More correctly, *also* per-token. Because as it turns out, while the top-level edition affects parsing (I think), the per-token edition affects escaping of identifiers/keywords.
I expected this to be faster (due to less allocations and better cache locality), but benchmarked it is not (neither it is slower). Memory usage, however, drops by ~50mb (of `analysis-stats .`). I guess tt construction is just not hot.
This also simplifies using even less memory for token trees by compressing equal span, which I plan to do right after.
Some workflows are more easily expressed with a flat tt, while some are better expressed with a tree. With the right helpers, though (which was mostly a matter of trial and error), even the worst workflows become very easy indeed.
Because it was a mess.
Previously, pretty much you had to handle all path diagnostics manually: remember to check for them and handle them. Now, we wrap the resolver in `TyLoweringContext` and ensure proper error reporting.
This means that you don't have to worry about them: most of the things are handled automatically, and things that cannot will create a compile-time error (forcing you top `drop(ty_lowering_context);`) if forgotten, instead of silently dropping the diagnostics.
The real place for error reporting is in the hir-def resolver, because there are other things resolving, both in hir-ty and in hir-def, and they all need to ensure proper diagnostics. But this is a good start, and future compatible.
This commit also ensures proper path diagnostics for value/pattern paths, which is why it's marked "feat".
The reason I did this is because I plan to add another field to this struct (indicating whether the item was cfg'ed out), but it seems worthy even separately and removes a bunch of one-letter variable names and tuple-indexing. It is also easy to separate from future changes, so it will be easier to review this way.
This mainly aids in error recovery but also makes it a bit easier to handle lifetime resolution.
While doing so it also came apparent that we were not actually lowering lifetime outlives relationships within lifetime parameter declaration bounds, so this fixes that.
Implement diagnostics in all places left: generics (predicates, defaults, const params' types), fields, and type aliases.
Unfortunately this results in a 20mb addition in `analysis-stats .` due to many type methods returning an addition diagnostics result now (even if it's `None` in most cases). I'm not sure if this can be improved.
An alternative strategy that can prevent the memory usage growth is to never produce diagnostics in hir-ty methods. Instead, lower all types in the hir crate when computing diagnostics from scratch (with diagnostics this time). But this has two serious disadvantages:
1. This can cause code duplication (although it can probably be not that bad, it will still mean a lot more code).
2. I believe we eventually want to compute diagnostics for the *entire* workspace (either on-type or on-save or something alike), so users can know when they have diagnostics even in inactive files. Choosing this approach will mean we lose all precomputed salsa queries. For one file this is fine, for the whole workspace this will be very slow.
