Archive for the ‘Uncategorized’ Category

Is std::string’s storage contiguous?

Take a look at the following code:

#include <cstddef>
#include <cstring>
#include <iostream>
#include <string>

/* Dummy recv function */
size_t recv(int socket, void *buffer, size_t length, int flags)
{
 std::memcpy(buffer, "Some stuff", 10);
 return 10;
}

int main()
{
 //Create a string with space for some characters
 std::string x(64, char());

 std::cout << "x.size() = " << x.size() << ", x.capacity() = "
           << x.capacity() << std::endl;

 //filling the string with data assuming that &x[0]
 //is pointing at the beginning a contiguous array
 std::size_t bytes_read = recv(0, &x[0], x.size(), 0);

 //Using the good old swap trick to free excess space
 std::string(&x[0], bytes_read).swap(x);

 std::cout << "x.size() = " << x.size() << ", x.capacity() = "
           << x.capacity() << std::endl;
 std::cout << x << std::endl;
}

Is that code… valid? I always thought it wasn’t since I heard lots of people telling me that storage for a std::string isn’t guaranteed to be contiguous (I’ve yet to see an implementation with storage that isn’t) so the preceding code would possibly cause undefined behavior. A few days ago, I was on Freenode’s ##C++ when some guy asked the same question, I was responding him with the  classical answer… when another guy “Tinodidriksen” replied that I was wrong. What?!? What if he’s right? I mean, I’ve never actually looked at the standard about it, I only took the usual answer as being the truth. Let’s see what the standard says about it.

1) basic_string constructor requirements(See tables 38-43 in the 14882:2003 standard):
Excerpt from table 39: “data() points at the first element of an allocated copy of rlen consecutive elements of the string controlled by str beginning at position pos”, rlen being equal to size() according to the same table.

2) 21.3.4 paragraph 1, basic_string indexed access:
Returns: If pos < size(), returns data()[pos]. Otherwise, if pos == size(), the const version returns charT().

3) 21.3.6 paragraph 4, const charT* data() const:
Requires: The program shall not alter any of the values stored in the character array.

Well, these are the three relevant excerpts I found, the first one guarantees that the storage is contiguous because it’s allocated in a single block of X consecutive elements. The second proves that when you modify the string using the index operator, it modifies the buffer pointed by data(). (data() in that specific sentence is probably meaning the pointer to the underlying storage and not the data() member function itself since using it would force the implementor to use an ugly const cast and would violate the third and last point).

My conclusion, the standard does not say explicitly “std::basic_string must have contiguous storage” like std::vector (ISO/IEC 14882:2003 only) but there are enough constraints available to state that it can’t be otherwise. If I missed something or you have a different opinion, I’d love to hear it.

Clang, a better compiler for C and its derivatives.

If there’s a project that caught my interest lately, it’s Clang, a front-end for the LLVM compiler created as a “drop-in” replacement for gcc. It’s a brand new, state of the art, modern, BSD-licensed open-source compiler that is a lot faster, optimizes more aggressively while using far less memory than the competition. For now, only the C(Yes, at last, a C99 compiler) and Objective C support is complete and are considered “production ready”. C++’s support is still only partial but the situation is improving at a very fast pace. See  http://clang.llvm.org/cxx_status.html for more information about it.

Having a compiler that produces good and optimized code is always welcome but… Is there anything left in it for the programmer? Yes, clang’s static analyzer provides the programmer with clearer, better diagnostic messages. It is also easier to integrate to IDEs and is already doing wonders in Apple’s XCode 3.2. Let’s see what it can do.

For a short example, see this ill-formed code:


#include <stdio.h>

typedef const char *LPCTSTR; /*Let's say they're ugly          */
typedef char *LPSTR;         /*typedefs from some other header */

void print(LPSTR toprint)
{
    printf("%s", toprint)    /*Omitted ';' on purpose*/
}

int main(void)
{
    LPCTSTR a = "Hello";
    LPCTSTR b = "World";
    LPCTSTR c = a + b; /*cannot concatenate strings with operator +*/
    print(c); /*Passing a const char * to a function taking a char * */
    return 0;
}

Let’s compare clang to gcc and cl (Visual C++). All three compiler detect two errors and a warning but let’s see how good are their diagnostics:

First error: Semi-colon missing at the end of line 8

GCC:
test.c: In function ‘print’:
test.c:9: error: expected ‘;’ before ‘}’ token
CL:
...\test.c(9) : error C2143: syntax error : missing ';' before '}'
CLang:
test.c:8:26: error: expected ';' after expression
    printf("%s", toprint)
                         ^
                         ;

Both gcc’s and cl’s diagnostic are wrong, the semi-colon should be placed at the end of line 8. On the other side, clang found that a semi-colon is missing at the 26th character of line 8 and even points it with ^, a really nice improvement.

Second Error: You can’t add two LPCTSTR(const char * in disguise)

GCC:
test.c: In function ‘main’:
test.c:15: error: invalid operands to binary + (have ‘LPCTSTR’ and
           ‘LPCTSTR’)
CL:
...\test.c(15) : error C2110: '+' : cannot add two pointers
CLang:
test.c:15:19: error: invalid operands to binary expression
              ('LPCTSTR' (aka 'char const *') and 'LPCTSTR'
              (aka 'char const *'))
    LPCTSTR c = a + b;
                ~ ^ ~

Now, all three compilers found the error but the diagnostics of gcc and cl are not really helpful, they only tell us that we cannot add two LPCTSTR but doesn’t give us a clue as to why. CLang’s diagnostic message is about the same but also shows us that LPCTSTRs are simply char pointers in disguise and where is the exact source of the problem.

Third Error: Passing a const char * to a function taking char * could be dangerous

GCC:
test.c:16: warning: passing argument 1 of ‘print’ discards
           qualifiers from pointer target type
test.c:6: note: expected ‘LPSTR’ but argument is of type ‘LPCTSTR’
CL:
.../test.c(16) : warning C4090: 'function' : different 'const' qualifiers
CLang:
test.c:16:11: warning: passing 'LPCTSTR' (aka 'char const *') discards
              qualifiers, expected 'LPSTR' (aka 'char *')
    print(c);
          ^

Now, gcc’s diagnostic is just plain vague, the fact is mentions “pointer target type” can give us a small hint that LPCTSTR might be a pointer but it’s not like “const” is the only possible qualifier in C. cl does a bit better telling us that it’s a constness issue but again, clang is clearly the winner by letting us know that LPCTSTR is in fact a char const * and showing us where is the problem.

Probably that some conservative “elitists” are going to say that gcc or cl diagnostics are fine the way they are but the way I see it, there’s always place for improvement and if a tool like clang can increase my productivity while generating fast and optimized code then… why not? I encourage everyone to at least give it a try( for C and Objective C ) and even contributing to it by improving it’s C++ support.

http://clang.llvm.org

C++ and functional programming idioms

If you’re curious like me, you probably ventured at least once in the scary and mind-bending world of functional programming, came back and told yourself: “It would be nice if I could do this or that in c++”. FP languages have been present for decades but only recently, we’ve been starting to see the adoption of some of their techniques in classical imperative languages… like higher order functions, closures/lambdas functions, currying and lazy evaluation. For example, Javascript supports closures since version 1.7 and C# from 3.0.

Seeing how useful these techniques are, it’s normal to want them in our favorite programming language. We’re already doing a bit a FP without knowing thanks to the standard library algorithms. Lots of them takes functors/predicates as arguments so they mimics fairly well the behavior of higher order functions. Besides that, C++ has no built-in support for other idioms like lambda functions or closures but we can achieve similar effects due to the lazy nature of templates and a technique known as “expression templates”. More on that technique on a future post…

To demonstrate my point, let’s take a small program that takes a string as input and return the most frequent character. In the old classical C++ way, it could be implemented as follow:

#include <iostream>
#include <locale>
#include <map>
#include <string>

namespace
{

    char most_frequent_letter(const std::string &str)
    {
        typedef std::map<char, unsigned int> char_counts_t;

        char_counts_t char_counts;

        for(std::string::const_iterator itr = str.begin();
                itr != str.end(); ++itr)
            if(std::isalpha(*itr, std::locale()))
                ++char_counts[*itr];

        for(char_counts_t::const_iterator itr = char_counts.begin();
                itr != char_counts.end(); ++itr)
            std::cout << itr->first << " => " << itr->second << std::endl;

        if(!char_counts.empty())
        {
            char_counts_t::const_iterator highest_count = char_counts.begin();
            for(char_counts_t::const_iterator itr = ++char_counts.begin();
                    itr != char_counts.end(); ++itr)
                if(itr->second > highest_count->second)
                    highest_count = itr;
            return highest_count->first;
        }
        return ' ';
    }

}

int main(int argc, char *argv[])
{
    if(argc > 1)
    {
        std::string some_string = argv[1];
        std::cout << "The string is: " << some_string << "\n" << std::endl;
        std::cout << "The most frequent letter is: " <<
            most_frequent_letter(some_string) << std::endl;
    }
    else
        std::cout << "Usage: " << argv&#91;0&#93; << " <string>" << std::endl;
}
&#91;/sourcecode&#93;

So far so good, it works and does the job. We're putting the characters in a map using the character as the key and the count as the value. Then, we print the content of the map and finally iterate through it to find the character with the highest value. The problems with this code is that we're reinventing parts already in the standard library and that code lacks expressiveness. Let's see how the code could look like if we used the standard algorithms.

&#91;sourcecode language='cpp'&#93;
namespace
{

    template <typename map_t>
    struct map_filler
    {
        typedef void result_type;

        map_filler(map_t &map):
            map_(map)
        {
        }
        template <typename T>
        result_type operator()(const T &t) const
        {
            if(std::isalpha(t, std::locale()))
                ++map_[t];
        }
    private:
        map_t &map_;
    };

    struct pair_printer
    {
        typedef void result_type;

        template <typename pair_t>
        result_type operator()(const pair_t &pair) const
        {
            std::cout << pair.first << " => " << pair.second << std::endl;
        }
    };

    struct pair_value_comparer
    {
        typedef bool result_type;

        template <typename pair_t>
        result_type operator()(const pair_t &a, const pair_t &b)
        {
            return a.second < b.second;
        }
    };

    char most_frequent_letter(const std::string &str)
    {
        typedef std::map<char, unsigned int> char_counts_t;

        char_counts_t char_counts;

        std::for_each(str.begin(), str.end(),
                map_filler<char_counts_t>(char_counts));

        std::for_each(char_counts.begin(), char_counts.end(),
                pair_printer());

        char_counts_t::const_iterator result = std::max_element(
                char_counts.begin(),
                char_counts.end(),
                pair_value_comparer());

        return (result != char_counts.end()) ? result->first : ' ';
    }

}

Hmm… Okay… Let’s see, our “most_frequent_letter” function is now using the standard library algorithms. It does make the function clearer and way more expressive but at the cost of around 40 lines of “support code” whose are in our case, functors. Even when thinking in terms of reusability, the chance of needing that same support code in the future is small, if not inexistant. What would we do in that case in a FP language? Use a small lambda functions/closure instead. For that example, I’m going to use boost::phoenix 2.0, an efficient FP library part of boost which is in my opinion the best general, multi-purpose C++ library and a must-have for any serious C++ programmer. Let’s see what phoenix can do:

namespace
{

	namespace phx = boost::phoenix;
	using namespace phx::arg_names;
	using namespace phx::local_names;
	using phx::at_c;

	char most_frequent_letter(const std::string &str)
	{
		typedef std::map<char, unsigned int> char_counts_t;

		char_counts_t char_counts;

		std::for_each(str.begin(), str.end(),
				phx::if_(phx::bind(std::isalpha<char>, _1,
						phx::construct<std::locale>()))
				[
					phx::let(_a = phx::ref(char_counts)[_1])
					[
						++_a
					]
				]);

		std::for_each(char_counts.begin(), char_counts.end(),
				std::cout << at_c<0>(_1) << " => " << at_c<1>(_1) << std::endl);

		char_counts_t::const_iterator result = std::max_element(
				char_counts.begin(), char_counts.end(),
				at_c<1>(_1) < at_c<1>(_2));

		return (result != char_counts.end()) ? result->first : ' ';
	}

}

I made a few using statements to make the code easier to understand. Let’s take a look the for_each statement, the 2 first arguments are the usual .begin() and .end() but then you see that strange if_ as the third argument. if_, like every phoenix statement, returns a functor object created at compile time via template composition (Expression templates). So with this library, you can create inline functors on the fly without the “support code” bloat. You can use your own functors as long as they’re lazy which means they don’t do anything before the operator () is called on them. Fortunately, the lib also provides wrapper for “normal” functions.

Now for that code, nothing much to say for the if_ statement, it’s just a lazy version of the classic if keyword. phx::bind is one of the included wrappers, it creates a lazy version of a function passed as the first argument binded with the arguments passed as additional parameters. _1 and _2 are placeholders, they’re the actual parameters passed by the algorithm to the functor and phx::construct returns a new object of the type passed as the template parameter. Knowing that, we can now understand that “phx::bind(std::isalpha, _1, phx::construct())” returns a lazy version of std::isalpha with the current argument from std::for_each binded as the first argument to std::isalpha and an object of type std::locale binded as the second. phx::let’s only purpose is to create scoped local variables. phx::ref returns a reference to the object passed as the parameter. phx::at_c is simple, on a std::pair phx::at_c returns .first and phx::at_c .second.

For more information, consult the boost::phoenix documentation.

With that new tool, we can now more easily than ever use C++ to imitate some FP idioms:

#include
#include
#include
#include
#include
#include

#include
#include
#include
#include
#include
#include
#include
#include

int main(int argc, char *argv[])
{
namespace phx = boost::phoenix;
using namespace phx::arg_names;
using namespace phx::local_names;

std::vector input;
input.push_back(1);
input.push_back(2);
input.push_back(3);
input.push_back(4);
input.push_back(5);
//map( Make a new sequence with all the elements multiplied by 2 )
std::transform(input.begin(),
input.end(),
std::ostream_iterator(std::cout, “, “),
_1 * 2);
std::cout << std::endl; //filter( Make a new sequence containing all the odd numbers ) std::remove_copy_if(input.begin(), input.end(), std::ostream_iterator(std::cout, “, “),
!(_1 % 2));
std::cout << std::endl; //fold/reduce (Builds up and returns a value based on the sequence) //I use std::string here because it makes it easier to show what is //going on exactly. std::vector words;
words.push_back(“H”);
words.push_back(“e”);
words.push_back(“l”);
words.push_back(“l”);
words.push_back(“o”);

//foldl
std::string result = std::accumulate(words.begin(),
words.end(),
static_cast(“”),
_1 + _2);
std::cout << result << std::endl; //foldr result = std::accumulate(words.rbegin(), words.rend(), static_cast(“”),
_1 + _2);
std::cout << result << std::endl; } [/sourcecode] In a near future, we'll probably see more FP techniques being applied to imperative languages because they can make the code cleaner and more expressive without penalties. In the case of C++, like we just saw, they can leverage existing standard library algorithms and make them more convenient.