Duplicate a Class in a Compiler Phase (Compiler Plugin)

OK, Here some rough example. I'm assuming you can fill-in the blanks
looking at a basic Plugin example. Please also realize that I'm only
recently started with Scala, so my coding style should not be taken as
a "good example".

First, in your TypingTransformer, you want to process classes that
belong to a package. The best way I found is to override
transformStats:

override def transformStats(stats: List[Tree], exprOwner: Symbol):
List[Tree] = {
// Default-process the trees
val transformedStats = super.transformStats(stats, exprOwner)
var newStats = List.empty[Tree]
// Check if any tree is an annotated trait, and if we can generate
a new wrapper for it
for (t <- transformedStats; newTree = genNewClass(t); if (newTree !
= null)) {
newStats = newTree :: newStats
}
// Add the new processed wrappers, if any, to the previous
content
super.transformStats(newStats, exprOwner) ::: transformedStats
}

Then, you could define genNewClass about like this:

private def genNewClass(tree: Tree): Tree = {
tree match {
case clsDef @ ClassDef(mods, name, tparams, impl) if SOME_CONDITION
=> {
val newName = newTypeName("MY_NEW_NAME")
val newImpl = treeCopy.Template(impl, impl.parents,
impl.self, impl.body)
treeCopy.ClassDef(clsDef , mods, newName, tparams,
newImpl)
}
case _ => null
}
}

My implementation of genNewClass is very different, but I think this
is the basic form. If it doesn't work, just post you Exception.

I suggest using the following scalac parameters:

-explaintypes -Ylog:MY-PHASE -Xprint:MY-PHASE

On Nov 1, 5:41 pm, Sciss wrote:
> ok, question re-formulated:
>
> how do i duplicate a class in compiler-phase (using a new class name)?
>
> thx, -sciss-

the problem seems to be that each class if full of symbols, and each is 'owned' by another symbol. so if you want to duplicate a class, you virtually need to clone each single statement, from type parameters to methods to ... so you need an omniscient pattern matcher for every possible body component.

the only other idea i have is that the cloning needs to be performed at a stage where the symbols not yet 'owned'. i don't know if this is possible. it would mean something like runsBefore( "namer" ) i guess -- not sure this is all possible, or nothing exists yet at that point...

the third alternative would be to create a source generator instead, something based on scala-refactoring that takes the annotated source and produces two sources (for the 'original' and the 'copy') before feeding them into the compiler.

i have given up with this for the moment, as either way takes more resources than i currently have.

best, -sciss-

On 2 Nov 2011, at 18:14, monster wrote:

> OK, Here some rough example. I'm assuming you can fill-in the blanks
> looking at a basic Plugin example. Please also realize that I'm only
> recently started with Scala, so my coding style should not be taken as
> a "good example".
>
> First, in your TypingTransformer, you want to process classes that
> belong to a package. The best way I found is to override
> transformStats:
>
> override def transformStats(stats: List[Tree], exprOwner: Symbol):
> List[Tree] = {
> // Default-process the trees
> val transformedStats = super.transformStats(stats, exprOwner)
> var newStats = List.empty[Tree]
> // Check if any tree is an annotated trait, and if we can generate
> a new wrapper for it
> for (t <- transformedStats; newTree = genNewClass(t); if (newTree !
> = null)) {
> newStats = newTree :: newStats
> }
> // Add the new processed wrappers, if any, to the previous
> content
> super.transformStats(newStats, exprOwner) ::: transformedStats
> }
>
> Then, you could define genNewClass about like this:
>
> private def genNewClass(tree: Tree): Tree = {
> tree match {
> case clsDef @ ClassDef(mods, name, tparams, impl) if SOME_CONDITION
> => {
> val newName = newTypeName("MY_NEW_NAME")
> val newImpl = treeCopy.Template(impl, impl.parents,
> impl.self, impl.body)
> treeCopy.ClassDef(clsDef , mods, newName, tparams,
> newImpl)
> }
> case _ => null
> }
> }
>
> My implementation of genNewClass is very different, but I think this
> is the basic form. If it doesn't work, just post you Exception.
>
> I suggest using the following scalac parameters:
>
> -explaintypes -Ylog:MY-PHASE -Xprint:MY-PHASE
>
> On Nov 1, 5:41 pm, Sciss wrote:
>> ok, question re-formulated:
>>
>> how do i duplicate a class in compiler-phase (using a new class name)?
>>
>> thx, -sciss-

On Wed, Nov 2, 2011 at 7:49 PM, Sciss wrote:
> the problem seems to be that each class if full of symbols, and each is 'owned' by another symbol. so if you want to duplicate a class, you virtually need to clone each single statement, from type parameters to methods to ... so you need an omniscient pattern matcher for every possible body component.

http://github.com/scala/scala/tree/master/src/compiler/scala/tools/nsc/t...

Even given that this task is probably harder than you think.

> the only other idea i have is that the cloning needs to be performed at a stage where the symbols not yet 'owned'. i don't know if this is possible.

It is not. One thing, almost the only thing you can feel pretty
confident a symbol will have is an owner. With the exception of a few
places where symbols are directly instantiated and which we should
probably get rid of, all symbols are created by calling a symbol
creation method on another symbol instance -- the other one being its
owner.

On 2 Nov 2011, at 20:51, Paul Phillips wrote:

> On Wed, Nov 2, 2011 at 7:49 PM, Sciss wrote:
>> the problem seems to be that each class if full of symbols, and each is 'owned' by another symbol. so if you want to duplicate a class, you virtually need to clone each single statement, from type parameters to methods to ... so you need an omniscient pattern matcher for every possible body component.
>
> http://github.com/scala/scala/tree/master/src/compiler/scala/tools/nsc/t...
>
> Even given that this task is probably harder than you think.

that looks like offering a solution, thanks!

>
>> the only other idea i have is that the cloning needs to be performed at a stage where the symbols not yet 'owned'. i don't know if this is possible.
>
> It is not. One thing, almost the only thing you can feel pretty
> confident a symbol will have is an owner. With the exception of a few
> places where symbols are directly instantiated and which we should
> probably get rid of, all symbols are created by calling a symbol
> creation method on another symbol instance -- the other one being its
> owner.