|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
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
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
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.
|