Remove `expr_parentheses_needed` from `ParseSess`
Not sure why this method needed to exist on `ParseSess`, but we can achieve the same behavior by just inlining it everywhere.
Fix `format_args` capture for macro expanded format strings
Since #100996 `format_args` capture for macro expanded strings aren't prevented when the span of the expansion points to a string literal, e.g.
```rust
// not a terribly realistic example, but also happens for proc_macros that set
// the span of the output to an input str literal, such as indoc
macro_rules! x {
($e:expr) => { $e }
}
fn main() {
let a = 1;
println!(x!("{a}"));
}
```
The tests didn't catch it as the span of `concat!()` points to the macro invocation
r? `@m-ou-se`
Move lint level source explanation to the bottom
So, uhhhhh
r? `@estebank`
## User-facing change
"note: `#[warn(...)]` on by default" and such are moved to the bottom of the diagnostic:
```diff
- = note: `#[warn(unsupported_calling_conventions)]` on by default
= warning: this was previously accepted by the compiler but is being phased out; it will become a hard error in a future release!
= note: for more information, see issue #87678 <https://github.com/rust-lang/rust/issues/87678>
+ = note: `#[warn(unsupported_calling_conventions)]` on by default
```
Why warning is enabled is the least important thing, so it shouldn't be the first note the user reads, IMO.
## Developer-facing change
`struct_span_lint` and similar methods have a different signature.
Before: `..., impl for<'a> FnOnce(LintDiagnosticBuilder<'a, ()>)`
After: `..., impl Into<DiagnosticMessage>, impl for<'a, 'b> FnOnce(&'b mut DiagnosticBuilder<'a, ()>) -> &'b mut DiagnosticBuilder<'a, ()>`
The reason for this is that `struct_span_lint` needs to edit the diagnostic _after_ `decorate` closure is called. This also makes lint code a little bit nicer in my opinion.
Another option is to use `impl for<'a> FnOnce(LintDiagnosticBuilder<'a, ()>) -> DiagnosticBuilder<'a, ()>` altough I don't _really_ see reasons to do `let lint = lint.build(message)` everywhere.
## Subtle problem
By moving the message outside of the closure (that may not be called if the lint is disabled) `format!(...)` is executed earlier, possibly formatting `Ty` which may call a query that trims paths that crashes the compiler if there were no warnings...
I don't think it's that big of a deal, considering that we move from `format!(...)` to `fluent` (which is lazy by-default) anyway, however this required adding a workaround which is unfortunate.
## P.S.
I'm sorry, I do not how to make this PR smaller/easier to review. Changes to the lint API affect SO MUCH 😢
Until now out-of-range integers in format string literals
were silently ignored. They wrapped around to zero at
usize::MAX, producing unexpected results.
When using debug builds of rustc, such integers in format string
literals even cause an 'attempt to add with overflow' panic in
rustc.
Fix this by producing an error diagnostic for integers in format
string literals which do not fit into usize.
Fixes#102528
Group together more size assertions.
Also add a few more assertions for some relevant token-related types.
And fix an erroneous comment in `rustc_errors`.
r? `@lqd`
Flush delayed bugs before codegen
Sometimes it can happen that invalid code like a TyKind::Error makes its way through the compiler without triggering any errors (this is always a bug in rustc but bugs do happen sometimes :)). These ICEs will manifest in the backend like as cg_llvm not being able to get the layout of `[type error]`, which makes it hard to debug. By flushing before codegen, we display all the delayed bugs, making it easier to trace it to the root of the problem.
I tried this on #102366 and it showed tons of of delayed bugs and no error in cg_llvm, so it seems to be working.
remove the unused :: between trait and type to give user correct diag…
…nostic information
modified: compiler/rustc_trait_selection/src/traits/error_reporting/mod.rs
new file: src/test/ui/type/issue-101866.rs
new file: src/test/ui/type/issue-101866.stderr
Manually order `DefId` on 64-bit big-endian
`DefId` uses different field orders on 64-bit big-endian vs. others, in
order to optimize its `Hash` implementation. However, that also made it
derive different lexical ordering for `PartialOrd` and `Ord`. That
caused spurious differences wherever `DefId`s are sorted, like the
candidate sources list in `report_method_error`.
Now we manually implement `PartialOrd` and `Ord` on 64-bit big-endian to
match the same lexical ordering as other targets, fixing at least one
test, `src/test/ui/methods/method-ambig-two-traits-cross-crate.rs`.
