This patch uses the same data structures for structs and enum variants in AST and HIR. These changes in data structures lead to noticeable simplification in most of code dealing with them.
I didn't touch the top level, i.e. `ItemStruct` is still `ItemStruct` and not `ItemEnum` with one variant, like in the type checker.
As part of this patch, structures and variants get the `kind` field making distinction between "normal" structs, tuple structs and unit structs explicit instead of relying on the number of fields and presence of constructor `NodeId`. In particular, we can now distinguish empty tuple structs from unit structs, which was impossible before! Comprehensive tests for empty structs are added and some improvements to empty struct feature gates are made. Some tests don't pass due to issue https://github.com/rust-lang/rust/issues/28692 , they are still there for completeness, but are commented out.
This patch fixes issue mentioned in https://github.com/rust-lang/rust/issues/16819#issuecomment-139509861, now emptiness of tuple structs is checked after expansion.
It also touches https://github.com/rust-lang/rust/issues/28750 by providing span for visit_struct_def
cc https://github.com/rust-lang/rust/pull/28336
r? @nrc
The link is broken here: <https://doc.rust-lang.org/std/io/#types>.
Looks like crate documentation generator uses only first paragraph of the module documentation and
so doesn't resolve the link defined below.
This commit adds #[derive(Clone)] to std::fs::Metadata, making that struct
cloneable. Although the exact contents of that struct differ between OSes,
they all have it contain only value types, meaning that the data can be re-used without repercussions.
It also adds #[derive(Clone)] to every type used by that struct across all
OSes, including the various Unix `stat` structs and Windows's
`WIN32_FILE_ATTRIBUTE_DATA`.
This rather crucial requirement was not checked. In most cases, that
didn't cause any trouble because the argument types are required to
outlive the call and are subtypes of a subformula of the callee type.
However, binary ops are taken by ref only indirectly, without it being
marked in the argument types, which led to the argument types not being
constrained anywhere causing spurious errors (as these are basically
unconstrainable, I don't think this change can break code). Of course,
the old way was also incorrent with contravariance, but that is still
unsound for other reasons.
This also improves rustc::front to get RUST_LOG to *somewhat* work.
Fixes#28999
