Bug #70232 [Ver->Ana]: Incorrect bump-along behavior with \K and empty string match

Edit report at https://bugs.php.net/bug.php?id=70232&edit=1 ID: 70232 Updated by: cmb@php.net Reported by: nhahtdh at gmail dot com Summary: Incorrect bump-along behavior with \K and empty string match -Status: Verified +Status: Analyzed Type: Bug Package: PCRE related PHP Version: 5.6.12 -Assigned To: +Assigned To: cmb Block user comment: N Private report: N New Comment: Great, thanks! PCRE_NOTEMPTY_ATSTART is exactly what is needed[1]. It's available as of PCRE 8.00 (2009-10-19)[2], so maybe we need to have a workaround for older libpcre, even though this bug would not be fixed for such versions. [1] <https://lists.exim.org/lurker/message/20090911.102109.6e80cce4.pt-BR.html> [2] <http://www.pcre.org/original/changelog.txt> Previous Comments: ------------------------------------------------------------------------ [2015-08-11 14:49:12] nhahtdh at gmail dot com Oh, you are right. I forgot that the exec function bumps the match along automatically, so relying on start_offset is a bad idea. Looking at the source code of pcretest, it seems that it uses a different option. Can you check whether the option is usable? http://vcs.pcre.org/pcre/code/branches/pcre16/pcretest.c?view=markup&pathrev=829#l4233 4233 if (use_offsets[0] == use_offsets[1]) 4234 { 4235 if (use_offsets[0] == len) break; 4236 g_notempty = PCRE_NOTEMPTY_ATSTART | PCRE_ANCHORED; 4237 } ------------------------------------------------------------------------ [2015-08-11 12:40:51] cmb@php.net I can confirm the issue, and that this case would be solved by your suggestion (checking for offsets[1] == start_offset). However, that would break some other cases, e.g. preg_replace('/\b/', '*', '(#11/19/2002#)') Expected: string(20) "(#*11*/*19*/*2002*#)" Actual: string(24) "(#**11**/*19**/*2002**#)" ------------------------------------------------------------------------ [2015-08-11 05:28:27] requinix@php.net Related To: Bug #70233 ------------------------------------------------------------------------ [2015-08-11 04:57:15] nhahtdh at gmail dot com Description: ------------ Please see the test case, the expected result and the actual result in the bug report below. From analyzing the code, I have identified that the error comes from the following line of code (plus similar code in other functions): https://github.com/php/php-src/blob/0787cd60ed3d0c8c8c8ff7e49b9bb3587bf33b64/ext/pcre/php_pcre.c#L891 g_notempty = (offsets[1] == offsets[0])? PCRE_NOTEMPTY | PCRE_ANCHORED : 0; /* Advance to the position right after the last full match */ start_offset = offsets[1]; This part of the code checks whether the match is empty. If the match it empty, it tries to perform another match from the last position of the match with PCRE_NOTEMPTY flag to force the match to be non-empty. However, it's incorrect to set the PCRE_NOTEMPTY flag by checking the length of the match based on the start and end indices of the match. For the regex in the test case ~(?: |\G)\d\B\K~ (which always results in empty string matches due to \K), the index where we execute the match from (start_offset) and the start index of the match (offsets[0]) are always different. Therefore, the regex actually advances ahead in the string, and setting PCRE_NOTEMPTY in such case rejects valid empty string matches. I believe (offsets[1] == offsets[0]) should be changed to (offsets[1] == start_offset). If start_offset <= offsets[0] <= offsets[1] is an invariant, then the change should be correct. Original report and analysis: http://chat.stackoverflow.com/transcript/message/24995536#24995536 Test script: --------------- <? $str = "123 a123 1234567 b123 123"; $str = preg_replace('~(?: |\G)\d\B\K~', "*", $str); echo $str; ?> Expected result: ---------------- 1*2*3 a123 1*2*3*4*5*6*7 b123 1*2*3 The expected result (10 matches/replacements) can be observed on - pcretest with PCRE version 8.35 2014-04-04 - regex101 https://regex101.com/r/oP9mZ7/3 Actual result: -------------- 1*23 a123 1*23*45*67 b123 1*23 ------------------------------------------------------------------------ -- Edit this bug report at https://bugs.php.net/bug.php?id=70232&edit=1