SI-7470 implements fundep materialization

This fix provides implicit macros with an ability to affect type inference
in a more or less sane manner. That's crucial for materialization of
multi-parametric type class instances (e.g. Iso's from shapeless).
Details of the technique can be found in comments.

1 files changed, 40 insertions, 2 deletions

diff --git a/test/files/run/t5923a/Macros_1.scala b/test/files/run/t5923a/Macros_1.scala
index 6d21362c4d..97076eb102 100644
--- a/test/files/run/t5923a/Macros_1.scala
+++ b/test/files/run/t5923a/Macros_1.scala
@@ -7,8 +7,46 @@ object C {
 }
 
 object Macros {
-  def impl[T: c.WeakTypeTag](c: Context) = {
+  def impl[T](c: Context)(ttag: c.WeakTypeTag[T]) = {
     import c.universe._
-    reify(C[T](c.literal(weakTypeOf[T].toString).splice))
+    val ttag0 = ttag;
+    {
+      // When we're expanding implicitly[C[Nothing]], the type inferencer will see
+      // that foo[T] returns C[T] and that we request an implicit of type C[Nothing].
+      //
+      // Then the type inferencer will try to match C[T] against C[Nothing] and infer everything it can infer
+      // from that match, but not more (e.g. if we were returning Iso[T, U] and the type we were looking at was Iso[Foo, L],
+      // we wouldn't want U to be auto-inferred to Nothing, as it usually happens with normal methods,
+      // but would rather want it to remain unknown, so that our macro could take a stab at inferring it:
+      // see the comments in this commit for more information).
+      //
+      // Equipped with common sense, in our case of C[T] and C[Nothing] we would expect T to be inferred as Nothing, and then we
+      // would expect T in the corresponding macro invocation to be Nothing. Unfortunately it is not that simple.
+      //
+      // Internally the type inferencer uses Nothing as a dummy value, which stands for "don't know how to
+      // infer this type parameter". In the Iso example, matching Iso[T, U] against Iso[Foo, L] would result in
+      // T being inferred as Foo and U being inferred as Nothing (!!). Then the type inferencer will think:
+      // "Aha! U ended up being Nothing. This means that I failed to infer it,
+      // therefore the result of my work is: T -> Foo, U -> still unknown".
+      //
+      // That's all very good and works very well until Nothing is a genuine result of type inference,
+      // as in our original example of inferring T in C[T] from C[Nothing]. In that case, the inferencer becomes confused
+      // and here in the macro implementation we get weakTypeOf[T] equal to some dummy type carrying a type parameter
+      // instead of Nothing.
+      //
+      // This eccentric behavior of the type inferencer is a long-standing problem in scalac,
+      // so the best one can do for now until it's fixed is to work around, manually converting
+      // suspicious T's into Nothings. Of course, this means that we would have to approximate,
+      // because there's no way to know whether having T here stands for a failed attempt to infer Nothing
+      // or for a failed attempt to infer anything, but at least we're in full control of making the best
+      // of this sad situation.
+      implicit def ttag: WeakTypeTag[T] = {
+        val tpe = ttag0.tpe
+        val sym = tpe.typeSymbol.asType
+        if (sym.isParameter && !sym.isSkolem) TypeTag.Nothing.asInstanceOf[TypeTag[T]]
+        else ttag0
+      }
+      reify(C[T](c.literal(weakTypeOf[T].toString).splice))
+    }
   }
 }
\ No newline at end of file