clarify comment in RawVec::into_box
On first reading I almost thought "len <= cap" would be all that there is to check here. Expand the comment to clarify that that is not the case.
Fix some aliasing issues in Vec
`Vec::extend` and `Vec::truncate` invalidated references into the vector even without reallocation, because they (implicitly) created a mutable reference covering the *entire* initialized part of the vector.
Fixes https://github.com/rust-lang/rust/issues/70301
I verified the fix by adding some new tests here that I ran in Miri.
Fix performance regression in debuginfo file_metadata.
Fixes performance regression caused by #69718.
Finding the `SourceFile` associated with a `FileName` called `get_source_file` on the `SourceMap`, which does a linear search through all files in the `SourceMap`.
This resolves the issue by passing the `SourceFile` in from the caller (which already had it available) instead of the `FileName`
Fixes#70785.
Finding the `SourceFile` associated with a `FileName` called `get_source_file` on
the `SourceMap`, which does a linear search through all files in the `SourceMap`.
This resolves the issue by passing the SourceFile in from the caller (which already
had it available).
use ManuallyDrop instead of forget inside collections
This PR changes some usage of `mem::forget` into `mem::ManuallyDrop` in some `Vec`, `VecDeque`, `BTreeMap` and `Box` methods.
Before the commit, the generated IR for some of the methods was longer, and even after optimization, some unwinding artifacts were still present.
Add slice::fill
Adds the `slice::fill` method to fill a slice with an item. This replaces manual for loops where items are copied one-by-one. This is a counterpart to C++20's [`std::fill`](https://en.cppreference.com/w/cpp/algorithm/fill) function.
## Usage
```rust
let mut buf = vec![0; 10];
buf.fill(1);
assert_eq!(buf, vec![1; 10]);
```
## Performance
When compiling in release mode, for `[u8]` and `[u16]` this method will optimize to a `memset(3)` call ([godbolt](https://godbolt.org/z/85El_c)). The initial implementation relies on LLVM's optimizer to make it as fast as possible for any given input. But as @jonas-schievink [pointed out](https://twitter.com/sheevink/status/1245756597453885442) this can later be optimized through specialization to guarantee it has a specific performance profile.
## Why now?
Conversations about adding `slice::fill` are not new. In fact, https://github.com/rust-lang/rfcs/issues/2067 was opened 3 years ago about this exact topic. However discussion stranded while discussing implementation details, and it's not seen much forward motion since.
In ["The Hunt for the Fastest Zero"](https://travisdowns.github.io/blog/2020/01/20/zero.html) Travis Downs provides disects C++'s `std::fill` performance profile on gcc, comparing it among others to `memset(3)`. Even though `memset(3)` outperforms `std::fill` in their tests, the author notes the following:
> That the optimization fails, perhaps unexpectedly, in some cases is unfortunate but it’s nice that you can fix it yourself. [...] Do we throw out modern C++ idioms, at least where performance matters, for example by replacing std::fill with memset? I don’t think so.
Much of the article focuses on how how to fix the performance of `std::fill` by providing specializations for specific input. In Rust we don't have any dedicated methods to fill slices with values, so it either needs to be optimized at the MIR layer, or more likely rely on LLVM's optimizer.
By adding a dedicated method for filling slices with values it opens up the ability for us to in the future guarantee that e.g. `Vec<u8>` will always optimize to `memset` even in debug mode. Or perhaps provide stronger guarantees about memory when zeroing values when a certain flag is passed. But regardless of that, it improves general ergonomics of working with slices by providing a dedicated method with documentation and examples.
## References
- [slice-fill prototype on docs.rs](https://docs.rs/slice-fill/1.0.1/slice_fill/)
- [The Hunt For The Fastest Zero](https://travisdowns.github.io/blog/2020/01/20/zero.html)
- [Safe memset for slices](https://github.com/rust-lang/rfcs/issues/2067)
- [C++20 std::fill](https://en.cppreference.com/w/cpp/algorithm/fill)
- [ASM output on Godbolt](https://godbolt.org/z/5-XU66)
Do not disable field reordering on enums with big discriminant
The field are always re-ordered to minimize padding, regardless of the
alignment of the discriminant
(spinoff from #70477)
This commit changes some usage of mem::forget into mem::ManuallyDrop
in some Vec, VecDeque, BTreeMap and Box methods.
Before the commit, the generated IR for some of the methods was
longer, and even after optimization, some unwinding artifacts were
still present.
typeck/type_of: let wfcheck handle generics in opaque types' substs.
I was working on #70164, and `type_of`'s handling of opaque types seemed to be, by far, the trickiest use of `Ty::walk`, but I believe it wasn't doing anything (see https://github.com/rust-lang/rust/pull/57896#discussion_r396064431 - I suspect, based on glancing at the PR discussion, that an early attempt was kept in, despite becoming just an overcomplicated way to do exactly the same as the previous simple type equality check).
I would've loved to remove `ResolvedOpaqueTy` (keep the `Ty` and lose the `Substs`), but it looks like the MIR borrowck part of the process needs that now, so it would've been added anyway since #57896, even if that PR hadn't happened.
<hr/>
In the process, I've moved the remaining substitution validation to `wfcheck`, which was already handling lifetimes, and kept only `delay_span_bug`s in `type_of`, as an insurance policy.
I've added tests for lifetime and const cases, they seem to be checked correctly now.
(and more uniform than they were in https://github.com/rust-lang/rust/issues/63063#issuecomment-602162804)
However, the quality of the errors is maybe a bit worse, and they don't trigger when there are other errors (not sure if this is due to compilation stop points or something more specific to one opaque type).
r? @nikomatsakis cc @matthewjasper @oli-obk @Aaron1011
add basic IP support in HermitCore
- add initial version to support sockets
- use TcpStream as test case
- HermitCore uses smoltcp as IP stack for pure Rust applications
- further functionalities (e.g. UDP support) will be added step by step
- in principle, the current PR is a revision of #69404
Add hash of source files in debug info
LLVM supports placing the hash of source files inside the debug info.
This information can be used by a debugger to verify that the source code matches
the executable.
This change adds support for both hash algorithms supported by LLVM, MD5 and SHA1, controlled by a target option.
* DWARF only supports MD5
* LLVM IR supports MD5 and SHA1 (and SHA256 in LLVM 11).
* CodeView (.PDB) supports MD5, SHA1, and SHA256.
Fixes#68980.
Tracking issue: #70401
rustc dev guide PR with further details: https://github.com/rust-lang/rustc-dev-guide/pull/623
Make the rustc respect the `-C codegen-units` flag in incremental mode.
This PR implements (the as of yet unapproved) major change proposal at https://github.com/rust-lang/compiler-team/issues/245. See the description there for background and rationale.
The changes are pretty straightforward and should be easy to rebase if the proposal gets accepted at some point.
r? @nikomatsakis cc @pnkfelix
Place TLS initializers with relocations in .tdata
Should fix#70673, although I'm not sure how to test this. Perhaps @joshlf could find a MCVE?
Also adds more context to the FIXME.
r? @oli-obk
Fix double-free and undefined behaviour in libstd::syn::unix::Thread::new
While working on concurrency support for Miri, I found that the `libstd::syn::unix::Thread::new` method has two potential problems: double-free and undefined behaviour.
**Double-free** could occur if the following events happened (credit for pointing this out goes to @RalfJung):
1. The call to `pthread_create` successfully launched a new thread that executed to completion and deallocated `p`.
2. The call to `pthread_attr_destroy` returned a non-zero value causing the `assert_eq!` to panic.
3. Since `mem::forget(p)` was not yet executed, the destructor of `p` would be executed and cause a double-free.
As far as I understand, this code also violates the stacked-borrows aliasing rules and thus would result in **undefined behaviour** if these rules were adopted. The problem is that the ownership of `p` is passed to the newly created thread before the call to `mem::forget`. Since the call to `mem::forget` is still a call, it counts as a use of `p` and triggers UB.
This pull request changes the code to use `mem::ManuallyDrop` instead of `mem::forget`. As a consequence, in case of a panic, `p` would be potentially leaked, which while undesirable is probably better than double-free or undefined behaviour.
Use associated numeric consts in documentation
Now when the associated constants on int/float types are stabilized and the recommended way of accessing said constants (#68952). We can start using it in this repository, and recommend it via documentation example code.
This PR is the reincarnation of #67913 minus the actual adding + stabilization of said constants. (EDIT: Now it's only changing the documentation. So users will see the new consts, but we don't yet update the internal code)
Because of how fast bit rot happens to PRs that touch this many files, it does not try to replace 100% of the old usage of the constants in the entire repo, but a good chunk of them.
bootstrap: add `--json-output` for rust-analyzer
Motivation is that this allows us to customize rust-analyzer's "cargo watch" integration to run x.py. You simply have to set the command to run to be `x.py --json-output`
r? @Mark-Simulacrum -- feel free to make changes, this is quick and dirty for sure
Miri engine: stronger type-based sanity check for assignments
r? @oli-obk @eddyb
Fixes https://github.com/rust-lang/rust/issues/70405
That issue says
> be sure to also add appropriate mutability checks to the patterns (mutable for the source, immutable for the dest)
I decided not to do that because I see no good reason to do it. The engine does not care either way, the assignment will happen correctly.
