Make queries thread safe
This makes queries thread safe by removing the query stack and making queries point to their parents. Queries write to the query map when starting and cycles are detected by checking if there's already an entry in the query map. This makes cycle detection O(1) instead of O(n), where `n` is the size of the query stack.
This is mostly corresponds to the method I described [here](https://internals.rust-lang.org/t/parallelizing-rustc-using-rayon/6606).
cc @rust-lang/compiler
r? @michaelwoerister
This crate moves the compiler's error reporting to using the `termcolor` crate
from crates.io. Previously rustc used a super-old version of the `term` crate
in-tree which is basically unmaintained at this point, but Cargo has been using
`termcolor` for some time now and tools like `rg` are using `termcolor` as well,
so it seems like a good strategy to take!
Note that the `term` crate remains in-tree for libtest. Changing libtest will be
a bit tricky due to how the build works, but we can always tackle that later.
cc #45728
Add approximate suggestions for rustfix
This adds `span_approximate_suggestion()` that lets you emit a
suggestion marked as "non-machine applicable" in the JSON output. UI
users see no difference. This is for when rustc and clippy wish to
emit suggestions which will make sense to the reader (e.g. they may
have placeholders like `<type>`) but are not source-applicable, so that
rustfix/etc can ignore these.
fixes#39254
This adds `span_approximate_suggestion()` that lets you emit a
suggestion marked as "approximate" in the JSON output. UI
users see no difference. This is for when rustc and clippy wish to
emit suggestions which will make sense to the reader (e.g. they may
have placeholders like `<type>`) but are not source-applicable, so that
rustfix/etc can ignore these.
fixes#39254
Use correct line offsets for doctests
Not yet tested.
This doesn't handle char positions. It could if I collected a map of char offsets and lines, but this is a bit more work and requires hooking into the parser much more (unsure if it's possible).
r? @QuietMisdreavus
(fixes#45868)
Don't emit the same compiler diagnostic twice.
This PR makes the compiler filter out diagnostic messages that have already been emitted during the same compilation session.
Make fields of `Span` private
I actually tried to intern spans and benchmark the result<sup>*</sup>, and this was a prerequisite.
This kind of encapsulation will be a prerequisite for any other attempt to compress span's representation, so I decided to submit this change alone.
The issue https://github.com/rust-lang/rust/issues/43088 seems relevant, but it looks like `SpanId` won't be able to reuse this interface, unless the tables are global (like interner that I tried) and are not a part of HIR.
r? @michaelwoerister anyway
<sup>*</sup> Interning means 2-3 times more space is required for a single span, but duplicates are free. In practice it turned out that duplicates are not *that* common, so more memory was wasted by interning rather than saved.
rustc: Start moving toward "try_get is a bug" for incremental
This PR is an effort to burn down some of the work items on #42633. The basic change here was to leave the `try_get` function exposed but have it return a `DiagnosticBuilder` instead of a `CycleError`. This means that it should be a compiler bug to *not* handle the error as dropping a diagnostic should result in a complier panic.
After that change it was then necessary to update the compiler's callsites of `try_get` to handle the error coming out. These were handled as:
* The `sized_constraint` and `needs_drop_raw` checks take the diagnostic and defer it as a compiler bug. This was a new piece of functionality added to the error handling infrastructure, and the idea is that for both these checks a "real" compiler error should be emitted elsewhere, so it's only a bug if we don't actually emit the complier error elsewhere.
* MIR inlining was updated to just ignore the diagnostic. This is being tracked by https://github.com/rust-lang/rust/issues/43542 which sounded like it either already had some work underway or was planning to change regardless.
* The final case, `item_path`, is still sort of up for debate. At the time of this writing this PR simply removes the invocations of `try_get` there, assuming that the query will always succeed. This turns out to be true for the test suite anyway! It sounds like, though, that this logic was intended to assist in "weird" situations like `RUST_LOG` where debug implementations can trigger at any time. This PR would therefore, however, break those implementations.
I'm unfortunately sort of out of ideas on how to handle `item_path`, but other thoughts would be welcome!
Closes#42633
This adds a function to `DiagnosticBuilder` to delay the entire diagnostic as a
bug to be emitted at a later time. This'll end up getting used in the compiler
in the subsequent commits...
This commit alters the `rustc::ty::maps` implementation to ensure that all
output diagnostics from the compiler are tracked for the duration of each query.
These are then intended to be replayed back the first time a cached value is
loaded, and otherwise the cache should operate the same as it does today.
Closes#42513
