Polonius: update to 0.12.1, fix more move errors false positives, update test expectations
This PR:
- updates `polonius-engine` to version 0.12.1 to fix some move errors false positives
- fixes a fact generation mistake creating the other move errors false positives
- updates the test expectations for the polonius compare-mode so that all (minus the 2 OOMs) ui tests pass again (matching the [analysis doc](https://hackmd.io/CjYB0fs4Q9CweyeTdKWyEg?view) starting at case 34)
In my opinion, this is safe to rollup.
r? @nikomatsakis
When a trait bound is not met and restricting a type parameter would
make the restriction hold, use a structured suggestion pointing at an
appropriate place (type param in param list or `where` clause).
Account for opaque parameters where instead of suggesting extending
the `where` clause, we suggest appending the new restriction:
`fn foo(impl Trait + UnmetTrait)`.
Fix most remaining Polonius test differences
This fixes most of the Polonius test differences and also avoids overflow in issue-38591.rs.
r? @nikomatsakis
As described in #57374, NLL currently produces unhelpful higher-ranked
trait bound (HRTB) errors when '-Zno-leak-check' is enabled.
This PR tackles one half of this issue - making the error message point
at the proper span. The error message itself is still the very generic
"higher-ranked subtype error", but this can be improved in a follow-up
PR.
The root cause of the bad spans lies in how NLL attempts to compute the
'blamed' region, for which it will retrieve a span for.
Consider the following code, which (correctly) does not compile:
```rust
let my_val: u8 = 25;
let a: &u8 = &my_val;
let b = a;
let c = b;
let d: &'static u8 = c;
```
This will cause NLL to generate the following subtype constraints:
d :< c
c :< b
b <: a
Since normal Rust lifetimes are covariant, this results in the following
region constraints (I'm using 'd to denote the lifetime of 'd',
'c to denote the lifetime of 'c, etc.):
'c: 'd
'b: 'c
'a: 'b
From this, we can derive that 'a: 'd holds, which implies that 'a: 'static
must hold. However, this is not the case, since 'a refers to 'my_val',
which does not outlive the current function.
When NLL attempts to infer regions for this code, it will see that the
region 'a has grown 'too large' - it will be inferred to outlive
'static, despite the fact that is not declared as outliving 'static
We can find the region responsible, 'd, by starting at the *end* of
the 'constraint chain' we generated above. This works because for normal
(non-higher-ranked) lifetimes, we generally build up a 'chain' of
lifetime constraints *away* from the original variable/lifetime.
That is, our original lifetime 'a is required to outlive progressively
more regions. If it ends up living for too long, we can look at the
'end' of this chain to determine the 'most recent' usage that caused
the lifetime to grow too large.
However, this logic does not work correctly when higher-ranked trait
bounds (HRTBs) come into play. This is because HRTBs have
*contravariance* with respect to their bound regions. For example,
this code snippet compiles:
```rust
let a: for<'a> fn(&'a ()) = |_| {};
let b: fn(&'static ()) = a;
```
Here, we require that 'a' is a subtype of 'b'. Because of
contravariance, we end up with the region constraint 'static: 'a,
*not* 'a: 'static
This means that our 'constraint chains' grow in the opposite direction
of 'normal lifetime' constraint chains. As we introduce subtypes, our
lifetime ends up being outlived by other lifetimes, rather than
outliving other lifetimes. Therefore, starting at the end of the
'constraint chain' will cause us to 'blame' a lifetime close to the original
definition of a variable, instead of close to where the bad lifetime
constraint is introduced.
This PR improves how we select the region to blame for 'too large'
universal lifetimes, when bound lifetimes are involved. If the region
we're checking is a 'placeholder' region (e.g. the region 'a' in
for<'a>, or the implicit region in fn(&())), we start traversing the
constraint chain from the beginning, rather than the end.
There are two (maybe more) different ways we generate region constraints for NLL:
requirements generated from trait queries, and requirements generated
from MIR subtype constraints. While the former always use explicit
placeholder regions, the latter is more tricky. In order to implement
contravariance for HRTBs, TypeRelating replaces placeholder regions with
existential regions. This requires us to keep track of whether or not an
existential region was originally a placeholder region. When we look for
a region to blame, we check if our starting region is either a
placeholder region or is an existential region created from a
placeholder region. If so, we start iterating from the beginning of the
constraint chain, rather than the end.
Polonius: fix some cases of `killed` fact generation, and most of the `ui` test suite
Since basic Polonius functionality was re-enabled by @matthewjasper in #54468, some tests were still failing in the polonius compare-mode.
This PR fixes all but one test in the `ui` suite by:
- fixing some bugs in the fact generation code, related to the `killed` relation: Polonius would incorrectly reject some NLL-accepted code, because of these missing `killed` facts.
- ignoring some tests in the polonius compare-mode: a lot of those manually test the NLL or migrate mode, and the failures were mostly artifacts of the test revisions, e.g. that `-Z polonius` requires full NLLs. Some others were also both failing with NLL and succeeding with Polonius, which we can't encode in tests at the moment.
- blessing the output of some tests: whenever Polonius and NLL have basically the same errors, except for diagnostics differences, the Polonius output is blessed. Whenever we've advanced into a less experimental phase, we'll want to revisit these cases (much like we did on the NLL test suite last year) to specifically work on diagnostics.
Fact generation changes:
- we now kill loans on the destination place of `Call` terminators
- we now kill loans on the locals destroyed by `StorageDead`
- we now also handle assignments to projections: killing the loans on a either a deref-ed local, or the ones whose `borrowed_place` conflicts with the current place.
One failing test remains: an overflow during fact generation, on a case of polymorphic recursion (and which I'll continue investigating later).
This adds some tests for the fact generation changes, with some simple Polonius cases similar to the existing smoke tests, but also for some cases encountered in the wild (in the `rand` crate for example).
A more detailed write-up is available [here](https://hackmd.io/CjYB0fs4Q9CweyeTdKWyEg?view) with an explanation for each test failure, the steps taken to resolve it (as a commit in the current PR), NLL and Polonius outputs (and diff), etc.
Since they've worked on this before, and we've discussed some of these failures together:
r? @matthewjasper
When mentioning lifetimes, only invert wording between the expected trait and the self type when the self type has the vid.
This way, the lifetimes always stay close to the self type or trait ref that actually contains them.
