ICE: `interpret const eval failure of Unevaluated(UnevaluatedConst ... which is not in required_consts` about rust

[RalfJung]

Caused by running the interpreter on code that doesn't even pass name resolution. Not sure where to fix this, but probably not in the interpreter.^^

Cc @oli-obk

[matthiaskrgr]

#121557 cc @RalfJung

[oli-obk]

this is a promoted failing to compile because it gets the layout of this broken type, which gives us back an ErrorGuaranteed. I'm certain we can prevent const eval from running by checking such layouts earlier in the MIR pipeline, but the fix isn't clear to me yet.

[RalfJung]

The issue is that in the sanity check here we only know it's an ErrorHandled::Reported, we don't know whether it was a typeck error on a const-eval error. So we can't really distinguish these cases.

If this is a LayoutError::ReferencesError, we could probably fix it by tweaking the error handling here. But there probably are similar cases where a promoted fails to evaluate due to a SizeOverflow and then we have to report the error somewhere.

The only option I see right now is to add another variant to ErrorHandled (or a flag in ErrorHandled::Reported) to track whether this was a failure inside the constant itself or a failure in something the constant referenced. But this distinction isn't always very clear anyway... it would basically have to be "is this an error that can occur even for supposedly infallible operations like &[]", i.e., match exactly the usecase of the const-eval / promoted sanity check. Which is not great.

[oli-obk]

We already got the tainted_by_errors flag that we can reuse for that