Server : Apache System : Linux host44.registrar-servers.com 4.18.0-513.18.1.lve.2.el8.x86_64 #1 SMP Sat Mar 30 15:36:11 UTC 2024 x86_64 User : vapecompany ( 2719) PHP Version : 7.4.33 Disable Function : NONE Directory : /opt/alt/ruby30/share/gems/gems/rdoc-6.3.4.1/lib/rdoc/parser/ |
Upload File : |
# frozen_string_literal: true ## # Parse a non-source file. We basically take the whole thing as one big # comment. class RDoc::Parser::Simple < RDoc::Parser include RDoc::Parser::Text parse_files_matching(//) attr_reader :content # :nodoc: ## # Prepare to parse a plain file def initialize(top_level, file_name, content, options, stats) super preprocess = RDoc::Markup::PreProcess.new @file_name, @options.rdoc_include @content = preprocess.handle @content, @top_level end ## # Extract the file contents and attach them to the TopLevel as a comment def scan comment = remove_coding_comment @content comment = remove_private_comment comment comment = RDoc::Comment.new comment, @top_level @top_level.comment = comment @top_level end ## # Removes the encoding magic comment from +text+ def remove_coding_comment text text.sub(/\A# .*coding[=:].*$/, '') end ## # Removes private comments. # # Unlike RDoc::Comment#remove_private this implementation only looks for two # dashes at the beginning of the line. Three or more dashes are considered # to be a rule and ignored. def remove_private_comment comment # Workaround for gsub encoding for Ruby 1.9.2 and earlier empty = '' empty = RDoc::Encoding.change_encoding empty, comment.encoding comment = comment.gsub(%r%^--\n.*?^\+\+\n?%m, empty) comment.sub(%r%^--\n.*%m, empty) end end