Change Log for PCRE2 -------------------- Version 10.21 xx-xxx-xxxx ------------------------- 1. Improve matching speed of patterns starting with + or * in JIT. 2. Use memchr() to find the first character in an unanchored match in 8-bit mode in the interpreter. This gives a significant speed improvement. Version 10.20 30-June-2015 -------------------------- 1. Callouts with string arguments have been added. 2. Assertion code generator in JIT has been optimized. 3. The invalid pattern (?(?C) has a missing assertion condition at the end. The pcre2_compile() function read past the end of the input before diagnosing an error. This bug was discovered by the LLVM fuzzer. 4. Implemented pcre2_callout_enumerate(). 5. Fix JIT compilation of conditional blocks whose assertion is converted to (*FAIL). E.g: /(?(?!))/. 6. The pattern /(?(?!)^)/ caused references to random memory. This bug was discovered by the LLVM fuzzer. 7. The assertion (?!) is optimized to (*FAIL). This was not handled correctly when this assertion was used as a condition, for example (?(?!)a|b). In pcre2_match() it worked by luck; in pcre2_dfa_match() it gave an incorrect error about an unsupported item. 8. For some types of pattern, for example /Z*(|d*){216}/, the auto- possessification code could take exponential time to complete. A recursion depth limit of 1000 has been imposed to limit the resources used by this optimization. This infelicity was discovered by the LLVM fuzzer. 9. A pattern such as /(*UTF)[\S\V\H]/, which contains a negated special class such as \S in non-UCP mode, explicit wide characters (> 255) can be ignored because \S ensures they are all in the class. The code for doing this was interacting badly with the code for computing the amount of space needed to compile the pattern, leading to a buffer overflow. This bug was discovered by the LLVM fuzzer. 10. A pattern such as /((?2)+)((?1))/ which has mutual recursion nested inside other kinds of group caused stack overflow at compile time. This bug was discovered by the LLVM fuzzer. 11. A pattern such as /(?1)(?#?'){8}(a)/ which had a parenthesized comment between a subroutine call and its quantifier was incorrectly compiled, leading to buffer overflow or other errors. This bug was discovered by the LLVM fuzzer. 12. The illegal pattern /(?(?.*!.*)?)/ was not being diagnosed as missing an assertion after (?(. The code was failing to check the character after (?(?< for the ! or = that would indicate a lookbehind assertion. This bug was discovered by the LLVM fuzzer. 13. A pattern such as /X((?2)()*+){2}+/ which has a possessive quantifier with a fixed maximum following a group that contains a subroutine reference was incorrectly compiled and could trigger buffer overflow. This bug was discovered by the LLVM fuzzer. 14. Negative relative recursive references such as (?-7) to non-existent subpatterns were not being diagnosed and could lead to unpredictable behaviour. This bug was discovered by the LLVM fuzzer. 15. The bug fixed in 14 was due to an integer variable that was unsigned when it should have been signed. Some other "int" variables, having been checked, have either been changed to uint32_t or commented as "must be signed". 16. A mutual recursion within a lookbehind assertion such as (?<=((?2))((?1))) caused a stack overflow instead of the diagnosis of a non-fixed length lookbehind assertion. This bug was discovered by the LLVM fuzzer. 17. The use of \K in a positive lookbehind assertion in a non-anchored pattern (e.g. /(?<=\Ka)/) could make pcre2grep loop. 18. There was a similar problem to 17 in pcre2test for global matches, though the code there did catch the loop. 19. If a greedy quantified \X was preceded by \C in UTF mode (e.g. \C\X*), and a subsequent item in the pattern caused a non-match, backtracking over the repeated \X did not stop, but carried on past the start of the subject, causing reference to random memory and/or a segfault. There were also some other cases where backtracking after \C could crash. This set of bugs was discovered by the LLVM fuzzer. 20. The function for finding the minimum length of a matching string could take a very long time if mutual recursion was present many times in a pattern, for example, /((?2){73}(?2))((?1))/. A better mutual recursion detection method has been implemented. This infelicity was discovered by the LLVM fuzzer. 21. Implemented PCRE2_NEVER_BACKSLASH_C. 22. The feature for string replication in pcre2test could read from freed memory if the replication required a buffer to be extended, and it was not working properly in 16-bit and 32-bit modes. This issue was discovered by a fuzzer: see http://lcamtuf.coredump.cx/afl/. 23. Added the PCRE2_ALT_CIRCUMFLEX option. 24. Adjust the treatment of \8 and \9 to be the same as the current Perl behaviour. 25. Static linking against the PCRE2 library using the pkg-config module was failing on missing pthread symbols. 26. If a group that contained a recursive back reference also contained a forward reference subroutine call followed by a non-forward-reference subroutine call, for example /.((?2)(?R)\1)()/, pcre2_compile() failed to compile correct code, leading to undefined behaviour or an internally detected error. This bug was discovered by the LLVM fuzzer. 27. Quantification of certain items (e.g. atomic back references) could cause incorrect code to be compiled when recursive forward references were involved. For example, in this pattern: /(?1)()((((((\1++))\x85)+)|))/. This bug was discovered by the LLVM fuzzer. 28. A repeated conditional group whose condition was a reference by name caused a buffer overflow if there was more than one group with the given name. This bug was discovered by the LLVM fuzzer. 29. A recursive back reference by name within a group that had the same name as another group caused a buffer overflow. For example: /(?J)(?'d'(?'d'\g{d}))/. This bug was discovered by the LLVM fuzzer. 30. A forward reference by name to a group whose number is the same as the current group, for example in this pattern: /(?|(\k'Pm')|(?'Pm'))/, caused a buffer overflow at compile time. This bug was discovered by the LLVM fuzzer. 31. Fix -fsanitize=undefined warnings for left shifts of 1 by 31 (it treats 1 as an int; fixed by writing it as 1u). 32. Fix pcre2grep compile when -std=c99 is used with gcc, though it still gives a warning for "fileno" unless -std=gnu99 us used. 33. A lookbehind assertion within a set of mutually recursive subpatterns could provoke a buffer overflow. This bug was discovered by the LLVM fuzzer. 34. Give an error for an empty subpattern name such as (?''). 35. Make pcre2test give an error if a pattern that follows #forbud_utf contains \P, \p, or \X. 36. The way named subpatterns are handled has been refactored. There is now a pre-pass over the regex which does nothing other than identify named subpatterns and count the total captures. This means that information about named patterns is known before the rest of the compile. In particular, it means that forward references can be checked as they are encountered. Previously, the code for handling forward references was contorted and led to several errors in computing the memory requirements for some patterns, leading to buffer overflows. 37. There was no check for integer overflow in subroutine calls such as (?123). 38. The table entry for \l in EBCDIC environments was incorrect, leading to its being treated as a literal 'l' instead of causing an error. 39. If a non-capturing group containing a conditional group that could match an empty string was repeated, it was not identified as matching an empty string itself. For example: /^(?:(?(1)x|)+)+$()/. 40. In an EBCDIC environment, pcretest was mishandling the escape sequences \a and \e in test subject lines. 41. In an EBCDIC environment, \a in a pattern was converted to the ASCII instead of the EBCDIC value. 42. The handling of \c in an EBCDIC environment has been revised so that it is now compatible with the specification in Perl's perlebcdic page. 43. Single character repetition in JIT has been improved. 20-30% speedup was achieved on certain patterns. 44. The EBCDIC character 0x41 is a non-breaking space, equivalent to 0xa0 in ASCII/Unicode. This has now been added to the list of characters that are recognized as white space in EBCDIC. 45. When PCRE2 was compiled without Unicode support, the use of \p and \P gave an error (correctly) when used outside a class, but did not give an error within a class. 46. \h within a class was incorrectly compiled in EBCDIC environments. 47. JIT should return with error when the compiled pattern requires more stack space than the maximum. 48. Fixed a memory leak in pcre2grep when a locale is set. Version 10.10 06-March-2015 --------------------------- 1. When a pattern is compiled, it remembers the highest back reference so that when matching, if the ovector is too small, extra memory can be obtained to use instead. A conditional subpattern whose condition is a check on a capture having happened, such as, for example in the pattern /^(?:(a)|b)(?(1)A|B)/, is another kind of back reference, but it was not setting the highest backreference number. This mattered only if pcre2_match() was called with an ovector that was too small to hold the capture, and there was no other kind of back reference (a situation which is probably quite rare). The effect of the bug was that the condition was always treated as FALSE when the capture could not be consulted, leading to a incorrect behaviour by pcre2_match(). This bug has been fixed. 2. Functions for serialization and deserialization of sets of compiled patterns have been added. 3. The value that is returned by PCRE2_INFO_SIZE has been corrected to remove excess code units at the end of the data block that may occasionally occur if the code for calculating the size over-estimates. This change stops the serialization code copying uninitialized data, to which valgrind objects. The documentation of PCRE2_INFO_SIZE was incorrect in stating that the size did not include the general overhead. This has been corrected. 4. All code units in every slot in the table of group names are now set, again in order to avoid accessing uninitialized data when serializing. 5. The (*NO_JIT) feature is implemented. 6. If a bug that caused pcre2_compile() to use more memory than allocated was triggered when using valgrind, the code in (3) above passed a stupidly large value to valgrind. This caused a crash instead of an "internal error" return. 7. A reference to a duplicated named group (either a back reference or a test for being set in a conditional) that occurred in a part of the pattern where PCRE2_DUPNAMES was not set caused the amount of memory needed for the pattern to be incorrectly calculated, leading to overwriting. 8. A mutually recursive set of back references such as (\2)(\1) caused a segfault at compile time (while trying to find the minimum matching length). The infinite loop is now broken (with the minimum length unset, that is, zero). 9. If an assertion that was used as a condition was quantified with a minimum of zero, matching went wrong. In particular, if the whole group had unlimited repetition and could match an empty string, a segfault was likely. The pattern (?(?=0)?)+ is an example that caused this. Perl allows assertions to be quantified, but not if they are being used as conditions, so the above pattern is faulted by Perl. PCRE2 has now been changed so that it also rejects such patterns. 10. The error message for an invalid quantifier has been changed from "nothing to repeat" to "quantifier does not follow a repeatable item". 11. If a bad UTF string is compiled with NO_UTF_CHECK, it may succeed, but scanning the compiled pattern in subsequent auto-possessification can get out of step and lead to an unknown opcode. Previously this could have caused an infinite loop. Now it generates an "internal error" error. This is a tidyup, not a bug fix; passing bad UTF with NO_UTF_CHECK is documented as having an undefined outcome. 12. A UTF pattern containing a "not" match of a non-ASCII character and a subroutine reference could loop at compile time. Example: /[^\xff]((?1))/. 13. The locale test (RunTest 3) has been upgraded. It now checks that a locale that is found in the output of "locale -a" can actually be set by pcre2test before it is accepted. Previously, in an environment where a locale was listed but would not set (an example does exist), the test would "pass" without actually doing anything. Also the fr_CA locale has been added to the list of locales that can be used. 14. Fixed a bug in pcre2_substitute(). If a replacement string ended in a capturing group number without parentheses, the last character was incorrectly literally included at the end of the replacement string. 15. A possessive capturing group such as (a)*+ with a minimum repeat of zero failed to allow the zero-repeat case if pcre2_match() was called with an ovector too small to capture the group. 16. Improved error message in pcre2test when setting the stack size (-S) fails. 17. Fixed two bugs in CMakeLists.txt: (1) Some lines had got lost in the transfer from PCRE1, meaning that CMake configuration failed if "build tests" was selected. (2) The file src/pcre2_serialize.c had not been added to the list of PCRE2 sources, which caused a failure to build pcre2test. 18. Fixed typo in pcre2_serialize.c (DECL instead of DEFN) that causes problems only on Windows. 19. Use binary input when reading back saved serialized patterns in pcre2test. 20. Added RunTest.bat for running the tests under Windows. 21. "make distclean" was not removing config.h, a file that may be created for use with CMake. 22. A pattern such as "((?2){0,1999}())?", which has a group containing a forward reference repeated a large (but limited) number of times within a repeated outer group that has a zero minimum quantifier, caused incorrect code to be compiled, leading to the error "internal error: previously-checked referenced subpattern not found" when an incorrect memory address was read. This bug was reported as "heap overflow", discovered by Kai Lu of Fortinet's FortiGuard Labs. (Added 24-March-2015: CVE-2015-2325 was given to this.) 23. A pattern such as "((?+1)(\1))/" containing a forward reference subroutine call within a group that also contained a recursive back reference caused incorrect code to be compiled. This bug was reported as "heap overflow", discovered by Kai Lu of Fortinet's FortiGuard Labs. (Added 24-March-2015: CVE-2015-2326 was given to this.) 24. Computing the size of the JIT read-only data in advance has been a source of various issues, and new ones are still appear unfortunately. To fix existing and future issues, size computation is eliminated from the code, and replaced by on-demand memory allocation. 25. A pattern such as /(?i)[A-`]/, where characters in the other case are adjacent to the end of the range, and the range contained characters with more than one other case, caused incorrect behaviour when compiled in UTF mode. In that example, the range a-j was left out of the class. Version 10.00 05-January-2015 ----------------------------- Version 10.00 is the first release of PCRE2, a revised API for the PCRE library. Changes prior to 10.00 are logged in the ChangeLog file for the old API, up to item 20 for release 8.36. The code of the library was heavily revised as part of the new API implementation. Details of each and every modification were not individually logged. In addition to the API changes, the following changes were made. They are either new functionality, or bug fixes and other noticeable changes of behaviour that were implemented after the code had been forked. 1. Including Unicode support at build time is now enabled by default, but it can optionally be disabled. It is not enabled by default at run time (no change). 2. The test program, now called pcre2test, was re-specified and almost completely re-written. Its input is not compatible with input for pcretest. 3. Patterns may start with (*NOTEMPTY) or (*NOTEMPTY_ATSTART) to set the PCRE2_NOTEMPTY or PCRE2_NOTEMPTY_ATSTART options for every subject line that is matched by that pattern. 4. For the benefit of those who use PCRE2 via some other application, that is, not writing the function calls themselves, it is possible to check the PCRE2 version by matching a pattern such as /(?(VERSION>=10)yes|no)/ against a string such as "yesno". 5. There are case-equivalent Unicode characters whose encodings use different numbers of code units in UTF-8. U+023A and U+2C65 are one example. (It is theoretically possible for this to happen in UTF-16 too.) If a backreference to a group containing one of these characters was greedily repeated, and during the match a backtrack occurred, the subject might be backtracked by the wrong number of code units. For example, if /^(\x{23a})\1*(.)/ is matched caselessly (and in UTF-8 mode) against "\x{23a}\x{2c65}\x{2c65}\x{2c65}", group 2 should capture the final character, which is the three bytes E2, B1, and A5 in UTF-8. Incorrect backtracking meant that group 2 captured only the last two bytes. This bug has been fixed; the new code is slower, but it is used only when the strings matched by the repetition are not all the same length. 6. A pattern such as /()a/ was not setting the "first character must be 'a'" information. This applied to any pattern with a group that matched no characters, for example: /(?:(?=.)|(?