As part of this change, we also extend the usage of RealVariable in more
places during DFA. Now mutable properties, property with custom getters,
delegated properties, etc are also treatd as a `RealVariable`. In
general this is needed in order to carry out smartcast computation in
order to report `SMARTCAST_IMPOSSIBLE`. It seems to also have side
effects that improves behavior of some test files.
This new kind of expression encompasses the nullability of the original
expression after null check (or equivalent `is Nothing?` check).
Unlike FirExpressionWithSmartcast, this expression won't be materialized
during conversion to backend IR. Also, Nothing? is discarded when
computing the intersection of possible types from smartcast info.
In that way, Nothing? is not used during resolution, while such
smartcast info is stored in it (and the expression kind itself).
Added checker for FirEqualityOperatorCall. It's surfaced as one of the
following diagnostics depending on the PSI structure and types under
comparison:
* INCOMPATIBLE_TYPES(_WARNING)
* EQUALITY_NOT_APPLICABLE(_WARNING)
* INCOMPATIBLE_ENUM_COMPARISON_ERROR
Comparing with FE1.0, the current implementation is more conservative
and only highlights error if the types are known to follow certain
contracts with `equals` method. Otherwise, the checker reports warnings
instead.
However, the current checker is more strict in the following situations:
1. it now rejects incompatible enum types like `Enum<E1>` and
`Enum<E2>`, which was previously accepted
2. it now rejects incompatible class types like `Class<String>` and
`Class<Int>`, which was previously accepted
3. the check now takes smart cast into consideration, so
`if (x is String) x == 3` is now rejected
Previously if an unsafe call is to an overloaded function, FIR checkers
report NONE_APPLICABLE. This change instead report them as UNSAFE_CALL
or its variants.
Previously unsafe call is reported as part of InapplicableWrongReceiver.
This makes it difficult for the downstream checkers to report different
diagnostics.
Specifically, the report the following 4 errors.
* NON_VARARG_SPREAD
* ARGUMENT_PASSED_TWICE
* TOO_MANY_ARGUMENTS
* NO_VALUE_FOR_PARAMETER
Also added/updated the following position strategies.
* NAME_OF_NAMED_ARGUMENT
* VALUE_ARGUMENTS
Currently, FIR reports errors caused by previous resolution failure. For
example with unresolved `a` and `b` in code `a.b`, both `a` and `b` are
highlighted. FE1.0 only highlights `a` since it's the root cause. This
change applies this heuristics when reporting FirDiagnostics.
Currently if there is an error in a function call, FIR would report the
entire expression if this call is qualified, but *only* the name if it's
not qualified. For example, assume the following two calls are all
contains some errors.
```
a.foo(1,2,3)
^^^^^^^^^^^^
bar(1,2,3)
^^^
```
The entire call of `foo` is reported since it's qualified. But only the
reference `bar` is reported since it's not qualified. This limits the
usage of position strategies because the IDE does not allow position
strategies to go outside of the initially reported PSI element
(org.jetbrains.kotlin.idea.fir.highlighter.KotlinHighLevelDiagnosticHighlightingPass#addDiagnostic).
This change passes both the original error named reference and the
surrounding qualified access expression and defer the decision of which
to use to the reporting logic.
For unresolved reference and checks on `super` keyword, the position
strategy should not highlight the surrounding parentheses. Hence a new
position strategy `REFERENCED_NAME_BY_QUALIFIED` is added.
In addition, this change also has the following side effect
* some diagnostics are no longer reported when there is a syntax error
since the higher level structure does not exist when there is a syntax
error
To do so, inside the root cause of inapplicable candidate errors,
we will record expected/actual type of receiver, if any.
That will help identifying inapplicable calls on nullable receiver.
This commit includes a lot of minor changes in testdata: new infrastructure
requires that each test file in one testdata file in one module must
have unique name, but a lot of existing spec tests didn't satisfy
this requirement
Before this commit, we used type after smart cast both for original
FirThisReceiverExpression and for wrapping FirExpressionWithSmartCast.
However, this makes FIR2IR implicit cast generator work incorrectly
(it decides not to insert implicit cast because original type is the same).
After this commit, expressions have different types and implicit cast
generator works properly.
Dmitriy Novozhilov