fix regression in latex2html

This commit is contained in:
Martin Monperrus
2016-08-20 17:52:51 +02:00
parent 2f5e6c7d5d
commit 6a54d0b2ec
2 changed files with 16 additions and 14 deletions
+13 -13
View File
@@ -1189,15 +1189,18 @@ class BibEntry {
}
function transformValue($value) {
// trim space
$value = xtrim($value);
if (c('BIBTEXBROWSER_USE_LATEX2HTML'))
{
// trim space
$value = xtrim($value);
// transform Latex markup to HTML entities (easier than a one to one mapping to each character)
// HTML entity is an intermediate format
$value = latex2html($value);
// transform to the target output encoding
$value = html_entity_decode($value, ENT_QUOTES|ENT_XHTML, OUTPUT_ENCODING);
// transform Latex markup to HTML entities (easier than a one to one mapping to each character)
// HTML entity is an intermediate format
$value = latex2html($value);
// transform to the target output encoding
$value = html_entity_decode($value, ENT_QUOTES|ENT_XHTML, OUTPUT_ENCODING);
}
return $value;
}
@@ -1210,11 +1213,8 @@ class BibEntry {
// we assume that "comment" is never latex code
// but instead could contain HTML code (with links using the character "~" for example)
// so "comment" is not transformed too
if ($name!='url' && $name!='comment') {
if (c('BIBTEXBROWSER_USE_LATEX2HTML')) {
if ($name!='url' && $name!='comment') {
$value = $this->transformValue($value);
}
// 4. transform existing encoded character in the new format
if (function_exists('mb_convert_encoding') && OUTPUT_ENCODING != BIBTEX_INPUT_ENCODING) {
@@ -3804,7 +3804,7 @@ class BibDataBase {
/** Adds a new bib entry to the database. */
function addEntry($entry) {
if (!$entry->hasField('key')) {
die('error: a bibliographic entry must have a key '.$entry->getText());
throw new Exception('error: a bibliographic entry must have a key '.$entry->getText());
}
// we keep its insertion order
$entry->order = count($this->bibdb);