Question

float mixValue = ... //in range -1.0f to 1.0f
for(... ; ... ; ...  ) //long loop
{
    float inputLevel = ... //in range -1.0f to 1.0f
    if(inputLevel < 0.0 && mixValue < 0.0)
    {
        mixValue = (mixValue + inputLevel) + (mixValue*inputLevel);
    }
    else
    {
        mixValue = (mixValue + inputLevel) - (mixValue*inputLevel);
    }
}

just a simple question, can we calculate mixValue without branching? or any other optimization suggestion, such as using SIMD?

edit:
just for more information, I ended up
using this solution, based on chosen answer:

const float sign[] = {-1, 1};
float mixValue = ... //in range -1.0f to 1.0f
for(... ; ... ; ...  ) //long loop
{
    float inputLevel = ... //in range -1.0f to 1.0f
    unsigned a = *(unsigned*)(&mixValue);
    unsigned b = *(unsigned*)(&inputLevel);

    float mulValue = mixValue * inputLevel * sign[(a & b) >> (8*sizeof(unsigned)-1)];
    float addValue = mixValue + inputLevel;
    mixValue = addValue + mulValue;
}

thank you.

Answer 1

Inspired by Roku’s answer (which on MSVC++10 branches), this doesn’t seem to branch:

#include <iostream>

using namespace std;
const float sign[] = {-1, 1};
int main() {
    const int N = 10;
    float mixValue = -0.5F;
    for(int i = 0; i < N; i++) {
        volatile float inputLevel = -0.3F;
        int bothNegative = ((((unsigned char*)&inputLevel)[3] & 0x80) & (((unsigned char*)&mixValue)[3] & 0x80)) >> 7;
        mixValue = (mixValue + inputLevel) + (sign[bothNegative]*mixValue*inputLevel);
    }

    std::cout << mixValue << std::endl;
}

Here’s the disassembly, as analyzed by IDA Pro (compiled on MSVC++10, Release mode):

Disassembly http://img248.imageshack.us/img248/6865/floattestbranchmine.png