Home  >  Article  >  Backend Development  >  About the optimization of string processing logic in PHP7!

About the optimization of string processing logic in PHP7!

藏色散人
藏色散人forward
2020-10-29 14:35:382309browse

推荐教程:《PHP7

一、 通过代码对比 PHP 5 和 PHP 7 中字符串处理的区别

   先看如下示例代码:

$a = 'foo';
$b = 'bar';

$c = "I like $a and $b";

   在 PHP 5.6 中执行代码,得到的 opcode 输出如下图:

   具体的执行过程:

  • 首先为 I like 申请内存
  • 然后为 I like foo 申请内存,然后将 I like 和 foo 复制到新申请的内存空间并返回
  • 继续为 I like foo and 申请内存,然后将 I like foo 和 and 复制到新申请的内存空间并返回
  • 最后为 I like foo and bar 申请内存,然后将 I like foo and 和 bar 复制到新申请的内存并返回
  • 释放字符串处理过程中申请的内存空间

   在 PHP 7 中执行代码,得到的 opcode 输出如下入:

   PHP 7 中对字符串的处理过程相对简单,首先创建一个堆栈 stack,然后将要连接的字符串片段存入栈空间,最后只需要分配一次内存空间,然后将结果从栈中移动到所分配的内存空间中。相较于 PHP 5,PHP 7 在处理过程中避免了反复申请内存的过程。而 PHP 7 在字符串处理方面性能的提升得益于数据结构 rope 的使用。

二、 Rope 数据结构介绍

⒈ Rope 介绍

   Rope 是一棵二叉树,其中每个叶子节点存储的是字符串的子串,每个非叶子节点存储的是位于该节点左侧的每个叶子节点所存储的子串中所包含的字符总数(方便了根据索引查找指定字符)。Rope 数据结构示意图

⒉ Rope 的优势和不足

   优势:

  • Rope makes string insertion, deletion and other operations faster. Compared with the traditional character array method, the time complexity of Rope is O(logN)O(\log{N}), characters The time complexity of the array is ##O(N)O(N)When operating a string as a character array, you first need to copy the array , which requires additional memory space
  • ##O(N)O(N)##OUsing the Rope data structure, the undo of string operations will be very convenient
  • Using Rope, no matter how big the size of the string being operated is, the performance is very stable
  •   Disadvantages:
  • Requires additional memory space to store non-leaf nodes, which increases memory space consumption compared to character arrays

Due to the different structure of binary trees Compared with character arrays, it is more complicated, so the performance is worse when operating strings with smaller sizes.

    Using Rope to operate strings increases the complexity of the code and also increases the risk of BUG
  • ⒊ Code implementation of Rope
  •  Rope is essentially a binary tree, and its basic insertion, deletion, and search operations are the same as those of a binary tree. Here we only introduce the concatenation (concat) and split (split) of strings.

In addition, in order to minimize the time complexity of the operation, Rope can be constructed as an AVL tree, so that self-balancing can be achieved after each operation is completed. But this may increase the complexity of some undo operations. For example, a new Rope formed by concating two Ropes with different heights may destroy the structure of the original two trees after self-balancing through node rotation. In this way, undoing the previous concat operation will become very complicated. .

⓵ concat

The operation of concat is relatively simple. You only need to join two Rope trees into a new Rope tree.

⓶ split
  When splitting a Rope tree, you will encounter two situations: the split position is exactly at the end of a node or the split position is in the middle of a leaf node. For the second case, we can split the corresponding leaf nodes to transform into the first case. As for the first situation, it can be divided into two situations according to the location of the node:

À ˆ a. The branch where the node is located on the far left or the far right

b. The branch whose node is located inside the Rope tree

## #

class Node:
    def __init__(self, data):
        self.parent = None
        self.left = None
        self.right = None
        self.data = data
        self.weight = 0

    def __repr__(self):
        return str(self.__dict__)

    def __str__(self):
        return str(self.__dict__)


class Rope:
    # 每个叶子节点最多存储 5 个字符,包括空字符在内
    LEAF_DATA_LEN = 5

    def __init__(self):
        self.root = None

    def create_rope(self, parent, data, left_index, right_index):
        """
        创建 rope 数据结构
        :param parent: 父节点(根节点的父节点为空)
        :param data: 用于创建 rope 数据结构的原始数据,这里只接受 list 和 str 两种类型
        :param left_index: 起始位置索引
        :param right_index: 结束位置索引
        :return: Node
        """
        if isinstance(data, str):
            data = list(data)
        elif not isinstance(data, list):
            return

        if right_index - left_index > self.LEAF_DATA_LEN:
            node = Node("")

            node.parent = parent
            middle_index = (left_index + right_index) // 2
            node.weight = middle_index - left_index
            node.left = self.create_rope(node, data, left_index, middle_index)
            node.right = self.create_rope(node, data, middle_index, right_index)
        else:
            node = Node(data[left_index: right_index])

            node.parent = parent
            node.weight = right_index - left_index

        if node.parent is None:
            self.root = node

        return node

    @staticmethod
    def calc_weight(node):
        """
        计算节点 weight 值
        :param node:
        :return:
        """
        if node is None:
            return 0

        init_weight = node.weight
        while node.right is not None:
            node = node.right
            init_weight += node.weight

        return init_weight

    def concat_rope(self, data1, data2):
        """
        字符串连接
        :param data1:
        :param data2:
        :return:
        """
        r1 = Rope()
        r1.create_rope(None, data1, 0, len(data1))
        r2 = Rope()
        r2.create_rope(None, data2, 0, len(data2))

        node = Node("")
        node.left = r1.root
        node.right = r2.root
        r1.root.parent = node
        r2.root.parent = node

        node.weight = self.calc_weight(r1)

        self.root = node

    def split_rope(self, data, index):
        """
        字符串拆分
        :param data: 要拆分的字符串
        :param index: 拆分的位置(字符串索引从 0 开始计算)
        :return: Rope
        """
        if index < 0 or index > len(data) - 1:
            return

        node = self.create_rope(None, data, 0, len(data))
        original_index = index

        if index == self.root.weight - 1:
            # 在根节点拆分
            rope_left = node.left
            rope_left.parent = None
            rope_right = node.right
            rope_right.parent = None
            return rope_left, rope_right
        elif index < self.root.weight - 1:
            while index < node.weight - 1 and node.data == "":
                node = node.left
        else:
            while index > node.weight - 1 and node.data == "":
                index -= node.weight
                node = node.right

        if node.data != "":
            # index 落在了最左侧和最右侧的两个叶子节点
            if original_index < self.root.weight - 1:
                # index 落在了最左侧的叶子节点
                rope_left = self.create_rope(None, node.data[0:index + 1], 0, index + 1)
                rope_right = self.root
                # 更新 rope_right 的 weight
                node.data = node.data[index + 1:]
                while node is not None:
                    node.weight -= (index + 1)
                    node = node.parent
            else:
                # index 落在了最右侧的叶子节点
                rope_left = self.root
                rope_right = self.create_rope(None, node.data[index + 1:], 0, len(node.data[index + 1:]))
                node.data = node.data[0:index + 1]
        elif index == node.weight - 1:
            # index 正好落在了节点的末尾
            if original_index < self.root.weight:
                # index 落在了最左侧分支中的非叶子节点的末尾
                weight_sub = node.weight
                rope_left = node.left
                rope_left.parent = None
                node.left = None
                rope_right = self.root
                # 更新节点 weight
                while node is not None:
                    node.weight -= weight_sub
                    node = node.parent
            else:
                # index 落在了最右侧分支中的非叶子节点的末尾
                rope_left = self.root
                rope_right = node.right
                rope_right.parent = None
                node.right = None
        else:
            stack = []
            if original_index < self.root.weight:
                # index 落在了左子树中的节点
                index -= node.weight
                rope_left = node
                rope_right = self.root
                node.parent.left = None
                node.parent = None
                node = node.right
            else:
                # index 落在了右子树中的节点
                rope_left = self.root
                stack.append(node.right)
                rope_right = None
                node.right = None
                node = node.left
            while node.data == "" and index >= 0:
                if index < node.weight - 1:
                    stack.append(node.right)
                    node.right = None
                    node = node.left
                elif index > node.weight - 1:
                    node = node.right
                    index -= node.weight
                else:
                    stack.append(node.right)
                    node.right = None
                    break
            if node.data != "":
                # 需要拆分叶子节点
                new_node = Node(node.data[index + 1:])
                new_node.weight = node.weight - index - 1
                stack.append(new_node)
                node.data = node.data[0:index + 1]
            # 更新节点的 weight 信息
            while node is not None:
                if node.data != "":
                    node.weight = len(node.data)
                else:
                    node.weight = self.calc_weight(node.left)
                node = node.parent
            # 组装 rope_right并更新节点的 weight 值
            left_node = None
            while len(stack) > 0:
                root_node = Node("")
                if left_node is None:
                    left_node = stack.pop()
                    root_node = left_node
                else:
                    root_node.left = left_node
                    left_node.parent = root_node
                    right_node = stack.pop()
                    root_node.right = right_node
                    right_node.parent = root_node
                    root_node.weight = self.calc_weight(root_node.left)
                    left_node = root_node

            if rope_right is None:
                # index > self.root.weight - 1
                rope_right = root_node
            else:
                # index < self.root.weight - 1
                tmp = rope_right
                while tmp.left is not None:
                    tmp = tmp.left
                tmp.left = root_node
                root_node.parent = tmp
                while tmp.parent is not None:
                    tmp.weight = self.calc_weight(tmp.left)
                    tmp = tmp.parent
                rope_right = tmp
                rope_right.weight = self.calc_weight(rope_right.left)

        return rope_left, rope_right


rope = Rope()
data = "php is a script language"
index = 18
left, right = rope.split_rope(data, index)
print(left)
print(right)

三、 PHP 5 和 PHP 7 底层字符串处理逻辑比较

⒈ PHP 5 中字符串处理逻辑以及存在的问题

⓵ 处理逻辑
  • PHP 5 中的字符串并没有固定的数据结构,而是采用 C 中以 NULL 结尾的字符数组的方式来处理
  • 为了支持二进制字符串(字符串中包含 NULL 字符),还需要额外记录字符串的长度
⓶ 存在的问题
  a. 字符串的长度

   PHP 5 中 zval 中定义的字符串的长度为 int(有符号整型) 类型,这就导致即使是在 64 为机器上字符串的长度也不能超过 2312^{31} (LP64 数据模型中,int 永远只有 32 位)。

  b. 格式不统一
// zval 中对字符串的定义,长度为 int 类型
typedef union _zvalue_value {
		long lval;
		double dval;
		struct {
			char *val;     /* C string buffer, NULL terminated */
			int len;      /* String length : num of ASCII chars */
		} str;            /* string structure */
		HashTable *ht;
		zend_object_value obj;
		zend_ast *ast;
} zvalue_value;
	
// 类实例结构体中对字符串的定义,此时字符串的长度为 zend_uint 类型,相较于 int 类型,字符串长度提升了一倍
struct _zend_class_entry {
	char type;
	const char *name;		/* C string buffer, NULL terminated */
	zend_uint name_length;		/* String length : num of ASCII chars */
	struct _zend_class_entry *parent;
	int refcount;
	zend_uint ce_flags;

	/*……*/
}	

// hashtable 的 key 中对字符串的定义,长度为 uint 类型
typedef struct _zend_hash_key {
		const char *arKey;  /* C string buffer, NULL terminated */
		uint nKeyLength;    /* String length : num of ASCII chars */
		ulong h;
} zend_hash_key;

    PHP 5 中很多地方都支持二进制字符串,由于字符串没有一个固定的结构体,这就导致很多地方都有对字符串的定义。同时,由于各处对字符串长度的定义不同,导致各处支持的字符串长度也不同。

  c. 内存消耗大

   在较老版本的 PHP 中,由于没有引入 interned string,同一个字符串如果在多处被使用,为了互不影响,就需要复制多份出来,这就造成了对内存空间大量消耗。

static PHP_FUNCTION(session_id)
	{
		char *name = NULL;
		int name_len, argc = ZEND_NUM_ARGS();

		if (zend_parse_parameters(argc TSRMLS_CC, "|s", &name, &name_len) == FAILURE) {
		    return;
		}
	
		/* …… */

		if (name) {
		    if (PS(id)) {
		        efree(PS(id));
		    }
		    PS(id) = estrndup(name, name_len);
		}
	}

   以 PHP 函数 session_id 为例,该函数最终将 name 完全复制一份然后存入 PS(id)。这样,后续代码对 name 进行的任何操作都不会影响 PS(id) 中存储的值。

⓷ interned string

   从 PHP 5.4 起,为了解决上述字符串复制导致内存消耗大的问题,PHP 引入了 interned string 。另外,由于 interned string 需要共享内存空间,所以在线程安全的 PHP 版本中并不被支持。

   所谓 interned string,即一个字符串在一个进程中只存储一次。当一个 php-fpm 进程启动时,会申请一块 size 为 1MB 的 buffer,持久化的字符串(字符串常量、函数名、变量名、类名、方法名……)会被存储到该缓冲并添加到一个 hashtable 中。由于 PHP 处理的 request 之间是相互独立的,所以一个 request 处理完成后,其间在内存中产生的数据都会被销毁。为了在处理 request 的过程中也能使用 interned string,在 request 到来时 PHP 会记录当前 interned string buffer 的 top 位置,在该 request 请求的处理完成后,top 位置之后消耗的空间又会被恢复。

interned string buffer 示意图

  • interned string 的初始化,只会发生在 PHP 启动以及 PHP 脚本的编译阶段
  • interned string 是只读的,既不能修改,也不能销毁
  • 由于 interned string buffer 只有 1MB 的空间,当 1MB 的空间存满后,后续的字符串处理还会回到引入 interned string 之前的状态

   interned string 的优势

  • 一个字符串只会存储一次,节省了内存空间
  • interned string 的 hash 只会计算一次,然后被多次使用
  • interned string 的比较最终转换成了指针地址的比较

   由于需要操作 hashtable,interned string 的初始化和创建会比较复杂,也正因为如此,并不是所有的字符串都适合存入 interned string。

⒉ PHP 7 中对字符串处理进行的优化

struct _zend_string {
	zend_refcounted_h gc;
	zend_ulong        h;                /* hash value */
	size_t            len;
	char              val[1];
};

typedef struct _zend_refcounted_h {
	uint32_t         refcount;			/* reference counter 32-bit */
	union {
		uint32_t type_info;
	} u;
} zend_refcounted_h;

   PHP 7 中字符串有了固定的结构 zend_string

   PHP 7 中字符串的长度类型为 size_t,字符串的长度不再局限于 PHP 5 中的 2312^{31} ,尤其在 64 位的机器上,字符串的长度取决于平台支持的最大长度。

  PHP 7 中字符串内容的存储不再像之前使用 char * ,而是使用了 struct hack ,使得字符串内容和 zend_string 一起存储在一块连续的内存空间。同时,zend_string 还内嵌了字符串的 hash 值,这样,对于一个指定的字符串只需要进行一次 hash 计算。

  PHP 7 中字符串引入了引用计数,这样,只要字符串还在被使用,就不用担心会被销毁。

static PHP_FUNCTION(session_id)
{
	zend_string *name = NULL;
	int argc = ZEND_NUM_ARGS();

	/* …… */

	if (name) {
	    if (PS(id)) {
	        zend_string_release(PS(id));
	    }
	    PS(id) = zend_string_copy(name);
	}
}

static zend_always_inline zend_string *zend_string_copy(zend_string *s)
{
	if (!ZSTR_IS_INTERNED(s)) {
	    GC_REFCOUNT(s)++;
	}
	return s;
}

static zend_always_inline void zend_string_release(zend_string *s)
{
	if (!ZSTR_IS_INTERNED(s)) {
	    if (--GC_REFCOUNT(s) == 0) {
	        pefree(s, GC_FLAGS(s) & IS_STR_PERSISTENT);
	    }
	}
}

  仍以函数 session_id 为例,设置 session_id 时不再需要将 name 完全复制,而只是将 name 的引用计数加 1。在删除字符串时也是同样,将字符串的引用计数减 1,只有引用计数为 0 时才会真正销毁字符串。

  PHP 7 中的 interned string 不再需要单独申请 interned string buffer 来存储,而是在 zend_string 的 gc 中将 type_info 标记为 IS_STR_INTERNED。这样,在销毁字符串时,zend_string_release 会检查字符串是否为 interned string,如果是则不会进行任何操作。


The above is the detailed content of About the optimization of string processing logic in PHP7!. For more information, please follow other related articles on the PHP Chinese website!

Statement:
This article is reproduced at:juejin.im. If there is any infringement, please contact admin@php.cn delete