I think that it is clear that can't do it just splitting existing control group in two parts, as such action could be done in many different ways and researcher could choose most favorable, and also because there could be some interactions inside control group, and also because smaler statistic power.

You can. Cross-validation, the bootstrap, permutation tests - these rely on that sort of procedure. They generate an empirical distribution of differences between groups or effect sizes which replace the assumption of being two normal distributions etc. It would be better to do those with both the experimental and control data, though.