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 $WHATEVER -- --important
27//! ```
28//!
29//! Similarly, to skip outputting progress to the command line, pass `-- --quiet`.
30//! These flags can be combined.
31//!
32//! ## Comparing runs
33//! Passing `--json=ident` will save per-crate run files in `.perf-runs`, e.g.
34//! `cargo perf-test -p gpui -- --json=blah` will result in `.perf-runs/blah.gpui.json`
35//! being created (unless no tests were run). These results can be automatically
36//! compared. To do so, run `cargo perf-compare new-ident old-ident`.
37//!
38//! To save the markdown output to a file instead, run `cargo perf-compare --save=$FILE
39//! new-ident old-ident`.
40//!
41//! NB: All files matching `.perf-runs/ident.*.json` will be considered when
42//! doing this comparison, so ensure there aren't leftover files in your `.perf-runs`
43//! directory that might match that!
44//!
45//! # Notes
46//! This should probably not be called manually unless you're working on the profiler
47//! itself; use the `cargo perf-test` alias (after building this crate) instead.
48
49use zed_perf::{FailKind, Importance, Output, TestMdata, Timings, consts};
50
51use std::{
52 fs::OpenOptions,
53 io::{Read, Write},
54 num::NonZero,
55 path::{Path, PathBuf},
56 process::{Command, Stdio},
57 sync::atomic::{AtomicBool, Ordering},
58 time::{Duration, Instant},
59};
60
61/// How many iterations to attempt the first time a test is run.
62const DEFAULT_ITER_COUNT: NonZero<usize> = NonZero::new(3).unwrap();
63/// Multiplier for the iteration count when a test doesn't pass the noise cutoff.
64const ITER_COUNT_MUL: NonZero<usize> = NonZero::new(4).unwrap();
65
66/// Do we keep stderr empty while running the tests?
67static QUIET: AtomicBool = AtomicBool::new(false);
68
69/// Report a failure into the output and skip an iteration.
70macro_rules! fail {
71 ($output:ident, $name:expr, $kind:expr) => {{
72 $output.failure($name, None, None, $kind);
73 continue;
74 }};
75 ($output:ident, $name:expr, $mdata:expr, $kind:expr) => {{
76 $output.failure($name, Some($mdata), None, $kind);
77 continue;
78 }};
79 ($output:ident, $name:expr, $mdata:expr, $count:expr, $kind:expr) => {{
80 $output.failure($name, Some($mdata), Some($count), $kind);
81 continue;
82 }};
83}
84
85/// How does this perf run return its output?
86enum OutputKind<'a> {
87 /// Print markdown to the terminal.
88 Markdown,
89 /// Save JSON to a file.
90 Json(&'a Path),
91}
92
93impl OutputKind<'_> {
94 /// Logs the output of a run as per the `OutputKind`.
95 fn log(&self, output: &Output, t_bin: &str) {
96 match self {
97 OutputKind::Markdown => println!("{output}"),
98 OutputKind::Json(ident) => {
99 // We're going to be in tooling/perf/$whatever.
100 let wspace_dir = PathBuf::from(std::env::var("CARGO_MANIFEST_DIR").unwrap())
101 .join("..")
102 .join("..");
103 let runs_dir = PathBuf::from(&wspace_dir).join(consts::RUNS_DIR);
104 std::fs::create_dir_all(&runs_dir).unwrap();
105 assert!(
106 !ident.to_string_lossy().is_empty(),
107 "FATAL: Empty filename specified!"
108 );
109 // Get the test binary's crate's name; a path like
110 // target/release-fast/deps/gpui-061ff76c9b7af5d7
111 // would be reduced to just "gpui".
112 let test_bin_stripped = Path::new(t_bin)
113 .file_name()
114 .unwrap()
115 .to_str()
116 .unwrap()
117 .rsplit_once('-')
118 .unwrap()
119 .0;
120 let mut file_path = runs_dir.join(ident);
121 file_path
122 .as_mut_os_string()
123 .push(format!(".{test_bin_stripped}.json"));
124 let mut out_file = OpenOptions::new()
125 .write(true)
126 .create(true)
127 .truncate(true)
128 .open(&file_path)
129 .unwrap();
130 out_file
131 .write_all(&serde_json::to_vec(&output).unwrap())
132 .unwrap();
133 if !QUIET.load(Ordering::Relaxed) {
134 eprintln!("JSON output written to {}", file_path.display());
135 }
136 }
137 }
138 }
139}
140
141/// Runs a given metadata-returning function from a test handler, parsing its
142/// output into a `TestMdata`.
143fn parse_mdata(t_bin: &str, mdata_fn: &str) -> Result<TestMdata, FailKind> {
144 let mut cmd = Command::new(t_bin);
145 cmd.args([mdata_fn, "--exact", "--nocapture"]);
146 let out = cmd
147 .output()
148 .expect("FATAL: Could not run test binary {t_bin}");
149 assert!(out.status.success());
150 let stdout = String::from_utf8_lossy(&out.stdout);
151 let mut version = None;
152 let mut iterations = None;
153 let mut importance = Importance::default();
154 let mut weight = consts::WEIGHT_DEFAULT;
155 for line in stdout
156 .lines()
157 .filter_map(|l| l.strip_prefix(consts::MDATA_LINE_PREF))
158 {
159 let mut items = line.split_whitespace();
160 // For v0, we know the ident always comes first, then one field.
161 match items.next().ok_or(FailKind::BadMetadata)? {
162 consts::VERSION_LINE_NAME => {
163 let v = items
164 .next()
165 .ok_or(FailKind::BadMetadata)?
166 .parse::<u32>()
167 .map_err(|_| FailKind::BadMetadata)?;
168 if v > consts::MDATA_VER {
169 return Err(FailKind::VersionMismatch);
170 }
171 version = Some(v);
172 }
173 consts::ITER_COUNT_LINE_NAME => {
174 // This should never be zero!
175 iterations = Some(
176 items
177 .next()
178 .ok_or(FailKind::BadMetadata)?
179 .parse::<usize>()
180 .map_err(|_| FailKind::BadMetadata)?
181 .try_into()
182 .map_err(|_| FailKind::BadMetadata)?,
183 );
184 }
185 consts::IMPORTANCE_LINE_NAME => {
186 importance = match items.next().ok_or(FailKind::BadMetadata)? {
187 "critical" => Importance::Critical,
188 "important" => Importance::Important,
189 "average" => Importance::Average,
190 "iffy" => Importance::Iffy,
191 "fluff" => Importance::Fluff,
192 _ => return Err(FailKind::BadMetadata),
193 };
194 }
195 consts::WEIGHT_LINE_NAME => {
196 weight = items
197 .next()
198 .ok_or(FailKind::BadMetadata)?
199 .parse::<u8>()
200 .map_err(|_| FailKind::BadMetadata)?;
201 }
202 _ => unreachable!(),
203 }
204 }
205
206 Ok(TestMdata {
207 version: version.ok_or(FailKind::BadMetadata)?,
208 // Iterations may be determined by us and thus left unspecified.
209 iterations,
210 // In principle this should always be set, but just for the sake of
211 // stability allow the potentially-breaking change of not reporting the
212 // importance without erroring. Maybe we want to change this.
213 importance,
214 // Same with weight.
215 weight,
216 })
217}
218
219/// Compares the perf results of two profiles as per the arguments passed in.
220fn compare_profiles(args: &[String]) {
221 let mut save_to = None;
222 let mut ident_idx = 0;
223 args.first().inspect(|a| {
224 if a.starts_with("--save") {
225 save_to = Some(
226 a.strip_prefix("--save=")
227 .expect("FATAL: save param formatted incorrectly"),
228 );
229 }
230 ident_idx = 1;
231 });
232 let ident_new = args
233 .get(ident_idx)
234 .expect("FATAL: missing identifier for new run");
235 let ident_old = args
236 .get(ident_idx + 1)
237 .expect("FATAL: missing identifier for old run");
238 let wspace_dir = std::env::var("CARGO_MANIFEST_DIR").unwrap();
239 let runs_dir = PathBuf::from(&wspace_dir)
240 .join("..")
241 .join("..")
242 .join(consts::RUNS_DIR);
243
244 // Use the blank outputs initially, so we can merge into these with prefixes.
245 let mut outputs_new = Output::blank();
246 let mut outputs_old = Output::blank();
247
248 for e in runs_dir.read_dir().unwrap() {
249 let Ok(entry) = e else {
250 continue;
251 };
252 let Ok(metadata) = entry.metadata() else {
253 continue;
254 };
255 if metadata.is_file() {
256 let Ok(name) = entry.file_name().into_string() else {
257 continue;
258 };
259
260 // A little helper to avoid code duplication. Reads the `output` from
261 // a json file, then merges it into what we have so far.
262 let read_into = |output: &mut Output| {
263 let mut elems = name.split('.').skip(1);
264 let prefix = elems.next().unwrap();
265 assert_eq!("json", elems.next().unwrap());
266 assert!(elems.next().is_none());
267 let mut buffer = Vec::new();
268 let _ = OpenOptions::new()
269 .read(true)
270 .open(entry.path())
271 .unwrap()
272 .read_to_end(&mut buffer)
273 .unwrap();
274 let o_other: Output = serde_json::from_slice(&buffer).unwrap();
275 output.merge(o_other, prefix);
276 };
277
278 if name.starts_with(ident_old) {
279 read_into(&mut outputs_old);
280 } else if name.starts_with(ident_new) {
281 read_into(&mut outputs_new);
282 }
283 }
284 }
285
286 let res = outputs_new.compare_perf(outputs_old);
287 if let Some(filename) = save_to {
288 let mut file = std::fs::OpenOptions::new()
289 .create(true)
290 .write(true)
291 .truncate(true)
292 .open(filename)
293 .expect("FATAL: couldn't save run results to file");
294 file.write_all(format!("{res}").as_bytes()).unwrap();
295 } else {
296 println!("{res}");
297 }
298}
299
300/// Runs a test binary, filtering out tests which aren't marked for perf triage
301/// and giving back the list of tests we care about.
302///
303/// The output of this is an iterator over `test_fn_name, test_mdata_name`.
304fn get_tests(t_bin: &str) -> impl ExactSizeIterator<Item = (String, String)> {
305 let mut cmd = Command::new(t_bin);
306 // --format=json is nightly-only :(
307 cmd.args(["--list", "--format=terse"]);
308 let out = cmd
309 .output()
310 .expect("FATAL: Could not run test binary {t_bin}");
311 assert!(
312 out.status.success(),
313 "FATAL: Cannot do perf check - test binary {t_bin} returned an error"
314 );
315 if !QUIET.load(Ordering::Relaxed) {
316 eprintln!("Test binary ran successfully; starting profile...");
317 }
318 // Parse the test harness output to look for tests we care about.
319 let stdout = String::from_utf8_lossy(&out.stdout);
320 let mut test_list: Vec<_> = stdout
321 .lines()
322 .filter_map(|line| {
323 // This should split only in two; e.g.,
324 // "app::test::test_arena: test" => "app::test::test_arena:", "test"
325 let line: Vec<_> = line.split_whitespace().collect();
326 match line[..] {
327 // Final byte of t_name is ":", which we need to ignore.
328 [t_name, kind] => (kind == "test").then(|| &t_name[..t_name.len() - 1]),
329 _ => None,
330 }
331 })
332 // Exclude tests that aren't marked for perf triage based on suffix.
333 .filter(|t_name| {
334 t_name.ends_with(consts::SUF_NORMAL) || t_name.ends_with(consts::SUF_MDATA)
335 })
336 .collect();
337
338 // Pulling itertools just for .dedup() would be quite a big dependency that's
339 // not used elsewhere, so do this on a vec instead.
340 test_list.sort_unstable();
341 test_list.dedup();
342
343 // Tests should come in pairs with their mdata fn!
344 assert!(
345 test_list.len().is_multiple_of(2),
346 "Malformed tests in test binary {t_bin}"
347 );
348
349 let out = test_list
350 .chunks_exact_mut(2)
351 .map(|pair| {
352 // Be resilient against changes to these constants.
353 if consts::SUF_NORMAL < consts::SUF_MDATA {
354 (pair[0].to_owned(), pair[1].to_owned())
355 } else {
356 (pair[1].to_owned(), pair[0].to_owned())
357 }
358 })
359 .collect::<Vec<_>>();
360 out.into_iter()
361}
362
363/// Runs the specified test `count` times, returning the time taken if the test
364/// succeeded.
365#[inline]
366fn spawn_and_iterate(t_bin: &str, t_name: &str, count: NonZero<usize>) -> Option<Duration> {
367 let mut cmd = Command::new(t_bin);
368 cmd.args([t_name, "--exact"]);
369 cmd.env(consts::ITER_ENV_VAR, format!("{count}"));
370 // Don't let the child muck up our stdin/out/err.
371 cmd.stdin(Stdio::null());
372 cmd.stdout(Stdio::null());
373 cmd.stderr(Stdio::null());
374 let pre = Instant::now();
375 // Discard the output beyond ensuring success.
376 let out = cmd.spawn().unwrap().wait();
377 let post = Instant::now();
378 out.iter().find_map(|s| s.success().then_some(post - pre))
379}
380
381/// Triage a test to determine the correct number of iterations that it should run.
382/// Specifically, repeatedly runs the given test until its execution time exceeds
383/// `thresh`, calling `step(iterations)` after every failed run to determine the new
384/// iteration count. Returns `None` if the test errored or `step` returned `None`,
385/// else `Some(iterations)`.
386///
387/// # Panics
388/// This will panic if `step(usize)` is not monotonically increasing, or if the test
389/// binary is invalid.
390fn triage_test(
391 t_bin: &str,
392 t_name: &str,
393 thresh: Duration,
394 mut step: impl FnMut(NonZero<usize>) -> Option<NonZero<usize>>,
395) -> Option<NonZero<usize>> {
396 let mut iter_count = DEFAULT_ITER_COUNT;
397 // It's possible that the first loop of a test might be an outlier (e.g. it's
398 // doing some caching), in which case we want to skip it.
399 let duration_once = spawn_and_iterate(t_bin, t_name, NonZero::new(1).unwrap())?;
400 loop {
401 let duration = spawn_and_iterate(t_bin, t_name, iter_count)?;
402 if duration.saturating_sub(duration_once) > thresh {
403 break Some(iter_count);
404 }
405 let new = step(iter_count)?;
406 assert!(
407 new > iter_count,
408 "FATAL: step must be monotonically increasing"
409 );
410 iter_count = new;
411 }
412}
413
414/// Profiles a given test with hyperfine, returning the mean and standard deviation
415/// for its runtime. If the test errors, returns `None` instead.
416fn hyp_profile(t_bin: &str, t_name: &str, iterations: NonZero<usize>) -> Option<Timings> {
417 let mut perf_cmd = Command::new("hyperfine");
418 // Warm up the cache and print markdown output to stdout, which we parse.
419 perf_cmd.args([
420 "--style",
421 "none",
422 "--warmup",
423 "1",
424 "--export-markdown",
425 "-",
426 // Parse json instead...
427 "--time-unit",
428 "millisecond",
429 &format!("{t_bin} --exact {t_name}"),
430 ]);
431 perf_cmd.env(consts::ITER_ENV_VAR, format!("{iterations}"));
432 let p_out = perf_cmd.output().unwrap();
433 if !p_out.status.success() {
434 return None;
435 }
436
437 let cmd_output = String::from_utf8_lossy(&p_out.stdout);
438 // Can't use .last() since we have a trailing newline. Sigh.
439 let results_line = cmd_output.lines().nth(3).unwrap();
440 // Grab the values out of the pretty-print.
441 // TODO: Parse json instead.
442 let mut res_iter = results_line.split_whitespace();
443 // Durations are given in milliseconds, so account for that.
444 let mean = Duration::from_secs_f64(res_iter.nth(5).unwrap().parse::<f64>().unwrap() / 1000.);
445 let stddev = Duration::from_secs_f64(res_iter.nth(1).unwrap().parse::<f64>().unwrap() / 1000.);
446
447 Some(Timings { mean, stddev })
448}
449
450fn main() {
451 let args = std::env::args().collect::<Vec<_>>();
452 // We get passed the test we need to run as the 1st argument after our own name.
453 let t_bin = args
454 .get(1)
455 .expect("FATAL: No test binary or command; this shouldn't be manually invoked!");
456
457 // We're being asked to compare two results, not run the profiler.
458 if t_bin == "compare" {
459 compare_profiles(&args[2..]);
460 return;
461 }
462
463 // Minimum test importance we care about this run.
464 let mut thresh = Importance::Iffy;
465 // Where to print the output of this run.
466 let mut out_kind = OutputKind::Markdown;
467
468 for arg in args.iter().skip(2) {
469 match arg.as_str() {
470 "--critical" => thresh = Importance::Critical,
471 "--important" => thresh = Importance::Important,
472 "--average" => thresh = Importance::Average,
473 "--iffy" => thresh = Importance::Iffy,
474 "--fluff" => thresh = Importance::Fluff,
475 "--quiet" => QUIET.store(true, Ordering::Relaxed),
476 s if s.starts_with("--json") => {
477 out_kind = OutputKind::Json(Path::new(
478 s.strip_prefix("--json=")
479 .expect("FATAL: Invalid json parameter; pass --json=ident"),
480 ));
481 }
482 _ => (),
483 }
484 }
485 if !QUIET.load(Ordering::Relaxed) {
486 eprintln!("Starting perf check");
487 }
488
489 let mut output = Output::default();
490
491 // Spawn and profile an instance of each perf-sensitive test, via hyperfine.
492 // Each test is a pair of (test, metadata-returning-fn), so grab both. We also
493 // know the list is sorted.
494 let i = get_tests(t_bin);
495 let len = i.len();
496 for (idx, (ref t_name, ref t_mdata)) in i.enumerate() {
497 if !QUIET.load(Ordering::Relaxed) {
498 eprint!("\rProfiling test {}/{}", idx + 1, len);
499 }
500 // Pretty-printable stripped name for the test.
501 let t_name_pretty = t_name.replace(consts::SUF_NORMAL, "");
502
503 // Get the metadata this test reports for us.
504 let t_mdata = match parse_mdata(t_bin, t_mdata) {
505 Ok(mdata) => mdata,
506 Err(err) => fail!(output, t_name_pretty, err),
507 };
508
509 if t_mdata.importance < thresh {
510 fail!(output, t_name_pretty, t_mdata, FailKind::Skipped);
511 }
512
513 // Time test execution to see how many iterations we need to do in order
514 // to account for random noise. This is skipped for tests with fixed
515 // iteration counts.
516 let final_iter_count = t_mdata.iterations.or_else(|| {
517 triage_test(t_bin, t_name, consts::NOISE_CUTOFF, |c| {
518 if let Some(c) = c.checked_mul(ITER_COUNT_MUL) {
519 Some(c)
520 } else {
521 // This should almost never happen, but maybe..?
522 eprintln!(
523 "WARNING: Ran nearly usize::MAX iterations of test {t_name_pretty}; skipping"
524 );
525 None
526 }
527 })
528 });
529
530 // Don't profile failing tests.
531 let Some(final_iter_count) = final_iter_count else {
532 fail!(output, t_name_pretty, t_mdata, FailKind::Triage);
533 };
534
535 // Now profile!
536 if let Some(timings) = hyp_profile(t_bin, t_name, final_iter_count) {
537 output.success(t_name_pretty, t_mdata, final_iter_count, timings);
538 } else {
539 fail!(
540 output,
541 t_name_pretty,
542 t_mdata,
543 final_iter_count,
544 FailKind::Profile
545 );
546 }
547 }
548 if !QUIET.load(Ordering::Relaxed) {
549 if output.is_empty() {
550 eprintln!("Nothing to do.");
551 } else {
552 // If stdout and stderr are on the same terminal, move us after the
553 // output from above.
554 eprintln!();
555 }
556 }
557
558 // No need making an empty json file on every empty test bin.
559 if output.is_empty() {
560 return;
561 }
562
563 out_kind.log(&output, t_bin);
564}