This is basically a workaround for a slightly different IR generated by
fir2ir vs psi2ir. Simplified, psi2ir generates something like this for
the sample from KT-59218:
TRY type=Unit
try: BLOCK type=Unit
VAR methodHandle [...]
TYPE_OP type=Unit origin=IMPLICIT_COERCION_TO_UNIT
CALL invokeExact [...]
While fir2ir generates the following:
TYPE_OP type=Unit origin=IMPLICIT_COERCION_TO_UNIT
TRY type=Any?
try: BLOCK type=Any?
VAR methodHandle [...]
CALL invokeExact [...]
The lowering relies on the fact that a polymorphic call (`invokeExact`
in this case) is a direct argument to the TYPE_OP, to determine the
correct return type (Unit in this case) to be generated in the bytecode.
The solution here is to push the type coercion "through" all the
block-like structures (`try`, `when`, container expression) so that if
the last statement in the block is a polymorphic call, it gets properly
converted even if the whole block is under a type coercion operation, as
it happens in fir2ir. We achieve that by using the "data" parameter of
the IR transformer: appropriate immediate children of
IrTypeOperatorCall/IrTry/IrWhen/IrContainerExpression get the type that
the expression needs to be coerced to, and all the other expressions
ignore that type and set it to null when transforming their children.
A proper solution would be to ensure fir2ir generates exactly the same
IR as psi2ir (KT-59781), but since PolymorphicSignatureLowering is the
only lowering affected so far, and polymorphic calls occur very rarely,
it seems safe to workaround it in the lowering for now.
#KT-59218 Fixed
This is important for IR lowerings like PolymorphicSignatureLowering
which are very sensitive about the correct types of expressions and
placement of coercions to Unit (KT-59218).
A boolean parameter to `insertImplicitCasts` is not the best solution to
ensure that coercion to Unit is added. The best solution would be to fix
the TODO and generate coercion to the block's type for the last
statement. But that will affect many other places and will need to be
done separately => KT-59781.
Code in IrInterpreter is uncommented to fix the FIR test
`compiler/testData/ir/interpreter/exceptions/tryFinally.kt`; otherwise
evaluation of the function `returnTryFinally` there crashes with
"NoSuchElementException: ArrayDeque is empty". No idea why this test
didn't fail for K1 though, since the created IR is exactly the same.
For some unknown reason this breaks WASM backend with K2, but not with
K1 => KT-59800.
Before this change, it could happen that `when` of type Unit has a
branch whose type is not Unit. This can lead to problems in IR
lowerings, for example PolymorphicSignatureLowering which is very
reliant on the correct types of expressions and placement of coercions
to Unit (KT-59218).
Otherwise, according to the spec, a type will be treated as a final,
which is not correct in many cases.
Later, we can generate final struct types when it's applicable.
#KT-60200 Fixed
In REPL it's possible that builtins will be created several times with
the same symbol table. And since IrBuiltinsPackageFragmentDescriptorImpl
has overridden equals, symbol for external package will be the same for
different descriptors (with same FQN). So we should create IrExternalPackageFragment
here only if it was not created before, on previous compilation
After this change SymbolTable (and ReferenceSymbolTable) contains only
methods with IdSignatures. All descriptors-related methods are moved
into DescriptorSymbolTableExtension, which automatically delegates to
the SymbolTable if needed
At this moment there are cross-references between SymbolTable, because
descriptor API is still actively used across backends. So SymbolTable
is accessible in some place then descriptor extension will be accessible
too
DescriptorSymbolTableExtension is an implementation of abstract SymbolTableExtension
which allows to implement different kinds of storages, e.g. FIR based
(it probably will be needed for FIR2IR)
For interfaces with custom serializer (@Serializer(`SerializerType::class)`) now the `serializer()` function is generated in the companion so that user can get a serializer by type
Merge-request: KT-MR-11040
Merged-by: Sergey Shanshin <Sergey.Shanshin@jetbrains.com>
Use it in diagnostic tests with JVM backend instead of the
RENDER_DIAGNOSTICS_FULL_TEXT directive.
This is needed because otherwise in
AbstractDiagnosticsTestWithJvmBackend there are two handlers trying to
dump full diagnostic texts: JvmBackendDiagnosticsHandler and
FirDiagnosticsHandler, which dump different diagnostics (reported by
backend vs frontend) and in slightly different formats.
In fact, this is why exceptionFromInterpreter.fir.diag.txt was added in
48484368c7: the existing .diag.txt was detected as incorrect by
FirDiagnosticsHandler (even though the diagnostics are the same, just
printed in a different way), which led to the creation of .fir.diag.txt.
For these tests, the behavior about checking diagnostic text in
FirDiagnosticsHandler is useless because it doesn't include backend
diagnostics. So we disable it by using another directive.
Remove the "_ir" suffix. It was originally used to distinguish tests on
the JVM IR backend from the old JVM backend ("_old"), but the latter are
removed now.
Also take into account diagnostics reported via
GenerationState.collectedExtraJvmDiagnostics, and render them in the
similar format to KtDiagnostic.
Don't make the test pass if the reported list of diagnostics is empty,
because that defeats the purpose of the test.
`FragmentLocalFunctionPatchLowering.visitSimpleFunction`
uses a transformer with an overridden `visitCall`.
Inside it visits every nested call recursively in two places:
in the beginning (1, line 59), and via super method
if the current call doesn't refer to a local function (2, line 64).
It leads to a combinatorial explosion: if there are N nested calls
in one expression, then the number of `visitCall` invocations
will be `pow(2, N)`.
Thus, it will take forever to compile certain expressions. A perfect
example of a long nested call is string interpolation: string templates
are being compiled to a chain of `StringBuilder::append` calls. Here's
why IDEA-325225 happens -- the code in the issue has large string
templates, and the resulting `StringBuilder::append` chains may contain
up to 30 links, making for around 1 billion `visitCall` invocations.
However, the visiting (2) is obsolete. First, the visiting (1) will
process all the children of an expression, replacing them with the
right local function calls, if necessary, recursively, up the stack.
Second, the expression itself will be transformed, if necessary, in
that same method, after all its children are processed. That means we
can get rid of one visiting and avoid the combinatorial explosion, yet
all the local function calls will still be correctly processed.
Fixes IDEA-325225, contributes to IDEA-325146
Merge-request: KT-MR-11104
Merged-by: Alexander Kuznetsov <Aleksander.Kuznetsov@jetbrains.com>
Ilya Kirillov