For large number of deserialized comments (~100k) the way how we try to attach them to declaration in completion comments is too expensive.
Currently we're sorting all the comments by source location (expensive) and later bisecting them for every interesting declaration to find the closest comment before and after the declaration documentation comment.
This patch tries to just iterate over unsorted comments and see if there's any of the decls we're interested in nearby.
Why not DenseMap?