Also make TypeParameterUpperBounds a "strong" incompatibility, meaning
that non-actual members from platform module are _not_ going to be
matched to the expected members if this incompatibility exists between
them, and therefore NO_ACTUAL_FOR_EXPECT will be reported on the
expected declaration, instead of ACTUAL_MISSING on the platform member.
This is needed because the difference in type parameter upper bounds can
have effect on the function signature on the platform (e.g. on JVM,
Array<T> -> T[], but Array<T> -> Comparable[] if T : Comparable<T>), and
it would be incorrect to report ACTUAL_MISSING on the member that has
nothing to do with the expected declaration that happens to coincide
with it in everything except type parameter bounds
#KT-21864 Fixed
In case the referred type is actually an enum that is not found in
dependencies due to a configuration problem, this usage could be valid.
So we can avoid reporting an error here, to reduce the number of
diagnostics.
Also do not report "default value of annotation parameter must be a
compile-time constant" in the same case for the same reason
Consider following case:
fun foo(): Unit = run { "hello" }
Previously, NI would analyze lambda body without expected type, because
it is a type variable 'R' from 'run', which hasn't been fixed yet. This
leads to treating "hello" as lambda-return argument and adding bogus
'String' constraint on 'R', and, consequently, type mismatch.
Now, we peek into current constraint system and check if return-type of
lambda is type variable with upper-Unit constraint (which is exactly
condition for its body to be Unit-coerced). If so, then we provide
expected Unit-type for body explicitly, and the rest will be done
automatically (in particular, in aforementioned example "hello" wouldn't
be treated as lambda return argument).
This commit introduces proper handling of recursion in scopes, which
could occur when some of companion object supertypes are members of
that companion owner:
```
class Container {
open class Base
companion object : Base()
}
```
To resolve `Base`, we have to build member scope for `Container`.
In the member scope of `Container`, we see all classifiers from
companion and his supertypes
So, we have to resolve companion objects supertype, which happens to be
`Base` again - therefore, we encounter recursion here.
Previously, we created `ThrowingLexicalScope` for such recursive calls,
but didn't checked for loop explicitly, which lead to a wide variety of
bugs (see https://jetbrains.quip.com/dc5aABhZoaQY and KT-10532).
To report such cyclic declarations properly, we first change
`ThrowingLexicalScope` to `ErrorLexicalScope` -- the main difference is
that latter doesn't throws ISE when someone tries to resolve type in it,
allowing us to report error instead of crashing with exception.
Then, we add additional fake edge in supertypes graph (from
host-class to companion object) which allows us to piggyback on existing
supertypes loops detection mechanism, and report such cycles for user.
The issue here is that OI infers correct types for (k, v) destructing
declaration, while NI infers errors for them. That happens because NI
resolves iterator() on nullable 'm', but rightfully reports it as
unsuccessful, while OI somehow manages to resolve it to success (and
thus getting nice expected type, allowing destructing declaration to be
resolved in a proper types).
Previously, there was receiver of type Nothing, which could case result
of the call to be inferred to Nothing too. This could case bogus
UNREACHABLE_CODE diagnostics in cases like this:
```
fun <T> id(x: T) = x
fun test() {
id(unresolvedReference) // type of statement is 'Nothing'
// ... everything here is marked as unreachable ...
}
```
This commit changes type of receiver for such calls form Nothing to
ErrorType.
This commits introduces testdata changes, where NI behaviour strictly
improved, after several previous fixes.
For some tests, just WITH_NEW_INFERENCE directive was added. It
indicates, that some of previous commits first introduced error in that
test, and then some other commit fixed it (netting no overall testdata
change). It is preferrably to keep those annotations until we will
migrate to NI completely, to prevent unexpected regressions.
This test introduces very special (for current implementation) case,
when we have smartcast indirectly, via some reified type parameter.
It covers recursive call inSmartCastManager.checkAndRecordPossibleCast(),
which wasn't previously covered by any test in testbase.
Expect members should always lose in resolution to non-expect members,
be it simple calls or callable references. Note that there should be
exactly one actual member for each expect member in correct code, so
both ways to check for expect vs non-expect are correct: either before
signature comparison, or after.
#KT-20903 Fixed
The problem is that when performing full analysis we do it in
a backward order while result for trivial vals is filled
in a forward one.
It turns out that reversedInstuctions might return a superset of
forward traversed instructions, e.g. in case of dead code in lambda.
At the same time result for trivial vals is constant
for any instruction, thus we can just return its constant value
and use it in the full analysis
#KT-20895 Fixed
From Kotlin's point of view, everything in annotation classes is
non-abstract. A class inheriting from an annotation has a non-abstract
fake override for each property of the annotation class constructor. But
because members of annotation classes themselves were considered as
abstract in the bridge-generating code (see
DescriptorBasedFunctionHandle.isAbstract), there was a situation where a
concrete fake override has only one declaration among overridden
descriptors and it was abstract. This situation is invalid (a concrete
fake override must have exactly one concrete super-declaration),
therefore an exception was thrown.
The fix is to avoid considering annotation class members abstract for
the purposes of bridge generation. It's reasonably safe because no
bridges should be ever generated for annotation subclasses anyway,
because annotations can only have members with simple return types
(final and non-generic).
Note that in KT-19928, the problem is reproducible because of an
incorrect "inexact analysis" in light classes where "Target" is resolved
to an annotation class kotlin.annotation.Target. This behavior of the
analysis in light classes seems to do no harm otherwise, so it's not a
goal of this commit to change anything in that regard
#KT-19928 Fixed
When recursion is detected while computing
`ClassResolutionScopesSupport.scopeForMemberDeclarationResolution`,
create 'ThrowingLexicalScope' (as with other scopes), instead of
throwing ISE immediately. That allows to report error properly in cases
like in KT-18514
#KT-18514 Fixed
Note that current behaviour is made similar to the case with
properties initializers/accessors, which means that more complex
cases are not covered yet (see KT-20801) #KT-20802 fixed.
The reason is that before dc02b2e3ab and 8a0dcca957,
TypeConstructor.isFinal for some class descriptors
(DeserializedClassDescriptor, LazyJavaClassDescriptor,
MutableClassDescriptor) were implemented as `isFinalClass` (which is
`modality == FINAL && kind != ENUM_CLASS`), and all others as
`modality == FINAL` or simply true/false. This led to differences in
behavior depending on the exact instance of the class descriptor.
Now that TypeConstructor.isFinal is always `modality == FINAL`, some
tests (PseudoValueTestGenerated) fail because the finality of some type
constructors changed and these tests render final vs non-final type
constructors differently.
In this commit, TypeConstructor.isFinal is now made to behave safer,
i.e. considering enum class type constructor to be non-final (as was the
case earlier for some ClassDescriptor instances). Some diagnostics might
disappear (e.g. FINAL_UPPER_BOUND) but it doesn't look like a big deal
Alexander Udalov