Because access to the lines and info fields of token.File are protected by the mutex of the parent FileSet, all concurrent calls to AddLine in a typical application contend for the same lock, which becomes a major bottleneck when parsing files in parallel, as done by golang.org/x/tools/go/loader for example.

Fine-grained locking, that is, an additional mutex in each token.File that protects these two fields, eliminates the bottleneck. I quickly prototyped and measured it on a realistic application, and execution time fell to 580ms from 700m, an improvement of nearly 20%.

The necessary change is straightforward except for two places:
(1) In its loop, (*FileSet).Write needs to make a copy of each File's lines and infos while holding the file's mutex.
(2) (*File).PositionFor seems to have a pre-existing data race, since it calls File.position without a lock, which calls unpack, which reads the lines and info fields.