// Copyright 2014 The Rust Project Developers. See the COPYRIGHT // file at the top-level directory of this distribution and at // http://rust-lang.org/COPYRIGHT. // // Licensed under the Apache License, Version 2.0 or the MIT license // , at your // option. This file may not be copied, modified, or distributed // except according to those terms. // This is an attempt at an implementation following the ideal // // ``` // struct BTreeMap { // height: usize, // root: Option>> // } // // struct Node { // keys: [K; 2 * B - 1], // vals: [V; 2 * B - 1], // edges: if height > 0 { // [Box>; 2 * B] // } else { () }, // parent: *const Node, // parent_idx: u16, // len: u16, // } // ``` // // Since Rust doesn't actually have dependent types and polymorphic recursion, // we make do with lots of unsafety. // A major goal of this module is to avoid complexity by treating the tree as a generic (if // weirdly shaped) container and avoiding dealing with most of the B-Tree invariants. As such, // this module doesn't care whether the entries are sorted, which nodes can be underfull, or // even what underfull means. However, we do rely on a few invariants: // // - Trees must have uniform depth/height. This means that every path down to a leaf from a // given node has exactly the same length. // - A node of length `n` has `n` keys, `n` values, and (in an internal node) `n + 1` edges. // This implies that even an empty internal node has at least one edge. use core::marker::PhantomData; use core::mem; use core::ptr::{self, Unique, NonNull}; use core::slice; use alloc::{Global, Alloc, Layout}; use boxed::Box; const B: usize = 6; pub const MIN_LEN: usize = B - 1; pub const CAPACITY: usize = 2 * B - 1; /// The underlying representation of leaf nodes. Note that it is often unsafe to actually store /// these, since only the first `len` keys and values are assumed to be initialized. As such, /// these should always be put behind pointers, and specifically behind `BoxedNode` in the owned /// case. /// /// See also rust-lang/rfcs#197, which would make this structure significantly more safe by /// avoiding accidentally dropping unused and uninitialized keys and values. /// /// We put the metadata first so that its position is the same for every `K` and `V`, in order /// to statically allocate a single dummy node to avoid allocations. This struct is `repr(C)` to /// prevent them from being reordered. #[repr(C)] struct LeafNode { /// We use `*const` as opposed to `*mut` so as to be covariant in `K` and `V`. /// This either points to an actual node or is null. parent: *const InternalNode, /// This node's index into the parent node's `edges` array. /// `*node.parent.edges[node.parent_idx]` should be the same thing as `node`. /// This is only guaranteed to be initialized when `parent` is nonnull. parent_idx: u16, /// The number of keys and values this node stores. /// /// This next to `parent_idx` to encourage the compiler to join `len` and /// `parent_idx` into the same 32-bit word, reducing space overhead. len: u16, /// The arrays storing the actual data of the node. Only the first `len` elements of each /// array are initialized and valid. keys: [K; CAPACITY], vals: [V; CAPACITY], } impl LeafNode { /// Creates a new `LeafNode`. Unsafe because all nodes should really be hidden behind /// `BoxedNode`, preventing accidental dropping of uninitialized keys and values. unsafe fn new() -> Self { LeafNode { // As a general policy, we leave fields uninitialized if they can be, as this should // be both slightly faster and easier to track in Valgrind. keys: mem::uninitialized(), vals: mem::uninitialized(), parent: ptr::null(), parent_idx: mem::uninitialized(), len: 0 } } fn is_shared_root(&self) -> bool { self as *const _ == &EMPTY_ROOT_NODE as *const _ as *const LeafNode } } // We need to implement Sync here in order to make a static instance. unsafe impl Sync for LeafNode<(), ()> {} // An empty node used as a placeholder for the root node, to avoid allocations. // We use () in order to save space, since no operation on an empty tree will // ever take a pointer past the first key. static EMPTY_ROOT_NODE: LeafNode<(), ()> = LeafNode { parent: ptr::null(), parent_idx: 0, len: 0, keys: [(); CAPACITY], vals: [(); CAPACITY], }; /// The underlying representation of internal nodes. As with `LeafNode`s, these should be hidden /// behind `BoxedNode`s to prevent dropping uninitialized keys and values. Any pointer to an /// `InternalNode` can be directly casted to a pointer to the underlying `LeafNode` portion of the /// node, allowing code to act on leaf and internal nodes generically without having to even check /// which of the two a pointer is pointing at. This property is enabled by the use of `repr(C)`. #[repr(C)] struct InternalNode { data: LeafNode, /// The pointers to the children of this node. `len + 1` of these are considered /// initialized and valid. edges: [BoxedNode; 2 * B], } impl InternalNode { /// Creates a new `InternalNode`. /// /// This is unsafe for two reasons. First, it returns an `InternalNode` by value, risking /// dropping of uninitialized fields. Second, an invariant of internal nodes is that `len + 1` /// edges are initialized and valid, meaning that even when the node is empty (having a /// `len` of 0), there must be one initialized and valid edge. This function does not set up /// such an edge. unsafe fn new() -> Self { InternalNode { data: LeafNode::new(), edges: mem::uninitialized() } } } /// An owned pointer to a node. This basically is either `Box>` or /// `Box>`. However, it contains no information as to which of the two types /// of nodes is actually behind the box, and, partially due to this lack of information, has no /// destructor. struct BoxedNode { ptr: Unique> } impl BoxedNode { fn from_leaf(node: Box>) -> Self { BoxedNode { ptr: Box::into_unique(node) } } fn from_internal(node: Box>) -> Self { unsafe { BoxedNode { ptr: Unique::new_unchecked(Box::into_raw(node) as *mut LeafNode) } } } unsafe fn from_ptr(ptr: NonNull>) -> Self { BoxedNode { ptr: Unique::from(ptr) } } fn as_ptr(&self) -> NonNull> { NonNull::from(self.ptr) } } /// An owned tree. Note that despite being owned, this does not have a destructor, /// and must be cleaned up manually. pub struct Root { node: BoxedNode, height: usize } unsafe impl Sync for Root { } unsafe impl Send for Root { } impl Root { pub fn is_shared_root(&self) -> bool { self.as_ref().is_shared_root() } pub fn shared_empty_root() -> Self { Root { node: unsafe { BoxedNode::from_ptr(NonNull::new_unchecked( &EMPTY_ROOT_NODE as *const _ as *const LeafNode as *mut _ )) }, height: 0, } } pub fn new_leaf() -> Self { Root { node: BoxedNode::from_leaf(Box::new(unsafe { LeafNode::new() })), height: 0 } } pub fn as_ref(&self) -> NodeRef { NodeRef { height: self.height, node: self.node.as_ptr(), root: self as *const _ as *mut _, _marker: PhantomData, } } pub fn as_mut(&mut self) -> NodeRef { NodeRef { height: self.height, node: self.node.as_ptr(), root: self as *mut _, _marker: PhantomData, } } pub fn into_ref(self) -> NodeRef { NodeRef { height: self.height, node: self.node.as_ptr(), root: ptr::null_mut(), // FIXME: Is there anything better to do here? _marker: PhantomData, } } /// Adds a new internal node with a single edge, pointing to the previous root, and make that /// new node the root. This increases the height by 1 and is the opposite of `pop_level`. pub fn push_level(&mut self) -> NodeRef { debug_assert!(!self.is_shared_root()); let mut new_node = Box::new(unsafe { InternalNode::new() }); new_node.edges[0] = unsafe { BoxedNode::from_ptr(self.node.as_ptr()) }; self.node = BoxedNode::from_internal(new_node); self.height += 1; let mut ret = NodeRef { height: self.height, node: self.node.as_ptr(), root: self as *mut _, _marker: PhantomData }; unsafe { ret.reborrow_mut().first_edge().correct_parent_link(); } ret } /// Removes the root node, using its first child as the new root. This cannot be called when /// the tree consists only of a leaf node. As it is intended only to be called when the root /// has only one edge, no cleanup is done on any of the other children are elements of the root. /// This decreases the height by 1 and is the opposite of `push_level`. pub fn pop_level(&mut self) { debug_assert!(self.height > 0); let top = self.node.ptr; self.node = unsafe { BoxedNode::from_ptr(self.as_mut() .cast_unchecked::() .first_edge() .descend() .node) }; self.height -= 1; self.as_mut().as_leaf_mut().parent = ptr::null(); unsafe { Global.dealloc(NonNull::from(top).cast(), Layout::new::>()); } } } // N.B. `NodeRef` is always covariant in `K` and `V`, even when the `BorrowType` // is `Mut`. This is technically wrong, but cannot result in any unsafety due to // internal use of `NodeRef` because we stay completely generic over `K` and `V`. // However, whenever a public type wraps `NodeRef`, make sure that it has the // correct variance. /// A reference to a node. /// /// This type has a number of parameters that controls how it acts: /// - `BorrowType`: This can be `Immut<'a>` or `Mut<'a>` for some `'a` or `Owned`. /// When this is `Immut<'a>`, the `NodeRef` acts roughly like `&'a Node`, /// when this is `Mut<'a>`, the `NodeRef` acts roughly like `&'a mut Node`, /// and when this is `Owned`, the `NodeRef` acts roughly like `Box`. /// - `K` and `V`: These control what types of things are stored in the nodes. /// - `Type`: This can be `Leaf`, `Internal`, or `LeafOrInternal`. When this is /// `Leaf`, the `NodeRef` points to a leaf node, when this is `Internal` the /// `NodeRef` points to an internal node, and when this is `LeafOrInternal` the /// `NodeRef` could be pointing to either type of node. pub struct NodeRef { height: usize, node: NonNull>, // This is null unless the borrow type is `Mut` root: *const Root, _marker: PhantomData<(BorrowType, Type)> } impl<'a, K: 'a, V: 'a, Type> Copy for NodeRef, K, V, Type> { } impl<'a, K: 'a, V: 'a, Type> Clone for NodeRef, K, V, Type> { fn clone(&self) -> Self { *self } } unsafe impl Sync for NodeRef { } unsafe impl<'a, K: Sync + 'a, V: Sync + 'a, Type> Send for NodeRef, K, V, Type> { } unsafe impl<'a, K: Send + 'a, V: Send + 'a, Type> Send for NodeRef, K, V, Type> { } unsafe impl Send for NodeRef { } impl NodeRef { fn as_internal(&self) -> &InternalNode { unsafe { &*(self.node.as_ptr() as *mut InternalNode) } } } impl<'a, K, V> NodeRef, K, V, marker::Internal> { fn as_internal_mut(&mut self) -> &mut InternalNode { unsafe { &mut *(self.node.as_ptr() as *mut InternalNode) } } } impl NodeRef { /// Finds the length of the node. This is the number of keys or values. In an /// internal node, the number of edges is `len() + 1`. pub fn len(&self) -> usize { self.as_leaf().len as usize } /// Returns the height of this node in the whole tree. Zero height denotes the /// leaf level. pub fn height(&self) -> usize { self.height } /// Removes any static information about whether this node is a `Leaf` or an /// `Internal` node. pub fn forget_type(self) -> NodeRef { NodeRef { height: self.height, node: self.node, root: self.root, _marker: PhantomData } } /// Temporarily takes out another, immutable reference to the same node. fn reborrow<'a>(&'a self) -> NodeRef, K, V, Type> { NodeRef { height: self.height, node: self.node, root: self.root, _marker: PhantomData } } fn as_leaf(&self) -> &LeafNode { unsafe { self.node.as_ref() } } pub fn is_shared_root(&self) -> bool { self.as_leaf().is_shared_root() } pub fn keys(&self) -> &[K] { self.reborrow().into_key_slice() } fn vals(&self) -> &[V] { self.reborrow().into_val_slice() } /// Finds the parent of the current node. Returns `Ok(handle)` if the current /// node actually has a parent, where `handle` points to the edge of the parent /// that points to the current node. Returns `Err(self)` if the current node has /// no parent, giving back the original `NodeRef`. /// /// `edge.descend().ascend().unwrap()` and `node.ascend().unwrap().descend()` should /// both, upon success, do nothing. pub fn ascend(self) -> Result< Handle< NodeRef< BorrowType, K, V, marker::Internal >, marker::Edge >, Self > { let parent_as_leaf = self.as_leaf().parent as *const LeafNode; if let Some(non_zero) = NonNull::new(parent_as_leaf as *mut _) { Ok(Handle { node: NodeRef { height: self.height + 1, node: non_zero, root: self.root, _marker: PhantomData }, idx: self.as_leaf().parent_idx as usize, _marker: PhantomData }) } else { Err(self) } } pub fn first_edge(self) -> Handle { Handle::new_edge(self, 0) } pub fn last_edge(self) -> Handle { let len = self.len(); Handle::new_edge(self, len) } /// Note that `self` must be nonempty. pub fn first_kv(self) -> Handle { debug_assert!(self.len() > 0); Handle::new_kv(self, 0) } /// Note that `self` must be nonempty. pub fn last_kv(self) -> Handle { let len = self.len(); debug_assert!(len > 0); Handle::new_kv(self, len - 1) } } impl NodeRef { /// Similar to `ascend`, gets a reference to a node's parent node, but also /// deallocate the current node in the process. This is unsafe because the /// current node will still be accessible despite being deallocated. pub unsafe fn deallocate_and_ascend(self) -> Option< Handle< NodeRef< marker::Owned, K, V, marker::Internal >, marker::Edge > > { debug_assert!(!self.is_shared_root()); let node = self.node; let ret = self.ascend().ok(); Global.dealloc(node.cast(), Layout::new::>()); ret } } impl NodeRef { /// Similar to `ascend`, gets a reference to a node's parent node, but also /// deallocate the current node in the process. This is unsafe because the /// current node will still be accessible despite being deallocated. pub unsafe fn deallocate_and_ascend(self) -> Option< Handle< NodeRef< marker::Owned, K, V, marker::Internal >, marker::Edge > > { let node = self.node; let ret = self.ascend().ok(); Global.dealloc(node.cast(), Layout::new::>()); ret } } impl<'a, K, V, Type> NodeRef, K, V, Type> { /// Unsafely asserts to the compiler some static information about whether this /// node is a `Leaf`. unsafe fn cast_unchecked(&mut self) -> NodeRef { NodeRef { height: self.height, node: self.node, root: self.root, _marker: PhantomData } } /// Temporarily takes out another, mutable reference to the same node. Beware, as /// this method is very dangerous, doubly so since it may not immediately appear /// dangerous. /// /// Because mutable pointers can roam anywhere around the tree and can even (through /// `into_root_mut`) mess with the root of the tree, the result of `reborrow_mut` /// can easily be used to make the original mutable pointer dangling, or, in the case /// of a reborrowed handle, out of bounds. // FIXME(@gereeter) consider adding yet another type parameter to `NodeRef` that restricts // the use of `ascend` and `into_root_mut` on reborrowed pointers, preventing this unsafety. unsafe fn reborrow_mut(&mut self) -> NodeRef { NodeRef { height: self.height, node: self.node, root: self.root, _marker: PhantomData } } fn as_leaf_mut(&mut self) -> &mut LeafNode { unsafe { self.node.as_mut() } } fn keys_mut(&mut self) -> &mut [K] { unsafe { self.reborrow_mut().into_key_slice_mut() } } fn vals_mut(&mut self) -> &mut [V] { unsafe { self.reborrow_mut().into_val_slice_mut() } } } impl<'a, K: 'a, V: 'a, Type> NodeRef, K, V, Type> { fn into_key_slice(self) -> &'a [K] { // When taking a pointer to the keys, if our key has a stricter // alignment requirement than the shared root does, then the pointer // would be out of bounds, which LLVM assumes will not happen. If the // alignment is more strict, we need to make an empty slice that doesn't // use an out of bounds pointer. if mem::align_of::() > mem::align_of::>() && self.is_shared_root() { &[] } else { // Here either it's not the root, or the alignment is less strict, // in which case the keys pointer will point "one-past-the-end" of // the node, which is allowed by LLVM. unsafe { slice::from_raw_parts( self.as_leaf().keys.as_ptr(), self.len() ) } } } fn into_val_slice(self) -> &'a [V] { debug_assert!(!self.is_shared_root()); unsafe { slice::from_raw_parts( self.as_leaf().vals.as_ptr(), self.len() ) } } fn into_slices(self) -> (&'a [K], &'a [V]) { let k = unsafe { ptr::read(&self) }; (k.into_key_slice(), self.into_val_slice()) } } impl<'a, K: 'a, V: 'a, Type> NodeRef, K, V, Type> { /// Gets a mutable reference to the root itself. This is useful primarily when the /// height of the tree needs to be adjusted. Never call this on a reborrowed pointer. pub fn into_root_mut(self) -> &'a mut Root { unsafe { &mut *(self.root as *mut Root) } } fn into_key_slice_mut(mut self) -> &'a mut [K] { if mem::align_of::() > mem::align_of::>() && self.is_shared_root() { &mut [] } else { unsafe { slice::from_raw_parts_mut( &mut self.as_leaf_mut().keys as *mut [K] as *mut K, self.len() ) } } } fn into_val_slice_mut(mut self) -> &'a mut [V] { debug_assert!(!self.is_shared_root()); unsafe { slice::from_raw_parts_mut( &mut self.as_leaf_mut().vals as *mut [V] as *mut V, self.len() ) } } fn into_slices_mut(self) -> (&'a mut [K], &'a mut [V]) { let k = unsafe { ptr::read(&self) }; (k.into_key_slice_mut(), self.into_val_slice_mut()) } } impl<'a, K, V> NodeRef, K, V, marker::Leaf> { /// Adds a key/value pair the end of the node. pub fn push(&mut self, key: K, val: V) { // Necessary for correctness, but this is an internal module debug_assert!(self.len() < CAPACITY); debug_assert!(!self.is_shared_root()); let idx = self.len(); unsafe { ptr::write(self.keys_mut().get_unchecked_mut(idx), key); ptr::write(self.vals_mut().get_unchecked_mut(idx), val); } self.as_leaf_mut().len += 1; } /// Adds a key/value pair to the beginning of the node. pub fn push_front(&mut self, key: K, val: V) { // Necessary for correctness, but this is an internal module debug_assert!(self.len() < CAPACITY); debug_assert!(!self.is_shared_root()); unsafe { slice_insert(self.keys_mut(), 0, key); slice_insert(self.vals_mut(), 0, val); } self.as_leaf_mut().len += 1; } } impl<'a, K, V> NodeRef, K, V, marker::Internal> { /// Adds a key/value pair and an edge to go to the right of that pair to /// the end of the node. pub fn push(&mut self, key: K, val: V, edge: Root) { // Necessary for correctness, but this is an internal module debug_assert!(edge.height == self.height - 1); debug_assert!(self.len() < CAPACITY); let idx = self.len(); unsafe { ptr::write(self.keys_mut().get_unchecked_mut(idx), key); ptr::write(self.vals_mut().get_unchecked_mut(idx), val); ptr::write(self.as_internal_mut().edges.get_unchecked_mut(idx + 1), edge.node); self.as_leaf_mut().len += 1; Handle::new_edge(self.reborrow_mut(), idx + 1).correct_parent_link(); } } fn correct_childrens_parent_links(&mut self, first: usize, after_last: usize) { for i in first..after_last { Handle::new_edge(unsafe { self.reborrow_mut() }, i).correct_parent_link(); } } fn correct_all_childrens_parent_links(&mut self) { let len = self.len(); self.correct_childrens_parent_links(0, len + 1); } /// Adds a key/value pair and an edge to go to the left of that pair to /// the beginning of the node. pub fn push_front(&mut self, key: K, val: V, edge: Root) { // Necessary for correctness, but this is an internal module debug_assert!(edge.height == self.height - 1); debug_assert!(self.len() < CAPACITY); unsafe { slice_insert(self.keys_mut(), 0, key); slice_insert(self.vals_mut(), 0, val); slice_insert( slice::from_raw_parts_mut( self.as_internal_mut().edges.as_mut_ptr(), self.len()+1 ), 0, edge.node ); self.as_leaf_mut().len += 1; self.correct_all_childrens_parent_links(); } } } impl<'a, K, V> NodeRef, K, V, marker::LeafOrInternal> { /// Removes a key/value pair from the end of this node. If this is an internal node, /// also removes the edge that was to the right of that pair. pub fn pop(&mut self) -> (K, V, Option>) { // Necessary for correctness, but this is an internal module debug_assert!(self.len() > 0); let idx = self.len() - 1; unsafe { let key = ptr::read(self.keys().get_unchecked(idx)); let val = ptr::read(self.vals().get_unchecked(idx)); let edge = match self.reborrow_mut().force() { ForceResult::Leaf(_) => None, ForceResult::Internal(internal) => { let edge = ptr::read(internal.as_internal().edges.get_unchecked(idx + 1)); let mut new_root = Root { node: edge, height: internal.height - 1 }; new_root.as_mut().as_leaf_mut().parent = ptr::null(); Some(new_root) } }; self.as_leaf_mut().len -= 1; (key, val, edge) } } /// Removes a key/value pair from the beginning of this node. If this is an internal node, /// also removes the edge that was to the left of that pair. pub fn pop_front(&mut self) -> (K, V, Option>) { // Necessary for correctness, but this is an internal module debug_assert!(self.len() > 0); let old_len = self.len(); unsafe { let key = slice_remove(self.keys_mut(), 0); let val = slice_remove(self.vals_mut(), 0); let edge = match self.reborrow_mut().force() { ForceResult::Leaf(_) => None, ForceResult::Internal(mut internal) => { let edge = slice_remove( slice::from_raw_parts_mut( internal.as_internal_mut().edges.as_mut_ptr(), old_len+1 ), 0 ); let mut new_root = Root { node: edge, height: internal.height - 1 }; new_root.as_mut().as_leaf_mut().parent = ptr::null(); for i in 0..old_len { Handle::new_edge(internal.reborrow_mut(), i).correct_parent_link(); } Some(new_root) } }; self.as_leaf_mut().len -= 1; (key, val, edge) } } fn into_kv_pointers_mut(mut self) -> (*mut K, *mut V) { ( self.keys_mut().as_mut_ptr(), self.vals_mut().as_mut_ptr() ) } } impl NodeRef { /// Checks whether a node is an `Internal` node or a `Leaf` node. pub fn force(self) -> ForceResult< NodeRef, NodeRef > { if self.height == 0 { ForceResult::Leaf(NodeRef { height: self.height, node: self.node, root: self.root, _marker: PhantomData }) } else { ForceResult::Internal(NodeRef { height: self.height, node: self.node, root: self.root, _marker: PhantomData }) } } } /// A reference to a specific key/value pair or edge within a node. The `Node` parameter /// must be a `NodeRef`, while the `Type` can either be `KV` (signifying a handle on a key/value /// pair) or `Edge` (signifying a handle on an edge). /// /// Note that even `Leaf` nodes can have `Edge` handles. Instead of representing a pointer to /// a child node, these represent the spaces where child pointers would go between the key/value /// pairs. For example, in a node with length 2, there would be 3 possible edge locations - one /// to the left of the node, one between the two pairs, and one at the right of the node. pub struct Handle { node: Node, idx: usize, _marker: PhantomData } impl Copy for Handle { } // We don't need the full generality of `#[derive(Clone)]`, as the only time `Node` will be // `Clone`able is when it is an immutable reference and therefore `Copy`. impl Clone for Handle { fn clone(&self) -> Self { *self } } impl Handle { /// Retrieves the node that contains the edge of key/value pair this handle points to. pub fn into_node(self) -> Node { self.node } } impl Handle, marker::KV> { /// Creates a new handle to a key/value pair in `node`. `idx` must be less than `node.len()`. pub fn new_kv(node: NodeRef, idx: usize) -> Self { // Necessary for correctness, but in a private module debug_assert!(idx < node.len()); Handle { node, idx, _marker: PhantomData } } pub fn left_edge(self) -> Handle, marker::Edge> { Handle::new_edge(self.node, self.idx) } pub fn right_edge(self) -> Handle, marker::Edge> { Handle::new_edge(self.node, self.idx + 1) } } impl PartialEq for Handle, HandleType> { fn eq(&self, other: &Self) -> bool { self.node.node == other.node.node && self.idx == other.idx } } impl Handle, HandleType> { /// Temporarily takes out another, immutable handle on the same location. pub fn reborrow(&self) -> Handle, HandleType> { // We can't use Handle::new_kv or Handle::new_edge because we don't know our type Handle { node: self.node.reborrow(), idx: self.idx, _marker: PhantomData } } } impl<'a, K, V, NodeType, HandleType> Handle, K, V, NodeType>, HandleType> { /// Temporarily takes out another, mutable handle on the same location. Beware, as /// this method is very dangerous, doubly so since it may not immediately appear /// dangerous. /// /// Because mutable pointers can roam anywhere around the tree and can even (through /// `into_root_mut`) mess with the root of the tree, the result of `reborrow_mut` /// can easily be used to make the original mutable pointer dangling, or, in the case /// of a reborrowed handle, out of bounds. // FIXME(@gereeter) consider adding yet another type parameter to `NodeRef` that restricts // the use of `ascend` and `into_root_mut` on reborrowed pointers, preventing this unsafety. pub unsafe fn reborrow_mut(&mut self) -> Handle, HandleType> { // We can't use Handle::new_kv or Handle::new_edge because we don't know our type Handle { node: self.node.reborrow_mut(), idx: self.idx, _marker: PhantomData } } } impl Handle, marker::Edge> { /// Creates a new handle to an edge in `node`. `idx` must be less than or equal to /// `node.len()`. pub fn new_edge(node: NodeRef, idx: usize) -> Self { // Necessary for correctness, but in a private module debug_assert!(idx <= node.len()); Handle { node, idx, _marker: PhantomData } } pub fn left_kv(self) -> Result, marker::KV>, Self> { if self.idx > 0 { Ok(Handle::new_kv(self.node, self.idx - 1)) } else { Err(self) } } pub fn right_kv(self) -> Result, marker::KV>, Self> { if self.idx < self.node.len() { Ok(Handle::new_kv(self.node, self.idx)) } else { Err(self) } } } impl<'a, K, V> Handle, K, V, marker::Leaf>, marker::Edge> { /// Inserts a new key/value pair between the key/value pairs to the right and left of /// this edge. This method assumes that there is enough space in the node for the new /// pair to fit. /// /// The returned pointer points to the inserted value. fn insert_fit(&mut self, key: K, val: V) -> *mut V { // Necessary for correctness, but in a private module debug_assert!(self.node.len() < CAPACITY); debug_assert!(!self.node.is_shared_root()); unsafe { slice_insert(self.node.keys_mut(), self.idx, key); slice_insert(self.node.vals_mut(), self.idx, val); self.node.as_leaf_mut().len += 1; self.node.vals_mut().get_unchecked_mut(self.idx) } } /// Inserts a new key/value pair between the key/value pairs to the right and left of /// this edge. This method splits the node if there isn't enough room. /// /// The returned pointer points to the inserted value. pub fn insert(mut self, key: K, val: V) -> (InsertResult<'a, K, V, marker::Leaf>, *mut V) { if self.node.len() < CAPACITY { let ptr = self.insert_fit(key, val); (InsertResult::Fit(Handle::new_kv(self.node, self.idx)), ptr) } else { let middle = Handle::new_kv(self.node, B); let (mut left, k, v, mut right) = middle.split(); let ptr = if self.idx <= B { unsafe { Handle::new_edge(left.reborrow_mut(), self.idx).insert_fit(key, val) } } else { unsafe { Handle::new_edge( right.as_mut().cast_unchecked::(), self.idx - (B + 1) ).insert_fit(key, val) } }; (InsertResult::Split(left, k, v, right), ptr) } } } impl<'a, K, V> Handle, K, V, marker::Internal>, marker::Edge> { /// Fixes the parent pointer and index in the child node below this edge. This is useful /// when the ordering of edges has been changed, such as in the various `insert` methods. fn correct_parent_link(mut self) { let idx = self.idx as u16; let ptr = self.node.as_internal_mut() as *mut _; let mut child = self.descend(); child.as_leaf_mut().parent = ptr; child.as_leaf_mut().parent_idx = idx; } /// Unsafely asserts to the compiler some static information about whether the underlying /// node of this handle is a `Leaf`. unsafe fn cast_unchecked(&mut self) -> Handle, marker::Edge> { Handle::new_edge(self.node.cast_unchecked(), self.idx) } /// Inserts a new key/value pair and an edge that will go to the right of that new pair /// between this edge and the key/value pair to the right of this edge. This method assumes /// that there is enough space in the node for the new pair to fit. fn insert_fit(&mut self, key: K, val: V, edge: Root) { // Necessary for correctness, but in an internal module debug_assert!(self.node.len() < CAPACITY); debug_assert!(edge.height == self.node.height - 1); unsafe { // This cast is a lie, but it allows us to reuse the key/value insertion logic. self.cast_unchecked::().insert_fit(key, val); slice_insert( slice::from_raw_parts_mut( self.node.as_internal_mut().edges.as_mut_ptr(), self.node.len() ), self.idx + 1, edge.node ); for i in (self.idx+1)..(self.node.len()+1) { Handle::new_edge(self.node.reborrow_mut(), i).correct_parent_link(); } } } /// Inserts a new key/value pair and an edge that will go to the right of that new pair /// between this edge and the key/value pair to the right of this edge. This method splits /// the node if there isn't enough room. pub fn insert(mut self, key: K, val: V, edge: Root) -> InsertResult<'a, K, V, marker::Internal> { // Necessary for correctness, but this is an internal module debug_assert!(edge.height == self.node.height - 1); if self.node.len() < CAPACITY { self.insert_fit(key, val, edge); InsertResult::Fit(Handle::new_kv(self.node, self.idx)) } else { let middle = Handle::new_kv(self.node, B); let (mut left, k, v, mut right) = middle.split(); if self.idx <= B { unsafe { Handle::new_edge(left.reborrow_mut(), self.idx).insert_fit(key, val, edge); } } else { unsafe { Handle::new_edge( right.as_mut().cast_unchecked::(), self.idx - (B + 1) ).insert_fit(key, val, edge); } } InsertResult::Split(left, k, v, right) } } } impl Handle, marker::Edge> { /// Finds the node pointed to by this edge. /// /// `edge.descend().ascend().unwrap()` and `node.ascend().unwrap().descend()` should /// both, upon success, do nothing. pub fn descend(self) -> NodeRef { NodeRef { height: self.node.height - 1, node: unsafe { self.node.as_internal().edges.get_unchecked(self.idx).as_ptr() }, root: self.node.root, _marker: PhantomData } } } impl<'a, K: 'a, V: 'a, NodeType> Handle, K, V, NodeType>, marker::KV> { pub fn into_kv(self) -> (&'a K, &'a V) { let (keys, vals) = self.node.into_slices(); unsafe { (keys.get_unchecked(self.idx), vals.get_unchecked(self.idx)) } } } impl<'a, K: 'a, V: 'a, NodeType> Handle, K, V, NodeType>, marker::KV> { pub fn into_kv_mut(self) -> (&'a mut K, &'a mut V) { let (keys, vals) = self.node.into_slices_mut(); unsafe { (keys.get_unchecked_mut(self.idx), vals.get_unchecked_mut(self.idx)) } } } impl<'a, K, V, NodeType> Handle, K, V, NodeType>, marker::KV> { pub fn kv_mut(&mut self) -> (&mut K, &mut V) { unsafe { let (keys, vals) = self.node.reborrow_mut().into_slices_mut(); (keys.get_unchecked_mut(self.idx), vals.get_unchecked_mut(self.idx)) } } } impl<'a, K, V> Handle, K, V, marker::Leaf>, marker::KV> { /// Splits the underlying node into three parts: /// /// - The node is truncated to only contain the key/value pairs to the right of /// this handle. /// - The key and value pointed to by this handle and extracted. /// - All the key/value pairs to the right of this handle are put into a newly /// allocated node. pub fn split(mut self) -> (NodeRef, K, V, marker::Leaf>, K, V, Root) { debug_assert!(!self.node.is_shared_root()); unsafe { let mut new_node = Box::new(LeafNode::new()); let k = ptr::read(self.node.keys().get_unchecked(self.idx)); let v = ptr::read(self.node.vals().get_unchecked(self.idx)); let new_len = self.node.len() - self.idx - 1; ptr::copy_nonoverlapping( self.node.keys().as_ptr().offset(self.idx as isize + 1), new_node.keys.as_mut_ptr(), new_len ); ptr::copy_nonoverlapping( self.node.vals().as_ptr().offset(self.idx as isize + 1), new_node.vals.as_mut_ptr(), new_len ); self.node.as_leaf_mut().len = self.idx as u16; new_node.len = new_len as u16; ( self.node, k, v, Root { node: BoxedNode::from_leaf(new_node), height: 0 } ) } } /// Removes the key/value pair pointed to by this handle, returning the edge between the /// now adjacent key/value pairs to the left and right of this handle. pub fn remove(mut self) -> (Handle, K, V, marker::Leaf>, marker::Edge>, K, V) { debug_assert!(!self.node.is_shared_root()); unsafe { let k = slice_remove(self.node.keys_mut(), self.idx); let v = slice_remove(self.node.vals_mut(), self.idx); self.node.as_leaf_mut().len -= 1; (self.left_edge(), k, v) } } } impl<'a, K, V> Handle, K, V, marker::Internal>, marker::KV> { /// Splits the underlying node into three parts: /// /// - The node is truncated to only contain the edges and key/value pairs to the /// right of this handle. /// - The key and value pointed to by this handle and extracted. /// - All the edges and key/value pairs to the right of this handle are put into /// a newly allocated node. pub fn split(mut self) -> (NodeRef, K, V, marker::Internal>, K, V, Root) { unsafe { let mut new_node = Box::new(InternalNode::new()); let k = ptr::read(self.node.keys().get_unchecked(self.idx)); let v = ptr::read(self.node.vals().get_unchecked(self.idx)); let height = self.node.height; let new_len = self.node.len() - self.idx - 1; ptr::copy_nonoverlapping( self.node.keys().as_ptr().offset(self.idx as isize + 1), new_node.data.keys.as_mut_ptr(), new_len ); ptr::copy_nonoverlapping( self.node.vals().as_ptr().offset(self.idx as isize + 1), new_node.data.vals.as_mut_ptr(), new_len ); ptr::copy_nonoverlapping( self.node.as_internal().edges.as_ptr().offset(self.idx as isize + 1), new_node.edges.as_mut_ptr(), new_len + 1 ); self.node.as_leaf_mut().len = self.idx as u16; new_node.data.len = new_len as u16; let mut new_root = Root { node: BoxedNode::from_internal(new_node), height, }; for i in 0..(new_len+1) { Handle::new_edge(new_root.as_mut().cast_unchecked(), i).correct_parent_link(); } ( self.node, k, v, new_root ) } } /// Returns whether it is valid to call `.merge()`, i.e., whether there is enough room in /// a node to hold the combination of the nodes to the left and right of this handle along /// with the key/value pair at this handle. pub fn can_merge(&self) -> bool { ( self.reborrow() .left_edge() .descend() .len() + self.reborrow() .right_edge() .descend() .len() + 1 ) <= CAPACITY } /// Combines the node immediately to the left of this handle, the key/value pair pointed /// to by this handle, and the node immediately to the right of this handle into one new /// child of the underlying node, returning an edge referencing that new child. /// /// Assumes that this edge `.can_merge()`. pub fn merge(mut self) -> Handle, K, V, marker::Internal>, marker::Edge> { let self1 = unsafe { ptr::read(&self) }; let self2 = unsafe { ptr::read(&self) }; let mut left_node = self1.left_edge().descend(); let left_len = left_node.len(); let mut right_node = self2.right_edge().descend(); let right_len = right_node.len(); // necessary for correctness, but in a private module debug_assert!(left_len + right_len + 1 <= CAPACITY); unsafe { ptr::write(left_node.keys_mut().get_unchecked_mut(left_len), slice_remove(self.node.keys_mut(), self.idx)); ptr::copy_nonoverlapping( right_node.keys().as_ptr(), left_node.keys_mut().as_mut_ptr().offset(left_len as isize + 1), right_len ); ptr::write(left_node.vals_mut().get_unchecked_mut(left_len), slice_remove(self.node.vals_mut(), self.idx)); ptr::copy_nonoverlapping( right_node.vals().as_ptr(), left_node.vals_mut().as_mut_ptr().offset(left_len as isize + 1), right_len ); slice_remove(&mut self.node.as_internal_mut().edges, self.idx + 1); for i in self.idx+1..self.node.len() { Handle::new_edge(self.node.reborrow_mut(), i).correct_parent_link(); } self.node.as_leaf_mut().len -= 1; left_node.as_leaf_mut().len += right_len as u16 + 1; if self.node.height > 1 { ptr::copy_nonoverlapping( right_node.cast_unchecked().as_internal().edges.as_ptr(), left_node.cast_unchecked() .as_internal_mut() .edges .as_mut_ptr() .offset(left_len as isize + 1), right_len + 1 ); for i in left_len+1..left_len+right_len+2 { Handle::new_edge( left_node.cast_unchecked().reborrow_mut(), i ).correct_parent_link(); } Global.dealloc( right_node.node.cast(), Layout::new::>(), ); } else { Global.dealloc( right_node.node.cast(), Layout::new::>(), ); } Handle::new_edge(self.node, self.idx) } } /// This removes a key/value pair from the left child and replaces it with the key/value pair /// pointed to by this handle while pushing the old key/value pair of this handle into the right /// child. pub fn steal_left(&mut self) { unsafe { let (k, v, edge) = self.reborrow_mut().left_edge().descend().pop(); let k = mem::replace(self.reborrow_mut().into_kv_mut().0, k); let v = mem::replace(self.reborrow_mut().into_kv_mut().1, v); match self.reborrow_mut().right_edge().descend().force() { ForceResult::Leaf(mut leaf) => leaf.push_front(k, v), ForceResult::Internal(mut internal) => internal.push_front(k, v, edge.unwrap()) } } } /// This removes a key/value pair from the right child and replaces it with the key/value pair /// pointed to by this handle while pushing the old key/value pair of this handle into the left /// child. pub fn steal_right(&mut self) { unsafe { let (k, v, edge) = self.reborrow_mut().right_edge().descend().pop_front(); let k = mem::replace(self.reborrow_mut().into_kv_mut().0, k); let v = mem::replace(self.reborrow_mut().into_kv_mut().1, v); match self.reborrow_mut().left_edge().descend().force() { ForceResult::Leaf(mut leaf) => leaf.push(k, v), ForceResult::Internal(mut internal) => internal.push(k, v, edge.unwrap()) } } } /// This does stealing similar to `steal_left` but steals multiple elements at once. pub fn bulk_steal_left(&mut self, count: usize) { unsafe { let mut left_node = ptr::read(self).left_edge().descend(); let left_len = left_node.len(); let mut right_node = ptr::read(self).right_edge().descend(); let right_len = right_node.len(); // Make sure that we may steal safely. debug_assert!(right_len + count <= CAPACITY); debug_assert!(left_len >= count); let new_left_len = left_len - count; // Move data. { let left_kv = left_node.reborrow_mut().into_kv_pointers_mut(); let right_kv = right_node.reborrow_mut().into_kv_pointers_mut(); let parent_kv = { let kv = self.reborrow_mut().into_kv_mut(); (kv.0 as *mut K, kv.1 as *mut V) }; // Make room for stolen elements in the right child. ptr::copy(right_kv.0, right_kv.0.offset(count as isize), right_len); ptr::copy(right_kv.1, right_kv.1.offset(count as isize), right_len); // Move elements from the left child to the right one. move_kv(left_kv, new_left_len + 1, right_kv, 0, count - 1); // Move parent's key/value pair to the right child. move_kv(parent_kv, 0, right_kv, count - 1, 1); // Move the left-most stolen pair to the parent. move_kv(left_kv, new_left_len, parent_kv, 0, 1); } left_node.reborrow_mut().as_leaf_mut().len -= count as u16; right_node.reborrow_mut().as_leaf_mut().len += count as u16; match (left_node.force(), right_node.force()) { (ForceResult::Internal(left), ForceResult::Internal(mut right)) => { // Make room for stolen edges. let right_edges = right.reborrow_mut().as_internal_mut().edges.as_mut_ptr(); ptr::copy(right_edges, right_edges.offset(count as isize), right_len + 1); right.correct_childrens_parent_links(count, count + right_len + 1); move_edges(left, new_left_len + 1, right, 0, count); }, (ForceResult::Leaf(_), ForceResult::Leaf(_)) => { } _ => { unreachable!(); } } } } /// The symmetric clone of `bulk_steal_left`. pub fn bulk_steal_right(&mut self, count: usize) { unsafe { let mut left_node = ptr::read(self).left_edge().descend(); let left_len = left_node.len(); let mut right_node = ptr::read(self).right_edge().descend(); let right_len = right_node.len(); // Make sure that we may steal safely. debug_assert!(left_len + count <= CAPACITY); debug_assert!(right_len >= count); let new_right_len = right_len - count; // Move data. { let left_kv = left_node.reborrow_mut().into_kv_pointers_mut(); let right_kv = right_node.reborrow_mut().into_kv_pointers_mut(); let parent_kv = { let kv = self.reborrow_mut().into_kv_mut(); (kv.0 as *mut K, kv.1 as *mut V) }; // Move parent's key/value pair to the left child. move_kv(parent_kv, 0, left_kv, left_len, 1); // Move elements from the right child to the left one. move_kv(right_kv, 0, left_kv, left_len + 1, count - 1); // Move the right-most stolen pair to the parent. move_kv(right_kv, count - 1, parent_kv, 0, 1); // Fix right indexing ptr::copy(right_kv.0.offset(count as isize), right_kv.0, new_right_len); ptr::copy(right_kv.1.offset(count as isize), right_kv.1, new_right_len); } left_node.reborrow_mut().as_leaf_mut().len += count as u16; right_node.reborrow_mut().as_leaf_mut().len -= count as u16; match (left_node.force(), right_node.force()) { (ForceResult::Internal(left), ForceResult::Internal(mut right)) => { move_edges(right.reborrow_mut(), 0, left, left_len + 1, count); // Fix right indexing. let right_edges = right.reborrow_mut().as_internal_mut().edges.as_mut_ptr(); ptr::copy(right_edges.offset(count as isize), right_edges, new_right_len + 1); right.correct_childrens_parent_links(0, new_right_len + 1); }, (ForceResult::Leaf(_), ForceResult::Leaf(_)) => { } _ => { unreachable!(); } } } } } unsafe fn move_kv( source: (*mut K, *mut V), source_offset: usize, dest: (*mut K, *mut V), dest_offset: usize, count: usize) { ptr::copy_nonoverlapping(source.0.offset(source_offset as isize), dest.0.offset(dest_offset as isize), count); ptr::copy_nonoverlapping(source.1.offset(source_offset as isize), dest.1.offset(dest_offset as isize), count); } // Source and destination must have the same height. unsafe fn move_edges( mut source: NodeRef, source_offset: usize, mut dest: NodeRef, dest_offset: usize, count: usize) { let source_ptr = source.as_internal_mut().edges.as_mut_ptr(); let dest_ptr = dest.as_internal_mut().edges.as_mut_ptr(); ptr::copy_nonoverlapping(source_ptr.offset(source_offset as isize), dest_ptr.offset(dest_offset as isize), count); dest.correct_childrens_parent_links(dest_offset, dest_offset + count); } impl Handle, HandleType> { /// Check whether the underlying node is an `Internal` node or a `Leaf` node. pub fn force(self) -> ForceResult< Handle, HandleType>, Handle, HandleType> > { match self.node.force() { ForceResult::Leaf(node) => ForceResult::Leaf(Handle { node, idx: self.idx, _marker: PhantomData }), ForceResult::Internal(node) => ForceResult::Internal(Handle { node, idx: self.idx, _marker: PhantomData }) } } } impl<'a, K, V> Handle, K, V, marker::LeafOrInternal>, marker::Edge> { /// Move the suffix after `self` from one node to another one. `right` must be empty. /// The first edge of `right` remains unchanged. pub fn move_suffix(&mut self, right: &mut NodeRef, K, V, marker::LeafOrInternal>) { unsafe { let left_new_len = self.idx; let mut left_node = self.reborrow_mut().into_node(); let right_new_len = left_node.len() - left_new_len; let mut right_node = right.reborrow_mut(); debug_assert!(right_node.len() == 0); debug_assert!(left_node.height == right_node.height); let left_kv = left_node.reborrow_mut().into_kv_pointers_mut(); let right_kv = right_node.reborrow_mut().into_kv_pointers_mut(); move_kv(left_kv, left_new_len, right_kv, 0, right_new_len); left_node.reborrow_mut().as_leaf_mut().len = left_new_len as u16; right_node.reborrow_mut().as_leaf_mut().len = right_new_len as u16; match (left_node.force(), right_node.force()) { (ForceResult::Internal(left), ForceResult::Internal(right)) => { move_edges(left, left_new_len + 1, right, 1, right_new_len); }, (ForceResult::Leaf(_), ForceResult::Leaf(_)) => { } _ => { unreachable!(); } } } } } pub enum ForceResult { Leaf(Leaf), Internal(Internal) } pub enum InsertResult<'a, K, V, Type> { Fit(Handle, K, V, Type>, marker::KV>), Split(NodeRef, K, V, Type>, K, V, Root) } pub mod marker { use core::marker::PhantomData; pub enum Leaf { } pub enum Internal { } pub enum LeafOrInternal { } pub enum Owned { } pub struct Immut<'a>(PhantomData<&'a ()>); pub struct Mut<'a>(PhantomData<&'a mut ()>); pub enum KV { } pub enum Edge { } } unsafe fn slice_insert(slice: &mut [T], idx: usize, val: T) { ptr::copy( slice.as_ptr().offset(idx as isize), slice.as_mut_ptr().offset(idx as isize + 1), slice.len() - idx ); ptr::write(slice.get_unchecked_mut(idx), val); } unsafe fn slice_remove(slice: &mut [T], idx: usize) -> T { let ret = ptr::read(slice.get_unchecked(idx)); ptr::copy( slice.as_ptr().offset(idx as isize + 1), slice.as_mut_ptr().offset(idx as isize), slice.len() - idx - 1 ); ret }