| Commit message (Collapse) | Author | Age | Files | Lines |
|---|
| ... | |
| |\| |
| | |
| | |
| | |
| | |
| | | |
Conflicts:
src/library/scala/collection/MapLike.scala
src/library/scala/collection/SortedMapLike.scala
|
| | |\ \
| | | |
| | | | |
Fix range positions when applying anonymous classes.
|
| | | | |
| | | |
| | | |
| | | | |
@odersky
|
| | |/ / |
|
| | |/
| |
| |
| | |
A singleton type is a type ripe for enumeration.
|
| | |\
| | |
| | | |
Fix SI-5284.
|
| | | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | | |
The problem was the false assumption that methods specialized
on their type parameter, such as this one:
class Foo[@spec(Int) T](val x: T) {
def bar[@spec(Int) S >: T](f: S => S) = f(x)
}
have their normalized versions (`bar$mIc$sp`) never called
from the base specialization class `Foo`.
This meant that the implementation of `bar$mIc$sp` in `Foo`
simply threw an exception.
This assumption is not true, however. See this:
object Baz {
def apply[T]() = new Foo[T]
}
Calling `Baz.apply[Int]()` will create an instance of the
base specialization class `Foo` at `Int`.
Calling `bar` on this instance will be rewritten by
specialization to calling `bar$mIc$sp`, hence the error.
So, we have to emit a valid implementation for `bar`,
obviously.
Problem is, such an implementation would have conflicting
type bounds in the base specialization class `Foo`, since
we don't know if `T` is a subtype of `S = Int`.
In other words, we cannot emit:
def bar$mIc$sp(f: Int => Int) = f(x) // x: T
without typechecking errors.
However, notice that the bounds are valid if and only if
`T = Int`. In the same time, invocations of `bar$mIc$sp` will only
be emitted in callsites where the type bounds hold.
This means we can cast the expressions in method applications
to the required specialized type bound.
The following changes have been made:
1) The decision of whether or not to create a normalized
version of the specialized method is not done on the
`conflicting` relation anymore.
Instead, it's done based on the `satisfiable` relation,
which is true if there is possibly an instantiation of
the type parameters where the bounds hold.
2) The `satisfiable` method has a new variant called
`satisfiableConstraints`, which does unification to
figure out how the type parameters should be instantiated
in order to satisfy the bounds.
3) The `Duplicators` are changed to transform a tree
using the `castType` method which just returns the tree
by default.
In specialization, the `castType` in `Duplicators` is overridden,
and uses a map from type parameters to types.
This map is obtained by `satisfiableConstraints` from 2).
If the type of the expression is not equal to the expected type,
and this map contains a mapping to the expected type, then
the tree is cast, as discussed above.
Additional tests added.
Review by @dragos
Review by @VladUreche
Conflicts:
src/compiler/scala/tools/nsc/transform/SpecializeTypes.scala
src/compiler/scala/tools/nsc/typechecker/Duplicators.scala
|
| | |\ \
| | |/
| |/| |
SI-4842 Forbid access to in-construction this in self-constructor args
|
| | | |
| | |
| | |
| | |
| | |
| | |
| | | |
The check was already in place for direct calls to the super constructor.
Without this additional check, ExplicitOuter crashes, as it doesn't create
an $outer pointer for the constructor-arg scoped inner object, but expects
one to exist when transforming the Outer.this reference.
|
| | | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | | |
As usual, .tpe -> .tpeHK. As a side note following an old theme,
if symbols of type parameters knew that they were symbols of type
parameters, they could call tpeHK themselves rather than every call
site having to do it. It's the operation which injects dummies which
should require explicit programmer action, not the operation which
faithfully reproduces the unapplied type. Were it that way, errors could
be caught much more quickly via ill-kindedness.
Seems like an improvement over lurking compiler crashes at every call
to tparam.tpe.
|
| | |/
| |
| |
| | |
Empty statements are A-OK. Closes SI-5910. Review by @dragos.
|
| |/ |
|
| |
|
|
|
| |
There doesn't seem to be any way to do that by adding
a synthetic annotation.
|
| |\
| |
| | |
SI-4831 Fix ambiguous import detection for renamed imports.
|
| | | |
|
| | | |
|
| |/ |
|
| |\
| |
| | |
Fix SI-5853.
|
| | |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| | |
This solves two issues.
First, up to now the newly generated symbols for normalized
members were not being added to the declaration list of the
owner during `specialize`. Now they are.
Second, during `extmethods`, the extension methods generated
get an additional curried parameter list for `$this`.
Trouble was, after that, during `uncurry` and before `specialize`,
these curried parameter lists were merged into one list.
Specialization afterwards treats extension methods just
like normal methods and generates new symbols without the
curried parameter list.
The `extensionMethod` now takes this into account by checking
if the first parameter of a potential extension method has
the name `$this`.
Review by @dragos.
Review by @odersky.
|
| | |
| |
| |
| |
| |
| |
| |
| |
| |
| | |
Reduce time spent in lubs by making depth more adaptive. It now takes into account separately the depth of the lub types and the maximal depth of theior base type sequences. It cuts depth more aggressively if it is the base types instead of the types themselves that grow deep.
The old truncation behavior is retained under option -Xfull-lubs
Another change is that we now track depth more precisely, which should also help or at least allow better statistics.
Also added statistics that measure #lubs and time spent in them.
|
| | |
| |
| |
| |
| | |
This protects everyone from the confusion caused by stuff like this:
https://issues.scala-lang.org/browse/SI-5884
|
| | |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| | |
Before 2.10 we had a notion of ClassManifest that could be used to retain
erasures of abstract types (type parameters, abstract type members) for
being used at runtime.
With the advent of ClassManifest (and its subtype Manifest)
it became possible to write:
def mkGenericArray[T: Manifest] = Array[T]()
When compiling array instantiation, scalac would use a ClassManifest
implicit parameter from scope (in this case, provided by a context bound)
to remember Ts that have been passed to invoke mkGenericArray and
use that information to instantiate arrays at runtime (via Java reflection).
When redesigning manifests into what is now known as type tags, we decided
to explore a notion of ArrayTags that would stand for abstract and pure array
creators. Sure, ClassManifests were perfectly fine for this job, but they did
too much - technically speaking, one doesn't necessarily need a java.lang.Class
to create an array. Depending on a platform, e.g. within JavaScript runtime,
one would want to use a different mechanism.
As tempting as this idea was, it has also proven to be problematic.
First, it created an extra abstraction inside the compiler. Along with class tags
and type tags, we had a third flavor of tags - array tags. This has threaded the
additional complexity though implicits and typers.
Second, consequently, when redesigning tags multiple times over the course of
Scala 2.10.0 development, we had to carry this extra abstraction with us, which
exacerbated the overall feeling towards array tags.
Finally, array tags didn't fit into the naming scheme we had for tags.
Both class tags and type tags sound logical, because, they are descriptors for
the things they are supposed to tag, according to their names.
However array tags are the odd ones, because they don't actually tag any arrays.
As funny as it might sound, the naming problem was the last straw
that made us do away with the array tags. Hence this commit.
|
| |/ |
|
| |\
| |
| | |
SI-5313 Revert to two traversals in substThisAndSym.
|
| | |
| |
| |
| | |
Partially reverts 334872e.
|
| |\ \
| | |
| | |
| | |
| | | |
Unreachability analysis for pattern matches
Thanks for reviewing, @retronym!
|
| | |/
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| | |
Analyze matches for unreachable cases.
A case is unreachable if it implies its preceding cases.
Call `C` the formula that is satisfiable if the considered case matches.
Call `P` the formula that is satisfiable if the cases preceding it match.
The case is reachable if there is a model for `-P /\ C`.
Thus, the case is unreachable if there is no model for `-(-P /\ C)`,
or, equivalently, `P \/ -C`, or `C => P`.
Unreachability needs a more precise model for type and value tests than exhaustivity.
Before, `{case _: Int => case 1 =>}` would be modeled as `X = Int \/ X = 1.type`,
and thus, the second case would be reachable if we can satisfy `X != Int /\ X = 1.type`.
Of course, the case isn't reachable, yet the formula is satisfiable, so we must augment
our model to take into account that `X = 1.type => X = Int`.
This is done by `removeVarEq`, which models the following axioms about equality.
It does so to retain the meaning of equality after replacing `V = C` (variable = constant)
by a literal (fresh symbol). For each variable:
1. a sealed type test must result in exactly one of its partitions being chosen
(the core of exhaustivity)
2. when a type test is true, tests of super types must also be true,
and unrelated type tests must be false
For example, `V : X ::= A | B | C`, and `A => B` (since `A extends B`).
Coverage (1) is formulated as: `A \/ B \/ C`, and the implications of (2) are simply
`V=A => V=B /\ V=X`, `V=B => V=X`, `V=C => V=X`.
Exclusion for unrelated types typically results from matches such as `{case SomeConst =>
case OtherConst => }`. Here, `V=SomeConst.type => !V=OtherConst.type`. This is a
conservative approximation. If these constants happen to be the same value dynamically
(but the types don't tell us this), the last case is actually unreachable. Of course we
must err on the safe side.
We simplify the equality axioms as follows (in principle this could be done by the
solver, but it's easy to do before solving). If we've already excluded a pair of
assignments of constants to a certain variable at some point, say `(-A \/ -B)`, then
don't exclude the symmetric one `(-B \/ -A)`. (Nor the positive implications `-B \/ A`,
or `-A \/ B`, which would entail the equality axioms falsifying the whole formula.)
TODO: We should also model dependencies between variables: if `V1` corresponds to
`x: List[_]` and `V2` is `x.hd`, `V2` cannot be assigned at all when `V1 = null` or
`V1 = Nil`. Right now this is implemented hackily by pruning counter-examples in exhaustivity.
Unreachability would also benefit from a more faithful representation.
I had to refactor some of the framework, but most of it is shared with exhaustivity. We
must allow approximating tree makers twice, sharing variables, but using different
approximations for values not statically known. When considering reachability of a case,
we must assume, for example, that its unknown guard succeeds (otherwise it would wrongly
be considered unreachable), whereas unknown guards in the preceding cases must be
considered to fail (otherwise we could never get to those case, and again, it would
falsely be considered unreachable).
Since this analysis is relatively expensive, you may opt-out using `-Xno-patmat-analysis`
(or annotating the selector with @unchecked). We hope to improve the performance in the
near future. -Ystatistics has also been extended to provide some numbers on time spent in
the equality-rewrite, solving and analyzing.
|
| |/ |
|
| |\
| |
| | |
SI-4579 Yoke the power of lisp.scala as a stress for the optimizer.
|
| | |
| |
| |
| | |
The reported bug was fixed between 2.10.0-M1 and 2.10.0-M2.
|
| |\ \
| | |
| | | |
fixes for exhaustivity
|
| | |/ |
|
| |\ \
| |/
|/| |
Fix a NSDNHAO in extension methods.
|
| | |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| | |
A bridge method, created when we override a method from
a superclass and refine the return type, was appearing
as an overloaded alternative. (`erasure` doesn't create
new scopes, so the bridges it builds are visible at
earlier phases.)
The problem was masked when compiling with specialization,
which *does* create a new scope, shielding the code in
question from the artefacts of erasure.
To fix the problem, we filter out bridge methods from
the overloaded alternatives returned by `.decl`, as would
happen internally in `.member`.
|
| |/
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Are these -msil checkfiles used in some secret fashion? The level of
activity suggest otherwise. Since scala-nightly-msil has been disabled
for over a year, it's an easy rm unless someone speaks up.
% tools/rm-orphan-checkfiles
Scanning for orphan check files...
rm 'test/disabled/run/code.check'
rm 'test/files/jvm/t1652.check'
rm 'test/files/neg/macro-argtype-mismatch.check'
rm 'test/files/neg/macro-noncompilertree.check'
rm 'test/files/neg/macro-nontree.check'
rm 'test/files/run/Course-2002-01-msil.check'
rm 'test/files/run/Course-2002-02-msil.check'
rm 'test/files/run/Course-2002-03-msil.check'
rm 'test/files/run/Course-2002-04-msil.check'
rm 'test/files/run/Course-2002-08-msil.check'
rm 'test/files/run/Course-2002-09-msil.check'
rm 'test/files/run/Course-2002-10-msil.check'
rm 'test/files/run/absoverride-msil.check'
rm 'test/files/run/bitsets-msil.check'
rm 'test/files/run/boolord-msil.check'
rm 'test/files/run/bugs-msil.check'
rm 'test/files/run/impconvtimes-msil.check'
rm 'test/files/run/infix-msil.check'
rm 'test/files/run/iq-msil.check'
rm 'test/files/run/macro-invalidret-doesnt-conform-to-impl-rettype.check'
rm 'test/files/run/macro-rettype-mismatch.check'
rm 'test/files/run/misc-msil.check'
rm 'test/files/run/promotion-msil.check'
rm 'test/files/run/richs-msil.check'
rm 'test/files/run/runtime-msil.check'
rm 'test/files/run/tuples-msil.check'
rm 'test/pending/jvm/t1464.check'
rm 'test/pending/run/subarray.check'
rm 'test/pending/run/t0446.check'
rm 'test/pending/run/t5629.check'
Scanning for orphan flags files...
rm 'test/files/neg/macro-argtype-mismatch.flags'
rm 'test/files/neg/macro-noncompilertree.flags'
rm 'test/files/neg/macro-nontree.flags'
rm 'test/files/pos/anyval-children.flags'
rm 'test/files/pos/t3097.flags'
rm 'test/files/run/macro-invalidret-doesnt-conform-to-impl-rettype.flags'
rm 'test/files/run/macro-rettype-mismatch.flags'
|
| |\
| |
| | |
fix SI-5829: refinement typeref has a prefix
|
| | | |
|
| |\ \
| | |
| | | |
Test case closes SI-5041.
|
| | |/ |
|
| |/
|
|
| |
The unchecked warning departed sometime between 4afae5be...278a225.
|
| |
|
|
|
|
|
|
| |
We were getting away with this somehow, but the types are wrong after
typer and that sort of thing is noticed by more people now. I took the
opportunity to add our first -Ycheck:all test, which is at least as much
about helping -Ycheck:all remain in good working order as it is about
this test.
|
| |
|
|
|
|
|
|
| |
Move now-passing SI-963 test into neg.
Test for partial specialization.
Pending test for SI-5008.
Pending test for SI-4649.
Abstract array type test.
|
| | |
|
| |\
| |
| | |
virtpatmat: treemaker approximation refactorings and exhaustivity
|
| | |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| | |
We check exhaustivity by representing a match as a formula in
finite-domain propositional logic (FDPL) that is false when
the match may fail. The variables in the formula represent
tested trees in the match (type tests/value equality tests).
The approximation uses the same framework as the CSE analysis.
A matrix of tree makers is turned into a DAG, where sharing
represents the same value/type being tested.
We reduce FDPL to Boolean PL as follows.
For all assignments, V_i = c_i_j, we introduce a proposition
P_i_j that is true iff V_i is equal to the constant c_i_j,
for a given i, and all j, P_i_j are mutually exclusive
(a variable cannot have multiple values).
If the variable's domain is closed, we assert that one of P_i_j
must be true for each i and some j. The conjunction of these
propositions constitutes the equality axioms.
After going through negational normal form to conjunctive normal
form, we use a small SAT solver (the DPLL algorithm) to find
a model under which the equational axioms hold but the match fails.
The formula: EqAxioms /\ -MatchSucceeds.
Note that match failure expresses nicely in CNF: the
negation of each case (which yields a disjunction) is anded.
We then turn this model into variable assignments
(what's the variable (not) equal to, along with recursive
assignments for its fields). Valid assignments (no ill-typed
field assignments) are then presented to the user as counter examples.
A counter example is a value, a type test, a constructor call
or a wildcard. We prune the example set and only report the
most general examples. (Finally, we sort the output to yield stable,
i.e. testable, warning messages.)
A match is only checked for exhaustivity when the type of the
selector is "checkable" (has a sealed type or is a tuple
with at least one component of sealed type).
We consider statically known guard outcomes, but generally back
off (don't check exhaustivity) when a match has guards
or user-defined extractor calls. (Sometimes constant folding
lets us statically decide a guard.)
We ignore possibly failing null checks (which are performed
before calling extractors, for example), though this could
be done easily in the current framework. The problem is false
positives. People don't usually put nulls in tuples or lists.
To improve the exhaustivity checks, we rewrite `List()` to Nil.
TODO: more general rewrite of List(a, b, ..., z) to `a :: b :: ... :: z`.
When presenting counter examples, we represent lists in the
user-friendly List(a,...,z) format. (Similarly for tuples.)
There are no exhaustivity checks for a match-defined PartialFunction.
misc notes:
- fix pure case of dpll solver
impure set (symbol that occurs both as a positive and negative literal)
was always empty since I was looking for literals (which are not equal if positivity is not equal)
but should have been looking for symbols
- FDPL -> BoolPL translation collects all syms in props
since propForEqualsTo generates an Or, must traverse the prop
rather than assuming only top-level Syms are relevant...
also, propForEqualsTo will not assume Or'ing a whole domain is equivalent to True
(which it isn't, since the type test may fail in general)
- improve counter example description
- treat as constructor call when we either have definite type information about a real class,
or we have no equality information at all, but the variable's type is a class
and we gathered constraints about its fields (typically when selector is a tuple)
- flatten a :: b :: ... :: Nil to List(a, b, ...)
- don't print tuple constructor names, so instead of "Tuple2(a, b)", say "(a, b)"
- filter out more statically impossible subtypes
the static types convey more information than is actually checkable at run time
this is good, as it allows us to narrow down the subtypes of a sealed type,
however, when modeling the corresponding run-time checks we need to "erase" the
uncheckable parts (using existentials so that the types are still well-kinded),
so that we can use static subtyping as a sound model for dynamic type checks
- experimental java enum handling seals enum class
before, we created a refinement class as the placeholder for the sealed children
it seems more direct to use the enum class for that
this way, the new pattern matcher's exhaustiveness checker just works for java enums
|
| | |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| | |
- TypeTestTreeMaker subsumes the old TypeTestTM and TypeAndEqualityTM
its type- and equality-testing logic is configurable so that it can:
- generate trees (main purpose)
- check whether this tree maker is a pure type test
- generate the proposition that models this tree maker
(for exhaustivity and other analyses)
- [CSE] subst binders of dropped tm's to stored ones
somehow the refactoring broke the replacement of the binder of dropped
treemakers by the binder of the reused treemaker
when replacing
TM1(x1 => ...) >> TM2(x2 => ...) >> TM3(x3 => ...) >> ...
TM1'(x1' => ...) >> TM2'(x2' => ...) >> TM3(x3' => ...) >> ...
by
Memo1(x1 => ...) >> TM2(x2 => ...) >> Memo2(x3 => ...) >> ...
Reuse(Memo2)...
you need to replace x1' and x2' by x1
since TM2 tested a shared condition but was not memo-ised,
that implies it simply passed x1 through to x3 unmodified,
and x2' can simply use the stored x1
- type of first uniqued binder sets type of tree
when approximating a tree of treemakers as a DAG,
where sharing indicates the same value is tested,
use the type of the binder that was first used to
create a unique tree as the type of that tree,
and thus all trees used for the same binder in the future
- track substitution of alternatives when approximating
- error on unswitchable @switch annotated matches
if we can't turn a match (with more than two cases) into a switch,
but the user insists, emit an error
misc notes:
- when all you need is nextBinder, FunTreeMaker is your guy
- must pass flag to TypeTestTM for extractorarg test
case TypeTestTreeMaker(prevBinder, testedBinder, expectedTp, nextBinderTp)
(prevBinder eq testedBinder) does not imply it's a pure type test for an extractor call
note that the expected type for an extractor argument is not a type pattern,
thus we only do a classic type test -- the idea was to detect that case by noticing we're
being called with the same previous and tested binder, but that case also arises
for Typed patterns
|
| |\ \
| | |
| | | |
SI-5779: Wrong warning message (comparing Number)
|
| | |/
| |
| |
| |
| |
| | |
Number using `==' will always yield false)
BoxesRuntime knows how to compare java.lang.Number, so we must not warn.
|
| |\ \
| | |
| | | |
Consider method-scoped companions in the implicit scope.
|
| | | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | | |
Fixes SI-4975.
I'll reproduce a relevant snippet of dialogue from Namers#companionSymbolOf:
/** The companion class or companion module of `original`.
* Calling .companionModule does not work for classes defined inside methods.
*
* !!! Then why don't we fix companionModule? Does the presence of these
* methods imply all the places in the compiler calling sym.companionModule are
* bugs waiting to be reported? If not, why not? When exactly do we need to
* call this method?
*/
def companionSymbolOf(original: Symbol, ctx: Context): Symbol = {
This was one such bug.
|
