@@ -0,0 +1,80 @@ |
| 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_array_plus_invariant_scalar_loop() { |
| 32 | + let source = fixture("do_loop_vectorize_scalar.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 | + |
| 43 | + // Either form is acceptable as long as the scalar+invariant |
| 44 | + // pattern was vectorized: |
| 45 | + // * NeonVectorize: vbroadcast in preheader + vload/vadd/vstore in body. |
| 46 | + // * Vectorize fallback: afs_array_add_scalar_i32 kernel call. |
| 47 | + let neon = o3_ir.contains("vbroadcast") |
| 48 | + && o3_ir.contains("vadd") |
| 49 | + && o3_ir.contains("vstore"); |
| 50 | + let kernel = o3_ir.contains("call @afs_array_add_scalar_i32("); |
| 51 | + assert!( |
| 52 | + neon || kernel, |
| 53 | + "O3 should vectorize a(i) = b(i) + scale (vbroadcast/vadd/vstore or afs_array_add_scalar_i32):\n{}", |
| 54 | + o3_ir |
| 55 | + ); |
| 56 | + // Prefer the real NEON path: it avoids the runtime call entirely. |
| 57 | + assert!( |
| 58 | + neon, |
| 59 | + "O3 should pick the NeonVectorize broadcast path over the runtime kernel:\n{}", |
| 60 | + o3_ir |
| 61 | + ); |
| 62 | + |
| 63 | + // Also verify runtime correctness: a(1) = 1+7 = 8, a(32) = 32+7 = 39. |
| 64 | + let stdout = capture_run_stdout(CaptureRequest { |
| 65 | + input: source, |
| 66 | + requested: BTreeSet::from([Stage::Run]), |
| 67 | + opt_level: OptLevel::O3, |
| 68 | + }); |
| 69 | + let trimmed: Vec<&str> = stdout |
| 70 | + .lines() |
| 71 | + .map(|l| l.trim()) |
| 72 | + .filter(|l| !l.is_empty()) |
| 73 | + .collect(); |
| 74 | + assert_eq!( |
| 75 | + trimmed, |
| 76 | + vec!["8", "39"], |
| 77 | + "vectorized a(i) = b(i) + scale should produce 8 then 39:\n{}", |
| 78 | + stdout |
| 79 | + ); |
| 80 | +} |