Why sqrt in global scope is much slower than std::sqrt in MinGW?

Question

Consider the following code:

#include <cmath>
#include <cstdio>

const int COUNT = 100000000;

int main()
{
    double sum = 0;
    for (int i = 1; i <= COUNT; ++i)
        sum += sqrt(i);
    printf("%f\n", sum);
    return 0;
}

It runs 5.5s on my computer. However, if I change sqrt into std::sqrt, It will run only 0.7s.

I know that if I use sqrt, I'm using the function from C library, and if I use std::sqrt, I'm using the one in <cmath>.

But <cmath> doesn't define one for int, and if I change the type of i into double, they will run for equal speed. So the compiler isn't optimizing for int. This seems to only happen to sqrt in Windows.

So why is std::sqrt much faster than sqrt, but not other functions? And why in Linux they are not?

Answer 1

This is the typical situation in which the -fdump-tree-* switch can give some insight into what is going on:

g++ -fdump-tree-optimized example1.cc

and you get the example1.cc.165t.optimized file. Somewhere inside:

<bb 3>:
_5 = (double) i_2;
_6 = sqrt (_5);
sum_7 = sum_1 + _6;
i_8 = i_2 + 1;

The compiler (gcc v4.8.3) is doing the math with doubles.

Replacing sqrt with std::sqrt what you get is:

<bb 3>:
_5 = std::sqrt<int> (i_2);
sum_6 = sum_1 + _5;
i_7 = i_2 + 1;

Now it uses a different sqrt overload for integers (i_2 is int and sum_6 is double).

As Mike Seymour says in a comment, GCC uses the new overload whether or not you specify C++11.

Anyway under Linux there isn't a sensible performance difference between the two implementations.

Under Windows (MinGW) this is different since sqrt(double) calls into msvcrt.