nginx データ構造 2 - 赤黒ツリーを自分で書き直す

早速ですが、今回は英語の復習を兼ねて英語で書かれたコメントを使ってコードを書き直してみます。

rbtree.h:

/*
 * Copyright (C) Bipedal Bit
 * Verson 1.0.0.1
 */

#ifndef _RBTREE_H_INCLUDED_
#define _RBTREE_H_INCLUDED_

/* the node structure of the red-black tree */
typedef struct rbtree_node_s rbtree_node_t;
/* Using type int means its range is -0x7fffffff-1~0x7fffffff. */
typedef int rbtree_key_t;
/* Abstract type is complicated to achieve with C so I use char* instead. */
typedef char* rbtree_data_t;

struct rbtree_node_s
{
	/* key of the node */
	rbtree_key_t	key;
	/* pointer of the parent of the node */
	rbtree_node_t*	parent;
	/* pointer of the left kid of the node */
	rbtree_node_t*	left;
	/* pointer of the right kid of the node */
	rbtree_node_t*	right;
	/* color of the node */
	unsigned char	color;
	/* pointer of the value of the node corresponding to the key */
	rbtree_data_t	value;
};

/* the tree object stucture of the red-black tree */
typedef struct rbtree_s rbtree_t;
/* foundational insert function pointer*/
typedef void (*rbtree_insert_p) (rbtree_t* root, rbtree_node_t* node);

struct rbtree_s
{
	/* the pointer of the root node of the tree */
	rbtree_node_t* root;
	/* black leaf nodes as sentinel */
	rbtree_node_t* sentinel;
	/* the polymorphic insert function pointer */
	rbtree_insert_p insert;
};

/* macros */
#define rbtree_init(tree, s, i)		\
rbtree_sentinel_init(s);			\
(tree)->root = s;				\
(tree)->sentinel = s;			\
(tree)->insert = i

#define rbtree_red(node)	((node)->color = 1)
#define rbtree_black(node)	((node)->color = 0)
#define rbtree_is_red(node)	((node)->color)
#define rbtree_is_black(node)	(!rbtree_is_red(node))
 /* copy n2's color to n1 */
#define rbtree_copy_color(n1, n2)	(n1->color = n2->color)
/* sentinel must be black cuz it's leaf node */
#define rbtree_sentinel_init(node)	rbtree_black(node)

/* statements of public methods */
void rbtree_insert_value(rbtree_t* tree, rbtree_node_t* node);
void rbtree_insert(rbtree_t* tree, rbtree_node_t* node);
void rbtree_delete(rbtree_t* tree, rbtree_node_t* node);
rbtree_node_t* rbtree_find(rbtree_t* tree, rbtree_key_t key);

#endif	/* _RBTREE_H_INCLUDED_ */

nginx のソース コードを読んだことがある方は、私のヘッダー ファイルが ngx_rbree.h と比べてあまり変わっておらず、非常によく似ていることがわかるでしょう。

キー rbtree.c:

/*
 * Copyright (C) Bipedal Bit
 * Verson 1.0.0.1
 */

#include <stddef.h>
#include "rbtree.h"

/* inline methods */
/* get the node with the minimum key in a subtree of the red-black tree */
static inline rbtree_node_t*
rbtree_subtree_min(rbtree_node_t* node, rbtree_node_t* sentinel)
{
    while(node->left != sentinel)
    {
        node = node->left;
    }

    return node;
}

/* replace the node "node" in the tree with node "tmp" */
static inline void rbtree_replace(rbtree_t* tree,
    rbtree_node_t* node, rbtree_node_t* tmp)
{
    /* upward: p[node] parent = node->parent;

    if (node == tree->root)
    {
        tree->root = tmp;
    }
    else if (node == node->parent->left)
    {
        /* downward: left[p[node]] parent->left = tmp;
    }
    else
    {
        /* downward: right[p[node]] parent->right = tmp;
    }

    node->parent = tmp;
}

/* change the topologic structure of the tree keeping the order of the nodes */
static inline void rbtree_left_rotate(rbtree_t* tree, rbtree_node_t* node)
{
    /* node as the var x in CLRS while tmp as the var y */
    rbtree_node_t* tmp = node->right;

    /* replace y with left[y] */
    /* downward: right[x] right = tmp->left;
    /* if left[[y] is not NIL it has a parent */
    if (tmp->left != tree->sentinel)
    {
        /* upward: p[left[y]] left->parent = node;
    }

    /* replace x with y */
    rbtree_replace(tree, node, tmp);
    tmp->left = node;
}

static inline void rbtree_right_rotate(rbtree_t* tree, rbtree_node_t* node)
{
    rbtree_node_t* tmp = node->left;

    /* replace y with right[y] */
    node->left = tmp->right;
    if (tmp->right != tree->sentinel)
    {
        tmp->right->parent = node;
    }

    /* replace x with y */
    rbtree_replace(tree, node, tmp);
    tmp->right = node;
}

/* static methods */
/* fix the red-black tree after the new node inserted */
static void rbtree_insert_fixup(rbtree_t* tree, rbtree_node_t* node)
{
    while(rbtree_is_red(node->parent))
    {
        if (node->parent == node->parent->parent->left)
        {
            /* case 1: node's uncle is red */
            if (rbtree_is_red(node->parent->parent->right))
            {
                rbtree_black(node->parent);
                rbtree_black(node->parent->parent->right);
                rbtree_red(node->parent->parent);
                node = node->parent->parent;
                /* Then we can consider the whole subtree */
                /* which is represented by the new "node" as the "node" before */
                /* and keep looping till "node" become the root. */
            }
            /* case 2: node's uncle is black */
            else
            {
                /* ensure node is the left kid of its parent */
                if (node == node->parent->right)
                {
                    node = node->parent;
                    rbtree_left_rotate(tree, node);
                }
                /* case 2 -> case 1 */
                rbtree_black(node->parent);
                rbtree_red(node->parent->parent);
                rbtree_right_rotate(tree, node->parent->parent);
            }
        }
        /* same as the "if" clause before with "left" and "right" exchanged */
        else
        {
            if (rbtree_is_red(node->parent->parent->left))
            {
                rbtree_black(node->parent);
                rbtree_black(node->parent->parent->left);
                rbtree_red(node->parent->parent);
                node = node->parent->parent;
            }
            else
            {
                if (node == node->parent->left)
                {
                    node = node->parent;
                    rbtree_right_rotate(tree, node);
                }
                rbtree_black(node->parent);
                rbtree_red(node->parent->parent);
                rbtree_left_rotate(tree, node->parent->parent);
            }
        }
    }
    /* ensure the root node being black */
    rbtree_black(tree->root);
}

static void rbtree_delete_fixup(rbtree_t* tree, rbtree_node_t* node)
{
    rbtree_node_t* brother = NULL;

    while(node != tree->root && rbtree_is_black(node))
    {
        if (node == node->parent->left)
        {
            brother = node->parent->right;
            if (rbtree_is_red(brother))
            {
                rbtree_black(brother);
                rbtree_red(node->parent);
                rbtree_left_rotate(tree, node->parent);
                /* update brother after topologic change of the tree */
                brother = node->parent->right;
            }

            if (rbtree_is_black(brother->left) && rbtree_is_black(brother->right))
            {
                rbtree_red(brother);
                /* go upward and keep on fixing color */
                node = node->parent;
            }
            else
            {
                if (rbtree_is_black(brother->right))
                {
                    rbtree_black(brother->left);
                    rbtree_red(brother);
                    rbtree_right_rotate(tree, brother);
                    /* update brother after topologic change of the tree */
                    brother = node->parent->right;
                }
                rbtree_copy_color(brother, node->parent);
                rbtree_black(node->parent);
                rbtree_black(brother->right);
                rbtree_left_rotate(tree, node->parent);
                /* end the loop and ensure root is black */
                node = tree->root;
            }
        }
        /* same as the "if" clause before with "left" and "right" exchanged */
        else
        {
            brother = node->parent->left;
            if (rbtree_is_red(brother))
            {
                rbtree_black(brother);
                rbtree_red(node->parent);
                rbtree_left_rotate(tree, node->parent);
                brother = node->parent->left;
            }

            if (rbtree_is_black(brother->left) && rbtree_is_black(brother->right))
            {
                rbtree_red(brother);
                node = node->parent;
            }
            else
            {
                if (rbtree_is_black(brother->left))
                {
                    rbtree_black(brother->right);
                    rbtree_red(brother);
                    rbtree_right_rotate(tree, brother);
                    brother = node->parent->left;
                }
                rbtree_copy_color(brother, node->parent);
                rbtree_black(node->parent);
                rbtree_black(brother->left);
                rbtree_left_rotate(tree, node->parent);
                node = tree->root;
            }
        }
    }

    rbtree_black(node);
}

/* public methods */
void rbtree_insert_value(rbtree_t* tree, rbtree_node_t* node)
{
    /* Using ** to know wether the new node will be a left kid */
    /* or a right kid of its parent node. */
    rbtree_node_t** tmp = &tree->root;
    rbtree_node_t* parent;

    while(*tmp != tree->sentinel)
    {
        parent = *tmp;
        tmp = (node->key key) ? &parent->left : &parent->right;
    }

    /* The pointer knows wether the node should be on the left side */
    /* or on the right one. */
    *tmp = node;
    node->parent = parent;
    node->left = tree->sentinel;
    node->right = tree->sentinel;
    rbtree_red(node);
}

void rbtree_insert(rbtree_t* tree, rbtree_node_t* node)
{
    rbtree_node_t* sentinel = tree->sentinel;

    /* if the tree is empty */
    if (tree->root == sentinel)
    {
        tree->root = node;
        node->parent = sentinel;
        node->left = sentinel;
        node->right = sentinel;
        rbtree_black(node);

        return;
    }

    /* generally */
    tree->insert(tree, node);
    rbtree_insert_fixup(tree, node);
}

void rbtree_delete(rbtree_t* tree, rbtree_node_t* node)
{
    rbtree_node_t* sentinel = tree->sentinel;
    /* wether "node" is on the left side or the right one */
    rbtree_node_t** ptr_to_node = NULL;
    /* "cover" is the node which is going to cover "node" */
    rbtree_node_t* cover = NULL;
    /* wether we lossing a red node on the edge of the tree */
    int loss_red = rbtree_is_red(node);
    int is_root = (node == tree->root);

    /* get "cover" & "loss_red"  */
    /* sentinel in "node"'s kids */
    if (node->left == sentinel)
    {
        cover = node->right;
    }
    else if (node->right == sentinel)
    {
        cover = node->left;
    }
    /* "node"'s kids are both non-sentinel */
    else
    {
        /* update "node" & "loss_red" & "is_root" & "cover" */
        cover = rbtree_subtree_min(node->right, sentinel);
        node->key = cover->key;
        node->value = cover->value;
        node = cover;
        loss_red = rbtree_is_red(node);
        is_root = 0;
        /* move "cover"'s kids */
        /* "cover" can only be a left kid */
        /* and can only have a right non-sentinel kid */
        /* because of function "rbtree_subtree_min" */
        cover = node->right;
    }

    if (is_root)
    {
        /* update root */
        tree->root = cover;
    }
    else
    {
        /* downward link */
        if (node == node->parent->left)
        {
            node->parent->left = cover;
        }
        else
        {
            node->parent->right = cover;
        }
    }
    /* upward link */
    cover->parent = node->parent;
    /* "cover" may be a sentinel */
    if (cover != sentinel)
    {
        /* set "cover" */
        cover->left = node->left;
        cover->right = node->right;
        rbtree_copy_color(cover, node);
    }

    /* clear "node" since it's useless */
    node->key = -1;
    node->parent = NULL;
    node->left = NULL;
    node->right = NULL;
    node->value = NULL;

    if (loss_red)
    {
        return;
    }

    /* When lossing a black node on edge */
    /* the fifth rule of red-black tree will be broke. */
    /* So the tree need to be fixed. */
    rbtree_delete_fixup(tree, cover);
}

/* find the node in the tree corresponding to the given key value */
rbtree_node_t* rbtree_find(rbtree_t* tree, rbtree_key_t key)
{
    rbtree_node_t* tmp = tree->root;
    int step_cnt = 0;

    /* search the binary tree */
    while(tmp != tree->sentinel)
    {
        /* next line is just fot test */
        // step_cnt++;
        if(key == tmp->key)
        {
            /* next line is just for test */
            // printf("step count: %d, color: %s, ", step_cnt, rbtree_is_red(tmp) ? "red" : "black");
            return tmp;
        }

        tmp = (key key) ? tmp->left : tmp->right;
    }

    return NULL;
}
 </stddef.h>

nginx ソース コード内の 100 行を超える長い関数本体も、時間とスペースのオーバーヘッドを増加させる関数呼び出しが多すぎるのを避けるための最適化であることは理解していますが、それでも分類し、すべての関数を次のように 100 行に分割します。可読性を高めることは一つのことですが、それは少し強迫的なものになる可能性もあります。その後、max、min、mid などのいくつかの統計手法が拡張され、トラバーサル手法も拡張されます。

以下は、test.c の呼び出しです:

#include <stdio.h>
#include "rbtree.h"

int main(int argc, char const *argv[])
{
    rbtree_t t = {};
    rbtree_node_t s = {};
    rbtree_init(&t, &s, rbtree_insert_value);

    const int cnt = 10;
    const int max_len = 15;

#define TEST_VALUES {"apple", "banana", "cherry", "grape", "lemon", "mango", "pear", "pineapple", "strawberry", "watermelon"}

    /* for gcc */
    char* v[] = TEST_VALUES;
    /* for g++ */
    // char v[][max_len] = TEST_VALUES;

    rbtree_node_t n[cnt];
    int i;
    for (i = 0; i value : "?");
    }

    rbtree_delete(&t, &n[5]);

    printf("\nafter delete 6->mango:\n\n");

    for (i = 1; i value : "?");
    }

    return 0;
}
</stdio.h>

rbtree_find メソッドのテスト行コメントのロックを解除し、スムーズに実行します:

key: 1
step count: 3, color: black, value: apple
key: 2
step count: 2, color: black, value: banana
key: 3
step count: 3, color: black, value: cherry
key: 4
step count: 1, color: black, value: grape
key: 5
step count: 3, color: black, value: lemon
key: 6
step count: 2, color: black, value: mango
key: 7
step count: 4, color: black, value: pear
key: 8
step count: 3, color: red, value: pineapple
key: 9
step count: 4, color: black, value: strawberry
key: 10
step count: 5, color: red, value: watermelon

after delete 6->mango:

key: 1
step count: 3, color: black, value: apple
key: 2
step count: 2, color: black, value: banana
key: 3
step count: 3, color: black, value: cherry
key: 4
step count: 1, color: black, value: grape
key: 5
step count: 3, color: black, value: lemon
key: 6
value: ?
key: 7
step count: 2, color: black, value: pear
key: 8
step count: 4, color: black, value: pineapple
key: 9
step count: 3, color: red, value: strawberry
key: 10
step count: 4, color: black, value: watermelon

以下は、6->mango と を削除する前の赤黒ツリーです。削除後の赤いツリー 黒いツリー図:

大量のデータに対してストレス テストを実行しましょう。 rbtree_find メソッドのテスト行をコメント アウトするように注意してください。そうしないと、恐ろしい結果が生じる可能性があります。

#include <stdio.h>
#include <stdlib.h>
#include <time.h>
#include "rbtree.h"

int main(int argc, char const *argv[])
{
    double duration;
    double room;

    rbtree_t t = {};
    rbtree_node_t s = {};
    rbtree_init(&t, &s, rbtree_insert_value);

    const int cnt = 1<br> 結果を見てみましょう: <p></p>
<pre name="code">Inserting 1048576 nodes costs 48.00MB and spends 0.425416 seconds.
Searching 1024 nodes among 1048576 spends 0.001140 seconds.
Hash 1024 times spends 0.000334 seconds.
Deleting 1024 nodes among 1048576 spends 0.000783 seconds.

削除は検索よりも速く、数百万件の挿入にも 0.5 秒もかかりません。かなり満足。

統計とトラバースメソッドを書いていきます。

著作権声明: この記事はブロガーによるオリジナルの記事であり、ブロガーの許可なく複製することはできません。

以上、nginxのデータ構造2～赤黒ツリーを自分で書き換える～を、関連内容も含めて紹介しましたので、PHPチュートリアルに興味のある方の参考になれば幸いです。