Make emitting an error necessary to acquire a `ErrorReported` token

[varkor]

ErrorReported indicates that a compiler error has been emitted, but currently it may be freely constructed. We should make it unconstructable, so that the only way to get one is to .emit() an error.

cc #68434 (comment) @eddyb @estebank

[mark-i-m]

We could have it set a thread_local! that panics if not defused by some diagnostic emission on the same thread or something, but it might not be worth it if calls can be removed.

Regarding the unchecked constructor of ErrorReported in #93222...

[eddyb]

We could have it set a thread_local! that panics if not defused by some diagnostic emission on the same thread or something, but it might not be worth it if calls can be removed.

I wrote that but now I believe it's unnecessary. Nothing I saw in that PR even requires the unchecked constructor, short of references_error.

[eddyb]

To make it more accessible, I'm replicating #93222 (comment) here:

quiet is from 5cf300d from #84176.

@GuillaumeGomez do you know if that change can be reverted? It shouldn't have been merged, it makes it look like it's legitimate to disable diagnostics (to be clear, this is very dangerous, especially as more things rely on delay_span_bug and error emission).

There is no way to keep rustdoc doing what that PR does and still e.g. create ty::Error. An error has to be emitted to have ty::Error.

I don't think compromising the integrity of the rest of the system is worth getting access to method call annotations in code, which seems to be the only use of this misfeature.

AFAICT the only way to get the information you want is to run a separate compilation pass, and sinkhole errors at the actual point where they would be printed to stderr, while letting the compiler properly track the fact that errors happened, and then scavenge whatever data is available, despite those errors.

TL;DR: there is now a quiet flag inside the diagnostic machinery that silences everything, including load-bearing delay_span_bugs, and it's really easy to turn on.

Progress on ErrorReported cannot be made until that change is reverted.

[eddyb]

Random bikeshed for ErrorReported's name if we take delay_span_bug into account: Doomed(Compilation).

[Manishearth]

I'm also looking at how it's used and ... that code should absolutely not be re-typechecking source, there should be better ways to get the information it needs. @GuillaumeGomez please talk to the compiler team and figure out how to best get the information you need there.

@eddyb The feature where this is used is unstable and behind a flag in large part due to the rustdoc team being quite wary of such a large change, so y'all should absolutely just rip out quiet in this PR, it's fine if the link-to-def feature is broken.

[GuillaumeGomez]

I opened a discussion on zulip. I'd really prefer to have an alternative (if there is one) before the PR is merged.