With this commit, if one bootstraps rust against system llvm then the
src/llvm submodule is not updated/checked-out. This saves considerable
network bandwith when starting from a fresh clone of rust-lang/rust as
the llvm submodule is never cloned.
cc #30107
memrchr: Correct aligned offset computation
The memrchr fallback did not compute the offset correctly. It was
intentioned to land on usize-aligned addresses but did not.
This was suspected to have resulted in a crash on ARMv7!
This bug affected non-linux platforms.
I think like this, if we have a slice with pointer `ptr` and length
`len`, we want to find the last usize-aligned offset in the slice.
The correct computation should be:
For example if ptr = 1 and len = 6, and `size_of::<usize>()` is 4:
```
[ x x x x x x ]
1 2 3 4 5 6
^-- last aligned address at offset 3 from the start.
```
The last aligned address is ptr + len - (ptr + len) % usize_size.
Compute offset from the start as:
offset = len - (ptr + len) % usize_size = 6 - (1 + 6) % 4 = 6 - 3 = 3.
I believe the function's return value was always correct previously, if
the platform supported unaligned addresses.
Fixes#35967
rustc_borrowck: Don't hash types in loan paths
1) Types for equal loan paths are not always equal, they can sometimes differ in lifetimes, making equal loan paths hash differently.
Example:
71bdeea561/src/libcollections/linked_list.rs (L835-L856)
One of `self.list`s has type
```
&ReFree(CodeExtent(15013/CallSiteScope { fn_id: 18907, body_id: 18912 }), BrNamed(0:DefIndex(3066), 'a(397), WontChange)) mut linked_list::LinkedList<T>
```
and other has type
```
&ReScope(CodeExtent(15018/Remainder(BlockRemainder { block: 18912, first_statement_index: 0 }))) mut linked_list::LinkedList<T>
```
(... but I'm not sure it's not a bug actually.)
2) Not hashing types is faster than hashing types.
r? @arielb1
Combine types and regions in Substs into one interleaved list.
Previously, `Substs` would contain types and regions, in two separate vectors, for example:
```rust
<X as Trait<'a, 'b, A, B>>::method::<'p, 'q, T, U>
/* corresponds to */
Substs { regions: ['a, 'b, 'p, 'q], types: [X, A, B, T, U] }
```
This PR continues the work started in #35605 by further removing the distinction.
A new abstraction over types and regions is introduced in the compiler, `Kind`.
Each `Kind` is a pointer (`&TyS` or `&Region`), with the lowest two bits used as a tag.
Two bits were used instead of just one (type = `0`, region = `1`) to allow adding more kinds.
`Substs` contain only a `Vec<Kind>`, with `Self` first, followed by regions and types (in the definition order):
```rust
Substs { params: [X, 'a, 'b, A, B, 'p, 'q, T, U] }
```
The resulting interleaved list has the property of being the concatenation of parameters for the (potentially) nested generic items it describes, and can be sliced back into those components:
```rust
params[0..5] = [X, 'a, 'b, A, B] // <X as Trait<'a, 'b, A, B>>
params[5..9] = ['p, 'q, T, U] // <_>::method::<'p, 'q, T, U>
```
r? @nikomatsakis
Improve error message when failing to parse a block
We want to catch this error:
```
if (foo)
bar;
```
as it's valid syntax in other languages, and say how to fix it.
Unfortunately it didn't care if the suggestion made sense and just
highlighted the unexpected token.
Now it attempts to parse a statement, and if it succeeds, it shows the
help message.
Fixes#35907
Remove style guide.
We originally imported this into the repository with the intent of
fixing it up. Instead, nothing happened.
Its appearance on rust-lang.org makes it seem semi-official, but it's
not. The rustfmt strike team will end up producing something like this
anyway, and leaving it around does nothing but mislead people.
r? @aturon
[MIR] track Location in MirVisitor, combine Location
All the users of MirVisitor::visit_statement implement their own statement index tracking. This PR move the tracking into MirVisitor itself.
Also, there were 2 separate implementations of Location that were identical. This PR eliminates one of them.
Do not emit "class method" debuginfo for types that are not DICompositeType.
Fixes#35991 by restricting the "class method" debuginfo sugar from #33358 to structs and enums only.
r? @michaelwoerister
Buffer unix and lock windows to prevent message interleaving
When cargo does a build on multiple processes, multiple crates may error at the same time. If this happens, currently you'll see interleaving of the error messages, which makes for an unreadable message.
Example:
```
--> --> src/bin/multithread-unix.rs:16:35src/bin/singlethread.rs:16:24
||
1616 | | Server::new(&addr).workers(8). Server::new(&addr).serveserve(|r: Request| {(|r: Request| {
| | ^^^^^^^^^^ expected struct `std::io::Error`, found () expected struct `std::io::Error`, found ()
||
= = notenote: expected type `std::io::Error`: expected type `std::io::Error`
= = notenote: found type `()`: found type `()`
= = notenote: required because of the requirements on the impl of `futures_minihttp::Service<futures_minihttp::Request, futures_minihttp::Response>` for `[closure@src/bin/multithread-unix.rs:16:41: 22:6]`: required because of the requirements on the impl of `futures_minihttp::Service<futures_minihttp::Request, futures_minihttp::Response>` for `[closure@src/bin/singlethread.rs:16:30: 22:6]`
error: aborting due to previous error
error: aborting due to previous error
```
This patch uses two techniques to prevent this interleaving. On Unix systems, since they use the text-based ANSI protocol for coloring, we can safely buffer up whole messages before we emit. This PR does this buffering, and emits the whole message at once.
On Windows, term must use the Windows terminal API to color the output. This disallows us from using the same buffering technique. Instead, here we grab a Windows mutex (thanks to @alexcrichton for the lock code). This lock only works in Windows and will hold a mutex for the duration of a message output.
r? @nikomatsakis
modify fds-are-cloexec test to open a file that exists
Fixes#35753.
Is it a valid assumption that the current directory is always the root of the repo when the tests are run?
r? @nagisa
We originally imported this into the repository with the intent of
fixing it up. Instead, nothing happened.
Its appearance on rust-lang.org makes it seem semi-official, but it's
not. The rustfmt strike team will end up producing something like this
anyway, and leaving it around does nothing but mislead people.
Deny (by default) transmuting from fn item types to pointer-sized types.
This sets the #19925 lint (transmute from zero-sized fn item type) to `deny` by default.
Technically a `[breaking-change]`, but will not affect dependent crates because of `--cap-lints`.
Use arc4rand(9) on FreeBSD
From rust-lang-nursery/rand#112:
>After reading through #30691 it seems that there's general agreement that using OS-provided facilities for seeding rust userland processes is fine as long as it doesn't use too much from libc. FreeBSD's `arc4random_buf(3)` is not only a whole lot of libc code, but also not even currently exposed in the libc crate. Fortunately, the mechanism `arc4random_buf(3)` et al. use for getting entropy from the kernel ([`arc4rand(9)`](https://www.freebsd.org/cgi/man.cgi?query=arc4random&apropos=0&sektion=9&manpath=FreeBSD+10.3-RELEASE&arch=default&format=html)) is exposed via `sysctl(3)` with constants that are already in the libc crate.
>I haven't found too much documentation on `KERN_ARND`—it's missing or only briefly described in most of the places that cover sysctl mibs. But, from digging through the kernel source, it appears that the sysctl used in this PR is very close to just calling `arc4rand(9)` directly (with `reseed` set to 0 and no way to change it).
I expected [rand](/rust-lang-nursery/rand) to reply quicker, so I tried submitting it there first. It's been a few weeks with no comment, so I don't know the state of it, but maybe someone will see it here and have an opinion. This is basically the same patch. It pains me to duplicate the code but I guess it hasn't been factored out into just one place yet.
rustc_trans: do not generate allocas for unused locals.
This fixes a regression observed in a [`mio` test](https://travis-ci.org/carllerche/mio/jobs/152142886) which was referencing a 4MB `const` array.
Even though MIR rvalue promotion would promote the borrow of the array, a dead temp was left behind.
As the array doesn't have an immediate type, an `alloca` was generated for it, even though it had no uses.
The fix is pretty dumb: assume that locals need to be borrowed or assigned before being used.
And if it can't be used, it doesn't get an `alloca`, even if the type would otherwise demand it.
This could change in the future, but all the MIR we generate now doesn't break that rule.
Fix debug line number info for macro expansions.
Macro expansions result in code tagged with completely different debug locations than the surrounding expressions. This wrecks havoc on debugger's ability the step over source lines.
This change fixes the problem by tagging expanded code as "inlined" at the macro expansion site, which allows the debugger to sort it out.
Note that only the outermost expansion is currently handled, stepping into a macro will still result in stepping craziness.
r? @eddyb
Macro expansions produce code tagged with debug locations that are completely different from the surrounding expressions. This wrecks havoc on debugger's ability the step over source lines.
In order to have a good line stepping behavior in debugger, we overwrite debug locations of macro expansions with that of the outermost expansion site.
rustc: Don't enable NEON by default on armv7 Linux
One of the primary platforms for the `armv7-unknown-linux-gnueabihf` target,
Linux distributions, do not enable NEON extensions by default. This PR disables
that feature by defualt but enables the `d16` feature which enables VFP3D16 that
distributions do enable.
Closes#35590
