Flag `-Z no-link` was previously introduced, which allows creating
an `.rlink` file to perform compilation without linking.
This change enables linking from an `.rlink` file.
Lockless LintStore
This removes mutability from the lint store after registration. Each commit stands alone, for the most part, though they don't make sense out of sequence.
The intent here is to move LintStore to a more parallel-friendly architecture, although also just a cleaner one from an implementation perspective. Specifically, this has the following changes:
* We no longer implicitly register lints when registering lint passes
* For the most part this means that registration calls now likely want to call something like:
`lint_store.register_lints(&Pass::get_lints())` as well as `register_*_pass`.
* In theory this is a simplification as it's much easier for folks to just register lints and then have passes that implement whichever lint however they want, rather than necessarily tying passes to lints.
* Lint passes still have a list of associated lints, but a followup PR could plausibly change that
* This list must be known for a given pass type, not instance, i.e., `fn get_lints()` is the signature instead of `fn get_lints(&self)` as before.
* We do not store pass objects, instead storing constructor functions. This means we always get new passes when running lints (this happens approximately once though for a given compiler session, so no behavior change is expected).
* Registration API is _much_ simpler: generally all functions are just taking `Fn() -> PassObject` rather than several different `bool`s.
Move to using Box<dyn Fn() -> ...> so that we can let plugins register
state.
This also adds a callback that'll get called from plugin registration so
that Clippy and other tools can register lints without using the plugin
API. The plugin API still works, but this new API is more compatible
with drivers other than rustc.
This changes the default parallelism for parallel compilers to one,
instead of the previous default, which was "num cpus". This is likely
not an optimal default long-term, but it is a good default for testing
whether parallel compilers are not a significant regression over a
sequential compiler.
Notably, this in theory makes a parallel-enabled compiler behave
exactly like a sequential compiler with respect to the jobserver.
