Do not unify borrowed locals in CopyProp.
Instead of trying yet another scheme to unify borrowed locals in CopyProp, let's just stop trying. We had already enough miscompilations because of this.
I'm convinced it's possible to have both unification of some borrowed locals and soundness, but I don't have a simple and convincing formulation yet.
Fixes https://github.com/rust-lang/rust/issues/143491
rustc_builtin_macros: Make sure registered attributes stay sorted
As with the list of builtin macros, use tidy to make sure the list of
builtin attributes stays sorted.
use `is_multiple_of` and `div_ceil`
In tricky logic, these functions are much more informative than the manual implementations. They also catch subtle bugs:
- the manual `is_multiple_of` often does not handle division by zero
- manual `div_ceil` often does not consider overflow
The transformation is free for `is_multiple_of` if the divisor is compile-time known to be non-zero. For `div_ceil` there is a small cost to considering overflow. Here is some assembly https://godbolt.org/z/5zP8KaE1d.
mbe: Defer checks for `compile_error!` until reporting an unused macro rule
The current MBE parser checks rules at initial parse time to see if their RHS has `compile_error!` in it, and returns a list of rule indexes and LHS spans that don't map to `compile_error!`, for use in unused macro rule checking.
Instead, have the unused macro rule reporting ask the macro for the rule to report, and let the macro check at that time. That avoids checking rules unless they're unused.
In the process, refactor the data structure used to store macro rules, to group the LHS and RHS (and LHS span) of each rule together, and refactor the unused rule tracking to only track rule indexes.
This builds atop a couple of minor MBE refactors. I would suggest reviewing commit-by-commit.
The overall result is a further simplification of the macro code.
The MBE parser checks rules at initial parse time to see if their RHS
has `compile_error!` in it, and returns a list of rule indexes and LHS
spans that don't map to `compile_error!`, for use in unused macro rule
checking.
Instead, have the unused macro rule reporting ask the macro for the rule
to report, and let the macro check at that time. That avoids checking
rules unless they're unused.
In the process, refactor the data structure used to store macro rules,
to group the LHS and RHS (and LHS span) of each rule together, and
refactor the unused rule tracking to only track rule indexes.
This ends up being a net simplification, and reduction in code size.
The parser repeatedly invokes the `parse` function, constructing a
one-entry vector, and assuming that the return value will be a one-entry
vector. Add a helper for that case. This will simplify adding additional
callers, and put all the logic in one place to allow potential future
simplification of the one-TT case.
Stop using `Key` trait unnecessarily
Few places where the `Key` trait was being used but not really for a useful reason. This fixes those usages.
Namely, `<Ty as Key>::default_span()` is `DUMMY_SP` anyways.
Canonicalize input ty/ct infer/placeholder in the root universe
We shouldn't care what universe the inputs are, since we only ever do the leak check on the universes instantiated after entering the canonical binder.
Remove `Symbol` from `Named` variant of `BoundRegionKind`/`LateParamRegionKind`
The `Symbol` is redundant, since we already store a `DefId` in the region variant. Instead, load the name via `item_name` when needed (which is almost always on the diagnostic path).
This introduces a `BoundRegionKind::NamedAnon` which is used for giving anonymous bound regions names, but which should only be used during pretty printing and error reporting.
Allow `enum` and `union` literals to also create SSA values
Today, `Some(x)` always goes through an `alloca`, even in trivial cases where the niching means the constructor doesn't even change the value.
For example, <https://rust.godbolt.org/z/6KG6PqoYz>
```rust
pub fn demo(r: &i32) -> Option<&i32> {
Some(r)
}
```
currently emits the IR
```llvm
define align 4 ptr `@demo(ptr` align 4 %r) unnamed_addr {
start:
%_0 = alloca [8 x i8], align 8
store ptr %r, ptr %_0, align 8
%0 = load ptr, ptr %_0, align 8
ret ptr %0
}
```
but with this PR it becomes just
```llvm
define align 4 ptr `@demo(ptr` align 4 %r) unnamed_addr {
start:
ret ptr %r
}
```
(Of course the optimizer can clean that up, but it'd be nice if it didn't have to -- especially in debug where it doesn't run. This is like rust-lang/rust#123886, but that only handled non-simd `struct`s -- this PR generalizes it to all non-simd ADTs.)
Doing this means handing variants other than `FIRST_VARIANT`, handling the active field for unions, refactoring the discriminant code so the Place and Operand parts can share the calculation, etc.
Other PRs that led up to this one:
- https://github.com/rust-lang/rust/pull/142005
- https://github.com/rust-lang/rust/pull/142103
- https://github.com/rust-lang/rust/pull/142324
- https://github.com/rust-lang/rust/pull/142383
---
try-job: aarch64-gnu
remove special-casing of boxes from match exhaustiveness/usefulness analysis
As a first step in replacing `box_patterns` with `deref_patterns`, this treats box patterns as deref patterns in the THIR and exhaustiveness analysis. This allows a bunch of special-casing to be removed. The emitted MIR is unchanged.
Incidentally, this fixes a bug caused by box patterns being treated like structs rather than pointers, where enabling `exhaustive_patterns` (rust-lang/rust#51085) could give rise to spurious `unreachable_patterns` lints on arms required for exhaustiveness. Following the lint's advice to remove the match arm would result in an error. I'm not sure what the current state of `exhaustive_patterns` is with regard to reference/box opsem, or whether there's any intention to have `unreachable_patterns` be more granular than the whole arm, but regardless this should hopefully make them easier to handle consistently.
Tracking issue for deref patterns: rust-lang/rust#87121
r? `@Nadrieril`
Make metadata a workproduct and reuse it
This PR aims to skip the generation of metadata by reusing the infrastructure that already exists for compiled codegen-units, namely "workproducts".
This can yield substantial gains (~10%) when we can demonstrate that metadata does not change between an incremental session and the next. This is the case if the crate is unchanged, or if all the changes are in upstream crates and have no effect on it. This latter case is most interesting, as it arises regularly for users with several crates in their workspace.
TODO:
- [x] Materialize the fact that metadata encoding relies on the relative order of definitions;
- [x] Refactor the handling of doc links.
