When a suspend lambda does not capture crossinline lambda, it is
generated with as state-machine, since it does not inline anything.
However, when regenerating, the inliner used to remove all DebugMetadata
annotations to avoid duplication. This lead to missing annotation if
the lambda is regenerated, but state-machine is not regenerated.
This change fixes the blind spot by readding the annotation after
regeneration.
#KT-41789 Fixed
The change is a bit complicated because the name mapping logic is
refactored so that we would compute the needed suffix first (either
module name, or multifile part class name), and then shove it into the
mangled name between the original function name and "$default", if the
function in question is the default arguments adapter.
The main motivation for this change was to fix KT-41809, but as a side
effect, private functions with default arguments in multifile parts now
also contain the file name, just like normal private functions.
#KT-41809 Fixed
Note that only irrelevantStaticProperty.kt failed before this change.
Having private declarations caused no problems, but it seems incorrect,
so it's fixed now and irrelevantPrivateDeclarations is added just in
case.
#KT-41848 Fixed
The problem here was that although the IR type of the expression was
primitive, the type of the actual expression in the bytecode generated
after type erasure was `Ljava/lang/Object;`, and we were trying to call
a non-existing method `Object.hashCode(Object)`.
#KT-41669 Fixed
Before this change, we stored the enclosing constructor in a map in
JvmBackendContext before moving lambdas and local classes out of
initializer blocks. However, in case the lambda was declared in an inner
class, we stored a reference to the unlowered constructor of the inner
class, whose JVM signature is "()V" instead of the correct "(LOuter;)V".
Java reflection then threw exception if we tried to call
`getEnclosingConstructor()` on such class at runtime. Proguard finished
with errors for the same reason.
It turns out that we can just store the fact that the class has been
moved, and load the matching constructor in codegen, where everything is
already lowered and guaranteed to match with the signatures of the
actual generated declarations.
#KT-41668 Fixed
There was a typo in JvmLocalClassPopupLowering which allowed the
EnclosingMethod for lambdas and anonymous classes in initializers to
become any function in a class, in case when there was no primary
constructor in that class. E.g. in the added test, `getIrrelevantField`
was the EnclosingMethod of the lambda class before this change.
In box tests, only check that Java reflection does not crash on the
EnclosingMethod attribute generated in these classes. If it doesn't
crash, most likely it returns the value that can be read from the class
file by ASM, which is what the newly added bytecode listing tests are
checking now.
Raw type Q is represented as a flexible type
Q<B1, ... Bn> .. Q<*, ... *>
where Bi is a representative upper bound of the corresponding ith type
parameter of Q.
When mapping generic signature, JVM takes type arguments of lower bound
(which is 'Q<B1, ..., Bn>').
There is still some difference in how JVM and JVM_IR handle raw type in
signature. It requires additional investigation.
Ilmir Usmanov