rustc: fix ICE when trait alias has bare Self

[seanmonstar]

Fixes #59029

[rust-highfive]

r? @oli-obk

(rust_highfive has picked a reviewer for you, use r? to override)

[seanmonstar]

cc @alexreg

[alexreg]

@seanmonstar Thanks! I'm glad the simple fix seems to have worked. Would you mind creating a new test under src/test/run-pass/issues/issue-59029.rs and writing a bit of code there that actually uses a trait alias with a Self bound in some meaningful way (simple is fine, but at least have an assert)? You can just revert the existing test file then.

[alexreg]

r? @alexreg

[alexreg]

Okay, so this fix isn't actually enough, it seems... Could you try the following locally and let me know if they succeed?

trait SelfWhere<One, Two> = Assoc<Two> where Assoc<Two>::Out: Assoc<One>;

and

trait SelfWhere<One, Two> = Assoc<Two> where <Self as Assoc<Two>>::Out: Assoc<One>;

[seanmonstar]

The second one does work, it's part of the initial bug report. I haven't tried the first yet.

If Self should be qualified, I can just change this to a compile-fail test checking for the right error.

[seanmonstar]

The first one (Assoc<Two> where Assoc<Two>::Out: Assoc<One>) complains about "abiguous associated type".

[alexreg]

As a loose analogy to this example (insofar as it involves trait bounds on an associated type), I'm of the opinion that it's fine for Self alone not to work (so long as it doesn't ICE, of course).

So, now that we've solved the ICE, the question is: do we want better diagnostics when, e.g., trait SelfWhere<One, Two> = Assoc<Two> where <Self as Assoc<Two>>::Out: Assoc<One>; is used. @estebank, what do you think about the error message in the above CI test failure. Should we be doing better?

@seanmonstar Let's have a ui test for the failing case (just Self), and a run-pass test for the case that should pass (<Self as ...>). Two new test files. You can put them both in the ui/issues/ dir though, maybe, and just annotate the passing one with // run-pass as its first line.

[estebank]

@alexreg it's probably fine. The ideal output would be to provide a suggestion to use <Self as Svc<Target>>::Res, but to do that you'd have to peek into Svc to see if the assoc type not found exists in it. I'm ok with merging this PR as is.

error[E0220]: associated type `Res` not found for `Self`
  --> $DIR/issue-59029-1.rs:5:46
   |
LL | trait MkSvc<Target, Req> = Svc<Target> where Self::Res: Svc<Req>;
   |                                              ^^^^^^^^^
   |                                              |
   |                                              associated type `Res` not found
   |                                              help: you might have meant to use the fully qualified path: `<Self as Svc<Target>>::Res`

[alexreg]

@estebank Great. I'm okay with merging this as-is too, in that case.

[alexreg]

@bors r+

Thanks @seanmonstar!

[bors]

📌 Commit df05fbf has been approved by alexreg