Closer to thr 5% . Between the intermediate code and final code writing there is an optimization stage. The compiler can reduce redundant code and adjust based on machine. i.e. my understanding is an old 4700 can have different instruction sets available than the latest intel.gen chip features. Rather than compile for generic x86 the optimization phase can tailor to the machine's hardware. The benefits are like car tuning, at some point you only get marginal gains. But if squeezing out every drop of performance and reducing bytes is your thing then the wasted compiling time may not been seen as waste.