g++ -O optimizations

Question

I have the following code:

#include <iostream>
int main()
{
        int n = 100;
        long a = 0;
        int *x = new int[n];

        for(int i = 0; i < n; ++i)
                for(int j = 0; j < n; ++j)
                        for(int k = 0; k < n; ++k)
                                for(int l = 0; l < n; ++l)
                                {   

                                        a += x[(j + k + l) % 100];
                                }   
        std::cout << a << std::endl;
        delete[] x;
        return 0;
}

If I compile without optimizations g++ test.cc and then run time ./a.out it will display 0.7s. However when I compile it with -O, the time decreases by 2 times.

Compiler used

g++ (Ubuntu/Linaro 4.5.2-8ubuntu4) 4.5.2

My question

How can I rewrite the code so when compiling without -O I can obtain same time ( or close to it ) ? In other words, how to optimize nested loops manually ?

Why I ask

I have a similar code which runs about 4 times faster if I use -O optimization.

PS: I've looked in the manual of the compiler but there are way too many flags there to test which one really makes the difference.

Answer 1

Most of the things the compiler optimizes with -O are things on a level below C++. For example, all variables live in memory, including your loop variables. Therefore without optimization, the compiler will most likely on each iteration of the inner loop first read the loop variable to compare it with 0, inside the loop load it again in order to use it for the index, and then at the end of the loop will read the value again, increment it, and write it back. With optimization, it will notice that the loop variable is not changed in the loop body, and therefore does not need to get re-read from memory each time. Moreover it will also note that the address of the variable is never taken, so no other code will ever access it, and therefore writing it to memory can also be omitted. That is, the loop variable will live only in memory. This optimization alone will save three hundred million memory reads and one hundred million memory writes during execution of your function. But since things like processor registers and memory reads/writes are not exposed at the language level, there's no way to optimize it at the language level.

Furthermore, it doesn't make sense to hand-optimize things which the compiler optimizes anyway. Better spend your time at optimizing things which the compiler cannot optimize.