Fostering the development of SPARQL interfaces to heterogeneous databases is a key to efficiently expose legacy data as RDF on the Web. To deal with the variety of modern database formats and query languages, this paper describes a two-step approach to translate a SPARQL query into an equivalent target database query. First, given an xR2RML mapping describing how native database entities can be mapped to RDF, a SPARQL query is translated into a pivot abstract query language independent of the database. In a second step, the pivot query is translated into the target database query language, considering the specific database capabilities. The paper focuses on the first step of the query translation, from SPARQL to a pivot query that takes into account join constraints and SPARQL filters, and embeds conditions entailed by matching SPARQL graph patterns with relevant mappings. It discusses the query optimisations that can be implemented at this level, and briefly describes an application to the case of MongoDB, a NoSQL document store.