Isn't that also the insight which led to HAMT (and friends) being so fast compared to older persistent datastructures? Turns out for a while now copying more but denser memory has been way cheaper than chasing pointers.
Brian Cantrill also noticed that in his comparison of his C and Rust versions of statemap[0]: Rust had a way, way better cache behaviour than C (96.9 L1 hit rate to 77.9%, and half the L2 misses although the better L1 behaviour also led to >90% less L2 interactions in the first place), and most of that was attributed to using a btreeset instead of an AVL BST.
Brian Cantrill also noticed that in his comparison of his C and Rust versions of statemap[0]: Rust had a way, way better cache behaviour than C (96.9 L1 hit rate to 77.9%, and half the L2 misses although the better L1 behaviour also led to >90% less L2 interactions in the first place), and most of that was attributed to using a btreeset instead of an AVL BST.
[0] http://dtrace.org/blogs/bmc/2018/09/28/the-relative-performa...