Normalization can skip WF

[CAD97]

I tried this code: [playground]

trait WellUnformed {
    type RequestNormalize;
}

impl<T: ?Sized> WellUnformed for T {
    type RequestNormalize = ();
}

const _: <[[[[[[u8]]]]]] as WellUnformed>::RequestNormalize = ();
const _: <Vec<str> as WellUnformed>::RequestNormalize = ();

I expected this to error; [[[[[[u8]]]]]] is clearly not a well-formed type, as

error[E0277]: the size for values of type [u8] cannot be known at compilation time
= help: the trait Sized is not implemented for [u8]
= note: slice and array elements must have Sized type

Vec<str> is perhaps even more cursed.

Instead: it compiles.

This is especially relevant for blanket/builtin traits which have associated types, such as marker::DiscriminantKind and ptr::Pointee.

• Can cause ICE: Normalization can skip WF #100041 (comment)
• Also causes us to sometimes ignore where-bounds of aliases Trait bound in associated type isn't checked when normalizing type that doesn't require inspecting the AT #154172
• Believed to not be unsound in isolation, as the non-WF type never "fully" exists.

Meta

1.64.0-nightly (2022-07-31 f9cba63)

@rustbot label +A-associated-items

cc @eddyb who said this was
@rustbot label +I-unsound

[eddyb]

cc @rust-lang/types

[CAD97]

@rustbot label +T-types

[compiler-errors]

I think I found the bug(s). This is:

1. partly due to Try to normalize associated types before processing obligations #90887, which normalizes WellFormed predicates before we can actually walk through the projection type and register a WF obligation for the projection's self ty,
2. and partly due to us doing a normalize call before registering a WellFormed predicate in wfcheck code.

Let me put up a PR.

[jackh726]

Is this fixed in #99217?

[compiler-errors]

Just checked out that PR, and it's not fixed by #99217.

[apiraino]

WG-prioritization assigning priority (Zulip discussion).

@rustbot label -I-prioritize +P-high