You can optimize one specific query (quite painstakingly, I believe) to beat a general db, but it is very far from clear that “for loops will beat higher level abstractions”, imo.
With my own set-up I can have my data in pre-calculated immutable tables that are memory mapped off disk. That means I can write lock-free code. My updates are large hour-long batch operations off a written journal.
For the lexicon I'm also rolling a special one-way compression scheme/hash function that guarantees uniqueness but only works due to quirks with the data.
General purpose databases can't do these things, because then they wouldn't be general purpose databases.
SQL databases perform really well on average given your query is relatively optimized, but they do not represent the pinnacle of performance. Any time you want the fastest most efficient code to deal with a particular use case, it pays enormous dividends to roll your own implementation that has domain awareness. It means you can cut corners general solutions simply can't.
In my experience programming programming with primitives and basic flow control operations frequently tends to be at least be order of magnitude faster than more complex state management paradigms. Compilers are very good at optimizing that style of code. Loops often get unrolled, the and the branch predictor is kept happy. A good compiler may use vector expressions.
In many cases with cold code it flat out doesn't matter, the CPU is plenty fast, but when it does matter, explicit if-and-for code absolutely mops the floor with the alternatives.