| 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_manual_sum_reduction_loop() { |
| 32 | let source = fixture("do_loop_vectorize_reduce_sum.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 reduction path must produce a VBroadcast in the preheader, |
| 43 | // VAdd of two <V x i32> vectors in the body, and a VReduceSum |
| 44 | // after the loop. |
| 45 | assert!( |
| 46 | o3_ir.contains("vbroadcast") |
| 47 | && o3_ir.contains("vadd") |
| 48 | && o3_ir.contains("vreduce_sum"), |
| 49 | "expected NeonVectorize reduction shape (vbroadcast + vadd + vreduce_sum):\n{}", |
| 50 | o3_ir |
| 51 | ); |
| 52 | |
| 53 | // Assembly must use `mov.16b` for the loop-param transfer rather |
| 54 | // than `fmov d` (which would clobber the upper lanes of the V128 |
| 55 | // accumulator and produce a wrong sum). |
| 56 | let o3_asm = capture_text( |
| 57 | CaptureRequest { |
| 58 | input: source.clone(), |
| 59 | requested: BTreeSet::from([Stage::Asm]), |
| 60 | opt_level: OptLevel::O3, |
| 61 | }, |
| 62 | Stage::Asm, |
| 63 | ); |
| 64 | assert!( |
| 65 | o3_asm.contains("mov.16b"), |
| 66 | "regalloc must materialise V128 block-param transfers via `mov.16b`, not `fmov d`:\n{}", |
| 67 | o3_asm |
| 68 | ); |
| 69 | assert!( |
| 70 | o3_asm.contains("addv.4s"), |
| 71 | "VReduceSum should lower to `addv.4s` for an i32 accumulator:\n{}", |
| 72 | o3_asm |
| 73 | ); |
| 74 | |
| 75 | // Runtime: sum(1..32) = 32*33/2 = 528. |
| 76 | let stdout = capture_run_stdout(CaptureRequest { |
| 77 | input: source, |
| 78 | requested: BTreeSet::from([Stage::Run]), |
| 79 | opt_level: OptLevel::O3, |
| 80 | }); |
| 81 | let trimmed: Vec<&str> = stdout |
| 82 | .lines() |
| 83 | .map(|l| l.trim()) |
| 84 | .filter(|l| !l.is_empty()) |
| 85 | .collect(); |
| 86 | assert_eq!( |
| 87 | trimmed, |
| 88 | vec!["528"], |
| 89 | "vectorized sum reduction should produce 528:\n{}", |
| 90 | stdout |
| 91 | ); |
| 92 | } |
| 93 |