I learned through three things:

1. writing some basic functions in assembly code by hand for a course (not many)
2. implementing a basic compiler back-end in llvm (any similar IR or assembly target would do)
3. learning the principles other people were using to write fast code (in my case game engine developers)

The first two things helped me understand how common code constructs are translated to assembly, so I can do a rough projection in my head when skimming a C function. Nowadays you can get quite far just by playing around on godbolt.

The third thing helps surface the less visible aspects of CPUs. After learning how a few low-level optimisations work, all the principles and explanations start to repeat, and 90% of them apply to every modern architecture. You can set out with specific high-level questions, like:

• why is iteration faster with an array than a linked list?
• what does vectorisation mean?
• what is a "struct of array" optimisation?
• why does the ECS pattern make game engines fast?

Very quickly you'll find lots of insightful articles and comments explaining things like CPU caching, prefetching, branch prediction, pipelining, etc.

I have no book recommendations for you. I've found all the best information is freely online in blogs and comment sections, and that the best way to direct my learning is to have a project (or get employed to do low-level stuff). Might be different for you though!