Effect of compiler options and coding techniques

    for (i=0; i<n; i++) {
	a[i] += b[i];
    }

Code and results

Code	Code with unroll pragma	Driver	Results
s1.c	s2.c	s1driver.c	file

The general recommendations based on these results are:

• C code must use pragma disjoint wherever possible
• pragma unroll can boost performance
• The __alignx pseudo function helps sometimes but hurts other times. In particular, in this example, it decreased performance when pragma unroll was used.
• const did not help (or hurt) when pragma disjoint was used.
• Using pragma unroll with -O3 gave performance that was almost as good as -O5, suggesting that codes that cannot be compiled correctly at -O5 may benefit from pragma unroll.

Interestingly enough, on this example no explicit code tranformations were necessary; the use of pragmas were sufficient. The winning version was (with -O5) was

void add2( double  *a, double *b, int n )
{
#pragma disjoint (*a,*b)
    int i;
#pragma unroll(10)
    for (i=0; i<n; i++) {
	a[i] += b[i];
    }
}