The Type-Traits Library: Optimization

The type-traits library has two main goals: correctness and optimization. Today, I write about optimization.

This post is my last post in my miniseries about the type-traits library. I have already written the following posts:

• The Type-Traits Library: Type Checks
• The Type-Traits Library: Type Comparisons
• The Type-Traits Library: std::is_base_of
• The Type-Traits Library: Correctness

Before I start to write about optimization in C++, I want to tell a short anecdote. I often have the following conversation with my students in my classes:

• Me: Why do we have the feature ABC in C++?
• Student: I don't know.
• Me: If you don't have an answer, just say performance. This always works in C++.

So, let me write about the type-traits library from the optimization perspective.

Optimization

The idea is quite straightforward and used in the Standard Template Library (STL). If the elements of a range are simple enough, the algorithms of the STL like std::copy, std::fill, or  std::equal are directly applied to memory. Instead of using std::copy to copy each element one by one, all is done in one big step. Internally, C functions like memcmp, memset, memcpy, or memmove are used. The small difference between memcpy and memmove is that memmove can deal with overlapping memory areas.

The implementations of the algorithm std::copy, std::fill, or std::equal use a simple strategy.  std::copy is like a wrapper. This wrapper checks if the elements are simple enough. If so, the wrapper will delegate the work to the optimized copy function. If not, the conservative copy algorithm is used. This conservative one copies each element after the other. To make the right decision, the functions of the type-traits library are heavily used.

The graphic shows the general strategy:

That was the theory, but here is the practice. Which strategy is used by std::fill?

std::fill

std::fill assigns each element in the range a value. The listing shows a GCC-inspired implementation of std::fill.

// fillGCC.cpp
 
#include <cstring>
#include <chrono>
#include <iostream>
#include <type_traits>

namespace my{

  template <typename I, typename T, bool b>
  void fill_impl(I first, I last, const T& val, const std::integral_constant<bool, b>&){
    while(first != last){
      *first = val;
      ++first;
    }
  }

  template <typename T>                                                              // (2)
  void fill_impl(T* first, T* last, const T& val, const std::true_type&){
    std::memset(first, val, last-first);
  }

  template <class I, class T>
  inline void fill(I first, I last, const T& val){
    typedef std::integral_constant<bool,std::is_trivially_copy_assignable<T>::value 
            && (sizeof(T) == 1)> boolType;                                           // (1)
    fill_impl(first, last, val, boolType());
  }
}

const int arraySize = 100'000'000;
char charArray1[arraySize]= {0,};
char charArray2[arraySize]= {0,};

int main(){

  std::cout << '\n';

  auto begin = std::chrono::steady_clock::now();
  my::fill(charArray1, charArray1 + arraySize,1);
  auto last =  std::chrono::steady_clock::now() - begin;
  std::cout <<  "charArray1: " << std::chrono::duration<double>(last).count() << " seconds\n";

  begin = std::chrono::steady_clock::now();
  my::fill(charArray2, charArray2 + arraySize, static_cast<char>(1));
  last=  std::chrono::steady_clock::now() - begin;
  std::cout <<  "charArray2: " << std::chrono::duration<double>(last).count() << " seconds\n";

  std::cout << '\n';

}

Back to the code example. If the expression boolType() in line (1) is true, the optimized version of my::fill_impl in lines 2 is used. This variant fills the entire memory of 100 million entries with the value 1.  sizeof(char) is 1.

What's about the performance of the program? I compiled the program without optimization to measure the non-optimized performance.

The optimized version in line (2) is about ten times faster. Interestingly, when I enable full optimization, both variants are equally fast, because the compiler generates the same code for both variants. Also, the generic version (line (3)) uses memset: fillGCC.cpp with maximum optimization on Compiler Explorer.

I presented an old GCC implementation of std::fill, because the newer ones are not so easy to read. Here are the essential parts of the GCC 6 implementation.

std::fill

// fill  
// Specialization: for char types we can use memset.                   
template<typename _Tp>
  inline typename
  __gnu_cxx::__enable_if<__is_byte<_Tp>::__value, void>::__type     // (1)
  __fill_a(_Tp* __first, _Tp* __last, const _Tp& __c)
  {
    const _Tp __tmp = __c;
    if (const size_t __len = __last - __first)
    __builtin_memset(__first, static_cast<unsigned char>(__tmp), __len);
  }

The GCC 6 implementation uses SFINAE. The full specialization of the function template __fill_a use __builtin_memset. The key part in this implementation is line (1): __gnu_cxx::__enable_if<__is_byte<_Tp>::__value, void>::__type. Let me rewrite this expression in a human-readable way and use the official names.

std::enable_if<std::is_byte<Tp>::value, void>::type

The expression checks first if the template parameter TP is a byte:std::is_byte<T>::value. If this expression evaluates to true thanks to std::enable_if from the type-traits library SFINAE kicks in. SFINAE stands for Substitution Failure Is Not An Error and applies during overload resolution of a function template. It means that if substituting the template parameter fails, the specialization is discarded from the overload set but this failure causes no compiler error. This means in this concrete case: When the condition std::is_byte<T>::value returns false, this full specialization is discarded and another version of __fill_a is used.

What's next?

First, I make a Christmas break of two weeks. My next post will be published on 10th January 2022. I will write about constexpr functions because they have a lot in common with templates and become more powerful with C++20.

Second, for a long time,I would like to improve my professional teaching of C++. Therefore, I plan to start a mentoring program for C++. Soon I will publish more details about my idea.