Previously each opaque type instantiation would create new inference vars, even for the same opaque type/substs combination. Now there is a global map in InferCtxt that gets filled whenever we encounter an opaque type.
Rollup of 9 pull requests
Successful merges:
- #87561 (thread set_name haiku implementation.)
- #87715 (Add long error explanation for E0625)
- #87727 (explicit_generic_args_with_impl_trait: fix min expected number of generics)
- #87742 (Validate FFI-safety warnings on naked functions)
- #87756 (Add back -Zno-profiler-runtime)
- #87759 (Re-use std::sealed::Sealed in os/linux/process.)
- #87760 (Promote `aarch64-apple-ios-sim` to Tier 2)
- #87770 (permit drop impls with generic constants in where clauses)
- #87780 (alloc: Use intra doc links for the reserve function)
Failed merges:
r? `@ghost`
`@rustbot` modify labels: rollup
alloc: Use intra doc links for the reserve function
The sentence exists to highlight the existence of a
performance footgun of repeated calls of the
reserve_exact function.
permit drop impls with generic constants in where clauses
Fixes#79248
`==` is not sufficient to check for equality between unevaluated consts which causes the above issue because the const in `[(); N - 1]:` on the impl and the const in `[(); N - 1]:` on the struct def are not seen as equal. Any predicate that can contain an unevaluated const cant use `==` here as it will cause us to incorrectly emit an error.
I dont know much about chalk but it seems like we ought to be relating the `TypeWellFormedFromEnv` instead of `==` as it contains a `Ty` so I added that too...
r? ``````@lcnr``````
Promote `aarch64-apple-ios-sim` to Tier 2
As per https://github.com/rust-lang/compiler-team/issues/428 this target becomes Tier 2.
---
I tested that the latest nightly has the target and I'm already building my project with that on CI and locally.
Re-use std::sealed::Sealed in os/linux/process.
This uses `std::sealed::Sealed` in `std::os::linux::process` instead of defining new `Sealed` traits there.
Add back -Zno-profiler-runtime
This was removed by #85284 in favor of `-Zprofiler-runtime=<name>`.However the suggested `-Zprofiler-runtime=None` doesn't work because`None` is treated as a crate name.
Validate FFI-safety warnings on naked functions
Test that FFI-safety warnings don't get accidentally dropped on naked
functions. The big picture is that if you implement a naked function
with the Rust ABI you'll get a warning. Further, if you implement a
naked function with a standardized ABI, but use non-FFI-safe types you
will still get a warning.
rust-lang/rfcs#2774rust-lang/rfcs#2972
cc ``````@joshtriplett`````` ``````@Amanieu`````` ``````@haraldh``````
explicit_generic_args_with_impl_trait: fix min expected number of generics
Fixes#87718
The problem was that `synth_type_param_count` was already subtracted from `named_type_param_count`, so this ended up being subtracted again. This caused `expected_min` to overflow, and ultimately resulting in weird and wrong behaviour.
I've also added another test not present in the original issue but caused by the same bug.
The indexes into the VaListImpl struct used on aarch64 ABI (not macos/ios) are hard-coded which is brittle so we replace them with the usual lookup.
The varargs ffi is tested in ui/abi/variadic-ffi.rs on aarch64 Linux.
Generate links to definition in rustdoc source code pages
## Description
This PR adds an option (disabled by default) to add links in the source code page on ident. So for for example:
```rust
mod other_module;
struct Foo;
fn bar() {}
fn x<T: other_module::Trait>(f: Foo, g: other_module::Whatever, t: &T) {
let f: Foo = Foo;
bar();
f.some_method();
}
```
In the example (mostly in the `x` function), `other_module::Trait`, `Foo`, `other_module::Whatever`, `bar` and `some_method` are now links (and `other_module` at the top too).
In case there is a type coming from another crate, it'll link to its documentation page and not its definition (but you can then click on `[src]` so I guess it's fine).
Another important detail: I voluntarily didn't add links for primitive types. I think we can discuss about adding links on them or not in a later PR (adding the support for them would require only a few lines).
Here is a video summing up everything I wrote above:
https://user-images.githubusercontent.com/3050060/114622354-21307b00-9cae-11eb-834d-f6d8178a37bd.mp4
## Performance impact
So, on my computer, the performance remains more or less the same (which is quite surprising but that's a nice surprise). Here are the numbers:
Without the option:
* core: 1m 21s
* alloc: 26.78s
* std: 27.30s
* proc_macro: 4.50s
With source to definition links generation (I enabled by default the option):
* core: 1m 25s
* alloc: 25.76s
* std: 27.07s
* proc_macro: 4.66s
So no real change here (again, I'm very surprised by this fact).
For the size of the generated source files (only taking into account the `src` folder here since it's the only one impacted) by running `du -shc .` (when I am in the source folder).
Without the option: 11.939 MB
With the option: 12.611 MB
So not a big change here either. In all those docs, I ran `grep -nR '<a class=' . | wc -l` and got 43917. So there are quite a lot of links added. :)
cc `@rust-lang/rustdoc`
r? `@jyn514`
Rather than relying on `getPointerElementType()` from LLVM function
pointers, we now pass the function type explicitly when building `call`
or `invoke` instructions.
