Treat them like lambdas. This means:
1) Like local suspend lambdas, which become "non-suspend" after creating
jvmSuspendView, they do this as well
2) They both are generated with the same codegen.
3) They are treated differently only on front-end side.
#KT-20364: Fixed
Sometimes instead of {POP, GETSTATIC Unit.INSTANCE, ARETURN} sequence
the codegen emits {CHECKCAST Unit, ARETURN} sequence, which breaks tail
call optimization. By replacing CHECKCAST with ARETURN we eliminate
this issue.
#KT-19790: Fixed
Consider call chain `foo.bar.baz()` which has expected type `Int`
Elements `foo` and `bar` doesn't have expected type `Int`,
this is especially important in NI, because we use expected type for
candidate applicability
Before this change, the codegen used to generate beforeInlineMarker
in order to spill variables later. However, since the tailrec call
is replaced with jump, no afterInlineMarker is generated, leading to
inconsistency of these markers. Thus, the spilling fails.
This change disables generating of beforeInlineMarker.
#KT-21521: Fixed
Unlike ordinary lambdas suspend lambdas has create method which invokes
the constructor of the lambda object (continuation).
The inliner could not cope with this.
The previous change fixed the case of suspend lambda without receiver.
This change adds support of suspend lambdas with receiver.
#KT-21605: Fixed
in redundant locals elimination.
Since IgnoringCopySourceInterpreter ignores ALOADs and ASTOREs,
the source instruction of ALOAD K in {ALOAD N, ASTORE K, ALOAD K}
sequence is not ASTORE K. In this case we cannot simply replace K
with N, since there can be multiple {ALOAD N, ASTORE K} sequences
in separate branches. After replacement we get different stack
frames.
This change resolves this.
However, in ReturnUnitMethodTransformer we want to ignore copies
of the same GETSTATIC kotlin/Unit.INSTANCE, since we do not mess
with local variables and just replace ASTORE with ARETURN to help
tail-call optimization.
#KT-23373: Fixed
Use fake continuation instead of ALOAD 0 while inlining
Do not generate state machine for inner lambdas and inner objects,
which capture crossinline suspend lambda.
#KT-19159: Fixed
Otherwise if Char value is a result of some erased generic call,
coercing it to I (primitive int) causes CCE at run-time.
#KT-23104 Fixed Target versions 1.2.40
Note that there are more places where assertions for inline classes should refined:
- lateinit vars
- values that come from Java
- type casts (interfaces to inline class type)
- Into PrimitiveNumberRangeLiteralRangeValue modifies how bounded
value are created by checking if the high bound range can be modified
to be exclusive instead of inclusive such as the generated code will
be optimized.
- Add black box tests checking that the specialization does not fail
on any kind of arrays.
- Add a bytecode test checking that the code is correctly optimized.
Fix of https://youtrack.jetbrains.com/issue/KT-22334
Probably, when NI is there this fix will become unnecessary because
there will be no approximation applied, thus the value parameter
will remain Hello<#Captured> instead of Nothing
#KT-17171 Fixed
Alexander Udalov