Program transformations that are able to rely on the presence of canonical properties of the program undergoing optimization can be written to be more robust and efficient than an equivalent but generalized transformation that also handles non-canonical programs. If a canonical property is required but broken earlier in an earlier transformation, it must be rebuilt (often from scratch). This additional work can be a dominating factor in compilation time when many transformations are applied over large programs. This dissertation introduces a methodology for constructing program transformations so that the program remains in an always-canonical form as the program is mutated, making only local changes to restore broken properties