nested RPIT and HRTB: unclear semantics and future incompatibility

[aliemjay]

(Original post in internals forum)

Given the following code (stable sin 1.30!):

fn f() -> impl for<'a> Tr<'a, Assoc = impl Copy> {}

It is undocumented whether the inner existential type, impl Copy, is allowed to access 'a or not. In other words, it is equivalent to which of the following?

type Assoc = impl Copy;
fn f() -> impl for<'a> Tr<'a, Assoc = Assoc> {}

or,

type Assoc<'a> = impl Copy;
fn f() -> impl for<'a> Tr<'a, Assoc = Assoc<'a>> {}

On stable

The behavior in current stable is consistent with the first as shown by:

• If Assoc uses the lifetime 'a, the compiler complains about Implementation is not general enough as expected.
• The following test on the return type passes

fn assert_unique_assoc<Assoc>(_: impl for<'a> Tr<'a, Assoc = Assoc>) {}
fn main() { assert_unique_assoc(f()); }

It looks like this was unintentionally stabilized and changing that now would break the above test (unless we leak the fact that impl Copy captures 'a ??).

On beta

#94081 makes #88236 pass but that seems to be inconsistent with the behavior in stable:

• If Assoc uses 'a, it gives ICE, see yet another ICE with HRTB and RPIT #95647.
• Using the above test on the return type fails, which is quite surprising since removing the lifetime from impl Copy + 'a makes the test pass. playground. And I believe adding a lifetime bound shouldn't change the semantics in this regard.
• It would be a backward compatibility risk if we ever want to reject the code in the future.

Meta

stable version: 1.60.0
beta version: 1.61.0-beta.3 2022-04-17 2431a974c2dcf82ba513)

@rustbot label +T-Lang +A-impl-Trait

[aliemjay]

Just now found out that nested impl Trait in argument position are allowed 🙈

fn test_argument_position(x: impl for<'a> Tr<'a, Assoc = impl Copy>)  {}

and it actually desugars to

T: for<'a> Tr<'a, Assoc = Assoc>,
Assoc: Copy,

instead of:

for<'a> T: Tr<'a>, 
for<'a> <T as Tr<'a>>::Assoc: Copy,

Now I believe this is a strong argument to reject impl for<'a> Tr<'a, Assoc = impl Copy + 'a> in return position!

[aliemjay]

It would be unfortunate if it hits stable with such ambiguous semantics before the language team decides on the issue.

Compiler team might want to intervene in this case.
@rustbot label T-compiler I-prioritize

[inquisitivecrystal]

@oli-obk: I just wanted to check, was this intended to be stabilized?

Edit: sorry, I didn't see the internals link! That makes it clear that it was unintentional.

[Mark-Simulacrum]

cc @nikomatsakis / @jackh726 for lack of better contacts in still-forming traits team, this is potentially another issue that is good to urgently-ish take a look at as an accidental stabilization.

[apiraino]

Assigning priority as discussed in the Zulip thread of the Prioritization Working Group.

@rustbot label -I-prioritize +P-critical