Once the global features are known (arch, os, format, etc.), then prune logic from the rules that won't ever match. This way, the same global feature are not re-evaluated over and over again, for each instance of the scope.

Edit:
I think it will be fine to prune logic away from the trees, but I don't think its easy to "pre-evaluate" global features. We would have to find a way to extend the logic tree nodes to cache a "pre-evaluated" result, which I think might be tricky/complex/fragile.