implied bounds from associated types may not actually get implied pt 2.

[lcnr]

revives #91068 which has been fixed by only considering implied bounds from projections if they don't normalize while typechecking the function itself

rust/compiler/rustc_borrowck/src/type_check/free_region_relations.rs

Lines 256 to 289 in 7125846

	// We only add implied bounds for the normalized type as the unnormalized
	// type may not actually get checked by the caller.
	//
	// Can otherwise be unsound, see #91068.
	let TypeOpOutput { output: norm_ty, constraints: constraints1, .. } = self
	.param_env
	.and(type_op::normalize::Normalize::new(ty))
	.fully_perform(self.infcx)
	.unwrap_or_else(|_| {
	self.infcx
	.tcx
	.sess
	.delay_span_bug(DUMMY_SP, &format!("failed to normalize {:?}", ty));
	TypeOpOutput {
	output: self.infcx.tcx.ty_error(),
	constraints: None,
	error_info: None,
	}
	});
	// Note: we need this in examples like
	// ```
	// trait Foo {
	// type Bar;
	// fn foo(&self) -> &Self::Bar;
	// }
	// impl Foo for () {
	// type Bar = ();
	// fn foo(&self) ->&() {}
	// }
	// ```
	// Both &Self::Bar and &() are WF
	let constraints_implied = self.add_implied_bounds(norm_ty);
	normalized_inputs_and_output.push(norm_ty);
	constraints1.into_iter().chain(constraints_implied)

This does not mean that the projection won't normalize when getting called:

trait Trait {
    type Type;
}

impl<T> Trait for T {
    type Type = ();
}

fn f<'a, 'b>(s: &'b str, _: <&'a &'b () as Trait>::Type) -> &'a str
where
    &'a &'b (): Trait, // <- adding this bound is the change from #91068 
{
    s
}

fn main() {
    let x = String::from("Hello World!");
    let y = f(&x, ());
    drop(x);
    println!("{}", y);
}

The added bound prevents <&'a &'b () as Trait>::Type from getting normalized while borrowchecking f, as we prefer param candidates over impl candidates. When calling f, we don't have the &'a &'b (): Trait in our param_env, so we can now freely use the impl candidate to normalize the projection. The caller therefore doesn't have to prove that &'a &'b () is well formed, causing unsoundness.

I am a bit surprised that the caller doesn't have to prove that the &'a &'b (): Trait obligation is well formed, which would cause this example to not be unsound. It doesn't remove the general unsoundness here though. Here's an alternative test where that isn't enough:

trait Trait {
    type Type;
}

impl<T> Trait for T {
    type Type = ();
}

fn f<'a, 'b>(s: &'b str, _: <&'a &'b () as Trait>::Type) -> &'a str
where
    for<'what, 'ever> &'what &'ever (): Trait,
{
    s
}

fn main() {
    let x = String::from("Hello World!");
    let y = f(&x, ());
    drop(x);
    println!("{}", y);
}

[steffahn]

Unlike #91068, these code examples compile “successfully” all the way since Rust 1.7.0.

[apiraino]

WG-prioritization assigning priority (Zulip discussion).

@rustbot label -I-prioritize +P-high

[lcnr]

while this currently does not work, we should also add a test for something like this. If we get stronger implied bounds, simply requiring the caller to prove wf for all projection components won't be enough anymore

#![feature(generic_associated_types)]

trait Trait<'a> {
    type Type
    where
        Self: 'a;
}

impl<'a, T> Trait<'a> for T {
    type Type = ()
    where
        Self: 'a; // once this bound gets implied, we have UB again
}

fn f<'a, 'b>(s: &'b str, _: <&'b () as Trait<'a>>::Type) -> &'a str
where
    &'b (): Trait<'a>, // <- adding this bound is the change from #91068 
{
    s
}

fn main() {
    let x = String::from("Hello World!");
    let y = f(&x, ());
    drop(x);
    println!("{}", y);
}

trait Trait<'a>: 'a {
    type Type;
}

impl<'a, T> Trait<'a> for T
where
    T: 'a // if this bound get's implied we would probably get ub here again
{
    type Type = ();
}

fn f<'a, 'b>(s: &'b str, _: <&'b () as Trait<'a>>::Type) -> &'a str
where
    &'b (): Trait<'a>, // <- adding this bound is the change from #91068 
{
    s
}

fn main() {
    let x = String::from("Hello World!");
    let y = f(&x, ());
    drop(x);
    println!("{}", y);
}

[jackh726]

There's a PR open that fixes this.