"constructing invalid value: wrong trait in wide pointer vtable" for seemingly identical traits

[Jules-Bertholet]

I don't have a minimal reduction atm, but the issue can be reproduced by cloning https://github.com/Jules-Bertholet/unsized-vec and running cargo miri test -p emplacable. This results in the following report of UB:

constructing invalid value: wrong trait in wide pointer vtable: expected `for<'b> std::ops::FnMut(std::alloc::Layout, <[std::boxed::Box<i32>] as std::ptr::Pointee>::Metadata, &'b mut (dyn std::ops::FnMut(*mut std::marker::PhantomData<[std::boxed::Box<i32>]>) + 'b))`, but encountered `for<'a> std::ops::FnMut<(std::alloc::Layout, usize, &'a mut (dyn std::ops::FnMut(*mut std::marker::PhantomData<[std::boxed::Box<i32>]>) + 'a))>`

But the two traits are the same! <[std::boxed::Box<i32>] as std::ptr::Pointee>::Metadata is usize, and for<'a> ... vs for<'b> ... should not make a difference either.

@rustbot label I-false-UB A-validation

[RalfJung]

Probably the use of != around here is wrong. @lcnr what is the proper way to compare two Option<ty::Binder<'tcx, ExistentialTraitRef<'tcx>>>?

[lcnr]

let ocx = ObligationCtxt::new(infcx);
let param_env = ty::ParamEnv::reveal_all(); // in an empty environment post borrowck
ocx.eq(param_env, lhs, rhs)?; // equate the two trait refs
ocx.select_all_or_error().is_empty() // and check any potential nested constraints

[RalfJung]

Thanks! We actually have a param_env in the interpreter so I guess it won't be reveal_all. But we don't have a infcx.

[lcnr]

you can create a local one, given that the input should not refer to either placeholders or inference variables