Chapter 7.  Strings

Table of Contents

String Classes
Simple Transformations
Case Sensitivity
Arbitrary Character Types
Tokenizing
Shrink to Fit
CString (MFC)

String Classes

Simple Transformations

Here are Standard, simple, and portable ways to perform common transformations on a string instance, such as "convert to all upper case." The word transformations is especially apt, because the standard template function transform<> is used.

This code will go through some iterations. Here's a simple version:

   #include <string>
   #include <algorithm>
   #include <cctype>      // old <ctype.h>

   struct ToLower
   {
     char operator() (char c) const  { return std::tolower(c); }
   };

   struct ToUpper
   {
     char operator() (char c) const  { return std::toupper(c); }
   };

   int main()
   {
     std::string  s ("Some Kind Of Initial Input Goes Here");

     // Change everything into upper case
     std::transform (s.begin(), s.end(), s.begin(), ToUpper());

     // Change everything into lower case
     std::transform (s.begin(), s.end(), s.begin(), ToLower());

     // Change everything back into upper case, but store the
     // result in a different string
     std::string  capital_s;
     capital_s.resize(s.size());
     std::transform (s.begin(), s.end(), capital_s.begin(), ToUpper());
   }
   

Note that these calls all involve the global C locale through the use of the C functions toupper/tolower. This is absolutely guaranteed to work -- but only if the string contains only characters from the basic source character set, and there are only 96 of those. Which means that not even all English text can be represented (certain British spellings, proper names, and so forth). So, if all your input forevermore consists of only those 96 characters (hahahahahaha), then you're done.

Note that the ToUpper and ToLower function objects are needed because toupper and tolower are overloaded names (declared in <cctype> and <locale>) so the template-arguments for transform<> cannot be deduced, as explained in this message. At minimum, you can write short wrappers like

   char toLower (char c)
   {
      return std::tolower(c);
   } 

(Thanks to James Kanze for assistance and suggestions on all of this.)

Another common operation is trimming off excess whitespace. Much like transformations, this task is trivial with the use of string's find family. These examples are broken into multiple statements for readability:

   std::string  str (" \t blah blah blah    \n ");

   // trim leading whitespace
   string::size_type  notwhite = str.find_first_not_of(" \t\n");
   str.erase(0,notwhite);

   // trim trailing whitespace
   notwhite = str.find_last_not_of(" \t\n");
   str.erase(notwhite+1); 

Obviously, the calls to find could be inserted directly into the calls to erase, in case your compiler does not optimize named temporaries out of existence.

Case Sensitivity

The well-known-and-if-it-isn't-well-known-it-ought-to-be Guru of the Week discussions held on Usenet covered this topic in January of 1998. Briefly, the challenge was, write a 'ci_string' class which is identical to the standard 'string' class, but is case-insensitive in the same way as the (common but nonstandard) C function stricmp().

   ci_string s( "AbCdE" );

   // case insensitive
   assert( s == "abcde" );
   assert( s == "ABCDE" );

   // still case-preserving, of course
   assert( strcmp( s.c_str(), "AbCdE" ) == 0 );
   assert( strcmp( s.c_str(), "abcde" ) != 0 ); 

The solution is surprisingly easy. The original answer was posted on Usenet, and a revised version appears in Herb Sutter's book Exceptional C++ and on his website as GotW 29.

See? Told you it was easy!

Added June 2000: The May 2000 issue of C++ Report contains a fascinating article by Matt Austern (yes, the Matt Austern) on why case-insensitive comparisons are not as easy as they seem, and why creating a class is the wrong way to go about it in production code. (The GotW answer mentions one of the principle difficulties; his article mentions more.)

Basically, this is "easy" only if you ignore some things, things which may be too important to your program to ignore. (I chose to ignore them when originally writing this entry, and am surprised that nobody ever called me on it...) The GotW question and answer remain useful instructional tools, however.

Added September 2000: James Kanze provided a link to a Unicode Technical Report discussing case handling, which provides some very good information.

Arbitrary Character Types

The std::basic_string is tantalizingly general, in that it is parameterized on the type of the characters which it holds. In theory, you could whip up a Unicode character class and instantiate std::basic_string<my_unicode_char>, or assuming that integers are wider than characters on your platform, maybe just declare variables of type std::basic_string<int>.

That's the theory. Remember however that basic_string has additional type parameters, which take default arguments based on the character type (called CharT here):

      template <typename CharT,
		typename Traits = char_traits<CharT>,
		typename Alloc = allocator<CharT> >
      class basic_string { .... };

Now, allocator<CharT> will probably Do The Right Thing by default, unless you need to implement your own allocator for your characters.

But char_traits takes more work. The char_traits template is declared but not defined. That means there is only

      template <typename CharT>
	struct char_traits
	{
	    static void foo (type1 x, type2 y);
	    ...
	};

and functions such as char_traits<CharT>::foo() are not actually defined anywhere for the general case. The C++ standard permits this, because writing such a definition to fit all possible CharT's cannot be done.

The C++ standard also requires that char_traits be specialized for instantiations of char and wchar_t, and it is these template specializations that permit entities like basic_string<char,char_traits<char>> to work.

If you want to use character types other than char and wchar_t, such as unsigned char and int, you will need suitable specializations for them. For a time, in earlier versions of GCC, there was a mostly-correct implementation that let programmers be lazy but it broke under many situations, so it was removed. GCC 3.4 introduced a new implementation that mostly works and can be specialized even for int and other built-in types.

If you want to use your own special character class, then you have a lot of work to do, especially if you with to use i18n features (facets require traits information but don't have a traits argument).

Another example of how to specialize char_traits was given on the mailing list and at a later date was put into the file include/ext/pod_char_traits.h. We agree that the way it's used with basic_string (scroll down to main()) doesn't look nice, but that's because the nice-looking first attempt turned out to not be conforming C++, due to the rule that CharT must be a POD. (See how tricky this is?)

Tokenizing

The Standard C (and C++) function strtok() leaves a lot to be desired in terms of user-friendliness. It's unintuitive, it destroys the character string on which it operates, and it requires you to handle all the memory problems. But it does let the client code decide what to use to break the string into pieces; it allows you to choose the "whitespace," so to speak.

A C++ implementation lets us keep the good things and fix those annoyances. The implementation here is more intuitive (you only call it once, not in a loop with varying argument), it does not affect the original string at all, and all the memory allocation is handled for you.

It's called stringtok, and it's a template function. Sources are as below, in a less-portable form than it could be, to keep this example simple (for example, see the comments on what kind of string it will accept).

#include <string>
template <typename Container>
void
stringtok(Container &container, string const &in,
	  const char * const delimiters = " \t\n")
{
    const string::size_type len = in.length();
	  string::size_type i = 0;

    while (i < len)
    {
	// Eat leading whitespace
	i = in.find_first_not_of(delimiters, i);
	if (i == string::npos)
	  return;   // Nothing left but white space

	// Find the end of the token
	string::size_type j = in.find_first_of(delimiters, i);

	// Push token
	if (j == string::npos)
	{
	  container.push_back(in.substr(i));
	  return;
	}
	else
	  container.push_back(in.substr(i, j-i));

	// Set up for next loop
	i = j + 1;
    }
}

The author uses a more general (but less readable) form of it for parsing command strings and the like. If you compiled and ran this code using it:

   std::list<string>  ls;
   stringtok (ls, " this  \t is\t\n  a test  ");
   for (std::list<string>const_iterator i = ls.begin();
	i != ls.end(); ++i)
   {
       std::cerr << ':' << (*i) << ":\n";
   } 

You would see this as output:

   :this:
   :is:
   :a:
   :test: 

with all the whitespace removed. The original s is still available for use, ls will clean up after itself, and ls.size() will return how many tokens there were.

As always, there is a price paid here, in that stringtok is not as fast as strtok. The other benefits usually outweigh that, however.

Added February 2001: Mark Wilden pointed out that the standard std::getline() function can be used with standard istringstreams to perform tokenizing as well. Build an istringstream from the input text, and then use std::getline with varying delimiters (the three-argument signature) to extract tokens into a string.

Shrink to Fit

From GCC 3.4 calling s.reserve(res) on a string s with res < s.capacity() will reduce the string's capacity to std::max(s.size(), res).

This behaviour is suggested, but not required by the standard. Prior to GCC 3.4 the following alternative can be used instead

      std::string(str.data(), str.size()).swap(str);
   

This is similar to the idiom for reducing a vector's memory usage (see this FAQ entry) but the regular copy constructor cannot be used because libstdc++'s string is Copy-On-Write.

In C++11 mode you can call s.shrink_to_fit() to achieve the same effect as s.reserve(s.size()).

CString (MFC)

A common lament seen in various newsgroups deals with the Standard string class as opposed to the Microsoft Foundation Class called CString. Often programmers realize that a standard portable answer is better than a proprietary nonportable one, but in porting their application from a Win32 platform, they discover that they are relying on special functions offered by the CString class.

Things are not as bad as they seem. In this message, Joe Buck points out a few very important things:

  • The Standard string supports all the operations that CString does, with three exceptions.

  • Two of those exceptions (whitespace trimming and case conversion) are trivial to implement. In fact, we do so on this page.

  • The third is CString::Format, which allows formatting in the style of sprintf. This deserves some mention:

The old libg++ library had a function called form(), which did much the same thing. But for a Standard solution, you should use the stringstream classes. These are the bridge between the iostream hierarchy and the string class, and they operate with regular streams seamlessly because they inherit from the iostream hierarchy. An quick example:

   #include <iostream>
   #include <string>
   #include <sstream>

   string f (string& incoming)     // incoming is "foo  N"
   {
       istringstream   incoming_stream(incoming);
       string          the_word;
       int             the_number;

       incoming_stream >> the_word        // extract "foo"
		       >> the_number;     // extract N

       ostringstream   output_stream;
       output_stream << "The word was " << the_word
		     << " and 3*N was " << (3*the_number);

       return output_stream.str();
   } 

A serious problem with CString is a design bug in its memory allocation. Specifically, quoting from that same message:

   CString suffers from a common programming error that results in
   poor performance.  Consider the following code:

   CString n_copies_of (const CString& foo, unsigned n)
   {
	   CString tmp;
	   for (unsigned i = 0; i < n; i++)
		   tmp += foo;
	   return tmp;
   }

   This function is O(n^2), not O(n).  The reason is that each +=
   causes a reallocation and copy of the existing string.  Microsoft
   applications are full of this kind of thing (quadratic performance
   on tasks that can be done in linear time) -- on the other hand,
   we should be thankful, as it's created such a big market for high-end
   ix86 hardware. :-)

   If you replace CString with string in the above function, the
   performance is O(n).
   

Joe Buck also pointed out some other things to keep in mind when comparing CString and the Standard string class:

  • CString permits access to its internal representation; coders who exploited that may have problems moving to string.

  • Microsoft ships the source to CString (in the files MFC\SRC\Str{core,ex}.cpp), so you could fix the allocation bug and rebuild your MFC libraries. Note: It looks like the CString shipped with VC++6.0 has fixed this, although it may in fact have been one of the VC++ SPs that did it.

  • string operations like this have O(n) complexity if the implementors do it correctly. The libstdc++ implementors did it correctly. Other vendors might not.

  • While parts of the SGI STL are used in libstdc++, their string class is not. The SGI string is essentially vector<char> and does not do any reference counting like libstdc++'s does. (It is O(n), though.) So if you're thinking about SGI's string or rope classes, you're now looking at four possibilities: CString, the libstdc++ string, the SGI string, and the SGI rope, and this is all before any allocator or traits customizations! (More choices than you can shake a stick at -- want fries with that?)