Strange performance behavior

[Bruce Adams]

Bill Baxter Wrote:

> Bill Baxter wrote:
> > Marius Muja wrote:
> >> I have noticed the following strange (at least for me) performance 
> >> behavior with one of my programs. It is a program that does some 
> >> scientific computations and while trying to optimize it I noticed that 
> >> the code from case_B below executes faster (almost twice as fast) as 
> >> the code in case_A. This is a bit counterintuitive for me, since in 
> >> case _B there is also the cost of the function call (or should be the 
> >> same if the function is inlined).
> >> Can anybody shed some light on why it's behaving this way?
> >>
> >> case_A:
> >> -------------------------------
> >> foreach (i,index; indices) {
> >>    foreach (k, inout value; centers[belongs_to[i]])
> >>     value += vecs[index][k];
> >> }
> >> ----------------------------------
> >>
> >> case_B:
> >> -------------------------------
> >> void addTo(T,U)(T[] a, U[] b) {
> >>    foreach(index, inout value; a) {
> >>       value += b[index];
> >>    }
> >> }
> >> ....
> >> foreach (i,index; indices) {
> >>    addTo(centers[belongs_to[i]],vecs[index]);
> >> }
> >> _______________________________
> > 
> > My guess would be the compiler's failure to optimize[*] away the [index] 
> > indexing in A.  So you do 2 lookups per iteration rather than just one. 
> >  If so then this should be just as fast as case_B:
> > 
> > foreach (i,index; indices) {
> >    auto vecs_i = vecs[index];
> >    foreach (k, inout value; centers[belongs_to[i]])
> >       value += vecs_i[k];
> > }
> > 
> Forgot the footnote!:
> 
> [*] It may not really be a failure of the optimizer -- maybe it's 
> unreasonably difficult to determine absolutely that vecs won't changed 
> by the operation '+= vecs[index][k]'.
> 
> --bb

<shields up>
This is exactly the sort of thing where const would help :)
</shields down>