Make data race detection more liberal.

Allow to rebind a NmedType to refer to a symbol in a subclass
of where the previous symbol was defined. This generalizes the
previous rule that we allow to rebind from a root method to
its synthetic implementation.

The change is not necessary to make the new scheme of computeDenot pass
the tests, but it seems useful to avoid spurious errors elsehere.

Author: odersky

src/dotty/tools/dotc/core/Definitions.scala

@@ -451,7 +451,6 @@ class Definitions {
 
   lazy val RootImports = List[Symbol](JavaLangPackageVal, ScalaPackageVal, ScalaPredefModule, DottyPredefModule)
 
-  lazy val overriddenBySynthetic = Set[Symbol](Any_equals, Any_hashCode, Any_toString, Product_canEqual)
   def isTupleType(tp: Type)(implicit ctx: Context) = {
     val arity = tp.dealias.argInfos.length
     arity <= MaxTupleArity && (tp isRef TupleClass(arity))

src/dotty/tools/dotc/core/Types.scala

@@ -1283,9 +1283,7 @@ object Types {
           (lastDefRunId == NoRunId)
         } ||
         (lastSymbol.infoOrCompleter == ErrorType ||
-        defn.overriddenBySynthetic.contains(lastSymbol)
-          // for overriddenBySynthetic symbols a TermRef such as SomeCaseClass.this.hashCode
-          // might be rewritten from Object#hashCode to the hashCode generated at SyntheticMethods
+        sym.owner.derivesFrom(lastSymbol.owner) && sym.owner != lastSymbol.owner
       ),
         s"data race? overwriting symbol of ${this.show} / $this / ${this.getClass} / ${lastSymbol.id} / ${sym.id} / ${sym.owner} / ${lastSymbol.owner} / ${ctx.phase}")