The idea here is to eventually get rid of `dyn Diagnostic` and
`DiagnosticSink` infrastructure altogether, and just have a `enum
hir::Diagnostic` instead.
The problem with `dyn Diagnostic` is that it is defined in the lowest
level of the stack (hir_expand), but is used by the highest level (ide).
As a first step, we free hir_expand and hir_def from `dyn Diagnostic`
and kick the can up to `hir_ty`, as an intermediate state. The plan is
then to move DiagnosticSink similarly to the hir crate, and, as final
third step, remove its usage from the ide.
One currently unsolved problem is testing. You can notice that the test
which checks precise diagnostic ranges, unresolved_import_in_use_tree,
was moved to the ide layer. Logically, only IDE should have the infra to
render a specific range.
At the same time, the range is determined with the data produced in
hir_def and hir crates, so this layering is rather unfortunate. Working
on hir_def shouldn't require compiling `ide` for testing.
8987: Fix lowering of FnOnce() without return type r=flodiebold a=flodiebold
This should result in an implicit `-> ()`, not leaving out the binding.
Co-authored-by: Florian Diebold <flodiebold@gmail.com>
Chalk's unification can sometimes create lifetime variables, which we
currently don't really deal with, but at least we don't want to leak
them outside of inference.
Should fix#8919.
We can't do the easy hack that we did before anymore, where we kept
track of whether any inference variables changed since the last time we
rechecked obligations. Instead, we store the obligations in
canonicalized form; that way we can easily check the inference variables
to see whether they have changed since the goal was canonicalized.
8813: Get some more array lengths! r=lf- a=lf-
This is built on #8799 and thus contains its changes. I'll rebase it onto master when that one gets merged. It adds support for r-a understanding the length of:
* `let a: [u8; 2] = ...`
* `let a = b"aaa"`
* `let a = [0u8; 4]`
I have added support for getting the values of byte strings, which was not previously there. I am least confident in the correctness of this part and it probably needs some more tests, as we currently have only one test that exercised that part (!).
Fixes#2922.
Co-authored-by: Jade <software@lfcode.ca>
Fix#2922: add unknown length as a condition for a type having unknown.
Incorporate reviews:
* Extract some of the const evaluation workings into functions
* Add fixmes on the hacks
* Add tests for impls on specific array lengths (these work!!! 😁)
* Add tests for const generics (indeed we don't support it yet)
8799: Add basic support for array lengths in types r=flodiebold a=lf-
This recognizes `let a = [1u8, 2, 3]` as having type `[u8; 3]` instead
of the previous `[u8; _]`. Byte strings and `[0u8; 2]` kinds of range
array declarations are unsupported as before.
I don't know why a bunch of our rustc tests had single quotes inside
strings un-escaped by `UPDATE_EXPECT=1 cargo t`, but I don't think it's
bad? Maybe something in a nightly?
Co-authored-by: Jade <software@lfcode.ca>
This recognizes `let a = [1u8, 2, 3]` as having type `[u8; 3]` instead
of the previous `[u8; _]`. Byte strings and `[0u8; 2]` kinds of range
array declarations are unsupported as before.
I don't know why a bunch of our rustc tests had single quotes inside
strings un-escaped by `UPDATE_EXPECT=1 cargo t`, but I don't think it's
bad? Maybe something in a nightly?
8432: decl_check: consider outer scopes' allows r=jonas-schievink a=lf-
Fix#8417. Also makes it less noisy about no_mangle annotated stuff the
user can do nothing about.
Note: this still is broken with bitfield! macros. A repro in an ignore
test is included here. I believe this bug is elsewhere, and I don't
think I can work around it here.
I would like help filing the remaining bug, as it does actually affect
users, but I don't know how to describe the behaviour (or even if it
is unintended).
Co-authored-by: Jade <software@lfcode.ca>
This condition should always be true for *valid* code, but of course
there might be invalid code or things that we can't currently resolve.
Fixes#8464.
8406: Improve indexing of impls r=flodiebold a=flodiebold
Store impls for e.g. &Foo with the ones for Foo instead of the big "other" bucket. This can improve performance and simplifies the HIR impl search a bit.
Co-authored-by: Florian Diebold <flodiebold@gmail.com>
Store impls for e.g. &Foo with the ones for Foo instead of the big
"other" bucket. This can improve performance and simplifies the HIR impl
search a bit.
If we get lifetime variables back in autoderef, just immediately replace
them by static lifetimes for now. Method resolution doesn't really deal
correctly with new variables being introduced (this needs to be fixed
more properly).
This fixes `rust-analyzer analysis-stats --with-deps` crashing in the RA
repo.
Fix#8417. Also makes it less noisy about no_mangle annotated stuff the
user can do nothing about.
Note: this still is broken with bitfield! macros. A repro in an ignore
test is included here. I believe this bug is elsewhere, and I don't
think I can work around it here.
I'd prefer getting rid of it, but it's used in the impl search and not
super easy to replace there (I think ideally the impl search would do
proper unification, but that's a bit more complicated).
Almost all uses actually only care about ADT substs, so it's better to
be explicit. The methods were a bad abstraction anyway since they
already didn't include the inner types of e.g. `TyKind::Ref` anymore.
8371: Don't use HirDisplayWrapper when displaying SourceCode r=matklad a=Veykril
The issue was basically that when displaying for `DisplayTarget::SourceCode` some `hir_fmt` functions would create `HirDisplayWrapper`s which would then `fmt` these triggering the Display panic since `fmt::Display` can't fail the same way as `HirDisplay`. Simple fix is to just use `hir_fmt` directly. Should probably write that down somewhere in source, looking for a good spot to put that right now.
Fixes#8077, Fixes#8370
Co-authored-by: Lukas Wirth <lukastw97@gmail.com>
8364: Memory usage improvements r=jonas-schievink a=alexmaco
These are mostly focused on splitting up enum variants with large size differences between variants by `Box`-ing things up.
In my testing this reduces the memory usage somewhere in the low percentages, even though the measurements are quite noisy.
Co-authored-by: Alexandru Macovei <alexnmaco@gmail.com>
Rationale: only a minority of variants used almost half the size.
By keeping large members (especially in Option) behind a box
the memory cost is only payed when the large variants are needed.
This reduces the size Vec<Expr> needs to allocate.
- don't shift in/out for Chalk mapping (we want to have the same
binders now)
- do shift in when creating the signature for a closure (though it
shouldn't matter much)
- do shift in when lowering a `fn()` type
- correctly deal with the implied binder in TypeWalk
8353: Replace hir_ty::Lifetime with chalk equivalent r=flodiebold a=Veykril
Our `Lifetime` isn't really used yet so this is a rather simple change
Co-authored-by: Lukas Wirth <lukastw97@gmail.com>
8284: Reduce memory usage by using global `Arc`-based interning r=jonas-schievink a=jonas-schievink
This saves around 50 mb when running `analysis-stats` on r-a itself. Not a lot, but this infra can be easily reused to intern more stuff.
Co-authored-by: Jonas Schievink <jonasschievink@gmail.com>
8285: Don't recheck obligations if we have learned nothing new r=matklad a=flodiebold
This is just the most trivial check: If no inference variables have been updated, and there are no new obligations, we can just skip trying to solve them again. We could be smarter about it, but this already helps quite a bit, and I don't want to touch this too much before we replace the inference table by Chalk's.
Fixes#8263 (well, improves it quite a bit).
Co-authored-by: Florian Diebold <flodiebold@gmail.com>
8283: Resolve associated types r=flodiebold a=Veykril
Prior we were only resolving paths until the first type was found, then discarding the result if the path wasn't fully consumed. That of course causes associated types to not resolve. Fixes#5003
Co-authored-by: Lukas Wirth <lukastw97@gmail.com>
This is just the most trivial check: If no inference variables have been
updated, and there are no new obligations, we can just skip trying to
solve them again. We could be smarter about it, but this already helps
quite a bit, and I don't want to touch this too much before we replace
the inference table by Chalk's.
Fixes#8263 (well, improves it quite a bit).
8266: Fix generic arguments being incorrectly offset in qualified trait casts r=flodiebold a=Veykril
We reverse the segments and generic args of the lowered path after building it, this wasn't accounted for when inserting the self parameter in `Type as Trait` segments.
Fixes#5886
Co-authored-by: Lukas Wirth <lukastw97@gmail.com>
We have a bug where type-checking `per_query_memory_usage` takes a
couple of seconds. It also reveals another bug: our type inference is
not cancellable.
8201: Fix recursive macro statements expansion r=edwin0cheng a=edwin0cheng
This PR attempts to properly handle macro statement expansion by implementing the following:
1. Merge macro expanded statements to parent scope statements.
2. Add a new hir `Expr::MacroStmts` for handle tail expression infer.
PS : The scope of macro expanded statements are so strange that it took more time than I thought to understand and implement it :(
Fixes #8171
Co-authored-by: Edwin Cheng <edwin0cheng@gmail.com>
7907: Autoderef with visibility r=cynecx a=cynecx
Fixes https://github.com/rust-analyzer/rust-analyzer/issues/7841.
I am not sure about the general approach here. Right now this simply tries to check whether the autoderef candidate is reachable from the current module. ~~However this doesn't exactly work with traits (see the `tests::macros::infer_derive_clone_in_core` test, which fails right now).~~ see comment below
Refs:
- `rustc_typeck` checking fields: 66ec64ccf3/compiler/rustc_typeck/src/check/expr.rs (L1610)
r? @flodiebold
Co-authored-by: cynecx <me@cynecx.net>
8136: Introduce QuantifiedWhereClause and DynTy analogous to Chalk r=flodiebold a=flodiebold
This introduces a bunch of new binders in lots of places, which we have to be careful about, but we had to add them at some point. There's a lot of skipping of the binders; once we're done with the Chalk move, we should review the remaining ones.
8146: Document patch policy r=matklad a=matklad
bors r+
🤖
Co-authored-by: Florian Diebold <flodiebold@gmail.com>
Co-authored-by: Aleksey Kladov <aleksey.kladov@gmail.com>
