Rollup of 4 pull requests
Successful merges:
- #91169 (Change cg_ssa's get_param to borrow the builder mutably)
- #91176 (If the thread does not get the lock in the short term, yield the CPU)
- #91212 (Fix ICE due to out-of-bounds statement index when reporting borrowck error)
- #91225 (Fix invalid scrollbar display on source code page)
Failed merges:
r? `@ghost`
`@rustbot` modify labels: rollup
Fix ICE due to out-of-bounds statement index when reporting borrowck error
Replace an `[index]` with a `.get` when `statement_index` points to a basic-block terminator (and is therefore out-of-bounds in the statements list).
Fixes#91206
Cc ``@camsteffen``
r? ``@oli-obk``
Change cg_ssa's get_param to borrow the builder mutably
This is a small change to make `get_param` more flexible for codegens that may need to modify things when retrieving function parameters.
This will currently only be used by [rustc_codegen_nvvm](https://github.com/Rust-GPU/Rust-CUDA) (my own project), but may be useful to more codegens in the future.
This is needed because cg_nvvm needs to remap certain types to libnvvm-friendly types, such as `i128` -> `<2 x i64>`. Because cg_ssa does not give mutable access to the builder, i resorted to using a mutex:
```rs
fn get_param(&self, index: usize) -> Self::Value {
let val = llvm::get_param(self.llfn(), index as c_uint);
trace!("Get param `{:?}`", val);
unsafe {
let llfnty = LLVMRustGetFunctionType(self.llfn());
let map = self.remapped_integer_args.borrow();
if let Some((_, key)) = map.get(llfnty) {
if let Some((_, new_ty)) = key.iter().find(|t| t.0 == index) {
trace!("Casting irregular param {:?} to {:?}", val, new_ty);
return transmute_llval(
*self.llbuilder.lock().unwrap(),
&self.cx,
val,
*new_ty,
);
}
}
val
}
}
```
However, i predict this is pretty bad for performance, considering how much builders are called during codegen, so i would greatly appreciate having a more flexible API for this.
Fix stack overflow in `usefulness.rs`
Fix#88747
Applied the suggestion from `@nbdd0121,` not sure if this has any drawbacks. The first call to `ensure_sufficient_stack` is not needed to fix the test case, but I added it to be safe.
Visit `param_env` field in Obligation's `TypeFoldable` impl
This oversight appears to have gone unnoticed for a long time
without causing issues, but it should still be fixed.
Diagnostic tweaks
* On type mismatch caused by assignment, point at the source of the expectation
* Hide redundant errors
* Suggest `while let` when `let` is missing in some cases
* Do not emit unnecessary E0308 after E0070
* Show fewer errors on `while let` missing `let`
* Hide redundant E0308 on `while let` missing `let`
* Point at binding definition when possible on invalid assignment
* do not point at closure twice
* do not suggest `if let` for literals in lhs
* account for parameter types
Do not visit attributes in `ItemLowerer`.
By default, AST visitors visit expressions that appear in key-value attributes.
Those expressions should not be lowered to HIR, as they do not correspond to actually compiled code.
Since an attribute cannot produce meaningful HIR, just skip them altogether.
Fixes https://github.com/rust-lang/rust/issues/81886
Fixes https://github.com/rust-lang/rust/issues/90873
r? `@michaelwoerister`
Print associated types on opaque `impl Trait` types
This PR generalizes #91021, printing associated types for all opaque `impl Trait` types instead of just special-casing for future.
before:
```
error[E0271]: type mismatch resolving `<impl Iterator as Iterator>::Item == u32`
```
after:
```
error[E0271]: type mismatch resolving `<impl Iterator<Item = usize> as Iterator>::Item == u32`
```
---
Questions:
1. I'm kinda lost in binders hell with this one. Is all of the `rebind`ing necessary?
2. Is there a map collection type that will give me a stable iteration order? Doesn't seem like TraitRef is Ord, so I can't just sort later..
3. I removed the logic that suppresses printing generator projection types. It creates outputs like this [gist](https://gist.github.com/compiler-errors/d6f12fb30079feb1ad1d5f1ab39a3a8d). Should I put that back?
4. I also added spaces between traits, `impl A+B` -> `impl A + B`. I quite like this change, but is there a good reason to keep it like that?
r? ````@estebank````
Link with default MACOSX_DEPLOYMENT_TARGET if not otherwise specified.
This PR sets the MACOSX_DEPLOYMENT_TARGET environment variable during the linking stage to our default, if it is not specified. This way it matches the deployment target we pass to llvm. If not set the the linker uses Xcode or Xcode commandline tools default which varies by version.
Fixes#90342, #91082.
Drive-by fixes to make Rust behave more like clang:
* Default to 11.0 deployment target for ARM64 which is the earliest version that had support for it.
* Set the llvm target to `arm64-apple-macosx<deployment target>` instead of `aarch64-apple-macosx<deployment target>`.
Various fixes for const_trait_impl
A few problems I found while making `Iterator` easier to const-implement.
1. More generous `~const Drop` check.
We check for nested fields with caller bounds.
For example, an ADT type with fields of types `A`, `B`, `C`, check if all of them are either:
- Bounded (`A: ~const Drop`, `B: Copy`)
- Known to be able to destruct at compile time (`C = i32`, `struct C(i32)`, `C = some_fn`)
2. Don't treat trait functions marked with `#[default_method_body_is_const]` as stable const fns when checking `const_for` and `const_try` feature gates.
I think anyone can review this, so no r? this time.
Tokenize emoji as if they were valid identifiers
In the lexer, consider emojis to be valid identifiers and reject
them later to avoid knock down parse errors.
Partially address #86102.
Restrict aarch64 outline atomics to glibc for now.
The introduced dependency on `getauxval` causes linking problems with musl, making compiling any binaries for `aarch64-unknown-linux-musl` impossible without workarounds such as using lld or adding liblibc.rlib again to the linker invocation, see #89626.
This is a workaround until libc>0.2.108 is merged.
Optimize live point computation
This refactors the live-point computation to lower per-MIR-instruction costs by operating on a largely per-block level. This doesn't fundamentally change the number of operations necessary, but it greatly improves the practical performance by aggregating bit manipulation into ranges rather than single-bit; this scales much better with larger blocks.
On the benchmark provided in #90445, with 100,000 array elements, walltime for a check build is improved from 143 seconds to 15.
I consider the tiny losses here acceptable given the many small wins on real world benchmarks and large wins on stress tests. The new code scales much better, but on some subset of inputs the slightly higher constant overheads decrease performance somewhat. Overall though, this is expected to be a big win for pathological cases (as illustrated by the test case motivating this work) and largely not material for non-pathological cases. I consider the new code somewhat easier to follow, too.
