fortrangoingonforty/armfortas / a08b208

Browse files

Test i64 sum reduction emits addp.2d + umov.d and prints 528

Authored by mfwolffe <wolffemf@dukes.jmu.edu>
SHA
a08b208be350de6376a70a3309320a2de496138a
Parents
5b39a81
Tree
3d997b4

1 changed file

StatusFile+-
A tests/vectorize_reduce_i64.rs 77 0
tests/vectorize_reduce_i64.rsadded
@@ -0,0 +1,77 @@
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_i64_sum_reduction() {
32
+    let source = fixture("do_loop_vectorize_reduce_i64.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
+    assert!(
43
+        o3_ir.contains("<2 x i64>")
44
+            && o3_ir.contains("vbroadcast")
45
+            && o3_ir.contains("vadd")
46
+            && o3_ir.contains("vreduce_sum"),
47
+        "expected i64 reduction shape (<2 x i64> + vbroadcast + vadd + vreduce_sum):\n{}",
48
+        o3_ir
49
+    );
50
+
51
+    // i64 reduce: cross-lane is `addp.2d` + `umov.d`, not addv.
52
+    let o3_asm = capture_text(
53
+        CaptureRequest {
54
+            input: source.clone(),
55
+            requested: BTreeSet::from([Stage::Asm]),
56
+            opt_level: OptLevel::O3,
57
+        },
58
+        Stage::Asm,
59
+    );
60
+    assert!(
61
+        o3_asm.contains("addp.2d") && o3_asm.contains("umov.d"),
62
+        "VReduceSum on i64 should lower via addp.2d + umov.d:\n{}",
63
+        o3_asm
64
+    );
65
+
66
+    let stdout = capture_run_stdout(CaptureRequest {
67
+        input: source,
68
+        requested: BTreeSet::from([Stage::Run]),
69
+        opt_level: OptLevel::O3,
70
+    });
71
+    let trimmed: Vec<&str> = stdout
72
+        .lines()
73
+        .map(|l| l.trim())
74
+        .filter(|l| !l.is_empty())
75
+        .collect();
76
+    assert_eq!(trimmed, vec!["528"], "i64 sum should be 528:\n{}", stdout);
77
+}