optimization differences between cmd line and code declarations

Apparently optimzations work differently when putting the -O<xxx> in the cmd line and into the code. This is version 4.7.4. (the current version supported by my tool vendor iSystem). Copy-and-paste the following code into a standalone C file and include it in any project with no -O option in the command line:

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

//#pragma GCC optimize ("Ofast")

unsigned long Fibonacci(unsigned long p_Arg)
//unsigned long __attribute__ ((optimize("Ofast"))) Fibonacci(unsigned long p_Arg)
{
    switch(p_Arg)
    {
    case 0:
    case 1: return p_Arg;
    default: return Fibonacci(p_Arg-1)+Fibonacci(p_Arg-2);
    };
}

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

In the second step, uncomment either the pragma or __attribute___ lines and rebuild. The resulting code will (expectedly) be different. In the third step, re-start from the original code (leaving all the commented lines commented) but add the option -Ofast to the command line. The code will be different from step 2, implying that -Ofast means different things when invoked from the command line and within the code.

Is this behavior by design and/or documented anywhere? Apologies if this should have been addressed before, I scanned all QAs regarding optimizations and couldn't find any reference.

Thanks!

Edit Feb 19, 2016: It appears that the optimization is done when -Ofast is within the source file and not the cmd line is the same performed as when -O2 is selected.

Edit Feb. 28, 2016: Reopened due to its still applicability. This is pretty severe in my opinion and possibly trivial to solve, in any case easy to reproduce, why are there no takers? Thanks!