I dunno, is that really true? The indices into the array can be 4 instead of 8 bytes on a 64-bit system for example, that would double the indices available in

I dunno, is that really true? The indices into the array can be 4 instead of 8 bytes on a 64-bit system for example, that would double the indices available in one cache line. And then you have the possibility of storing the in/out node references as indices also, halving the memory there. And I know that work lists and stacks of nodes are common for certain transformations, so that would also have an impact. And that's just one thing, having an index effectively means that you can split your data structure into multiple arrays ('struct of arrays' fashion), which might mean that you're doing fewer accesses and jumps over all. Considering all of that, I'm really not convinced that plain pointers will beat indices. I also don't see what alias analysis has to do with it in this case, sounds like a compiler could easily do alias analysis on A being accessed only as a T by virtue of A having the type Array<T,N>.