Optimizing sequential memory access in loops

["Alexander Monakov" <monoid at ispras dot ru>]

Hi,

I have noticed that in examples like these:

void test1(int N, double *a)
{
  int i;
  for (i=0; i<N-1; i++)
    a[i] += a[i+1];
}
void test2(int N, double *a)
{
  int i;
  for (i=0; i<N-1; i++)
    a[i+1] += a[i];
}

GCC generates two memory loads per iteration (I checked svn HEAD on ia64  
and 4.0.3 on i386). If loop unrolling is enabled, GCC is able to eliminate  
unneeded stores in second function, but not in first.

Should GCC be able to eliminate extraneous loads, given its current  
analysis/optimization abilities (i.e. is it a bug, or such analysis is not  
implemented yet)? If not, how would you estimate the difficulty of  
implementing an analysis pass that would improve such cases, and how  
useful it would be (how "real-world" applications would benefit from it)?

Thanks in advance.

Alexander Monakov