When there are multiple implementors for the same trait that is present
in an unmet binding, modify the E0277 error to refer to the parent
obligation and verify whether borrowing the argument being passed in
would satisfy the unmet bound. If it would, suggest it.
The previous ordering of the sentences kept switching between the return
value and the value of `self` after execution, making it hard to follow.
Additionally, as rendered in the browser, the period in "`Self`. `self`"
was difficult to make out as being a sentence separator and not one code
block.
Fully integrate derive helpers into name resolution
```rust
#[derive(Foo)]
#[foo_helper] // already goes through name resolution
struct S {
#[foo_helper] // goes through name resolution after this PR
field: u8
}
```
How name resolution normally works:
- We have an identifier we need to resolve, take its location (in some sense) and look what names are in scope in that location.
How derive helper attributes are "resolved" (before this PR):
- After resolving the derive `Foo` we visit the derive's input (`struct S { ... } `) as a piece of AST and mark attributes textually matching one of the derive's helper attributes (`foo_helper`) as "known", so they never go through name resolution.
This PR changes the rules for derive helpers, so they are not proactively marked as known (which is a big hack ignoring any ambiguities or hygiene), but go through regular name resolution instead.
This change was previously blocked by attributes not being resolved in some positions at all (fixed in https://github.com/rust-lang/rust/pull/63468).
"Where exactly are derive helpers in scope?" is an interesting question, and I need some feedback from proc macro authors to answer it, see the post below (https://github.com/rust-lang/rust/pull/64694#issuecomment-533925160).
It is used in the case where a variable-length slice pattern is used to
match on an array of known size. This allows considering only those
entries in the array that are captured by one of the patterns.
As a side-effect, diagnostics improve a bit for those cases.
These helpers are resolved before their respective derives through a kind of look ahead into future expansions.
Some of these will migrate to proper resolution, others will be deprecated.
```
#[trait_helper] // Deprecate
#[derive(Trait)]
#[trait_helper] // Migrate to proper resolution
```
Update cc, git2, num_cpus.
This updates the `cc` crate, bringing in better parallel building support. Also updates `git2` which enables the parallel feature. (Note: I don't expect it will have a significant impact on build time, but seems good to update anyways.)
The main thorn is that `cc` gained knowledge about RISC-V architectures (https://github.com/alexcrichton/cc-rs/pull/428, https://github.com/alexcrichton/cc-rs/pull/429, https://github.com/alexcrichton/cc-rs/pull/430), but the builders on CI do not have the riscv C compiler installed. This means that bootstraps' cc detection was finding a C compiler that isn't installed, and fails.
The solution here is to override the cc detection to `false`. The C compiler isn't actually used on riscv platforms. AFAIK, the only location would be compiler_builtins, and it currently forces C support off (a533ae9c5a/build.rs (L49-L55)).
Other possible solutions:
- Add the override in cc_detect for riscv (or any "no-C" platform like wasm32 and nvptx)
- Install and use the appropriate c compiler. I tried this the `g++-riscv64-linux-gnu` package, but it failed missing some header file.
Closes#66232
Rollup of 5 pull requests
Successful merges:
- #66350 (protect creation of destructors by a mutex)
- #66407 (Add more tests for fixed ICEs)
- #66415 (Add --force-run-in-process unstable option to libtest)
- #66427 (Move the JSON error emitter to librustc_errors)
- #66441 (libpanic_unwind for Miri: make sure we have the SEH lang items when needed)
Failed merges:
r? @ghost
Refactor integer range handling in the usefulness algorithm
Integer range handling had accumulated a lot of debt. This cleans up a lot of it.
In particular this:
- removes unnecessary conversions between `Const` and `u128`, and between `Constructor` and `IntRange`
- clearly distinguishes between on the one hand ranges of integers that may or may not be matched exhaustively, and on the other hand ranges of non-integers that are never matched exhaustively and are compared using Const-based shenanigans
- cleans up some overly complicated code paths
- generally tries to be more idiomatic.
As a nice side-effect, I measured a 10% perf increase on `unicode_normalization`.
There's one thing that I feel remains to clean up: the [overlapping range check](https://github.com/rust-lang/rust/pull/64007), which is currently quite ad-hoc. But that is intricate enough that I'm leaving it out of this PR.
There's also one little thing I'm not sure I understand: can `try_eval_bits` fail for an integer constant value in that code ? What would that mean, and how do I construct a test case for this possibility ?
