[mmiao]

But we need extra computation to move forward/backward, which is not suitable for the problem

[numba888]

one xor, yes, but on most systems smaller memory footprint is better. there are many more nodes than iterators. so savings can be significant.

another possible optimization is when 'key' fits in pointer or two. which is usually the case.

wonder if r1 or 03 can do it, with the right instructions of course.

[numba888]

> wonder if r1 or 03 can do it, with the right instructions of course.

for o3-mini-high it's too difficult to write even simple ebt. not many implementations on the web to learn from(?) reasoning isn't strong enough for this. besides what it produced is very inefficient.