Rss-Bridge/lib/html.php
2018-11-05 12:55:58 +01:00

176 lines
6.3 KiB
PHP

<?php
function sanitize($textToSanitize,
$removedTags = array('script', 'iframe', 'input', 'form'),
$keptAttributes = array('title', 'href', 'src'),
$keptText = array()){
$htmlContent = str_get_html($textToSanitize);
foreach($htmlContent->find('*[!b38fd2b1fe7f4747d6b1c1254ccd055e]') as $element) {
if(in_array($element->tag, $keptText)) {
$element->outertext = $element->plaintext;
} elseif(in_array($element->tag, $removedTags)) {
$element->outertext = '';
} else {
foreach($element->getAllAttributes() as $attributeName => $attribute) {
if(!in_array($attributeName, $keptAttributes))
$element->removeAttribute($attributeName);
}
}
}
return $htmlContent;
}
function backgroundToImg($htmlContent) {
$regex = '/background-image[ ]{0,}:[ ]{0,}url\([\'"]{0,}(.*?)[\'"]{0,}\)/';
$htmlContent = str_get_html($htmlContent);
foreach($htmlContent->find('*[!b38fd2b1fe7f4747d6b1c1254ccd055e]') as $element) {
if(preg_match($regex, $element->style, $matches) > 0) {
$element->outertext = '<img style="display:block;" src="' . $matches[1] . '" />';
}
}
return $htmlContent;
}
/**
* Convert relative links in HTML into absolute links
* @param $content HTML content to fix. Supports HTML objects or string objects
* @param $server full URL to the page containing relative links
* @return content with fixed URLs, as HTML object or string depending on input type
*/
function defaultLinkTo($content, $server){
$string_convert = false;
if (is_string($content)) {
$string_convert = true;
$content = str_get_html($content);
}
foreach($content->find('img') as $image) {
$image->src = urljoin($server, $image->src);
}
foreach($content->find('a') as $anchor) {
$anchor->href = urljoin($server, $anchor->href);
}
if ($string_convert) {
$content = $content->outertext;
}
return $content;
}
/**
* Extract the first part of a string matching the specified start and end delimiters
* @param $string input string, e.g. '<div>Post author: John Doe</div>'
* @param $start start delimiter, e.g. 'author: '
* @param $end end delimiter, e.g. '<'
* @return extracted string, e.g. 'John Doe', or false if the delimiters were not found.
*/
function extractFromDelimiters($string, $start, $end) {
if (strpos($string, $start) !== false) {
$section_retrieved = substr($string, strpos($string, $start) + strlen($start));
$section_retrieved = substr($section_retrieved, 0, strpos($section_retrieved, $end));
return $section_retrieved;
} return false;
}
/**
* Remove one or more part(s) of a string using a start and end delmiters
* @param $string input string, e.g. 'foo<script>superscript()</script>bar'
* @param $start start delimiter, e.g. '<script'
* @param $end end delimiter, e.g. '</script>'
* @return cleaned string, e.g. 'foobar'
*/
function stripWithDelimiters($string, $start, $end) {
while(strpos($string, $start) !== false) {
$section_to_remove = substr($string, strpos($string, $start));
$section_to_remove = substr($section_to_remove, 0, strpos($section_to_remove, $end) + strlen($end));
$string = str_replace($section_to_remove, '', $string);
}
return $string;
}
/**
* Remove HTML sections containing one or more sections using the same HTML tag
* @param $string input string, e.g. 'foo<div class="ads"><div>ads</div>ads</div>bar'
* @param $tag_name name of the HTML tag, e.g. 'div'
* @param $tag_start start of the HTML tag to remove, e.g. '<div class="ads">'
* @return cleaned string, e.g. 'foobar'
*/
function stripRecursiveHTMLSection($string, $tag_name, $tag_start){
$open_tag = '<' . $tag_name;
$close_tag = '</' . $tag_name . '>';
$close_tag_length = strlen($close_tag);
if(strpos($tag_start, $open_tag) === 0) {
while(strpos($string, $tag_start) !== false) {
$max_recursion = 100;
$section_to_remove = null;
$section_start = strpos($string, $tag_start);
$search_offset = $section_start;
do {
$max_recursion--;
$section_end = strpos($string, $close_tag, $search_offset);
$search_offset = $section_end + $close_tag_length;
$section_to_remove = substr($string, $section_start, $section_end - $section_start + $close_tag_length);
$open_tag_count = substr_count($section_to_remove, $open_tag);
$close_tag_count = substr_count($section_to_remove, $close_tag);
} while ($open_tag_count > $close_tag_count && $max_recursion > 0);
$string = str_replace($section_to_remove, '', $string);
}
}
return $string;
}
/**
* Convert Markdown tags into HTML tags. Only a subset of the Markdown syntax is implemented.
* @param $string input string in Markdown format
* @return output string in HTML format
*/
function markdownToHtml($string) {
//For more details about how these regex work:
// https://github.com/RSS-Bridge/rss-bridge/pull/802#discussion_r216138702
// Images: https://regex101.com/r/JW9Evr/1
// Links: https://regex101.com/r/eRGVe7/1
// Bold: https://regex101.com/r/2p40Y0/1
// Italic: https://regex101.com/r/xJkET9/1
// Separator: https://regex101.com/r/ZBEqFP/1
// Plain URL: https://regex101.com/r/2JHYwb/1
// Site name: https://regex101.com/r/qIuKYE/1
$string = preg_replace('/\!\[([^\]]+)\]\(([^\) ]+)(?: [^\)]+)?\)/', '<img src="$2" alt="$1" />', $string);
$string = preg_replace('/\[([^\]]+)\]\(([^\)]+)\)/', '<a href="$2">$1</a>', $string);
$string = preg_replace('/\*\*(.*)\*\*/U', '<b>$1</b>', $string);
$string = preg_replace('/\*(.*)\*/U', '<i>$1</i>', $string);
$string = preg_replace('/__(.*)__/U', '<b>$1</b>', $string);
$string = preg_replace('/_(.*)_/U', '<i>$1</i>', $string);
$string = preg_replace('/[-]{6,99}/', '<hr />', $string);
$string = str_replace('&#10;', '<br />', $string);
$string = preg_replace('/([^"])(https?:\/\/[^ "<]+)([^"])/', '$1<a href="$2">$2</a>$3', $string . ' ');
$string = preg_replace('/([^"\/])(www\.[^ "<]+)([^"])/', '$1<a href="http://$2">$2</a>$3', $string . ' ');
//As the regex are not perfect, we need to fix <i> and </i> that are introduced in URLs
// Fixup regex <i>: https://regex101.com/r/NTRPf6/1
// Fixup regex </i>: https://regex101.com/r/aNklRp/1
$count = 1;
while($count > 0) {
$string = preg_replace('/ (src|href)="([^"]+)<i>([^"]+)"/U', ' $1="$2_$3"', $string, -1, $count);
}
$count = 1;
while($count > 0) {
$string = preg_replace('/ (src|href)="([^"]+)<\/i>([^"]+)"/U', ' $1="$2_$3"', $string, -1, $count);
}
return '<div>' . trim($string) . '</div>';
}