//! This file builds up the `ScopeTree`, which describes //! the parent links in the region hierarchy. //! //! For more information about how MIR-based region-checking works, //! see the [rustc guide]. //! //! [rustc guide]: https://rust-lang.github.io/rustc-guide/mir/borrowck.html use crate::ich::{StableHashingContext, NodeIdHashingMode}; use crate::util::nodemap::{FxHashMap, FxHashSet}; use crate::ty; use std::mem; use std::fmt; use rustc_macros::HashStable; use syntax::source_map; use syntax::ast; use syntax_pos::{Span, DUMMY_SP}; use crate::ty::{DefIdTree, TyCtxt}; use crate::ty::query::Providers; use crate::hir; use crate::hir::Node; use crate::hir::def_id::DefId; use crate::hir::intravisit::{self, Visitor, NestedVisitorMap}; use crate::hir::{Block, Arm, Pat, PatKind, Stmt, Expr, Local}; use rustc_data_structures::indexed_vec::Idx; use rustc_data_structures::stable_hasher::{HashStable, StableHasher, StableHasherResult}; /// Scope represents a statically-describable scope that can be /// used to bound the lifetime/region for values. /// /// `Node(node_id)`: Any AST node that has any scope at all has the /// `Node(node_id)` scope. Other variants represent special cases not /// immediately derivable from the abstract syntax tree structure. /// /// `DestructionScope(node_id)` represents the scope of destructors /// implicitly-attached to `node_id` that run immediately after the /// expression for `node_id` itself. Not every AST node carries a /// `DestructionScope`, but those that are `terminating_scopes` do; /// see discussion with `ScopeTree`. /// /// `Remainder { block, statement_index }` represents /// the scope of user code running immediately after the initializer /// expression for the indexed statement, until the end of the block. /// /// So: the following code can be broken down into the scopes beneath: /// /// ```text /// let a = f().g( 'b: { let x = d(); let y = d(); x.h(y) } ) ; /// /// +-+ (D12.) /// +-+ (D11.) /// +---------+ (R10.) /// +-+ (D9.) /// +----------+ (M8.) /// +----------------------+ (R7.) /// +-+ (D6.) /// +----------+ (M5.) /// +-----------------------------------+ (M4.) /// +--------------------------------------------------+ (M3.) /// +--+ (M2.) /// +-----------------------------------------------------------+ (M1.) /// /// (M1.): Node scope of the whole `let a = ...;` statement. /// (M2.): Node scope of the `f()` expression. /// (M3.): Node scope of the `f().g(..)` expression. /// (M4.): Node scope of the block labeled `'b:`. /// (M5.): Node scope of the `let x = d();` statement /// (D6.): DestructionScope for temporaries created during M5. /// (R7.): Remainder scope for block `'b:`, stmt 0 (let x = ...). /// (M8.): Node scope of the `let y = d();` statement. /// (D9.): DestructionScope for temporaries created during M8. /// (R10.): Remainder scope for block `'b:`, stmt 1 (let y = ...). /// (D11.): DestructionScope for temporaries and bindings from block `'b:`. /// (D12.): DestructionScope for temporaries created during M1 (e.g., f()). /// ``` /// /// Note that while the above picture shows the destruction scopes /// as following their corresponding node scopes, in the internal /// data structures of the compiler the destruction scopes are /// represented as enclosing parents. This is sound because we use the /// enclosing parent relationship just to ensure that referenced /// values live long enough; phrased another way, the starting point /// of each range is not really the important thing in the above /// picture, but rather the ending point. // // FIXME(pnkfelix): this currently derives `PartialOrd` and `Ord` to // placate the same deriving in `ty::FreeRegion`, but we may want to // actually attach a more meaningful ordering to scopes than the one // generated via deriving here. #[derive(Clone, PartialEq, PartialOrd, Eq, Ord, Hash, Copy, RustcEncodable, RustcDecodable, HashStable)] pub struct Scope { pub id: hir::ItemLocalId, pub data: ScopeData, } impl fmt::Debug for Scope { fn fmt(&self, fmt: &mut fmt::Formatter<'_>) -> fmt::Result { match self.data { ScopeData::Node => write!(fmt, "Node({:?})", self.id), ScopeData::CallSite => write!(fmt, "CallSite({:?})", self.id), ScopeData::Arguments => write!(fmt, "Arguments({:?})", self.id), ScopeData::Destruction => write!(fmt, "Destruction({:?})", self.id), ScopeData::Remainder(fsi) => write!( fmt, "Remainder {{ block: {:?}, first_statement_index: {}}}", self.id, fsi.as_u32(), ), } } } #[derive(Clone, PartialEq, PartialOrd, Eq, Ord, Hash, Debug, Copy, RustcEncodable, RustcDecodable, HashStable)] pub enum ScopeData { Node, // Scope of the call-site for a function or closure // (outlives the arguments as well as the body). CallSite, // Scope of arguments passed to a function or closure // (they outlive its body). Arguments, // Scope of destructors for temporaries of node-id. Destruction, // Scope following a `let id = expr;` binding in a block. Remainder(FirstStatementIndex) } newtype_index! { /// Represents a subscope of `block` for a binding that is introduced /// by `block.stmts[first_statement_index]`. Such subscopes represent /// a suffix of the block. Note that each subscope does not include /// the initializer expression, if any, for the statement indexed by /// `first_statement_index`. /// /// For example, given `{ let (a, b) = EXPR_1; let c = EXPR_2; ... }`: /// /// * The subscope with `first_statement_index == 0` is scope of both /// `a` and `b`; it does not include EXPR_1, but does include /// everything after that first `let`. (If you want a scope that /// includes EXPR_1 as well, then do not use `Scope::Remainder`, /// but instead another `Scope` that encompasses the whole block, /// e.g., `Scope::Node`. /// /// * The subscope with `first_statement_index == 1` is scope of `c`, /// and thus does not include EXPR_2, but covers the `...`. pub struct FirstStatementIndex { .. } } impl_stable_hash_for!(struct crate::middle::region::FirstStatementIndex { private }); // compilation error if size of `ScopeData` is not the same as a `u32` static_assert!(ASSERT_SCOPE_DATA: mem::size_of::() == 4); impl Scope { /// Returns a item-local ID associated with this scope. /// /// N.B., likely to be replaced as API is refined; e.g., pnkfelix /// anticipates `fn entry_node_id` and `fn each_exit_node_id`. pub fn item_local_id(&self) -> hir::ItemLocalId { self.id } pub fn node_id(&self, tcx: TyCtxt<'_, '_, '_>, scope_tree: &ScopeTree) -> ast::NodeId { match scope_tree.root_body { Some(hir_id) => { tcx.hir().hir_to_node_id(hir::HirId { owner: hir_id.owner, local_id: self.item_local_id() }) } None => ast::DUMMY_NODE_ID } } /// Returns the span of this `Scope`. Note that in general the /// returned span may not correspond to the span of any `NodeId` in /// the AST. pub fn span(&self, tcx: TyCtxt<'_, '_, '_>, scope_tree: &ScopeTree) -> Span { let node_id = self.node_id(tcx, scope_tree); if node_id == ast::DUMMY_NODE_ID { return DUMMY_SP; } let span = tcx.hir().span(node_id); if let ScopeData::Remainder(first_statement_index) = self.data { if let Node::Block(ref blk) = tcx.hir().get(node_id) { // Want span for scope starting after the // indexed statement and ending at end of // `blk`; reuse span of `blk` and shift `lo` // forward to end of indexed statement. // // (This is the special case aluded to in the // doc-comment for this method) let stmt_span = blk.stmts[first_statement_index.index()].span; // To avoid issues with macro-generated spans, the span // of the statement must be nested in that of the block. if span.lo() <= stmt_span.lo() && stmt_span.lo() <= span.hi() { return Span::new(stmt_span.lo(), span.hi(), span.ctxt()); } } } span } } pub type ScopeDepth = u32; /// The region scope tree encodes information about region relationships. #[derive(Default, Debug)] pub struct ScopeTree { /// If not empty, this body is the root of this region hierarchy. root_body: Option, /// The parent of the root body owner, if the latter is an /// an associated const or method, as impls/traits can also /// have lifetime parameters free in this body. root_parent: Option, /// `parent_map` maps from a scope ID to the enclosing scope id; /// this is usually corresponding to the lexical nesting, though /// in the case of closures the parent scope is the innermost /// conditional expression or repeating block. (Note that the /// enclosing scope ID for the block associated with a closure is /// the closure itself.) parent_map: FxHashMap, /// `var_map` maps from a variable or binding ID to the block in /// which that variable is declared. var_map: FxHashMap, /// maps from a `NodeId` to the associated destruction scope (if any) destruction_scopes: FxHashMap, /// `rvalue_scopes` includes entries for those expressions whose cleanup scope is /// larger than the default. The map goes from the expression id /// to the cleanup scope id. For rvalues not present in this /// table, the appropriate cleanup scope is the innermost /// enclosing statement, conditional expression, or repeating /// block (see `terminating_scopes`). /// In constants, None is used to indicate that certain expressions /// escape into 'static and should have no local cleanup scope. rvalue_scopes: FxHashMap>, /// Encodes the hierarchy of fn bodies. Every fn body (including /// closures) forms its own distinct region hierarchy, rooted in /// the block that is the fn body. This map points from the ID of /// that root block to the ID of the root block for the enclosing /// fn, if any. Thus the map structures the fn bodies into a /// hierarchy based on their lexical mapping. This is used to /// handle the relationships between regions in a fn and in a /// closure defined by that fn. See the "Modeling closures" /// section of the README in infer::region_constraints for /// more details. closure_tree: FxHashMap, /// If there are any `yield` nested within a scope, this map /// stores the `Span` of the last one and its index in the /// postorder of the Visitor traversal on the HIR. /// /// HIR Visitor postorder indexes might seem like a peculiar /// thing to care about. but it turns out that HIR bindings /// and the temporary results of HIR expressions are never /// storage-live at the end of HIR nodes with postorder indexes /// lower than theirs, and therefore don't need to be suspended /// at yield-points at these indexes. /// /// For an example, suppose we have some code such as: /// ```rust,ignore (example) /// foo(f(), yield y, bar(g())) /// ``` /// /// With the HIR tree (calls numbered for expository purposes) /// ``` /// Call#0(foo, [Call#1(f), Yield(y), Call#2(bar, Call#3(g))]) /// ``` /// /// Obviously, the result of `f()` was created before the yield /// (and therefore needs to be kept valid over the yield) while /// the result of `g()` occurs after the yield (and therefore /// doesn't). If we want to infer that, we can look at the /// postorder traversal: /// ```plain,ignore /// `foo` `f` Call#1 `y` Yield `bar` `g` Call#3 Call#2 Call#0 /// ``` /// /// In which we can easily see that `Call#1` occurs before the yield, /// and `Call#3` after it. /// /// To see that this method works, consider: /// /// Let `D` be our binding/temporary and `U` be our other HIR node, with /// `HIR-postorder(U) < HIR-postorder(D)` (in our example, U would be /// the yield and D would be one of the calls). Let's show that /// `D` is storage-dead at `U`. /// /// Remember that storage-live/storage-dead refers to the state of /// the *storage*, and does not consider moves/drop flags. /// /// Then: /// 1. From the ordering guarantee of HIR visitors (see /// `rustc::hir::intravisit`), `D` does not dominate `U`. /// 2. Therefore, `D` is *potentially* storage-dead at `U` (because /// we might visit `U` without ever getting to `D`). /// 3. However, we guarantee that at each HIR point, each /// binding/temporary is always either always storage-live /// or always storage-dead. This is what is being guaranteed /// by `terminating_scopes` including all blocks where the /// count of executions is not guaranteed. /// 4. By `2.` and `3.`, `D` is *statically* storage-dead at `U`, /// QED. /// /// I don't think this property relies on `3.` in an essential way - it /// is probably still correct even if we have "unrestricted" terminating /// scopes. However, why use the complicated proof when a simple one /// works? /// /// A subtle thing: `box` expressions, such as `box (&x, yield 2, &y)`. It /// might seem that a `box` expression creates a `Box` temporary /// when it *starts* executing, at `HIR-preorder(BOX-EXPR)`. That might /// be true in the MIR desugaring, but it is not important in the semantics. /// /// The reason is that semantically, until the `box` expression returns, /// the values are still owned by their containing expressions. So /// we'll see that `&x`. yield_in_scope: FxHashMap, /// The number of visit_expr and visit_pat calls done in the body. /// Used to sanity check visit_expr/visit_pat call count when /// calculating generator interiors. body_expr_count: FxHashMap, } #[derive(Debug, Copy, Clone)] pub struct Context { /// the root of the current region tree. This is typically the id /// of the innermost fn body. Each fn forms its own disjoint tree /// in the region hierarchy. These fn bodies are themselves /// arranged into a tree. See the "Modeling closures" section of /// the README in infer::region_constraints for more /// details. root_id: Option, /// The scope that contains any new variables declared, plus its depth in /// the scope tree. var_parent: Option<(Scope, ScopeDepth)>, /// Region parent of expressions, etc., plus its depth in the scope tree. parent: Option<(Scope, ScopeDepth)>, } struct RegionResolutionVisitor<'a, 'tcx: 'a> { tcx: TyCtxt<'a, 'tcx, 'tcx>, // The number of expressions and patterns visited in the current body expr_and_pat_count: usize, // Generated scope tree: scope_tree: ScopeTree, cx: Context, /// `terminating_scopes` is a set containing the ids of each /// statement, or conditional/repeating expression. These scopes /// are calling "terminating scopes" because, when attempting to /// find the scope of a temporary, by default we search up the /// enclosing scopes until we encounter the terminating scope. A /// conditional/repeating expression is one which is not /// guaranteed to execute exactly once upon entering the parent /// scope. This could be because the expression only executes /// conditionally, such as the expression `b` in `a && b`, or /// because the expression may execute many times, such as a loop /// body. The reason that we distinguish such expressions is that, /// upon exiting the parent scope, we cannot statically know how /// many times the expression executed, and thus if the expression /// creates temporaries we cannot know statically how many such /// temporaries we would have to cleanup. Therefore, we ensure that /// the temporaries never outlast the conditional/repeating /// expression, preventing the need for dynamic checks and/or /// arbitrary amounts of stack space. Terminating scopes end /// up being contained in a DestructionScope that contains the /// destructor's execution. terminating_scopes: FxHashSet, } struct ExprLocatorVisitor { hir_id: hir::HirId, result: Option, expr_and_pat_count: usize, } // This visitor has to have the same visit_expr calls as RegionResolutionVisitor // since `expr_count` is compared against the results there. impl<'tcx> Visitor<'tcx> for ExprLocatorVisitor { fn nested_visit_map<'this>(&'this mut self) -> NestedVisitorMap<'this, 'tcx> { NestedVisitorMap::None } fn visit_pat(&mut self, pat: &'tcx Pat) { intravisit::walk_pat(self, pat); self.expr_and_pat_count += 1; if pat.hir_id == self.hir_id { self.result = Some(self.expr_and_pat_count); } } fn visit_expr(&mut self, expr: &'tcx Expr) { debug!("ExprLocatorVisitor - pre-increment {} expr = {:?}", self.expr_and_pat_count, expr); intravisit::walk_expr(self, expr); self.expr_and_pat_count += 1; debug!("ExprLocatorVisitor - post-increment {} expr = {:?}", self.expr_and_pat_count, expr); if expr.hir_id == self.hir_id { self.result = Some(self.expr_and_pat_count); } } } impl<'tcx> ScopeTree { pub fn record_scope_parent(&mut self, child: Scope, parent: Option<(Scope, ScopeDepth)>) { debug!("{:?}.parent = {:?}", child, parent); if let Some(p) = parent { let prev = self.parent_map.insert(child, p); assert!(prev.is_none()); } // record the destruction scopes for later so we can query them if let ScopeData::Destruction = child.data { self.destruction_scopes.insert(child.item_local_id(), child); } } pub fn each_encl_scope(&self, mut e: E) where E: FnMut(Scope, Scope) { for (&child, &parent) in &self.parent_map { e(child, parent.0) } } pub fn each_var_scope(&self, mut e: E) where E: FnMut(&hir::ItemLocalId, Scope) { for (child, &parent) in self.var_map.iter() { e(child, parent) } } pub fn opt_destruction_scope(&self, n: hir::ItemLocalId) -> Option { self.destruction_scopes.get(&n).cloned() } /// Records that `sub_closure` is defined within `sup_closure`. These ids /// should be the ID of the block that is the fn body, which is /// also the root of the region hierarchy for that fn. fn record_closure_parent(&mut self, sub_closure: hir::ItemLocalId, sup_closure: hir::ItemLocalId) { debug!("record_closure_parent(sub_closure={:?}, sup_closure={:?})", sub_closure, sup_closure); assert!(sub_closure != sup_closure); let previous = self.closure_tree.insert(sub_closure, sup_closure); assert!(previous.is_none()); } fn record_var_scope(&mut self, var: hir::ItemLocalId, lifetime: Scope) { debug!("record_var_scope(sub={:?}, sup={:?})", var, lifetime); assert!(var != lifetime.item_local_id()); self.var_map.insert(var, lifetime); } fn record_rvalue_scope(&mut self, var: hir::ItemLocalId, lifetime: Option) { debug!("record_rvalue_scope(sub={:?}, sup={:?})", var, lifetime); if let Some(lifetime) = lifetime { assert!(var != lifetime.item_local_id()); } self.rvalue_scopes.insert(var, lifetime); } pub fn opt_encl_scope(&self, id: Scope) -> Option { //! Returns the narrowest scope that encloses `id`, if any. self.parent_map.get(&id).cloned().map(|(p, _)| p) } #[allow(dead_code)] // used in cfg pub fn encl_scope(&self, id: Scope) -> Scope { //! Returns the narrowest scope that encloses `id`, if any. self.opt_encl_scope(id).unwrap() } /// Returns the lifetime of the local variable `var_id` pub fn var_scope(&self, var_id: hir::ItemLocalId) -> Scope { self.var_map.get(&var_id).cloned().unwrap_or_else(|| bug!("no enclosing scope for id {:?}", var_id)) } pub fn temporary_scope(&self, expr_id: hir::ItemLocalId) -> Option { //! Returns the scope when temp created by expr_id will be cleaned up // check for a designated rvalue scope if let Some(&s) = self.rvalue_scopes.get(&expr_id) { debug!("temporary_scope({:?}) = {:?} [custom]", expr_id, s); return s; } // else, locate the innermost terminating scope // if there's one. Static items, for instance, won't // have an enclosing scope, hence no scope will be // returned. let mut id = Scope { id: expr_id, data: ScopeData::Node }; while let Some(&(p, _)) = self.parent_map.get(&id) { match p.data { ScopeData::Destruction => { debug!("temporary_scope({:?}) = {:?} [enclosing]", expr_id, id); return Some(id); } _ => id = p } } debug!("temporary_scope({:?}) = None", expr_id); return None; } pub fn var_region(&self, id: hir::ItemLocalId) -> ty::RegionKind { //! Returns the lifetime of the variable `id`. let scope = ty::ReScope(self.var_scope(id)); debug!("var_region({:?}) = {:?}", id, scope); scope } pub fn scopes_intersect(&self, scope1: Scope, scope2: Scope) -> bool { self.is_subscope_of(scope1, scope2) || self.is_subscope_of(scope2, scope1) } /// Returns `true` if `subscope` is equal to or is lexically nested inside `superscope`, and /// `false` otherwise. pub fn is_subscope_of(&self, subscope: Scope, superscope: Scope) -> bool { let mut s = subscope; debug!("is_subscope_of({:?}, {:?})", subscope, superscope); while superscope != s { match self.opt_encl_scope(s) { None => { debug!("is_subscope_of({:?}, {:?}, s={:?})=false", subscope, superscope, s); return false; } Some(scope) => s = scope } } debug!("is_subscope_of({:?}, {:?})=true", subscope, superscope); return true; } /// Returns the ID of the innermost containing body pub fn containing_body(&self, mut scope: Scope) -> Option { loop { if let ScopeData::CallSite = scope.data { return Some(scope.item_local_id()); } scope = self.opt_encl_scope(scope)?; } } /// Finds the nearest common ancestor of two scopes. That is, finds the /// smallest scope which is greater than or equal to both `scope_a` and /// `scope_b`. pub fn nearest_common_ancestor(&self, scope_a: Scope, scope_b: Scope) -> Scope { if scope_a == scope_b { return scope_a; } let mut a = scope_a; let mut b = scope_b; // Get the depth of each scope's parent. If either scope has no parent, // it must be the root, which means we can stop immediately because the // root must be the nearest common ancestor. (In practice, this is // moderately common.) let (parent_a, parent_a_depth) = match self.parent_map.get(&a) { Some(pd) => *pd, None => return a, }; let (parent_b, parent_b_depth) = match self.parent_map.get(&b) { Some(pd) => *pd, None => return b, }; if parent_a_depth > parent_b_depth { // `a` is lower than `b`. Move `a` up until it's at the same depth // as `b`. The first move up is trivial because we already found // `parent_a` above; the loop does the remaining N-1 moves. a = parent_a; for _ in 0..(parent_a_depth - parent_b_depth - 1) { a = self.parent_map.get(&a).unwrap().0; } } else if parent_b_depth > parent_a_depth { // `b` is lower than `a`. b = parent_b; for _ in 0..(parent_b_depth - parent_a_depth - 1) { b = self.parent_map.get(&b).unwrap().0; } } else { // Both scopes are at the same depth, and we know they're not equal // because that case was tested for at the top of this function. So // we can trivially move them both up one level now. assert!(parent_a_depth != 0); a = parent_a; b = parent_b; } // Now both scopes are at the same level. We move upwards in lockstep // until they match. In practice, this loop is almost always executed // zero times because `a` is almost always a direct ancestor of `b` or // vice versa. while a != b { a = self.parent_map.get(&a).unwrap().0; b = self.parent_map.get(&b).unwrap().0; }; a } /// Assuming that the provided region was defined within this `ScopeTree`, /// returns the outermost `Scope` that the region outlives. pub fn early_free_scope<'a, 'gcx>(&self, tcx: TyCtxt<'a, 'gcx, 'tcx>, br: &ty::EarlyBoundRegion) -> Scope { let param_owner = tcx.parent(br.def_id).unwrap(); let param_owner_id = tcx.hir().as_local_hir_id(param_owner).unwrap(); let scope = tcx.hir().maybe_body_owned_by_by_hir_id(param_owner_id).map(|body_id| { tcx.hir().body(body_id).value.hir_id.local_id }).unwrap_or_else(|| { // The lifetime was defined on node that doesn't own a body, // which in practice can only mean a trait or an impl, that // is the parent of a method, and that is enforced below. assert_eq!(Some(param_owner_id), self.root_parent, "free_scope: {:?} not recognized by the \ region scope tree for {:?} / {:?}", param_owner, self.root_parent.map(|id| tcx.hir().local_def_id_from_hir_id(id)), self.root_body.map(|hir_id| DefId::local(hir_id.owner))); // The trait/impl lifetime is in scope for the method's body. self.root_body.unwrap().local_id }); Scope { id: scope, data: ScopeData::CallSite } } /// Assuming that the provided region was defined within this `ScopeTree`, /// returns the outermost `Scope` that the region outlives. pub fn free_scope<'a, 'gcx>(&self, tcx: TyCtxt<'a, 'gcx, 'tcx>, fr: &ty::FreeRegion) -> Scope { let param_owner = match fr.bound_region { ty::BoundRegion::BrNamed(def_id, _) => { tcx.parent(def_id).unwrap() } _ => fr.scope }; // Ensure that the named late-bound lifetimes were defined // on the same function that they ended up being freed in. assert_eq!(param_owner, fr.scope); let param_owner_id = tcx.hir().as_local_hir_id(param_owner).unwrap(); let body_id = tcx.hir().body_owned_by(param_owner_id); Scope { id: tcx.hir().body(body_id).value.hir_id.local_id, data: ScopeData::CallSite } } /// Checks whether the given scope contains a `yield`. If so, /// returns `Some((span, expr_count))` with the span of a yield we found and /// the number of expressions and patterns appearing before the `yield` in the body + 1. /// If there a are multiple yields in a scope, the one with the highest number is returned. pub fn yield_in_scope(&self, scope: Scope) -> Option<(Span, usize)> { self.yield_in_scope.get(&scope).cloned() } /// Checks whether the given scope contains a `yield` and if that yield could execute /// after `expr`. If so, it returns the span of that `yield`. /// `scope` must be inside the body. pub fn yield_in_scope_for_expr(&self, scope: Scope, expr_hir_id: hir::HirId, body: &'tcx hir::Body) -> Option { self.yield_in_scope(scope).and_then(|(span, count)| { let mut visitor = ExprLocatorVisitor { hir_id: expr_hir_id, result: None, expr_and_pat_count: 0, }; visitor.visit_body(body); if count >= visitor.result.unwrap() { Some(span) } else { None } }) } /// Gives the number of expressions visited in a body. /// Used to sanity check visit_expr call count when /// calculating generator interiors. pub fn body_expr_count(&self, body_id: hir::BodyId) -> Option { self.body_expr_count.get(&body_id).map(|r| *r) } } /// Records the lifetime of a local variable as `cx.var_parent` fn record_var_lifetime(visitor: &mut RegionResolutionVisitor<'_, '_>, var_id: hir::ItemLocalId, _sp: Span) { match visitor.cx.var_parent { None => { // this can happen in extern fn declarations like // // extern fn isalnum(c: c_int) -> c_int } Some((parent_scope, _)) => visitor.scope_tree.record_var_scope(var_id, parent_scope), } } fn resolve_block<'a, 'tcx>(visitor: &mut RegionResolutionVisitor<'a, 'tcx>, blk: &'tcx hir::Block) { debug!("resolve_block(blk.hir_id={:?})", blk.hir_id); let prev_cx = visitor.cx; // We treat the tail expression in the block (if any) somewhat // differently from the statements. The issue has to do with // temporary lifetimes. Consider the following: // // quux({ // let inner = ... (&bar()) ...; // // (... (&foo()) ...) // (the tail expression) // }, other_argument()); // // Each of the statements within the block is a terminating // scope, and thus a temporary (e.g., the result of calling // `bar()` in the initializer expression for `let inner = ...;`) // will be cleaned up immediately after its corresponding // statement (i.e., `let inner = ...;`) executes. // // On the other hand, temporaries associated with evaluating the // tail expression for the block are assigned lifetimes so that // they will be cleaned up as part of the terminating scope // *surrounding* the block expression. Here, the terminating // scope for the block expression is the `quux(..)` call; so // those temporaries will only be cleaned up *after* both // `other_argument()` has run and also the call to `quux(..)` // itself has returned. visitor.enter_node_scope_with_dtor(blk.hir_id.local_id); visitor.cx.var_parent = visitor.cx.parent; { // This block should be kept approximately in sync with // `intravisit::walk_block`. (We manually walk the block, rather // than call `walk_block`, in order to maintain precise // index information.) for (i, statement) in blk.stmts.iter().enumerate() { match statement.node { hir::StmtKind::Local(..) | hir::StmtKind::Item(..) => { // Each declaration introduces a subscope for bindings // introduced by the declaration; this subscope covers a // suffix of the block. Each subscope in a block has the // previous subscope in the block as a parent, except for // the first such subscope, which has the block itself as a // parent. visitor.enter_scope( Scope { id: blk.hir_id.local_id, data: ScopeData::Remainder(FirstStatementIndex::new(i)) } ); visitor.cx.var_parent = visitor.cx.parent; } hir::StmtKind::Expr(..) | hir::StmtKind::Semi(..) => {} } visitor.visit_stmt(statement) } walk_list!(visitor, visit_expr, &blk.expr); } visitor.cx = prev_cx; } fn resolve_arm<'a, 'tcx>(visitor: &mut RegionResolutionVisitor<'a, 'tcx>, arm: &'tcx hir::Arm) { visitor.terminating_scopes.insert(arm.body.hir_id.local_id); if let Some(hir::Guard::If(ref expr)) = arm.guard { visitor.terminating_scopes.insert(expr.hir_id.local_id); } intravisit::walk_arm(visitor, arm); } fn resolve_pat<'a, 'tcx>(visitor: &mut RegionResolutionVisitor<'a, 'tcx>, pat: &'tcx hir::Pat) { visitor.record_child_scope(Scope { id: pat.hir_id.local_id, data: ScopeData::Node }); // If this is a binding then record the lifetime of that binding. if let PatKind::Binding(..) = pat.node { record_var_lifetime(visitor, pat.hir_id.local_id, pat.span); } debug!("resolve_pat - pre-increment {} pat = {:?}", visitor.expr_and_pat_count, pat); intravisit::walk_pat(visitor, pat); visitor.expr_and_pat_count += 1; debug!("resolve_pat - post-increment {} pat = {:?}", visitor.expr_and_pat_count, pat); } fn resolve_stmt<'a, 'tcx>(visitor: &mut RegionResolutionVisitor<'a, 'tcx>, stmt: &'tcx hir::Stmt) { let stmt_id = stmt.hir_id.local_id; debug!("resolve_stmt(stmt.id={:?})", stmt_id); // Every statement will clean up the temporaries created during // execution of that statement. Therefore each statement has an // associated destruction scope that represents the scope of the // statement plus its destructors, and thus the scope for which // regions referenced by the destructors need to survive. visitor.terminating_scopes.insert(stmt_id); let prev_parent = visitor.cx.parent; visitor.enter_node_scope_with_dtor(stmt_id); intravisit::walk_stmt(visitor, stmt); visitor.cx.parent = prev_parent; } fn resolve_expr<'a, 'tcx>(visitor: &mut RegionResolutionVisitor<'a, 'tcx>, expr: &'tcx hir::Expr) { debug!("resolve_expr - pre-increment {} expr = {:?}", visitor.expr_and_pat_count, expr); let prev_cx = visitor.cx; visitor.enter_node_scope_with_dtor(expr.hir_id.local_id); { let terminating_scopes = &mut visitor.terminating_scopes; let mut terminating = |id: hir::ItemLocalId| { terminating_scopes.insert(id); }; match expr.node { // Conditional or repeating scopes are always terminating // scopes, meaning that temporaries cannot outlive them. // This ensures fixed size stacks. hir::ExprKind::Binary( source_map::Spanned { node: hir::BinOpKind::And, .. }, _, ref r) | hir::ExprKind::Binary( source_map::Spanned { node: hir::BinOpKind::Or, .. }, _, ref r) => { // For shortcircuiting operators, mark the RHS as a terminating // scope since it only executes conditionally. terminating(r.hir_id.local_id); } hir::ExprKind::If(ref expr, ref then, Some(ref otherwise)) => { terminating(expr.hir_id.local_id); terminating(then.hir_id.local_id); terminating(otherwise.hir_id.local_id); } hir::ExprKind::If(ref expr, ref then, None) => { terminating(expr.hir_id.local_id); terminating(then.hir_id.local_id); } hir::ExprKind::Loop(ref body, _, _) => { terminating(body.hir_id.local_id); } hir::ExprKind::While(ref expr, ref body, _) => { terminating(expr.hir_id.local_id); terminating(body.hir_id.local_id); } hir::ExprKind::Match(..) => { visitor.cx.var_parent = visitor.cx.parent; } hir::ExprKind::Use(ref expr) => { // `Use(expr)` does not denote a conditional scope. // Rather, we want to achieve the same behavior as `{ let _t = expr; _t }`. terminating(expr.hir_id.local_id); } hir::ExprKind::AssignOp(..) | hir::ExprKind::Index(..) | hir::ExprKind::Unary(..) | hir::ExprKind::Call(..) | hir::ExprKind::MethodCall(..) => { // FIXME(https://github.com/rust-lang/rfcs/issues/811) Nested method calls // // The lifetimes for a call or method call look as follows: // // call.id // - arg0.id // - ... // - argN.id // - call.callee_id // // The idea is that call.callee_id represents *the time when // the invoked function is actually running* and call.id // represents *the time to prepare the arguments and make the // call*. See the section "Borrows in Calls" borrowck/README.md // for an extended explanation of why this distinction is // important. // // record_superlifetime(new_cx, expr.callee_id); } _ => {} } } match expr.node { // Manually recurse over closures, because they are the only // case of nested bodies that share the parent environment. hir::ExprKind::Closure(.., body, _, _) => { let body = visitor.tcx.hir().body(body); visitor.visit_body(body); } _ => intravisit::walk_expr(visitor, expr) } visitor.expr_and_pat_count += 1; debug!("resolve_expr post-increment {}, expr = {:?}", visitor.expr_and_pat_count, expr); if let hir::ExprKind::Yield(..) = expr.node { // Mark this expr's scope and all parent scopes as containing `yield`. let mut scope = Scope { id: expr.hir_id.local_id, data: ScopeData::Node }; loop { visitor.scope_tree.yield_in_scope.insert(scope, (expr.span, visitor.expr_and_pat_count)); // Keep traversing up while we can. match visitor.scope_tree.parent_map.get(&scope) { // Don't cross from closure bodies to their parent. Some(&(superscope, _)) => match superscope.data { ScopeData::CallSite => break, _ => scope = superscope }, None => break } } } visitor.cx = prev_cx; } fn resolve_local<'a, 'tcx>(visitor: &mut RegionResolutionVisitor<'a, 'tcx>, pat: Option<&'tcx hir::Pat>, init: Option<&'tcx hir::Expr>) { debug!("resolve_local(pat={:?}, init={:?})", pat, init); let blk_scope = visitor.cx.var_parent.map(|(p, _)| p); // As an exception to the normal rules governing temporary // lifetimes, initializers in a let have a temporary lifetime // of the enclosing block. This means that e.g., a program // like the following is legal: // // let ref x = HashMap::new(); // // Because the hash map will be freed in the enclosing block. // // We express the rules more formally based on 3 grammars (defined // fully in the helpers below that implement them): // // 1. `E&`, which matches expressions like `&` that // own a pointer into the stack. // // 2. `P&`, which matches patterns like `ref x` or `(ref x, ref // y)` that produce ref bindings into the value they are // matched against or something (at least partially) owned by // the value they are matched against. (By partially owned, // I mean that creating a binding into a ref-counted or managed value // would still count.) // // 3. `ET`, which matches both rvalues like `foo()` as well as places // based on rvalues like `foo().x[2].y`. // // A subexpression `` that appears in a let initializer // `let pat [: ty] = expr` has an extended temporary lifetime if // any of the following conditions are met: // // A. `pat` matches `P&` and `expr` matches `ET` // (covers cases where `pat` creates ref bindings into an rvalue // produced by `expr`) // B. `ty` is a borrowed pointer and `expr` matches `ET` // (covers cases where coercion creates a borrow) // C. `expr` matches `E&` // (covers cases `expr` borrows an rvalue that is then assigned // to memory (at least partially) owned by the binding) // // Here are some examples hopefully giving an intuition where each // rule comes into play and why: // // Rule A. `let (ref x, ref y) = (foo().x, 44)`. The rvalue `(22, 44)` // would have an extended lifetime, but not `foo()`. // // Rule B. `let x = &foo().x`. The rvalue ``foo()` would have extended // lifetime. // // In some cases, multiple rules may apply (though not to the same // rvalue). For example: // // let ref x = [&a(), &b()]; // // Here, the expression `[...]` has an extended lifetime due to rule // A, but the inner rvalues `a()` and `b()` have an extended lifetime // due to rule C. if let Some(expr) = init { record_rvalue_scope_if_borrow_expr(visitor, &expr, blk_scope); if let Some(pat) = pat { if is_binding_pat(pat) { record_rvalue_scope(visitor, &expr, blk_scope); } } } // Make sure we visit the initializer first, so expr_and_pat_count remains correct if let Some(expr) = init { visitor.visit_expr(expr); } if let Some(pat) = pat { visitor.visit_pat(pat); } /// Returns `true` if `pat` match the `P&` non-terminal. /// /// P& = ref X /// | StructName { ..., P&, ... } /// | VariantName(..., P&, ...) /// | [ ..., P&, ... ] /// | ( ..., P&, ... ) /// | box P& fn is_binding_pat(pat: &hir::Pat) -> bool { // Note that the code below looks for *explicit* refs only, that is, it won't // know about *implicit* refs as introduced in #42640. // // This is not a problem. For example, consider // // let (ref x, ref y) = (Foo { .. }, Bar { .. }); // // Due to the explicit refs on the left hand side, the below code would signal // that the temporary value on the right hand side should live until the end of // the enclosing block (as opposed to being dropped after the let is complete). // // To create an implicit ref, however, you must have a borrowed value on the RHS // already, as in this example (which won't compile before #42640): // // let Foo { x, .. } = &Foo { x: ..., ... }; // // in place of // // let Foo { ref x, .. } = Foo { ... }; // // In the former case (the implicit ref version), the temporary is created by the // & expression, and its lifetime would be extended to the end of the block (due // to a different rule, not the below code). match pat.node { PatKind::Binding(hir::BindingAnnotation::Ref, ..) | PatKind::Binding(hir::BindingAnnotation::RefMut, ..) => true, PatKind::Struct(_, ref field_pats, _) => { field_pats.iter().any(|fp| is_binding_pat(&fp.node.pat)) } PatKind::Slice(ref pats1, ref pats2, ref pats3) => { pats1.iter().any(|p| is_binding_pat(&p)) || pats2.iter().any(|p| is_binding_pat(&p)) || pats3.iter().any(|p| is_binding_pat(&p)) } PatKind::TupleStruct(_, ref subpats, _) | PatKind::Tuple(ref subpats, _) => { subpats.iter().any(|p| is_binding_pat(&p)) } PatKind::Box(ref subpat) => { is_binding_pat(&subpat) } _ => false, } } /// If `expr` matches the `E&` grammar, then records an extended rvalue scope as appropriate: /// /// E& = & ET /// | StructName { ..., f: E&, ... } /// | [ ..., E&, ... ] /// | ( ..., E&, ... ) /// | {...; E&} /// | box E& /// | E& as ... /// | ( E& ) fn record_rvalue_scope_if_borrow_expr<'a, 'tcx>( visitor: &mut RegionResolutionVisitor<'a, 'tcx>, expr: &hir::Expr, blk_id: Option) { match expr.node { hir::ExprKind::AddrOf(_, ref subexpr) => { record_rvalue_scope_if_borrow_expr(visitor, &subexpr, blk_id); record_rvalue_scope(visitor, &subexpr, blk_id); } hir::ExprKind::Struct(_, ref fields, _) => { for field in fields { record_rvalue_scope_if_borrow_expr( visitor, &field.expr, blk_id); } } hir::ExprKind::Array(ref subexprs) | hir::ExprKind::Tup(ref subexprs) => { for subexpr in subexprs { record_rvalue_scope_if_borrow_expr( visitor, &subexpr, blk_id); } } hir::ExprKind::Cast(ref subexpr, _) => { record_rvalue_scope_if_borrow_expr(visitor, &subexpr, blk_id) } hir::ExprKind::Block(ref block, _) => { if let Some(ref subexpr) = block.expr { record_rvalue_scope_if_borrow_expr( visitor, &subexpr, blk_id); } } _ => {} } } /// Applied to an expression `expr` if `expr` -- or something owned or partially owned by /// `expr` -- is going to be indirectly referenced by a variable in a let statement. In that /// case, the "temporary lifetime" or `expr` is extended to be the block enclosing the `let` /// statement. /// /// More formally, if `expr` matches the grammar `ET`, record the rvalue scope of the matching /// `` as `blk_id`: /// /// ET = *ET /// | ET[...] /// | ET.f /// | (ET) /// | /// /// Note: ET is intended to match "rvalues or places based on rvalues". fn record_rvalue_scope<'a, 'tcx>(visitor: &mut RegionResolutionVisitor<'a, 'tcx>, expr: &hir::Expr, blk_scope: Option) { let mut expr = expr; loop { // Note: give all the expressions matching `ET` with the // extended temporary lifetime, not just the innermost rvalue, // because in codegen if we must compile e.g., `*rvalue()` // into a temporary, we request the temporary scope of the // outer expression. visitor.scope_tree.record_rvalue_scope(expr.hir_id.local_id, blk_scope); match expr.node { hir::ExprKind::AddrOf(_, ref subexpr) | hir::ExprKind::Unary(hir::UnDeref, ref subexpr) | hir::ExprKind::Field(ref subexpr, _) | hir::ExprKind::Index(ref subexpr, _) => { expr = &subexpr; } _ => { return; } } } } } impl<'a, 'tcx> RegionResolutionVisitor<'a, 'tcx> { /// Records the current parent (if any) as the parent of `child_scope`. /// Returns the depth of `child_scope`. fn record_child_scope(&mut self, child_scope: Scope) -> ScopeDepth { let parent = self.cx.parent; self.scope_tree.record_scope_parent(child_scope, parent); // If `child_scope` has no parent, it must be the root node, and so has // a depth of 1. Otherwise, its depth is one more than its parent's. parent.map_or(1, |(_p, d)| d + 1) } /// Records the current parent (if any) as the parent of `child_scope`, /// and sets `child_scope` as the new current parent. fn enter_scope(&mut self, child_scope: Scope) { let child_depth = self.record_child_scope(child_scope); self.cx.parent = Some((child_scope, child_depth)); } fn enter_node_scope_with_dtor(&mut self, id: hir::ItemLocalId) { // If node was previously marked as a terminating scope during the // recursive visit of its parent node in the AST, then we need to // account for the destruction scope representing the scope of // the destructors that run immediately after it completes. if self.terminating_scopes.contains(&id) { self.enter_scope(Scope { id, data: ScopeData::Destruction }); } self.enter_scope(Scope { id, data: ScopeData::Node }); } } impl<'a, 'tcx> Visitor<'tcx> for RegionResolutionVisitor<'a, 'tcx> { fn nested_visit_map<'this>(&'this mut self) -> NestedVisitorMap<'this, 'tcx> { NestedVisitorMap::None } fn visit_block(&mut self, b: &'tcx Block) { resolve_block(self, b); } fn visit_body(&mut self, body: &'tcx hir::Body) { let body_id = body.id(); let owner_id = self.tcx.hir().body_owner(body_id); debug!("visit_body(id={:?}, span={:?}, body.id={:?}, cx.parent={:?})", owner_id, self.tcx.sess.source_map().span_to_string(body.value.span), body_id, self.cx.parent); let outer_ec = mem::replace(&mut self.expr_and_pat_count, 0); let outer_cx = self.cx; let outer_ts = mem::replace(&mut self.terminating_scopes, FxHashSet::default()); self.terminating_scopes.insert(body.value.hir_id.local_id); if let Some(root_id) = self.cx.root_id { self.scope_tree.record_closure_parent(body.value.hir_id.local_id, root_id); } self.cx.root_id = Some(body.value.hir_id.local_id); self.enter_scope(Scope { id: body.value.hir_id.local_id, data: ScopeData::CallSite }); self.enter_scope(Scope { id: body.value.hir_id.local_id, data: ScopeData::Arguments }); // The arguments and `self` are parented to the fn. self.cx.var_parent = self.cx.parent.take(); for argument in &body.arguments { self.visit_pat(&argument.pat); } // The body of the every fn is a root scope. self.cx.parent = self.cx.var_parent; if self.tcx.hir().body_owner_kind(owner_id).is_fn_or_closure() { self.visit_expr(&body.value) } else { // Only functions have an outer terminating (drop) scope, while // temporaries in constant initializers may be 'static, but only // according to rvalue lifetime semantics, using the same // syntactical rules used for let initializers. // // e.g., in `let x = &f();`, the temporary holding the result from // the `f()` call lives for the entirety of the surrounding block. // // Similarly, `const X: ... = &f();` would have the result of `f()` // live for `'static`, implying (if Drop restrictions on constants // ever get lifted) that the value *could* have a destructor, but // it'd get leaked instead of the destructor running during the // evaluation of `X` (if at all allowed by CTFE). // // However, `const Y: ... = g(&f());`, like `let y = g(&f());`, // would *not* let the `f()` temporary escape into an outer scope // (i.e., `'static`), which means that after `g` returns, it drops, // and all the associated destruction scope rules apply. self.cx.var_parent = None; resolve_local(self, None, Some(&body.value)); } if body.is_generator { self.scope_tree.body_expr_count.insert(body_id, self.expr_and_pat_count); } // Restore context we had at the start. self.expr_and_pat_count = outer_ec; self.cx = outer_cx; self.terminating_scopes = outer_ts; } fn visit_arm(&mut self, a: &'tcx Arm) { resolve_arm(self, a); } fn visit_pat(&mut self, p: &'tcx Pat) { resolve_pat(self, p); } fn visit_stmt(&mut self, s: &'tcx Stmt) { resolve_stmt(self, s); } fn visit_expr(&mut self, ex: &'tcx Expr) { resolve_expr(self, ex); } fn visit_local(&mut self, l: &'tcx Local) { resolve_local(self, Some(&l.pat), l.init.as_ref().map(|e| &**e)); } } fn region_scope_tree<'a, 'tcx>(tcx: TyCtxt<'a, 'tcx, 'tcx>, def_id: DefId) -> &'tcx ScopeTree { let closure_base_def_id = tcx.closure_base_def_id(def_id); if closure_base_def_id != def_id { return tcx.region_scope_tree(closure_base_def_id); } let id = tcx.hir().as_local_hir_id(def_id).unwrap(); let scope_tree = if let Some(body_id) = tcx.hir().maybe_body_owned_by_by_hir_id(id) { let mut visitor = RegionResolutionVisitor { tcx, scope_tree: ScopeTree::default(), expr_and_pat_count: 0, cx: Context { root_id: None, parent: None, var_parent: None, }, terminating_scopes: Default::default(), }; let body = tcx.hir().body(body_id); visitor.scope_tree.root_body = Some(body.value.hir_id); // If the item is an associated const or a method, // record its impl/trait parent, as it can also have // lifetime parameters free in this body. match tcx.hir().get_by_hir_id(id) { Node::ImplItem(_) | Node::TraitItem(_) => { visitor.scope_tree.root_parent = Some(tcx.hir().get_parent_item(id)); } _ => {} } visitor.visit_body(body); visitor.scope_tree } else { ScopeTree::default() }; tcx.arena.alloc(scope_tree) } pub fn provide(providers: &mut Providers<'_>) { *providers = Providers { region_scope_tree, ..*providers }; } impl<'a> HashStable> for ScopeTree { fn hash_stable(&self, hcx: &mut StableHashingContext<'a>, hasher: &mut StableHasher) { let ScopeTree { root_body, root_parent, ref body_expr_count, ref parent_map, ref var_map, ref destruction_scopes, ref rvalue_scopes, ref closure_tree, ref yield_in_scope, } = *self; hcx.with_node_id_hashing_mode(NodeIdHashingMode::HashDefPath, |hcx| { root_body.hash_stable(hcx, hasher); root_parent.hash_stable(hcx, hasher); }); body_expr_count.hash_stable(hcx, hasher); parent_map.hash_stable(hcx, hasher); var_map.hash_stable(hcx, hasher); destruction_scopes.hash_stable(hcx, hasher); rvalue_scopes.hash_stable(hcx, hasher); closure_tree.hash_stable(hcx, hasher); yield_in_scope.hash_stable(hcx, hasher); } }