fix foldMap stack safety

https://github.com/non/cats/pull/702

core/src/main/scala/scalaz/Free.scala

@@ -193,12 +193,16 @@ sealed abstract class Free[S[_], A] {
    * Runs to completion, mapping the suspension with the given transformation at each step and
    * accumulating into the monad `M`.
    */
+  @annotation.tailrec
   final def foldMap[M[_]](f: S ~> M)(implicit M: Monad[M]): M[A] =
-    step match {
+    this match {
       case Return(a) => M.pure(a)
       case Suspend(s) => f(s)
-      // This is stack safe because `step` ensures right-associativity of Gosub
-      case Gosub(x, g) => M.bind(x foldMap f)(c => g(c) foldMap f)
+      case Gosub(x, g) => x match {
+        case Suspend(s) => g(f(s)).foldMap(f)
+        case Gosub(cSub, h) => cSub.flatMap(cc => h(cc).flatMap(g)).foldMap(f)
+        case Return(a) => g(a).foldMap(f)
+      }
     }
 
   import Id._

tests/src/test/scala/scalaz/FreeTest.scala

@@ -91,6 +91,25 @@ object FreeTest extends SpecLite {
     checkAll(bindRec.laws[FreeOption])
   }
 
+  "foldMap is stack safe" ! {
+    val n = 1000000
+    trait FTestApi[A]
+    case class TB(i: Int) extends FTestApi[Int]
+
+    def a(i: Int): Free[FTestApi, Int] = for {
+      j <- Free.liftF(TB(i))
+      z <- if (j < n) a(j) else Free.pure[FTestApi, Int](j)
+    } yield z
+
+    val runner = new (FTestApi ~> Id.Id) {
+      def apply[A](fa: FTestApi[A]) = fa match {
+        case TB(i) => i + 1
+      }
+    }
+
+    a(0).foldMap(runner) must_=== n
+  }
+
   "List" should {
     "not stack overflow with 50k binds" in {
       val expected = Applicative[FreeList].point(())