RE: [boost] RE: Re: [program_options] Re: Unicode support
Hi Volodya,
From: Vladimir Prus [mailto:ghost@cs.msu.su]
UCS-2 held all characters in Unicode 1.1, There was a need for more unique numbers and UCS-4 was introduced in Unicode 2.0. Unfortunately there is no 4-byte character specialization for basic_string<> in STL yet.
Or, to be exact, there's no agreement if wchar_t should be 32-bit or 16-bit. Linux (or gcc specifically) uses 32 bits, and Windows 16, which means wstring only suitable for UCS-2.
Wow, I did not know that about gcc, what a smart compiler... ;-) MSVC does it usually because Win32 API id ANSI/UCS-2, so the usage is straightforward then.
Besides, UCS-2 has a mechanism to represent characters outside of 16-bit space with two elements, which, I suspect, won't work if wchar_t is 16 bit.
Actually, not UCS-2, but another variable-sized character encoding was introduced in Unicode 2.0 - UTF-16. It serves for the 16-bit characters similarly like UTF-8 for the 8-bit ones. Using numbers from the unused (and forbidden) space it allows to encode characters in more than one element and so break the constraint of the 16-bits. It works with wchar_t, if only UCS-2 characters are stored in it; because the special characters are forbidden in UCS-2 anyway. However, UTF-16 brings the inconvenience of variable-sized character and those, who used UCS-2 and "enjoyed" the fixed character size should switch to UCS-4 to carry on... Not all applications need it, indeed, UCS-2 is enough, usually.
Generally speaking UTF encodings (e.g. UTF-8 or UTF-16) are not suitable for the fast in-memory usage because some operations, for example at(int) and length() have not O(1) but O(n) to give a result.
I believe UTF-8 is popular party because such operations are believed to be rare.
Really? string.length() ? You must have been joking... :-) UTF-8 is popular because the vast amount of texts contains much more ASCII characters than the accented ones. It saves space to store them in this encoding. It was with the UTF-7 encoding also the first one, thus the oldest. It can be also comfortable for the application not to perform any conversion and work with the text in its native encoding - here can help the stringized interface working with UTF-8. Not for all applications, however. For example, XML parsers tend to work in UTF-16 internally, because UCS-4 eats much memory and UTF-8 as well for East Asian encodings. I also feel, that the applications, which parse texts and extracts string parameters to be used in some structures or as user input are processed usually with fixed-size strings, because the string operations are easier.
They (UTF, T=Transformation) are better for storing texts as they save place by using variable number of bytes for a character. However UCS (e.g. UCS-2 or UCS-4) are fast for memory operations because they use fixed character size.
What about representing values with two 16-bit values (that's what I've mentioned above)?
You mean UTF-16. Yes, it is also a possiblity, but having wchar_t not fixed sized, you would have to provide basic_string<char16_t>... I see it pretty the same like your UTF-8 option, because both encodings use variable-sized character, being so incompatible with the shared implementation of basic_string<> for char and wchar_t. Moreover, if basic_string<char> from the API would produce UTF-8 string, that it would be only consistent, if basic_string<wchar_t> produced UTF-16 and not UCS-2. To say nothing of support of 4-byte wchar_t and UTF-32 :-)
(BTW, for reference to those interested, http://www.cl.cam.ac.uk/~mgk25/unicode.html#ucsutf talks about different encodings).
That is why I would not like to use basic_string<utf8char> in memory, rather basic_string<ucs4char> instead, but I would not generalize it for all possible applications.
In case of program options, I suspect that everything will work for UTF-8 strings. IOW, the library does not care that any given 'char' might be part of some Unicode character. That's why it's attractive for me.
It depends on the way of usage of basic_string<> inside and outside of the program_options. The implemenation of basic_string<> is done for fixed-size characters. It does not do searching/iterating/length counting with regards to the character size, which can vary from 1 to 6 bytes in UTF-8. For example string.find() would produce unexpected results if in haystack was stored UTF-8 and a needle was in ANSI, where all 255 numbers are allowed. Not only the library must be aware of it, but also the users, that for example basic_string<char>.size() does not deliver the count of characters but the count of bytes in a string then. That is why I do not like usage of basic_string<> for variable sized characters; basic_string<char> would not have consistent behavior for different encodings; as long as it is used for different ANSI encodings, the only misunderstanding is when two different characters in two local alphabets share the same character above ASCII. For other methods is basic_string<char> consistent.
You can expect initialization from (const char * argv []) on all platforms or (const wchar_t * argv []) on Windows in UCS-2. With the basic_string<> you already have support for the parameters from the current locale (char) and for parameters in UCS-2.
What do you mean by 'parameters from the current locale'?
If you set LC_CTYPE to something like "cs_CZ.iso8859-2" on UN*Xes or choose "Czech" in Windows Control Panel, command line shell (well, not only it) will start to accept and deliver in (char * argv []) also characters from a local alphabet (here Czech, for example). Such string you would have convert into UTF-8, if you wanted to have an UTF-8 interface. It means, you would have to do more than basic_string<char>(argv[x])...
I am not sure that ctype::widen is required to care about user-selected character encoding. Not do I think it's requires from default codecvt facet.
AFAIK it must care; the method widen() needs locale to provide an extension from char to a templated char_type. STL in MSVC uses correctly mbcstowc() to perform the conversion from the local alphabet to the UCS-2. I hope that no-one simply casts to wchar_t, being so reliable only for 7-bit characters. Anyway, it is always possible to write an own converting facet and force its usage for widen by with imbue.
If my locale on Linux is ru_RU.KOI8-R I don't think standard requires codecvt<wchar_t, char> facet in default instance of 'locale' to do meaningfull conversion from KOI8-R into unicode.
Hmm, I am not sure about the ISO/C++ definition. If the standard implementation were something like "return (wchar_t) x" then it would be useless for anything except 7-bit ASCII. MSVC works with locale and it should be according to the standard, hopefully. Yet delivers the shell the variable "char * argv []" in ru_RU.KOI8-R to you.
If we take an option to read parameters into basic_string<wchar_t> or basic_string<ucs4char_t>, where the character size or encoding is not the same as the native encoding on the command line, there is an affinity to streams. Some shells allow usage of UTF-8 encoded parameters or, generally, usage of characters out of the current locale. It means, that a program can choose the way, how to encode all characters from Unicode to char. UTF-7/8, etc. I would like to have a solution similar to streams: imbue(). Having this, you could convert internally every argv[x] using imbue(y) applied to a stringstream, where the facet y provides the caller.
That's right. There should be some mechanism to convert from current 8-bit encoding and source code encoding into unicode. At least at linux there's a mechanism to do the first thing, but I'm not aware about Windows.
There are such methods in stdlib.h, which produce a wide character or a string of wide characters: int mbtowc(wchar_t *pwc, const char *s, size_t n); size_t mbstowcs(wchar_t *pwcs, const char *s, size_t n); These methods work with the locale supported in the operating system and thus you can always get a wchar_t from a char. And these should be used in the default facet in STL.
On the other hand, such a conversion can be performed also by a user. The parameters sent to main() are char* or wchar_t* and thus program_options can give them back just as they are in basic_string<char> and basic_string<wchar_t>. The client can use his facet to imbue a basic_stringstream<> initialized with the parameter from program_options. Or a conversion library could be used to perform the conversion, something like lexical_cast<> does for type converions. It is a matter of convenience only - separate converions library (no support for encoding in program_options) or imbue(), performing the conversion inside the program_options. However, the conversion should not be implemented for program_options only, that is why I suggested an existing interface - facets.
I agree with basing the mechanism on facets. OTOH, it's really can be made orthogonal to program options. So initially program_options can support only ascii (strict 7-bit) and unicode, for which conversion is trivial.
Hmm, if it was orthogonal to program options, like facets are to streams, then there is no need to declare support of encodings. char * and basic_string<char> are simply in the local encoding and wchar_t * and basic_string<wchar_t> are wide according to the understanding of "wideness" on a platform (Unicode). It will also support the input, which you can expect for program_options from main(): char * in the current locale encoding. If someone wants something else, he can use facet and streams to convert it (which could be hidden in program_options with imbue, but not necessarily). I think the point is not to support UTF-8 or UCS-2 but to support encodings explicitely declared or not. I would not implicitely return UTF-8 or UTF-16 basic_string<>s, not only because STL handles their chars implicitely according to the current locale, but also because basic_string<char>(argv[x]) does not behave this way. Having the encoding independent on program_options using some external conversions, possibly imbuable into program_options, make the library thinner and more concentrated on the problem: parse the command line. Like the streams - read characters from input. Those, who say "we need full Unicode, UTF-8 or UCS-4" have the same option as if working with streams - imbue or convert. Otherwise they have currebnt locale char from istream or wchar_t from wistream.
You can find come converting facets for UTF-8 raedy for imbue() to a stream in the files section on yahoo. Unfortunately not finished or not reviewed...
Yea, I can find some facets, including one written by myself ;-( And yea, unfortunately they are not in Boost yet.
Great! You can probly prepare it for review, if your time allows it... :-) Ferda
- Volodya
_______________________________________________ Unsubscribe & other changes: http://lists.boost.org/mailman/listinfo.cgi/boost
Hi Ferda,
Linux (or gcc specifically) uses 32 bits, and Windows 16, which means wstring only suitable for UCS-2.
Wow, I did not know that about gcc, what a smart compiler... ;-)
MSVC does it usually because Win32 API id ANSI/UCS-2, so the usage is straightforward then.
Maybe gcc is just not burdened with compatibility with older APIs ;-)
Actually, not UCS-2, but another variable-sized character encoding was introduced in Unicode 2.0 - UTF-16.
Thanks for correction.
I believe UTF-8 is popular party because such operations are believed to be rare.
Really? string.length() ? You must have been joking... :-)
No, I was not. The point is that you rarely need string.length() to return the number of real characters. You need it when doing output, but then wstring::length() is not accurate either, because there are composing characters which don't take space, and there are characters which takes twice as much space on screen/terminal and ascii characters. You probably need it when working with unicode characters. But if you want to find 7-bit ascii substring in Unicode string, you can just do iterate over all characters -- it does not matter if you'll check the bytes which are part of unicode character. At least, that's what I think is the case ;-) For program_options this is true.
What about representing values with two 16-bit values (that's what I've mentioned above)?
You mean UTF-16. Yes, it is also a possiblity, but having wchar_t not fixed sized, you would have to provide basic_string<char16_t>... I see it pretty the same like your UTF-8 option, because both encodings use variable-sized character, being so incompatible with the shared implementation of basic_string<> for char and wchar_t.
Moreover, if basic_string<char> from the API would produce UTF-8 string, that it would be only consistent, if basic_string<wchar_t> produced UTF-16 and not UCS-2. To say nothing of support of 4-byte wchar_t and UTF-32 :-)
In fact, given what Mira said in another message, I'm entirely lost what's the really right representation for Unicode string. So it's better to focus on what representation can be used inside program_options.
In case of program options, I suspect that everything will work for UTF-8 strings. IOW, the library does not care that any given 'char' might be part of some Unicode character. That's why it's attractive for me.
It depends on the way of usage of basic_string<> inside and outside of the program_options. The implemenation of basic_string<> is done for fixed-size characters. It does not do searching/iterating/length counting with regards to the character size, which can vary from 1 to 6 bytes in UTF-8.
Luckily, it won't expose UTF-8 string, so if all searches use 7-bit string we're safe.
You can expect initialization from (const char * argv []) on all platforms or (const wchar_t * argv []) on Windows in UCS-2. With the basic_string<> you already have support for the parameters from the current locale (char) and for parameters in UCS-2.
What do you mean by 'parameters from the current locale'?
If you set LC_CTYPE to something like "cs_CZ.iso8859-2" on UN*Xes or choose "Czech" in Windows Control Panel, command line shell (well, not only it) will start to accept and deliver in (char * argv []) also characters from a local alphabet (here Czech, for example). Such string you would have convert into UTF-8, if you wanted to have an UTF-8 interface. It means, you would have to do more than basic_string<char>(argv[x])...
Yes, you right. But I think I'd still have to do conversion. E.g. user creates an option with a type of wstring but parsers char** argv.
I am not sure that ctype::widen is required to care about user-selected character encoding. Not do I think it's requires from default codecvt facet.
AFAIK it must care; the method widen() needs locale to provide an extension from char to a templated char_type.
The standard says The char argument of do_widen is intended to accept value derived from character literals... and encoding of character literals usually does not change when user changes locale. So implementation might handle only 7-bit values.
STL in MSVC uses correctly mbcstowc() to perform the conversion from the local alphabet to the UCS-2. I hope that no-one simply casts to wchar_t, being so reliable only for 7-bit characters. Anyway, it is always possible to write an own converting facet and force its usage for widen by with imbue.
I've just looked at implementation in one of the gcc versions I have and it just casts to wchar_t.
That's right. There should be some mechanism to convert from current 8-bit encoding and source code encoding into unicode. At least at linux there's a mechanism to do the first thing, but I'm not aware about Windows.
There are such methods in stdlib.h, which produce a wide character or a string of wide characters: int mbtowc(wchar_t *pwc, const char *s, size_t n); size_t mbstowcs(wchar_t *pwcs, const char *s, size_t n);
Yes, that's the mechanism I was talking about.
These methods work with the locale supported in the operating system and thus you can always get a wchar_t from a char. And these should be used in the default facet in STL.
I've tried to find implementation in gcc, but failed :-(
I agree with basing the mechanism on facets. OTOH, it's really can be made orthogonal to program options. So initially program_options can support only ascii (strict 7-bit) and unicode, for which conversion is trivial.
Hmm, if it was orthogonal to program options, like facets are to streams, then there is no need to declare support of encodings. char * and basic_string<char> are simply in the local encoding and wchar_t * and basic_string<wchar_t> are wide according to the understanding of "wideness" on a platform (Unicode).
It will also support the input, which you can expect for program_options from main(): char * in the current locale encoding. If someone wants something else, he can use facet and streams to convert it (which could be hidden in program_options with imbue, but not necessarily).
I think the point is not to support UTF-8 or UCS-2 but to support encodings explicitely declared or not. I would not implicitely return UTF-8 or UTF-16 basic_string<>s, not only because STL handles their chars implicitely according to the current locale, but also because basic_string<char>(argv[x]) does not behave this way. Having the encoding independent on program_options using some external conversions, possibly imbuable into program_options, make the library thinner and more concentrated on the problem: parse the command line. Like the streams - read characters from input.
So, can we agree on this: 1. Sometimes, the program_options library need to make conversion from local encoding to unicode and from unicode to local encoding. 2. This conversion should be done via codecvt facet 3. It might be possible to allow user to imbue/specify codecvt facet when using program_options.
Those, who say "we need full Unicode, UTF-8 or UCS-4" have the same option as if working with streams - imbue or convert. Otherwise they have currebnt locale char from istream or wchar_t from wistream.
For example if one excepts char** argv to be in UTF-8 he just changes locale accordingly and all std::string values that program_options *returns* will be in that encoding automatically.
You can find come converting facets for UTF-8 raedy for imbue() to a stream in the files section on yahoo. Unfortunately not finished or not reviewed...
Yea, I can find some facets, including one written by myself ;-( And yea, unfortunately they are not in Boost yet.
Great! You can probly prepare it for review, if your time allows it... :-)
Let's see ;-) - Volodya
participants (2)
-
Ferdinand Prantl -
Vladimir Prus