pcre/maint/README

MAINTENANCE README FOR PCRE2
============================

The files in the "maint" directory of the PCRE2 source contain data, scripts,
and programs that are used for the maintenance of PCRE2, but which do not form
part of the PCRE2 distribution tarballs. This document describes these files
and also contains some notes for maintainers. Its contents are:

  Files in the maint directory
  Updating to a new Unicode release
  Preparing for a PCRE2 release
  Making a PCRE2 release
  Long-term ideas (wish list)


Files in the maint directory
============================

GenerateUtt.py   A Python script to generate part of the pcre2_tables.c file
                 that contains Unicode script names in a long string with
                 offsets, which is tedious to maintain by hand.

ManyConfigTests  A shell script that runs "configure, make, test" a number of
                 times with different configuration settings.

MultiStage2.py   A Python script that generates the file pcre2_ucd.c from six
                 Unicode data files, which are themselves downloaded from the
                 Unicode web site. Run this script in the "maint" directory.
                 The generated file is written to stdout. It contains the
                 tables for a 2-stage lookup of Unicode properties, along with
                 some auxiliary tables.

pcre2_chartables.c.non-standard
                 This is a set of character tables that came from a Windows
                 system. It has characters greater than 128 that are set as
                 spaces, amongst other things. I kept it so that it can be
                 used for testing from time to time.

README           This file.

Unicode.tables   The files in this directory were downloaded from the Unicode
                 web site. They contain information about Unicode characters
                 and scripts. The ones used by the MultiStage2.py script are
                 CaseFolding.txt, DerivedGeneralCategory.txt, Scripts.txt,
                 ScriptExtensions.txt, GraphemeBreakProperty.txt, and
                 emoji-data.txt. I've kept UnicodeData.txt (which is no longer
                 used by the script) because it is useful occasionally for
                 manually looking up the details of certain characters.
                 However, note that character names in this file such as
                 "Arabic sign sanah" do NOT mean that the character is in a
                 particular script (in this case, Arabic). Scripts.txt and
                 ScriptExtensions.txt are where to look for script information.

ucptest.c        A short C program for testing the Unicode property macros
                 that do lookups in the pcre2_ucd.c data, mainly useful after
                 rebuilding the Unicode property table. Compile and run this in
                 the "maint" directory (see comments at its head). This program 
                 can also be used to find characters with specific properties. 

ucptestdata      A directory containing four files, testinput{1,2} and
                 testoutput{1,2}, for use in conjunction with the ucptest
                 program.

utf8.c           A short, freestanding C program for converting a Unicode code
                 point into a sequence of bytes in the UTF-8 encoding, and vice
                 versa. If its argument is a hex number such as 0x1234, it
                 outputs a list of the equivalent UTF-8 bytes. If its argument
                 is a sequence of concatenated UTF-8 bytes (e.g. e188b4) it
                 treats them as a UTF-8 character and outputs the equivalent
                 code point in hex. See comments at its head for details.


Updating to a new Unicode release
=================================

When there is a new release of Unicode, the files in Unicode.tables must be
refreshed from the web site. If the new version of Unicode adds new character
scripts, the source file pcre2_ucp.h and both the MultiStage2.py and the
GenerateUtt.py scripts must be edited to add the new names. I have been adding
each new group at the end of the relevant list, with a comment. Note also that
both the pcre2syntax.3 and pcre2pattern.3 man pages contain lists of Unicode
script names.

MultiStage2.py has two lists: the full names and the abbreviations that are
found in the ScriptExtensions.txt file. A list of script names and their
abbreviations can be found in the PropertyValueAliases.txt file on the
Unicode web site. There is also a Wikipedia page that lists them, and notes the
Unicode version in which they were introduced:

https://en.wikipedia.org/wiki/Unicode_scripts#Table_of_Unicode_scripts

Once the script name lists have been updated, MultiStage2.py can be run to
generate a new version of pcre2_ucd.c, and GenerateUtt.py can be run to
generate the tricky tables for inclusion in pcre2_tables.c (which must be
hand-edited). If MultiStage2.py gives the error "ValueError: list.index(x): x
not in list", the cause is usually a missing (or misspelt) name in one of the
lists of scripts.

The ucptest program can be compiled and used to check that the new tables in
pcre2_ucd.c work properly, using the data files in ucptestdata to check a
number of test characters. It used to be necessary to update the source
ucptest.c whenever new Unicode scripts were added, but this is no longer
required because that program now uses the lists in the PCRE2 source. However,
adding a few tests for new scripts to the files in ucptestdata is a good idea.


Preparing for a PCRE2 release
=============================

This section contains a checklist of things that I consult before building a
distribution for a new release.

. Ensure that the version number and version date are correct in configure.ac.

. Update the library version numbers in configure.ac according to the rules
  given below.

. If new build options or new source files have been added, ensure that they
  are added to the CMake files as well as to the autoconf files. The relevant
  files are CMakeLists.txt and config-cmake.h.in. After making a release
  tarball, test it out with CMake if there have been changes here.

. Run ./autogen.sh to ensure everything is up-to-date.

. Compile and test with many different config options, and combinations of
  options. Also, test with valgrind by running "RunTest valgrind" and
  "RunGrepTest valgrind" (which takes quite a long time). The script
  maint/ManyConfigTests now encapsulates this testing. It runs tests with
  different configurations, and it also runs some of them with valgrind, all of
  which can take quite some time.

. Run tests in both 32-bit and 64-bit environments if possible.

. Run tests with two or more different compilers (e.g. clang and gcc), and
  make use of -fsanitize=address and friends where possible. For gcc,
  -fsanitize=undefined -std=gnu99 picks up undefined behaviour at runtime, but
  needs -fno-sanitize=shift to get rid of warnings for shifts of negative
  numbers in the JIT compiler. For clang, -fsanitize=address,undefined,integer
  can be used but -fno-sanitize=alignment,shift,unsigned-integer-overflow must
  be added when compiling with JIT. Another useful clang option is
  -fsanitize=signed-integer-overflow

. Do a test build using CMake.

. Run perltest.sh on the test data for tests 1 and 4. The output should match
  the PCRE2 test output, apart from the version identification at the start of
  each test. Sometimes there are other differences in test 4 if PCRE2 and Perl
  are using different Unicode releases. The other tests are not Perl-compatible
  (they use various PCRE2-specific features or options).

. It is possible to test with the emulated memmove() function by undefining
  HAVE_MEMMOVE and HAVE_BCOPY in config.h, though I do not do this often.

. Documentation: check AUTHORS, ChangeLog (check version and date), LICENCE,
  NEWS (check version and date), NON-AUTOTOOLS-BUILD, and README. Many of these
  won't need changing, but over the long term things do change.

. I used to test new releases myself on a number of different operating
  systems. For example, on Solaris it is helpful to test using Sun's cc
  compiler as a change from gcc. Adding -xarch=v9 to the cc options does a
  64-bit test, but it also needs -S 64 for pcre2test to increase the stack size
  for test 2. Since I retired I can no longer do much of this, but instead I
  rely on putting out release candidates for folks on the pcre-dev list to
  test.

. The buildbots at http://buildfarm.opencsw.org/ do some automated testing
  of PCRE2 and should be checked before putting out a release.


Updating version info for libtool
=================================

This set of rules for updating library version information came from a web page
whose URL I have forgotten. The version information consists of three parts:
(current, revision, age).

1. Start with version information of 0:0:0 for each libtool library.

2. Update the version information only immediately before a public release of
   your software. More frequent updates are unnecessary, and only guarantee
   that the current interface number gets larger faster.

3. If the library source code has changed at all since the last update, then
   increment revision; c:r:a becomes c:r+1:a.

4. If any interfaces have been added, removed, or changed since the last
   update, increment current, and set revision to 0.

5. If any interfaces have been added since the last public release, then
   increment age.

6. If any interfaces have been removed or changed since the last public
   release, then set age to 0.

The following explanation may help in understanding the above rules a bit
better. Consider that there are three possible kinds of reaction from users to
changes in a shared library:

1. Programs using the previous version may use the new version as a drop-in
   replacement, and programs using the new version can also work with the
   previous one. In other words, no recompiling nor relinking is needed. In
   this case, increment revision only, don't touch current or age.

2. Programs using the previous version may use the new version as a drop-in
   replacement, but programs using the new version may use APIs not present in
   the previous one. In other words, a program linking against the new version
   may fail if linked against the old version at run time. In this case, set
   revision to 0, increment current and age.

3. Programs may need to be changed, recompiled, relinked in order to use the
   new version. Increment current, set revision and age to 0.


Making a PCRE2 release
======================

Run PrepareRelease and commit the files that it changes (by removing trailing
spaces). The first thing this script does is to run CheckMan on the man pages;
if it finds any markup errors, it reports them and then aborts.

Once PrepareRelease has run clean, run "make distcheck" to create the tarballs
and the zipball. Double-check with "svn status", then create an SVN tagged
copy:

  svn copy svn://vcs.exim.org/pcre2/code/trunk \
           svn://vcs.exim.org/pcre2/code/tags/pcre2-10.xx

When the new release is out, don't forget to tell webmaster@pcre.org and the
mailing list. Also, update the list of version numbers in Bugzilla
(administration > products > PCRE > Edit versions).


Future ideas (wish list)
========================

This section records a list of ideas so that they do not get forgotten. They
vary enormously in their usefulness and potential for implementation. Some are
very sensible; some are rather wacky. Some have been on this list for years.

. Optimization

  There are always ideas for new optimizations so as to speed up pattern
  matching. Most of them try to save work by recognizing a non-match without
  having to scan all the possibilities. These are some that I've recorded:

  * /((A{0,5}){0,5}){0,5}(something complex)/ on a non-matching string is very
    slow, though Perl is fast. Can we speed up somehow? Convert to {0,125}?
    OTOH, this is pathological - the user could easily fix it.

  * Turn ={4} into ==== ? (for speed). I once did an experiment, and it seems
    to have little effect, and maybe makes things worse.

  * "Ends with literal string" - note that a single character doesn't gain much
    over the existing "required code unit" feature that just remembers one code
    unit.

  * Remember an initial string rather than just 1 code unit.

  * A required code unit from alternatives - not just the last unit, but an
    earlier one if common to all alternatives.

  * Friedl contains other ideas.

  * The code does not set initial code unit flags for Unicode property types
    such as \p; I don't know how much benefit there would be for, for example,
    setting the bits for 0-9 and all values >= xC0 (in 8-bit mode) when a
    pattern starts with \p{N}.

. If Perl gets to a consistent state over the settings of capturing sub-
  patterns inside repeats, see if we can match it. One example of the
  difference is the matching of /(main(O)?)+/ against mainOmain, where PCRE2
  leaves $2 set. In Perl, it's unset. Changing this in PCRE2 will be very hard
  because I think it needs much more state to be remembered.

. A feature to suspend a match via a callout was once requested.

. An option to convert results into character offsets and character lengths.

. An option for pcre2grep to scan only the start of a file. I am not keen -
  this is the job of "head".

. A (non-Unix) user wanted pcregrep options to (a) list a file name just once,
  preceded by a blank line, instead of adding it to every matched line, and (b)
  support --outputfile=name.

. Define a union for the results from pcre2_pattern_info().

. Provide a "random access to the subject" facility so that the way in which it
  is stored is independent of PCRE2. For efficiency, it probably isn't possible
  to switch this dynamically. It would have to be specified when PCRE2 was
  compiled. PCRE2 would then call a function every time it wanted a character.

. pcre2grep: add -rs for a sorted recurse. Having to store file names and sort
  them will of course slow it down.

. Someone suggested --disable-callout to save code space when callouts are
  never wanted. This seems rather marginal.

. A user suggested a parameter to limit the length of string matched, for
  example if the parameter is N, the current match should fail if the matched
  substring exceeds N. This could apply to both match functions. The value
  could be a new field in the match context. Compare the offset_limit feature,
  which limits where a match must start.

. Write a function that generates random matching strings for a compiled
  pattern.

. Pcre2grep: an option to specify the output line separator, either as a string
  or select from a fixed list. This is not straightforward, because at the
  moment it outputs whatever is in the input file.

. Improve the code for duplicate checking in pcre2_dfa_match(). An incomplete,
  non-thread-safe patch showed that this can help performance for patterns
  where there are many alternatives. However, a simple thread-safe
  implementation that I tried made things worse in many simple cases, so this
  is not an obviously good thing.

. PCRE2 cannot at present distinguish between subpatterns with different names,
  but the same number (created by the use of ?|). In order to do so, a way of
  remembering *which* subpattern numbered n matched is needed. Bugzilla #760.
  (*MARK) can perhaps be used as a way round this problem. However, note that
  Perl does not distinguish: like PCRE2, a name is just an alias for a number
  in Perl.

. Instead of having #ifdef HAVE_CONFIG_H in each module, put #include
  "something" and the the #ifdef appears only in one place, in "something".

. Implement something like (?(R2+)... to check outer recursions.

. If Perl ever supports the POSIX notation [[.something.]] PCRE2 should try
  to follow.

. A user wanted a way of ignoring all Unicode "mark" characters so that, for
  example "a" followed by an accent would, together, match "a". This can only
  be done clumsily at present by using a lookahead such as /(?=a)\X/, which
  works for "combining" characters.

. Perl supports [\N{x}-\N{y}] as a Unicode range, even in EBCDIC. PCRE2
  supports \N{U+dd..} everywhere, but not in EBCDIC.

. Unicode stuff from Perl:

    \b{gcb} or \b{g}    grapheme cluster boundary
    \b{sb}              sentence boundary
    \b{wb}              word boundary

  See Unicode TR 29. The last two are very much aimed at natural language.

. (?[...]) extended classes: big project.

. Bugzilla #1694 requests backwards searching.

. Allow a callout to specify a number of characters to skip. This can be done
  compatibly via an extra callout field.

. Allow callouts to return *PRUNE, *COMMIT, *THEN, *SKIP, with and without
  continuing (that is, with and without an implied *FAIL). A new option,
  PCRE2_CALLOUT_EXTENDED say, would be needed. This is unlikely ever to be
  implemented by JIT, so this could be an option for pcre2_match().

. A limit on substitutions: a user suggested somehow finding a way of making
  match_limit apply to the whole operation instead of each match separately.

. Redesign handling of class/nclass/xclass because the compile code logic is
  currently very contorted and obscure.

. Some #defines could be replaced with enums to improve robustness.

. There was a request for an option for pcre2_match() to return the longest
  match. This would mean searching for all possible matches, of course.

. Perl's /a modifier sets Unicode, but restricts \d etc to ASCII characters,
  which is the PCRE2 default for PCRE2_UTF (use PCRE2_UCP to change). However,
  Perl also has /aa, which in addition, disables ASCII/non-ASCII caseless
  matching. Perhaps we need a new option PCRE2_CASELESS_RESTRICT_ASCII. In
  practice, this just means not using the ucd_caseless_sets[] table.

. There is more that could be done to the oss-fuzz setup (needs some research).
  A seed corpus could be built. I noted something about $LIB_FUZZING_ENGINE.
  The test function could make use of get_substrings() to cover more code.

. A neater way of handling recursion file names in pcre2grep, e.g. a single
  buffer that can grow.

. A user suggested that before/after parameters in pcre2grep could have
  negative values, to list lines near to the matched line, but not necessarily
  the line itself. For example, --before-context=-1 would list the line *after*
  each matched line, without showing the matched line. The problem here is what
  to do with matches that are close together. Maybe a simpler way would be a
  flag to disable showing matched lines, only valid with either -A or -B?

. There was a suggestiong for a pcre2grep colour default, or possibly a more
  general PCRE2GREP_OPT, but only for some options - not file names or patterns.

. Breaking loops that match an empty string: perhaps find a way of continuing
  if *something* has changed, but this might mean remembering additional data.
  "Something" could be a capture value, but then a list of previous values
  would be needed to avoid a cycle of changes. Bugzilla #2182.

. The use of \K in assertions is problematic. There was some talk of Perl
  banning this, but it hasn't happened. Some problems could be avoided by
  not allowing it to set a value before the match start; others by not allowing
  it to set a value after the match end. This could be controlled by an option
  such as PCRE2_SANE_BACKSLASH_K, for compatibility (or possibly make the sane
  behaviour the default and implement PCRE2_INSANE_BACKSLASH_K).

. If a function could be written to find 3-character (or other length) fixed
  strings, at least one of which must be present for a match, efficient
  pre-searching of large datasets could be implemented.

. If pcre2grep had --first-line (match only in the first line) it could be
  efficiently used to find files "starting with xxx". What about --last-line?

. A user requested a means of determining whether a failed match was failed by
  the start-of-match optimizations, or by running the match engine. Easy enough
  to define a bit in the match data, but all three matchers would need work.

. Would inlining "simple" recursions provide a useful performance boost for the
  interpreters? JIT already does some of this, but it may not be worth it for
  the interpreters.

. There was a request for a way of re-defining \w (and therefore \W, \b, and
  \B). An in-pattern sequence such as (?w=[...]) was suggested. Easiest way
  would be simply to inline the class, with lookarounds for \b and \B. Ideally
  the setting should last till the end of the group, which means remembering
  all previous settings; maybe a fixed amount of stack would do - how deep
  would anyone want to nest these things? Bugzilla #2301.

. Recognize the short script names. They are already listed in maint/
  Multistage2.py because they are needed for scanning the script extensions
  file.

. Use script extensions for \p?

. A user suggested something like --with-build-info to set a build information
  string that could be retrieved by pcre2_config(). However, there's no
  facility for a length limit in pcre2_config(), and what would be the
  encoding?

Philip Hazel
Email local part: ph10
Email domain: cam.ac.uk
Last updated: 01 April 2020