The one big exception to this I've found is working with image data. When you have over a million pixels, you'll have a slow time transforming them unless you can process each one within nanoseconds. Even a compiled language like Rust will struggle with image data if you build it in debug mode. So there's really always going to be a place for optimized implementations of those sorts of things.

Then caring about a microptimisation over readability would still be the wrong call