| Commit message (Collapse) | Author | Age | Files | Lines |
|
|
|
|
|
|
| |
Before this commit, ParSetLike compiled fine as part of compileStdLib
but crashed when compiled by itself because we tried to force a LazyRef
while forcing the same LazyRef. This commit fixes this by being slightly
more lazy where it matters.
|
|\
| |
| | |
Fix #2066: Don't qualify private members in SelectionProto's...
|
| |
| |
| |
| |
| | |
Now we never match `? { name: T }` with types that
have only a private `name` member. This is what scalac does, too.
|
| |
| |
| |
| | |
... unless they would be accessible in the given context.
|
| |
| |
| |
| | |
You can still have positions printed by explictly passing -Yprintpos
|
|\ \
| | |
| | | |
Fix #360: Improve avoidance algorithm
|
| | | |
|
| | |
| | |
| | |
| | | |
No more try-again business necessary.
|
| |/
| |
| |
| |
| |
| | |
The essential change is that we do not throw away more
precise info of the avoided type if the expected type
is fully defined.
|
| |
| |
| |
| |
| |
| |
| | |
This was a mistake introduced in the previous commit, installAfter is
only safe to use in `IdentityDenotTransformer` phases, otherwise it
means that the phase denotation transformer is not run at all for this
particular denotation, this caused Ycheck to fail.
|
|/
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
If A method like:
override def foo(x: Object*)
overrides a Java method, it needs to be rewritten as:
def foo(x: Seq[Object])
override def foo(x: Array[Object]): Object = foo(Predef.wrapRefArray(x))
This should be handled by ElimRepeated but there were two bugs:
- `addVarArgsBridge` was called at phase `thisTransformer.next`, this is
too late to create the bridge since `T*` has already been rewritten as
`Seq[T]`
- The original method symbol needs to have the `override` flag dropped,
since it doesn't override anything.
Furthermore, RefChecks had to be moved after ElimRepeated, otherwise the
testcase would fail the overriding checks.
|
|\
| |
| | |
Fix bug in erasedLub leading to incorrect signatures
|
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| | |
Before this commit, the added testcase failed in a strange way:
14 | def bla(foo: Foo) = orElse2(identity).apply(foo)
| ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
|value of type <nonsensical><notype></nonsensical> does not take parameters
This happened because the TermRef for the apply method had an incorrect
signature, therefore its underlying type was NoType.
According to the documentation of `erasedLub`, the erasure should be:
"a common superclass or trait S of the argument classes, with the
following two properties:
S is minimal: no other common superclass or trait derives from S]
S is last : in the linearization of the first argument type `tp1`
there are no minimal common superclasses or traits that
come after S.
(the reason to pick last is that we prefer classes over traits that way)."
I'm not convinced that the implementation satisfies either of these two
properties, but this commit at least makes S closer to being minimal by
making sure that the last best candidate never derives from it.
|
|\ \
| | |
| | | |
Fix #2064: Avoid illegal types in OrDominator
|
| | | |
|
| | |
| | |
| | |
| | |
| | | |
`superType` should never return a `TypeBounds` in the first place.
This commit makes sure it doesn't.
|
| | |
| | |
| | |
| | |
| | | |
Need to skip type bounds in `underlying` chain, since
TypeBounds is not a legal operand type for OrType.
|
|\ \ \
| | | |
| | | | |
Fix type inference for HLists and HMaps
|
| | | | |
|
| | | |
| | | |
| | | |
| | | | |
and a typo fixed
|
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | | |
Turned out hmaps.scala requires the arg alignment to compile.
So we have our first counterexample that we cannot drop this
hack. Now it is made safe in the sense that no constraints get
lost anymore.
|
| | | |
| | | |
| | | |
| | | |
| | | | |
Need to use fresh PolyParams instead of WildcardTypes
if constraint is committable.
|
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | | |
We previously tried to force S1 and S2 be the same type
when encountering a lub like `T1 { A = S1 } & T2 { A = S2 }`.
The comments in this commit explain why this is unsound, so
this rewrite is now made subject to a new config option, which
is off by default.
I verified that the new behavior does not affect the performance of
the junit tests.
|
| | | |
| | | |
| | | |
| | | |
| | | |
| | | | |
We approximate dependencies to parameters by Wildcards. This was already
done in one place, is now done in other places as well, instead of doing nothing for
dependent methods.
|
| | | |
| | | |
| | | |
| | | |
| | | | |
No reason why we should not - normalize handles implicit
methods just fine. This fixes type errors in test HLists.scala.
|
| | | |
| | | |
| | | |
| | | |
| | | |
| | | | |
If an applied type has a refinement, it was
printed before as one large refinement type
including the type parameter bindings.
|
|\ \ \ \
| | | | |
| | | | | |
Change "recursive/cyclic definitions needs type" errors to Message
|
| | | | |
| | | | |
| | | | |
| | | | | |
Fall-back to reporting "overloaded or recursive needs type".
|
| | | | |
| | | | |
| | | | |
| | | | |
| | | | |
| | | | | |
Split error messages for recursive method and overloaded method needs type into two (but
did not solve the analysis which to show).
Make CyclicReference type error construct corresponding error message.
|
| | | | | |
|
| |\ \ \ \
| | | | | |
| | | | | |
| | | | | | |
ennru_RecursiveNeedsType
|
| | | | | | |
|
|\ \ \ \ \ \
| | | | | | |
| | | | | | | |
Disallow subtypes of Function1 acting as implicit conversions
|
| | | | | | |
| | | | | | |
| | | | | | |
| | | | | | |
| | | | | | |
| | | | | | | |
Now only Scala2 mode treats Function1's as implicit conversions. Instead we introduce
a new subclass ImplicitConverter of Function1, instances of which are turned into
implicit conversions.
|
| | | | | | |
| | | | | | |
| | | | | | |
| | | | | | | |
I introduced an error in the refactoring two commits ago.
|
| | | | | | | |
|
| | |_|_|_|/
| |/| | | |
| | | | | |
| | | | | |
| | | | | |
| | | | | |
| | | | | |
| | | | | |
| | | | | |
| | | | | | |
The new test `falseView.scala` shows the problem. We might create
an implicit value of some type that happens to be a subtype of Function1.
We might now expect that this gives us an implicit conversion, this
is most often unintended and surprising.
See the comment in Implicits#discardForView for a discussion why
we picked the particular scheme implemented here.
|
|\ \ \ \ \ \
| | | | | | |
| | | | | | | |
support merging companion objects in expanded trees
|
| | | | | | | |
|
| | | | | | |
| | | | | | |
| | | | | | |
| | | | | | |
| | | | | | |
| | | | | | |
| | | | | | |
| | | | | | |
| | | | | | |
| | | | | | |
| | | | | | |
| | | | | | |
| | | | | | |
| | | | | | |
| | | | | | |
| | | | | | |
| | | | | | |
| | | | | | | |
Previous implementation is problematic when there are multiple
transforms before typer:
1. There might be objects needing merging that only exist in the
expanded trees, which cannot be handled by the previous algorithm.
2. There may be companion object and class defined only in the
expanded trees, the previous algorithm doesn't create links for them.
This PR simplifies the companion object merging logic and
fixes the problems above.
In fact, this PR supports multiple pre-typer transform during expansion.
The protocol is that the expansion of a MemberDef is either a flattened
thicket or a non-thicket tree.
|
| |/ / / / /
|/| | | | |
| | | | | |
| | | | | |
| | | | | |
| | | | | |
| | | | | | |
It's hard to predict for defKind all the desugarings that can be
applied to a definition. Better to do it once the desugarings have been
applied. NoInit and PureInterface are both tested only after Typer,
so it's OK to delay their initialization until regular Typer.
|
| | | | | |
| | | | | |
| | | | | |
| | | | | |
| | | | | |
| | | | | |
| | | | | |
| | | | | |
| | | | | |
| | | | | |
| | | | | | |
The default param will be desugared into a method with a body, so
setting PureInterface would be wrong. The enclosed test previously
failed with a pickling difference, because the unpickler correctly
decided to not set the PureInterface flag since it saw the default param
method.
This fixes the tasty_dotc_util which failed since the last commit
because FreshNameCreator was now incorrectly recognized as a PureInterface.
|
| | | | | | |
|
| | | | | | |
|
| | | | | | |
|
| | | | | | |
|
| | | | | | |
|
| | | | | | |
|
| | | | | | |
|
| |_|_|_|/
|/| | | | |
|