| Commit message (Collapse) | Author | Age | Files | Lines |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
In 65340ed4ad2e, parts of RefChecks were disabled when
we traversed into the results of the new pattern matcher.
Similar logic existed for the old pattern matcher, but in
that case the Match / CaseDef nodes still existed in the tree.
The new approach was too broad: important checks no longer
scrutinized the body of cases.
This commit turns the checks back on when it finds the remnants
of a case body, which appears as an application to a label def.
Conflicts:
src/compiler/scala/tools/nsc/typechecker/RefChecks.scala
Cherry pick of 3df1d77fc984b976efa68098206e801cf3b83a9e
|
|\
| |
| | |
[nomaster] Fix non-deterministic <:< for deeply nested types
|
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| | |
Backported from master. This is a squashed commmit comprising:
SI-8146 Pending test, diagnosis for bug in decidability of <:<
(cherry picked from commit 8beeef339ad65f3308ece6fb0440cdb31b1ad404)
SI-8146 Test cases for typechecking decidability
Taken from "On Decidability of Nominal Subtyping with Variance"
(Pierce, Kennedy), which was implemented in 152563b.
Part of the implementation (SubTypePair) will be changed in the
following commit to fix the non-deterministic errors typechecking
heavily nested types involving aliases or annotations.
(cherry picked from commit 2e28cf7f76c3d5fd0c2df4274f1af9acb42de699)
SI-8146 Fix non-deterministic <:< for deeply nested types
In the interests of keeping subtyping decidable [1], 152563b
added some bookkeeping to `isSubType` to detect cycles.
However, this was based on a hash set containing instances of
`SubTypePair`, and that class had inconsistencies between its
`hashCode` (in terms of `Type#hashCode`) and `equals`
(in terms of `=:=`).
This inconsistency can be seen in:
scala> trait C { def apply: (Int @unchecked) }
defined trait C
scala> val intUnchecked = typeOf[C].decls.head.info.finalResultType
intUnchecked: $r.intp.global.Type = Int @unchecked
scala> val p1 = new SubTypePair(intUnchecked, intUnchecked)
p1: $r.intp.global.SubTypePair = Int @unchecked <:<? Int @unchecked
scala> val p2 = new SubTypePair(intUnchecked.withoutAnnotations, intUnchecked.withoutAnnotations)
p2: $r.intp.global.SubTypePair = Int <:<? Int
scala> p1 == p2
res0: Boolean = true
scala> p1.hashCode == p2.hashCode
res1: Boolean = false
This commit switches to using `Type#==`, by way of the standard
case class equality.
The risk here is that you could find a subtyping computation that
progresses in such a manner that we don't detect the cycle. It would
need to produce an infinite stream of representations for types that
were `=:=` but not `==`. If that happened, we'd fail to terminate,
rather than judging the relationship as `false`.
[1] http://research.microsoft.com/pubs/64041/fool2007.pdf
(cherry picked from commit a09e143b7fd1c6b433386d45e9c5ae3548819b59)
Conflicts:
src/reflect/scala/reflect/internal/tpe/TypeComparers.scala
src/reflect/scala/reflect/runtime/JavaUniverseForce.scala
|
|\ \
| | |
| | | |
Fix bug with super-accessors / dependent types
|
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | | |
Super-accessors are generated as `DefDef`'s with `EmptyTree` as a
placeholder for the RHS. This is filled in later in `Mixin` in
`completeSuperAccessor`.
A change in `Uncurry` (SI-6443 / 493197f), however, converted this
to a `{ EmptyTree }`, which evaded the pattern match in mixin.
This commit adds a special case to the dependent method treatment
in Uncurry to avoid generating redundant blocks.
|
|\ \ \
| |_|/
|/| | |
[nomaster] Backport variance validator performance fix
|
| |/
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| | |
% time qbin/scalac test/files/pos/t8146-performance.scala
real 0m2.015s
user 0m2.892s
sys 0m0.215s
% time scalac-hash v2.10.3 test/files/pos/t8146-performance.scala
real 1m13.652s
user 1m14.245s
sys 0m0.508s
Cherry-picks one hunk from 882f8e64.
|
|\ \
| |/
|/| |
Repair symbol owners after abandoned named-/default-args
|
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| | |
Names/Defaults eagerly transforms an application with temporaries
to maintain evaluation order, and dutifully changes owners of
symbols along the way.
However, if this approach doesn't work out, we throw away this
and try a auto-tupling. However, we an still witness symbols
owned by the temporaries.
This commit records which symbols are owned by the context.owner
before `transformNamedApplication`, and rolls back the changes
before `tryTupleApply`.
Perhaps a better approach would be to separate the names/defaults
applicability checks from the evaluation-order-preserving transform,
and only call the latter after we have decided to go that way.
|
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| | |
We can't backport SI-7120 to 2.10.x as it changes erased signatures,
which can lead to interop problems between 2.10.3 and 2.10.4.
But, we can detect one of the nasty symptoms -- a bridge method
with the same signature as its target -- and treat that.
This commit detects duplicate bridges in the ASM (only) backend
and removes them.
|
|/
|
|
| |
Progressed to working in SI-7636 / c4bf1d5.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
As it was discovered in SI-8104, whiteboxity doesn’t apply equally to
type parameters and type members of materialized type classes.
During implicit search and subsequent type inference, whitebox type parameters
are consistently erased to wildcards, whereas whitebox type members sometimes
remain as is and get in the way of signature conformance checks.
Unfortunately, 2.10.x can’t make use of type parameter whiteboxity, because
it requires fundep materializers that were only merged into 2.11:
https://github.com/scala/scala/pull/2499, and therefore Si-8104 seems to be
a hard blocker for 2.10.x at the moment. Stay tuned for updates.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
A source file like:
import foo.bar
package object baz
Is parsed into:
package <empty> {
import foo.bar
package baz {
object `package`
}
}
A special case in Namers compensates by adjusting the owner of
`baz` to be `<root>`, rather than `<empty>`.
This wasn't being accounted for in `BrowserTraverser`, which
underpins `-sourcepath`, and allows the presentation compiler to
load top level symbols from sources outside those passes as
the list of sources to compile.
This bug did not appear in sources like:
package p1
package object p2 { ... }
... because the parser does not wrap this in the `package <empty> {}`
This goes some way to explaining why it has gone unnoticed for
so long.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
* The presentation compiler sourcepath is now correctly set-up.
* Amazingly enough (for me at least), the outer import in the
package object seem to be responsible of the faulty behavior.
In fact, if you move the import clause *inside* the package object,
the test succeed!
The test's output does provide the correct hint of this:
```
% diff /Users/mirco/Projects/ide/scala/test/files/presentation/t8085-presentation.log /Users/mirco/Projects/ide/scala/test/files/presentation/t8085.check
@@ -1,3 +1,2 @@
reload: NodeScalaSuite.scala
-prefixes differ: <empty>.nodescala,nodescala
-value always is not a member of object scala.concurrent.Future
+Test OK
```
Notice the ``-prefixes differ: <empty>.nodescala,nodescala``. And compare
it with the output when the import clause is placed inside the package
object:
```
% diff /Users/mirco/Projects/ide/scala/test/files/presentation/t8085-presentation.log /Users/mirco/Projects/ide/scala/test/files/presentation/t8085.check
@@ -1,4 +1,2 @@
reload: NodeScalaSuite.scala
-reload: NodeScalaSuite.scala
-open package module: package object nodescala
Test OK
```
Notice now the ``-open package module: package object nodescala``!
|
|
|
|
|
|
|
| |
We used to silently skip class files that would exceed the JVM's size limits.
While rare, this should still be an error.
While I was at it, also included the name of the offending method.
|
|\
| |
| | |
Revert ", importable _."
|
| |
| |
| |
| |
| |
| |
| | |
This reverts commit d2316df920ffa4804fe51e8f8780240c46efa982.
We can't make `_` an illegal identifier -- it's legal in Java,
so we must be able to name these Java underscores.
|
|\ \
| |/
|/| |
Fix inliner cycle with recursion, separate compilation
|
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| | |
ICodeReaders, which decompiles JVM bytecode to ICode, was not
setting the `recursive` attribute of `IMethod`. This meant that
the inliner got into a cycle, repeatedly inlining the recursive
call.
The method name `filter` was needed to trigger this as the inliner
heuristically treats that as a more attractive inlining candidate,
based on `isMonadicMethod`.
This commit:
- refactors the checking / setting of `virtual`
- adds this to ICodeReaders
- tests the case involving `invokevirtual`
I'm not sure how to setup a test that fails without the other changes
to `ICodeReader` (for invokestatic and invokespecial).
|
|\ \
| | |
| | | |
SI-7912 Be defensive calling `toString` in `MatchError#getMessage`
|
| | |
| | |
| | |
| | |
| | | |
Otherwise, objects with exception-throwing `toString` lead to a
cascading error far removed from the originally failed match.
|
|\ \ \
| | | |
| | | | |
SI-8060 Avoid infinite loop with higher kinded type alias
|
| |/ /
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | | |
The `dealiasLocals` map was assuming that:
tp.isAliasType implies (tp.dealias ne tp)
This isn't true if `!typeParamsMatchArgs`.
This commit avoids the infinite loop by checking whether or not
dealiasing progresses.
|
|\ \ \
| |/ /
|/| | |
Clean up Scala distribution a bit.
|
| | | |
|
|\ \ \
| |/ /
|/| | |
SI-7995 completion imported vars and vals
|
| | |
| | |
| | |
| | |
| | |
| | | |
Imported member vals and vars were always marked inaccessible, even
if referencing them at the location of the completion is valid in code.
The accessible flag is now set accordingly to the accessibility of the getter.
|
|\ \ \
| | | |
| | | | |
SI-8029 Avoid multi-run cyclic error with companions, package object
|
| |/ /
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | | |
The bug report suggests this problem only became visible in the
IDE *after* 2.10.3, but I tested an IDE with exactly 2.10.3 and
saw the same problem.
In fact, my test case fails as far back as 2.10.0. I suspect the
problem showed up after 816cecf9a9, which checks that pairs of
companion symbols are codefined eagerly in Namers.
This commit modifies the check of `rawInfo ne NoType`
in `isCoDefinedWith` to avoid triggering adaptation of types from
the previous run. I'm not sure of the precise intent of that check.
I looked at c9861cd198 (genesis of isCoDefinedWith). Before that we
get back to 3761cb4b3a1 (the dawn of Subversion.)
|
|\ \ \
| | | |
| | | | |
[backport] SI-7439 Avoid NPE in `isMonomorphicType` with stub symbols.
|
| | |/
| |/|
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | | |
`originalInfo` can return null for stub symbols; deal with that
as we used to before a regression in 016bc3db.
After this change, we can once again delete A_1.class and still compile
code instantiating B_1. (A_1 is only referred to in a method signature
of B_1 which is not called from our code.)
scala> new B_1
warning: Class A_1 not found - continuing with a stub.
res0: B_1 = B_1@5284b8f9
In practice, this situation arises when someone uses a third
party class that was compiled against other libraries not avaialable
on the current compilation classpath.
(cherry picked from commit a95432168204964e4f6c4571e781559c1640f2d8)
|
|\ \ \
| | | |
| | | | |
SI-8010 Fix regression in erasure double definition checks
|
| | |/
| |/|
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | | |
Calls to `Symbol#info` during scope iteration considered harmful.
Looks like calling `info` during this Scope iteration is triggering
the ExplicitOuter info transformer, which "makes all super accessors
and modules in traits non-private, mangling their names.". This name
change necessitates a rehashing of the owning scope, which I suspect
is enough to corrupt the ScopeEntry-s being traversed in
`checkNoDeclaredDoubleDefs`.
The upshot was that we encountered the same symbol twice, which was
reported as being a double-definition.
This problem only showed up after 086702d8a74, which did nothing
worse then change the order in which `{e, e1}.sym.info` were
forced.
I inspected SymbolPairs/OverridingPairs which *appear* to be immune
as they only test flags during scope iteration; infos are not used
until later, at which point we're iterating a temporary scope that
isn't part of the type of the owner of the symbols.
|
| |/
|/|
| |
| | |
backport of 1d3ec4e708154ec05554f540d7d68ed55dc12426
|
|\ \
| | |
| | | |
make askLoadedType unload arguments out of the PC by default,
|
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | | |
The rationale for not keeping units loaded by default is that the more
units are loaded, the slower is background compilation. For instance, in
the Scala IDE for Eclipse (which uses the presentation compiler),
typechecking occurs every time the reconciler kicks-in (~500millis after
you stop typing), hence it is important that units are not kept loaded
unless strictly necessary (for some extra information about this, see
https://www.assembla.com/spaces/scala-ide/tickets/1001388)
While I agree that using a boolean argument (`keepLoaded`) for deciding
if a unit should be loaded isn't a great design, other methods in
`CompilerControl` also have a keepLoaded parameter, so at least we have
some consistency. For the future, I'm thinking we should be able to
remove the `keepLoaded` flag altogether, and change the implementation
of `askLoadedType` to preserve the same units loaded in the presentation
compiler before and after its execution. Basically, if you want a unit
to be kept loaded, you should call `askReload` first, and then
`askLoadedType`. However, to reduce impact, I think the changes carried
by this commit will help us estimate if the solution I just outlined is
viable (because `askLoadeType` won't be keeping units loaded by default,
which wasn't the case with the former implementation).
(While the patch was mostly contributed by @huitseeker, @dotta has edited the
commit message to preserve the comments in the PR
https://github.com/scala/scala/pull/3209)
|
| |/
|/|
| |
| | |
Origin: viktorklang@1bbe854
|
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| | |
Otherwise we can think that `+` in `1 + BigInt(2)` refers
to a method in `Int`.
In general, this protects the IDE from observing results from
"exploratory" typing which is discarded as the compiler backtracks
to another possibility.
This protection subsumes the condition that checked for overloaded
types: presentation/t7458 now passes without this.
|
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| | |
We shouldn't observe tree types under silent mode.
The enclosed test is a standalone version of `1 + BigInt(2)`,
a standard example of exploratory typing in Scala. Once we determine
that none of the `+` methods in `Int` accepts (possibly implicitly
coerced `BigInt`), we have to backtrack and look for a view from
`Int => { +(_: BigInt): ? }`.
The next commit will correct the problem.
|
|/
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
targeted type-checked
When asking for targeted typecheck, the located tree may have overloaded types
is the source isn't yet fully typechecked (e.g., a select tree for an
overloaded method). This is problematic as it can lead to unknown 'hovers',
broken hyperlinking that suddenly starts working, unresolved ScalaDoc comments,
and similar, in the Scala IDE.
With this commit we are hardening the contract of `askTypeAt` to return the
same type whether the file was fully type-checked or targeted type-checked.
This is done by preventing the typechecker to stop too early if the `located`
tree has an overloaded type. Furthermore, I'm assuming that if `located.tpe`
is of type `OverloadedType`, by letting the compiler carry-on the typechecking,
the `located.tpe` will eventually be resolved to a non-overloaded type. Said
otherwise, I expect the targeted typechecking will always terminate (if my
reasoning isn't sound, please say so).
The test provided with this commit demonstrates the new behavior (the position
used to execute the test is resolved to the `foo` method's call). In fact,
before this commit, executing the test returned the following:
(x: Int, y: String)Unit <and> (x: String)Unit <and> (x: Int)Unit
Showing that the tree's type is an overloaded type. The ambiguity is fixed by
this commit, and in fact the test's output is now:
(x: Int)Unit
|
|\
| |
| | |
teaches toolbox about -Yrangepos
|
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| | |
Unlike in master, in 2.10.x enabling -Yrangepos requires instantiating
Global with mixed in RangePositions trait.
Same story for toolboxes. Just setting Yrangepos is not enough - one
needs to mix in RangePositions into ToolboxGlobal. I didn’t know that
back then, so now I’m fixing the oversight.
The commit is marked as [nomaster], because -Yrangepos doesn’t need
special treatment in master.
|
| |
| |
| |
| |
| |
| |
| |
| | |
Previously there occurred a NoPosition error when one asks for position
information in the AST because no positions were set to the trees
created during the transformation for updateDynamic calls. This commit
applies range positions to the trees in order to being able to highlight
them inside of the scala-ide.
|
| |
| |
| |
| |
| |
| |
| |
| |
| | |
Previously, there were no positions created for the tuples that are
generated while doing the transformation for an applyDynamicNamed call.
This led to an NoPosition error in scalac when one tries to show
position information in the AST. Furthermore, this simplifies semantic
highlighting in the scala-ide because no position information for color
ranges have to be created anymore.
|
|/
|
|
|
|
|
|
|
|
|
|
| |
The new positions are range positions that directly refer to the
beginning and the end of the method calls in the sources instead of
simply point to the beginning of the expression. This allows the
scala-ide to semantically highlight select- and applyDynamic method
calls, because it has only to traverse the tree and apply the color
ranges to the given position ranges.
This also fixes the position marker of an error messages related
to a wrong Dynamic method signature.
|
|\
| |
| | |
Backport of
|
| |
| |
| |
| |
| |
| |
| |
| |
| | |
Updates localeContext() to return the best context possible when there are none directly
associated with the given position. It happens when an expression cannot be
successfully typed, as no precise ContextTree covers the expression location, or if the
position is not inside any expression.
Adds corresponding tests
(cherry picked from commit 3028327e2a2b553b12ee45519413515c8aa0865f)
|
| |
| |
| |
| | |
(cherry picked from commit 3d55fe723f1af91f4d2db421f0e0965c583346dc)
|
| |
| |
| |
| |
| |
| | |
Adds a new marker /*_*/ to trigger scope completion test.
Original type completion test oracles update for the tweaked output
(cherry picked from commit 9c7c66ff7907e3ab814f0f4375eeaf6cdd230d5e)
|
|/
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The tree created during expansion of default arguments contained trees with the
wrong type of positions. Let's discuss this with an example. Below is the tree
generated for the `foo` method in the test class included in this commit.
Before this commit:
```
[54:94]def foo(): [58]Unit = <70:90>{
[70:79]<artifact> val qual$1: [70]Bar = [70:79][70:79][70:79]new [74:77]Bar();
[80]<artifact> val x$1: [80]Int = [80]qual$1.bar$default$1;
<70:90><70:83>qual$1.bar([80]x$1)
}
```
Now:
```
[54:99]def foo(): [58]Unit = <70:95>{
<70:84><artifact> val qual$1: [70]Bar = [70:84][70:84][70:84]new [74:77]Bar();
[85]<artifact> val x$1: [85]Int = [85]qual$1.bar$default$1;
<70:95>[84:88]qual$1.bar([85]x$1)
}
```
Here are the list of changes:
* The synthetic `qual$1` has a transparent position, instead of a range position.
* The new Select tree (i.e., `qual$1.bar`) should always have a range position,
because `selected` (i.e., the called method) is always visible in the source
(in fact, this is the whole point of the fix, we need a range position or
hyperlinking request from the Scala IDE won't work).
* The Block that contains the expanded default arguments is forced to have a
transparent position, as it never exist in the original source.
The tricky part of the fix is the position assigned to the new Select tree,
which needs to respect the range position's invariants. In the specific case,
we ought to make sure that range positions don't overlap. Therefore, the position
assigned to the new Select tree is computed by intersecting the original Select
position (i.e., `baseFun`'s position) and the original qualifier's position
(i.e., `qual`'s position).
If you take a closer look at the range positions assigned in the tree after this
commit, you'll notice that the range position of the `qual$1`'s rhs (i.e.,
[70:84]), and `qual$1.bar` (i.e., [84:88]) might seem to overlap, because the
former ends where the latter begins. However, this not the case because of the
range position's invariant 2, which states:
> Invariant 2: in a range position, start <= point < end
Hence, the above two positions aren't overlapping as far as the compiler is
concerned.
One additional aspect (that may look like a detail) is that we make sure to
never generate a position such that its start is after its end. This is why we
take the position with the smallest end point.
Furthermore, calling `withStart` would turn any position in a range position,
which isn't desiderable in general (and, even worse, this can lead to
generation of invalid positions - bad offsets - if the computation is performed
on offset positions). Hence, the position's computation is only performed when
both `baseFun` and `qual` positions are range positions. Indeed, I expect this to
be always the case if the compiler is started with -Yrangepos.
(cherry picked from commit 3009a525b58a4c7865ff524899b85518884ee5f7)
|