-Zvalidate-mir: check that all referenced locals exist

[jonas-schievink]

When encountering a Local, there must be a corresponding LocalDecl in the MIR body. If there isn't, later passes can panic due to indexing out of bounds of the LocalDecl array. The validation pass could detect this problem earlier.

[jonas-schievink]

Instructions:

Extend the validation pass in src/librustc_mir/transform/validate.rs by implementing visit_local for the TypeChecker MIR visitor:

rust/src/librustc_mir/transform/validate.rs

Line 88 in 4fb54ed

impl<'a, 'tcx> Visitor<'tcx> for TypeChecker<'a, 'tcx> {

The Visitor trait is documented here.

The method should check that the local exists in self.body.local_decls, and fail with a descriptive error message if not (by calling self.fail, not by panicking directly).

[yerke]

@rustbot claim

[jonas-schievink]

@yerke FYI, the validation pass currently fails on even libcore until #73175 is merged, so you need to either wait for that, or cherry-pick some commits from there to get it to work.

You can test your changes by running RUSTFLAGS_NOT_BOOTSTRAP="-Zvalidate-mir" ./x.py test --stage=1 src/test/ui (the validator only runs on very little code by default, -Zvalidate-mir turns it on for everything).

[yerke]

@jonas-schievink I wanted to double check, running RUSTFLAGS_NOT_BOOTSTRAP="-Zvalidate-mir" ./x.py test --stage=1 src/test/ui should fail when I run it, right? It actually passed for me after I merged in your branch from #73359.
I wonder if I should try clean and build stage 1 again. :\

[jonas-schievink]

It should pass with #73359 applied, but fail on current master.

[yerke]

@jonas-schievink So your PR fixes this issue (#73356) as well?

[jonas-schievink]

Ah, no, this issue is just about adding additional validation. It shouldn't actually detect any locals that don't exist. So the command should work both before and after implementing this. If it fails after implementing this, something else is going on that we should look into.

[pitaj]

@yerke are you still working on this?

[yerke]

No, sorry. Didn’t have time for it. Go ahead.

[jonas-schievink]

@rustbot release-assignment