incr.comp.: Do not hard-wire DepGraph into Session and CrateStore.

[michaelwoerister]

In the current setup, we create the DepGraph very early, so it can be put into the CrateStore and the Session object. This means that we construct the DepGraph long before we have enough information to know where to load the previous DepGraph from and, because the CrateStore is used very early too, we are registering dep-nodes and edges before the graph is fully operational. This is ugly but harmless in the current implementation. In the new tracking system it would become a problem.

So the plan is to construct the DepGraph only when we can already load the previous DepGraph. This needs some refactoring:

• Remove the dep_graph field from Session, - rustc: Remove Session::dep_graph #44502
• Remove the dep_graph field from CStore and don't do any dependency tracking in the regular CStore implementation. - rustc: Remove DepGraph handling from rustc_metadata #44418
• Make sure that we do perform dependency tracking for anything within the query system, via a tasteful blend of: - Migrate a slew of metadata methods to queries #44142 and rustc: Make CrateStore private to TyCtxt #44420
  • creating a CrateStore wrapper that does dependency tracking and make only that available via tcx.cstore(), and - Migrate a slew of metadata methods to queries #44142
  • making all CrateStore access go through queries. - Migrate a slew of metadata methods to queries #44142
• Adapt librustc_incremental::persist::fs::prepare_session_directory() to not need the tcx and invoke it right after parsing, when we know the definitive crate name. - rustc: Remove Session::dep_graph #44502
• Move DepGraph construction to right after librustc_incremental::persist::fs::prepare_session_directory() has been called. - rustc: Remove Session::dep_graph #44502

Any work in this direction will want to be based on top of #44341 the master branch.

[michaelwoerister]

cc @nikomatsakis @alexcrichton

[alexcrichton]

@michaelwoerister should the DepGraph argument all throughout Decoder also be removed? For example in this function.

After #44142 lands then 100% of the existing cstore methods are now queries (yay!) or otherwise don't need to be tracked.

[michaelwoerister]

@michaelwoerister should the DepGraph argument all throughout Decoder also be removed? For example in this function.

Ideally the "raw" CrateStore implementation would not know anything about the DepGraph anymore, just the queries (if that's possible).

[alexcrichton]

Ok! Sounds like a "yes"

[alexcrichton]

I have a commit for completely removing DepGraph from the CrateStore at alexcrichton@74f6a22, just waiting on #44142 to land