Initial UnsafePinned impl [Part 2: Lowering]

[Sky9x]

Coroutine lowering part of RFC 3467.
Tracking issue: #125735
Part 1: #137043

• Coroutines now track which of their stored locals are borrowed across suspension points and thus must be considered pinned.
• UnsafeUnpin for coroutines:
  • if the coro is movable, it always implements UnsafeUnpin. (movable coros never have pinned fields)
  • if it is immovable (static), it implements UnsafeUnpin iff none of its fields are pinned.

Still TODO (for later PRs):

• Coroutine layout (restore niche opts)
• Codegen (noalias metadata etc)
• Miri

---

@rustbot label F-unsafe_pinned F-coroutines A-coroutines

cc @RalfJung @compiler-errors @traviscross

[rustbot]

r? @lcnr

rustbot has assigned @lcnr.
They will have a look at your PR within the next two weeks and either review your PR or reassign to another reviewer.

Use r? to explicitly pick a reviewer

[rustbot]

Some changes occurred to MIR optimizations

cc @rust-lang/wg-mir-opt

[RalfJung]

Movable coroutines now always implement UnsafeUnpin and immovable coroutines now never implement UnsafeUnpin.

"always" sounds wrong; a movable coroutine could have an UnsafeUnpin type as a field, right?

I understand this is consistent with what we do with Unpin today, but I wonder why we don't just let the auto trait do its thing: if all variables for all yield points are UnsafeUnpin, we just propagate that. So even immotable coroutines could be UnsafeUnpin if they happen to not have any references across yield points.

Coroutine layout (restore niche opts)

This is missing context -- what is the interaction with niche ops?

[Sky9x]

Movable coroutines now always implement UnsafeUnpin and immovable coroutines now never implement UnsafeUnpin.

"always" sounds wrong; a movable coroutine could have an UnsafeUnpin type as a field, right?

I understand this is consistent with what we do with Unpin today, but I wonder why we don't just let the auto trait do its thing: if all variables for all yield points are UnsafeUnpin, we just propagate that. So even immotable coroutines could be UnsafeUnpin if they happen to not have any references across yield points.

Updated the trait solving logic to only implement UnsafeUnpin if no fields are pinned. This allows immovable (static) coros with no pinned fields to implement UnsafeUnpin.

Coroutine layout (restore niche opts)

This is missing context -- what is the interaction with niche ops?

#129313 Disabled niches in coroutines because of #63818. The planned "Part 3: Codegen" will restore this.

[Sky9x]

I'm not sure how to write tests checking impls of UnsafeUnpin as the trait is private in libcore

[Sky9x]

This should prob use its own dataflow analysis to avoid false positives (see borrow_not_held_across_yield mir test) (EDIT: WIP)

[RalfJung]

#129313 Disabled niches in coroutines because of #63818. The planned "Part 3: Codegen" will restore this.

Ah, but only outside pinned fields I assume. :)