Add flag to forbid recovery in the parser
To start the effort of fixing #103534, this adds a new flag to the parser, which forbids the parser from doing recovery, which it shouldn't do in macros.
This doesn't add any new checks for recoveries yet and is just here to bikeshed the names for the functions here before doing more.
r? `@compiler-errors`
rustc_metadata: Add struct and variant constructors to module children at encoding time
instead of decoding time.
Continuation of https://github.com/rust-lang/rust/pull/95899.
The last time it caused some ICEs from generator use, but not everything seems ok.
Clean up hidden type registration
work on https://github.com/rust-lang/rust/issues/101186
Actually passing down the relation and using it instead of `eq` for the hidden type comparison has *no* effect whatsoever and allows for no further improvements at the call sites. I decided the increased complexity was not worth it and thus did not include that change in this PR.
r? `@compiler-errors`
Rollup of 9 pull requests
Successful merges:
- #103035 (Even nicer errors from assert_unsafe_precondition)
- #103106 (Try to say that memory outside the AM is always exposed)
- #103475 (Make param index generation a bit more robust)
- #103525 (Move a wf-check into the site where the value is instantiated)
- #103564 (library: allow some unused things in Miri)
- #103586 (Process registered region obligation in `resolve_regions_with_wf_tys`)
- #103592 (rustdoc: remove redundant CSS selector `.notable-traits .notable`)
- #103593 (Remove an unused parser function (`Expr::returns`))
- #103611 (Add test for issue 103574)
Failed merges:
r? `@ghost`
`@rustbot` modify labels: rollup
Remove `commit_if_ok` probe from NLL type relation
It was not really necessary to add the `commit_if_ok` in #100092 -- I added it to protect us against weird inference error messages due to recursive RPIT calls, but we are always on the error path when this happens anyways, and I can't come up with an example that makes this manifest.
Fixes#103599
r? `@oli-obk` since you reviewed #100092, feel free to re-roll.
🅱️📢 beta-nominating this since it's on beta (which forks in ~a week~ two days 😨) -- worst case we could revert the original PR on beta and land this on nightly, to give it some extra soak time...
The current logic to ignore ThinLTO when `-Ccodegen-units=1` makes sense
for local ThinLTO but even in this scenario, a user may still want
(non-local) ThinLTO for the purpose of optimizing dependencies into the
final crate which is being compiled with 1 CGU.
The previous behavior was even more confusing because if you were
generating a binary (`--emit=link`), then you would get ThinLTO but if
you asked for LLVM IR or bytecode, then it would silently change to
using regular LTO.
With this change, we only override the defaults for local ThinLTO if you
ask for a single output such as LLVM IR or bytecode and in all other
cases honor the requested LTO setting.
Diagnostic derives: allow specifying multiple alternative suggestions
This allows porting `span_suggestions()` to diagnostic structs.
Doesn't work for `multipart_suggestions()` because the rank would be reversed - the struct would specify multiple spans, each of which has multiple possible replacements, while `multipart_suggestions()` creates multiple possible replacements, each with multiple spans.
suggest type annotation for local statement initialed by ref expression
In a local statement with a type declaration, if a ref expression is used on the right side and not used on the left side, in addition to removing the `&` and `&mut` on the right side, we can add them on the left side alternatively
Fixes#102892
Add eval hack in `super_relate_consts` back
Partially reverts 01adb7e98d.
This extra eval call *still* needs to happen, for example, in `normalize_param_env_or_error` when a param-env predicate has an unnormalized constant, since the param-env candidates never get normalized during candidate assembly (everywhere else we can assume that they are normalized fully).
r? `@lcnr,` though I feel like I've assigned quite a few PRs to you in the last few days, so feel free to reassign to someone else familiar with this code if you're busy!
cc #103243 (fixes the issue, but don't want to auto-close that until a backport is performed).
Assert if inference vars are leaking from `InferCtxt::fully_resolve`
`InferCtxt::fully_resolve` shouldn't return unresolved inference vars without us at least being aware of it, so make it an assertion now. This should only happen in cases where we used to be returning `ReEmpty`...
cc `@jackh726`
Split phase change from `MirPass`
The main goal here is to simplify the pass manager logic. `MirPass` no longer contains the `phase_change` method, and `run_passes` instead accepts an `Option<PhaseChange>`. The hope is that this addresses the comments (and maybe perf regression) from #99102 .
r? `@oli-obk` cc `@RalfJung`
This allows writing the following function signatures:
```rust
fn f0() -> impl Fn() -> impl Trait;
fn f3() -> &'static dyn Fn() -> impl Trait;
```
These signatures were already allowed for common traits and associated
types, there is no reason why `Fn*` traits should be special in this
regard.
rustc: Use `unix_sigpipe` instead of `rustc_driver::set_sigpipe_handler`
This is the first (known) step towards starting to use `unix_sigpipe` in the wild. Eventually, `rustc_driver::set_sigpipe_handler` can be removed and all clients can use `unix_sigpipe` instead.
For now we just start using `unix_sigpipe` in one place: `rustc` itself.
It is easy to manually verify this change. If you remove `#[unix_sigpipe = "sig_dfl"]` and run `./x.py build` you will get an ICE when you do `./build/x86_64-unknown-linux-gnu/stage1/bin/rustc --help | false`. Add back `#[unix_sigpipe = "sig_dfl"]` and the ICE disappears again.
PR that added `set_sigpipe_handler`: https://github.com/rust-lang/rust/pull/49606
Tracking issue for `unix_sigpipe`: #97889
Not sure exactly how to label this PR. Going with T-libs for now since this is a T-libs feature.
````@rustdoc```` labels +T-libs
Support raw-dylib functions being used inside inlined functions
Fixes#102714
Issue Details:
When generating the import library for `raw-dylib` symbols, we currently only use the functions and variables declared within the current crate. This works fine if all crates are static libraries or `rlib`s as the generated import library will be contained in the static library or `rlib` itself, but if a dependency is a dynamic library AND the use of a `raw-dylib` function or variable is inlined or part of a generic instantiation then the current crate won't see its dependency's import library and so linking will fail.
Fix Details:
Instead, when we generate the import library for a `dylib` or `bin` crate, we will now generate it for the symbols both for the current crate and all upstream crates. We do this in two steps so that the import library for the current crate is passed into the linker first, thus it is preferred if there are any ambiguous symbols.
`TerminatorCodegenHelper` has three methods `llblock`, `llbb`, and
`lltarget`. They're all similar, but the names given no indication of
the differences.
This commit renames `lltarget` as `llbb_with_landing_pad`, and `llblock`
as `llbb_with_cleanup`. These aren't fantastic names, but at least it's
now clear that `llbb` is the lowest-level of the three and the other two
wrap it.
Ensure:
- builders always have a `bx` suffix;
- backend basic blocks always have an `llbb` suffix,
- paired builders and basic blocks have consistent prefixes.
Delay span bug when we can't map lifetimes back in `collect_trait_impl_trait_tys`
When a lifetime is late-bound in a trait signature, but early-bound in an impl signature, we already emit an error -- however, we also ICE in `collect_trait_impl_trait_tys`, so just delay a bug here.
Fixes#103407
Don't ICE when reporting borrowck errors involving regions from `anonymous_lifetime_in_impl_trait`
The issue here is that when we have:
```
trait Trait<'a> { .. }
fn foo(arg: impl Trait) { .. }
```
The anonymous lifetime `'_` that we generate for `arg: impl Trait` doesn't end up in the argument type (which is a param) but in a where-clause of the function, in a predicate whose self type is that param ty.
Fixes#101660
r? ``@cjgillot``
Only apply `ProceduralMasquerade` hack to older versions of `rental`
The latest version of `rental` (v0.5.6) contains a fix that allows it to
compile without relying on the pretty-print back-compat hack.
Hopefully, there are no longer any crates relying on the affected
versions of the (much less popular) `procedural-masquerade` crate. This
should allow us to target the pretty-print back-compat hack specifically
to older versions of `rental`, and specifically mention upgrading to
`rental` v0.5.6 in the lint message.
Sort tests at compile time, not at startup
Recently, another Miri user was trying to run `cargo miri test` on the crate `iced-x86` with `--features=code_asm,mvex`. This configuration has a startup time of ~18 minutes. That's ~18 minutes before any tests even start to run. The fact that this crate has over 26,000 tests and Miri is slow makes a lot of code which is otherwise a bit sloppy but fine into a huge runtime issue.
Sorting the tests when the test harness is created instead of at startup time knocks just under 4 minutes out of those ~18 minutes. I have ways to remove most of the rest of the startup time, but this change requires coordinating changes of both the compiler and libtest, so I'm sending it separately.
(except for doctests, because there is no compile-time harness)
Shorten the `lookup_line` code slightly
The `match` looks like it's exactly the same as `checked_sub(1)`, so we might as well see if perf says we can just do that to save a couple lines.
This patch makes it possible to use varargs for calling conventions,
which are either based on C (like efiapi) or C is based
on them (for example sysv64 and win64).
Pretty print lifetimes captured by RPIT
This specifically makes the output in #103409 change from:
```diff
error: `impl` item signature doesn't match `trait` item signature
--> $DIR/signature-mismatch.rs:15:5
|
LL | fn async_fn(&self, buff: &[u8]) -> impl Future<Output = Vec<u8>>;
| ----------------------------------------------------------------- expected `fn(&'1 Struct, &'2 [u8]) -> impl Future<Output = Vec<u8>>`
...
LL | fn async_fn<'a>(&self, buff: &'a [u8]) -> impl Future<Output = Vec<u8>> + 'a {
- | ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ found `fn(&'1 Struct, &'2 [u8]) -> impl Future<Output = Vec<u8>>`
+ | ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ found `fn(&'1 Struct, &'2 [u8]) -> impl Future<Output = Vec<u8>> + '2`
|
= note: expected `fn(&'1 Struct, &'2 [u8]) -> impl Future<Output = Vec<u8>>`
- found `fn(&'1 Struct, &'2 [u8]) -> impl Future<Output = Vec<u8>>`
+ found `fn(&'1 Struct, &'2 [u8]) -> impl Future<Output = Vec<u8>> + '2`
= help: the lifetime requirements from the `impl` do not correspond to the requirements in the `trait`
= help: verify the lifetime relationships in the `trait` and `impl` between the `self` argument, the other inputs and its output
error: aborting due to previous error
```
Along with the UI tests in this PR, which I think are all improvements!
r? `@oli-obk` though feel free to re-roll
Fix wrapped valid-range handling in ty_find_init_error
Rust's niche handling allows for wrapping valid ranges with end < start;
for instance, a valid range with start=43 and end=41 means a niche of
42. Most places in the compiler handle this correctly, but
`ty_find_init_error` assumed that `lo > 0` means the type cannot contain a
zero.
Fix it to handle wrapping ranges.
This allows porting uses of span_suggestions() to diagnostic structs.
Doesn't work for multipart_suggestions() because the rank would be
reversed - the struct would specify multiple spans, each of which has
multiple possible replacements, while multipart_suggestions() creates
multiple possible replacements, each with multiple spans.
Enable LTO for rustc_driver.so
Alternative to https://github.com/rust-lang/rust/pull/97154
This enables LTO'ing dylibs behind a feature flag and uses this feature for compiling rustc_driver.so.
Rollup of 6 pull requests
Successful merges:
- #101293 (Recover when unclosed char literal is parsed as a lifetime in some positions)
- #101908 (Suggest let for assignment, and some code refactor)
- #103192 (rustdoc: Eliminate uses of `EarlyDocLinkResolver::all_traits`)
- #103226 (Check `needs_infer` before `needs_drop` during HIR generator analysis)
- #103249 (resolve: Revert "Set effective visibilities for imports more precisely")
- #103305 (Move some tests to more reasonable places)
Failed merges:
r? `@ghost`
`@rustbot` modify labels: rollup
resolve: Revert "Set effective visibilities for imports more precisely"
In theory the change was correct, but in practice the use of import items in HIR is limited and hacky, and it expects that (effective) visibilities for all (up to) 3 IDs of the import are set to the value reflecting (effective) visibility of the whole syntactic `use` item rather than its individual components.
Fixes https://github.com/rust-lang/rust/issues/102352
r? `@oli-obk`
Check `needs_infer` before `needs_drop` during HIR generator analysis
This is kinda a revival of #103036, but with the understanding that after fallback, a generator-interior type will only have `needs_infer` true if there's an error that prevented int or float variable fallback to occur (modulo region variables, which are erased).
Therefore the best choice here is to delay a bug and skip the `needs_drop` call altogether.
r? `@lcnr` feel free to reassign though
Flatten diagnostic slug modules
This makes it easier to grep for the slugs in the code.
See https://rust-lang.zulipchat.com/#narrow/stream/131828-t-compiler/topic/Localization.20infra.20interferes.20with.20grepping.20for.20error for more discussion about it.
This was mostly done with a few regexes and a bunch of manual work. This also exposes a pretty annoying inconsistency for the extra labels. Some of the extra labels are defined as additional properties in the fluent message (which makes them not prefixed with the crate name) and some of them are new fluent messages themselves (which makes them prefixed with the crate name). I don't know whether we want to clean this up at some point but it's useful to know.
r? `@davidtwco`
Change `unknown_lint` applicability to `MaybeIncorrect`
This small PR changes the applicability of `unknown_lint` to `MaybeIncorrect`, because the suggested lint might not be the correct one.
Here is one example where the current applicability causes a problem. Clippy has a set of internal lints guarded by a feature called `internal`. If the feature is not enabled, then the internal lints are "unknown." In that case, running `cargo clippy --fix ...` on `clippy_utils` causes lines such as the followig
26c96e3416/src/tools/clippy/clippy_utils/src/paths.rs (L51-L52)
to be changed to
```rust
#[expect(clippy::invalid_regex)] // internal lints do not know about all external crates
pub const FUTURES_IO_ASYNCREADEXT: [&str; 3] = ["futures_util", "io", "AsyncReadExt"];
```
which is not correct.
Delay ambiguity span bug in normalize query iff not rustdoc
Oli and I decided that the compiler debt of adding another usage of `tcx.sess.opts.actually_rustdoc` is fine, because we don't really want to add more complexity to the normalize query, and moving rustdoc to use fulfill normalization (`fully_normalize`, i.e. not use the normalize query) is unnecessary overhead given that it's skipping binders and stuff.
r? oli-obk
Fixes#102827Fixes#103181
Handle return-position `impl Trait` in traits properly in `register_hidden_type`
The bounds that we get by calling `bound_explicit_item_bounds` from an RPITIT have projections, not opaques, but when we're *registering* an opaque, we want to treat it like an opaque.
Coincidentally fixes#102688 as well, which makes sense, since that was failing because we were inferring an opaque type to be equal to itself (opaque cycle error => "cannot resolve opaque type").
Fixes#103352
r? ```@oli-obk```
Do not suggest trivially false const predicates
Pass through constness to `predicate_can_apply` and don't suggest other impls if it's satisfied but not const.
Fixes#103267
Introduce `subst_iter` and `subst_iter_copied` on `EarlyBinder`
Makes working with bounds lists a bit easier, which I seem to do a lot.
Specifically, means that we don't need to do `.transpose_iter().map(|(pred, _)| *pred)` every time we want to iterate through an `EarlyBinder<&'tcx [(Predicate, Span)]>` (and even then, still have to call `subst` later), which was a very awkward idiom imo.
Rust's niche handling allows for wrapping valid ranges with end < start;
for instance, a valid range with start=43 and end=41 means a niche of
42. Most places in the compiler handle this correctly, but
ty_find_init_error assumed that `lo > 0` means the type cannot contain a
zero.
Fix it to handle wrapping ranges.
Add a test to cover this case.
Allow semicolon after closure within parentheses in macros
#88546 added some parsing logic that if we're parsing a closure, and we're within parentheses, and a semicolon follows, then we must be parsing something erroneous like: `f(|| a; b)`, so it replaces the closure body with an error expression. However, it's valid to parse those tokens if we're within a macro, as in #103222.
This is a bit unsatisfying fix. Is there a more robust way of checking that we're within a macro?
I would also be open to removing this "_It is likely that the closure body is a block but where the braces have been removed_" check altogether at the expense of more verbose errors, since it seems very suspicious in the first place...
Fixes#103222.
Slightly tweak comments wrt `lint_overflowing_range_endpoint`
From the review: https://github.com/rust-lang/rust/pull/101986#discussion_r975610611
It _seemed_ that the lint was not emitted when the `if` check failed, but _actually_ this happens already in a special case and the lint is emitted outside of this function, if this function doesn't. I've cleared up the code/comments a bit, so it's more obvious :)
r? ```@estebank```
Remove byte swap of valtree hash on big endian
This addresses problem reported in #103183. The code was originally introduced in e14b34c386. (see https://github.com/rust-lang/rust/pull/96591)
On big-endian environment, this operation sequence actually put the other half from 128-bit result, thus we got different hash result on LE and BE.
stop using `ty::UnevaluatedConst` directly
best reviewed commit by commit.
simplifies #99798 because we now don't have to expand `ty::UnevaluatedConst` to `ty::Const`.
I also remember some other places where using `ty::UnevaluatedConst` directly was annoying and caused issues, though I don't quite remember what they were rn '^^
r? `@oli-obk` cc `@JulianKnodt`
Get rid of native_library projection queries
They don't seem particularly useful as I don't expect native libraries to change frequently.
Maybe they do provide significant value of keeping incremental compilation green though, I'm not sure.
The latest version of `rental` (v0.5.6) contains a fix that allows it to
compile without relying on the pretty-print back-compat hack.
Hopefully, there are no longer any crates relying on the affected
versions of the (much less popular) `procedural-masquerade` crate. This
should allow us to target the pretty-print back-compat hack specifically
to older versions of `rental`, and specifically mention upgrading to
`rental` v0.5.6 in the lint message.
Introduce deduced parameter attributes, and use them for deducing `readonly` on indirect immutable freeze by-value function parameters.
Introduce deduced parameter attributes, and use them for deducing `readonly` on
indirect immutable freeze by-value function parameters.
Right now, `rustc` only examines function signatures and the platform ABI when
determining the LLVM attributes to apply to parameters. This results in missed
optimizations, because there are some attributes that can be determined via
analysis of the MIR making up the function body. In particular, `readonly`
could be applied to most indirectly-passed by-value function arguments
(specifically, those that are freeze and are observed not to be mutated), but
it currently is not.
This patch introduces the machinery that allows `rustc` to determine those
attributes. It consists of a query, `deduced_param_attrs`, that, when
evaluated, analyzes the MIR of the function to determine supplementary
attributes. The results of this query for each function are written into the
crate metadata so that the deduced parameter attributes can be applied to
cross-crate functions. In this patch, we simply check the parameter for
mutations to determine whether the `readonly` attribute should be applied to
parameters that are indirect immutable freeze by-value. More attributes could
conceivably be deduced in the future: `nocapture` and `noalias` come to mind.
Adding `readonly` to indirect function parameters where applicable enables some
potential optimizations in LLVM that are discussed in [issue 103103] and [PR
103070] around avoiding stack-to-stack memory copies that appear in functions
like `core::fmt::Write::write_fmt` and `core::panicking::assert_failed`. These
functions pass a large structure unchanged by value to a subfunction that also
doesn't mutate it. Since the structure in this case is passed as an indirect
parameter, it's a pointer from LLVM's perspective. As a result, the
intermediate copy of the structure that our codegen emits could be optimized
away by LLVM's MemCpyOptimizer if it knew that the pointer is `readonly
nocapture noalias` in both the caller and callee. We already pass `nocapture
noalias`, but we're missing `readonly`, as we can't determine whether a
by-value parameter is mutated by examining the signature in Rust. I didn't have
much success with having LLVM infer the `readonly` attribute, even with fat
LTO; it seems that deducing it at the MIR level is necessary.
No large benefits should be expected from this optimization *now*; LLVM needs
some changes (discussed in [PR 103070]) to more aggressively use the `noalias
nocapture readonly` combination in its alias analysis. I have some LLVM patches
for these optimizations and have had them looked over. With all the patches
applied locally, I enabled LLVM to remove all the `memcpy`s from the following
code:
```rust
fn main() {
println!("Hello {}", 3);
}
```
which is a significant codegen improvement over the status quo. I expect that if this optimization kicks in in multiple places even for such a simple program, then it will apply to Rust code all over the place.
[issue 103103]: https://github.com/rust-lang/rust/issues/103103
[PR 103070]: https://github.com/rust-lang/rust/pull/103070
make `order_dependent_trait_objects` show up in future-breakage reports
tried to change it to a hard error in #102474 but breaking the more than 1000 dependents of `traitobject` doesn't feel great 😅
This lint has existed since more than 3 years now and the way this is currently implemented is buggy and will break with #102472. imo we should upgrade it to also report for dependencies and maybe also backport this to beta. Then after maybe 2-3 stable versions I would like to finally convert this lint to a hard error.
Change process spawning to inherit the parent's signal mask by default
Previously, the signal mask was always reset when a child process is
started. This breaks tools like `nohup` which expect `SIGHUP` to be
blocked for all transitive processes.
With this change, the default behavior changes to inherit the signal mask.
This also changes the signal disposition for `SIGPIPE` to only be changed if the `#[unix_sigpipe]` attribute isn't set.
Remove more attributes from metadata
A lot of the attributes that are currently stored in the metadata aren't used at all. The biggest metadata usage comes from the doc attributes currently but they are needed by rustdoc so we only removed the ones that cannot be used in downstream crates (doc comments on private items).
r? `@ghost`
Rollup of 6 pull requests
Successful merges:
- #102287 (Elaborate supertrait bounds when triggering `unused_must_use` on `impl Trait`)
- #102922 (Filtering spans when emitting json)
- #103051 (translation: doc comments with derives, subdiagnostic-less enum variants, more derive use)
- #103111 (Account for hygiene in typo suggestions, and use them to point to shadowed names)
- #103260 (Fixup a few tests needing asm support)
- #103321 (rustdoc: improve appearance of source page navigation bar)
Failed merges:
- #103209 (Diagnostic derives: allow specifying multiple alternative suggestions)
r? `@ghost`
`@rustbot` modify labels: rollup
translation: doc comments with derives, subdiagnostic-less enum variants, more derive use
- Adds support for `doc` attributes in the diagnostic derives so that documentation comments don't result in the derive failing.
- Adds support for enum variants in the subdiagnostic derive to not actually correspond to an addition to a diagnostic.
- Made use of the derive in more places in the `rustc_ast_lowering`, `rustc_ast_passes`, `rustc_lint`, `rustc_session`, `rustc_infer` - taking advantage of recent additions like eager subdiagnostics, multispan suggestions, etc.
cc #100717
Elaborate supertrait bounds when triggering `unused_must_use` on `impl Trait`
Given `impl Trait`, if one of its supertraits has a `#[must_use]`, then trigger the lint. This means that, for example, `-> impl ExactSizeIterator` also triggers the `must_use` on `trait Iterator`, which fixes#102183.
This might need `@rust-lang/lang` sign-off, since it changes the behavior of the lint, so cc'ing them.
move hir typeck into separate crate
second part https://github.com/rust-lang/compiler-team/issues/529
I avoided pretty much anything that wasn't just a simple move + path adjustment. Left fixmes for methods which are at an odd place
r? `@compiler-errors`
indirect immutable freeze by-value function parameters.
Right now, `rustc` only examines function signatures and the platform ABI when
determining the LLVM attributes to apply to parameters. This results in missed
optimizations, because there are some attributes that can be determined via
analysis of the MIR making up the function body. In particular, `readonly`
could be applied to most indirectly-passed by-value function arguments
(specifically, those that are freeze and are observed not to be mutated), but
it currently is not.
This patch introduces the machinery that allows `rustc` to determine those
attributes. It consists of a query, `deduced_param_attrs`, that, when
evaluated, analyzes the MIR of the function to determine supplementary
attributes. The results of this query for each function are written into the
crate metadata so that the deduced parameter attributes can be applied to
cross-crate functions. In this patch, we simply check the parameter for
mutations to determine whether the `readonly` attribute should be applied to
parameters that are indirect immutable freeze by-value. More attributes could
conceivably be deduced in the future: `nocapture` and `noalias` come to mind.
Adding `readonly` to indirect function parameters where applicable enables some
potential optimizations in LLVM that are discussed in [issue 103103] and [PR
103070] around avoiding stack-to-stack memory copies that appear in functions
like `core::fmt::Write::write_fmt` and `core::panicking::assert_failed`. These
functions pass a large structure unchanged by value to a subfunction that also
doesn't mutate it. Since the structure in this case is passed as an indirect
parameter, it's a pointer from LLVM's perspective. As a result, the
intermediate copy of the structure that our codegen emits could be optimized
away by LLVM's MemCpyOptimizer if it knew that the pointer is `readonly
nocapture noalias` in both the caller and callee. We already pass `nocapture
noalias`, but we're missing `readonly`, as we can't determine whether a
by-value parameter is mutated by examining the signature in Rust. I didn't have
much success with having LLVM infer the `readonly` attribute, even with fat
LTO; it seems that deducing it at the MIR level is necessary.
No large benefits should be expected from this optimization *now*; LLVM needs
some changes (discussed in [PR 103070]) to more aggressively use the `noalias
nocapture readonly` combination in its alias analysis. I have some LLVM patches
for these optimizations and have had them looked over. With all the patches
applied locally, I enabled LLVM to remove all the `memcpy`s from the following
code:
```rust
fn main() {
println!("Hello {}", 3);
}
```
which is a significant codegen improvement over the status quo. I expect that
if this optimization kicks in in multiple places even for such a simple
program, then it will apply to Rust code all over the place.
[issue 103103]: https://github.com/rust-lang/rust/issues/103103
[PR 103070]: https://github.com/rust-lang/rust/pull/103070
Previously, the signal mask is always reset when a child process is
started. This breaks tools like `nohup` which expect `SIGHUP` to be
blocked.
With this change, the default behavior changes to inherit the signal mask.
This also changes the signal disposition for `SIGPIPE` to only be
changed if the `#[unix_sigpipe]` attribute isn't set.
interpret: remove an incorrect assertion
This fixes an ICE in Miri, [reported](https://rust-lang.zulipchat.com/#narrow/stream/269128-miri/topic/SwitchInt.20with.20no.20targets.3F) by `@saethlin.` The faulty assertion was introduced by 432535da2b, when a previously correct assertion checking that the `otherwise` target exists got replaced by this assertion checking that at least one more target beyond `otherwise` exists.
Sadly we don't have a small reproducer so I don't think we can easily add a testcase.
linker: Fix weak lang item linking with combination windows-gnu + LLD + LTO
In https://github.com/rust-lang/rust/pull/100404 this logic was originally disabled for MSVC due to issues with LTO, but the same issues appear on windows-gnu with LLD because that LLD uses the same underlying logic as MSVC LLD, just with re-syntaxed command line options.
So this PR just disables it for LTO builds in general.
The canonical form of a usub.with.overflow check in LLVM are
separate sub + icmp instructions, rather than a usub.with.overflow
intrinsic. Using usub.with.overflow will generally result in worse
optimization potential.
The backend will attempt to form usub.with.overflow when it comes
to actual instruction selection. This is not fully reliable, but
I believe this is a better tradeoff than using the intrinsic in
IR.
Fixes#103285.
Fix the bug of next_point in source_map
There is a bug in `next_point`, the new span won't move to next position when be called in the first time.
For this reason, our current code is working like this:
1. When we really want to move to the next position, we called two times of `next_point`
2. Some code which use `next_point` actually done the same thing with `shrink_to_hi`
This fix make sure when `next_point` is called, span will move with the width at least 1, and also work correctly in the scenario of multiple bytes.
Ref: https://github.com/rust-lang/rust/pull/103140#discussion_r997710998
r? `@davidtwco`
Deny hashing ty/re/ct inference variables
cc `@cjgillot` and https://github.com/rust-lang/rust/pull/102695#issuecomment-1275706528
r? `@lcnr`
best reviewed one commit at a time, mostly because the second commit that fixes `ClosureOutlivesRequirement` is mostly noise because of losing its `<'tcx>` lifetime parameter.
Do anonymous lifetimes remapping correctly for nested rpits
Closes#103141
r? `@cjgillot` `@nikomatsakis`
This fixes a stable to stable regression that in my opinion is `P-critical` so, we probably want to backport it all the way up to stable.
Use already checked RHS ty for LHS deref suggestions
There's no reason to do the `check_lhs_assignable` and RHS `check_expr_with_hint` in that order, so invert them and use the typeck results to avoid exponential blowup on error.
Fixes#103219
Use Set instead of Vec in transitive_relation
Helps with #103195. It doesn't fix the underlying quadraticness but it makes it _a lot_ faster to an extent where even doubling the amount of nested references still takes less than two seconds (50s on nightly).
I want to see whether this causes regressions (because the vec was usually quite small) or improvements (as lookup for bigger sets is now much faster) in real code.
Let expressions on RHS shouldn't be terminating scopes
Fixes#100276.
Before this PR, we were unconditionally marking the RHS of short-circuiting binary expressions as a terminating scope.
In the case of a let chain where the `let` expression was on the RHS, this meant that temporaries within the `let` expr would only live until the end of the expression. Since this only affected the RHS, this led to surprising behavior ([example](https://play.rust-lang.org/?version=nightly&mode=debug&edition=2021&gist=d1b0a5d1f01882f9c89c2194a75eb19f)).
After this PR, we only mark the RHS as a terminating scope if it is not a `let` expression.
Standardize "use parentheses to call" suggestions between typeck and trait selection
1. Suggest calling constructors, since they're basically `FnDef`s but they have a different def kind and hir representation, so we were leaving them out.
2. Standardize the call suggestions between trait fulfillment errors and type mismatch. In the type mismatch suggestion, we suggest `/* Ty */` as the placeholder for an arg, and not the parameter's name, which is less helpful.
3. Use `predicate_must_hold_modulo_regions` instead of matching on `EvaluationResult` -- this might cause some suggestions to be filtered out, but we really shouldn't be suggesting a call if it "may" hold, only when it "must" hold.
4. Borrow some logic from `extract_callable_info` to generalize this suggestion to fn pointers, type parameters, and opaque types.
Fixes#102852
- Make the structure of the two variants more similar.
- Add some comments.
- Move various conditional `use` items inside the function that uses
them.
- Inline some closures.
