Skip to main content

XOR (Exclusive OR) for branchless coding

The following example shows the array reversing using the XOR operator. No need to take any additional variable to reverse the array. 
int main(int argc, _TCHAR* argv[])
{
    char str[] = "I AM STUDENT";
    int length = strlen(str);
    for(int i = 0; i < ((length/2)); i++)
    {
        str[i] ^= str[length - (1+i)];
        str[length - (1+i)] ^= str[i];
        str[i] ^= str[length - (1+i)];
    }
    cout << str << endl;
    return 0;
}

The above example is one of the uses of XOR but XOR comes in handy when we can do branchless coding methods like butterfly switch etc. Sometimes this is very effective in speeding up the execution. Let's see one of the uses of XOR in branchless coding. I am taking a simple example of Y = | X |. Yes, I am generating abs of a supplied number. So, my function signature/definition in C++ looks like below:

int absoluteBranch(int x) 
{
    if (x < 0) 
    {
        return -x;
    }
    else 
    {
        return x;
    }
}
From the C++ code, we can see the branching in code and until runtime, we can't 
definitely say which branch could be executed. If we look at the assembly generated 
by the code also shows branching (Without optimization):
absoluteBranch(int):
        push    rbp
        mov     rbp, rsp
        mov     DWORD PTR [rbp-4], edi
        cmp     DWORD PTR [rbp-4], 0
        jns     .L4
        mov     eax, DWORD PTR [rbp-4]
        neg     eax
        jmp     .L5
.L4:
        mov     eax, DWORD PTR [rbp-4]
.L5:
        pop     rbp
        ret
We have instructions like cmp(compare), jns(jump if not signed) and 
jmp (unconditional jump). With the help of XOR we can completely remove 
this branching of code while calculating abosolute of a signed number.
int absoluteBranchless(int x)
{
    int y = x >> (sizeof(int) * 8 - 1);
    return (x ^ y) - y;
}

Now I don't have any branch in the C++ code and assembly generated out of
it also branchless. Here goes the assembly (Without optimization):

absoluteBranchless(int):
        push    rbp
        mov     rbp, rsp
        mov     DWORD PTR [rbp-20], edi
        mov     eax, DWORD PTR [rbp-20]
        sar     eax, 31
        mov     DWORD PTR [rbp-4], eax
        mov     eax, DWORD PTR [rbp-20]
        xor     eax, DWORD PTR [rbp-4]
        sub     eax, DWORD PTR [rbp-4]
        pop     rbp
        ret


The SAR instruction shifts the MSB and it became all FFs. For positive number it's all 00s.
So for negative number mask became -1 (in two's complement) and for positive number 
it will always be 0. We then XOR the number with mask and subtract with mask, which 
effectively adds +1 in case of negative number and +0 for positive number.
In above example variable y is the mask value.

It's nice, we have no branching!
Labels:

Comments

Rahul Bhattacharjee said…
Remember you have shown this in office and we had some discussions on this as well.

If you cannot afford any extra memory for a new array then this XOR-ing is a good solution.I believe its used in hardware which cannot afford huge memory.
February 4, 2009 at 5:02 PM
Post a Comment

Popular posts from this blog

Reversing char array without splitting the array to tokens

 I was reading about strdup, a C++ function and suddenly an idea came to my mind if this can be leveraged to aid in reversing a character array without splitting the array into words and reconstructing it again by placing spaces and removing trailing spaces. Again, I wanted an array to be passed as a function argument and an array size to be passed implicitly with the array to the function. Assumed, a well-formed char array has been passed into the function. No malformed array checking is done inside the function. So, the function signature and definition are like below: Below is the call from the client code to reverse the array without splitting tokens and reconstructing it. Finally, copy the reversed array to the destination.  For GNU C++, we should use strdup instead _strdup . On run, we get the following output: Demo code
Post a Comment
Read more

A simple approach to generate Fibonacci series via multi-threading

T his is a very simple approach taken to generate the Fibonacci series through multithreading. Here instead of a function, used a function object. The code is very simple and self-explanatory.  #include <iostream> #include <mutex> #include <thread> class Fib { public:     Fib() : _num0(1), _num1(1) {}     unsigned long operator()(); private:     unsigned long _num0, _num1;     std::mutex mu; }; unsigned long Fib::operator()() {     mu.lock(); // critical section, exclusive access to the below code by locking the mutex     unsigned long  temp = _num0;     _num0 = _num1;     _num1 = temp + _num0;     mu.unlock();     return temp; } int main() {     Fib f;          int i = 0;     unsigned long res = 0, res2= 0, res3 = 0;     std::cout << "Fibonacci series: ";     while (i <= 15) {         std::thread t1([&] { res = f(); }); // Capturing result to respective variable via lambda         std::thread t2([&] { res2 = f(); });         std::thread t3(
Post a Comment
Read more

Variadic template class to add numbers recursively during compilation

 The idea of having a class to add numbers (variable parameters) during compilation time recursively. Also wanted to restrict types to a single type while sending parameters to class member function. That said, if we mix int, float and double types to add function shall result in compilation error. How do we achieve this. The below is the code which actually helps to achieve this: <code> #include < fmt/format.h > template < typename T> class MyVarSumClass{     private :         T _sum = 0 ;     public :         template < typename ... TRest>         T add(T num, TRest... nums){             static_assert (std::conjunction<std::is_same<TRest, T>...>{}); /* Assert fails                if types are different */             _sum += num;             return add(nums...); // Next parameter packs gets picked recursively         }         // Base case         T add(T num){             _sum += num;             return _sum;         } }; int main() {     My
Post a Comment
Read more