switch `jemalloc-sys` back to `tikv-jemalloc-sys`, and update to 0.6.0
Some context:
- we used to use jemalloc bindings from https://github.com/gnzlbg/jemallocator, since #55238
- that crate was abandoned, picked up as a fork in https://github.com/tikv/jemallocator, so we switched to that in #83152.
- then they were able to publish to the original `jemalloc-sys` bindings crate, and `jemalloc-sys` and `tikv-jemalloc-sys` became the same thing -- so I switched back to the OG crate in #96790
- they're now having publishing problems again: I've been waiting for https://github.com/tikv/jemallocator/pull/96 for the `jemalloc-sys` 0.6.0 update for a few months, but `tikv-jemalloc-sys` is already updated to 0.6.0.
A perf run showed some improvements, so this PR switches back to `tikv-jemalloc-sys` to update to 0.6.0.
Unify `sysroot_target_{bin,lib}dir` handling
Follow-up to https://github.com/rust-lang/rust/pull/131405#discussion_r1826558962 where `sysroot_target_bindir` had to do some dancing because the sysroot ensure logic embedded in `sysroot_target_libdir` returned `$sysroot/$relative_lib/rustlib/$target/lib` and not the `rustlib` parent `$sysroot/$relative_lib/rustlib/`.
This PR pulls out the sysroot ensure logic into a helper, and return `$sysroot/$relative_lib/rustlib/` instead so `sysroot_target_bindir` doesn't have to do parent traversal from the path returned from `sysroot_target_libdir`, and also make them easier to follow in that they are now clearly closely related based on the common target sysroot ensure logic.
Change `AttrArgs::Eq` to a struct variant
Cleanups for simplifying https://github.com/rust-lang/rust/pull/131808
Basically changes `AttrArgs::Eq` to a struct variant and then avoids several matches on `AttrArgsEq` in favor of methods on it. This will make future refactorings simpler, as they can either keep methods or switch to field accesses without having to restructure code
Reducing `target_feature` check-cfg merge conflicts
It was rightfully pointed in https://github.com/rust-lang/rust/pull/133099#discussion_r1862490542 that the expected values for the `target_feature` cfg are regularly updated and unfortunately the check-cfg tests for it are very merge-conflict prone.
This PR aims at drastically reducing the likely-hood of those, by normalizing the "and X more" diagnostic, as well as making the full expected list multi-line instead of being on a single one.
cc `@RalfJung`
r? `@jieyouxu`
add "profiler" and "optimized-compiler-builtins" option coverage for ci-rustc
Adds "profiler" and "optimized-compiler-builtins" option coverage in CI-rustc config compatibility check.
Resolves#133675
Use c"lit" for CStrings without unwrap
I've reviewed uses of `CString::new("lit")`.
Some could be changed to `c"lit"`. Some could be changed to `c"lit".to_owned()`, avoiding an `unwrap()`.
Many `CString` documentation examples could be simplified. I deliberately haven't changed all the examples to use the exact same expression, so that they can demonstrate many ways of creating `CString`s.
I've left UI tests mostly unchanged, because `c""` requires edition 2021, but most UI tests use 2015, and I didn't want to accidentally change what the tests are testing.
build `rustc` with 1 CGU on `aarch64-apple-darwin`
Distribute `aarch64-apple-darwin` artifacts built with `rust.codegen-units=1`, like we already do on Linux/Windows/macOS Intel.
1. Performance results (only wall-time on mac as usual)
- I only did some `ripgrep` check builds (the revision from rustc-perf), because we had noticeable wins in wall-time there back then on linux. It's a ~2-4% (mostly 3%) improvement on local builds.
```console
Benchmark 1: cargo +caa81728c3 check -q
Time (mean ± σ): 5.800 s ± 0.087 s [User: 16.048 s, System: 2.294 s]
Range (min … max): 5.725 s … 6.028 s 10 runs
Benchmark 2: cargo +41f40c97bbff6c8642f5510d0be283551b095b70 check -q
Time (mean ± σ): 5.551 s ± 0.037 s [User: 15.451 s, System: 2.252 s]
Range (min … max): 5.477 s … 5.602 s 10 runs
Summary
'cargo +41f40c97bbff6c8642f5510d0be283551b095b70 check -q' ran
1.04 ± 0.02 times faster than 'cargo +caa81728c3 check -q'
```
```console
Benchmark 1: CARGO_INCREMENTAL=0 cargo +caa81728c3 check -q
Time (mean ± σ): 5.743 s ± 0.030 s [User: 16.005 s, System: 2.249 s]
Range (min … max): 5.720 s … 5.792 s 5 runs
Benchmark 2: CARGO_INCREMENTAL=0 cargo +41f40c97bbff6c8642f5510d0be283551b095b70 check -q
Time (mean ± σ): 5.469 s ± 0.055 s [User: 15.244 s, System: 2.110 s]
Range (min … max): 5.404 s … 5.524 s 5 runs
Summary
'CARGO_INCREMENTAL=0 cargo +41f40c97bbff6c8642f5510d0be283551b095b70 check -q' ran
1.05 ± 0.01 times faster than 'CARGO_INCREMENTAL=0 cargo +caa81728c3 check -q'
```
```console
Benchmark 1: CARGO_INCREMENTAL=0 cargo +caa81728c3 check -q -j1
Time (mean ± σ): 15.092 s ± 0.049 s [User: 11.969 s, System: 1.665 s]
Range (min … max): 15.052 s … 15.165 s 5 runs
Benchmark 2: CARGO_INCREMENTAL=0 cargo +41f40c97bbff6c8642f5510d0be283551b095b70 check -q -j1
Time (mean ± σ): 14.623 s ± 0.035 s [User: 11.520 s, System: 1.619 s]
Range (min … max): 14.593 s … 14.682 s 5 runs
Summary
'CARGO_INCREMENTAL=0 cargo +41f40c97bbff6c8642f5510d0be283551b095b70 check -q -j1' ran
1.03 ± 0.00 times faster than 'CARGO_INCREMENTAL=0 cargo +caa81728c3 check -q -j1'
```
2. Effects on code size on `aarch64-apple-darwin`: it's a 13.24% reduction on `librustc_driver.dylib`
- [before](https://ci-artifacts.rust-lang.org/rustc-builds/caa81728c37f5ccfa9a0979574b9272a67f8a286/rustc-nightly-aarch64-apple-darwin.tar.xz): 173452888 (57.3 MB compressed).
- [after](https://ci-artifacts.rust-lang.org/rustc-builds/41f40c97bbff6c8642f5510d0be283551b095b70/rustc-nightly-aarch64-apple-darwin.tar.xz): 150471368 (55.2 MB compressed, -3.7% here).
3. Effects on CI
- the [1st try build](https://github.com/rust-lang-ci/rust/actions/runs/12118946198/job/33784512526) took 1h31
- the [2nd try build](https://github.com/rust-lang-ci/rust/actions/runs/12121452825/job/33792429553) took 1h28
- I don't know how long the builder usually takes
remove `Ty::is_copy_modulo_regions`
Using these functions is likely incorrect if an `InferCtxt` is available, I moved this function to `TyCtxt` (and added it to `LateContext`) and added a note to the documentation that one should prefer `Infer::type_is_copy_modulo_regions` instead.
I didn't yet move `is_sized` and `is_freeze`, though I think we should move these as well.
r? `@compiler-errors` cc #132279
Add `needs-target-has-atomic` directive
Before this PR, the test writer has to specify platforms and architectures by hand for targets that have differing atomic width support. `#[cfg(target_has_atomic="...")]` is not quite the same because (1) you may have to specify additional matchers manually which has to be maintained individually, and (2) the `#[cfg]` blocks does not communicate to compiletest that a test would be ignored for a given target.
This PR implements a `//@ needs-target-has-atomic` directive which admits a comma-separated list of required atomic widths that the target must satisfy in order for the test to run.
```
//@ needs-target-has-atomic: 8, 16, ptr
```
See <https://github.com/rust-lang/rust/issues/87377>.
This PR supersedes #133095 and is co-authored by `@kei519,` because it was somewhat subtle, and it turned out easier to implement than to review.
rustc-dev-guide docs PR: https://github.com/rust-lang/rustc-dev-guide/pull/2154
Remove `//@ compare-output-lines-by-subset`
There was only ever one test which used this flag, and it was removed in https://github.com/rust-lang/rust/pull/132244. I think this is a bad flag that should never have been added; comparing by subset makes the test failures extremely hard to debug. Any test that needs complicated output filtering like this should just use run-make instead.
Note that this does not remove the underlying comparison code, because it's still used if `runner` is set. I don't quite understand what's going on there, but since we still test on other platforms and in CI that the full output is accurate, I think it will be easier to debug than a test that uses compare-by-subset unconditionally.
rustc-dev-guide update PR: https://github.com/rust-lang/rustc-dev-guide/pull/2151
rustdoc-json: Include safety of `static`s
`static`s in an `extern` block can have an associated safety annotation ["because there is nothing guaranteeing that the bit pattern at the static’s memory is valid for the type it is declared with"](https://doc.rust-lang.org/reference/items/external-blocks.html#statics). Rustdoc already knows this and displays in for HTML. This PR also includes it in JSON.
Inspired by https://github.com/obi1kenobi/cargo-semver-checks/issues/975 which needs this, but it's probably useful in other places.
r? `@GuillaumeGomez.` Possibly easier to review commit-by-commit.
Eliminate magic numbers from expression precedence
Context: see https://github.com/rust-lang/rust/pull/133140.
This PR continues on backporting Syn's expression precedence design into rustc. Rustc's design used mysterious integer quantities represented variously as `i8` or `usize` (e.g. `PREC_CLOSURE = -40i8`), a special significance around `0` that is never named, and an extra `PREC_FORCE_PAREN` precedence level that does not correspond to any expression. Syn's design uses a C-like enum with variants that clearly correspond to specific sets of expression kinds.
This PR is a refactoring that has no intended behavior change on its own, but it unblocks other precedence work that rustc's precedence design was poorly suited to accommodate.
- Asymmetrical precedence, so that a pretty-printer can tell `(return 1) + 1` needs parens but `1 + return 1` does not.
- Squashing the `Closure` and `Jump` cases into a single precedence level.
- Numerous remaining false positives and false negatives in rustc pretty-printer's parenthesization of macro metavariables, for example in `$e < rhs` where $e is `lhs as Thing<T>`.
FYI `@fmease` — you don't need to review if rustbot picks someone else, but you mentioned being interested in the followup PRs.
Before this commit, the test writer has to specify platforms and
architectures by hand for targets that have differing atomic width
support. `#[cfg(target_has_atomic)]` is not quite the same because (1)
you may have to specify additional matchers manually which has to be
maintained individually, and (2) the `#[cfg]` blocks does not
communicate to compiletest that a test would be ignored for a given
target.
This commit implements a `//@ needs-target-has-atomic` directive which
admits a comma-separated list of required atomic widths that the target
must satisfy in order for the test to run.
```
//@ needs-target-has-atomic: 8, 16, ptr
```
See <https://github.com/rust-lang/rust/issues/87377>.
Co-authored-by: kei519 <masaki.keigo.q00@kyoto-u.jp>
Remove `hir::ArrayLen`
This refactoring removes `hir::ArrayLen`, replacing it with `hir::ConstArg`. To represent inferred array lengths (previously `hir::ArrayLen::Infer`), a new variant `ConstArgKind::Infer` is added.
r? `@BoxyUwU`
There was only ever one test which used this flag, and it was removed in https://github.com/rust-lang/rust/pull/132244. I think this is a bad flag that should never have been added; comparing by subset makes the test failures extremely hard to debug. Any test that needs complicated output filtering like this should just use run-make instead.
Note that this does not remove the underlying comparison code, because it's still used if `runner` is set. I don't quite understand what's going on there, but since we still test on other platforms and in CI that the full output is accurate, I think it will be easier to debug than a test that uses compare-by-subset unconditionally.
rust_analyzer_settings: force use of 'nightly' toolchain
The cranelift folder contains a rust-toolchain file. That means when RA opens `compiler/rustc_codegen_cranelift/Cargo.toml`, it will try to use that toolchain or fail. (Maybe that toolchain gets auto-installed for others? On my system, it just fails, but I also run vscodium in a sandbox.)
However, it shouldn't be necessary to use more than one toolchain for the rustc workspace. So we can set the `RUSTUP_TOOLCHAIN` variable on the server side to force the same toolchain to be used everywhere.
Suggested by `@ChayimFriedman2` in https://github.com/rust-lang/rust-analyzer/issues/18585.
Stop cloning `Context` so much
This is a first step for https://github.com/rust-lang/rust/issues/82381.
It's already big enough so I'll continue in a follow-up once this PR is merged. Next step will be to get rid of `SharedContext` by inlining it directly into `Context`.
cc `@camelid`
r? `@notriddle`
Rollup of 7 pull requests
Successful merges:
- #132974 (Properly pass linker arguments that contain commas)
- #133403 (Make `adjust_fulfillment_errors` work with `HostEffectPredicate` and `const_conditions`)
- #133482 (Only error raw lifetime followed by `\'` in edition 2021+)
- #133595 (Do not emit `missing_doc_code_examples` rustdoc lint on module and a few other items)
- #133669 (Move some functions out of const_swap feature gate)
- #133674 (Fix chaining `carrying_add`s)
- #133691 (Check let source before suggesting annotation)
r? `@ghost`
`@rustbot` modify labels: rollup
Do not emit `missing_doc_code_examples` rustdoc lint on module and a few other items
It doesn't make sense to expect modules to have code examples. Same goes for:
* Trait aliases
* Foreign items
* Associated types and constants
Should make the use of this lint a bit nicer.
r? ``@notriddle``
