fortrangoingonforty/armfortas / b874b02

Browse files

Test sum-reduction scalar tail across all four element types

Authored by mfwolffe <wolffemf@dukes.jmu.edu>
SHA
b874b020a421147e5d8625bf065d4ecd8b8bf77b
Parents
0a92ecd
Tree
48ca6bb

1 changed file

StatusFile+-
M tests/vectorize_reduce_sum_tail.rs 21 8
tests/vectorize_reduce_sum_tail.rsmodified
@@ -39,18 +39,21 @@ fn o3_vectorizes_sum_reductions_with_scalar_tail() {
3939
         },
4040
         Stage::OptIr,
4141
     );
42
-    // Both reductions should leave a VAdd in the body and a
43
-    // vreduce_sum at exit followed by peeled scalar iadd ops that
44
-    // chain from the reduce result.
42
+    // All four reductions should leave a vreduce_sum at exit
43
+    // followed by peeled scalar iadd/fadd ops chaining from the
44
+    // reduce result.
4545
     assert_eq!(
4646
         o3_ir.matches("vreduce_sum").count(),
47
-        2,
48
-        "expected two vreduce_sum (i32 + i64):\n{}",
47
+        4,
48
+        "expected four vreduce_sum (i32, i64, f32, f64):\n{}",
4949
         o3_ir
5050
     );
5151
     assert!(
52
-        o3_ir.contains("<4 x i32>") && o3_ir.contains("<2 x i64>"),
53
-        "expected i32 and i64 vector accumulators in IR:\n{}",
52
+        o3_ir.contains("<4 x i32>")
53
+            && o3_ir.contains("<2 x i64>")
54
+            && o3_ir.contains("<4 x f32>")
55
+            && o3_ir.contains("<2 x f64>"),
56
+        "expected i32/i64/f32/f64 vector accumulators in IR:\n{}",
5457
         o3_ir
5558
     );
5659
@@ -64,8 +67,18 @@ fn o3_vectorizes_sum_reductions_with_scalar_tail() {
6467
         .map(|l| l.trim())
6568
         .filter(|l| !l.is_empty())
6669
         .collect();
67
-    assert_eq!(trimmed.len(), 2, "expected two output lines:\n{}", stdout);
70
+    assert_eq!(trimmed.len(), 4, "expected four output lines:\n{}", stdout);
6871
     // 1 + 2 + ... + 31 = 31 * 32 / 2 = 496.
6972
     assert_eq!(trimmed[0], "496", "i32 sum wrong: got {:?}", trimmed[0]);
7073
     assert_eq!(trimmed[1], "496", "i64 sum wrong: got {:?}", trimmed[1]);
74
+    assert!(
75
+        trimmed[2].starts_with("4.96"),
76
+        "f32 sum wrong: got {:?}",
77
+        trimmed[2]
78
+    );
79
+    assert!(
80
+        trimmed[3].starts_with("4.96"),
81
+        "f64 sum wrong: got {:?}",
82
+        trimmed[3]
83
+    );
7184
 }