Alternative mbstring implementation using ICU

Hi there,

I almost finished an alternative implementation of mbstring that uses
ICU instead of the exotic libmbfl in hope of replacing the current one
for 5.4 (and possibly, 6.0.)

Although there are admittingly some known incompatibilities that need
extra libraries to resolve them besides a number of missing functions
that are intentionally removed for simplicity's sake, frequently used
functions are fully usable, and more compliant with the standard (e.g.
case insensitive matches).

Any comments are appreciated.

The source is ready in the following location:

http://github.com/moriyoshi/mbstring-ng/

Implemented functions:

• mb_convert_encoding()
• mb_detect_encoding()
• mb_ereg()
• mb_ereg_replace()
• mb_internal_encoding()
• mb_list_encodings()
• mb_output_handler()
• mb_parse_str()
• mb_preferred_mime_name()
• mb_regex_set_options()
• mb_split()
• mb_strcut()
• mb_strimwidth()
• mb_stripos()
• mb_stristr()
• mb_strlen()
• mb_strpos()
• mb_strripos()
• mb_strrpos()
• mb_strstr()
• mb_strtolower()
• mb_strtotitle()
• mb_strtoupper()
• mb_strwidth()
• mb_substr()
• mb_substr_count()

Removed functions and reasons behind it:

• mb_check_encoding()
  Not that usable as it is advertised, period. First of all, validation
  in terms of encoding is just as same as filtering through the
  converter supplied with the same value for the input and output
  encoding. Thus just use mb_convert_encoding().

• mb_convert_case()
  Use mb_strtoupper(), mb_strtolower() and mb_strtotitle()

• mb_convert_kana()
  This can't be standard-compliant. In addition, part of the
  functionality is already covered by Normalizer of intl extension, so
  we need to carefully consider what is actually needed here again.

• mb_convert_variables()
  This can be implemented as a script.

• mb_decode_mimeheader(), mb_encode_mimeheader()
  Non-standard compliancy.

• mb_decode_numericentity()
  Removed in favor of html_entity_decode().

• mb_encode_numericentity()
  Removed in favor of htmlentities() and htmlspecialchars().

• mb_encoding_aliases()
  Just unnecessary.

• mb_ereg_match()
  Use mb_ereg().

• mb_ereg_search(), mb_ereg_search_getpos(), mb_ereg_search_getregs(),
  mb_ereg_search_init(), mb_ereg_search_pos(), mb_ereg_search_regs() and
  mb_ereg_search_setpos()
  I rarely heard a script that actively uses these functions. They
  involve an internal state that is not visible to users, and thus it
  most likely causes confusion when used across the function calls.
  Need to be reimplemented as a class.

• mb_eregi()
  Use mb_regex_options() and mb_ereg()

• mb_eregi_replace()
  I wonder why this function was added in the first place because giving
  'i' option to mb_ereg_replace() works in the same way.

• mb_detect_order(), mb_get_info(), mb_http_input(), mb_http_output(),
  mb_language() and mb_substitute_character()
ini_set() and ini_get() are your friend, I guess...

• mb_regex_encoding()
  It is really confusing that the current mbstring allows two different
  encoding defaults that are applied to regex functions and the rest.
  Those settings are unified in the alternative version and so this is
  no longer necessary.

• mb_send_mail()
  The behavior of this function relies on the pseudo-locale setting
  called "mbstring.language" that supports just a limited set of
  possible locales. As not everyone can benefit from the function and
  most significant applications implement their own mail functions, I
  suppose this is no longer wanted.

• mb_strrchr()
  Use mb_strrpos().

• mb_strrichr()
  Use mb_strripos().

Known limitations and incompatibilities:

• mb_detect_encoding() doesn't work well anymore due to the
  inaccuracy of ICU's encoding detection facility.

• Request encoding translator now takes advantage of SAPI filter,
  therefore the name parts of the query components are not to be
  converted anymore.

• The group reference placeholders for mb_ereg_replace() is now
  $0, $1, $2... instead of \0, \1, \2. This can be avoided if we
  don't use uregex_replaceAll() and implement our own.

• ILP64 :-p

Regards,
Moriyoshi

and that there's nothing you can do with Oniguruma that you can't
also practically do with PCRE (to the best of my knowledge),
http://www.geocities.jp/kosako3/oniguruma/doc/RE.txt - Paragraph 8,
example 2 - specifying the nest level for subroutines/back
references doesn't work with pcre. That's one example I know from
the top of my head ...

Granted :). But it would be possible to rewrite that example into a
PCRE regexp, just a bit less efficiently. Subroutine calls are one
thing I like about Oniguruma, but not enough to counter all the other
arguments in favor of PCRE. Honestly, that sort of extension to the
syntax verges on kicking regexp out of the domain it was intended for;
function calls are verging on making regexp trivially Turing-complete.
If you need that level of matching, it's probably time to consider
more aggressive string parsing methodology, such as tokenization.
Regexp is basically a VM (at least as most implementations handle it),
and on nontrivial patterns will often be slower than other methods.

Certainly that particular example, matching an XML tag, will never
deal with all the esoteric cases. You can approach perfect matching as
pattern complexity approaches infinite (i.e. lim complexity->inf
correctness = inf), but the pattern is only realistically useful in
specific cases where you can be guaranteed a valid input stream. And
supposing it doesn't match, there's no reasonable way to determine
why.

Also, more to the point, Oniguruma hasn't been updated in two years
and counting, which is enough to count it as unmaintained. PCRE is
still seeing releases on a regular basis.

-- Gwynne