SI-8129 Plug a leak in perRunCaches

I guess we've never leaned to heavily on these, because the
registry of caches was forgetting about most of them
immediately if they compared == at the point of creation.

Consequently, maps like `treatedInfos` in Mixin were holding
onto old types and symbols and leaking memory.

You can reproduce the leak with:

	(for i in {1..200}; do echo src/library/scala/collection/*.scala; done; printf "\n") | scalac-hash v2.11.0-M7 -J-Xmx256M -Xresident

	nsc> warning: there were 6 deprecation warning(s); re-run with -deprecation for details

	(... 15 times)

	nsc> java.lang.OutOfMemoryError: GC overhead limit exceeded
	Dumping heap to java_pid90362.hprof ...
	Heap dump file created [340635728 bytes in 7.993 secs]
	^CException in thread "main" java.lang.OutOfMemoryError: GC overhead limit exceeded

As a consequence of this change, I had to remove eager clearing of the
uniques types cache. The comment in the code elaborates further.

Author: retronym

src/reflect/scala/reflect/internal/SymbolTable.scala

@@ -344,16 +344,18 @@ abstract class SymbolTable extends macros.Universe
 
     // Weak references so the garbage collector will take care of
     // letting us know when a cache is really out of commission.
-    private val caches = WeakHashSet[Clearable]()
+    import java.lang.ref.WeakReference
+    private var caches = List[WeakReference[Clearable]]()
 
     def recordCache[T <: Clearable](cache: T): T = {
-      caches += cache
+      caches ::= new WeakReference(cache)
       cache
     }
 
     def clearAll() = {
       debuglog("Clearing " + caches.size + " caches.")
-      caches foreach (_.clear)
+      caches foreach (ref => Option(ref.get).foreach(_.clear))
+      caches = caches.filterNot(_.get == null)
     }
 
     def newWeakMap[K, V]()        = recordCache(mutable.WeakHashMap[K, V]())
@@ -364,9 +366,9 @@ abstract class SymbolTable extends macros.Universe
       val NoCached: T = null.asInstanceOf[T]
       var cached: T = NoCached
       var cachedRunId = NoRunId
-      caches += new Clearable {
+      recordCache(new Clearable {
         def clear(): Unit = cached = NoCached
-      }
+      })
       () => {
         if (currentRunId != cachedRunId || cached == NoCached) {
           cached = f

src/reflect/scala/reflect/internal/Types.scala

@@ -8,7 +8,6 @@ package reflect
 package internal
 
 import scala.collection.{ mutable, immutable, generic }
-import generic.Clearable
 import scala.ref.WeakReference
 import mutable.ListBuffer
 import Flags._
@@ -3660,7 +3659,11 @@ trait Types
     if (Statistics.canEnable) Statistics.incCounter(rawTypeCount)
     if (uniqueRunId != currentRunId) {
       uniques = util.WeakHashSet[Type](initialUniquesCapacity)
-      perRunCaches.recordCache(uniques)
+      // JZ: We used to register this as a perRunCache so it would be cleared eagerly at
+      // the end of the compilation run. But, that facility didn't actually clear this map (SI-8129)!
+      // When i fixed that bug, run/tpeCache-tyconCache.scala started failing. Why was that?
+      // I've removed the registration for now. I don't think its particularly harmful anymore
+      // as a) this is now a weak set, and b) it is discarded completely before the next run.
       uniqueRunId = currentRunId
     }
     (uniques findEntryOrUpdate tp).asInstanceOf[T]