Which one is fastest in C: fabsf(-3) or (-1)*(-3)?

Question

Which one is fastest in C to do the L1-norm?

float x = fabsf(-3.0f);

or

float x = -1.0f * (-3.0f);

?

Both return the number x = 3

Answer 1

depends on the library used.

Some use the x86 instruction "fabs" as in Apple's fabfs implementation:

#define __FABSF(x) \
({ \
    float __value, __arg = (x); \
    asm volatile ("fabs %0,%1" : "=f" (__value): "f" (__arg)); \
    __value; \
}) 

just clear the negative bit using a bitmask, as in

float fabsf(const float x) _MATH_REENTRANT
{
    FS_STATIC union float_long fl;

    fl.f = x;
    fl.l &= 0x7fffffff;
    return fl.f;
}

or ReactOS

{
    /* Load the value as uint */
    unsigned int u32 = *(unsigned int*)&x;
 
    /* Clear the sign bit */
    u32 &= ~(1 << 31);
 
    /* Check for NAN */
    if (u32 > 0x7F800000)
    {
        /* Set error bit */
        *(unsigned int*)&x |= 0x00400000;
        return x;
    }
 
    /* Convert back to float */
    return *(float*)&u32;
}

and no doubt some multiply by -1.

so you'll have to write your own benchmark and tell us!