wasm32: Default to a "static" relocation model
LLVM 9 is adding support for a "pic" relocation model for wasm code,
which is quite different than the current model. In order to preserve
the mode of compilation that we have today default to "static" to ensure
that we don't accidentally start creating experimental relocatable
binaries.
improve worst-case performance of HashSet.is_subset
One more simple optimization opportunity for HashSet that was applied in BTreeSet in #59186 (and wasn't in #57043). Already covered by the existing unit test.
r? @KodrAus
std: Upgrade `compiler_builtins` to fix wasi linkage
Turns out we needed to exclude a number of math functions on the
`wasm32-unknown-wasi` target, and this was fixed in 0.1.9 of
compiler-builtins and this is pulling in the fix to libstd's own build.
ci: Update FreeBSD tarball downloads
These appear to have disappeared from the original server, so I acquired
the contents from a different mirror and uploaded them to our S3 bucket
LLVM 9 is adding support for a "pic" relocation model for wasm code,
which is quite different than the current model. In order to preserve
the mode of compilation that we have today default to "static" to ensure
that we don't accidentally start creating experimental relocatable
binaries.
Turns out we needed to exclude a number of math functions on the
`wasm32-unknown-wasi` target, and this was fixed in 0.1.9 of
compiler-builtins and this is pulling in the fix to libstd's own build.
std: Avoid usage of `Once` in `Instant`
This commit removes usage of `Once` from the internal implementation of
time utilities on OSX and Windows. It turns out that we accidentally hit
a deadlock today (#59020) via events that look like:
* A thread invokes `park_timeout`
* Internally, only on OSX, `park_timeout` calls `Instant::elapsed`
* Inside of `Instant::elapsed` on OSX we enter a `Once` to initialize
global timer data
* Inside of `Once`, it attempts to `park`
This means on the same stack frame, when there's contention, we're
calling `park` from inside `park_timeout`, causing a deadlock!
The solution implemented in this commit was to remove usage of `Once`
and instead just do a small dance with atomics. There's no real need we
need to guarantee that the global information is only learned once, only
that it's only *stored* once. This implementation may have multiple
threads invoke `mach_timebase_info`, but only one will store the global
information which will amortize the cost for all other threads.
A similar fix has been applied to windows to be uniform across our
implementations, but looking at the code on Windows no deadlock was
possible. This is purely just a consistency update for Windows and in
theory a slightly leaner implementation.
Closes#59020
This commit removes usage of `Once` from the internal implementation of
time utilities on OSX and Windows. It turns out that we accidentally hit
a deadlock today (#59020) via events that look like:
* A thread invokes `park_timeout`
* Internally, only on OSX, `park_timeout` calls `Instant::elapsed`
* Inside of `Instant::elapsed` on OSX we enter a `Once` to initialize
global timer data
* Inside of `Once`, it attempts to `park`
This means on the same stack frame, when there's contention, we're
calling `park` from inside `park_timeout`, causing a deadlock!
The solution implemented in this commit was to remove usage of `Once`
and instead just do a small dance with atomics. There's no real need we
need to guarantee that the global information is only learned once, only
that it's only *stored* once. This implementation may have multiple
threads invoke `mach_timebase_info`, but only one will store the global
information which will amortize the cost for all other threads.
A similar fix has been applied to windows to be uniform across our
implementations, but looking at the code on Windows no deadlock was
possible. This is purely just a consistency update for Windows and in
theory a slightly leaner implementation.
Closes#59020
Make some of lexer's API private
Lexer is a `pub` type, so it feels wrong that its fields are just pub (I guess it wasn't exported initially), so let's minimize visibility.
Context: I am looking into extracting rust-lexer into a library, which can be shared by rust-analyzer and rustc. I hope that a simple interface like `fn next_token(src: &str) -> (TokenKind, usize)` would work, but to try this out I need to understand what is the current API of the lexer.
Never return uninhabited values at all
Functions with uninhabited return values are already marked `noreturn`,
but we were still generating return instructions for this. When running
with `-C passes=lint`, LLVM prints:
Unusual: Return statement in function with noreturn attribute
The LLVM manual makes a stronger statement about `noreturn` though:
> This produces undefined behavior at runtime if the function ever does
dynamically return.
We now emit an `abort` anywhere that would have tried to return an
uninhabited value.
Fixes#48227
cc #7463#48229
r? @eddyb
Forward formatter settings to bounds of `Range<T>` in `fmt::Debug` impl
Before this change, formatter settings were lost when printing a `Range`. For example, printing a `Range<f32>` with `{:.2?}` would not apply the precision modifier when printing the floats. Now the `Debug` impls look a bit more verbose, but modifier are not lost.
---
I assume the exact output of `Debug` impls in `std` cannot be relied on by users and thus can change, right?
wasi: Implement more of the standard library
This commit fills out more of the `wasm32-unknown-wasi` target's standard library, notably the `std::fs` module and all of its internals. A few tweaks were made along the way to non-`fs` modules, but the last commit contains the bulk of the work which is to wire up all APIs to their equivalent on WASI targets instead of unconditionally returning "unsupported". After this some basic filesystem operations and such should all be working in WASI!
Support using LLVM's libunwind as the unwinder implementation
This avoids the dependency on host libraries such as libgcc_s which
may be undesirable in some deployment environments where these aren't
available.
Rollup of 6 pull requests
Successful merges:
- #59316 (Internal lints take 2)
- #59663 (Be more direct about borrow contract)
- #59664 (Updated the documentation of spin_loop and spin_loop_hint)
- #59666 (Updated the environment description in rustc.)
- #59669 (Reduce repetition in librustc(_lint) wrt. impl LintPass by using macros)
- #59677 (rustfix coverage: Skip UI tests with non-json error-format)
Failed merges:
r? @ghost
rustfix coverage: Skip UI tests with non-json error-format
When using the `rustfix-coverage` flag, some tests currently fail
because they define a different error-format than `json`.
The current implementation crashes when encountering those tests. Since
we don't care about non-json test output when collecting the coverage
data, we handle those tests by returning an empty `Vec` instead.
r? @oli-obk
Updated the environment description in rustc.
# Description
- Updated the "environment" description in the `rustc` man pages
The old wording suggested that all the mentioned flags influenced the output of the compiler,
where this was not the case.
closes#59504
Updated the documentation of spin_loop and spin_loop_hint
# Description
- Updated the description of `core::hints::spin_loop`
- Updated the description of `core::async::spin_loop_hint`
Both documentation is rewritten to better reflect when one should prefer using a busy-wait spin-loop (and the `spin_loop` and `spin_loop_hint` functions) over `yield_now`. It also dives a little bit deeper on what the function actually does.
closes#55418
Be more direct about borrow contract
I always was confused by the difference between Borrow and AsRef, despite the fact that I've read all available docs at least a dozen of times.
I finally grokked the difference between the two when I realized the Borrow invariant:
> If you implement Borrow, you **must** make sure that Eq, Ord and Hash implementations are equivalent for borrowed and owned data
My problem was that this invariant is not stated explicitly in documentation, and instead some vague and philosophical notions are used.
So I suggest to mention the requirements of `Borrow` very explicitly: instead of "use Borrow when X and use AsRef when Y", let's phrase this as `Borrow` differs from `AsRef` in `W`, so that's why `Borrow` is for `X` and `AsRef` is for `Y`.
Note that this change could be seen as tightening contract of the Borrow. Let's say Alice has written the following code:
```rust
#[derive(PartialEq, Eq, Hash, PartialOrd, Ord)]
struct Person {
first_name: String,
last_name: String,
}
impl Borrow<str> for Person {
fn borrow(&self) -> &str { self.first_name.as_str() }
}
```
Now Bob uses this `Person` struct, puts it into `HashMap` and tries to look it up using `&str` for the first name. Bob's code naturally fails.
The question is, who is to blame: Alice, who has written the impl, or Bob, who uses the HashMap. If I read the current docs literally, I would say that `Bob` is to blame: `Eq` and `Hash` bounds appear on HashMap, so it is the HashMap which requires that they are consistent. By using a type for which the `Borrow` impl does not yield well-behaved `Eq`, Bob is violating contract of HashMap.
If, as this PR proposes, we unconditionally require that Eq & friends for borrow should be valid, then the blame shifts to Alice, which I think is more reasonable.
closes https://github.com/rust-lang/rust/issues/44868
Functions with uninhabited return values are already marked `noreturn`,
but we were still generating return instructions for this. When running
with `-C passes=lint`, LLVM prints:
Unusual: Return statement in function with noreturn attribute
The LLVM manual makes a stronger statement about `noreturn` though:
> This produces undefined behavior at runtime if the function ever does
dynamically return.
We now emit an `abort` anywhere that would have tried to return an
uninhabited value.
When using the `rustfix-coverage` flag, some tests currently fail
because they define a different error-format than `json`.
The current implementation crashes when encountering those tests. Since
we don't care about non-json test output when collecting the coverage
data, we handle those tests by returning an empty `Vec` instead.
