At first reading I thought that was a neat idea, but on reflection (and sadly experience), what runs on one machine and one set of data does not mean it could come up with the "optimal" plan for another machine and its data. The whole purpose of recompile is take advantage of changing data and its statistics and change the query plan accordingly. Data normally not only changes in quantity, but usually in its distribution and relationship to each other. What works now may not be the optimal plan 10 minutes later.