main.rs

  1//! Perf profiler for Zed tests. Outputs timings of tests marked with the `#[perf]`
  2//! attribute to stdout in Markdown. See the documentation of `util_macros::perf`
  3//! for usage details on the actual attribute.
  4//!
  5//! # Setup
  6//! Make sure `hyperfine` is installed and in the shell path.
  7//!
  8//! # Usage
  9//! Calling this tool rebuilds the targeted crate(s) with some cfg flags set for the
 10//! perf proc macro *and* enables optimisations (`release-fast` profile), so expect
 11//! it to take a little while.
 12//!
 13//! To test an individual crate, run:
 14//! ```sh
 15//! cargo perf-test -p $CRATE
 16//! ```
 17//!
 18//! To test everything (which will be **VERY SLOW**), run:
 19//! ```sh
 20//! cargo perf-test --workspace
 21//! ```
 22//!
 23//! Some command-line parameters are also recognised by this profiler. To filter
 24//! out all tests below a certain importance (e.g. `important`), run:
 25//! ```sh
 26//! cargo perf-test $WHATEVER -- --important
 27//! ```
 28//!
 29//! Similarly, to skip outputting progress to the command line, pass `-- --quiet`.
 30//! These flags can be combined.
 31//!
 32//! ## Comparing runs
 33//! Passing `--json=ident` will save per-crate run files in `.perf-runs`, e.g.
 34//! `cargo perf-test -p gpui -- --json=blah` will result in `.perf-runs/blah.gpui.json`
 35//! being created (unless no tests were run). These results can be automatically
 36//! compared. To do so, run `cargo perf-compare new-ident old-ident`.
 37//!
 38//! NB: All files matching `.perf-runs/ident.*.json` will be considered when
 39//! doing this comparison, so ensure there aren't leftover files in your `.perf-runs`
 40//! directory that might match that!
 41//!
 42//! # Notes
 43//! This should probably not be called manually unless you're working on the profiler
 44//! itself; use the `cargo perf-test` alias (after building this crate) instead.
 45
 46use perf::{FailKind, Importance, Output, TestMdata, Timings, consts};
 47
 48use std::{
 49    fs::OpenOptions,
 50    io::Write,
 51    num::NonZero,
 52    path::{Path, PathBuf},
 53    process::{Command, Stdio},
 54    sync::atomic::{AtomicBool, Ordering},
 55    time::{Duration, Instant},
 56};
 57
 58/// How many iterations to attempt the first time a test is run.
 59const DEFAULT_ITER_COUNT: NonZero<usize> = NonZero::new(3).unwrap();
 60/// Multiplier for the iteration count when a test doesn't pass the noise cutoff.
 61const ITER_COUNT_MUL: NonZero<usize> = NonZero::new(4).unwrap();
 62
 63/// Do we keep stderr empty while running the tests?
 64static QUIET: AtomicBool = AtomicBool::new(false);
 65
 66/// Report a failure into the output and skip an iteration.
 67macro_rules! fail {
 68    ($output:ident, $name:expr, $kind:expr) => {{
 69        $output.failure($name, None, None, $kind);
 70        continue;
 71    }};
 72    ($output:ident, $name:expr, $mdata:expr, $kind:expr) => {{
 73        $output.failure($name, Some($mdata), None, $kind);
 74        continue;
 75    }};
 76    ($output:ident, $name:expr, $mdata:expr, $count:expr, $kind:expr) => {{
 77        $output.failure($name, Some($mdata), Some($count), $kind);
 78        continue;
 79    }};
 80}
 81
 82/// How does this perf run return its output?
 83enum OutputKind<'a> {
 84    /// Print markdown to the terminal.
 85    Markdown,
 86    /// Save JSON to a file.
 87    Json(&'a Path),
 88}
 89
 90impl OutputKind<'_> {
 91    /// Logs the output of a run as per the `OutputKind`.
 92    fn log(&self, output: &Output, t_bin: &str) {
 93        match self {
 94            OutputKind::Markdown => print!("{output}"),
 95            OutputKind::Json(ident) => {
 96                // We're going to be in tooling/perf/$whatever.
 97                let wspace_dir = PathBuf::from(std::env::var("CARGO_MANIFEST_DIR").unwrap())
 98                    .join("..")
 99                    .join("..");
100                let runs_dir = PathBuf::from(&wspace_dir).join(consts::RUNS_DIR);
101                std::fs::create_dir_all(&runs_dir).unwrap();
102                assert!(
103                    !ident.to_string_lossy().is_empty(),
104                    "FATAL: Empty filename specified!"
105                );
106                // Get the test binary's crate's name; a path like
107                // target/release-fast/deps/gpui-061ff76c9b7af5d7
108                // would be reduced to just "gpui".
109                let test_bin_stripped = Path::new(t_bin)
110                    .file_name()
111                    .unwrap()
112                    .to_str()
113                    .unwrap()
114                    .rsplit_once('-')
115                    .unwrap()
116                    .0;
117                let mut file_path = runs_dir.join(ident);
118                file_path
119                    .as_mut_os_string()
120                    .push(format!(".{test_bin_stripped}.json"));
121                let mut out_file = OpenOptions::new()
122                    .write(true)
123                    .create(true)
124                    .truncate(true)
125                    .open(&file_path)
126                    .unwrap();
127                out_file
128                    .write_all(&serde_json::to_vec(&output).unwrap())
129                    .unwrap();
130                if !QUIET.load(Ordering::Relaxed) {
131                    eprintln!("JSON output written to {}", file_path.display());
132                }
133            }
134        }
135    }
136}
137
138/// Runs a given metadata-returning function from a test handler, parsing its
139/// output into a `TestMdata`.
140fn parse_mdata(t_bin: &str, mdata_fn: &str) -> Result<TestMdata, FailKind> {
141    let mut cmd = Command::new(t_bin);
142    cmd.args([mdata_fn, "--exact", "--nocapture"]);
143    let out = cmd
144        .output()
145        .expect("FATAL: Could not run test binary {t_bin}");
146    assert!(out.status.success());
147    let stdout = String::from_utf8_lossy(&out.stdout);
148    let mut version = None;
149    let mut iterations = None;
150    let mut importance = Importance::default();
151    let mut weight = consts::WEIGHT_DEFAULT;
152    for line in stdout
153        .lines()
154        .filter_map(|l| l.strip_prefix(consts::MDATA_LINE_PREF))
155    {
156        let mut items = line.split_whitespace();
157        // For v0, we know the ident always comes first, then one field.
158        match items.next().ok_or(FailKind::BadMetadata)? {
159            consts::VERSION_LINE_NAME => {
160                let v = items
161                    .next()
162                    .ok_or(FailKind::BadMetadata)?
163                    .parse::<u32>()
164                    .map_err(|_| FailKind::BadMetadata)?;
165                if v > consts::MDATA_VER {
166                    return Err(FailKind::VersionMismatch);
167                }
168                version = Some(v);
169            }
170            consts::ITER_COUNT_LINE_NAME => {
171                // This should never be zero!
172                iterations = Some(
173                    items
174                        .next()
175                        .ok_or(FailKind::BadMetadata)?
176                        .parse::<usize>()
177                        .map_err(|_| FailKind::BadMetadata)?
178                        .try_into()
179                        .map_err(|_| FailKind::BadMetadata)?,
180                );
181            }
182            consts::IMPORTANCE_LINE_NAME => {
183                importance = match items.next().ok_or(FailKind::BadMetadata)? {
184                    "critical" => Importance::Critical,
185                    "important" => Importance::Important,
186                    "average" => Importance::Average,
187                    "iffy" => Importance::Iffy,
188                    "fluff" => Importance::Fluff,
189                    _ => return Err(FailKind::BadMetadata),
190                };
191            }
192            consts::WEIGHT_LINE_NAME => {
193                weight = items
194                    .next()
195                    .ok_or(FailKind::BadMetadata)?
196                    .parse::<u8>()
197                    .map_err(|_| FailKind::BadMetadata)?;
198            }
199            _ => unreachable!(),
200        }
201    }
202
203    Ok(TestMdata {
204        version: version.ok_or(FailKind::BadMetadata)?,
205        // Iterations may be determined by us and thus left unspecified.
206        iterations,
207        // In principle this should always be set, but just for the sake of
208        // stability allow the potentially-breaking change of not reporting the
209        // importance without erroring. Maybe we want to change this.
210        importance,
211        // Same with weight.
212        weight,
213    })
214}
215
216/// Compares the perf results of two profiles as per the arguments passed in.
217fn compare_profiles(args: &[String]) {
218    let ident_new = args.first().expect("FATAL: missing identifier for new run");
219    let ident_old = args.get(1).expect("FATAL: missing identifier for old run");
220    let wspace_dir = std::env::var("CARGO_MANIFEST_DIR").unwrap();
221    let runs_dir = PathBuf::from(&wspace_dir)
222        .join("..")
223        .join("..")
224        .join(consts::RUNS_DIR);
225
226    // Use the blank outputs initially, so we can merge into these with prefixes.
227    let mut outputs_new = Output::blank();
228    let mut outputs_old = Output::blank();
229
230    for e in runs_dir.read_dir().unwrap() {
231        let Ok(entry) = e else {
232            continue;
233        };
234        let Ok(metadata) = entry.metadata() else {
235            continue;
236        };
237        if metadata.is_file() {
238            let Ok(name) = entry.file_name().into_string() else {
239                continue;
240            };
241
242            // A little helper to avoid code duplication. Reads the `output` from
243            // a json file, then merges it into what we have so far.
244            let read_into = |output: &mut Output| {
245                let mut elems = name.split('.').skip(1);
246                let prefix = elems.next().unwrap();
247                assert_eq!("json", elems.next().unwrap());
248                assert!(elems.next().is_none());
249                let handle = OpenOptions::new().read(true).open(entry.path()).unwrap();
250                let o_other: Output = serde_json::from_reader(handle).unwrap();
251                output.merge(o_other, prefix);
252            };
253
254            if name.starts_with(ident_old) {
255                read_into(&mut outputs_old);
256            } else if name.starts_with(ident_new) {
257                read_into(&mut outputs_new);
258            }
259        }
260    }
261
262    let res = outputs_new.compare_perf(outputs_old);
263    println!("{res}");
264}
265
266/// Runs a test binary, filtering out tests which aren't marked for perf triage
267/// and giving back the list of tests we care about.
268///
269/// The output of this is an iterator over `test_fn_name, test_mdata_name`.
270fn get_tests(t_bin: &str) -> impl ExactSizeIterator<Item = (String, String)> {
271    let mut cmd = Command::new(t_bin);
272    // --format=json is nightly-only :(
273    cmd.args(["--list", "--format=terse"]);
274    let out = cmd
275        .output()
276        .expect("FATAL: Could not run test binary {t_bin}");
277    assert!(
278        out.status.success(),
279        "FATAL: Cannot do perf check - test binary {t_bin} returned an error"
280    );
281    if !QUIET.load(Ordering::Relaxed) {
282        eprintln!("Test binary ran successfully; starting profile...");
283    }
284    // Parse the test harness output to look for tests we care about.
285    let stdout = String::from_utf8_lossy(&out.stdout);
286    let mut test_list: Vec<_> = stdout
287        .lines()
288        .filter_map(|line| {
289            // This should split only in two; e.g.,
290            // "app::test::test_arena: test" => "app::test::test_arena:", "test"
291            let line: Vec<_> = line.split_whitespace().collect();
292            match line[..] {
293                // Final byte of t_name is ":", which we need to ignore.
294                [t_name, kind] => (kind == "test").then(|| &t_name[..t_name.len() - 1]),
295                _ => None,
296            }
297        })
298        // Exclude tests that aren't marked for perf triage based on suffix.
299        .filter(|t_name| {
300            t_name.ends_with(consts::SUF_NORMAL) || t_name.ends_with(consts::SUF_MDATA)
301        })
302        .collect();
303
304    // Pulling itertools just for .dedup() would be quite a big dependency that's
305    // not used elsewhere, so do this on a vec instead.
306    test_list.sort_unstable();
307    test_list.dedup();
308
309    // Tests should come in pairs with their mdata fn!
310    assert!(
311        test_list.len().is_multiple_of(2),
312        "Malformed tests in test binary {t_bin}"
313    );
314
315    let out = test_list
316        .chunks_exact_mut(2)
317        .map(|pair| {
318            // Be resilient against changes to these constants.
319            if consts::SUF_NORMAL < consts::SUF_MDATA {
320                (pair[0].to_owned(), pair[1].to_owned())
321            } else {
322                (pair[1].to_owned(), pair[0].to_owned())
323            }
324        })
325        .collect::<Vec<_>>();
326    out.into_iter()
327}
328
329/// Runs the specified test `count` times, returning the time taken if the test
330/// succeeded.
331#[inline]
332fn spawn_and_iterate(t_bin: &str, t_name: &str, count: NonZero<usize>) -> Option<Duration> {
333    let mut cmd = Command::new(t_bin);
334    cmd.args([t_name, "--exact"]);
335    cmd.env(consts::ITER_ENV_VAR, format!("{count}"));
336    // Don't let the child muck up our stdin/out/err.
337    cmd.stdin(Stdio::null());
338    cmd.stdout(Stdio::null());
339    cmd.stderr(Stdio::null());
340    let pre = Instant::now();
341    // Discard the output beyond ensuring success.
342    let out = cmd.spawn().unwrap().wait();
343    let post = Instant::now();
344    out.iter().find_map(|s| s.success().then_some(post - pre))
345}
346
347/// Triage a test to determine the correct number of iterations that it should run.
348/// Specifically, repeatedly runs the given test until its execution time exceeds
349/// `thresh`, calling `step(iterations)` after every failed run to determine the new
350/// iteration count. Returns `None` if the test errored or `step` returned `None`,
351/// else `Some(iterations)`.
352///
353/// # Panics
354/// This will panic if `step(usize)` is not monotonically increasing, or if the test
355/// binary is invalid.
356fn triage_test(
357    t_bin: &str,
358    t_name: &str,
359    thresh: Duration,
360    mut step: impl FnMut(NonZero<usize>) -> Option<NonZero<usize>>,
361) -> Option<NonZero<usize>> {
362    let mut iter_count = DEFAULT_ITER_COUNT;
363    // It's possible that the first loop of a test might be an outlier (e.g. it's
364    // doing some caching), in which case we want to skip it.
365    let duration_once = spawn_and_iterate(t_bin, t_name, NonZero::new(1).unwrap())?;
366    loop {
367        let duration = spawn_and_iterate(t_bin, t_name, iter_count)?;
368        if duration.saturating_sub(duration_once) > thresh {
369            break Some(iter_count);
370        }
371        let new = step(iter_count)?;
372        assert!(
373            new > iter_count,
374            "FATAL: step must be monotonically increasing"
375        );
376        iter_count = new;
377    }
378}
379
380/// Profiles a given test with hyperfine, returning the mean and standard deviation
381/// for its runtime. If the test errors, returns `None` instead.
382fn hyp_profile(t_bin: &str, t_name: &str, iterations: NonZero<usize>) -> Option<Timings> {
383    let mut perf_cmd = Command::new("hyperfine");
384    // Warm up the cache and print markdown output to stdout, which we parse.
385    perf_cmd.args([
386        "--style",
387        "none",
388        "--warmup",
389        "1",
390        "--export-markdown",
391        "-",
392        // Parse json instead...
393        "--time-unit",
394        "millisecond",
395        &format!("{t_bin} --exact {t_name}"),
396    ]);
397    perf_cmd.env(consts::ITER_ENV_VAR, format!("{iterations}"));
398    let p_out = perf_cmd.output().unwrap();
399    if !p_out.status.success() {
400        return None;
401    }
402
403    let cmd_output = String::from_utf8_lossy(&p_out.stdout);
404    // Can't use .last() since we have a trailing newline. Sigh.
405    let results_line = cmd_output.lines().nth(3).unwrap();
406    // Grab the values out of the pretty-print.
407    // TODO: Parse json instead.
408    let mut res_iter = results_line.split_whitespace();
409    // Durations are given in milliseconds, so account for that.
410    let mean = Duration::from_secs_f64(res_iter.nth(5).unwrap().parse::<f64>().unwrap() / 1000.);
411    let stddev = Duration::from_secs_f64(res_iter.nth(1).unwrap().parse::<f64>().unwrap() / 1000.);
412
413    Some(Timings { mean, stddev })
414}
415
416fn main() {
417    let args = std::env::args().collect::<Vec<_>>();
418    // We get passed the test we need to run as the 1st argument after our own name.
419    let t_bin = args
420        .get(1)
421        .expect("FATAL: No test binary or command; this shouldn't be manually invoked!");
422
423    // We're being asked to compare two results, not run the profiler.
424    if t_bin == "compare" {
425        compare_profiles(&args[2..]);
426        return;
427    }
428
429    // Minimum test importance we care about this run.
430    let mut thresh = Importance::Iffy;
431    // Where to print the output of this run.
432    let mut out_kind = OutputKind::Markdown;
433
434    for arg in args.iter().skip(2) {
435        match arg.as_str() {
436            "--critical" => thresh = Importance::Critical,
437            "--important" => thresh = Importance::Important,
438            "--average" => thresh = Importance::Average,
439            "--iffy" => thresh = Importance::Iffy,
440            "--fluff" => thresh = Importance::Fluff,
441            "--quiet" => QUIET.store(true, Ordering::Relaxed),
442            s if s.starts_with("--json") => {
443                out_kind = OutputKind::Json(Path::new(
444                    s.strip_prefix("--json=")
445                        .expect("FATAL: Invalid json parameter; pass --json=ident"),
446                ));
447            }
448            _ => (),
449        }
450    }
451    if !QUIET.load(Ordering::Relaxed) {
452        eprintln!("Starting perf check");
453    }
454
455    let mut output = Output::default();
456
457    // Spawn and profile an instance of each perf-sensitive test, via hyperfine.
458    // Each test is a pair of (test, metadata-returning-fn), so grab both. We also
459    // know the list is sorted.
460    let i = get_tests(t_bin);
461    let len = i.len();
462    for (idx, (ref t_name, ref t_mdata)) in i.enumerate() {
463        if !QUIET.load(Ordering::Relaxed) {
464            eprint!("\rProfiling test {}/{}", idx + 1, len);
465        }
466        // Pretty-printable stripped name for the test.
467        let t_name_pretty = t_name.replace(consts::SUF_NORMAL, "");
468
469        // Get the metadata this test reports for us.
470        let t_mdata = match parse_mdata(t_bin, t_mdata) {
471            Ok(mdata) => mdata,
472            Err(err) => fail!(output, t_name_pretty, err),
473        };
474
475        if t_mdata.importance < thresh {
476            fail!(output, t_name_pretty, t_mdata, FailKind::Skipped);
477        }
478
479        // Time test execution to see how many iterations we need to do in order
480        // to account for random noise. This is skipped for tests with fixed
481        // iteration counts.
482        let final_iter_count = t_mdata.iterations.or_else(|| {
483            triage_test(t_bin, t_name, consts::NOISE_CUTOFF, |c| {
484                if let Some(c) = c.checked_mul(ITER_COUNT_MUL) {
485                    Some(c)
486                } else {
487                    // This should almost never happen, but maybe..?
488                    eprintln!(
489                        "WARNING: Ran nearly usize::MAX iterations of test {t_name_pretty}; skipping"
490                    );
491                    None
492                }
493            })
494        });
495
496        // Don't profile failing tests.
497        let Some(final_iter_count) = final_iter_count else {
498            fail!(output, t_name_pretty, t_mdata, FailKind::Triage);
499        };
500
501        // Now profile!
502        if let Some(timings) = hyp_profile(t_bin, t_name, final_iter_count) {
503            output.success(t_name_pretty, t_mdata, final_iter_count, timings);
504        } else {
505            fail!(
506                output,
507                t_name_pretty,
508                t_mdata,
509                final_iter_count,
510                FailKind::Profile
511            );
512        }
513    }
514    if !QUIET.load(Ordering::Relaxed) {
515        if output.is_empty() {
516            eprintln!("Nothing to do.");
517        } else {
518            // If stdout and stderr are on the same terminal, move us after the
519            // output from above.
520            eprintln!();
521        }
522    }
523
524    // No need making an empty json file on every empty test bin.
525    if output.is_empty() {
526        return;
527    }
528
529    out_kind.log(&output, t_bin);
530}