Previously normalization was broken, which caused a lot of fake errors.
This fix most type mismatches of the new solver, and it also reverts many test regressions.
The downside is that now `chalk_ir::TyKind::AssociatedType`/`chalk_ir::TyKind::Alias` cannot be trusted anymore with their roles, namely: `AssociatedType` is always fully normalized and `Alias` only if it can possibly be normalized further. That seems okay as the new solver does not have this distinction at all (due to it being a lazy normalizer), so this will only hold for the migration time. This does mean we have to change some APIs, notably `hir::Type::walk()` and `TyFingerprint`, to treat `Alias` the same as `AssociatedType`.
Another small thing this commit does is to isolate processing of user-written types (currently involving replacing error types and normalizing, but in the future validation will also be needed) to separate functions.
More correctly, a `TyKind::AssociatedType` is not the same as `TyKind::Projection`.
We used to map next-solver `TyKind::Alias` to Chalk's `TyKind::AssociatedType`. This is very incorrect, as `AssociatedType` is assumed to be fully normalized, and caused a lot of type mismatches.
Unfortunately fixing this causes a lot of stack overflows, because the next solver doesn't have something akin to `AssociatedType` so we normalize again and again. The reason is that is the lazy-normalization nature of the next solver, which means we need to stop normalizing everything. This will be fixed in the next commit.
They have to do with diagnostics, we could probably not support them but we will also someday want good diagnostics.
The code is mostly copied from rustc.
The main changes are (there are some other small changes):
- Using a specific type for trait IDs in the new solver, allowing us to simplify a lot of code.
- Add `BoundConst` similar to `BoundTy` and `BoundRegion` (previously consts used `BoundVar` directly), due to a new trait requirement.
To the extent possible.
Previously they were confused. Sometimes generic params were treated as `Param` and sometimes as `Placeholder`. A completely redundant (in the new solver) mapping of salsa::Id to ints to intern some info where we could just store it uninterned (not in Chalk though, for some weird reason).
Plus fix a cute bug in closure substitution that was caught by the assertions of Chalk but the next solver did not have such assertions. Do we need more assertions?
Chalk represents dyn types as a list of predicate, the self type should be there. The next solver represents them quite differently. The `Self` was forgotten for the auto trait case.
Replace it with normal `SolverDefId::TypeAliasId`.
The split caused a very funny bug where code was getting `TypeAliasId` where it expected `ForeignId`, because `TypeAliasId` had a `From` impl from `hir_def::TypeAliasId` and `ForeignId` had not, plus a careless `into()`.
I could've fixed this specific bug but opted to remove the split instead; currently, it just provides more room for bugs, as we don't have typed IDs for the solver anyway, and even when we'll have (hopefully), that doesn't seem like a very useful distinction, for example in hir-def foreign types are just `TypeAliasId` with some flags.
Constructing a test for this isn't trivial; the trivial test (creating a foreign type, even proving a trait bound for it) fails to fail before the change, probably because we don't use the new solver everywhere yet so we don't trigger this specific code path.
The new solver does not eagerly normalize, but things after inference expect types to be normalized. rustc does the same.
Also, I'm afraid other things in r-a don't expect results of the solver to be unnormalized. We'll need to handle that.
Caching trait solving can do a lot to speed. Unfortunately it also consume a huge amount of memory. Therefore, as part of the migration to the new solver Jack Huey disabled caching of trait solving (he made the query transparent).
The PR proposes a middle ground: do cache trait solving, but only for the same revision. This allows us to be safe because during a revision the inputs cannot change.
The result is hopefully much better performance to features that tend to do a bulk of trait solving, and also repeat the same query (e.g. inference then IDE features).
There is another limitation: results are only cached in the same thread, to remove the need for synchronization which will be expensive. More measurements are required to check whether it's better to use a synchronized global cache, or maybe stay with a thread-local cache but batch multiple feature requests (highlighting, inlay hints etc.) of the same file to the same thread.
Alongside the actual cache we store the revision, because we need to verify it (we can't eagerly clear caches when incrementing the revision), and also the address of the db to prevent multiple dbs from interleaving (this is mostly relevant in tests, although injected highlighting also uses a new db, therefore maybe it's better to move it to a separate thread).
This "games" analysis-stats to both be way faster and use way more memory; the former is because analysis-stats doesn't increment revisions, therefore all queries share the cache and hit ratio is way too good, the latter is because analysis-stats doesn't increment revisions and therefore the cache isn't cleared. Both are not representative of a typical IDE scenario.
The parent generics were incorrectly not considered for TAIT.
I'm not convinced we should follow rustc here, also there are items (opaques) with more than 1 parent (opaque -> fn/type alias -> impl/trait) and I'm not sure we properly account for that in all places, but for now I left it as-is.
Also fix a bug where lifetimes' indices were incorrect when there is a self param (they started from 0 instead of 1).
