And when you do identify the 1%, you need to be testing optimizations with a profiler constantly while optimizing. Profile. Do some optimization. Profile again. Roll back if not successful. Repeat until done. It's impossible to optimize well if you're not doing profiling.
The ultimate tools would be either Intel's profiling tools, or ARM's profiling suite (both very expensive). But MSVC and GCC do a fantastic job of scheduling instructions to avoid pipeline stalls these days, so these deep profiling tools are unlikely to gain more than 2 or 3% performance increases these days. (Worthwhile pretty much only if you're writing GPU drivers for NVidia or AMD).
Taken as a given that 100% of your code is at least algorithmically correct in the first place. (Appropriately better than O(N^2) whenever possible).
- Former writer of graphics drivers, currently audio DSP engineer.