| Commit message (Collapse) | Author | Age | Files | Lines |
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Warning! Warning! Yes, that's what's in this commit. Why are you
panicking?
Mostly new command line options:
-Xlint // basically, the ones which aren't noisy Ywarn-all
-Ywarn-dead-code Ywarn-inaccessible // try this one on the library:
-it makes some good points Ywarn-nullary-override Ywarn-nullary-unit
-Ywarn-numeric-widen Ywarn-value-discard
Some accumulated motivations:
The wontfix resolution of ticket #4506 indicates that "def foo" and "def
foo()" are always going to be treated differently in some situations
and the same in others without users having any way to fix it. Summary
expressed in latest comment with which I agree (and quite sadly, given
that I've done a lot of work to try to make them usable) is "avoid using
structural types like the plague." But the least we can do is warn if
you're using parentheses "wrong".
I think it would be better if the warning about "def foo()" overriding
"def foo" were an error instead. If we have to live with this...
trait Me { def f(): Int }
class A { def f: Int = 5 }
class C extends A with Me { }
// error: Int does not take parameters
def f(x: C) = x.f()
// compiles
def f(x: Me) = x.f()
// error: Int does not take parameters. Mmph, how can a method be
// legal with parameter "Foo" and illegal with parameter "Foo with
// Bar" ?
def f(x: Me with C) = x.f()
The warning about a method contains a reference to a type which is less
accessible than the method itself is obviously to those who recall it
a response to GenTraversable being private and appearing in flatMap's
signature during the 2.9.0 RCs. It avoids warning in the case where the
unnormalized type is inaccessible but the normalized version would be,
but it could use further refinement.
|
|
|
|
|
| |
overeager optimisation caused timeout in neg/t3399. no review
|
|
|
|
|
|
| |
Use the correct encoding in the browsing loader for the presentation
compiler. This should be backported to 2.9 and 2.8. no review.
|
| |
|
|
|
|
|
|
|
|
| |
Revert "Check that 'info' is only called on the presentation compiler
thread. This is a temporary debugging option"
This reverts commit 29329aa84a1f8cff4ff866044faf2490a39dadd5.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
slight improvement to lubList so that the simple case of lubbing type
constructors works. review by extempore
the strategy is to detect when the ts in lub(ts) are actually type
constructors and remember their type parameters the BTS of a type
constructor is a list of proper types (the type constructors have been
applied to their dummy arguments, which are simply type refs to the
original type parameters) in lubList, we undo this damage by stripping
these dummy arguments (they refer to type parameters that are meant to
be bound)
a better fix would be to actually bind those type parameters that
appear free in error, but that would require major changes to the BTS
infrastructure
example that only kindasorta works now...
given:
trait Container[+T]
trait Template[+CC[X] <: Container[X]]
class C1[T] extends Template[Container] with Container[T]
C1's BTS contains Template[Container] with Container[T], but that
should really be [T] => Template[Container] with Container[T]
instead of wrapping it in a polytype, the current approach uses
elimHOTparams to patch up this type so that it looks more like a type
ctor: Template[Container] with Container, but this is ill-kinded as
Template[Container] is a proper type, whereas Container is not
the performance impact should be minimal, but caveat reviewer
|
|
|
|
|
|
|
| |
Check that 'info' is only called on the presentation compiler thread.
This is a temporary debugging option for the presentation compiler. no
review.
|
|
|
|
|
|
|
|
| |
Profiler suggested it would be worthwhile to short-circuit
allOverriddenSymbols.isEmpty, and one thing led to another. I don't
know how much I accomplished performancewise, but the cosmetology is
outstanding. Review by moors.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Continuing to press on RefChecks, discovered what seemed a pretty
spectacular performance bug, at least at first. I was disappointed to
see that the time gain didn't measure up. Still, nanos are nanos.
According to YourKit, millis spent in List#foreach on a compilation of
src/library broke down like this:
| +---LinearSeqOptimized$class.foreach | 173,123 100 % |
| +---RefChecks$RefCheckTransformer.register$1 | 117,658 68 % |
I think that is largely a profiler lie, but there is enough truth. After
adjusting to prune before recursing, the number of invocations of method
register in validateBaseTypes drops by 98%.
Review by odersky.
|
|
|
|
|
|
|
| |
Suppressing a failing assert which isn't accomplishing anything but
annoyance. Why is this suddenly tripping with frequency? I do not know.
No review.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
possible, just far enough to avoid all kinds of undesirable consequences
which accompany the preservation of too much type information. (The
problems are akin to inferring the singleton type too freely.)
// Example of code which did not compile, but now does class A class
B[T <: A](cons: T)
object C extends B(new A {})
Closes #4110, #3048. I already ran this by moors, so review by odersky.
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Profiling revealed a suspiciously heavy consumer of CPU time during
refchecks called checkOverloadedRestrictions. Turns out it was looking
for overloads where more than one method defined default arguments, but
with N^2 carved on its checking stick.
Time spent in checkOverloadedRestrictions in quick.lib drops from 5800ms
to 300ms.
Review by rytz.
|
|
|
|
|
|
|
|
| |
Avoided a ton of unnecessary garbage by creating singletons for <:<
and =:= and giving them to all comers. Mother Earth is pleased. Thanks
to Michael Bayne for pointing out the suboptimality of giving everyone
their own instance. Review by moors.
|
|
|
|
|
|
|
| |
Generalizing some TreeGen machinery on an opportunistic basis. Better
documenting how Class types and classOf are implemented. Cleaning up the
manifest code. No review.
|
| |
|
|
|
|
|
| |
Cosmetic removal of redundant toList call on iterable target, no review.
|
| |
|
|
|
|
|
|
|
| |
Defined tpeHK in a manner consistent with tpe so it has a better chance
of being found/understood, and tried to clean up its documentation a
little bit. Review by moors.
|
|
|
|
|
|
| |
Integrating feedback from adriaan, changed another call to tpe to tpeHK
and simplified the surrounding method. No review.
|
| |
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Added sanity check to lub calculation to prevent invalid lubs from
emerging. The underlying cause of said lubs is that higher-order
type parameters are not handled correctly: this is why the issue
is seen so frequently in the collections. See pending test
pending/pos/those-kinds-are-high.scala for a demonstration. Until that's
fixed, we can at least raise the bar a bit.
Closes #2094, #2322, #4501. Also, some test cases in neg have been
promoted into working programs: #2179, #3774. (They're not in neg for
the "shouldn't work" reason, but out of despair.)
In some cases, such as the original reported ticket in #3528, this
only pushes the problem downfield: it still fails due to inferred type
parameters not conforming to bounds. I believe a similar issue with
higher-order type parameters underlies that.
Look at how far this takes us though. All kinds of stuff which did not
work, now works. None of these even compiled until now:
scala> :type List(mutable.Map(1 -> 1), immutable.Map(1 -> 1))
List[scala.collection.Map[Int,Int]]
scala> :type Set(List(1), mutable.Map(1 -> 1))
scala.collection.Set[Iterable[Any] with PartialFunction[Int,Int]]
scala> :type Stream(List(1), Set(1), 1 to 5)
Stream[Iterable[Int] with Int => AnyVal{def getClass(): Class[_ >: Int with Boolean <: AnyVal]}]
scala> :type Map(1 -> (1 to 10), 2 -> (1 to 10).toList)
scala.collection.immutable.Map[Int,scala.collection.immutable.Seq[Int]
]
PERFORMANCE: compiling quick.lib and quick.comp, this patch results in
an extra 27 subtype tests. Total. Time difference too small to measure.
However to be on the safe side I made it really easy to disable.
private final val verifyLubs = true // set to false
Review by moors, odersky.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
hunch by adriaan (needed to change Object to Any in strategic location),
location + fix determined by paul,
menial work (reverts of obsolete spears and introduction of fix) by
adriaan
review by extempore
Revert "A line missed from spear thrust, no review. Revert "
"Thrusting spear into darkened alcove attempting to slay java5 "
Revert "New theory: fails running on java 1.5. Put in hack to "
discover Revert "Everything builds for me, but apparently not for "
jenkins. First "
|
|
|
|
|
|
|
|
| |
Overrode contains and apply in mutable.HashMap to avoid allocating
an unnecessary Some on every call to either of them. Fruit looks a
little better defended in immutable.HashMap, so I deleted a bunch of old
debugging code instead. Closes #4469, no review.
|
|
|
|
|
|
|
|
| |
Notice when erasure is trying to cast something to Unit, and don't
do it. I'm not sure at what prior point such things should have been
caught, but for now we can have a sanity check. Closes #4731, review by
odersky.
|
|
|
|
|
|
|
|
|
| |
Added a -Y option to resolve namespace collisions between package and
object. It's a blunt instrument: if people have lots of these conflicts
they need to resolve in individually nuanced fashion, they'll probably
remain out of luck. But now people can use YourKit probes without a
custom hacked compiler. Let's say this closes #2089. Review by odersky.
|
|
|
|
|
| |
Playing defense against NoSymbol in scaladoc, no review.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
implementations for $hash$hash can be inlined directly, with this kind
of beautiful change:
// source
def hash(x1: Int, x2: Int) = x1.## + x2.##
// before
0: getstatic #11; //Field scala/runtime/ScalaRunTime$.MODULE$:Lscala/runtime/ScalaRunTime$;
3: iload_1
4: invokevirtual #27; //Method scala/runtime/ScalaRunTime$.hash:(I)I
7: getstatic #11; //Field scala/runtime/ScalaRunTime$.MODULE$:Lscala/runtime/ScalaRunTime$;
10: iload_2
11: invokevirtual #27; //Method scala/runtime/ScalaRunTime$.hash:(I)I
14: iadd
15: ireturn
// after
0: iload_1
1: iload_2
2: iadd
3: ireturn
No review.
|
| |
|
|
|
|
|
|
| |
Thrusting spear into darkened alcove attempting to slay java5 build
failure. No review.
|
|
|
|
|
| |
truth of theory. No review.
|
|
|
|
|
| |
Moved test for #4727 to where it was supposed to go, no review.
|
|
|
|
|
|
|
|
|
| |
Everything builds for me, but apparently not for jenkins. First attempt
to solve mystery: explicitly set return type of Any#getClass() to
Class[_ <: Any] rather than allowing java's to be used. I'm guessing
that somehow it materializes as Class[_ <: Any] sometimes and Class[_ <:
AnyRef] other times. Review by moors.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
A total rewrite of "runtimeClass", discarding the user-space approach in
favor of simply fixing getClass.
def f1 = 5.getClass // Class[Int]
def f2 = (5: AnyVal).getClass // Class[_ <: AnyVal]
def f3 = (5: java.lang.Integer).getClass // Class[_ <: java.lang.Integer]
class A
class B extends A
def f1 = (new B: Any).getClass().newInstance() // Any
def f2 = (new B: AnyRef).getClass().newInstance() // AnyRef
def f3 = (new B: A).getClass().newInstance() // A
def f4 = (new B: B).getClass().newInstance() // B
But that's not all!
def f0[T >: B] = (new B: T).getClass().newInstance()
def f5 = f0[Any] // Any
def f6 = f0[AnyRef] // AnyRef
def f7 = f0[A] // A
def f8 = f0[B] // B
Closes #490, #896, #4696. Review by moors. (Note: I think this is pretty
good, but picky review requested.)
|
| |
|
| |
|
|
|
|
|
|
|
| |
Removing the package-private modifiers on Gen*Like traits.
No review.
|
|
|
|
|
|
|
|
| |
Removed "isInterpreterWrapper" from the conditions which imply an
effective root. This was an old hack from before my time which is no
longer necessary and then recently became actively hostile. Closes
#4710, no review.
|
|
|
|
|
|
|
|
| |
Relaxes a typer check which fails valid code with NoCommonType. If the
instantiated types or type bounds do not conform, it tries normalizing
the type before throwing the exception. Closes #4553. I wrote this patch
with adriaan already, but bonus review by moors.
|
|
|
|
|
|
| |
Documented what happened with the implementation of narrow for future
performance detectives. No review.
|
|
|
|
|
|
| |
Routed Stream.+: through cons to avoid some truly pathological
performance. Closes #4697, no review.
|
| |
|
| |
|
| |
|
| |
|
|
|
|
|
|
|
|
|
|
|
| |
Performance-motivated work: going through SpecializeTypes and
eliminating unnecessary intermediate data structures and multiple
traversals of lists. Also trying to get that code under control: it
dwarfs all other phases in terms of debugging output, and a too-large
percentage of the source is commented out debugging code which looks
past its sell-by date. I realize the patch is a little big for a very
thorough review, but review by dragos, prokopec.
|
|
|
|
|
|
| |
Quoted the windows runner path so "c:\program files (x86)" has some
chance of working. Closes #4636, no review.
|
| |
|
| |
|