This PR allows using methods from traits that are visible but are defined in an inaccessible module (fixes#18241). For example,
```rust
mod foo {
pub use foo::bar::Tr;
mod bar { // This module is inaccessible from `g`
pub trait Tr { fn f(&self) {} }
}
}
fn g<T: foo::Tr>(t: T) {
t.f(); // Currently, this is a privacy error even though `foo::Tr` is visible
}
```
After this PR, it will continue to be a privacy error to use a method from a trait that is not visible. This can happen when a public trait inherits from a private trait (in violation of the `public_in_private` lint) -- see @petrochenkov's example in #28504.
r? @nikomatsakis
Now that 767d85061a is upstream, the relevant deadlocking issue which prompted our downgrade has been resolved. As a result, there's no known issue to *not* upgrade! This also re-enables jemalloc for the pc-windows-gnu target as known issues with that have also been fixed.
Closes#31030
This PR improves the import resolution algorithm.
First, it records that an import succeeded or failed for one namespace (by calling `decrement_outstanding_references_for` and `try_define_child` if successful) even if it is still indeterminate in the other namespace, fixing #31444.
Second, it starts importing bindings from globs as soon as the glob path is determined.
It maintains links from imported modules to their importers so that when a resolution becomes successful in an imported module, a corresponding binding will be added to the importer module.
It also maintains links from importer modules to imported modules so that we can determine if an undefined name is indeterminate or failing by recursively checking this in the imported modules.
This allows, for example:
```rust
mod foo {
pub mod baz {}
pub use bar::baz::*;
}
mod bar {
pub use foo::*;
}
```
It also allows cycles of pub glob imports, although by to the current shadowing rules, the only way for such a cycle to compile is if each participating module defines no names. Incidentally, this PR lays the groundwork for more permissive feature-gated shadowing rules.
Finally, this PR encapsulates almost all implementation details of import resolution in `resolve_imports` (some of which used to be in `lib.rs`) and refactors reexport recording, shadowed trait collecting, some duplicate checking, and the `private_in_public` lint out of the core import resolution algorithm and into a post-processing pass in `resolve_imports`.
r? @nrc
The string may be arbitrarily long, but we want to limit the panic
message to a reasonable length. Truncate the string if it is too long
(simply to char boundary).
Also add details to the start <= end message. I think it's ok to flesh
out the code here, since it's in a cold function.
The file it generates had been modified, but instead the generator
should have been modified, and the file regenerated. This merges the
modifications into the template in the generator.
This defines the following intrinsics for 128 and 256 bit vectors of f32
and f64:
* `fmadd`
* `fmaddsub`
* `fmsub`
* `fmsubadd`
* `fnmadd`
* `fnmsub`
The `_sd` and `_ss` variants are not included yet.
Intel intrinsic reference: https://software.intel.com/en-us/node/523929
The intrinsics there are listed under AVX2, but in the Intel Intrinsic
Guide they are part of the "FMA" technology, and LLVM puts them under
FMA, not AVX2.
Adding -Wno-error is more reliable and simple than trying to modify existing
flags. We've been using this in Debian already for the past few releases.
Making this change also encourages future maintainers towards "best practises".
Also take the opportunity to use the same method at all places in the file.
If `v` were a type with some kind of indirection, so that `v += 1` would
have an effect even if `v` were not used anymore, the unused_assignments lint
would mark a false positive.
This exempts overloaded (non-primitive) assign ops from being treated as
assignments (they are method calls).
The previous compile-fail tests that ensure x += 1 can trigger for
primitive types continue to pass. Added a representative test for the
"view" indirection.
Right now there's just a smattering of `// ignore-foo` platforms which is ever
expanding as new ones are added. Instead switch to only running these tests on
Linux/OSX and then use a guaranteed-to-work but not-as-well-tested alternative
on other platforms.
