Modern C++ for C Programmers: Part 3

Welcome back! In part 2 I discussed basic classes, threading, atomic operations, smart pointers, resource acquisition and (very briefly) namespaces.

In this part we continue with further C++ features that you can use to spice up your code ’line by line’, without immediately having to use all 1400 pages of ‘The C++ Programming Language’.

Various code samples discussed here can be found on GitHub.

If you have any favorite things you’d like to see discussed or questions, please hit me up on @bert_hu_bert or bert@hubertnet.nl

Inheritance & polymorphism

These are features that are sometimes useful, but by no means the first things to reach for. It is entirely possible to write C++ that never uses any kind of inheritance. In other words, there is no pressure or need to be “all object oriented”.

But, it does have its uses. For example, an event-processing library may have to deal with arbitrary kinds of events, all of which have to pass through a single API. It works like this:

class Event
{
public:
	std::string getType();
	struct timeval getTime();
	virtual std::string getDescription()=0;

private:
	std::string m_type;
	struct timeval m_time;
};

This is the base class called Event. It has three methods, two of which are normal member functions: getType() and getTime(). These return data from the two member variables that every Event shares.

The third one getDescription() is virtual which means it is different in derived classes. In addition, we have set the function to zero which means that every derived class will have to define this function. In other words, you can’t do this:

Event e; // will error on 'pure virtual getDescription'

To make an actual Event that works, we do:


class PortScanEvent : public Event
{
public:
	virtual std::string getDescription() override
	{
		return "Port scan from "+m_fromIP;
	}	
private:
	std::string m_fromIP;
};

This defines a derived class that inherits from Event, which is its ‘base class’. Note how we define getDescription here, and that it is flagged with override which means the compiler will error out unless we are actually overriding a base class method.

Let’s assume we’ve also created an ICMPEvent, we could now write:

  PortScanEvent pse;
  cout << pse.getType() << endl; // "Portscan"

  ICMPEvent ice;
  cout << ice.getType() << endl; // "ICMP"

This is all entirely conventional, and does not do any magic, except that we are using an object that is partially defined in its base class.

However, we can also do:

  Event* e = &ice;
  cout << e->getDescription() << endl; // "ICMP of type 7"

  e = &pse;
  cout << e->getDescription() << endl; // "Portscan from 1.2.3.4"

This defines a pointer to an Event, in which we first store a pointer to an ICMPEvent. And lo, it continues to function as an ICMPEvent, even when stored in an Event pointer. The next two lines demonstrate how this also works for a PortScanEvent.

Event contains metadata which lets us know at runtime what type it really holds, and this metadata is consulted before doing any call that needs to know the actual class in there. This represents overhead, but is is the same kind of overhead generated if classes are simulated, as many C projects end up doing.

Interestingly enough, in the case above, compilers are sometimes able to ‘devirtualise’ calls since from the control flow, they know at compile time what the actual type of Event will be - an optimization not typically implemented in simulated classes in C.

Finally, if you ever need to know, you can find out what the real type of an Event is like this:

  auto ptr = dynamic_cast<PortScanEvent*>(e);
  if(ptr) {
    cout << "This is a PortScanEvent" << endl;
  }

If you got it wrong, ptr will be 0 (or nullptr in modern C++ lingo).

I personally rarely use runtime polymorphism in a project, and then almost exclusively for APIs that need to receive/respond with records or events of different types. It is of great use whenever you have a collection of different ’things’ that need to be stored in a single data structure.

Of special note, the moment you find yourself writing functions that start with a switch statement depending on the type of your struct, you are likely better off using actual C++ inheritance.

A brief note on references

void f(int& x) 
{
	x=0;
}

...

int i = 0;
int& j = i; 
j = 2;
cout << i << endl; // prints 2

f(i); // both i and j are now 0

Up to now, I have neglected to describe references, which made it to two examples in parts 1 and 2 of this series. Technically, a reference is nothing other than a pointer. There is no overhead. They are so much the same one may wonder why C++ bothered to provide this alternate syntax. Pointers already provided for pass by reference semantics.

There are some finer points to be made, but references do save some typing. Functions can now return things ‘by reference’ and not force you to add * or -> to every use for example.

This makes it possible for containers to implement: v[12]=4 for example, which underneath is value_type& operator[](size_t offset). If this returned a value_type* we’d have to type *(v[12])=12 everywhere.

Some discussion on pointers versus references can be found here and here on Stackoverflow.

Templates

As noted in part 1, C++ was designed with the “Zero Overhead” principle in mind, which in its second part states “[no] feature [should] be added for which the compiler would generate code that is not as good as a programmer would create without using the feature”. This is a bold statement.

Most programming languages called “Object Oriented” have made all objects descend (or inherit) from a magic Object base class. This means that to write a container in such a language means writing a data structure that hosts Object instances. If we store an ICMPEvent in there, we do so as an Object.

A problem with this technique is that for C++, it violates the “Zero Overhead” principle. Storing a billion 32 bit numbers in C uses 4GB of memory. Storing a billion Object instances will use no less than 16GB - and likely more.

To adhere to its “Zero Overhead” promise, C++ implemented ’templates’. Templates are like macros on steroids and come with tremendous power and, it has to be said, complication. They are worth it however, as they enable the library to provide generic containers that are as good and likely better than what you could write - and leave open the possibility of creating better versions too.

As a brief example:

template<typename T>
struct Vector
{
  void push_back(const T& t)
  {
    if(size + 1 > capacity) {
      if(capacity == 0) 
        capacity = 1;
      auto newcontents = new T[capacity *= 2];
      for(size_t a = 0 ; a < size ; ++a)
        newcontents[a]=contents[a];
      delete[] contents;
      contents=newcontents;
    }
    contents[size++] = t;
  }
  T& operator[](size_t pos)
  {
    return contents[pos];
  }       

  ~Vector()
  {
    delete[] contents;
  }

  size_t size{0}, capacity{0};
  T* contents{nullptr};
};

This implements a simplistic auto-growing vector of arbitrary type. It is used like this:

  Vector<uint32_t> v;
  for(unsigned int n = 0 ; n < 1000000000; ++n) {
    v.push_back(n);
  }

When the compiler encounters the first line, it triggers the instantiation of the Vector with T replaced by uint32_t. This delivers the exact same code as if you had written it by hand. There is no overhead.

Similarly, functions can be templatized, for example:

// within Vector
  template<typename C>
  bool isSorted(C pred)
  {
    if(!size)
      return true;
    for(size_t n=0; n < size - 1; ++n)
      if(!pred(contents[n], contents[n+1]))
        return false;
    return true;
  }
}

And a sample use:

struct User
{
	std::string name;
	int uid;
};

Vector<User> users;
// fill users

if(users.isSorted([](const auto& a, const auto& b) {
	return a.uid < b.uid;}) 
{
	// do things
}

This code compiles down to be as efficient as if you had written a.uid < b.uid in there yourself. Incidentally, because templates are incredibly generic, you could also pass function pointers or whole object to isSorted, as long as it is something the compiler can call. Incidentally, as shown in part 1, passing a “predicate” like this is also what makes std::sort faster than C qsort.

Based on these templates, C++ offers an array of powerful containers to store data in. Each of these containers comes with an API but also with a performance (scaling) guarantee. This in turn makes sure that implementors have to use state of the art algorithms - and they do.

Note: This also means none of the sample code above should ever see production - std::vector and the associated std::is_sorted are already there and do a far better job.

You may find yourself using a lot of these templated containers and associated algorithms, but very rarely writing any templated functions yourself. And this is pretty good news in two ways - first, writing templated code is harder than you’d think, and it has some surprising syntactical inconveniences. But secondly, almost everything you’d want to write a template for has been written already, so there rarely is a need.

A worked example

Our hallowed “The C Programming Language” contains example code to count the use of C keywords in C programs. The good book rightfully spends a lot of time creating the relevant data structures from scratch.

Within our C++ environment however, we know our journey starts with very capable data structures already, so let’s not only count words in code, let’s also index everything. What follows is a 100 line example that indexes the entire Linux kernel source code (692MB) in 10 seconds, and performs instant lookups.

First, let’s read which files to index:

int main(int argc, char** argv)
{
	vector<string> filenames;

	ifstream ifs(argv[1]);
	std::string line;

	while(getline(ifs, line))
		filenames.push_back(line);

	cout<<"Have "<<filenames.size()<<" files"<<endl;

We read a file with filenames to index into a std::vector. Earlier I noted the C++ iostreams are in no way mandatory and that you can continue to use C stdio, but in this case iostreams are a convenient way to read lines of text of arbitrary length.

Next up, we read and index each file in turn:

	unsigned int fileno=0;
	std::string word;

	for(const auto& fname : filenames) {   // "range-based for"
		size_t offset=0;
    		SmartFP sfp(fname.c_str(), "r");
		while(getWord(sfp.d_fp, word, offset)) {
			allWords[word].push_back({fileno, offset});
    		}
		++fileno;
  	}

This special form of for loop iterates over the contents of the std::vector filenames. This syntax works on all standard C++ containers, and will work on yours too if you adhere to some simple rules.

In the next line we use SmartFP as defined in part 2 of this series. SmartFP internally carries a C FILE*, which means we get the raw speed of C stdio.

getWord can be found in our sample code GitHub, the only special thing to note there is that std::string word gets passed as a reference. This means we can keep reusing the same string instance, which saves a ton of malloc traffic. It also passes offset by reference, and this gets updated to the offset where this word was found in the file.

The allWords line is where the action happens. allWords is defined like this:

struct Location
{
	unsigned int fileno;
	size_t offset;
};

std::unordered_map<string, vector<Location>> allWords;

What this does is create an unordered associative container, keyed on a string. Per entry it contains a vector of Location objects, each of which represents a place where this word was found.

Internally, the unordered containers are based on the hash of the key and by default C++ knows how to hash most primitive data types already. Compared to an std::map, which is ordered, the unordered variant is twice as fast for this usecase.

To actually look something up, we do:

while(getline(cin, line)) {
	auto iter = allWords.find(line);

	if(iter == allWords.end()) {
		cout<<"Word '"<<line<<"' was not found"<<endl;
		continue;
	}

	cout<<"Word '"<<line<<"' occurred "<<iter->second.size()<<" times: "<<endl;

	for(const auto& l : iter->second) {
		cout<<"\tFile "<<filenames[l.fileno]<<", offset "<<l.offset<<endl;
	}
}

This introduces the concept of an iterator. Sometimes an iterator is nothing more than a pointer, sometimes it is more complex, but it always denotes a ‘place’ within a container. There are two magic places, one begin(), one end(). To denote that nothing was found, the end() iterator is returned, and this is what we check against in the next line.

If we have found something, we use iterator syntax to print how many hits we found: iter->second.size(). For an associative container, an iterator has two parts, conveniently called first and second. The first one has the key of the item, the other one provides access to the value found there (or, the actual item).

In the final three lines, we again use range-based for syntax to loop over all the Locations for the word we searched for, and print the details.

Now, to calibrate, we spent less than 50 unique lines on this using only default functionality, is this any good?

$ find ~/linux-4.13 -name "*.c" -o -name "*.h" > linux
$ /usr/bin/time ./windex linux < /dev/null
Have 45000 files
...
45000/45000 /home/ahu/linux-4.13/tools/usb/ffs-test.c, 103302249 words, 607209 different
Read 692542148 bytes
Done indexing
9.62user 1.09system 0:10.73elapsed 99%CPU (0avgtext+0avgdata 2047712maxresident)k

This represents an indexing speed of around 65MB/s, covering 692MB of data. Lookups are instantaneous. Memory usage, at 2GB is around 20 bytes per word which is not too shabby, since this includes storage of the word itself. If we add __attribute__((packed)) or the equivalent to our struct Location, storage requirements decrease to 1.4GB or 15 bytes per word - close to the theoretical minimum of 12.

On GitHub you can find the source code of the indexer - with the additional functionality that it can do prefix searches as well.

Summarising

C++ offers inheritance and polymorphic classes, and these are sometimes useful, definitely better than the hand-coded equivalent, but absolutely not mandatory to use. As an alternative to deriving everything from Object, as some languages do, C++ offers templates, which provide generic code that works on arbitrary types. This is what enables std::sort to be faster than qsort in C.

Templates are very powerful, and underlie the useful array of C++ standard containers like std::map, std::unordered_map, std::vector. Using range-based for loops and iterators, containers can be interrogated and modified.

In part 4 we continue to explore C++ and containers, and if you have any favorite things you’d like to see discussed or questions, please do hit me up on @bert_hu_bert or bert@hubertnet.nl.