Techniques for composing architectural specifications are emerging in order to facilitate software architecture evolution. However, there is little empirical understanding on whether such techniques scale when they are used to express different types of architectural changes. This paper presents a first comparative evaluation of two significantly-different composition techniques for architectural descriptions. The first technique is fully based on heuristic composition operators, while the second one demands explicit composition specification. Several releases of a software product line were used in our evaluation, and their designs were expressed with an architectural description language, called ACME. Some metrics were used to compute the number of required modifications, syntactic conflicts, and semantic conflicts in composed (output) models produced with both heuristic and non-heuristic compositions. We observed that, in general, the heuristic composition approach outperformed the non-heuristic one mainly due to the narrow set of composition operators supported by the latter.