The new strategy is as follows. First, the `!` type is assigned
in two cases:
- a block with a diverging statement and no tail expression (e.g.,
`{return;}`);
- any expression with the type `!` is considered diverging.
Second, we track when we are in a diverging state, and we permit a value
of any type to be coerced **into** `!` if the expression that produced
it is diverging. This means that `fn foo() -> ! { panic!(); 22 }`
type-checks, even though the block has a type of `usize`.
Finally, coercions **from** the `!` type to any other are always
permitted.
Fixes#39808.
Remove internal liblog
This commit deletes the internal liblog in favor of the implementation that
lives on crates.io. Similarly it's also setting a convention for adding crates
to the compiler. The main restriction right now is that we want compiler
implementation details to be unreachable from normal Rust code (e.g. requires a
feature), and by default everything in the sysroot is reachable via `extern
crate`.
The proposal here is to require that crates pulled in have these lines in their
`src/lib.rs`:
#![cfg_attr(rustbuild, feature(staged_api, rustc_private))]
#![cfg_attr(rustbuild, unstable(feature = "rustc_private", issue = "27812"))]
This'll mean that by default they're not using these attributes but when
compiled as part of the compiler they do a few things:
* Mark themselves as entirely unstable via the `staged_api` feature and the
`#![unstable]` attribute.
* Allow usage of other unstable crates via `feature(rustc_private)` which is
required if the crate relies on any other crates to compile (other than std).
Revert #39485, fixing type-inference regressions
This reverts PR #39485, which should fix the immediate regressions. Eventually I'd like to land https://github.com/rust-lang/rust/pull/40224 -- or some variant of it -- which revisits the question fo dead-code and inference.
r? @eddyb
cc @canndrew
update LLVM with fix for PR32379
Fixes#40593.
The "root" codegen bug fixed here is that, when generating ARM code, unpatched LLVM 3.9/3.9.1 miscompiles bit operations in rare circumstances - this can cause user code compiled via LLVM (through both `rustc` and `clang`) to subtly return incorrect results - for more details, see the test in this PR or in the LLVM rare report.
One effect of that LLVM bug is that `rustc` 1.17 (and possibly other versions) is miscompiled on ARM. The code generated by a miscompiled `rustc` lacks destructor calls in many circumstances.
Users who run an affected/miscompiled `rustc` - 1.17 or above - on an ARM build machine will be affected by the (fairly blatant) missing destructor bug, regardless of the target architecture (this includes the official `1.17.0-beta.1`, `1.17.0-beta.2`, and some official 1.17/1.18 nightlies).
Users who use an affected LLVM (that's any unpatched LLVM 3.9/3.9.1), whether through `rustc` (in any version that supports 3.9 - that's 1.12 or above) or through `clang`, who compile code to an ARM target architecture might be affected by the (fairly hard to hit) bit operation bug, regardless of the build machine.
Distributors and user who want to compile rustc using their own LLVM should apply the [patch](cdc303e5ed) to avoid miscompilations.
r? @alexcrichton
Beta-nominating because regression (rustc 1.16 is not blatantly miscompiled). This also picks a fix for the (MSVC-affecting) PR29151.
In debugging #40546 I was able to reproduce locally finally using
the literal toolchain that the bots were using. I reproduced the error maybe 4
in 10 builds. I also have the 6.3.0 toolchain installed through `pacman` which
has yet to have a failed build.
When attempting to reproduce the bug with the toolchain that this commit
switches to I was unable to reproduce anything after a few builds. I have no
idea what the original problem was, but I'm hoping that it was just some random
bug fixed somewhere along the way.
I don't currently know of a technical reason to stick to the 4.9.2 toolchains we
were previously using. Historcal 5.3.* toolchains would cause llvm to segfault
(maybe a miscompile?) but this seems to have been fixed recently. To me if it
passes CI then I think we're good.
Closes#40546
This commit deletes the internal liblog in favor of the implementation that
lives on crates.io. Similarly it's also setting a convention for adding crates
to the compiler. The main restriction right now is that we want compiler
implementation details to be unreachable from normal Rust code (e.g. requires a
feature), and by default everything in the sysroot is reachable via `extern
crate`.
The proposal here is to require that crates pulled in have these lines in their
`src/lib.rs`:
#![cfg_attr(rustbuild, feature(staged_api, rustc_private))]
#![cfg_attr(rustbuild, unstable(feature = "rustc_private", issue = "27812"))]
This'll mean that by default they're not using these attributes but when
compiled as part of the compiler they do a few things:
* Mark themselves as entirely unstable via the `staged_api` feature and the
`#![unstable]` attribute.
* Allow usage of other unstable crates via `feature(rustc_private)` which is
required if the crate relies on any other crates to compile (other than std).
This reverts commit dc0bb3f283, reversing
changes made to e879aa43ef.
This is a temporary step intended to fix regressions. A more
comprehensive fix for type inference and dead-code is in the works.
Propagate expected type hints through struct literals.
Partial fix for #31260 to maximize backwards-compatibility, i.e. the hint is provided but not coerced to.
The added test works because `{...; x}` with a hint of `T` coerces `x` to `T`, and the reasoning why that is slightly different has to do with DSTs: `&Struct { tail: [x] }: &Struct<[T]>` has a hint of `[T]` for `[x]`, but the inferred type should be `[T; 1]` to succeed later, so `[x]` shouldn't be *forced* to be `[T]`.
*However*, implementing that complete behavior in a backwards-compatible way may be non-trivial, and has not yet been fully investigated, while this PR fixes#40355 and can be backported.
r? @nikomatsakis
Change how the `0` flag works in format!
Now it always implies right-alignment, so that padding zeroes are placed after the sign (if any) and before the digits. In other words, it always takes precedence over explicitly specified `[[fill]align]`. This also affects the '#' flag: zeroes are placed after the prefix (0b, 0o, 0x) and before the digits.
Here's a short summary of how similar format strings work in Python and Rust:
```
:05 :<05 :>05 :^05
Python 3.6 |-0001| |-1000| |000-1| |0-100|
Rust before |-0001| |-1000| |-0001| |-0100|
Rust after |-0001| |-0001| |-0001| |-0001|
:#05x :<#05x :>#05x :^#05x
Python 3.6 |0x001| |0x100| |000x1| |00x10|
Rust before |0x001| |0x100| |000x1| |0x010|
Rust after |0x001| |0x001| |0x001| |0x001|
```
Fixes#39997 [breaking-change]
Add feature gate for rvalue-static-promotion
Probably needs more tests (which ones?) and there may be other things that need to be done. Also not sure whether the version that introduces the flag is really `1.15.1`.
See https://github.com/rust-lang/rfcs/pull/1414.
Updates #38865.
Now it always implies right-alignment, so that padding zeroes are placed after the sign (if any) and before the digits. In other words, it always takes precedence over explicitly specified `[[fill]align]`.
:06 :<06 :>06 :^06
before |-001.2| |-1.200| |-001.2| |-01.20|
after |-001.2| |-001.2| |-001.2| |-001.2|
Now it always implies right-alignment, so that padding zeroes are placed after the sign (if any) and before the digits. In other words, it always takes precedence over explicitly specified `[[fill]align]`. This also affects the '#' flag: zeroes are placed after the prefix (0b, 0o, 0x) and before the digits.
:05 :<05 :>05 :^05
before |-0001| |-1000| |-0001| |-0100|
after |-0001| |-0001| |-0001| |-0001|
:#05x :<#05x :>#05x :^#05x
before |0x001| |0x100| |000x1| |0x010|
after |0x001| |0x001| |0x001| |0x001|
Fixes#39997 [breaking-change]
Disallow subtyping between T and U in T: Unsize<U>.
Because `&mut T` can be coerced to `&mut U`, `T` and `U` must be unified invariantly. Fixes#40288.
E.g. coercing `&mut [&'a X; N]` to `&mut [&'b X]` must require `'a` be equal to `'b`, otherwise you can convert between `&'a X` and `&'b X` (in either direction), potentially unsoundly lengthening lifetimes.
Subtyping here was introduced with `Unsize` in #24619 (landed in 1.1, original PR is #23785).
