interesting test

juno_samples/fork_join_tests/src/cpu.sch

@@ -5,6 +5,7 @@ dce(*);
 let out = auto-outline(*);
 cpu(out.test1);
 cpu(out.test2);
+cpu(out.test3);
 
 ip-sroa(*);
 sroa(*);

juno_samples/fork_join_tests/src/fork_join_tests.jn

@@ -21,3 +21,26 @@ fn test2(input : i32) -> i32[4, 4] {
   }
   return arr;
 }
+
+#[entry]
+fn test3(input : i32) -> i32[3, 3] {
+  let arr1 : i32[3, 3];
+  for i = 0 to 3 {
+    for j = 0 to 3 {
+      arr1[i, j] = (i + j) as i32 + input;
+    }
+  }
+  let arr2 : i32[3, 3];
+  for i = 0 to 3 {
+    for j = 0 to 3 {
+      arr2[i, j] = arr1[3 - i, 3 - j];
+    }
+  }
+  let arr3 : i32[3, 3];
+  for i = 0 to 3 {
+    for j = 0 to 3 {
+      arr3[i, j] = arr2[i, j] + 7;
+    }
+  }
+  return arr3;
+}

juno_samples/fork_join_tests/src/gpu.sch

@@ -5,6 +5,7 @@ dce(*);
 let out = auto-outline(*);
 gpu(out.test1);
 gpu(out.test2);
+gpu(out.test3);
 
 ip-sroa(*);
 sroa(*);

juno_samples/fork_join_tests/src/main.rs

@@ -27,6 +27,11 @@ fn main() {
         let output = r.run(3).await;
         let correct = vec![24i32; 16];
         assert(correct, output);
+
+        let mut r = runner!(test3);
+        let output = r.run(0).await;
+        let correct = vec![11, 10, 9, 10, 9, 8, 9, 8, 7];
+        assert(correct, output);
     });
 }