Why virtual calls are harder to optimise

Why do virtual function calls limit what an optimising compiler can do, and which optimisations are open for non-virtual calls?

The target of a virtual call is unknown at compile time, so it usually cannot be inlined. That blocks the follow-on optimisations inlining enables: constant propagation, dead-code removal, loop vectorisation and register allocation across the call, plus the indirect branch can mispredict. Non-virtual calls have a fixed target and can be inlined and optimised freely; compilers may devirtualise when the dynamic type is provable (e.g. final classes).