TL;DR: Don't optimize anything until you have a real-use scenario benchmark.

Problems

Readability
Bijection and MAPPER violations
Premature Optimization

Solutions

Cover your scenarios with tests.
Write readable (and possibly non-performant) code.
Do a real benchmark with real user data. (No, iterating your code 100,000 times might not be a real use case).
If you have conclusive data, you need to improve the benchmark's found bottlenecks using the Pareto principle.
Attack the worst 20% of problems causing 80% bad performance.

Context

In university and online courses, we learn algorithms, data structures, and computational complexity before good design rules.

We tend to overestimate the (possible) performance problems and underestimate code readability and software lifetime.

Premature optimization often has no evidence of solving real problems.

We need to surgically improve our code when the facts tell us we have a real issue.

Sample Code

Wrong

for (k = 0; k < 3 * 3; ++k) {
     i = Math.floor(k / 3);
     j = k % 3;
     console.log(i + ' ' +  j);
  }

// This cryptic piece of code iterates a 
//two-dimensional array
// We don't have proof this will be useful
// In real contexts

Right

for (innerIterator = 0; innerIterator < 3; innerIterator++) {
  for (outerIterator = 0; outerIterator < 3; outerIterator++) {
   console.log(innerIterator + ' ' +  outerIterator);
  }
 }

// This is a readable double for-loop
// 3 is a small number
// No performance issues (by now)
// We will wait for real evidence