Cannot resolve TAIT when used with impl/dyn Trait inside RPIT

[xxchan]

The condition seems complex, but a practical instance is async fn foo() -> Result<TAIT>

I tried this code:

#![feature(type_alias_impl_trait)]

trait Trait {}

type TAIT = impl Trait;

type Result<T> = std::result::Result<T, Box<dyn std::error::Error + Send + Sync>>;

fn foo() -> TAIT {
    struct Bar;
    impl Trait for Bar {}
    Bar
}

async fn bar() -> Result<TAIT> {
    Ok(foo())
}

I expected to see this happen: it compiles

Instead, this happened: error:

error: concrete type differs from previous defining opaque type use
  --> src/tmp.rs:15:32
   |
15 |   async fn bar() -> Result<TAIT> {
   |  ________________________________^
16 | |     Ok(foo())
17 | | }
   | |_^ expected `Bar`, got `TAIT`
   |
note: previous use here
  --> src/tmp.rs:12:5
   |
12 |     Bar
   |     ^^^

Notes:

• It compiles in nightly-2022-08-12, but failed in nightly-2022-09-13
• fn() -> Result<TAIT> works, and async fn() -> Option<TAIT> works

Meta

rustc --version --verbose:

rustc 1.65.0-nightly (fa6ee9375 2022-09-12)
binary: rustc
commit-hash: fa6ee9375242ae784dab1837dfc0b92f43e787ce
commit-date: 2022-09-12
host: aarch64-apple-darwin
release: 1.65.0-nightly
LLVM version: 15.0.0

[xxchan]

@rustbot label F-type_alias_impl_trait

[xxchan]

nightly-2022-08-18 works, and nightly-2022-08-19 breaks

https://github.com/rust-lang/rust/commits/master?since=2022-08-18&until=2022-08-18

[xxchan]

trait Trait {}

type TAIT = impl Trait;

struct Bar;
impl Trait for Bar {}

fn foo() -> TAIT {
    Bar
}

// ok
async fn bar() -> (TAIT, Bar) {
   (foo(), Bar)
}

// err: concrete type differs
// async fn bar() -> (TAIT, impl Trait) {
//     (foo(), Bar)
// }

// err: concrete type differs
// async fn bar() -> (TAIT, Box<dyn Trait>) {
//     (foo(), Box::new(Bar))
// }

[xxchan]

Caused by #99860

Any ideas? cc @oli-obk @compiler-errors

[xxchan]

nightly-2022-05-24 also works (a version before #97346 which is reverted by #99860

[oli-obk]

TAIT and RPIT were made to work the same in the revert, thus TAIT was changed from before #97346 instead of just reverting to the old behaviour.

I will need to look at some typeck logs to see what's goin on here

[TennyZhuang]

FYI, it seems that the bug is not related to async but to every nested impl trait.

// err: concrete type differs
fn bar() -> impl Iterator<Item = (TAIT, impl Trait)> {
    std::iter::once((foo(), Bar))
}

[TennyZhuang]

I've tested with the following codes:

#![feature(type_alias_impl_trait)]

trait Trait {}

type TAIT = impl Trait;

struct Bar;
impl Trait for Bar {}

fn foo() -> TAIT {
    Bar
}

fn haha() -> (impl Iterator<Item = TAIT>, impl Trait) {
    (std::iter::empty::<Bar>(), Bar)
}

fn bar() -> impl Iterator<Item = (TAIT, impl Trait)> {
    std::iter::once((foo(), Bar))
}

It seems that the result of mir_borrowck for bar is wrong:

│ │ │ ├─106ms DEBUG rustc_borrowck do_mir_borrowck: result = BorrowCheckResult {
│ │ │ │     concrete_opaque_types: VecMap(
│ │ │ │         [
│ │ │ │             (
│ │ │ │                 DefId(0:13 ~ lib[af95]::bar::{opaque#0}),
│ │ │ │                 OpaqueHiddenType {
│ │ │ │                     span: ../playground/tait-bug/src/lib.rs:19:5: 19:34 (#0),
│ │ │ │                     ty: std::iter::Once<(TAIT, Bar)>,
│ │ │ │                 },
│ │ │ │             ),
│ │ │ │             (
│ │ │ │                 DefId(0:11 ~ lib[af95]::TAIT::{opaque#0}),
│ │ │ │                 OpaqueHiddenType {
│ │ │ │                     span: ../playground/tait-bug/src/lib.rs:19:5: 19:34 (#0),
│ │ │ │                     ty: TAIT,
│ │ │ │                 },
│ │ │ │             ),
│ │ │ │             (
│ │ │ │                 DefId(0:14 ~ lib[af95]::bar::{opaque#1}),
│ │ │ │                 OpaqueHiddenType {
│ │ │ │                     span: ../playground/tait-bug/src/lib.rs:19:5: 19:34 (#0),
│ │ │ │                     ty: Bar,
│ │ │ │                 },
│ │ │ │             ),
│ │ │ │         ],
│ │ │ │     ),
│ │ │ │     closure_requirements: None,
│ │ │ │     used_mut_upvars: [],
│ │ │ │     tainted_by_errors: None,
│ │ │ │ }

It seems that 0:11 wasn't be resolved correctly.

[compiler-errors]

I think @oli-obk fixed this (inadvertently?) in #102700

[xxchan]

Just tested it and I can confirm it works now! Thanks!

BTW, what about adding this issue as a test cast?

[compiler-errors]

Yes a minimized test PR would be appreciated.

[xxchan]

@rustbot claim