Inference issues with trait constification

[clarfonthey]

I'm just going to file an issue for this so it's tracked somewhere.

Tracking for const_convert feature: #143773

Essentially, Implementations for Cow were reverted (#148011 and #148016) due to inference failures on beta (see #147964):

impl<T: ?Sized + ToOwned> const AsRef<T> for Cow<'_, T> where T::Owned: [const] Borrow<T> { /* ... */ }
impl<T: ?Sized + ToOwned> const Borrow<T> for Cow<'_, T> where T::Owned: [const] Borrow<T> { /* ... */ }
impl<T: ?Sized + ToOwned> const Deref for Cow<'_, T> where T::Owned: [const] Borrow<T> { /* ... */ }

The main issue here is the extra where clauses adding bounds which are technically redundant, but required for constness. We could fix these by making these contingent on const ToOwned, but I intentionally didn't do that in the initial implementation because it wasn't necessary: only const borrowing is necessary here, not const allocation.

Perhaps the trait solver could correctly identify these where clauses as being "redundant" except for constness to avoid the issues, or maybe next-solver fixes this. Either way, constification of these is probably blocked in the meantime, and I doubt these are going to be the only decisions going forward.

The reality is that constifying traits generally involves a lot of where clauses that might mess with the trait solver, and something needs to be decided on that.

---

Extra note: Deref was specifically not de-constified properly in 1.57.0, so it has been changed a bit.

(note, I'm using modern syntax for this for ease of reading, but technically the const trait syntax changed between 1.57.0 and 1.90.0)

// 1.56.1
impl<T: ?Sized + ToOwned> Deref for Cow<'_, T> { /* ... */ }

// 1.57.0
impl<T: ?Sized + ToOwned> const Deref for Cow<'_, T> where T::Owned: [const] Borrow<T> { /* ... */ }

// 1.70.0 - 1.90.0
impl<T: ?Sized + ToOwned> Deref for Cow<'_, T> where T::Owned: Borrow<T> { /* ... */ }

// 1.91.0 beta
impl<T: ?Sized + ToOwned> Deref for Cow<'_, T> { /* ... */ }

The decision was made to keep the status quo this close to the 1.90.0 stable release, but the revert to the pre-1.57 version is still on the 1.91.0 beta for now.

[cuviper]

I think this may have been a weakness even without the [const] part:

• The original regression: the method as_ref exists for reference #147964 found a problem with AsRef, which didn't have an associate type bound at all before.
• Then it was noticed that problem also existed if you tried similar test with Borrow or Deref.
• Revert constification of Borrow and Deref for Cow due to inference failure #148016 dropped the remaining associated type bounds, and added a test for all three traits.
• regression: the method as_ref exists for reference #147964 (comment) noticed that Deref did use to have a non-const associated type bound, so Revert constification of Borrow and Deref for Cow due to inference failure #148016 wasn't fully going back to the state we had before.
• I tried to further that during the stable preparation, but that failed the new test from Revert constification of Borrow and Deref for Cow due to inference failure #148016!
  [stable] Prepare Rust 1.91.0 release #148179 (comment)

So, I think we had just never tried that kind of code with Deref before, const or not.

[clarfonthey]

Summarised the changes on the Deref trait in the description based upon what was decided upon for the upcoming stable/beta releases. Also opened #148228 to make sure the regression tests run on the next solver too, at least until a decision is made on what these should be doing.