Previously, it was led to plainly adding NullableType <: T constraint
which silently led to successful call completion.
What is suggested is just marking such initial constraint
as unsuccessful.
In K1, the error was reported just via additional type checking
mechanism being run after call completion.
^KT-58665 Fixed
When BI/incomplete call is present in return argument, it will be
added to the main call-tree, leading to requirement violation in
`ConstraintSystemCompleter.getOrderedAllTypeVariables`
as after `FirBuilderInferenceSession.inferPostponedVariables` it will
be completed, and its type variables couldn't be found anymore
In K1, we actually do the same, but we are able to find type variables
of such calls due to architecture difference:
In K2, `getOrderedAllTypeVariables` uses FIR as the main source to
lookup
In K1, `getOrderedAllTypeVariables` uses resolution atom tree, where
candidate/CS of the call is still available after
`inferPostponedVariables`
In fact, all incomplete calls were analyzed in FULL mode according to
the contract of `FirBuilderInferenceSession.addPartiallyResolvedCall`.
Thus, it means we normally shouldn't add them to the main call-tree, but
accidentally do it as incomplete calls contain non-completed candidate
This particular commit addresses the problem partially, only
in cases when the expected return type for the lambda is Unit and when
the incomplete call is located in the last expression of the lambda
In such cases, we can skip the call from the last expression completely,
since all potential calls there were analyzed in FULL mode,
and couldn't introduce any useful info to the CS of the main call-tree
^KT-54294
Previously, when a candidate was found with an applicability that is
better than the current best applicability, all previous candidates were
thrown away. Now we keep them, unless the new applicability is
successful. If no successful candidates are found, we fully resolve all
the unsuccessful ones and select the ones with the least bad
applicability. This improves diagnostics for unresolved calls.
#KT-57844 Fixed
Before this commit, for property candidates in K2 their types wasn't
inferred/susbtituted properly.
So, when candidate for fooBar.liveLoaded.invoke() was created,
the type of `fooBar.liveLoaded` was just X type parameter for which
there is no any `bar()` functions in its member scope.
While proposed semantics is a bit different from K1, where
both property and invoke candidates are united into common system,
it doesn't contradict to the specification (https://kotlinlang.org/spec/overload-resolution.html#callables-and-invoke-convention)
which says explicitly that invoke-convention should be desugared as
`r.foo.invoke()`, thus `r.foo` should be completed independently.
Also, this strategy supports some reasonable use-cases like KT-58259
while it's still a breaking change but for more artificial-looking
situations (see KT-58260) and should be passed through
the language committee.
The changes in stubTypeReceiverRestriction* tests looks consistent
because of how `genericLambda` now works
(with full completion of property call).
NB: The code is going to be red once KT-54667 is fixed and also there's
already similar diagnostic in K1 (INFERRED_INTO_DECLARED_UPPER_BOUNDS)
^KT-58142 Fixed
^KT-58259 Fixed
^KT-58260 Related
There's a heuristic for approximation of a captured type that once
it has non-trivial lower bound (other than Nothing), it's worth
approximating it to sub-type even while the containing
top-level type is being approximated to super-type.
And that sounds reasonable in case the lower bound is indeed non-trivial,
but that's not the case because nullability here comes from
the nullability of captured type position.
So, the fix is basically not to treat such approximations as non-trivial.
And while that seems to be a bit of a change in the language semantics,
it still looks reasonable (see other changes in test data and KT-58087)
^KT-57958 Fixed
^KT-58087 Fixed
While plainly repeating K1 behavior might be a dumb solution,
but inventing another one might be quite complicated
because we need to stop fixing variables into Nothing in many cases,
but probably not in all of them (see KT-58232).
^KT-58149 Fixed
^KT-58232 Related
The change is needed for the parallel resolution (^KT-55750), so we can resolve the declaration
under a lock that is specific to this declaration.
Previously, if LL FIR was resolving some FirClass, LL FIR resolved all its children too, and it had no control over what parts of the FIR tree were modified.
The same applied to the designation path, sometimes the classes on the designation path
might be unexpectedly (and without lock) modified.
This commit introduces LLFirResolveTarget, which specifies which exact declarations should be resolved during the lazy resolution of the declaration.
All elements outside the declarations specified for resolve in LLFirResolveTarget, should not be modified.
The logic of lazy transformers is the following:
- Go to target declaration collecting all scopes from the file and containing classes
- Resolve only declarations that are specified by the LLFirResolveTarget, performing the resolve under a separate lock for each declaration
^KT-56543
^KT-57619 Fixed
Result of the `checkNotNull` calls should always be a non-nullable
values.
The simplest idea how to acheive it is adding not-nullable Any bound
to the type parameter declaration.
Existing comment stating about impossibility of such bound seems to be
not 100% correct because it doesn't take into account presence of
definitely-non-nullable X & Any types that allow described case with
nullable generic.
^KT-55804 Fixed
fix checking for intersection types in CheckIncompatibleTypeVariableUpperBounds
we need this resolve, because getEmptyIntersectionTypeKind under
the hood uses org.jetbrains.kotlin.resolve.checkers.EmptyIntersectionTypeChecker.computeEmptyIntersectionEmptiness
which uses computeByHavingCommonSubtype where we have
isFinalClassConstructor() call, so we need to resolve to STATUS phase
to get the correct modality
^KT-56543
The compiler should only report diagnostics for
comparisons over builtins and identity-less types,
other incompatibilities should be reported
via inspections.
It's ok that in `equalityChecksOnIntegerTypes`
instead of `EQUALITY_NOT_APPLICABLE_WARNING` we get
`EQUALITY_NOT_APPLICABLE`, because
`ProperEqualityChecksInBuilderInferenceCalls`
is already active by default.
This change also replaces the notion of a representative superclass
with the least upper bound.
This makes complex types like
intersection/flexible transparent to
RULES1-based compatibility checks.
One way to look at it is to think
that this is an automatic way of handling
type parameters: automatic picking of
"interesting" bounds, and checking them against one another.
Note that `TypeIntersector.intersectTypes`
for `Int` and `T` where `T` is a type parameter
may return both `{Int & T}` or `null`
depending on `T`-s bounds. At the same time,
for type parameters `T` and `K` it will
always return `{T & K}`.
`ConeTypeIntersector.intersectTypes`, on the
other hand, will always return `{Int & T}`
irrespectively of the bounds. Meaning, the two
intersectors differ in corner cases.
`lowerBoundIfFlexible` call in `isLiterallyTypeParameter` is backed by
the `equalityOfFlexibleTypeParameters` test.
^KT-35134 #fixed-in-k2
^KT-22499 #fixed-in-k2
^KT-46383 #fixed-in-k2
The only case when behavior is change is described at
computeNonTrivialTypeArgumentForScopeSubstitutor
The idea is to avoid depending on the presence of @UnsafeVariance
and instead approximate captured types in covariant argument positions
before building substitution scopes
It's correct because for Captured(*) <: Supertype,
Out<Captured(*)> <: Out<Supertype> and when we've got @UnsafeVariance
value parameters at Out, it's ok to allow passing Supertype there.
^KT-57602 Fixed
^KT-54894 Fixed
Note: here we set sinceVersion = null for this feature.
However, we plan in 1.9.* - 2.0 time frame to solve KT-56377
and to enable this feature in 2.0, the latest in 2.1.
This change allows to revert adding `WITH_STDLIB` directive
to tests which happened at `a9343aeb`.
Co-authored-by: Alexander Udalov <Alexander.Udalov@jetbrains.com>
The job is already done at CheckExtensionReceiver resolution stage
And repeating it might only lead to incorrect errors caused by
double-capture of receiver type that leads to contradiction because
in previous commit we started assuming different capture instantiations
as different types.
This is mostly a revert of 2f61a2f56f
There, we erroneously assumed that we may take captured types as equal
if they are based on the same-typed projections.
Each instance of capturing defines its own captured type,
that should not be equal to any other type captured in other place.
Initial motivation was brought by FP Ultimate, where a piece of code
from the new test was found that started working differently after
recent changes.
The most obvious consequence is the change in addAllProjection.fir.kt:
one cannot use an instance as an argument when expected type
is captured type based on the same instance.
Otherwise, it would lead to CCE if we allowed to put arbitrary charsequences
to the list that initially was a MutableList<String>
All other test data changes (but addAllProjection.fir.kt and differentCapturedTypes.kt)
are irrelevant and will be fixed in the subsequent commits
Namely, do not choose `Nothing?` result type when fixing a variable
that has other constraints besides the ones that came from
the relevant type parameter's upper bounds.
See more details in KT-55691.
In K1, the case from specialCallWithMaterializeAndExpectedType.kt
was working (inferred to String?) just because the branches
were analyzed independently with `String?` expected type.
This change became necessary after the previous commit when we united
inference subsystems for if/when branches (see motivation there).
NB: For K1, the behavior is left the same, but the code
was refactored a bit.
^KT-55691 Fixed
^KT-56448 Fixed
Otherwise, it leads to branches inference run fully independent,
while there are cases when it's necessary to flow type information from
one of the branch to another (see the new test).
NB. In K1, it worked differently: if branches were inferred altogether
only for Any/Any? expect types (otherwise they're analyzed independently)
See foo2/foo4 in the test.
To avoid breaking change we need to support foo1/foo3, but we're trying
not to have some special rule for Any, so we've got a new resolution mode
that provides expect type, but doesn't require full completion.
^KT-45989 Fixed
^KT-56563 Fixed
^KT-54709 Related
For change in specialCallWithMaterializeAndExpectedType.kt
At first, see at KT-36776
Long time ago, it's been decided that if/when resolution
should look similar to similar "select()" calls,
but it's a breaking change (see KT-36776), and we were ready for that back then.
But then, there were too many broken cases found, thus we reverted it at
100a6f70ca
But probably, it would be better to try to infer `String?`
instead of `Nothing?` (see next commits)
Note that change in specialCallWithMaterializeAndExpectedType.kt
will be addressed in later commits, too
When expected type is known, use it as expected type for branch bodies.
While it indeed becomes different from the usual select call resolution,
where expected type is applied only after completion starts,
it helps to support, e.g. callable references resolution just as powerful
as it was in K1.
Also, in some cases where diagnostics have been changed, they become
a bit more helpful since they are reported closer
to the problematic places
cannotCastToFunction.kt test has been removed because it relied
on the case erroneously supported by the hack removed from
the FirCallResolver in this commit.
^KT-45989 Fixed
^KT-55936 Fixed
^KT-56445 Fixed
^KT-54709 Related
^KT-55931 Related
Some of the changed tests may duplicate other existing diagnostics,
but that should not be reason not to report them at all.
There might be another job to be done to avoid diagnostic duplications
For example, NEW_INFERENCE_NO_INFORMATION_FOR_PARAMETER
It became especially relevant after 0e84bf2053
that together with later commits bring a lot of unnecessary
NEW_INFERENCE_NO_INFORMATION_FOR_PARAMETER diagnostic
Apparently if you add an empty line at the start of the FIR file, that's
not enough of a difference for the test suite to complain about, but
enough for it to not add the FIR_IDENTICAL directive...
While it is theoretically useful to know that `{ while(true) {} }`
returns Nothing, CFG node deadness is not precise enough to do that: if
the entire lambda is dead, it's no longer possible to find out whether
the loop is terminating. Besides, `while (true)` and `if (true)` are
pretty much the only constructs like that anyway.
Note that this commit does not affect resolution for lambdas that end in
a Nothing-returning expression, e.g. `throw`.
* `return` should only be added to the last statement if the return
type is not Unit
* If there is a `return` without an argument, then the expected return
type is Unit and the last expression is not a return argument (unless
it's an incomplete call, in which case it is inferred to return Unit;
this behavior is questionable, but inherited from K1)
* There should be a constraint on return arguments even if the expected
type is Unit, otherwise errors will be missed
* When the expected type is known, using the call completion results
writer is pointless (and probably subtly wrong).
^KT-54742 Fixed
Brian Norman