FIR simply ignores anything after the first comma if the when expression
doesn't have a subject. Hence, the checker has to rely on PSI structure
instead.
In try blocks, last call won't be completed when building node for it
This is workaround to partially reconstruct nothing stub node for such
calls. Should work for non-local returns in try only.
#KT-48160 Fixed
This change touches the following diagnostics to make them behave closer
to FE1.0
* SUPER_NOT_AVAILABLE
* SUPER_IS_NOT_AN_EXPRESSION
* INSTANCE_ACCESS_BEFORE_SUPER_CALL
* NOT_A_SUPERTYPE
Other than tweaking the diagnostics, this change also alters resolution
by consider marking `super` with mismatched type parameter as
errorenous. As a result, the following code no longer resolves.
```
class A: B() {
fun test() {
super<String>.length
// ^^^^^^ FIR currently resolves this to `String.length`.
// With this change, `length` becomes unresolved
// instead
}
}
```
Also, now we report `UNRESOLVED_LABEL` on unresolved label on `super`
reference, though FE1.0 reports `UNRESOLVED_REFERENCE`.
All the errors above are reported as ConeDiagnostics and hence some
checkers are deleted.
In addition, it also suppresses more downstream (mostly unresolved)
errors if the receiver has errors. FE1.0 doesn't do it for all the cases
we have here. But it seems nicer to reduce these "redundant" unresolved
errors.
First, the order of arguments is swapped.
Second, projection erasure for argType was too aggressive. We should
instead retain the projections that are under an invariant position.
FE1.0 only reports SENSELESS_COMPARISON if one of the operand is `null`.
This change makes FIR reports also in case one of the operand has type
`Nothing?`.
In addition, fix handling of type alias in ConeTypeContext#isNullableType
In order to make resolution still work for members not available from
`Nothing`, we track the type without `Nothing?` and use that for
resolution instead.
The logic should clear back aliases as well. To ensure all back aliases
don't lose any information, statements on the original variable are
copied over to its aliases.
Previously, if user compares an `it<String, Int>` with `String`, the
checker reports it since the flattened types `[String, Int, Int]` are
incompatible. But technically, before flattening, the intersection type
actually contain the other side, so they should really be compatible.
1. throw goes to catches instead of main exist block
2. return goes via finally (single level only supported atm)
3. collect non-direct return to retrieve all return expressions easier
Previously, the checker executed for every declaration i.e. every
declaration was considered as expect declaration. Because of that in
some cases this checker could eat 6% of compilation time.
After this commit only declarations marked with expect or actual
are checked. To achieve that, logic, that do reporting about missing
actual modifier was moved to the Actual part.
Please note, that in cases where there is no expect/actual modifier at
all other errors (like redeclaration and missing body on "actual"
declaration) would be reported.
Useful nodes:
- In this checker reportOn is always the same as
descriptor.sourceElement.ktElement. This is because the only case when
it isn't true is PropertyAccessors and they are filtered
- Annotation constructor descriptor isActual == true all the time
- previously for weak incompatible members ACTUAL_MISSING
was not reported
- the logic here is super complicated and crazy, but I don't think that
there is sense to refactor it in the old FE
Andrey Zinovyev