Otherwise if the member lookup gets simplified immediately and we
have a recursive dynamic member lookup we will crash since we wouldn't
have introduced the corresponding applicable function constraint.
rdar://164321858
In #65125 (and beyond) `matchTypes`, has logic to attempt to wrap an
incoming parameter in a tuple under certain conditions that might help
with type expansion.
In the case the incoming type was backed by a `var`, it would be wrapped
by an `LValueType` then be subsequently mis-diagnosed as not-a-tuple.
More details in #85924 , this this is also the cause of (and fix for)
#85837 as well...
Impact for an unknown property access was frequently higher than other options
on ambiguous selections, by 3 to 5 points, causing fix selections that were
farther away and frequently noted to be in accurate. This commit lowers the
impact to be in a similar range to other fixes and this causes property accesses
to be selected more proprotionaly.
In the existing test suite, this changed the diagnostic only in the case of
protocol composition, which was also discovered to be a flawed binding lookup.
Tests added for the property lookup, tests updated for protocol composition
(Including correcting a likely error in a test specification)
We noticed that a project failed to build with prepared overloads enabled,
but built successfully with prepared overloads disabled.
It turns out that the code clearly should never have been accepted in the
first place, because the type checker was making an arbitrary choice between
two nominal type declarations with the same name.
Further inspection revealed this was because of a FIXME added in 2013 which
was far too broad. Tighten up the logic here to only disambiguate if at least
one of the two declarations is a type alias, and it has the same underlying
type as the other one. A choice between unrelated nominal type declarations
should always be ambiguous.
This still isn't perfect, because we might have two generic type aliases
that are referenced in both solutions with different substitution maps,
in which case we will still erroneously pick the first one. But this
is better than the old logic, at least.
Fixes rdar://165863775.
This fixes an ambiguity introduced by the stdlib change in
0f99458900.
Since (borrowing T) -> () and (T) -> () both convert to
each other, we could end up with ambiguous solutions where
neither one was better than the other. Generalize the
existing trick we use for labeled vs unlabeled tuples to
also strip off ownership specifiers and @convention(...)
from function types. This fixes the regression, as well
an existing FIXME in a test I added a while ago where
the same problem arises with @convention(block).
We already had bookkeeping to track which statement in a multi-statement
closure we were looking at, but this was only used for the 'reasonable time'
diagnostic in the case that we hit the expression timer, which was almost
never hit, and is now off by default. The scope, memory, and trial limits
couldn't use this information, so they would always diagnose the entire
target being type checked.
Move it up from ExpressionTimer to ConstraintSystem, so that we get the
right source location there too. Also, factor out some code duplication
in BuilderTransform to ensure we get the same benefit for result builders
applied to function bodies too.
Handle the following situation:
```swift
struct S {
func test() {}
static func test(_: S) {}
}
```
Calling `S.test(s)` where `s` has a type `S` without any other context
should prefer a complete call to a static member over a partial
application of an instance once based on the choice of the base type.
The behavior is consistent for double-applies as well i.e.
`S.test(s)()` if static method produced a function type it would be
preferred.
Resolves: rdar://165862285
We know this is where the issue is so we can immediately bind to a hole,
ensuring we don't produce unnecessary downstream diagnostics from
things we can't infer.
`??` operator is overloaded on optionality of its result. When the
first argument matches exactly, the ranking is going to be skewed
towards selecting an overload choice that returns a non-optional type.
This is not always correct i.e. when operator is involved in optional
chaining. To avoid producing an incorrect favoring, let's skip the this
disjunction when constraints associated with result type indicate
that it should be optional.
Simply adding it as a binding won't work because if the second argument
is non-optional the overload that returns `T?` would still have a lower
score.
Resolves: rdar://164201746
Missed this in my original patch, handle pack parameters the same as
regular generic parameters, ensuring we don't prefer a pack overload
over a generic overload just because there are pointer conversions
involved. Note this doesn't fix the wider issue of rdar://122011759,
I'm planning on looking into that in a follow-up.
This is a fix for the ported "calls with a single unlabeled argument"
hack. If overload doesn't match context on async effect, let's not favor
it because that is more important than defaulted parameters.
Resolves: rdar://164269641
I missed upgrading this to an error for Swift 6 mode, let's upgrade it
to an error for a future language mode. It's important we reject these
cases since we're otherwise allowing subtyping to be a weaker constraint
than conversion.
Update special favoring logic for unlabeled unary calls to support
non-overloads member references in argument positions.
The original hack missed a case where a type of a member is known
in advance (i.e. a property without overloads) because there as
another hack (shrink) for that.
This helps in situations like `Double(x)` where `x` is a property
of some type that is referenced using an implicit `self.` injected
by the compiler.
Resolves: rdar://161419917
This works around the fact that existential opening does not currently
work correctly in cases where the argument isn't resolved before the
applicable fn is solved.
Add the application constraint before the member constraint, which
allows us to get rid of the special-cased delaying logic for dynamic
member subscripts. The diagnostic change here is due to the fact that
we no longer have a simplified type for the result in CSGen, which
would also be the case if we had a disjunction for the member.
We ought to consider outright banning these conversions if the
destination is a generic parameter type, but for now let's penalize
them such that we don't end up with ambiguities if you're doing an
implicit pointer conversion through an identity generic function.
rdar://161205293
The pack expansion type variable may be a nested in the fixed type of
another type variable, and as such we unfortunately need to fully
`simplifyType` here.
rdar://162545380
This appears to only support synthesized conformances because
operators in such cases use different names - `__derived_*`.
These days devirtualization like this is performed as part of
mandatory inlining. And this approach doesn't stack well with
features like `MemberImportVisibility` because there is change
to check whether witness is available or not.
This helps avoid producing more downstream errors. This changes
`GenericSignature::forInvalid` to produce the same signature as e.g
`<T where T == Undefined>`. This subsumes the need to introduce
conformance requirements for invertible protocols.
Dependent members cannot be simplified if base type contains unresolved
pack expansion type variables because they don't give enough information
to substitution logic to form a correct type. For example:
```
protocol P { associatedtype V }
struct S<each T> : P { typealias V = (repeat (each T)?) }
```
If pack expansion is represented as `$T1` and its pattern is `$T2`, a
reference to `V` would get a type `S<Pack{$T}>.V` and simplified version
would be `Optional<Pack{$T1}>` instead of `Pack{repeat Optional<$T2>}`
because `$T1` is treated as a substitution for `each T` until bound.
Resolves: rdar://161207705
This adds the -verify-ignore-unrelated flag. When -verify is used without -verify-ignore-unrelated, diagnostics emitted in buffers other than the main file and those passed with -verify-additional-file (except diagnostics emitted at <unknown>:0) will now result in an error. They were previously ignored. The old behaviour is still available as opt-in using -verify-ignore-unrelated, but by being strict by default it should make it harder to accidentally miss diagnostics.
To avoid unnecessary performance overhead, -verify-additional-file is still required to parse the expected-* directives in files other than the main file.
