His third point is addressable by using LLVM as your target. Optimization can occur at compile time, at link time, and at runtime, if you decide to do the final LLVM-to-machine-language translation at the latter point.

Inter-module optimizations (e.g., removing dead code) can occur at link time as well.

So there are better tools now that compiler writers can target to start making progress in at least one of these areas.