/[pcre]/code/trunk/ChangeLog
ViewVC logotype

Diff of /code/trunk/ChangeLog

Parent Directory Parent Directory | Revision Log Revision Log | View Patch Patch

revision 45 by nigel, Sat Feb 24 21:39:25 2007 UTC revision 1599 by zherczeg, Wed Sep 2 08:50:31 2015 UTC
# Line 1  Line 1 
1  ChangeLog for PCRE  ChangeLog for PCRE
2  ------------------  ------------------
3    
4    Note that the PCRE 8.xx series (PCRE1) is now in a bugfix-only state. All
5    development is happening in the PCRE2 10.xx series.
6    
7    Version 8.38 xx-xxx-xxxx
8    ------------------------
9    
10    1.  If a group that contained a recursive back reference also contained a
11        forward reference subroutine call followed by a non-forward-reference
12        subroutine call, for example /.((?2)(?R)\1)()/, pcre2_compile() failed to
13        compile correct code, leading to undefined behaviour or an internally
14        detected error. This bug was discovered by the LLVM fuzzer.
15    
16    2.  Quantification of certain items (e.g. atomic back references) could cause
17        incorrect code to be compiled when recursive forward references were
18        involved. For example, in this pattern: /(?1)()((((((\1++))\x85)+)|))/.
19        This bug was discovered by the LLVM fuzzer.
20    
21    3.  A repeated conditional group whose condition was a reference by name caused
22        a buffer overflow if there was more than one group with the given name.
23        This bug was discovered by the LLVM fuzzer.
24    
25    4.  A recursive back reference by name within a group that had the same name as
26        another group caused a buffer overflow. For example:
27        /(?J)(?'d'(?'d'\g{d}))/. This bug was discovered by the LLVM fuzzer.
28    
29    5.  A forward reference by name to a group whose number is the same as the
30        current group, for example in this pattern: /(?|(\k'Pm')|(?'Pm'))/, caused
31        a buffer overflow at compile time. This bug was discovered by the LLVM
32        fuzzer.
33    
34    6.  A lookbehind assertion within a set of mutually recursive subpatterns could
35        provoke a buffer overflow. This bug was discovered by the LLVM fuzzer.
36    
37    7.  Another buffer overflow bug involved duplicate named groups with a
38        reference between their definition, with a group that reset capture
39        numbers, for example: /(?J:(?|(?'R')(\k'R')|((?'R'))))/. This has been
40        fixed by always allowing for more memory, even if not needed. (A proper fix
41        is implemented in PCRE2, but it involves more refactoring.)
42    
43    8.  There was no check for integer overflow in subroutine calls such as (?123).
44    
45    9.  The table entry for \l in EBCDIC environments was incorrect, leading to its
46        being treated as a literal 'l' instead of causing an error.
47    
48    10. There was a buffer overflow if pcre_exec() was called with an ovector of
49        size 1. This bug was found by american fuzzy lop.
50    
51    11. If a non-capturing group containing a conditional group that could match
52        an empty string was repeated, it was not identified as matching an empty
53        string itself. For example: /^(?:(?(1)x|)+)+$()/.
54    
55    12. In an EBCDIC environment, pcretest was mishandling the escape sequences
56        \a and \e in test subject lines.
57    
58    13. In an EBCDIC environment, \a in a pattern was converted to the ASCII
59        instead of the EBCDIC value.
60    
61    14. The handling of \c in an EBCDIC environment has been revised so that it is
62        now compatible with the specification in Perl's perlebcdic page.
63    
64    15. The EBCDIC character 0x41 is a non-breaking space, equivalent to 0xa0 in
65        ASCII/Unicode. This has now been added to the list of characters that are
66        recognized as white space in EBCDIC.
67    
68    16. When PCRE was compiled without UCP support, the use of \p and \P gave an
69        error (correctly) when used outside a class, but did not give an error
70        within a class.
71    
72    17. \h within a class was incorrectly compiled in EBCDIC environments.
73    
74    18. A pattern with an unmatched closing parenthesis that contained a backward
75        assertion which itself contained a forward reference caused buffer
76        overflow. And example pattern is: /(?=di(?<=(?1))|(?=(.))))/.
77    
78    19. JIT should return with error when the compiled pattern requires more stack
79        space than the maximum.
80    
81    20. A possessively repeated conditional group that could match an empty string,
82        for example, /(?(R))*+/, was incorrectly compiled.
83    
84    21. Fix infinite recursion in the JIT compiler when certain patterns such as
85        /(?:|a|){100}x/ are analysed.
86    
87    22. Some patterns with character classes involving [: and \\ were incorrectly
88        compiled and could cause reading from uninitialized memory or an incorrect
89        error diagnosis.
90    
91    23. Pathological patterns containing many nested occurrences of [: caused
92        pcre_compile() to run for a very long time.
93    
94    24. A conditional group with only one branch has an implicit empty alternative
95        branch and must therefore be treated as potentially matching an empty
96        string.
97    
98    25. If (?R was followed by - or + incorrect behaviour happened instead of a
99        diagnostic.
100    
101    26. Arrange to give up on finding the minimum matching length for overly
102        complex patterns.
103    
104    27. Similar to (4) above: in a pattern with duplicated named groups and an
105        occurrence of (?| it is possible for an apparently non-recursive back
106        reference to become recursive if a later named group with the relevant
107        number is encountered. This could lead to a buffer overflow. Wen Guanxing
108        from Venustech ADLAB discovered this bug.
109    
110    28. If pcregrep was given the -q option with -c or -l, or when handling a
111        binary file, it incorrectly wrote output to stdout.
112    
113    29. The JIT compiler did not restore the control verb head in case of *THEN
114        control verbs. This issue was found by Karl Skomski with a custom LLVM
115        fuzzer.
116    
117    30. Error messages for syntax errors following \g and \k were giving inaccurate
118        offsets in the pattern.
119    
120    31. Added a check for integer overflow in conditions (?(<digits>) and
121        (?(R<digits>). This omission was discovered by Karl Skomski with the LLVM
122        fuzzer.
123    
124    32. Handling recursive references such as (?2) when the reference is to a group
125        later in the pattern uses code that is very hacked about and error-prone.
126        It has been re-written for PCRE2. Here in PCRE1, a check has been added to
127        give an internal error if it is obvious that compiling has gone wrong.
128    
129    33. The JIT compiler should not check repeats after a {0,1} repeat byte code.
130        This issue was found by Karl Skomski with a custom LLVM fuzzer.
131    
132    34. The JIT compiler should restore the control chain for empty possessive
133        repeats. This issue was found by Karl Skomski with a custom LLVM fuzzer.
134    
135    35. Match limit check added to JIT recursion. This issue was found by Karl
136        Skomski with a custom LLVM fuzzer.
137    
138    36. Yet another case similar to 27 above has been circumvented by an
139        unconditional allocation of extra memory. This issue is fixed "properly" in
140        PCRE2 by refactoring the way references are handled. Wen Guanxing
141        from Venustech ADLAB discovered this bug.
142    
143    37. Fix two assertion fails in JIT. These issues were found by Karl Skomski
144        with a custom LLVM fuzzer.
145    
146    38. Fixed a corner case of range optimization in JIT.
147    
148    39. An incorrect error "overran compiling workspace" was given if there were
149        exactly enough group forward references such that the last one extended
150        into the workspace safety margin. The next one would have expanded the
151        workspace. The test for overflow was not including the safety margin.
152    
153    40. A match limit issue is fixed in JIT which was found by Karl Skomski
154        with a custom LLVM fuzzer.
155    
156    
157    Version 8.37 28-April-2015
158    --------------------------
159    
160    1.  When an (*ACCEPT) is triggered inside capturing parentheses, it arranges
161        for those parentheses to be closed with whatever has been captured so far.
162        However, it was failing to mark any other groups between the hightest
163        capture so far and the currrent group as "unset". Thus, the ovector for
164        those groups contained whatever was previously there. An example is the
165        pattern /(x)|((*ACCEPT))/ when matched against "abcd".
166    
167    2.  If an assertion condition was quantified with a minimum of zero (an odd
168        thing to do, but it happened), SIGSEGV or other misbehaviour could occur.
169    
170    3.  If a pattern in pcretest input had the P (POSIX) modifier followed by an
171        unrecognized modifier, a crash could occur.
172    
173    4.  An attempt to do global matching in pcretest with a zero-length ovector
174        caused a crash.
175    
176    5.  Fixed a memory leak during matching that could occur for a subpattern
177        subroutine call (recursive or otherwise) if the number of captured groups
178        that had to be saved was greater than ten.
179    
180    6.  Catch a bad opcode during auto-possessification after compiling a bad UTF
181        string with NO_UTF_CHECK. This is a tidyup, not a bug fix, as passing bad
182        UTF with NO_UTF_CHECK is documented as having an undefined outcome.
183    
184    7.  A UTF pattern containing a "not" match of a non-ASCII character and a
185        subroutine reference could loop at compile time. Example: /[^\xff]((?1))/.
186    
187    8. When a pattern is compiled, it remembers the highest back reference so that
188       when matching, if the ovector is too small, extra memory can be obtained to
189       use instead. A conditional subpattern whose condition is a check on a
190       capture having happened, such as, for example in the pattern
191       /^(?:(a)|b)(?(1)A|B)/, is another kind of back reference, but it was not
192       setting the highest backreference number. This mattered only if pcre_exec()
193       was called with an ovector that was too small to hold the capture, and there
194       was no other kind of back reference (a situation which is probably quite
195       rare). The effect of the bug was that the condition was always treated as
196       FALSE when the capture could not be consulted, leading to a incorrect
197       behaviour by pcre_exec(). This bug has been fixed.
198    
199    9. A reference to a duplicated named group (either a back reference or a test
200       for being set in a conditional) that occurred in a part of the pattern where
201       PCRE_DUPNAMES was not set caused the amount of memory needed for the pattern
202       to be incorrectly calculated, leading to overwriting.
203    
204    10. A mutually recursive set of back references such as (\2)(\1) caused a
205        segfault at study time (while trying to find the minimum matching length).
206        The infinite loop is now broken (with the minimum length unset, that is,
207        zero).
208    
209    11. If an assertion that was used as a condition was quantified with a minimum
210        of zero, matching went wrong. In particular, if the whole group had
211        unlimited repetition and could match an empty string, a segfault was
212        likely. The pattern (?(?=0)?)+ is an example that caused this. Perl allows
213        assertions to be quantified, but not if they are being used as conditions,
214        so the above pattern is faulted by Perl. PCRE has now been changed so that
215        it also rejects such patterns.
216    
217    12. A possessive capturing group such as (a)*+ with a minimum repeat of zero
218        failed to allow the zero-repeat case if pcre2_exec() was called with an
219        ovector too small to capture the group.
220    
221    13. Fixed two bugs in pcretest that were discovered by fuzzing and reported by
222        Red Hat Product Security:
223    
224        (a) A crash if /K and /F were both set with the option to save the compiled
225        pattern.
226    
227        (b) Another crash if the option to print captured substrings in a callout
228        was combined with setting a null ovector, for example \O\C+ as a subject
229        string.
230    
231    14. A pattern such as "((?2){0,1999}())?", which has a group containing a
232        forward reference repeated a large (but limited) number of times within a
233        repeated outer group that has a zero minimum quantifier, caused incorrect
234        code to be compiled, leading to the error "internal error:
235        previously-checked referenced subpattern not found" when an incorrect
236        memory address was read. This bug was reported as "heap overflow",
237        discovered by Kai Lu of Fortinet's FortiGuard Labs and given the CVE number
238        CVE-2015-2325.
239    
240    23. A pattern such as "((?+1)(\1))/" containing a forward reference subroutine
241        call within a group that also contained a recursive back reference caused
242        incorrect code to be compiled. This bug was reported as "heap overflow",
243        discovered by Kai Lu of Fortinet's FortiGuard Labs, and given the CVE
244        number CVE-2015-2326.
245    
246    24. Computing the size of the JIT read-only data in advance has been a source
247        of various issues, and new ones are still appear unfortunately. To fix
248        existing and future issues, size computation is eliminated from the code,
249        and replaced by on-demand memory allocation.
250    
251    25. A pattern such as /(?i)[A-`]/, where characters in the other case are
252        adjacent to the end of the range, and the range contained characters with
253        more than one other case, caused incorrect behaviour when compiled in UTF
254        mode. In that example, the range a-j was left out of the class.
255    
256    26. Fix JIT compilation of conditional blocks, which assertion
257        is converted to (*FAIL). E.g: /(?(?!))/.
258    
259    27. The pattern /(?(?!)^)/ caused references to random memory. This bug was
260        discovered by the LLVM fuzzer.
261    
262    28. The assertion (?!) is optimized to (*FAIL). This was not handled correctly
263        when this assertion was used as a condition, for example (?(?!)a|b). In
264        pcre2_match() it worked by luck; in pcre2_dfa_match() it gave an incorrect
265        error about an unsupported item.
266    
267    29. For some types of pattern, for example /Z*(|d*){216}/, the auto-
268        possessification code could take exponential time to complete. A recursion
269        depth limit of 1000 has been imposed to limit the resources used by this
270        optimization.
271    
272    30. A pattern such as /(*UTF)[\S\V\H]/, which contains a negated special class
273        such as \S in non-UCP mode, explicit wide characters (> 255) can be ignored
274        because \S ensures they are all in the class. The code for doing this was
275        interacting badly with the code for computing the amount of space needed to
276        compile the pattern, leading to a buffer overflow. This bug was discovered
277        by the LLVM fuzzer.
278    
279    31. A pattern such as /((?2)+)((?1))/ which has mutual recursion nested inside
280        other kinds of group caused stack overflow at compile time. This bug was
281        discovered by the LLVM fuzzer.
282    
283    32. A pattern such as /(?1)(?#?'){8}(a)/ which had a parenthesized comment
284        between a subroutine call and its quantifier was incorrectly compiled,
285        leading to buffer overflow or other errors. This bug was discovered by the
286        LLVM fuzzer.
287    
288    33. The illegal pattern /(?(?<E>.*!.*)?)/ was not being diagnosed as missing an
289        assertion after (?(. The code was failing to check the character after
290        (?(?< for the ! or = that would indicate a lookbehind assertion. This bug
291        was discovered by the LLVM fuzzer.
292    
293    34. A pattern such as /X((?2)()*+){2}+/ which has a possessive quantifier with
294        a fixed maximum following a group that contains a subroutine reference was
295        incorrectly compiled and could trigger buffer overflow. This bug was
296        discovered by the LLVM fuzzer.
297    
298    35. A mutual recursion within a lookbehind assertion such as (?<=((?2))((?1)))
299        caused a stack overflow instead of the diagnosis of a non-fixed length
300        lookbehind assertion. This bug was discovered by the LLVM fuzzer.
301    
302    36. The use of \K in a positive lookbehind assertion in a non-anchored pattern
303        (e.g. /(?<=\Ka)/) could make pcregrep loop.
304    
305    37. There was a similar problem to 36 in pcretest for global matches.
306    
307    38. If a greedy quantified \X was preceded by \C in UTF mode (e.g. \C\X*),
308        and a subsequent item in the pattern caused a non-match, backtracking over
309        the repeated \X did not stop, but carried on past the start of the subject,
310        causing reference to random memory and/or a segfault. There were also some
311        other cases where backtracking after \C could crash. This set of bugs was
312        discovered by the LLVM fuzzer.
313    
314    39. The function for finding the minimum length of a matching string could take
315        a very long time if mutual recursion was present many times in a pattern,
316        for example, /((?2){73}(?2))((?1))/. A better mutual recursion detection
317        method has been implemented. This infelicity was discovered by the LLVM
318        fuzzer.
319    
320    40. Static linking against the PCRE library using the pkg-config module was
321        failing on missing pthread symbols.
322    
323    
324    Version 8.36 26-September-2014
325    ------------------------------
326    
327    1.  Got rid of some compiler warnings in the C++ modules that were shown up by
328        -Wmissing-field-initializers and -Wunused-parameter.
329    
330    2.  The tests for quantifiers being too big (greater than 65535) were being
331        applied after reading the number, and stupidly assuming that integer
332        overflow would give a negative number. The tests are now applied as the
333        numbers are read.
334    
335    3.  Tidy code in pcre_exec.c where two branches that used to be different are
336        now the same.
337    
338    4.  The JIT compiler did not generate match limit checks for certain
339        bracketed expressions with quantifiers. This may lead to exponential
340        backtracking, instead of returning with PCRE_ERROR_MATCHLIMIT. This
341        issue should be resolved now.
342    
343    5.  Fixed an issue, which occures when nested alternatives are optimized
344        with table jumps.
345    
346    6.  Inserted two casts and changed some ints to size_t in the light of some
347        reported 64-bit compiler warnings (Bugzilla 1477).
348    
349    7.  Fixed a bug concerned with zero-minimum possessive groups that could match
350        an empty string, which sometimes were behaving incorrectly in the
351        interpreter (though correctly in the JIT matcher). This pcretest input is
352        an example:
353    
354          '\A(?:[^"]++|"(?:[^"]*+|"")*+")++'
355          NON QUOTED "QUOT""ED" AFTER "NOT MATCHED
356    
357        the interpreter was reporting a match of 'NON QUOTED ' only, whereas the
358        JIT matcher and Perl both matched 'NON QUOTED "QUOT""ED" AFTER '. The test
359        for an empty string was breaking the inner loop and carrying on at a lower
360        level, when possessive repeated groups should always return to a higher
361        level as they have no backtrack points in them. The empty string test now
362        occurs at the outer level.
363    
364    8.  Fixed a bug that was incorrectly auto-possessifying \w+ in the pattern
365        ^\w+(?>\s*)(?<=\w) which caused it not to match "test test".
366    
367    9.  Give a compile-time error for \o{} (as Perl does) and for \x{} (which Perl
368        doesn't).
369    
370    10. Change 8.34/15 introduced a bug that caused the amount of memory needed
371        to hold a pattern to be incorrectly computed (too small) when there were
372        named back references to duplicated names. This could cause "internal
373        error: code overflow" or "double free or corruption" or other memory
374        handling errors.
375    
376    11. When named subpatterns had the same prefixes, back references could be
377        confused. For example, in this pattern:
378    
379          /(?P<Name>a)?(?P<Name2>b)?(?(<Name>)c|d)*l/
380    
381        the reference to 'Name' was incorrectly treated as a reference to a
382        duplicate name.
383    
384    12. A pattern such as /^s?c/mi8 where the optional character has more than
385        one "other case" was incorrectly compiled such that it would only try to
386        match starting at "c".
387    
388    13. When a pattern starting with \s was studied, VT was not included in the
389        list of possible starting characters; this should have been part of the
390        8.34/18 patch.
391    
392    14. If a character class started [\Qx]... where x is any character, the class
393        was incorrectly terminated at the ].
394    
395    15. If a pattern that started with a caseless match for a character with more
396        than one "other case" was studied, PCRE did not set up the starting code
397        unit bit map for the list of possible characters. Now it does. This is an
398        optimization improvement, not a bug fix.
399    
400    16. The Unicode data tables have been updated to Unicode 7.0.0.
401    
402    17. Fixed a number of memory leaks in pcregrep.
403    
404    18. Avoid a compiler warning (from some compilers) for a function call with
405        a cast that removes "const" from an lvalue by using an intermediate
406        variable (to which the compiler does not object).
407    
408    19. Incorrect code was compiled if a group that contained an internal recursive
409        back reference was optional (had quantifier with a minimum of zero). This
410        example compiled incorrect code: /(((a\2)|(a*)\g<-1>))*/ and other examples
411        caused segmentation faults because of stack overflows at compile time.
412    
413    20. A pattern such as /((?(R)a|(?1)))+/, which contains a recursion within a
414        group that is quantified with an indefinite repeat, caused a compile-time
415        loop which used up all the system stack and provoked a segmentation fault.
416        This was not the same bug as 19 above.
417    
418    21. Add PCRECPP_EXP_DECL declaration to operator<< in pcre_stringpiece.h.
419        Patch by Mike Frysinger.
420    
421    
422    Version 8.35 04-April-2014
423    --------------------------
424    
425    1.  A new flag is set, when property checks are present in an XCLASS.
426        When this flag is not set, PCRE can perform certain optimizations
427        such as studying these XCLASS-es.
428    
429    2.  The auto-possessification of character sets were improved: a normal
430        and an extended character set can be compared now. Furthermore
431        the JIT compiler optimizes more character set checks.
432    
433    3.  Got rid of some compiler warnings for potentially uninitialized variables
434        that show up only when compiled with -O2.
435    
436    4.  A pattern such as (?=ab\K) that uses \K in an assertion can set the start
437        of a match later then the end of the match. The pcretest program was not
438        handling the case sensibly - it was outputting from the start to the next
439        binary zero. It now reports this situation in a message, and outputs the
440        text from the end to the start.
441    
442    5.  Fast forward search is improved in JIT. Instead of the first three
443        characters, any three characters with fixed position can be searched.
444        Search order: first, last, middle.
445    
446    6.  Improve character range checks in JIT. Characters are read by an inprecise
447        function now, which returns with an unknown value if the character code is
448        above a certain threshold (e.g: 256). The only limitation is that the value
449        must be bigger than the threshold as well. This function is useful when
450        the characters above the threshold are handled in the same way.
451    
452    7.  The macros whose names start with RAWUCHAR are placeholders for a future
453        mode in which only the bottom 21 bits of 32-bit data items are used. To
454        make this more memorable for those maintaining the code, the names have
455        been changed to start with UCHAR21, and an extensive comment has been added
456        to their definition.
457    
458    8.  Add missing (new) files sljitNativeTILEGX.c and sljitNativeTILEGX-encoder.c
459        to the export list in Makefile.am (they were accidentally omitted from the
460        8.34 tarball).
461    
462    9.  The informational output from pcretest used the phrase "starting byte set"
463        which is inappropriate for the 16-bit and 32-bit libraries. As the output
464        for "first char" and "need char" really means "non-UTF-char", I've changed
465        "byte" to "char", and slightly reworded the output. The documentation about
466        these values has also been (I hope) clarified.
467    
468    10. Another JIT related optimization: use table jumps for selecting the correct
469        backtracking path, when more than four alternatives are present inside a
470        bracket.
471    
472    11. Empty match is not possible, when the minimum length is greater than zero,
473        and there is no \K in the pattern. JIT should avoid empty match checks in
474        such cases.
475    
476    12. In a caseless character class with UCP support, when a character with more
477        than one alternative case was not the first character of a range, not all
478        the alternative cases were added to the class. For example, s and \x{17f}
479        are both alternative cases for S: the class [RST] was handled correctly,
480        but [R-T] was not.
481    
482    13. The configure.ac file always checked for pthread support when JIT was
483        enabled. This is not used in Windows, so I have put this test inside a
484        check for the presence of windows.h (which was already tested for).
485    
486    14. Improve pattern prefix search by a simplified Boyer-Moore algorithm in JIT.
487        The algorithm provides a way to skip certain starting offsets, and usually
488        faster than linear prefix searches.
489    
490    15. Change 13 for 8.20 updated RunTest to check for the 'fr' locale as well
491        as for 'fr_FR' and 'french'. For some reason, however, it then used the
492        Windows-specific input and output files, which have 'french' screwed in.
493        So this could never have worked. One of the problems with locales is that
494        they aren't always the same. I have now updated RunTest so that it checks
495        the output of the locale test (test 3) against three different output
496        files, and it allows the test to pass if any one of them matches. With luck
497        this should make the test pass on some versions of Solaris where it was
498        failing. Because of the uncertainty, the script did not used to stop if
499        test 3 failed; it now does. If further versions of a French locale ever
500        come to light, they can now easily be added.
501    
502    16. If --with-pcregrep-bufsize was given a non-integer value such as "50K",
503        there was a message during ./configure, but it did not stop. This now
504        provokes an error. The invalid example in README has been corrected.
505        If a value less than the minimum is given, the minimum value has always
506        been used, but now a warning is given.
507    
508    17. If --enable-bsr-anycrlf was set, the special 16/32-bit test failed. This
509        was a bug in the test system, which is now fixed. Also, the list of various
510        configurations that are tested for each release did not have one with both
511        16/32 bits and --enable-bar-anycrlf. It now does.
512    
513    18. pcretest was missing "-C bsr" for displaying the \R default setting.
514    
515    19. Little endian PowerPC systems are supported now by the JIT compiler.
516    
517    20. The fast forward newline mechanism could enter to an infinite loop on
518        certain invalid UTF-8 input. Although we don't support these cases
519        this issue can be fixed by a performance optimization.
520    
521    21. Change 33 of 8.34 is not sufficient to ensure stack safety because it does
522        not take account if existing stack usage. There is now a new global
523        variable called pcre_stack_guard that can be set to point to an external
524        function to check stack availability. It is called at the start of
525        processing every parenthesized group.
526    
527    22. A typo in the code meant that in ungreedy mode the max/min qualifier
528        behaved like a min-possessive qualifier, and, for example, /a{1,3}b/U did
529        not match "ab".
530    
531    23. When UTF was disabled, the JIT program reported some incorrect compile
532        errors. These messages are silenced now.
533    
534    24. Experimental support for ARM-64 and MIPS-64 has been added to the JIT
535        compiler.
536    
537    25. Change all the temporary files used in RunGrepTest to be different to those
538        used by RunTest so that the tests can be run simultaneously, for example by
539        "make -j check".
540    
541    
542    Version 8.34 15-December-2013
543    -----------------------------
544    
545    1.  Add pcre[16|32]_jit_free_unused_memory to forcibly free unused JIT
546        executable memory. Patch inspired by Carsten Klein.
547    
548    2.  ./configure --enable-coverage defined SUPPORT_GCOV in config.h, although
549        this macro is never tested and has no effect, because the work to support
550        coverage involves only compiling and linking options and special targets in
551        the Makefile. The comment in config.h implied that defining the macro would
552        enable coverage support, which is totally false. There was also support for
553        setting this macro in the CMake files (my fault, I just copied it from
554        configure). SUPPORT_GCOV has now been removed.
555    
556    3.  Make a small performance improvement in strlen16() and strlen32() in
557        pcretest.
558    
559    4.  Change 36 for 8.33 left some unreachable statements in pcre_exec.c,
560        detected by the Solaris compiler (gcc doesn't seem to be able to diagnose
561        these cases). There was also one in pcretest.c.
562    
563    5.  Cleaned up a "may be uninitialized" compiler warning in pcre_exec.c.
564    
565    6.  In UTF mode, the code for checking whether a group could match an empty
566        string (which is used for indefinitely repeated groups to allow for
567        breaking an infinite loop) was broken when the group contained a repeated
568        negated single-character class with a character that occupied more than one
569        data item and had a minimum repetition of zero (for example, [^\x{100}]* in
570        UTF-8 mode). The effect was undefined: the group might or might not be
571        deemed as matching an empty string, or the program might have crashed.
572    
573    7.  The code for checking whether a group could match an empty string was not
574        recognizing that \h, \H, \v, \V, and \R must match a character.
575    
576    8.  Implemented PCRE_INFO_MATCH_EMPTY, which yields 1 if the pattern can match
577        an empty string. If it can, pcretest shows this in its information output.
578    
579    9.  Fixed two related bugs that applied to Unicode extended grapheme clusters
580        that were repeated with a maximizing qualifier (e.g. \X* or \X{2,5}) when
581        matched by pcre_exec() without using JIT:
582    
583        (a) If the rest of the pattern did not match after a maximal run of
584            grapheme clusters, the code for backing up to try with fewer of them
585            did not always back up over a full grapheme when characters that do not
586            have the modifier quality were involved, e.g. Hangul syllables.
587    
588        (b) If the match point in a subject started with modifier character, and
589            there was no match, the code could incorrectly back up beyond the match
590            point, and potentially beyond the first character in the subject,
591            leading to a segfault or an incorrect match result.
592    
593    10. A conditional group with an assertion condition could lead to PCRE
594        recording an incorrect first data item for a match if no other first data
595        item was recorded. For example, the pattern (?(?=ab)ab) recorded "a" as a
596        first data item, and therefore matched "ca" after "c" instead of at the
597        start.
598    
599    11. Change 40 for 8.33 (allowing pcregrep to find empty strings) showed up a
600        bug that caused the command "echo a | ./pcregrep -M '|a'" to loop.
601    
602    12. The source of pcregrep now includes z/OS-specific code so that it can be
603        compiled for z/OS as part of the special z/OS distribution.
604    
605    13. Added the -T and -TM options to pcretest.
606    
607    14. The code in pcre_compile.c for creating the table of named capturing groups
608        has been refactored. Instead of creating the table dynamically during the
609        actual compiling pass, the information is remembered during the pre-compile
610        pass (on the stack unless there are more than 20 named groups, in which
611        case malloc() is used) and the whole table is created before the actual
612        compile happens. This has simplified the code (it is now nearly 150 lines
613        shorter) and prepared the way for better handling of references to groups
614        with duplicate names.
615    
616    15. A back reference to a named subpattern when there is more than one of the
617        same name now checks them in the order in which they appear in the pattern.
618        The first one that is set is used for the reference. Previously only the
619        first one was inspected. This change makes PCRE more compatible with Perl.
620    
621    16. Unicode character properties were updated from Unicode 6.3.0.
622    
623    17. The compile-time code for auto-possessification has been refactored, based
624        on a patch by Zoltan Herczeg. It now happens after instead of during
625        compilation. The code is cleaner, and more cases are handled. The option
626        PCRE_NO_AUTO_POSSESS is added for testing purposes, and the -O and /O
627        options in pcretest are provided to set it. It can also be set by
628        (*NO_AUTO_POSSESS) at the start of a pattern.
629    
630    18. The character VT has been added to the default ("C" locale) set of
631        characters that match \s and are generally treated as white space,
632        following this same change in Perl 5.18. There is now no difference between
633        "Perl space" and "POSIX space". Whether VT is treated as white space in
634        other locales depends on the locale.
635    
636    19. The code for checking named groups as conditions, either for being set or
637        for being recursed, has been refactored (this is related to 14 and 15
638        above). Processing unduplicated named groups should now be as fast at
639        numerical groups, and processing duplicated groups should be faster than
640        before.
641    
642    20. Two patches to the CMake build system, by Alexander Barkov:
643    
644          (1) Replace the "source" command by "." in CMakeLists.txt because
645              "source" is a bash-ism.
646    
647          (2) Add missing HAVE_STDINT_H and HAVE_INTTYPES_H to config-cmake.h.in;
648              without these the CMake build does not work on Solaris.
649    
650    21. Perl has changed its handling of \8 and \9. If there is no previously
651        encountered capturing group of those numbers, they are treated as the
652        literal characters 8 and 9 instead of a binary zero followed by the
653        literals. PCRE now does the same.
654    
655    22. Following Perl, added \o{} to specify codepoints in octal, making it
656        possible to specify values greater than 0777 and also making them
657        unambiguous.
658    
659    23. Perl now gives an error for missing closing braces after \x{... instead of
660        treating the string as literal. PCRE now does the same.
661    
662    24. RunTest used to grumble if an inappropriate test was selected explicitly,
663        but just skip it when running all tests. This make it awkward to run ranges
664        of tests when one of them was inappropriate. Now it just skips any
665        inappropriate tests, as it always did when running all tests.
666    
667    25. If PCRE_AUTO_CALLOUT and PCRE_UCP were set for a pattern that contained
668        character types such as \d or \w, too many callouts were inserted, and the
669        data that they returned was rubbish.
670    
671    26. In UCP mode, \s was not matching two of the characters that Perl matches,
672        namely NEL (U+0085) and MONGOLIAN VOWEL SEPARATOR (U+180E), though they
673        were matched by \h. The code has now been refactored so that the lists of
674        the horizontal and vertical whitespace characters used for \h and \v (which
675        are defined only in one place) are now also used for \s.
676    
677    27. Add JIT support for the 64 bit TileGX architecture.
678        Patch by Jiong Wang (Tilera Corporation).
679    
680    28. Possessive quantifiers for classes (both explicit and automatically
681        generated) now use special opcodes instead of wrapping in ONCE brackets.
682    
683    29. Whereas an item such as A{4}+ ignored the possessivenes of the quantifier
684        (because it's meaningless), this was not happening when PCRE_CASELESS was
685        set. Not wrong, but inefficient.
686    
687    30. Updated perltest.pl to add /u (force Unicode mode) when /W (use Unicode
688        properties for \w, \d, etc) is present in a test regex. Otherwise if the
689        test contains no characters greater than 255, Perl doesn't realise it
690        should be using Unicode semantics.
691    
692    31. Upgraded the handling of the POSIX classes [:graph:], [:print:], and
693        [:punct:] when PCRE_UCP is set so as to include the same characters as Perl
694        does in Unicode mode.
695    
696    32. Added the "forbid" facility to pcretest so that putting tests into the
697        wrong test files can sometimes be quickly detected.
698    
699    33. There is now a limit (default 250) on the depth of nesting of parentheses.
700        This limit is imposed to control the amount of system stack used at compile
701        time. It can be changed at build time by --with-parens-nest-limit=xxx or
702        the equivalent in CMake.
703    
704    34. Character classes such as [A-\d] or [a-[:digit:]] now cause compile-time
705        errors. Perl warns for these when in warning mode, but PCRE has no facility
706        for giving warnings.
707    
708    35. Change 34 for 8.13 allowed quantifiers on assertions, because Perl does.
709        However, this was not working for (?!) because it is optimized to (*FAIL),
710        for which PCRE does not allow quantifiers. The optimization is now disabled
711        when a quantifier follows (?!). I can't see any use for this, but it makes
712        things uniform.
713    
714    36. Perl no longer allows group names to start with digits, so I have made this
715        change also in PCRE. It simplifies the code a bit.
716    
717    37. In extended mode, Perl ignores spaces before a + that indicates a
718        possessive quantifier. PCRE allowed a space before the quantifier, but not
719        before the possessive +. It now does.
720    
721    38. The use of \K (reset reported match start) within a repeated possessive
722        group such as (a\Kb)*+ was not working.
723    
724    40. Document that the same character tables must be used at compile time and
725        run time, and that the facility to pass tables to pcre_exec() and
726        pcre_dfa_exec() is for use only with saved/restored patterns.
727    
728    41. Applied Jeff Trawick's patch CMakeLists.txt, which "provides two new
729        features for Builds with MSVC:
730    
731        1. Support pcre.rc and/or pcreposix.rc (as is already done for MinGW
732           builds). The .rc files can be used to set FileDescription and many other
733           attributes.
734    
735        2. Add an option (-DINSTALL_MSVC_PDB) to enable installation of .pdb files.
736           This allows higher-level build scripts which want .pdb files to avoid
737           hard-coding the exact files needed."
738    
739    42. Added support for [[:<:]] and [[:>:]] as used in the BSD POSIX library to
740        mean "start of word" and "end of word", respectively, as a transition aid.
741    
742    43. A minimizing repeat of a class containing codepoints greater than 255 in
743        non-UTF 16-bit or 32-bit modes caused an internal error when PCRE was
744        compiled to use the heap for recursion.
745    
746    44. Got rid of some compiler warnings for unused variables when UTF but not UCP
747        is configured.
748    
749    
750    Version 8.33 28-May-2013
751    ------------------------
752    
753    1.  Added 'U' to some constants that are compared to unsigned integers, to
754        avoid compiler signed/unsigned warnings. Added (int) casts to unsigned
755        variables that are added to signed variables, to ensure the result is
756        signed and can be negated.
757    
758    2.  Applied patch by Daniel Richard G for quashing MSVC warnings to the
759        CMake config files.
760    
761    3.  Revise the creation of config.h.generic so that all boolean macros are
762        #undefined, whereas non-boolean macros are #ifndef/#endif-ed. This makes
763        overriding via -D on the command line possible.
764    
765    4.  Changing the definition of the variable "op" in pcre_exec.c from pcre_uchar
766        to unsigned int is reported to make a quite noticeable speed difference in
767        a specific Windows environment. Testing on Linux did also appear to show
768        some benefit (and it is clearly not harmful). Also fixed the definition of
769        Xop which should be unsigned.
770    
771    5.  Related to (4), changing the definition of the intermediate variable cc
772        in repeated character loops from pcre_uchar to pcre_uint32 also gave speed
773        improvements.
774    
775    6.  Fix forward search in JIT when link size is 3 or greater. Also removed some
776        unnecessary spaces.
777    
778    7.  Adjust autogen.sh and configure.ac to lose warnings given by automake 1.12
779        and later.
780    
781    8.  Fix two buffer over read issues in 16 and 32 bit modes. Affects JIT only.
782    
783    9.  Optimizing fast_forward_start_bits in JIT.
784    
785    10. Adding support for callouts in JIT, and fixing some issues revealed
786        during this work. Namely:
787    
788        (a) Unoptimized capturing brackets incorrectly reset on backtrack.
789    
790        (b) Minimum length was not checked before the matching is started.
791    
792    11. The value of capture_last that is passed to callouts was incorrect in some
793        cases when there was a capture on one path that was subsequently abandoned
794        after a backtrack. Also, the capture_last value is now reset after a
795        recursion, since all captures are also reset in this case.
796    
797    12. The interpreter no longer returns the "too many substrings" error in the
798        case when an overflowing capture is in a branch that is subsequently
799        abandoned after a backtrack.
800    
801    13. In the pathological case when an offset vector of size 2 is used, pcretest
802        now prints out the matched string after a yield of 0 or 1.
803    
804    14. Inlining subpatterns in recursions, when certain conditions are fulfilled.
805        Only supported by the JIT compiler at the moment.
806    
807    15. JIT compiler now supports 32 bit Macs thanks to Lawrence Velazquez.
808    
809    16. Partial matches now set offsets[2] to the "bumpalong" value, that is, the
810        offset of the starting point of the matching process, provided the offsets
811        vector is large enough.
812    
813    17. The \A escape now records a lookbehind value of 1, though its execution
814        does not actually inspect the previous character. This is to ensure that,
815        in partial multi-segment matching, at least one character from the old
816        segment is retained when a new segment is processed. Otherwise, if there
817        are no lookbehinds in the pattern, \A might match incorrectly at the start
818        of a new segment.
819    
820    18. Added some #ifdef __VMS code into pcretest.c to help VMS implementations.
821    
822    19. Redefined some pcre_uchar variables in pcre_exec.c as pcre_uint32; this
823        gives some modest performance improvement in 8-bit mode.
824    
825    20. Added the PCRE-specific property \p{Xuc} for matching characters that can
826        be expressed in certain programming languages using Universal Character
827        Names.
828    
829    21. Unicode validation has been updated in the light of Unicode Corrigendum #9,
830        which points out that "non characters" are not "characters that may not
831        appear in Unicode strings" but rather "characters that are reserved for
832        internal use and have only local meaning".
833    
834    22. When a pattern was compiled with automatic callouts (PCRE_AUTO_CALLOUT) and
835        there was a conditional group that depended on an assertion, if the
836        assertion was false, the callout that immediately followed the alternation
837        in the condition was skipped when pcre_exec() was used for matching.
838    
839    23. Allow an explicit callout to be inserted before an assertion that is the
840        condition for a conditional group, for compatibility with automatic
841        callouts, which always insert a callout at this point.
842    
843    24. In 8.31, (*COMMIT) was confined to within a recursive subpattern. Perl also
844        confines (*SKIP) and (*PRUNE) in the same way, and this has now been done.
845    
846    25. (*PRUNE) is now supported by the JIT compiler.
847    
848    26. Fix infinite loop when /(?<=(*SKIP)ac)a/ is matched against aa.
849    
850    27. Fix the case where there are two or more SKIPs with arguments that may be
851        ignored.
852    
853    28. (*SKIP) is now supported by the JIT compiler.
854    
855    29. (*THEN) is now supported by the JIT compiler.
856    
857    30. Update RunTest with additional test selector options.
858    
859    31. The way PCRE handles backtracking verbs has been changed in two ways.
860    
861        (1) Previously, in something like (*COMMIT)(*SKIP), COMMIT would override
862        SKIP. Now, PCRE acts on whichever backtracking verb is reached first by
863        backtracking. In some cases this makes it more Perl-compatible, but Perl's
864        rather obscure rules do not always do the same thing.
865    
866        (2) Previously, backtracking verbs were confined within assertions. This is
867        no longer the case for positive assertions, except for (*ACCEPT). Again,
868        this sometimes improves Perl compatibility, and sometimes does not.
869    
870    32. A number of tests that were in test 2 because Perl did things differently
871        have been moved to test 1, because either Perl or PCRE has changed, and
872        these tests are now compatible.
873    
874    32. Backtracking control verbs are now handled in the same way in JIT and
875        interpreter.
876    
877    33. An opening parenthesis in a MARK/PRUNE/SKIP/THEN name in a pattern that
878        contained a forward subroutine reference caused a compile error.
879    
880    34. Auto-detect and optimize limited repetitions in JIT.
881    
882    35. Implement PCRE_NEVER_UTF to lock out the use of UTF, in particular,
883        blocking (*UTF) etc.
884    
885    36. In the interpreter, maximizing pattern repetitions for characters and
886        character types now use tail recursion, which reduces stack usage.
887    
888    37. The value of the max lookbehind was not correctly preserved if a compiled
889        and saved regex was reloaded on a host of different endianness.
890    
891    38. Implemented (*LIMIT_MATCH) and (*LIMIT_RECURSION). As part of the extension
892        of the compiled pattern block, expand the flags field from 16 to 32 bits
893        because it was almost full.
894    
895    39. Try madvise first before posix_madvise.
896    
897    40. Change 7 for PCRE 7.9 made it impossible for pcregrep to find empty lines
898        with a pattern such as ^$. It has taken 4 years for anybody to notice! The
899        original change locked out all matches of empty strings. This has been
900        changed so that one match of an empty string per line is recognized.
901        Subsequent searches on the same line (for colouring or for --only-matching,
902        for example) do not recognize empty strings.
903    
904    41. Applied a user patch to fix a number of spelling mistakes in comments.
905    
906    42. Data lines longer than 65536 caused pcretest to crash.
907    
908    43. Clarified the data type for length and startoffset arguments for pcre_exec
909        and pcre_dfa_exec in the function-specific man pages, where they were
910        explicitly stated to be in bytes, never having been updated. I also added
911        some clarification to the pcreapi man page.
912    
913    44. A call to pcre_dfa_exec() with an output vector size less than 2 caused
914        a segmentation fault.
915    
916    
917    Version 8.32 30-November-2012
918    -----------------------------
919    
920    1.  Improved JIT compiler optimizations for first character search and single
921        character iterators.
922    
923    2.  Supporting IBM XL C compilers for PPC architectures in the JIT compiler.
924        Patch by Daniel Richard G.
925    
926    3.  Single character iterator optimizations in the JIT compiler.
927    
928    4.  Improved JIT compiler optimizations for character ranges.
929    
930    5.  Rename the "leave" variable names to "quit" to improve WinCE compatibility.
931        Reported by Giuseppe D'Angelo.
932    
933    6.  The PCRE_STARTLINE bit, indicating that a match can occur only at the start
934        of a line, was being set incorrectly in cases where .* appeared inside
935        atomic brackets at the start of a pattern, or where there was a subsequent
936        *PRUNE or *SKIP.
937    
938    7.  Improved instruction cache flush for POWER/PowerPC.
939        Patch by Daniel Richard G.
940    
941    8.  Fixed a number of issues in pcregrep, making it more compatible with GNU
942        grep:
943    
944        (a) There is now no limit to the number of patterns to be matched.
945    
946        (b) An error is given if a pattern is too long.
947    
948        (c) Multiple uses of --exclude, --exclude-dir, --include, and --include-dir
949            are now supported.
950    
951        (d) --exclude-from and --include-from (multiple use) have been added.
952    
953        (e) Exclusions and inclusions now apply to all files and directories, not
954            just to those obtained from scanning a directory recursively.
955    
956        (f) Multiple uses of -f and --file-list are now supported.
957    
958        (g) In a Windows environment, the default for -d has been changed from
959            "read" (the GNU grep default) to "skip", because otherwise the presence
960            of a directory in the file list provokes an error.
961    
962        (h) The documentation has been revised and clarified in places.
963    
964    9.  Improve the matching speed of capturing brackets.
965    
966    10. Changed the meaning of \X so that it now matches a Unicode extended
967        grapheme cluster.
968    
969    11. Patch by Daniel Richard G to the autoconf files to add a macro for sorting
970        out POSIX threads when JIT support is configured.
971    
972    12. Added support for PCRE_STUDY_EXTRA_NEEDED.
973    
974    13. In the POSIX wrapper regcomp() function, setting re_nsub field in the preg
975        structure could go wrong in environments where size_t is not the same size
976        as int.
977    
978    14. Applied user-supplied patch to pcrecpp.cc to allow PCRE_NO_UTF8_CHECK to be
979        set.
980    
981    15. The EBCDIC support had decayed; later updates to the code had included
982        explicit references to (e.g.) \x0a instead of CHAR_LF. There has been a
983        general tidy up of EBCDIC-related issues, and the documentation was also
984        not quite right. There is now a test that can be run on ASCII systems to
985        check some of the EBCDIC-related things (but is it not a full test).
986    
987    16. The new PCRE_STUDY_EXTRA_NEEDED option is now used by pcregrep, resulting
988        in a small tidy to the code.
989    
990    17. Fix JIT tests when UTF is disabled and both 8 and 16 bit mode are enabled.
991    
992    18. If the --only-matching (-o) option in pcregrep is specified multiple
993        times, each one causes appropriate output. For example, -o1 -o2 outputs the
994        substrings matched by the 1st and 2nd capturing parentheses. A separating
995        string can be specified by --om-separator (default empty).
996    
997    19. Improving the first n character searches.
998    
999    20. Turn case lists for horizontal and vertical white space into macros so that
1000        they are defined only once.
1001    
1002    21. This set of changes together give more compatible Unicode case-folding
1003        behaviour for characters that have more than one other case when UCP
1004        support is available.
1005    
1006        (a) The Unicode property table now has offsets into a new table of sets of
1007            three or more characters that are case-equivalent. The MultiStage2.py
1008            script that generates these tables (the pcre_ucd.c file) now scans
1009            CaseFolding.txt instead of UnicodeData.txt for character case
1010            information.
1011    
1012        (b) The code for adding characters or ranges of characters to a character
1013            class has been abstracted into a generalized function that also handles
1014            case-independence. In UTF-mode with UCP support, this uses the new data
1015            to handle characters with more than one other case.
1016    
1017        (c) A bug that is fixed as a result of (b) is that codepoints less than 256
1018            whose other case is greater than 256 are now correctly matched
1019            caselessly. Previously, the high codepoint matched the low one, but not
1020            vice versa.
1021    
1022        (d) The processing of \h, \H, \v, and \ in character classes now makes use
1023            of the new class addition function, using character lists defined as
1024            macros alongside the case definitions of 20 above.
1025    
1026        (e) Caseless back references now work with characters that have more than
1027            one other case.
1028    
1029        (f) General caseless matching of characters with more than one other case
1030            is supported.
1031    
1032    22. Unicode character properties were updated from Unicode 6.2.0
1033    
1034    23. Improved CMake support under Windows. Patch by Daniel Richard G.
1035    
1036    24. Add support for 32-bit character strings, and UTF-32
1037    
1038    25. Major JIT compiler update (code refactoring and bugfixing).
1039        Experimental Sparc 32 support is added.
1040    
1041    26. Applied a modified version of Daniel Richard G's patch to create
1042        pcre.h.generic and config.h.generic by "make" instead of in the
1043        PrepareRelease script.
1044    
1045    27. Added a definition for CHAR_NULL (helpful for the z/OS port), and use it in
1046        pcre_compile.c when checking for a zero character.
1047    
1048    28. Introducing a native interface for JIT. Through this interface, the compiled
1049        machine code can be directly executed. The purpose of this interface is to
1050        provide fast pattern matching, so several sanity checks are not performed.
1051        However, feature tests are still performed. The new interface provides
1052        1.4x speedup compared to the old one.
1053    
1054    29. If pcre_exec() or pcre_dfa_exec() was called with a negative value for
1055        the subject string length, the error given was PCRE_ERROR_BADOFFSET, which
1056        was confusing. There is now a new error PCRE_ERROR_BADLENGTH for this case.
1057    
1058    30. In 8-bit UTF-8 mode, pcretest failed to give an error for data codepoints
1059        greater than 0x7fffffff (which cannot be represented in UTF-8, even under
1060        the "old" RFC 2279). Instead, it ended up passing a negative length to
1061        pcre_exec().
1062    
1063    31. Add support for GCC's visibility feature to hide internal functions.
1064    
1065    32. Running "pcretest -C pcre8" or "pcretest -C pcre16" gave a spurious error
1066        "unknown -C option" after outputting 0 or 1.
1067    
1068    33. There is now support for generating a code coverage report for the test
1069        suite in environments where gcc is the compiler and lcov is installed. This
1070        is mainly for the benefit of the developers.
1071    
1072    34. If PCRE is built with --enable-valgrind, certain memory regions are marked
1073        unaddressable using valgrind annotations, allowing valgrind to detect
1074        invalid memory accesses. This is mainly for the benefit of the developers.
1075    
1076    25. (*UTF) can now be used to start a pattern in any of the three libraries.
1077    
1078    26. Give configure error if --enable-cpp but no C++ compiler found.
1079    
1080    
1081    Version 8.31 06-July-2012
1082    -------------------------
1083    
1084    1.  Fixing a wrong JIT test case and some compiler warnings.
1085    
1086    2.  Removed a bashism from the RunTest script.
1087    
1088    3.  Add a cast to pcre_exec.c to fix the warning "unary minus operator applied
1089        to unsigned type, result still unsigned" that was given by an MS compiler
1090        on encountering the code "-sizeof(xxx)".
1091    
1092    4.  Partial matching support is added to the JIT compiler.
1093    
1094    5.  Fixed several bugs concerned with partial matching of items that consist
1095        of more than one character:
1096    
1097        (a) /^(..)\1/ did not partially match "aba" because checking references was
1098            done on an "all or nothing" basis. This also applied to repeated
1099            references.
1100    
1101        (b) \R did not give a hard partial match if \r was found at the end of the
1102            subject.
1103    
1104        (c) \X did not give a hard partial match after matching one or more
1105            characters at the end of the subject.
1106    
1107        (d) When newline was set to CRLF, a pattern such as /a$/ did not recognize
1108            a partial match for the string "\r".
1109    
1110        (e) When newline was set to CRLF, the metacharacter "." did not recognize
1111            a partial match for a CR character at the end of the subject string.
1112    
1113    6.  If JIT is requested using /S++ or -s++ (instead of just /S+ or -s+) when
1114        running pcretest, the text "(JIT)" added to the output whenever JIT is
1115        actually used to run the match.
1116    
1117    7.  Individual JIT compile options can be set in pcretest by following -s+[+]
1118        or /S+[+] with a digit between 1 and 7.
1119    
1120    8.  OP_NOT now supports any UTF character not just single-byte ones.
1121    
1122    9.  (*MARK) control verb is now supported by the JIT compiler.
1123    
1124    10. The command "./RunTest list" lists the available tests without actually
1125        running any of them. (Because I keep forgetting what they all are.)
1126    
1127    11. Add PCRE_INFO_MAXLOOKBEHIND.
1128    
1129    12. Applied a (slightly modified) user-supplied patch that improves performance
1130        when the heap is used for recursion (compiled with --disable-stack-for-
1131        recursion). Instead of malloc and free for each heap frame each time a
1132        logical recursion happens, frames are retained on a chain and re-used where
1133        possible. This sometimes gives as much as 30% improvement.
1134    
1135    13. As documented, (*COMMIT) is now confined to within a recursive subpattern
1136        call.
1137    
1138    14. As documented, (*COMMIT) is now confined to within a positive assertion.
1139    
1140    15. It is now possible to link pcretest with libedit as an alternative to
1141        libreadline.
1142    
1143    16. (*COMMIT) control verb is now supported by the JIT compiler.
1144    
1145    17. The Unicode data tables have been updated to Unicode 6.1.0.
1146    
1147    18. Added --file-list option to pcregrep.
1148    
1149    19. Added binary file support to pcregrep, including the -a, --binary-files,
1150        -I, and --text options.
1151    
1152    20. The madvise function is renamed for posix_madvise for QNX compatibility
1153        reasons. Fixed by Giuseppe D'Angelo.
1154    
1155    21. Fixed a bug for backward assertions with REVERSE 0 in the JIT compiler.
1156    
1157    22. Changed the option for creating symbolic links for 16-bit man pages from
1158        -s to -sf so that re-installing does not cause issues.
1159    
1160    23. Support PCRE_NO_START_OPTIMIZE in JIT as (*MARK) support requires it.
1161    
1162    24. Fixed a very old bug in pcretest that caused errors with restarted DFA
1163        matches in certain environments (the workspace was not being correctly
1164        retained). Also added to pcre_dfa_exec() a simple plausibility check on
1165        some of the workspace data at the beginning of a restart.
1166    
1167    25. \s*\R was auto-possessifying the \s* when it should not, whereas \S*\R
1168        was not doing so when it should - probably a typo introduced by SVN 528
1169        (change 8.10/14).
1170    
1171    26. When PCRE_UCP was not set, \w+\x{c4} was incorrectly auto-possessifying the
1172        \w+ when the character tables indicated that \x{c4} was a word character.
1173        There were several related cases, all because the tests for doing a table
1174        lookup were testing for characters less than 127 instead of 255.
1175    
1176    27. If a pattern contains capturing parentheses that are not used in a match,
1177        their slots in the ovector are set to -1. For those that are higher than
1178        any matched groups, this happens at the end of processing. In the case when
1179        there were back references that the ovector was too small to contain
1180        (causing temporary malloc'd memory to be used during matching), and the
1181        highest capturing number was not used, memory off the end of the ovector
1182        was incorrectly being set to -1. (It was using the size of the temporary
1183        memory instead of the true size.)
1184    
1185    28. To catch bugs like 27 using valgrind, when pcretest is asked to specify an
1186        ovector size, it uses memory at the end of the block that it has got.
1187    
1188    29. Check for an overlong MARK name and give an error at compile time. The
1189        limit is 255 for the 8-bit library and 65535 for the 16-bit library.
1190    
1191    30. JIT compiler update.
1192    
1193    31. JIT is now supported on jailbroken iOS devices. Thanks for Ruiger
1194        Rill for the patch.
1195    
1196    32. Put spaces around SLJIT_PRINT_D in the JIT compiler. Required by CXX11.
1197    
1198    33. Variable renamings in the PCRE-JIT compiler. No functionality change.
1199    
1200    34. Fixed typos in pcregrep: in two places there was SUPPORT_LIBZ2 instead of
1201        SUPPORT_LIBBZ2. This caused a build problem when bzip2 but not gzip (zlib)
1202        was enabled.
1203    
1204    35. Improve JIT code generation for greedy plus quantifier.
1205    
1206    36. When /((?:a?)*)*c/ or /((?>a?)*)*c/ was matched against "aac", it set group
1207        1 to "aa" instead of to an empty string. The bug affected repeated groups
1208        that could potentially match an empty string.
1209    
1210    37. Optimizing single character iterators in JIT.
1211    
1212    38. Wide characters specified with \uxxxx in JavaScript mode are now subject to
1213        the same checks as \x{...} characters in non-JavaScript mode. Specifically,
1214        codepoints that are too big for the mode are faulted, and in a UTF mode,
1215        disallowed codepoints are also faulted.
1216    
1217    39. If PCRE was compiled with UTF support, in three places in the DFA
1218        matcher there was code that should only have been obeyed in UTF mode, but
1219        was being obeyed unconditionally. In 8-bit mode this could cause incorrect
1220        processing when bytes with values greater than 127 were present. In 16-bit
1221        mode the bug would be provoked by values in the range 0xfc00 to 0xdc00. In
1222        both cases the values are those that cannot be the first data item in a UTF
1223        character. The three items that might have provoked this were recursions,
1224        possessively repeated groups, and atomic groups.
1225    
1226    40. Ensure that libpcre is explicitly listed in the link commands for pcretest
1227        and pcregrep, because some OS require shared objects to be explicitly
1228        passed to ld, causing the link step to fail if they are not.
1229    
1230    41. There were two incorrect #ifdefs in pcre_study.c, meaning that, in 16-bit
1231        mode, patterns that started with \h* or \R* might be incorrectly matched.
1232    
1233    
1234    Version 8.30 04-February-2012
1235    -----------------------------
1236    
1237    1.  Renamed "isnumber" as "is_a_number" because in some Mac environments this
1238        name is defined in ctype.h.
1239    
1240    2.  Fixed a bug in fixed-length calculation for lookbehinds that would show up
1241        only in quite long subpatterns.
1242    
1243    3.  Removed the function pcre_info(), which has been obsolete and deprecated
1244        since it was replaced by pcre_fullinfo() in February 2000.
1245    
1246    4.  For a non-anchored pattern, if (*SKIP) was given with a name that did not
1247        match a (*MARK), and the match failed at the start of the subject, a
1248        reference to memory before the start of the subject could occur. This bug
1249        was introduced by fix 17 of release 8.21.
1250    
1251    5.  A reference to an unset group with zero minimum repetition was giving
1252        totally wrong answers (in non-JavaScript-compatibility mode). For example,
1253        /(another)?(\1?)test/ matched against "hello world test". This bug was
1254        introduced in release 8.13.
1255    
1256    6.  Add support for 16-bit character strings (a large amount of work involving
1257        many changes and refactorings).
1258    
1259    7.  RunGrepTest failed on msys because \r\n was replaced by whitespace when the
1260        command "pattern=`printf 'xxx\r\njkl'`" was run. The pattern is now taken
1261        from a file.
1262    
1263    8.  Ovector size of 2 is also supported by JIT based pcre_exec (the ovector size
1264        rounding is not applied in this particular case).
1265    
1266    9.  The invalid Unicode surrogate codepoints U+D800 to U+DFFF are now rejected
1267        if they appear, or are escaped, in patterns.
1268    
1269    10. Get rid of a number of -Wunused-but-set-variable warnings.
1270    
1271    11. The pattern /(?=(*:x))(q|)/ matches an empty string, and returns the mark
1272        "x". The similar pattern /(?=(*:x))((*:y)q|)/ did not return a mark at all.
1273        Oddly, Perl behaves the same way. PCRE has been fixed so that this pattern
1274        also returns the mark "x". This bug applied to capturing parentheses,
1275        non-capturing parentheses, and atomic parentheses. It also applied to some
1276        assertions.
1277    
1278    12. Stephen Kelly's patch to CMakeLists.txt allows it to parse the version
1279        information out of configure.ac instead of relying on pcre.h.generic, which
1280        is not stored in the repository.
1281    
1282    13. Applied Dmitry V. Levin's patch for a more portable method for linking with
1283        -lreadline.
1284    
1285    14. ZH added PCRE_CONFIG_JITTARGET; added its output to pcretest -C.
1286    
1287    15. Applied Graycode's patch to put the top-level frame on the stack rather
1288        than the heap when not using the stack for recursion. This gives a
1289        performance improvement in many cases when recursion is not deep.
1290    
1291    16. Experimental code added to "pcretest -C" to output the stack frame size.
1292    
1293    
1294    Version 8.21 12-Dec-2011
1295    ------------------------
1296    
1297    1.  Updating the JIT compiler.
1298    
1299    2.  JIT compiler now supports OP_NCREF, OP_RREF and OP_NRREF. New test cases
1300        are added as well.
1301    
1302    3.  Fix cache-flush issue on PowerPC (It is still an experimental JIT port).
1303        PCRE_EXTRA_TABLES is not suported by JIT, and should be checked before
1304        calling _pcre_jit_exec. Some extra comments are added.
1305    
1306    4.  (*MARK) settings inside atomic groups that do not contain any capturing
1307        parentheses, for example, (?>a(*:m)), were not being passed out. This bug
1308        was introduced by change 18 for 8.20.
1309    
1310    5.  Supporting of \x, \U and \u in JavaScript compatibility mode based on the
1311        ECMA-262 standard.
1312    
1313    6.  Lookbehinds such as (?<=a{2}b) that contained a fixed repetition were
1314        erroneously being rejected as "not fixed length" if PCRE_CASELESS was set.
1315        This bug was probably introduced by change 9 of 8.13.
1316    
1317    7.  While fixing 6 above, I noticed that a number of other items were being
1318        incorrectly rejected as "not fixed length". This arose partly because newer
1319        opcodes had not been added to the fixed-length checking code. I have (a)
1320        corrected the bug and added tests for these items, and (b) arranged for an
1321        error to occur if an unknown opcode is encountered while checking for fixed
1322        length instead of just assuming "not fixed length". The items that were
1323        rejected were: (*ACCEPT), (*COMMIT), (*FAIL), (*MARK), (*PRUNE), (*SKIP),
1324        (*THEN), \h, \H, \v, \V, and single character negative classes with fixed
1325        repetitions, e.g. [^a]{3}, with and without PCRE_CASELESS.
1326    
1327    8.  A possessively repeated conditional subpattern such as (?(?=c)c|d)++ was
1328        being incorrectly compiled and would have given unpredicatble results.
1329    
1330    9.  A possessively repeated subpattern with minimum repeat count greater than
1331        one behaved incorrectly. For example, (A){2,}+ behaved as if it was
1332        (A)(A)++ which meant that, after a subsequent mismatch, backtracking into
1333        the first (A) could occur when it should not.
1334    
1335    10. Add a cast and remove a redundant test from the code.
1336    
1337    11. JIT should use pcre_malloc/pcre_free for allocation.
1338    
1339    12. Updated pcre-config so that it no longer shows -L/usr/lib, which seems
1340        best practice nowadays, and helps with cross-compiling. (If the exec_prefix
1341        is anything other than /usr, -L is still shown).
1342    
1343    13. In non-UTF-8 mode, \C is now supported in lookbehinds and DFA matching.
1344    
1345    14. Perl does not support \N without a following name in a [] class; PCRE now
1346        also gives an error.
1347    
1348    15. If a forward reference was repeated with an upper limit of around 2000,
1349        it caused the error "internal error: overran compiling workspace". The
1350        maximum number of forward references (including repeats) was limited by the
1351        internal workspace, and dependent on the LINK_SIZE. The code has been
1352        rewritten so that the workspace expands (via pcre_malloc) if necessary, and
1353        the default depends on LINK_SIZE. There is a new upper limit (for safety)
1354        of around 200,000 forward references. While doing this, I also speeded up
1355        the filling in of repeated forward references.
1356    
1357    16. A repeated forward reference in a pattern such as (a)(?2){2}(.) was
1358        incorrectly expecting the subject to contain another "a" after the start.
1359    
1360    17. When (*SKIP:name) is activated without a corresponding (*MARK:name) earlier
1361        in the match, the SKIP should be ignored. This was not happening; instead
1362        the SKIP was being treated as NOMATCH. For patterns such as
1363        /A(*MARK:A)A+(*SKIP:B)Z|AAC/ this meant that the AAC branch was never
1364        tested.
1365    
1366    18. The behaviour of (*MARK), (*PRUNE), and (*THEN) has been reworked and is
1367        now much more compatible with Perl, in particular in cases where the result
1368        is a non-match for a non-anchored pattern. For example, if
1369        /b(*:m)f|a(*:n)w/ is matched against "abc", the non-match returns the name
1370        "m", where previously it did not return a name. A side effect of this
1371        change is that for partial matches, the last encountered mark name is
1372        returned, as for non matches. A number of tests that were previously not
1373        Perl-compatible have been moved into the Perl-compatible test files. The
1374        refactoring has had the pleasing side effect of removing one argument from
1375        the match() function, thus reducing its stack requirements.
1376    
1377    19. If the /S+ option was used in pcretest to study a pattern using JIT,
1378        subsequent uses of /S (without +) incorrectly behaved like /S+.
1379    
1380    21. Retrieve executable code size support for the JIT compiler and fixing
1381        some warnings.
1382    
1383    22. A caseless match of a UTF-8 character whose other case uses fewer bytes did
1384        not work when the shorter character appeared right at the end of the
1385        subject string.
1386    
1387    23. Added some (int) casts to non-JIT modules to reduce warnings on 64-bit
1388        systems.
1389    
1390    24. Added PCRE_INFO_JITSIZE to pass on the value from (21) above, and also
1391        output it when the /M option is used in pcretest.
1392    
1393    25. The CheckMan script was not being included in the distribution. Also, added
1394        an explicit "perl" to run Perl scripts from the PrepareRelease script
1395        because this is reportedly needed in Windows.
1396    
1397    26. If study data was being save in a file and studying had not found a set of
1398        "starts with" bytes for the pattern, the data written to the file (though
1399        never used) was taken from uninitialized memory and so caused valgrind to
1400        complain.
1401    
1402    27. Updated RunTest.bat as provided by Sheri Pierce.
1403    
1404    28. Fixed a possible uninitialized memory bug in pcre_jit_compile.c.
1405    
1406    29. Computation of memory usage for the table of capturing group names was
1407        giving an unnecessarily large value.
1408    
1409    
1410    Version 8.20 21-Oct-2011
1411    ------------------------
1412    
1413    1.  Change 37 of 8.13 broke patterns like [:a]...[b:] because it thought it had
1414        a POSIX class. After further experiments with Perl, which convinced me that
1415        Perl has bugs and confusions, a closing square bracket is no longer allowed
1416        in a POSIX name. This bug also affected patterns with classes that started
1417        with full stops.
1418    
1419    2.  If a pattern such as /(a)b|ac/ is matched against "ac", there is no
1420        captured substring, but while checking the failing first alternative,
1421        substring 1 is temporarily captured. If the output vector supplied to
1422        pcre_exec() was not big enough for this capture, the yield of the function
1423        was still zero ("insufficient space for captured substrings"). This cannot
1424        be totally fixed without adding another stack variable, which seems a lot
1425        of expense for a edge case. However, I have improved the situation in cases
1426        such as /(a)(b)x|abc/ matched against "abc", where the return code
1427        indicates that fewer than the maximum number of slots in the ovector have
1428        been set.
1429    
1430    3.  Related to (2) above: when there are more back references in a pattern than
1431        slots in the output vector, pcre_exec() uses temporary memory during
1432        matching, and copies in the captures as far as possible afterwards. It was
1433        using the entire output vector, but this conflicts with the specification
1434        that only 2/3 is used for passing back captured substrings. Now it uses
1435        only the first 2/3, for compatibility. This is, of course, another edge
1436        case.
1437    
1438    4.  Zoltan Herczeg's just-in-time compiler support has been integrated into the
1439        main code base, and can be used by building with --enable-jit. When this is
1440        done, pcregrep automatically uses it unless --disable-pcregrep-jit or the
1441        runtime --no-jit option is given.
1442    
1443    5.  When the number of matches in a pcre_dfa_exec() run exactly filled the
1444        ovector, the return from the function was zero, implying that there were
1445        other matches that did not fit. The correct "exactly full" value is now
1446        returned.
1447    
1448    6.  If a subpattern that was called recursively or as a subroutine contained
1449        (*PRUNE) or any other control that caused it to give a non-standard return,
1450        invalid errors such as "Error -26 (nested recursion at the same subject
1451        position)" or even infinite loops could occur.
1452    
1453    7.  If a pattern such as /a(*SKIP)c|b(*ACCEPT)|/ was studied, it stopped
1454        computing the minimum length on reaching *ACCEPT, and so ended up with the
1455        wrong value of 1 rather than 0. Further investigation indicates that
1456        computing a minimum subject length in the presence of *ACCEPT is difficult
1457        (think back references, subroutine calls), and so I have changed the code
1458        so that no minimum is registered for a pattern that contains *ACCEPT.
1459    
1460    8.  If (*THEN) was present in the first (true) branch of a conditional group,
1461        it was not handled as intended. [But see 16 below.]
1462    
1463    9.  Replaced RunTest.bat and CMakeLists.txt with improved versions provided by
1464        Sheri Pierce.
1465    
1466    10. A pathological pattern such as /(*ACCEPT)a/ was miscompiled, thinking that
1467        the first byte in a match must be "a".
1468    
1469    11. Change 17 for 8.13 increased the recursion depth for patterns like
1470        /a(?:.)*?a/ drastically. I've improved things by remembering whether a
1471        pattern contains any instances of (*THEN). If it does not, the old
1472        optimizations are restored. It would be nice to do this on a per-group
1473        basis, but at the moment that is not feasible.
1474    
1475    12. In some environments, the output of pcretest -C is CRLF terminated. This
1476        broke RunTest's code that checks for the link size. A single white space
1477        character after the value is now allowed for.
1478    
1479    13. RunTest now checks for the "fr" locale as well as for "fr_FR" and "french".
1480        For "fr", it uses the Windows-specific input and output files.
1481    
1482    14. If (*THEN) appeared in a group that was called recursively or as a
1483        subroutine, it did not work as intended. [But see next item.]
1484    
1485    15. Consider the pattern /A (B(*THEN)C) | D/ where A, B, C, and D are complex
1486        pattern fragments (but not containing any | characters). If A and B are
1487        matched, but there is a failure in C so that it backtracks to (*THEN), PCRE
1488        was behaving differently to Perl. PCRE backtracked into A, but Perl goes to
1489        D. In other words, Perl considers parentheses that do not contain any |
1490        characters to be part of a surrounding alternative, whereas PCRE was
1491        treading (B(*THEN)C) the same as (B(*THEN)C|(*FAIL)) -- which Perl handles
1492        differently. PCRE now behaves in the same way as Perl, except in the case
1493        of subroutine/recursion calls such as (?1) which have in any case always
1494        been different (but PCRE had them first :-).
1495    
1496    16. Related to 15 above: Perl does not treat the | in a conditional group as
1497        creating alternatives. Such a group is treated in the same way as an
1498        ordinary group without any | characters when processing (*THEN). PCRE has
1499        been changed to match Perl's behaviour.
1500    
1501    17. If a user had set PCREGREP_COLO(U)R to something other than 1:31, the
1502        RunGrepTest script failed.
1503    
1504    18. Change 22 for version 13 caused atomic groups to use more stack. This is
1505        inevitable for groups that contain captures, but it can lead to a lot of
1506        stack use in large patterns. The old behaviour has been restored for atomic
1507        groups that do not contain any capturing parentheses.
1508    
1509    19. If the PCRE_NO_START_OPTIMIZE option was set for pcre_compile(), it did not
1510        suppress the check for a minimum subject length at run time. (If it was
1511        given to pcre_exec() or pcre_dfa_exec() it did work.)
1512    
1513    20. Fixed an ASCII-dependent infelicity in pcretest that would have made it
1514        fail to work when decoding hex characters in data strings in EBCDIC
1515        environments.
1516    
1517    21. It appears that in at least one Mac OS environment, the isxdigit() function
1518        is implemented as a macro that evaluates to its argument more than once,
1519        contravening the C 90 Standard (I haven't checked a later standard). There
1520        was an instance in pcretest which caused it to go wrong when processing
1521        \x{...} escapes in subject strings. The has been rewritten to avoid using
1522        things like p++ in the argument of isxdigit().
1523    
1524    
1525    Version 8.13 16-Aug-2011
1526    ------------------------
1527    
1528    1.  The Unicode data tables have been updated to Unicode 6.0.0.
1529    
1530    2.  Two minor typos in pcre_internal.h have been fixed.
1531    
1532    3.  Added #include <string.h> to pcre_scanner_unittest.cc, pcrecpp.cc, and
1533        pcrecpp_unittest.cc. They are needed for strcmp(), memset(), and strchr()
1534        in some environments (e.g. Solaris 10/SPARC using Sun Studio 12U2).
1535    
1536    4.  There were a number of related bugs in the code for matching backrefences
1537        caselessly in UTF-8 mode when codes for the characters concerned were
1538        different numbers of bytes. For example, U+023A and U+2C65 are an upper
1539        and lower case pair, using 2 and 3 bytes, respectively. The main bugs were:
1540        (a) A reference to 3 copies of a 2-byte code matched only 2 of a 3-byte
1541        code. (b) A reference to 2 copies of a 3-byte code would not match 2 of a
1542        2-byte code at the end of the subject (it thought there wasn't enough data
1543        left).
1544    
1545    5.  Comprehensive information about what went wrong is now returned by
1546        pcre_exec() and pcre_dfa_exec() when the UTF-8 string check fails, as long
1547        as the output vector has at least 2 elements. The offset of the start of
1548        the failing character and a reason code are placed in the vector.
1549    
1550    6.  When the UTF-8 string check fails for pcre_compile(), the offset that is
1551        now returned is for the first byte of the failing character, instead of the
1552        last byte inspected. This is an incompatible change, but I hope it is small
1553        enough not to be a problem. It makes the returned offset consistent with
1554        pcre_exec() and pcre_dfa_exec().
1555    
1556    7.  pcretest now gives a text phrase as well as the error number when
1557        pcre_exec() or pcre_dfa_exec() fails; if the error is a UTF-8 check
1558        failure, the offset and reason code are output.
1559    
1560    8.  When \R was used with a maximizing quantifier it failed to skip backwards
1561        over a \r\n pair if the subsequent match failed. Instead, it just skipped
1562        back over a single character (\n). This seems wrong (because it treated the
1563        two characters as a single entity when going forwards), conflicts with the
1564        documentation that \R is equivalent to (?>\r\n|\n|...etc), and makes the
1565        behaviour of \R* different to (\R)*, which also seems wrong. The behaviour
1566        has been changed.
1567    
1568    9.  Some internal refactoring has changed the processing so that the handling
1569        of the PCRE_CASELESS and PCRE_MULTILINE options is done entirely at compile
1570        time (the PCRE_DOTALL option was changed this way some time ago: version
1571        7.7 change 16). This has made it possible to abolish the OP_OPT op code,
1572        which was always a bit of a fudge. It also means that there is one less
1573        argument for the match() function, which reduces its stack requirements
1574        slightly. This change also fixes an incompatibility with Perl: the pattern
1575        (?i:([^b]))(?1) should not match "ab", but previously PCRE gave a match.
1576    
1577    10. More internal refactoring has drastically reduced the number of recursive
1578        calls to match() for possessively repeated groups such as (abc)++ when
1579        using pcre_exec().
1580    
1581    11. While implementing 10, a number of bugs in the handling of groups were
1582        discovered and fixed:
1583    
1584        (?<=(a)+) was not diagnosed as invalid (non-fixed-length lookbehind).
1585        (a|)*(?1) gave a compile-time internal error.
1586        ((a|)+)+  did not notice that the outer group could match an empty string.
1587        (^a|^)+   was not marked as anchored.
1588        (.*a|.*)+ was not marked as matching at start or after a newline.
1589    
1590    12. Yet more internal refactoring has removed another argument from the match()
1591        function. Special calls to this function are now indicated by setting a
1592        value in a variable in the "match data" data block.
1593    
1594    13. Be more explicit in pcre_study() instead of relying on "default" for
1595        opcodes that mean there is no starting character; this means that when new
1596        ones are added and accidentally left out of pcre_study(), testing should
1597        pick them up.
1598    
1599    14. The -s option of pcretest has been documented for ages as being an old
1600        synonym of -m (show memory usage). I have changed it to mean "force study
1601        for every regex", that is, assume /S for every regex. This is similar to -i
1602        and -d etc. It's slightly incompatible, but I'm hoping nobody is still
1603        using it. It makes it easier to run collections of tests with and without
1604        study enabled, and thereby test pcre_study() more easily. All the standard
1605        tests are now run with and without -s (but some patterns can be marked as
1606        "never study" - see 20 below).
1607    
1608    15. When (*ACCEPT) was used in a subpattern that was called recursively, the
1609        restoration of the capturing data to the outer values was not happening
1610        correctly.
1611    
1612    16. If a recursively called subpattern ended with (*ACCEPT) and matched an
1613        empty string, and PCRE_NOTEMPTY was set, pcre_exec() thought the whole
1614        pattern had matched an empty string, and so incorrectly returned a no
1615        match.
1616    
1617    17. There was optimizing code for the last branch of non-capturing parentheses,
1618        and also for the obeyed branch of a conditional subexpression, which used
1619        tail recursion to cut down on stack usage. Unfortunately, now that there is
1620        the possibility of (*THEN) occurring in these branches, tail recursion is
1621        no longer possible because the return has to be checked for (*THEN). These
1622        two optimizations have therefore been removed. [But see 8.20/11 above.]
1623    
1624    18. If a pattern containing \R was studied, it was assumed that \R always
1625        matched two bytes, thus causing the minimum subject length to be
1626        incorrectly computed because \R can also match just one byte.
1627    
1628    19. If a pattern containing (*ACCEPT) was studied, the minimum subject length
1629        was incorrectly computed.
1630    
1631    20. If /S is present twice on a test pattern in pcretest input, it now
1632        *disables* studying, thereby overriding the use of -s on the command line
1633        (see 14 above). This is necessary for one or two tests to keep the output
1634        identical in both cases.
1635    
1636    21. When (*ACCEPT) was used in an assertion that matched an empty string and
1637        PCRE_NOTEMPTY was set, PCRE applied the non-empty test to the assertion.
1638    
1639    22. When an atomic group that contained a capturing parenthesis was
1640        successfully matched, but the branch in which it appeared failed, the
1641        capturing was not being forgotten if a higher numbered group was later
1642        captured. For example, /(?>(a))b|(a)c/ when matching "ac" set capturing
1643        group 1 to "a", when in fact it should be unset. This applied to multi-
1644        branched capturing and non-capturing groups, repeated or not, and also to
1645        positive assertions (capturing in negative assertions does not happen
1646        in PCRE) and also to nested atomic groups.
1647    
1648    23. Add the ++ qualifier feature to pcretest, to show the remainder of the
1649        subject after a captured substring, to make it easier to tell which of a
1650        number of identical substrings has been captured.
1651    
1652    24. The way atomic groups are processed by pcre_exec() has been changed so that
1653        if they are repeated, backtracking one repetition now resets captured
1654        values correctly. For example, if ((?>(a+)b)+aabab) is matched against
1655        "aaaabaaabaabab" the value of captured group 2 is now correctly recorded as
1656        "aaa". Previously, it would have been "a". As part of this code
1657        refactoring, the way recursive calls are handled has also been changed.
1658    
1659    25. If an assertion condition captured any substrings, they were not passed
1660        back unless some other capturing happened later. For example, if
1661        (?(?=(a))a) was matched against "a", no capturing was returned.
1662    
1663    26. When studying a pattern that contained subroutine calls or assertions,
1664        the code for finding the minimum length of a possible match was handling
1665        direct recursions such as (xxx(?1)|yyy) but not mutual recursions (where
1666        group 1 called group 2 while simultaneously a separate group 2 called group
1667        1). A stack overflow occurred in this case. I have fixed this by limiting
1668        the recursion depth to 10.
1669    
1670    27. Updated RunTest.bat in the distribution to the version supplied by Tom
1671        Fortmann. This supports explicit test numbers on the command line, and has
1672        argument validation and error reporting.
1673    
1674    28. An instance of \X with an unlimited repeat could fail if at any point the
1675        first character it looked at was a mark character.
1676    
1677    29. Some minor code refactoring concerning Unicode properties and scripts
1678        should reduce the stack requirement of match() slightly.
1679    
1680    30. Added the '=' option to pcretest to check the setting of unused capturing
1681        slots at the end of the pattern, which are documented as being -1, but are
1682        not included in the return count.
1683    
1684    31. If \k was not followed by a braced, angle-bracketed, or quoted name, PCRE
1685        compiled something random. Now it gives a compile-time error (as does
1686        Perl).
1687    
1688    32. A *MARK encountered during the processing of a positive assertion is now
1689        recorded and passed back (compatible with Perl).
1690    
1691    33. If --only-matching or --colour was set on a pcregrep call whose pattern
1692        had alternative anchored branches, the search for a second match in a line
1693        was done as if at the line start. Thus, for example, /^01|^02/ incorrectly
1694        matched the line "0102" twice. The same bug affected patterns that started
1695        with a backwards assertion. For example /\b01|\b02/ also matched "0102"
1696        twice.
1697    
1698    34. Previously, PCRE did not allow quantification of assertions. However, Perl
1699        does, and because of capturing effects, quantifying parenthesized
1700        assertions may at times be useful. Quantifiers are now allowed for
1701        parenthesized assertions.
1702    
1703    35. A minor code tidy in pcre_compile() when checking options for \R usage.
1704    
1705    36. \g was being checked for fancy things in a character class, when it should
1706        just be a literal "g".
1707    
1708    37. PCRE was rejecting [:a[:digit:]] whereas Perl was not. It seems that the
1709        appearance of a nested POSIX class supersedes an apparent external class.
1710        For example, [:a[:digit:]b:] matches "a", "b", ":", or a digit. Also,
1711        unescaped square brackets may also appear as part of class names. For
1712        example, [:a[:abc]b:] gives unknown class "[:abc]b:]". PCRE now behaves
1713        more like Perl. (But see 8.20/1 above.)
1714    
1715    38. PCRE was giving an error for \N with a braced quantifier such as {1,} (this
1716        was because it thought it was \N{name}, which is not supported).
1717    
1718    39. Add minix to OS list not supporting the -S option in pcretest.
1719    
1720    40. PCRE tries to detect cases of infinite recursion at compile time, but it
1721        cannot analyze patterns in sufficient detail to catch mutual recursions
1722        such as ((?1))((?2)). There is now a runtime test that gives an error if a
1723        subgroup is called recursively as a subpattern for a second time at the
1724        same position in the subject string. In previous releases this might have
1725        been caught by the recursion limit, or it might have run out of stack.
1726    
1727    41. A pattern such as /(?(R)a+|(?R)b)/ is quite safe, as the recursion can
1728        happen only once. PCRE was, however incorrectly giving a compile time error
1729        "recursive call could loop indefinitely" because it cannot analyze the
1730        pattern in sufficient detail. The compile time test no longer happens when
1731        PCRE is compiling a conditional subpattern, but actual runaway loops are
1732        now caught at runtime (see 40 above).
1733    
1734    42. It seems that Perl allows any characters other than a closing parenthesis
1735        to be part of the NAME in (*MARK:NAME) and other backtracking verbs. PCRE
1736        has been changed to be the same.
1737    
1738    43. Updated configure.ac to put in more quoting round AC_LANG_PROGRAM etc. so
1739        as not to get warnings when autogen.sh is called. Also changed
1740        AC_PROG_LIBTOOL (deprecated) to LT_INIT (the current macro).
1741    
1742    44. To help people who use pcregrep to scan files containing exceedingly long
1743        lines, the following changes have been made:
1744    
1745        (a) The default value of the buffer size parameter has been increased from
1746            8K to 20K. (The actual buffer used is three times this size.)
1747    
1748        (b) The default can be changed by ./configure --with-pcregrep-bufsize when
1749            PCRE is built.
1750    
1751        (c) A --buffer-size=n option has been added to pcregrep, to allow the size
1752            to be set at run time.
1753    
1754        (d) Numerical values in pcregrep options can be followed by K or M, for
1755            example --buffer-size=50K.
1756    
1757        (e) If a line being scanned overflows pcregrep's buffer, an error is now
1758            given and the return code is set to 2.
1759    
1760    45. Add a pointer to the latest mark to the callout data block.
1761    
1762    46. The pattern /.(*F)/, when applied to "abc" with PCRE_PARTIAL_HARD, gave a
1763        partial match of an empty string instead of no match. This was specific to
1764        the use of ".".
1765    
1766    47. The pattern /f.*/8s, when applied to "for" with PCRE_PARTIAL_HARD, gave a
1767        complete match instead of a partial match. This bug was dependent on both
1768        the PCRE_UTF8 and PCRE_DOTALL options being set.
1769    
1770    48. For a pattern such as /\babc|\bdef/ pcre_study() was failing to set up the
1771        starting byte set, because \b was not being ignored.
1772    
1773    
1774    Version 8.12 15-Jan-2011
1775    ------------------------
1776    
1777    1.  Fixed some typos in the markup of the man pages, and wrote a script that
1778        checks for such things as part of the documentation building process.
1779    
1780    2.  On a big-endian 64-bit system, pcregrep did not correctly process the
1781        --match-limit and --recursion-limit options (added for 8.11). In
1782        particular, this made one of the standard tests fail. (The integer value
1783        went into the wrong half of a long int.)
1784    
1785    3.  If the --colour option was given to pcregrep with -v (invert match), it
1786        did strange things, either producing crazy output, or crashing. It should,
1787        of course, ignore a request for colour when reporting lines that do not
1788        match.
1789    
1790    4.  Another pcregrep bug caused similar problems if --colour was specified with
1791        -M (multiline) and the pattern match finished with a line ending.
1792    
1793    5.  In pcregrep, when a pattern that ended with a literal newline sequence was
1794        matched in multiline mode, the following line was shown as part of the
1795        match. This seems wrong, so I have changed it.
1796    
1797    6.  Another pcregrep bug in multiline mode, when --colour was specified, caused
1798        the check for further matches in the same line (so they could be coloured)
1799        to overrun the end of the current line. If another match was found, it was
1800        incorrectly shown (and then shown again when found in the next line).
1801    
1802    7.  If pcregrep was compiled under Windows, there was a reference to the
1803        function pcregrep_exit() before it was defined. I am assuming this was
1804        the cause of the "error C2371: 'pcregrep_exit' : redefinition;" that was
1805        reported by a user. I've moved the definition above the reference.
1806    
1807    
1808    Version 8.11 10-Dec-2010
1809    ------------------------
1810    
1811    1.  (*THEN) was not working properly if there were untried alternatives prior
1812        to it in the current branch. For example, in ((a|b)(*THEN)(*F)|c..) it
1813        backtracked to try for "b" instead of moving to the next alternative branch
1814        at the same level (in this case, to look for "c"). The Perl documentation
1815        is clear that when (*THEN) is backtracked onto, it goes to the "next
1816        alternative in the innermost enclosing group".
1817    
1818    2.  (*COMMIT) was not overriding (*THEN), as it does in Perl. In a pattern
1819        such as   (A(*COMMIT)B(*THEN)C|D)  any failure after matching A should
1820        result in overall failure. Similarly, (*COMMIT) now overrides (*PRUNE) and
1821        (*SKIP), (*SKIP) overrides (*PRUNE) and (*THEN), and (*PRUNE) overrides
1822        (*THEN).
1823    
1824    3.  If \s appeared in a character class, it removed the VT character from
1825        the class, even if it had been included by some previous item, for example
1826        in [\x00-\xff\s]. (This was a bug related to the fact that VT is not part
1827        of \s, but is part of the POSIX "space" class.)
1828    
1829    4.  A partial match never returns an empty string (because you can always
1830        match an empty string at the end of the subject); however the checking for
1831        an empty string was starting at the "start of match" point. This has been
1832        changed to the "earliest inspected character" point, because the returned
1833        data for a partial match starts at this character. This means that, for
1834        example, /(?<=abc)def/ gives a partial match for the subject "abc"
1835        (previously it gave "no match").
1836    
1837    5.  Changes have been made to the way PCRE_PARTIAL_HARD affects the matching
1838        of $, \z, \Z, \b, and \B. If the match point is at the end of the string,
1839        previously a full match would be given. However, setting PCRE_PARTIAL_HARD
1840        has an implication that the given string is incomplete (because a partial
1841        match is preferred over a full match). For this reason, these items now
1842        give a partial match in this situation. [Aside: previously, the one case
1843        /t\b/ matched against "cat" with PCRE_PARTIAL_HARD set did return a partial
1844        match rather than a full match, which was wrong by the old rules, but is
1845        now correct.]
1846    
1847    6.  There was a bug in the handling of #-introduced comments, recognized when
1848        PCRE_EXTENDED is set, when PCRE_NEWLINE_ANY and PCRE_UTF8 were also set.
1849        If a UTF-8 multi-byte character included the byte 0x85 (e.g. +U0445, whose
1850        UTF-8 encoding is 0xd1,0x85), this was misinterpreted as a newline when
1851        scanning for the end of the comment. (*Character* 0x85 is an "any" newline,
1852        but *byte* 0x85 is not, in UTF-8 mode). This bug was present in several
1853        places in pcre_compile().
1854    
1855    7.  Related to (6) above, when pcre_compile() was skipping #-introduced
1856        comments when looking ahead for named forward references to subpatterns,
1857        the only newline sequence it recognized was NL. It now handles newlines
1858        according to the set newline convention.
1859    
1860    8.  SunOS4 doesn't have strerror() or strtoul(); pcregrep dealt with the
1861        former, but used strtoul(), whereas pcretest avoided strtoul() but did not
1862        cater for a lack of strerror(). These oversights have been fixed.
1863    
1864    9.  Added --match-limit and --recursion-limit to pcregrep.
1865    
1866    10. Added two casts needed to build with Visual Studio when NO_RECURSE is set.
1867    
1868    11. When the -o option was used, pcregrep was setting a return code of 1, even
1869        when matches were found, and --line-buffered was not being honoured.
1870    
1871    12. Added an optional parentheses number to the -o and --only-matching options
1872        of pcregrep.
1873    
1874    13. Imitating Perl's /g action for multiple matches is tricky when the pattern
1875        can match an empty string. The code to do it in pcretest and pcredemo
1876        needed fixing:
1877    
1878        (a) When the newline convention was "crlf", pcretest got it wrong, skipping
1879            only one byte after an empty string match just before CRLF (this case
1880            just got forgotten; "any" and "anycrlf" were OK).
1881    
1882        (b) The pcretest code also had a bug, causing it to loop forever in UTF-8
1883            mode when an empty string match preceded an ASCII character followed by
1884            a non-ASCII character. (The code for advancing by one character rather
1885            than one byte was nonsense.)
1886    
1887        (c) The pcredemo.c sample program did not have any code at all to handle
1888            the cases when CRLF is a valid newline sequence.
1889    
1890    14. Neither pcre_exec() nor pcre_dfa_exec() was checking that the value given
1891        as a starting offset was within the subject string. There is now a new
1892        error, PCRE_ERROR_BADOFFSET, which is returned if the starting offset is
1893        negative or greater than the length of the string. In order to test this,
1894        pcretest is extended to allow the setting of negative starting offsets.
1895    
1896    15. In both pcre_exec() and pcre_dfa_exec() the code for checking that the
1897        starting offset points to the beginning of a UTF-8 character was
1898        unnecessarily clumsy. I tidied it up.
1899    
1900    16. Added PCRE_ERROR_SHORTUTF8 to make it possible to distinguish between a
1901        bad UTF-8 sequence and one that is incomplete when using PCRE_PARTIAL_HARD.
1902    
1903    17. Nobody had reported that the --include_dir option, which was added in
1904        release 7.7 should have been called --include-dir (hyphen, not underscore)
1905        for compatibility with GNU grep. I have changed it to --include-dir, but
1906        left --include_dir as an undocumented synonym, and the same for
1907        --exclude-dir, though that is not available in GNU grep, at least as of
1908        release 2.5.4.
1909    
1910    18. At a user's suggestion, the macros GETCHAR and friends (which pick up UTF-8
1911        characters from a string of bytes) have been redefined so as not to use
1912        loops, in order to improve performance in some environments. At the same
1913        time, I abstracted some of the common code into auxiliary macros to save
1914        repetition (this should not affect the compiled code).
1915    
1916    19. If \c was followed by a multibyte UTF-8 character, bad things happened. A
1917        compile-time error is now given if \c is not followed by an ASCII
1918        character, that is, a byte less than 128. (In EBCDIC mode, the code is
1919        different, and any byte value is allowed.)
1920    
1921    20. Recognize (*NO_START_OPT) at the start of a pattern to set the PCRE_NO_
1922        START_OPTIMIZE option, which is now allowed at compile time - but just
1923        passed through to pcre_exec() or pcre_dfa_exec(). This makes it available
1924        to pcregrep and other applications that have no direct access to PCRE
1925        options. The new /Y option in pcretest sets this option when calling
1926        pcre_compile().
1927    
1928    21. Change 18 of release 8.01 broke the use of named subpatterns for recursive
1929        back references. Groups containing recursive back references were forced to
1930        be atomic by that change, but in the case of named groups, the amount of
1931        memory required was incorrectly computed, leading to "Failed: internal
1932        error: code overflow". This has been fixed.
1933    
1934    22. Some patches to pcre_stringpiece.h, pcre_stringpiece_unittest.cc, and
1935        pcretest.c, to avoid build problems in some Borland environments.
1936    
1937    
1938    Version 8.10 25-Jun-2010
1939    ------------------------
1940    
1941    1.  Added support for (*MARK:ARG) and for ARG additions to PRUNE, SKIP, and
1942        THEN.
1943    
1944    2.  (*ACCEPT) was not working when inside an atomic group.
1945    
1946    3.  Inside a character class, \B is treated as a literal by default, but
1947        faulted if PCRE_EXTRA is set. This mimics Perl's behaviour (the -w option
1948        causes the error). The code is unchanged, but I tidied the documentation.
1949    
1950    4.  Inside a character class, PCRE always treated \R and \X as literals,
1951        whereas Perl faults them if its -w option is set. I have changed PCRE so
1952        that it faults them when PCRE_EXTRA is set.
1953    
1954    5.  Added support for \N, which always matches any character other than
1955        newline. (It is the same as "." when PCRE_DOTALL is not set.)
1956    
1957    6.  When compiling pcregrep with newer versions of gcc which may have
1958        FORTIFY_SOURCE set, several warnings "ignoring return value of 'fwrite',
1959        declared with attribute warn_unused_result" were given. Just casting the
1960        result to (void) does not stop the warnings; a more elaborate fudge is
1961        needed. I've used a macro to implement this.
1962    
1963    7.  Minor change to pcretest.c to avoid a compiler warning.
1964    
1965    8.  Added four artifical Unicode properties to help with an option to make
1966        \s etc use properties (see next item). The new properties are: Xan
1967        (alphanumeric), Xsp (Perl space), Xps (POSIX space), and Xwd (word).
1968    
1969    9.  Added PCRE_UCP to make \b, \d, \s, \w, and certain POSIX character classes
1970        use Unicode properties. (*UCP) at the start of a pattern can be used to set
1971        this option. Modified pcretest to add /W to test this facility. Added
1972        REG_UCP to make it available via the POSIX interface.
1973    
1974    10. Added --line-buffered to pcregrep.
1975    
1976    11. In UTF-8 mode, if a pattern that was compiled with PCRE_CASELESS was
1977        studied, and the match started with a letter with a code point greater than
1978        127 whose first byte was different to the first byte of the other case of
1979        the letter, the other case of this starting letter was not recognized
1980        (#976).
1981    
1982    12. If a pattern that was studied started with a repeated Unicode property
1983        test, for example, \p{Nd}+, there was the theoretical possibility of
1984        setting up an incorrect bitmap of starting bytes, but fortunately it could
1985        not have actually happened in practice until change 8 above was made (it
1986        added property types that matched character-matching opcodes).
1987    
1988    13. pcre_study() now recognizes \h, \v, and \R when constructing a bit map of
1989        possible starting bytes for non-anchored patterns.
1990    
1991    14. Extended the "auto-possessify" feature of pcre_compile(). It now recognizes
1992        \R, and also a number of cases that involve Unicode properties, both
1993        explicit and implicit when PCRE_UCP is set.
1994    
1995    15. If a repeated Unicode property match (e.g. \p{Lu}*) was used with non-UTF-8
1996        input, it could crash or give wrong results if characters with values
1997        greater than 0xc0 were present in the subject string. (Detail: it assumed
1998        UTF-8 input when processing these items.)
1999    
2000    16. Added a lot of (int) casts to avoid compiler warnings in systems where
2001        size_t is 64-bit (#991).
2002    
2003    17. Added a check for running out of memory when PCRE is compiled with
2004        --disable-stack-for-recursion (#990).
2005    
2006    18. If the last data line in a file for pcretest does not have a newline on
2007        the end, a newline was missing in the output.
2008    
2009    19. The default pcre_chartables.c file recognizes only ASCII characters (values
2010        less than 128) in its various bitmaps. However, there is a facility for
2011        generating tables according to the current locale when PCRE is compiled. It
2012        turns out that in some environments, 0x85 and 0xa0, which are Unicode space
2013        characters, are recognized by isspace() and therefore were getting set in
2014        these tables, and indeed these tables seem to approximate to ISO 8859. This
2015        caused a problem in UTF-8 mode when pcre_study() was used to create a list
2016        of bytes that can start a match. For \s, it was including 0x85 and 0xa0,
2017        which of course cannot start UTF-8 characters. I have changed the code so
2018        that only real ASCII characters (less than 128) and the correct starting
2019        bytes for UTF-8 encodings are set for characters greater than 127 when in
2020        UTF-8 mode. (When PCRE_UCP is set - see 9 above - the code is different
2021        altogether.)
2022    
2023    20. Added the /T option to pcretest so as to be able to run tests with non-
2024        standard character tables, thus making it possible to include the tests
2025        used for 19 above in the standard set of tests.
2026    
2027    21. A pattern such as (?&t)(?#()(?(DEFINE)(?<t>a)) which has a forward
2028        reference to a subpattern the other side of a comment that contains an
2029        opening parenthesis caused either an internal compiling error, or a
2030        reference to the wrong subpattern.
2031    
2032    
2033    Version 8.02 19-Mar-2010
2034    ------------------------
2035    
2036    1.  The Unicode data tables have been updated to Unicode 5.2.0.
2037    
2038    2.  Added the option --libs-cpp to pcre-config, but only when C++ support is
2039        configured.
2040    
2041    3.  Updated the licensing terms in the pcregexp.pas file, as agreed with the
2042        original author of that file, following a query about its status.
2043    
2044    4.  On systems that do not have stdint.h (e.g. Solaris), check for and include
2045        inttypes.h instead. This fixes a bug that was introduced by change 8.01/8.
2046    
2047    5.  A pattern such as (?&t)*+(?(DEFINE)(?<t>.)) which has a possessive
2048        quantifier applied to a forward-referencing subroutine call, could compile
2049        incorrect code or give the error "internal error: previously-checked
2050        referenced subpattern not found".
2051    
2052    6.  Both MS Visual Studio and Symbian OS have problems with initializing
2053        variables to point to external functions. For these systems, therefore,
2054        pcre_malloc etc. are now initialized to local functions that call the
2055        relevant global functions.
2056    
2057    7.  There were two entries missing in the vectors called coptable and poptable
2058        in pcre_dfa_exec.c. This could lead to memory accesses outsize the vectors.
2059        I've fixed the data, and added a kludgy way of testing at compile time that
2060        the lengths are correct (equal to the number of opcodes).
2061    
2062    8.  Following on from 7, I added a similar kludge to check the length of the
2063        eint vector in pcreposix.c.
2064    
2065    9.  Error texts for pcre_compile() are held as one long string to avoid too
2066        much relocation at load time. To find a text, the string is searched,
2067        counting zeros. There was no check for running off the end of the string,
2068        which could happen if a new error number was added without updating the
2069        string.
2070    
2071    10. \K gave a compile-time error if it appeared in a lookbehind assersion.
2072    
2073    11. \K was not working if it appeared in an atomic group or in a group that
2074        was called as a "subroutine", or in an assertion. Perl 5.11 documents that
2075        \K is "not well defined" if used in an assertion. PCRE now accepts it if
2076        the assertion is positive, but not if it is negative.
2077    
2078    12. Change 11 fortuitously reduced the size of the stack frame used in the
2079        "match()" function of pcre_exec.c by one pointer. Forthcoming
2080        implementation of support for (*MARK) will need an extra pointer on the
2081        stack; I have reserved it now, so that the stack frame size does not
2082        decrease.
2083    
2084    13. A pattern such as (?P<L1>(?P<L2>0)|(?P>L2)(?P>L1)) in which the only other
2085        item in branch that calls a recursion is a subroutine call - as in the
2086        second branch in the above example - was incorrectly given the compile-
2087        time error "recursive call could loop indefinitely" because pcre_compile()
2088        was not correctly checking the subroutine for matching a non-empty string.
2089    
2090    14. The checks for overrunning compiling workspace could trigger after an
2091        overrun had occurred. This is a "should never occur" error, but it can be
2092        triggered by pathological patterns such as hundreds of nested parentheses.
2093        The checks now trigger 100 bytes before the end of the workspace.
2094    
2095    15. Fix typo in configure.ac: "srtoq" should be "strtoq".
2096    
2097    
2098    Version 8.01 19-Jan-2010
2099    ------------------------
2100    
2101    1.  If a pattern contained a conditional subpattern with only one branch (in
2102        particular, this includes all (*DEFINE) patterns), a call to pcre_study()
2103        computed the wrong minimum data length (which is of course zero for such
2104        subpatterns). This could cause incorrect "no match" results.
2105    
2106    2.  For patterns such as (?i)a(?-i)b|c where an option setting at the start of
2107        the pattern is reset in the first branch, pcre_compile() failed with
2108        "internal error: code overflow at offset...". This happened only when
2109        the reset was to the original external option setting. (An optimization
2110        abstracts leading options settings into an external setting, which was the
2111        cause of this.)
2112    
2113    3.  A pattern such as ^(?!a(*SKIP)b) where a negative assertion contained one
2114        of the verbs SKIP, PRUNE, or COMMIT, did not work correctly. When the
2115        assertion pattern did not match (meaning that the assertion was true), it
2116        was incorrectly treated as false if the SKIP had been reached during the
2117        matching. This also applied to assertions used as conditions.
2118    
2119    4.  If an item that is not supported by pcre_dfa_exec() was encountered in an
2120        assertion subpattern, including such a pattern used as a condition,
2121        unpredictable results occurred, instead of the error return
2122        PCRE_ERROR_DFA_UITEM.
2123    
2124    5.  The C++ GlobalReplace function was not working like Perl for the special
2125        situation when an empty string is matched. It now does the fancy magic
2126        stuff that is necessary.
2127    
2128    6.  In pcre_internal.h, obsolete includes to setjmp.h and stdarg.h have been
2129        removed. (These were left over from very, very early versions of PCRE.)
2130    
2131    7.  Some cosmetic changes to the code to make life easier when compiling it
2132        as part of something else:
2133    
2134        (a) Change DEBUG to PCRE_DEBUG.
2135    
2136        (b) In pcre_compile(), rename the member of the "branch_chain" structure
2137            called "current" as "current_branch", to prevent a collision with the
2138            Linux macro when compiled as a kernel module.
2139    
2140        (c) In pcre_study(), rename the function set_bit() as set_table_bit(), to
2141            prevent a collision with the Linux macro when compiled as a kernel
2142            module.
2143    
2144    8.  In pcre_compile() there are some checks for integer overflows that used to
2145        cast potentially large values to (double). This has been changed to that
2146        when building, a check for int64_t is made, and if it is found, it is used
2147        instead, thus avoiding the use of floating point arithmetic. (There is no
2148        other use of FP in PCRE.) If int64_t is not found, the fallback is to
2149        double.
2150    
2151    9.  Added two casts to avoid signed/unsigned warnings from VS Studio Express
2152        2005 (difference between two addresses compared to an unsigned value).
2153    
2154    10. Change the standard AC_CHECK_LIB test for libbz2 in configure.ac to a
2155        custom one, because of the following reported problem in Windows:
2156    
2157          - libbz2 uses the Pascal calling convention (WINAPI) for the functions
2158              under Win32.
2159          - The standard autoconf AC_CHECK_LIB fails to include "bzlib.h",
2160              therefore missing the function definition.
2161          - The compiler thus generates a "C" signature for the test function.
2162          - The linker fails to find the "C" function.
2163          - PCRE fails to configure if asked to do so against libbz2.
2164    
2165    11. When running libtoolize from libtool-2.2.6b as part of autogen.sh, these
2166        messages were output:
2167    
2168          Consider adding `AC_CONFIG_MACRO_DIR([m4])' to configure.ac and
2169          rerunning libtoolize, to keep the correct libtool macros in-tree.
2170          Consider adding `-I m4' to ACLOCAL_AMFLAGS in Makefile.am.
2171    
2172        I have done both of these things.
2173    
2174    12. Although pcre_dfa_exec() does not use nearly as much stack as pcre_exec()
2175        most of the time, it *can* run out if it is given a pattern that contains a
2176        runaway infinite recursion. I updated the discussion in the pcrestack man
2177        page.
2178    
2179    13. Now that we have gone to the x.xx style of version numbers, the minor
2180        version may start with zero. Using 08 or 09 is a bad idea because users
2181        might check the value of PCRE_MINOR in their code, and 08 or 09 may be
2182        interpreted as invalid octal numbers. I've updated the previous comment in
2183        configure.ac, and also added a check that gives an error if 08 or 09 are
2184        used.
2185    
2186    14. Change 8.00/11 was not quite complete: code had been accidentally omitted,
2187        causing partial matching to fail when the end of the subject matched \W
2188        in a UTF-8 pattern where \W was quantified with a minimum of 3.
2189    
2190    15. There were some discrepancies between the declarations in pcre_internal.h
2191        of _pcre_is_newline(), _pcre_was_newline(), and _pcre_valid_utf8() and
2192        their definitions. The declarations used "const uschar *" and the
2193        definitions used USPTR. Even though USPTR is normally defined as "const
2194        unsigned char *" (and uschar is typedeffed as "unsigned char"), it was
2195        reported that: "This difference in casting confuses some C++ compilers, for
2196        example, SunCC recognizes above declarations as different functions and
2197        generates broken code for hbpcre." I have changed the declarations to use
2198        USPTR.
2199    
2200    16. GNU libtool is named differently on some systems. The autogen.sh script now
2201        tries several variants such as glibtoolize (MacOSX) and libtoolize1x
2202        (FreeBSD).
2203    
2204    17. Applied Craig's patch that fixes an HP aCC compile error in pcre 8.00
2205        (strtoXX undefined when compiling pcrecpp.cc). The patch contains this
2206        comment: "Figure out how to create a longlong from a string: strtoll and
2207        equivalent. It's not enough to call AC_CHECK_FUNCS: hpux has a strtoll, for
2208        instance, but it only takes 2 args instead of 3!"
2209    
2210    18. A subtle bug concerned with back references has been fixed by a change of
2211        specification, with a corresponding code fix. A pattern such as
2212        ^(xa|=?\1a)+$ which contains a back reference inside the group to which it
2213        refers, was giving matches when it shouldn't. For example, xa=xaaa would
2214        match that pattern. Interestingly, Perl (at least up to 5.11.3) has the
2215        same bug. Such groups have to be quantified to be useful, or contained
2216        inside another quantified group. (If there's no repetition, the reference
2217        can never match.) The problem arises because, having left the group and
2218        moved on to the rest of the pattern, a later failure that backtracks into
2219        the group uses the captured value from the final iteration of the group
2220        rather than the correct earlier one. I have fixed this in PCRE by forcing
2221        any group that contains a reference to itself to be an atomic group; that
2222        is, there cannot be any backtracking into it once it has completed. This is
2223        similar to recursive and subroutine calls.
2224    
2225    
2226    Version 8.00 19-Oct-09
2227    ----------------------
2228    
2229    1.  The table for translating pcre_compile() error codes into POSIX error codes
2230        was out-of-date, and there was no check on the pcre_compile() error code
2231        being within the table. This could lead to an OK return being given in
2232        error.
2233    
2234    2.  Changed the call to open a subject file in pcregrep from fopen(pathname,
2235        "r") to fopen(pathname, "rb"), which fixed a problem with some of the tests
2236        in a Windows environment.
2237    
2238    3.  The pcregrep --count option prints the count for each file even when it is
2239        zero, as does GNU grep. However, pcregrep was also printing all files when
2240        --files-with-matches was added. Now, when both options are given, it prints
2241        counts only for those files that have at least one match. (GNU grep just
2242        prints the file name in this circumstance, but including the count seems
2243        more useful - otherwise, why use --count?) Also ensured that the
2244        combination -clh just lists non-zero counts, with no names.
2245    
2246    4.  The long form of the pcregrep -F option was incorrectly implemented as
2247        --fixed_strings instead of --fixed-strings. This is an incompatible change,
2248        but it seems right to fix it, and I didn't think it was worth preserving
2249        the old behaviour.
2250    
2251    5.  The command line items --regex=pattern and --regexp=pattern were not
2252        recognized by pcregrep, which required --regex pattern or --regexp pattern
2253        (with a space rather than an '='). The man page documented the '=' forms,
2254        which are compatible with GNU grep; these now work.
2255    
2256    6.  No libpcreposix.pc file was created for pkg-config; there was just
2257        libpcre.pc and libpcrecpp.pc. The omission has been rectified.
2258    
2259    7.  Added #ifndef SUPPORT_UCP into the pcre_ucd.c module, to reduce its size
2260        when UCP support is not needed, by modifying the Python script that
2261        generates it from Unicode data files. This should not matter if the module
2262        is correctly used as a library, but I received one complaint about 50K of
2263        unwanted data. My guess is that the person linked everything into his
2264        program rather than using a library. Anyway, it does no harm.
2265    
2266    8.  A pattern such as /\x{123}{2,2}+/8 was incorrectly compiled; the trigger
2267        was a minimum greater than 1 for a wide character in a possessive
2268        repetition. The same bug could also affect patterns like /(\x{ff}{0,2})*/8
2269        which had an unlimited repeat of a nested, fixed maximum repeat of a wide
2270        character. Chaos in the form of incorrect output or a compiling loop could
2271        result.
2272    
2273    9.  The restrictions on what a pattern can contain when partial matching is
2274        requested for pcre_exec() have been removed. All patterns can now be
2275        partially matched by this function. In addition, if there are at least two
2276        slots in the offset vector, the offset of the earliest inspected character
2277        for the match and the offset of the end of the subject are set in them when
2278        PCRE_ERROR_PARTIAL is returned.
2279    
2280    10. Partial matching has been split into two forms: PCRE_PARTIAL_SOFT, which is
2281        synonymous with PCRE_PARTIAL, for backwards compatibility, and
2282        PCRE_PARTIAL_HARD, which causes a partial match to supersede a full match,
2283        and may be more useful for multi-segment matching.
2284    
2285    11. Partial matching with pcre_exec() is now more intuitive. A partial match
2286        used to be given if ever the end of the subject was reached; now it is
2287        given only if matching could not proceed because another character was
2288        needed. This makes a difference in some odd cases such as Z(*FAIL) with the
2289        string "Z", which now yields "no match" instead of "partial match". In the
2290        case of pcre_dfa_exec(), "no match" is given if every matching path for the
2291        final character ended with (*FAIL).
2292    
2293    12. Restarting a match using pcre_dfa_exec() after a partial match did not work
2294        if the pattern had a "must contain" character that was already found in the
2295        earlier partial match, unless partial matching was again requested. For
2296        example, with the pattern /dog.(body)?/, the "must contain" character is
2297        "g". If the first part-match was for the string "dog", restarting with
2298        "sbody" failed. This bug has been fixed.
2299    
2300    13. The string returned by pcre_dfa_exec() after a partial match has been
2301        changed so that it starts at the first inspected character rather than the
2302        first character of the match. This makes a difference only if the pattern
2303        starts with a lookbehind assertion or \b or \B (\K is not supported by
2304        pcre_dfa_exec()). It's an incompatible change, but it makes the two
2305        matching functions compatible, and I think it's the right thing to do.
2306    
2307    14. Added a pcredemo man page, created automatically from the pcredemo.c file,
2308        so that the demonstration program is easily available in environments where
2309        PCRE has not been installed from source.
2310    
2311    15. Arranged to add -DPCRE_STATIC to cflags in libpcre.pc, libpcreposix.cp,
2312        libpcrecpp.pc and pcre-config when PCRE is not compiled as a shared
2313        library.
2314    
2315    16. Added REG_UNGREEDY to the pcreposix interface, at the request of a user.
2316        It maps to PCRE_UNGREEDY. It is not, of course, POSIX-compatible, but it
2317        is not the first non-POSIX option to be added. Clearly some people find
2318        these options useful.
2319    
2320    17. If a caller to the POSIX matching function regexec() passes a non-zero
2321        value for nmatch with a NULL value for pmatch, the value of
2322        nmatch is forced to zero.
2323    
2324    18. RunGrepTest did not have a test for the availability of the -u option of
2325        the diff command, as RunTest does. It now checks in the same way as
2326        RunTest, and also checks for the -b option.
2327    
2328    19. If an odd number of negated classes containing just a single character
2329        interposed, within parentheses, between a forward reference to a named
2330        subpattern and the definition of the subpattern, compilation crashed with
2331        an internal error, complaining that it could not find the referenced
2332        subpattern. An example of a crashing pattern is /(?&A)(([^m])(?<A>))/.
2333        [The bug was that it was starting one character too far in when skipping
2334        over the character class, thus treating the ] as data rather than
2335        terminating the class. This meant it could skip too much.]
2336    
2337    20. Added PCRE_NOTEMPTY_ATSTART in order to be able to correctly implement the
2338        /g option in pcretest when the pattern contains \K, which makes it possible
2339        to have an empty string match not at the start, even when the pattern is
2340        anchored. Updated pcretest and pcredemo to use this option.
2341    
2342    21. If the maximum number of capturing subpatterns in a recursion was greater
2343        than the maximum at the outer level, the higher number was returned, but
2344        with unset values at the outer level. The correct (outer level) value is
2345        now given.
2346    
2347    22. If (*ACCEPT) appeared inside capturing parentheses, previous releases of
2348        PCRE did not set those parentheses (unlike Perl). I have now found a way to
2349        make it do so. The string so far is captured, making this feature
2350        compatible with Perl.
2351    
2352    23. The tests have been re-organized, adding tests 11 and 12, to make it
2353        possible to check the Perl 5.10 features against Perl 5.10.
2354    
2355    24. Perl 5.10 allows subroutine calls in lookbehinds, as long as the subroutine
2356        pattern matches a fixed length string. PCRE did not allow this; now it
2357        does. Neither allows recursion.
2358    
2359    25. I finally figured out how to implement a request to provide the minimum
2360        length of subject string that was needed in order to match a given pattern.
2361        (It was back references and recursion that I had previously got hung up
2362        on.) This code has now been added to pcre_study(); it finds a lower bound
2363        to the length of subject needed. It is not necessarily the greatest lower
2364        bound, but using it to avoid searching strings that are too short does give
2365        some useful speed-ups. The value is available to calling programs via
2366        pcre_fullinfo().
2367    
2368    26. While implementing 25, I discovered to my embarrassment that pcretest had
2369        not been passing the result of pcre_study() to pcre_dfa_exec(), so the
2370        study optimizations had never been tested with that matching function.
2371        Oops. What is worse, even when it was passed study data, there was a bug in
2372        pcre_dfa_exec() that meant it never actually used it. Double oops. There
2373        were also very few tests of studied patterns with pcre_dfa_exec().
2374    
2375    27. If (?| is used to create subpatterns with duplicate numbers, they are now
2376        allowed to have the same name, even if PCRE_DUPNAMES is not set. However,
2377        on the other side of the coin, they are no longer allowed to have different
2378        names, because these cannot be distinguished in PCRE, and this has caused
2379        confusion. (This is a difference from Perl.)
2380    
2381    28. When duplicate subpattern names are present (necessarily with different
2382        numbers, as required by 27 above), and a test is made by name in a
2383        conditional pattern, either for a subpattern having been matched, or for
2384        recursion in such a pattern, all the associated numbered subpatterns are
2385        tested, and the overall condition is true if the condition is true for any
2386        one of them. This is the way Perl works, and is also more like the way
2387        testing by number works.
2388    
2389    
2390    Version 7.9 11-Apr-09
2391    ---------------------
2392    
2393    1.  When building with support for bzlib/zlib (pcregrep) and/or readline
2394        (pcretest), all targets were linked against these libraries. This included
2395        libpcre, libpcreposix, and libpcrecpp, even though they do not use these
2396        libraries. This caused unwanted dependencies to be created. This problem
2397        has been fixed, and now only pcregrep is linked with bzlib/zlib and only
2398        pcretest is linked with readline.
2399    
2400    2.  The "typedef int BOOL" in pcre_internal.h that was included inside the
2401        "#ifndef FALSE" condition by an earlier change (probably 7.8/18) has been
2402        moved outside it again, because FALSE and TRUE are already defined in AIX,
2403        but BOOL is not.
2404    
2405    3.  The pcre_config() function was treating the PCRE_MATCH_LIMIT and
2406        PCRE_MATCH_LIMIT_RECURSION values as ints, when they should be long ints.
2407    
2408    4.  The pcregrep documentation said spaces were inserted as well as colons (or
2409        hyphens) following file names and line numbers when outputting matching
2410        lines. This is not true; no spaces are inserted. I have also clarified the
2411        wording for the --colour (or --color) option.
2412    
2413    5.  In pcregrep, when --colour was used with -o, the list of matching strings
2414        was not coloured; this is different to GNU grep, so I have changed it to be
2415        the same.
2416    
2417    6.  When --colo(u)r was used in pcregrep, only the first matching substring in
2418        each matching line was coloured. Now it goes on to look for further matches
2419        of any of the test patterns, which is the same behaviour as GNU grep.
2420    
2421    7.  A pattern that could match an empty string could cause pcregrep to loop; it
2422        doesn't make sense to accept an empty string match in pcregrep, so I have
2423        locked it out (using PCRE's PCRE_NOTEMPTY option). By experiment, this
2424        seems to be how GNU grep behaves. [But see later change 40 for release
2425        8.33.]
2426    
2427    8.  The pattern (?(?=.*b)b|^) was incorrectly compiled as "match must be at
2428        start or after a newline", because the conditional assertion was not being
2429        correctly handled. The rule now is that both the assertion and what follows
2430        in the first alternative must satisfy the test.
2431    
2432    9.  If auto-callout was enabled in a pattern with a conditional group whose
2433        condition was an assertion, PCRE could crash during matching, both with
2434        pcre_exec() and pcre_dfa_exec().
2435    
2436    10. The PCRE_DOLLAR_ENDONLY option was not working when pcre_dfa_exec() was
2437        used for matching.
2438    
2439    11. Unicode property support in character classes was not working for
2440        characters (bytes) greater than 127 when not in UTF-8 mode.
2441    
2442    12. Added the -M command line option to pcretest.
2443    
2444    14. Added the non-standard REG_NOTEMPTY option to the POSIX interface.
2445    
2446    15. Added the PCRE_NO_START_OPTIMIZE match-time option.
2447    
2448    16. Added comments and documentation about mis-use of no_arg in the C++
2449        wrapper.
2450    
2451    17. Implemented support for UTF-8 encoding in EBCDIC environments, a patch
2452        from Martin Jerabek that uses macro names for all relevant character and
2453        string constants.
2454    
2455    18. Added to pcre_internal.h two configuration checks: (a) If both EBCDIC and
2456        SUPPORT_UTF8 are set, give an error; (b) If SUPPORT_UCP is set without
2457        SUPPORT_UTF8, define SUPPORT_UTF8. The "configure" script handles both of
2458        these, but not everybody uses configure.
2459    
2460    19. A conditional group that had only one branch was not being correctly
2461        recognized as an item that could match an empty string. This meant that an
2462        enclosing group might also not be so recognized, causing infinite looping
2463        (and probably a segfault) for patterns such as ^"((?(?=[a])[^"])|b)*"$
2464        with the subject "ab", where knowledge that the repeated group can match
2465        nothing is needed in order to break the loop.
2466    
2467    20. If a pattern that was compiled with callouts was matched using pcre_dfa_
2468        exec(), but without supplying a callout function, matching went wrong.
2469    
2470    21. If PCRE_ERROR_MATCHLIMIT occurred during a recursion, there was a memory
2471        leak if the size of the offset vector was greater than 30. When the vector
2472        is smaller, the saved offsets during recursion go onto a local stack
2473        vector, but for larger vectors malloc() is used. It was failing to free
2474        when the recursion yielded PCRE_ERROR_MATCH_LIMIT (or any other "abnormal"
2475        error, in fact).
2476    
2477    22. There was a missing #ifdef SUPPORT_UTF8 round one of the variables in the
2478        heapframe that is used only when UTF-8 support is enabled. This caused no
2479        problem, but was untidy.
2480    
2481    23. Steven Van Ingelgem's patch to CMakeLists.txt to change the name
2482        CMAKE_BINARY_DIR to PROJECT_BINARY_DIR so that it works when PCRE is
2483        included within another project.
2484    
2485    24. Steven Van Ingelgem's patches to add more options to the CMake support,
2486        slightly modified by me:
2487    
2488          (a) PCRE_BUILD_TESTS can be set OFF not to build the tests, including
2489              not building pcregrep.
2490    
2491          (b) PCRE_BUILD_PCREGREP can be see OFF not to build pcregrep, but only
2492              if PCRE_BUILD_TESTS is also set OFF, because the tests use pcregrep.
2493    
2494    25. Forward references, both numeric and by name, in patterns that made use of
2495        duplicate group numbers, could behave incorrectly or give incorrect errors,
2496        because when scanning forward to find the reference group, PCRE was not
2497        taking into account the duplicate group numbers. A pattern such as
2498        ^X(?3)(a)(?|(b)|(q))(Y) is an example.
2499    
2500    26. Changed a few more instances of "const unsigned char *" to USPTR, making
2501        the feature of a custom pointer more persuasive (as requested by a user).
2502    
2503    27. Wrapped the definitions of fileno and isatty for Windows, which appear in
2504        pcretest.c, inside #ifndefs, because it seems they are sometimes already
2505        pre-defined.
2506    
2507    28. Added support for (*UTF8) at the start of a pattern.
2508    
2509    29. Arrange for flags added by the "release type" setting in CMake to be shown
2510        in the configuration summary.
2511    
2512    
2513    Version 7.8 05-Sep-08
2514    ---------------------
2515    
2516    1.  Replaced UCP searching code with optimized version as implemented for Ad
2517        Muncher (http://www.admuncher.com/) by Peter Kankowski. This uses a two-
2518        stage table and inline lookup instead of a function, giving speed ups of 2
2519        to 5 times on some simple patterns that I tested. Permission was given to
2520        distribute the MultiStage2.py script that generates the tables (it's not in
2521        the tarball, but is in the Subversion repository).
2522    
2523    2.  Updated the Unicode datatables to Unicode 5.1.0. This adds yet more
2524        scripts.
2525    
2526    3.  Change 12 for 7.7 introduced a bug in pcre_study() when a pattern contained
2527        a group with a zero qualifier. The result of the study could be incorrect,
2528        or the function might crash, depending on the pattern.
2529    
2530    4.  Caseless matching was not working for non-ASCII characters in back
2531        references. For example, /(\x{de})\1/8i was not matching \x{de}\x{fe}.
2532        It now works when Unicode Property Support is available.
2533    
2534    5.  In pcretest, an escape such as \x{de} in the data was always generating
2535        a UTF-8 string, even in non-UTF-8 mode. Now it generates a single byte in
2536        non-UTF-8 mode. If the value is greater than 255, it gives a warning about
2537        truncation.
2538    
2539    6.  Minor bugfix in pcrecpp.cc (change "" == ... to NULL == ...).
2540    
2541    7.  Added two (int) casts to pcregrep when printing the difference of two
2542        pointers, in case they are 64-bit values.
2543    
2544    8.  Added comments about Mac OS X stack usage to the pcrestack man page and to
2545        test 2 if it fails.
2546    
2547    9.  Added PCRE_CALL_CONVENTION just before the names of all exported functions,
2548        and a #define of that name to empty if it is not externally set. This is to
2549        allow users of MSVC to set it if necessary.
2550    
2551    10. The PCRE_EXP_DEFN macro which precedes exported functions was missing from
2552        the convenience functions in the pcre_get.c source file.
2553    
2554    11. An option change at the start of a pattern that had top-level alternatives
2555        could cause overwriting and/or a crash. This command provoked a crash in
2556        some environments:
2557    
2558          printf "/(?i)[\xc3\xa9\xc3\xbd]|[\xc3\xa9\xc3\xbdA]/8\n" | pcretest
2559    
2560        This potential security problem was recorded as CVE-2008-2371.
2561    
2562    12. For a pattern where the match had to start at the beginning or immediately
2563        after a newline (e.g /.*anything/ without the DOTALL flag), pcre_exec() and
2564        pcre_dfa_exec() could read past the end of the passed subject if there was
2565        no match. To help with detecting such bugs (e.g. with valgrind), I modified
2566        pcretest so that it places the subject at the end of its malloc-ed buffer.
2567    
2568    13. The change to pcretest in 12 above threw up a couple more cases when pcre_
2569        exec() might read past the end of the data buffer in UTF-8 mode.
2570    
2571    14. A similar bug to 7.3/2 existed when the PCRE_FIRSTLINE option was set and
2572        the data contained the byte 0x85 as part of a UTF-8 character within its
2573        first line. This applied both to normal and DFA matching.
2574    
2575    15. Lazy qualifiers were not working in some cases in UTF-8 mode. For example,
2576        /^[^d]*?$/8 failed to match "abc".
2577    
2578    16. Added a missing copyright notice to pcrecpp_internal.h.
2579    
2580    17. Make it more clear in the documentation that values returned from
2581        pcre_exec() in ovector are byte offsets, not character counts.
2582    
2583    18. Tidied a few places to stop certain compilers from issuing warnings.
2584    
2585    19. Updated the Virtual Pascal + BCC files to compile the latest v7.7, as
2586        supplied by Stefan Weber. I made a further small update for 7.8 because
2587        there is a change of source arrangements: the pcre_searchfuncs.c module is
2588        replaced by pcre_ucd.c.
2589    
2590    
2591    Version 7.7 07-May-08
2592    ---------------------
2593    
2594    1.  Applied Craig's patch to sort out a long long problem: "If we can't convert
2595        a string to a long long, pretend we don't even have a long long." This is
2596        done by checking for the strtoq, strtoll, and _strtoi64 functions.
2597    
2598    2.  Applied Craig's patch to pcrecpp.cc to restore ABI compatibility with
2599        pre-7.6 versions, which defined a global no_arg variable instead of putting
2600        it in the RE class. (See also #8 below.)
2601    
2602    3.  Remove a line of dead code, identified by coverity and reported by Nuno
2603        Lopes.
2604    
2605    4.  Fixed two related pcregrep bugs involving -r with --include or --exclude:
2606    
2607        (1) The include/exclude patterns were being applied to the whole pathnames
2608            of files, instead of just to the final components.
2609    
2610        (2) If there was more than one level of directory, the subdirectories were
2611            skipped unless they satisfied the include/exclude conditions. This is
2612            inconsistent with GNU grep (and could even be seen as contrary to the
2613            pcregrep specification - which I improved to make it absolutely clear).
2614            The action now is always to scan all levels of directory, and just
2615            apply the include/exclude patterns to regular files.
2616    
2617    5.  Added the --include_dir and --exclude_dir patterns to pcregrep, and used
2618        --exclude_dir in the tests to avoid scanning .svn directories.
2619    
2620    6.  Applied Craig's patch to the QuoteMeta function so that it escapes the
2621        NUL character as backslash + 0 rather than backslash + NUL, because PCRE
2622        doesn't support NULs in patterns.
2623    
2624    7.  Added some missing "const"s to declarations of static tables in
2625        pcre_compile.c and pcre_dfa_exec.c.
2626    
2627    8.  Applied Craig's patch to pcrecpp.cc to fix a problem in OS X that was
2628        caused by fix #2  above. (Subsequently also a second patch to fix the
2629        first patch. And a third patch - this was a messy problem.)
2630    
2631    9.  Applied Craig's patch to remove the use of push_back().
2632    
2633    10. Applied Alan Lehotsky's patch to add REG_STARTEND support to the POSIX
2634        matching function regexec().
2635    
2636    11. Added support for the Oniguruma syntax \g<name>, \g<n>, \g'name', \g'n',
2637        which, however, unlike Perl's \g{...}, are subroutine calls, not back
2638        references. PCRE supports relative numbers with this syntax (I don't think
2639        Oniguruma does).
2640    
2641    12. Previously, a group with a zero repeat such as (...){0} was completely
2642        omitted from the compiled regex. However, this means that if the group
2643        was called as a subroutine from elsewhere in the pattern, things went wrong
2644        (an internal error was given). Such groups are now left in the compiled
2645        pattern, with a new opcode that causes them to be skipped at execution
2646        time.
2647    
2648    13. Added the PCRE_JAVASCRIPT_COMPAT option. This makes the following changes
2649        to the way PCRE behaves:
2650    
2651        (a) A lone ] character is dis-allowed (Perl treats it as data).
2652    
2653        (b) A back reference to an unmatched subpattern matches an empty string
2654            (Perl fails the current match path).
2655    
2656        (c) A data ] in a character class must be notated as \] because if the
2657            first data character in a class is ], it defines an empty class. (In
2658            Perl it is not possible to have an empty class.) The empty class []
2659            never matches; it forces failure and is equivalent to (*FAIL) or (?!).
2660            The negative empty class [^] matches any one character, independently
2661            of the DOTALL setting.
2662    
2663    14. A pattern such as /(?2)[]a()b](abc)/ which had a forward reference to a
2664        non-existent subpattern following a character class starting with ']' and
2665        containing () gave an internal compiling error instead of "reference to
2666        non-existent subpattern". Fortunately, when the pattern did exist, the
2667        compiled code was correct. (When scanning forwards to check for the
2668        existence of the subpattern, it was treating the data ']' as terminating
2669        the class, so got the count wrong. When actually compiling, the reference
2670        was subsequently set up correctly.)
2671    
2672    15. The "always fail" assertion (?!) is optimzed to (*FAIL) by pcre_compile;
2673        it was being rejected as not supported by pcre_dfa_exec(), even though
2674        other assertions are supported. I have made pcre_dfa_exec() support
2675        (*FAIL).
2676    
2677    16. The implementation of 13c above involved the invention of a new opcode,
2678        OP_ALLANY, which is like OP_ANY but doesn't check the /s flag. Since /s
2679        cannot be changed at match time, I realized I could make a small
2680        improvement to matching performance by compiling OP_ALLANY instead of
2681        OP_ANY for "." when DOTALL was set, and then removing the runtime tests
2682        on the OP_ANY path.
2683    
2684    17. Compiling pcretest on Windows with readline support failed without the
2685        following two fixes: (1) Make the unistd.h include conditional on
2686        HAVE_UNISTD_H; (2) #define isatty and fileno as _isatty and _fileno.
2687    
2688    18. Changed CMakeLists.txt and cmake/FindReadline.cmake to arrange for the
2689        ncurses library to be included for pcretest when ReadLine support is
2690        requested, but also to allow for it to be overridden. This patch came from
2691        Daniel Bergström.
2692    
2693    19. There was a typo in the file ucpinternal.h where f0_rangeflag was defined
2694        as 0x00f00000 instead of 0x00800000. Luckily, this would not have caused
2695        any errors with the current Unicode tables. Thanks to Peter Kankowski for
2696        spotting this.
2697    
2698    
2699    Version 7.6 28-Jan-08
2700    ---------------------
2701    
2702    1.  A character class containing a very large number of characters with
2703        codepoints greater than 255 (in UTF-8 mode, of course) caused a buffer
2704        overflow.
2705    
2706    2.  Patch to cut out the "long long" test in pcrecpp_unittest when
2707        HAVE_LONG_LONG is not defined.
2708    
2709    3.  Applied Christian Ehrlicher's patch to update the CMake build files to
2710        bring them up to date and include new features. This patch includes:
2711    
2712        - Fixed PH's badly added libz and libbz2 support.
2713        - Fixed a problem with static linking.
2714        - Added pcredemo. [But later removed - see 7 below.]
2715        - Fixed dftables problem and added an option.
2716        - Added a number of HAVE_XXX tests, including HAVE_WINDOWS_H and
2717            HAVE_LONG_LONG.
2718        - Added readline support for pcretest.
2719        - Added an listing of the option settings after cmake has run.
2720    
2721    4.  A user submitted a patch to Makefile that makes it easy to create
2722        "pcre.dll" under mingw when using Configure/Make. I added stuff to
2723        Makefile.am that cause it to include this special target, without
2724        affecting anything else. Note that the same mingw target plus all
2725        the other distribution libraries and programs are now supported
2726        when configuring with CMake (see 6 below) instead of with
2727        Configure/Make.
2728    
2729    5.  Applied Craig's patch that moves no_arg into the RE class in the C++ code.
2730        This is an attempt to solve the reported problem "pcrecpp::no_arg is not
2731        exported in the Windows port". It has not yet been confirmed that the patch
2732        solves the problem, but it does no harm.
2733    
2734    6.  Applied Sheri's patch to CMakeLists.txt to add NON_STANDARD_LIB_PREFIX and
2735        NON_STANDARD_LIB_SUFFIX for dll names built with mingw when configured
2736        with CMake, and also correct the comment about stack recursion.
2737    
2738    7.  Remove the automatic building of pcredemo from the ./configure system and
2739        from CMakeLists.txt. The whole idea of pcredemo.c is that it is an example
2740        of a program that users should build themselves after PCRE is installed, so
2741        building it automatically is not really right. What is more, it gave
2742        trouble in some build environments.
2743    
2744    8.  Further tidies to CMakeLists.txt from Sheri and Christian.
2745    
2746    
2747    Version 7.5 10-Jan-08
2748    ---------------------
2749    
2750    1.  Applied a patch from Craig: "This patch makes it possible to 'ignore'
2751        values in parens when parsing an RE using the C++ wrapper."
2752    
2753    2.  Negative specials like \S did not work in character classes in UTF-8 mode.
2754        Characters greater than 255 were excluded from the class instead of being
2755        included.
2756    
2757    3.  The same bug as (2) above applied to negated POSIX classes such as
2758        [:^space:].
2759    
2760    4.  PCRECPP_STATIC was referenced in pcrecpp_internal.h, but nowhere was it
2761        defined or documented. It seems to have been a typo for PCRE_STATIC, so
2762        I have changed it.
2763    
2764    5.  The construct (?&) was not diagnosed as a syntax error (it referenced the
2765        first named subpattern) and a construct such as (?&a) would reference the
2766        first named subpattern whose name started with "a" (in other words, the
2767        length check was missing). Both these problems are fixed. "Subpattern name
2768        expected" is now given for (?&) (a zero-length name), and this patch also
2769        makes it give the same error for \k'' (previously it complained that that
2770        was a reference to a non-existent subpattern).
2771    
2772    6.  The erroneous patterns (?+-a) and (?-+a) give different error messages;
2773        this is right because (?- can be followed by option settings as well as by
2774        digits. I have, however, made the messages clearer.
2775    
2776    7.  Patterns such as (?(1)a|b) (a pattern that contains fewer subpatterns
2777        than the number used in the conditional) now cause a compile-time error.
2778        This is actually not compatible with Perl, which accepts such patterns, but
2779        treats the conditional as always being FALSE (as PCRE used to), but it
2780        seems to me that giving a diagnostic is better.
2781    
2782    8.  Change "alphameric" to the more common word "alphanumeric" in comments
2783        and messages.
2784    
2785    9.  Fix two occurrences of "backslash" in comments that should have been
2786        "backspace".
2787    
2788    10. Remove two redundant lines of code that can never be obeyed (their function
2789        was moved elsewhere).
2790    
2791    11. The program that makes PCRE's Unicode character property table had a bug
2792        which caused it to generate incorrect table entries for sequences of
2793        characters that have the same character type, but are in different scripts.
2794        It amalgamated them into a single range, with the script of the first of
2795        them. In other words, some characters were in the wrong script. There were
2796        thirteen such cases, affecting characters in the following ranges:
2797    
2798          U+002b0 - U+002c1
2799          U+0060c - U+0060d
2800          U+0061e - U+00612
2801          U+0064b - U+0065e
2802          U+0074d - U+0076d
2803          U+01800 - U+01805
2804          U+01d00 - U+01d77
2805          U+01d9b - U+01dbf
2806          U+0200b - U+0200f
2807          U+030fc - U+030fe
2808          U+03260 - U+0327f
2809          U+0fb46 - U+0fbb1
2810          U+10450 - U+1049d
2811    
2812    12. The -o option (show only the matching part of a line) for pcregrep was not
2813        compatible with GNU grep in that, if there was more than one match in a
2814        line, it showed only the first of them. It now behaves in the same way as
2815        GNU grep.
2816    
2817    13. If the -o and -v options were combined for pcregrep, it printed a blank
2818        line for every non-matching line. GNU grep prints nothing, and pcregrep now
2819        does the same. The return code can be used to tell if there were any
2820        non-matching lines.
2821    
2822    14. Added --file-offsets and --line-offsets to pcregrep.
2823    
2824    15. The pattern (?=something)(?R) was not being diagnosed as a potentially
2825        infinitely looping recursion. The bug was that positive lookaheads were not
2826        being skipped when checking for a possible empty match (negative lookaheads
2827        and both kinds of lookbehind were skipped).
2828    
2829    16. Fixed two typos in the Windows-only code in pcregrep.c, and moved the
2830        inclusion of <windows.h> to before rather than after the definition of
2831        INVALID_FILE_ATTRIBUTES (patch from David Byron).
2832    
2833    17. Specifying a possessive quantifier with a specific limit for a Unicode
2834        character property caused pcre_compile() to compile bad code, which led at
2835        runtime to PCRE_ERROR_INTERNAL (-14). Examples of patterns that caused this
2836        are: /\p{Zl}{2,3}+/8 and /\p{Cc}{2}+/8. It was the possessive "+" that
2837        caused the error; without that there was no problem.
2838    
2839    18. Added --enable-pcregrep-libz and --enable-pcregrep-libbz2.
2840    
2841    19. Added --enable-pcretest-libreadline.
2842    
2843    20. In pcrecpp.cc, the variable 'count' was incremented twice in
2844        RE::GlobalReplace(). As a result, the number of replacements returned was
2845        double what it should be. I removed one of the increments, but Craig sent a
2846        later patch that removed the other one (the right fix) and added unit tests
2847        that check the return values (which was not done before).
2848    
2849    21. Several CMake things:
2850    
2851        (1) Arranged that, when cmake is used on Unix, the libraries end up with
2852            the names libpcre and libpcreposix, not just pcre and pcreposix.
2853    
2854        (2) The above change means that pcretest and pcregrep are now correctly
2855            linked with the newly-built libraries, not previously installed ones.
2856    
2857        (3) Added PCRE_SUPPORT_LIBREADLINE, PCRE_SUPPORT_LIBZ, PCRE_SUPPORT_LIBBZ2.
2858    
2859    22. In UTF-8 mode, with newline set to "any", a pattern such as .*a.*=.b.*
2860        crashed when matching a string such as a\x{2029}b (note that \x{2029} is a
2861        UTF-8 newline character). The key issue is that the pattern starts .*;
2862        this means that the match must be either at the beginning, or after a
2863        newline. The bug was in the code for advancing after a failed match and
2864        checking that the new position followed a newline. It was not taking
2865        account of UTF-8 characters correctly.
2866    
2867    23. PCRE was behaving differently from Perl in the way it recognized POSIX
2868        character classes. PCRE was not treating the sequence [:...:] as a
2869        character class unless the ... were all letters. Perl, however, seems to
2870        allow any characters between [: and :], though of course it rejects as
2871        unknown any "names" that contain non-letters, because all the known class
2872        names consist only of letters. Thus, Perl gives an error for [[:1234:]],
2873        for example, whereas PCRE did not - it did not recognize a POSIX character
2874        class. This seemed a bit dangerous, so the code has been changed to be
2875        closer to Perl. The behaviour is not identical to Perl, because PCRE will
2876        diagnose an unknown class for, for example, [[:l\ower:]] where Perl will
2877        treat it as [[:lower:]]. However, PCRE does now give "unknown" errors where
2878        Perl does, and where it didn't before.
2879    
2880    24. Rewrite so as to remove the single use of %n from pcregrep because in some
2881        Windows environments %n is disabled by default.
2882    
2883    
2884    Version 7.4 21-Sep-07
2885    ---------------------
2886    
2887    1.  Change 7.3/28 was implemented for classes by looking at the bitmap. This
2888        means that a class such as [\s] counted as "explicit reference to CR or
2889        LF". That isn't really right - the whole point of the change was to try to
2890        help when there was an actual mention of one of the two characters. So now
2891        the change happens only if \r or \n (or a literal CR or LF) character is
2892        encountered.
2893    
2894    2.  The 32-bit options word was also used for 6 internal flags, but the numbers
2895        of both had grown to the point where there were only 3 bits left.
2896        Fortunately, there was spare space in the data structure, and so I have
2897        moved the internal flags into a new 16-bit field to free up more option
2898        bits.
2899    
2900    3.  The appearance of (?J) at the start of a pattern set the DUPNAMES option,
2901        but did not set the internal JCHANGED flag - either of these is enough to
2902        control the way the "get" function works - but the PCRE_INFO_JCHANGED
2903        facility is supposed to tell if (?J) was ever used, so now (?J) at the
2904        start sets both bits.
2905    
2906    4.  Added options (at build time, compile time, exec time) to change \R from
2907        matching any Unicode line ending sequence to just matching CR, LF, or CRLF.
2908    
2909    5.  doc/pcresyntax.html was missing from the distribution.
2910    
2911    6.  Put back the definition of PCRE_ERROR_NULLWSLIMIT, for backward
2912        compatibility, even though it is no longer used.
2913    
2914    7.  Added macro for snprintf to pcrecpp_unittest.cc and also for strtoll and
2915        strtoull to pcrecpp.cc to select the available functions in WIN32 when the
2916        windows.h file is present (where different names are used). [This was
2917        reversed later after testing - see 16 below.]
2918    
2919    8.  Changed all #include <config.h> to #include "config.h". There were also
2920        some further <pcre.h> cases that I changed to "pcre.h".
2921    
2922    9.  When pcregrep was used with the --colour option, it missed the line ending
2923        sequence off the lines that it output.
2924    
2925    10. It was pointed out to me that arrays of string pointers cause lots of
2926        relocations when a shared library is dynamically loaded. A technique of
2927        using a single long string with a table of offsets can drastically reduce
2928        these. I have refactored PCRE in four places to do this. The result is
2929        dramatic:
2930    
2931          Originally:                          290
2932          After changing UCP table:            187
2933          After changing error message table:   43
2934          After changing table of "verbs"       36
2935          After changing table of Posix names   22
2936    
2937        Thanks to the folks working on Gregex for glib for this insight.
2938    
2939    11. --disable-stack-for-recursion caused compiling to fail unless -enable-
2940        unicode-properties was also set.
2941    
2942    12. Updated the tests so that they work when \R is defaulted to ANYCRLF.
2943    
2944    13. Added checks for ANY and ANYCRLF to pcrecpp.cc where it previously
2945        checked only for CRLF.
2946    
2947    14. Added casts to pcretest.c to avoid compiler warnings.
2948    
2949    15. Added Craig's patch to various pcrecpp modules to avoid compiler warnings.
2950    
2951    16. Added Craig's patch to remove the WINDOWS_H tests, that were not working,
2952        and instead check for _strtoi64 explicitly, and avoid the use of snprintf()
2953        entirely. This removes changes made in 7 above.
2954    
2955    17. The CMake files have been updated, and there is now more information about
2956        building with CMake in the NON-UNIX-USE document.
2957    
2958    
2959    Version 7.3 28-Aug-07
2960    ---------------------
2961    
2962     1. In the rejigging of the build system that eventually resulted in 7.1, the
2963        line "#include <pcre.h>" was included in pcre_internal.h. The use of angle
2964        brackets there is not right, since it causes compilers to look for an
2965        installed pcre.h, not the version that is in the source that is being
2966        compiled (which of course may be different). I have changed it back to:
2967    
2968          #include "pcre.h"
2969    
2970        I have a vague recollection that the change was concerned with compiling in
2971        different directories, but in the new build system, that is taken care of
2972        by the VPATH setting the Makefile.
2973    
2974     2. The pattern .*$ when run in not-DOTALL UTF-8 mode with newline=any failed
2975        when the subject happened to end in the byte 0x85 (e.g. if the last
2976        character was \x{1ec5}). *Character* 0x85 is one of the "any" newline
2977        characters but of course it shouldn't be taken as a newline when it is part
2978        of another character. The bug was that, for an unlimited repeat of . in
2979        not-DOTALL UTF-8 mode, PCRE was advancing by bytes rather than by
2980        characters when looking for a newline.
2981    
2982     3. A small performance improvement in the DOTALL UTF-8 mode .* case.
2983    
2984     4. Debugging: adjusted the names of opcodes for different kinds of parentheses
2985        in debug output.
2986    
2987     5. Arrange to use "%I64d" instead of "%lld" and "%I64u" instead of "%llu" for
2988        long printing in the pcrecpp unittest when running under MinGW.
2989    
2990     6. ESC_K was left out of the EBCDIC table.
2991    
2992     7. Change 7.0/38 introduced a new limit on the number of nested non-capturing
2993        parentheses; I made it 1000, which seemed large enough. Unfortunately, the
2994        limit also applies to "virtual nesting" when a pattern is recursive, and in
2995        this case 1000 isn't so big. I have been able to remove this limit at the
2996        expense of backing off one optimization in certain circumstances. Normally,
2997        when pcre_exec() would call its internal match() function recursively and
2998        immediately return the result unconditionally, it uses a "tail recursion"
2999        feature to save stack. However, when a subpattern that can match an empty
3000        string has an unlimited repetition quantifier, it no longer makes this
3001        optimization. That gives it a stack frame in which to save the data for
3002        checking that an empty string has been matched. Previously this was taken
3003        from the 1000-entry workspace that had been reserved. So now there is no
3004        explicit limit, but more stack is used.
3005    
3006     8. Applied Daniel's patches to solve problems with the import/export magic
3007        syntax that is required for Windows, and which was going wrong for the
3008        pcreposix and pcrecpp parts of the library. These were overlooked when this
3009        problem was solved for the main library.
3010    
3011     9. There were some crude static tests to avoid integer overflow when computing
3012        the size of patterns that contain repeated groups with explicit upper
3013        limits. As the maximum quantifier is 65535, the maximum group length was
3014        set at 30,000 so that the product of these two numbers did not overflow a
3015        32-bit integer. However, it turns out that people want to use groups that
3016        are longer than 30,000 bytes (though not repeat them that many times).
3017        Change 7.0/17 (the refactoring of the way the pattern size is computed) has
3018        made it possible to implement the integer overflow checks in a much more
3019        dynamic way, which I have now done. The artificial limitation on group
3020        length has been removed - we now have only the limit on the total length of
3021        the compiled pattern, which depends on the LINK_SIZE setting.
3022    
3023    10. Fixed a bug in the documentation for get/copy named substring when
3024        duplicate names are permitted. If none of the named substrings are set, the
3025        functions return PCRE_ERROR_NOSUBSTRING (7); the doc said they returned an
3026        empty string.
3027    
3028    11. Because Perl interprets \Q...\E at a high level, and ignores orphan \E
3029        instances, patterns such as [\Q\E] or [\E] or even [^\E] cause an error,
3030        because the ] is interpreted as the first data character and the
3031        terminating ] is not found. PCRE has been made compatible with Perl in this
3032        regard. Previously, it interpreted [\Q\E] as an empty class, and [\E] could
3033        cause memory overwriting.
3034    
3035    10. Like Perl, PCRE automatically breaks an unlimited repeat after an empty
3036        string has been matched (to stop an infinite loop). It was not recognizing
3037        a conditional subpattern that could match an empty string if that
3038        subpattern was within another subpattern. For example, it looped when
3039        trying to match  (((?(1)X|))*)  but it was OK with  ((?(1)X|)*)  where the
3040        condition was not nested. This bug has been fixed.
3041    
3042    12. A pattern like \X?\d or \P{L}?\d in non-UTF-8 mode could cause a backtrack
3043        past the start of the subject in the presence of bytes with the top bit
3044        set, for example "\x8aBCD".
3045    
3046    13. Added Perl 5.10 experimental backtracking controls (*FAIL), (*F), (*PRUNE),
3047        (*SKIP), (*THEN), (*COMMIT), and (*ACCEPT).
3048    
3049    14. Optimized (?!) to (*FAIL).
3050    
3051    15. Updated the test for a valid UTF-8 string to conform to the later RFC 3629.
3052        This restricts code points to be within the range 0 to 0x10FFFF, excluding
3053        the "low surrogate" sequence 0xD800 to 0xDFFF. Previously, PCRE allowed the
3054        full range 0 to 0x7FFFFFFF, as defined by RFC 2279. Internally, it still
3055        does: it's just the validity check that is more restrictive.
3056    
3057    16. Inserted checks for integer overflows during escape sequence (backslash)
3058        processing, and also fixed erroneous offset values for syntax errors during
3059        backslash processing.
3060    
3061    17. Fixed another case of looking too far back in non-UTF-8 mode (cf 12 above)
3062        for patterns like [\PPP\x8a]{1,}\x80 with the subject "A\x80".
3063    
3064    18. An unterminated class in a pattern like (?1)\c[ with a "forward reference"
3065        caused an overrun.
3066    
3067    19. A pattern like (?:[\PPa*]*){8,} which had an "extended class" (one with
3068        something other than just ASCII characters) inside a group that had an
3069        unlimited repeat caused a loop at compile time (while checking to see
3070        whether the group could match an empty string).
3071    
3072    20. Debugging a pattern containing \p or \P could cause a crash. For example,
3073        [\P{Any}] did so. (Error in the code for printing property names.)
3074    
3075    21. An orphan \E inside a character class could cause a crash.
3076    
3077    22. A repeated capturing bracket such as (A)? could cause a wild memory
3078        reference during compilation.
3079    
3080    23. There are several functions in pcre_compile() that scan along a compiled
3081        expression for various reasons (e.g. to see if it's fixed length for look
3082        behind). There were bugs in these functions when a repeated \p or \P was
3083        present in the pattern. These operators have additional parameters compared
3084        with \d, etc, and these were not being taken into account when moving along
3085        the compiled data. Specifically:
3086    
3087        (a) A item such as \p{Yi}{3} in a lookbehind was not treated as fixed
3088            length.
3089    
3090        (b) An item such as \pL+ within a repeated group could cause crashes or
3091            loops.
3092    
3093        (c) A pattern such as \p{Yi}+(\P{Yi}+)(?1) could give an incorrect
3094            "reference to non-existent subpattern" error.
3095    
3096        (d) A pattern like (\P{Yi}{2}\277)? could loop at compile time.
3097    
3098    24. A repeated \S or \W in UTF-8 mode could give wrong answers when multibyte
3099        characters were involved (for example /\S{2}/8g with "A\x{a3}BC").
3100    
3101    25. Using pcregrep in multiline, inverted mode (-Mv) caused it to loop.
3102    
3103    26. Patterns such as [\P{Yi}A] which include \p or \P and just one other
3104        character were causing crashes (broken optimization).
3105    
3106    27. Patterns such as (\P{Yi}*\277)* (group with possible zero repeat containing
3107        \p or \P) caused a compile-time loop.
3108    
3109    28. More problems have arisen in unanchored patterns when CRLF is a valid line
3110        break. For example, the unstudied pattern [\r\n]A does not match the string
3111        "\r\nA" because change 7.0/46 below moves the current point on by two
3112        characters after failing to match at the start. However, the pattern \nA
3113        *does* match, because it doesn't start till \n, and if [\r\n]A is studied,
3114        the same is true. There doesn't seem any very clean way out of this, but
3115        what I have chosen to do makes the common cases work: PCRE now takes note
3116        of whether there can be an explicit match for \r or \n anywhere in the
3117        pattern, and if so, 7.0/46 no longer applies. As part of this change,
3118        there's a new PCRE_INFO_HASCRORLF option for finding out whether a compiled
3119        pattern has explicit CR or LF references.
3120    
3121    29. Added (*CR) etc for changing newline setting at start of pattern.
3122    
3123    
3124    Version 7.2 19-Jun-07
3125    ---------------------
3126    
3127     1. If the fr_FR locale cannot be found for test 3, try the "french" locale,
3128        which is apparently normally available under Windows.
3129    
3130     2. Re-jig the pcregrep tests with different newline settings in an attempt
3131        to make them independent of the local environment's newline setting.
3132    
3133     3. Add code to configure.ac to remove -g from the CFLAGS default settings.
3134    
3135     4. Some of the "internals" tests were previously cut out when the link size
3136        was not 2, because the output contained actual offsets. The recent new
3137        "Z" feature of pcretest means that these can be cut out, making the tests
3138        usable with all link sizes.
3139    
3140     5. Implemented Stan Switzer's goto replacement for longjmp() when not using
3141        stack recursion. This gives a massive performance boost under BSD, but just
3142        a small improvement under Linux. However, it saves one field in the frame
3143        in all cases.
3144    
3145     6. Added more features from the forthcoming Perl 5.10:
3146    
3147        (a) (?-n) (where n is a string of digits) is a relative subroutine or
3148            recursion call. It refers to the nth most recently opened parentheses.
3149    
3150        (b) (?+n) is also a relative subroutine call; it refers to the nth next
3151            to be opened parentheses.
3152    
3153        (c) Conditions that refer to capturing parentheses can be specified
3154            relatively, for example, (?(-2)... or (?(+3)...
3155    
3156        (d) \K resets the start of the current match so that everything before
3157            is not part of it.
3158    
3159        (e) \k{name} is synonymous with \k<name> and \k'name' (.NET compatible).
3160    
3161        (f) \g{name} is another synonym - part of Perl 5.10's unification of
3162            reference syntax.
3163    
3164        (g) (?| introduces a group in which the numbering of parentheses in each
3165            alternative starts with the same number.
3166    
3167        (h) \h, \H, \v, and \V match horizontal and vertical whitespace.
3168    
3169     7. Added two new calls to pcre_fullinfo(): PCRE_INFO_OKPARTIAL and
3170        PCRE_INFO_JCHANGED.
3171    
3172     8. A pattern such as  (.*(.)?)*  caused pcre_exec() to fail by either not
3173        terminating or by crashing. Diagnosed by Viktor Griph; it was in the code
3174        for detecting groups that can match an empty string.
3175    
3176     9. A pattern with a very large number of alternatives (more than several
3177        hundred) was running out of internal workspace during the pre-compile
3178        phase, where pcre_compile() figures out how much memory will be needed. A
3179        bit of new cunning has reduced the workspace needed for groups with
3180        alternatives. The 1000-alternative test pattern now uses 12 bytes of
3181        workspace instead of running out of the 4096 that are available.
3182    
3183    10. Inserted some missing (unsigned int) casts to get rid of compiler warnings.
3184    
3185    11. Applied patch from Google to remove an optimization that didn't quite work.
3186        The report of the bug said:
3187    
3188          pcrecpp::RE("a*").FullMatch("aaa") matches, while
3189          pcrecpp::RE("a*?").FullMatch("aaa") does not, and
3190          pcrecpp::RE("a*?\\z").FullMatch("aaa") does again.
3191    
3192    12. If \p or \P was used in non-UTF-8 mode on a character greater than 127
3193        it matched the wrong number of bytes.
3194    
3195    
3196    Version 7.1 24-Apr-07
3197    ---------------------
3198    
3199     1. Applied Bob Rossi and Daniel G's patches to convert the build system to one
3200        that is more "standard", making use of automake and other Autotools. There
3201        is some re-arrangement of the files and adjustment of comments consequent
3202        on this.
3203    
3204     2. Part of the patch fixed a problem with the pcregrep tests. The test of -r
3205        for recursive directory scanning broke on some systems because the files
3206        are not scanned in any specific order and on different systems the order
3207        was different. A call to "sort" has been inserted into RunGrepTest for the
3208        approprate test as a short-term fix. In the longer term there may be an
3209        alternative.
3210    
3211     3. I had an email from Eric Raymond about problems translating some of PCRE's
3212        man pages to HTML (despite the fact that I distribute HTML pages, some
3213        people do their own conversions for various reasons). The problems
3214        concerned the use of low-level troff macros .br and .in. I have therefore
3215        removed all such uses from the man pages (some were redundant, some could
3216        be replaced by .nf/.fi pairs). The 132html script that I use to generate
3217        HTML has been updated to handle .nf/.fi and to complain if it encounters
3218        .br or .in.
3219    
3220     4. Updated comments in configure.ac that get placed in config.h.in and also
3221        arranged for config.h to be included in the distribution, with the name
3222        config.h.generic, for the benefit of those who have to compile without
3223        Autotools (compare pcre.h, which is now distributed as pcre.h.generic).
3224    
3225     5. Updated the support (such as it is) for Virtual Pascal, thanks to Stefan
3226        Weber: (1) pcre_internal.h was missing some function renames; (2) updated
3227        makevp.bat for the current PCRE, using the additional files
3228        makevp_c.txt, makevp_l.txt, and pcregexp.pas.
3229    
3230     6. A Windows user reported a minor discrepancy with test 2, which turned out
3231        to be caused by a trailing space on an input line that had got lost in his
3232        copy. The trailing space was an accident, so I've just removed it.
3233    
3234     7. Add -Wl,-R... flags in pcre-config.in for *BSD* systems, as I'm told
3235        that is needed.
3236    
3237     8. Mark ucp_table (in ucptable.h) and ucp_gentype (in pcre_ucp_searchfuncs.c)
3238        as "const" (a) because they are and (b) because it helps the PHP
3239        maintainers who have recently made a script to detect big data structures
3240        in the php code that should be moved to the .rodata section. I remembered
3241        to update Builducptable as well, so it won't revert if ucptable.h is ever
3242        re-created.
3243    
3244     9. Added some extra #ifdef SUPPORT_UTF8 conditionals into pcretest.c,
3245        pcre_printint.src, pcre_compile.c, pcre_study.c, and pcre_tables.c, in
3246        order to be able to cut out the UTF-8 tables in the latter when UTF-8
3247        support is not required. This saves 1.5-2K of code, which is important in
3248        some applications.
3249    
3250        Later: more #ifdefs are needed in pcre_ord2utf8.c and pcre_valid_utf8.c
3251        so as not to refer to the tables, even though these functions will never be
3252        called when UTF-8 support is disabled. Otherwise there are problems with a
3253        shared library.
3254    
3255    10. Fixed two bugs in the emulated memmove() function in pcre_internal.h:
3256    
3257        (a) It was defining its arguments as char * instead of void *.
3258    
3259        (b) It was assuming that all moves were upwards in memory; this was true
3260            a long time ago when I wrote it, but is no longer the case.
3261    
3262        The emulated memove() is provided for those environments that have neither
3263        memmove() nor bcopy(). I didn't think anyone used it these days, but that
3264        is clearly not the case, as these two bugs were recently reported.
3265    
3266    11. The script PrepareRelease is now distributed: it calls 132html, CleanTxt,
3267        and Detrail to create the HTML documentation, the .txt form of the man
3268        pages, and it removes trailing spaces from listed files. It also creates
3269        pcre.h.generic and config.h.generic from pcre.h and config.h. In the latter
3270        case, it wraps all the #defines with #ifndefs. This script should be run
3271        before "make dist".
3272    
3273    12. Fixed two fairly obscure bugs concerned with quantified caseless matching
3274        with Unicode property support.
3275    
3276        (a) For a maximizing quantifier, if the two different cases of the
3277            character were of different lengths in their UTF-8 codings (there are
3278            some cases like this - I found 11), and the matching function had to
3279            back up over a mixture of the two cases, it incorrectly assumed they
3280            were both the same length.
3281    
3282        (b) When PCRE was configured to use the heap rather than the stack for
3283            recursion during matching, it was not correctly preserving the data for
3284            the other case of a UTF-8 character when checking ahead for a match
3285            while processing a minimizing repeat. If the check also involved
3286            matching a wide character, but failed, corruption could cause an
3287            erroneous result when trying to check for a repeat of the original
3288            character.
3289    
3290    13. Some tidying changes to the testing mechanism:
3291    
3292        (a) The RunTest script now detects the internal link size and whether there
3293            is UTF-8 and UCP support by running ./pcretest -C instead of relying on
3294            values substituted by "configure". (The RunGrepTest script already did
3295            this for UTF-8.) The configure.ac script no longer substitutes the
3296            relevant variables.
3297    
3298        (b) The debugging options /B and /D in pcretest show the compiled bytecode
3299            with length and offset values. This means that the output is different
3300            for different internal link sizes. Test 2 is skipped for link sizes
3301            other than 2 because of this, bypassing the problem. Unfortunately,
3302            there was also a test in test 3 (the locale tests) that used /B and
3303            failed for link sizes other than 2. Rather than cut the whole test out,
3304            I have added a new /Z option to pcretest that replaces the length and
3305            offset values with spaces. This is now used to make test 3 independent
3306            of link size. (Test 2 will be tidied up later.)
3307    
3308    14. If erroroffset was passed as NULL to pcre_compile, it provoked a
3309        segmentation fault instead of returning the appropriate error message.
3310    
3311    15. In multiline mode when the newline sequence was set to "any", the pattern
3312        ^$ would give a match between the \r and \n of a subject such as "A\r\nB".
3313        This doesn't seem right; it now treats the CRLF combination as the line
3314        ending, and so does not match in that case. It's only a pattern such as ^$
3315        that would hit this one: something like ^ABC$ would have failed after \r
3316        and then tried again after \r\n.
3317    
3318    16. Changed the comparison command for RunGrepTest from "diff -u" to "diff -ub"
3319        in an attempt to make files that differ only in their line terminators
3320        compare equal. This works on Linux.
3321    
3322    17. Under certain error circumstances pcregrep might try to free random memory
3323        as it exited. This is now fixed, thanks to valgrind.
3324    
3325    19. In pcretest, if the pattern /(?m)^$/g<any> was matched against the string
3326        "abc\r\n\r\n", it found an unwanted second match after the second \r. This
3327        was because its rules for how to advance for /g after matching an empty
3328        string at the end of a line did not allow for this case. They now check for
3329        it specially.
3330    
3331    20. pcretest is supposed to handle patterns and data of any length, by
3332        extending its buffers when necessary. It was getting this wrong when the
3333        buffer for a data line had to be extended.
3334    
3335    21. Added PCRE_NEWLINE_ANYCRLF which is like ANY, but matches only CR, LF, or
3336        CRLF as a newline sequence.
3337    
3338    22. Code for handling Unicode properties in pcre_dfa_exec() wasn't being cut
3339        out by #ifdef SUPPORT_UCP. This did no harm, as it could never be used, but
3340        I have nevertheless tidied it up.
3341    
3342    23. Added some casts to kill warnings from HP-UX ia64 compiler.
3343    
3344    24. Added a man page for pcre-config.
3345    
3346    
3347    Version 7.0 19-Dec-06
3348    ---------------------
3349    
3350     1. Fixed a signed/unsigned compiler warning in pcre_compile.c, shown up by
3351        moving to gcc 4.1.1.
3352    
3353     2. The -S option for pcretest uses setrlimit(); I had omitted to #include
3354        sys/time.h, which is documented as needed for this function. It doesn't
3355        seem to matter on Linux, but it showed up on some releases of OS X.
3356    
3357     3. It seems that there are systems where bytes whose values are greater than
3358        127 match isprint() in the "C" locale. The "C" locale should be the
3359        default when a C program starts up. In most systems, only ASCII printing
3360        characters match isprint(). This difference caused the output from pcretest
3361        to vary, making some of the tests fail. I have changed pcretest so that:
3362    
3363        (a) When it is outputting text in the compiled version of a pattern, bytes
3364            other than 32-126 are always shown as hex escapes.
3365    
3366        (b) When it is outputting text that is a matched part of a subject string,
3367            it does the same, unless a different locale has been set for the match
3368            (using the /L modifier). In this case, it uses isprint() to decide.
3369    
3370     4. Fixed a major bug that caused incorrect computation of the amount of memory
3371        required for a compiled pattern when options that changed within the
3372        pattern affected the logic of the preliminary scan that determines the
3373        length. The relevant options are -x, and -i in UTF-8 mode. The result was
3374        that the computed length was too small. The symptoms of this bug were
3375        either the PCRE error "internal error: code overflow" from pcre_compile(),
3376        or a glibc crash with a message such as "pcretest: free(): invalid next
3377        size (fast)". Examples of patterns that provoked this bug (shown in
3378        pcretest format) are:
3379    
3380          /(?-x: )/x
3381          /(?x)(?-x: \s*#\s*)/
3382          /((?i)[\x{c0}])/8
3383          /(?i:[\x{c0}])/8
3384    
3385        HOWEVER: Change 17 below makes this fix obsolete as the memory computation
3386        is now done differently.
3387    
3388     5. Applied patches from Google to: (a) add a QuoteMeta function to the C++
3389        wrapper classes; (b) implement a new function in the C++ scanner that is
3390        more efficient than the old way of doing things because it avoids levels of
3391        recursion in the regex matching; (c) add a paragraph to the documentation
3392        for the FullMatch() function.
3393    
3394     6. The escape sequence \n was being treated as whatever was defined as
3395        "newline". Not only was this contrary to the documentation, which states
3396        that \n is character 10 (hex 0A), but it also went horribly wrong when
3397        "newline" was defined as CRLF. This has been fixed.
3398    
3399     7. In pcre_dfa_exec.c the value of an unsigned integer (the variable called c)
3400        was being set to -1 for the "end of line" case (supposedly a value that no
3401        character can have). Though this value is never used (the check for end of
3402        line is "zero bytes in current character"), it caused compiler complaints.
3403        I've changed it to 0xffffffff.
3404    
3405     8. In pcre_version.c, the version string was being built by a sequence of
3406        C macros that, in the event of PCRE_PRERELEASE being defined as an empty
3407        string (as it is for production releases) called a macro with an empty
3408        argument. The C standard says the result of this is undefined. The gcc
3409        compiler treats it as an empty string (which was what was wanted) but it is
3410        reported that Visual C gives an error. The source has been hacked around to
3411        avoid this problem.
3412    
3413     9. On the advice of a Windows user, included <io.h> and <fcntl.h> in Windows
3414        builds of pcretest, and changed the call to _setmode() to use _O_BINARY
3415        instead of 0x8000. Made all the #ifdefs test both _WIN32 and WIN32 (not all
3416        of them did).
3417    
3418    10. Originally, pcretest opened its input and output without "b"; then I was
3419        told that "b" was needed in some environments, so it was added for release
3420        5.0 to both the input and output. (It makes no difference on Unix-like
3421        systems.) Later I was told that it is wrong for the input on Windows. I've
3422        now abstracted the modes into two macros, to make it easier to fiddle with
3423        them, and removed "b" from the input mode under Windows.
3424    
3425    11. Added pkgconfig support for the C++ wrapper library, libpcrecpp.
3426    
3427    12. Added -help and --help to pcretest as an official way of being reminded
3428        of the options.
3429    
3430    13. Removed some redundant semicolons after macro calls in pcrecpparg.h.in
3431        and pcrecpp.cc because they annoy compilers at high warning levels.
3432    
3433    14. A bit of tidying/refactoring in pcre_exec.c in the main bumpalong loop.
3434    
3435    15. Fixed an occurrence of == in configure.ac that should have been = (shell
3436        scripts are not C programs :-) and which was not noticed because it works
3437        on Linux.
3438    
3439    16. pcretest is supposed to handle any length of pattern and data line (as one
3440        line or as a continued sequence of lines) by extending its input buffer if
3441        necessary. This feature was broken for very long pattern lines, leading to
3442        a string of junk being passed to pcre_compile() if the pattern was longer
3443        than about 50K.
3444    
3445    17. I have done a major re-factoring of the way pcre_compile() computes the
3446        amount of memory needed for a compiled pattern. Previously, there was code
3447        that made a preliminary scan of the pattern in order to do this. That was
3448        OK when PCRE was new, but as the facilities have expanded, it has become
3449        harder and harder to keep it in step with the real compile phase, and there
3450        have been a number of bugs (see for example, 4 above). I have now found a
3451        cunning way of running the real compile function in a "fake" mode that
3452        enables it to compute how much memory it would need, while actually only
3453        ever using a few hundred bytes of working memory and without too many
3454        tests of the mode. This should make future maintenance and development
3455        easier. A side effect of this work is that the limit of 200 on the nesting
3456        depth of parentheses has been removed (though this was never a serious
3457        limitation, I suspect). However, there is a downside: pcre_compile() now
3458        runs more slowly than before (30% or more, depending on the pattern). I
3459        hope this isn't a big issue. There is no effect on runtime performance.
3460    
3461    18. Fixed a minor bug in pcretest: if a pattern line was not terminated by a
3462        newline (only possible for the last line of a file) and it was a
3463        pattern that set a locale (followed by /Lsomething), pcretest crashed.
3464    
3465    19. Added additional timing features to pcretest. (1) The -tm option now times
3466        matching only, not compiling. (2) Both -t and -tm can be followed, as a
3467        separate command line item, by a number that specifies the number of
3468        repeats to use when timing. The default is 50000; this gives better
3469        precision, but takes uncomfortably long for very large patterns.
3470    
3471    20. Extended pcre_study() to be more clever in cases where a branch of a
3472        subpattern has no definite first character. For example, (a*|b*)[cd] would
3473        previously give no result from pcre_study(). Now it recognizes that the
3474        first character must be a, b, c, or d.
3475    
3476    21. There was an incorrect error "recursive call could loop indefinitely" if
3477        a subpattern (or the entire pattern) that was being tested for matching an
3478        empty string contained only one non-empty item after a nested subpattern.
3479        For example, the pattern (?>\x{100}*)\d(?R) provoked this error
3480        incorrectly, because the \d was being skipped in the check.
3481    
3482    22. The pcretest program now has a new pattern option /B and a command line
3483        option -b, which is equivalent to adding /B to every pattern. This causes
3484        it to show the compiled bytecode, without the additional information that
3485        -d shows. The effect of -d is now the same as -b with -i (and similarly, /D
3486        is the same as /B/I).
3487    
3488    23. A new optimization is now able automatically to treat some sequences such
3489        as a*b as a*+b. More specifically, if something simple (such as a character
3490        or a simple class like \d) has an unlimited quantifier, and is followed by
3491        something that cannot possibly match the quantified thing, the quantifier
3492        is automatically "possessified".
3493    
3494    24. A recursive reference to a subpattern whose number was greater than 39
3495        went wrong under certain circumstances in UTF-8 mode. This bug could also
3496        have affected the operation of pcre_study().
3497    
3498    25. Realized that a little bit of performance could be had by replacing
3499        (c & 0xc0) == 0xc0 with c >= 0xc0 when processing UTF-8 characters.
3500    
3501    26. Timing data from pcretest is now shown to 4 decimal places instead of 3.
3502    
3503    27. Possessive quantifiers such as a++ were previously implemented by turning
3504        them into atomic groups such as ($>a+). Now they have their own opcodes,
3505        which improves performance. This includes the automatically created ones
3506        from 23 above.
3507    
3508    28. A pattern such as (?=(\w+))\1: which simulates an atomic group using a
3509        lookahead was broken if it was not anchored. PCRE was mistakenly expecting
3510        the first matched character to be a colon. This applied both to named and
3511        numbered groups.
3512    
3513    29. The ucpinternal.h header file was missing its idempotency #ifdef.
3514    
3515    30. I was sent a "project" file called libpcre.a.dev which I understand makes
3516        building PCRE on Windows easier, so I have included it in the distribution.
3517    
3518    31. There is now a check in pcretest against a ridiculously large number being
3519        returned by pcre_exec() or pcre_dfa_exec(). If this happens in a /g or /G
3520        loop, the loop is abandoned.
3521    
3522    32. Forward references to subpatterns in conditions such as (?(2)...) where
3523        subpattern 2 is defined later cause pcre_compile() to search forwards in
3524        the pattern for the relevant set of parentheses. This search went wrong
3525        when there were unescaped parentheses in a character class, parentheses
3526        escaped with \Q...\E, or parentheses in a #-comment in /x mode.
3527    
3528    33. "Subroutine" calls and backreferences were previously restricted to
3529        referencing subpatterns earlier in the regex. This restriction has now
3530        been removed.
3531    
3532    34. Added a number of extra features that are going to be in Perl 5.10. On the
3533        whole, these are just syntactic alternatives for features that PCRE had
3534        previously implemented using the Python syntax or my own invention. The
3535        other formats are all retained for compatibility.
3536    
3537        (a) Named groups can now be defined as (?<name>...) or (?'name'...) as well
3538            as (?P<name>...). The new forms, as well as being in Perl 5.10, are
3539            also .NET compatible.
3540    
3541        (b) A recursion or subroutine call to a named group can now be defined as
3542            (?&name) as well as (?P>name).
3543    
3544        (c) A backreference to a named group can now be defined as \k<name> or
3545            \k'name' as well as (?P=name). The new forms, as well as being in Perl
3546            5.10, are also .NET compatible.
3547    
3548        (d) A conditional reference to a named group can now use the syntax
3549            (?(<name>) or (?('name') as well as (?(name).
3550    
3551        (e) A "conditional group" of the form (?(DEFINE)...) can be used to define
3552            groups (named and numbered) that are never evaluated inline, but can be
3553            called as "subroutines" from elsewhere. In effect, the DEFINE condition
3554            is always false. There may be only one alternative in such a group.
3555    
3556        (f) A test for recursion can be given as (?(R1).. or (?(R&name)... as well
3557            as the simple (?(R). The condition is true only if the most recent
3558            recursion is that of the given number or name. It does not search out
3559            through the entire recursion stack.
3560    
3561        (g) The escape \gN or \g{N} has been added, where N is a positive or
3562            negative number, specifying an absolute or relative reference.
3563    
3564    35. Tidied to get rid of some further signed/unsigned compiler warnings and
3565        some "unreachable code" warnings.
3566    
3567    36. Updated the Unicode property tables to Unicode version 5.0.0. Amongst other
3568        things, this adds five new scripts.
3569    
3570    37. Perl ignores orphaned \E escapes completely. PCRE now does the same.
3571        There were also incompatibilities regarding the handling of \Q..\E inside
3572        character classes, for example with patterns like [\Qa\E-\Qz\E] where the
3573        hyphen was adjacent to \Q or \E. I hope I've cleared all this up now.
3574    
3575    38. Like Perl, PCRE detects when an indefinitely repeated parenthesized group
3576        matches an empty string, and forcibly breaks the loop. There were bugs in
3577        this code in non-simple cases. For a pattern such as  ^(a()*)*  matched
3578        against  aaaa  the result was just "a" rather than "aaaa", for example. Two
3579        separate and independent bugs (that affected different cases) have been
3580        fixed.
3581    
3582    39. Refactored the code to abolish the use of different opcodes for small
3583        capturing bracket numbers. This is a tidy that I avoided doing when I
3584        removed the limit on the number of capturing brackets for 3.5 back in 2001.
3585        The new approach is not only tidier, it mak