make `output_filenames` a real query
part of #105462
This may be a perf regression and is not obviously the right way forward. We may store this information in the resolver after freezing it for example.
Move format_args!() into AST (and expand it during AST lowering)
Implements https://github.com/rust-lang/compiler-team/issues/541
This moves FormatArgs from rustc_builtin_macros to rustc_ast_lowering. For now, the end result is the same. But this allows for future changes to do smarter things with format_args!(). It also allows Clippy to directly access the ast::FormatArgs, making things a lot easier.
This change turns the format args types into lang items. The builtin macro used to refer to them by their path. After this change, the path is no longer relevant, making it easier to make changes in `core`.
This updates clippy to use the new language items, but this doesn't yet make clippy use the ast::FormatArgs structure that's now available. That should be done after this is merged.
Convert all the crates that have had their diagnostic migration
completed (except save_analysis because that will be deleted soon and
apfloat because of the licensing problem).
Encode spans relative to the enclosing item -- enable on nightly
Follow-up to #84373 with the flag `-Zincremental-relative-spans` set by default.
This PR seeks to remove one of the main shortcomings of incremental: the handling of spans.
Changing the contents of a function may require redoing part of the compilation process for another function in another file because of span information is changed.
Within one file: all the spans in HIR change, so typechecking had to be re-done.
Between files: spans of associated types/consts/functions change, so type-based resolution needs to be re-done (hygiene information is stored in the span).
The flag `-Zincremental-relative-spans` encodes local spans relative to the span of an item, stored inside the `source_span` query.
Trap: stashed diagnostics are referenced by the "raw" span, so stealing them requires to remove the span's parent.
In order to avoid too much traffic in the span interner, span encoding uses the `ctxt_or_tag` field to encode:
- the parent when the `SyntaxContext` is 0;
- the `SyntaxContext` when the parent is `None`.
Even with this, the PR creates a lot of traffic to the Span interner, when a Span has both a LocalDefId parent and a non-root SyntaxContext. They appear in lowering, when we add a parent to all spans, including those which come from macros, and during inlining when we mark inlined spans.
The last commit changes how queries of `LocalDefId` manage their cache. I can put this in a separate PR if required.
Possible future directions:
- validate that all spans are marked in HIR validation;
- mark macro-expanded spans relative to the def-site and not the use-site.
Remove `token::Lit` from `ast::MetaItemLit`.
Currently `ast::MetaItemLit` represents the literal kind twice. This PR removes that redundancy. Best reviewed one commit at a time.
r? `@petrochenkov`
Add help for `#![feature(impl_trait_in_fn_trait_return)]`
This adds a new variant `ImplTraitContext::FeatureGated`, so we can
generalize the help for `return_position_impl_trait_in_trait` to also
work for `impl_trait_in_fn_trait_return`.
cc #99697
Simplify attribute handling in rustc_ast_lowering
Given that attributes is stored in a separate BTreeMap, it's not necessary to pass it in when constructing `hir::Expr`. We can just construct `hir::Expr` and then call `self.lower_attrs` later if it needs attributes.
As most desugaring code don't use attributes, this allows some code cleanup.
This adds a new variant `ImplTraitContext::FeatureGated`, so we can
generalize the help for `return_position_impl_trait_in_trait` to also
work for `impl_trait_in_fn_trait_return`.
Replaces using `ResumeTy` / `get_context` in favor of using `&'static mut Context<'_>`.
Usage of the `'static` lifetime here is technically "cheating", and replaces
the raw pointer in `ResumeTy` and the `get_context` fn that pulls the
correct lifetimes out of thin air.
Given that attributes is stored in a separate BTreeMap, it's not necessary
to pass it in when constructing `hir::Expr`. We can just construct
`hir::Expr` and then call `self.lower_attrs` later if it needs attributes.
As most desugaring code don't use attributes, this allows some code cleanup.
`token::Lit` contains a `kind` field that indicates what kind of literal
it is. `ast::MetaItemLit` currently wraps a `token::Lit` but also has
its own `kind` field. This means that `ast::MetaItemLit` encodes the
literal kind in two different ways.
This commit changes `ast::MetaItemLit` so it no longer wraps
`token::Lit`. It now contains the `symbol` and `suffix` fields from
`token::Lit`, but not the `kind` field, eliminating the redundancy.
This is required to distinguish between cooked and raw byte string
literals in an `ast::LitKind`, without referring to an adjacent
`token::Lit`. It's a prerequisite for the next commit.
Lower them into a single item with multiple resolutions instead.
This also allows to remove additional `NodId`s and `DefId`s related to those additional items.
This makes sure that ICEing because of def ids created outside of ast lowering will be able to produce a query backtrace and not cause a double panic because of trying to call the `def_span` query
There is code for converting `Attribute` (syntactic) to `MetaItem`
(semantic). There is also code for the reverse direction. The reverse
direction isn't really necessary; it's currently only used when
generating attributes, e.g. in `derive` code.
This commit adds some new functions for creating `Attributes`s directly,
without involving `MetaItem`s: `mk_attr_word`, `mk_attr_name_value_str`,
`mk_attr_nested_word`, and
`ExtCtxt::attr_{word,name_value_str,nested_word}`.
These new methods replace the old functions for creating `Attribute`s:
`mk_attr_inner`, `mk_attr_outer`, and `ExtCtxt::attribute`. Those
functions took `MetaItem`s as input, and relied on many other functions
that created `MetaItems`, which are also removed: `mk_name_value_item`,
`mk_list_item`, `mk_word_item`, `mk_nested_word_item`,
`{MetaItem,MetaItemKind,NestedMetaItem}::token_trees`,
`MetaItemKind::attr_args`, `MetaItemLit::{from_lit_kind,to_token}`,
`ExtCtxt::meta_word`.
Overall this cuts more than 100 lines of code and makes thing simpler.
Separate lifetime ident from lifetime resolution in HIR
Drive-by: change how suggested generic args are computed.
Fixes https://github.com/rust-lang/rust/issues/103815
I recommend reviewing commit-by-commit.
Avoid `GenFuture` shim when compiling async constructs
Previously, async constructs would be lowered to "normal" generators, with an additional `from_generator` / `GenFuture` shim in between to convert from `Generator` to `Future`.
The compiler will now special-case these generators internally so that async constructs will *directly* implement `Future` without the need to go through the `from_generator` / `GenFuture` shim.
The primary motivation for this change was hiding this implementation detail in stack traces and debuginfo, but it can in theory also help the optimizer as there is less abstractions to see through.
---
Given this demo code:
```rust
pub async fn a(arg: u32) -> Backtrace {
let bt = b().await;
let _arg = arg;
bt
}
pub async fn b() -> Backtrace {
Backtrace::force_capture()
}
```
I would get the following with the latest stable compiler (on Windows):
```
4: async_codegen:🅱️:async_fn$0
at .\src\lib.rs:10
5: core::future::from_generator::impl$1::poll<enum2$<async_codegen:🅱️:async_fn_env$0> >
at /rustc/897e37553bba8b42751c67658967889d11ecd120\library\core\src\future\mod.rs:91
6: async_codegen:🅰️:async_fn$0
at .\src\lib.rs:4
7: core::future::from_generator::impl$1::poll<enum2$<async_codegen:🅰️:async_fn_env$0> >
at /rustc/897e37553bba8b42751c67658967889d11ecd120\library\core\src\future\mod.rs:91
```
whereas now I get a much cleaner stack trace:
```
3: async_codegen:🅱️:async_fn$0
at .\src\lib.rs:10
4: async_codegen:🅰️:async_fn$0
at .\src\lib.rs:4
```
Previously, async constructs would be lowered to "normal" generators,
with an additional `from_generator` / `GenFuture` shim in between to
convert from `Generator` to `Future`.
The compiler will now special-case these generators internally so that
async constructs will *directly* implement `Future` without the need
to go through the `from_generator` / `GenFuture` shim.
The primary motivation for this change was hiding this implementation
detail in stack traces and debuginfo, but it can in theory also help
the optimizer as there is less abstractions to see through.
Lower return type outside async block creation
This allows feeding a different output type to async blocks with a different `ImplTraitContext`. Spotted this while working on #104321
`MacArgs` is an enum with three variants: `Empty`, `Delimited`, and `Eq`. It's
used in two ways:
- For representing attribute macro arguments (e.g. in `AttrItem`), where all
three variants are used.
- For representing function-like macros (e.g. in `MacCall` and `MacroDef`),
where only the `Delimited` variant is used.
In other words, `MacArgs` is used in two quite different places due to them
having partial overlap. I find this makes the code hard to read. It also leads
to various unreachable code paths, and allows invalid values (such as
accidentally using `MacArgs::Empty` in a `MacCall`).
This commit splits `MacArgs` in two:
- `DelimArgs` is a new struct just for the "delimited arguments" case. It is
now used in `MacCall` and `MacroDef`.
- `AttrArgs` is a renaming of the old `MacArgs` enum for the attribute macro
case. Its `Delimited` variant now contains a `DelimArgs`.
Various other related things are renamed as well.
These changes make the code clearer, avoids several unreachable paths, and
disallows the invalid values.
Support using `Self` or projections inside an RPIT/async fn
I reuse the same idea as https://github.com/rust-lang/rust/pull/103449 to use variances to encode whether a lifetime parameter is captured by impl-trait.
The current implementation of async and RPIT replace all lifetimes from the parent generics by `'static`. This PR changes the scheme
```rust
impl<'a> Foo<'a> {
fn foo<'b, T>() -> impl Into<Self> + 'b { ... }
}
opaque Foo::<'_a>::foo::<'_b, T>::opaque<'b>: Into<Foo<'_a>> + 'b;
impl<'a> Foo<'a> {
// OLD
fn foo<'b, T>() -> Foo::<'static>::foo::<'static, T>::opaque::<'b> { ... }
^^^^^^^ the `Self` becomes `Foo<'static>`
// NEW
fn foo<'b, T>() -> Foo::<'a>::foo::<'b, T>::opaque::<'b> { ... }
^^ the `Self` stays `Foo<'a>`
}
```
There is the same issue with projections. In the example, substitute `Self` by `<T as Trait<'b>>::Assoc` in the sugared version, and `Foo<'_a>` by `<T as Trait<'_b>>::Assoc` in the desugared one.
This allows to support `Self` in impl-trait, since we do not replace lifetimes by `'static` any more. The same trick allows to use projections like `T::Assoc` where `Self` is allowed. The feature is gated behind a `impl_trait_projections` feature gate.
The implementation relies on 2 tweaking rules for opaques in 2 places:
- we only relate substs that correspond to captured lifetimes during TypeRelation;
- we only list captured lifetimes in choice region computation.
For simplicity, I encoded the "capturedness" of lifetimes as a variance, `Bivariant` vs `Invariant` for unused vs captured lifetimes. The `variances_of` query used to ICE for opaques.
Impl-trait that do not reference `Self` or projections will have their variances as:
- `o` (invariant) for each parent type or const;
- `*` (bivariant) for each parent lifetime --> will not participate in borrowck;
- `o` (invariant) for each own lifetime.
Impl-trait that does reference `Self` and/or projections will have some parent lifetimes marked as `o` (as the example above), and participate in type relation and borrowck. In the example above, `variances_of(opaque) = ['_a: o, '_b: *, T: o, 'b: o]`.
r? types
cc `@compiler-errors` , as you asked about the issue with `Self` and projections.
