The main reasoning for the module is to avoid running any compiler tests while executing run configuration that searches tests across module dependencies.
Do not call `isError()` on property's type right after creation of
PropertyGetterDescriptor because the property has no getter yet (it's created
but not yet stored to the property, that happens a bit later), and `isError()`
leads to computation of the delegate type, which for delegated properties
performs some complex resolution (see `VariableTypeResolver#process`) which
relies on the fact that the property already has a getter.
Since the purpose of the original change (883e2e4d) was to support a quick fix
which would add the type to a property in an expression like "val x get() =
...", check the type (or initializer) presence in the PSI instead, this is
safer and is still suitable for the quick fix.
Also fix arguments to "wrong getter type" diagnostic: previously something
useless like "expected Int, actual Int" was reported
#KT-11809 Fixed
POP operations are backward-propagated.
Operation can't be transformed if its result is moved within stack
(by DUP, DUP_X1, DUP_X2, DUP2, DUP2_X1, DUP2_X2, or SWAP).
Unnecessary operations are replaced with NOPs (for debugger).
Unnecessary NOPs are removed.
KT-9922 Suboptimal generation for simple safe call with unused return value
KT-11116 Optimize methods returning Unit
Preserve static initialization semantics for parts by introducing a special "clinit trigger" class.
Insert "static initialization trigger" call to every method of a part class, remove this call on inline.
Always mangle names for private functions in multifile class parts to avoid resolution clashes on inheritance.
NB in codegen tests initializers for all non-const vals are wrapped in 'run { ... }',
so that the initializer is not a constant expression, and some static initialization code should be generated.
Although a test is present, it doesn't check the behavior because it seems that
in the test environment both streams are necessarily flushed
#KT-9546 Fixed
Rework backing field generation logic in PropertyCodegen to switch the
ClassBuilder instance for a multifile part to that of the corresponding facade
class. This became needed because multifile parts, and their metadata, are
generated _before_ the multifile facade class and otherwise we would never
record that there's a synthetic '$annotations' method for a const val and would
not write that to the protobuf message for the property.
See also bad83200
#KT-10892 Fixed