要将一个32位数字除以3，可以将其乘以0x55555556，然后取64位结果的前32位。

现在剩下要做的就是使用位运算和移位来实现乘法…

第一:

x/3 = (x/4) / (1-1/4)

然后求x/(1 - y)

x/(1-1/y)
  = x * (1+y) / (1-y^2)
  = x * (1+y) * (1+y^2) / (1-y^4)
  = ...
  = x * (1+y) * (1+y^2) * (1+y^4) * ... * (1+y^(2^i)) / (1-y^(2^(i+i))
  = x * (1+y) * (1+y^2) * (1+y^4) * ... * (1+y^(2^i))

y = 1/4:

int div3(int x) {
    x <<= 6;    // need more precise
    x += x>>2;  // x = x * (1+(1/2)^2)
    x += x>>4;  // x = x * (1+(1/2)^4)
    x += x>>8;  // x = x * (1+(1/2)^8)
    x += x>>16; // x = x * (1+(1/2)^16)
    return (x+1)>>8; // as (1-(1/2)^32) very near 1,
                     // we plus 1 instead of div (1-(1/2)^32)
}

虽然它使用了+，但有人已经实现了按位操作的add。

以下是我的解决方案:

public static int div_by_3(long a) {
    a <<= 30;
    for(int i = 2; i <= 32 ; i <<= 1) {
        a = add(a, a >> i);
    }
    return (int) (a >> 32);
}

public static long add(long a, long b) {
    long carry = (a & b) << 1;
    long sum = (a ^ b);
    return carry == 0 ? sum : add(carry, sum);
}

首先，请注意

1/3 = 1/4 + 1/16 + 1/64 + ...

现在，剩下的很简单!

a/3 = a * 1/3  
a/3 = a * (1/4 + 1/16 + 1/64 + ...)
a/3 = a/4 + a/16 + 1/64 + ...
a/3 = a >> 2 + a >> 4 + a >> 6 + ...

现在我们要做的就是把a的这些位移位值加在一起!哦!但是我们不能做加法，所以我们必须使用位操作符来编写一个加法函数!如果您熟悉逐位操作符，那么我的解决方案应该看起来相当简单……但以防你不懂，我会在最后讲一个例子。

另一件需要注意的事情是，首先我左移30!这是为了确保分数不会四舍五入。

11 + 6

1011 + 0110  
sum = 1011 ^ 0110 = 1101  
carry = (1011 & 0110) << 1 = 0010 << 1 = 0100  
Now you recurse!

1101 + 0100  
sum = 1101 ^ 0100 = 1001  
carry = (1101 & 0100) << 1 = 0100 << 1 = 1000  
Again!

1001 + 1000  
sum = 1001 ^ 1000 = 0001  
carry = (1001 & 1000) << 1 = 1000 << 1 = 10000  
One last time!

0001 + 10000
sum = 0001 ^ 10000 = 10001 = 17  
carry = (0001 & 10000) << 1 = 0

Done!

这就是你小时候学过的简单加法!

111
 1011
+0110
-----
10001

这个实现失败了，因为我们不能把方程的所有项相加:

a / 3 = a/4 + a/4^2 + a/4^3 + ... + a/4^i + ... = f(a, i) + a * 1/3 * 1/4^i
f(a, i) = a/4 + a/4^2 + ... + a/4^i

假设div_by_3(a) = x的结果，则x <= floor(f(a, i)) < a / 3。当a = 3k时，我们得到错误的答案。

Yet another solution. This should handle all ints (including negative ints) except the min value of an int, which would need to be handled as a hard coded exception. This basically does division by subtraction but only using bit operators (shifts, xor, & and complement). For faster speed, it subtracts 3 * (decreasing powers of 2). In c#, it executes around 444 of these DivideBy3 calls per millisecond (2.2 seconds for 1,000,000 divides), so not horrendously slow, but no where near as fast as a simple x/3. By comparison, Coodey's nice solution is about 5 times faster than this one.

public static int DivideBy3(int a) {
    bool negative = a < 0;
    if (negative) a = Negate(a);
    int result;
    int sub = 3 << 29;
    int threes = 1 << 29;
    result = 0;
    while (threes > 0) {
        if (a >= sub) {
            a = Add(a, Negate(sub));
            result = Add(result, threes);
        }
        sub >>= 1;
        threes >>= 1;
    }
    if (negative) result = Negate(result);
    return result;
}
public static int Negate(int a) {
    return Add(~a, 1);
}
public static int Add(int a, int b) {
    int x = 0;
    x = a ^ b;
    while ((a & b) != 0) {
        b = (a & b) << 1;
        a = x;
        x = a ^ b;
    }
    return x;
}

这是c#，因为这是我手边的东西，但与c的区别应该很小。

使用fma()库函数的解决方案，适用于任何正数:

#include <stdio.h>
#include <math.h>

int main()
{
    int number = 8;//Any +ve no.
    int temp = 3, result = 0;
    while(temp <= number){
        temp = fma(temp, 1, 3); //fma(a, b, c) is a library function and returns (a*b) + c.
        result = fma(result, 1, 1);
    } 
    printf("\n\n%d divided by 3 = %d\n", number, result);
}

请看我的另一个答案。

#include <stdio.h>

typedef struct { char a,b,c; } Triple;

unsigned long div3(Triple *v, char *r) {
  if ((long)v <= 2)  
    return (unsigned long)r;
  return div3(&v[-1], &r[1]);
}

int main() {
  unsigned long v = 21; 
  int r = div3((Triple*)v, 0); 
  printf("%ld / 3 = %d\n", v, r); 
  return 0;
}