1 /* Copyright (C) 1995-2015 Free Software Foundation, Inc.
2 This file is part of the GNU C Library.
3 Contributed by Bernd Schmidt <crux@Pool.Informatik.RWTH-Aachen.DE>, 1997.
5 The GNU C Library is free software; you can redistribute it and/or
6 modify it under the terms of the GNU Lesser General Public
7 License as published by the Free Software Foundation; either
8 version 2.1 of the License, or (at your option) any later version.
10 The GNU C Library is distributed in the hope that it will be useful,
11 but WITHOUT ANY WARRANTY; without even the implied warranty of
12 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 Lesser General Public License for more details.
15 You should have received a copy of the GNU Lesser General Public
16 License along with the GNU C Library; if not, see
17 <http://www.gnu.org/licenses/>. */
19 /* Tree search for red/black trees.
20 The algorithm for adding nodes is taken from one of the many "Algorithms"
21 books by Robert Sedgewick, although the implementation differs.
22 The algorithm for deleting nodes can probably be found in a book named
23 "Introduction to Algorithms" by Cormen/Leiserson/Rivest. At least that's
24 the book that my professor took most algorithms from during the "Data
27 Totally public domain. */
29 /* Red/black trees are binary trees in which the edges are colored either red
30 or black. They have the following properties:
31 1. The number of black edges on every path from the root to a leaf is
33 2. No two red edges are adjacent.
34 Therefore there is an upper bound on the length of every path, it's
35 O(log n) where n is the number of nodes in the tree. No path can be longer
36 than 1+2*P where P is the length of the shortest path in the tree.
37 Useful for the implementation:
38 3. If one of the children of a node is NULL, then the other one is red
41 In the implementation, not the edges are colored, but the nodes. The color
42 interpreted as the color of the edge leading to this node. The color is
43 meaningless for the root node, but we color the root node black for
44 convenience. All added nodes are red initially.
46 Adding to a red/black tree is rather easy. The right place is searched
47 with a usual binary tree search. Additionally, whenever a node N is
48 reached that has two red successors, the successors are colored black and
49 the node itself colored red. This moves red edges up the tree where they
50 pose less of a problem once we get to really insert the new node. Changing
51 N's color to red may violate rule 2, however, so rotations may become
52 necessary to restore the invariants. Adding a new red leaf may violate
53 the same rule, so afterwards an additional check is run and the tree
56 Deleting is hairy. There are mainly two nodes involved: the node to be
57 deleted (n1), and another node that is to be unchained from the tree (n2).
58 If n1 has a successor (the node with a smallest key that is larger than
59 n1), then the successor becomes n2 and its contents are copied into n1,
60 otherwise n1 becomes n2.
61 Unchaining a node may violate rule 1: if n2 is black, one subtree is
62 missing one black edge afterwards. The algorithm must try to move this
63 error upwards towards the root, so that the subtree that does not have
64 enough black edges becomes the whole tree. Once that happens, the error
65 has disappeared. It may not be necessary to go all the way up, since it
66 is possible that rotations and recoloring can fix the error before that.
68 Although the deletion algorithm must walk upwards through the tree, we
69 do not store parent pointers in the nodes. Instead, delete allocates a
70 small array of parent pointers and fills it while descending the tree.
71 Since we know that the length of a path is O(log n), where n is the number
72 of nodes, this is likely to use less memory. */
74 /* Tree rotations look like this:
83 In this case, A has been rotated left. This preserves the ordering of the
92 /* Callers expect this to be the first element in the structure - do not
99 typedef const struct node_t
*const_node
;
105 /* Routines to check tree invariants. */
109 #define CHECK_TREE(a) check_tree(a)
112 check_tree_recurse (node p
, int d_sofar
, int d_total
)
116 assert (d_sofar
== d_total
);
120 check_tree_recurse (p
->left
, d_sofar
+ (p
->left
&& !p
->left
->red
), d_total
);
121 check_tree_recurse (p
->right
, d_sofar
+ (p
->right
&& !p
->right
->red
), d_total
);
123 assert (!(p
->left
->red
&& p
->red
));
125 assert (!(p
->right
->red
&& p
->red
));
129 check_tree (node root
)
136 for(p
= root
->left
; p
; p
= p
->left
)
138 check_tree_recurse (root
, 0, cnt
);
144 #define CHECK_TREE(a)
148 /* Possibly "split" a node with two red successors, and/or fix up two red
149 edges in a row. ROOTP is a pointer to the lowest node we visited, PARENTP
150 and GPARENTP pointers to its parent/grandparent. P_R and GP_R contain the
151 comparison values that determined which way was taken in the tree to reach
152 ROOTP. MODE is 1 if we need not do the split, but must check for two red
153 edges between GPARENTP and ROOTP. */
155 maybe_split_for_insert (node
*rootp
, node
*parentp
, node
*gparentp
,
156 int p_r
, int gp_r
, int mode
)
160 rp
= &(*rootp
)->right
;
161 lp
= &(*rootp
)->left
;
163 /* See if we have to split this node (both successors red). */
165 || ((*rp
) != NULL
&& (*lp
) != NULL
&& (*rp
)->red
&& (*lp
)->red
))
167 /* This node becomes red, its successors black. */
174 /* If the parent of this node is also red, we have to do
176 if (parentp
!= NULL
&& (*parentp
)->red
)
180 /* There are two main cases:
181 1. The edge types (left or right) of the two red edges differ.
182 2. Both red edges are of the same type.
183 There exist two symmetries of each case, so there is a total of
185 if ((p_r
> 0) != (gp_r
> 0))
187 /* Put the child at the top of the tree, with its parent
188 and grandparent as successors. */
194 /* Child is left of parent. */
202 /* Child is right of parent. */
212 *gparentp
= *parentp
;
213 /* Parent becomes the top of the tree, grandparent and
214 child are its successors. */
234 /* Find or insert datum into search tree.
235 KEY is the key to be located, ROOTP is the address of tree root,
236 COMPAR the ordering function. */
238 __tsearch (const void *key
, void **vrootp
, __compar_fn_t compar
)
241 node
*parentp
= NULL
, *gparentp
= NULL
;
242 node
*rootp
= (node
*) vrootp
;
244 int r
= 0, p_r
= 0, gp_r
= 0; /* No they might not, Mr Compiler. */
249 /* This saves some additional tests below. */
256 while (*nextp
!= NULL
)
259 r
= (*compar
) (key
, root
->key
);
263 maybe_split_for_insert (rootp
, parentp
, gparentp
, p_r
, gp_r
, 0);
264 /* If that did any rotations, parentp and gparentp are now garbage.
265 That doesn't matter, because the values they contain are never
266 used again in that case. */
268 nextp
= r
< 0 ? &root
->left
: &root
->right
;
280 q
= (struct node_t
*) malloc (sizeof (struct node_t
));
283 *nextp
= q
; /* link new node to old */
284 q
->key
= key
; /* initialize new node */
286 q
->left
= q
->right
= NULL
;
289 /* There may be two red edges in a row now, which we must avoid by
290 rotating the tree. */
291 maybe_split_for_insert (nextp
, rootp
, parentp
, r
, p_r
, 1);
296 weak_alias (__tsearch
, tsearch
)
299 /* Find datum in search tree.
300 KEY is the key to be located, ROOTP is the address of tree root,
301 COMPAR the ordering function. */
303 __tfind (key
, vrootp
, compar
)
306 __compar_fn_t compar
;
308 node
*rootp
= (node
*) vrootp
;
315 while (*rootp
!= NULL
)
320 r
= (*compar
) (key
, root
->key
);
324 rootp
= r
< 0 ? &root
->left
: &root
->right
;
328 weak_alias (__tfind
, tfind
)
331 /* Delete node with given key.
332 KEY is the key to be deleted, ROOTP is the address of the root of tree,
333 COMPAR the comparison function. */
335 __tdelete (const void *key
, void **vrootp
, __compar_fn_t compar
)
337 node p
, q
, r
, retval
;
339 node
*rootp
= (node
*) vrootp
;
340 node root
, unchained
;
341 /* Stack of nodes so we remember the parents without recursion. It's
342 _very_ unlikely that there are paths longer than 40 nodes. The tree
343 would need to have around 250.000 nodes. */
346 node
**nodestack
= alloca (sizeof (node
*) * stacksize
);
356 while ((cmp
= (*compar
) (key
, (*rootp
)->key
)) != 0)
362 newstack
= alloca (sizeof (node
*) * stacksize
);
363 nodestack
= memcpy (newstack
, nodestack
, sp
* sizeof (node
*));
366 nodestack
[sp
++] = rootp
;
375 /* This is bogus if the node to be deleted is the root... this routine
376 really should return an integer with 0 for success, -1 for failure
377 and errno = ESRCH or something. */
380 /* We don't unchain the node we want to delete. Instead, we overwrite
381 it with its successor and unchain the successor. If there is no
382 successor, we really unchain the node to be deleted. */
389 if (q
== NULL
|| r
== NULL
)
393 node
*parent
= rootp
, *up
= &root
->right
;
400 newstack
= alloca (sizeof (node
*) * stacksize
);
401 nodestack
= memcpy (newstack
, nodestack
, sp
* sizeof (node
*));
403 nodestack
[sp
++] = parent
;
405 if ((*up
)->left
== NULL
)
412 /* We know that either the left or right successor of UNCHAINED is NULL.
413 R becomes the other one, it is chained into the parent of UNCHAINED. */
416 r
= unchained
->right
;
421 q
= *nodestack
[sp
-1];
422 if (unchained
== q
->right
)
428 if (unchained
!= root
)
429 root
->key
= unchained
->key
;
432 /* Now we lost a black edge, which means that the number of black
433 edges on every path is no longer constant. We must balance the
435 /* NODESTACK now contains all parents of R. R is likely to be NULL
436 in the first iteration. */
437 /* NULL nodes are considered black throughout - this is necessary for
439 while (sp
> 0 && (r
== NULL
|| !r
->red
))
441 node
*pp
= nodestack
[sp
- 1];
443 /* Two symmetric cases. */
446 /* Q is R's brother, P is R's parent. The subtree with root
447 R has one black edge less than the subtree with root Q. */
451 /* If Q is red, we know that P is black. We rotate P left
452 so that Q becomes the top node in the tree, with P below
453 it. P is colored red, Q is colored black.
454 This action does not change the black edge count for any
455 leaf in the tree, but we will be able to recognize one
456 of the following situations, which all require that Q
464 /* Make sure pp is right if the case below tries to use
466 nodestack
[sp
++] = pp
= &q
->left
;
469 /* We know that Q can't be NULL here. We also know that Q is
471 if ((q
->left
== NULL
|| !q
->left
->red
)
472 && (q
->right
== NULL
|| !q
->right
->red
))
474 /* Q has two black successors. We can simply color Q red.
475 The whole subtree with root P is now missing one black
476 edge. Note that this action can temporarily make the
477 tree invalid (if P is red). But we will exit the loop
478 in that case and set P black, which both makes the tree
479 valid and also makes the black edge count come out
480 right. If P is black, we are at least one step closer
481 to the root and we'll try again the next iteration. */
487 /* Q is black, one of Q's successors is red. We can
488 repair the tree with one operation and will exit the
490 if (q
->right
== NULL
|| !q
->right
->red
)
492 /* The left one is red. We perform the same action as
493 in maybe_split_for_insert where two red edges are
494 adjacent but point in different directions:
495 Q's left successor (let's call it Q2) becomes the
496 top of the subtree we are looking at, its parent (Q)
497 and grandparent (P) become its successors. The former
498 successors of Q2 are placed below P and Q.
499 P becomes black, and Q2 gets the color that P had.
500 This changes the black edge count only for node R and
513 /* It's the right one. Rotate P left. P becomes black,
514 and Q gets the color that P had. Q's right successor
515 also becomes black. This changes the black edge
516 count only for node R and its successors. */
535 /* Comments: see above. */
544 nodestack
[sp
++] = pp
= &q
->right
;
547 if ((q
->right
== NULL
|| !q
->right
->red
)
548 && (q
->left
== NULL
|| !q
->left
->red
))
555 if (q
->left
== NULL
|| !q
->left
->red
)
588 weak_alias (__tdelete
, tdelete
)
591 /* Walk the nodes of a tree.
592 ROOT is the root of the tree to be walked, ACTION the function to be
593 called at each node. LEVEL is the level of ROOT in the whole tree. */
596 trecurse (const void *vroot
, __action_fn_t action
, int level
)
598 const_node root
= (const_node
) vroot
;
600 if (root
->left
== NULL
&& root
->right
== NULL
)
601 (*action
) (root
, leaf
, level
);
604 (*action
) (root
, preorder
, level
);
605 if (root
->left
!= NULL
)
606 trecurse (root
->left
, action
, level
+ 1);
607 (*action
) (root
, postorder
, level
);
608 if (root
->right
!= NULL
)
609 trecurse (root
->right
, action
, level
+ 1);
610 (*action
) (root
, endorder
, level
);
615 /* Walk the nodes of a tree.
616 ROOT is the root of the tree to be walked, ACTION the function to be
617 called at each node. */
619 __twalk (const void *vroot
, __action_fn_t action
)
621 const_node root
= (const_node
) vroot
;
625 if (root
!= NULL
&& action
!= NULL
)
626 trecurse (root
, action
, 0);
628 weak_alias (__twalk
, twalk
)
632 /* The standardized functions miss an important functionality: the
633 tree cannot be removed easily. We provide a function to do this. */
636 tdestroy_recurse (node root
, __free_fn_t freefct
)
638 if (root
->left
!= NULL
)
639 tdestroy_recurse (root
->left
, freefct
);
640 if (root
->right
!= NULL
)
641 tdestroy_recurse (root
->right
, freefct
);
642 (*freefct
) ((void *) root
->key
);
643 /* Free the node itself. */
648 __tdestroy (void *vroot
, __free_fn_t freefct
)
650 node root
= (node
) vroot
;
655 tdestroy_recurse (root
, freefct
);
657 weak_alias (__tdestroy
, tdestroy
)