Build the first LLVM without LTO in try builds
Currently, we perform three LLVM builds in the Linux x64 dist builder, which is used for `try` builds:
1) "Normal" LLVM - takes ~5s to compile thanks to `sccache`, but ~8 minutes to link because of ThinLTO
2) PGO instrumented LLVM - same timings as 1)
3) PGO optimized LLVM - takes about 20 minutes to build
When I tried to disable LTO for build 1), it suddenly takes only about a minute to build, because the linking step is much faster. The first LLVM doesn't really need LTO all that much. Without it, it will be a bit slower to build `rustc` in two subsequent steps, but it seems that the ~7 minutes saved on linking it do win that back.
Btw, we can't use the host LLVM for build 1), because this LLVM then builds `rustc` in PGO instrumented mode, and we need the same compiler when later PGO optimizing `rustc`. And we want to use our in-house LLVM for that I think.
[rustdoc] If re-export is private, get the next item until a public one is found or expose the private item directly
Fixes#81141.
If we have:
```rust
use Private as Something;
pub fn foo() -> Something {}
```
Then `Something` will be replaced by `Private`.
r? `@notriddle`
Turns out opaque types can have hidden types registered during mir validation
See the newly added test's documentation for an explanation.
fixes#114121
Restore region uniquification in the new solver 🎉
All of the bugs that were "due" to uniquification have been settled via other means (e.g. bidirectional alias-relate, param-env incompleteness, etc).
Firstly, revert the functional changes in #110180. 😸
Secondly, we need to ignore regions when considering if a goal has changed (the "has_changed" boolean returned from `evaluate_goal`) -- otherwise, because we're doing region uniquification, we may perpetually consider a goal to be changed. See the UI test I committed for an explanation.
docs: fmt::Debug*: Fix comments for finish method.
In the code sample for the `finish` method on `DebugList`, `DebugMap`, and `DebugSet`, refer to finishing the list, map, or set, rather than struct as it did.
rustdoc: fix cross-crate `impl Sized` & `impl ?Sized`
Previously, cross-crate impl-Trait (APIT, RPIT, etc.) that only consists of a single `Sized` bound (modulo outlives-bounds) and ones that are `?Sized` were incorrectly rendered. To give you a taste (before vs. after):
```diff
- fn sized(x: impl ) -> impl
+ fn sized(x: impl Sized) -> impl Sized
- fn sized_outlives<'a>(x: impl 'a) -> impl 'a
+ fn sized_outlives<'a>(x: impl Sized + 'a) -> impl Sized + 'a
- fn maybe_sized(x: &impl ) -> &impl
+ fn maybe_sized(x: &impl ?Sized) -> &impl ?Sized
- fn debug_maybe_sized(x: &impl Debug) -> &impl ?Sized + Debug
+ fn debug_maybe_sized(x: &(impl Debug + ?Sized)) -> &(impl Debug + ?Sized)
```
Moreover, we now surround impl-Trait that has multiple bounds with parentheses if they're the pointee of a reference or raw pointer type. This affects both local and cross-crate docs. The current output isn't correct (rustc would emit the error *ambiguous `+` in a type* if we fed the rendered code back to it).
---
Best reviewed commit by commit :)
`@rustbot` label A-cross-crate-reexports
Implement diagnostic translation for rustc-errors
This is my first PR to rustc yeah~
I'm going to implement diagnostic translation on rustc-errors crate.
This PR is WIP, the reason of opening this as draft, I want to show my code to prevent the issue caused by misunderstanding and also I have few questions.
Some error messages are processed by `pluralize!` macro which determines to use plural word or not. From now, I make two kinds of keys and combine with enum but I'm not sure is this best method to do it.
Is there any prefered method to do this? => This resolved on conversation on PR.
I'll remain to perform force-push until my first implementation looks good to me
proc-macros are processed early in the compiler pipeline. There is no
need to involve codegen. So change to check-pass.
I have also looked through each changed test and to me it is
sufficiently clear that codegen is not needed for the purpose of the
test.
I skipped changing tests/ui/proc-macro/no-missing-docs.rs in this commit
because it was not clear to me that it can be changed to check-pass.
The purpose of the test is to make sure that compiling hello world
produces no compiler output. To properly test that, we need to run the
entire compiler pipeline. We don't want the test to pass if codegen
accidentally starts writing to stdout. So keep it as build-pass.
Don't treat negative trait predicates as always knowable
We don't need this. It was added in #90104 but I don't really know why. It's not sound afaict -- negative trait predicates need the same coherence-ambiguity/orphan check rules as positive ones.
r? `@lcnr`
cc `@spastorino,` do you remember why?
Tweak CGU sorting in a couple of places.
In `base.rs`, tweak how the CGU size interleaving works. Since #113777, it's much more common to have multiple CGUs with identical sizes. With the existing code these same-sized items ended up in the opposite-to-desired order due to the stable sorting. The code now starts with a reverse sort (like is done in `partitioning.rs`) which gives the behaviour we want. This doesn't matter much for perf, but makes profiles in `samply` look more like what we expect.
In `partitioning.rs`, we can use `sort_by_key` instead of `sort_by_cached_key` because `CGU::size_estimate()` is cheap. (There is an identical CGU sort earlier in that function that already uses `sort_by_key`.)
r? `@pnkfelix`
This is surprising, but the new comment explains why. It's a logical
conclusion in the drive to avoid `TokenTree` clones.
`TokenTreeCursor` is now only used within `Parser`. It's still needed
due to `replace_prev_and_rewind`.