CODEWITHSUNDEEP
c++compiler-optimization
sks
sks

Reputation: 934

Release mode skipping parts of code

When trying to test the speed of functions, I found that not all parts of my code work in Release mode. The same code works perfectly in Debug mode, though.

I'm using VC++ compiler with /O2 optimization.

Here's the cut out part, which doesn't work.

int main()
{
    boost::timer::auto_cpu_timer t;

    for(int i = 0; i < 1000000; i++)
        gcdb(i, 5);
    return 0;
}

The generated assembly in release mode, the code for the for loop is missing only in this mode.

int main()
{
000000013F8E1280  sub         rsp,88h  
000000013F8E1287  mov         rax,qword ptr [__security_cookie (013F8E7150h)]  
000000013F8E128E  xor         rax,rsp  
000000013F8E1291  mov         qword ptr [rsp+70h],rax  
    boost::timer::auto_cpu_timer t;
000000013F8E1296  lea         rcx,[t]  
000000013F8E129B  mov         edx,6  
000000013F8E12A0  call        boost::timer::auto_cpu_timer::auto_cpu_timer (013F8E2DA0h)  

    for(int i = 0; i < 1000000; i++)
        gcdb(i, 5);
    return 0;
000000013F8E12A5  lea         rcx,[t]  
000000013F8E12AA  call        boost::timer::auto_cpu_timer::~auto_cpu_timer (013F8E2810h)  
000000013F8E12AF  xor         eax,eax  
}

gcdb() is just a function for the finding the GCD of two numbers.

What could be causing this code skipping?

Upvotes: 4

Views: 1369

Answers (2)

Mysticial
Mysticial

Reputation: 471419

What you're seeing here is a compiler optimization known as Dead Code Elimination.

When the compiler sees that the result of some code is not needed, it is free to eliminate it. This is a standard optimization employed by all modern compilers.

A work-around to keep the compiler from optimizing it out is to actually use the output in some way:

int main()
{
    boost::timer::auto_cpu_timer t;

    int sum = 0;

    for(int i = 0; i < 1000000; i++)
        sum += gcdb(i, 5);

    cout << sum << endl;
    return 0;
}

Related: How does GCC optimize out an unused variable incremented inside a loop?

Upvotes: 8

Useless
Useless

Reputation: 67802

If the compiler can prove there are no side-effects and you don't use the result anywhere, it's free to prune that loop entirely.

Try summing the results and then returning the integer sum from main - that way there's an observable side-effect which should stop the optimizer getting too clever.

Upvotes: 2

Related Questions