C++ Core Guidelines: Rules to Performance

Before I write about the rules of performance I will do a very simple job. Accessing the elements of a container one by one.



Here is the last rule to arithmetic.

ES.107: Don’t use unsigned for subscripts, prefer gsl::index

Did I say that this is a simple job? Honestly, this was a lie.  See what all can go wrong. Here is an example of an std::vector.

vector<int> vec = /*...*/;

for (int i = 0; i < vec.size(); i += 2)                    // may not be big enough (2)
    cout << vec[i] << '\n';
for (unsigned i = 0; i < vec.size(); i += 2)               // risk wraparound       (3)
    cout << vec[i] << '\n';
for (auto i = 0; i < vec.size(); i += 2)                   // may not be big enough (2)
    cout << vec[i] << '\n';
for (vector<int>::size_type i = 0; i < vec.size(); i += 2) // verbose               (1)
    cout << vec[i] << '\n';
for (auto i = vec.size()-1; i >= 0; i -= 2)                // bug                   (4) 
    cout << vec[i] << '\n';
for (int i = vec.size()-1; i >= 0; i -= 2)                 // may not be big enough (2)
    cout << vec[i] << '\n';


Scary? Right! Only line (1) is correct. It may happen in the lines (2) that the variable i is too small. The result may be an overflow. This will not hold for the line (3) because i is unsigned. Instead of an overflow, you will get a modulo operation. I wrote about this nice effect in my last post: C++ Core Guidelines: Rules to Statements and Arithmetic. To be more specific, it was rule ES.106.

Line 4 is left. This is my favourite on. What is the problem? The problem is that vec.size() is of type std::size_t. std::size_t is an unsigned type and can, therefore, not represent negative numbers. Imagine what would happen if the vector is empty. This means that vec.size() -1 is -1. The result is that we get the maximum value of type std::size_t.

The program index.cpp shows this strange behaviour.


// index.cpp

#include <iostream>
#include <vector>

int main(){
    std::cout << std::endl;
    std::vector<int> vec{};
    auto ind1 = vec.size() - 1 ;
    int ind2 = vec.size() -1 ;
    std::cout << "ind1: " << ind1 << std::endl;
    std::cout << "ind2: " << ind2 << std::endl;
    std::cout << std::endl;


 And here is the output:


The guidelines suggest that the variable i should be of type gsl::index.


for (gsl::index i = 0; i < vec.size(); i += 2)             // ok
    cout << vec[i] << '\n';
for (gsl::index i = vec.size()-1; i >= 0; i -= 2)          // ok
    cout << vec[i] << '\n';

If this is not an option for you, use the type std::vector<int>::size_type for i.

Performance is the domain of C++! Right? Therefore I was quite curious to write about the rules to performance. But this is hardly possible because most of the rules lack the beef. They just consist of a title and a reason. Sometimes even the reason is missing.

Anyway. Here are the first rules:

 Instead of writing general remarks to general rules I will provide a few examples to this rules. Let's start with rules Per.4, Per.5, and Per.6

Per.4: Don’t assume that complicated code is necessarily faster than simple code

Per.5: Don’t assume that low-level code is necessarily faster than high-level code

Per.6: Don’t make claims about performance without measurements


Before I continue to write I have to make a disclaimer: I do not recommend to use the singleton pattern. I only want to show that complicated and low-level code does not always pay off. To prove my point I have to measure the performance.

Long, long ago I wrote about the thread-safe initialisation of the singleton pattern in my post: Thread-safe initialization of a singleton.The key idea of the post was, to invoke the singleton pattern 40.000.000 times from four threads and measure the execution time. The singleton pattern will be initialised in a lazy fashion; therefore, the first call has to initialise it.

I implemented the singleton pattern in various ways. I did it with a std::lock_guard and the function std::call_once in combination with the std::once_flag. I did it with a static variable. I even used atomics and broke the sequential consistency for performance reasons.

To make my pointer clear. I want to show you the easiest implementation and the most challenging one.

The easiest implementation is the so-called Meyers singleton. It is thread-safe because of the C++11-standard guarantees that a static variable with block scope will be initialised in a thread-safe way.


// singletonMeyers.cpp

#include <chrono>
#include <iostream>
#include <future>

constexpr auto tenMill= 10000000;

class MySingleton{
  static MySingleton& getInstance(){
    static MySingleton instance;                         // (1)
    // volatile int dummy{};
    return instance;
  MySingleton()= default;
  ~MySingleton()= default;
  MySingleton(const MySingleton&)= delete;
  MySingleton& operator=(const MySingleton&)= delete;


std::chrono::duration<double> getTime(){

  auto begin= std::chrono::system_clock::now();
  for (size_t i= 0; i < tenMill; ++i){
      MySingleton::getInstance();                        // (2)
  return std::chrono::system_clock::now() - begin;

int main(){
    auto fut1= std::async(std::launch::async,getTime);
    auto fut2= std::async(std::launch::async,getTime);
    auto fut3= std::async(std::launch::async,getTime);
    auto fut4= std::async(std::launch::async,getTime);
    auto total= fut1.get() + fut2.get() + fut3.get() + fut4.get();
    std::cout << total.count() << std::endl;



Line (1) uses the guarantee of the C++11-runtime that the singleton will be initialised in a thread-safe way. Each of the four threads in the main function invokes 10 million times the singleton in line (2). In total, this makes 40 million calls.

But I can do better. This time I use atomics to make the singleton pattern thread-safe. My implementation is based on the infamous double-checked locking pattern. For the sake of simplicity, I will only show the implementation of the class MySingleton.

class MySingleton{
  static MySingleton* getInstance(){
    MySingleton* sin= instance.load(std::memory_order_acquire);
    if ( !sin ){
      std::lock_guard<std::mutex> myLock(myMutex);
      sin= instance.load(std::memory_order_relaxed);
      if( !sin ){
        sin= new MySingleton();
    // volatile int dummy{};
    return sin;
  MySingleton()= default;
  ~MySingleton()= default;
  MySingleton(const MySingleton&)= delete;
  MySingleton& operator=(const MySingleton&)= delete;

  static std::atomic<MySingleton*> instance;
  static std::mutex myMutex;

std::atomic<MySingleton*> MySingleton::instance;
std::mutex MySingleton::myMutex;


Maybe you heard that the double-checked locking pattern is broken.  Of course, not my implementation! If you don't believe me, prove it to me. First, you have to study the memory model, think about the acquire-release semantic and think about the synchronisation and ordering constraint that will hold in this implementation. This is not an easy job. But you know, high sophisticated code pays off.

Damn. I forgot the rule Per.6: Here are the performance numbers for the Meyers singleton on Linux. I compiled the program with maximum optimisation. The numbers on Windows were in the same ballpark.singletonMeyers opt

Now I'm curious. What are the numbers for my highly sophisticated code? Let's see which performance we will get with atomics.

singletonAcquireRelease opt

50% percent slower! 50% percent slower and we don't even know if the implementation is correct. Disclaimer: The implementation is correct.

Indeed, the Meyers singleton was the fastest and the easiest way to get a thread-safe implementation of the singleton pattern. If you are curious about the details, read my post: Thread-safe initialization of a singleton.

What's next?

There are more than 10 rules to performance left in the guidelines. Although it is quite challenging to write about such general rules I have for my next post a few ideas in mind.




Thanks a lot to my Patreon Supporters: Eric Pederson, Paul Baxter and Sai Raghavendra Prasad Poosa.



Get your e-book at leanpub:

The C++ Standard Library


Concurrency With Modern C++


Get Both as one Bundle

cover   ConcurrencyCoverFrame   bundle
With C++11, C++14, and C++17 we got a lot of new C++ libraries. In addition, the existing ones are greatly improved. The key idea of my book is to give you the necessary information to the current C++ libraries in about 200 pages.  

C++11 is the first C++ standard that deals with concurrency. The story goes on with C++17 and will continue with C++20.

I'll give you a detailed insight in the current and the upcoming concurrency in C++. This insight includes the theory and a lot of practice with more the 100 source files.


Get my books "The C++ Standard Library" (including C++17) and "Concurrency with Modern C++" in a bundle.

In sum, you get more than 550 pages full of modern C++ and more than 100 source files presenting concurrency in practice.


My Newest E-Books

Course: Modern C++ Concurrency in Practice

Course: C++ Standard Library including C++14 & C++17

Course: Embedded Programming with Modern C++

Course: Generic Programming (Templates)

Course: C++ Fundamentals for Professionals

Subscribe to the newsletter (+ pdf bundle)

Blog archive

Source Code


Today 6832

Yesterday 7978

Week 31397

Month 185992

All 4806886

Currently are 226 guests and no members online

Kubik-Rubik Joomla! Extensions

Latest comments