Garbage collection

[nikomatsakis]

As currently implemented, old query results are going to pile up in the hash tables. We need a method that allows users to specify a root set of query results -- that is, queries plus keys. We can then trace all the values reachable from there and evict the rest.

[matklad]

It might be useful to GC dependencies of roots as well. For example, when processing a rarely-changed crate, you might want to parse all files, compute available items and symbol indexes, and then throw-away intermediate results (syntax trees) to save memory.

[nikomatsakis]

Yes, good point.

[nikomatsakis]

So what are the options we need here:

• Some way to specify a set of root queries (probably root queries + keys?)
• Some kind of modes:
  • Keep anything reachable from roots (default)
  • Memory saver (reduces ability to do incremental updates, but saves memory)

If trying to save memory, there are a few more options:

• Drop memoized values from intermediate nodes but keep other info
• Drop all intermediate nodes (no ability to do incremental updates, but saves more memory)
• Perhaps some way to specify on a per-query basis whether they get dropped?

[matklad]

We can also implement LRU using the existing verified_at field.