@@ -0,0 +1,124 @@ |
| 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_fp_minmax_reductions() { |
| 32 | + let source = fixture("do_loop_vectorize_reduce_fp_minmax.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 | + // Each min/max loop should leave behind one vmax/vmin in the |
| 43 | + // body and one vreduce_max/vreduce_min at exit. |
| 44 | + assert_eq!( |
| 45 | + o3_ir.matches("vreduce_max").count(), |
| 46 | + 2, |
| 47 | + "expected two vreduce_max (f32 + f64):\n{}", |
| 48 | + o3_ir |
| 49 | + ); |
| 50 | + assert_eq!( |
| 51 | + o3_ir.matches("vreduce_min").count(), |
| 52 | + 2, |
| 53 | + "expected two vreduce_min (f32 + f64):\n{}", |
| 54 | + o3_ir |
| 55 | + ); |
| 56 | + assert!( |
| 57 | + o3_ir.contains("<4 x f32>") && o3_ir.contains("<2 x f64>"), |
| 58 | + "expected both f32 and f64 vector accumulators in IR:\n{}", |
| 59 | + o3_ir |
| 60 | + ); |
| 61 | + |
| 62 | + let o3_asm = capture_text( |
| 63 | + CaptureRequest { |
| 64 | + input: source.clone(), |
| 65 | + requested: BTreeSet::from([Stage::Asm]), |
| 66 | + opt_level: OptLevel::O3, |
| 67 | + }, |
| 68 | + Stage::Asm, |
| 69 | + ); |
| 70 | + assert!( |
| 71 | + o3_asm.contains("fmaxv.4s"), |
| 72 | + "f32 max reduce should use fmaxv.4s:\n{}", |
| 73 | + o3_asm |
| 74 | + ); |
| 75 | + assert!( |
| 76 | + o3_asm.contains("fminv.4s"), |
| 77 | + "f32 min reduce should use fminv.4s:\n{}", |
| 78 | + o3_asm |
| 79 | + ); |
| 80 | + // f64: NEON has no fmaxv.2d, the pairwise scalar form is the |
| 81 | + // across-lane reduce for two f64 lanes. |
| 82 | + assert!( |
| 83 | + o3_asm.contains("fmaxp.2d"), |
| 84 | + "f64 max reduce should use fmaxp.2d:\n{}", |
| 85 | + o3_asm |
| 86 | + ); |
| 87 | + assert!( |
| 88 | + o3_asm.contains("fminp.2d"), |
| 89 | + "f64 min reduce should use fminp.2d:\n{}", |
| 90 | + o3_asm |
| 91 | + ); |
| 92 | + |
| 93 | + let stdout = capture_run_stdout(CaptureRequest { |
| 94 | + input: source, |
| 95 | + requested: BTreeSet::from([Stage::Run]), |
| 96 | + opt_level: OptLevel::O3, |
| 97 | + }); |
| 98 | + let trimmed: Vec<&str> = stdout |
| 99 | + .lines() |
| 100 | + .map(|l| l.trim()) |
| 101 | + .filter(|l| !l.is_empty()) |
| 102 | + .collect(); |
| 103 | + assert_eq!(trimmed.len(), 4, "expected four output lines:\n{}", stdout); |
| 104 | + assert!( |
| 105 | + trimmed[0].starts_with("3.2"), |
| 106 | + "f32 max should be 32.0, got {:?}", |
| 107 | + trimmed[0] |
| 108 | + ); |
| 109 | + assert!( |
| 110 | + trimmed[1].starts_with("1.0"), |
| 111 | + "f32 min should be 1.0, got {:?}", |
| 112 | + trimmed[1] |
| 113 | + ); |
| 114 | + assert!( |
| 115 | + trimmed[2].starts_with("3.2"), |
| 116 | + "f64 max should be 32.0, got {:?}", |
| 117 | + trimmed[2] |
| 118 | + ); |
| 119 | + assert!( |
| 120 | + trimmed[3].starts_with("1.0"), |
| 121 | + "f64 min should be 1.0, got {:?}", |
| 122 | + trimmed[3] |
| 123 | + ); |
| 124 | +} |