Ignore coroutine witness type region args in auto trait confirmation

[compiler-errors]

The problem

Consider code like:

async fn process<'a>() {
    Box::pin(process()).await;
}

fn require_send(_: impl Send) {}

fn main() {
    require_send(process());
}

When proving that the coroutine {coroutine@process}::<'?0>: Send, we end up instantiating a nested goal {witness@process}::<'?0>: Send by synthesizing a witness type from the coroutine's args:

Proving a coroutine witness type implements an auto trait requires looking up the coroutine's witness types. The witness types are a binder that look like for<'r> { Pin<Box<{coroutine@process}::<'r>>> }. We instantiate this binder with placeholders and prove Send on the witness types. This ends up eventually needing to prove something like {coroutine@process}::<'!1>: Send. Repeat this process, and we end up in an overflow during fulfillment, since fulfillment does not use freshening.

This can be visualized with a trait stack that ends up looking like:

• {coroutine@process}::<'?0>: Send
  • {witness@process}::<'?0>: Send
    • Pin<Box<{coroutine@process}::<'!1>>>: Send
      • {coroutine@process}::<'!1>: Send
        • ...
          • {coroutine@process}::<'!2>: Send
            • {witness@process}::<'!2>: Send
              • ...
                • overflow!

The problem here specifically comes from the first step: synthesizing a witness type from the coroutine's args.

Why wasn't this an issue before?

Specifically, before 63f6845, this wasn't an issue because we were instead extracting the witness from the coroutine type itself. It turns out that given some {coroutine@process}::<'?0>, the witness type was actually something like {witness@process}::<'erased>!

So why do we end up with a witness type with 'erased in its args? This is due to the fact that opaque type inference erases all regions from the witness. This is actually explicitly part of opaque type inference -- changing this to actually visit the witness types actually replicates this overflow even with 63f6845 reverted:

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

Lines 303 to 313 in ca77504

	ty::Coroutine(_, args) => {
	// Skip lifetime parameters of the enclosing item(s)
	// Also skip the witness type, because that has no free regions.
	for upvar in args.as_coroutine().upvar_tys() {
	upvar.visit_with(self);
	}
	args.as_coroutine().return_ty().visit_with(self);
	args.as_coroutine().yield_ty().visit_with(self);
	args.as_coroutine().resume_ty().visit_with(self);
	}

To better understand this difference and how it avoids a cycle, if you look at the trait stack before 63f6845, we end up with something like:

• {coroutine@process}::<'?0>: Send
  • {witness@process}::<'erased>: Send <-- THIS CHANGED
    • Pin<Box<{coroutine@process}::<'!1>>>: Send
      • {coroutine@process}::<'!1>: Send
        • ...
          • {coroutine@process}::<'erased>: Send <-- THIS CHANGED
            • coinductive cycle! 🎉

So what's the fix?

This hack replicates the behavior in opaque type inference to erase regions from the witness type, but instead erasing the regions during auto trait confirmation. This is kinda a hack, but is sound. It does not need to be replicated in the new trait solver, of course.

---

I hope this explanation makes sense.

We could beta backport this instead of the revert #145193, but then I'd like to un-revert that on master in this PR along with landing this this hack. Thoughts?

r? lcnr

[lcnr]

what a mess, the reason this works is even more subtle. While adding member constraints does ignore witnesses, any lifetime that occurs both in the witness and somewhere else also causes lifetimes in the witness to get mapped back to that lifetime, see https://rust-lang.zulipchat.com/#narrow/channel/144729-t-types/topic/nested.20bodies.20in.20opaque.20types/near/523055535 where I explain how this manifests for closures.

However, we never actually check that the witness args are equal to the coroutine args, so in MIR borrowck, coroutines are all coroutine<'parent_a, witness<'unconstrained_lt>>. So that's why witness args have to remain static.

Manually erasing the regions in the trait solver sgtm

I think I would just backport this PR

@bors r+

[bors]

📌 Commit b4aa629 has been approved by lcnr

It is now in the queue for this repository.

[lcnr]

need generic args for typeck_root, not the coroutine here

gonna fix soon