• 欢迎访问开心洋葱网站,在线教程,推荐使用最新版火狐浏览器和Chrome浏览器访问本网站,欢迎加入开心洋葱 QQ群
  • 为方便开心洋葱网用户,开心洋葱官网已经开启复制功能!
  • 欢迎访问开心洋葱网站,手机也能访问哦~欢迎加入开心洋葱多维思维学习平台 QQ群
  • 如果您觉得本站非常有看点,那么赶紧使用Ctrl+D 收藏开心洋葱吧~~~~~~~~~~~~~!
  • 由于近期流量激增,小站的ECS没能经的起亲们的访问,本站依然没有盈利,如果各位看如果觉着文字不错,还请看官给小站打个赏~~~~~~~~~~~~~!

一个python编写的用于搜索文件并进行内容替换的类

PHP 水墨上仙 2590次浏览 0个评论

一个python编写的用于搜索文件并进行内容替换的类

#!/usr/bin/python -O
# coding: UTF-8
"""
-replace string in files (recursive)
-display the difference.
v0.2
 - search_string can be a re.compile() object -> use re.sub for replacing
 
v0.1
 - initial version
 
    Useable by a small "client" script, e.g.:
-------------------------------------------------------------------------------
#!/usr/bin/python -O
# coding: UTF-8
import sys, re
#sys.path.insert(0,"/path/to/git/repro/") # Please change path
from replace_in_files import SearchAndReplace
SearchAndReplace(
    search_path = "/to/the/files/",
    # e.g.: simple string replace:
    search_string  = 'the old string',
    replace_string = 'the new string',
    # e.g.: Regular expression replacing (used re.sub)
    #search_string  = re.compile('{% url (.*?) %}'),
    #replace_string = "{% url '\g<1>' %}",
    search_only = True, # Display only the difference
    #search_only = False, # write the new content
    file_filter=("*.py",), # fnmatch-Filter
)
-------------------------------------------------------------------------------
:copyleft: 2009-2011 by Jens Diemer
"""
__author__ = "Jens Diemer"
__license__ = """GNU General Public License v3 or above -
 http://www.opensource.org/licenses/gpl-license.php"""
__url__ = "http://www.jensdiemer.de"
__version__ = "0.2"
import os, re, time, fnmatch, difflib
# FIXME: see http://stackoverflow.com/questions/4730121/cant-get-an-objects-class-name-in-python
RE_TYPE = type(re.compile(""))
class SearchAndReplace(object):
    def __init__(self, search_path, search_string, replace_string,
                                        search_only=True, file_filter=("*.*",)):
        self.search_path = search_path
        self.search_string = search_string
        self.replace_string = replace_string
        self.search_only = search_only
        self.file_filter = file_filter
        assert isinstance(self.file_filter, (list, tuple))
        # FIXME: see http://stackoverflow.com/questions/4730121/cant-get-an-objects-class-name-in-python
        self.is_re = isinstance(self.search_string, RE_TYPE)
        print "Search '%s' in [%s]..." % (
            self.search_string, self.search_path
        )
        print "_" * 80
        time_begin = time.time()
        file_count = self.walk()
        print "_" * 80
        print "%s files searched in %0.2fsec." % (
            file_count, (time.time() - time_begin)
        )
    def walk(self):
        file_count = 0
        for root, dirlist, filelist in os.walk(self.search_path):
            if ".svn" in root:
                continue
            for filename in filelist:
                for file_filter in self.file_filter:
                    if fnmatch.fnmatch(filename, file_filter):
                        self.search_file(os.path.join(root, filename))
                        file_count += 1
        return file_count
    def search_file(self, filepath):
        f = file(filepath, "r")
        old_content = f.read()
        f.close()
        if self.is_re or self.search_string in old_content:
            new_content = self.replace_content(old_content, filepath)
            if self.is_re and new_content == old_content:
                return
            print filepath
            self.display_plaintext_diff(old_content, new_content)
    def replace_content(self, old_content, filepath):
        if self.is_re:
            new_content = self.search_string.sub(self.replace_string, old_content)
            if new_content == old_content:
                return old_content
        else:
            new_content = old_content.replace(
                self.search_string, self.replace_string
            )
        if self.search_only != False:
            return new_content
        print "Write new content into %s..." % filepath,
        try:
            f = file(filepath, "w")
            f.write(new_content)
            f.close()
        except IOError, msg:
            print "Error:", msg
        else:
            print "OK"
        print
        return new_content
    def display_plaintext_diff(self, content1, content2):
        """
        Display a diff.
        """
        content1 = content1.splitlines()
        content2 = content2.splitlines()
        diff = difflib.Differ().compare(content1, content2)
        def is_diff_line(line):
            for char in ("-", "+", "?"):
                if line.startswith(char):
                    return True
            return False
        print "line | text\n-------------------------------------------"
        old_line = ""
        in_block = False
        old_lineno = lineno = 0
        for line in diff:
            if line.startswith(" ") or line.startswith("+"):
                lineno += 1
            if old_lineno == lineno:
                display_line = "%4s | %s" % ("", line.rstrip())
            else:
                display_line = "%4s | %s" % (lineno, line.rstrip())
            if is_diff_line(line):
                if not in_block:
                    print "..."
                    # Display previous line
                    print old_line
                    in_block = True
                print display_line
            else:
                if in_block:
                    # Display the next line aber a diff-block
                    print display_line
                in_block = False
            old_line = display_line
            old_lineno = lineno
        print "..."
if __name__ == "__main__":
    SearchAndReplace(
        search_path=".",
        # e.g.: simple string replace:
        search_string='the old string',
        replace_string='the new string',
        # e.g.: Regular expression replacing (used re.sub)
        #search_string   = re.compile('{% url (.*?) %}'),
        #replace_string  = "{% url '\g<1>' %}",
        search_only=True, # Display only the difference
#        search_only     = False, # write the new content
        file_filter=("*.py",), # fnmatch-Filter
    )


开心洋葱 , 版权所有丨如未注明 , 均为原创丨未经授权请勿修改 , 转载请注明一个python编写的用于搜索文件并进行内容替换的类
喜欢 (0)

您必须 登录 才能发表评论!

加载中……