Introduce macro sub-namespaces and `macro_use` prelude
This PR implements two mechanisms needed for correct macro name resolution: macro sub-namespace and `macro_use` prelude.
- [macro sub-namespaces][subns-ref]
Macros have two sub-namespaces: one for function-like macro and the other for those in attributes (including custom derive macros). When we're resolving a macro name for function-like macro, we should ignore non-function-like macros, and vice versa.
This helps resolve single-segment macro names because we can (and should, as rustc does) fallback to names in preludes when the name in the current module scope is in different sub-namespace.
- [`macro_use` prelude][prelude-ref]
`#[macro_use]`'d extern crate declarations (including the standard library) bring their macros into scope, but they should not be prioritized over local macros (those defined in place and those explicitly imported).
We have been bringing them into legacy (textual) macro scope, which has the highest precedence in name resolution. This PR introduces the `macro_use` prelude in crate-level `DefMap`s, whose precedence is lower than local macros but higher than the standard library prelude.
The first 3 commits are drive-by fixes/refactors.
Fixes#8828 (prelude)
Fixes#12505 (prelude)
Fixes#12734 (prelude)
Fixes#13683 (prelude)
Fixes#13821 (prelude)
Fixes#13974 (prelude)
Fixes#14254 (namespace)
[subns-ref]: https://doc.rust-lang.org/reference/names/namespaces.html#sub-namespaces
[prelude-ref]: https://doc.rust-lang.org/reference/names/preludes.html#macro_use-prelude
Beginning of MIR
This pull request introduces the initial implementation of MIR lowering and interpreting in Rust Analyzer.
The implementation of MIR has potential to bring several benefits:
- Executing a unit test without compiling it: This is my main goal. It can be useful for quickly testing code changes and print-debugging unit tests without the need for a full compilation (ideally in almost zero time, similar to languages like python and js). There is a probability that it goes nowhere, it might become slower than rustc, or it might need some unreasonable amount of memory, or we may fail to support a common pattern/function that make it unusable for most of the codes.
- Constant evaluation: MIR allows for easier and more correct constant evaluation, on par with rustc. If r-a wants to fully support the type system, it needs full const eval, which means arbitrary code execution, which needs MIR or something similar.
- Supporting more diagnostics: MIR can be used to detect errors, most famously borrow checker and lifetime errors, but also mutability errors and uninitialized variables, which can be difficult/impossible to detect in HIR.
- Lowering closures: With MIR we can find out closure capture modes, which is useful in detecting if a closure implements the `FnMut` or `Fn` traits, and calculating its size and data layout.
But the current PR implements no diagnostics and doesn't support closures. About const eval, I removed the old const eval code and it now uses the mir interpreter. Everything that is supported in stable rustc is either implemented or is super easy to implement. About interpreting unit tests, I added an experimental config, disabled by default, that shows a `pass` or `fail` on hover of unit tests (ideally it should be a button similar to `Run test` button, but I didn't figured out how to add them). Currently, no real world test works, due to missing features including closures, heap allocation, `dyn Trait` and ... so at this point it is only useful for me selecting what to implement next.
The implementation of MIR is based on the design of rustc, the data structures are almost copy paste (so it should be easy to migrate it to a possible future stable-mir), but the lowering and interpreting code is from me.
I am not certain if this will improve performance,
but it seems having a .clone() without any need should be removed.
This was done with clippy, and manually reviewed:
```
cargo clippy --fix -- -A clippy::all -D clippy::redundant_clone
```
This commit "inverts" the order of generic parameters/arguments of an
item and its parent. This is to fulfill chalk's expectation on the
order of `Substitution` for generic associated types and it's one step
forward for their support (hopefully).
Although chalk doesn't put any constraint on the order of `Substitution`
for other items, it feels natural to get everything aligned rather than
special casing GATs.
One complication is that `TyBuilder` now demands its users to pass in
parent's `Substitution` upon construction unless it's obvious that the
the item has no parent (e.g. an ADT never has parent). All users
*should* already know the parent of the item in question, and without
this, it cannot be easily reasoned about whether we're pushing the
argument for the item or for its parent.
Quick comparison of how this commit changes `Substitution`:
```rust
trait Trait<TP, const CP: usize> {
type Type<TC, const CC: usize> = ();
fn f<TC, const CC: usize>() {}
}
```
- before this commit: `[Self, TP, CP, TC, CC]` for each trait item
- after this commit: `[TC, CC, Self, TP, CP]` for each trait item
fix: sort and deduplicate auto traits in trait object types
Fixes#12739
Chalk solver doesn't sort and deduplicate auto traits in trait object types, so we need to handle them ourselves in the lowering phase, just like [`rustc`](880416180b/compiler/rustc_typeck/src/astconv/mod.rs (L1487-L1488)) and [`chalk-integration`](https://github.com/rust-lang/chalk/blob/master/chalk-integration/src/lowering.rs#L575) do.
Quoting from [the Chalk book](https://rust-lang.github.io/chalk/book/types/rust_types.html#dyn-types):
> Note that -- for this purpose -- ordering of bounds is significant. That means that if you create a `dyn Foo + Send` and a `dyn Send + Foo`, chalk would consider them distinct types. The assumption is that bounds are ordered in some canonical fashion somewhere else.
Also, trait object types with more than one non-auto traits were previously allowed, but are now disallowed with this patch.
fix: Fix panics on GATs involving const generics
This workaround avoids constant crashing of rust analyzer when using GATs with const generics,
even when the const generics are only on the `impl` block.
The workaround treats GATs as non-existing if either itself or the parent has const generics and
removes relevant panicking code-paths.
Fixes#11989, fixes#12193
This workaround avoids constant crashing of rust analyzer when using GATs with const generics,
even when the const generics are only on the `impl` block.
The workaround treats GATs as non-existing if either itself or the parent has const generics and
removes relevant panicking code-paths.
