One of the usage guidelines for the SFDC WS API (http://blog.sforce.com/sforce/2005/04/performance_tip.html) says

Get familiar with the retrieve call. If you are pulling data from related tables, the retrieve call is a high performance call that you should be using

So I looked up retrieve() in the API doc (http://www.salesforce.com/us/developer/docs/api/Content/sforce_api_calls_retrieve.htm)

It looks like retrieve() is a special purpose call that gives high performance when you want to retrieve a list of records from an object by object Id. It lets you select which fields you want returned (and I believe their order), but does not allow any functions on those fields, or any additional criteria beyond this object list.

At first, this sounds like a very limited case, but in fact this is essentially the case we have in the SFDC connector when MM plans out a query as a dep join, with an SFDC table as the dependent source. In fact the API doc specifically recommends retrieve() for this case:

Client applications can use retrieve() to perform a client-side join. For example, a client application can run a query() to obtain a set of Opportunity records, iterate through the returned opportunity records, obtain the accountId for each opportunity, and then call retrieve() to obtain Account information for those accountIds.

We may want to implement this special case in the SFDC connector - if a query has as criteria only a single IN clause, and that IN clause is on the Id column, then use retrieve rather than query.

If we implement relationship joins in the SFDC connector, then this optimization is largely irrelevant - so this is a potential fallback or stopgap if we don't want to tackle relationship joins just yet.