Download a more recent LLVM version if `src/version` is modified
When bumping the bootstrap version, the name of the generated LLVM
shared object file is changed, even though it's the same contents as
before. If bootstrap tries to use an older version, it will get linking
errors:
```
Building rustdoc for stage1 (x86_64-unknown-linux-gnu)
Compiling rustdoc-tool v0.0.0 (/home/joshua/rustc/src/tools/rustdoc)
error: linking with `cc` failed: exit code: 1
|
= note: "cc" "-Wl,--as-needed" ... lots of args ...
= note: /usr/bin/ld: cannot find -lLLVM-12-rust-1.53.0-nightly
clang: error: linker command failed with exit code 1 (use -v to see invocation)
error: could not compile `rustdoc-tool`
```
Helps with https://github.com/rust-lang/rust/issues/81930.
Move `std::sys::unix::platform` to `std::sys::unix::ext`
This moves the operating system dependent alias `platform` (`std::os::{linux, android, ...}`) from `std::sys::unix` to `std::sys::unix::ext` (a.k.a. `std::os::unix`), removing the need for compatibility code in `unix_ext` when documenting on another platform.
This is also a step in making it possible to properly move `std::sys::unix::ext` to `std::os::unix`, as ideally `std::sys` should not depend on the rest of `std`.
Fix invalid slice access in String::retain
As noted in #78499, the previous fix was technically still unsound because it accessed elements of a slice outside its bounds (even though they were still inside the same allocation). This PR addresses that concern by switching to a dropguard approach.
Implement TrustedLen and TrustedRandomAccess for Range<integer>, array::IntoIter, VecDequeue's iterators
This should make some `FromIterator` and `.zip()` specializations applicable in a few more cases.
``@rustbot`` label libs-impl
Make NonNull::as_ref (and friends) return refs with unbound lifetimes
# Rationale:
1. The documentation for all of these functions claims that this is what the functions already do, as they all come with this comment:
> You must enforce Rust's aliasing rules, *since the returned lifetime 'a is arbitrarily chosen* and does not necessarily reflect the actual lifetime of the data...
So I think it's just a bug that they weren't this way already. Note that had it not been for this part, I wouldn't be making this PR, so if we decide we won't take this change, I'll follow it up with a docs PR to fix this.
2. This is how the equivalent raw pointer functions behave.
They also take `self` and not `&self`/`&mut self`, but that can't be changed compatibly at this point. This is the next best thing.
3. Without this fix, often code that uses these methods will find it has to expand the lifetime of the result.
(I can't speak for others but even in unsafe-heavy code, needing to do this unexpectedly is a huge red flag -- if Rust thinks something should have a specific lifetime, I assume it's for a reason)
### Can this cause existing code to be unsound?
I'm confident this can't cause new unsoundness since the reference exists for at most its lifetime, but you get a borrow checker error if you do something that would require/allow the reference to exist past its lifetime.
Additionally, the aliasing rules of a reference only applies while the reference exists.
This *must* be the case, as it is required by the rules used by safe code. (That said, the documentation in this file sort of contradicts it, but I think it's just ambiguity between the lifetime `'a` in `&'a T` and lifetime of the `&'a T` reference itself...)
We are increasing the lifetime of these references, but they should already have hard bounds on that lifetime, or they'd have borrow checker errors.
(CC ``@RalfJung`` because I have gone and done the mistake where I say something definitive about aliasing in Rust which is honestly outside the group of things I should make definitive comments about).
# Caveats
1. This is insta-stable (except for on the unstable functions ofc). I don't think there's any other alternative.
2. I don't believe this is a breaking change in practice. In theory someone could be assigning `NonNull::as_ref` to a function pointer of type `fn(&NonNull<T>) -> &T`. Now they'd need to use a slightly different function pointer type which is (probably) incompatible. This seems pathological, but I guess crater could be used if there are concerns.
3. This has no tests. The old version didn't either that I saw. I could add some stuff that fails to compile without it, if that would be useful.
4. Sometimes the NLL borrow checker gives up and decides lifetimes live till the end of the scope, as opposed to the range where they're used. If this change can cause this to happen more, then my soundness rationale is wrong, and it's likely breaking.
In practice this seems super unlikely.
Anyway. That was a lot of typing.
Fixes https://github.com/rust-lang/rust/issues/80183
Use TrustedRandomAccess for in-place iterators where possible
This can speed up in-place iterators containing simple casts and transmutes from `Copy` types to any type of same size. `!Copy` types can't be optimized since `TrustedRandomAccess` isn't implemented for those iterators.
```
name on.b ns/iter o1.b ns/iter diff ns/iter diff % speedup
vec::bench_transmute 20 (40000 MB/s) 12 (66666 MB/s) -8 -40.00% x 1.67
```
Enable mutable noalias for LLVM >= 12
Enable mutable noalias by default on LLVM 12, as previously known miscompiles have been resolved. Now it's time to find the next one ;)
* The `-Z mutable-noalias` option no longer has an explicit default and accepts `-Z mutable-noalias=yes` and `-Z mutable-noalias=no` to override the LLVM version based default behavior.
* The decision on whether to apply the noalias attribute is moved into rustc_codegen_llvm. rustc_middle only provides us with the necessary information to make the decision.
* `noalias` is not emitted for types that are `!Unpin`, as a heuristic for self-referential structures (see #54878 and #63818).
This saves us both the Freeze/Unpin queries, and avoids placing
noalias attributes, which have a compile-time impact on LLVM
even in optnone builds (due to always_inline functions).
This allows the optimizer to turn certain iterator pipelines such as
```rust
let vec = vec![0usize; 100];
vec.into_iter().map(|e| e as isize).collect::<Vec<_>>()
```
into a noop.
The optimization only applies when iterator sources are `T: Copy`
since `impl TrustedRandomAccess for IntoIter<T>`.
No such requirement applies to the output type (`Iterator::Item`).
Such structures may contain self-references, in which case the
same location may be accessible through a pointer that is not
based-on the noalias pointer.
This is still grey area as far as language semantics are concerned,
but checking for !Unpin as an indicator for self-referential
sturctures seems like a good approach for the meantime.
The frontend shouldn't be deciding whether or not to use mutable
noalias attributes, as this is a pure LLVM concern. Only provide
the necessary information and do the actual decision in
codegen_llvm.
When bumping the bootstrap version, the name of the generated LLVM
shared object file is changed, even though it's the same contents as
before. If bootstrap tries to use an older version, it will get linking
errors:
```
Building rustdoc for stage1 (x86_64-unknown-linux-gnu)
Compiling rustdoc-tool v0.0.0 (/home/joshua/rustc/src/tools/rustdoc)
error: linking with `cc` failed: exit code: 1
|
= note: "cc" "-Wl,--as-needed" ... lots of args ...
= note: /usr/bin/ld: cannot find -lLLVM-12-rust-1.53.0-nightly
clang: error: linker command failed with exit code 1 (use -v to see invocation)
error: could not compile `rustdoc-tool`
```
Fix suggestion with generics for `field_reassign_with_default` lint
Fix bad suggestion where `::` is missing after type if generics are involved
Fixes#6944
changelog: none
Perform 'deep recollection' in test helper macros
Currently, the print helper macro performs 'recollection' by doing
`token_stream.into_iter().collect()`. However, this will not affect
nonterminals that occur nested inside delimited groups, since the
wrapping delimited group will be left untouched.
This commit adds 'deep recollection', which recursively recollects every
delimited group in the token stream. As with normal recollection, we
only print out something if deep recollection results in a different
stringified token stream.
This is useful for catching bugs where we update the AST of a
nonterminal (which affects pretty-printing), but do not update the
attatched `TokenStream`
Currently, the print helper macro performs 'recollection' by doing
`token_stream.into_iter().collect()`. However, this will not affect
nonterminals that occur nested inside delimited groups, since the
wrapping delimited group will be left untouched.
This commit adds 'deep recollection', which recursively recollects every
delimited group in the token stream. As with normal recollection, we
only print out something if deep recollection results in a different
stringified token stream.
This is useful for catching bugs where we update the AST of a
nonterminal (which affects pretty-printing), but do not update the
attatched `TokenStream`
Extend `proc_macro_back_compat` lint to `js-sys`
With this PR, we now lint for all cases where we perform some kind of
proc-macro back-compat hack.
The `js-sys` had an internal fix made to properly handle
`None`-delimited groups, so we need to manually check the version in the
filename. As a result, we no longer apply the back-compat hack to cases
where the version number is missing file the file path. This should not
affect any users of the `crates.io` crate.
