1 //! This file declares the `ScopeTree` type, which describes
2 //! the parent links in the region hierarchy.
4 //! For more information about how MIR-based region-checking works,
5 //! see the [rustc dev guide].
7 //! [rustc dev guide]: https://rustc-dev-guide.rust-lang.org/borrow_check.html
10 use rustc_data_structures::fx::{FxHashMap, FxIndexMap};
11 use rustc_data_structures::stable_hasher::{HashStable, StableHasher};
14 use rustc_macros::HashStable;
15 use rustc_query_system::ich::StableHashingContext;
16 use rustc_span::{Span, DUMMY_SP};
21 /// Represents a statically-describable scope that can be used to
22 /// bound the lifetime/region for values.
24 /// `Node(node_id)`: Any AST node that has any scope at all has the
25 /// `Node(node_id)` scope. Other variants represent special cases not
26 /// immediately derivable from the abstract syntax tree structure.
28 /// `DestructionScope(node_id)` represents the scope of destructors
29 /// implicitly-attached to `node_id` that run immediately after the
30 /// expression for `node_id` itself. Not every AST node carries a
31 /// `DestructionScope`, but those that are `terminating_scopes` do;
32 /// see discussion with `ScopeTree`.
34 /// `Remainder { block, statement_index }` represents
35 /// the scope of user code running immediately after the initializer
36 /// expression for the indexed statement, until the end of the block.
38 /// So: the following code can be broken down into the scopes beneath:
41 /// let a = f().g( 'b: { let x = d(); let y = d(); x.h(y) } ) ;
45 /// +---------+ (R10.)
47 /// +----------+ (M8.)
48 /// +----------------------+ (R7.)
50 /// +----------+ (M5.)
51 /// +-----------------------------------+ (M4.)
52 /// +--------------------------------------------------+ (M3.)
54 /// +-----------------------------------------------------------+ (M1.)
56 /// (M1.): Node scope of the whole `let a = ...;` statement.
57 /// (M2.): Node scope of the `f()` expression.
58 /// (M3.): Node scope of the `f().g(..)` expression.
59 /// (M4.): Node scope of the block labeled `'b:`.
60 /// (M5.): Node scope of the `let x = d();` statement
61 /// (D6.): DestructionScope for temporaries created during M5.
62 /// (R7.): Remainder scope for block `'b:`, stmt 0 (let x = ...).
63 /// (M8.): Node scope of the `let y = d();` statement.
64 /// (D9.): DestructionScope for temporaries created during M8.
65 /// (R10.): Remainder scope for block `'b:`, stmt 1 (let y = ...).
66 /// (D11.): DestructionScope for temporaries and bindings from block `'b:`.
67 /// (D12.): DestructionScope for temporaries created during M1 (e.g., f()).
70 /// Note that while the above picture shows the destruction scopes
71 /// as following their corresponding node scopes, in the internal
72 /// data structures of the compiler the destruction scopes are
73 /// represented as enclosing parents. This is sound because we use the
74 /// enclosing parent relationship just to ensure that referenced
75 /// values live long enough; phrased another way, the starting point
76 /// of each range is not really the important thing in the above
77 /// picture, but rather the ending point.
79 // FIXME(pnkfelix): this currently derives `PartialOrd` and `Ord` to
80 // placate the same deriving in `ty::FreeRegion`, but we may want to
81 // actually attach a more meaningful ordering to scopes than the one
82 // generated via deriving here.
83 #[derive(Clone, PartialEq, PartialOrd, Eq, Ord, Hash, Copy, TyEncodable, TyDecodable)]
86 pub id: hir::ItemLocalId,
90 impl fmt::Debug for Scope {
91 fn fmt(&self, fmt: &mut fmt::Formatter<'_>) -> fmt::Result {
93 ScopeData::Node => write!(fmt, "Node({:?})", self.id),
94 ScopeData::CallSite => write!(fmt, "CallSite({:?})", self.id),
95 ScopeData::Arguments => write!(fmt, "Arguments({:?})", self.id),
96 ScopeData::Destruction => write!(fmt, "Destruction({:?})", self.id),
97 ScopeData::IfThen => write!(fmt, "IfThen({:?})", self.id),
98 ScopeData::Remainder(fsi) => write!(
100 "Remainder {{ block: {:?}, first_statement_index: {}}}",
108 #[derive(Clone, PartialEq, PartialOrd, Eq, Ord, Hash, Debug, Copy, TyEncodable, TyDecodable)]
109 #[derive(HashStable)]
113 /// Scope of the call-site for a function or closure
114 /// (outlives the arguments as well as the body).
117 /// Scope of arguments passed to a function or closure
118 /// (they outlive its body).
121 /// Scope of destructors for temporaries of node-id.
124 /// Scope of the condition and then block of an if expression
125 /// Used for variables introduced in an if-let expression.
128 /// Scope following a `let id = expr;` binding in a block.
129 Remainder(FirstStatementIndex),
132 rustc_index::newtype_index! {
133 /// Represents a subscope of `block` for a binding that is introduced
134 /// by `block.stmts[first_statement_index]`. Such subscopes represent
135 /// a suffix of the block. Note that each subscope does not include
136 /// the initializer expression, if any, for the statement indexed by
137 /// `first_statement_index`.
139 /// For example, given `{ let (a, b) = EXPR_1; let c = EXPR_2; ... }`:
141 /// * The subscope with `first_statement_index == 0` is scope of both
142 /// `a` and `b`; it does not include EXPR_1, but does include
143 /// everything after that first `let`. (If you want a scope that
144 /// includes EXPR_1 as well, then do not use `Scope::Remainder`,
145 /// but instead another `Scope` that encompasses the whole block,
146 /// e.g., `Scope::Node`.
148 /// * The subscope with `first_statement_index == 1` is scope of `c`,
149 /// and thus does not include EXPR_2, but covers the `...`.
150 pub struct FirstStatementIndex {
155 // compilation error if size of `ScopeData` is not the same as a `u32`
156 static_assert_size!(ScopeData, 4);
159 /// Returns an item-local ID associated with this scope.
161 /// N.B., likely to be replaced as API is refined; e.g., pnkfelix
162 /// anticipates `fn entry_node_id` and `fn each_exit_node_id`.
163 pub fn item_local_id(&self) -> hir::ItemLocalId {
167 pub fn hir_id(&self, scope_tree: &ScopeTree) -> Option<hir::HirId> {
170 .map(|hir_id| hir::HirId { owner: hir_id.owner, local_id: self.item_local_id() })
173 /// Returns the span of this `Scope`. Note that in general the
174 /// returned span may not correspond to the span of any `NodeId` in
176 pub fn span(&self, tcx: TyCtxt<'_>, scope_tree: &ScopeTree) -> Span {
177 let Some(hir_id) = self.hir_id(scope_tree) else {
180 let span = tcx.hir().span(hir_id);
181 if let ScopeData::Remainder(first_statement_index) = self.data {
182 if let Node::Block(ref blk) = tcx.hir().get(hir_id) {
183 // Want span for scope starting after the
184 // indexed statement and ending at end of
185 // `blk`; reuse span of `blk` and shift `lo`
186 // forward to end of indexed statement.
188 // (This is the special case alluded to in the
189 // doc-comment for this method)
191 let stmt_span = blk.stmts[first_statement_index.index()].span;
193 // To avoid issues with macro-generated spans, the span
194 // of the statement must be nested in that of the block.
195 if span.lo() <= stmt_span.lo() && stmt_span.lo() <= span.hi() {
196 return span.with_lo(stmt_span.lo());
204 pub type ScopeDepth = u32;
206 /// The region scope tree encodes information about region relationships.
207 #[derive(TyEncodable, TyDecodable, Default, Debug)]
208 pub struct ScopeTree {
209 /// If not empty, this body is the root of this region hierarchy.
210 pub root_body: Option<hir::HirId>,
212 /// Maps from a scope ID to the enclosing scope id;
213 /// this is usually corresponding to the lexical nesting, though
214 /// in the case of closures the parent scope is the innermost
215 /// conditional expression or repeating block. (Note that the
216 /// enclosing scope ID for the block associated with a closure is
217 /// the closure itself.)
218 pub parent_map: FxIndexMap<Scope, (Scope, ScopeDepth)>,
220 /// Maps from a variable or binding ID to the block in which that
221 /// variable is declared.
222 var_map: FxIndexMap<hir::ItemLocalId, Scope>,
224 /// Maps from a `NodeId` to the associated destruction scope (if any).
225 destruction_scopes: FxIndexMap<hir::ItemLocalId, Scope>,
227 /// Identifies expressions which, if captured into a temporary, ought to
228 /// have a temporary whose lifetime extends to the end of the enclosing *block*,
229 /// and not the enclosing *statement*. Expressions that are not present in this
230 /// table are not rvalue candidates. The set of rvalue candidates is computed
231 /// during type check based on a traversal of the AST.
232 pub rvalue_candidates: FxHashMap<hir::HirId, RvalueCandidateType>,
234 /// If there are any `yield` nested within a scope, this map
235 /// stores the `Span` of the last one and its index in the
236 /// postorder of the Visitor traversal on the HIR.
238 /// HIR Visitor postorder indexes might seem like a peculiar
239 /// thing to care about. but it turns out that HIR bindings
240 /// and the temporary results of HIR expressions are never
241 /// storage-live at the end of HIR nodes with postorder indexes
242 /// lower than theirs, and therefore don't need to be suspended
243 /// at yield-points at these indexes.
245 /// For an example, suppose we have some code such as:
246 /// ```rust,ignore (example)
247 /// foo(f(), yield y, bar(g()))
250 /// With the HIR tree (calls numbered for expository purposes)
253 /// Call#0(foo, [Call#1(f), Yield(y), Call#2(bar, Call#3(g))])
256 /// Obviously, the result of `f()` was created before the yield
257 /// (and therefore needs to be kept valid over the yield) while
258 /// the result of `g()` occurs after the yield (and therefore
259 /// doesn't). If we want to infer that, we can look at the
260 /// postorder traversal:
262 /// `foo` `f` Call#1 `y` Yield `bar` `g` Call#3 Call#2 Call#0
265 /// In which we can easily see that `Call#1` occurs before the yield,
266 /// and `Call#3` after it.
268 /// To see that this method works, consider:
270 /// Let `D` be our binding/temporary and `U` be our other HIR node, with
271 /// `HIR-postorder(U) < HIR-postorder(D)`. Suppose, as in our example,
272 /// U is the yield and D is one of the calls.
273 /// Let's show that `D` is storage-dead at `U`.
275 /// Remember that storage-live/storage-dead refers to the state of
276 /// the *storage*, and does not consider moves/drop flags.
280 /// 1. From the ordering guarantee of HIR visitors (see
281 /// `rustc_hir::intravisit`), `D` does not dominate `U`.
283 /// 2. Therefore, `D` is *potentially* storage-dead at `U` (because
284 /// we might visit `U` without ever getting to `D`).
286 /// 3. However, we guarantee that at each HIR point, each
287 /// binding/temporary is always either always storage-live
288 /// or always storage-dead. This is what is being guaranteed
289 /// by `terminating_scopes` including all blocks where the
290 /// count of executions is not guaranteed.
292 /// 4. By `2.` and `3.`, `D` is *statically* storage-dead at `U`,
295 /// This property ought to not on (3) in an essential way -- it
296 /// is probably still correct even if we have "unrestricted" terminating
297 /// scopes. However, why use the complicated proof when a simple one
300 /// A subtle thing: `box` expressions, such as `box (&x, yield 2, &y)`. It
301 /// might seem that a `box` expression creates a `Box<T>` temporary
302 /// when it *starts* executing, at `HIR-preorder(BOX-EXPR)`. That might
303 /// be true in the MIR desugaring, but it is not important in the semantics.
305 /// The reason is that semantically, until the `box` expression returns,
306 /// the values are still owned by their containing expressions. So
307 /// we'll see that `&x`.
308 pub yield_in_scope: FxHashMap<Scope, Vec<YieldData>>,
310 /// The number of visit_expr and visit_pat calls done in the body.
311 /// Used to sanity check visit_expr/visit_pat call count when
312 /// calculating generator interiors.
313 pub body_expr_count: FxHashMap<hir::BodyId, usize>,
316 /// Identifies the reason that a given expression is an rvalue candidate
317 /// (see the `rvalue_candidates` field for more information what rvalue
318 /// candidates in general). In constants, the `lifetime` field is None
319 /// to indicate that certain expressions escape into 'static and
320 /// should have no local cleanup scope.
321 #[derive(Debug, Copy, Clone, TyEncodable, TyDecodable, HashStable)]
322 pub enum RvalueCandidateType {
323 Borrow { target: hir::ItemLocalId, lifetime: Option<Scope> },
324 Pattern { target: hir::ItemLocalId, lifetime: Option<Scope> },
327 #[derive(Debug, Copy, Clone, TyEncodable, TyDecodable, HashStable)]
328 pub struct YieldData {
329 /// The `Span` of the yield.
331 /// The number of expressions and patterns appearing before the `yield` in the body, plus one.
332 pub expr_and_pat_count: usize,
333 pub source: hir::YieldSource,
337 pub fn record_scope_parent(&mut self, child: Scope, parent: Option<(Scope, ScopeDepth)>) {
338 debug!("{:?}.parent = {:?}", child, parent);
340 if let Some(p) = parent {
341 let prev = self.parent_map.insert(child, p);
342 assert!(prev.is_none());
345 // Record the destruction scopes for later so we can query them.
346 if let ScopeData::Destruction = child.data {
347 self.destruction_scopes.insert(child.item_local_id(), child);
351 pub fn opt_destruction_scope(&self, n: hir::ItemLocalId) -> Option<Scope> {
352 self.destruction_scopes.get(&n).cloned()
355 pub fn record_var_scope(&mut self, var: hir::ItemLocalId, lifetime: Scope) {
356 debug!("record_var_scope(sub={:?}, sup={:?})", var, lifetime);
357 assert!(var != lifetime.item_local_id());
358 self.var_map.insert(var, lifetime);
361 pub fn record_rvalue_candidate(
364 candidate_type: RvalueCandidateType,
366 debug!("record_rvalue_candidate(var={var:?}, type={candidate_type:?})");
367 match &candidate_type {
368 RvalueCandidateType::Borrow { lifetime: Some(lifetime), .. }
369 | RvalueCandidateType::Pattern { lifetime: Some(lifetime), .. } => {
370 assert!(var.local_id != lifetime.item_local_id())
374 self.rvalue_candidates.insert(var, candidate_type);
377 /// Returns the narrowest scope that encloses `id`, if any.
378 pub fn opt_encl_scope(&self, id: Scope) -> Option<Scope> {
379 self.parent_map.get(&id).cloned().map(|(p, _)| p)
382 /// Returns the lifetime of the local variable `var_id`, if any.
383 pub fn var_scope(&self, var_id: hir::ItemLocalId) -> Option<Scope> {
384 self.var_map.get(&var_id).cloned()
387 /// Returns `true` if `subscope` is equal to or is lexically nested inside `superscope`, and
388 /// `false` otherwise.
391 pub fn is_subscope_of(&self, subscope: Scope, superscope: Scope) -> bool {
392 let mut s = subscope;
393 debug!("is_subscope_of({:?}, {:?})", subscope, superscope);
394 while superscope != s {
395 match self.opt_encl_scope(s) {
397 debug!("is_subscope_of({:?}, {:?}, s={:?})=false", subscope, superscope, s);
400 Some(scope) => s = scope,
404 debug!("is_subscope_of({:?}, {:?})=true", subscope, superscope);
409 /// Checks whether the given scope contains a `yield`. If so,
410 /// returns `Some(YieldData)`. If not, returns `None`.
411 pub fn yield_in_scope(&self, scope: Scope) -> Option<&[YieldData]> {
412 self.yield_in_scope.get(&scope).map(Deref::deref)
415 /// Gives the number of expressions visited in a body.
416 /// Used to sanity check visit_expr call count when
417 /// calculating generator interiors.
418 pub fn body_expr_count(&self, body_id: hir::BodyId) -> Option<usize> {
419 self.body_expr_count.get(&body_id).copied()
423 impl<'a> HashStable<StableHashingContext<'a>> for ScopeTree {
424 fn hash_stable(&self, hcx: &mut StableHashingContext<'a>, hasher: &mut StableHasher) {
430 ref destruction_scopes,
431 ref rvalue_candidates,
435 root_body.hash_stable(hcx, hasher);
436 body_expr_count.hash_stable(hcx, hasher);
437 parent_map.hash_stable(hcx, hasher);
438 var_map.hash_stable(hcx, hasher);
439 destruction_scopes.hash_stable(hcx, hasher);
440 rvalue_candidates.hash_stable(hcx, hasher);
441 yield_in_scope.hash_stable(hcx, hasher);