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