Register allocation is an NP-hard problem. It's not feasible to reach a single "good" decision, you'll always have to choose how much compiler time should be assigned to it, and what heuristics to use.
To "lose" 5% I imagine one has to only be "worse" than some other heuristic that they measured the results of.
This isn’t a problem in practice - you don’t need a single best solution, because you can get a good-enough solution very easily - linear scan. Classic example of an over-theorised problem.