Don't split return type if there is a comment after function metadata.

lib/src/front_end/piece_factory.dart

@@ -616,6 +616,19 @@ mixin PieceFactory {
       return;
     }
 
+    // Hoist any comments before the function so they don't force a split
+    // between the return type and function. In most cases, this doesn't matter
+    // because the [SequenceBuilder] for the surrounding code will separate out
+    // the leading comment. But if there is a metadata annotation followed by
+    // a comment, then the function, then the comment doesn't get captured by
+    // the [SequenceBuilder], as in:
+    //
+    //     @meta
+    //     // Weird place for comment.
+    //     int f() {}
+    var leadingComments =
+        pieces.takeCommentsBefore(returnType.firstNonCommentToken);
+
     var returnTypePiece = pieces.build(() {
       for (var keyword in modifiers) {
         pieces.modifier(keyword);
@@ -628,7 +641,8 @@ mixin PieceFactory {
       writeFunction();
     });
 
-    pieces.add(VariablePiece(returnTypePiece, [signature], hasType: true));
+    pieces.add(prependLeadingComments(leadingComments,
+        VariablePiece(returnTypePiece, [signature], hasType: true)));
   }
 
   /// If [parameter] has a [defaultValue] then writes a piece for the parameter

test/tall/function/comment.unit

@@ -21,4 +21,12 @@ main() {
       // nested
     }
   }
+}
+>>> Don't split return type if comment before.
+// Comment.
+int f() {;}
+<<<
+// Comment.
+int f() {
+  ;
 }

test/tall/function/metadata.unit

@@ -95,4 +95,14 @@ f([
   @metadata
   @another(argument, argument)
   callback() = constantFunction,
-]) {}
+]) {}
+>>> Don't split return type if comment after metadata.
+@meta
+// Comment.
+int f() {;}
+<<<
+@meta
+// Comment.
+int f() {
+  ;
+}

test/tall/regression/other/dart.unit

@@ -39,4 +39,19 @@ main() {
           .transform(utf8.decoder)
           .transform(const LineSplitter())
           .asBroadcastStream();
+}
+>>> Don't split return type if comment after metadata annotation.
+class Benchmark {
+  @override
+  // A rate of one run per 2s, with a millisecond of noise.  Some variation is
+  // needed for Golem's noise-based filtering and regression detection.
+  double
+  measure() => (2000 + Random().nextDouble() - 0.5) * 1000;
+}
+<<<
+class Benchmark {
+  @override
+  // A rate of one run per 2s, with a millisecond of noise.  Some variation is
+  // needed for Golem's noise-based filtering and regression detection.
+  double measure() => (2000 + Random().nextDouble() - 0.5) * 1000;
 }