| Commit message (Collapse) | Author | Age | Files | Lines |
|
|
|
|
|
|
|
|
| |
are in tail position)
Reverts commit 0ada0706746c9c603bf5bc8a0e6780e5783297cf.
Reverts commit 51c92f02229098d0b402a65a72267f7a17984022.
Reverts commit cdfbe8e39fbbec00c969cd74f117ae410b98b40b.
Reverts commit 796024c7429a03e974a7d8e1dc5c80b84f82467d.
|
|\
| |
| | |
Don't crash if cpsParam is not on the classpath.
|
| |
| |
| | |
When checking if a piece of code needs the continuations plugin, the cpsParam classes may not be on the class path. Assume it does not need it in that case.
|
| | |
|
| | |
|
|/
|
|
|
|
|
|
|
|
|
| |
Adds a stack of context trees to AnnotationChecker(s). Here, it is used to
enforce that adaptAnnotations will only adapt the annotation of a return
expression if the expected type is a CPS type.
The remove-tail-return transform is reasonably general, covering cases such as
try-catch-finally. Moreover, an error is thrown if, in a CPS method, a return
is encountered which is not in a tail position such that it will be removed
subsequently.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
A must read: "SIP: Scala Reflection":
https://docs.google.com/document/d/1Z1VhhNPplbUpaZPIYdc0_EUv5RiGQ2X4oqp0i-vz1qw/edit
Highlights:
* Architecture has undergone a dramatic rehash.
* Universes and mirrors are now separate entities:
universes host reflection artifacts (trees, symbols, types, etc),
mirrors abstract loading of those artifacts (e.g. JavaMirror loads stuff
using a classloader and annotation unpickler, while GlobalMirror uses
internal compiler classreader to achieve the same goal).
* No static reflection mirror is imposed on the user.
One is free to choose between lightweight mirrors and full-blown
classloader-based mirror (read below).
* Public reflection API is split into scala.reflect.base and scala.reflect.api.
The former represents a minimalistic snapshot that is exactly enough to
build reified trees and types. To build, but not to analyze - everything smart
(for example, getting a type signature) is implemented in scala.reflect.api.
* Both reflection domains have their own universe: scala.reflect.basis and
scala.reflect.runtime.universe. The former is super lightweight and doesn't
involve any classloaders, while the latter represents a stripped down compiler.
* Classloader problems from 2.10.0-M3 are solved.
* Exprs and type tags are now bound to a mirror upon creation.
* However there is an easy way to migrate exprs and type tags between mirrors
and even between universes.
* This means that no classloader is imposed on the user of type tags and exprs.
If one doesn't like a classloader that's there (associated with tag's mirror),
one can create a custom mirror and migrate the tag or the expr to it.
* There is a shortcut that works in most cases. Requesting a type tag from
a full-blown universe will create that tag in a mirror that corresponds to
the callsite classloader aka `getClass.getClassLoader`. This imposes no
obligations on the programmer, since Type construction is lazy, so one
can always migrate a tag into a different mirror.
Migration notes for 2.10.0-M3 users:
* Incantations in Predef are gone, some of them have moved to scala.reflect.
* Everything path-dependent requires implicit prefix (for example, to refer
to a type tag, you need to explicitly specify the universe it belongs to,
e.g. reflect.basis.TypeTag or reflect.runtime.universe.TypeTag).
* ArrayTags have been removed, ConcreteTypeTag have been renamed to TypeTags,
TypeTags have been renamed to AbsTypeTags. Look for the reasoning in the
nearby children of this commit. Why not in this commit? Scroll this message
to the very bottom to find out the reason.
* Some of the functions have been renamed or moved around.
The rule of thumb is to look for anything non-trivial in scala.reflect.api.
Some of tree build utils have been moved to Universe.build.
* staticModule and staticClass have been moved from universes to mirrors
* ClassTag.erasure => ClassTag.runtimeClass
* For the sake of purity, type tags no longer have erasures.
Use multiple context bounds (e.g. def foo[T: ru.TypeTag : ClassTag](...) = ...)
if you're interested in having both erasures and types for type parameters.
* reify now rolls back macro applications.
* Runtime evaluation is now explicit, requires import scala.tools.reflect.Eval
and scala-compiler.jar on the classpath.
* Macro context now has separate universe and mirror fields.
* Most of the useful stuff is declared in c.universe,
so be sure to change your "import c.universe._" to "import c.mirror._".
* Due to the changes in expressions and type tags, their regular factories
are now really difficult to use. We acknowledge that macro users need to
frequently create exprs and tags, so we added old-style factories to context.
Bottom line: almost always prepend Expr(...)/TypeTag(...) with "c.".
* Expr.eval has been renamed to Expr.splice.
* Expr.value no longer splices (it can still be used to express cross-stage
path-dependent types as specified in SIP-16).
* c.reifyTree now has a mirror parameter that lets one customize the initial
mirror the resulting Expr will be bound to. If you provide EmptyTree, then
the reifier will automatically pick a reasonable mirror (callsite classloader
mirror for a full-blown universe and rootMirror for a basis universe).
Bottom line: this parameter should be EmptyTree in 99% of cases.
* c.reifyErasure => c.reifyRuntimeClass.
Known issues:
* API is really raw, need your feedback.
* All reflection artifacts are now represented by abstract types.
This means that pattern matching against them will emit unchecked warnings.
Adriaan is working on a patch that will fix that.
WARNING, FELLOW CODE EXPLORER! You have entered a turbulence zone.
For this commit and its nearby parents and children
tests are not guaranteed to work. Things get back to normal only after
the "repairs the tests after the refactoring spree" commit.
Why so weird? These twentish changesets were once parts of a humongous blob,
which spanned 1200 files and 15 kLOC. I did my best to split up the blob,
so that the individual parts of the code compile and make sense in isolation.
However doing the same for tests would be too much work.
|
|
|
|
|
|
|
| |
And for a bunch of other tickets where we unleash a stack
trace rather than printing a sensible error message. But
SI-3718 is a continuations plugin crash, now a reasonable
if somewhat vague error.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
sort field accessors, necessary after typers -- apparently...
don't throw TypeError, use issueTypeError
don't run patmat phase when -Xoldpatmat
only virtualize matches when -Xexperimental
recycle cps type of match for re-typechecking:
when one of the internal cps-type-state annotations is present, strip all CPS annotations
a cps-type-state-annotated type makes no sense as an expected type (matchX.tpe is used as pt in translateMatch)
don't synth FunctionN impls during typer, only do this for PartialFunction
updated check now function synth for match is deferred until uncurry
patmat-transform try/catch with match in cps
cleanup in selective anf
remove TODO: can there be cases that are not CaseDefs -- nope
|
|
|
|
|
|
|
|
|
| |
detect partialfunction in cpsannotationchecker
emit apply/isDefinedAt if PF has @cps targs (applyOrElse can't be typed)
further hacky improvements to selective anf
better try/catch support in selective cps using freshly minted anonfun match
make virtpatmat resilient to scaladoc (after uncurry, don't translate matches
TODO: factor out translation all together so presentation compiler/scaladoc can skip it)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
typers&patmatvirtualizer have ad-hoc support for dropping annotations
in a way that makes the CPS plugins happy... this is not ideal,
but unless virtpatmat runs after the plugin phases, I don't see how to solve it
running virtpatmat after the CPS plugin would mean the pattern matching evaluation
cannot be captured by CPS, so it's not even desirable to move it to a later phase
- typedIf must lub annotated types
- drop selector.tpe's annotations
- drop annots in matchEnd's argument type
- deal with annots in casts synth by in virtpatmat
(drop them from type arg to asInstanceof, recover them using type ascription)
- workaround skolemize existential dropping annots
CPS is the main reason why typedMatchAnonFun is not used anymore,
and PartialFunction synthesis is moved back to uncurry (which is quite painful due to labeldefs being so broken)
we can't synth partialfunction during typer since T @cps[U] does not conform to Any,
so we can't pass it as a type arg to PartialFunction, so we can't type a cps-transformed PF
after the CPS plugin, T @cps[U] becomes ControlContext[...], which is a type we can pass
to PartialFunction
virtpatmat is now also run until right before uncurry
(so, can't use isPastTyper, although it means more or less the same thing -- we don't run after uncurry)
the main functional improvements are in the selective ANF transform
its treatment of labeldefs was broken: for example,
LabelDef L1; LabelDef L2 --> DefDef L1; L1(); DefDef L2; L2()
but this does not take into account L1 may jump over L2 to another label
since methods always return (or fail), and the ANF transform generates ValDefs
to store the result of those method calls, both L1 and L2 would always be executed
(so you would run a match with N cases N times, with each partial run starting at a later case)
also fixed a couple of weird bugs in selective anf that caused matches to be duplicated
(with the duplicate being nested in the original)
since label defs are turned into method defs, and later defs
will be nested in the flatMap calls on the controlcontext
yielded by earlier statements, we reverse the list of method
definitions, so that earlier (in the control flow sense) methods
are visible in later ones
selective CPS now generates a catch that's directly digestible by backend
|
|
|
|
| |
This reverts commit da35106f81a5c24e78ff51c95e10052ad4f23b18.
|
|
|
|
|
|
|
|
| |
Avoid explicit type arguments which don't conform to bounds
where they could be successfully inferred.
I had to disable one "neg" test which is no longer neg.
Can anyone clue me in as to whether it is important?
|
| |
|
|
|
|
|
|
|
| |
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.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
I guess these are self-explanatory.
@inline final def afterErasure[T](op: => T): T = afterPhase(currentRun.erasurePhase)(op)
@inline final def afterExplicitOuter[T](op: => T): T = afterPhase(currentRun.explicitouterPhase)(op)
...
@inline final def beforeTyper[T](op: => T): T = beforePhase(currentRun.typerPhase)(op)
@inline final def beforeUncurry[T](op: => T): T = beforePhase(currentRun.uncurryPhase)(op)
This commit is basically pure substitution. To get anywhere interesting
with all the phase-related bugs we have to determine why we use atPhase
and capture that reasoning directly. With the exception of erasure, most
phases don't have much meaning outside of the compiler. How can anyone
know why a block of code which says atPhase(explicitouter.prev) { ... }
needs to run there? Too much cargo cult, and it should stop. Every usage
of atPhase should be commented as to intention.
It's easy to find bugs like
atPhase(uncurryPhase.prev)
which was probably intended to run before uncurry, but actually runs
before whatever happens to be before uncurry - which, luckily enough, is
non-deterministic because the continuations plugin inserts phases
between refchecks and uncurry.
% scalac -Xplugin-disable:continuations -Xshow-phases
refchecks 7 reference/override checking, translate nested objects
uncurry 8 uncurry, translate function values to anonymous classes
% scalac -Xshow-phases
selectivecps 9
uncurry 10 uncurry, translate function values to anonymous classes
Expressions like atPhase(somePhase.prev) are never right or are at best
highly suboptimal, because most of the time you have no guarantees about
what phase precedes you. Anyway, I think most or all atPhases expressed
that way only wanted to run before somePhase, and because one can never
be too sure without searching for documentation whether "atPhase" means
before or after, people err on the side of caution and overshoot by a
phase. Unfortunately, this usually works. (I prefer bugs which never work.)
|
|\ |
|
| |
| |
| |
| | |
multi-argument list methods.
|
|/ |
|
|
|
|
|
|
| |
Changed most symbol creations to be consistent with all the others.
Opportunistically streamlined various call sites. Moved some phase-specific
methods out of Symbol to somewhere more appropriate (like that phase.)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
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.
|
|
|
|
|
|
| |
We're not buying anything with it off by default. Added a
-P:continuations:disable option to disable. -P:continuations:enable
lives on as a no-op. Review by @odersky.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
AnnotationInfo inertia takes me into continuations.
And kept carrying me until I was carried away. The changes are mostly of
the janitorial variety, just doing my part to make the interesting logic
visible without being buried in low level compiler plumbing.
Added at least one seriously convenient convenience method:
tree modifyType fn
// equivalent to if (tree.tpe == null) tree else tree setType fn(tree.tpe)
This is the analogue to the recently added:
symbol modifyInfo fn
// same idea
It's like having our carpets steam cleaned when we can keep pushing
until machinery stays in the machine and the relevant logic stands
gloriously on top. You'll eventually exclaim, "I didn't even know these
carpets were that color!"
|
| |
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Reversed the values of "is" and "is not" in recent for comprehension
deprecation.
DO NOT BLOW HATCH
REPEAT
DO NOT BLOW HATCH
"Roger! Hatch blown."
Events reveal it was all baby, no bathwater. It turns out that the
specification is merely a document, not infallible holy writ as we
had all previously believed. So it is not the ABSENCE of val in a for
comprehension assignment which is deprecated, it is the PRESENCE of val.
Summarizing again, more accurately perhaps:
for (x <- 1 to 5 ; y = x) yield x+y // THAT's the one
for (val x <- 1 to 5 ; y = x) yield x+y // fail
for (val x <- 1 to 5 ; val y = x) yield x+y // fail
for (x <- 1 to 5 ; val y = x) yield x+y // deprecated
No review.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Adding some Sets/Maps to perRunCaches, and eliminating ambiguously named
imports.
Did a tour of the compiler adding a few longer-lived mutable structures
to the per-run cache clearing mechanism. Some of these were not a big
threat, but there is (almost) literally no cost to tracking them and the
fewer mutable structures which are created "lone wolf style" the easier
it is to spot the one playing by his own rules.
While I was at it I followed through on long held ambition to eliminate
the importing of highly ambiguous names like "Map" and "HashSet" from
the mutable and immutable packages. I didn't quite manage elimination
but it's pretty close. Something potentially as pernicious which I
didn't do much about is this import:
import scala.collection._
Imagine coming across that one on lines 407 and 474 of a 1271 file.
That's not cool. Some poor future programmer will be on line 1100 and
use "Map[A, B]" in some function and only after the product has shipped
will it be discovered that the signature is wrong and the rocket will
now be crashing into the mountainside straightaway. No review.
|
|
|
|
|
|
| |
Refactored reflection into reflect.api and reflect.internal. Severed the
last remaining dependency on reflect.generic. Review by extempore.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
introduce NullaryMethodType to disambiguate PolyType
motivation:
given `def foo[T]: (T, T)` and `type Foo[T] = (T, T)`,
`foo.info` and `TypeRef(_, Foo, Nil).normalize` are both
`PolyType(List(T), Pair[T, T])`
uncurry has been relying on an ugly hack to distinguish these
cases based on ad-hoc kind inference without this distinction,
the type alias's info (a type function) would be transformed to
`PolyType(List(T), MethodType(Nil, Pair[T, T]))`
anonymous type functions are being used more often (see #2741, #4017,
#4079, #3443, #3106), which makes a proper treatment of PolyTypes more
pressing
change to type representation:
PolyType(Nil, tp) -> NullaryMethodType(tp)
PolyType(tps, tp) -> PolyType(tps, NullaryMethodType(tp)) (if the polytype denoted a polymorphic nullary method)
PolyType(Nil, tp) is now invalid
the kind of a PolyType is * iff its resulttype is a NullaryMethodType or a MethodType (i.e., it's a polymorphic value)
in all other cases a PolyType now denotes a type constructor
NullaryMethodType is eliminated during uncurry
pickling:
for backwards compatibility, a NullaryMethodType(tp) is still pickled as a PolyType(Nil, tp),
unpickling rewrites pre-2.9-pickled PolyTypes according to the expected kind of the unpickled type (similar to what we used to do in uncurry)
a pickled PolyType(Nil, restpe) is unpickled to NullaryMethodType(restpe)
a pickled PolyType(tps, restpe) is unpickled to PolyType(tps, NullaryMethodType(restpe)) when the type is expected to have kind *
the rewrite probably isn't complete, but was validated by compiling
against the old scalacheck jar (which has plenty of polymorphic nullary
methods) nevertheless, this commit includes a new scalacheck jar
summary of the refactoring:
* PolyType(List(), tp) or PolyType(Nil, tp) or PolyType(parms, tp) if params.isEmpty ==> NullaryMethodType(tp)
* whenever there was a case PolyType(tps, tp) (irrespective of tps isEmpty), now need to consider the
case PolyType(tps, NullaryMethodType(tp)); just add a case NullaryMethodType(tp), since usually:
- there already is a PolyType case that recurses on the result type,
- the polytype case applied to empty and non-empty type parameter lists alike
* tp.resultType, where tp was assumed to be a PolyType that represents a polymorphic nullary method type
before, tp == PolyType(tps, res), now tp == PolyType(tps, NullaryMethodType(res))
* got bitten again (last time was dependent-method types refactoring)
by a TypeMap not being the identity when dropNonConstraintAnnotations
is true (despite having an identity apply method). Since asSeenFrom
is skipped when isTrivial, the annotations aren't dropped. The
cps plugin relies on asSeenFrom dropping these annotations for
trivial types though. Therefore, NullaryMethodType pretends to
never be trivial. Better fix(?) in AsSeenFromMap: `if(tp.isTrivial)
dropNonContraintAnnotations(tp) else ...`
TODO: scalap and eclipse
review by odersky, rytz
|
|
|
|
|
|
| |
Warded off an NPE (which only arose on erroneous code) in the
continuations plugin. Closes #4132, no review.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Mopping up after the deprecation of exit and error. It is decidedly
non-trivial (at least for the IDE-impaired) to be completely sure of
which error function was being called when there were about twenty with
the same signature in trunk and they are being variously inherited,
imported, shadowed, etc. So although I was careful, the possibility
exists that something is now calling a different "error" function than
before. Caveat programmer.
(And let's all make it our policy not to name anything "error" or "exit"
from here on out....) No review.
|
|
|
|
|
|
|
|
|
| |
the same issues as JavaConversions with respect to overloading implicit
methods making them inaccessible to view bounds. Fixed JavaConverters.
Added a warning for when people overload parameterized implicits: in
almost all cases the name is irrelevant so there's little point in
unwittingly suffering degraded functionality. No review.
|
|
|
|
|
|
|
|
|
|
|
| |
Determined that half a dozen ways of checking for varargs and
by-name-ness in param lists exceeded the legal limit. Also assessed that
names which are only used as type names would be a lot easier to deal
with if we created them as type names up front. Performed the changes
implied by the preceding along with a partial cleanup on TreeInfo which
one can see hasn't had a good look in a long time. (And still hasn't.)
No review.
|
|
|
|
|
|
|
|
|
| |
The inliner now looks harder for method implementations. In addition
to the receiver, it will find those in directly mixed in traits and
in superclasses. It still won't find those in traits mixed only into
superclasses, as that didn't come quite so easily. Closes #3234 but I'll
be opening up another ticket. Review by dragos.
|
|
|
|
|
|
|
|
| |
transformer. It now aligns tree nodes that contain
names to the symbol name that was substituted. Before
this change identifiers may refer to one symbol, while
the name they carry would resovlve to another one.
|
| |
|
|
|
|
|
|
| |
fixes the unfounded "name clash between inherited members" error. review
by dragos.
|
| |
|
| |
|
| |
|
| |
|
|
|
|
|
|
| |
continuations plugin will now report a nice error message if it is not
enabled and encounters an @cps expression. review by rytz
|
|
|
|
|
|
|
|
|
| |
method on Iterator called collect which I had to remove, because if
the method is overloaded it puts a bullet in the type inference, an
intolerable result for a function which takes a partial function as its
argument. I don't think there's much chance of confusion, but I put a
migration warning on collect just in case. No review.
|
| |
|
|
|
|
|
|
| |
added support for continuations in try/catch blocks. review by
community.
|
|
|