fortrangoingonforty/armfortas / 879adc0

Browse files

Test WHERE with array+array body vectorizes at O3

Authored by mfwolffe <wolffemf@dukes.jmu.edu>
SHA
879adc09943876d05bbda8f5d7fd6ef11eb62679
Parents
f4d73bc
Tree
b960784

2 changed files

StatusFile+-
A test_programs/do_loop_vectorize_where_2arr.f90 25 0
A tests/vectorize_where_2arr.rs 78 0
test_programs/do_loop_vectorize_where_2arr.f90added
@@ -0,0 +1,25 @@
1
+! WHERE with an array+array body: `where (a > 0) c = a + b`. Both
2
+! `a` and `b` are arrays loaded per iteration; the matcher must
3
+! recognize the second array load (in the then_block) as a
4
+! vector load and feed both into a VAdd that drives vselect's
5
+! true arm.
6
+!
7
+! a(i) = i - 16 (range -15..16); b(i) = 100; c(i) = i.
8
+! Where a > 0 (lanes 17..32): c ← a + b = (i-16) + 100 = i + 84.
9
+!   c(1) = 1, c(16) = 16, c(17) = 101, c(32) = 116.
10
+!
11
+! CHECK: 1.0000000E0     1.6000000E1     1.0100000E2     1.1600000E2
12
+program test_do_loop_vectorize_where_2arr
13
+  implicit none
14
+  integer :: i
15
+  real(4) :: a(32), b(32), c(32)
16
+  do i = 1, 32
17
+    a(i) = real(i - 16, 4)
18
+    b(i) = 100.0
19
+    c(i) = real(i, 4)
20
+  end do
21
+  where (a > 0.0)
22
+    c = a + b
23
+  end where
24
+  print *, c(1), c(16), c(17), c(32)
25
+end program test_do_loop_vectorize_where_2arr
tests/vectorize_where_2arr.rsadded
@@ -0,0 +1,78 @@
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_with_array_plus_array_body() {
32
+    let source = fixture("do_loop_vectorize_where_2arr.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 array+array WHERE body must lower to two vloads (a and b)
43
+    // feeding a vadd that drives vselect's true arm.
44
+    assert_eq!(
45
+        o3_ir.matches("vselect").count(),
46
+        1,
47
+        "expected exactly one vselect:\n{}",
48
+        o3_ir
49
+    );
50
+    assert!(
51
+        o3_ir.matches("vload").count() >= 3,
52
+        "expected >=3 vload (a, dest, b):\n{}",
53
+        o3_ir
54
+    );
55
+    assert!(
56
+        o3_ir.contains("vadd"),
57
+        "expected vadd lifted from a+b in WHERE body:\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
+    assert_eq!(
73
+        trimmed[0],
74
+        "1.0000000E0     1.6000000E1     1.0100000E2     1.1600000E2",
75
+        "array+array WHERE wrong: {:?}",
76
+        trimmed[0]
77
+    );
78
+}