| Commit message (Collapse) | Author | Age | Files | Lines |
|
|
|
|
|
| |
Closes SI-3569, SI-3770.
Also threw in experimental -Yoverride-vars. It's not robust.
|
|
|
|
|
| |
Tried to paint a picture of how one might synthesize an implicit
method to accompany an implicit class.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
(Looks like there is more range position breakage yet, but
this gets the outermost layer.)
Channeling my struggles into a slightly easier future.
% scalac -Ypos-debug -d /tmp ./src/library/scala/Predef.scala
./src/library/scala/Predef.scala:222: warning: Positioned tree has unpositioned child in phase extmethods
def x = __resultOfEnsuring
^
parent: #7109 line 222 Select // (value __resultOfEnsuring in class Ensuring)
child: #7108 Ident // (value $this)
./src/library/scala/Predef.scala:258: warning: Positioned tree has unpositioned child in phase extmethods
def x = __leftOfArrow
^
parent: #7280 line 258 Select // (value __leftOfArrow in class ArrowAssoc)
child: #7279 Ident // (value $this)
two warnings found
Or try this to really see some output:
% scalac -Yrangepos -Ypos-debug
|
|\
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| | |
Conflicts:
lib/scala-compiler.jar.desired.sha1
lib/scala-library-src.jar.desired.sha1
lib/scala-library.jar.desired.sha1
src/compiler/scala/reflect/internal/Definitions.scala
src/compiler/scala/reflect/internal/Symbols.scala
src/compiler/scala/tools/nsc/Global.scala
src/compiler/scala/tools/nsc/transform/Constructors.scala
src/compiler/scala/tools/nsc/transform/Erasure.scala
src/compiler/scala/tools/nsc/transform/SpecializeTypes.scala
src/compiler/scala/tools/nsc/typechecker/Contexts.scala
src/compiler/scala/tools/nsc/typechecker/RefChecks.scala
src/compiler/scala/tools/nsc/typechecker/SyntheticMethods.scala
src/library/scala/Function0.scala
src/library/scala/Function1.scala
src/library/scala/Function10.scala
src/library/scala/Function11.scala
src/library/scala/Function12.scala
src/library/scala/Function13.scala
src/library/scala/Function14.scala
src/library/scala/Function15.scala
src/library/scala/Function16.scala
src/library/scala/Function17.scala
src/library/scala/Function18.scala
src/library/scala/Function19.scala
src/library/scala/Function2.scala
src/library/scala/Function20.scala
src/library/scala/Function21.scala
src/library/scala/Function22.scala
src/library/scala/Function3.scala
src/library/scala/Function4.scala
src/library/scala/Function5.scala
src/library/scala/Function6.scala
src/library/scala/Function7.scala
src/library/scala/Function8.scala
src/library/scala/Function9.scala
test/files/codelib/code.jar.desired.sha1
test/files/neg/anyval-children-2.check
test/files/run/programmatic-main.check
|
| |
| |
| |
| | |
new STARR!
|
| |
| |
| |
| | |
generic value classes.
|
| | |
|
| | |
|
| |\
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | | |
Conflicts:
src/compiler/scala/reflect/internal/Definitions.scala
src/compiler/scala/tools/nsc/typechecker/Namers.scala
src/library/scala/AnyVal.scala
test/files/run/primitive-sigs-2.check
test/files/run/t4172.check
test/files/scalap/abstractClass/result.test
test/files/scalap/abstractMethod/result.test
test/files/scalap/caseClass/result.test
test/files/scalap/cbnParam/result.test
test/files/scalap/classPrivate/result.test
test/files/scalap/classWithExistential/result.test
test/files/scalap/classWithSelfAnnotation/result.test
test/files/scalap/covariantParam/result.test
test/files/scalap/implicitParam/result.test
test/files/scalap/paramClauses/result.test
test/files/scalap/paramNames/result.test
test/files/scalap/sequenceParam/result.test
test/files/scalap/simpleClass/result.test
test/files/scalap/traitObject/result.test
test/files/scalap/typeAnnotations/result.test
test/files/scalap/valAndVar/result.test
test/files/scalap/wildcardType/result.test
|
| | |
| | |
| | |
| | | |
ScalaObject. Undoing wrong fix in ExtensionMethods.
|
|\ \ \
| |/ /
|/| |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | | |
Conflicts:
lib/scala-compiler.jar.desired.sha1
lib/scala-library-src.jar.desired.sha1
lib/scala-library.jar.desired.sha1
src/compiler/scala/reflect/internal/Definitions.scala
src/compiler/scala/reflect/internal/Importers.scala
src/compiler/scala/reflect/internal/Symbols.scala
src/compiler/scala/reflect/internal/Trees.scala
src/compiler/scala/reflect/internal/Types.scala
src/compiler/scala/tools/nsc/Global.scala
src/compiler/scala/tools/nsc/transform/Erasure.scala
src/compiler/scala/tools/nsc/transform/LiftCode.scala
src/compiler/scala/tools/nsc/transform/UnCurry.scala
src/compiler/scala/tools/nsc/typechecker/RefChecks.scala
src/compiler/scala/tools/nsc/typechecker/Typers.scala
test/files/run/programmatic-main.check
test/files/speclib/instrumented.jar.desired.sha1
|
| | |
| | |
| | |
| | |
| | |
| | |
| | | |
Removed all the trailing whitespace to make eugene happier.
Will try to keep it that way by protecting at the merge level.
Left the tabs in place because they can't be uniformly changed
to spaces, some are 2, some are 4, some are 8, whee.
|
| | |
| | |
| | |
| | |
| | | |
Quieting things down. Fixed some things revealed by quieter
logs, like forwarders being generated for superaccessors.
|
| | |
| | |
| | |
| | |
| | |
| | | |
Skolemization code doesn't belong in Typers. I carved
out a little place for it. Also simplifications in
specialization.
|
| | |
| | |
| | |
| | |
| | | |
Fix for trait/impl renaming in 5cbd7d06eb was incomplete.
Looks more complete now.
|
| | |
| | |
| | |
| | | |
"This too shall pass."
|
| |/
|/| |
|
| | |
|
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| | |
Fail compile if AnyVal is inherited by a trait, a non-@inline
class, or a class with an AnyRef parent somewhere. Added tests.
Added logging, like
[log extmethods] Inline class class Bippy spawns extension method.
Old: def getClass: Class[_ <: Bippy]
New: final def extension$getClass($this: Bippy): Class[_ <: Bippy]
Fixed what I hope was a bug in ExtensionMethods where the original
method params were dropped.
Since adding a NonNull parent was also inflicting an AnyRef on AnyVal
subclasses, suppressed that for those. Had the bright idea that AnyVal
could extend NotNull. It doesn't seem to accomplish much, but then,
I don't think NotNull accomplishes much. Still, maybe it's time to
restrict the ways one can use AnyVal so one can't do this:
scala> var x: AnyVal = _
x: AnyVal = null
|
| | |
|
|\ \
| | |
| | |
| | |
| | |
| | |
| | | |
Temporarily removed getClass from AnyVal to get build going.
Disabled anyval-childen test.
Fixed some other build problems.
Implemented step 1 + 2 of inline classes proposal.
|
|/ / |
|
|\| |
|
| |\
| | |
| | |
| | |
| | |
| | |
| | |
| | | |
Conflicts:
src/compiler/scala/reflect/internal/Symbols.scala
src/compiler/scala/reflect/internal/Types.scala
src/compiler/scala/tools/nsc/typechecker/Typers.scala
src/library/scala/reflect/api/Trees.scala
|
| | |
| | |
| | |
| | | |
The idea is that all operations that need to be synchronized are overriden in classes reflect.runtime.Synchronized*. Sometimes this applies to operations defined in SymbolTable, which can be directly overridden. Sometimes it is more convenient to generate SynchronizedClazz subclasses of SymbolTable classes Clazz. In the latter case, all instance creation must go over factory methods that can be overridden in the Synchronized traits.
|
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | | |
given moment (instead of throwing type errors). This avoids previous problems where we were creating fake error trees in some incorrect places like in type completers in Namers etc. Implicits relied heavily on type errors being thrown but performance should stay the same due to some explicit checks/returns.
Some of the problems involved how ambiguous error messages were collected/reported because it was very random (similarly for divergent implicits). This should be more explicit now. Reduced the number of unnecessary cyclic references being thrown (apart from those in Symbols/Types which don't have a context and need to stay for now as is).
Review by @paulp, @odersky.
|
| |/
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| | |
A short recap:
* Macro expansion now works finely for instance macro invocations
* Macros are now hidden behind -Xmacros
* Bodies of macros now have "import _context._" in their preamble
* Macros are now loaded from classpath, much like regular libraries
* Macros can now override methods (in that case macro expansion
does not crash if macro is not found, it just falls back to super)
Review by @odersky.
|
|/ |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Since I have established (no small effort, this) that there is no need
for certain important flags to be mutable for the entire lifetime of a
symbol, I have codified this knowledge by moving it out of the flags
entirely and into the inheritance hierarchy where its constant nature
can find true happiness.
AliasTypeSymbol ... it's an alias (forever!)
AbstractTypeSymbol ... it's an abstract type (forever!)
The only time DEFERRED is inspected is if the generic creation function
is called (e.g. "newTypeSymbol", not "newAliasType") in which case the
presence of the flag is used to determine which symbol to create. This
is mostly for legacy support.
Certain symbols were being created inconsistently with the others. Now
every symbol is created by invoking a creation method on the owner, with
the exception of FreeVar. I changed the owner of those from RootClass
to NoSymbol because there is no reason for them to pollute the symbol
hierarchy at the root.
The signature of cloneSymbolImpl is now
def cloneSymbolImpl(owner: Symbol, newFlags: Long): Symbol
There is an overload with the old signature which calls that one with no
flags. With this step, every symbol creation in trunk is performed with
knowledge of the initial flags, opening the door to many optimizations
in the Symbol and Type logic, not to mention boosting my sanity by at
least five sanity points.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
There are too many potential optimizations unavailable to us due to the
lack of bright lines among different kinds of symbols. For instance the
difference between a TypeSymbol which represents a type alias and one
which represents an abstract type is only whether the DEFERRED flag
is set. This creates issues.
1) There are many (many) places where tests are performed on every symbol
which could be done more efficiently and (especially) more verifiably
correctly with polymorphism.
2) TypeRefs based on those symbols are also checking that flag
constantly, in perpetuity. A symbol created as an alias is never (to the
best of my knowledge) going to intentionally morph into one representing
an abstract type, nor vice versa.
3) One has no guarantees, because anyone can set or reset the DEFERRED
flag at any time.
So tackling more than one problem at once herein:
1) I created canonical symbol creation points which take the flags as
an argument, so that there can be a difference between initializing a
symbol's flags and setting/resetting them at arbitrary times.
2) I structured all the symbol creators to take arguments in the
same order, which is:
def newXXX(name: Name, ..., pos: Position = NoPosition, flags: Long = 0L)
(Where "..." is for those symbols which require something
beyond the name to create, such as a TypeSkolem's origin.)
The name is first because it's the only always required argument.
I left but deprecated the variations which take (pos, name).
3) I created subclasses of TypeRef based on the information which
should be stable from creation time onward:
- args or no args?
- abstract type, type alias, or class?
2x3 == 6 and that's how many subclasses of TypeRef there are now. So
now, for example, every TypeRef doesn't have to carry null symInfoCache
and thisInfoCache fields for the benefit of the minority which use them.
I still intend to realize the gain possible once we can evade the fields
for pre and args without losing pattern matcher efficiency.
|
|
|
|
|
|
|
|
| |
Several large helpings of tedium later, fewer strings are being
discarded like so much refuse. Some names now cache a String, but only
"named Names", so it's not very many and they pay for themselves pretty
quickly. Many fewer name-related implicit conversions now taking place.
A number of efficiency related measures.
|
|
|
|
| |
Streamlining some of our more common operations.
|
|
|
|
|
| |
Deprecation warnings, unchecked warnings, "that's not the value you
think it is" warnings. Also eliminated a warning by fixing a warning bug.
|
| |
|
| |
|
|
|
|
|
|
|
|
| |
Took a more ambitious swing based on input from martin. Eliminated
the external map and gave annotations a more useful inheritance
hierarchy. Eliminated AnnotationInfoBase and made LazyAnnotationInfo an
AnnotationInfo (just like LazyType is a Type.) Review by odersky.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
I noticed a long time ago that calls to def annotations in Symbols
figured way, way too high in profiling output, but my earlier efforts to
modify it failed because I didn't understand the "accidental" service
it was supplying. Here is the key piece of the former implementation of
annotations:
- val annots1 = initialize.rawannots map {
- case x: LazyAnnotationInfo => x.annot()
- case x: AnnotationInfo => x
- } filterNot (_.atp.isError)
The first thing you might notice is that because it calls initialize,
any call to either annotations or (more frequently) a method like
"hasAnnotation" causes a symbol to be initialized. The upshot is that
taking away tens of thousands of calls to initialize means a certain
amount of "free lunch" is over.
The second thing is that this implementation lead to the allocation of
a new list on every call to annotations. 99.999% of the time it's the
same elements in the list. The fact that rawannots is typed as a list of
"AnnotationInfoBase" which may as well be AnyRef means you can't even
use mapConserve, but even mapConserve would be an abuse of the garbage
collector given how infrequently there is any change.
So here's what we have now:
1) Annotations are delivered from trees to symbols by way of an
externally positioned map, not a field on the symbol. It's done once.
The only overhead on a call to annotations now is a null check.
2) I added a small sprinkling of calls to initialize in sensible
locations.
3) The profiler impact is hard to believe, but this is reproducible.
For whatever reason the non-profiler wall clock time impact is not as
impressive.
My profiling target was the compilation of these 15 files:
src/library/scala/collection/generic/G*.scala
Before this patch, heap usage peaked at 60MB. After, 35MB. 40% drop in
profiler measured time elapsed. (Again, it's not like that outside the
profiler.) About a 55% drop in number of allocations. About a 40% drop
in total size of allocations.
+----------------------+------------------+-----------------+-----------------+
| Name | Time Diff (ms) | Old Time (ms) | New Time (ms) |
+----------------------+------------------+-----------------+-----------------+
| +---<All threads> | -19,569 | 52,496 | 32,926 |
+----------------------+------------------+-----------------+-----------------+
+----------------------------+--------------------+-----------------------+
| Packages and Classes | Objects (+/-) | Size (+/-) |
+----------------------------+--------------------+-----------------------+
| +---<Objects by classes> | -877,387 -56 % | -26,425,512 -37 % |
| | | | |
| +---char[] | -43,308 -2 % | -2,756,744 -3 % |
| | | | |
| +---java | -67,064 -3 % | -2,027,264 -2 % |
| | | | |
| +---scala | -745,099 -48 % | -19,021,760 -26 % |
+----------------------------+--------------------+-----------------------+
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Painstakingly winnowed out the most frequently duplicated code sequences
related to symbol cloning and substitution. Created canonical methods to
perform these actions and documented them. Key methods include:
def createFromClonedSymbols[T](syms: List[Symbol], tpe: Type)(creator: (List[Symbol], Type) => T): T
def deriveSymbols(syms: List[Symbol], symFn: Symbol => Symbol): List[Symbol]
def deriveType(syms: List[Symbol], symFn: Symbol => Symbol)(tpe: Type): Type
Many example usages enclosed with commit.
I did lots of timing tests, I find no material difference before and
after. Actually I won by four seconds in this incarnation:
Before - Total time: 7 minutes 55 seconds
After - Total time: 7 minutes 51 seconds
Review by moors.
|
|
|
|
|
| |
No review.
|
| |
|
|
|
|
|
|
|
|
| |
It represents a lot of work because the mutation flies fast and furious
and you can't even blink at these things without upsetting them. They're
a little hardier now, or at least we stand a better chance of changing
them. Open season on review.
|
|
|
|
|
|
| |
Starting to see a glimmer of the other side now. I nudged a few things
into more sensible places. No review.
|
|
|
|
|
| |
One little tiny oversight, those positions are finicky. No review.
|
|
|
|
|
|
|
| |
I was hoping to be further than this by now ("Accursed mutants!" *shakes
fist*) but you can't deny we've come a long way. Review is welcome, but
no review.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
A digression into motivations:
It's not there yet but the future looks bright. I have winnowed the
number of mutation points down and I will take it down much further.
There are only a few fundamental state changes which take place and
these can be made to happen in a systematic fashion at well-known
junctions.
1) Fresh symbols are allocated and (usually) assigned to a tree. 2)
Symbol flags and access are manipulated. 3) A (possibly lazy) info is
assigned to a symbol. 4) Synthetics are created or lazily positioned
for creation
Among the complications is that the symbol's info often cannot be
determined in a straightforward fashion lest cycles develop. Type
inference is obviously dependent on having some type information, so the
black art is to pursue a) only the right information and b) only as far
as necessary to avoid cycles.
Compounding the difficulty is that synthetic methods must be introduced
before the typer phase. Over time a variety of ad-hoc mechanisms have
evolved to deal with these difficulties, and they have gotten us a good
distance (we have bean setter/getters, case classes, copy methods,
default getters, and more) but there are big disadvantages:
- they are heavily custom fitted to the specific uses
- because of the intertwingling of namer and typer, it is all
only possible from the inside. Compiler plugins are shut out.
A particularly difficult scenario has recently arisen with case classes.
They now receive a ProductN parent, but because a class's parents must
be completed before it can be completed, we encounter this:
object Foo { type T = String }
case class Foo(x: Foo.T, y: Foo.T) { }
Now one of class Foo's parents is Product2[T, T]. So class Foo cannot be
completed without information from object Foo. But object Foo needs to
be given these synthetic methods:
def apply(x: T, y: T): Foo
def unapply(x: Foo): Option[(T, T)]
You can see these two have their hands all over one another.
The good news is that I have established in principle that the problem
can be overcome, for that use and I think in a sufficiently general way
that plugins will be able to perform this kind of method synthesis,
with the following approach. For synthetic methods which require type
information before they can be created (this is the catch-22: once type
information is available, it's too late to add new synthetic methods) we
create a "stub symbol" like so:
val sym = owner.newMethod("nameOfMethod")
sym setInfo stubMethodInfo
stubMethodInfo will be some very general method type like Any* => Any
(or Nothing => Any, it really depends on how or whether it is used),
whatever it takes to pass type checking. At the same time, we enter the
stub symbol into a map along with a thunk which creates the symbol and
tree the way we would if we had full type information.
Then, immediately after a class is typed, the template is examined for
stub method symbols and when found, they are updated with the symbol
info found in the map, assigned to the associated tree, and added to the
class template. This approach will probably break down under some uses,
but I think it can take us a long way.
So all these changes in Namers are about evolving it to a point where I
can roll out a principled division of responsibility for those entities
which cannot be naively typed, and to unify the several different
approaches to lazy typing under a consistent and predictable mechanism.
If anyone would like to review this, please be my guest, but you might
want to wait one more commit.
|
| |
|
|
|
|
|
|
|
|
|
|
| |
for now, left the old if(settings.YdepMethTpes.value) guards in comments
removed *.flags containing -Ydependent-method-types also updated one
check file with one fewer error
no review
|
|
|
|
|
|
|
|
| |
Having I think established this isn't merely a dance around the maypole,
I'm going to continue breaking down Namers because it's hard to figure
out where my stubbing mechanism is colliding with the existing logic.
Getting there, no review.
|
|
|
|
|
| |
Added some reverse engineered documentation to Namers, no review.
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
I can't go bear hunting without a clean gun. Basically I iterated over
Namers until I could understand it. I added a variety of documentation
there and elsewhere. There shouldn't be anything particularly behavioral
changing in this commit, but I did delete some years-old code (having
huge commented out blocks of way-out-of-date code is not a boon to
understanding) and the debugging output will look different. Better, one
can hope.
How about, review by moors.
|