 and before From fe03349f87ec84916f3ff1cfe813ea0fd40e6700 Mon Sep 17 00:00:00 2001 From: Aman Arora Date: Mon, 2 Aug 2021 05:12:31 -0400 Subject: [PATCH 7/9] Add details on truncation and ref-uniq --- src/types/closure.md | 27 +++++++++++++++++---------- 1 file changed, 17 insertions(+), 10 deletions(-) diff --git a/src/types/closure.md b/src/types/closure.md index 440756735..c37f18f5e 100644 --- a/src/types/closure.md +++ b/src/types/closure.md @@ -337,12 +337,12 @@ If a closure captures a field of a composite types such as structs, tuples, and * Helper functions: * `unsafe_check(Place, Mode) -> (Place, Mode)` * "Ensure unsafe accesses occur within the closure" - * If Place contains a deref of a raw pointer: - * Let Place1 = Place truncated just before the deref - * Return (Place1, Mode) - * If Mode is `ref *` and the place contains a field of a packed struct: - * Let Place1 = Place truncated just before the field - * Return (Place1, Mode) + * If Place contains a deref (at index `i`) of a raw pointer: + * Let `(Place1, Mode1) = truncate_place(Place, Mode, i)` + * Return (Place1, Mode1) + * If Mode is `ref *` and the place contains a field of a packed struct at index `i`: + * Let `(Place1, Mode1) = truncate_place(Place, Mode, i)` + * Return (Place1, Mode1) * Else * Return (Place, Mode) * `move_xform(Place, Mode) -> (Place, Mode)` (For move closures) @@ -353,16 +353,16 @@ If a closure captures a field of a composite types such as structs, tuples, and * Return (Place, Mode) * Else if Mode is `ref *` and the place contains a deref of an `&`: * Return (Place, Mode) - * Else if place contains a deref: - * Let Place1 = Place truncated just before the deref + * Else if place contains a deref at index `i`: + * Let `(Place1, _) = truncate_place(Place, Mode, i)` * Return (Place1, ByValue) * Else: * Return (Place, ByValue) * `ref_xform(Place, Mode) -> (Place, Mode)` (for ref closures) * "If taking ownership of data, only move data from enclosing stack frame." * Generate C' by mapping each (Mode, Place) in C - * If Mode is ByValue and place contains a deref: - * Let Place1 = Place truncated just before the deref + * If Mode is ByValue and place contains a deref at index `i`: + * Let `(Place1, _) = truncate_place(Place, Mode, i)` * Return (Place1, ByValue) * Else: * Return (Place, Mode) @@ -378,6 +378,13 @@ If a closure captures a field of a composite types such as structs, tuples, and and `Place.type_before_projection(l) = ty::Ref(.., Mutability::Not)` * Let Place1 = (Base, Projections[0..=l]) * Return (Place1, Ref) + * `truncate_place(Place, Mode, len) -> (Place, Mode)` + * "Truncate the place to length `len`, i.e. upto but not including index `len`" + * "If during truncation we drop Deref of a `&mut` and the place was being used by `ref mut`, the access to the truncated place must be unique" + * Let (Proj_before, Proj_after) = Place.split_before(len) + * If Mode == `ref mut` and there exists `Deref` in `Proj_after` at index `i` such that `Place.type_before_projection(len + i)` is `&mut T` + * Return (Place(Proj_before, ..InputPlace), `ref-uniq`) + * Else Return (Place(Proj_before, ..InputPlace), Mode) ## Key examples From 702a71f187d74defc6a134cb8e48680d33bfb399 Mon Sep 17 00:00:00 2001 From: Aman Arora Date: Wed, 22 Sep 2021 00:54:10 -0700 Subject: [PATCH 8/9] Update closure algorithm per #88467 #88477 --- src/types/closure.md | 37 +++++++++++++++++++++---------------- 1 file changed, 21 insertions(+), 16 deletions(-) diff --git a/src/types/closure.md b/src/types/closure.md index c37f18f5e..ed5538cc5 100644 --- a/src/types/closure.md +++ b/src/types/closure.md @@ -106,14 +106,14 @@ let c = || match x { ### Capturing references in move contexts -Moving fields out of references is not allowed. As a result, in the case of move closures, when values accessed through a shared references are moved into the closure body, the compiler, instead of moving the values out of the reference, would reborrow the data. +Moving fields out of references is not allowed. As a result, in the case of move closures, when values accessed through a shared references are moved into the closure body, the compiler will truncate right before a dereference. ```rust struct T(String, String); let mut t = T(String::from("foo"), String::from("bar")); let t = &mut t; -let c = move || t.0.truncate(0); // closure captures (&mut t.0) +let c = move || t.0.truncate(0); // closure captures `t` ``` ### Raw pointer dereference @@ -325,10 +325,11 @@ If a closure captures a field of a composite types such as structs, tuples, and * Cleanup and truncation * Generate C' by mapping each (Mode, Place) in C: * `(Place1, Mode1) = ref_opt(unsafe_check(Place, Mode))` - * if this is a ref closure: - * Add `ref_xform(Place1, Mode1)` to C' + * `(Place2, Mode2)` = if this is a ref closure: + * `ref_xform(Place1, Mode1)` * else: - * Add `move_xform(Place1, Mode1)` to C' + * `move_xform(Place1, Mode1)` + * Add `(Place3, Mode3) = truncate_move_through_drop(Place2, Mode2)` to C'. * Minimization * Until no rules apply: * For each two places (P1, M1), (P2, M2) where P1 is a prefix of P2: @@ -346,18 +347,12 @@ If a closure captures a field of a composite types such as structs, tuples, and * Else * Return (Place, Mode) * `move_xform(Place, Mode) -> (Place, Mode)` (For move closures) - * "Take ownership if data being accessed is owned by the variable used to access it (or if closure attempts to move data that it doesn't own)." - * "When taking ownership, only capture data found on the stack." - * "Otherwise, reborrow the reference." - * If Mode is `ref mut` and the place contains a deref of an `&mut`: - * Return (Place, Mode) - * Else if Mode is `ref *` and the place contains a deref of an `&`: - * Return (Place, Mode) - * Else if place contains a deref at index `i`: + * If place contains a deref at index `i`: * Let `(Place1, _) = truncate_place(Place, Mode, i)` * Return (Place1, ByValue) * Else: * Return (Place, ByValue) + * Note that initially we had considered an approach where "Take ownership if data being accessed is owned by the variable used to access it (or if closure attempts to move data that it doesn't own). That is when taking ownership only capture data that is found on the stack otherwise reborrow the reference.". This cause a bug around lifetimes, check [rust-lang/rust#88431](https://github.com/rust-lang/rust/issues/88431). * `ref_xform(Place, Mode) -> (Place, Mode)` (for ref closures) * "If taking ownership of data, only move data from enclosing stack frame." * Generate C' by mapping each (Mode, Place) in C @@ -378,6 +373,16 @@ If a closure captures a field of a composite types such as structs, tuples, and and `Place.type_before_projection(l) = ty::Ref(.., Mutability::Not)` * Let Place1 = (Base, Projections[0..=l]) * Return (Place1, Ref) + * `truncate_move_through_drop(Place1, Mode1) -> (Place, Mode)` + * Rust doesn't permit moving out of a type that implements drop + * In the case where we do a disjoint capture in a move closure, we might end up trying to move out of drop type + * We truncate move of not-Copy types + * If Mode1 != ByBalue + * return (Place1, Mode1) + * If there exists `i` such that `Place1.before_projection(i): Drop` and `Place1.ty()` doesn't impl `Copy` + * then return `truncate_place(Place1, Mode1, i)` + * Else return (Place1, Mode1) + * Check [rust-lang/rust#88476](https://github.com/rust-lang/rust/issues/88476) for examples. * `truncate_place(Place, Mode, len) -> (Place, Mode)` * "Truncate the place to length `len`, i.e. upto but not including index `len`" * "If during truncation we drop Deref of a `&mut` and the place was being used by `ref mut`, the access to the truncated place must be unique" @@ -397,11 +402,11 @@ struct Foo { x: i32 } fn box_mut() { let mut s = Foo { x: 0 } ; - + let px = &mut s; let bx = Box::new(px); - - + + let c = move || bx.x += 10; // Mutable reference to this place: // (*(*bx)).x From 44f6e1e20f40dddf1b5e9aa002d5cd6e7a629663 Mon Sep 17 00:00:00 2001 From: Aman Arora Date: Fri, 19 Nov 2021 14:06:04 -0800 Subject: [PATCH 9/9] Update src/types/closure.md Co-authored-by: Josh Triplett --- src/types/closure.md | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/src/types/closure.md b/src/types/closure.md index ed5538cc5..ecdb33719 100644 --- a/src/types/closure.md +++ b/src/types/closure.md @@ -7,7 +7,7 @@ closure: ```rust #[derive(Debug)] -struct Point { x:i32, y:i32 } +struct Point { x: i32, y: i32 } struct Rectangle { left_top: Point, right_bottom: Point } fn f String> (g: F) {