Subtract One With Only Bit Operations In C/++

In the tutoring session, I was asked to swap the exponent part of a (32-bit) floating point number to be a new value without any arithmetic operations. That is, let $x$ be a floating number in the binary format like the following:

$$1.\dots\times 2^{\texttt{m}}$$

; we want to change the exponent $m$ to be some new value $n$. The result of this swapping is

$$1.\dots\times 2^{\texttt{n}}$$

In the IEEE standard, a floating number is represented as

$$x = \underbrace{\texttt{\_}}_{\texttt{Sign}}\underbrace{\texttt{\_\_\_\_\_\_\_\_}}_{\texttt{Exponent}}\underbrace{\texttt{\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_}}_{\texttt{Significand}}$$

We can use a bit mask to extract the exponent part of $x$:

union float_bits {
    float f;
    unsigned u;
};

float swap_exp(float f, int8_t exp) {
    union float_bits x;
    x.f = f;
    uint32_t mask = 0x7f800000;
    uint32_t old_exp = x.u & mask;

    // more stuff needs to be done here
    // so that we can swap the old_exp
    // to be the exp

    return x.f;
}

Recall that the exponent of the binary representation of a floating number needs to be offset by a bias number before it can be the real exponent part in the floating point bit representation. In this case, the bias number is $127$.

We can simply take the previous code and add the bias number to the new exponent $n$ to get the new exponent part of the floating number.

union float_bits {
    float f;
    unsigned u;
};

float swap_exp(float f, int8_t exp) {
    union float_bits x;
    x.f = f;
    uint32_t mask = 0x7f800000;
    uint32_t old_exp = x.u & mask;
    uint32_t new_exp = exp + 127;

    x.u = (x.u & ~old_exp) | (new_exp << 23);

    return x.f;
}

But NO ARITHMETIC OPERATIONS are allowed. So we can’t just add the exp with the bias number. We need something trickier here. Ryan and I thought of a solution that uses only bit operations.

For any positive $n$ added with the biased number $127$ is equal to $128 + n - 1$. That is,

$$\begin{aligned} \texttt{0x0111\_1111} + n &= \texttt{0x1000\_0000} + n - 1 \\ &= \texttt{0x1000\_0000 | (n - 1) } \end{aligned}$$

Therefore, if we know what $n - 1$ is, we can | it with the mask 0x1000_0000.

So how do we compute $n-1$ without arithmetic operations? Well,

uint32_t sub_one(uint32_t n) {
    uint32_t i;
    for (i = 1; i != (i & n) ; n |= i, i <<= 1) ;
    return n & ~i;
}

When we subtract one from a binary, if the $0$ th place is $0$, we need to burrow from higher places and the $0$ th place with turn to $1$. For example, if we want to subtract one from $1010$, we need to borrow from the $1$st place and the result will be $1001$. This means, when we subtract one from a binary number, we substitute all zeros we see to be $1$ before we see any $1$; then we substitute the first $1$ to be $0$. The for loop above does exactly this. It stops when it finds the first $1$ but before that it substitutes any $0$ in n to be $1$. After the for loop, we set the first $1$ we saw to be $0$. Then we subtracted $1$!

For any negative $n$, we can get the real exponent bits by

$$(\texttt{0x0111\_1111 \& n}) - 1$$

So the final code is

union float_bits {
    float f;
    unsigned u;
};

uint32_t sub_one(uint32_t n) {
    uint32_t i;
    for (i = 1; i != (i & n) ; n |= i, i <<= 1) ;
    return n & ~i;
}

float swap_exp(float f, int8_t exp) {
    union float_bits x;
    x.f = f;
    uint32_t mask = 0x7f800000;
    uint32_t old_exp = x.u & mask;
    uint32_t new_exp = exp == 0 ? 127 :
                (
                    exp > 0 ?
                    (0b10000000 | sub_one(exp)) :
                    sub_one(0b01111111 & exp)
                );

    x.u = (x.u & ~old_exp) | (new_exp << 23);

    return x.f;
}

int main() {
    float f = 16.0;  // 1.0 * 2^4
    printf(
        "The output is %.5f and the expected output is 32.0\n",
        swap_exp(f, 5)
    );
    printf(
        "The output is %.5f and the expected output is 0.5\n",
        swap_exp(f, -1)
    );
    printf(
        "The output is %.5f and the expected output is 1.0\n",
        swap_exp(f, 0)
    );
    return 0;
}

There are inf and nan but we are not going to deal with them here.

The code format is not working very well. I will fix it later! :)