-o3 compiler option gives wrong output

Abhishek,

The compiler performs various optimization steps when you enable -o option. See: http://focus.ti.com/lit/ug/spru187o/spru187o.pdf chapter 3 for details. I'll list couple common coding issues that can cause -o3 generated code to not give correct output

• Have a global variable that is modified outside the scope of the file. Define the variable with volatile keyword
• Make sure dependent memory pointers are not declared as 'restrict/const'. This can cause compiler to assume no dependency between pointers and produce incorrect result

I don't know how big your application is, but below is what you can try:

• Build the entire pjt with no -o option. Ensure all is working well
• Selectively enable -o option for individual file. You can do it using CCS Pjt option of adding file specific compiler options
• If you are lucky, you will identify the one file that has issue
• Split the file and try and come to the one loop that has the issue (if you are lucky!)
• Review loop carefully for logic that can go wrong due to assumptions made by compiler when you enable -o option.

Gagan

I cant get this "You can do it using CCS Pjt option of adding file specific compiler options ". Can you explain a bit more about enabling -o3 option for one file at a time. Can you post ur e-mail id too?

see: http://focus.ti.com/lit/ug/spru509h/spru509h.pdf section 4.3.2.2

Gagan

Thanks a lot for the suggestion Gagan!

I ran into a similar issue and I fixed mine by selective optimization of the files and identifying the file which needed volatile variables. 

Thanks,

Arun

I use the C6748,CCS5.3,my software is below ,when I use the -o2 or -O3 ,the calculate result c1 is wrong while I use the -o1 the result is right ,is there any wrong in my software ? How can I deal with?

//****************************************************************************

 float DSPF_dp_sp_dotprod( float * x,   float * y,  int nx)
{
     int i;
     float sum = 0;

    _nassert(nx >= 0);
    _nassert(nx % 4 == 0);

    for(i=0;i<nx;i++)
    {
     sum += x[i]*y[i];
    }

    return (sum);
}

float a[600],b[600];
volatile float c1=0,c2=0,yt1,yt2,yt;

void main()
{
 int i;
 for(i=0;i<600;i++)
 {
  a[i]=1.0/(i+1.0);
  b[i]=i*2.1;
 }

 c1=DSPF_dp_sp_dot(a,b,600);
}

I tried your test code. I get 1245.352417 for O=2 and  1245.352539 for O=3. Moving either the array init code or the DSPF_dp_sp_dotprod() to another file seems to fix it. Or making arrays a and b volatile. I would guess it is a bug with file level optimizations. You should submit a complete test case to the "Development Tools"->"TI C/C++ Compiler " forum.