| 1 | //! Determinism sweep — compile every test program twice at -O2 and |
| 2 | //! verify byte-identical assembly output. Catches non-deterministic |
| 3 | //! codegen (HashMap iteration order, unstable sorts, stale regalloc |
| 4 | //! state) across the entire test corpus, not just the subset with |
| 5 | //! explicit REPRO_CHECK annotations. |
| 6 | //! |
| 7 | //! Multi-file tests (those containing `!--- file:` markers) and |
| 8 | //! error-expected tests are skipped — they either need special |
| 9 | //! handling or don't produce assembly. |
| 10 | |
| 11 | use std::fs; |
| 12 | use std::path::{Path, PathBuf}; |
| 13 | use std::process::Command; |
| 14 | use std::sync::atomic::{AtomicUsize, Ordering}; |
| 15 | |
| 16 | static NEXT_ID: AtomicUsize = AtomicUsize::new(0); |
| 17 | |
| 18 | fn find_compiler() -> PathBuf { |
| 19 | for c in &["target/release/armfortas", "target/debug/armfortas"] { |
| 20 | let p = PathBuf::from(c); |
| 21 | if p.exists() { |
| 22 | return fs::canonicalize(&p).unwrap(); |
| 23 | } |
| 24 | } |
| 25 | panic!("armfortas binary not found — run `cargo build` first"); |
| 26 | } |
| 27 | |
| 28 | fn find_test_programs() -> PathBuf { |
| 29 | for c in &["test_programs", "../test_programs"] { |
| 30 | let p = PathBuf::from(c); |
| 31 | if p.is_dir() { |
| 32 | return p; |
| 33 | } |
| 34 | } |
| 35 | panic!("cannot find test_programs/ directory"); |
| 36 | } |
| 37 | |
| 38 | fn unique_path(prefix: &str, ext: &str) -> PathBuf { |
| 39 | let id = NEXT_ID.fetch_add(1, Ordering::Relaxed); |
| 40 | std::env::temp_dir().join(format!( |
| 41 | "afs_det_{}_{}_{}{}", |
| 42 | prefix, |
| 43 | std::process::id(), |
| 44 | id, |
| 45 | ext |
| 46 | )) |
| 47 | } |
| 48 | |
| 49 | /// Compile a source to assembly, returning the bytes. |
| 50 | /// Returns None if compilation fails (e.g., error-expected tests). |
| 51 | fn compile_to_asm(compiler: &Path, source: &Path, opt: &str) -> Option<Vec<u8>> { |
| 52 | let out = unique_path("sweep", ".s"); |
| 53 | let result = Command::new(compiler) |
| 54 | .args([ |
| 55 | source.to_str().unwrap(), |
| 56 | opt, |
| 57 | "-S", |
| 58 | "-o", |
| 59 | out.to_str().unwrap(), |
| 60 | ]) |
| 61 | .output() |
| 62 | .expect("compiler launch failed"); |
| 63 | if !result.status.success() { |
| 64 | let _ = fs::remove_file(&out); |
| 65 | return None; |
| 66 | } |
| 67 | let bytes = fs::read(&out).expect("cannot read .s output"); |
| 68 | let _ = fs::remove_file(&out); |
| 69 | Some(bytes) |
| 70 | } |
| 71 | |
| 72 | fn should_skip(source_text: &str) -> bool { |
| 73 | // Skip multi-file tests. |
| 74 | if source_text.contains("!--- file:") { |
| 75 | return true; |
| 76 | } |
| 77 | // Skip error-expected tests (they're supposed to fail compilation). |
| 78 | if source_text.contains("! ERROR_EXPECTED:") { |
| 79 | return true; |
| 80 | } |
| 81 | // Skip XFAIL tests (they have known failures). |
| 82 | if source_text.contains("! XFAIL:") { |
| 83 | return true; |
| 84 | } |
| 85 | false |
| 86 | } |
| 87 | |
| 88 | #[test] |
| 89 | fn all_programs_deterministic_at_o2() { |
| 90 | let compiler = find_compiler(); |
| 91 | let test_dir = find_test_programs(); |
| 92 | |
| 93 | let mut sources: Vec<PathBuf> = fs::read_dir(&test_dir) |
| 94 | .expect("cannot read test_programs/") |
| 95 | .filter_map(|e| e.ok()) |
| 96 | .map(|e| e.path()) |
| 97 | .filter(|p| p.extension().map(|e| e == "f90").unwrap_or(false)) |
| 98 | .collect(); |
| 99 | sources.sort(); |
| 100 | |
| 101 | let mut checked = 0; |
| 102 | let mut skipped = 0; |
| 103 | let mut failures = Vec::new(); |
| 104 | |
| 105 | for source in &sources { |
| 106 | let name = source.file_name().unwrap().to_str().unwrap(); |
| 107 | let text = fs::read_to_string(source).unwrap(); |
| 108 | if should_skip(&text) { |
| 109 | skipped += 1; |
| 110 | continue; |
| 111 | } |
| 112 | |
| 113 | let first = match compile_to_asm(&compiler, source, "-O2") { |
| 114 | Some(bytes) => bytes, |
| 115 | None => { |
| 116 | // Compilation failed — this test has other issues; skip. |
| 117 | skipped += 1; |
| 118 | continue; |
| 119 | } |
| 120 | }; |
| 121 | let second = compile_to_asm(&compiler, source, "-O2") |
| 122 | .expect("second compile failed but first succeeded — flaky"); |
| 123 | |
| 124 | if first != second { |
| 125 | failures.push(format!( |
| 126 | "{}: assembly differs between two -O2 compilations ({} vs {} bytes)", |
| 127 | name, |
| 128 | first.len(), |
| 129 | second.len(), |
| 130 | )); |
| 131 | } else { |
| 132 | checked += 1; |
| 133 | } |
| 134 | } |
| 135 | |
| 136 | eprintln!( |
| 137 | "\nDeterminism sweep: {} checked, {} skipped, {} failures out of {} programs", |
| 138 | checked, |
| 139 | skipped, |
| 140 | failures.len(), |
| 141 | sources.len(), |
| 142 | ); |
| 143 | |
| 144 | if !failures.is_empty() { |
| 145 | panic!("Determinism failures:\n{}", failures.join("\n"),); |
| 146 | } |
| 147 | |
| 148 | // Sanity: we should have checked a substantial portion. |
| 149 | assert!( |
| 150 | checked >= 150, |
| 151 | "only {} programs checked — expected at least 150", |
| 152 | checked, |
| 153 | ); |
| 154 | } |
| 155 |