Why does not std::sqrt() work correctly on double data in C++?

Question

Here is a simple code in C++ which I am compiling using TDM-GCC 5.1 32bit (gcc-5.1.0-tdm-1-core) on a Windows 8.1 64bit Intel core i-7 machine.

#include <iostream>
#include <cmath>

using namespace std;

int main()
{
    double a, b;
    a = 0.002;
    b = 0.004;
    cout << std::sqrt(a*a+b*b)<< endl;
    return 0;
}

But on running, it gives the following output

nan

However, it works correctly with TDM-GCC 32bit compiler on a 32 bit machine.

Also, when I am changing the datatype to float, it is working properly on both 32bit and 64 bit computers.

What is wrong here?

Use `-S` and compare the assembly output between the two GCCs. Or, at least, attach the two executables. — Matteo Italia, Dec 01 '15 at 10:09
Seems to [work as indented](http://coliru.stacked-crooked.com/a/f0bc69e84e0009b5). — Kerrek SB, Dec 01 '15 at 10:09
Interestingly, something like Clang won't even [execute the computation](https://goo.gl/9WD1V1) at runtime. — Kerrek SB, Dec 01 '15 at 10:11
Compile with `g++ -O -fverbose-asm -S` and look inside the produced `.s` assembler file — Basile Starynkevitch, Dec 01 '15 at 10:11
Here is the assembly output link: [link](http://pastebin.com/ikdqVfxM) from `g++ -O -fverbose-asm -S` — , Dec 01 '15 at 10:21
I need some time to compare the two versions since I don't have the 32bit computer right now. — , Dec 01 '15 at 10:27
Works well for me - Apple LLVM version 7.0.0 (clang-700.1.76) — Jean-Baptiste Yunès, Dec 01 '15 at 10:45
Glad you found what the problem was. Can you post your final update as an answer? — Not a real meerkat, Dec 01 '15 at 12:50

score 1 · Accepted Answer · answered Dec 01 '15 at 13:26

The problem is as follows:

It was compiled inherently in the Code::Blocks with the following command (narrowed):

g++.exe -L"C:\Program Files (x86)\GnuWin32\lib" -o main.exe main.cpp -lm

where C:\Program Files (x86)\GnuWin32\lib has an old math library named libm.a.

So, on putting -lm along with -L"C:\Program Files (x86)\GnuWin32\lib", it is found using Process Explorer that main.exe is using that corresponding old libm5.dll which seems to be incompatible.

So, if at least one of the options -lm or -L"C:\Program Files (x86)\GnuWin32\lib" is removed and compiled as follows:

g++.exe -L"C:\Program Files (x86)\GnuWin32\lib" -o main.exe main.cpp

or

g++.exe -o main.exe main.cpp -lm

or

g++.exe -o main.exe main.cpp

it is not linking the old libm.a and hence running correctly.

Why does not std::sqrt() work correctly on double data in C++?

1 Answers1