| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The body of lambdas is compiled into a synthetic method
in the enclosing class. Previously, this method was a public
virtual method named `fully$qualified$Class$$anonfun$n`.
For lambdas that didn't capture a `this` reference, a static
method was used.
This commit changes two aspects.
Firstly, all lambda impl methods are now emitted static.
An extra parameter is added to those that require a this
reference.
This is an improvement as it:
- allows, shorter, more readable names for the lambda impl method
- avoids pollution of the vtable of the class. Note that javac uses
private instance methods, rather than public static methods. If
we followed its lead, we would be unable to support important use
cases in our inliner
Secondly, the name of the enclosing method has been included in
the name of the lambda impl method to improve debuggability and
to improve serialization compatibility. The serialization improvement
comes from the way that fresh names for the impl methods are
allocated: adding or removing lambdas in methods not named "foo" won't
change the numbering of the `anonfun$foo$n` impl methods from methods
named "foo". This is in line with user expectations about anonymous
class and lambda serialization stability. Brian Goetz has described
this tricky area well in:
http://cr.openjdk.java.net/~briangoetz/eg-attachments/lambda-serialization.html
This commit doesn't go as far a Javac, we don't use the hash of the
lambda type info, param names, etc to map to a lambda impl method name.
As such, we are more prone to the type-1 and -2 failures described there.
However, our Scala 2.11.8 has similar characteristics, so we aren't going
backwards.
Special case in the naming: Use "new" rather than "<init>" for constructor enclosed
lambdas, as javac does.
I have also changed the way that "delambdafy target" methods are identifed.
Rather than relying on the naming convention, I have switched to using a
symbol attachment. The assumption is that we only need to identify them
from within the same compilation unit.
This means we can distinguish impl metbods for expanded functions
(ones called from an `apply` method of an ahead-of-time expanded
anonfun class), from those that truly end up as targets for lambda
metafactory. Only the latter are translated to static methods in
this patch.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The Delambdafy phase generates its `FunctionN` subclasses after
the specialization phase. As such, `((x: Int) => x).apply(42)` incurs
boxing.
This commit falls back to the `-Ydelambdafy:inline` in this case.
This is done by running the specialization type map over the
type of the function, and seeing if anything changes. To make this
work robustly, we first need to ensure that the specialization info
transformer has processed all the function types.
This is not a fundamental limitation; we could in principle generate
the specialized code.
A followup change will use `-Ydelambdafy:method` as the basis for
invokedymnamic lambdas. As part of that stream of
work, we will synthesize specialization-aware lambdas, and remove
the fallback to `-Ydelambdafy:inline`.
I have updated some tests that intend to test the delambdafy transform
to avoid use of specialized function types.
|
