Allow LintPasses to run after trans

[kmcallister]

We currently have one lint VARIANT_SIZE_DIFFERENCE that's hard-wired into trans. We should generalize this and make it available to lint plugins.

[sfackler]

cc #14406

[steveklabnik]

Triage: looks like no change.

[steveklabnik]

Triage: still no change

[hanna-kruppe]

Some relevant changes since the OP:

• VARIANT_SIZE_DIFFERENCES isn't hardwired into trans any more as of Move variant_size_differences out of trans #34755
• lint plugins are on the way out, the Right Way is a custom driver now
• we don't even keep HIR around at codegen time any more, let alone walk it

There might still be use cases for emitting lint warnings from monomorphized MIR (if we can find a way to accurately preserve the allow(..), deny(..), etc. annotations in metadata) but AFAIK that's hypothetical and I believe custom drivers can already achieve it by hooking into the relevant queries.

[oli-obk]

monomorphized MIR

That's not a thing. We don't monomorphize MIR

I think we should just close this issue. It's very problematic to only report lints in case the trans backend decides on doing something, especially once we get to MIR-only rlibs

[hanna-kruppe]

That's not a thing. We don't monomorphize MIR

Sorry, I should have been more clear, I meant the (generic MIR, substs) pair that codegen item collection produces.

I think we should just close this issue. It's very problematic to only report lints in case the trans backend decides on doing something, especially once we get to MIR-only rlibs

That's always been the case, but yeah it's very sensitive to improvements to the collector or even MIR optimizations. I think it might be OK for some desirable warnings that are impossible to emit pre-monomorphization but YMMV and certainly requires more motivation than is given here.

[hanna-kruppe]

To be clear, I agree that we should close. Although I think the idea of lints emitted from codegen might have some merit,

1. having one in-tree would be quite controversial and thus should go through the RFC process
2. writing one out of tree is already possible w/ custom drivers
3. such lints are entirely hypothetical at this point

So let's close this.