Merge pull request #665 from dotty-staging/fix/#647-expected-type-overloading

Take expected result type into account for overloading resolution

Author: odersky

src/dotty/tools/dotc/typer/Applications.scala

@@ -915,7 +915,9 @@ trait Applications extends Compatibility { self: Typer =>
   }}
 
   def narrowMostSpecific(alts: List[TermRef])(implicit ctx: Context): List[TermRef] = track("narrowMostSpecific") {
-    (alts: @unchecked) match {
+    alts match {
+      case Nil => alts
+      case _ :: Nil => alts
       case alt :: alts1 =>
         def winner(bestSoFar: TermRef, alts: List[TermRef]): TermRef = alts match {
           case alt :: alts1 =>
@@ -966,6 +968,51 @@ trait Applications extends Compatibility { self: Typer =>
     def narrowByTypes(alts: List[TermRef], argTypes: List[Type], resultType: Type): List[TermRef] =
       alts filter (isApplicable(_, argTypes, resultType))
 
+    /** Is `alt` a method or polytype whose result type after the first value parameter
+     *  section conforms to the expected type `resultType`? If `resultType`
+     *  is a `IgnoredProto`, pick the underlying type instead.
+     */
+    def resultConforms(alt: Type, resultType: Type)(implicit ctx: Context): Boolean = resultType match {
+      case IgnoredProto(ignored) => resultConforms(alt, ignored)
+      case _: ValueType =>
+        alt.widen match {
+          case tp: PolyType => resultConforms(constrained(tp).resultType, resultType)
+          case tp: MethodType => constrainResult(tp.resultType, resultType)
+          case _ => true
+        }
+      case _ => true
+    }
+
+    /** If the `chosen` alternative has a result type incompatible with the expected result
+     *  type `pt`, run overloading resolution again on all alternatives that do match `pt`.
+     *  If the latter succeeds with a single alternative, return it, otherwise
+     *  fallback to `chosen`.
+     *
+     *  Note this order of events is done for speed. One might be tempted to
+     *  preselect alternatives by result type. But is slower, because it discriminates
+     *  less. The idea is when searching for a best solution, as is the case in overloading
+     *  resolution, we should first try criteria which are cheap and which have a high
+     *  probability of pruning the search. result type comparisons are neither cheap nor
+     *  do they prune much, on average.
+     */
+    def adaptByResult(alts: List[TermRef], chosen: TermRef) = {
+        def nestedCtx = ctx.fresh.setExploreTyperState
+        pt match {
+          case pt: FunProto if !resultConforms(chosen, pt.resultType)(nestedCtx) =>
+            alts.filter(alt =>
+              (alt ne chosen) && resultConforms(alt, pt.resultType)(nestedCtx)) match {
+              case Nil => chosen
+              case alt2 :: Nil => alt2
+              case alts2 =>
+                resolveOverloaded(alts2, pt) match {
+                  case alt2 :: Nil => alt2
+                  case _ => chosen
+                }
+            }
+          case _ => chosen
+        }
+      }
+
     val candidates = pt match {
       case pt @ FunProto(args, resultType, _) =>
         val numArgs = args.length
@@ -1027,15 +1074,20 @@ trait Applications extends Compatibility { self: Typer =>
       case pt =>
         alts filter (normalizedCompatible(_, pt))
     }
-    if (isDetermined(candidates)) candidates
-    else narrowMostSpecific(candidates) match {
-      case result @ (alt1 :: alt2 :: _) =>
-//        overload.println(i"ambiguous $alt1 $alt2")
+    narrowMostSpecific(candidates) match {
+      case Nil => Nil
+      case alt :: Nil =>
+        adaptByResult(alts, alt) :: Nil
+        // why `alts` and not `candidates`? pos/array-overload.scala gives a test case.
+        // Here, only the Int-apply is a candidate, but it is not compatible with the result
+        // type. Picking the Byte-apply as the only result-compatible solution then forces
+        // the arguments (which are constants) to be adapted to Byte. If we had picked
+        // `candidates` instead, no solution would have been found.
+      case alts =>
+//      overload.println(i"ambiguous $alts%, %")
         val deepPt = pt.deepenProto
         if (deepPt ne pt) resolveOverloaded(alts, deepPt, targs)
-        else result
-      case result =>
-        result
+        else alts
     }
   }
 

src/dotty/tools/dotc/typer/ProtoTypes.scala

@@ -46,7 +46,7 @@ object ProtoTypes {
      *  fits the given expected result type.
      */
     def constrainResult(mt: Type, pt: Type)(implicit ctx: Context): Boolean = pt match {
-      case _: FunProto =>
+      case pt: FunProto =>
         mt match {
           case mt: MethodType =>
             mt.isDependent || constrainResult(mt.resultType, pt.resultType)

tests/pending/pos/t8230a.scala

@@ -0,0 +1,26 @@
+trait Arr[T]
+object Arr {
+  def apply[T](xs: T): Arr[T]    = null
+  def apply(x: Long) : Arr[Long] = null
+}
+
+object I {
+  implicit def arrToTrav[T] (a: Arr[T])   : Traversable[T]    = null
+  implicit def longArrToTrav(a: Arr[Long]): Traversable[Long] = null
+}
+
+object Test {
+  def foo(t: Traversable[Any]) = {}
+
+  object Okay {
+    Arr("1")
+
+    import I.{ arrToTrav, longArrToTrav }
+    foo(Arr("2"))
+  }
+
+  object Fail {
+    import I.arrToTrav
+    foo(Arr("3")) // found String, expected Long
+  }
+}

tests/pos/array-overload.scala

@@ -0,0 +1,8 @@
+class Test {
+  object A {
+    def apply(x: Byte, xs: Byte*): Array[Byte] = ???
+    def apply(x: Int, xs: Int*): Array[Int] = ???
+  }
+
+  val x: Array[Byte] = A.apply(1, 2)
+}

tests/pending/run/arrayclone-new.scala renamed to tests/run/arrayclone-new.scala

@@ -1,4 +1,4 @@
-import scala.reflect.{ClassTag, classTag}
+import scala.reflect.ClassTag
 
 object Test extends dotty.runtime.LegacyApp{
   BooleanArrayClone;
@@ -106,3 +106,4 @@ object PolymorphicArrayClone{
 
   testIt(mangled.it, 0, 1);
 }
+