org.organicdesign.fp.collections.RrbTree Maven / Gradle / Ivy
// Copyright 2016-05-28 PlanBase Inc. & Glen Peterson
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
package org.organicdesign.fp.collections;
import java.io.IOException;
import java.io.InvalidObjectException;
import java.io.ObjectInputStream;
import java.io.ObjectOutputStream;
import java.io.Serializable;
import java.util.Arrays;
import java.util.List;
import org.organicdesign.fp.tuple.Tuple2;
import org.organicdesign.fp.tuple.Tuple4;
import static org.organicdesign.fp.collections.Cowry.*;
import static org.organicdesign.fp.collections.Indented.arrayString;
import static org.organicdesign.fp.collections.Indented.indentSpace;
/**
An RRB Tree is an immutable List (like Clojure's PersistentVector) that also supports random inserts, deletes,
and can be split and joined back together in logarithmic time.
This is based on the paper, "RRB-Trees: Efficient Immutable Vectors" by Phil Bagwell and
Tiark Rompf, with the following differences:
- The Relaxed nodes can be sized between n/3 and 2n/3 (Bagwell/Rompf specify n and n-1)
- The Join operation sticks the shorter tree unaltered into the larger tree (except for very
small trees which just get concatenated).
Details were filled in from the Cormen, Leiserson, Rivest & Stein Algorithms book entry
on B-Trees. Also with an awareness of the Clojure PersistentVector by Rich Hickey. All errors
are by Glen Peterson.
History (what little I know):
1972: B-Tree: Rudolf Bayer and Ed McCreight
1998: Purely Functional Data Structures: Chris Okasaki
2007: Clojure's Persistent Vector (and HashMap) implementations: Rich Hickey
2012: RRB-Tree: Phil Bagwell and Tiark Rompf
Compared to other collections (timings summary from 2017-06-11):
- append() - {@link ImRrbt} varies between 90% and 100% of the speed of {@link PersistentVector} (biggest difference above 100K).
{@link MutableRrbt} varies between 45% and 80% of the speed of
{@link org.organicdesign.fp.collections.PersistentVector.MutableVector} (biggest difference from 100 to 1M).
- get() - varies between 50% and 150% of the speed of PersistentVector (PV wins above 1K) if you build RRB using append().
If you build rrb using random inserts (worst case), it goes from 90% at 10 items down to 15% of the speed of the PV at 1M items.
- iterate() - is about the same speed as PersistentVector
- insert(0, item) - beats ArrayList above 1K items (worse than ArrayList below 100 items).
- insert(random, item) - beats ArrayList above 10K items (worse than ArrayList until then).
- O(log n) split(), join(), and remove() (not timed yet).
Latest detailed timing results are
here.
*/
@SuppressWarnings("WeakerAccess")
public abstract class RrbTree implements BaseList, Indented {
// Focus is like the tail in Rich Hickey's Persistent Vector, but named after the structure
// in Scala's implementation. Tail and focus are both designed to allow repeated appends or
// inserts to the same area of a vector to be done in constant time. Tail only handles appends
// but this can handle repeated inserts to any area of a vector.
/** Mutable version of an {@link RrbTree}. Timing information is available there. */
public static class MutableRrbt extends RrbTree implements MutableList {
private E[] focus;
private int focusStartIndex;
private int focusLength;
private Node root;
private int size;
MutableRrbt(E[] f, int fi, int fl, Node r, int s) {
focus = f; focusStartIndex = fi; focusLength = fl; root = r; size = s;
}
/** {@inheritDoc} */
@SuppressWarnings("unchecked")
@Override public MutableRrbt append(E val) {
// If our focus isn't set up for appends or if it's full, insert it into the data structure
// where it belongs. Then make a new focus
if ( (focusLength >= STRICT_NODE_LENGTH) ||
((focusLength > 0) &&
(focusStartIndex < (size - focusLength))) ) {
root = root.pushFocus(focusStartIndex, arrayCopy(focus, focusLength, null));
focus = (E[]) new Object[STRICT_NODE_LENGTH];
focus[0] = val;
focusStartIndex = size;
focusLength = 1;
size++;
return this;
}
// TODO: 3. Make the root the first argument to RrbTree, MutableRrbt and ImRrbt.
if (focus.length <= focusLength) {
focus = arrayCopy(focus, STRICT_NODE_LENGTH, null);
}
focus[focusLength] = val;
focusLength++;
size++;
return this;
}
/** {@inheritDoc} */
@Override public MutableRrbt concat(Iterable es) {
return (MutableRrbt) MutableList.super.concat(es);
}
void debugValidate() {
if (focusLength > STRICT_NODE_LENGTH) {
throw new IllegalStateException("focus len:" + focusLength +
" gt STRICT_NODE_LENGTH:" + STRICT_NODE_LENGTH +
"\n" + this.indentedStr(0));
}
int sz = root.debugValidate();
if (sz != size - focusLength) {
throw new IllegalStateException("Size incorrect. Root size: " + root.size() +
" RrbSize: " + size +
" focusLen: " + focusLength + "\n" +
this.indentedStr(0));
}
if ( (focusStartIndex < 0) || (focusStartIndex > size) ) {
throw new IllegalStateException("focusStartIndex out of bounds!\n" +
this.indentedStr(0));
}
if (!root.equals(eliminateUnnecessaryAncestors(root))) {
throw new IllegalStateException("Unnecessary ancestors!\n" +
this.indentedStr(0));
}
}
/** {@inheritDoc} */
@Override public E get(int i) {
if ( (i < 0) || (i > size) ) {
throw new IndexOutOfBoundsException("Index: " + i + " size: " + size);
}
// This is a debugging assertion - can't be covered by a test.
// if ( (focusStartIndex < 0) || (focusStartIndex > size) ) {
// throw new IllegalStateException("focusStartIndex: " + focusStartIndex +
// " size: " + size);
// }
if (i >= focusStartIndex) {
int focusOffset = i - focusStartIndex;
if (focusOffset < focusLength) {
return focus[focusOffset];
}
i -= focusLength;
}
return root.get(i);
}
/** {@inheritDoc} */
@Override public ImRrbt immutable() {
return new ImRrbt<>(arrayCopy(focus, focusLength, null),
focusStartIndex,
root, size);
}
/** {@inheritDoc} */
@Override public String indentedStr(int indent) {
return "RrbTree(size=" + size +
" fsi=" + focusStartIndex +
" focus=" + arrayString(focus) + "\n" +
indentSpace(indent + 8) + "root=" +
(root == null ? "null" : root.indentedStr(indent + 13)) +
")";
}
/** {@inheritDoc} */
@Override public MutableRrbt insert(int idx, E element) {
// If the focus is full, push it into the tree and make a new one with the new element.
if (focusLength >= STRICT_NODE_LENGTH) {
root = root.pushFocus(focusStartIndex,
arrayCopy(focus, focusLength, null));
focus = singleElementArray(element);
focusStartIndex = idx;
focusLength = 1;
size++;
return this;
}
// If we have no focus, add a new one at the ideal spot.
// TODO: Make sure Immutable does this too.
if (focusLength == 0) {
focus = singleElementArray(element);
focusStartIndex = idx;
focusLength = 1;
size++;
return this;
}
// If the index is within the focus, add the item there.
int diff = idx - focusStartIndex;
if ( (diff >= 0) && (diff <= focusLength) ) {
// Here focus length cannot be zero!
// We want to double the length each time up to STRICT_NODE_LENGTH
// because there is no guarantee that the next insert will be in the same
// place, so this hedges our bets.
if (focus.length <= focusLength) {
int newLen = (focusLength >= HALF_STRICT_NODE_LENGTH)
? STRICT_NODE_LENGTH
: focusLength << 1; // double size.
focus = arrayCopy(focus, newLen, null);
}
// Shift existing items past insertion index to the right
int numItemsToShift = focusLength - diff;
// src, srcPos, dest, destPos, length
if (numItemsToShift > 0) {
System.arraycopy(focus, diff, focus, diff + 1, numItemsToShift);
}
// Put new item into the focus.
focus[diff] = element;
focusLength++;
size++;
return this;
}
// Here we are left with an insert somewhere else than the current focus.
// Here the mutable version has a focus that's longer than the number of items used,
// So we need to shorten it before pushing it into the tree.
if (focusLength > 0) {
root = root.pushFocus(focusStartIndex, arrayCopy(focus, focusLength, null));
}
focus = singleElementArray(element);
focusStartIndex = idx;
focusLength = 1;
size++;
return this;
}
/** {@inheritDoc} */
@Override public UnmodSortedIterator iterator() {
return new Iter(pushFocus());
}
/** {@inheritDoc} */
@Override Node pushFocus() {
return (focusLength == 0)
? root
: root.pushFocus(focusStartIndex,
arrayCopy(focus, focusLength, null));
}
/** {@inheritDoc} */
@Override public String toString() {
return UnmodIterable.toString("MutableRrbt", this);
}
/**
Joins the given tree to the right side of this tree (or this to the left side of that one) in
something like O(log n) time.
*/
@SuppressWarnings("unchecked")
public RrbTree join(RrbTree that) {
// We don't want to wonder below if we're inserting leaves or branch-nodes.
// Also, it leaves the tree cleaner to just smash leaves onto the bigger tree.
// Ultimately, we might want to see if we can grab the tail and stick it where it belongs
// but for now, this should be alright.
if (that.size() < MAX_NODE_LENGTH) {
return concat(that);
}
if (this.size < MAX_NODE_LENGTH) {
for (int i = 0; i < size; i++) {
that = that.insert(i, this.get(i));
}
return that;
}
// OK, here we've eliminated the case of merging a leaf into a tree. We only have to
// deal with tree-into-tree merges below.
//
// Note that if the right-hand tree is bigger, we'll effectively add this tree to the
// left-hand side of that one. It's logically the same as adding that tree to the right
// of this, but the mechanism by which it happens is a little different.
Node leftRoot = pushFocus();
Node rightRoot = that.pushFocus();
// if (leftRoot != eliminateUnnecessaryAncestors(leftRoot)) {
// throw new IllegalStateException("Left had unnecessary ancestors!");
// }
// if (rightRoot != eliminateUnnecessaryAncestors(rightRoot)) {
// throw new IllegalStateException("Right had unnecessary ancestors!");
// }
// Whether to add the right tree to the left one (true) or vice-versa (false).
// True also means left is taller, false: right is taller.
boolean leftIntoRight = leftRoot.height() < rightRoot.height();
Node taller = leftIntoRight ? rightRoot : leftRoot;
Node shorter = leftIntoRight ? leftRoot : rightRoot;
// Most compact: Descend the taller tree to shorter.height and find room for all
// shorter children as children of that node.
//
// Next: add the shorter node, unchanged, as a child to the taller tree at
// shorter.height + 1
//
// If that level of the taller tree is full, add an ancestor to the shorter node and try to
// fit at the next level up in the taller tree.
//
// If this brings us to the top of the taller tree (both trees are the same height), add a
// new parent node with leftRoot and rightRoot as children
// Walk down the taller tree to one below the shorter, remembering ancestors.
Node n = taller;
// This is the maximum we can descend into the taller tree (before running out of tree)
// int maxDescent = taller.height() - 1;
// Actual amount we're going to descend.
int descentDepth = taller.height() - shorter.height();
// if ( (descentDepth < 0) || (descentDepth >= taller.height()) ) {
// throw new IllegalStateException("Illegal descent depth: " + descentDepth);
// }
Node[] ancestors = genericNodeArray(descentDepth);
int i = 0;
for (; i < ancestors.length; i++) {
// Add an ancestor to array
ancestors[i] = n;
// if (n instanceof Leaf) {
// throw new IllegalStateException("Somehow found a leaf node");
// }
n = n.endChild(leftIntoRight);
}
// i is incremented before leaving the loop, so decrement it here to make it point
// to ancestors.length - 1;
i--;
// if (n.height() != shorter.height()) {
// throw new IllegalStateException("Didn't get to proper height");
// }
// Most compact: Descend the taller tree to shorter.height and find room for all
// shorter children as children of that node.
if (n.thisNodeHasRelaxedCapacity(shorter.numChildren())) {
// Adding kids of shorter to proper level of taller...
Node[] kids;
if (shorter instanceof Strict) {
kids = ((Strict) shorter).nodes;
} else if (shorter instanceof Relaxed) {
kids = ((Relaxed) shorter).nodes;
} else {
throw new IllegalStateException("Expected a strict or relaxed, but found " +
shorter.getClass());
}
n = n.addEndChildren(leftIntoRight, kids);
}
if (i >= 0) {
// Go back up one after lowest check.
n = ancestors[i];
i--;
// if (n.height() != shorter.height() + 1) {
// throw new IllegalStateException("Didn't go back up enough");
// }
}
// TODO: Is this used?
// While nodes in the taller are full, add a parent to the shorter and try the next level
// up.
while (!n.thisNodeHasRelaxedCapacity(1) &&
(i >= 0) ) {
// no room for short at this level (n has too many kids)
n = ancestors[i];
i--;
shorter = addAncestor(shorter);
// shorter.debugValidate();
// Sometimes we care about which is shorter and sometimes about left and right.
// Since we fixed the shorter tree, we have to update the left/right
// pointer to point to the new shorter.
if (leftIntoRight) {
leftRoot = shorter;
} else {
rightRoot = shorter;
}
}
// Here we either have 2 trees of equal height, or
// we have room in n for the shorter as a child.
if (shorter.height() == (n.height() - 1)) {
// if (!n.thisNodeHasRelaxedCapacity(1)) {
// throw new IllegalStateException("somehow got here without relaxed capacity...");
// }
// Shorter one level below n and there's room
// Trees are not equal height and there's room somewhere.
n = n.addEndChild(leftIntoRight, shorter);
// n.debugValidate();
} else if (i < 0) {
// 2 trees of equal height so we make a new parent
// if (shorter.height() != n.height()) {
// throw new IllegalStateException("Expected trees of equal height");
// }
@SuppressWarnings("unchecked")
Node[] newRootArray = new Node[] {leftRoot, rightRoot};
int leftSize = leftRoot.size();
Node newRoot =
new Relaxed<>(new int[] {leftSize, leftSize + rightRoot.size()}, newRootArray);
// newRoot.debugValidate();
return new MutableRrbt<>(emptyArray(), 0, 0, newRoot, newRoot.size());
} else {
throw new IllegalStateException("How did we get here?");
}
// We've merged the nodes. Now see if we need to create new parents
// to hold the changed sub-nodes...
while (i >= 0) {
Node anc = ancestors[i];
// By definition, I think that if we need a new root node, then we aren't dealing with
// leaf nodes, but I could be wrong.
// I also think we should get rid of relaxed nodes and everything will be much
// easier.
Relaxed rel = (anc instanceof Strict) ? ((Strict) anc).relax()
: (Relaxed) anc;
int repIdx = leftIntoRight ? 0 : rel.numChildren() - 1;
n = Relaxed.replaceInRelaxedAt(rel.cumulativeSizes, rel.nodes, n, repIdx,
n.size() - rel.nodes[repIdx].size());
i--;
}
// n.debugValidate();
return new MutableRrbt<>(emptyArray(), 0, 0, n, n.size());
}
/** {@inheritDoc} */
@Override public MutableRrbt replace(int index, E item) {
if ( (index < 0) || (index > size) ) {
throw new IndexOutOfBoundsException("Index: " + index + " size: " + size);
}
if (index >= focusStartIndex) {
int focusOffset = index - focusStartIndex;
if (focusOffset < focusLength) {
focus[focusOffset] = item;
return this;
}
index -= focusLength;
}
// About to do replace with maybe-adjusted index
root = root.replace(index, item);
return this;
}
/** {@inheritDoc} */
public MutableRrbt without(int index) { return (MutableRrbt) super.without(index); }
@Override public int size() { return size; }
/**
Divides this RRB-Tree such that every index less-than the given index ends up in the
left-hand tree and the indexed item and all subsequent ones end up in the right-hand tree.
@param splitIndex the split point (excluded from the left-tree, included in the right one)
@return two new sub-trees as determined by the split point. If the point is 0 or
this.size() one tree will be empty (but never null).
*/
public Tuple2,MutableRrbt> split(int splitIndex) {
if ( (splitIndex < 1) || (splitIndex > size) ) {
throw new IndexOutOfBoundsException(
"Constraint violation failed: 1 <= splitIndex <= size");
}
// Push the focus before splitting.
Node newRoot = pushFocus();
// If a leaf-node is split, the fragments become the new focus for each side of the split.
// Otherwise, the focus can be left empty, or the last node of each side can be made into
// the focus.
SplitNode split = newRoot.splitAt(splitIndex);
// split.left().debugValidate();
// split.right().debugValidate();
E[] lFocus = split.leftFocus();
Node left = eliminateUnnecessaryAncestors(split.left());
E[] rFocus = split.rightFocus();
Node right = eliminateUnnecessaryAncestors(split.right());
// These branches are identical, just different classes.
return Tuple2.of(new MutableRrbt<>(lFocus, left.size(), lFocus.length,
left, left.size() + lFocus.length),
new MutableRrbt<>(rFocus, 0, rFocus.length,
right, right.size() + rFocus.length));
}
}
/** Immutable version of an {@link RrbTree}. Timing information is available there. */
public static class ImRrbt extends RrbTree implements ImList, Serializable {
private final E[] focus;
private final int focusStartIndex;
private final Node root;
private final int size;
ImRrbt(E[] f, int fi, Node r, int s) {
focus = f; focusStartIndex = fi; root = r; size = s;
}
// ===================================== Serialization =====================================
// This class has a custom serialized form designed to be as small as possible. It does not
// have the same internal structure as an instance of this class.
// For serializable. Make sure to change whenever internal data format changes.
private static final long serialVersionUID = 20170625165600L;
// Check out Josh Bloch Item 78, p. 312 for an explanation of what's going on here.
private static class SerializationProxy implements Serializable {
// For serializable. Make sure to change whenever internal data format changes.
private static final long serialVersionUID = 20160904155600L;
private final int size;
private transient RrbTree rrbTree;
SerializationProxy(RrbTree v) {
size = v.size();
rrbTree = v;
}
// Taken from Josh Bloch Item 75, p. 298
private void writeObject(ObjectOutputStream s) throws IOException {
s.defaultWriteObject();
// Write out all elements in the proper order
for (E entry : rrbTree) {
s.writeObject(entry);
}
}
@SuppressWarnings("unchecked")
private void readObject(ObjectInputStream s) throws IOException, ClassNotFoundException {
s.defaultReadObject();
MutableRrbt temp = emptyMutable();
for (int i = 0; i < size; i++) {
temp.append((E) s.readObject());
}
rrbTree = temp.immutable();
}
private Object readResolve() { return rrbTree; }
}
private Object writeReplace() { return new SerializationProxy<>(this); }
private void readObject(java.io.ObjectInputStream in) throws IOException,
ClassNotFoundException {
throw new InvalidObjectException("Proxy required");
}
// =================================== Instance Methods ===================================
/** {@inheritDoc} */
@Override public ImRrbt append(E val) {
// If our focus isn't set up for appends or if it's full, insert it into the data
// structure where it belongs. Then make a new focus
if ( (focus.length >= STRICT_NODE_LENGTH) ||
((focus.length > 0) &&
(focusStartIndex < (size - focus.length))) ) {
Node newRoot = root.pushFocus(focusStartIndex, focus);
return new ImRrbt<>(singleElementArray(val), size, newRoot,
size + 1);
}
return new ImRrbt<>(insertIntoArrayAt(val, focus, focus.length, null),
focusStartIndex, root,
size + 1);
}
/** {@inheritDoc} */
@Override public ImRrbt concat(Iterable es) {
return this.mutable().concat(es).immutable();
}
void debugValidate() {
if (focus.length > STRICT_NODE_LENGTH) {
throw new IllegalStateException("focus len:" + focus.length +
" gt STRICT_NODE_LENGTH:" + STRICT_NODE_LENGTH +
"\n" + this.indentedStr(0));
}
int sz = root.debugValidate();
if (sz != size - focus.length) {
throw new IllegalStateException("Size incorrect. Root size: " + root.size() +
" RrbSize: " + size +
" focusLen: " + focus.length + "\n" +
this.indentedStr(0));
}
if ( (focusStartIndex < 0) || (focusStartIndex > size) ) {
throw new IllegalStateException("focusStartIndex out of bounds!\n" +
this.indentedStr(0));
}
if (!root.equals(eliminateUnnecessaryAncestors(root))) {
throw new IllegalStateException("Unnecessary ancestors!\n" +
this.indentedStr(0));
}
}
/** {@inheritDoc} */
@Override public E get(int i) {
if ( (i < 0) || (i > size) ) {
throw new IndexOutOfBoundsException("Index: " + i + " size: " + size);
}
// This is a debugging assertion - can't be covered by a test.
// if ( (focusStartIndex < 0) || (focusStartIndex > size) ) {
// throw new IllegalStateException("focusStartIndex: " + focusStartIndex +
// " size: " + size);
// }
if (i >= focusStartIndex) {
int focusOffset = i - focusStartIndex;
if (focusOffset < focus.length) {
return focus[focusOffset];
}
i -= focus.length;
}
return root.get(i);
}
/** {@inheritDoc} */
@Override public ImRrbt insert(int idx, E element) {
// If the focus is full, push it into the tree and make a new one with the new element.
if (focus.length >= STRICT_NODE_LENGTH) {
Node newRoot = root.pushFocus(focusStartIndex, focus);
E[] newFocus = singleElementArray(element);
return new ImRrbt<>(newFocus, idx, newRoot, size + 1);
}
// If the index is within the focus, add the item there.
int diff = idx - focusStartIndex;
if ( (diff >= 0) && (diff <= focus.length) ) {
// new focus
E[] newFocus = insertIntoArrayAt(element, focus, diff, null);
return new ImRrbt<>(newFocus, focusStartIndex, root, size + 1) ;
}
// Here we are left with an insert somewhere else than the current focus.
Node newRoot = focus.length > 0 ? root.pushFocus(focusStartIndex,
focus)
: root;
E[] newFocus = singleElementArray(element);
return new ImRrbt<>(newFocus, idx, newRoot, size + 1);
}
/** {@inheritDoc} */
@Override public MutableRrbt mutable() {
// TODO: Should we defensively copy the root as well?
return new MutableRrbt<>(arrayCopy(focus, focus.length, null),
focusStartIndex, focus.length,
root, size);
}
/** {@inheritDoc} */
@Override public UnmodSortedIterator iterator() {
return new Iter(pushFocus());
}
/** {@inheritDoc} */
@Override Node pushFocus() {
return (focus.length == 0)
? root
: root.pushFocus(focusStartIndex, focus);
}
/**
Joins the given tree to the right side of this tree (or this to the left side of that one)
in something like O(log n) time.
*/
@SuppressWarnings("unchecked")
public RrbTree join(RrbTree that) {
// We don't want to wonder below if we're inserting leaves or branch-nodes.
// Also, it leaves the tree cleaner to just smash leaves onto the bigger tree.
// Ultimately, we might want to see if we can grab the tail and stick it where it
// belongs but for now, this should be alright.
if (that.size() < MAX_NODE_LENGTH) {
return concat(that);
}
if (this.size < MAX_NODE_LENGTH) {
for (int i = 0; i < size; i++) {
that = that.insert(i, this.get(i));
}
return that;
}
// OK, here we've eliminated the case of merging a leaf into a tree. We only have to
// deal with tree-into-tree merges below.
//
// Note that if the right-hand tree is bigger, we'll effectively add this tree to the
// left-hand side of that one. It's logically the same as adding that tree to the right
// of this, but the mechanism by which it happens is a little different.
Node leftRoot = pushFocus();
Node rightRoot = that.pushFocus();
// if (leftRoot != eliminateUnnecessaryAncestors(leftRoot)) {
// throw new IllegalStateException("Left had unnecessary ancestors!");
// }
// if (rightRoot != eliminateUnnecessaryAncestors(rightRoot)) {
// throw new IllegalStateException("Right had unnecessary ancestors!");
// }
// Whether to add the right tree to the left one (true) or vice-versa (false).
// True also means left is taller, false: right is taller.
boolean leftIntoRight = leftRoot.height() < rightRoot.height();
Node taller = leftIntoRight ? rightRoot : leftRoot;
Node shorter = leftIntoRight ? leftRoot : rightRoot;
// Most compact: Descend the taller tree to shorter.height and find room for all
// shorter children as children of that node.
//
// Next: add the shorter node, unchanged, as a child to the taller tree at
// shorter.height + 1
//
// If that level of the taller tree is full, add an ancestor to the shorter node and try to
// fit at the next level up in the taller tree.
//
// If this brings us to the top of the taller tree (both trees are the same height), add a
// new parent node with leftRoot and rightRoot as children
// Walk down the taller tree to one below the shorter, remembering ancestors.
Node n = taller;
// This is the maximum we can descend into the taller tree (before running out of tree)
// int maxDescent = taller.height() - 1;
// Actual amount we're going to descend.
int descentDepth = taller.height() - shorter.height();
// if ( (descentDepth < 0) || (descentDepth >= taller.height()) ) {
// throw new IllegalStateException("Illegal descent depth: " + descentDepth);
// }
Node[] ancestors = genericNodeArray(descentDepth);
int i = 0;
for (; i < ancestors.length; i++) {
// Add an ancestor to array
ancestors[i] = n;
// if (n instanceof Leaf) {
// throw new IllegalStateException("Somehow found a leaf node");
// }
n = n.endChild(leftIntoRight);
}
// i is incremented before leaving the loop, so decrement it here to make it point
// to ancestors.length - 1;
i--;
// if (n.height() != shorter.height()) {
// throw new IllegalStateException("Didn't get to proper height");
// }
// Most compact: Descend the taller tree to shorter.height and find room for all
// shorter children as children of that node.
if (n.thisNodeHasRelaxedCapacity(shorter.numChildren())) {
// Adding kids of shorter to proper level of taller...
Node[] kids;
if (shorter instanceof Strict) {
kids = ((Strict) shorter).nodes;
} else if (shorter instanceof Relaxed) {
kids = ((Relaxed) shorter).nodes;
} else {
throw new IllegalStateException("Expected a strict or relaxed, but found " +
shorter.getClass());
}
n = n.addEndChildren(leftIntoRight, kids);
}
if (i >= 0) {
// Go back up one after lowest check.
n = ancestors[i];
i--;
// if (n.height() != shorter.height() + 1) {
// throw new IllegalStateException("Didn't go back up enough");
// }
}
// TODO: Is this used?
// While nodes in the taller are full, add a parent to the shorter and try the next level
// up.
while (!n.thisNodeHasRelaxedCapacity(1) &&
(i >= 0) ) {
// no room for short at this level (n has too many kids)
n = ancestors[i];
i--;
shorter = addAncestor(shorter);
// shorter.debugValidate();
// Sometimes we care about which is shorter and sometimes about left and right.
// Since we fixed the shorter tree, we have to update the left/right
// pointer to point to the new shorter.
if (leftIntoRight) {
leftRoot = shorter;
} else {
rightRoot = shorter;
}
}
// Here we either have 2 trees of equal height, or
// we have room in n for the shorter as a child.
if (shorter.height() == (n.height() - 1)) {
// if (!n.thisNodeHasRelaxedCapacity(1)) {
// throw new IllegalStateException("somehow got here without relaxed capacity...");
// }
// Shorter one level below n and there's room
// Trees are not equal height and there's room somewhere.
n = n.addEndChild(leftIntoRight, shorter);
// n.debugValidate();
} else if (i < 0) {
// 2 trees of equal height so we make a new parent
// if (shorter.height() != n.height()) {
// throw new IllegalStateException("Expected trees of equal height");
// }
@SuppressWarnings("unchecked")
Node[] newRootArray = new Node[] {leftRoot, rightRoot};
int leftSize = leftRoot.size();
Node newRoot =
new Relaxed<>(new int[] {leftSize, leftSize + rightRoot.size()}, newRootArray);
// newRoot.debugValidate();
return new ImRrbt<>(emptyArray(), 0, newRoot, newRoot.size());
} else {
throw new IllegalStateException("How did we get here?");
}
// We've merged the nodes. Now see if we need to create new parents
// to hold the changed sub-nodes...
while (i >= 0) {
Node anc = ancestors[i];
// By definition, I think that if we need a new root node, then we aren't dealing with
// leaf nodes, but I could be wrong.
// I also think we should get rid of relaxed nodes and everything will be much
// easier.
Relaxed rel = (anc instanceof Strict) ? ((Strict) anc).relax()
: (Relaxed) anc;
int repIdx = leftIntoRight ? 0 : rel.numChildren() - 1;
n = Relaxed.replaceInRelaxedAt(rel.cumulativeSizes, rel.nodes, n, repIdx,
n.size() - rel.nodes[repIdx].size());
i--;
}
// n.debugValidate();
return new ImRrbt<>(emptyArray(), 0, n, n.size());
}
/** {@inheritDoc} */
@Override public ImRrbt replace(int index, E item) {
if ( (index < 0) || (index > size) ) {
throw new IndexOutOfBoundsException("Index: " + index + " size: " + size);
}
if (index >= focusStartIndex) {
int focusOffset = index - focusStartIndex;
if (focusOffset < focus.length) {
return new ImRrbt<>(replaceInArrayAt(item, focus, focusOffset, null),
focusStartIndex, root, size);
}
index -= focus.length;
}
// About to do replace with maybe-adjusted index
return new ImRrbt<>(focus, focusStartIndex, root.replace(index, item), size);
}
/** {@inheritDoc} */
public ImRrbt without(int index) { return (ImRrbt) super.without(index); }
@Override public int size() { return size; }
/**
Divides this RRB-Tree such that every index less-than the given index ends up in the left-hand
tree and the indexed item and all subsequent ones end up in the right-hand tree.
@param splitIndex the split point (excluded from the left-tree, included in the right one)
@return two new sub-trees as determined by the split point. If the point is 0 or this.size()
one tree will be empty (but never null).
*/
public Tuple2,ImRrbt> split(int splitIndex) {
if ( (splitIndex < 1) || (splitIndex > size) ) {
throw new IndexOutOfBoundsException(
"Constraint violation failed: 1 <= splitIndex <= size");
}
// Push the focus before splitting.
Node newRoot = pushFocus();
// If a leaf-node is split, the fragments become the new focus for each side of the split.
// Otherwise, the focus can be left empty, or the last node of each side can be made into
// the focus.
SplitNode split = newRoot.splitAt(splitIndex);
// split.left().debugValidate();
// split.right().debugValidate();
E[] lFocus = split.leftFocus();
Node left = eliminateUnnecessaryAncestors(split.left());
E[] rFocus = split.rightFocus();
Node right = eliminateUnnecessaryAncestors(split.right());
// These branches are identical, just different classes.
return Tuple2.of(new ImRrbt<>(lFocus, left.size(),
left, left.size() + lFocus.length),
new ImRrbt<>(rFocus, 0,
right, right.size() + rFocus.length));
}
/** {@inheritDoc} */
@Override public String indentedStr(int indent) {
return "RrbTree(size=" + size +
" fsi=" + focusStartIndex +
" focus=" + arrayString(focus) + "\n" +
indentSpace(indent + 8) + "root=" +
(root == null ? "null" : root.indentedStr(indent + 13)) +
")";
}
/** {@inheritDoc} */
@Override public String toString() {
return UnmodIterable.toString("ImRrbt", this);
}
private static final ImRrbt EMPTY_IM_RRBT =
new ImRrbt<>(emptyArray(), 0, emptyLeaf(), 0);
}
/** Returns the empty, immutable RRB-Tree (there is only one) */
@SuppressWarnings("unchecked")
public static ImRrbt empty() { return (ImRrbt) ImRrbt.EMPTY_IM_RRBT; }
/** Returns the empty, mutable RRB-Tree (there is only one) */
@SuppressWarnings("unchecked")
public static MutableRrbt emptyMutable() {
return (MutableRrbt) empty().mutable();
}
// ===================================== Instance Methods =====================================
/** {@inheritDoc} */
@Override abstract public RrbTree append(E t);
/** Internal validation method for testing. */
abstract void debugValidate();
/** {@inheritDoc} */
@Override abstract public E get(int i);
/**
Inserts an item in the RRB tree pushing the current element at that index and all subsequent
elements to the right.
@param idx the insertion point
@param element the item to insert
@return a new RRB-Tree with the item inserted.
*/
@SuppressWarnings("WeakerAccess")
public abstract RrbTree insert(int idx, E element);
/** {@inheritDoc} */
@Override abstract public UnmodSortedIterator iterator();
/*
I'm implementing something like the [Bagwell/Rompf RRB-Tree][1] and I'm a little unsatisfied with
the details of the join/merge algorithm. I wonder if there's a standard way to do this that they
assume that I know (I don't), or if someone has come up with a better way to do this.
I'm thinking the signature is something like:
public RrbTree join(RrbTree that)
My basic approach was to fit the shorter tree into the left-most or right-most leg of the taller
tree at the correct height. For `a.join(b)`, if `b` is taller, fit `a` into `b`'s left-most leg at
the right height, otherwise fit `b` into `a`'s right-most leg at the right height
Overview:
1. Push the focus of both trees so we don't have to worry about it.
2. Find the height of each tree.
3. Stick the shorter tree into the proper level of the larger tree (on the left or right as
appropriate). If the leftmost/rightmost node at the proper level level of the larger tree is full,
add a "skinny leg" (a new root with a single child) to the short tree and stick it on the left or
right one level up in the large one. If the large tree is packed, several skinny-leg insertion
attempts may be required, or even a new root added to the large tree with 2 children: the old
large tree and the smaller tree on the appropriate side of it.
Optimization: If one tree is really small, we could do an append or prepend.
I don't see the advantage of zipping nodes together at every level the very complicated
way it seems to do in the paper. Even after all that work, it still has nodes of varying sizes and
involves changing more nodes than maybe necessary.
[1]: https://infoscience.epfl.ch/record/169879/files/RMTrees.pdf
*/
/**
Joins the given tree to the right side of this tree (or this to the left side of that one) in
something like O(log n) time.
*/
public abstract RrbTree join(RrbTree that);
/** Internal method - do not use. */
abstract Node pushFocus();
/** {@inheritDoc} */
@Override abstract public RrbTree replace(int index, E item);
/** {@inheritDoc} */
@Override abstract public int size();
/**
Divides this RRB-Tree such that every index less-than the given index ends up in the left-hand
tree and the indexed item and all subsequent ones end up in the right-hand tree.
@param splitIndex the split point (excluded from the left-tree, included in the right one)
@return two new sub-trees as determined by the split point. If the point is 0 or this.size()
one tree will be empty (but never null).
*/
abstract public Tuple2,? extends RrbTree> split(int splitIndex);
/**
Returns a new RrbTree minus the given item (all items to the right are shifted left one)
This is O(log n).
*/
public RrbTree without(int index) {
if ( (index > 0) && (index < size() - 1) ) {
Tuple2,? extends RrbTree> s1 = split(index);
Tuple2,? extends RrbTree> s2 = s1._2().split(1);
return s1._1().join(s2._2());
} else if (index == 0) {
return split(1)._2();
} else if (index == size() - 1) {
return split(size() - 1)._1();
} else {
throw new IndexOutOfBoundsException("Failed test: 0 <= index < size");
}
}
private static Node eliminateUnnecessaryAncestors(Node n) {
while ( !(n instanceof Leaf) &&
(n.numChildren() == 1) ) {
n = n.child(0);
}
return n;
}
@SuppressWarnings("unchecked")
private static Node addAncestor(Node n) {
return ( (n instanceof Leaf) &&
(n.size() == STRICT_NODE_LENGTH) ) ? new Strict<>(NODE_LENGTH_POW_2,
n.size(),
(Node[]) new Node[]{n}) :
(n instanceof Strict) ? new Strict<>(((Strict) n).shift + NODE_LENGTH_POW_2,
n.size(),
(Node[]) new Node[]{n}) :
new Relaxed<>(new int[] { n.size() },
(Node[]) new Node[]{n});
}
// ================================== Standard Object Methods ==================================
/** {@inheritDoc} */
@SuppressWarnings("unchecked")
@Override public boolean equals(Object other) {
if (this == other) { return true; }
if ( !(other instanceof List) ) { return false; }
List that = (List) other;
return (this.size() == that.size()) &&
UnmodSortedIterable.equal(this, UnmodSortedIterable.castFromList(that));
}
/** This implementation is correct and compatible with java.util.AbstractList, but O(n). */
@Override public int hashCode() {
int ret = 1;
for (E item : this) {
ret *= 31;
if (item != null) {
ret += item.hashCode();
}
}
return ret;
}
/** {@inheritDoc} */
@Override abstract public String indentedStr(int indent);
// ================================== Implementation Details ==================================
// Definitions:
// Strict: Short for "Strict Radix." Strict nodes have leaf widths of exactly
// STRICT_NODE_LENGTH and are left-filled and packed up to the last full node. This
// lets us use a power of 2 to take advantage of bit shifting to exactly index which
// sub-node an item is found in. Does not support inserts (only appends).
// Relaxed: short for "Relaxed Radix." Relaxed nodes are of somewhat varying sizes, ranging
// from MIN_NODE_LENGTH (Cormen et al calls this "Minimum Degree") to MAX_NODE_LENGTH.
// This requires linear interpolation, a bit of searching, and subtraction to find an
// index into a sub-node, but supports inserts, split, and combine (with another
// RrbTree)
// There's bit shifting going on here because it's a very fast operation.
// Shifting right by 5 is eons faster than dividing by 32.
private static final int NODE_LENGTH_POW_2 = 5; // 2 for testing, 5 for real
// 0b00000000000000000000000000100000 = 0x20 = 32
static final int STRICT_NODE_LENGTH = 1 << NODE_LENGTH_POW_2;
private static final int HALF_STRICT_NODE_LENGTH = STRICT_NODE_LENGTH >> 1;
// (MIN_NODE_LENGTH + MAX_NODE_LENGTH) / 2 should equal STRICT_NODE_LENGTH so that they have the
// same average node size to make the index interpolation easier.
private static final int MIN_NODE_LENGTH = (STRICT_NODE_LENGTH+1) * 2 / 3;
// Always check if less-than this. Never less-than-or-equal. Cormen adds a -1 here and tests
// for <= (I think!).
private static final int MAX_NODE_LENGTH = ( (STRICT_NODE_LENGTH+1) * 4 / 3);
private static final Leaf EMPTY_LEAF = new Leaf<>(EMPTY_ARRAY);
@SuppressWarnings("unchecked")
private static Leaf emptyLeaf() { return (Leaf) EMPTY_LEAF; }
// ================================ Node private inner classes ================================
private interface Node extends Indented {
/** Returns the immediate child node at the given index. */
Node child(int childIdx);
int debugValidate();
/** Returns the leftMost (first) or right-most (last) child */
Node endChild(boolean leftMost);
/** Adds a node as the first/leftmost or last/rightmost child */
Node addEndChild(boolean leftMost, Node shorter);
/** Adds kids as leftmost or rightmost of current children */
Node addEndChildren(boolean leftMost, Node[] newKids);
/** Return the item at the given index */
T get(int i);
/** Returns true if this strict-Radix tree can take another 32 items. */
boolean hasStrictCapacity();
/** Returns the maximum depth below this node. Leaf nodes are height 1. */
int height();
// /** Try to add all sub-nodes to this one. */
// Node join(Node that);
/** Number of items stored in this node */
int size();
// /** Returns true if this node's array is not full */
// boolean thisNodeHasCapacity();
/** Can this node take the specified number of children? */
boolean thisNodeHasRelaxedCapacity(int numItems);
/**
Can we put focus at the given index without reshuffling nodes?
@param index the index we want to insert at
@param size the number of items to insert. Must be size < MAX_NODE_LENGTH
@return true if we can do so without otherwise adjusting the tree.
*/
boolean hasRelaxedCapacity(int index, int size);
// Splitting a strict node yields an invalid Relaxed node (too short).
// We don't yet split Leaf nodes.
// So this needs to only be implemented on Relaxed for now.
// Relaxed[] split();
/** Returns the number of immediate children of this node, not all descendants. */
int numChildren();
// Because we want to append/insert into the focus as much as possible, we will treat
// the insert or append of a single item as a degenerate case. Instead, the primary way
// to add to the internal data structure will be to push the entire focus array into it
Node pushFocus(int index, T[] oldFocus);
Node replace(int idx, T t);
SplitNode splitAt(int splitIndex);
}
// private interface BranchNode extends Node {
// }
// /** For calcCumulativeSizes */
// private static final class CumulativeSizes {
// int szSoFar; // Size so far (of all things to left)
// int srcOffset; // offset in source array
// int[] destArray; // the destination array
// int destPos; // offset in destArray to copy to
// int length; // number of items to copy.
// }
/** This class is the return type when splitting a node. */
private static class SplitNode extends Tuple4,T[],Node,T[]> implements Indented {
/**
Constructor.
@param ln Left-hand whole-node
@param lf Left-focus (leftover items from left node)
@param rn Right-hand whole-node
@param rf Right-focus (leftover items from right node)
*/
SplitNode(Node ln, T[] lf, Node rn, T[] rf) {
super(ln, lf, rn, rf);
// if (lf.length > STRICT_NODE_LENGTH) {
// throw new IllegalStateException("Left focus too long: " + arrayString(lf));
// }
// if (rf.length > STRICT_NODE_LENGTH) {
// throw new IllegalStateException("Right focus too long: " + arrayString(rf));
// }
}
public Node left() { return _1; }
public T[] leftFocus() { return _2; }
public Node right() { return _3; }
public T[] rightFocus() { return _4; }
public int size() { return _1.size() + _2.length + _3.size() + _4.length; }
@Override public String indentedStr(int indent) {
StringBuilder sB = new StringBuilder() // indentSpace(indent)
.append("SplitNode(");
int nextIndent = indent + sB.length();
String nextIndentStr = indentSpace(nextIndent).toString();
return sB.append("left=").append(left().indentedStr(nextIndent + 5)).append(",\n")
.append(nextIndentStr).append("leftFocus=").append(arrayString(leftFocus()))
.append(",\n")
.append(nextIndentStr).append("right=")
.append(right().indentedStr(nextIndent + 6)).append(",\n")
.append(nextIndentStr).append("rightFocus=")
.append(arrayString(rightFocus())).append(")")
.toString();
}
@Override public String toString() { return indentedStr(0); }
}
private static class Leaf implements Node {
final T[] items;
// It can only be Strict if items.length == STRICT_NODE_LENGTH and if its parents
// are strict.
// boolean isStrict;
Leaf(T[] ts) { items = ts; }
@Override public Node child(int childIdx) {
throw new UnsupportedOperationException("Don't call this on a leaf");
}
@Override public int debugValidate() {
if (items.length == 0) {
return 0;
}
if (items.length < MIN_NODE_LENGTH) {
throw new IllegalStateException("Leaf too short!\n" +
this.indentedStr(0));
} else if (items.length >= MAX_NODE_LENGTH) {
throw new IllegalStateException("Leaf too long!\n" +
this.indentedStr(0));
}
return items.length;
}
/** Returns the leftMost (first) or right-most (last) child */
@Override public Node endChild(boolean leftMost) {
throw new UnsupportedOperationException("Don't call this on a leaf");
}
/** Adds a node as the first/leftmost or last/rightmost child */
@Override public Node addEndChild(boolean leftMost, Node shorter) {
throw new UnsupportedOperationException("Don't call this on a leaf");
}
/** Adds kids as leftmost or rightmost of current children */
@Override public Node addEndChildren(boolean leftMost, Node[] newKids) {
throw new UnsupportedOperationException("Don't call this on a leaf");
}
@Override public T get(int i) { return items[i]; }
@Override public int height() { return 1; }
@Override public int size() { return items.length; }
// If we want to add one more to an existing leaf node, it must already be part of a
// relaxed tree.
// public boolean thisNodeHasCapacity() { return items.length < MAX_NODE_LENGTH; }
@Override public boolean hasStrictCapacity() { return false; }
@Override public boolean hasRelaxedCapacity(int index, int size) {
// Appends and prepends need to be a good size, but random inserts do not.
// if ( (size < 1) || (size >= MAX_NODE_LENGTH) ) {
// throw new IllegalArgumentException("Bad size: " + size);
// // + " MIN_NODE_LENGTH=" + MIN_NODE_LENGTH + " MAX_NODE_LENGTH=" + MAX_NODE_LENGTH);
// }
return (items.length + size) < MAX_NODE_LENGTH;
}
@Override
public SplitNode splitAt(int splitIndex) {
// if (splitIndex < 0) {
// throw new IllegalArgumentException("Called splitAt when splitIndex < 0");
// }
// if (splitIndex > items.length - 1) {
// throw new IllegalArgumentException("Called splitAt when splitIndex > orig.length - 1");
// }
// Should we just ensure that the split is between 1 and items.length (exclusive)?
if (splitIndex == 0) {
return new SplitNode<>(emptyLeaf(), emptyArray(), emptyLeaf(), items);
}
if (splitIndex == items.length) {
return new SplitNode<>(emptyLeaf(), items, emptyLeaf(), emptyArray());
}
Tuple2 split = splitArray(items, splitIndex);
T[] splitL = split._1();
T[] splitR = split._2();
Leaf leafL = emptyLeaf();
Leaf leafR = emptyLeaf();
if (splitL.length > STRICT_NODE_LENGTH) {
leafL = new Leaf<>(splitL);
splitL = emptyArray();
}
if (splitR.length > STRICT_NODE_LENGTH) {
leafR = new Leaf<>(splitR);
splitR = emptyArray();
}
return new SplitNode<>(leafL, splitL, leafR, splitR);
}
@SuppressWarnings("unchecked")
private Leaf[] spliceAndSplit(T[] oldFocus, int splitIndex) {
// Consider optimizing:
T[] newItems = spliceIntoArrayAt(oldFocus, items, splitIndex,
(Class) items[0].getClass());
// Shift right one is divide-by 2.
Tuple2 split = splitArray(newItems, newItems.length >> 1);
return new Leaf[] {new Leaf<>(split._1()), new Leaf<>(split._2())};
}
@Override public int numChildren() { return size(); }
// I think this can only be called when the root node is a leaf.
@SuppressWarnings("unchecked")
@Override public Node pushFocus(int index, T[] oldFocus) {
// if (oldFocus.length == 0) {
// throw new IllegalStateException("Never call this with an empty focus!");
// }
// We put the empty Leaf as the root of the empty vector and it stays there
// until the first call to this method, at which point, the oldFocus becomes the
// new root.
if (items.length == 0) {
return new Leaf<>(oldFocus);
}
// Try first to yield a Strict node. For a leaf like this, that means both this node
// and the pushed focus are STRICT_NODE_LENGTH. It also means the old focus is being
// pushed at either the beginning or the end of this node (not anywhere in-between).
if ( (items.length == STRICT_NODE_LENGTH) &&
(oldFocus.length == STRICT_NODE_LENGTH) &&
((index == STRICT_NODE_LENGTH) || (index == 0)) ) {
Leaf[] newNodes = (index == STRICT_NODE_LENGTH)
? new Leaf[] { this,
new Leaf<>(oldFocus)}
: new Leaf[] { new Leaf<>(oldFocus),
this };
// Size is twice STRICT_NODE_LENGTH, so shift left 1 to double.
return new Strict<>(NODE_LENGTH_POW_2, STRICT_NODE_LENGTH << 1, newNodes);
}
if ((items.length + oldFocus.length) < MAX_NODE_LENGTH) {
return new Leaf<>(spliceIntoArrayAt(oldFocus, items, index,
(Class) items[0].getClass()));
}
// We should only get here when the root node is a leaf.
// Maybe we should be more circumspect with our array creation, but for now, just jam
// jam it into one big array, then split it up for simplicity
Leaf[] res = spliceAndSplit(oldFocus, index);
Leaf leftLeaf = res[0];
Leaf rightLeaf = res[1];
int leftSize = leftLeaf.size();
return new Relaxed<>(new int[] { leftSize,
leftSize + rightLeaf.size() },
res);
}
@Override
public Node replace(int idx, T t) {
// if (idx >= size()) {
// throw new IllegalArgumentException("Invalid index " + idx + " >= " + size());
// }
return new Leaf<>(replaceInArrayAt(t, items, idx, null));
}
@Override public boolean thisNodeHasRelaxedCapacity(int numItems) {
// if ( (numItems < 1) || (numItems >= MAX_NODE_LENGTH) ) {
// throw new IllegalArgumentException("Bad size: " + numItems);
// }
return items.length + numItems < MAX_NODE_LENGTH;
}
@Override public String toString() {
// return "Leaf("+ arrayString(items) + ")";
return arrayString(items);
}
@Override public String indentedStr(int indent) {
return arrayString(items);
}
} // end class Leaf
// Contains a left-packed tree of exactly 32-item nodes.
private static class Strict implements Node {
// This is the number of levels below this node (height) times NODE_LENGTH
// For speed, we calculate it as height << NODE_LENGTH_POW_2
// TODO: Can we store shift at the top-level Strict only?
final int shift;
final int size;
// These are the child nodes
final Node[] nodes;
// Constructor
Strict(int sh, int sz, Node[] ns) {
shift = sh; size = sz; nodes = ns;
}
@Override public Node child(int childIdx) { return nodes[childIdx]; }
@Override public int debugValidate() {
if (nodes.length > STRICT_NODE_LENGTH) {
throw new IllegalStateException("Too many child nodes!\n" +
this.indentedStr(0));
}
int sz = 0;
int height = height() - 1;
int sh = shift - NODE_LENGTH_POW_2;
for (int i = 0; i < nodes.length; i++) {
Node n = nodes[i];
if ( !(n instanceof Strict) &&
!(n instanceof Leaf) ) {
throw new IllegalStateException(
"Strict nodes can only have strict or leaf children!\n" +
this.indentedStr(0));
}
if (n.height() != height) {
throw new IllegalStateException("Unequal height! My height = " + height() + "\n" + this.indentedStr(0));
}
if ( (n instanceof Strict) &&
((Strict) n).shift != sh ) {
throw new IllegalStateException(
"Unexpected shift difference between levels!\n" + this.indentedStr(0));
}
if (i < nodes.length - 1) {
if (n.hasStrictCapacity()) {
throw new IllegalStateException("Non-last strict node is not full!\n" +
this.indentedStr(0));
}
if ((n.size() % STRICT_NODE_LENGTH) != 0){
throw new IllegalStateException("Non-last strict node has a weird size!\n" +
this.indentedStr(0));
}
}
if (n instanceof Strict) {
n.debugValidate();
}
sz += n.size();
}
return sz;
}
/** Returns the leftMost (first) or right-most (last) child */
@Override public Node endChild(boolean leftMost) {
return nodes[leftMost ? 0 : nodes.length - 1];
}
/** Adds a node as the first/leftmost or last/rightmost child */
@SuppressWarnings("unchecked")
@Override public Node addEndChild(boolean leftMost, Node shorter) {
if (leftMost || !(shorter instanceof Strict)) {
return relax().addEndChild(leftMost, shorter);
}
return new Strict<>(shift, size + shorter.size(),
insertIntoArrayAt(shorter, nodes, nodes.length, Node.class));
}
/** Adds kids as leftmost or rightmost of current children */
@Override public Node addEndChildren(boolean leftMost, Node[] newKids) {
// if (!thisNodeHasRelaxedCapacity(newKids.length)) {
// throw new IllegalStateException("Can't add enough kids");
// }
return relax().addEndChildren(leftMost, newKids);
}
@Override public int height() {
return (shift / NODE_LENGTH_POW_2) + 1;
}
/**
Returns the highest bits which we use to index into our array - the index of the immediate
child of this node. This is the simplicity (and
speed) of Strict indexing. When everything works, this can be inlined for performance.
This could maybe yield a good guess for Relaxed nodes?
Shifting right by a number is equivalent to dividing by: 2 raised to the power of that
number.
i >> n is equivalent to i / (2^n)
*/
private int highBits(int i) { return i >> shift; }
/**
Returns the low bits of the index (the part Strict sub-nodes need to know about). This
only works because the leaf nodes are all the same size and that size is a power of 2
(the radix). All branch must have the same radix (branching factor or number of immediate
sub-nodes).
Bit shifting is faster than addition or multiplication, but perhaps more importantly, it
means we don't have to store the sizes of the nodes which means we don't have to fetch
those sizes from memory or use up cache space. All of this helps make this data structure
simple and fast.
When everything works, this function can be inlined for performance (if that even helps).
Contrast this with how Relaxed nodes work: they use subtraction instead!
*/
private int lowBits(int i) {
// Little trick: -1 in binary is all ones: 0b11111111111111111111111111111111
// We shift it left, filling the right-most bits with zeros and creating a bit-mask
// with ones on the left and zeros on the right
int shifter = -1 << shift;
// Now we take the inverse so our bit-mask has zeros on the left and ones on the right
int invShifter = ~shifter;
// Finally, we bitwise-and the mask with the index to leave only the low bits.
return i & invShifter;
}
@Override public T get(int i) {
// Find the node indexed by the high bits (for this height).
// Send the low bits on to our sub-nodes.
return nodes[highBits(i)].get(lowBits(i));
}
@Override public int size() { return size; }
// private boolean thisNodeHasCapacity() { return nodes.length < STRICT_NODE_LENGTH; }
@Override public boolean hasStrictCapacity() {
// boolean ret = thisNodeHasCapacity() || nodes[nodes.length - 1].hasStrictCapacity();
// boolean ret2 = highBits(size) != STRICT_NODE_LENGTH;
// if (ret != ret2) {
// System.out.println("size: " + size);
// System.out.println("hasStrict: " + ret);
// System.out.println("highBits(size)" + highBits(size));
// System.out.println("lowBits(size)" + lowBits(size));
// throw new IllegalStateException("Won't work!");
// }
// This works because when a strict node is not full, it's highest child index is
// STRICT_NODE_LENGTH - 1 (which is what highBits(size) will return.
// This used to then walk down the right hand side of the tree checking for room.
// But it turns out that the highest index would have lowBits() equals all ones.
// size is the maxIndex + 1, so that flips the lowest bit of highBits() and makes
// all lowBits() zeros so that the following line works:
return highBits(size) != STRICT_NODE_LENGTH;
}
@Override public boolean hasRelaxedCapacity(int index, int size) {
// if ( (size < 1) || (size >= MAX_NODE_LENGTH) ) {
// throw new IllegalArgumentException("Bad size: " + size);
// }
// It has relaxed capacity because a Relaxed node could have up to MAX_NODE_LENGTH nodes
// and by definition this Strict node has exactly STRICT_NODE_LENGTH items.
return size < MAX_NODE_LENGTH - STRICT_NODE_LENGTH;
}
@SuppressWarnings("unchecked")
@Override
public SplitNode splitAt(int splitIndex) {
// int size = size();
// if ( (splitIndex < 0) || (splitIndex > size) ) {
// throw new IllegalArgumentException("Bad splitIndex: " + splitIndex);
// }
if (splitIndex == 0) {
return new SplitNode<>(emptyLeaf(), emptyArray(), this, emptyArray());
}
if (splitIndex == size) {
return new SplitNode<>(this, emptyArray(), emptyLeaf(), emptyArray());
}
// Not split on a child boundary, so find which child to split and pass it the
// appropriate index.
int subNodeIndex = highBits(splitIndex);
Node subNode = nodes[subNodeIndex];
int subNodeAdjustedIndex = lowBits(splitIndex);
SplitNode split = subNode.splitAt(subNodeAdjustedIndex);
final Node left;
final Node splitLeft = split.left();
if (subNodeIndex == 0) {
left = new Strict<>(shift, splitLeft.size(), new Node[] {splitLeft});
} else {
boolean haveLeft = (splitLeft.size() > 0);
int numLeftItems = subNodeIndex + (haveLeft ? 1 : 0);
Node[] leftNodes = genericNodeArray(numLeftItems);
// Copy one less item if we are going to add the split one in a moment.
// I could have written:
// haveLeft ? numLeftItems - 1
// : numLeftItems
// but that's always equal to subNodeIndex.
System.arraycopy(nodes, 0, leftNodes, 0, subNodeIndex);
if (haveLeft) {
leftNodes[numLeftItems - 1] = splitLeft;
}
int newSize = 0;
for (Node n : leftNodes) {
newSize += n.size();
}
left = new Strict<>(shift, newSize, leftNodes);
}
// left.debugValidate();
// split.right().debugValidate();
// if ( (split.right().size() > 0) &&
// (nodes[0].height() != split.right().height()) ) {
// throw new IllegalStateException(
// "Have a right node of a different height!" +
// "nodes:" + showSubNodes(new StringBuilder(), nodes, 6) +
// "\nright:" + split.right().indentedStr(6));
// }
final Node right = Relaxed.fixRight(nodes, split.right(), subNodeIndex);
// right.debugValidate();
// if ( (left.size() > 0) &&
// (right.size() > 0) &&
// (left.height() != right.height()) ) {
// throw new IllegalStateException("Unequal heights of split!\n" +
// "left: " + left.indentedStr(6) +
// "\nright:" + right.indentedStr(6));
// }
// if (this.size() != ret.size()) {
// throw new IllegalStateException(
// "Split on " + this.size() + " items returned " + ret.size() + " items");
// }
return new SplitNode<>(left, split.leftFocus(),
right, split.rightFocus());
}
Relaxed relax() {
int[] newCumSizes = new int[nodes.length];
int cumulativeSize = 0;
// We know all sub-nodes (except the last) have the same size because they are
// packed-left.
int subNodeSize = nodes[0].size();
for (int i = 0; i < nodes.length - 1; i++) {
cumulativeSize += subNodeSize;
newCumSizes[i] = cumulativeSize;
}
// Final node may not be packed, so it could have a different size
cumulativeSize += nodes[nodes.length - 1].size();
newCumSizes[newCumSizes.length - 1] = cumulativeSize;
return new Relaxed<>(newCumSizes, nodes);
}
@Override public int numChildren() { return nodes.length; }
@SuppressWarnings("unchecked")
@Override
public Node pushFocus(int index, T[] oldFocus) {
// If the proper sub-node can take the additional array, let it!
int subNodeIndex = highBits(index);
// It's a strict-compatible addition if the focus being pushed is of
// STRICT_NODE_LENGTH and the index it's pushed to falls on the final leaf-node boundary
// and the children of this node are leaves and this node is not full.
if (oldFocus.length == STRICT_NODE_LENGTH) {
if (index == size()) {
Node lastNode = nodes[nodes.length - 1];
if (lastNode.hasStrictCapacity()) {
// Pushing focus down to lower-level node with capacity.
// TODO: This line appears to be the slowest part.
// This variable is my attempt to prevent dynamic dispatch on the call
// to pushFocus()
Strict strict = (Strict) lastNode;
Node newNode = strict.pushFocus(lowBits(index), oldFocus);
Node[] newNodes = replaceInArrayAt(newNode, nodes, nodes.length - 1,
Node.class);
return new Strict<>(shift, size + oldFocus.length, newNodes);
}
// Regardless of what else happens, we're going to add a new node.
Node newNode = new Leaf<>(oldFocus);
// Make a skinny branch of a tree by walking up from the leaf node until our
// new branch is at the same level as the old one. We have to build evenly
// (like hotels in Monopoly) in order to keep the tree balanced. Even height,
// but left-packed (the lower indices must all be filled before adding new
// nodes to the right).
int newShift = NODE_LENGTH_POW_2;
// If we've got space in our array, we just have to add skinny-branch nodes up
// to the level below ours. But if we don't have space, we have to add a
// single-element strict node at the same level as ours here too.
int maxShift = (nodes.length < STRICT_NODE_LENGTH) ? shift : shift + 1;
// Make the skinny-branch of single-element strict nodes:
while (newShift < maxShift) {
// Add a skinny branch node
newNode = new Strict<>(newShift, oldFocus.length, singleElementArray(newNode, Node.class));
newShift += NODE_LENGTH_POW_2;
}
if ((nodes.length < STRICT_NODE_LENGTH)) {
// Add a node to the existing array
Node[] newNodes =
insertIntoArrayAt(newNode, nodes, subNodeIndex, Node.class);
// This could allow cheap strict inserts on any leaf-node boundary...
return new Strict<>(shift, size + oldFocus.length, newNodes);
} else {
// Add a level to the Strict tree
return new Strict(shift + NODE_LENGTH_POW_2,
size + oldFocus.length,
new Node[]{this, newNode});
}
} else if ( (shift == NODE_LENGTH_POW_2) &&
(lowBits(index) == 0) &&
(nodes.length < STRICT_NODE_LENGTH) ) {
// Here we are:
// Pushing a STRICT_NODE_LENGTH focus
// At the level above the leaf nodes
// Inserting *between* existing leaf nodes (or before or after)
// Have room for at least one more leaf child
// That makes it free and legal to insert a new STRICT_NODE_LENGTH leaf node and
// still yield a Strict (as opposed to Relaxed).
// Regardless of what else happens, we're going to add a new node.
Node newNode = new Leaf<>(oldFocus);
Node[] newNodes =
insertIntoArrayAt(newNode, nodes, subNodeIndex, Node.class);
// This allows cheap strict inserts on any leaf-node boundary...
return new Strict<>(shift, size + oldFocus.length, newNodes);
}
} // end if oldFocus.length == STRICT_NODE_LENGTH
// Here we're going to yield a Relaxed Radix node, so punt to that (slower) logic.
return relax().pushFocus(index, oldFocus);
}
@SuppressWarnings("unchecked")
@Override
public Node replace(int idx, T t) {
// Find the node indexed by the high bits (for this height).
// Send the low bits on to our sub-nodes.
int thisNodeIdx = highBits(idx);
Node newNode = nodes[thisNodeIdx].replace(lowBits(idx), t);
return new Strict<>(shift, size,
replaceInArrayAt(newNode, nodes, thisNodeIdx, Node.class));
}
@Override public boolean thisNodeHasRelaxedCapacity(int numNodes) {
return nodes.length + numNodes < MAX_NODE_LENGTH;
}
@Override public String toString() {
// return "Strict(nodes.length="+ nodes.length + ", shift=" + shift + ")";
return "Strict" + shift + arrayString(nodes);
}
@Override public String indentedStr(int indent) {
StringBuilder sB = new StringBuilder() // indentSpace(indent)
.append("Strict").append(shift).append("(");
int len = sB.length();
sB.append("size=").append(size).append("\n");
sB.append(indentSpace(len + indent));
return showSubNodes(sB, nodes, indent + len)
.append(")")
.toString();
}
}
// Contains a relaxed tree of nodes that average around 32 items each.
private static class Relaxed implements Node {
// Holds the size of each sub-node and plus all nodes to its left. You could think of this
// as maxIndex + 1. This is a separate array so it can be retrieved in a single memory
// fetch. Note that this is a 1-based count, not a zero-based index.
final int[] cumulativeSizes;
// The sub nodes
final Node[] nodes;
// Constructor
Relaxed(int[] szs, Node[] ns) {
cumulativeSizes = szs;
nodes = ns;
// Consider removing constraint validations before shipping for performance
// if (cumulativeSizes.length < 1) {
// throw new IllegalArgumentException("cumulativeSizes.length < 1");
// }
// if (nodes.length < 1) {
// throw new IllegalArgumentException("nodes.length < 1");
// }
// if (cumulativeSizes.length != nodes.length) {
// throw new IllegalArgumentException(
// "cumulativeSizes.length:" + cumulativeSizes.length +
// " != nodes.length:" + nodes.length);
// }
// int cumulativeSize = 0;
// for (int i = 0; i < nodes.length; i++) {
// cumulativeSize += nodes[i].size();
// if (cumulativeSize != cumulativeSizes[i]) {
// throw new IllegalArgumentException(
// "nodes[" + i + "].size() was " +
// nodes[i].size() +
// " which is not compatable with cumulativeSizes[" +
// i + "] which was " + cumulativeSizes[i] +
// "\n\tcumulativeSizes=" + arrayString(cumulativeSizes) +
// "\n\tnodes=" + arrayString(nodes));
// }
// }
}
@Override public Node child(int childIdx) { return nodes[childIdx]; }
@Override public int debugValidate() {
int sz = 0;
int height = height() - 1;
if (nodes.length != cumulativeSizes.length) {
throw new IllegalStateException("Unequal size of nodes and sizes!\n" +
this.indentedStr(0));
}
for (int i = 0; i < nodes.length; i++) {
Node n = nodes[i];
if (n.height() != height) {
throw new IllegalStateException("Unequal height!\n" +
this.indentedStr(0));
}
sz += n.size();
if (sz != cumulativeSizes[i]) {
throw new IllegalStateException("Cumulative Sizes are wrong!\n" +
this.indentedStr(0));
}
}
return sz;
}
/** Returns the leftMost (first) or right-most (last) child */
@Override public Node endChild(boolean leftMost) {
return nodes[leftMost ? 0 : nodes.length - 1];
}
/** Adds a node as the first/leftmost or last/rightmost child */
@Override public Node addEndChild(boolean leftMost, Node shorter) {
return insertInRelaxedAt(cumulativeSizes, nodes, shorter,
leftMost ? 0 : nodes.length);
}
/** Adds kids as leftmost or rightmost of current children */
@Override public Node addEndChildren(boolean leftMost, Node[] newKids) {
// if (!thisNodeHasRelaxedCapacity(newKids.length)) {
// throw new IllegalStateException("Can't add enough kids");
// }
// if (nodes[0].height() != newKids[0].height()) {
// throw new IllegalStateException("Kids not same height");
// }
@SuppressWarnings("unchecked")
Node[] res = spliceIntoArrayAt(newKids, nodes,
leftMost ? 0
: nodes.length, Node.class);
// TODO: Figure out which side we inserted on and do the math to adjust counts instead
// of looking them up.
return new Relaxed<>(makeSizeArray(res), res);
}
@Override public int height() { return nodes[0].height() + 1; }
@Override public int size() {
return cumulativeSizes[cumulativeSizes.length - 1];
}
/**
Converts the index of an item into the index of the sub-node containing that item.
@param treeIndex The index of the item in the tree
@return The index of the immediate child of this node that the desired node resides in.
*/
private int subNodeIndex(int treeIndex) {
// For radix=4 this is actually faster, or at least as fast...
// for (int i = 0; i < cumulativeSizes.length; i++) {
// if (treeIndex < cumulativeSizes[i]) {
// return i;
// }
// }
// if (treeIndex == size()) {
// return cumulativeSizes.length - 1;
// }
// treeSize = cumulativeSizes[cumulativeSizes.length - 1]
// range of sub-node indices: 0 to cumulativeSizes.length - 1
// range of tree indices: 0 to treeSize
// liner interpolation:
// treeIndex / treeSize ~= subNodeIndex / cumulativeSizes.length
// solve for endIdxSlot
// cumulativeSizes.length * (treeIndex / treeSize) = subNodeIndex
// Put division last
// subNodeIndex = cumulativeSizes.length * treeIndex / treeSize
//
// Now guess the sub-node index (quickly).
int guess = (cumulativeSizes.length * treeIndex) / size();
if (guess >= cumulativeSizes.length) {
// Guessed beyond end length - returning last item.
return cumulativeSizes.length - 1;
}
int guessedCumSize = cumulativeSizes[guess];
// Now we must check the guess. The cumulativeSizes we store are slightly misnamed because
// the max valid treeIndex for a node is its size - 1. If our guessedCumSize is
// - less than the treeIndex
// Increment guess and check result again until greater, then return
// that guess
// - greater than (treeIndex + MIN_NODE_SIZE)
// Decrement guess and check result again until less, then return PREVIOUS guess
// - equal to the treeIndex (see note about size)
// If treeIndex == size Return guess
// Else return guess + 1
// guessedCumSize less than the treeIndex
// Increment guess and check result again until greater, then return
// that guess
if (guessedCumSize < treeIndex) {
while (guess < (cumulativeSizes.length - 1)) {
// System.out.println(" Too low. Check higher...");
guessedCumSize = cumulativeSizes[++guess];
if (guessedCumSize >= treeIndex) {
// System.out.println(" RIGHT!");
// See note in equal case below...
return (guessedCumSize == treeIndex) ? guess + 1
: guess;
}
// System.out.println(" ==== Guess higher...");
}
throw new IllegalStateException("Can we get here? If so, how?");
} else if (guessedCumSize > (treeIndex + MIN_NODE_LENGTH)) {
// guessedCumSize greater than (treeIndex + MIN_NODE_LENGTH)
// Decrement guess and check result again until less, then return PREVIOUS guess
while (guess > 0) {
// System.out.println(" Maybe too high. Check lower...");
int nextGuess = guess - 1;
guessedCumSize = cumulativeSizes[nextGuess];
if (guessedCumSize <= treeIndex) {
// System.out.println(" RIGHT!");
return guess;
}
// System.out.println(" ======= Guess lower...");
guess = nextGuess;
}
// System.out.println(" Returning lower: " + guess);
return guess;
} else if (guessedCumSize == treeIndex) {
// guessedCumSize equal to the treeIndex (see note about size)
// If treeIndex == size Return guess
// Else return guess + 1
// System.out.println(" Equal, so should be simple...");
// For an append just one element beyond the end of the existing data structure,
// just try to add it to the last node. This might seem overly permissive to accept
// these as inserts or appends without differentiating between the two, but it flows
// naturally with this data structure and I think makes it easier to use without
// encouraging user programming errors.
// Hopefully this still leads to a relatively balanced tree...
return (treeIndex == size()) ? guess : guess + 1;
} else {
// System.out.println(" First guess: " + guess);
return guess;
}
}
/**
Converts the index of an item into the index to pass to the sub-node containing that item.
@param index The index of the item in the entire tree
@param subNodeIndex the index into this node's array of sub-nodes.
@return The index to pass to the sub-branch the item resides in
*/
// Better name might be: nextLevelIndex? subNodeSubIndex?
private int subNodeAdjustedIndex(int index, int subNodeIndex) {
return (subNodeIndex == 0) ? index
: index - cumulativeSizes[subNodeIndex - 1];
}
@Override public T get(int index) {
int subNodeIndex = subNodeIndex(index);
return nodes[subNodeIndex].get(subNodeAdjustedIndex(index, subNodeIndex));
}
@Override public boolean thisNodeHasRelaxedCapacity(int numNodes) {
return nodes.length + numNodes < MAX_NODE_LENGTH;
}
// I don't think this should ever be called. Should this throw an exception instead?
@Override public boolean hasStrictCapacity() {
throw new UnsupportedOperationException("I don't think this should ever be called.");
// return false;
}
@Override public boolean hasRelaxedCapacity(int index, int size) {
// I think we can add any number of items (less than MAX_NODE_LENGTH)
// if ( (size < MIN_NODE_LENGTH) || (size > MAX_NODE_LENGTH) ) {
// if ( (size < 1) || (size > MAX_NODE_LENGTH) ) {
// throw new IllegalArgumentException("Bad size: " + size);
// }
if (thisNodeHasRelaxedCapacity(1)) { return true; }
int subNodeIndex = subNodeIndex(index);
return nodes[subNodeIndex].hasRelaxedCapacity(subNodeAdjustedIndex(index, subNodeIndex),
size);
}
@SuppressWarnings("unchecked")
Relaxed[] split() {
int midpoint = nodes.length >> 1; // Shift-right one is the same as dividing by 2.
Relaxed left = new Relaxed<>(Arrays.copyOf(cumulativeSizes, midpoint),
Arrays.copyOf(nodes, midpoint));
int[] rightCumSizes = new int[nodes.length - midpoint];
int leftCumSizes = cumulativeSizes[midpoint - 1];
for (int j = 0; j < rightCumSizes.length; j++) {
rightCumSizes[j] = cumulativeSizes[midpoint + j] - leftCumSizes;
}
// I checked this at javaRepl and indeed this starts from the correct item.
Relaxed right = new Relaxed<>(rightCumSizes,
Arrays.copyOfRange(nodes, midpoint, nodes.length));
return new Relaxed[] {left, right};
}
@Override
public SplitNode splitAt(int splitIndex) {
int size = size();
// if ( (splitIndex < 0) || (splitIndex > size) ) {
// throw new IllegalArgumentException("Bad splitIndex: " + splitIndex);
// }
if (splitIndex == 0) {
return new SplitNode<>(emptyLeaf(), emptyArray(), emptyLeaf(), emptyArray());
}
if (splitIndex == size) {
return new SplitNode<>(this, emptyArray(), emptyLeaf(), emptyArray());
}
int subNodeIndex = subNodeIndex(splitIndex);
Node subNode = nodes[subNodeIndex];
if ( (subNodeIndex > 0) && (splitIndex == cumulativeSizes[subNodeIndex - 1]) ) {
// Falls on an existing node boundary
Tuple2[],Node[]> splitNodes = splitArray(nodes, subNodeIndex);
int[][] splitCumSizes = splitArray(cumulativeSizes, subNodeIndex);
int[] leftCumSizes = splitCumSizes[0];
int[] rightCumSizes = splitCumSizes[1];
int bias = leftCumSizes[leftCumSizes.length - 1];
for (int i = 0; i < rightCumSizes.length; i++) {
rightCumSizes[i] = rightCumSizes[i] - bias;
}
Node left = new Relaxed<>(leftCumSizes, splitNodes._1());
// left.debugValidate();
Node right = new Relaxed<>(rightCumSizes, splitNodes._2());
// right.debugValidate();
return new SplitNode<>(left, emptyArray(),
right, emptyArray());
}
int subNodeAdjustedIndex = subNodeAdjustedIndex(splitIndex, subNodeIndex);
SplitNode split = subNode.splitAt(subNodeAdjustedIndex);
final Node left;
Node splitLeft = split.left();
// splitLeft.debugValidate();
if (subNodeIndex == 0) {
// debug("If we have a single left node, it doesn't need a parent.");
left = splitLeft;
} else {
boolean haveLeft = (splitLeft.size() > 0);
int numLeftItems = subNodeIndex + (haveLeft ? 1 : 0);
int[] leftCumSizes = new int[numLeftItems];
Node[] leftNodes = genericNodeArray(numLeftItems);
// src, srcPos, dest,destPos, length
System.arraycopy(cumulativeSizes, 0, leftCumSizes, 0, numLeftItems);
if (haveLeft) {
int cumulativeSize = (numLeftItems > 1) ? leftCumSizes[numLeftItems - 2] : 0;
leftCumSizes[numLeftItems - 1] = cumulativeSize + splitLeft.size();
}
// debug("leftCumSizes=" + arrayString(leftCumSizes));
// Copy one less item if we are going to add the split one in a moment.
// I could have written:
// haveLeft ? numLeftItems - 1
// : numLeftItems
// but that's always equal to subNodeIndex.
System.arraycopy(nodes, 0, leftNodes, 0, subNodeIndex);
if (haveLeft) {
// I don't know why I have to fix this here.
while (splitLeft.height() < this.height() - 1) {
splitLeft = addAncestor(splitLeft);
}
// if ( (leftNodes.length > 0) &&
// (leftNodes[0].height() != splitLeft.height()) ) {
// throw new IllegalStateException("nodesHeight: " + leftNodes[0].height() +
// " splitLeftHeight: " + splitLeft.height());
// }
leftNodes[numLeftItems - 1] = splitLeft;
}
left = new Relaxed<>(leftCumSizes, leftNodes);
// left.debugValidate();
}
final Node right = fixRight(nodes, split.right(), subNodeIndex);
// right.debugValidate();
// debug("RETURNING=", ret);
// if (this.size() != ret.size()) {
// throw new IllegalStateException(
// "Split on " + this.size() + " items returned " +
// ret.size() + " items\n" +
// "original=" + this.indentedStr(9) + "\n" +
// "splitIndex=" + splitIndex + "\n" +
// "leftFocus=" + arrayString(split.leftFocus()) + "\n" +
// "left=" + left.indentedStr(5) + "\n" +
// "rightFocus=" + arrayString(split.rightFocus()) + "\n" +
// "right=" + right.indentedStr(6));
// }
return new SplitNode<>(left, split.leftFocus(),
right, split.rightFocus());
}
@Override public int numChildren() { return nodes.length; }
@SuppressWarnings("unchecked")
@Override public Node pushFocus(int index, T[] oldFocus) {
// TODO: Review this entire method.
int subNodeIndex = subNodeIndex(index);
Node subNode = nodes[subNodeIndex];
int subNodeAdjustedIndex = subNodeAdjustedIndex(index, subNodeIndex);
// 1st choice: insert into the subNode if it has enough space enough to handle it
if (subNode.hasRelaxedCapacity(subNodeAdjustedIndex, oldFocus.length)) {
// Push the focus down to a lower-level node w. capacity.
Node newNode = subNode.pushFocus(subNodeAdjustedIndex, oldFocus);
// Make a copy of our nodesArray, replacing the old node at subNodeIndex with the
// new node
return replaceInRelaxedAt(cumulativeSizes, nodes, newNode, subNodeIndex,
oldFocus.length);
}
// I think this is a root node thing.
if (!thisNodeHasRelaxedCapacity(1)) {
// For now, split at half of size.
Relaxed[] split = split();
int max1 = split[0].size();
Relaxed newRelaxed =
new Relaxed<>(new int[] {max1,
max1 + split[1].size()},
split);
return newRelaxed.pushFocus(index, oldFocus);
}
if (subNode instanceof Leaf) {
// Here we already know:
// - the leaf doesn't have capacity
// - We don't need to split ourselves
// Therefore:
// - If the focus is big enough to be its own leaf and the index is on a leaf
// boundary and , make it one.
// - Else, insert into the array and replace one leaf with two.
final Node[] newNodes;
final int[] newCumSizes;
final int numToSkip;
// If the focus is big enough to be its own leaf and the index is on a leaf
// boundary, make it one.
if ( (oldFocus.length >= MIN_NODE_LENGTH) &&
(subNodeAdjustedIndex == 0 || subNodeAdjustedIndex == subNode.size()) ) {
// Insert-between
// Just add a new leaf
Leaf newNode = new Leaf<>(oldFocus);
// If we aren't inserting before the existing leaf node, we're inserting after.
if (subNodeAdjustedIndex != 0) {
subNodeIndex++;
}
newNodes = insertIntoArrayAt(newNode, nodes, subNodeIndex, Node.class);
// Increment newCumSizes for the changed item and all items to the right.
newCumSizes = new int[cumulativeSizes.length + 1];
int cumulativeSize = 0;
if (subNodeIndex > 0) {
System.arraycopy(cumulativeSizes, 0, newCumSizes, 0, subNodeIndex);
cumulativeSize = newCumSizes[subNodeIndex - 1];
}
newCumSizes[subNodeIndex] = cumulativeSize + oldFocus.length;
numToSkip = 1;
// for (int i = subNodeIndex + 1; i < newCumSizes.length; i++) {
// newCumSizes[i] = cumulativeSizes[i - 1] + oldFocus.length;
// }
} else {
// Grab the array from the existing leaf node, make the insert, and yield two
// new leaf nodes.
// Split-to-insert
Leaf[] res =
((Leaf) subNode).spliceAndSplit(oldFocus, subNodeAdjustedIndex);
Leaf leftLeaf = res[0];
Leaf rightLeaf = res[1];
newNodes = new Node[nodes.length + 1];
// Increment newCumSizes for the changed item and all items to the right.
newCumSizes = new int[cumulativeSizes.length + 1];
int leftSize = 0;
// Copy nodes and cumulativeSizes before split
if (subNodeIndex > 0) {
// src,srcPos,dest,destPos,length
System.arraycopy(nodes, 0, newNodes, 0, subNodeIndex);
// src, srcPos, dest, destPos, length
System.arraycopy(cumulativeSizes, 0, newCumSizes, 0, subNodeIndex);
leftSize = cumulativeSizes[subNodeIndex - 1];
}
// Copy split nodes and cumulativeSizes
newNodes[subNodeIndex] = leftLeaf;
newNodes[subNodeIndex + 1] = rightLeaf;
leftSize += leftLeaf.size();
newCumSizes[subNodeIndex] = leftSize;
newCumSizes[subNodeIndex + 1] = leftSize + rightLeaf.size();
if (subNodeIndex < (nodes.length - 1)) {
// src,srcPos,dest,destPos,length
System.arraycopy(nodes, subNodeIndex + 1, newNodes, subNodeIndex + 2,
nodes.length - subNodeIndex - 1);
}
numToSkip = 2;
}
for (int i = subNodeIndex + numToSkip; i < newCumSizes.length; i++) {
newCumSizes[i] = cumulativeSizes[i - 1] + oldFocus.length;
}
return new Relaxed<>(newCumSizes, newNodes);
// end if subNode instanceof Leaf
} else if (subNode instanceof Strict) {
// Convert Strict to Relaxed
Relaxed relaxed = ((Strict) subNode).relax();
Node newNode = relaxed.pushFocus(subNodeAdjustedIndex, oldFocus);
return replaceInRelaxedAt(cumulativeSizes, nodes, newNode, subNodeIndex,
oldFocus.length);
}
// Here we have capacity and the full sub-node is not a leaf or strict, so we have to
// split the appropriate sub-node.
// For now, split at half of size.
Relaxed[] newSubNode = ((Relaxed) subNode).split();
Relaxed node1 = newSubNode[0];
Relaxed node2 = newSubNode[1];
Node[] newNodes = genericNodeArray(nodes.length + 1);
// If we aren't inserting at the first item, array-copy the nodes before the insert
// point.
if (subNodeIndex > 0) {
System.arraycopy(nodes, 0, newNodes, 0, subNodeIndex);
}
// Insert the new item.
newNodes[subNodeIndex] = node1;
newNodes[subNodeIndex + 1] = node2;
// If we aren't inserting at the last item, array-copy the nodes after the insert
// point.
if (subNodeIndex < nodes.length) {
System.arraycopy(nodes, subNodeIndex + 1, newNodes, subNodeIndex + 2,
nodes.length - subNodeIndex - 1);
}
int[] newCumSizes = new int[cumulativeSizes.length + 1];
int cumulativeSize = 0;
if (subNodeIndex > 0) {
System.arraycopy(cumulativeSizes, 0, newCumSizes, 0, subNodeIndex);
cumulativeSize = cumulativeSizes[subNodeIndex - 1];
}
for (int i = subNodeIndex; i < newCumSizes.length; i++) {
// TODO: Calculate instead of loading into memory. See splitAt calculation above.
cumulativeSize += newNodes[i].size();
newCumSizes[i] = cumulativeSize;
}
Relaxed newRelaxed = new Relaxed<>(newCumSizes, newNodes);
// debug("newRelaxed2:\n" + newRelaxed.indentedStr(0));
return newRelaxed.pushFocus(index, oldFocus);
// debug("Parent after:" + after.indentedStr(0));
}
@SuppressWarnings("unchecked")
@Override public Node replace(int index, T t) {
int subNodeIndex = subNodeIndex(index);
Node alteredNode =
nodes[subNodeIndex].replace(subNodeAdjustedIndex(index, subNodeIndex), t);
Node[] newNodes = replaceInArrayAt(alteredNode, nodes, subNodeIndex, Node.class);
return new Relaxed<>(cumulativeSizes, newNodes);
}
@Override public String indentedStr(int indent) {
StringBuilder sB = new StringBuilder() // indentSpace(indent)
.append("Relaxed(");
int nextIndent = indent + sB.length();
sB.append("cumulativeSizes=").append(arrayString(cumulativeSizes)).append("\n")
.append(indentSpace(nextIndent)).append("nodes=[");
// + 6 for "nodes="
return showSubNodes(sB, nodes, nextIndent + 7)
.append("])")
.toString();
}
@Override public String toString() { return indentedStr(0); }
/**
This asks each node what size it is, then puts the cumulative sizes into a new array.
In theory, it might be faster to figure out what side we added/removed nodes and do
some addition/subtraction (in the amount of the added/removed nodes). But I want to
optimize other things first and be sure everything is correct before experimenting with
that. After all, it might not even be faster!
@param newNodes the nodes to take sizes from.
@return An array of cumulative sizes of each node in the passed array.
*/
private static int[] makeSizeArray(Node[] newNodes) {
int[] newCumSizes = new int[newNodes.length];
int cumulativeSize = 0;
for (int i = 0; i < newCumSizes.length; i++) {
cumulativeSize += newNodes[i].size();
newCumSizes[i] = cumulativeSize;
}
return newCumSizes;
}
// TODO: Search for more opportunities to use this
/**
Replace a node in a relaxed node by recalculating the cumulative sizes and copying
all sub nodes.
@param is original cumulative sizes
@param ns original nodes
@param newNode replacement node
@param subNodeIndex index to replace in this node's immediate children
@param insertSize the difference in size between the original node and the new node.
@return a new immutable Relaxed node with the immediate child node replaced.
*/
static Relaxed replaceInRelaxedAt(int[] is, Node[] ns, Node newNode,
int subNodeIndex, int insertSize) {
@SuppressWarnings("unchecked")
Node[] newNodes = replaceInArrayAt(newNode, ns, subNodeIndex, Node.class);
// Increment newCumSizes for the changed item and all items to the right.
int[] newCumSizes = new int[is.length];
if (subNodeIndex > 0) {
System.arraycopy(is, 0, newCumSizes, 0, subNodeIndex);
}
for (int i = subNodeIndex; i < is.length; i++) {
newCumSizes[i] = is[i] + insertSize;
}
return new Relaxed<>(newCumSizes, newNodes);
}
/**
Insert a node in a relaxed node by recalculating the cumulative sizes and copying
all sub nodes.
@param oldCumSizes original cumulative sizes
@param ns original nodes
@param newNode replacement node
@param subNodeIndex index to insert in this node's immediate children
@return a new immutable Relaxed node with the immediate child node inserted.
*/
static Relaxed insertInRelaxedAt(int[] oldCumSizes, Node[] ns, Node newNode,
int subNodeIndex) {
@SuppressWarnings("unchecked")
Node[] newNodes = insertIntoArrayAt(newNode, ns, subNodeIndex, Node.class);
int oldLen = oldCumSizes.length;
// if (subNodeIndex > oldLen) {
// throw new IllegalStateException("subNodeIndex > oldCumSizes.length");
// }
int[] newCumSizes = new int[oldLen + 1];
// Copy unchanged cumulative sizes
if (subNodeIndex > 0) {
System.arraycopy(oldCumSizes, 0, newCumSizes, 0, subNodeIndex);
}
// insert the cumulative size of the new node
int newNodeSize = newNode.size();
// Find cumulative size of previous node
int prevNodeTotal =
(subNodeIndex == 0) ? 0
: oldCumSizes[subNodeIndex - 1];
newCumSizes[subNodeIndex] = newNodeSize + prevNodeTotal;
for (int i = subNodeIndex; i < oldCumSizes.length; i++) {
newCumSizes[i + 1] = oldCumSizes[i] + newNodeSize;
}
return new Relaxed<>(newCumSizes, newNodes);
}
/**
Fixes up nodes on the right-hand side of the split. Might want to explain this better...
@param origNodes the immediate children of the node we're splitting
@param splitRight the pre-split right node
@param subNodeIndex the index to split children at?
@return a copy of this node with only the right-hand side of the split.
*/
@SuppressWarnings("unchecked")
public static Node fixRight(Node[] origNodes, Node splitRight,
int subNodeIndex) {
// if ( (splitRight.size() > 0) &&
// (origNodes[0].height() != splitRight.height()) ) {
// throw new IllegalStateException("Passed a splitRight node of a different height" +
// " than the origNodes!");
// }
Node right;
if (subNodeIndex == (origNodes.length - 1)) {
// right = splitRight;
right = new Relaxed<>(new int[] { splitRight.size() }, new Node[]{ splitRight });
} else {
boolean haveRightSubNode = splitRight.size() > 0;
// If we have a rightSubNode, it's going to need a space in our new node array.
int numRightNodes = (origNodes.length - subNodeIndex) - (haveRightSubNode ? 0 : 1);
// Here the first (leftmost) node of the right-hand side was turned into the focus
// and we have additional right-hand origNodes to adjust the parent for.
int[] rightCumSizes = new int[numRightNodes];
Node[] rightNodes = genericNodeArray(numRightNodes);
int cumulativeSize = 0;
int destCopyStartIdx = 0;
if (haveRightSubNode) {
// src, srcPos, dest, destPos, length
System.arraycopy(origNodes, subNodeIndex + 1, rightNodes, 1, numRightNodes - 1);
rightNodes[0] = splitRight;
cumulativeSize = splitRight.size();
rightCumSizes[0] = cumulativeSize;
destCopyStartIdx = 1;
} else {
// src, srcPos, dest, destPos, length
System.arraycopy(origNodes, subNodeIndex + 1, rightNodes, 0, numRightNodes);
}
// For relaxed nodes, we could calculate from previous cumulativeSizes instead of
// calling .size() on each one. For strict, we could just add a strict amount.
// For now, this works.
for (int i = destCopyStartIdx; i < numRightNodes; i++) {
cumulativeSize += rightNodes[i].size();
rightCumSizes[i] = cumulativeSize;
}
right = new Relaxed<>(rightCumSizes, rightNodes);
// right.debugValidate();
}
return right;
} // end fixRight()
} // end class Relaxed
// =================================== Tree-walking Iterator ==================================
/** Holds a node and the index of the child node we are currently iterating in. */
private static final class IdxNode {
int idx = 0;
final Node node;
IdxNode(Node n) { node = n; }
public boolean hasNext() { return idx < node.numChildren(); }
public Node