Fix #2201: Less aggressive type application reduction for better inference

Previously we believed that reducing type applications did not affect
type inference as long as the reduced type constructor had the same
arity as the unreduced one, for example reducing `Foo[X, Y]` is fine
when `Foo` is defined as:

type Foo[A, B] = Bar[A, B]

but not when it's defined as:

type Foo[A] = Bar[A, A]

But this is not a sufficient condition: the bounds of the type
constructor arguments also matter for type inference, so we need to be
more strict and disallow reductions in cases like:

type Foo[A, B] = Bar[B, A]

and:

type Foo[A, B] = Bar[A, Int]

1 files changed, 6 insertions, 3 deletions

diff --git a/compiler/src/dotty/tools/dotc/core/TypeApplications.scala b/compiler/src/dotty/tools/dotc/core/TypeApplications.scala
index 23c3f96cc..94b726491 100644
--- a/compiler/src/dotty/tools/dotc/core/TypeApplications.scala
+++ b/compiler/src/dotty/tools/dotc/core/TypeApplications.scala
@@ -391,9 +391,12 @@ class TypeApplications(val self: Type) extends AnyVal {
           if (!args.exists(_.isInstanceOf[TypeBounds])) {
             val followAlias = Config.simplifyApplications && {
               dealiased.resType match {
-                case AppliedType(tyconBody, _) =>
-                  sameLength(dealiased.typeParams, tyconBody.typeParams)
-                    // Reducing is safe for type inference, as kind arity of type constructor does not change
+                case AppliedType(tyconBody, dealiasedArgs) =>
+                  // Reduction should not affect type inference when it's
+                  // just eta-reduction (ignoring variance annotations).
+                  // See i2201*.scala for examples where more aggressive
+                  // reduction would break type inference.
+                  dealiased.paramRefs == dealiasedArgs
                 case _ => false
               }
             }