How can std::bitset be faster than std::vector?

Measurements on Visual Studio 2010 show that std::bitset is not generally faster than std::vector<bool>. What the exact reason for this is I cannot say — only that bitset is implemented significantly different from the std::vector full specialization.

std::bitset stores it’s full content in the object via a

template<size_t _Bits>
    class bitset .....

    _Ty _Array[_Words + 1]; // the set of bits
    };

array and that makes large bitset unsuitable to be put on the stack — which isn’t a performance argument per se.

vector<bool> doesn’t suffer from the stack problem, and testing with a size of 1e6 and 1e7 it seems that on my box here querying values in a loop is actually 2x faster with a vector.

Well. I guess the usual timing caveats apply and YMMV, but here’s the test code I used should anyone care to try himself:

The output on my box is:

1
vector<bool> loop with a size of 10000000 and 10 iterations*n: 11187 ms
bitset<10000000> loop with 10 iterations*n: 22719 ms
101250010
Press any key to continue . . .

BitMap.cpp

#include "stdafx.h"
#include "BitMap.h"

using namespace std;

// Global var to prevent optimizer from messing things up
volatile size_t ext;

volatile clock_t t1;
volatile clock_t t2;
double delta1;
double delta2;

int main(int argc, _TCHAR* argv[])
{
  ext = 1;
  printf("%d\n", ext);

  vb_t *const vec = new vb_t(bssz);
  bs_t *const bits = new bs_t(); // must put large bitset on heap

  const int iter = 10;
  delta1=0;
  delta2=0;
  for(int o=0; o<5; ++o) {
    t1 = clock();
    for(int i=0; i!=5; ++i)
      bs_loop(iter, *vec);
    t2 = clock();
    delta1 += t2-t1;
    t1 = clock();
    for(int i=0; i!=5; ++i)
      bs_loop(iter, *bits);
    t2 = clock();
    delta2 += t2-t1;
  }

  delta1 /= CLOCKS_PER_SEC;
  delta2 /= CLOCKS_PER_SEC;
  delta1 *= 1000;
  delta2 *= 1000;

  cout << "vector<bool> loop with a size of " << bssz << " and " << iter << " iterations*n: " << delta1 << " ms\n";
  cout << "bitset<" << bssz << "> loop with " << iter << " iterations*n: " << delta2 << " ms\n";

  printf("%d\n", ext);
  delete vec;
  delete bits;
  return 0;
}

BitMap.h

#pragma once
#include <vector>
#include <bitset>

extern volatile size_t ext;
const size_t bssz = size_t(1e7); // 1e7 ca 10m

using namespace std; // Test code, using here is OK.
typedef vector<bool> vb_t;
typedef bitset<bssz> bs_t;

template<class COLL>
void bs_loop(const int iterations, COLL const& v);

bs_loop.cpp

#include "stdafx.h"
#include "BitMap.h"

template<class COLL>
void bs_loop(const int iterations, COLL const& v)
{
  ext = sizeof(COLL);
  for(size_t i=0; i!=iterations; ++i) {
    ++ext;
    for(size_t j=0, e=v.size(); j!=e; ++j) {
      if(v[j]) {
        --ext;
      }
      else {
        ++ext;
      }
    }
  }
}

template
void bs_loop(const int iterations, vb_t const& v);

template
void bs_loop(const int iterations, bs_t const& v);

Compiler command line:

/Zi /nologo /W3 /WX- /O2 /Oi /Oy- /D "WIN32" /D "NDEBUG"
/D "_CONSOLE" /D "_UNICODE" /D "UNICODE" /Gm- /EHsc /GS /Gy 
/fp:precise /Zc:wchar_t /Zc:forScope /Yu"StdAfx.h" /Fp"Release\BitMap.pch" 
/Fa"Release\" /Fo"Release\" /Fd"Release\vc100.pdb" /Gd /analyze- 
/errorReport:queue 

note the /O2 and the missing /GL (no whole prg opt).