Change some search index update code to do two regexps backwards.

These tail-anchored regexps are insanely slow. The worst case comes
when Japanese or Chinese text (ie, no word spacing) is written on
a wiki configured for Western UTF-8 mode: the Unicode characters are
expanded to hex codes and the "words" are very long paragraph-length
monstrosities. On a large page the previous regexps may take over 20
seconds *each* on a 1GHz-level processor on such pages of non-trivial
length...

The reversed versions which are consistently much faster, for instance
by a factor of 10000 for the pathological case and a factor of 10 for
more typical English-style pages.

diff --git a/includes/SearchUpdate.php b/includes/SearchUpdate.php
index fd0a10a..4c8a860 100644 (file)
--- a/includes/SearchUpdate.php
+++ b/includes/SearchUpdate.php
@@ -30,6 +30,9 @@ class SearchUpdate {
                if( $wgDisableSearchUpdate || !$this->mId ) {
                        return false;
                }
+               $fname = "SearchUpdate::doUpdate";
+               wfProfileIn( $fname );
+
                require_once( 'SearchEngine.php' );
                $lc = SearchEngine::legalSearchChars() . "&#;";
                $db =& wfGetDB( DB_MASTER );
@@ -42,12 +45,14 @@ class SearchUpdate {
                          $db->strencode( Title::indexTitle( $this->mNamespace, $this->mTitle ) ) .
                          "' WHERE si_page={$this->mId}";
                        $db->query( $sql, "SearchUpdate::doUpdate" );
+                       wfProfileOut( $fname );
                        return;
                }
 
                # Language-specific strip/conversion
                $text = $wgLang->stripForSearch( $this->mText );
 
+               wfProfileIn( "$fname-regexps" );
                $text = preg_replace( "/<\\/?\\s*[A-Za-z][A-Za-z0-9]*\\s*([^>]*?)>/",
                  " ", strtolower( " " . $text /*$this->mText*/ . " " ) ); # Strip HTML markup
                $text = preg_replace( "/(^|\\n)\\s*==\\s+([^\\n]+)\\s+==\\s/sD",
@@ -75,16 +80,32 @@ class SearchUpdate {
                $text = preg_replace( "/[^{$lc}]+/", " ", $text );
 
                # Handle 's, s'
-               $text = preg_replace( "/([{$lc}]+)'s /", "\\1 \\1's ", $text );
-               $text = preg_replace( "/([{$lc}]+)s' /", "\\1s ", $text );
+               #
+               #   $text = preg_replace( "/([{$lc}]+)'s /", "\\1 \\1's ", $text );
+               #   $text = preg_replace( "/([{$lc}]+)s' /", "\\1s ", $text );
+               #
+               # These tail-anchored regexps are insanely slow. The worst case comes
+               # when Japanese or Chinese text (ie, no word spacing) is written on
+               # a wiki configured for Western UTF-8 mode. The Unicode characters are
+               # expanded to hex codes and the "words" are very long paragraph-length
+               # monstrosities. On a large page the above regexps may take over 20
+               # seconds *each* on a 1GHz-level processor.
+               #
+               # Following are reversed versions which are consistently fast
+               # (about 3 milliseconds on 1GHz-level processor).
+               #
+               $text = strrev( preg_replace( "/ s'([{$lc}]+)/", " s'\\1 \\1", strrev( $text ) ) );
+               $text = strrev( preg_replace( "/ 's([{$lc}]+)/", " s\\1", strrev( $text ) ) );
 
                # Strip wiki '' and '''
                $text = preg_replace( "/''[']*/", " ", $text );
+               wfProfileOut( "$fname-regexps" );
                
                $sql = "REPLACE  INTO $searchindex (si_page,si_title,si_text) VALUES ({$this->mId},'" .
                  $db->strencode( Title::indexTitle( $this->mNamespace, $this->mTitle ) ) . "','" .
                  $db->strencode( $text ) . "')";
                $db->query( $sql, "SearchUpdate::doUpdate" );
+               wfProfileOut( $fname );
        }
 }