| 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 function_section<'a>(ir: &'a str, name: &str) -> &'a str { |
| 23 | let header = format!("func @{}", name); |
| 24 | let start = ir |
| 25 | .find(&header) |
| 26 | .unwrap_or_else(|| panic!("missing function section for {}", name)); |
| 27 | let rest = &ir[start..]; |
| 28 | let end = rest |
| 29 | .find("\n }\n") |
| 30 | .unwrap_or_else(|| panic!("unterminated function section for {}", name)); |
| 31 | &rest[..end + "\n }".len()] |
| 32 | } |
| 33 | |
| 34 | fn block_section<'a>(func_section: &'a str, prefix: &str) -> &'a str { |
| 35 | let mut start = None; |
| 36 | let mut end = None; |
| 37 | |
| 38 | for (idx, _line) in func_section.match_indices('\n') { |
| 39 | let line_start = idx + 1; |
| 40 | let tail = &func_section[line_start..]; |
| 41 | let line_text = tail.split_once('\n').map(|(line, _)| line).unwrap_or(tail); |
| 42 | |
| 43 | if start.is_none() { |
| 44 | if line_text.starts_with(" ") |
| 45 | && !line_text.starts_with(" ") |
| 46 | && line_text[4..].starts_with(prefix) |
| 47 | { |
| 48 | start = Some(line_start); |
| 49 | } |
| 50 | continue; |
| 51 | } |
| 52 | |
| 53 | if line_text.starts_with(" ") && !line_text.starts_with(" ") { |
| 54 | end = Some(idx); |
| 55 | break; |
| 56 | } |
| 57 | if line_text == " }" { |
| 58 | end = Some(idx); |
| 59 | break; |
| 60 | } |
| 61 | } |
| 62 | |
| 63 | let start = start.unwrap_or_else(|| panic!("missing block with prefix {}", prefix)); |
| 64 | let end = end.unwrap_or(func_section.len()); |
| 65 | &func_section[start..end] |
| 66 | } |
| 67 | |
| 68 | fn last_block_section<'a>(func_section: &'a str, prefix: &str) -> &'a str { |
| 69 | let mut starts = Vec::new(); |
| 70 | |
| 71 | for (idx, _line) in func_section.match_indices('\n') { |
| 72 | let line_start = idx + 1; |
| 73 | let tail = &func_section[line_start..]; |
| 74 | let line_text = tail.split_once('\n').map(|(line, _)| line).unwrap_or(tail); |
| 75 | if line_text.starts_with(" ") |
| 76 | && !line_text.starts_with(" ") |
| 77 | && line_text[4..].starts_with(prefix) |
| 78 | { |
| 79 | starts.push(line_start); |
| 80 | } |
| 81 | } |
| 82 | |
| 83 | let start = *starts |
| 84 | .last() |
| 85 | .unwrap_or_else(|| panic!("missing block with prefix {}", prefix)); |
| 86 | let tail = &func_section[start..]; |
| 87 | let end_rel = tail |
| 88 | .match_indices('\n') |
| 89 | .find_map(|(idx, _)| { |
| 90 | let line_start = idx + 1; |
| 91 | let line_tail = &tail[line_start..]; |
| 92 | let line_text = line_tail |
| 93 | .split_once('\n') |
| 94 | .map(|(line, _)| line) |
| 95 | .unwrap_or(line_tail); |
| 96 | if (line_text.starts_with(" ") && !line_text.starts_with(" ")) |
| 97 | || line_text == " }" |
| 98 | { |
| 99 | Some(idx) |
| 100 | } else { |
| 101 | None |
| 102 | } |
| 103 | }) |
| 104 | .unwrap_or(tail.len()); |
| 105 | &tail[..end_rel] |
| 106 | } |
| 107 | |
| 108 | #[test] |
| 109 | fn o2_reuses_branch_join_affine_expression() { |
| 110 | let source = fixture("realworld_join_bias_sum.f90"); |
| 111 | |
| 112 | let raw_ir = capture_text( |
| 113 | CaptureRequest { |
| 114 | input: source.clone(), |
| 115 | requested: BTreeSet::from([Stage::Ir]), |
| 116 | opt_level: OptLevel::O0, |
| 117 | }, |
| 118 | Stage::Ir, |
| 119 | ); |
| 120 | let opt_ir = capture_text( |
| 121 | CaptureRequest { |
| 122 | input: source, |
| 123 | requested: BTreeSet::from([Stage::OptIr]), |
| 124 | opt_level: OptLevel::O2, |
| 125 | }, |
| 126 | Stage::OptIr, |
| 127 | ); |
| 128 | |
| 129 | let raw_tally = function_section(&raw_ir, "tally"); |
| 130 | let raw_join = last_block_section(raw_tally, "if_end_"); |
| 131 | |
| 132 | assert!( |
| 133 | raw_join.matches("call @offset_value").count() >= 2, |
| 134 | "raw join block should still recompute the repeated branch-join PURE helper call:\n{}", |
| 135 | raw_join |
| 136 | ); |
| 137 | assert!( |
| 138 | opt_ir.matches("call @offset_value").count() < raw_ir.matches("call @offset_value").count(), |
| 139 | "O2 should reduce duplicated branch-join PURE helper calls:\n{}", |
| 140 | opt_ir |
| 141 | ); |
| 142 | } |
| 143 | |
| 144 | #[test] |
| 145 | fn o2_removes_dead_seed_store_across_noalias_call() { |
| 146 | let source = fixture("realworld_seed_overwrite.f90"); |
| 147 | |
| 148 | let raw_ir = capture_text( |
| 149 | CaptureRequest { |
| 150 | input: source.clone(), |
| 151 | requested: BTreeSet::from([Stage::Ir]), |
| 152 | opt_level: OptLevel::O0, |
| 153 | }, |
| 154 | Stage::Ir, |
| 155 | ); |
| 156 | let opt_ir = capture_text( |
| 157 | CaptureRequest { |
| 158 | input: source, |
| 159 | requested: BTreeSet::from([Stage::OptIr]), |
| 160 | opt_level: OptLevel::O2, |
| 161 | }, |
| 162 | Stage::OptIr, |
| 163 | ); |
| 164 | |
| 165 | let raw_fill = function_section(&raw_ir, "seed_and_fill"); |
| 166 | let opt_fill = function_section(&opt_ir, "seed_and_fill"); |
| 167 | let raw_body = block_section(raw_fill, "do_body_"); |
| 168 | let opt_body = block_section(opt_fill, "do_body_"); |
| 169 | |
| 170 | assert!( |
| 171 | raw_body.matches("store ").count() >= 2, |
| 172 | "raw loop body should still contain the seed store and the real fill store:\n{}", |
| 173 | raw_body |
| 174 | ); |
| 175 | assert!( |
| 176 | opt_body.matches("store ").count() < raw_body.matches("store ").count(), |
| 177 | "O2 should remove the dead seed store while keeping the real fill:\n{}", |
| 178 | opt_body |
| 179 | ); |
| 180 | } |
| 181 |