|
|
This bug is a result of a subtle interplay of the stackable modifications
mechanism and specialization.
Prior to this commit, using `abstract override` with specialization was
broken in the sense that specialization did not create a specialized
version of the super accessor. Observe the following code:
trait A[@specialized(Int) T] {
def foo(t: T)
}
trait B extends A[Int] {
def foo(t: Int) {
println("B.foo")
}
}
trait M extends B {
abstract override def foo(t: Int) {
super.foo(t)
println("M.foo")
}
}
object C extends B with M
During the `superaccessors` phase, the following stub is generated
in `M`:
private <artifact> <superaccessor> def super$foo(t: Int)
Note that `foo` is a method that will later need to be specialized.
During the `specialize` phase, `A.foo` gets a *special overload*
`A.foo$mcI$sp`, which is a bridge to `A.foo`.
`B` extends `A$mcI$sp` (previously `A[Int]`), so `B.foo` gets a
*special override* `B.foo$mcI$sp`, which contains the implementation.
`B.foo` is overridden to become a bridge to `B.foo$mcI$sp`.
`M` extends `B`, so `M.foo` gets a special override `M.foo$mcI$sp`,
and `M.foo` itself is turned into a bridge to `M.foo$mcI$sp`, just
as was the case with `B`.
This is where the first problem arises - `M.foo$mcI$sp` does not
get an `ABSOVERRIDE` flag after being created. This commit fixes
that.
As mentioned earlier, `M` has a super accessor `super$foo`.
A super accessor (naturally) does not override anything, thus,
according to the standing specialization criteria it does not
need a special override (see `needsSpecialOverride`).
Its type does not contain any specialized type parameters, thus,
it is not eligible to obtain a special overload.
So, our `super$foo` stays the way it is, subsequently being
renamed to `M$$super$foo`.
Later, during `mixin`, it is implemented in `C` as a forwarder
to `B$class.foo` (Not `B.foo`, because the implementation
classes are generated in `mixin`).
Lets see what we have now (I omit the `$this` parameter
for methods in implementation classes `B$class` and `M$class`):
class B$class
def foo(t: Int) = ------------\ <-\
def foo$mcI$sp(t: Int) = | |
| |
trait M$class | |
def foo(t: Int) = -------\ | |
| | |
def foo$mcI$sp(t: Int) = <-/<-\ | |
{ | | |
/---- M$$super$foo(t) | | |
| ... | | |
| } | | |
| | | |
| object C | | |
| def foo$mcI$sp(t: Int) = -----/ <-/ |
\-> def M$$super$foo(t: Int) = |
{ |
------------------------------------/
Now, call `C.foo`. This call is rewritten to
`C.foo$mcI$sp`, which is a bridge to
`M$class.foo$mcI$sp`.
Follow the lines on the diagram to enter
an endless loop. Boom! Stack overflow.
The culprit is the super accessor `C.M$$super$foo`, which
should have forwarded to the special override
`B$class.foo$mcI$sp`, instead of to `B$class.foo`.
So, we have 2 options:
1) Make `C.M$$super$foo` forward to the proper special
override, where the implementation actually is.
2) During specialization, create a specialized version
of `M$$super$foo` called `M$$super$foo$mcI$sp`, and set
its alias to the special overload `foo$mcI$sp`. Later,
during `mixin`, this super accessor will be appropriately
resolved in concrete classes.
Option 1) involves cluttering `mixin` with specialization
logic.
Furthermore, the specialization already does create
specialized versions of super accessors when the super
accessor type contains specialized type parameters
(in other words, it generates a special overload).
Thus, 2) seems an ideal solution. We cannot deduct
if a super accessor should get a special override
directly, but we can see if its alias needs a
special override. If it does, then we generate
a special override for the super accessor.
This commit introduces the following changes:
1) The `ABSOVERRIDE` flag is now retained, as mentioned
earlier.
2) A super accessor gets a special override if its
alias needs a special override.
3) The super calls in the methods bodies are rewritten
to their specialized variants if they exist.
4) Newly generated special overrides and special overloads
are now entered into the declaration list of the owner
during specialization.
Strangely, this was not done before, but it is necessary for
`mixin` to detect the generated special overload/override
which is an alias for the super accessor.
|
