Auto merge of rust-lang#132325 - lcnr:damn-son, r=oli-obk

rework winnowing to sensibly handle global where-bounds

There may be multiple ways to prove a given trait-bound. In case there are multiple such applicable candidates we need to somehow merge them or bail with ambiguity. When merging candidates we prefer some over others for multiple reasons:

- we want to guide inference during typeck, even if not strictly necessary
- to avoid ambiguity if there if there are at most lifetime differences
    - old solver needs exactly one candidate
    - new solver only needs to handle lifetime differences
- we disable normalization via impls if the goal is proven by using a where-bound

## The approach in this PR[^1]

- always prefer trivial builtin impls[^6]
- then prefer non-global[^global] where-bounds
    - if there exists exactly one where-bound, guide inference
    - if there are multiple where-bounds even if some of them are global, ambig
- then prefer alias bounds[^2] and builtin trait object candidates[^3][^2]
- merge everything ignoring global where-bounds
- if there are no other candidates, try using global where-bounds[^5]

**We disable normalization via impls when using non-global where-bounds or alias-bounds, even if we're unable to normalize by using the where-bound.**

[^1]: see the source for more details
[^2]: [we arbitrary select a single object and alias-bound candidate in case multiple apply and they don't impact inference](https://github.com/rust-lang/rust/blob/a4cedecc9ec76b46dcbb954750068c832cf2dd43/compiler/rustc_trait_selection/src/traits/select/mod.rs#L1906-L1911). This should be unnecessary in the new solver.
[^3]: Necessary for `dyn Any` and rust-lang#57893
[^global]: a where-bound is global if it is not higher-ranked and doesn't contain any generic parameters, `'static` is ok
[^5]: global where-bounds are only used if they are unsatisfiable, i.e. no impl candidate exists
[^6]: they don't constrain inference and don't add any lifetime constraints

## Why this behavior?

### inference guidance via where-bounds and alias-bounds

#### where-bounds

```rust
fn method_selection<T: Into<u64>>(x: T) -> Option<u32> {
    x.into().try_into().ok()
    // prove `T: Into<?0>` and then select a method `?0`,
    // needs eager inference.
}
```

While the above pattern exists in the wild, I think that most inference guidance due to where-bounds is actually unintended. I believe we may want to restrict inference guidance in the future, e.g. limit it to where-bounds whose self-type is a param.

#### alias-bounds

```rust
pub trait Dyn {
    type Word: Into<u64>;
    fn d_tag(&self) -> Self::Word;
    fn tag32(&self) -> Option<u32> {
        self.d_tag().into().try_into().ok()
        // prove `Self::Word: Into<?0>` and then select a method
        // on `?0`, needs eager inference.
    }
}
```

### Disable normalization via impls when using where-bounds

cc rust-lang/trait-system-refactor-initiative#125

```rust
trait Trait<'a> {
    type Assoc;
}

impl<T> Trait<'static> for T {
    type Assoc = ();
}

// normalizing requires `'a == 'static`, the trait bound does not.
fn foo<'a, T: Trait<'a>>(_: T::Assoc) {}
```

If an impl adds constraints not required by a where-bound, using the impl may cause compilation failure avoided by treating the associated type as rigid.

This is also why we can always use trivial builtin impls, even for normalization. They are guaranteed to never add any requirements.

### Lower priority for global where-bounds

A where-bound is considered global if it does not refer to any generic parameters and is not higher-ranked. It may refer to `'static`.

This means global where-bounds are either always fully implied by an impl or unsatisfiable. We don't really care about the inference behavior of unsatisfiable where-bounds :3

If a where-bound is fully implied then using an applicable impl for normalization cannot result in additional constraints. As this is the - afaict only - reason why we disable normalization via impls in the first place, we don't have to disable normalization via impls when encountering global where-bounds.

### Consider global where-bounds at all

Given that we just use impls even if there exists a global where-bounds, you may ask why we don't just ignore these global where-bounds entirely: we use them to weaken the inference guidance from non-global where-bounds.

Without a global where-bound, we currently prefer non-global where bounds even though there would be an applicable impl as well. By adding a non-global where-bound, this *unnecessary* inference guidance is disabled, allowing the following to compile:
```rust
fn check<Color>(color: Color)
where
    Vec: Into<Color> + Into<f32>,
{
    let _: f32 = Vec.into();
    // Without the global `Vec: Into<f32>`  bound we'd
    // eagerly use the non-global `Vec: Into<Color>` bound
    // here, causing this to fail.
}

struct Vec;
impl From<Vec> for f32 {
    fn from(_: Vec) -> Self {
        loop {}
    }
}
```
[There exist multiple crates which rely on this behavior](rust-lang#124592 (comment)).

## Design considerations

We would like to be able to normalize via impls as much as possible. Disabling normalization simply because there exists a where-bound is undesirable.

For the sake of backwards compatability I intend to mostly mirror the current inference guidance rules and then explore possible improvements once the new solver is done. I do believe that removing unnecessary inference guidance where possible is desirable however.

Whether a where-bound is global depends on whether used lifetimes are `'static`. The where-bound `u32: Trait<'static>` is either entirely implied by an impl, meaning that it does not have to disable normalization via impls, **while `u32: Trait<'a>` needs to disable normalization via impls as the impl may only hold for `'static`**. Considering all where-bounds to be non-global once they contain any region is unfortunately a breaking change.

## How does this differ from stable

The currently stable approach is order dependent:
- it prefers impls over global where-bounds: impl > global
- it prefers non-global where-bounds over impls: non-global > impl
- it treats all where-bounds equally: global = non-global

This means that whether we bail with ambiguity or simply use the non-global where bound depending on the *order of where-clauses* and *number of applicable impl candidates*. See the tests added in the first commit for more details. With this PR we now always bail with ambiguity.

I've previously tried to always use the non-global candidate, causing unnecessary inference guidance and undesirable breakage. This already went through an FCP in rust-lang#124592. However, I consider the new approach to be preferable as it exclusively removes incompleteness. It also doesn't cause any crater breakage.

## How to support this in the new solver :o

**This is separately implemented in rust-lang#133643 and not part of this FCP!**

To implement the global vs non-global where-bound distinction, we have to either keep `'static` in the `param_env` when canonicalizing, or eagerly distinguish global from non-global where-bounds and provide that information to the canonical query.

The old solver currently keeps `'static` only the `param_env`, replacing it with an inference variable in the `value`.
https://github.com/rust-lang/rust/blob/a4cedecc9ec76b46dcbb954750068c832cf2dd43/compiler/rustc_infer/src/infer/canonical/canonicalizer.rs#L49-L64

I dislike that based on *vibes* and it may end up being a problem once we extend the environment inside of the solver as [we must not rely on `'static` in the `predicate` as it would get erased in MIR typeck](rust-lang/trait-system-refactor-initiative#30).

An alternative would be to eagerly detect trivial where-bounds when constructing the `ParamEnv`. We can't entirely drop them [as explained above](https://hackmd.io/qoesqyzVTe2v9cOgFXd2SQ#Consider-true-global-where-bounds-at-all), so we'd instead replace them with a new clause kind `TraitImpliedByImpl` which gets entirely ignored except when checking whether we should eagerly guide inference via a where-bound. This approach can be extended to where-bounds which are currently not considered global to stop disabling normalization for them as well.

Keeping `'static` in the `param_env` is the simpler solution here and we should be able to move to the second approach without any breakage. I therefore propose to keep `'static` in the environment for now.

---

r? `@compiler-errors`

Author: bors

compiler/rustc_trait_selection/src/lib.rs

@@ -23,6 +23,7 @@
 #![feature(extract_if)]
 #![feature(if_let_guard)]
 #![feature(iter_intersperse)]
+#![feature(iterator_try_reduce)]
 #![feature(let_chains)]
 #![feature(never_type)]
 #![feature(rustdoc_internals)]

compiler/rustc_trait_selection/src/traits/select/mod.rs

@@ -0,0 +1,23 @@
+//@ check-pass
+
+// This caused a regression in a crater run in #132325.
+//
+// The underlying issue is a really subtle implementation detail.
+//
+// When building the `param_env` for `Trait` we start out with its
+// explicit predicates `Self: Trait` and `Self: for<'a> Super<'a, { 1 + 1 }>`.
+//
+// When normalizing the environment we also elaborate. This implicitly
+// deduplicates its returned predicates. We currently first eagerly
+// normalize constants in the unnormalized param env to avoid issues
+// caused by our lack of deferred alias equality.
+//
+// So we actually elaborate `Self: Trait` and `Self: for<'a> Super<'a, 2>`,
+// resulting in a third `Self: for<'a> Super<'a, { 1 + 1 }>` predicate which
+// then gets normalized to  `Self: for<'a> Super<'a, 2>` at which point we
+// do not deduplicate however. By failing to handle equal where-bounds in
+// candidate selection, this caused ambiguity when checking that `Trait` is
+// well-formed.
+trait Super<'a, const N: usize> {}
+trait Trait: for<'a> Super<'a, { 1 + 1 }> {}
+fn main() {}

tests/ui/const-generics/min_const_generics/param-env-eager-norm-dedup.rs

@@ -0,0 +1,18 @@
+//@ check-pass
+
+// A regression test for an edge case of candidate selection
+// in the old trait solver, see #132325 for more details.
+
+trait Trait<T> {}
+impl<T> Trait<T> for () {}
+
+fn impls_trait<T: Trait<U>, U>(_: T) -> U { todo!() }
+fn foo<T>() -> u32
+where
+    (): Trait<u32>,
+    (): Trait<T>,
+{
+    impls_trait(())
+}
+
+fn main() {}

tests/ui/traits/winnowing/global-non-global-env-1.rs

@@ -0,0 +1,20 @@
+//@ check-pass
+
+// A regression test for an edge case of candidate selection
+// in the old trait solver, see #132325 for more details. Unlike
+// the first test, this one has two impl candidates.
+
+trait Trait<T> {}
+impl Trait<u32> for () {}
+impl Trait<u64> for () {}
+
+fn impls_trait<T: Trait<U>, U>(_: T) -> U { todo!() }
+fn foo<T>() -> u32
+where
+    (): Trait<u32>,
+    (): Trait<T>,
+{
+    impls_trait(())
+}
+
+fn main() {}

tests/ui/traits/winnowing/global-non-global-env-2.rs

@@ -0,0 +1,20 @@
+//@ check-pass
+
+// A regression test for an edge case of candidate selection
+// in the old trait solver, see #132325 for more details. Unlike
+// the second test, the where-bounds are in a different order.
+
+trait Trait<T> {}
+impl Trait<u32> for () {}
+impl Trait<u64> for () {}
+
+fn impls_trait<T: Trait<U>, U>(_: T) -> U { todo!() }
+fn foo<T>() -> u32
+where
+    (): Trait<T>,
+    (): Trait<u32>,
+{
+    impls_trait(())
+}
+
+fn main() {}

tests/ui/traits/winnowing/global-non-global-env-3.rs

@@ -0,0 +1,21 @@
+//@ check-pass
+
+// A regression test for an edge case of candidate selection
+// in the old trait solver, see #132325 for more details. Unlike
+// the third test, this one has 3 impl candidates.
+
+trait Trait<T> {}
+impl Trait<u32> for () {}
+impl Trait<u64> for () {}
+impl Trait<u128> for () {}
+
+fn impls_trait<T: Trait<U>, U>(_: T) -> U { todo!() }
+fn foo<T>() -> u32
+where
+    (): Trait<T>,
+    (): Trait<u32>,
+{
+    impls_trait(())
+}
+
+fn main() {}