fortrangoingonforty/armfortas / e8d2691

Browse files

Test WHERE-block masking produces vselect + correct lane outputs

Authored by mfwolffe <wolffemf@dukes.jmu.edu>
SHA
e8d269142a194364491cb127c7ba78d4ae97d2a7
Parents
8bc8c29
Tree
2d23784

1 changed file

StatusFile+-
A tests/vectorize_where.rs 87 0
tests/vectorize_where.rsadded
@@ -0,0 +1,87 @@
1
+use std::collections::BTreeSet;
2
+use std::path::PathBuf;
3
+
4
+use armfortas::driver::OptLevel;
5
+use armfortas::testing::{capture_from_path, CaptureRequest, CapturedStage, Stage};
6
+
7
+fn fixture(name: &str) -> PathBuf {
8
+    let path = PathBuf::from("test_programs").join(name);
9
+    assert!(path.exists(), "missing test fixture {}", path.display());
10
+    path
11
+}
12
+
13
+fn capture_text(request: CaptureRequest, stage: Stage) -> String {
14
+    let result = capture_from_path(&request).expect("capture should succeed");
15
+    match result.get(stage) {
16
+        Some(CapturedStage::Text(text)) => text.clone(),
17
+        Some(CapturedStage::Run(_)) => panic!("expected text stage for {}", stage.as_str()),
18
+        None => panic!("missing requested stage {}", stage.as_str()),
19
+    }
20
+}
21
+
22
+fn capture_run_stdout(request: CaptureRequest) -> String {
23
+    let result = capture_from_path(&request).expect("capture should succeed");
24
+    match result.get(Stage::Run) {
25
+        Some(CapturedStage::Run(run)) => run.stdout.clone(),
26
+        _ => panic!("missing run stage"),
27
+    }
28
+}
29
+
30
+#[test]
31
+fn o3_vectorizes_where_block_masked_assign() {
32
+    let source = fixture("do_loop_vectorize_where.f90");
33
+
34
+    let o3_ir = capture_text(
35
+        CaptureRequest {
36
+            input: source.clone(),
37
+            requested: BTreeSet::from([Stage::OptIr]),
38
+            opt_level: OptLevel::O3,
39
+        },
40
+        Stage::OptIr,
41
+    );
42
+    // The diamond should be rewritten into a single body with vload a,
43
+    // vload b_old, vfcmp, vselect, vstore.
44
+    assert!(
45
+        o3_ir.contains("vselect"),
46
+        "expected vselect:\n{}",
47
+        o3_ir
48
+    );
49
+    assert!(
50
+        o3_ir.contains("vfcmp"),
51
+        "expected vfcmp:\n{}",
52
+        o3_ir
53
+    );
54
+    // Two vloads in the WHERE body: source `a` and old `b`.
55
+    assert!(
56
+        o3_ir.matches("vload").count() >= 2,
57
+        "expected >= 2 vload:\n{}",
58
+        o3_ir
59
+    );
60
+
61
+    let stdout = capture_run_stdout(CaptureRequest {
62
+        input: source,
63
+        requested: BTreeSet::from([Stage::Run]),
64
+        opt_level: OptLevel::O3,
65
+    });
66
+    let trimmed: Vec<&str> = stdout
67
+        .lines()
68
+        .map(|l| l.trim())
69
+        .filter(|l| !l.is_empty())
70
+        .collect();
71
+    assert_eq!(trimmed.len(), 1, "expected one output line:\n{}", stdout);
72
+    // b(1)=1 (mask=false), b(16)=16 (mask=false, a(16)=0), b(17)=1
73
+    // (mask=true, b ← a = 17-16 = 1), b(32)=16 (mask=true, b ← a = 16).
74
+    assert!(
75
+        trimmed[0].contains("1.0000000E0"),
76
+        "missing b(1)=1: {:?}",
77
+        trimmed[0]
78
+    );
79
+    assert!(
80
+        trimmed[0].contains("1.6000000E1"),
81
+        "missing b(16)=16 / b(32)=16: {:?}",
82
+        trimmed[0]
83
+    );
84
+    // Order: b(1)=1, b(16)=16, b(17)=1, b(32)=16.
85
+    let expected = "1.0000000E0     1.6000000E1     1.0000000E0     1.6000000E1";
86
+    assert_eq!(trimmed[0], expected, "WHERE result wrong: {:?}", trimmed[0]);
87
+}